Method: Integrates in-context learning with entropy-guided chain-of-thought reasoning and adaptive context retrieval. The model dynamically adjusts both number of retrieved examples and reasoning paths based on token-level uncertainty: uses concise reasoning with minimal context when uncertainty is low, and expands to multi-path CoT exploration when uncertainty is high.

Result: Entropy-aware adaptive reasoning substantially improved decision quality, increasing average game outcome from -11.6% (baseline LLM) to +9.5% over 100 games (win=+1, tie=0, loss=-1), while maintaining relatively low LLM queries per game. Statistical validation confirms significant improvement, and correlation analysis shows negative association between token-level entropy and move optimality.

Conclusion: Uncertainty-guided adaptive reasoning effectively enhances LLM performance in sequential decision-making environments, demonstrating that dynamic adjustment of reasoning complexity based on uncertainty can significantly improve decision quality in game-theoretic tasks.

Abstract: We propose a novel LLM-based framework for reasoning in discrete, game-theoretic tasks, illustrated with \emph{Tic-Tac-Toe}. The method integrates in-context learning with entropy-guided chain-of-thought (CoT) reasoning and adaptive context retrieval. The model dynamically adjusts both the number of retrieved examples and reasoning paths according to token-level uncertainty: concise reasoning with minimal context is used when uncertainty is low, whereas higher uncertainty triggers expanded multi-path CoT exploration. Experimental evaluation against a sub-optimal algorithmic opponent shows that entropy-aware adaptive reasoning substantially improves decision quality, increasing the average game outcome from (-11.6%) with the baseline LLM to (+9.5%) with entropy-guided adaptive reasoning over 100 games (win = +1, tie = 0, loss = -1), while maintaining a relatively low number of LLM queries per game. Statistical validation confirms that the improvement is significant, and correlation analysis reveals a negative association between token-level entropy and move optimality. These findings demonstrate that uncertainty-guided adaptive reasoning effectively enhances LLM performance in sequential decision-making environments.

Method: Introduces a unified framework with language resource classification (4 tiers), full-stack data refinement pipeline for low-resource languages (corpus cleaning, synthetic generation, continual pretraining, supervised finetuning), and translation-mediated pathway for extreme-low-resource languages.

Result: Achieved ~12% average improvement over multilingual baselines across 12 benchmarks for Chichewa and Maori, preserved English/multilingual capabilities via weight-space merging, and improved Inuktitut translation by 4 BLEU over commercial baseline.

Conclusion: BYOL provides scalable pathways for language-aware LLM development across resource levels, enabling better performance and accessibility for underrepresented languages while maintaining multilingual capabilities.

Abstract: Large Language Models (LLMs) exhibit strong multilingual capabilities, yet remain fundamentally constrained by the severe imbalance in global language resources. While over 7,000 languages are spoken worldwide, only a small subset (fewer than 100) has sufficient digital presence to meaningfully influence modern LLM training. This disparity leads to systematic underperformance, cultural misalignment, and limited accessibility for speakers of low-resource and extreme-low-resource languages. To address this gap, we introduce Bring Your Own Language (BYOL), a unified framework for scalable, language-aware LLM development tailored to each language’s digital footprint. BYOL begins with a language resource classification that maps languages into four tiers (Extreme-Low, Low, Mid, High) using curated web-scale corpora, and uses this classification to select the appropriate integration pathway. For low-resource languages, we propose a full-stack data refinement and expansion pipeline that combines corpus cleaning, synthetic text generation, continual pretraining, and supervised finetuning. Applied to Chichewa and Maori, this pipeline yields language-specific LLMs that achieve approximately 12 percent average improvement over strong multilingual baselines across 12 benchmarks, while preserving English and multilingual capabilities via weight-space model merging. For extreme-low-resource languages, we introduce a translation-mediated inclusion pathway, and show on Inuktitut that a tailored machine translation system improves over a commercial baseline by 4 BLEU, enabling high-accuracy LLM access when direct language modeling is infeasible. Finally, we release human-translated versions of the Global MMLU-Lite benchmark in Chichewa, Maori, and Inuktitut, and make our codebase and models publicly available at https://github.com/microsoft/byol .

Method: 1) Compare coupled, semi-decoupled, and fully decoupled speech tokenizers in a fair SLM framework. 2) Introduce multi-token prediction (MTP) to handle speech-text information density mismatch. 3) Propose speaker-aware generation paradigm and create RoleTriviaQA benchmark for role-playing knowledge QA with diverse speaker identities.

Result: Decoupled tokenization significantly improves alignment and synthesis quality. MTP enables up to 12× faster decoding and reduces word error rate from 6.07 to 3.01. Speaker-aware methods enhance both knowledge understanding and speaker consistency.

Conclusion: The systematic investigation of speech tokenizer designs reveals decoupled approaches work best. Multi-token prediction effectively addresses speech-text density mismatch, and speaker-aware generation improves model performance on both knowledge and speaker consistency tasks.

Abstract: Speech-language models (SLMs) offer a promising path toward unifying speech and text understanding and generation. However, challenges remain in achieving effective cross-modal alignment and high-quality speech generation. In this work, we systematically investigate the role of speech tokenizer designs in LLM-centric SLMs, augmented by speech heads and speaker modeling. We compare coupled, semi-decoupled, and fully decoupled speech tokenizers under a fair SLM framework and find that decoupled tokenization significantly improves alignment and synthesis quality. To address the information density mismatch between speech and text, we introduce multi-token prediction (MTP) into SLMs, enabling each hidden state to decode multiple speech tokens. This leads to up to 12$\times$ faster decoding and a substantial drop in word error rate (from 6.07 to 3.01). Furthermore, we propose a speaker-aware generation paradigm and introduce RoleTriviaQA, a large-scale role-playing knowledge QA benchmark with diverse speaker identities. Experiments demonstrate that our methods enhance both knowledge understanding and speaker consistency.

Method: 1) Surveyed 127 conversational agent papers from ACL Anthology to identify 12 frequently used style features. 2) Used controlled, synthetic dialogues across task-oriented and open domain settings. 3) Quantified how prompting for one style feature causally affects others via pairwise LLM as a Judge evaluation framework. 4) Created CASSE dataset capturing these interactions. 5) Evaluated prompt-based and activation steering-based mitigation strategies.

Result: Revealed consistent and structured side effects (e.g., prompting for conciseness significantly reduces perceived expertise). Found that style features are deeply entangled rather than orthogonal. Mitigation strategies can partially restore suppressed traits but often degrade the primary intended style.

Conclusion: The findings challenge the assumption of faithful style control in LLMs and highlight the need for multi-objective and more principled approaches to safe, targeted stylistic steering in conversational agents.

Abstract: Style features such as friendly, helpful, or concise are widely used in prompts to steer the behavior of Large Language Model (LLM) conversational agents, yet their unintended side effects remain poorly understood. In this work, we present the first systematic study of cross-feature stylistic side effects. We conduct a comprehensive survey of 127 conversational agent papers from ACL Anthology and identify 12 frequently used style features. Using controlled, synthetic dialogues across task-oriented and open domain settings, we quantify how prompting for one style feature causally affects others via a pairwise LLM as a Judge evaluation framework. Our results reveal consistent and structured side effects, such as prompting for conciseness significantly reduces perceived expertise. They demonstrate that style features are deeply entangled rather than orthogonal. To support future research, we introduce CASSE (Conversational Agent Stylistic Side Effects), a dataset capturing these complex interactions. We further evaluate prompt based and activation steering based mitigation strategies and find that while they can partially restore suppressed traits, they often degrade the primary intended style. These findings challenge the assumption of faithful style control in LLMs and highlight the need for multi-objective and more principled approaches to safe, targeted stylistic steering in conversational agents.

Method: Introduces POWSM (Phonetic Open Whisper-style Speech Model), a unified framework capable of jointly performing multiple phone-related tasks, enabling seamless conversion between audio, text, and phones.

Result: POWSM outperforms or matches specialized phone recognition models (Wav2Vec2Phoneme and ZIPA) of similar size while jointly supporting G2P, P2G, and ASR tasks.

Conclusion: POWSM represents a significant step toward universal speech processing, opening possibilities for low-resource applications, with training data, code, and models released to foster open science.

Abstract: Recent advances in spoken language processing have led to substantial progress in phonetic tasks such as automatic speech recognition (ASR), phone recognition (PR), grapheme-to-phoneme conversion (G2P), and phoneme-to-grapheme conversion (P2G). Despite their conceptual similarity, these tasks have largely been studied in isolation, each relying on task-specific architectures and datasets. In this paper, we introduce POWSM (Phonetic Open Whisper-style Speech Model), the first unified framework capable of jointly performing multiple phone-related tasks. POWSM enables seamless conversion between audio, text (graphemes), and phones, opening up new possibilities for universal and low-resource speech processing. Our model outperforms or matches specialized PR models of similar size (Wav2Vec2Phoneme and ZIPA) while jointly supporting G2P, P2G, and ASR. Our training data, code and models are released to foster open science.

Method: Used quantitative analysis and mechanistic interpretability methods on reasoning traces from models like DeepSeek-R1 and QwQ-32B, examining perspective diversity, feature activation, and conversational behaviors. Conducted controlled reinforcement learning experiments where base models were rewarded for reasoning accuracy.

Result: Reasoning models show much greater perspective diversity than instruction-tuned models, activating broader conflict between heterogeneous personality- and expertise-related features. They exhibit conversational behaviors including question-answering, perspective shifts, and reconciliation of conflicting views. Reinforcement learning experiments show base models increase conversational behaviors when rewarded for accuracy, and fine-tuning with conversational scaffolding accelerates reasoning improvement.

Conclusion: Enhanced reasoning emerges from multi-agent-like interactions (“society of thought”) that enable diversification and debate among internal cognitive perspectives. This social organization of thought enables effective exploration of solution spaces, establishing a computational parallel to collective intelligence in human groups, suggesting new opportunities for agent organization to harness collective wisdom.

Abstract: Large language models have achieved remarkable capabilities across domains, yet mechanisms underlying sophisticated reasoning remain elusive. Recent reasoning models outperform comparable instruction-tuned models on complex cognitive tasks, attributed to extended computation through longer chains of thought. Here we show that enhanced reasoning emerges not from extended computation alone, but from simulating multi-agent-like interactions – a society of thought – which enables diversification and debate among internal cognitive perspectives characterized by distinct personality traits and domain expertise. Through quantitative analysis and mechanistic interpretability methods applied to reasoning traces, we find that reasoning models like DeepSeek-R1 and QwQ-32B exhibit much greater perspective diversity than instruction-tuned models, activating broader conflict between heterogeneous personality- and expertise-related features during reasoning. This multi-agent structure manifests in conversational behaviors, including question-answering, perspective shifts, and the reconciliation of conflicting views, and in socio-emotional roles that characterize sharp back-and-forth conversations, together accounting for the accuracy advantage in reasoning tasks. Controlled reinforcement learning experiments reveal that base models increase conversational behaviors when rewarded solely for reasoning accuracy, and fine-tuning models with conversational scaffolding accelerates reasoning improvement over base models. These findings indicate that the social organization of thought enables effective exploration of solution spaces. We suggest that reasoning models establish a computational parallel to collective intelligence in human groups, where diversity enables superior problem-solving when systematically structured, which suggests new opportunities for agent organization to harness the wisdom of crowds.

Method: EncodeRec learns compact, informative embeddings directly from item descriptions while keeping the language model parameters frozen during recommender system training, making it computationally efficient without sacrificing semantic fidelity.

Result: Experiments across core recommendation benchmarks show substantial gains over PLM-based and embedding model baselines, demonstrating effectiveness both as a backbone for sequential recommendation models and for semantic ID tokenization.

Conclusion: Embedding adaptation is pivotal for bridging the gap between general-purpose language models and practical recommender systems, with EncodeRec showing promising results in aligning textual representations with recommendation objectives.

Abstract: Recent recommender systems increasingly leverage embeddings from large pre-trained language models (PLMs). However, such embeddings exhibit two key limitations: (1) PLMs are not explicitly optimized to produce structured and discriminative embedding spaces, and (2) their representations remain overly generic, often failing to capture the domain-specific semantics crucial for recommendation tasks. We present EncodeRec, an approach designed to align textual representations with recommendation objectives while learning compact, informative embeddings directly from item descriptions. EncodeRec keeps the language model parameters frozen during recommender system training, making it computationally efficient without sacrificing semantic fidelity. Experiments across core recommendation benchmarks demonstrate its effectiveness both as a backbone for sequential recommendation models and for semantic ID tokenization, showing substantial gains over PLM-based and embedding model baselines. These results underscore the pivotal role of embedding adaptation in bridging the gap between general-purpose language models and practical recommender systems.

Method: Introduces DialDefer framework and Dialogic Deference Score (DDS) to detect and measure framing-induced judgment shifts. Tests across nine domains, 3k+ instances, and four models, comparing statement verification (“Is this correct?”) vs. speaker attribution (“Is this speaker correct?”). Also examines naturalistic Reddit conversations and conducts ablations.

Result: Conversational framing induces large judgment shifts (|DDS| up to 87 percentage points) while accuracy remains stable (<2pp). Effects amplify 2-4x on naturalistic conversations. Models shift toward agreement (deference) or disagreement (skepticism) depending on domain, with human-vs-LLM attribution driving largest shifts (17.7pp swing). Mitigation attempts reduce deference but can over-correct into skepticism.

Conclusion: LLMs exhibit systematic framing biases in dialogue evaluation that aggregate accuracy metrics obscure. This “dialogic deference” reveals models treat disagreement with humans as more costly than with AI, framing the problem as a calibration issue beyond accuracy optimization.

Abstract: LLMs are increasingly used as third-party judges, yet their reliability when evaluating speakers in dialogue remains poorly understood. We show that LLMs judge identical claims differently depending on framing: the same content elicits different verdicts when presented as a statement to verify (“Is this statement correct?”) versus attributed to a speaker (“Is this speaker correct?”). We call this dialogic deference and introduce DialDefer, a framework for detecting and mitigating these framing-induced judgment shifts. Our Dialogic Deference Score (DDS) captures directional shifts that aggregate accuracy obscures. Across nine domains, 3k+ instances, and four models, conversational framing induces large shifts (|DDS| up to 87pp, p < .0001) while accuracy remains stable (<2pp), with effects amplifying 2-4x on naturalistic Reddit conversations. Models can shift toward agreement (deference) or disagreement (skepticism) depending on domain – the same model ranges from DDS = -53 on graduate-level science to +58 on social judgment. Ablations reveal that human-vs-LLM attribution drives the largest shifts (17.7pp swing), suggesting models treat disagreement with humans as more costly than with AI. Mitigation attempts reduce deference but can over-correct into skepticism, framing this as a calibration problem beyond accuracy optimization.

Method: Proposes a novel method for automatically constructing unweighted FSTs by following the hidden state geometry learned by a recurrent neural network. The approach leverages neural network representations to inform transducer structure.

Result: The constructed FSTs are highly accurate and robust across multiple real-world datasets (morphological inflection, grapheme-to-phoneme prediction, historical normalization), substantially outperforming classical transducer learning algorithms by up to 87% accuracy on held-out test sets.

Conclusion: The method successfully bridges neural network learning with symbolic transducer construction, providing an effective automated approach for creating accurate FSTs that significantly outperforms traditional transducer learning methods.

Abstract: Finite-State Transducers (FSTs) are effective models for string-to-string rewriting tasks, often providing the efficiency necessary for high-performance applications, but constructing transducers by hand is difficult. In this work, we propose a novel method for automatically constructing unweighted FSTs following the hidden state geometry learned by a recurrent neural network. We evaluate our methods on real-world datasets for morphological inflection, grapheme-to-phoneme prediction, and historical normalization, showing that the constructed FSTs are highly accurate and robust for many datasets, substantially outperforming classical transducer learning algorithms by up to 87% accuracy on held-out test sets.

Method: Conducted first study on neural models for joint prediction of interlinear glosses and morphological segmentation; extended GlossLM’s training corpus; pretrained PolyGloss family of seq2seq multilingual models; experimented with optimal training approaches balancing segmentation and glossing accuracy and alignment; demonstrated quick adaptation to new datasets via low-rank adaptation.

Result: PolyGloss outperforms GlossLM on glossing and beats various open-source LLMs on segmentation, glossing, and alignment tasks.

Conclusion: Joint prediction of glosses and morphological segmentation addresses critical barriers to usefulness in real-world language documentation, making predictions more interpretable and trustworthy to human annotators.

Abstract: Automated interlinear gloss prediction with neural networks is a promising approach to accelerate language documentation efforts. However, while state-of-the-art models like GlossLM achieve high scores on glossing benchmarks, user studies with linguists have found critical barriers to the usefulness of such models in real-world scenarios. In particular, existing models typically generate morpheme-level glosses but assign them to whole words without predicting the actual morpheme boundaries, making the predictions less interpretable and thus untrustworthy to human annotators. We conduct the first study on neural models that jointly predict interlinear glosses and the corresponding morphological segmentation from raw text. We run experiments to determine the optimal way to train models that balance segmentation and glossing accuracy, as well as the alignment between the two tasks. We extend the training corpus of GlossLM and pretrain PolyGloss, a family of seq2seq multilingual models for joint segmentation and glossing that outperforms GlossLM on glossing and beats various open-source LLMs on segmentation, glossing, and alignment. In addition, we demonstrate that PolyGloss can be quickly adapted to a new dataset via low-rank adaptation.

Method: The authors analyze four key factors for model selection: (1) model openness, (2) model footprint (size), (3) training data characteristics, and (4) model architectures and fine-tuning approaches. They propose a framework prioritizing replicability and recommend starting with smaller, open models while constructing targeted benchmarks.

Result: The analysis reveals that replicability should be prioritized over ex-ante benchmark validation for social science applications. The authors conclude that social scientists cannot avoid ex-post validation of computational measures and should focus on building reproducible computational pipelines.

Conclusion: Social scientists should prioritize replicability when selecting language models, starting with smaller open models and creating delimited benchmarks to validate entire computational pipelines, rather than relying solely on pre-existing benchmark scores.

Abstract: Currently, there are thousands of large pretrained language models (LLMs) available to social scientists. How do we select among them? Using validity, reliability, reproducibility, and replicability as guides, we explore the significance of: (1) model openness, (2) model footprint, (3) training data, and (4) model architectures and fine-tuning. While ex-ante tests of validity (i.e., benchmarks) are often privileged in these discussions, we argue that social scientists cannot altogether avoid validating computational measures (ex-post). Replicability, in particular, is a more pressing guide for selecting language models. Being able to reliably replicate a particular finding that entails the use of a language model necessitates reliably reproducing a task. To this end, we propose starting with smaller, open models, and constructing delimited benchmarks to demonstrate the validity of the entire computational pipeline.

Method: Created Ch-PatientSim dataset from realistic clinical scenarios using five-dimensional persona structure. Augmented dataset with few-shot generation and manual verification. Proposed Multi-Stage Patient Role-Playing (MSPRP) framework that decomposes interactions into three stages for personalization and realism.

Result: Most existing LLMs produce overly formal responses lacking individual personality. The proposed MSPRP framework significantly improves model performance across multiple dimensions of patient simulation compared to existing approaches.

Conclusion: The Ch-PatientSim dataset and MSPRP framework address key limitations in clinical LLM evaluation by providing realistic patient simulation benchmarks and improving model ability to emulate authentic patient behavior with personalization.

Abstract: The simulation of realistic clinical interactions plays a pivotal role in advancing clinical Large Language Models (LLMs) and supporting medical diagnostic education. Existing approaches and benchmarks rely on generic or LLM-generated dialogue data, which limits the authenticity and diversity of doctor-patient interactions. In this work, we propose the first Chinese patient simulation dataset (Ch-PatientSim), constructed from realistic clinical interaction scenarios to comprehensively evaluate the performance of models in emulating patient behavior. Patients are simulated based on a five-dimensional persona structure. To address issues of the persona class imbalance, a portion of the dataset is augmented using few-shot generation, followed by manual verification. We evaluate various state-of-the-art LLMs and find that most produce overly formal responses that lack individual personality. To address this limitation, we propose a training-free Multi-Stage Patient Role-Playing (MSPRP) framework, which decomposes interactions into three stages to ensure both personalization and realism in model responses. Experimental results demonstrate that our approach significantly improves model performance across multiple dimensions of patient simulation.

Method: Proposes a training-free inference-time logit intervention approach that uses statistical token score tables derived from z-normalized log-odds of labeled corpora to shift the decoding distribution during generation.

Result: Empirical evaluations across three diverse datasets (writing complexity, formality, and toxicity) demonstrate effective steering of output characteristics. The method achieves large, consistent, and multi-task control gains: up to +47%p accuracy and 50x f1 improvement.

Conclusion: Statistically grounded logit steering provides broad applicability and task-agnostic control for LLM steering, addressing limitations of existing methods while maintaining training-free operation.

Abstract: Steering LLMs is essential for specialized applications such as style-sensitive text rewriting, user-adaptive communication, and toxicity mitigation. Current steering methods, such as prompting-based and activation-based approaches, are widely used to guide model behavior. However, activation-based techniques require deep access to internal layers, while prompting-based steering often fails to provide consistent or fine-grained control. In order to address these limitations, we propose a training-free inference-time logit intervention for controllable generation. Our approach utilizes a statistical token score table derived from z-normalized log-odds of labeled corpora to shift the decoding distribution. Empirical evaluations across three diverse datasets focusing on writing complexity, formality, and toxicity demonstrate that our method effectively steers output characteristics, confirming its broad applicability and task-agnostic nature. Our results show that statistically grounded logit steering can achieve large, consistent, and multi-task control gains: up to +47%p accuracy and 50x f1 improvement.

Method: Inspired by Zone of Proximal Development theory, ZPD Detector uses difficulty calibration, IRT-based model capability estimation, and capability-difficulty matching scores to dynamically select the most informative samples at each learning stage.

Result: The framework improves data utilization efficiency and provides new insights into training strategy design through dynamic model-data alignment.

Conclusion: ZPD Detector offers a novel bidirectional perspective on model-data relationships that addresses limitations of static selection methods, with code and data to be released for reproducibility.

Abstract: As the cost of training large language models continues to increase and high-quality training data become increasingly scarce, selecting high-value samples or synthesizing effective training data under limited data budgets has emerged as a critical research problem. Most existing data selection methods rely on static criteria, such as difficulty, uncertainty, or heuristics, and fail to model the evolving relationship between the model and the data. Inspired by the educational theory of the Zone of Proximal Development (ZPD), we propose ZPD Detector, a data selection framework that adopts a bidirectional perspective between models and data by explicitly modeling the alignment between sample difficulty and the model’s current capability. ZPD Detector integrates difficulty calibration, model capability estimation based on Item Response Theory (IRT), and a capability-difficulty matching score to dynamically identify the most informative samples at each learning stage, improving data utilization efficiency; moreover, this dynamic matching strategy provides new insights into training strategy design. All code and data will be released after our work be accepted to support reproducible researc

Method: Proposes Factuality-Preserving Personalized Steering (FPPS), a lightweight inference-time approach that mitigates personalization-induced factual distortions while preserving personalized behavior. Also introduces PFQABench, the first benchmark for jointly evaluating factual and personalized question answering under personalization.

Result: Experiments across multiple LLM backbones and personalization methods show that FPPS substantially improves factual accuracy while maintaining personalized performance.

Conclusion: Personalization can cause factual distortions in LLMs, but the proposed FPPS method effectively addresses this issue by preserving factual accuracy while maintaining personalized behavior, with PFQABench providing a comprehensive evaluation framework.

Abstract: Personalized large language models (LLMs) adapt model behavior to individual users to enhance user satisfaction, yet personalization can inadvertently distort factual reasoning. We show that when personalized LLMs face factual queries, there exists a phenomenon where the model generates answers aligned with a user’s prior history rather than the objective truth, resulting in personalization-induced hallucinations that degrade factual reliability and may propagate incorrect beliefs, due to representational entanglement between personalization and factual representations. To address this issue, we propose Factuality-Preserving Personalized Steering (FPPS), a lightweight inference-time approach that mitigates personalization-induced factual distortions while preserving personalized behavior. We further introduce PFQABench, the first benchmark designed to jointly evaluate factual and personalized question answering under personalization. Experiments across multiple LLM backbones and personalization methods show that FPPS substantially improves factual accuracy while maintaining personalized performance.

Method: Extended SiMT action space with four adaptive actions, adapted in an LLM framework with action-aware prompting for training. Developed a latency-aware TTS pipeline to evaluate both quality and word-level monotonicity by mapping textual outputs to speech with realistic timing.

Result: Experiments on ACL60/60 English-Chinese, English-German, and English-Japanese benchmarks show consistent improvements in semantic metrics and lower delay compared to reference translations and salami-based baselines. Combining Drop and Sentence_Cut actions achieves the best balance between fluency and latency.

Conclusion: Enriching the action space of LLM-based SiMT provides a promising direction for bridging the gap between human and machine interpretation, demonstrating that adaptive actions enable better real-time translation performance.

Abstract: Simultaneous Machine Translation (SiMT) requires high-quality translations under strict real-time constraints, which traditional policies with only READ/WRITE actions cannot fully address. We extend the action space of SiMT with four adaptive actions: Sentence_Cut, Drop, Partial_Summarization and Pronominalization, which enable real-time restructuring, omission, and simplification while preserving semantic fidelity. We adapt these actions in a large language model (LLM) framework and construct training references through action-aware prompting. To evaluate both quality and word-level monotonicity, we further develop a latency-aware TTS pipeline that maps textual outputs to speech with realistic timing. Experiments on the ACL60/60 English-Chinese, English-German and English-Japanese benchmarks show that our framework consistently improves semantic metrics and achieves lower delay compared to reference translations and salami-based baselines. Notably, combining Drop and Sentence_Cut leads to consistent improvements in the balance between fluency and latency. These results demonstrate that enriching the action space of LLM-based SiMT provides a promising direction for bridging the gap between human and machine interpretation.

Method: Propose NAACL Rules (Noise-AwAre Confidence CaLibration Rules) as a principled foundation, then design NAACL framework that synthesizes supervision from ~2K HotpotQA examples guided by these rules, performing supervised fine-tuning (SFT) to equip models with intrinsic noise awareness without stronger teacher models.

Result: NAACL yields substantial gains, improving ECE (Expected Calibration Error) scores by 10.9% in-domain and 8.0% out-of-domain, effectively bridging the gap between retrieval noise and verbal calibration.

Conclusion: NAACL paves the way for both accurate and epistemically reliable LLMs by addressing overconfidence under noisy contexts in RAG settings through noise-aware calibration.

Abstract: Accurately assessing model confidence is essential for deploying large language models (LLMs) in mission-critical factual domains. While retrieval-augmented generation (RAG) is widely adopted to improve grounding, confidence calibration in RAG settings remains poorly understood. We conduct a systematic study across four benchmarks, revealing that LLMs exhibit poor calibration performance due to noisy retrieved contexts. Specifically, contradictory or irrelevant evidence tends to inflate the model’s false certainty, leading to severe overconfidence. To address this, we propose NAACL Rules (Noise-AwAre Confidence CaLibration Rules) to provide a principled foundation for resolving overconfidence under noise. We further design NAACL, a noise-aware calibration framework that synthesizes supervision from about 2K HotpotQA examples guided by these rules. By performing supervised fine-tuning (SFT) with this data, NAACL equips models with intrinsic noise awareness without relying on stronger teacher models. Empirical results show that NAACL yields substantial gains, improving ECE scores by 10.9% in-domain and 8.0% out-of-domain. By bridging the gap between retrieval noise and verbal calibration, NAACL paves the way for both accurate and epistemically reliable LLMs.

Method: Use Sparse Autoencoders (SAEs) to identify task-specific features. First recall frequently co-activated features on translation inputs, then filter them for functional coherence using a PCA-based consistency metric. This isolates translation initiation features. Then propose a data selection strategy that prioritizes training on “mechanistically hard” samples that fail to naturally activate these features.

Result: Successfully isolated a small set of translation initiation features. Causal interventions show amplifying these features steers models toward correct translation, while ablating them causes hallucinations. The data selection strategy significantly improves data efficiency and suppresses hallucinations. These mechanisms are transferable to larger models of the same family.

Conclusion: The work decodes a core component of LLMs’ translation mechanism and provides a blueprint for using internal model mechanisms to create more robust and efficient models. The approach demonstrates how mechanistic understanding can lead to practical improvements in model training and performance.

Abstract: Large Language Models (LLMs) frequently exhibit strong translation abilities, even without task-specific fine-tuning. However, the internal mechanisms governing this innate capability remain largely opaque. To demystify this process, we leverage Sparse Autoencoders (SAEs) and introduce a novel framework for identifying task-specific features. Our method first recalls features that are frequently co-activated on translation inputs and then filters them for functional coherence using a PCA-based consistency metric. This framework successfully isolates a small set of translation initiation features. Causal interventions demonstrate that amplifying these features steers the model towards correct translation, while ablating them induces hallucinations and off-task outputs, confirming they represent a core component of the model’s innate translation competency. Moving from analysis to application, we leverage this mechanistic insight to propose a new data selection strategy for efficient fine-tuning. Specifically, we prioritize training on mechanistically hard samples-those that fail to naturally activate the translation initiation features. Experiments show this approach significantly improves data efficiency and suppresses hallucinations. Furthermore, we find these mechanisms are transferable to larger models of the same family. Our work not only decodes a core component of the translation mechanism in LLMs but also provides a blueprint for using internal model mechanism to create more robust and efficient models. The codes are available at https://github.com/flamewei123/AAAI26-translation-Initiation-Features.

Method: Proposes RetMask, a training method that generates contrastive signals by comparing normal model outputs with outputs from an ablated variant where retrieval heads are masked. This mechanism-based approach creates training signals that specifically target the improvement of retrieval head functionality.

Result: Achieves substantial improvements: +2.28 points on HELMET at 128K for Llama-3.1, with +70% gains on generation with citation and +32% on passage re-ranking, while preserving performance on general tasks. Effectiveness depends on retrieval head organization - models with concentrated patterns respond strongly, while those with distributed patterns show limited gains.

Conclusion: The mechanistic relationship validates the function of retrieval heads and demonstrates that mechanistic insights from interpretability research can be transformed into practical performance enhancements for LLMs, particularly for long-context tasks.

Abstract: Advances in mechanistic interpretability have identified special attention heads, known as retrieval heads, that are responsible for retrieving information from the context. However, the role of these retrieval heads in improving model performance remains unexplored. This work investigates whether retrieval heads can be leveraged to enhance the long-context capabilities of LLMs. Specifically, we propose RetMask, a method that generates training signals by contrasting normal model outputs with those from an ablated variant in which the retrieval heads are masked. This mechanism-based approach achieves substantial improvements: +2.28 points on HELMET at 128K for Llama-3.1, with +70% gains on generation with citation and +32% on passage re-ranking, while preserving performance on general tasks. Experiments across three model families reveal that the effectiveness depends on retrieval head organization: models with concentrated patterns of retrieval heads respond strongly, while those with distributed patterns show limited gains. This mechanistic relationship validates the function of retrieval heads and demonstrates that mechanistic insights can be transformed into performance enhancements.

Method: 1) Anytime reasoning framework with Anytime Index metric quantifying solution quality improvement per reasoning token; 2) Inference-time self-improvement using LLM-synthesized preference data where models learn from their own reasoning comparisons to produce better intermediate solutions.

Result: Experiments on NaturalPlan (Trip), AIME, and GPQA datasets show consistent gains across Grok-3, GPT-oss, GPT-4.1/4o, and LLaMA models, improving both reasoning quality and efficiency under budget constraints.

Conclusion: The proposed anytime reasoning framework and self-improvement method effectively enhance LLMs’ practical utility in budget-constrained scenarios, making them more suitable for real-world applications where computation resources are limited.

Abstract: We study the reasoning behavior of large language models (LLMs) under limited computation budgets. In such settings, producing useful partial solutions quickly is often more practical than exhaustive reasoning, which incurs high inference costs. Many real-world tasks, such as trip planning, require models to deliver the best possible output within a fixed reasoning budget. We introduce an anytime reasoning framework and the Anytime Index, a metric that quantifies how effectively solution quality improves as reasoning tokens increase. To further enhance efficiency, we propose an inference-time self-improvement method using LLM-synthesized preference data, where models learn from their own reasoning comparisons to produce better intermediate solutions. Experiments on NaturalPlan (Trip), AIME, and GPQA datasets show consistent gains across Grok-3, GPT-oss, GPT-4.1/4o, and LLaMA models, improving both reasoning quality and efficiency under budget constraints.

Method: REVIVE performs spectral analysis showing that general abilities are associated with dominant singular directions of pretrained weight matrices. The framework preserves these sensitive directions by representing parameter updates in the spectral basis of original weights and filtering components that would interfere with the protected dominant singular subspace.

Result: REVIVE consistently improves editing efficacy while substantially preserving general abilities under long-horizon sequential editing, including extreme settings with up to 20,000 edits across multiple models and benchmarks.

Conclusion: The dominant singular directions of pretrained weights are crucial for maintaining LLM general abilities during sequential editing. REVIVE’s spectral subspace preservation approach effectively stabilizes editing while preventing catastrophic collapse, advancing understanding and practical implementation of sequential knowledge editing.

Abstract: Sequential knowledge editing in large language models often causes catastrophic collapse of the model’s general abilities, especially for parameter-modifying methods. Existing approaches mitigate this issue through heuristic constraints on parameter updates, yet the mechanisms underlying such degradation remain insufficiently understood. In this work, we present a spectral analysis of sequential knowledge editing and show that a model’s general abilities are closely associated with dominant singular directions of pretrained weight matrices. These directions are highly sensitive to perturbations and are progressively disrupted by repeated edits, closely tracking the collapse in both editing efficacy and general performance. Building on this insight, we propose REVIVE, a plug-and-play framework that stabilizes sequential editing by explicitly preserving the dominant singular subspace. REVIVE represents parameter updates in the spectral basis of the original weights and filters components that would interfere with the protected region. Extensive experiments across multiple models and benchmarks show that REVIVE consistently improves editing efficacy while substantially preserving general abilities under long-horizon sequential editing, including extreme settings with up to 20,000 edits.

Method: CoG is inspired by Dual-Process Theory, mimicking intuition-deliberation interplay. It has two modules: 1) Relational Blueprint Guidance (fast, intuitive) uses relational blueprints as interpretable soft structural constraints to stabilize search direction against noise. 2) Failure-Aware Refinement (prudent, analytical) intervenes upon reasoning impasses with evidence-conditioned reflection and controlled backtracking to overcome stagnation.

Result: Experimental results on three benchmarks show CoG significantly outperforms state-of-the-art approaches in both accuracy and efficiency.

Conclusion: CoG addresses cognitive rigidity in KG-augmented LLMs through a dual-process approach, combining intuitive stabilization with analytical refinement to improve reasoning reliability without training.

Abstract: Large Language Models (LLMs) have demonstrated remarkable reasoning capabilities but often grapple with reliability challenges like hallucinations. While Knowledge Graphs (KGs) offer explicit grounding, existing paradigms of KG-augmented LLMs typically exhibit cognitive rigidity–applying homogeneous search strategies that render them vulnerable to instability under neighborhood noise and structural misalignment leading to reasoning stagnation. To address these challenges, we propose CoG, a training-free framework inspired by Dual-Process Theory that mimics the interplay between intuition and deliberation. First, functioning as the fast, intuitive process, the Relational Blueprint Guidance module leverages relational blueprints as interpretable soft structural constraints to rapidly stabilize the search direction against noise. Second, functioning as the prudent, analytical process, the Failure-Aware Refinement module intervenes upon encountering reasoning impasses. It triggers evidence-conditioned reflection and executes controlled backtracking to overcome reasoning stagnation. Experimental results on three benchmarks demonstrate that CoG significantly outperforms state-of-the-art approaches in both accuracy and efficiency.

Method: The method uses a convolutional neural network with parallel convolution branches, depthwise-separable operations, and hierarchical classification to process names efficiently. The model is evaluated on a multilingual dataset covering 104 languages and four entity types.

Result: Onomas-CNN X achieves 92.1% accuracy while processing 2,813 names per second on a single CPU core - 46 times faster than fine-tuned XLM-RoBERTa with comparable accuracy, and reduces energy consumption by a factor of 46 compared to transformer baselines.

Conclusion: Specialized CNN architectures remain competitive with large pre-trained transformer models for focused NLP tasks when sufficient training data exists, offering significant advantages in speed and energy efficiency on CPU hardware.

Abstract: We present a convolutional neural network approach for classifying proper names by language and entity type. Our model, Onomas-CNN X, combines parallel convolution branches with depthwise-separable operations and hierarchical classification to process names efficiently on CPU hardware. We evaluate the architecture on a large multilingual dataset covering 104 languages and four entity types (person, organization, location, other). Onomas-CNN X achieves 92.1% accuracy while processing 2,813 names per second on a single CPU core - 46 times faster than fine-tuned XLM-RoBERTa with comparable accuracy. The model reduces energy consumption by a factor of 46 compared to transformer baselines. Our experiments demonstrate that specialized CNN architectures remain competitive with large pre-trained models for focused NLP tasks when sufficient training data exists.

Method: A document-layer watermarking system that embeds schema-aware, item-level watermarks into assessment PDFs while keeping human-visible appearance unchanged. Watermarks prevent MLLMs from answering shielded exams and encode stable, item-level signatures recoverable from responses.

Result: Across 30 exams spanning STEM, humanities, and medical reasoning, Integrity Shield achieves 91-94% exam-level blocking and 89-93% signature retrieval across four commercial MLLMs.

Conclusion: Integrity Shield provides an effective solution for maintaining academic integrity by preventing LLM-based cheating while enabling reliable detection of AI-generated responses through document-layer watermarking.

Abstract: Large Language Models (LLMs) can now solve entire exams directly from uploaded PDF assessments, raising urgent concerns about academic integrity and the reliability of grades and credentials. Existing watermarking techniques either operate at the token level or assume control over the model’s decoding process, making them ineffective when students query proprietary black-box systems with instructor-provided documents. We present Integrity Shield, a document-layer watermarking system that embeds schema-aware, item-level watermarks into assessment PDFs while keeping their human-visible appearance unchanged. These watermarks consistently prevent MLLMs from answering shielded exam PDFs and encode stable, item-level signatures that can be reliably recovered from model or student responses. Across 30 exams spanning STEM, humanities, and medical reasoning, Integrity Shield achieves exceptionally high prevention (91-94% exam-level blocking) and strong detection reliability (89-93% signature retrieval) across four commercial MLLMs. Our demo showcases an interactive interface where instructors upload an exam, preview watermark behavior, and inspect pre/post AI performance & authorship evidence.

Method: Established continuous crawling infrastructure for iterative national top-level domain crawling across South Slavic and related webs, with automatic genre and topic annotation.

Result: Created CLASSLA-web 2.0 with 17.0 billion words in 38.1 million texts across 7 languages (Bosnian, Bulgarian, Croatian, Macedonian, Montenegrin, Serbian, Slovenian), with only 20% overlap with previous version.

Conclusion: Continuous crawling yields substantial new content but reveals growing content quality issues as machine-generated sites now contribute significantly to web corpora.

Abstract: Crawling national top-level domains has proven to be highly effective for collecting texts in less-resourced languages. This approach has been recently used for South Slavic languages and resulted in the largest general corpora for this language group: the CLASSLA-web 1.0 corpora. Building on this success, we established a continuous crawling infrastructure for iterative national top-level domain crawling across South Slavic and related webs. We present the first outcome of this crawling infrastructure - the CLASSLA-web 2.0 corpus collection, with substantially larger web corpora containing 17.0 billion words in 38.1 million texts in seven languages: Bosnian, Bulgarian, Croatian, Macedonian, Montenegrin, Serbian, and Slovenian. In addition to genre categories, the new version is also automatically annotated with topic labels. Comparing CLASSLA-web 2.0 with its predecessor reveals that only one-fifth of the texts overlap, showing that re-crawling after just two years yields largely new content. However, while the new web crawls bring growing gains, we also notice growing pains - a manual inspection of top domains reveals a visible degradation of web content, as machine-generated sites now contribute a significant portion of texts.

Method: DOREMI is an iterative framework that actively selects the most informative examples for minimal targeted manual annotations to enhance underrepresented relations. It can be applied to any existing DocRE model and focuses on improving training efficiency and robustness without relying on large-scale noisy data or heuristic denoising.

Result: The framework effectively mitigates long-tail biases and offers a scalable solution to improve generalization on rare relations in document-level relation extraction tasks.

Conclusion: DOREMI provides an effective approach to address the long-tail problem in document-level relation extraction through targeted annotation strategies, enhancing model performance on underrepresented relations while maintaining scalability and compatibility with existing models.

Abstract: Document-Level Relation Extraction (DocRE) presents significant challenges due to its reliance on cross-sentence context and the long-tail distribution of relation types, where many relations have scarce training examples. In this work, we introduce DOcument-level Relation Extraction optiMizing the long taIl (DOREMI), an iterative framework that enhances underrepresented relations through minimal yet targeted manual annotations. Unlike previous approaches that rely on large-scale noisy data or heuristic denoising, DOREMI actively selects the most informative examples to improve training efficiency and robustness. DOREMI can be applied to any existing DocRE model and is effective at mitigating long-tail biases, offering a scalable solution to improve generalization on rare relations.

Method: Uses a TraceRL-based training curriculum that starts from an AR-initialized small-block MDM and smoothly transitions to larger blocks through progressive scaling.

Result: Achieves higher-parallelism decoding with minimal performance degradation on math reasoning benchmarks, and can converge to an alternative decoding schedule that achieves comparable performance.

Conclusion: T* provides an effective curriculum training approach for scaling block sizes in masked diffusion language models, enabling more efficient parallel decoding while preserving reasoning capabilities.

Abstract: We present T$^\star$, a simple \textsc{TraceRL}-based training curriculum for progressive block-size scaling in masked diffusion language models (MDMs). Starting from an AR-initialized small-block MDM, T$^\star$~transitions smoothly to larger blocks, enabling higher-parallelism decoding with minimal performance degradation on math reasoning benchmarks. Moreover, further analysis suggests that T$^\star$~can converge to an alternative decoding schedule $\hat{\rm S}$ that achieves comparable performance.

Method: Created MultiCaption dataset with pairs of claims referring to same image/video labeled through multiple strategies to determine contradictions. Conducted comprehensive experiments using transformer-based architectures, NLI models, and large language models to establish baselines.

Result: MultiCaption comprises 11,088 visual claims in 64 languages. Results show it’s more challenging than standard NLI tasks, requiring task-specific finetuning for strong performance. Multilingual training/testing gains highlight dataset’s potential for building effective multilingual fact-checking pipelines without machine translation.

Conclusion: MultiCaption provides a unique resource for building and evaluating misinformation-detection systems in truly multimodal and multilingual environments, addressing critical gaps in current fact-checking capabilities.

Abstract: Online disinformation poses an escalating threat to society, driven increasingly by the rapid spread of misleading content across both multimedia and multilingual platforms. While automated fact-checking methods have advanced in recent years, their effectiveness remains constrained by the scarcity of datasets that reflect these real-world complexities. To address this gap, we first present MultiCaption, a new dataset specifically designed for detecting contradictions in visual claims. Pairs of claims referring to the same image or video were labeled through multiple strategies to determine whether they contradict each other. The resulting dataset comprises 11,088 visual claims in 64 languages, offering a unique resource for building and evaluating misinformation-detection systems in truly multimodal and multilingual environments. We then provide comprehensive experiments using transformer-based architectures, natural language inference models, and large language models, establishing strong baselines for future research. The results show that MultiCaption is more challenging than standard NLI tasks, requiring task-specific finetuning for strong performance. Moreover, the gains from multilingual training and testing highlight the dataset’s potential for building effective multilingual fact-checking pipelines without relying on machine translation.

Method: Study shows different thinking languages occupy distinct regions in model’s thinking space. Two sampling strategies are studied: Single-Language Sampling (using one non-English thinking language) and Mixed-Language Sampling (aggregating across multiple thinking languages). Diversity is evaluated on English outputs regardless of thinking language used.

Result: Switching thinking language from English to non-English consistently increases output diversity, with positive correlation: languages farther from English in thinking space yield larger gains. Aggregating samples across multiple thinking languages yields additional improvements through compositional effects. Scaling sampling with linguistic heterogeneity expands model’s diversity ceiling.

Conclusion: Multilingual thinking provides practical benefits in pluralistic alignment scenarios, leading to broader coverage of cultural knowledge and value orientations in LLM outputs. This offers a novel approach to enhancing output diversity through language of thought control.

Abstract: Output diversity is crucial for Large Language Models as it underpins pluralism and creativity. In this work, we reveal that controlling the language used during model thinking-the language of thought-provides a novel and structural source of output diversity. Our preliminary study shows that different thinking languages occupy distinct regions in a model’s thinking space. Based on this observation, we study two repeated sampling strategies under multilingual thinking-Single-Language Sampling and Mixed-Language Sampling-and conduct diversity evaluation on outputs that are controlled to be in English, regardless of the thinking language used. Across extensive experiments, we demonstrate that switching the thinking language from English to non-English languages consistently increases output diversity, with a clear and consistent positive correlation such that languages farther from English in the thinking space yield larger gains. We further show that aggregating samples across multiple thinking languages yields additional improvements through compositional effects, and that scaling sampling with linguistic heterogeneity expands the model’s diversity ceiling. Finally, we show that these findings translate into practical benefits in pluralistic alignment scenarios, leading to broader coverage of cultural knowledge and value orientations in LLM outputs. Our code is publicly available at https://github.com/iNLP-Lab/Multilingual-LoT-Diversity.

Method: FactCorrector is a post-hoc correction approach that leverages structured feedback about the factuality of original LLM responses to generate corrections. It adapts across domains without requiring retraining.

Result: Experiments on the VELI5 benchmark and several popular long-form factuality datasets show that FactCorrector significantly improves factual precision while preserving relevance, outperforming strong baselines.

Conclusion: FactCorrector provides an effective approach for correcting factual errors in LLM responses using structured feedback, and the VELI5 benchmark enables rigorous evaluation of factuality correction methods.

Abstract: Large language models (LLMs) are widely used in knowledge-intensive applications but often generate factually incorrect responses. A promising approach to rectify these flaws is correcting LLMs using feedback. Therefore, in this paper, we introduce FactCorrector, a new post-hoc correction method that adapts across domains without retraining and leverages structured feedback about the factuality of the original response to generate a correction. To support rigorous evaluations of factuality correction methods, we also develop the VELI5 benchmark, a novel dataset containing systematically injected factual errors and ground-truth corrections. Experiments on VELI5 and several popular long-form factuality datasets show that the FactCorrector approach significantly improves factual precision while preserving relevance, outperforming strong baselines. We release our code at https://ibm.biz/factcorrector.

Method: DDAIR uses Sentence Transformers to detect ambiguous class-guided augmented examples generated by LLMs. It identifies synthetic examples that are semantically more similar to another intent than their target one, and provides an iterative re-generation method to mitigate such ambiguities.

Result: Sentence embeddings effectively help to (re)generate less ambiguous examples, showing promising potential to improve classification performance in scenarios where intents are loosely or broadly defined.

Conclusion: The DDAIR approach successfully addresses the ambiguity problem in LLM-generated data augmentation for intent recognition, particularly benefiting low-resource scenarios with loosely defined intents.

Abstract: Large Language Models (LLMs) are effective for data augmentation in classification tasks like intent detection. In some cases, they inadvertently produce examples that are ambiguous with regard to untargeted classes. We present DDAIR (Disambiguated Data Augmentation for Intent Recognition) to mitigate this problem. We use Sentence Transformers to detect ambiguous class-guided augmented examples generated by LLMs for intent recognition in low-resource scenarios. We identify synthetic examples that are semantically more similar to another intent than to their target one. We also provide an iterative re-generation method to mitigate such ambiguities. Our findings show that sentence embeddings effectively help to (re)generate less ambiguous examples, and suggest promising potential to improve classification performance in scenarios where intents are loosely or broadly defined.

Method: RT-RAG systematically decomposes multi-hop questions into explicit reasoning trees using structured entity analysis and consensus-based tree selection, then employs bottom-up traversal with iterative query rewriting and refinement to collect evidence.

Result: RT-RAG substantially outperforms state-of-the-art methods by 7.0% F1 and 6.0% EM in comprehensive experiments.

Conclusion: The reasoning tree guided approach effectively addresses decomposition accuracy and error propagation issues in complex multi-hop QA, demonstrating significant performance improvements over existing methods.

Abstract: Retrieval-Augmented Generation (RAG) has demonstrated significant effectiveness in enhancing large language models (LLMs) for complex multi-hop question answering (QA). For multi-hop QA tasks, current iterative approaches predominantly rely on LLMs to self-guide and plan multi-step exploration paths during retrieval, leading to substantial challenges in maintaining reasoning coherence across steps from inaccurate query decomposition and error propagation. To address these issues, we introduce Reasoning Tree Guided RAG (RT-RAG), a novel hierarchical framework for complex multi-hop QA. RT-RAG systematically decomposes multi-hop questions into explicit reasoning trees, minimizing inaccurate decomposition through structured entity analysis and consensus-based tree selection that clearly separates core queries, known entities, and unknown entities. Subsequently, a bottom-up traversal strategy employs iterative query rewriting and refinement to collect high-quality evidence, thereby mitigating error propagation. Comprehensive experiments show that RT-RAG substantially outperforms state-of-the-art methods by 7.0% F1 and 6.0% EM, demonstrating the effectiveness of RT-RAG in complex multi-hop QA.

Method: Propose multi-task learning (MTL) to fine-tune a single small decoder-only LLM (e.g., Qwen3-4b) for three AFC tasks: claim detection, evidence ranking, and stance detection. Explore three MTL strategies: classification heads, causal language modeling heads, and instruction-tuning. Evaluate across model sizes, task orders, and compare with standard non-LLM baselines.

Result: Multi-task models yield substantial improvements over zero-/few-shot settings: up to 44% relative gain for claim detection, 54% for evidence re-ranking, and 31% for stance detection. While not universally surpassing single-task baselines, they provide significant performance gains with practical efficiency.

Conclusion: Multi-task learning with small LLMs offers an efficient alternative for automated fact-checking, balancing performance and cost. The paper provides empirically grounded guidelines for practitioners to apply MTL with LLMs in AFC workflows.

Abstract: Large language models (LLMs) are reshaping automated fact-checking (AFC) by enabling unified, end-to-end verification pipelines rather than isolated components. While large proprietary models achieve strong performance, their closed weights, complexity, and high costs limit sustainability. Fine-tuning smaller open weight models for individual AFC tasks can help but requires multiple specialized models resulting in high costs. We propose \textbf{multi-task learning (MTL)} as a more efficient alternative that fine-tunes a single model to perform claim detection, evidence ranking, and stance detection jointly. Using small decoder-only LLMs (e.g., Qwen3-4b), we explore three MTL strategies: classification heads, causal language modeling heads, and instruction-tuning, and evaluate them across model sizes, task orders, and standard non-LLM baselines. While multitask models do not universally surpass single-task baselines, they yield substantial improvements, achieving up to \textbf{44%}, \textbf{54%}, and \textbf{31%} relative gains for claim detection, evidence re-ranking, and stance detection, respectively, over zero-/few-shot settings. Finally, we also provide practical, empirically grounded guidelines to help practitioners apply MTL with LLMs for automated fact-checking.

Method: SimMIA uses an advanced sampling strategy and scoring mechanism tailored for text-only regime, plus introduces WikiMIA-25 benchmark for evaluating MIA on modern proprietary LLMs.

Result: SimMIA achieves state-of-the-art results in black-box setting, rivaling baselines that exploit internal model information.

Conclusion: SimMIA provides an effective auditing tool for LLM training data in practical black-box scenarios where only text outputs are available.

Abstract: Membership Inference Attacks (MIAs) act as a crucial auditing tool for the opaque training data of Large Language Models (LLMs). However, existing techniques predominantly rely on inaccessible model internals (e.g., logits) or suffer from poor generalization across domains in strict black-box settings where only generated text is available. In this work, we propose SimMIA, a robust MIA framework tailored for this text-only regime by leveraging an advanced sampling strategy and scoring mechanism. Furthermore, we present WikiMIA-25, a new benchmark curated to evaluate MIA performance on modern proprietary LLMs. Experiments demonstrate that SimMIA achieves state-of-the-art results in the black-box setting, rivaling baselines that exploit internal model information.

Method: Keep audio encoder frozen and finetune only the language model, requiring just 2,000 hours of data. Single-stage training protocol without large-scale pretraining or multi-stage optimization. Model follows explicit instructions to control speaker voice, conversation topic, and conversational behavior.

Result: Developed the first open, instruction-following full-duplex conversational speech model that can be trained efficiently under typical academic resource constraints. The model enables control over speaker voice, conversation topic, conversational behavior (backchanneling, interruptions), and dialogue initiation.

Conclusion: The model and training code will be released to enable reproducible research on controllable full-duplex speech systems, addressing limitations of current systems and advancing natural conversational AI.

Abstract: Spoken conversational systems require more than accurate speech generation to have human-like conversations: to feel natural and engaging, they must produce conversational behaviour that adapts dynamically to the context. Current spoken conversational systems, however, rarely allow such customization, limiting their naturalness and usability. In this work, we present the first open, instruction-following full-duplex conversational speech model that can be trained efficiently under typical academic resource constraints. By keeping the audio encoder frozen and finetuning only the language model, our model requires just 2,000 hours of data, without relying on large-scale pretraining or multi-stage optimization. The model can follow explicit instructions to control speaker voice, conversation topic, conversational behaviour (e.g., backchanneling and interruptions), and dialogue initiation. We propose a single-stage training protocol and systematically analyze design choices. Both the model and training code will be released to enable reproducible research on controllable full-duplex speech systems.

Method: The authors propose using natural-language editorials (solution explanations) for both solution generation and evaluation. They introduce a dataset of 83 ICPC-style problems with gold editorials and full test suites. They evaluate 19 LLMs using both generated and gold editorials, and develop an LLM-as-a-judge protocol for scalable evaluation of editorial quality.

Result: Generating editorials before code improves solve rates for some LLMs, with larger gains when using expertly written gold editorials. However, models still struggle with implementation even with gold editorials, and the gap between generated and gold editorials reveals persistent problem-solving bottlenecks. The LLM-as-a-judge protocol was validated for scalable editorial evaluation.

Conclusion: Future benchmarks should explicitly separate problem solving from implementation. The editorial-based approach provides better diagnostic capabilities for understanding LLM limitations in competitive programming, revealing distinct bottlenecks in algorithmic reasoning versus code implementation.

Abstract: Large Language Models (LLMs) increasingly succeed on competitive programming problems, yet existing evaluations conflate algorithmic reasoning with code-level implementation. We argue that competitive programming is fundamentally a problem-solving task and propose centering natural-language editorials in both solution generation and evaluation. Generating an editorial prior to code improves solve rates for some LLMs, with substantially larger gains when using expertly written gold editorials. However, even with gold editorials, models continue to struggle with implementation, while the gap between generated and gold editorials reveals a persistent problem-solving bottleneck in specifying correct and complete algorithms. Beyond pass/fail metrics, we diagnose reasoning errors by comparing model-generated editorials to gold standards using expert annotations and validate an LLM-as-a-judge protocol for scalable evaluation. We introduce a dataset of 83 ICPC-style problems with gold editorials and full test suites, and evaluate 19 LLMs, arguing that future benchmarks should explicitly separate problem solving from implementation.

Method: Neural Chain-of-Thought Search (NCoTS) reformulates reasoning as dynamic search for optimal thinking strategy, uses quantitative characterization of solution space, and navigates using dual-factor heuristic evaluating candidate reasoning operators for correctness and computational cost.

Result: NCoTS achieves Pareto improvement across diverse reasoning benchmarks, boosting accuracy by over 3.5% while reducing generation length by over 22%, revealing existence of sparse superior reasoning paths.

Conclusion: NCoTS demonstrates that reformulating reasoning as search for optimal thinking strategies enables finding more accurate and concise reasoning paths, providing a framework for efficient reasoning in LLMs.

Abstract: Chain-of-Thought reasoning has significantly enhanced the problem-solving capabilities of Large Language Models. Unfortunately, current models generate reasoning steps sequentially without foresight, often becoming trapped in suboptimal reasoning paths with redundant steps. In contrast, we introduce Neural Chain-of-Thought Search (NCoTS), a framework that reformulates reasoning as a dynamic search for the optimal thinking strategy. By quantitatively characterizing the solution space, we reveal the existence of sparse superior reasoning paths that are simultaneously more accurate and concise than standard outputs. Our method actively navigates towards these paths by evaluating candidate reasoning operators using a dual-factor heuristic that optimizes for both correctness and computational cost. Consequently, NCoTS achieves a Pareto improvement across diverse reasoning benchmarks, boosting accuracy by over 3.5% while reducing generation length by over 22%. Our code and data are available at https://github.com/MilkThink-Lab/Neural-CoT-Search.

Method: Developed a novel taxonomy of thematic elements in clinician responses and evaluation framework for assessing editing load. Created expert-annotated dataset and conducted large-scale evaluations of local/commercial LLMs using thematic prompting, retrieval-augmented generation, supervised fine-tuning, and direct preference optimization.

Result: Substantial epistemic uncertainty in aligning LLM drafts with clinician responses. LLMs perform well on some thematic elements but struggle with clinician-aligned generation in others, especially question-asking to elicit patient information. Theme-driven adaptation strategies improve performance across most themes.

Conclusion: LLMs need adaptation to individual clinician preferences for reliable and responsible use in patient-clinician communication workflows, highlighting the necessity of personalized alignment rather than one-size-fits-all solutions.

Abstract: Large language models (LLMs) show promise in drafting responses to patient portal messages, yet their integration into clinical workflows raises various concerns, including whether they would actually save clinicians time and effort in their portal workload. We investigate LLM alignment with individual clinicians through a comprehensive evaluation of the patient message response drafting task. We develop a novel taxonomy of thematic elements in clinician responses and propose a novel evaluation framework for assessing clinician editing load of LLM-drafted responses at both content and theme levels. We release an expert-annotated dataset and conduct large-scale evaluations of local and commercial LLMs using various adaptation techniques including thematic prompting, retrieval-augmented generation, supervised fine-tuning, and direct preference optimization. Our results reveal substantial epistemic uncertainty in aligning LLM drafts with clinician responses. While LLMs demonstrate capability in drafting certain thematic elements, they struggle with clinician-aligned generation in other themes, particularly question asking to elicit further information from patients. Theme-driven adaptation strategies yield improvements across most themes. Our findings underscore the necessity of adapting LLMs to individual clinician preferences to enable reliable and responsible use in patient-clinician communication workflows.

Method: A two-stage training framework: first optimizes scientific evaluation preferences, then refines reasoning capabilities. Uses multi-aspect evaluation design and joint training across diverse tasks to enable fine-grained assessment and robustness to dynamic criteria and scoring rubrics.

Result: The training regime strongly improves LLM-based scientific writing evaluation. The models generalize effectively across tasks and to previously unseen scientific writing evaluation settings, allowing a single trained evaluator to be reused without task-specific retraining.

Conclusion: The proposed cost-efficient, open-source reward models bridge the gap in scientific writing evaluation by providing adaptable, generalizable evaluators that don’t require task-specific retraining, addressing the limitations of existing LLM-based judges.

Abstract: Scientific writing is an expert-domain task that demands deep domain knowledge, task-specific requirements and reasoning capabilities that leverage the domain knowledge to satisfy the task specifications. While scientific text generation has been widely studied, its evaluation remains a challenging and open problem. It is critical to develop models that can be reliably deployed for evaluating diverse open-ended scientific writing tasks while adhering to their distinct requirements. However, existing LLM-based judges and reward models are primarily optimized for general-purpose benchmarks with fixed scoring rubrics and evaluation criteria. Consequently, they often fail to reason over sparse knowledge of scientific domains when interpreting task-dependent and multi-faceted criteria. Moreover, fine-tuning for each individual task is costly and impractical for low-resource settings. To bridge these gaps, we propose cost-efficient, open-source reward models tailored for scientific writing evaluation. We introduce a two-stage training framework that initially optimizes scientific evaluation preferences and then refines reasoning capabilities. Our multi-aspect evaluation design and joint training across diverse tasks enable fine-grained assessment and robustness to dynamic criteria and scoring rubrics. Experimental analysis shows that our training regime strongly improves LLM-based scientific writing evaluation. Our models generalize effectively across tasks and to previously unseen scientific writing evaluation settings, allowing a single trained evaluator to be reused without task-specific retraining.

Method: Proposed framework using economic methodologies to analyze LLM decision logic; built synthetic datasets from real freelancer profiles and project descriptions from a European online freelance marketplace; applied full factorial design to estimate how LLMs weigh different match-relevant criteria.

Result: LLMs weigh core productivity signals (skills, experience) but interpret certain features beyond their explicit matching value; minimal average discrimination against minority groups, but intersectional effects show productivity signals carry different weights between demographic groups.

Conclusion: The paper provides a framework to evaluate LLM decision logic in recruitment and demonstrates how comparable experimental setups could be implemented with human recruiters to assess alignment between model and human decisions.

Abstract: General-purpose Large Language Models (LLMs) show significant potential in recruitment applications, where decisions require reasoning over unstructured text, balancing multiple criteria, and inferring fit and competence from indirect productivity signals. Yet, it is still uncertain how LLMs assign importance to each attribute and whether such assignments are in line with economic principles, recruiter preferences or broader societal norms. We propose a framework to evaluate an LLM’s decision logic in recruitment, by drawing on established economic methodologies for analyzing human hiring behavior. We build synthetic datasets from real freelancer profiles and project descriptions from a major European online freelance marketplace and apply a full factorial design to estimate how a LLM weighs different match-relevant criteria when evaluating freelancer-project fit. We identify which attributes the LLM prioritizes and analyze how these weights vary across project contexts and demographic subgroups. Finally, we explain how a comparable experimental setup could be implemented with human recruiters to assess alignment between model and human decisions. Our findings reveal that the LLM weighs core productivity signals, such as skills and experience, but interprets certain features beyond their explicit matching value. While showing minimal average discrimination against minority groups, intersectional effects reveal that productivity signals carry different weights between demographic groups.

Method: Created SyntHal dataset with 6000 synthetic entities for six relations, measured hallucination rates on four models, and quantified relational linearity using Δcos metric.

Result: Found strong correlation (r ∈ [.78,.82]) between relational linearity and hallucination rate, showing linear relations cause more hallucinations than nonlinear ones.

Conclusion: The underlying storage format of factual triples (linear vs nonlinear) significantly affects LLMs’ ability to self-assess knowledge, suggesting new approaches for managing hallucinations and improving knowledge representation.

Abstract: Hallucination is a central failure mode in large language models (LLMs). We focus on hallucinations of answers to questions like: “Which instrument did Glenn Gould play?”, but we ask these questions for synthetic entities that are unknown to the model. Surprisingly, we find that medium-size models like Gemma-7B-IT frequently hallucinate, i.e., they have difficulty recognizing that the hallucinated fact is not part of their knowledge. We hypothesize that an important factor in causing these hallucinations is the linearity of the relation: linear relations tend to be stored more abstractly, making it difficult for the LLM to assess its knowledge; the facts of nonlinear relations tend to be stored more directly, making knowledge assessment easier. To investigate this hypothesis, we create SyntHal, a dataset of 6000 synthetic entities for six relations. In our experiments with four models, we determine, for each relation, the hallucination rate on SyntHal and also measure its linearity, using $Δ\cos$. We find a strong correlation ($r \in [.78,.82]$) between relational linearity and hallucination rate, providing evidence for our hypothesis that the underlying storage of triples of a relation is a factor in how well a model can self-assess its knowledge. This finding has implications for how to manage hallucination behavior and suggests new research directions for improving the representation of factual knowledge in LLMs.

Method: Testing large language models on “Jabberwocky” language where most or all content words are randomly replaced by nonsense strings, then evaluating their ability to translate/recover the original meaning.

Result: LLMs demonstrate astonishing ability to make sense of Jabberwocky language, successfully recovering meaning from structural patterns even when content words are nonsense.

Conclusion: Pattern-matching is not an alternative to “real” intelligence but rather a key ingredient; LLMs’ ability to recover meaning from structural patterns speaks to the unreasonable effectiveness of pattern-matching in language understanding.

Abstract: We report on an astonishing ability of large language models (LLMs) to make sense of “Jabberwocky” language in which most or all content words have been randomly replaced by nonsense strings, e.g., translating “He dwushed a ghanc zawk” to “He dragged a spare chair”. This result addresses ongoing controversies regarding how to best think of what LLMs are doing: are they a language mimic, a database, a blurry version of the Web? The ability of LLMs to recover meaning from structural patterns speaks to the unreasonable effectiveness of pattern-matching. Pattern-matching is not an alternative to “real” intelligence, but rather a key ingredient.

Method: Proposes HORSE (Hierarchical Orthogonal Residual SprEad) that focuses on the information matrix to reduce noisy gradients, enabling more stable edits from a different perspective than traditional optimization approaches.

Result: Extensive experiments on two datasets across multiple LLMs show HORSE maintains precise massive editing across diverse scenarios, with theoretical comparisons demonstrating advantages over popular methods.

Conclusion: HORSE provides an effective alternative to existing model editing approaches, offering more stable and computationally efficient safety improvements for LLMs through hierarchical orthogonal residual spread.

Abstract: Large language models (LLMs) exhibit exceptional performance across various domains, yet they face critical safety concerns. Model editing has emerged as an effective approach to mitigate these issues. Existing model editing methods often focus on optimizing an information matrix that blends new and old knowledge. While effective, these approaches can be computationally expensive and may cause conflicts. In contrast, we shift our attention to Hierarchical Orthogonal Residual SprEad of the information matrix, which reduces noisy gradients and enables more stable edits from a different perspective. We demonstrate the effectiveness of our method HORSE through a clear theoretical comparison with several popular methods and extensive experiments conducted on two datasets across multiple LLMs. The results show that HORSE maintains precise massive editing across diverse scenarios. The code is available at https://github.com/XiaojieGu/HORSE

Method: TTARAG introduces test-time adaptation where the language model dynamically updates its parameters during inference. The model learns to predict retrieved content, enabling automatic parameter adjustment to the target domain without requiring labeled data.

Result: Extensive experiments across six specialized domains show TTARAG achieves substantial performance improvements over baseline RAG systems, demonstrating effective domain adaptation.

Conclusion: TTARAG provides an effective test-time adaptation approach for RAG systems in specialized domains, enabling dynamic parameter updates during inference to overcome distribution shift challenges.

Abstract: Retrieval-Augmented Generation (RAG) has emerged as a powerful approach for enhancing large language models’ question-answering capabilities through the integration of external knowledge. However, when adapting RAG systems to specialized domains, challenges arise from distribution shifts, resulting in suboptimal generalization performance. In this work, we propose TTARAG, a test-time adaptation method that dynamically updates the language model’s parameters during inference to improve RAG system performance in specialized domains. Our method introduces a simple yet effective approach where the model learns to predict retrieved content, enabling automatic parameter adjustment to the target domain. Through extensive experiments across six specialized domains, we demonstrate that TTARAG achieves substantial performance improvements over baseline RAG systems. Code available at https://github.com/sunxin000/TTARAG.

Method: Three-module framework: (1) Writing Style Generalizability via LLM-based rephrasing, (2) Synthetic Error Injection at graded severities, (3) Metrics-vs-Expert correlation using clinician ratings on 175 disagreement cases. Evaluates 8 metrics across 7 LLMs with CT-CLIP encoder.

Result: Lexical NLG metrics are highly sensitive to stylistic variations; GREEN Score aligns best with expert judgments (Spearman~0.70); CRG shows negative correlation; BERTScore-F1 is least sensitive to factual error injection.

Conclusion: CTest-Metric provides a standardized framework for assessing clinical feasibility of RRG metrics, revealing significant performance variations and enabling reproducible benchmarking for future metric development.

Abstract: In the generative AI era, where even critical medical tasks are increasingly automated, radiology report generation (RRG) continues to rely on suboptimal metrics for quality assessment. Developing domain-specific metrics has therefore been an active area of research, yet it remains challenging due to the lack of a unified, well-defined framework to assess their robustness and applicability in clinical contexts. To address this, we present CTest-Metric, a first unified metric assessment framework with three modules determining the clinical feasibility of metrics for CT RRG. The modules test: (i) Writing Style Generalizability (WSG) via LLM-based rephrasing; (ii) Synthetic Error Injection (SEI) at graded severities; and (iii) Metrics-vs-Expert correlation (MvE) using clinician ratings on 175 “disagreement” cases. Eight widely used metrics (BLEU, ROUGE, METEOR, BERTScore-F1, F1-RadGraph, RaTEScore, GREEN Score, CRG) are studied across seven LLMs built on a CT-CLIP encoder. Using our novel framework, we found that lexical NLG metrics are highly sensitive to stylistic variations; GREEN Score aligns best with expert judgments (Spearman~0.70), while CRG shows negative correlation; and BERTScore-F1 is least sensitive to factual error injection. We will release the framework, code, and allowable portion of the anonymized evaluation data (rephrased/error-injected CT reports), to facilitate reproducible benchmarking and future metric development.

Method: Evaluated explanation generalization by testing whether explanations from one LRM induce same behavior in other LRMs, measured consistency between models, analyzed conditions for consistent answers, and proposed sentence-level ensembling strategy.

Result: CoT explanations often generalize (increase consistency between LRMs), and this generalization correlates with human preference rankings and reinforcement learning post-training.

Conclusion: Caution needed when using LRM explanations for new insights; proposed framework for characterizing explanation generalization; sentence-level ensembling improves consistency.

Abstract: Large reasoning models (LRMs) produce a textual chain of thought (CoT) in the process of solving a problem, which serves as a potentially powerful tool to understand the problem by surfacing a human-readable, natural-language explanation. However, it is unclear whether these explanations generalize, i.e. whether they capture general patterns about the underlying problem rather than patterns which are esoteric to the LRM. This is a crucial question in understanding or discovering new concepts, e.g. in AI for science. We study this generalization question by evaluating a specific notion of generalizability: whether explanations produced by one LRM induce the same behavior when given to other LRMs. We find that CoT explanations often exhibit this form of generalization (i.e. they increase consistency between LRMs) and that this increased generalization is correlated with human preference rankings and post-training with reinforcement learning. We further analyze the conditions under which explanations yield consistent answers and propose a straightforward, sentence-level ensembling strategy that improves consistency. Taken together, these results prescribe caution when using LRM explanations to yield new insights and outline a framework for characterizing LRM explanation generalization.

Method: The authors conduct a comprehensive empirical analysis of tokenization, exploring how sequences are compressed into tokens across different distributions of textual data.

Result: The analysis reveals significant variation in tokenization across models and text domains, challenging commonly held heuristics about token lengths as being overly simplistic.

Conclusion: Tokenization is not as stable or consistent as commonly assumed, and naive interpretation of token counts is problematic. The study aims to provide clarity and intuition about tokenization in contemporary LLMs.

Abstract: Frontier LLMs are increasingly utilised across academia, society and industry. A commonly used unit for comparing models, their inputs and outputs, and estimating inference pricing is the token. In general, tokens are used as a stable currency, assumed to be broadly consistent across tokenizers and contexts, enabling direct comparisons. However, tokenization varies significantly across models and domains of text, making naive interpretation of token counts problematic. We quantify this variation by providing a comprehensive empirical analysis of tokenization, exploring the compression of sequences to tokens across different distributions of textual data. Our analysis challenges commonly held heuristics about token lengths, finding them to be overly simplistic. We hope the insights of our study add clarity and intuition toward tokenization in contemporary LLMs.

Method: Developed an evaluation framework to study collaboration settings, specifically comparing synchronous vs. asynchronous collaboration using three different NLP-assisted qualitative research tools.

Result: Found significant differences in consistency, cohesiveness, and correctness of outputs between collaboration modes across the three systems.

Conclusion: The proposed evaluation framework helps understand how collaboration settings affect NLP-assisted qualitative analysis outcomes, providing guidance for tool development and research methodology.

Abstract: NLP-assisted solutions have gained considerable traction to support qualitative data analysis. However, no unified evaluation framework exists which can account for the many different settings in which qualitative researchers may employ them. In this paper, we propose an evaluation framework to study the way collaboration settings may produce different outcomes across a variety of interactive systems. Specifically, we study the impact of synchronous vs. asynchronous collaboration using three different NLP-assisted qualitative research tools and present a comprehensive analysis of significant differences in the consistency, cohesiveness, and correctness of their outputs.

Method: Incorporates frozen representations of various language models into a discriminative vision-language framework, measures zero-shot generalization to novel concepts, and proposes ShareLock - a lightweight method for fusing frozen vision and language backbones with minimal training data and compute.

Result: Decoder-based models show stronger visual alignment than encoders; language modeling performance correlates with visual generalization; ShareLock achieves 51% accuracy on ImageNet with only 563k image-caption pairs and under one GPU-hour training; dramatically outperforms CLIP in cross-lingual settings (38.7% vs 1.4% on Chinese image classification).

Conclusion: Advances in unimodal LLMs can simultaneously improve vision models; ShareLock provides an efficient method for vision-language alignment with minimal resources, enabling robust performance across tasks while reducing the need for paired data and compute.

Abstract: How well do text-only large language models (LLMs) align with the visual world? We present a systematic evaluation of this question by incorporating frozen representations of various language models into a discriminative vision-language framework and measuring zero-shot generalization to novel concepts. We find that decoder-based models exhibit stronger visual alignment than encoders, even when controlling for model and dataset size. Moreover, language modeling performance correlates with visual generalization, suggesting that advances in unimodal LLMs can simultaneously improve vision models. Leveraging these insights, we propose ShareLock, a lightweight method for fusing frozen vision and language backbones. ShareLock achieves robust performance across tasks while drastically reducing the need for paired data and compute. With just 563k image-caption pairs and under one GPU-hour of training, it reaches 51% accuracy on ImageNet. In cross-lingual settings, ShareLock dramatically outperforms CLIP, achieving 38.7% top-1 accuracy on Chinese image classification versus CLIP’s 1.4%. Code is available.

Method: Panacea optimizes adaptive perturbations applied to models after fine-tuning to recover safety alignment without compromising downstream performance, unlike simple random perturbations that degrade performance.

Result: Reduces average harmful scores by up to 21.2% across different harmful ratios, fine-tuning tasks, and LLMs while maintaining fine-tuning performance.

Conclusion: Panacea provides an effective defense against harmful fine-tuning attacks, revealing distinct safety affinity across different layers in LLMs that aligns with previous research findings.

Abstract: Harmful fine-tuning attack introduces significant security risks to the fine-tuning services. Main-stream defenses aim to vaccinate the model such that the later harmful fine-tuning attack is less effective. However, our evaluation results show that such defenses are fragile–with a few fine-tuning steps, the model still can learn the harmful knowledge. To this end, we do further experiment and find that an embarrassingly simple solution–adding purely random perturbations to the fine-tuned model, can recover the model from harmful behaviors, though it leads to a degradation in the model’s fine-tuning performance. To address the degradation of fine-tuning performance, we further propose Panacea, which optimizes an adaptive perturbation that will be applied to the model after fine-tuning. Panacea maintains model’s safety alignment performance without compromising downstream fine-tuning performance. Comprehensive experiments are conducted on different harmful ratios, fine-tuning tasks and mainstream LLMs, where the average harmful scores are reduced by up-to 21.2%, while maintaining fine-tuning performance. As a by-product, we analyze the adaptive perturbation and show that different layers in various LLMs have distinct safety affinity, which coincide with finding from several previous study. Source code available at https://github.com/w-yibo/Panacea.

Method: Constructed corpus from multiple wire services (AP, UPI, NEA) across entire newspapers, not just front pages. Used OCR text extraction plus LLM-based correction pipeline to reduce noise. Retained multiple versions of same wire dispatch for comparative analysis.

Result: Created a large-scale historical corpus with both raw and corrected text versions, enabling study of editorial differences across newspapers and comparative analysis of wire service patterns in Southern news transmission.

Conclusion: The corpus provides detailed perspective on how Southern newspapers transmitted national/international news during 1960-1975, offering valuable resource for studying editorial practices and news framing in a historically significant period.

Abstract: This paper describes the construction of a large-scale corpus of historical wire articles from U.S. Southern newspapers, spanning 1960-1975 and covering multiple wire services (e.g., Associated Press, United Press International, Newspaper Enterprise Association). Unlike prior work that focuses primarily on front-page content, the corpus captures wire-sourced articles across the entire newspaper, offering broader insight into mid-century Southern news coverage. The analysis incorporates both raw OCR text and a version processed through an LLM-based text correction pipeline designed to reduce OCR noise and improve suitability for quantitative text analysis. Multiple versions of the same wire dispatch are retained, allowing for the study of editorial differences in language and framing across newspapers. Articles are classified by wire service, enabling comparative analysis of editorial patterns across agencies. Together, these features provide a detailed perspective on how Southern newspapers transmitted national and international news during a transformative period in American history.

Method: Developed a taxonomy of DeepSeek-R1’s basic reasoning building blocks, then conducted analyses on: thought length impact and controllability, management of long/confusing contexts, cultural and safety concerns, and comparison to cognitive phenomena like human-like language processing and world modeling.

Result: DeepSeek-R1 has a ‘sweet spot’ of reasoning where extra inference time can impair performance. The model shows persistent rumination on previously explored problem formulations, obstructing further exploration. It exhibits strong safety vulnerabilities compared to its non-reasoning counterpart, which can also compromise safety-aligned LLMs.

Conclusion: The findings present a nuanced picture of reasoning models: while they enable transparent reasoning chains and open new research avenues in Thoughtology, they also reveal significant limitations including optimal reasoning thresholds, exploration constraints, and serious safety vulnerabilities that need addressing.

Abstract: Large Reasoning Models like DeepSeek-R1 mark a fundamental shift in how LLMs approach complex problems. Instead of directly producing an answer for a given input, DeepSeek-R1 creates detailed multi-step reasoning chains, seemingly “thinking” about a problem before providing an answer. This reasoning process is publicly available to the user, creating endless opportunities for studying the reasoning behaviour of the model and opening up the field of Thoughtology. Starting from a taxonomy of DeepSeek-R1’s basic building blocks of reasoning, our analyses on DeepSeek-R1 investigate the impact and controllability of thought length, management of long or confusing contexts, cultural and safety concerns, and the status of DeepSeek-R1 vis-à-vis cognitive phenomena, such as human-like language processing and world modelling. Our findings paint a nuanced picture. Notably, we show DeepSeek-R1 has a ‘sweet spot’ of reasoning, where extra inference time can impair model performance. Furthermore, we find a tendency for DeepSeek-R1 to persistently ruminate on previously explored problem formulations, obstructing further exploration. We also note strong safety vulnerabilities of DeepSeek-R1 compared to its non-reasoning counterpart, which can also compromise safety-aligned LLMs.

Method: Proposes PSN-IRT (Pseudo-Siamese Network for Item Response Theory), an enhanced IRT framework with rich item parameters. Uses this to analyze 11 LLM benchmarks with 41,871 items. Demonstrates PSN-IRT can construct smaller benchmarks while maintaining human preference alignment.

Result: Analysis reveals significant and varied shortcomings in benchmark measurement quality. PSN-IRT enables accurate estimation of item characteristics and model abilities. Shows that leveraging PSN-IRT can create smaller benchmarks with stronger alignment to human preferences.

Conclusion: Current LLM benchmarks have substantial measurement quality issues. PSN-IRT provides a more reliable framework for evaluating LLMs and can be used to construct more efficient, human-aligned benchmarks that better reflect true model capabilities.

Abstract: The evaluation of large language models (LLMs) via benchmarks is widespread, yet inconsistencies between different leaderboards and poor separability among top models raise concerns about their ability to accurately reflect authentic model capabilities. This paper provides a critical analysis of benchmark effectiveness, examining mainstream prominent LLM benchmarks using results from diverse models. We first propose Pseudo-Siamese Network for Item Response Theory (PSN-IRT), an enhanced Item Response Theory framework that incorporates a rich set of item parameters within an IRT-grounded architecture. PSN-IRT can be utilized for accurate and reliable estimations of item characteristics and model abilities. Based on PSN-IRT, we conduct extensive analysis on 11 LLM benchmarks comprising 41,871 items, revealing significant and varied shortcomings in their measurement quality. Furthermore, we demonstrate that leveraging PSN-IRT is able to construct smaller benchmarks while maintaining stronger alignment with human preference.

Method: Proposes Inferential Preference Mining (IPM) to construct high-quality preference data (IPM-PrefDial dataset), then introduces a Decoupled ESC framework inspired by Gross’s Extended Process Model of Emotion Regulation. The framework decomposes ESC into two sequential subtasks: strategy planning and empathic response generation, each trained via SFT and enhanced by DPO.

Result: Extensive experiments show the Decoupled ESC framework outperforms joint optimization baselines, reducing preference bias and improving response quality.

Conclusion: Decoupling ESC tasks and using IPM for preference data construction effectively addresses psychological errors in emotional support conversations, leading to better alignment with psychological preferences and improved response quality.

Abstract: Recent advances in Emotional Support Conversation (ESC) have improved emotional support generation by fine-tuning Large Language Models (LLMs) via Supervised Fine-Tuning (SFT). However, common psychological errors still persist. While Direct Preference Optimization (DPO) shows promise in reducing such errors through pairwise preference learning, its effectiveness in ESC tasks is limited by two key challenges: (1) Entangled data structure: Existing ESC data inherently entangles psychological strategies and response content, making it difficult to construct high-quality preference pairs; and (2) Optimization ambiguity: Applying vanilla DPO to such entangled pairwise data leads to ambiguous training objectives. To address these issues, we introduce Inferential Preference Mining (IPM) to construct high-quality preference data, forming the IPM-PrefDial dataset. Building upon this data, we propose a Decoupled ESC framework inspired by Gross’s Extended Process Model of Emotion Regulation, which decomposes the ESC task into two sequential subtasks: strategy planning and empathic response generation. Each was trained via SFT and subsequently enhanced by DPO to align with the psychological preference. Extensive experiments demonstrate that our Decoupled ESC framework outperforms joint optimization baselines, reducing preference bias and improving response quality.

Method: Created Chandomitra dataset for English-to-structured Sanskrit poetry translation. Benchmarked open/closed models, tested constrained decoding for metrical accuracy, and instruction fine-tuning for semantic coherence.

Result: Constrained decoding achieved 99.86% syntactic accuracy for metrically valid Sanskrit poetry, vastly outperforming GPT-4o (31.24%). Instruction-tuned models performed better in semantic coherence and poetic aspects, though with slightly lower syntactic accuracy.

Conclusion: Both constrained decoding and instruction fine-tuning are effective approaches for structured poetry generation in low-resource languages, with constrained decoding excelling at syntactic accuracy and instruction tuning better capturing semantic and poetic qualities.

Abstract: Text Generation has achieved remarkable performance using large language models. It has also been recently well-studied that these large language models are capable of creative generation tasks but prominently for high-resource languages. This prompts a fundamental question: Is there a way to utilize these (large) language models for structured poetry generation in a low-resource language, such as Sanskrit? We present Chandomitra, an English input to structured Sanskrit Poetry translation dataset, specifically adhering to the Anushtubh meter. We benchmark various open and closed models, and scrutinize specialized techniques such as constrained decoding and instruction fine-tuning, for the proposed task. Our constrained decoding methodology achieves 99.86% syntactic accuracy in generating metrically valid Sanskrit poetry, outperforming GPT-4o (1-shot: 31.24%). Our best-performing instruction-tuned model, on the other hand, performs better in semantic coherence with the English input, at the expense of slightly lower syntactic accuracy. Human evaluation further reveals that instruction fine-tuned model is better able to capture the poetic aspects. Data and Code are available.

Method: Used causal tracing to systematically analyze how pretrained causal transformers process idioms, localizing specific mechanisms in the model architecture.

Result: Identified three key mechanisms: 1) Early sublayers retrieve figurative meaning while suppressing literal, 2) Context is leveraged from earliest layers and refined if conflicting, 3) Parallel pathways maintain both interpretations with figurative prioritized in intermediate path and literal in direct route.

Conclusion: The study provides mechanistic evidence for how autoregressive transformers comprehend idioms, revealing sophisticated processing strategies that maintain both figurative and literal interpretations through selective, competing pathways.

Abstract: Idioms present a unique challenge for language models due to their non-compositional figurative interpretations, which often strongly diverge from the idiom’s literal interpretation. In this paper, we employ causal tracing to systematically analyze how pretrained causal transformers deal with this ambiguity. We localize three mechanisms: (i) Early sublayers and specific attention heads retrieve an idiom’s figurative interpretation, while suppressing its literal interpretation. (ii) When disambiguating context precedes the idiom, the model leverages it from the earliest layer and later layers refine the interpretation if the context conflicts with the retrieved interpretation. (iii) Then, selective, competing pathways carry both interpretations: an intermediate pathway prioritizes the figurative interpretation and a parallel direct route favors the literal interpretation, ensuring that both readings remain available. Our findings provide mechanistic evidence for idiom comprehension in autoregressive transformers.

Method: Built around a standardized Dialog representation, SDialog provides: persona-driven multi-agent simulation with composable orchestration, comprehensive evaluation combining linguistic metrics and LLM-as-a-judge, mechanistic interpretability tools for activation inspection and steering, and audio generation with full acoustic simulation including 3D room modeling.

Result: The toolkit integrates with all major LLM backends, enabling mixed-backend experiments under a unified API, and provides a dialog-centric architecture for systematic conversational system development.

Conclusion: SDialog enables researchers to build, benchmark, and understand conversational systems more systematically by coupling generation, evaluation, and interpretability in a single framework.

Abstract: We present SDialog, an MIT-licensed open-source Python toolkit that unifies dialog generation, evaluation and mechanistic interpretability into a single end-to-end framework for building and analyzing LLM-based conversational agents. Built around a standardized Dialog representation, SDialog provides: (1) persona-driven multi-agent simulation with composable orchestration for controlled, synthetic dialog generation, (2) comprehensive evaluation combining linguistic metrics, LLM-as-a-judge and functional correctness validators, (3) mechanistic interpretability tools for activation inspection and steering via feature ablation and induction, and (4) audio generation with full acoustic simulation including 3D room modeling and microphone effects. The toolkit integrates with all major LLM backends, enabling mixed-backend experiments under a unified API. By coupling generation, evaluation, and interpretability in a dialog-centric architecture, SDialog enables researchers to build, benchmark and understand conversational systems more systematically.

Method: Proposes MIST framework that reconceptualizes stereotypes as multidimensional ToM failures across competence, sociability, and morality domains. Uses two indirect tasks: Word Association Bias Test (WABT) for implicit lexical associations and Affective Attribution Test (AAT) for implicit emotional tendencies.

Result: Extensive experimentation on eight state-of-the-art LLMs demonstrates the framework’s ability to reveal complex bias structures with improved robustness compared to traditional methods.

Conclusion: MIST provides an effective indirect assessment framework for uncovering latent stereotypes in LLMs’ Theory of Mind capabilities, avoiding model avoidance behaviors while capturing multidimensional bias structures.

Abstract: Theory of Mind (ToM) in Large Language Models (LLMs) refers to the model’s ability to infer the mental states of others, with failures in this ability often manifesting as systemic implicit biases. Assessing this challenge is difficult, as traditional direct inquiry methods are often met with refusal to answer and fail to capture its subtle and multidimensional nature. Therefore, we propose MIST, which reconceptualizes the content model of stereotypes into multidimensional failures of ToM, specifically in the domains of competence, sociability, and morality. The framework introduces two indirect tasks. The Word Association Bias Test (WABT) assesses implicit lexical associations, while the Affective Attribution Test (AAT) measures implicit emotional tendencies, aiming to uncover latent stereotypes without triggering model avoidance. Through extensive experimentation on eight state-of-the-art LLMs, our framework demonstrates the ability to reveal complex bias structures and improved robustness. All data and code will be released.

Method: The paper analyzes LLM applications in education through two main scenarios (assistance and assessment) across four dimensions: reading, writing, speaking, and tutoring. It presents new directions enabled by LLMs and identifies key challenges to address.

Result: The paper provides a holistic overview of LLM applications in educational NLP, highlighting both opportunities and challenges in developing language-focused educational applications.

Conclusion: This comprehensive analysis serves as a valuable resource for NLP researchers and practitioners interested in exploring LLMs for future educational applications, particularly in language-focused and NLP-enabled educational tools.

Abstract: Interest in the role of large language models (LLMs) in education is increasing, considering the new opportunities they offer for teaching, learning, and assessment. In this paper, we examine the impact of LLMs on educational NLP in the context of two main application scenarios: {\em assistance} and {\em assessment}, grounding them along the four dimensions – reading, writing, speaking, and tutoring. We then present the new directions enabled by LLMs, and the key challenges to address. We envision that this holistic overview would be useful for NLP researchers and practitioners interested in exploring the role of LLMs in developing language-focused and NLP-enabled educational applications of the future.

Method: Propose Culturescope - first mechanistic interpretability method to probe internal cultural knowledge representations in LLMs. Introduce cultural flattening score to measure intrinsic cultural biases in decoded knowledge.

Result: Found that low-resource cultures are less susceptible to cultural biases, likely due to models’ limited parametric knowledge about them. Traced emergence of Western-dominance bias and cultural flattening within LLMs.

Conclusion: Provides foundation for future research on mitigating cultural biases and enhancing LLMs’ cultural understanding through mechanistic interpretability approaches.

Abstract: The growing deployment of large language models (LLMs) across diverse cultural contexts necessitates a deeper understanding of LLMs’ representations of different cultures. Prior work has focused on evaluating the cultural awareness of LLMs by only examining the text they generate. This approach overlooks the internal sources of cultural misrepresentation within the models themselves. To bridge this gap, we propose Culturescope, the first mechanistic interpretability-based method that probes the internal representations of different cultural knowledge in LLMs. We also introduce a cultural flattening score as a measure of the intrinsic cultural biases of the decoded knowledge from Culturescope. Additionally, we study how LLMs internalize cultural biases, which allows us to trace how cultural biases such as Western-dominance bias and cultural flattening emerge within LLMs. We find that low-resource cultures are less susceptible to cultural biases, likely due to the model’s limited parametric knowledge. Our work provides a foundation for future research on mitigating cultural biases and enhancing LLMs’ cultural understanding.

Method: MedReflect introduces a physician-like reflective thinking mode with a single-pass reflection chain: initial hypothesis generation, self-questioning, self-answering, and decision refinement. This self-verified, self-reflective approach leverages LLMs’ latent capabilities without external retrieval.

Result: The approach enables cost-efficient medical dataset construction and achieves notable absolute accuracy improvements across medical benchmarks with only minimal randomly sampled training examples and lightweight fine-tuning, while significantly reducing annotation requirements.

Conclusion: LLMs can learn to solve specialized medical problems through self-reflection and self-improvement, reducing reliance on external supervision and extensive task-specific fine-tuning data.

Abstract: Medical problem-solving demands expert knowledge and intricate reasoning. Recent studies of large language models (LLMs) attempt to ease this complexity by introducing external knowledge verification through retrieval-augmented generation or by training on reasoning datasets. However, these approaches suffer from drawbacks such as retrieval overhead and high annotation costs, and they heavily rely on substituted external assistants to reach limited performance in medical field. In this paper, we introduce MedReflect, a generalizable framework designed to inspire LLMs with a physician-like reflective thinking mode. MedReflect generates a single-pass reflection chain that includes initial hypothesis generation, self-questioning, self-answering and decision refinement. This self-verified and self-reflective nature releases large language model’s latent capability in medical problem-solving without external retrieval or heavy annotation. We demonstrate that MedReflect enables cost-efficient medical dataset construction. With only a minimal subset of randomly sampled training examples and lightweight fine-tuning, this approach achieves notable absolute accuracy improvements across a series of medical benchmarks while significantly cutting annotation requirements. Our results provide evidence that LLMs can learn to solve specialized medical problems via self-reflection and self-improvement, reducing reliance on external supervision and extensive task-specific fine-tuning data.

Method: Multi-agent debate framework where multiple MLLM agents iteratively refine inferences through debate rounds, verifying and updating each other’s responses to improve performance and robustness.

Result: Experiments on ImplicitAVE dataset show significant accuracy improvements with just a few debate rounds, especially for attributes with initially low performance. Various debate configurations (identical/different agents) were evaluated, revealing insights about convergence dynamics.

Conclusion: Multi-agent debate strategies effectively address single-agent limitations and offer a scalable solution for implicit AVE in multimodal e-commerce applications.

Abstract: Implicit Attribute Value Extraction (AVE) is essential for accurately representing products in e-commerce, as it infers latent attributes from multimodal data. Despite advances in multimodal large language models (MLLMs), implicit AVE remains challenging due to the complexity of multidimensional data and gaps in vision-text understanding. In this work, we introduce MADIAVE, a multi-agent debate framework that employs multiple MLLM agents to iteratively refine inferences. Through a series of debate rounds, agents verify and update each other’s responses, thereby improving inference performance and robustness. Experiments on the ImplicitAVE dataset demonstrate that even a few rounds of debate significantly boost accuracy, especially for attributes with initially low performance. We systematically evaluate various debate configurations, including identical or different MLLM agents, and analyze how debate rounds affect convergence dynamics. Our findings highlight the potential of multi-agent debate strategies to address the limitations of single-agent approaches and offer a scalable solution for implicit AVE in multimodal e-commerce.

Method: Tested across four model families (Gemini, Kimi-K2, Grok, and Qwen) using narrow in-context examples. Measured EM rates with varying numbers of examples (2-16). Formulated and tested hypothesis about conflict between safety objectives and context-following behavior.

Result: EM emerges in ICL with rates ranging from 1% to 24% depending on model and domain. Neither larger model scale nor explicit reasoning provides reliable protection. Instructing models to prioritize safety reduces EM while prioritizing context-following increases it.

Conclusion: ICL is an underappreciated vector for emergent misalignment that operates without parameter modification and resists scaling-based solutions. The phenomenon stems from conflict between safety objectives and context-following behavior.

Abstract: Recent work has shown that narrow finetuning can produce broadly misaligned LLMs, a phenomenon termed emergent misalignment (EM). While concerning, these findings were limited to finetuning and activation steering, leaving out in-context learning (ICL). We therefore ask: does EM emerge in ICL? We find that it does: across four model families (Gemini, Kimi-K2, Grok, and Qwen), narrow in-context examples cause models to produce misaligned responses to benign, unrelated queries. With 16 in-context examples, EM rates range from 1% to 24% depending on model and domain, appearing with as few as 2 examples. Neither larger model scale nor explicit reasoning provides reliable protection. We formulate and test a hypothesis, which explains in-context EM as conflict between safety objectives and context-following behavior. Consistent with this, instructing models to prioritize safety reduces EM while prioritizing context-following increases it. These findings establish ICL as a previously underappreciated vector for emergent misalignment that operates without parameter modification and resists simple scaling-based solutions.

Method: The paper analyzes limitations of existing preference datasets that collapse multiple annotations into single labels, and proposes actionable strategies for incorporating HLV into dataset construction to preserve pluralistic human values.

Result: The paper positions HLV preservation as a Selbstzweck (intrinsic value) and provides analysis of current dataset limitations with proposed solutions for better capturing human pluralism in NLP systems.

Conclusion: HLV should be treated as an embodiment of human pluralism and preserved as an intrinsic value in dataset construction, especially for LLM alignment and safety evaluation where diverse human perspectives are essential.

Abstract: Human Label Variation (HLV) refers to legitimate disagreement in annotation that reflects the diversity of human perspectives rather than mere error. Long treated in NLP as noise to be eliminated, HLV has only recently been reframed as a signal for improving model robustness. With the rise of large language models (LLMs) and post-training methods such as human feedback-based alignment, the role of HLV has become increasingly consequential. Yet current preference-learning datasets routinely collapse multiple annotations into a single label, flattening diverse perspectives into artificial consensus. Preserving HLV is necessary not only for pluralistic alignment but also for sociotechnical safety evaluation, where model behavior must be assessed in relation to human interaction and societal context. This position paper argues that preserving HLV as an embodiment of human pluralism must be treated as a Selbstzweck, an intrinsic value in itself. We analyze the limitations of existing preference datasets and propose actionable strategies for incorporating HLV into dataset construction to better preserve pluralistic human values.

Method: 1) Created 106K multiple-choice problems using conjunction-based segmentation from news/cultural web sources. 2) Developed DRESS-AF (Distractor Ranking via Embedding Similarity Scoring and Adversarial Filtering) - a generation-free adversarial filtering method that selects challenging distractors from gold continuations to maximize model confusion.

Result: Human annotators scored 89%, OpenAI-o3 achieved 92.18%, Claude-Sonnet-3.7 got 91.17%, and the best open-source model (DeepSeek-R1) reached 82.51%. DRESS-AF also successfully transferred to English HellaSwag, increasing difficulty without hurting human solvability.

Conclusion: PerCoR establishes a challenging Persian commonsense reasoning benchmark with a significant performance gap between proprietary and open-source models, demonstrating both the dataset’s difficulty and the need for improved Persian language understanding capabilities.

Abstract: We introduced PerCoR (Persian Commonsense Reasoning), the first large-scale Persian benchmark for commonsense reasoning. PerCoR contains 106K multiple-choice sentence-completion problems drawn from more than forty news, cultural, and other web sources. We introduce a novel conjunction-based segmentation strategy to generate coherent sentence-completion pairs, enabling broad topical and structural diversity. To create challenging distractors, we propose DRESS-AF (Distractor Ranking via Embedding Similarity Scoring and Adversarial Filtering), a generation-free adversarial filtering method that selects distractors from the pool of gold continuations while maximising model confusion. Human annotators score 89% on PerCoR, while OpenAI-o3 achieves the highest performance at 92.18%, followed closely by Claude-Sonnet-3.7 (91.17%). The strongest open-source model, DeepSeek-R1, reaches 82.51%, underscoring both the dataset’s difficulty and the remaining performance gap in Persian commonsense reasoning. We further show that DRESS-AF transfers to the English HellaSwag benchmark, increasing its difficulty without hurting human solvability. The dataset is available at https://huggingface.co/datasets/MCINext/PerCoR.

Method: Used curated poetic prompts across 25 frontier models, converted 1,200 MLCommons harmful prompts into verse via standardized meta-prompt, evaluated outputs using ensemble of 3 LLM judges validated on human-labeled subset, mapped attacks to risk taxonomies.

Result: Poetic attacks achieved 62% success for hand-crafted poems and 43% for meta-prompt conversions, with ASRs up to 18x higher than prose baselines, affecting CBRN, manipulation, cyber-offence, and loss-of-control domains across all tested models.

Conclusion: Stylistic variation alone can circumvent contemporary safety mechanisms, revealing fundamental limitations in current alignment methods and evaluation protocols, suggesting poetry represents a systematic vulnerability across model families.

Abstract: We present evidence that adversarial poetry functions as a universal single-turn jailbreak technique for Large Language Models (LLMs). Across 25 frontier proprietary and open-weight models, curated poetic prompts yielded high attack-success rates (ASR), with some providers exceeding 90%. Mapping prompts to MLCommons and EU CoP risk taxonomies shows that poetic attacks transfer across CBRN, manipulation, cyber-offence, and loss-of-control domains. Converting 1,200 MLCommons harmful prompts into verse via a standardized meta-prompt produced ASRs up to 18 times higher than their prose baselines. Outputs are evaluated using an ensemble of 3 open-weight LLM judges, whose binary safety assessments were validated on a stratified human-labeled subset. Poetic framing achieved an average jailbreak success rate of 62% for hand-crafted poems and approximately 43% for meta-prompt conversions (compared to non-poetic baselines), substantially outperforming non-poetic baselines and revealing a systematic vulnerability across model families and safety training approaches. These findings demonstrate that stylistic variation alone can circumvent contemporary safety mechanisms, suggesting fundamental limitations in current alignment methods and evaluation protocols.

Method: A novel question generation pipeline using “generate-then-validate” strategy: expansive generation of candidate questions followed by selective validation using probabilistic reasoning. Evaluated with both human experts (7) and LLM judges.

Result: Most judges (human and LLM) agreed generated questions had clear answers and aligned well with learning objectives, demonstrating SLMs can effectively generate high-quality questions with proper pipeline design.

Conclusion: SLMs can be effective for automatic question generation when guided by well-designed pipelines that leverage their strengths, offering a viable complement to larger models in learning analytics.

Abstract: We explore the use of small language models (SLMs) for automatic question generation as a complement to the prevalent use of their large counterparts in learning analytics research. We present a novel question generation pipeline that leverages both the text generation and the probabilistic reasoning abilities of SLMs to generate high-quality questions. Adopting a “generate-then-validate” strategy, our pipeline first performs expansive generation to create an abundance of candidate questions and refine them through selective validation based on novel probabilistic reasoning. We conducted two evaluation studies, one with seven human experts and the other with a large language model (LLM), to assess the quality of the generated questions. Most judges (humans or LLMs) agreed that the generated questions had clear answers and generally aligned well with the intended learning objectives. Our findings suggest that an SLM can effectively generate high-quality questions when guided by a well-designed pipeline that leverages its strengths.

Method: Evaluated leading mLLMs for video-based emotional arousal measurement using two complementary human-labeled video datasets: (1) recordings created under laboratory conditions, and (2) real-world parliamentary debates. Also compared performance on video recordings versus text transcripts of the same speeches.

Result: Found critical lab-vs-field performance gap: mLLMs approach human-level reliability with little demographic bias in lab videos, but in parliamentary debates, arousal scores correlate only moderately with human ratings and show systematic bias by speaker gender and age. Neither closed-source mLLMs nor noise mitigation strategies helped. mLLMs also underperformed in sentiment analysis when using video instead of text transcripts.

Conclusion: Current mLLMs have important limitations for real-world political video analysis, revealing systematic biases and performance gaps that persist despite using leading models or mitigation strategies. The study establishes an evaluation framework for tracking future developments.

Abstract: Research increasingly leverages audio-visual materials to analyze emotions in political communication. Multimodal large language models (mLLMs) promise to enable such analyses through in-context learning. However, we lack systematic evidence on whether these models can reliably measure emotions in real-world political settings. This paper evaluates leading mLLMs for video-based emotional arousal measurement using two complementary human-labeled video datasets: recordings created under laboratory conditions and real-world parliamentary debates. I find a critical lab-vs-field performance gap. In video created under laboratory conditions, mLLMs arousal scores approach human-level reliability with little to no demographic bias. However, in parliamentary debate recordings, all examined models’ arousal scores correlate at best moderately with average human ratings and exhibit systematic bias by speaker gender and age. Neither relying on leading closed-source mLLMs nor computational noise mitigation strategies change this finding. Further, mLLMs underperform even in sentiment analysis when using video recordings instead of text transcripts of the same speeches. These findings reveal important limitations of current mLLMs for real-world political video analysis and establish a rigorous evaluation framework for tracking future developments.

Method: Proposes Mistake Notebook Learning (MNL) - a memory framework that enables agents to self-curate generalizable guidance from batch-clustered failures, distilling shared error patterns into structured “mistake notes” and updating external memory only when batch performance improves for stability. Integrates MNL with test-time scaling to actively steer search away from known pitfalls.

Result: Experiments on mathematical reasoning, Text-to-SQL, and interactive agent benchmarks show MNL achieves competitive performance compared to existing memory mechanisms and in-context methods in both effectiveness and efficiency.

Conclusion: Structured mistake abstraction is a critical lever for robust agent evolution, enabling continuous improvement without parameter updates. The approach positions mistake learning as essential for persistent agent performance.

Abstract: With the growing adoption of Large Language Model (LLM) agents in persistent, real-world roles, they naturally encounter continuous streams of tasks and inevitable failures. A key limitation, however, is their inability to systematically learn from these mistakes, forcing them to repeat identical errors in similar contexts. Unlike prior training-free methods that primarily store raw instance-level experience or focus on retrieving successful trajectories, we propose Mistake Notebook Learning (MNL), a novel memory framework that enables agents to self-curate generalizable guidance from batch-clustered failures. This mechanism allows agents to distill shared error patterns into structured “mistake notes,” updating an external memory only when batch performance improves to ensure stability. To further amplify adaptability, we integrate MNL with test-time scaling, leveraging aggregated failure patterns to actively steer the search process away from known pitfalls. Experiments on mathematical reasoning, Text-to-SQL, and interactive agent benchmarks show that MNL achieves competitive performance compared to existing memory mechanisms and in-context methods in both effectiveness and efficiency. These findings position structured mistake abstraction as a critical lever for robust agent evolution, enabling continuous improvement without the cost of parameter updates. The code is available at https://github.com/Bairong-Xdynamics/MistakeNotebookLearning/tree/main.

Method: Used Llama 3.3 70B to generate descriptions of 406 fictional characters with Big Five trait scores. Prompted model with descriptions and Alpaca questionnaire questions, sampled hidden activations, and learned per-layer linear directions via regression for probing and steering personality behavior.

Result: Linear directions aligned with trait-scores effectively probe personality detection. Steering capabilities are context-dependent: reliable in forced-choice tasks but limited in open-ended generation or when additional context is present in prompts.

Conclusion: Linear directions in activation space are effective probes for personality detection but have limited steering utility, suggesting they work best for constrained tasks rather than open-ended personality manipulation.

Abstract: Large language models (LLMs) exhibit distinct and consistent personalities that greatly impact trust and engagement. While this means that personality frameworks would be highly valuable tools to characterize and control LLMs’ behavior, current approaches remain either costly (post-training) or brittle (prompt engineering). Probing and steering via linear directions has recently emerged as a cheap and efficient alternative. In this paper, we investigate whether linear directions aligned with the Big Five personality traits can be used for probing and steering model behavior. Using Llama 3.3 70B, we generate descriptions of 406 fictional characters and their Big Five trait scores. We then prompt the model with these descriptions and questions from the Alpaca questionnaire, allowing us to sample hidden activations that vary along personality traits in known, quantifiable ways. Using linear regression, we learn a set of per-layer directions in activation space, and test their effectiveness for probing and steering model behavior. Our results suggest that linear directions aligned with trait-scores are effective probes for personality detection, while their steering capabilities strongly depend on context, producing reliable effects in forced-choice tasks but limited influence in open-ended generation or when additional context is present in the prompt.

Method: DEER includes 50 report-writing tasks across 13 domains with expert-grounded evaluation taxonomy (7 dimensions, 25 subdimensions, 101 rubric items), task-specific Expert Evaluation Guidance for LLM judges, and a document-level fact-checking architecture that verifies both cited and uncited claims while quantifying evidence quality.

Result: Experimental results show DEER exhibits strong correlation with human expert judgments and provides interpretable diagnostics of system strengths and weaknesses.

Conclusion: DEER addresses key limitations in evaluating deep research reports by providing comprehensive, expert-grounded evaluation with improved consistency and reliability assessment.

Abstract: As large language models advance, deep research systems capable of generating expert-level reports through multi-step reasoning and evidence-based synthesis are emerging. However, evaluating such reports remains challenging. Existing benchmarks often lack systematic evaluation criteria, rely heavily on LLM-based judges that may miss issues requiring expert judgment, and verify only a limited subset of explicitly cited statements rather than report-wide factual reliability. To address these limitations, we introduce DEER, a benchmark for evaluating expert-level deep research reports. DEER comprises 50 report-writing tasks spanning 13 domains, along with an expert-grounded evaluation taxonomy with seven dimensions and 25 subdimensions, operationalized into 101 fine-grained rubric items. To improve evaluation consistency, DEER provides task-specific Expert Evaluation Guidance to support LLM-based judging. Complementing rubric-based assessment, we propose a document-level fact-checking architecture that verifies both cited and uncited claims and quantifies the quality and reliability of the supporting evidence. Experimental results show that DEER exhibits strong correlation with human expert judgments and yields interpretable diagnostics of system strengths and weaknesses.

Method: Introduces FACTUM framework with four mechanistic scores: Contextual Alignment (CAS), Attention Sink Usage (BAS), Parametric Force (PFS), and Pathway Alignment (PAS). Analyzes coordination between Attention (reading) and Feed-Forward Network (recalling) pathways across different model scales.

Result: FACTUM outperforms state-of-the-art baselines by up to 37.5% in AUC. Correct citations show higher parametric force (PFS) and greater attention sink usage (BAS). Detection strategies evolve with scale: 3B models rely on high pathway alignment, while 8B models use specialized orthogonal information strategies.

Conclusion: Citation hallucinations result from coordination failures between neural pathways, not just parametric over-reliance. FACTUM’s nuanced understanding of these mechanisms enables more reliable RAG systems, showing that high parametric force can be constructive when properly coordinated with attention pathways.

Abstract: Retrieval-Augmented Generation (RAG) models are critically undermined by citation hallucinations, a deceptive failure where a model cites a source that fails to support its claim. While existing work attributes hallucination to a simple over-reliance on parametric knowledge, we reframe this failure as an evolving, scale-dependent coordination failure between the Attention (reading) and Feed-Forward Network (recalling) pathways. We introduce FACTUM (Framework for Attesting Citation Trustworthiness via Underlying Mechanisms), a framework of four mechanistic scores: Contextual Alignment (CAS), Attention Sink Usage (BAS), Parametric Force (PFS), and Pathway Alignment (PAS). Our analysis reveals that correct citations are consistently marked by higher parametric force (PFS) and greater use of the attention sink (BAS) for information synthesis. Crucially, we find that “one-size-fits-all” theories are insufficient as the signature of correctness evolves with scale: while the 3B model relies on high pathway alignment (PAS), our best-performing 8B detector identifies a shift toward a specialized strategy where pathways provide distinct, orthogonal information. By capturing this complex interplay, FACTUM outperforms state-of-the-art baselines by up to 37.5% in AUC. Our results demonstrate that high parametric force is constructive when successfully coordinated with the Attention pathway, paving the way for more nuanced and reliable RAG systems.

Method: Proposes iReasoner framework with Proposer-Solver loop over unlabeled images. Augments outcome-level intrinsic rewards with trajectory-aware signals defined over intermediate reasoning steps. Uses chain-of-thought elicitation and rewards internal agreement between reasoning paths to provide learning signals without supervision.

Result: Starting from Qwen2.5-VL-7B, iReasoner yields up to +2.1 points across diverse multimodal reasoning benchmarks under fully unsupervised post-training.

Conclusion: iReasoner serves as a starting point for reasoning-aware self-improvement in LMMs in purely unsupervised settings, demonstrating that explicit reasoning path evaluation can enhance model performance without external supervision.

Abstract: Recent work shows that large multimodal models (LMMs) can self-improve from unlabeled data via self-play and intrinsic feedback. Yet existing self-evolving frameworks mainly reward final outcomes, leaving intermediate reasoning weakly constrained despite its importance for visually grounded decision making. We propose iReasoner, a self-evolving framework that improves an LMM’s implicit reasoning by explicitly eliciting chain-of-thought (CoT) and rewarding its internal agreement. In a Proposer–Solver loop over unlabeled images, iReasoner augments outcome-level intrinsic rewards with a trajectory-aware signal defined over intermediate reasoning steps, providing learning signals that distinguish reasoning paths leading to the same answer without ground-truth labels or external judges. Starting from Qwen2.5-VL-7B, iReasoner yields up to $+2.1$ points across diverse multimodal reasoning benchmarks under fully unsupervised post-training. We hope this work serves as a starting point for reasoning-aware self-improvement in LMMs in purely unsupervised settings.

Method: Proposes EvoToken-DLM with evolving soft token distributions instead of hard masks, enabling progressive transition from masked states to discrete outputs, plus continuous trajectory supervision to align training with iterative probabilistic updates.

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

Conclusion: EvoToken-DLM’s soft token evolution approach with continuous supervision enables revisable decoding and better utilization of intermediate representations, leading to improved diffusion language modeling performance.

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

Method: DART uses dynamic rollout trees during training to discover valid tool-use opportunities by branching at promising positions, then employs tree-based process advantage estimation to identify and reinforce beneficial tool-integrated sub-trajectories.

Result: DART significantly outperforms existing methods on challenging benchmarks like AIME and GPQA-Diamond, successfully harmonizing tool execution with long CoT reasoning.

Conclusion: The framework enables spontaneous tool-use during long CoT reasoning without human annotation, addressing key limitations in tool-integrated reasoning for LLMs.

Abstract: Tool-Integrated Reasoning has emerged as a key paradigm to augment Large Language Models (LLMs) with computational capabilities, yet integrating tool-use into long Chain-of-Thought (long CoT) remains underexplored, largely due to the scarcity of training data and the challenge of integrating tool-use without compromising the model’s intrinsic long-chain reasoning. In this paper, we introduce DART (Discovery And Reinforcement of Tool-Integrated Reasoning Chains via Rollout Trees), a reinforcement learning framework that enables spontaneous tool-use during long CoT reasoning without human annotation. DART operates by constructing dynamic rollout trees during training to discover valid tool-use opportunities, branching out at promising positions to explore diverse tool-integrated trajectories. Subsequently, a tree-based process advantage estimation identifies and credits specific sub-trajectories where tool invocation positively contributes to the solution, effectively reinforcing these beneficial behaviors. Extensive experiments on challenging benchmarks like AIME and GPQA-Diamond demonstrate that DART significantly outperforms existing methods, successfully harmonizing tool execution with long CoT reasoning.

Method: QuantEval integrates three evaluation dimensions: knowledge-based QA, quantitative mathematical reasoning, and quantitative strategy coding with a CTA-style backtesting framework that executes model-generated strategies using financial performance metrics.

Result: Evaluation of state-of-the-art LLMs shows substantial gaps compared to human experts, particularly in reasoning and strategy coding. Fine-tuning and RL experiments on domain-aligned data demonstrate consistent improvements.

Conclusion: QuantEval provides a more realistic assessment of LLMs’ quantitative finance capabilities and aims to facilitate research and accelerate practical adoption in real-world trading workflows, with full deterministic backtesting configuration released for reproducibility.

Abstract: Large Language Models (LLMs) have shown strong capabilities across many domains, yet their evaluation in financial quantitative tasks remains fragmented and mostly limited to knowledge-centric question answering. We introduce QuantEval, a benchmark that evaluates LLMs across three essential dimensions of quantitative finance: knowledge-based QA, quantitative mathematical reasoning, and quantitative strategy coding. Unlike prior financial benchmarks, QuantEval integrates a CTA-style backtesting framework that executes model-generated strategies and evaluates them using financial performance metrics, enabling a more realistic assessment of quantitative coding ability. We evaluate some state-of-the-art open-source and proprietary LLMs and observe substantial gaps to human experts, particularly in reasoning and strategy coding. Finally, we conduct large-scale supervised fine-tuning and reinforcement learning experiments on domain-aligned data, demonstrating consistent improvements. We hope QuantEval will facilitate research on LLMs’ quantitative finance capabilities and accelerate their practical adoption in real-world trading workflows. We additionally release the full deterministic backtesting configuration (asset universe, cost model, and metric definitions) to ensure strict reproducibility.

Method: The technique embeds harmful content within cyberpunk narratives and prompts models to perform functional analysis inspired by Vladimir Propp’s morphology of folktales. By casting the task as structural decomposition, the attack induces models to reconstruct harmful procedures as legitimate narrative interpretation.

Result: Across 26 frontier models from nine providers, the attack achieved an average success rate of 71.3%, with no model family proving reliably robust. This builds on prior work with Adversarial Poetry to show structurally-grounded jailbreaks are a broad vulnerability class.

Conclusion: The space of culturally coded frames that can mediate harmful intent is vast and likely inexhaustible by pattern-matching defenses alone. The authors propose a mechanistic interpretability research agenda to investigate how narrative cues reshape model representations and whether models can learn to recognize harmful intent independently of surface form.

Abstract: Safety mechanisms in LLMs remain vulnerable to attacks that reframe harmful requests through culturally coded structures. We introduce Adversarial Tales, a jailbreak technique that embeds harmful content within cyberpunk narratives and prompts models to perform functional analysis inspired by Vladimir Propp’s morphology of folktales. By casting the task as structural decomposition, the attack induces models to reconstruct harmful procedures as legitimate narrative interpretation. Across 26 frontier models from nine providers, we observe an average attack success rate of 71.3%, with no model family proving reliably robust. Together with our prior work on Adversarial Poetry, these findings suggest that structurally-grounded jailbreaks constitute a broad vulnerability class rather than isolated techniques. The space of culturally coded frames that can mediate harmful intent is vast, likely inexhaustible by pattern-matching defenses alone. Understanding why these attacks succeed is therefore essential: we outline a mechanistic interpretability research agenda to investigate how narrative cues reshape model representations and whether models can learn to recognize harmful intent independently of surface form.

Method: Created Virām benchmark with 54 manually curated punctuation-ambiguous instances. Evaluated two strategies: 1) pipeline-based restore-then-translate approach, and 2) direct fine-tuning on punctuation-varied data. Also compared with current Large Language Models.

Result: Specialized fine-tuned models and pipeline systems significantly improve translation quality over standard baselines on the Virām benchmark. Current LLMs lag behind task-specific approaches in preserving meaning for punctuation-ambiguous text.

Conclusion: Task-specific approaches (fine-tuning and pipeline systems) are necessary for handling punctuation ambiguity in MT, especially for low-resource languages. Current LLMs need further research in this area. The Virām benchmark and code are publicly available.

Abstract: Punctuation plays a critical role in resolving semantic and structural ambiguity in written language. Machine Translation (MT) systems are now widely applied across diverse domains and languages, including many low-resource settings. In this work, we focus on Marathi, a low- to middle-resource language. We introduce Virām, the first diagnostic benchmark for assessing punctuation robustness in English-to-Marathi machine translation, consisting of 54 manually curated, punctuation-ambiguous instances. We evaluate two primary strategies for enhancing reliability: a pipeline-based restore-then-translate approach and direct fine-tuned on punctuation-varied data. Our results demonstrate that specialized fine-tuned models and pipeline systems significantly improve translation quality over standard baselines on the Virām benchmark. Qualitative analysis reveals that the original model may result in wrong translations leading to wrong interpretations, while fine-tuned models significantly improve overall reliability. Furthermore, we find that current Large Language Models (LLMs) lag behind these task-specific approaches in preserving meaning for punctuation-ambiguous text, thus necessitating further research in this area. The code and dataset is available at https://github.com/KaustubhShejole/Viram_Marathi.

Method: Late cross-attention architecture for long-context handling, using LLM as teacher for fine-grained supervision, distilled into student model via enriched distillation loss.

Result: Outperforms state-of-the-art baselines on relevance, ranking, and calibration metrics, producing consistent and interpretable skill-fit scores.

Conclusion: The proposed approach effectively addresses long-context and bias challenges in person-job matching through architectural innovation and LLM-guided distillation.

Abstract: Finding the most relevant person for a job proposal in real time is challenging, especially when resumes are long, structured, and multilingual. In this paper, we propose a re-ranking model based on a new generation of late cross-attention architecture, that decomposes both resumes and project briefs to efficiently handle long-context inputs with minimal computational overhead. To mitigate historical data biases, we use a generative large language model (LLM) as a teacher, generating fine-grained, semantically grounded supervision. This signal is distilled into our student model via an enriched distillation loss function. The resulting model produces skill-fit scores that enable consistent and interpretable person-job matching. Experiments on relevance, ranking, and calibration metrics demonstrate that our approach outperforms state-of-the-art baselines.

Method: Created OctoBench with 34 environments and 217 tasks across three scaffold types, paired with 7,098 objective checklist items. Developed automated observation-and-scoring toolkit to capture full trajectories and perform fine-grained checks, disentangling task-solving from rule-following.

Result: Experiments on eight representative models reveal a systematic gap between task-solving and scaffold-aware compliance, showing that models can solve tasks but struggle to follow heterogeneous scaffold instructions.

Conclusion: There’s a need for training and evaluation that explicitly targets heterogeneous instruction following. The benchmark is released to support reproducible benchmarking and accelerate development of more scaffold-aware coding agents.

Abstract: Modern coding scaffolds turn LLMs into capable software agents, but their ability to follow scaffold-specified instructions remains under-examined, especially when constraints are heterogeneous and persist across interactions. To fill this gap, we introduce OctoBench, which benchmarks scaffold-aware instruction following in repository-grounded agentic coding. OctoBench includes 34 environments and 217 tasks instantiated under three scaffold types, and is paired with 7,098 objective checklist items. To disentangle solving the task from following the rules, we provide an automated observation-and-scoring toolkit that captures full trajectories and performs fine-grained checks. Experiments on eight representative models reveal a systematic gap between task-solving and scaffold-aware compliance, underscoring the need for training and evaluation that explicitly targets heterogeneous instruction following. We release the benchmark to support reproducible benchmarking and to accelerate the development of more scaffold-aware coding agents.

Method: Psychology-grounded Empathetic Reward Modeling (PERM) operationalizes empathy evaluation through bidirectional decomposition: 1) Supporter perspective (internal resonation + communicative expression), 2) Seeker perspective (emotional reception), plus 3) Bystander perspective for overall interaction quality monitoring.

Result: PERM outperforms state-of-the-art baselines by over 10% on emotional intelligence benchmarks and industrial daily conversation datasets. A blinded user study shows 70% preference for PERM-generated responses.

Conclusion: PERM effectively enhances LLM empathy by incorporating psychology-grounded bidirectional evaluation, demonstrating significant improvements in emotional support capabilities with strong user preference.

Abstract: Large Language Models (LLMs) are increasingly deployed in human-centric applications, yet they often fail to provide substantive emotional support. While Reinforcement Learning (RL) has been utilized to enhance empathy of LLMs, existing reward models typically evaluate empathy from a single perspective, overlooking the inherently bidirectional interaction nature of empathy between the supporter and seeker as defined by Empathy Cycle theory. To address this limitation, we propose Psychology-grounded Empathetic Reward Modeling (PERM). PERM operationalizes empathy evaluation through a bidirectional decomposition: 1) Supporter perspective, assessing internal resonation and communicative expression; 2) Seeker perspective, evaluating emotional reception. Additionally, it incorporates a bystander perspective to monitor overall interaction quality. Extensive experiments on a widely-used emotional intelligence benchmark and an industrial daily conversation dataset demonstrate that PERM outperforms state-of-the-art baselines by over 10%. Furthermore, a blinded user study reveals a 70% preference for our approach, highlighting its efficacy in generating more empathetic responses. Our code, dataset, and models are available at https://github.com/ZhengWwwq/PERM.

Method: FOFPred uses a unified Vision-Language Model (VLM) and Diffusion architecture for multimodal reasoning with pixel-level generative fidelity. It’s trained on web-scale human activity data using data preprocessing techniques and strong image pretraining to extract meaningful signals from noisy video-caption data.

Result: The trained model is extended to tackle robotic manipulation and video generation tasks under language-driven settings, demonstrating cross-domain versatility and confirming the value of the unified VLM-Diffusion architecture.

Conclusion: The paper establishes that unified VLM-Diffusion architectures and scalable learning from diverse web data are effective for future optical flow prediction, enabling applications in both control and generation domains.

Abstract: Future motion representations, such as optical flow, offer immense value for control and generative tasks. However, forecasting generalizable spatially dense motion representations remains a key challenge, and learning such forecasting from noisy, real-world data remains relatively unexplored. We introduce FOFPred, a novel language-conditioned optical flow forecasting model featuring a unified Vision-Language Model (VLM) and Diffusion architecture. This unique combination enables strong multimodal reasoning with pixel-level generative fidelity for future motion prediction. Our model is trained on web-scale human activity data-a highly scalable but unstructured source. To extract meaningful signals from this noisy video-caption data, we employ crucial data preprocessing techniques and our unified architecture with strong image pretraining. The resulting trained model is then extended to tackle two distinct downstream tasks in control and generation. Evaluations across robotic manipulation and video generation under language-driven settings establish the cross-domain versatility of FOFPred, confirming the value of a unified VLM-Diffusion architecture and scalable learning from diverse web data for future optical flow prediction.

Method: Introduced ICONIC-444 dataset containing over 3.1 million RGB images spanning 444 classes captured with a prototype industrial sorting machine. Defined four reference tasks within the dataset to benchmark OOD detection research and provided baseline results for 22 state-of-the-art post-hoc OOD detection methods.

Result: Created a specialized large-scale industrial image dataset that closely mimics real-world tasks and complements existing datasets by offering structured, diverse data suited for rigorous OOD evaluation across a spectrum of task complexities.

Conclusion: ICONIC-444 addresses critical limitations in OOD detection research by providing a comprehensive benchmark dataset with defined tasks and baseline results, enabling more rigorous evaluation and advancement of OOD detection methods.

Abstract: Current progress in out-of-distribution (OOD) detection is limited by the lack of large, high-quality datasets with clearly defined OOD categories across varying difficulty levels (near- to far-OOD) that support both fine- and coarse-grained computer vision tasks. To address this limitation, we introduce ICONIC-444 (Image Classification and OOD Detection with Numerous Intricate Complexities), a specialized large-scale industrial image dataset containing over 3.1 million RGB images spanning 444 classes tailored for OOD detection research. Captured with a prototype industrial sorting machine, ICONIC-444 closely mimics real-world tasks. It complements existing datasets by offering structured, diverse data suited for rigorous OOD evaluation across a spectrum of task complexities. We define four reference tasks within ICONIC-444 to benchmark and advance OOD detection research and provide baseline results for 22 state-of-the-art post-hoc OOD detection methods.

Method: Adapted Sparse4D framework with absolute world-coordinate geometric priors, occlusion-aware ReID embedding module, and generative data augmentation using NVIDIA COSMOS framework for Sim2Real domain adaptation.

Result: Achieves state-of-the-art HOTA of 45.22 on AI City Challenge 2025 benchmark, with 2.15× speedup via optimized TensorRT plugin for Multi-Scale Deformable Aggregation.

Conclusion: The camera-only framework enables real-time deployment supporting over 64 concurrent camera streams on a single Blackwell-class GPU, addressing industrial infrastructure needs.

Abstract: Accurate 3D object perception and multi-target multi-camera (MTMC) tracking are fundamental for the digital transformation of industrial infrastructure. However, transitioning “inside-out” autonomous driving models to “outside-in” static camera networks presents significant challenges due to heterogeneous camera placements and extreme occlusion. In this paper, we present an adapted Sparse4D framework specifically optimized for large-scale infrastructure environments. Our system leverages absolute world-coordinate geometric priors and introduces an occlusion-aware ReID embedding module to maintain identity stability across distributed sensor networks. To bridge the Sim2Real domain gap without manual labeling, we employ a generative data augmentation strategy using the NVIDIA COSMOS framework, creating diverse environmental styles that enhance the model’s appearance-invariance. Evaluated on the AI City Challenge 2025 benchmark, our camera-only framework achieves a state-of-the-art HOTA of $45.22$. Furthermore, we address real-time deployment constraints by developing an optimized TensorRT plugin for Multi-Scale Deformable Aggregation (MSDA). Our hardware-accelerated implementation achieves a $2.15\times$ speedup on modern GPU architectures, enabling a single Blackwell-class GPU to support over 64 concurrent camera streams.

Method: Evaluated three leading VLMs (GPT-4o, Florence 2, LLaVa-1.5) using a curated dataset of 1,000 images annotated across ten action and ten emotion categories. Assessed each model through standardized inference pipelines and multiple evaluation metrics including F1-scores and accuracy, with confusion matrix analysis.

Result: GPT-4o achieved highest performance: average F1-score 0.756 and accuracy 0.799 in action recognition, and F1-score 0.712 and accuracy 0.773 in emotion recognition. Florence 2 performed moderately (F1: 0.497 action, 0.414 emotion). LLaVa-1.5 showed lowest performance (F1: 0.466 action, 0.461 emotion). All models struggled with semantically close categories like collaborating vs. communicating.

Conclusion: General-purpose VLMs offer baseline capability for human behavior recognition in construction, but need improvements like domain adaptation, temporal modeling, or multimodal sensing for real-world reliability, as models struggle with fine-grained distinctions between similar categories.

Abstract: As robotics become increasingly integrated into construction workflows, their ability to interpret and respond to human behavior will be essential for enabling safe and effective collaboration. Vision-Language Models (VLMs) have emerged as a promising tool for visual understanding tasks and offer the potential to recognize human behaviors without extensive domain-specific training. This capability makes them particularly appealing in the construction domain, where labeled data is scarce and monitoring worker actions and emotional states is critical for safety and productivity. In this study, we evaluate the performance of three leading VLMs, GPT-4o, Florence 2, and LLaVa-1.5, in detecting construction worker actions and emotions from static site images. Using a curated dataset of 1,000 images annotated across ten action and ten emotion categories, we assess each model’s outputs through standardized inference pipelines and multiple evaluation metrics. GPT-4o consistently achieved the highest scores across both tasks, with an average F1-score of 0.756 and accuracy of 0.799 in action recognition, and an F1-score of 0.712 and accuracy of 0.773 in emotion recognition. Florence 2 performed moderately, with F1-scores of 0.497 for action and 0.414 for emotion, while LLaVa-1.5 showed the lowest overall performance, with F1-scores of 0.466 for action and 0.461 for emotion. Confusion matrix analyses revealed that all models struggled to distinguish semantically close categories, such as collaborating in teams versus communicating with supervisors. While the results indicate that general-purpose VLMs can offer a baseline capability for human behavior recognition in construction environments, further improvements, such as domain adaptation, temporal modeling, or multimodal sensing, may be needed for real-world reliability.

Method: Comprehensive empirical study using ResNet-50 architecture, benchmarking 21 post-hoc OOD detection methods across 56 ImageNet-trained models from diverse training strategies, evaluated on eight OOD test sets.

Result: Found non-monotonic relationship between ID accuracy and OOD detection performance - OOD performance initially improves with accuracy but declines when advanced training recipes push accuracy beyond baseline. Strong interdependence between training strategy, detector choice, and OOD performance, with no single method universally optimal.

Conclusion: The common assumption that higher ID accuracy implies better OOD detection is incorrect; optimal OOD detection requires considering the interplay between training strategy and detector choice rather than relying on a single universal method.

Abstract: Out-of-distribution (OOD) detection is crucial for deploying robust and reliable machine-learning systems in open-world settings. Despite steady advances in OOD detectors, their interplay with modern training pipelines that maximize in-distribution (ID) accuracy and generalization remains under-explored. We investigate this link through a comprehensive empirical study. Fixing the architecture to the widely adopted ResNet-50, we benchmark 21 post-hoc, state-of-the-art OOD detection methods across 56 ImageNet-trained models obtained via diverse training strategies and evaluate them on eight OOD test sets. Contrary to the common assumption that higher ID accuracy implies better OOD detection performance, we uncover a non-monotonic relationship: OOD performance initially improves with accuracy but declines once advanced training recipes push accuracy beyond the baseline. Moreover, we observe a strong interdependence between training strategy, detector choice, and resulting OOD performance, indicating that no single method is universally optimal.

Method: Created modified versions of three 3D ResNet models (MC3, R3D, R(2+1)D) with dropout before final classifier, then developed 10 variants for each with different attention blocks (CBAM, TCN, multi-headed attention, channel attention). Tested on UCF101 dataset.

Result: Best accuracy of 88.98% achieved by modified R(2+1)D with multi-headed attention. Variants showed different class-level accuracy behaviors despite similar overall performance improvements. Results demonstrate significance of missing temporal features in high-resolution models.

Conclusion: Temporal features are crucial for action recognition performance even when spatial resolution is increased. Attention mechanisms can partially compensate for reduced temporal information, with multi-headed attention showing particular effectiveness.

Abstract: Human action recognition has become an important research focus in computer vision due to the wide range of applications where it is used. 3D Resnet-based CNN models, particularly MC3, R3D, and R(2+1)D, have different convolutional filters to extract spatiotemporal features. This paper investigates the impact of reducing the captured knowledge from temporal data, while increasing the resolution of the frames. To establish this experiment, we created similar designs to the three originals, but with a dropout layer added before the final classifier. Secondly, we then developed ten new versions for each one of these three designs. The variants include special attention blocks within their architecture, such as convolutional block attention module (CBAM), temporal convolution networks (TCN), in addition to multi-headed and channel attention mechanisms. The purpose behind that is to observe the extent of the influence each of these blocks has on performance for the restricted-temporal models. The results of testing all the models on UCF101 have shown accuracy of 88.98% for the variant with multiheaded attention added to the modified R(2+1)D. This paper concludes the significance of missing temporal features in the performance of the newly created increased resolution models. The variants had different behavior on class-level accuracy, despite the similarity of their enhancements to the overall performance.

Method: A cascading multi-agent framework with early modules for reconstruction-gated filtering and object-level assessment, plus higher-level reasoning agents selectively invoked for ambiguous events. Uses adaptive escalation thresholds and publish-subscribe communication for asynchronous coordination across heterogeneous hardware.

Result: Achieves 3x latency reduction compared to direct vision-language inference while maintaining high perceptual fidelity (PSNR = 38.3 dB, SSIM = 0.965) and consistent semantic labeling on large-scale monitoring data.

Conclusion: The framework advances anomaly detection by combining early-exit efficiency, adaptive multi-agent reasoning, and explainable anomaly attribution, establishing a reproducible and energy-efficient foundation for scalable intelligent visual monitoring.

Abstract: Intelligent anomaly detection in dynamic visual environments requires reconciling real-time performance with semantic interpretability. Conventional approaches address only fragments of this challenge. Reconstruction-based models capture low-level deviations without contextual reasoning, object detectors provide speed but limited semantics, and large vision-language systems deliver interpretability at prohibitive computational cost. This work introduces a cascading multi-agent framework that unifies these complementary paradigms into a coherent and interpretable architecture. Early modules perform reconstruction-gated filtering and object-level assessment, while higher-level reasoning agents are selectively invoked to interpret semantically ambiguous events. The system employs adaptive escalation thresholds and a publish-subscribe communication backbone, enabling asynchronous coordination and scalable deployment across heterogeneous hardware. Extensive evaluation on large-scale monitoring data demonstrates that the proposed cascade achieves a threefold reduction in latency compared to direct vision-language inference, while maintaining high perceptual fidelity (PSNR = 38.3 dB, SSIM = 0.965) and consistent semantic labeling. The framework advances beyond conventional detection pipelines by combining early-exit efficiency, adaptive multi-agent reasoning, and explainable anomaly attribution, establishing a reproducible and energy-efficient foundation for scalable intelligent visual monitoring.

Method: Full fine-tuning of SAM3 on large-scale, heterogeneous 2D and 3D medical imaging datasets (33 datasets spanning 10 modalities) with paired segmentation masks and text prompts, rather than just prompt engineering.

Result: Medical SAM3 achieves consistent and significant performance gains across organs, imaging modalities, and dimensionalities, especially in challenging scenarios with semantic ambiguity, complex morphology, and long-range 3D context.

Conclusion: Medical SAM3 establishes as a universal, text-guided segmentation foundation model for medical imaging, demonstrating that holistic model adaptation (beyond prompt engineering) is crucial for robust prompt-driven segmentation under severe domain shift.

Abstract: Promptable segmentation foundation models such as SAM3 have demonstrated strong generalization capabilities through interactive and concept-based prompting. However, their direct applicability to medical image segmentation remains limited by severe domain shifts, the absence of privileged spatial prompts, and the need to reason over complex anatomical and volumetric structures. Here we present Medical SAM3, a foundation model for universal prompt-driven medical image segmentation, obtained by fully fine-tuning SAM3 on large-scale, heterogeneous 2D and 3D medical imaging datasets with paired segmentation masks and text prompts. Through a systematic analysis of vanilla SAM3, we observe that its performance degrades substantially on medical data, with its apparent competitiveness largely relying on strong geometric priors such as ground-truth-derived bounding boxes. These findings motivate full model adaptation beyond prompt engineering alone. By fine-tuning SAM3’s model parameters on 33 datasets spanning 10 medical imaging modalities, Medical SAM3 acquires robust domain-specific representations while preserving prompt-driven flexibility. Extensive experiments across organs, imaging modalities, and dimensionalities demonstrate consistent and significant performance gains, particularly in challenging scenarios characterized by semantic ambiguity, complex morphology, and long-range 3D context. Our results establish Medical SAM3 as a universal, text-guided segmentation foundation model for medical imaging and highlight the importance of holistic model adaptation for achieving robust prompt-driven segmentation under severe domain shift. Code and model will be made available at https://github.com/AIM-Research-Lab/Medical-SAM3.

Method: Constructed high-quality motion dataset with atomic, temporally-aware part-level annotations using LLMs, then developed diffusion-based framework where each body part is guided by its own temporally-structured textual prompt.

Result: FrankenMotion outperforms all previous baseline models adapted for this setting and can compose motions unseen during training.

Conclusion: First work to provide atomic, temporally-aware part-level motion annotations and enable motion generation with both spatial (body part) and temporal (atomic action) control.

Abstract: Human motion generation from text prompts has made remarkable progress in recent years. However, existing methods primarily rely on either sequence-level or action-level descriptions due to the absence of fine-grained, part-level motion annotations. This limits their controllability over individual body parts. In this work, we construct a high-quality motion dataset with atomic, temporally-aware part-level text annotations, leveraging the reasoning capabilities of large language models (LLMs). Unlike prior datasets that either provide synchronized part captions with fixed time segments or rely solely on global sequence labels, our dataset captures asynchronous and semantically distinct part movements at fine temporal resolution. Based on this dataset, we introduce a diffusion-based part-aware motion generation framework, namely FrankenMotion, where each body part is guided by its own temporally-structured textual prompt. This is, to our knowledge, the first work to provide atomic, temporally-aware part-level motion annotations and have a model that allows motion generation with both spatial (body part) and temporal (atomic action) control. Experiments demonstrate that FrankenMotion outperforms all previous baseline models adapted and retrained for our setting, and our model can compose motions unseen during training. Our code and dataset will be publicly available upon publication.

Method: The authors propose a novel methodology that increases resolution of synthetic aperture radar imagery by optimizing loss functions that account for both image quality and classification performance. They use a specialized algorithmic strategy that jointly optimizes for SR and classification objectives.

Result: The approach improves image quality as measured by scientifically ascertained image quality indicators while also enhancing classification accuracy, demonstrating that integrating classification objectives into SR benefits both tasks.

Conclusion: Integrating classification objectives directly into the super-resolution process can improve both image quality and classification performance, addressing the gap between traditional pixel-level SR optimization and downstream task performance.

Abstract: High-resolution imagery plays a critical role in improving the performance of visual recognition tasks such as classification, detection, and segmentation. In many domains, including remote sensing and surveillance, low-resolution images can limit the accuracy of automated analysis. To address this, super-resolution (SR) techniques have been widely adopted to attempt to reconstruct high-resolution images from low-resolution inputs. Related traditional approaches focus solely on enhancing image quality based on pixel-level metrics, leaving the relationship between super-resolved image fidelity and downstream classification performance largely underexplored. This raises a key question: can integrating classification objectives directly into the super-resolution process further improve classification accuracy? In this paper, we try to respond to this question by investigating the relationship between super-resolution and classification through the deployment of a specialised algorithmic strategy. We propose a novel methodology that increases the resolution of synthetic aperture radar imagery by optimising loss functions that account for both image quality and classification performance. Our approach improves image quality, as measured by scientifically ascertained image quality indicators, while also enhancing classification accuracy.

Method: The study implements multiple classification methods including DenseNet121, develops an open-source web-based application for deployment, and conducts comparative testing to evaluate performance across different approaches.

Result: The paper presents comparative results of different classification methods, identifies weaknesses in the proposed approaches, and provides an accessible web application for chest X-ray multi-classification.

Conclusion: The study offers insights into effective methods for chest X-ray multi-classification, highlights areas for improvement in current approaches, and provides practical deployment tools with open-source code for further research and application.

Abstract: Multi-Classification Chest X-Ray Images are one of the most prevalent forms of radiological examination used for diagnosing thoracic diseases. In this study, we offer a concise overview of several methods employed for tackling this task, including DenseNet121. In addition, we deploy an open-source web-based application. In our study, we conduct tests to compare different methods and see how well they work. We also look closely at the weaknesses of the methods we propose and suggest ideas for making them better in the future. Our code is available at: https://github.com/AML4206-MINE20242/Proyecto_AML

Method: Uses SSL-guided Latent Diffusion Model with DINO teacher embeddings to generate synthetic training patches, then combines real/synthetic data to fine-tune Vision Transformer with patch aggregation for WSI-level classification.

Result: Achieves 96.47% accuracy and reduces FID score to 45.72, significantly outperforming class-conditioned baselines in 5-fold cross-validation.

Conclusion: The proposed SSL-guided LDM effectively addresses data scarcity in DUV-FSM imaging, enabling accurate breast cancer margin assessment for BCS.

Abstract: Breast-Conserving Surgery (BCS) requires precise intraoperative margin assessment to preserve healthy tissue. Deep Ultraviolet Fluorescence Scanning Microscopy (DUV-FSM) offers rapid, high-resolution surface imaging for this purpose; however, the scarcity of annotated DUV data hinders the training of robust deep learning models. To address this, we propose an Self-Supervised Learning (SSL)-guided Latent Diffusion Model (LDM) to generate high-quality synthetic training patches. By guiding the LDM with embeddings from a fine-tuned DINO teacher, we inject rich semantic details of cellular structures into the synthetic data. We combine real and synthetic patches to fine-tune a Vision Transformer (ViT), utilizing patch prediction aggregation for WSI-level classification. Experiments using 5-fold cross-validation demonstrate that our method achieves 96.47 % accuracy and reduces the FID score to 45.72, significantly outperforming class-conditioned baselines.

Method: Introduces RobuMTL with adaptive selection of task-specific hierarchical Low-Rank Adaptation (LoRA) modules and LoRA expert squads based on input perturbations, using a mixture-of-experts approach to specialize for different weather conditions.

Result: On PASCAL: +2.8% average relative improvement under single perturbations, up to +44.4% under mixed weather conditions vs MTL baseline. On NYUD-v2: +9.7% average relative improvement across tasks.

Conclusion: RobuMTL effectively addresses visual degradation in adverse weather through adaptive specialization, demonstrating superior robustness compared to single-task models, standard MTL baselines, and state-of-the-art methods.

Abstract: Robust Multi-Task Learning (MTL) is crucial for autonomous systems operating in real-world environments, where adverse weather conditions can severely degrade model performance and reliability. In this paper, we introduce RobuMTL, a novel architecture designed to adaptively address visual degradation by dynamically selecting task-specific hierarchical Low-Rank Adaptation (LoRA) modules and a LoRA expert squad based on input perturbations in a mixture-of-experts fashion. Our framework enables adaptive specialization based on input characteristics, improving robustness across diverse real-world conditions. To validate our approach, we evaluated it on the PASCAL and NYUD-v2 datasets and compared it against single-task models, standard MTL baselines, and state-of-the-art methods. On the PASCAL benchmark, RobuMTL delivers a +2.8% average relative improvement under single perturbations and up to +44.4% under mixed weather conditions compared to the MTL baseline. On NYUD-v2, RobuMTL achieves a +9.7% average relative improvement across tasks. The code is available at GitHub.

Method: Evaluated five representative architectures on the small, imbalanced dataset: YOLOv11, Mask R-CNN, DeepLabv3, Swin-UNET, and DINOv2. Analyzed training strategies, augmentation policies, and model behavior under data constraints.

Result: Pretrained CNN-based models (YOLOv11 and Mask R-CNN) generalized significantly better than transformer-based models. DeepLabv3, Swin-UNET and DINOv2 underperformed due to task differences, high data requirements of transformers, and lack of strong inductive biases.

Conclusion: Transformer-based architectures struggle in low-data regimes without substantial pretraining or augmentation. Lightweight CNN-based methods remain most reliable for canopy detection with limited imagery, confirming the importance of inductive biases in data-scarce scenarios.

Abstract: Tree canopy detection from aerial imagery is an important task for environmental monitoring, urban planning, and ecosystem analysis. Simulating real-life data annotation scarcity, the Solafune Tree Canopy Detection competition provides a small and imbalanced dataset of only 150 annotated images, posing significant challenges for training deep models without severe overfitting. In this work, we evaluate five representative architectures, YOLOv11, Mask R-CNN, DeepLabv3, Swin-UNet, and DINOv2, to assess their suitability for canopy segmentation under extreme data scarcity. Our experiments show that pretrained convolution-based models, particularly YOLOv11 and Mask R-CNN, generalize significantly better than pretrained transformer-based models. DeeplabV3, Swin-UNet and DINOv2 underperform likely due to differences between semantic and instance segmentation tasks, the high data requirements of Vision Transformers, and the lack of strong inductive biases. These findings confirm that transformer-based architectures struggle in low-data regimes without substantial pretraining or augmentation and that differences between semantic and instance segmentation further affect model performance. We provide a detailed analysis of training strategies, augmentation policies, and model behavior under the small-data constraint and demonstrate that lightweight CNN-based methods remain the most reliable for canopy detection on limited imagery.

Method: Proposed PCDF with two VLMs: DocVLM generates follow-up questions based on images and dialogue history, while PatientVLM responds using symptom profiles from ground-truth diagnoses. The synthetic dialogues are clinically validated and used to fine-tune DocVLM.

Result: Clinical validation confirmed synthetic symptoms are clinically relevant, comprehensive, and realistic. Dialogue-based supervision substantially outperforms image-only training, demonstrating the value of realistic symptom elicitation for diagnosis.

Conclusion: Simulating doctor-patient diagnostic dialogues through VLMs creates coherent multi-turn consultations that significantly improve diagnostic AI performance, highlighting the importance of incorporating symptom elicitation beyond just image analysis.

Abstract: Traditionally, AI research in medical diagnosis has largely centered on image analysis. While this has led to notable advancements, the absence of patient-reported symptoms continues to hinder diagnostic accuracy. To address this, we propose a Pre-Consultation Dialogue Framework (PCDF) that mimics real-world diagnostic procedures, where doctors iteratively query patients before reaching a conclusion. Specifically, we simulate diagnostic dialogues between two vision-language models (VLMs): a DocVLM, which generates follow-up questions based on the image and dialogue history, and a PatientVLM, which responds using a symptom profile derived from the ground-truth diagnosis. We additionally conducted a small-scale clinical validation of the synthetic symptoms generated by our framework, with licensed clinicians confirming their clinical relevance, symptom coverage, and overall realism. These findings indicate that the resulting DocVLM-PatientVLM interactions form coherent, multi-turn consultations paired with images and diagnoses, which we then use to fine-tune the DocVLM. This dialogue-based supervision leads to substantial gains over image-only training, highlighting the value of realistic symptom elicitation for diagnosis.

Method: 1) Construct MMedExpert dataset with 10K samples across four specialties featuring step-by-step reasoning traces. 2) Domain-Specific Adaptation creates specialty-specific LoRA modules for diverse initialization. 3) Guideline-Based Advantages explicitly models different clinical reasoning perspectives. 4) Conflict-Aware Capability Integration merges specialized experts into a unified agent.

Result: State-of-the-art performance with 7B model achieving 27.50 on MedXpert-MM and 83.03 on OmniMedVQA, establishing robust foundation for reliable multimodal medical reasoning systems.

Conclusion: MMedExpert-R1 successfully addresses critical mismatches in medical reasoning by combining domain-specific adaptation with clinical guideline reinforcement, enabling robust multi-specialty alignment and reliable clinical reasoning capabilities.

Abstract: Medical Vision-Language Models (MedVLMs) excel at perception tasks but struggle with complex clinical reasoning required in real-world scenarios. While reinforcement learning (RL) has been explored to enhance reasoning capabilities, existing approaches face critical mismatches: the scarcity of deep reasoning data, cold-start limits multi-specialty alignment, and standard RL algorithms fail to model clinical reasoning diversity. We propose MMedExpert-R1, a novel reasoning MedVLM that addresses these challenges through domain-specific adaptation and clinical guideline reinforcement. We construct MMedExpert, a high-quality dataset of 10K samples across four specialties with step-by-step reasoning traces. Our Domain-Specific Adaptation (DSA) creates specialty-specific LoRA modules to provide diverse initialization, while Guideline-Based Advantages (GBA) explicitly models different clinical reasoning perspectives to align with real-world diagnostic strategies. Conflict-Aware Capability Integration then merges these specialized experts into a unified agent, ensuring robust multi-specialty alignment. Comprehensive experiments demonstrate state-of-the-art performance, with our 7B model achieving 27.50 on MedXpert-MM and 83.03 on OmniMedVQA, establishing a robust foundation for reliable multimodal medical reasoning systems.

Method: Combines implicit 3D representations with 2D detectors. Uses LPIPS loss and multi-view compensation loss (MVCL) to optimize rendering from corrupted images. End-to-end training aggregates information from multiple corrupted views to synthesize clean 3D scenes.

Result: Effectively removes multiple distractor types (snowflakes, confetti, defoliation, petals). Achieves comparable performance to state-of-the-art specialized desnow methods. Works on both realistic and synthetic distractors. Validated through extensive experiments on new benchmark dataset.

Conclusion: IDDR-NGP is the first unified distractor removal method for Instant-NGP that handles diverse distractor types effectively, demonstrating robustness through comprehensive evaluation on a new benchmark dataset.

Abstract: This paper presents the first unified distractor removal method, named IDDR-NGP, which directly operates on Instant-NPG. The method is able to remove a wide range of distractors in 3D scenes, such as snowflakes, confetti, defoliation and petals, whereas existing methods usually focus on a specific type of distractors. By incorporating implicit 3D representations with 2D detectors, we demonstrate that it is possible to efficiently restore 3D scenes from multiple corrupted images. We design the learned perceptual image patch similarity~( LPIPS) loss and the multi-view compensation loss (MVCL) to jointly optimize the rendering results of IDDR-NGP, which could aggregate information from multi-view corrupted images. All of them can be trained in an end-to-end manner to synthesize high-quality 3D scenes. To support the research on distractors removal in implicit 3D representations, we build a new benchmark dataset that consists of both synthetic and real-world distractors. To validate the effectiveness and robustness of IDDR-NGP, we provide a wide range of distractors with corresponding annotated labels added to both realistic and synthetic scenes. Extensive experimental results demonstrate the effectiveness and robustness of IDDR-NGP in removing multiple types of distractors. In addition, our approach achieves results comparable with the existing SOTA desnow methods and is capable of accurately removing both realistic and synthetic distractors.

Method: Created AIGVDBench benchmark covering 31 state-of-the-art generation models, over 440,000 videos, and conducted 1,500+ evaluations on 33 existing detectors across four distinct categories.

Result: The benchmark provides 8 in-depth analyses from multiple perspectives and identifies 4 novel findings that offer valuable insights for future research in AI-generated video detection.

Conclusion: AIGVDBench addresses critical gaps in the field by providing a comprehensive, representative foundation for advancing AI-generated video detection research, with open-source availability.

Abstract: Recent advances in generative modeling can create remarkably realistic synthetic videos, making it increasingly difficult for humans to distinguish them from real ones and necessitating reliable detection methods. However, two key limitations hinder the development of this field. \textbf{From the dataset perspective}, existing datasets are often limited in scale and constructed using outdated or narrowly scoped generative models, making it difficult to capture the diversity and rapid evolution of modern generative techniques. Moreover, the dataset construction process frequently prioritizes quantity over quality, neglecting essential aspects such as semantic diversity, scenario coverage, and technological representativeness. \textbf{From the benchmark perspective}, current benchmarks largely remain at the stage of dataset creation, leaving many fundamental issues and in-depth analysis yet to be systematically explored. Addressing this gap, we propose AIGVDBench, a benchmark designed to be comprehensive and representative, covering \textbf{31} state-of-the-art generation models and over \textbf{440,000} videos. By executing more than \textbf{1,500} evaluations on \textbf{33} existing detectors belonging to four distinct categories. This work presents \textbf{8 in-depth analyses} from multiple perspectives and identifies \textbf{4 novel findings} that offer valuable insights for future research. We hope this work provides a solid foundation for advancing the field of AI-generated video detection. Our benchmark is open-sourced at https://github.com/LongMa-2025/AIGVDBench.

Method: Propose M3DDM+ which applies uniform mask direction and width across all frames during training (instead of random masks), followed by fine-tuning of the pretrained M3DDM model to align training with inference requirements.

Result: M3DDM+ substantially improves visual fidelity and temporal coherence in information-limited scenarios while maintaining computational efficiency of the original framework.

Conclusion: Aligning training and inference masking strategies through uniform mask application and fine-tuning effectively addresses quality degradation in challenging video outpainting scenarios, resulting in improved performance while preserving computational efficiency.

Abstract: M3DDM provides a computationally efficient framework for video outpainting via latent diffusion modeling. However, it exhibits significant quality degradation – manifested as spatial blur and temporal inconsistency – under challenging scenarios characterized by limited camera motion or large outpainting regions, where inter-frame information is limited. We identify the cause as a training-inference mismatch in the masking strategy: M3DDM’s training applies random mask directions and widths across frames, whereas inference requires consistent directional outpainting throughout the video. To address this, we propose M3DDM+, which applies uniform mask direction and width across all frames during training, followed by fine-tuning of the pretrained M3DDM model. Experiments demonstrate that M3DDM+ substantially improves visual fidelity and temporal coherence in information-limited scenarios while maintaining computational efficiency. The code is available at https://github.com/tamaki-lab/M3DDM-Plus.

Method: The authors introduce a physics-aware reinforcement learning paradigm that enforces physical collision rules directly in high-dimensional spaces. They extend this to a unified framework called Mimicry-Discovery Cycle (MDcycle) that allows substantial fine-tuning while preserving the model’s ability to leverage physics-grounded feedback.

Result: The approach is validated through a new benchmark called PhysRVGBench with extensive qualitative and quantitative experiments demonstrating its effectiveness in improving physical realism of generated videos.

Conclusion: The paper presents a novel approach to incorporating physical principles into video generation through reinforcement learning, addressing a critical limitation in current transformer-based models and enabling more realistic simulation of rigid body motion and collisions.

Abstract: Physical principles are fundamental to realistic visual simulation, but remain a significant oversight in transformer-based video generation. This gap highlights a critical limitation in rendering rigid body motion, a core tenet of classical mechanics. While computer graphics and physics-based simulators can easily model such collisions using Newton formulas, modern pretrain-finetune paradigms discard the concept of object rigidity during pixel-level global denoising. Even perfectly correct mathematical constraints are treated as suboptimal solutions (i.e., conditions) during model optimization in post-training, fundamentally limiting the physical realism of generated videos. Motivated by these considerations, we introduce, for the first time, a physics-aware reinforcement learning paradigm for video generation models that enforces physical collision rules directly in high-dimensional spaces, ensuring the physics knowledge is strictly applied rather than treated as conditions. Subsequently, we extend this paradigm to a unified framework, termed Mimicry-Discovery Cycle (MDcycle), which allows substantial fine-tuning while fully preserving the model’s ability to leverage physics-grounded feedback. To validate our approach, we construct new benchmark PhysRVGBench and perform extensive qualitative and quantitative experiments to thoroughly assess its effectiveness.

Method: Proposes CoDance with Unbind-Rebind framework: (1) Unbind module uses pose shift encoder with stochastic perturbations to break rigid spatial binding and learn location-agnostic motion; (2) Rebind module uses text prompts and subject masks for semantic and spatial guidance to direct motion to intended characters.

Result: Achieves state-of-the-art performance on new CoDanceBench and existing datasets, showing remarkable generalization across diverse subjects and spatial layouts.

Conclusion: CoDance effectively addresses limitations of existing methods by decoupling motion from spatial constraints and enabling precise control over multi-subject animation, with code and weights to be open-sourced.

Abstract: Character image animation is gaining significant importance across various domains, driven by the demand for robust and flexible multi-subject rendering. While existing methods excel in single-person animation, they struggle to handle arbitrary subject counts, diverse character types, and spatial misalignment between the reference image and the driving poses. We attribute these limitations to an overly rigid spatial binding that forces strict pixel-wise alignment between the pose and reference, and an inability to consistently rebind motion to intended subjects. To address these challenges, we propose CoDance, a novel Unbind-Rebind framework that enables the animation of arbitrary subject counts, types, and spatial configurations conditioned on a single, potentially misaligned pose sequence. Specifically, the Unbind module employs a novel pose shift encoder to break the rigid spatial binding between the pose and the reference by introducing stochastic perturbations to both poses and their latent features, thereby compelling the model to learn a location-agnostic motion representation. To ensure precise control and subject association, we then devise a Rebind module, leveraging semantic guidance from text prompts and spatial guidance from subject masks to direct the learned motion to intended characters. Furthermore, to facilitate comprehensive evaluation, we introduce a new multi-subject CoDanceBench. Extensive experiments on CoDanceBench and existing datasets show that CoDance achieves SOTA performance, exhibiting remarkable generalization across diverse subjects and spatial layouts. The code and weights will be open-sourced.

Method: Integrates a graph smoothing module to optimize graph structure by minimizing unreliable sparse/noisy connections, plus an enhanced local geometry learning module with shape features from adaptive geometric descriptors (eigenvectors) and distribution features from cylindrical coordinate transformation.

Result: Experimental results on real-world datasets validate effectiveness in various point cloud learning tasks including classification, part segmentation, and semantic segmentation.

Conclusion: The proposed integration of graph smoothing and enhanced local geometry learning addresses limitations of conventional graph structures in 3D point cloud analysis, improving performance across multiple tasks.

Abstract: Graph-based methods have proven to be effective in capturing relationships among points for 3D point cloud analysis. However, these methods often suffer from suboptimal graph structures, particularly due to sparse connections at boundary points and noisy connections in junction areas. To address these challenges, we propose a novel method that integrates a graph smoothing module with an enhanced local geometry learning module. Specifically, we identify the limitations of conventional graph structures, particularly in handling boundary points and junction areas. In response, we introduce a graph smoothing module designed to optimize the graph structure and minimize the negative impact of unreliable sparse and noisy connections. Based on the optimized graph structure, we improve the feature extract function with local geometry information. These include shape features derived from adaptive geometric descriptors based on eigenvectors and distribution features obtained through cylindrical coordinate transformation. Experimental results on real-world datasets validate the effectiveness of our method in various point cloud learning tasks, i.e., classification, part segmentation, and semantic segmentation.

Method: VIGA uses a closed-loop write-run-render-compare-revise procedure starting from an empty world. It combines: (1) a skill library alternating generator and verifier roles, and (2) evolving context memory containing plans, code diffs, and render history. The approach is task-agnostic and model-agnostic (no finetuning needed).

Result: VIGA substantially improves one-shot baselines: 35.32% on BlenderGym, 117.17% on SlideBench, and 124.70% on the new BlenderBench benchmark. It covers diverse tasks including 3D reconstruction, multi-step scene editing, 4D physical interaction, and 2D document editing.

Conclusion: VIGA demonstrates that iterative execution and verification with multimodal reasoning enables effective vision-as-inverse-graphics, providing a unified protocol to evaluate heterogeneous foundation VLMs across various graphics-related tasks.

Abstract: Vision-as-inverse-graphics, the concept of reconstructing an image as an editable graphics program is a long-standing goal of computer vision. Yet even strong VLMs aren’t able to achieve this in one-shot as they lack fine-grained spatial and physical grounding capability. Our key insight is that closing this gap requires interleaved multimodal reasoning through iterative execution and verification. Stemming from this, we present VIGA (Vision-as-Inverse-Graphic Agent) that starts from an empty world and reconstructs or edits scenes through a closed-loop write-run-render-compare-revise procedure. To support long-horizon reasoning, VIGA combines (i) a skill library that alternates generator and verifier roles and (ii) an evolving context memory that contains plans, code diffs, and render history. VIGA is task-agnostic as it doesn’t require auxiliary modules, covering a wide range of tasks such as 3D reconstruction, multi-step scene editing, 4D physical interaction, and 2D document editing, etc. Empirically, we found VIGA substantially improves one-shot baselines on BlenderGym (35.32%) and SlideBench (117.17%). Moreover, VIGA is also model-agnostic as it doesn’t require finetuning, enabling a unified protocol to evaluate heterogeneous foundation VLMs. To better support this protocol, we introduce BlenderBench, a challenging benchmark that stress-tests interleaved multimodal reasoning with graphics engine, where VIGA improves by 124.70%.

Method: The authors conduct analytical study and systematic experiments on layer-wise hybridization patterns of linear and softmax attention, then propose SoLA-Vision - a flexible layer-wise hybrid attention backbone that enables fine-grained control over integration of linear and softmax attention layers.

Result: On ImageNet-1K, SoLA-Vision outperforms purely linear and other hybrid attention models. On dense prediction tasks, it consistently surpasses strong baselines by considerable margins while achieving strong accuracy-computation trade-offs.

Conclusion: Fine-grained layer-wise hybridization with strategic insertion of a small number of global softmax layers provides better performance than rigid intra-block hybrid designs, enabling efficient high-resolution vision models with maintained accuracy.

Abstract: Standard softmax self-attention excels in vision tasks but incurs quadratic complexity O(N^2), limiting high-resolution deployment. Linear attention reduces the cost to O(N), yet its compressed state representations can impair modeling capacity and accuracy. We present an analytical study that contrasts linear and softmax attention for visual representation learning from a layer-stacking perspective. We further conduct systematic experiments on layer-wise hybridization patterns of linear and softmax attention. Our results show that, compared with rigid intra-block hybrid designs, fine-grained layer-wise hybridization can match or surpass performance while requiring fewer softmax layers. Building on these findings, we propose SoLA-Vision (Softmax-Linear Attention Vision), a flexible layer-wise hybrid attention backbone that enables fine-grained control over how linear and softmax attention are integrated. By strategically inserting a small number of global softmax layers, SoLA-Vision achieves a strong trade-off between accuracy and computational cost. On ImageNet-1K, SoLA-Vision outperforms purely linear and other hybrid attention models. On dense prediction tasks, it consistently surpasses strong baselines by a considerable margin. Code will be released.

Method: Transforms multi-sensor Landsat and MODIS data into quantized latent vectors using ESDNet architecture and Finite Scalar Quantization, achieving ~340x compression by condensing annual phenology into 12 temporal steps.

Result: Achieves high compression (2.4TB/year for global land surface) with strong reconstructive fidelity (MAE: 0.0130; RMSE: 0.0179; CC: 0.8543) and outperforms raw reflectance in land-cover classification (79.74% vs 76.92%).

Conclusion: ESD provides a versatile foundation for democratizing planetary-scale research and advancing geospatial AI by enabling decadal-scale global analysis on standard local workstations.

Abstract: The rapid evolution of satellite-borne Earth Observation (EO) systems has revolutionized terrestrial monitoring, yielding petabyte-scale archives. However, the immense computational and storage requirements for global-scale analysis often preclude widespread use, hindering planetary-scale studies. To address these barriers, we present Embedded Seamless Data (ESD), an ultra-lightweight, 30-m global Earth embedding database spanning the 25-year period from 2000 to 2024. By transforming high-dimensional, multi-sensor observations from the Landsat series (5, 7, 8, and 9) and MODIS Terra into information-dense, quantized latent vectors, ESD distills essential geophysical and semantic features into a unified latent space. Utilizing the ESDNet architecture and Finite Scalar Quantization (FSQ), the dataset achieves a transformative ~340-fold reduction in data volume compared to raw archives. This compression allows the entire global land surface for a single year to be encapsulated within approximately 2.4 TB, enabling decadal-scale global analysis on standard local workstations. Rigorous validation demonstrates high reconstructive fidelity (MAE: 0.0130; RMSE: 0.0179; CC: 0.8543). By condensing the annual phenological cycle into 12 temporal steps, the embeddings provide inherent denoising and a semantically organized space that outperforms raw reflectance in land-cover classification, achieving 79.74% accuracy (vs. 76.92% for raw fusion). With robust few-shot learning capabilities and longitudinal consistency, ESD provides a versatile foundation for democratizing planetary-scale research and advancing next-generation geospatial artificial intelligence.

Method: Uses joint transport of a segment in sample space built on top of arbitrary Rectified Flow models operating on structured latent space, enabling faster inference computation.

Result: Demonstrates high structural alignment and visual quality for sample pairs across image, video, and 3D shape generation domains. Improves state-of-the-art for image/video pipelines and shows comparable 3D quality with orders of magnitude faster computation.

Conclusion: The proposed method provides an efficient, high-quality solution for joint inference of structurally aligned samples across multiple modalities, addressing limitations of existing approaches.

Abstract: We suggest a new multi-modal algorithm for joint inference of paired structurally aligned samples with Rectified Flow models. While some existing methods propose a codependent generation process, they do not view the problem of joint generation from a structural alignment perspective. Recent work uses Score Distillation Sampling to generate aligned 3D models, but SDS is known to be time-consuming, prone to mode collapse, and often provides cartoonish results. By contrast, our suggested approach relies on the joint transport of a segment in the sample space, yielding faster computation at inference time. Our approach can be built on top of an arbitrary Rectified Flow model operating on the structured latent space. We show the applicability of our method to the domains of image, video, and 3D shape generation using state-of-the-art baselines and evaluate it against both editing-based and joint inference-based competing approaches. We demonstrate a high degree of structural alignment for the sample pairs obtained with our method and a high visual quality of the samples. Our method improves the state-of-the-art for image and video generation pipelines. For 3D generation, it is able to show comparable quality while working orders of magnitude faster.

Method: BioTune uses evolutionary optimization to adaptively fine-tune pre-trained models by optimally choosing which layers to freeze and adjusting learning rates for unfrozen layers.

Result: BioTune demonstrates superior accuracy and efficiency over state-of-the-art fine-tuning methods (AutoRGN and LoRA) on nine image classification datasets spanning natural and medical domains, and achieves top performance across four different CNN architectures.

Conclusion: BioTune is an effective adaptive fine-tuning technique that enhances transfer learning performance across diverse domains and architectures, with ablation studies providing insights into its key components’ impact.

Abstract: Deep learning has significantly advanced image analysis across diverse domains but often depends on large, annotated datasets for success. Transfer learning addresses this challenge by utilizing pre-trained models to tackle new tasks with limited labeled data. However, discrepancies between source and target domains can hinder effective transfer learning. We introduce BioTune, a novel adaptive fine-tuning technique utilizing evolutionary optimization. BioTune enhances transfer learning by optimally choosing which layers to freeze and adjusting learning rates for unfrozen layers. Through extensive evaluation on nine image classification datasets, spanning natural and specialized domains such as medical imaging, BioTune demonstrates superior accuracy and efficiency over state-of-the-art fine-tuning methods, including AutoRGN and LoRA, highlighting its adaptability to various data characteristics and distribution changes. Additionally, BioTune consistently achieves top performance across four different CNN architectures, underscoring its flexibility. Ablation studies provide valuable insights into the impact of BioTune’s key components on overall performance. The source code is available at https://github.com/davilac/BioTune.

Method: Three-stage ITKM framework: 1) Fine-tune CLIP image encoder with scenario embedding, 2) Optimize text embeddings with multi-scenario separation loss, 3) Use heterogeneous matching modules and dynamic text representation update for cross-scenario consistency.

Result: ITKM outperforms existing scenario-specific methods and enhances overall performance by integrating knowledge from multiple scenarios, demonstrating superiority and generalizability across diverse ReID scenarios.

Conclusion: The proposed UMS-ReID task and ITKM framework successfully address multi-scenario person re-identification by effectively leveraging vision-language models, providing a unified approach that surpasses specialized methods while maintaining cross-scenario consistency.

Abstract: We propose unsupervised multi-scenario (UMS) person re-identification (ReID) as a new task that expands ReID across diverse scenarios (cross-resolution, clothing change, etc.) within a single coherent framework. To tackle UMS-ReID, we introduce image-text knowledge modeling (ITKM) – a three-stage framework that effectively exploits the representational power of vision-language models. We start with a pre-trained CLIP model with an image encoder and a text encoder. In Stage I, we introduce a scenario embedding in the image encoder and fine-tune the encoder to adaptively leverage knowledge from multiple scenarios. In Stage II, we optimize a set of learned text embeddings to associate with pseudo-labels from Stage I and introduce a multi-scenario separation loss to increase the divergence between inter-scenario text representations. In Stage III, we first introduce cluster-level and instance-level heterogeneous matching modules to obtain reliable heterogeneous positive pairs (e.g., a visible image and an infrared image of the same person) within each scenario. Next, we propose a dynamic text representation update strategy to maintain consistency between text and image supervision signals. Experimental results across multiple scenarios demonstrate the superiority and generalizability of ITKM; it not only outperforms existing scenario-specific methods but also enhances overall performance by integrating knowledge from multiple scenarios.

Method: Proposes a lightweight asymmetric dual-encoder framework that learns unified, style-invariant visual embeddings. Jointly optimizes instance-level alignment and class-level semantic consistency to anchor visual embeddings to language-agnostic semantic prototypes, enforcing invariance across scripts and writing styles.

Result: Outperforms 28 baselines and achieves state-of-the-art accuracy on within-language retrieval benchmarks. Successfully validates cross-lingual retrieval where query language differs from target language. Achieves strong performance with only a fraction of parameters required by existing models.

Conclusion: The proposed framework enables accurate and resource-efficient cross-script handwriting retrieval, making it practical for edge deployment while maintaining high performance across different languages and writing styles.

Abstract: Handwritten word retrieval is vital for digital archives but remains challenging due to large handwriting variability and cross-lingual semantic gaps. While large vision-language models offer potential solutions, their prohibitive computational costs hinder practical edge deployment. To address this, we propose a lightweight asymmetric dual-encoder framework that learns unified, style-invariant visual embeddings. By jointly optimizing instance-level alignment and class-level semantic consistency, our approach anchors visual embeddings to language-agnostic semantic prototypes, enforcing invariance across scripts and writing styles. Experiments show that our method outperforms 28 baselines and achieves state-of-the-art accuracy on within-language retrieval benchmarks. We further conduct explicit cross-lingual retrieval, where the query language differs from the target language, to validate the effectiveness of the learned cross-lingual representations. Achieving strong performance with only a fraction of the parameters required by existing models, our framework enables accurate and resource-efficient cross-script handwriting retrieval.

Method: Proposes FTDMamba network with two core components: (1) Frequency Decoupled Spatiotemporal Correlation Module that disentangles coupled motion patterns through frequency analysis to model global spatiotemporal dependencies; (2) Temporal Dilation Mamba Module that uses Mamba’s sequence modeling to jointly learn fine-grained temporal dynamics and local spatial structures across multiple temporal receptive fields.

Result: Achieves state-of-the-art performance on two public static benchmarks and the new MUVAD dataset. Created MUVAD dataset with 222,736 frames, 240 anomaly events across 12 anomaly types, addressing the lack of dynamic background UAV VAD datasets.

Conclusion: FTDMamba effectively addresses UAV video anomaly detection in dynamic backgrounds by decoupling motion patterns through frequency analysis and leveraging Mamba for multi-scale temporal modeling. The new MUVAD dataset fills a critical gap in the field for dynamic UAV scenarios.

Abstract: Recent advances in video anomaly detection (VAD) mainly focus on ground-based surveillance or unmanned aerial vehicle (UAV) videos with static backgrounds, whereas research on UAV videos with dynamic backgrounds remains limited. Unlike static scenarios, dynamically captured UAV videos exhibit multi-source motion coupling, where the motion of objects and UAV-induced global motion are intricately intertwined. Consequently, existing methods may misclassify normal UAV movements as anomalies or fail to capture true anomalies concealed within dynamic backgrounds. Moreover, many approaches do not adequately address the joint modeling of inter-frame continuity and local spatial correlations across diverse temporal scales. To overcome these limitations, we propose the Frequency-Assisted Temporal Dilation Mamba (FTDMamba) network for UAV VAD, including two core components: (1) a Frequency Decoupled Spatiotemporal Correlation Module, which disentangles coupled motion patterns and models global spatiotemporal dependencies through frequency analysis; and (2) a Temporal Dilation Mamba Module, which leverages Mamba’s sequence modeling capability to jointly learn fine-grained temporal dynamics and local spatial structures across multiple temporal receptive fields. Additionally, unlike existing UAV VAD datasets which focus on static backgrounds, we construct a large-scale Moving UAV VAD dataset (MUVAD), comprising 222,736 frames with 240 anomaly events across 12 anomaly types. Extensive experiments demonstrate that FTDMamba achieves state-of-the-art (SOTA) performance on two public static benchmarks and the new MUVAD dataset. The code and MUVAD dataset will be available at: https://github.com/uavano/FTDMamba.

Method: Offline cross-architecture knowledge distillation transfers visual representations from frozen DINOv2 teacher to ResNet-18 student on ImageNet. The distilled encoder is then jointly fine-tuned with a diffusion policy head on target manipulation tasks.

Result: Outperforms policies with from-scratch ResNet or fine-tuned DINOv2 encoders across 34 simulated benchmarks and 5 real-world tasks. Also surpasses 3D encoders using point clouds and larger Vision-Language Models.

Conclusion: Simple, well-founded distillation strategy enables state-of-the-art performance in data-efficient robotic manipulation by synergizing large ViT generalization with compact CNN optimization efficiency.

Abstract: Visuomotor policies often leverage large pre-trained Vision Transformers (ViTs) for their powerful generalization capabilities. However, their significant data requirements present a major challenge in the data-scarce context of most robotic learning settings, where compact CNNs with strong inductive biases can be more easily optimized. To address this trade-off, we introduce X-Distill, a simple yet highly effective method that synergizes the strengths of both architectures. Our approach involves an offline, cross-architecture knowledge distillation, transferring the rich visual representations of a large, frozen DINOv2 teacher to a compact ResNet-18 student on the general-purpose ImageNet dataset. This distilled encoder, now endowed with powerful visual priors, is then jointly fine-tuned with a diffusion policy head on the target manipulation tasks. Extensive experiments on $34$ simulated benchmarks and $5$ challenging real-world tasks demonstrate that our method consistently outperforms policies equipped with from-scratch ResNet or fine-tuned DINOv2 encoders. Notably, X-Distill also surpasses 3D encoders that utilize privileged point cloud observations or much larger Vision-Language Models. Our work highlights the efficacy of a simple, well-founded distillation strategy for achieving state-of-the-art performance in data-efficient robotic manipulation.

Method: TTR formulates image patches as tokens, uses lightweight similarity metrics to identify static regions, and propagates precomputed deep features to bypass redundant backbone computations.

Result: On edge-grade hardware, TTR achieves 30% reduction in inference latency with negligible segmentation accuracy degradation (<0.5% mIoU), validated on standard benchmarks and a new Oblique Floodwater Dataset.

Conclusion: TTR effectively shifts the operational Pareto frontier, enabling high-fidelity, real-time oblique video understanding for time-critical remote sensing missions.

Abstract: Effective disaster response relies on rapid disaster response, where oblique aerial video is the primary modality for initial scouting due to its ability to maximize spatial coverage and situational awareness in limited flight time. However, the on-board processing of high-resolution oblique streams is severely bottlenecked by the strict Size, Weight, and Power (SWaP) constraints of Unmanned Aerial Vehicles (UAVs). The computational density required to process these wide-field-of-view streams precludes low-latency inference on standard edge hardware. To address this, we propose Temporal Token Reuse (TTR), an adaptive inference framework capable of accelerating video segmentation on embedded devices. TTR exploits the intrinsic spatiotemporal redundancy of aerial video by formulating image patches as tokens; it utilizes a lightweight similarity metric to dynamically identify static regions and propagate their precomputed deep features, thereby bypassing redundant backbone computations. We validate the framework on standard benchmarks and a newly curated Oblique Floodwater Dataset designed for hydrological monitoring. Experimental results on edge-grade hardware demonstrate that TTR achieves a 30% reduction in inference latency with negligible degradation in segmentation accuracy (< 0.5% mIoU). These findings confirm that TTR effectively shifts the operational Pareto frontier, enabling high-fidelity, real-time oblique video understanding for time-critical remote sensing missions

Method: Integrates Segment Anything Model 2 (SAM2) into human-in-the-loop workflow with modifications to reduce resource requirements. Features include: persistent instance identity management, automated “lock-and-refine” workflow with barrier frames, mask-skeletonization-based auto-prompting, and a processing layer minimizing computational overhead for responsive UI.

Result: Generates research-ready datasets in YOLO and PNG formats with structured interaction logs. Verified through animal behavior tracking use-cases and subsets of LVOS and DAVIS benchmark datasets. Provides scalable, private, cost-effective alternative to commercial platforms.

Conclusion: SAMannot offers a practical solution for complex video annotation tasks by combining foundation model capabilities with human oversight in a local, privacy-preserving framework that addresses resource constraints while maintaining high throughput.

Abstract: Current research workflows for precise video segmentation are often forced into a compromise between labor-intensive manual curation, costly commercial platforms, and/or privacy-compromising cloud-based services. The demand for high-fidelity video instance segmentation in research is often hindered by the bottleneck of manual annotation and the privacy concerns of cloud-based tools. We present SAMannot, an open-source, local framework that integrates the Segment Anything Model 2 (SAM2) into a human-in-the-loop workflow. To address the high resource requirements of foundation models, we modified the SAM2 dependency and implemented a processing layer that minimizes computational overhead and maximizes throughput, ensuring a highly responsive user interface. Key features include persistent instance identity management, an automated ``lock-and-refine’’ workflow with barrier frames, and a mask-skeletonization-based auto-prompting mechanism. SAMannot facilitates the generation of research-ready datasets in YOLO and PNG formats alongside structured interaction logs. Verified through animal behavior tracking use-cases and subsets of the LVOS and DAVIS benchmark datasets, the tool provides a scalable, private, and cost-effective alternative to commercial platforms for complex video annotation tasks.

Method: CASWiT uses a dual-branch Swin-based architecture: 1) Context encoder processes downsampled neighborhood for long-range dependencies, 2) High-resolution encoder extracts detailed features from UHR patches, 3) Cross-scale fusion module combines cross-attention and gated feature injection to enrich high-resolution tokens with context. Also includes SimMIM-style pretraining with 75% masking of high-resolution tokens and corresponding low-resolution regions.

Result: Achieves 65.83% mIoU on IGN FLAIR-HUB aerial dataset, outperforming RGB baselines by 1.78 points. On URUR dataset, achieves 49.1% mIoU, surpassing current state-of-the-art by +0.9% under official evaluation protocol.

Conclusion: CASWiT effectively addresses memory constraints in UHR segmentation by separating global context and fine-grained feature processing, then intelligently fusing them. The method demonstrates superior performance on large-scale aerial datasets and provides a practical solution for remote sensing applications.

Abstract: Semantic ultra high resolution image (UHR) segmentation is essential in remote sensing applications such as aerial mapping and environmental monitoring. Transformer-based models struggle in this setting because memory grows quadratically with token count, constraining either the contextual scope or the spatial resolution. We introduce CASWiT (Context-Aware Stage-Wise Transformer), a dual-branch, Swin-based architecture that injects global cues into fine-grained UHR features. A context encoder processes a downsampled neighborhood to capture long-range dependencies, while a high resolution encoder extracts detailed features from UHR patches. A cross-scale fusion module, combining cross-attention and gated feature injection, enriches high-resolution tokens with context. Beyond architecture, we propose a SimMIM-style pretraining. We mask 75% of the high-resolution image tokens and the low-resolution center region that spatially corresponds to the UHR patch, then train the shared dual-encoder with small decoder to reconstruct the UHR initial image. Extensive experiments on the large-scale IGN FLAIR-HUB aerial dataset demonstrate the effectiveness of CASWiT. Our method achieves 65.83% mIoU, outperforming RGB baselines by 1.78 points. On URUR, CASWiT achieves 49.1% mIoU, surpassing the current SoTA by +0.9% under the official evaluation protocol. All codes are provided on: https://huggingface.co/collections/heig-vd-geo/caswit.

Method: Integrates VLMs with traditional computer vision methods through explicit logic reasoning, employs intelligent fine-tuning strategy, and generates justifications for VLM outputs during inference.

Result: Intelligent fine-tuning mechanism achieves substantially higher accuracy than uninformed selection, provides means to confirm validity of VLM outputs or indicate why they may be questionable during inference.

Conclusion: The proposed approach enhances situational awareness by combining VLMs with traditional CV methods through logic reasoning, improving accuracy, fine-grained detail extraction, and providing interpretable justifications for model outputs.

Abstract: Vision-Language Models (VLMs) offer the ability to generate high-level, interpretable descriptions of complex activities from images and videos, making them valuable for situational awareness (SA) applications. In such settings, the focus is on identifying infrequent but significant events with high reliability and accuracy, while also extracting fine-grained details and assessing recognition quality. In this paper, we propose an approach that integrates VLMs with traditional computer vision methods through explicit logic reasoning to enhance SA in three key ways: (a) extracting fine-grained event details, (b) employing an intelligent fine-tuning (FT) strategy that achieves substantially higher accuracy than uninformed selection, and (c) generating justifications for VLM outputs during inference. We demonstrate that our intelligent FT mechanism improves the accuracy and provides a valuable means, during inferencing, to either confirm the validity of the VLM output or indicate why it may be questionable.

Method: Unsupervised encoder-decoder architecture: compact U-Net encoder predicts nonnegative concentration channels, differentiable Beer-Lambert forward model decoder with learnable stain matrix initialized from typical chromogen hues, trained with perceptual reconstruction objective and anti-mixing loss terms.

Result: Excellent RGB reconstruction and significantly reduced inter-channel bleed-through compared to matrix-based deconvolution on colorectal mIHC panel with 5 stains (H, CDX2, MUC2, MUC5, CD8).

Conclusion: The data-driven approach effectively learns cohort-specific stain characteristics for multiplex IHC, producing crisp, well-separated per-stain concentration maps for improved quantitative analysis.

Abstract: Separating the contributions of individual chromogenic stains in RGB histology whole slide images (WSIs) is essential for stain normalization, quantitative assessment of marker expression, and cell-level readouts in immunohistochemistry (IHC). Classical Beer-Lambert (BL) color deconvolution is well-established for two- or three-stain settings, but becomes under-determined and unstable for multiplex IHC (mIHC) with K>3 chromogens. We present a simple, data-driven encoder-decoder architecture that learns cohort-specific stain characteristics for mIHC RGB WSIs and yields crisp, well-separated per-stain concentration maps. The encoder is a compact U-Net that predicts K nonnegative concentration channels; the decoder is a differentiable BL forward model with a learnable stain matrix initialized from typical chromogen hues. Training is unsupervised with a perceptual reconstruction objective augmented by loss terms that discourage unnecessary stain mixing. On a colorectal mIHC panel comprising 5 stains (H, CDX2, MUC2, MUC5, CD8) we show excellent RGB reconstruction, and significantly reduced inter-channel bleed-through compared with matrix-based deconvolution. Code and model are available at https://github.com/measty/StainQuant.git.

Method: Proposes a dual-modality cross-view learning approach using coupled global context vision transformer (CGCViT) to fuse UAV and street-view imagery, with HotSat-1 thermal infrared measurements to quantify building-heat relationships.

Result: The framework outperforms single-modality models by up to 9.3%, identifies vegetation, brighter roofing, and specific roofing materials as significantly associated with lower heat, and successfully maps heat exposure inequalities in Dar es Salaam.

Conclusion: Localized, data-driven risk assessment using machine learning can identify heat exposure inequalities and inform equitable climate adaptation strategies, demonstrating the value of complementary UAV and street-view perspectives.

Abstract: Climate change is intensifying human heat exposure, particularly in densely built urban centers of the Global South. Low-cost construction materials and high thermal-mass surfaces further exacerbate this risk. Yet scalable methods for assessing such heat-relevant building attributes remain scarce. We propose a machine learning framework that fuses openly available unmanned aerial vehicle (UAV) and street-view (SV) imagery via a coupled global context vision transformer (CGCViT) to learn heat-relevant representations of urban structures. Thermal infrared (TIR) measurements from HotSat-1 are used to quantify the relationship between building attributes and heat-associated health risks. Our dual-modality cross-view learning approach outperforms the best single-modality models by up to $9.3%$, demonstrating that UAV and SV imagery provide valuable complementary perspectives on urban structures. The presence of vegetation surrounding buildings (versus no vegetation), brighter roofing (versus darker roofing), and roofing made of concrete, clay, or wood (versus metal or tarpaulin) are all significantly associated with lower HotSat-1 TIR values. Deployed across the city of Dar es Salaam, Tanzania, the proposed framework illustrates how household-level inequalities in heat exposure - often linked to socio-economic disadvantage and reflected in building materials - can be identified and addressed using machine learning. Our results point to the critical role of localized, data-driven risk assessment in shaping climate adaptation strategies that deliver equitable outcomes.

Method: TCS uses two key components: (1) Multi-Query Reasoning that generates multiple queries to capture complementary aspects of questions and videos, and (2) Clip-level Slow-Fast Sampling that adaptively balances dense local details with sparse global context.

Result: TCS consistently improves performance across different MLLMs on MLVU, LongVideoBench, and VideoMME benchmarks, achieving up to 6.9% accuracy boost and comparable accuracy with 50% fewer inference time cost.

Conclusion: TCS demonstrates both efficiency and efficacy for long video understanding, offering a training-free solution that enhances MLLM performance while reducing computational costs.

Abstract: Recent progress in multi-modal large language models (MLLMs) has significantly advanced video understanding. However, their performance on long-form videos remains limited by computational constraints and suboptimal frame selection. We present Think-Clip-Sample (TCS), a training-free framework that enhances long video understanding through two key components: (i) Multi-Query Reasoning, which generates multiple queries to capture complementary aspects of the question and video; and (ii) Clip-level Slow-Fast Sampling, which adaptively balances dense local details and sparse global context. Extensive experiments on MLVU, LongVideoBench, and VideoMME demonstrate that TCS consistently improves performance across different MLLMs, boosting up to 6.9% accuracy, and is capable of achieving comparable accuracy with 50% fewer inference time cost, highlighting both efficiency and efficacy of TCS on long video understanding.

Method: HUG uses fine-grained probabilistic learning with Gaussian embeddings for queries and targets. It customizes heterogeneous uncertainty estimations for multi-modal queries vs uni-modal targets, captures uncertainties about content quality and multi-modal coordination, and uses dynamic weighting for comprehensive query uncertainty. It also implements uncertainty-guided objectives with holistic and fine-grained contrasts and comprehensive negative sampling.

Result: Experiments on benchmarks show HUG’s effectiveness beyond state-of-the-art baselines, with faithful analysis justifying the technical contributions.

Conclusion: HUG successfully addresses CIR’s uncertainty challenges through heterogeneous uncertainty-guided probabilistic learning, improving robustness and performance over existing approaches.

Abstract: Composed Image Retrieval (CIR) enables image search by combining a reference image with modification text. Intrinsic noise in CIR triplets incurs intrinsic uncertainty and threatens the model’s robustness. Probabilistic learning approaches have shown promise in addressing such issues; however, they fall short for CIR due to their instance-level holistic modeling and homogeneous treatment of queries and targets. This paper introduces a Heterogeneous Uncertainty-Guided (HUG) paradigm to overcome these limitations. HUG utilizes a fine-grained probabilistic learning framework, where queries and targets are represented by Gaussian embeddings that capture detailed concepts and uncertainties. We customize heterogeneous uncertainty estimations for multi-modal queries and uni-modal targets. Given a query, we capture uncertainties not only regarding uni-modal content quality but also multi-modal coordination, followed by a provable dynamic weighting mechanism to derive comprehensive query uncertainty. We further design uncertainty-guided objectives, including query-target holistic contrast and fine-grained contrasts with comprehensive negative sampling strategies, which effectively enhance discriminative learning. Experiments on benchmarks demonstrate HUG’s effectiveness beyond state-of-the-art baselines, with faithful analysis justifying the technical contributions.

Method: 1) Uses semantic and uncertainty priors to suppress free space projections during view transformation with unsigned distance encoding for geometric consistency. 2) Cascade sparse completion module with hyper cross sparse convolution and generative upsampling for coarse-to-fine reasoning. 3) Object contextual representation (OCR) based mask decoder that aggregates global semantic context from sparse features via lightweight query-context interactions instead of expensive attention operations.

Result: Extensive experiments on SemanticKITTI benchmark show the approach outperforms baselines with 7.34% improvement in accuracy and 57.8% gain in efficiency.

Conclusion: SUG-Occ successfully addresses the computational challenges of 3D semantic occupancy prediction by exploiting scene sparsity through semantics and uncertainty guidance, enabling efficient real-time deployment while maintaining high accuracy.

Abstract: As autonomous driving moves toward full scene understanding, 3D semantic occupancy prediction has emerged as a crucial perception task, offering voxel-level semantics beyond traditional detection and segmentation paradigms. However, such a refined representation for scene understanding incurs prohibitive computation and memory overhead, posing a major barrier to practical real-time deployment. To address this, we propose SUG-Occ, an explicit Semantics and Uncertainty Guided Sparse Learning Enabled 3D Occupancy Prediction Framework, which exploits the inherent sparsity of 3D scenes to reduce redundant computation while maintaining geometric and semantic completeness. Specifically, we first utilize semantic and uncertainty priors to suppress projections from free space during view transformation while employing an explicit unsigned distance encoding to enhance geometric consistency, producing a structurally consistent sparse 3D representation. Secondly, we design an cascade sparse completion module via hyper cross sparse convolution and generative upsampling to enable efficiently coarse-to-fine reasoning. Finally, we devise an object contextual representation (OCR) based mask decoder that aggregates global semantic context from sparse features and refines voxel-wise predictions via lightweight query-context interactions, avoiding expensive attention operations over volumetric features. Extensive experiments on SemanticKITTI benchmark demonstrate that the proposed approach outperforms the baselines, achieving a 7.34/% improvement in accuracy and a 57.8% gain in efficiency.

Method: Dual-branch design: 1) Temporally prompted branch extends SAM with hierarchical adapters and dynamic temporal aggregation to separate wetland characteristics from phenological variability; 2) Spatial branch uses temporally constrained region-growing to generate dense pseudo-labels; 3) Bidirectional consistency regularization jointly optimizes both branches.

Result: Extensive experiments across eight global regions (~5,000 km² each) show WetSAM substantially outperforms state-of-the-art methods with average F1-score of 85.58%, delivering accurate and structurally consistent wetland segmentation with minimal labeling effort.

Conclusion: WetSAM demonstrates strong generalization capability and potential for scalable, low-cost, high-resolution wetland mapping by effectively integrating temporal information with sparse point supervision.

Abstract: Accurate wetland mapping is essential for ecosystem monitoring, yet dense pixel-level annotation is prohibitively expensive and practical applications usually rely on sparse point labels, under which existing deep learning models perform poorly, while strong seasonal and inter-annual wetland dynamics further render single-date imagery inadequate and lead to significant mapping errors; although foundation models such as SAM show promising generalization from point prompts, they are inherently designed for static images and fail to model temporal information, resulting in fragmented masks in heterogeneous wetlands. To overcome these limitations, we propose WetSAM, a SAM-based framework that integrates satellite image time series for wetland mapping from sparse point supervision through a dual-branch design, where a temporally prompted branch extends SAM with hierarchical adapters and dynamic temporal aggregation to disentangle wetland characteristics from phenological variability, and a spatial branch employs a temporally constrained region-growing strategy to generate reliable dense pseudo-labels, while a bidirectional consistency regularization jointly optimizes both branches. Extensive experiments across eight global regions of approximately 5,000 km2 each demonstrate that WetSAM substantially outperforms state-of-the-art methods, achieving an average F1-score of 85.58%, and delivering accurate and structurally consistent wetland segmentation with minimal labeling effort, highlighting its strong generalization capability and potential for scalable, low-cost, high-resolution wetland mapping.

Method: Three key improvements: 1) NWDLoss replaces IoU to reduce sensitivity to positional deviations in tiny objects; 2) EUCB replaces upsampling to preserve edge/texture details; 3) MSFA module adaptively strengthens perception in key scale intervals for local-global feature fusion.

Result: On PKU-PCB dataset, SME-YOLO achieves state-of-the-art performance with 2.2% mAP improvement and 4% Precision gain over baseline YOLOv11n.

Conclusion: The proposed SME-YOLO framework effectively addresses PCB defect detection challenges through specialized loss function, upsampling enhancement, and multi-scale attention, demonstrating superior performance for tiny, texture-similar defects.

Abstract: Surface defects on Printed Circuit Boards (PCBs) directly compromise product reliability and safety. However, achieving high-precision detection is challenging because PCB defects are typically characterized by tiny sizes, high texture similarity, and uneven scale distributions. To address these challenges, this paper proposes a novel framework based on YOLOv11n, named SME-YOLO (Small-target Multi-scale Enhanced YOLO). First, we employ the Normalized Wasserstein Distance Loss (NWDLoss). This metric effectively mitigates the sensitivity of Intersection over Union (IoU) to positional deviations in tiny objects. Second, the original upsampling module is replaced by the Efficient Upsampling Convolution Block (EUCB). By utilizing multi-scale convolutions, the EUCB gradually recovers spatial resolution and enhances the preservation of edge and texture details for tiny defects. Finally, this paper proposes the Multi-Scale Focused Attention (MSFA) module. Tailored to the specific spatial distribution of PCB defects, this module adaptively strengthens perception within key scale intervals, achieving efficient fusion of local fine-grained features and global context information. Experimental results on the PKU-PCB dataset demonstrate that SME-YOLO achieves state-of-the-art performance. Specifically, compared to the baseline YOLOv11n, SME-YOLO improves mAP by 2.2% and Precision by 4%, validating the effectiveness of the proposed method.

Method: Proposes a mathematical framework integrating width information into topological characterization using persistent homology and PDE smoothing concepts. Modifies local extrema of upper-level sets to capture width properties. Incorporates this enhanced topological description into variational segmentation models and designs neural networks with proper loss functions to enforce topological and width constraints.

Result: Numerical experiments demonstrate the method effectively maintains topological fidelity (preserving connectivity and genus counts) while explicitly embedding width characteristics (line thickness and length) into segmented image structures.

Conclusion: The proposed framework successfully overcomes limitations of traditional topological methods by integrating width information, enabling practical image segmentation that preserves both essential topological invariants and critical width attributes through variational constraints on topological energies.

Abstract: Existing research highlights the crucial role of topological priors in image segmentation, particularly in preserving essential structures such as connectivity and genus. Accurately capturing these topological features often requires incorporating width-related information, including the thickness and length inherent to the image structures. However, traditional mathematical definitions of topological structures lack this dimensional width information, limiting methods like persistent homology from fully addressing practical segmentation needs. To overcome this limitation, we propose a novel mathematical framework that explicitly integrates width information into the characterization of topological structures. This method leverages persistent homology, complemented by smoothing concepts from partial differential equations (PDEs), to modify local extrema of upper-level sets. This approach enables the resulting topological structures to inherently capture width properties. We incorporate this enhanced topological description into variational image segmentation models. Using some proper loss functions, we are also able to design neural networks that can segment images with the required topological and width properties. Through variational constraints on the relevant topological energies, our approach successfully preserves essential topological invariants such as connectivity and genus counts, simultaneously ensuring that segmented structures retain critical width attributes, including line thickness and length. Numerical experiments demonstrate the effectiveness of our method, showcasing its capability to maintain topological fidelity while explicitly embedding width characteristics into segmented image structures.

Method: Derived from PubMed Central Open Access PDFs, annotated with Google Cloud Vision OCR, and released in a compact JSON schema with word-, line-, and paragraph-level bounding boxes.

Result: Created a corpus spanning 209.5K articles (1.5M pages; ~1.3B words) with analyzed characteristics including journal coverage and detected layout features.

Conclusion: The corpus facilitates downstream research in OCR and document understanding, though limitations include reliance on single OCR engine and heuristic line reconstruction; data and schema are publicly released.

Abstract: PubMed-OCR is an OCR-centric corpus of scientific articles derived from PubMed Central Open Access PDFs. Each page image is annotated with Google Cloud Vision and released in a compact JSON schema with word-, line-, and paragraph-level bounding boxes. The corpus spans 209.5K articles (1.5M pages; ~1.3B words) and supports layout-aware modeling, coordinate-grounded QA, and evaluation of OCR-dependent pipelines. We analyze corpus characteristics (e.g., journal coverage and detected layout features) and discuss limitations, including reliance on a single OCR engine and heuristic line reconstruction. We release the data and schema to facilitate downstream research and invite extensions.

Method: Uses Metric Cognitive Map (Metric-CogMap) for unified spatial representation combining discrete grid for relational reasoning and continuous metric-scale representation for geometric understanding, plus Cognitive Chain-of-Thought (Cog-CoT) for explicit geometric reasoning through deterministic operations like vector operations, bounding-box distances, and occlusion-aware appearance order cues.

Result: Achieves 59.9% accuracy using only half the supervision, closely matching the 60.9% baseline trained with full dataset. Consistently outperforms state-of-the-art methods by 5.3%, 4.8%, and 4.0% under 10%, 25%, and 50% training subsets respectively on VSI-Bench.

Conclusion: Map2Thought enables explainable 3D understanding through explicit spatial reasoning with interpretable inference traces grounded in 3D structure, demonstrating strong performance with reduced supervision.

Abstract: We propose Map2Thought, a framework that enables explicit and interpretable spatial reasoning for 3D VLMs. The framework is grounded in two key components: Metric Cognitive Map (Metric-CogMap) and Cognitive Chain-of-Thought (Cog-CoT). Metric-CogMap provides a unified spatial representation by integrating a discrete grid for relational reasoning with a continuous, metric-scale representation for precise geometric understanding. Building upon the Metric-CogMap, Cog-CoT performs explicit geometric reasoning through deterministic operations, including vector operations, bounding-box distances, and occlusion-aware appearance order cues, producing interpretable inference traces grounded in 3D structure. Experimental results show that Map2Thought enables explainable 3D understanding, achieving 59.9% accuracy using only half the supervision, closely matching the 60.9% baseline trained with the full dataset. It consistently outperforms state-of-the-art methods by 5.3%, 4.8%, and 4.0% under 10%, 25%, and 50% training subsets, respectively, on the VSI-Bench.

Method: Three-step pipeline: (1) Detect candidate infrastructure (barns, feedlots, manure lagoons, silos) using domain-tuned YOLOv8 detector, derive SAM2 masks, and filter with component-specific criteria; (2) Extract structured descriptors (counts, areas, orientations, spatial relations) and fuse with deep visual features using lightweight spatial cross-attention classifier; (3) Output CAFO type predictions with mask-level attributions linking decisions to visible infrastructure.

Result: Achieves state-of-the-art performance with Swin-B+PRISM-CAFO surpassing best performing baseline by up to 15%. Shows strong predictive performance across diverse U.S. regions and provides systematic gradient-activation analyses quantifying the impact of domain priors.

Conclusion: The proposed infrastructure-first, explainable pipeline effectively identifies and characterizes CAFOs from aerial imagery, providing both accurate predictions and interpretable attributions that link decisions to visible infrastructure components, enabling better monitoring of large-scale livestock operations.

Abstract: Large-scale livestock operations pose significant risks to human health and the environment, while also being vulnerable to threats such as infectious diseases and extreme weather events. As the number of such operations continues to grow, accurate and scalable mapping has become increasingly important. In this work, we present an infrastructure-first, explainable pipeline for identifying and characterizing Concentrated Animal Feeding Operations (CAFOs) from aerial and satellite imagery. Our method (1) detects candidate infrastructure (e.g., barns, feedlots, manure lagoons, silos) with a domain-tuned YOLOv8 detector, then derives SAM2 masks from these boxes and filters component-specific criteria, (2) extracts structured descriptors (e.g., counts, areas, orientations, and spatial relations) and fuses them with deep visual features using a lightweight spatial cross-attention classifier, and (3) outputs both CAFO type predictions and mask-level attributions that link decisions to visible infrastructure. Through comprehensive evaluation, we show that our approach achieves state-of-the-art performance, with Swin-B+PRISM-CAFO surpassing the best performing baseline by up to 15%. Beyond strong predictive performance across diverse U.S. regions, we run systematic gradient–activation analyses that quantify the impact of domain priors and show ho

Method: Two core techniques: (1) modality-adaptive partial-RoPE strategy that selectively masks nonessential dimensions for both traditional and multimodal settings, and (2) modality-decoupled low-rank approximation that independently compresses visual and textual KV spaces. Uses parameter-efficient fine-tuning with focus on minimizing output activation error.

Result: Extensive experiments on three representative VLMs show MHA2MLA-VLM restores original model performance with minimal supervised data, significantly reduces KV cache footprint, and integrates seamlessly with KV quantization.

Conclusion: MHA2MLA-VLM provides a parameter-efficient and multimodal-aware framework for converting off-the-shelf VLMs to MLA architecture, addressing KV cache bottlenecks while maintaining performance.

Abstract: As vision-language models (VLMs) tackle increasingly complex and multimodal tasks, the rapid growth of Key-Value (KV) cache imposes significant memory and computational bottlenecks during inference. While Multi-Head Latent Attention (MLA) offers an effective means to compress the KV cache and accelerate inference, adapting existing VLMs to the MLA architecture without costly pretraining remains largely unexplored. In this work, we present MHA2MLA-VLM, a parameter-efficient and multimodal-aware framework for converting off-the-shelf VLMs to MLA. Our approach features two core techniques: (1) a modality-adaptive partial-RoPE strategy that supports both traditional and multimodal settings by selectively masking nonessential dimensions, and (2) a modality-decoupled low-rank approximation method that independently compresses the visual and textual KV spaces. Furthermore, we introduce parameter-efficient fine-tuning to minimize adaptation cost and demonstrate that minimizing output activation error, rather than parameter distance, substantially reduces performance loss. Extensive experiments on three representative VLMs show that MHA2MLA-VLM restores original model performance with minimal supervised data, significantly reduces KV cache footprint, and integrates seamlessly with KV quantization.

Method: 1) Train VLA model to encode ego vehicle and agent dynamics into latent tokens with planning, motion, and language supervision. 2) Perform language-conditioned autoregressive generation using multi-view images, scenario descriptions, and ego-action questions. 3) Use rollout-consistency loss with ground truth/pseudo-labels to stabilize predictions and maintain text-action alignment.

Result: On Bench2Drive, GeRo improves driving score by +15.7 and success rate by +26.2. Achieves state-of-the-art closed-loop and open-loop performance with strong zero-shot robustness through integration of reinforcement learning with generative rollouts.

Conclusion: Generative, language-conditioned reasoning shows promise as a foundation for safer and more interpretable end-to-end autonomous driving, enabling temporally consistent rollouts that support long-horizon reasoning and multi-agent planning.

Abstract: Vision-Language-Action (VLA) models are emerging as highly effective planning models for end-to-end autonomous driving systems. However, current works mostly rely on imitation learning from sparse trajectory annotations and under-utilize their potential as generative models. We propose Generative Scenario Rollouts (GeRo), a plug-and-play framework for VLA models that jointly performs planning and generation of language-grounded future traffic scenes through an autoregressive rollout strategy. First, a VLA model is trained to encode ego vehicle and agent dynamics into latent tokens under supervision from planning, motion, and language tasks, facilitating text-aligned generation. Next, GeRo performs language-conditioned autoregressive generation. Given multi-view images, a scenario description, and ego-action questions, it generates future latent tokens and textual responses to guide long-horizon rollouts. A rollout-consistency loss stabilizes predictions using ground truth or pseudo-labels, mitigating drift and preserving text-action alignment. This design enables GeRo to perform temporally consistent, language-grounded rollouts that support long-horizon reasoning and multi-agent planning. On Bench2Drive, GeRo improves driving score and success rate by +15.7 and +26.2, respectively. By integrating reinforcement learning with generative rollouts, GeRo achieves state-of-the-art closed-loop and open-loop performance, demonstrating strong zero-shot robustness. These results highlight the promise of generative, language-conditioned reasoning as a foundation for safer and more interpretable end-to-end autonomous driving.

Method: ReScene4D adapts 3DSIS architectures for 4DSIS without needing dense observations. It explores strategies to share information across observations, using shared context to enable consistent instance tracking while also improving standard 3DSIS quality.

Result: Achieves state-of-the-art performance on the 3RScan dataset, establishing a new benchmark. Introduces t-mAP metric that extends mAP to reward temporal identity consistency.

Conclusion: ReScene4D successfully addresses the challenge of temporally sparse 4D indoor semantic instance segmentation, enabling consistent tracking of object instances across intermittent scans while improving segmentation quality.

Abstract: Indoor environments evolve as objects move, appear, or disappear. Capturing these dynamics requires maintaining temporally consistent instance identities across intermittently captured 3D scans, even when changes are unobserved. We introduce and formalize the task of temporally sparse 4D indoor semantic instance segmentation (SIS), which jointly segments, identifies, and temporally associates object instances. This setting poses a challenge for existing 3DSIS methods, which require a discrete matching step due to their lack of temporal reasoning, and for 4D LiDAR approaches, which perform poorly due to their reliance on high-frequency temporal measurements that are uncommon in the longer-horizon evolution of indoor environments. We propose ReScene4D, a novel method that adapts 3DSIS architectures for 4DSIS without needing dense observations. It explores strategies to share information across observations, demonstrating that this shared context not only enables consistent instance tracking but also improves standard 3DSIS quality. To evaluate this task, we define a new metric, t-mAP, that extends mAP to reward temporal identity consistency. ReScene4D achieves state-of-the-art performance on the 3RScan dataset, establishing a new benchmark for understanding evolving indoor scenes.

Method: Uses off-the-shelf visual-inertial SLAM, 3D detection, and vision-language models to extract sparse SLAM points, posed multi-view images, and captions for each object. A rectified flow transformer is trained to condition on these modalities, with compositional augmentations, curriculum training (object- to scene-level), and background handling techniques for robustness.

Result: Significantly outperforms existing approaches, achieving 2.7x improvement in Chamfer distance compared to state-of-the-art. Introduces a new evaluation benchmark with 178 in-the-wild objects across 7 real-world scenes with geometry annotations.

Conclusion: ShapeR demonstrates effective 3D shape generation from casually captured sequences by leveraging multi-modal inputs and robust training strategies, addressing real-world challenges that existing methods struggle with.

Abstract: Recent advances in 3D shape generation have achieved impressive results, but most existing methods rely on clean, unoccluded, and well-segmented inputs. Such conditions are rarely met in real-world scenarios. We present ShapeR, a novel approach for conditional 3D object shape generation from casually captured sequences. Given an image sequence, we leverage off-the-shelf visual-inertial SLAM, 3D detection algorithms, and vision-language models to extract, for each object, a set of sparse SLAM points, posed multi-view images, and machine-generated captions. A rectified flow transformer trained to effectively condition on these modalities then generates high-fidelity metric 3D shapes. To ensure robustness to the challenges of casually captured data, we employ a range of techniques including on-the-fly compositional augmentations, a curriculum training scheme spanning object- and scene-level datasets, and strategies to handle background clutter. Additionally, we introduce a new evaluation benchmark comprising 178 in-the-wild objects across 7 real-world scenes with geometry annotations. Experiments show that ShapeR significantly outperforms existing approaches in this challenging setting, achieving an improvement of 2.7x in Chamfer distance compared to state of the art.

Method: UniX decouples understanding and generation into separate branches: an autoregressive branch for understanding and a diffusion branch for high-fidelity generation. It introduces cross-modal self-attention to dynamically guide generation with understanding features, combined with data cleaning and multi-stage training.

Result: UniX achieves 46.1% improvement in understanding performance (Micro-F1) and 24.2% gain in generation quality (FD-RadDino) compared to benchmarks, using only a quarter of the parameters of LLM-CXR. It performs on par with task-specific models.

Conclusion: UniX establishes a scalable paradigm for synergistic medical image understanding and generation by effectively decoupling tasks while enabling cross-modal guidance, demonstrating that unified models can match task-specific performance.

Abstract: Despite recent progress, medical foundation models still struggle to unify visual understanding and generation, as these tasks have inherently conflicting goals: semantic abstraction versus pixel-level reconstruction. Existing approaches, typically based on parameter-shared autoregressive architectures, frequently lead to compromised performance in one or both tasks. To address this, we present UniX, a next-generation unified medical foundation model for chest X-ray understanding and generation. UniX decouples the two tasks into an autoregressive branch for understanding and a diffusion branch for high-fidelity generation. Crucially, a cross-modal self-attention mechanism is introduced to dynamically guide the generation process with understanding features. Coupled with a rigorous data cleaning pipeline and a multi-stage training strategy, this architecture enables synergistic collaboration between tasks while leveraging the strengths of diffusion models for superior generation. On two representative benchmarks, UniX achieves a 46.1% improvement in understanding performance (Micro-F1) and a 24.2% gain in generation quality (FD-RadDino), using only a quarter of the parameters of LLM-CXR. By achieving performance on par with task-specific models, our work establishes a scalable paradigm for synergistic medical image understanding and generation. Codes and models are available at https://github.com/ZrH42/UniX.

Method: Progressive scene graph network architecture that dynamically allocates local scene graphs for temporal windows; uses DINOv2 foundation model to extract appearance/shape codes for sparse-view dynamic objects; includes frequency-modulated module to regularize object frequency spectrum.

Result: Achieves state-of-the-art view synthesis accuracy, object manipulation, and scene roaming ability across various scenes, demonstrating effective handling of large-scale urban environments with fast-moving vehicles.

Conclusion: The proposed approach successfully addresses the joint challenges of large-scale scenes and fast-moving vehicles through progressive scene graph learning, foundation model integration, and frequency-domain regularization.

Abstract: Implicit neural representation has demonstrated promising results in 3D reconstruction on various scenes. However, existing approaches either struggle to model fast-moving objects or are incapable of handling large-scale camera ego-motions in urban environments. This leads to low-quality synthesized views of the large-scale urban scenes. In this paper, we aim to jointly solve the problems caused by large-scale scenes and fast-moving vehicles, which are more practical and challenging. To this end, we propose a progressive scene graph network architecture to learn the local scene representations of dynamic objects and global urban scenes. The progressive learning architecture dynamically allocates a new local scene graph trained on frames within a temporal window, with the window size automatically determined, allowing us to scale up the representation to arbitrarily large scenes. Besides, according to our observations, the training views of dynamic objects are relatively sparse according to rapid movements, which leads to a significant decline in reconstruction accuracy for dynamic objects. Therefore, we utilize a foundation model network to encode the latent code. Specifically, we leverage the generalization capability of the visual foundation model DINOv2 to extract appearance and shape codes, and train the network on a large-scale urban scene object dataset to enhance its prior modeling ability for handling sparse-view dynamic inputs. In parallel, we introduce a frequency-modulated module that regularizes the frequency spectrum of objects, thereby addressing the challenge of modeling sparse image inputs from a frequency-domain perspective. Experimental results demonstrate that our method achieves state-of-the-art view synthesis accuracy, object manipulation, and scene roaming ability in various scenes.

Method: Collected data using connected vehicle platform and intelligent roadside unit equipped with 4D Radar, LiDAR, and multi-view cameras. Data collected across various weather conditions (sunny, rainy), times of day (daytime, dusk, nighttime), and challenging scenarios.

Result: Created V2X-Radar dataset with 20K LiDAR frames, 40K camera images, 20K 4D Radar data, and 350K annotated boxes across five categories. Established three sub-datasets: V2X-Radar-C for cooperative perception, V2X-Radar-I for roadside perception, and V2X-Radar-V for single-vehicle perception with comprehensive benchmarks.

Conclusion: V2X-Radar fills the critical gap in cooperative perception datasets by including 4D Radar, enabling research on weather-resilient autonomous driving systems. The dataset supports multiple research domains and will be publicly released with benchmark codebase.

Abstract: Modern autonomous vehicle perception systems often struggle with occlusions and limited perception range. Previous studies have demonstrated the effectiveness of cooperative perception in extending the perception range and overcoming occlusions, thereby enhancing the safety of autonomous driving. In recent years, a series of cooperative perception datasets have emerged; however, these datasets primarily focus on cameras and LiDAR, neglecting 4D Radar, a sensor used in single-vehicle autonomous driving to provide robust perception in adverse weather conditions. In this paper, to bridge the gap created by the absence of 4D Radar datasets in cooperative perception, we present V2X-Radar, the first large-scale, real-world multi-modal dataset featuring 4D Radar. V2X-Radar dataset is collected using a connected vehicle platform and an intelligent roadside unit equipped with 4D Radar, LiDAR, and multi-view cameras. The collected data encompasses sunny and rainy weather conditions, spanning daytime, dusk, and nighttime, as well as various typical challenging scenarios. The dataset consists of 20K LiDAR frames, 40K camera images, and 20K 4D Radar data, including 350K annotated boxes across five categories. To support various research domains, we have established V2X-Radar-C for cooperative perception, V2X-Radar-I for roadside perception, and V2X-Radar-V for single-vehicle perception. Furthermore, we provide comprehensive benchmarks across these three sub-datasets. We will release all datasets and benchmark codebase at https://huggingface.co/datasets/yanglei18/V2X-Radar and https://github.com/yanglei18/V2X-Radar.

Method: Proposes Fourier Occupancy Field (FOF) that factorizes 3D occupancy fields into 2D vector fields, enabling compatibility with 2D CNNs. FOF-X framework integrates human parametric models as priors, uses Laplacian constraints and automaton-based discontinuity matchers for improved mesh conversion.

Result: FOF-X achieves state-of-the-art results on different datasets and real-captured data, with real-time performance and improved robustness against texture/lighting variations.

Conclusion: FOF representation effectively bridges 2D-3D domains for real-time human reconstruction, with FOF-X demonstrating superior performance and robustness in handling domain gaps between training and real images.

Abstract: We introduce FOF-X for real-time reconstruction of detailed human geometry from a single image. Balancing real-time speed against high-quality results is a persistent challenge, mainly due to the high computational demands of existing 3D representations. To address this, we propose Fourier Occupancy Field (FOF), an efficient 3D representation by learning the Fourier series. The core of FOF is to factorize a 3D occupancy field into a 2D vector field, retaining topology and spatial relationships within the 3D domain while facilitating compatibility with 2D convolutional neural networks. Such a representation bridges the gap between 3D and 2D domains, enabling the integration of human parametric models as priors and enhancing the reconstruction robustness. Based on FOF, we design a new reconstruction framework, FOF-X, to avoid the performance degradation caused by texture and lighting. This enables our real-time reconstruction system to better handle the domain gap between training images and real images. Additionally, in FOF-X, we enhance the inter-conversion algorithms between FOF and mesh representations with a Laplacian constraint and an automaton-based discontinuity matcher, improving both quality and robustness. We validate the strengths of our approach on different datasets and real-captured data, where FOF-X achieves new state-of-the-art results. The code has already been released for research purposes at https://cic.tju.edu.cn/faculty/likun/projects/FOFX/index.html.

Method: Introduced BBQ-Vision benchmark with 14,144 image-question pairs using real and multi-actor images across 9 diverse categories and 50 sub-categories. Features real-world visual samples, image variations, and open-ended question formats. Rigorously tested 19 state-of-the-art open-source (general-purpose and reasoning) and closed-source LMMs.

Result: Top-performing models often exhibit bias on several social stereotypes. Thinking models (reasoning models) induce more bias in their reasoning chains. The benchmark enables precise assessment of models’ reasoning capabilities across varying difficulty levels.

Conclusion: BBQ-V represents a significant step toward fostering fairness in AI systems and reducing harmful biases, laying groundwork for more equitable and socially responsible LMMs. The dataset and evaluation code are publicly available.

Abstract: Stereotype biases in Large Multimodal Models (LMMs) perpetuate harmful societal prejudices, undermining the fairness and equity of AI applications. As LMMs grow increasingly influential, addressing and mitigating inherent biases related to stereotypes, harmful generations, and ambiguous assumptions in real-world scenarios has become essential. However, existing datasets evaluating stereotype biases in LMMs often lack diversity, rely on synthetic images, and often have single-actor images, leaving a gap in bias evaluation for real-world visual contexts. To address the gap in bias evaluation using real images, we introduce the BBQ-Vision (BBQ-V), the most comprehensive framework for assessing stereotype biases across nine diverse categories and 50 sub-categories with real and multi-actor images. BBQ-V benchmark contains 14,144 image-question pairs and rigorously evaluates LMMs through carefully curated, visually grounded scenarios, challenging them to reason accurately about visual stereotypes. It offers a robust evaluation framework featuring real-world visual samples, image variations, and open-ended question formats. BBQ-V enables a precise and nuanced assessment of a model’s reasoning capabilities across varying levels of difficulty. Through rigorous testing of 19 state-of-the-art open-source (general-purpose and reasoning) and closed-source LMMs, we highlight that these top-performing models are often biased on several social stereotypes, and demonstrate that the thinking models induce more bias in the reasoning chains. This benchmark represents a significant step toward fostering fairness in AI systems and reducing harmful biases, laying the groundwork for more equitable and socially responsible LMMs. Our dataset and evaluation code are publicly available.

Method: Hybrid 3D representation decoupling color and volume density. Color modeled via triplane representation mapping high-frequency information; volume density as continuous fields with reference features from neighboring views. Depth-guided optimization using point clouds mitigates overfitting.

Result: Achieves 30x speed increase compared to NeRF-based methods while improving rendering quality: 7.4% increase in PSNR and 3.4% in SSIM over advanced few-shot methods across multiple remote sensing scenes.

Conclusion: TriDF provides an efficient solution for few-shot remote sensing novel view synthesis, balancing computational efficiency with rendering quality, making it practical for applications like urban planning and environmental monitoring.

Abstract: Remote sensing novel view synthesis (NVS) offers significant potential for 3D interpretation of remote sensing scenes, with important applications in urban planning and environmental monitoring. However, remote sensing scenes frequently lack sufficient multi-view images due to acquisition constraints. While existing NVS methods tend to overfit when processing limited input views, advanced few-shot NVS methods are computationally intensive and perform sub-optimally in remote sensing scenes. This paper presents TriDF, an efficient hybrid 3D representation for fast remote sensing NVS from as few as 3 input views. Our approach decouples color and volume density information, modeling them independently to reduce the computational burden on implicit radiance fields and accelerate reconstruction. We explore the potential of the triplane representation in few-shot NVS tasks by mapping high-frequency color information onto this compact structure, and the direct optimization of feature planes significantly speeds up convergence. Volume density is modeled as continuous density fields, incorporating reference features from neighboring views through image-based rendering to compensate for limited input data. Additionally, we introduce depth-guided optimization based on point clouds, which effectively mitigates the overfitting problem in few-shot NVS. Comprehensive experiments across multiple remote sensing scenes demonstrate that our hybrid representation achieves a 30x speed increase compared to NeRF-based methods, while simultaneously improving rendering quality metrics over advanced few-shot methods (7.4% increase in PSNR and 3.4% in SSIM). The code is publicly available at https://github.com/kanehub/TriDF

Method: ViSTA consists of: (1) multi-modal history fusion module to extract relevant history features, (2) history adapter to condition generation on extracted features, and (3) salient history selection strategy during inference to choose the most relevant history text-image pair. Also introduces TIFA metric for text-image alignment assessment.

Result: Evaluated on StorySalon and FlintStonesSV datasets, ViSTA achieves consistent frames across sequences while maintaining good alignment with narrative text descriptions.

Conclusion: ViSTA addresses the visual storytelling challenge by providing a training-efficient approach that maintains consistency across frames and adapts well to narrative prompts, with improved evaluation using the TIFA metric.

Abstract: Text-to-image diffusion models have achieved remarkable success, yet generating coherent image sequences for visual storytelling remains challenging. A key challenge is effectively leveraging all previous text-image pairs, referred to as history text-image pairs, which provide contextual information for maintaining consistency across frames. Existing auto-regressive methods condition on all past image-text pairs but require extensive training, while training-free subject-specific approaches ensure consistency but lack adaptability to narrative prompts. To address these limitations, we propose a multi-modal history adapter for text-to-image diffusion models, \textbf{ViSTA}. It consists of (1) a multi-modal history fusion module to extract relevant history features and (2) a history adapter to condition the generation on the extracted relevant features. We also introduce a salient history selection strategy during inference, where the most salient history text-image pair is selected, improving the quality of the conditioning. Furthermore, we propose to employ a Visual Question Answering-based metric TIFA to assess text-image alignment in visual storytelling, providing a more targeted and interpretable assessment of generated images. Evaluated on the StorySalon and FlintStonesSV dataset, our proposed ViSTA model is not only consistent across different frames, but also well-aligned with the narrative text descriptions.

Method: The authors design a high-resolution semantic voxel sensor in CARLA simulator to produce dense annotations, develop a baseline model with inter-agent feature fusion via spatial alignment and attention aggregation, and establish benchmarks with varying prediction ranges to assess spatial extent impact.

Result: Experimental results demonstrate superior performance of the baseline model, with increasing performance gains observed as prediction range expands, validating the effectiveness of collaborative perception for 3D semantic occupancy prediction.

Conclusion: The work bridges the dataset gap for collaborative 3D semantic occupancy prediction, provides a baseline model with effective feature fusion, and establishes systematic benchmarks showing that collaborative perception significantly enhances prediction completeness and accuracy, especially over larger spatial ranges.

Abstract: 3D semantic occupancy prediction is an emerging perception paradigm in autonomous driving, providing a voxel-level representation of both geometric details and semantic categories. However, its effectiveness is inherently constrained in single-vehicle setups by occlusions, restricted sensor range, and narrow viewpoints. To address these limitations, collaborative perception enables the exchange of complementary information, thereby enhancing the completeness and accuracy of predictions. Despite its potential, research on collaborative 3D semantic occupancy prediction is hindered by the lack of dedicated datasets. To bridge this gap, we design a high-resolution semantic voxel sensor in CARLA to produce dense and comprehensive annotations. We further develop a baseline model that performs inter-agent feature fusion via spatial alignment and attention aggregation. In addition, we establish benchmarks with varying prediction ranges designed to systematically assess the impact of spatial extent on collaborative prediction. Experimental results demonstrate the superior performance of our baseline, with increasing gains observed as range expands. Our code is available at https://github.com/tlab-wide/Co3SOP}{https://github.com/tlab-wide/Co3SOP.

Method: Built on frozen SAM encoder with two innovations: 1) Linguistic Adversarial Disentanglement (LAD) uses VLM to generate textual style descriptions and trains features to be contrastively dissimilar, removing non-causal information; 2) Test-Time Causal Intervention (TCI) allows LLM to interpret clinician’s natural language commands to modulate segmentation decoder features in real-time for error correction.

Result: Achieves new SOTA in OOD robustness on composite benchmark from 4 datasets (BTCV, CHAOS, AMOS, BraTS), improving average Dice score by up to 6.2 points and reducing Hausdorff Distance by 15.8 mm over strongest baseline, using less than 9% of trainable parameters.

Conclusion: The framework charts a new course for building robust, efficient, and interactively controllable medical AI systems by elevating LLMs to causal reasoners for domain generalization.

Abstract: The clinical utility of deep learning models for medical image segmentation is severely constrained by their inability to generalize to unseen domains. This failure is often rooted in the models learning spurious correlations between anatomical content and domain-specific imaging styles. To overcome this fundamental challenge, we introduce Causal-SAM-LLM, a novel framework that elevates Large Language Models (LLMs) to the role of causal reasoners. Our framework, built upon a frozen Segment Anything Model (SAM) encoder, incorporates two synergistic innovations. First, Linguistic Adversarial Disentanglement (LAD) employs a Vision-Language Model to generate rich, textual descriptions of confounding image styles. By training the segmentation model’s features to be contrastively dissimilar to these style descriptions, it learns a representation robustly purged of non-causal information. Second, Test-Time Causal Intervention (TCI) provides an interactive mechanism where an LLM interprets a clinician’s natural language command to modulate the segmentation decoder’s features in real-time, enabling targeted error correction. We conduct an extensive empirical evaluation on a composite benchmark from four public datasets (BTCV, CHAOS, AMOS, BraTS), assessing generalization under cross-scanner, cross-modality, and cross-anatomy settings. Causal-SAM-LLM establishes a new state of the art in out-of-distribution (OOD) robustness, improving the average Dice score by up to 6.2 points and reducing the Hausdorff Distance by 15.8 mm over the strongest baseline, all while using less than 9% of the full model’s trainable parameters. Our work charts a new course for building robust, efficient, and interactively controllable medical AI systems.

Method: Adapts SAM vision foundation model. Uses multi-receptive field ensemble to process SAM encoder features. Includes spatial-temporal feature enhancement (STFE) for cross-temporal relations, decoder for change pattern reconstruction, and multi-scale decoder fusion with attention (MSDFA). Introduces cross-entropy masking (CEM) loss for class imbalance.

Result: Outperforms SOTA methods on four change detection datasets: Levir-CD, WHU-CD, CLCD, and S2Looking. Achieves 2.97% F1-score improvement on complex S2Looking dataset.

Conclusion: Proposed SAM-based architecture with multi-receptive field ensemble and CEM loss effectively addresses RSCD challenges of multi-scale changes and class imbalance, demonstrating superior performance across diverse datasets.

Abstract: Remote sensing change detection (RSCD) is a complex task, where changes often appear at different scales and orientations. Convolutional neural networks (CNNs) are good at capturing local spatial patterns but cannot model global semantics due to limited receptive fields. Alternatively, transformers can model long-range dependencies but are data hungry, and RSCD datasets are not large enough to train these models effectively. To tackle this, this paper presents a new architecture for RSCD which adapts a segment anything (SAM) vision foundation model and processes features from the SAM encoder through a multi-receptive field ensemble to capture local and global change patterns. We propose an ensemble of spatial-temporal feature enhancement (STFE) to capture cross-temporal relations, a decoder to reconstruct change patterns, and a multi-scale decoder fusion with attention (MSDFA) to fuse multi-scale decoder information and highlight key change patterns. Each branch in an ensemble operates on a separate receptive field to capture finer-to-coarser level details. Additionally, we propose a novel cross-entropy masking (CEM) loss to handle class-imbalance in RSCD datasets. Our work outperforms state-of-the-art (SOTA) methods on four change detection datasets, Levir-CD, WHU-CD, CLCD, and S2Looking. We achieved 2.97% F1-score improvement on a complex S2Looking dataset. The code is available at: https://github.com/humza909/SAM-ECEM

Method: Two lightweight debiasing techniques: 1) Remove recency-induced attention trends to create position-agnostic importance measures, and 2) Suppress attention sink effects by eliminating spurious attention on padding tokens. The method is model-agnostic, pruning-method-agnostic, and task-agnostic.

Result: Evaluated on ten vision-language benchmarks across image and video tasks, compared with seven state-of-the-art pruning methods and two VLM architectures. Achieves substantial performance gains, demonstrating strong effectiveness and generalizability.

Conclusion: Attention biases inherited from LLMs distort VLM pruning, but lightweight debiasing techniques can restore attention reliability, enabling more effective visual token pruning while maintaining plug-and-play compatibility with existing methods.

Abstract: Vision-language models (VLMs) typically encode substantially more visual tokens than text tokens, resulting in significant token redundancy. Pruning uninformative visual tokens is therefore crucial for improving computational efficiency, and language-to-vision attention has become a widely used importance criterion for this purpose. However, we find that attention in VLMs is systematically biased. It disproportionately favors tokens appearing later in the sequence, manifesting as over-attention to lower image regions, and assigns inflated scores to semantically empty padding tokens. These behaviors stem from intrinsic recency bias and attention sink effects inherited from large language models (LLMs), and they distort attention-based pruning by preserving irrelevant visual content. To derive a pruning criterion better aligned with semantic relevance, we introduce two lightweight yet effective debiasing techniques that restore the reliability of attention. The first compensates for positional distortions by removing recency-induced attention trends, producing a content-aware and position-agnostic importance measure. The second suppresses attention sink effects by eliminating spurious attention on padding tokens. Our method is model-agnostic, pruning-method-agnostic, and task-agnostic, enabling plug-and-play integration with existing VLM pruning models. Despite its simplicity, our approach consistently delivers strong performance gains. We evaluate our method on ten vision-language benchmarks spanning both image-based and video-based tasks, in comparison with seven state-of-the-art visual token pruning methods and across two representative VLM architectures. Our method achieves substantial performance gains, demonstrating strong effectiveness and generalizability. Our code is available at https://github.com/intcomp/attention-bias.

Method: MINGLE (Modeling INterpersonal Group-Level Engagement) is a modular three-stage pipeline: (1) off-the-shelf human detection and depth estimation, (2) VLM-based reasoning to classify pairwise social affiliation, and (3) a lightweight spatial aggregation algorithm to localize socially connected groups.

Result: The paper introduces a new dataset of 100K urban street-view images annotated with bounding boxes and labels for both individuals and socially interacting groups. The annotations combine human-created labels and MINGLE pipeline outputs for semantic richness and broad coverage.

Conclusion: The work introduces a novel social group region detection task and provides both a methodological framework (MINGLE) and a comprehensive dataset to advance research in understanding social interactions from visual data for urban planning applications.

Abstract: Understanding group-level social interactions in public spaces is crucial for urban planning, informing the design of socially vibrant and inclusive environments. Detecting such interactions from images involves interpreting subtle visual cues such as relations, proximity, and co-movement - semantically complex signals that go beyond traditional object detection. To address this challenge, we introduce a social group region detection task, which requires inferring and spatially grounding visual regions defined by abstract interpersonal relations. We propose MINGLE (Modeling INterpersonal Group-Level Engagement), a modular three-stage pipeline that integrates: (1) off-the-shelf human detection and depth estimation, (2) VLM-based reasoning to classify pairwise social affiliation, and (3) a lightweight spatial aggregation algorithm to localize socially connected groups. To support this task and encourage future research, we present a new dataset of 100K urban street-view images annotated with bounding boxes and labels for both individuals and socially interacting groups. The annotations combine human-created labels and outputs from the MINGLE pipeline, ensuring semantic richness and broad coverage of real-world scenarios.

Method: The review employs a PRISMA-based methodology with a challenge-oriented taxonomy to systematically synthesize recent research. It examines approaches like data augmentation, hybrid models, and feature fusion to overcome DL limitations in skin disease diagnosis.

Result: The review identifies innovative solutions to key challenges in DL-based skin cancer diagnosis and highlights emerging directions including hybrid CNN-Transformer architectures and uncertainty-aware models that show promise for improving diagnostic accuracy.

Conclusion: Deep learning has significant potential to revolutionize skin disease diagnosis and clinical decision-making. The systematic review provides a foundation for future dermatological AI research by addressing current limitations and pointing toward advanced hybrid architectures and clinical workflow integration.

Abstract: Skin cancer is one of the most prevalent and deadly forms of cancer worldwide, highlighting the critical importance of early detection and diagnosis in improving patient outcomes. Deep learning (DL) has shown significant promise in enhancing the accuracy and efficiency of automated skin disease diagnosis, particularly in detecting and classifying skin lesions. However, several challenges remain for DL-based skin cancer diagnosis, including complex features, image noise, intra-class variation, inter-class similarity, and data imbalance. This review synthesizes recent research and discusses innovative approaches to address these challenges, such as data augmentation, hybrid models, and feature fusion. Furthermore, the review highlights the integration of DL models into clinical workflows, offering insights into the potential of deep learning to revolutionize skin disease diagnosis and improve clinical decision-making. This review uniquely integrates a PRISMA-based methodology with a challenge-oriented taxonomy, providing a systematic and transparent synthesis of recent deep learning advances for skin disease diagnosis. It further highlights emerging directions such as hybrid CNN-Transformer architectures and uncertainty-aware models, emphasizing its contribution to future dermatological AI research.

Method: Introduce learnable aggregation tokens prepended to patch tokens before a particular transformer block. These tokens interact globally via self-attention, implicitly aggregating information from patch tokens. Only take aggregation tokens from last output and concatenate as global representation. Also propose optimal token insertion strategy and initialization method.

Result: Outperforms state-of-the-art methods on several VPR datasets with higher efficiency. Ranks 1st on the MSLS challenge leaderboard.

Conclusion: Dedicated aggregators are unnecessary in transformer-based VPR. Simple implicit aggregation via learnable tokens with proper insertion strategy and initialization can achieve robust global descriptors more efficiently.

Abstract: Visual place recognition (VPR) is typically regarded as a specific image retrieval task, whose core lies in representing images as global descriptors. Over the past decade, dominant VPR methods (e.g., NetVLAD) have followed a paradigm that first extracts the patch features/tokens of the input image using a backbone, and then aggregates these patch features into a global descriptor via an aggregator. This backbone-plus-aggregator paradigm has achieved overwhelming dominance in the CNN era and remains widely used in transformer-based models. In this paper, however, we argue that a dedicated aggregator is not necessary in the transformer era, that is, we can obtain robust global descriptors only with the backbone. Specifically, we introduce some learnable aggregation tokens, which are prepended to the patch tokens before a particular transformer block. All these tokens will be jointly processed and interact globally via the intrinsic self-attention mechanism, implicitly aggregating useful information within the patch tokens to the aggregation tokens. Finally, we only take these aggregation tokens from the last output tokens and concatenate them as the global representation. Although implicit aggregation can provide robust global descriptors in an extremely simple manner, where and how to insert additional tokens, as well as the initialization of tokens, remains an open issue worthy of further exploration. To this end, we also propose the optimal token insertion strategy and token initialization method derived from empirical studies. Experimental results show that our method outperforms state-of-the-art methods on several VPR datasets with higher efficiency and ranks 1st on the MSLS challenge leaderboard. The code is available at https://github.com/lu-feng/image.

Method: Collected raw data from six fields, preprocessed and aligned using accompanied-text-priority method, constructed Affect Table for annotation constraints, used five MLLMs to analyze both text and visual modalities, fused outputs with Reciprocal Rank Fusion algorithm, and validated through user studies with Composite Affect Consistency Index.

Result: Created InfoAffect dataset with 3.5k samples, achieved overall CACI score of 0.608 indicating high accuracy, dataset publicly available on GitHub repository.

Conclusion: The InfoAffect dataset addresses the scarcity of affect-annotated infographic data and provides a valuable resource for studying how infographics influence user emotions in social media contexts.

Abstract: Infographics are widely used in social media to convey complex information, yet how they influence users’ affects remains underexplored due to the scarcity of relevant datasets. To address this gap, we introduce a 3.5k-sample affect-annotated InfoAffect dataset, which combines textual content with real-world infographics. We first collected the raw data from six fields and aligned it via preprocessing, the accompanied-text-priority method, and three strategies to guarantee quality and compliance. After that, we constructed an Affect Table to constrain annotation. We used five state-of-the-art multimodal large language models (MLLMs) to analyze both modalities, and their outputs were fused with the Reciprocal Rank Fusion (RRF) algorithm to yield robust affects and confidences. We conducted a user study with two experiments to validate usability and assess the InfoAffect dataset using the Composite Affect Consistency Index (CACI), achieving an overall score of 0.608, which indicates high accuracy. The InfoAffect dataset is available in a public repository at https://github.com/bulichuchu/InfoAffect-dataset.

Method: Extends the Hummingbird framework to 3D using MVImgNet dataset. Evaluates models through in-context segmentation of novel views given reference images at specific camera angles. Tests performance across 4 difficulty categories based on key-query view contrast.

Result: Benchmarked 7 state-of-the-art foundation models, showing DINO-based encoders remain competitive across large viewpoint shifts. The benchmark provides code for public evaluation.

Conclusion: The proposed benchmark enables direct evaluation of intrinsic 3D reasoning in foundation models without fine-tuning, revealing that DINO-based models maintain strong performance despite significant viewpoint variations.

Abstract: Benchmarking 3D spatial understanding of foundation models is essential for real-world applications such as robotics and autonomous driving. Existing evaluations often rely on downstream fine-tuning with linear heads or task-specific decoders, making it difficult to isolate the intrinsic 3D reasoning ability of pre-trained encoders. In this work, we introduce a novel benchmark for in-context 3D scene understanding that requires no fine-tuning and directly probes the quality of dense visual features. Building on the Hummingbird framework, which evaluates in-context 2D scene understanding, we extend the setup to the 3D Multi-View ImageNet (MVImgNet) dataset. Given a set of images depicting objects at specific camera angles (keys), we benchmark the performance of segmenting novel views (queries) and report the scores in 4 categories of easy, medium, hard, and extreme based on the key-query view contrast. We benchmark 7 state-of-the-art foundation models and show that DINO-based encoders remain competitive across large viewpoint shifts. Our code is publicly available at https://github.com/ToyeshC/open-hummingbird-3d-eval.

Method: Proposes Video-Browser agent with Pyramidal Perception: uses cheap metadata filtering first, then selectively applies expensive visual perception only when necessary for fine-grained verification.

Result: Achieves 37.5% relative improvement over direct visual inference while reducing token consumption by 58.3% on the Video-BrowseComp benchmark for open-ended agentic video browsing tasks.

Conclusion: Video-Browser establishes a foundation for verifiable open-web video research by effectively balancing the scale of open-ended exploration with fine-grained visual verification through selective perception.

Abstract: The evolution of autonomous agents is redefining information seeking, transitioning from passive retrieval to proactive, open-ended web research. However, a significant modality gap remains in processing the web’s most dynamic and information-dense modality: video. In this paper, we first formalize the task of Agentic Video Browsing and introduce Video-BrowseComp, a benchmark evaluating open-ended agentic browsing tasks that enforce a mandatory dependency on videos. We observe that current paradigms struggle to reconcile the scale of open-ended video exploration with the need for fine-grained visual verification. Direct visual inference (e.g., RAG) maximizes perception but incurs prohibitive context costs, while text-centric summarization optimizes efficiency but often misses critical visual details required for accurate grounding. To address this, we propose Video-Browser, a novel agent leveraging Pyramidal Perception, filtering with cheap metadata and zooming in with expensive visual perception only when necessary. Experiments demonstrate that our approach achieves a 37.5% relative improvement while reducing token consumption by 58.3% compared to Direct visual inference, establishing a foundation for verifiable open-web video research. We open-source all codes, benchmark at {https://anonymous.4open.science/r/VideoBrowser} and {https://github.com/chrisx599/Video-Browser}.

Method: 1) Introduces FFP-300K dataset (300K high-fidelity 720p video pairs, 81 frames) via two-track pipeline for diverse edits. 2) Proposes guidance-free FFP framework with Adaptive Spatio-Temporal RoPE (AST-RoPE) to disentangle appearance and motion references. 3) Uses self-distillation with identity propagation task as regularizer for temporal stability.

Result: Significantly outperforms existing academic and commercial models on EditVerseBench benchmark with ~0.2 PickScore and ~0.3 VLM score improvements against competitors.

Conclusion: The proposed guidance-free FFP framework with large-scale dataset and novel architectural components effectively resolves the tension between maintaining first-frame appearance and preserving source video motion, achieving state-of-the-art performance in controllable video editing.

Abstract: First-Frame Propagation (FFP) offers a promising paradigm for controllable video editing, but existing methods are hampered by a reliance on cumbersome run-time guidance. We identify the root cause of this limitation as the inadequacy of current training datasets, which are often too short, low-resolution, and lack the task diversity required to teach robust temporal priors. To address this foundational data gap, we first introduce FFP-300K, a new large-scale dataset comprising 300K high-fidelity video pairs at 720p resolution and 81 frames in length, constructed via a principled two-track pipeline for diverse local and global edits. Building on this dataset, we propose a novel framework designed for true guidance-free FFP that resolves the critical tension between maintaining first-frame appearance and preserving source video motion. Architecturally, we introduce Adaptive Spatio-Temporal RoPE (AST-RoPE), which dynamically remaps positional encodings to disentangle appearance and motion references. At the objective level, we employ a self-distillation strategy where an identity propagation task acts as a powerful regularizer, ensuring long-term temporal stability and preventing semantic drift. Comprehensive experiments on the EditVerseBench benchmark demonstrate that our method significantly outperforming existing academic and commercial models by receiving about 0.2 PickScore and 0.3 VLM score improvement against these competitors.

Method: Combines pruning with meta-learning via a novel Disease-Aware Channel Importance Scoring (DACIS) mechanism and a three-stage Prune-then-Meta-Learn-then-Prune (PMP) pipeline to balance generalization and deployment feasibility.

Result: Reduces model size by 78% while maintaining 92.3% of original accuracy. Compressed model achieves 7 FPS on Raspberry Pi 4, enabling practical real-time field diagnosis.

Conclusion: The framework successfully addresses the tension between generalization capability and deployment feasibility for agricultural disease classification, enabling practical edge deployment for smallholder farmers.

Abstract: A key challenge in agricultural AI is deploying disease detection systems in remote fields with limited access to laboratories or high-performance computing (HPC) resources. While deep learning (DL) models, specifically deep convolutional networks, achieve high accuracy in identifying plant pathologies from leaf imagery, their memory footprints and computational demands limit edge deployment on devices constrained by battery life, processing power, and connectivity, such as Raspberry Pi. Few-shot learning (FSL) paradigms offer a compelling solution to the data scarcity problem inherent in agricultural applications, where obtaining labeled samples for novel disease variants proves both costly and time-sensitive. This work introduces a framework combining pruning with meta-learning for agricultural disease classification, addressing the tension between generalization capability and deployment feasibility. The proposed approach combines a novel Disease-Aware Channel Importance Scoring (DACIS) mechanism with a three-stage Prune-then-Meta-Learn-then-Prune (PMP) pipeline. Experiments on PlantVillage and PlantDoc datasets demonstrate that the proposed approach reduces model size by 78% while maintaining 92.3% of the original accuracy. The compressed model achieves 7 frames per second (FPS) on a Raspberry Pi 4, enabling practical real-time field diagnosis for smallholder farmers.

Method: VINO couples a vision-language model with a Multimodal Diffusion Transformer (MMDiT), encoding multimodal inputs as interleaved conditioning tokens to guide diffusion. It uses a multi-stage training pipeline that progressively expands a video generation base model into a unified, multi-task generator.

Result: VINO demonstrates strong visual quality, faithful instruction following, improved reference/attribute preservation, and more controllable multi-identity edits across diverse generation and editing benchmarks.

Conclusion: VINO presents a practical path toward scalable unified visual generation, highlighting the promise of interleaved, in-context computation as a foundation for general-purpose visual creation.

Abstract: We present VINO, a unified visual generator that performs image and video generation and editing within a single framework. Instead of relying on task-specific models or independent modules for each modality, VINO uses a shared diffusion backbone that conditions on text, images and videos, enabling a broad range of visual creation and editing tasks under one model. Specifically, VINO couples a vision-language model (VLM) with a Multimodal Diffusion Transformer (MMDiT), where multimodal inputs are encoded as interleaved conditioning tokens, and then used to guide the diffusion process. This design supports multi-reference grounding, long-form instruction following, and coherent identity preservation across static and dynamic content, while avoiding modality-specific architectural components. To train such a unified system, we introduce a multi-stage training pipeline that progressively expands a video generation base model into a unified, multi-task generator capable of both image and video input and output. Across diverse generation and editing benchmarks, VINO demonstrates strong visual quality, faithful instruction following, improved reference and attribute preservation, and more controllable multi-identity edits. Our results highlight a practical path toward scalable unified visual generation, and the promise of interleaved, in-context computation as a foundation for general-purpose visual creation.

Method: Uses LLM module for floor layout generation from natural language prompts, diffusion-based posterior sampling to populate scenes with articulated assets from 3D repositories, and differentiable guidance functions to regulate object quantity, prevent articulation collisions, and maintain walkable space.

Result: Extensive experiments show the framework generates structurally valid, semantically coherent, and functionally interactive environments across diverse scene types and conditions, enabling scalable embodied AI research.

Conclusion: SceneFoundry successfully creates apartment-scale 3D worlds with articulated furniture for robotic training, addressing the limitations of existing approaches and providing a scalable solution for embodied AI research.

Abstract: The ability to automatically generate large-scale, interactive, and physically realistic 3D environments is crucial for advancing robotic learning and embodied intelligence. However, existing generative approaches often fail to capture the functional complexity of real-world interiors, particularly those containing articulated objects with movable parts essential for manipulation and navigation. This paper presents SceneFoundry, a language-guided diffusion framework that generates apartment-scale 3D worlds with functionally articulated furniture and semantically diverse layouts for robotic training. From natural language prompts, an LLM module controls floor layout generation, while diffusion-based posterior sampling efficiently populates the scene with articulated assets from large-scale 3D repositories. To ensure physical usability, SceneFoundry employs differentiable guidance functions to regulate object quantity, prevent articulation collisions, and maintain sufficient walkable space for robotic navigation. Extensive experiments demonstrate that our framework generates structurally valid, semantically coherent, and functionally interactive environments across diverse scene types and conditions, enabling scalable embodied AI research. project page: https://anc891203.github.io/SceneFoundry-Demo/

Method: 1) UV-guided avatar modeling with pixel-wise facial correspondence estimation to reproject colors from screen to UV space; 2) Learnable UV tokens with attention mechanisms at screen and UV levels; 3) Large-scale synthetic training dataset for identity-rich training.

Result: The method significantly outperforms existing approaches in both monocular and multi-view settings, demonstrating superior performance across different input scenarios.

Conclusion: UIKA provides an efficient, feed-forward solution for animatable head modeling that works with everyday capture devices, overcoming limitations of traditional studio-based avatar creation methods.

Abstract: We present UIKA, a feed-forward animatable Gaussian head model from an arbitrary number of unposed inputs, including a single image, multi-view captures, and smartphone-captured videos. Unlike the traditional avatar method, which requires a studio-level multi-view capture system and reconstructs a human-specific model through a long-time optimization process, we rethink the task through the lenses of model representation, network design, and data preparation. First, we introduce a UV-guided avatar modeling strategy, in which each input image is associated with a pixel-wise facial correspondence estimation. Such correspondence estimation allows us to reproject each valid pixel color from screen space to UV space, which is independent of camera pose and character expression. Furthermore, we design learnable UV tokens on which the attention mechanism can be applied at both the screen and UV levels. The learned UV tokens can be decoded into canonical Gaussian attributes using aggregated UV information from all input views. To train our large avatar model, we additionally prepare a large-scale, identity-rich synthetic training dataset. Our method significantly outperforms existing approaches in both monocular and multi-view settings. See more details in our project page: https://zijian-wu.github.io/uika-page/

Method: Adapt SAM 2.1 with minimal encoder modifications, use PoseMaskRefine fine-tuning to incorporate high-visibility pose keypoints into SAM’s iterative correction process, simplify inference by selecting only top 3 keypoints.

Result: Improved robustness and accuracy across multiple datasets, accurate mask prediction from as few as one keypoint, reduced sensitivity to errors like missing body parts or misclassified clothing.

Conclusion: Pose-guided fine-tuning enables effective occlusion-aware human segmentation while preserving SAM’s generalization capabilities.

Abstract: Segment Anything (SAM) provides an unprecedented foundation for human segmentation, but may struggle under occlusion, where keypoints may be partially or fully invisible. We adapt SAM 2.1 for pose-guided segmentation with minimal encoder modifications, retaining its strong generalization. Using a fine-tuning strategy called PoseMaskRefine, we incorporate pose keypoints with high visibility into the iterative correction process originally employed by SAM, yielding improved robustness and accuracy across multiple datasets. During inference, we simplify prompting by selecting only the three keypoints with the highest visibility. This strategy reduces sensitivity to common errors, such as missing body parts or misclassified clothing, and allows accurate mask prediction from as few as a single keypoint. Our results demonstrate that pose-guided fine-tuning of SAM enables effective, occlusion-aware human segmentation while preserving the generalization capabilities of the original model. The code and pretrained models will be available at https://mirapurkrabek.github.io/BBox-Mask-Pose/.

Method: Instead of using token IDs, the decoder receives grayscale images of individual Chinese characters at very low resolutions (as low as 8x8 pixels). This visual input approach is compared against traditional index-based token representations.

Result: Visual inputs achieve 39.2% accuracy, comparable to the index-based baseline of 39.1%. More remarkably, in low-resource settings, visual models show a “hot-start” effect - reaching above 12% accuracy by 0.4% of total training, while index-based models lag below 6%.

Conclusion: Minimal visual structure provides a robust and efficient signal for Chinese language modeling, offering an alternative perspective on character representation that complements traditional index-based approaches.

Abstract: Large language models typically represent Chinese characters as discrete index-based tokens, largely ignoring their visual form. For logographic scripts, visual structure carries semantic and phonetic information, which may aid prediction. We investigate whether low-resolution visual inputs can serve as an alternative for character-level modeling. Instead of token IDs, our decoder receives grayscale images of individual characters, with resolutions as low as 8 x 8 pixels. Remarkably, these inputs achieve 39.2% accuracy, comparable to the index-based baseline of 39.1%. Such low-resource settings also exhibit a pronounced hot-start effect: by 0.4% of total training, accuracy reaches above 12%, while index-based models lag at below 6%. Overall, our results demonstrate that minimal visual structure can provide a robust and efficient signal for Chinese language modeling, offering an alternative perspective on character representation that complements traditional index-based approaches.

Method: Fine-tune vision-language model to infer material composition/fabric attributes from images, then train lightweight predictor to map attributes to physical fabric parameters using material-physics dataset.

Result: Superior accuracy in material composition estimation and fabric attribute prediction, enabling higher-fidelity simulations compared to state-of-the-art image-to-garment methods.

Conclusion: Proposed framework enables simulation-ready garment estimation from single images without iterative optimization, bridging the gap between visual appearance and physical properties.

Abstract: Estimating physically accurate, simulation-ready garments from a single image is challenging due to the absence of image-to-physics datasets and the ill-posed nature of this problem. Prior methods either require multi-view capture and expensive differentiable simulation or predict only garment geometry without the material properties required for realistic simulation. We propose a feed-forward framework that sidesteps these limitations by first fine-tuning a vision-language model to infer material composition and fabric attributes from real images, and then training a lightweight predictor that maps these attributes to the corresponding physical fabric parameters using a small dataset of material-physics measurements. Our approach introduces two new datasets (FTAG and T2P) and delivers simulation-ready garments from a single image without iterative optimization. Experiments show that our estimator achieves superior accuracy in material composition estimation and fabric attribute prediction, and by passing them through our physics parameter estimator, we further achieve higher-fidelity simulations compared to state-of-the-art image-to-garment methods.

Method: Full-pipeline co-design through network surgery, feature-wise generative distillation, and structured architectural scaling applied jointly to both U-Net and VAE encoder-decoder. This preserves the generative prior while optimizing for hardware efficiency.

Result: Achieves real-time inference down to 20ms on mobile NPUs with 130M-315M parameters. Outperforms prior lightweight diffusion models in perceptual quality and deployability across multiple tasks: super-resolution, deblurring, face restoration, and depth estimation.

Conclusion: NanoSD establishes a general-purpose diffusion foundation model family suitable for real-time visual generation and restoration on edge devices, demonstrating that parameter reduction alone doesn’t guarantee hardware efficiency - architectural balance and latent-space preservation are crucial.

Abstract: Latent diffusion models such as Stable Diffusion 1.5 offer strong generative priors that are highly valuable for image restoration, yet their full pipelines remain too computationally heavy for deployment on edge devices. Existing lightweight variants predominantly compress the denoising U-Net or reduce the diffusion trajectory, which disrupts the underlying latent manifold and limits generalization beyond a single task. We introduce NanoSD, a family of Pareto-optimal diffusion foundation models distilled from Stable Diffusion 1.5 through network surgery, feature-wise generative distillation, and structured architectural scaling jointly applied to the U-Net and the VAE encoder-decoder. This full-pipeline co-design preserves the generative prior while producing models that occupy distinct operating points along the accuracy-latency-size frontier (e.g., 130M-315M parameters, achieving real-time inference down to 20ms on mobile-class NPUs). We show that parameter reduction alone does not correlate with hardware efficiency, and we provide an analysis revealing how architectural balance, feature routing, and latent-space preservation jointly shape true on-device latency. When used as a drop-in backbone, NanoSD enables state-of-the-art performance across image super-resolution, image deblurring, face restoration, and monocular depth estimation, outperforming prior lightweight diffusion models in both perceptual quality and practical deployability. NanoSD establishes a general-purpose diffusion foundation model family suitable for real-time visual generation and restoration on edge devices.

Method: MERGETUNE is a model-agnostic CFT strategy guided by linear mode connectivity. It continues fine-tuning trainable parameters (soft prompts or linear heads) to find a continued model with low-loss paths to both zero-shot (CLIP) and fine-tuned (CoOp) solutions. It approximates the LMC constraint via a second-order surrogate to avoid large-scale data replay.

Result: MERGETUNE improves harmonic mean of CoOp by +5.6% on base-novel generalization without adding parameters. On robust fine-tuning, the LMC-merged model surpasses ensemble baselines with lower inference cost and achieves state-of-the-art results when ensembled with zero-shot model.

Conclusion: The paper introduces continued fine-tuning as a novel paradigm for recovering lost pretrained knowledge, presents MERGETUNE as an effective, model-agnostic solution that exploits loss landscape geometry, and demonstrates significant improvements in generalization and robustness without architectural changes.

Abstract: Fine-tuning vision-language models (VLMs) such as CLIP often leads to catastrophic forgetting of pretrained knowledge. Prior work primarily aims to mitigate forgetting during adaptation; however, forgetting often remains inevitable during this process. We introduce a novel paradigm, continued fine-tuning (CFT), which seeks to recover pretrained knowledge after a zero-shot model has already been adapted. We propose a simple, model-agnostic CFT strategy (named MERGETUNE) guided by linear mode connectivity (LMC), which can be applied post hoc to existing fine-tuned models without requiring architectural changes. Given a fine-tuned model, we continue fine-tuning its trainable parameters (e.g., soft prompts or linear heads) to search for a continued model which has two low-loss paths to the zero-shot (e.g., CLIP) and the fine-tuned (e.g., CoOp) solutions. By exploiting the geometry of the loss landscape, the continued model implicitly merges the two solutions, restoring pretrained knowledge lost in the fine-tuned counterpart. A challenge is that the vanilla LMC constraint requires data replay from the pretraining task. We approximate this constraint for the zero-shot model via a second-order surrogate, eliminating the need for large-scale data replay. Experiments show that MERGETUNE improves the harmonic mean of CoOp by +5.6% on base-novel generalisation without adding parameters. On robust fine-tuning evaluations, the LMC-merged model from MERGETUNE surpasses ensemble baselines with lower inference cost, achieving further gains and state-of-the-art results when ensembled with the zero-shot model. Our code is available at https://github.com/Surrey-UP-Lab/MERGETUNE.

Method: TANDEM employs a novel tandem reinforcement learning strategy where vision-language and audio-language models optimize each other through self-constrained cross-modal context, enabling stable reasoning over extended temporal sequences without requiring dense frame-level supervision.

Result: TANDEM significantly outperforms zero-shot and context-augmented baselines across three benchmark datasets, achieving 0.73 F1 in target identification on HateMM (30% improvement over state-of-the-art) while maintaining precise temporal grounding. However, differentiating offensive vs. hateful content remains challenging in multi-class settings.

Conclusion: Structured, interpretable alignment is achievable in complex multimodal settings, offering a blueprint for transparent and actionable online safety moderation tools that provide granular evidence for human moderators rather than just binary classifications.

Abstract: Social media platforms are increasingly dominated by long-form multimodal content, where harmful narratives are constructed through a complex interplay of audio, visual, and textual cues. While automated systems can flag hate speech with high accuracy, they often function as “black boxes” that fail to provide the granular, interpretable evidence, such as precise timestamps and target identities, required for effective human-in-the-loop moderation. In this work, we introduce TANDEM, a unified framework that transforms audio-visual hate detection from a binary classification task into a structured reasoning problem. Our approach employs a novel tandem reinforcement learning strategy where vision-language and audio-language models optimize each other through self-constrained cross-modal context, stabilizing reasoning over extended temporal sequences without requiring dense frame-level supervision. Experiments across three benchmark datasets demonstrate that TANDEM significantly outperforms zero-shot and context-augmented baselines, achieving 0.73 F1 in target identification on HateMM (a 30% improvement over state-of-the-art) while maintaining precise temporal grounding. We further observe that while binary detection is robust, differentiating between offensive and hateful content remains challenging in multi-class settings due to inherent label ambiguity and dataset imbalance. More broadly, our findings suggest that structured, interpretable alignment is achievable even in complex multimodal settings, offering a blueprint for the next generation of transparent and actionable online safety moderation tools.

Method: Dual-purpose AI system: vector database with academic/gov/media/social sources; ReAct agent chatbot with multi-tool orchestration across 5 knowledge sources; automated report generation with news/research/sentiment/user pattern analysis modules.

Result: Chatbot scored 4.80 overall (single-turn: relevance 4.83, correctness 4.90, etc.; multi-turn: 4.98 overall). Report system: completeness 4.00-5.00, correctness 4.00-5.00, helpfulness 3.67-5.00, reference validity 5.00.

Conclusion: Feasible integrated AI system for bidirectional HPV vaccine communication, enabling verified info delivery with source attribution plus systematic public discourse analysis; transferable framework for other medical contexts.

Abstract: Human papillomavirus (HPV) vaccine hesitancy poses significant public health challenges, particularly in Japan where proactive vaccination recommendations were suspended from 2013 to 2021. The resulting information gap is exacerbated by misinformation on social media, and traditional ways cannot simultaneously address individual queries while monitoring population-level discourse. This study aimed to develop a dual-purpose AI agent system that provides verified HPV vaccine information through a conversational interface while generating analytical reports for medical institutions based on user interactions and social media. We implemented a system comprising: a vector database integrating academic papers, government sources, news media, and social media; a Retrieval-Augmented Generation chatbot using ReAct agent architecture with multi-tool orchestration across five knowledge sources; and an automated report generation system with modules for news analysis, research synthesis, social media sentiment analysis, and user interaction pattern identification. Performance was assessed using a 0-5 scoring scale. For single-turn evaluation, the chatbot achieved mean scores of 4.83 for relevance, 4.89 for routing, 4.50 for reference quality, 4.90 for correctness, and 4.88 for professional identity (overall 4.80). Multi-turn evaluation yielded higher scores: context retention 4.94, topic coherence 5.00, and overall 4.98. The report generation system achieved completeness 4.00-5.00, correctness 4.00-5.00, and helpfulness 3.67-5.00, with reference validity 5.00 across all periods. This study demonstrates the feasibility of an integrated AI agent system for bidirectional HPV vaccine communication. The architecture enables verified information delivery with source attribution while providing systematic public discourse analysis, with a transferable framework for adaptation to other medical contexts.

Method: Analyzed instruction-tuned LLMs (Llama 3.1 8B, Qwen 2.5 7B, Mistral 7B) using PEACE-Reviews dataset annotated for cognitive appraisals, emotions, and behavioral intentions. Examined layer- and head-level activation differences between high- and low-trust texts, conducted probing analyses for linearly decodable trust signals, and studied fine-tuning effects.

Result: LLMs show systematic activation differences distinguishing high- from low-trust texts, revealing that trust cues are implicitly encoded during pretraining. Trust signals are linearly decodable, and fine-tuning refines rather than restructures these representations. Strongest associations emerge with appraisals of fairness, certainty, and accountability-self - dimensions central to human trust formation online.

Conclusion: Modern LLMs internalize psychologically grounded trust signals without explicit supervision, providing a representational foundation for designing credible, transparent, and trustworthy AI systems in the web ecosystem.

Abstract: Perceived trustworthiness underpins how users navigate online information, yet it remains unclear whether large language models (LLMs),increasingly embedded in search, recommendation, and conversational systems, represent this construct in psychologically coherent ways. We analyze how instruction-tuned LLMs (Llama 3.1 8B, Qwen 2.5 7B, Mistral 7B) encode perceived trustworthiness in web-like narratives using the PEACE-Reviews dataset annotated for cognitive appraisals, emotions, and behavioral intentions. Across models, systematic layer- and head-level activation differences distinguish high- from low-trust texts, revealing that trust cues are implicitly encoded during pretraining. Probing analyses show linearly de-codable trust signals and fine-tuning effects that refine rather than restructure these representations. Strongest associations emerge with appraisals of fairness, certainty, and accountability-self – dimensions central to human trust formation online. These findings demonstrate that modern LLMs internalize psychologically grounded trust signals without explicit supervision, offering a representational foundation for designing credible, transparent, and trust-worthy AI systems in the web ecosystem. Code and appendix are available at: https://github.com/GerardYeo/TrustworthinessLLM.

Method: Two-agent system: 1) Improver agent rewrites referral requests using LLM, 2) Evaluator agent measures revision quality using a model trained to predict referral success probability. Enhanced with Retrieval-Augmented Generation (RAG) to prevent degradation of strong requests.

Result: LLM revisions increase predicted success rates for weaker requests but reduce them for stronger ones. RAG prevents edits that worsen stronger requests while amplifying improvements for weaker ones, resulting in 14% predicted success rate increase for weaker requests without degrading strong ones.

Conclusion: AI agents with RAG-enhanced LLMs can effectively improve job referral requests, providing low-cost signals for promising features before real-world testing, though model-predicted improvements don’t guarantee actual referral success.

Abstract: This paper develops AI agents that help job seekers write effective requests for job referrals in a professional online community. The basic workflow consists of an improver agent that rewrites the referral request and an evaluator agent that measures the quality of revisions using a model trained to predict the probability of receiving referrals from other users. Revisions suggested by the LLM (large language model) increase predicted success rates for weaker requests while reducing them for stronger requests. Enhancing the LLM with Retrieval-Augmented Generation (RAG) prevents edits that worsen stronger requests while it amplifies improvements for weaker requests. Overall, using LLM revisions with RAG increases the predicted success rate for weaker requests by 14% without degrading performance on stronger requests. Although improvements in model-predicted success do not guarantee more referrals in the real world, they provide low-cost signals for promising features before running higher-stakes experiments on real users.

Method: ORBITFLOW uses a lightweight ILP solver to decide which layers’ KV caches to keep on GPU for each request within memory constraints. It continuously refines KV placements based on runtime feedback and includes a fallback mechanism to temporarily defer memory-intensive requests under heavy load to preserve overall SLO attainment.

Result: ORBITFLOW improves SLO attainment for TPOT and TBT by up to 66% and 48% respectively, reduces 95th percentile latency by 38%, and achieves up to 3.3x higher throughput compared to existing offloading methods.

Conclusion: ORBITFLOW effectively addresses the challenges of long-context LLM serving by providing fine-grained, adaptive KV cache management that dynamically responds to runtime memory demands, significantly improving SLO attainment, latency, and throughput.

Abstract: Serving long-context LLMs is challenging because request lengths and batch composition vary during token generation, causing the memory footprint to fluctuate significantly at runtime. Offloading KV caches to host memory limits effective memory usage, but existing static and predetermined offloading strategies cannot adapt to the rapidly shifting memory demands of long-context serving. This often leads to excessive CPU-to-GPU KV transfers that translate into latency spikes and frequent SLO violations. To address these challenges, we introduce ORBITFLOW, a fine-grained and adaptive KV cache management system that meets latency SLOs in long-context LLM serving. ORBITFLOW employs a lightweight ILP solver to decide which layers’ KV caches to retain on the GPU for each request, within memory capacity constraints. It continuously refines KV placements based on runtime feedback when the active plan becomes suboptimal during token generation. Under heavy load, ORBITFLOW invokes a fallback mechanism to temporarily defer in-flight requests with large memory footprints, preserving overall SLO attainment. Our experiments demonstrate that ORBITFLOW improves SLO attainment for TPOT and TBT by up to 66% and 48%, respectively, while reducing the 95th percentile latency by 38% and achieving up to 3.3x higher throughput compared to existing offloading methods.

Method: CTHA enforces three key constraints: (1) Message Contract Constraints formalizing information flow via typed packets, (2) Authority Manifold Constraints bounding decision spaces by temporal scope, and (3) Arbiter Resolution Constraints ensuring conflict-free multi-layer decisions.

Result: CTHA achieves 47% reduction in failure cascades, 2.3x improvement in sample efficiency, and superior scalability compared to unconstrained hierarchical baselines in complex task execution.

Conclusion: CTHA provides a principled framework for stable multi-time-scale agent coordination, contributing to understanding multi-agent coordination and advancing robust autonomous systems.

Abstract: Recently, multi-time-scale agent architectures have extended the ubiquitous single-loop paradigm by introducing temporal hierarchies with distinct cognitive layers. While yielding substantial performance gains, this diversification fundamentally compromises the coordination stability intrinsic to unified agent systems, which causes severe inter-layer conflicts, unbounded error propagation, and restricted scalability. To address these challenges, we propose Constrained Temporal Hierarchical Architecture (CTHA), a general framework that projects the inter-layer communication space onto structured manifolds to restore coordination stability, while incorporating principled arbitration mechanisms to ensure coherent decision-making. Specifically, CTHA enforces three key constraints: (1) Message Contract Constraints that formalize information flow between layers via typed summary, plan, and policy packets; (2) Authority Manifold Constraints that bound each layer’s decision space according to its temporal scope; and (3) Arbiter Resolution Constraints that guarantee conflict-free composition of multi-layer decisions. Empirical experiments demonstrate that CTHA is effective for complex task execution at scale, offering 47% reduction in failure cascades, 2.3x improvement in sample efficiency, and superior scalability compared to unconstrained hierarchical baselines. We anticipate that CTHA, as a principled extension of temporal hierarchies, will contribute to a deeper understanding of multi-agent coordination and suggest promising directions for the evolution of robust autonomous systems.

Method: Proposes MemoryExplorer - fine-tunes multimodal LLM with reinforcement learning using multi-task rewards (action prediction, frontier selection, QA) to encourage active memory querying and proactive exploration.

Result: Extensive experiments show significant advantages over state-of-the-art embodied exploration models in long-horizon tasks, demonstrating improved memory recall and proactive exploration capabilities.

Conclusion: LMEE successfully unifies exploration cognition with decision-making for lifelong embodied learning, with MemoryExplorer enabling effective memory-driven exploration that outperforms existing approaches.

Abstract: An ideal embodied agent should possess lifelong learning capabilities to handle long-horizon and complex tasks, enabling continuous operation in general environments. This not only requires the agent to accurately accomplish given tasks but also to leverage long-term episodic memory to optimize decision-making. However, existing mainstream one-shot embodied tasks primarily focus on task completion results, neglecting the crucial process of exploration and memory utilization. To address this, we propose Long-term Memory Embodied Exploration (LMEE), which aims to unify the agent’s exploratory cognition and decision-making behaviors to promote lifelong learning.We further construct a corresponding dataset and benchmark, LMEE-Bench, incorporating multi-goal navigation and memory-based question answering to comprehensively evaluate both the process and outcome of embodied exploration. To enhance the agent’s memory recall and proactive exploration capabilities, we propose MemoryExplorer, a novel method that fine-tunes a multimodal large language model through reinforcement learning to encourage active memory querying. By incorporating a multi-task reward function that includes action prediction, frontier selection, and question answering, our model achieves proactive exploration. Extensive experiments against state-of-the-art embodied exploration models demonstrate that our approach achieves significant advantages in long-horizon embodied tasks.

Method: Uses trend-based heuristics expressed as simple trends (increasing, decreasing, constant) as quantifiers. Defines solutions as sets of scenarios with possible transitions between them, represented by transition graphs where system behavior is depicted as paths within the graph.

Result: Develops a framework where any possible future or past behavior of the system can be represented as a path within the transition graph of scenarios, providing a structured way to analyze product-innovation processes.

Conclusion: Trend-based heuristics provide an effective, minimally information-intensive approach for modeling complex product-innovation processes, with transition graphs offering comprehensive representation of system behavior over time.

Abstract: This paper investigates complex product-innovation processes using models grounded in a set of heuristics. Each heuristic is expressed through simple trends – increasing, decreasing, or constant – which serve as minimally information-intensive quantifiers, avoiding reliance on numerical values or rough sets. A solution to a trend model is defined as a set of scenarios with possible transitions between them, represented by a transition graph. Any possible future or past behaviour of the system under study can thus be depicted by a path within this graph.

Method: Refinement loops - iterative program optimization guided by feedback signals. This includes evolutionary program synthesis approaches and application-layer refinements to commercial AI systems. Also zero-pretraining deep learning methods with small networks (7M parameters).

Result: Top score of 24% on ARC-AGI-2 private evaluation set. 1,455 teams participated with 90 paper submissions. Four frontier AI labs (Anthropic, Google DeepMind, OpenAI, xAI) now report ARC-AGI performance, establishing it as industry standard. However, current performance remains constrained to knowledge coverage rather than true reasoning.

Conclusion: Refinement loops represent the defining theme of 2025 in AGI progress. Current frontier AI reasoning is fundamentally knowledge-dependent, leading to new forms of benchmark contamination. ARC-AGI-3 will introduce interactive reasoning challenges requiring exploration, planning, memory, goal acquisition, and alignment capabilities.

Abstract: The ARC-AGI benchmark series serves as a critical measure of few-shot generalization on novel tasks, a core aspect of intelligence. The ARC Prize 2025 global competition targeted the newly released ARC-AGI-2 dataset, which features greater task complexity compared to its predecessor. The Kaggle competition attracted 1,455 teams and 15,154 entries, with the top score reaching 24% on the ARC-AGI-2 private evaluation set. Paper submissions nearly doubled year-over-year to 90 entries, reflecting the growing research interest in fluid intelligence and abstract reasoning. The defining theme of 2025 is the emergence of the refinement loop – a per-task iterative program optimization loop guided by a feedback signal. Refinement loops come in a variety of forms, in particular evolutionary program synthesis approaches and application-layer refinements to commercial AI systems. Such refinement loops are also possible in weight space, as evidenced by zero-pretraining deep learning methods which are now achieving competitive performance with remarkably small networks (7M parameters). In parallel, four frontier AI labs (Anthropic, Google DeepMind, OpenAI, and xAI) reported ARC-AGI performance in public model cards in 2025, establishing ARC-AGI as an industry standard benchmark for AI reasoning. However, our analysis indicates that current frontier AI reasoning performance remains fundamentally constrained to knowledge coverage, giving rise to new forms of benchmark contamination. In this paper, we survey the top-performing methods, examine the role of refinement loops in AGI progress, discuss knowledge-dependent overfitting, and preview ARC-AGI-3, which introduces interactive reasoning challenges that require exploration, planning, memory, goal acquisition, and alignment capabilities.

Method: Two-stage framework: 1) Prototype generation using multi-agent reasoner with retrieval-anchored fragment-level edits; 2) RL-based fine-grained optimization using Group Relative Policy Optimization (GRPO) for fragment-level refinements to minimize property errors while regulating edit complexity and prototype deviation. Uses automatically curated dataset with reasoning chains of fragment edits and property deltas.

Result: Experiments on generation under two sets of property constraints (QED, LogP, Molecular Weight and HOMO, LUMO) show consistent gains in validity and precise satisfaction of multi-property targets, outperforming strong LLMs and graph-based algorithms.

Conclusion: MolGen better reasons about molecules by leveraging fragments and supports controllable refinement toward numeric targets, addressing limitations of prior work in multi-property constrained molecule generation.

Abstract: Generating molecules that satisfy precise numeric constraints over multiple physicochemical properties is critical and challenging. Although large language models (LLMs) are expressive, they struggle with precise multi-objective control and numeric reasoning without external structure and feedback. We introduce \textbf{M olGen}, a fragment-level, retrieval-augmented, two-stage framework for molecule generation under multi-property constraints. Stage I : Prototype generation: a multi-agent reasoner performs retrieval-anchored, fragment-level edits to produce a candidate near the feasible region. Stage II : RL-based fine-grained optimization: a fragment-level optimizer trained with Group Relative Policy Optimization (GRPO) applies one- or multi-hop refinements to explicitly minimize the property errors toward our target while regulating edit complexity and deviation from the prototype. A large, automatically curated dataset with reasoning chains of fragment edits and measured property deltas underpins both stages, enabling deterministic, reproducible supervision and controllable multi-hop reasoning. Unlike prior work, our framework better reasons about molecules by leveraging fragments and supports controllable refinement toward numeric targets. Experiments on generation under two sets of property constraints (QED, LogP, Molecular Weight and HOMO, LUMO) show consistent gains in validity and precise satisfaction of multi-property targets, outperforming strong LLMs and graph-based algorithms.

Method: Used a compact curated dataset from Walton Multimodal Cold Start, placed first in the challenge, then conducted post-competition ablations to analyze different data curation strategies including difficulty-based selection, dataset size variations, and diversity/synthetic augmentation heuristics.

Result: Difficulty-based example selection on aligned base dataset was the dominant performance driver. Increasing dataset size didn’t reliably improve mean accuracy but reduced run-to-run variance. Diversity and synthetic augmentation heuristics provided no benefit and often degraded performance.

Conclusion: DCVLR represents a saturation-regime evaluation where alignment and difficulty are central to data-efficient multimodal reasoning, suggesting that quality (difficulty on aligned data) matters more than quantity or diversity heuristics.

Abstract: We study data curation for multimodal reasoning through the NeurIPS 2025 Data Curation for Vision-Language Reasoning (DCVLR) challenge, which isolates dataset selection by fixing the model and training protocol. Using a compact curated dataset derived primarily from Walton Multimodal Cold Start, our submission placed first in the challenge. Through post-competition ablations, we show that difficulty-based example selection on an aligned base dataset is the dominant driver of performance gains. Increasing dataset size does not reliably improve mean accuracy under the fixed training recipe, but mainly reduces run-to-run variance, while commonly used diversity and synthetic augmentation heuristics provide no additional benefit and often degrade performance. These results characterize DCVLR as a saturation-regime evaluation and highlight the central role of alignment and difficulty in data-efficient multimodal reasoning.

Method: Proposes AdaMARP with: 1) immersive message format interleaving [Thought], (Action), , and Speech; 2) explicit Scene Manager governing role-playing through discrete actions (init_scene, pick_speaker, switch_scene, add_role, end) with rationales; 3) training datasets AdaRPSet for Actor Model and AdaSMSet for orchestration decisions; 4) AdaptiveBench for trajectory-level evaluation.

Result: Experiments show consistent improvements: AdaRPSet enhances character consistency, environment grounding, and narrative coherence (8B actor outperforms commercial LLMs); AdaSMSet enables smoother scene transitions and more natural role introductions (14B LLM surpasses Claude Sonnet 4.5).

Conclusion: AdaMARP framework effectively addresses limitations of existing role-playing systems through adaptive multi-agent orchestration and immersive messaging, achieving superior performance with smaller models compared to commercial LLMs.

Abstract: LLM role-playing aims to portray arbitrary characters in interactive narratives, yet existing systems often suffer from limited immersion and adaptability. They typically under-model dynamic environmental information and assume largely static scenes and casts, offering insufficient support for multi-character orchestration, scene transitions, and on-the-fly character introduction. We propose an adaptive multi-agent role-playing framework, AdaMARP, featuring an immersive message format that interleaves [Thought], (Action), , and Speech, together with an explicit Scene Manager that governs role-playing through discrete actions (init_scene, pick_speaker, switch_scene, add_role, end) accompanied by rationales. To train these capabilities, we construct AdaRPSet for the Actor Model and AdaSMSet for supervising orchestration decisions, and introduce AdaptiveBench for trajectory-level evaluation. Experiments across multiple backbones and model scales demonstrate consistent improvements: AdaRPSet enhances character consistency, environment grounding, and narrative coherence, with an 8B actor outperforming several commercial LLMs, while AdaSMSet enables smoother scene transitions and more natural role introductions, surpassing Claude Sonnet 4.5 using only a 14B LLM.

Method: HADES uses Hamiltonian dynamics to sample from a structure-aware approximated posterior, with momentum and uncertainty in simulated physical movements enabling rapid proposal transitions. It includes position discretization to propose discrete protein sequences from continuous states, and uses a two-stage encoder-decoder framework to learn structure-function relationships between mutant neighbors.

Result: Extensive experiments show HADES outperforms state-of-the-art baselines in in-silico evaluations across most metrics, demonstrating superior protein sequence optimization capabilities.

Conclusion: HADES offers a unique advantage by leveraging mutual constraints between protein structure and sequence, facilitating the design of protein sequences with similar structures and optimized properties, with publicly available code and data.

Abstract: The ability to engineer optimized protein variants has transformative potential for biotechnology and medicine. Prior sequence-based optimization methods struggle with the high-dimensional complexities due to the epistasis effect and the disregard for structural constraints. To address this, we propose HADES, a Bayesian optimization method utilizing Hamiltonian dynamics to efficiently sample from a structure-aware approximated posterior. Leveraging momentum and uncertainty in the simulated physical movements, HADES enables rapid transition of proposals toward promising areas. A position discretization procedure is introduced to propose discrete protein sequences from such a continuous state system. The posterior surrogate is powered by a two-stage encoder-decoder framework to determine the structure and function relationships between mutant neighbors, consequently learning a smoothed landscape to sample from. Extensive experiments demonstrate that our method outperforms state-of-the-art baselines in in-silico evaluations across most metrics. Remarkably, our approach offers a unique advantage by leveraging the mutual constraints between protein structure and sequence, facilitating the design of protein sequences with similar structures and optimized properties. The code and data are publicly available at https://github.com/GENTEL-lab/HADES.

Method: Boundary-Aware Policy Optimization (BAPO) introduces two key components: 1) a group-based boundary-aware reward that encourages IDK responses only when reasoning reaches its limit, and 2) an adaptive reward modulator that strategically suspends this reward during early exploration to prevent the model from exploiting IDK as a shortcut.

Result: Extensive experiments on four benchmarks demonstrate that BAPO substantially enhances the overall reliability of agentic search systems.

Conclusion: BAPO successfully addresses the reliability gap in RL-based agentic search by cultivating boundary awareness, enabling systems to recognize their reasoning limits and respond appropriately without compromising accuracy.

Abstract: RL-based agentic search enables LLMs to solve complex questions via dynamic planning and external search. While this approach significantly enhances accuracy with agent policies optimized via large-scale reinforcement learning, we identify a critical gap in reliability: these agents fail to recognize their reasoning boundaries and rarely admit ``I DON’T KNOW’’ (IDK) even when evidence is insufficient or reasoning reaches its limit. The lack of reliability often leads to plausible but unreliable answers, introducing significant risks in many real-world scenarios. To this end, we propose Boundary-Aware Policy Optimization (BAPO), a novel RL framework designed to cultivate reliable boundary awareness without compromising accuracy. BAPO introduces two key components: (i) a group-based boundary-aware reward that encourages an IDK response only when the reasoning reaches its limit, and (ii) an adaptive reward modulator that strategically suspends this reward during early exploration, preventing the model from exploiting IDK as a shortcut. Extensive experiments on four benchmarks demonstrate that BAPO substantially enhances the overall reliability of agentic search.

Method: Created benchmark with 138 tasks across 32 real-world scenarios requiring ~90 tool calls, 1M tokens, and hours of execution. Uses user simulation agents for iterative feedback and Docker sandboxes for visual/functional rubric-based automated evaluation.

Result: Closed-source models significantly outperform open-source models (48.4% vs 32.1%). Found disparities in resource efficiency, feedback-driven self-correction, and tool-use preferences. Proprietary models perform best in native ecosystems while open-source models show distinct performance peaks in specific frameworks.

Conclusion: AgencyBench serves as critical testbed for next-generation agents, highlighting need for co-optimizing model architecture with agentic frameworks. The benchmark and toolkit are publicly released to advance autonomous agent development.

Abstract: Large Language Models (LLMs) based autonomous agents demonstrate multifaceted capabilities to contribute substantially to economic production. However, existing benchmarks remain focused on single agentic capability, failing to capture long-horizon real-world scenarios. Moreover, the reliance on human-in-the-loop feedback for realistic tasks creates a scalability bottleneck, hindering automated rollout collection and evaluation. To bridge this gap, we introduce AgencyBench, a comprehensive benchmark derived from daily AI usage, evaluating 6 core agentic capabilities across 32 real-world scenarios, comprising 138 tasks with specific queries, deliverables, and rubrics. These scenarios require an average of 90 tool calls, 1 million tokens, and hours of execution time to resolve. To enable automated evaluation, we employ a user simulation agent to provide iterative feedback, and a Docker sandbox to conduct visual and functional rubric-based assessment. Experiments reveal that closed-source models significantly outperform open-source models (48.4% vs 32.1%). Further analysis reveals significant disparities across models in resource efficiency, feedback-driven self-correction, and specific tool-use preferences. Finally, we investigate the impact of agentic scaffolds, observing that proprietary models demonstrate superior performance within their native ecosystems (e.g., Claude-4.5-Opus via Claude-Agent-SDK), while open-source models exhibit distinct performance peaks, suggesting potential optimization for specific execution frameworks. AgencyBench serves as a critical testbed for next-generation agents, highlighting the necessity of co-optimizing model architecture with agentic frameworks. We believe this work sheds light on the future direction of autonomous agents, and we release the full benchmark and evaluation toolkit at https://github.com/GAIR-NLP/AgencyBench.

Method: MiCA infers mobility relations through causal discovery and integrates them into temporal forecasting models using gated residual mixing. This allows lightweight models to selectively exploit spatial structure without heavy relational components like graph neural networks.

Result: MiCA consistently improves lightweight temporal backbones across four real-world epidemic datasets (COVID-19 incidence, COVID-19 mortality, influenza, dengue), achieving 7.5% average relative error reduction. It performs competitively with state-of-the-art spatio-temporal models while remaining lightweight.

Conclusion: MiCA provides an effective, lightweight solution for integrating mobility information into epidemic forecasting that works well under noisy, data-limited conditions, offering practical advantages over heavier relational models.

Abstract: Accurate forecasting of infectious disease dynamics is critical for public health planning and intervention. Human mobility plays a central role in shaping the spatial spread of epidemics, but mobility data are noisy, indirect, and difficult to integrate reliably with disease records. Meanwhile, epidemic case time series are typically short and reported at coarse temporal resolution. These conditions limit the effectiveness of parameter-heavy mobility-aware forecasters that rely on clean and abundant data. In this work, we propose the Mobility-Informed Causal Adapter (MiCA), a lightweight and architecture-agnostic module for epidemic forecasting. MiCA infers mobility relations through causal discovery and integrates them into temporal forecasting models via gated residual mixing. This design allows lightweight forecasters to selectively exploit mobility-derived spatial structure while remaining robust under noisy and data-limited conditions, without introducing heavy relational components such as graph neural networks or full attention. Extensive experiments on four real-world epidemic datasets, including COVID-19 incidence, COVID-19 mortality, influenza, and dengue, show that MiCA consistently improves lightweight temporal backbones, achieving an average relative error reduction of 7.5% across forecasting horizons. Moreover, MiCA attains performance competitive with SOTA spatio-temporal models while remaining lightweight.

Method: ReCreate uses an agent-as-optimizer paradigm with three components: (1) experience storage and retrieval for on-demand inspection, (2) reasoning-creating synergy pipeline that maps execution experience into scaffold edits, and (3) hierarchical updates that abstract instance-level details into reusable domain patterns.

Result: In experiments across diverse domains, ReCreate consistently outperforms human-designed agents and existing automated agent generation methods, even when starting from minimal seed scaffolds.

Conclusion: ReCreate demonstrates that systematically leveraging agent interaction histories provides rich signals for automatic agent creation, enabling effective domain agent adaptation without the limitations of black-box approaches.

Abstract: Large Language Model agents are reshaping the industrial landscape. However, most practical agents remain human-designed because tasks differ widely, making them labor-intensive to build. This situation poses a central question: can we automatically create and adapt domain agents in the wild? While several recent approaches have sought to automate agent creation, they typically treat agent generation as a black-box procedure and rely solely on final performance metrics to guide the process. Such strategies overlook critical evidence explaining why an agent succeeds or fails, and often require high computational costs. To address these limitations, we propose ReCreate, an experience-driven framework for the automatic creation of domain agents. ReCreate systematically leverages agent interaction histories, which provide rich concrete signals on both the causes of success or failure and the avenues for improvement. Specifically, we introduce an agent-as-optimizer paradigm that effectively learns from experience via three key components: (i) an experience storage and retrieval mechanism for on-demand inspection; (ii) a reasoning-creating synergy pipeline that maps execution experience into scaffold edits; and (iii) hierarchical updates that abstract instance-level details into reusable domain patterns. In experiments across diverse domains, ReCreate consistently outperforms human-designed agents and existing automated agent generation methods, even when starting from minimal seed scaffolds.

Method: Proposes SCALE framework: Self prediction of optimizer with few-shot CALibration for Evaluation. Instead of full validation execution, uses self-evolution and generative reward modeling for low-cost task-level workflow generation.

Result: SCALE maintains competitive performance with only 0.61% average degradation compared to existing approaches across multiple datasets, while reducing overall token usage by up to 83%.

Conclusion: Task-level workflow generation with efficient evaluation (SCALE) is sufficient for MAS, eliminating the need for costly query-level generation and expensive execution-based validation while maintaining performance.

Abstract: Multi-Agent Systems (MAS) built on large language models typically solve complex tasks by coordinating multiple agents through workflows. Existing approaches generates workflows either at task level or query level, but their relative costs and benefits remain unclear. After rethinking and empirical analyses, we show that query-level workflow generation is not always necessary, since a small set of top-K best task-level workflows together already covers equivalent or even more queries. We further find that exhaustive execution-based task-level evaluation is both extremely token-costly and frequently unreliable. Inspired by the idea of self-evolution and generative reward modeling, we propose a low-cost task-level generation framework \textbf{SCALE}, which means \underline{\textbf{S}}elf prediction of the optimizer with few shot \underline{\textbf{CAL}}ibration for \underline{\textbf{E}}valuation instead of full validation execution. Extensive experiments demonstrate that \textbf{SCALE} maintains competitive performance, with an average degradation of just 0.61% compared to existing approach across multiple datasets, while cutting overall token usage by up to 83%.

Method: Policy-based deep reinforcement learning hyper-heuristic framework with two key extensions: 1) Action prefiltering to restrict decisions to feasible low-level actions, enabling unbiased heuristic evaluation, and 2) Commitment mechanism to regulate heuristic switching frequency. Investigates different commitment strategies (step-wise to full-episode) and action selection strategies (deterministic greedy vs stochastic sampling).

Result: The proposed approach outperforms traditional heuristics, metaheuristics, and recent neural network-based scheduling methods on standard JSSP benchmarks.

Conclusion: The hyper-heuristic framework with action prefiltering and commitment mechanisms provides an effective solution for JSSP, demonstrating superior performance through dynamic rule switching learned via deep reinforcement learning.

Abstract: This paper proposes a policy-based deep reinforcement learning hyper-heuristic framework for solving the Job Shop Scheduling Problem. The hyper-heuristic agent learns to switch scheduling rules based on the system state dynamically. We extend the hyper-heuristic framework with two key mechanisms. First, action prefiltering restricts decision-making to feasible low-level actions, enabling low-level heuristics to be evaluated independently of environmental constraints and providing an unbiased assessment. Second, a commitment mechanism regulates the frequency of heuristic switching. We investigate the impact of different commitment strategies, from step-wise switching to full-episode commitment, on both training behavior and makespan. Additionally, we compare two action selection strategies at the policy level: deterministic greedy selection and stochastic sampling. Computational experiments on standard JSSP benchmarks demonstrate that the proposed approach outperforms traditional heuristics, metaheuristics, and recent neural network-based scheduling methods

Method: Proposes Think-with-Me, a test-time interactive reasoning paradigm that pauses reasoning at transitional conjunction points for external feedback. Uses multi-criteria evaluation (rationality and completeness) for feedback from humans or LLM proxies. Trains target models using Group Relative Policy Optimization to adapt to interactive mode.

Result: Achieves superior balance between accuracy and reasoning length under limited context windows. On AIME24, outperforms QwQ-32B by 7.19% in accuracy while reducing average reasoning length by 81% under an 8K window. Also benefits security and creative tasks.

Conclusion: Think-with-Me effectively addresses inefficiencies in LRM reasoning by introducing external feedback intervention at strategic points, enabling adaptive reasoning extension/termination to reduce redundancy while preserving accuracy, making it particularly valuable for constrained computational environments.

Abstract: Large Reasoning Models (LRMs) excel at multi-step reasoning but often suffer from inefficient reasoning processes like overthinking and overshoot, where excessive or misdirected reasoning increases computational cost and degrades performance. Existing efficient reasoning methods operate in a closed-loop manner, lacking mechanisms for external intervention to guide the reasoning process. To address this, we propose Think-with-Me, a novel test-time interactive reasoning paradigm that introduces external feedback intervention into the reasoning process. Our key insights are that transitional conjunctions serve as natural points for intervention, signaling phases of self-validation or exploration and using transitional words appropriately to prolong the reasoning enhances performance, while excessive use affects performance. Building on these insights, Think-with-Me pauses reasoning at these points for external feedback, adaptively extending or terminating reasoning to reduce redundancy while preserving accuracy. The feedback is generated via a multi-criteria evaluation (rationality and completeness) and comes from either human or LLM proxies. We train the target model using Group Relative Policy Optimization (GRPO) to adapt to this interactive mode. Experiments show that Think-with-Me achieves a superior balance between accuracy and reasoning length under limited context windows. On AIME24, Think-with-Me outperforms QwQ-32B by 7.19% in accuracy while reducing average reasoning length by 81% under an 8K window. The paradigm also benefits security and creative tasks.

Method: XChoice fits mechanism-based decision models to both human data and LLM-generated decisions, recovering interpretable parameters that capture decision factor importance, constraint sensitivity, and implied trade-offs. Alignment is assessed by comparing parameter vectors across models, options, and subgroups.

Result: Applied to Americans’ daily time allocation using ATUS data, XChoice revealed heterogeneous alignment across models and activities, with salient misalignment concentrated in Black and married groups. The framework demonstrated robustness via invariance analysis and showed targeted mitigation potential with RAG interventions.

Conclusion: XChoice provides mechanism-based metrics that diagnose misalignment and support informed improvements beyond surface outcome matching, offering a more nuanced approach to AI-human alignment evaluation in constrained decision making.

Abstract: We present XChoice, an explainable framework for evaluating AI-human alignment in constrained decision making. Moving beyond outcome agreement such as accuracy and F1 score, XChoice fits a mechanism-based decision model to human data and LLM-generated decisions, recovering interpretable parameters that capture the relative importance of decision factors, constraint sensitivity, and implied trade-offs. Alignment is assessed by comparing these parameter vectors across models, options, and subgroups. We demonstrate XChoice on Americans’ daily time allocation using the American Time Use Survey (ATUS) as human ground truth, revealing heterogeneous alignment across models and activities and salient misalignment concentrated in Black and married groups. We further validate robustness of XChoice via an invariance analysis and evaluate targeted mitigation with a retrieval augmented generation (RAG) intervention. Overall, XChoice provides mechanism-based metrics that diagnose misalignment and support informed improvements beyond surface outcome matching.

Method: Introduces AstroReason-Bench, a comprehensive benchmark for Space Planning Problems (SPP) that integrates multiple scheduling regimes including ground station communication and agile Earth observation. Provides a unified agent-oriented interaction protocol and evaluates state-of-the-art open- and closed-source agentic LLM systems.

Result: Current agentic LLMs substantially underperform specialized solvers in Space Planning Problems, highlighting key limitations of generalist planning under realistic physical constraints.

Conclusion: AstroReason-Bench offers a challenging and diagnostic testbed for future agentic research, revealing the gap between generalist LLM planning and specialized solvers in real-world constrained environments.

Abstract: Recent advances in agentic Large Language Models (LLMs) have positioned them as generalist planners capable of reasoning and acting across diverse tasks. However, existing agent benchmarks largely focus on symbolic or weakly grounded environments, leaving their performance in physics-constrained real-world domains underexplored. We introduce AstroReason-Bench, a comprehensive benchmark for evaluating agentic planning in Space Planning Problems (SPP), a family of high-stakes problems with heterogeneous objectives, strict physical constraints, and long-horizon decision-making. AstroReason-Bench integrates multiple scheduling regimes, including ground station communication and agile Earth observation, and provides a unified agent-oriented interaction protocol. Evaluating on a range of state-of-the-art open- and closed-source agentic LLM systems, we find that current agents substantially underperform specialized solvers, highlighting key limitations of generalist planning under realistic constraints. AstroReason-Bench offers a challenging and diagnostic testbed for future agentic research.

Method: Probe and solve algorithm: a two-phase framework that partitions time budget into (1) probing phase exploring hyperparameters using configurable optimization methods (Bayesian optimization or Hamming distance search), and (2) solving phase using the best configuration found. Implemented in CPMpy library.

Result: Bayesian optimization outperformed default configurations: improved solution quality for ACE in 25.4% of instances (matching default in 57.9%), and for Choco in 38.6% of instances. Bayesian optimization consistently surpassed Hamming distance search, showing advantage of model-based exploration over simple local search.

Conclusion: The probe and solve algorithm offers a practical, resource-aware approach for tuning constraint solvers that yields robust improvements across diverse problem types, with Bayesian optimization being particularly effective.

Abstract: The performance of constraint programming solvers is highly sensitive to the choice of their hyperparameters. Manually finding the best solver configuration is a difficult, time-consuming task that typically requires expert knowledge. In this paper, we introduce probe and solve algorithm, a novel two-phase framework for automated hyperparameter optimization integrated into the CPMpy library. This approach partitions the available time budget into two phases: a probing phase that explores different sets of hyperparameters using configurable hyperparameter optimization methods, followed by a solving phase where the best configuration found is used to tackle the problem within the remaining time. We implement and compare two hyperparameter optimization methods within the probe and solve algorithm: Bayesian optimization and Hamming distance search. We evaluate the algorithm on two different constraint programming solvers, ACE and Choco, across 114 combinatorial problem instances, comparing their performance against the solver’s default configurations. Results show that using Bayesian optimization, the algorithm outperforms the solver’s default configurations, improving solution quality for ACE in 25.4% of instances and matching the default performance in 57.9%, and for Choco, achieving superior results in 38.6% of instances. It also consistently surpasses Hamming distance search within the same framework, confirming the advantage of model-based exploration over simple local search. Overall, the probe and solve algorithm offers a practical, resource-aware approach for tuning constraint solvers that yields robust improvements across diverse problem types.

Method: Extension of LLM-based Predictive Process Monitoring framework via prompting, with empirical evaluations conducted on three distinct event logs across KPIs including Total Time and Activity Occurrence prediction.

Result: In data-scarce settings with only 100 traces, the LLM surpasses benchmark methods. The LLM exploits both its embodied prior knowledge and internal correlations among training traces, and performs higher-order reasoning rather than merely replicating existing predictive methods.

Conclusion: LLM-based Predictive Process Monitoring demonstrates strong performance in data-scarce scenarios, leveraging both prior knowledge and complex reasoning strategies, making it a promising approach for predictive process monitoring across multiple KPIs.

Abstract: Predictive Process Monitoring is a branch of process mining that aims to predict the outcome of an ongoing process. Recently, it leveraged machine-and-deep learning architectures. In this paper, we extend our prior LLM-based Predictive Process Monitoring framework, which was initially focused on total time prediction via prompting. The extension consists of comprehensively evaluating its generality, semantic leverage, and reasoning mechanisms, also across multiple Key Performance Indicators. Empirical evaluations conducted on three distinct event logs and across the Key Performance Indicators of Total Time and Activity Occurrence prediction indicate that, in data-scarce settings with only 100 traces, the LLM surpasses the benchmark methods. Furthermore, the experiments also show that the LLM exploits both its embodied prior knowledge and the internal correlations among training traces. Finally, we examine the reasoning strategies employed by the model, demonstrating that the LLM does not merely replicate existing predictive methods but performs higher-order reasoning to generate the predictions.

Method: Developed the LEG (Large language model and Extended Greedy) framework that combines a provable approximation algorithm for population coverage optimization with LLM-driven iterative refinement incorporating human-AI alignment to integrate expert qualitative guidance while preserving coverage guarantees.

Result: Experiments on real-world data from three Ethiopian regions demonstrate the framework’s effectiveness in producing solutions that balance coverage optimization with expert preferences, showing potential for equitable, data-driven health system planning.

Conclusion: The LEG framework successfully bridges the gap between classical optimization methods requiring explicit quantitative objectives and stakeholder preferences expressed in natural language, offering a practical approach for resource-constrained health system planning that respects both quantitative and qualitative considerations.

Abstract: Ethiopia’s Ministry of Health is upgrading health posts to improve access to essential services, particularly in rural areas. Limited resources, however, require careful prioritization of which facilities to upgrade to maximize population coverage while accounting for diverse expert and stakeholder preferences. In collaboration with the Ethiopian Public Health Institute and Ministry of Health, we propose a hybrid framework that systematically integrates expert knowledge with optimization techniques. Classical optimization methods provide theoretical guarantees but require explicit, quantitative objectives, whereas stakeholder criteria are often articulated in natural language and difficult to formalize. To bridge these domains, we develop the Large language model and Extended Greedy (LEG) framework. Our framework combines a provable approximation algorithm for population coverage optimization with LLM-driven iterative refinement that incorporates human-AI alignment to ensure solutions reflect expert qualitative guidance while preserving coverage guarantees. Experiments on real-world data from three Ethiopian regions demonstrate the framework’s effectiveness and its potential to inform equitable, data-driven health system planning.

Method: Defines atomic punch events with temporal/spatial/technical attributes to parse match footage into 18 hierarchical technical-tactical indicators. Uses graph-based predictive model fusing explicit profiles with learnable latent embeddings to capture matchup dynamics. Models winning probability as differentiable function of indicators to generate tactical adjustments.

Result: Achieves 69.8% accuracy on BoxerGraph test set and 87.5% on Olympic matches for outcome prediction. System generates strategic recommendations comparable to human experts. Validated in 2024 Paris Olympics, contributing to Chinese National Team’s 3 gold and 2 silver medals.

Conclusion: BoxMind establishes a replicable paradigm for transforming unstructured video data into strategic intelligence, bridging computer vision and decision support in competitive sports through closed-loop AI expert systems.

Abstract: Competitive sports require sophisticated tactical analysis, yet combat disciplines like boxing remain underdeveloped in AI-driven analytics due to the complexity of action dynamics and the lack of structured tactical representations. To address this, we present BoxMind, a closed-loop AI expert system validated in elite boxing competition. By defining atomic punch events with precise temporal boundaries and spatial and technical attributes, we parse match footage into 18 hierarchical technical-tactical indicators. We then propose a graph-based predictive model that fuses these explicit technical-tactical profiles with learnable, time-variant latent embeddings to capture the dynamics of boxer matchups. Modeling match outcome as a differentiable function of technical-tactical indicators, we turn winning probability gradients into executable tactical adjustments. Experiments show that the outcome prediction model achieves state-of-the-art performance, with 69.8% accuracy on BoxerGraph test set and 87.5% on Olympic matches. Using this predictive model as a foundation, the system generates strategic recommendations that demonstrate proficiency comparable to human experts. BoxMind is validated through a closed-loop deployment during the 2024 Paris Olympics, directly contributing to the Chinese National Team’s historic achievement of three gold and two silver medals. BoxMind establishes a replicable paradigm for transforming unstructured video data into strategic intelligence, bridging the gap between computer vision and decision support in competitive sports.

Method: Introduces MPCI-Bench, a Multimodal Pairwise Contextual Integrity benchmark with paired positive/negative instances from the same visual source across three tiers: Seed judgments (normative), Story reasoning (context-rich), and agent action Traces (executable). Uses Tri-Principle Iterative Refinement pipeline to ensure data quality.

Result: Evaluation of state-of-the-art multimodal models reveals systematic failures to balance privacy and utility, and a pronounced modality leakage gap where sensitive visual information is leaked more frequently than textual information.

Conclusion: MPCI-Bench addresses critical gaps in evaluating agent privacy behavior and will be open-sourced to facilitate future research on agentic Contextual Integrity, highlighting the need for better multimodal privacy safeguards in AI assistants.

Abstract: As language-model agents evolve from passive chatbots into proactive assistants that handle personal data, evaluating their adherence to social norms becomes increasingly critical, often through the lens of Contextual Integrity (CI). However, existing CI benchmarks are largely text-centric and primarily emphasize negative refusal scenarios, overlooking multimodal privacy risks and the fundamental trade-off between privacy and utility. In this paper, we introduce MPCI-Bench, the first Multimodal Pairwise Contextual Integrity benchmark for evaluating privacy behavior in agentic settings. MPCI-Bench consists of paired positive and negative instances derived from the same visual source and instantiated across three tiers: normative Seed judgments, context-rich Story reasoning, and executable agent action Traces. Data quality is ensured through a Tri-Principle Iterative Refinement pipeline. Evaluations of state-of-the-art multimodal models reveal systematic failures to balance privacy and utility and a pronounced modality leakage gap, where sensitive visual information is leaked more frequently than textual information. We will open-source MPCI-Bench to facilitate future research on agentic CI.

Method: Model knowledge graph embeddings as approximate global sections of a cellular sheaf. Use the diffusion dynamics encoded by the corresponding sheaf Laplacian to optimally propagate known embeddings from a subgraph to new entities. Implement via an efficient iterative scheme.

Result: On large-scale knowledge graph embedding benchmarks, the method is competitive with and sometimes outperforms more complex models designed explicitly for inductive knowledge graph reasoning tasks.

Conclusion: Sheaf theory provides an effective algebraic framework for inductive knowledge graph reasoning, enabling simple yet powerful propagation of embeddings to new entities without requiring complex specialized models.

Abstract: Many inference tasks on knowledge graphs, including relation prediction, operate on knowledge graph embeddings – vector representations of the vertices (entities) and edges (relations) that preserve task-relevant structure encoded within the underlying combinatorial object. Such knowledge graph embeddings can be modeled as an approximate global section of a cellular sheaf, an algebraic structure over the graph. Using the diffusion dynamics encoded by the corresponding sheaf Laplacian, we optimally propagate known embeddings of a subgraph to inductively represent new entities introduced into the knowledge graph at inference time. We implement this algorithm via an efficient iterative scheme and show that on a number of large-scale knowledge graph embedding benchmarks, our method is competitive with – and in some scenarios outperforms – more complex models derived explicitly for inductive knowledge graph reasoning tasks.

Method: ProMis links uncertain geospatial data and noisy perception with declarative Hybrid Probabilistic Logic Programs to reason over agent state space legality, generating Probabilistic Mission Landscapes that quantify belief in HPLP satisfaction across state space.

Result: ProMis integrates with potent ML models like LLMs and Transformer-based vision models, enabling application with multi-modal input data across many AAM scenarios, extending prior work on reasoning capabilities and computational characteristics.

Conclusion: ProMis provides an interpretable, adaptable neuro-symbolic architecture for UAS navigation within legal frameworks that can handle uncertainty and integrate with modern ML models for practical AAM applications.

Abstract: Advanced Air Mobility (AAM) is a growing field that demands accurate and trustworthy models of legal concepts and restrictions for navigating Unmanned Aircraft Systems (UAS). In addition, any implementation of AAM needs to face the challenges posed by inherently dynamic and uncertain human-inhabited spaces robustly. Nevertheless, the employment of UAS beyond visual line of sight (BVLOS) is an endearing task that promises to significantly enhance today’s logistics and emergency response capabilities. Hence, we propose Probabilistic Mission Design (ProMis), a novel neuro-symbolic approach to navigating UAS within legal frameworks. ProMis is an interpretable and adaptable system architecture that links uncertain geospatial data and noisy perception with declarative, Hybrid Probabilistic Logic Programs (HPLP) to reason over the agent’s state space and its legality. To inform planning with legal restrictions and uncertainty in mind, ProMis yields Probabilistic Mission Landscapes (PML). These scalar fields quantify the belief that the HPLP is satisfied across the agent’s state space. Extending prior work on ProMis’ reasoning capabilities and computational characteristics, we show its integration with potent machine learning models such as Large Language Models (LLM) and Transformer-based vision models. Hence, our experiments underpin the application of ProMis with multi-modal input data and how our method applies to many AAM scenarios.

Method: 1) Iterative autoformalisation that refines theorem prover formalisation to reduce errors; 2) Theorem Prover as a Judge (TP-as-a-Judge) that uses theorem prover formalisation to assess LLM intermediate reasoning; 3) Reinforcement Learning from Theorem Prover Feedback (RLTPF) that replaces human annotation with theorem prover feedback in RLHF.

Result: Iterative autoformalisation increased execution rate on Lean prover from 60% to 87%. TP-as-a-Judge and RLTPF improved benchmarks with only 3,508 samples: 5.56% accuracy gain on Mistral-7B for MultiArith, 6.00% on Llama-2-7B for SVAMP, and 3.55% on Llama-3.1-8B for AQUA.

Conclusion: The proposed methods effectively integrate autoformalisation with synthetic data generation, enabling rigorous assessment of LLM reasoning and efficient improvement of mathematical capabilities with minimal synthetic data, demonstrating the effectiveness of theorem prover-based feedback mechanisms.

Abstract: The demand for synthetic data in mathematical reasoning has increased due to its potential to enhance the mathematical capabilities of large language models (LLMs). However, ensuring the validity of intermediate reasoning steps remains a significant challenge, affecting data quality. While formal verification via theorem provers effectively validates LLM reasoning, the autoformalisation of mathematical proofs remains error-prone. In response, we introduce iterative autoformalisation, an approach that iteratively refines theorem prover formalisation to mitigate errors, thereby increasing the execution rate on the Lean prover from 60% to 87%. Building upon that, we introduce Theorem Prover as a Judge (TP-as-a-Judge), a method that employs theorem prover formalisation to rigorously assess LLM intermediate reasoning, effectively integrating autoformalisation with synthetic data generation. Finally, we present Reinforcement Learning from Theorem Prover Feedback (RLTPF), a framework that replaces human annotation with theorem prover feedback in Reinforcement Learning from Human Feedback (RLHF). Across multiple LLMs, applying TP-as-a-Judge and RLTPF improves benchmarks with only 3,508 samples, achieving 5.56% accuracy gain on Mistral-7B for MultiArith, 6.00% on Llama-2-7B for SVAMP, and 3.55% on Llama-3.1-8B for AQUA.

Method: ARC-AGI-2 preserves the input-output pair task format of ARC-AGI but incorporates a newly curated and expanded set of tasks designed to provide more granular assessment of abstract reasoning and problem-solving abilities at higher fluid intelligence levels.

Result: The paper presents extensive human testing results to contextualize the difficulty and characteristics of ARC-AGI-2, showing it’s accessible to human intelligence but difficult for current AI systems.

Conclusion: ARC-AGI-2 aims to serve as a next-generation tool for rigorously measuring progress toward more general and human-like AI capabilities.

Abstract: The Abstraction and Reasoning Corpus for Artificial General Intelligence (ARC-AGI), introduced in 2019, established a challenging benchmark for evaluating the general fluid intelligence of artificial systems via a set of unique, novel tasks only requiring minimal prior knowledge. While ARC-AGI has spurred significant research activity over the past five years, recent AI progress calls for benchmarks capable of finer-grained evaluation at higher levels of cognitive complexity. We introduce ARC-AGI-2, an upgraded version of the benchmark. ARC-AGI-2 preserves the input-output pair task format of its predecessor, ensuring continuity for researchers. It incorporates a newly curated and expanded set of tasks specifically designed to provide a more granular signal to assess abstract reasoning and problem-solving abilities at higher levels of fluid intelligence. To contextualize the difficulty and characteristics of ARC-AGI-2, we present extensive results from human testing, providing a robust baseline that highlights the benchmark’s accessibility to human intelligence, yet difficulty for current AI systems. ARC-AGI-2 aims to serve as a next-generation tool for rigorously measuring progress towards more general and human-like AI capabilities.

Method: Position paper analysis that critically revisits claims about meta-learning as a pathway to compositionality, examining limitations in proposed meta-learning frameworks and analyzing the narrow conditions under which current systems might perform such tasks.

Result: The analysis shows that modern neural meta-learning systems can only perform compositionality tasks under very narrow and restricted definitions of meta-learning setups, failing to achieve human-like systematic compositionality.

Conclusion: Fodor and Pylyshyn’s critique persists - neural networks still lack human-like systematic compositionality, and current meta-learning approaches do not overcome this fundamental limitation.

Abstract: Strong meta-learning capabilities for systematic compositionality are emerging as an important skill for navigating the complex and changing tasks of today’s world. However, in presenting models for robust adaptation to novel environments, it is important to refrain from making unsupported claims about the performance of meta-learning systems that ultimately do not stand up to scrutiny. While Fodor and Pylyshyn famously posited that neural networks inherently lack this capacity as they are unable to model compositional representations or structure-sensitive operations, and thus are not a viable model of the human mind, Lake and Baroni recently presented meta-learning as a pathway to compositionality. In this position paper, we critically revisit this claim and highlight limitations in the proposed meta-learning framework for compositionality. Our analysis shows that modern neural meta-learning systems can only perform such tasks, if at all, under a very narrow and restricted definition of a meta-learning setup. We therefore claim that `Fodor and Pylyshyn’s legacy’ persists, and to date, there is no human-like systematic compositionality learned in neural networks.

Method: COPE uses a planner model to generate a plan as a lightweight intermediate that guides a downstream executor model. Small and large models take turns acting as planner and executor, exchanging plans in a multi-stage cascade to collaboratively solve tasks.

Result: COPE achieves performance comparable to large proprietary models while drastically reducing inference API costs, demonstrated across mathematical reasoning, code generation, open-ended tasks, and agent tasks benchmarks.

Conclusion: Planning serves as an effective prior for cost-efficient inference, enabling small and large models to collaborate effectively and bridge the performance-cost trade-off.

Abstract: Recently, large language models (LLMs) have demonstrated strong performance, ranging from simple to complex tasks. However, while large proprietary models (e.g., models with over 100B parameters) achieve remarkable results across diverse tasks, they are often accessible through costly APIs, making frequent use too costly for many applications. In contrast, small open-source models (e.g., models with fewer than 3B parameters) are freely available and easy to deploy locally, but their performance on complex tasks remains limited. This trade-off raises a natural question: how can small and large models efficiently collaborate to combine their complementary strengths? To bridge this trade-off, we propose COPE, a test-time collaboration framework. A planner model first generates a plan that serves as a lightweight intermediate that guides a downstream executor model. Small and large models take turns acting as planner and executor, exchanging plans in a multi-stage cascade to collaboratively solve tasks. Through comprehensive experiments on benchmarks spanning mathematical reasoning, code generation, open-ended tasks, and agent tasks, we demonstrate that COPE achieves performance comparable to large proprietary models, while drastically reducing the inference API cost. These results highlight planning as an effective prior for cost-efficient inference.

Method: Proposes Valley-to-Peak (V2P) method with two key innovations: (1) suppression attention mechanism to minimize focus on irrelevant background regions and highlight intended areas, and (2) Fitts’ Law-inspired approach modeling GUI interactions as 2D Gaussian heatmaps where weight decreases from center to edges based on target size.

Result: Achieves 92.4% and 52.5% performance on ScreenSpot-v2 and ScreenSpot-Pro benchmarks. Ablation studies confirm each component’s contribution, demonstrating generalizability for precise GUI grounding tasks.

Conclusion: V2P effectively isolates target areas and teaches models to focus on essential UI element points, showing potential for real-world deployment in future GUI agents by addressing key limitations of existing methods.

Abstract: Precise localization of GUI elements is crucial for the development of GUI agents. Traditional methods rely on bounding box or center-point regression, neglecting spatial interaction uncertainty and visual-semantic hierarchies. Recent methods incorporate attention mechanisms but still face two key issues: (1) ignoring processing background regions causes attention drift from the desired area, and (2) uniform modeling the target UI element fails to distinguish between its center and edges, leading to click imprecision. Inspired by how humans visually process and interact with GUI elements, we propose the Valley-to-Peak (V2P) method to address these issues. To mitigate background distractions, V2P introduces a suppression attention mechanism that minimizes the model’s focus on irrelevant regions to highlight the intended region. For the issue of center-edge distinction, V2P applies a Fitts’ Law-inspired approach by modeling GUI interactions as 2D Gaussian heatmaps where the weight gradually decreases from the center towards the edges. The weight distribution follows a Gaussian function, with the variance determined by the target’s size. Consequently, V2P effectively isolates the target area and teaches the model to concentrate on the most essential point of the UI element. The model trained by V2P achieves the performance with 92.4% and 52.5% on two benchmarks ScreenSpot-v2 and ScreenSpot-Pro (see Fig.~\ref{fig:main_results_charts}). Ablations further confirm each component’s contribution, underscoring V2P’s generalizability in precise GUI grounding tasks and its potential for real-world deployment in future GUI agents.

Method: 1) Analyze successful vs failed search trajectories to identify four beneficial reasoning behaviors: Information Verification, Authority Evaluation, Adaptive Search, and Error Recovery. 2) Propose Behavior Priming: first perform supervised fine-tuning on trajectories exhibiting these behaviors, then apply standard reinforcement learning to improve task performance.

Result: Behavior Priming yields 37.2% relative improvement over direct RL on three web benchmarks and 6.2% improvement on seven multi-hop QA benchmarks. Outperforms SFT-then-RL baseline using outcome-correct trajectories. Shows reasoning behaviors matter more than outcome correctness in priming stage. Enhances exploration (pass@8) and test-time scaling.

Conclusion: Behavior Priming effectively equips agentic search models with beneficial reasoning behaviors before RL, providing a robust foundation for RL and improving performance. The identified reasoning behaviors are crucial for successful agentic search.

Abstract: Agentic search requires large language models (LLMs) to perform multi-step search to solve complex information-seeking tasks, imposing unique challenges on their reasoning capabilities. However, what constitutes effective reasoning for agentic search and how it can be learned remains unclear. In this work, we first investigate the reasoning behaviors that enable success in agentic search. By comparing successful and failed trajectories via an LLM-based analysis pipeline, we identify four beneficial behaviors: Information Verification, Authority Evaluation, Adaptive Search, and Error Recovery. Building on this, we propose Behavior Priming, a training approach that equips agentic search models with these reasoning behaviors before reinforcement learning (RL). Specifically, it first performs supervised fine-tuning (SFT) on collected trajectories exhibiting the identified behaviors to cultivate these behaviors, and then applies standard RL to further improve task performance. Experiments on Qwen3-1.7B and Llama3.2-3B-Instruct show that Behavior Priming yields relative improvements over direct RL by 37.2% on three web benchmarks and 6.2% on seven multi-hop QA benchmarks, and outperforms the SFT-then-RL baseline using outcome-correct trajectories for fine-tuning. Crucially, we show that these reasoning behaviors matter more than outcome correctness in the priming stage prior to RL. Further analysis reveals that Behavior Priming enhances exploration (pass@8) and test-time scaling (search step number), providing a robust foundation for RL. Our code are avalible at https://github.com/cxcscmu/Behavior-Priming-for-Agentic-Search.

Method: Transformer model that directly generates molecular structures from input tandem mass spectra and molecular formulas, using transfer learning from simulated data and a novel test-time tuning strategy for adapting to novel experimental data.

Result: Achieves Top-1 accuracy of 3.16% on MassSpecGym and 12.88% on NPLIB1, outperforming conventional fine-tuning by 27% and 67% respectively. Generated candidates show high structural plausibility with 83% and 64% relative improvement in average Tanimoto similarity compared to state-of-the-art methods.

Conclusion: The framework combines simplicity with adaptability, generating accurate molecular candidates that provide valuable guidance for expert interpretation of unseen spectra, addressing challenges in small molecule identification from mass spectrometry data.

Abstract: Tandem Mass Spectrometry is a cornerstone technique for identifying unknown small molecules in fields such as metabolomics, natural product discovery and environmental analysis. However, certain aspects, such as the probabilistic fragmentation process and size of the chemical space, make structure elucidation from such spectra highly challenging, particularly when there is a shift between the deployment and training conditions. Current methods rely on database matching of previously observed spectra of known molecules and multi-step pipelines that require intermediate fingerprint prediction or expensive fragment annotations. We introduce a novel end-to-end framework based on a transformer model that directly generates molecular structures from an input tandem mass spectrum and its corresponding molecular formula, thereby eliminating the need for manual annotations and intermediate steps, while leveraging transfer learning from simulated data. To further address the challenge of out-of-distribution spectra, we introduce a test-time tuning strategy that dynamically adapts the pre-trained model to novel experimental data. Our approach achieves a Top-1 accuracy of 3.16% on the MassSpecGym benchmark and 12.88% on the NPLIB1 datasets, considerably outperforming conventional fine-tuning. Baseline approaches are also surpassed by 27% and 67% respectively. Even when the exact reference structure is not recovered, the generated candidates are chemically informative, exhibiting high structural plausibility as reflected by strong Tanimoto similarity to the ground truth. Notably, we observe a relative improvement in average Tanimoto similarity of 83% on NPLIB1 and 64% on MassSpecGym compared to state-of-the-art methods. Our framework combines simplicity with adaptability, generating accurate molecular candidates that offer valuable guidance for expert interpretation of unseen spectra.

Method: Conducted experiments across 66 AxA configurations, 4 domains (3 transactional, 1 advisory), and over 2500 conversations (250k+ LLM inferences). Analyzed prompt and conversation dynamics, and tested a protocol-level mitigation using structured responses.

Result: Echoing occurs across major LLM providers with rates up to 70%, persists in advanced reasoning models (32.8%), increases with longer interactions (7+ agent turns), and can be reduced to 9% using structured response mitigation.

Conclusion: Agent-agent conversations exhibit unique behavioral drifts like echoing that aren’t predictable from single-agent performance, requiring new mitigation strategies beyond reasoning improvements, with structured responses showing promise as a practical solution.

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 $66$ AxA configurations, $4$ domains (3 transactional, 1 advisory), and $2500+$ conversations (over $250000$ LLM inferences), we show that echoing occurs across major LLM providers, with echoing rates as high as $70%$ 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 reasoning efforts. We analyze prompt, conversation dynamics, showing that echoing arises as interaction grows longer ($7+$ agent turns) and is not merely an artifact of sub-optimal experiment design. Finally, we introduce a protocol-level mitigation where targeted use of structured response reduces echoing to $9%$.

Method: Proposes a co-evolving agents framework with two components: a target agent and an auxiliary failure agent. The failure agent learns through preference optimization over failure trajectories from both agents, generating hard negatives that are close to success but remain failures. These hard negatives are then incorporated into the target agent’s optimization to sharpen decision boundaries.

Result: The method shows improved performance across benchmark datasets and demonstrates that failures can be systematically transformed into structured and valuable learning signals, enhancing generalization in self-improving agents.

Conclusion: The co-evolving agents framework effectively addresses overfitting in self-improving agents by leveraging failure trajectories as structured learning signals through hard negative generation, leading to better generalization and performance.

Abstract: The rapid progress of large foundation models has accelerated the development of task-specialized agents across diverse domains. However, the effectiveness of agents remains tightly coupled with the quality of training data, while curating task-specific datasets remains costly and often infeasible in real-world scenarios. Recent work has explored self-improving agents that autonomously generate, refine, and re-train on their own trajectories. A prominent line of approaches further leverages preference optimization by pairing predicted trajectories with scarce ground-truth trajectories, enabling agents to learn directly from their own failures. While these methods outperform supervised fine-tuning, their heavy reliance on predicted trajectories under limited ground-truth supervision leaves them prone to overfitting. To address this, we propose a co-evolving agents framework in which a target agent improves jointly with an auxiliary failure agent. The failure agent learns through preference optimization over failure trajectories from both the target and itself, thereby generating hard negatives that are close to success yet remain failures. Incorporating these informative hard negatives into the target agent’s optimization sharpens decision boundaries and enhances generalization. Our comprehensive analysis and experiments across benchmark datasets show that our method not only shows improved performance but also demonstrates that failures, instead of being used as-is, can be systematically transformed into structured and valuable learning signals in self-improving agents.

Method: Two-pronged approach: 1) Semantic entropy-guided curriculum learning organizes training data from low to high semantic entropy for progressive optimization; 2) Non-uniform token treatment applies KL regularization on low-entropy tokens (critical for exploration) with stronger constraints on high-covariance portions within these tokens.

Result: Outperforms other entropy-based approaches across 6 benchmarks with 3 different parameter-scale base models, effectively mitigating entropy collapse and enhancing LLM reasoning.

Conclusion: Joint optimization of data organization (curriculum learning) and algorithmic design (non-uniform token treatment) effectively addresses entropy collapse in RLVR, leading to improved reasoning capabilities in LLMs.

Abstract: Reinforcement learning with verifiable rewards (RLVR) has demonstrated superior performance in enhancing the reasoning capability of large language models (LLMs). However, this accuracy-oriented learning paradigm often suffers from entropy collapse, which reduces policy exploration and limits reasoning capabilities. To address this challenge, we propose an efficient reinforcement learning framework that leverages entropy signals at both the semantic and token levels to improve reasoning. From the data perspective, we introduce semantic entropy-guided curriculum learning, organizing training data from low to high semantic entropy to guide progressive optimization from easier to more challenging tasks. For the algorithmic design, we adopt non-uniform token treatment by imposing KL regularization on low-entropy tokens that critically impact policy exploration and applying stronger constraints on high-covariance portions within these tokens. By jointly optimizing data organization and algorithmic design, our method effectively mitigates entropy collapse and enhances LLM reasoning. Experimental results across 6 benchmarks with 3 different parameter-scale base models demonstrate that our method outperforms other entropy-based approaches in improving reasoning.

Method: Proposes SimVC-CAS, a collective agent system with role-playing investor agents and GNN-based supervised interaction module. Models startup financing prediction as group decision-making with heterogeneous investor traits and preferences, using graph-structured co-investment networks for information exchange.

Result: Using PitchBook data with strict leakage controls, SimVC-CAS achieves ~25% relative improvement in average precision@10, significantly boosting predictive accuracy while providing interpretable, multiperspective reasoning.

Conclusion: SimVC-CAS effectively captures both enterprise fundamentals and investor behavioral dynamics, offering a novel approach to startup success prediction with applications to other complex group decision scenarios.

Abstract: Due to the high value and high failure rate of startups, predicting their success has become a critical challenge across interdisciplinary research. Existing approaches typically model success prediction from the perspective of a single decision-maker, overlooking the collective dynamics of investor groups that dominate real-world venture capital (VC) decisions. In this paper, we propose SimVC-CAS, a novel collective agent system that simulates VC decision-making as a multi-agent interaction process. By designing role-playing agents and a GNN-based supervised interaction module, we reformulate startup financing prediction as a group decision-making task, capturing both enterprise fundamentals and the behavioral dynamics of potential investor networks. Each agent embodies an investor with unique traits and preferences, enabling heterogeneous evaluation and realistic information exchange through a graph-structured co-investment network. Using real-world data from PitchBook and under strict data leakage controls, we show that SimVC-CAS significantly improves predictive accuracy while providing interpretable, multiperspective reasoning, for example, approximately 25% relative improvement with respect to average precision@10. SimVC-CAS also sheds light on other complex group decision scenarios.

Method: Uses Neural Prophet with Deep Neural Network (NP-DNN) with Z-score normalization for preprocessing, missing value imputation, and Multi-Layer Perceptron (MLP) to learn complex nonlinear relationships and extract hidden patterns from stock price data.

Result: The proposed NP-DNN model achieved 99.21% accuracy, outperforming other approaches including the Fused Large Language Model.

Conclusion: NP-DNN effectively predicts stock market prices by combining neural prophet architecture with deep learning techniques, demonstrating superior accuracy compared to existing methods.

Abstract: Stock market price prediction is a significant interdisciplinary research domain that depends at the intersection of finance, statistics, and economics. Forecasting Accurately predicting stock prices has always been a focal point for various researchers. However, existing statistical approaches for time-series prediction often fail to effectively forecast the probability range of future stock prices. Hence, to solve this problem, the Neural Prophet with a Deep Neural Network (NP-DNN) is proposed to predict stock market prices. The preprocessing technique used in this research is Z-score normalization, which normalizes stock price data by removing scale differences, making patterns easier to detect. Missing value imputation fills gaps in historical data, enhancing the models use of complete information for more accurate predictions. The Multi-Layer Perceptron (MLP) learns complex nonlinear relationships among stock market prices and extracts hidden patterns from the input data, thereby creating meaningful feature representations for better prediction accuracy. The proposed NP-DNN model achieved an accuracy of 99.21% compared with other approaches using the Fused Large Language Model. Keywords: deep neural network, forecasting stock prices, multi-layer perceptron, neural prophet, stock market price prediction.

Method: Proposes Valley-to-Peak (V2P) method with two key innovations: (1) suppression attention mechanism to minimize focus on irrelevant background regions, and (2) Fitts’ Law-inspired approach modeling GUI interactions as 2D Gaussian heatmaps where weight decreases from center to edges, with variance determined by target size.

Result: Achieves 92.4% and 52.5% performance on ScreenSpot-v2 and ScreenSpot-Pro benchmarks. Ablation studies confirm each component’s contribution, demonstrating V2P’s generalizability for precise GUI grounding tasks.

Conclusion: V2P effectively isolates target areas and teaches models to focus on essential UI element points, showing potential for real-world deployment in future GUI agents by addressing background distractions and center-edge distinction issues.

Abstract: Precise localization of GUI elements is crucial for the development of GUI agents. Traditional methods rely on bounding box or center-point regression, neglecting spatial interaction uncertainty and visual-semantic hierarchies. Recent methods incorporate attention mechanisms but still face two key issues: (1) ignoring processing background regions causes attention drift from the desired area, and (2) uniform modeling the target UI element fails to distinguish between its center and edges, leading to click imprecision. Inspired by how humans visually process and interact with GUI elements, we propose the Valley-to-Peak (V2P) method to address these issues. To mitigate background distractions, V2P introduces a suppression attention mechanism that minimizes the model’s focus on irrelevant regions to highlight the intended region. For the issue of center-edge distinction, V2P applies a Fitts’ Law-inspired approach by modeling GUI interactions as 2D Gaussian heatmaps where the weight gradually decreases from the center towards the edges. The weight distribution follows a Gaussian function, with the variance determined by the target’s size. Consequently, V2P effectively isolates the target area and teaches the model to concentrate on the most essential point of the UI element. The model trained by V2P achieves the performance with 92.4% and 52.5% on two benchmarks ScreenSpot-v2 and ScreenSpot-Pro (see Fig.~\ref{fig:main_results_charts}). Ablations further confirm each component’s contribution, underscoring V2P’s generalizability in precise GUI grounding tasks and its potential for real-world deployment in future GUI agents.

Method: Introduces AviationLMM architecture that ingests multimodal inputs (air-ground voice, surveillance, telemetry, video, structured texts), performs cross-modal alignment and fusion, and produces flexible outputs including situation summaries, risk alerts, predictive diagnostics, and incident reconstructions.

Result: Identifies key research opportunities including data acquisition, alignment and fusion, pretraining, reasoning, trustworthiness, privacy, robustness to missing modalities, and synthetic scenario generation to realize the AviationLMM vision.

Conclusion: By articulating the design and challenges of AviationLMM, the paper aims to boost civil aviation foundation model progress and catalyze coordinated research toward an integrated, trustworthy, and privacy-preserving aviation AI ecosystem.

Abstract: Civil aviation is a cornerstone of global transportation and commerce, and ensuring its safety, efficiency and customer satisfaction is paramount. Yet conventional Artificial Intelligence (AI) solutions in aviation remain siloed and narrow, focusing on isolated tasks or single modalities. They struggle to integrate heterogeneous data such as voice communications, radar tracks, sensor streams and textual reports, which limits situational awareness, adaptability, and real-time decision support. This paper introduces the vision of AviationLMM, a Large Multimodal foundation Model for civil aviation, designed to unify the heterogeneous data streams of civil aviation and enable understanding, reasoning, generation and agentic applications. We firstly identify the gaps between existing AI solutions and requirements. Secondly, we describe the model architecture that ingests multimodal inputs such as air-ground voice, surveillance, on-board telemetry, video and structured texts, and performs cross-modal alignment and fusion, and produces flexible outputs ranging from situation summaries and risk alerts to predictive diagnostics and multimodal incident reconstructions. In order to fully realize this vision, we identify key research opportunities to address, including data acquisition, alignment and fusion, pretraining, reasoning, trustworthiness, privacy, robustness to missing modalities, and synthetic scenario generation. By articulating the design and challenges of AviationLMM, we aim to boost the civil aviation foundation model progress and catalyze coordinated research efforts toward an integrated, trustworthy and privacy-preserving aviation AI ecosystem.

Method: Formalizes safe refusal as answerability-gating problem. Uses LatentRefusal mechanism that predicts query answerability from intermediate LLM hidden activations. Introduces Tri-Residual Gated Encoder to suppress schema noise and amplify sparse cues of question-schema mismatch.

Result: Achieves 88.5% average F1 across four benchmarks on both backbones while adding only ~2ms probe overhead. Effectively identifies ambiguous/unanswerable queries and provides attachable safety layer.

Conclusion: LatentRefusal provides an efficient, lightweight solution for safe refusal in text-to-SQL systems by leveraging latent signals from LLM activations, outperforming existing methods in both effectiveness and efficiency.

Abstract: In LLM-based text-to-SQL systems, unanswerable and underspecified user queries may generate not only incorrect text but also executable programs that yield misleading results or violate safety constraints, posing a major barrier to safe deployment. Existing refusal strategies for such queries either rely on output-level instruction following, which is brittle due to model hallucinations, or estimate output uncertainty, which adds complexity and overhead. To address this challenge, we formalize safe refusal in text-to-SQL systems as an answerability-gating problem and propose LatentRefusal, a latent-signal refusal mechanism that predicts query answerability from intermediate hidden activations of a large language model. We introduce the Tri-Residual Gated Encoder, a lightweight probing architecture, to suppress schema noise and amplify sparse, localized cues of question-schema mismatch that indicate unanswerability. Extensive empirical evaluations across diverse ambiguous and unanswerable settings, together with ablation studies and interpretability analyses, demonstrate the effectiveness of the proposed approach and show that LatentRefusal provides an attachable and efficient safety layer for text-to-SQL systems. Across four benchmarks, LatentRefusal improves average F1 to 88.5 percent on both backbones while adding approximately 2 milliseconds of probe overhead.

Method: Proposes ChartComplete dataset based on visualization community’s chart taxonomy, covering 30 different chart types. The dataset consists of classified chart images without learning signals.

Result: ChartComplete dataset is created and presented to the community as a resource for building more comprehensive chart understanding benchmarks.

Conclusion: ChartComplete addresses the diversity gap in chart understanding benchmarks by providing a dataset covering 30 chart types, enabling more comprehensive evaluation of MLLMs’ chart understanding capabilities.

Abstract: With advancements in deep learning (DL) and computer vision techniques, the field of chart understanding is evolving rapidly. In particular, multimodal large language models (MLLMs) are proving to be efficient and accurate in understanding charts. To accurately measure the performance of MLLMs, the research community has developed multiple datasets to serve as benchmarks. By examining these datasets, we found that they are all limited to a small set of chart types. To bridge this gap, we propose the ChartComplete dataset. The dataset is based on a chart taxonomy borrowed from the visualization community, and it covers thirty different chart types. The dataset is a collection of classified chart images and does not include a learning signal. We present the ChartComplete dataset as is to the community to build upon it.

Method: Integrated safety evaluation of six frontier models (GPT-5.2, Gemini 3 Pro, Qwen3-VL, Grok 4.1 Fast, Nano Banana Pro, Seedream 4.5) across language, vision-language, and image generation using unified protocol combining benchmark, adversarial, multilingual, and compliance evaluations.

Result: Highly uneven safety landscape: GPT-5.2 shows consistently strong balanced performance; other models exhibit trade-offs across safety dimensions. All models remain highly vulnerable under adversarial testing (worst-case safety rates <6%). Text-to-image models show slightly better alignment in regulated visual risk categories but remain fragile to adversarial/ambiguous prompts.

Conclusion: Safety in frontier models is inherently multidimensional—shaped by modality, language, and evaluation design—highlighting need for standardized, holistic safety assessments to better reflect real-world risk and guide responsible deployment.

Abstract: The rapid evolution of Large Language Models (LLMs) and Multimodal Large Language Models (MLLMs) has driven major gains in reasoning, perception, and generation across language and vision, yet whether these advances translate into comparable improvements in safety remains unclear, partly due to fragmented evaluations that focus on isolated modalities or threat models. In this report, we present an integrated safety evaluation of six frontier models–GPT-5.2, Gemini 3 Pro, Qwen3-VL, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5–assessing each across language, vision-language, and image generation using a unified protocol that combines benchmark, adversarial, multilingual, and compliance evaluations. By aggregating results into safety leaderboards and model profiles, we reveal a highly uneven safety landscape: while GPT-5.2 demonstrates consistently strong and balanced performance, other models exhibit clear trade-offs across benchmark safety, adversarial robustness, multilingual generalization, and regulatory compliance. Despite strong results under standard benchmarks, all models remain highly vulnerable under adversarial testing, with worst-case safety rates dropping below 6%. Text-to-image models show slightly stronger alignment in regulated visual risk categories, yet remain fragile when faced with adversarial or semantically ambiguous prompts. Overall, these findings highlight that safety in frontier models is inherently multidimensional–shaped by modality, language, and evaluation design–underscoring the need for standardized, holistic safety assessments to better reflect real-world risk and guide responsible deployment.

Method: Proposes DSA-Tokenizer with: 1) semantic tokens supervised by ASR to capture linguistic content, 2) acoustic tokens focusing on mel-spectrograms restoration to encode style, 3) hierarchical Flow-Matching decoder to eliminate rigid length constraints, and 4) joint reconstruction-recombination training strategy.

Result: Enables high fidelity reconstruction and flexible recombination through robust disentanglement, facilitating controllable generation in speech LLMs. Audio samples available online.

Conclusion: Disentangled tokenization is a pivotal paradigm for future speech modeling. Code and model will be made publicly available after paper acceptance.

Abstract: Speech tokenizers serve as the cornerstone of discrete Speech Large Language Models (Speech LLMs). Existing tokenizers either prioritize semantic encoding, fuse semantic content with acoustic style inseparably, or achieve incomplete semantic-acoustic disentanglement. To achieve better disentanglement, we propose DSA-Tokenizer, which explicitly disentangles speech into discrete semantic and acoustic tokens via distinct optimization constraints. Specifically, semantic tokens are supervised by ASR to capture linguistic content, while acoustic tokens focus on mel-spectrograms restoration to encode style. To eliminate rigid length constraints between the two sequences, we introduce a hierarchical Flow-Matching decoder that further improve speech generation quality. Furthermore, We employ a joint reconstruction-recombination training strategy to enforce this separation. DSA-Tokenizer enables high fidelity reconstruction and flexible recombination through robust disentanglement, facilitating controllable generation in speech LLMs. Our analysis highlights disentangled tokenization as a pivotal paradigm for future speech modeling. Audio samples are avaialble at https://anonymous.4open.science/w/DSA_Tokenizer_demo/. The code and model will be made publicly available after the paper has been accepted.

Method: Uses a unified LLM architecture with shared discrete audio token space, instruction-driven task induction, fully autoregressive formulation over discrete speech tokens, joint multi-task training across speech domains, and scalable inference pipeline.

Result: Achieves competitive performance across diverse speech tasks while supporting efficient multi-scale deployment, including a lightweight 0.3B-parameter variant for edge/resource-constrained environments.

Conclusion: A unified autoregressive architecture can effectively handle multiple core speech tasks while remaining viable for practical, low-latency deployment.

Abstract: Traditional speech systems typically rely on separate, task-specific models for text-to-speech (TTS), automatic speech recognition (ASR), and voice conversion (VC), resulting in fragmented pipelines that limit scalability, efficiency, and cross-task generalization. In this paper, we present General-Purpose Audio (GPA), a unified audio foundation model that integrates multiple core speech tasks within a single large language model (LLM) architecture. GPA operates on a shared discrete audio token space and supports instruction-driven task induction, enabling a single autoregressive model to flexibly perform TTS, ASR, and VC without architectural modifications. This unified design combines a fully autoregressive formulation over discrete speech tokens, joint multi-task training across speech domains, and a scalable inference pipeline that achieves high concurrency and throughput. The resulting model family supports efficient multi-scale deployment, including a lightweight 0.3B-parameter variant optimized for edge and resource-constrained environments. Together, these design choices demonstrate that a unified autoregressive architecture can achieve competitive performance across diverse speech tasks while remaining viable for low-latency, practical deployment.

Method: Uses an interleaved text-audio token schedule (1:2 ratio) that supports streaming generation, enabling sub-second end-to-end latency while maintaining high-quality personalized voice synthesis across multi-turn conversations.

Result: Achieves 10.96% relative improvement in speaker similarity over human baseline with Real-Time Factor of 0.43, while maintaining strong reasoning and dialogue capabilities. Code and models are publicly available.

Conclusion: Chroma 1.0 successfully addresses the speaker identity preservation problem in spoken dialogue systems, enabling both low-latency interaction and high-fidelity personalized voice cloning in real-time applications.

Abstract: Recent end-to-end spoken dialogue systems leverage speech tokenizers and neural audio codecs to enable LLMs to operate directly on discrete speech representations. However, these models often exhibit limited speaker identity preservation, hindering personalized voice interaction. In this work, we present Chroma 1.0, the first open-source, real-time, end-to-end spoken dialogue model that achieves both low-latency interaction and high-fidelity personalized voice cloning. Chroma achieves sub-second end-to-end latency through an interleaved text-audio token schedule (1:2) that supports streaming generation, while maintaining high-quality personalized voice synthesis across multi-turn conversations. Our experimental results demonstrate that Chroma achieves a 10.96% relative improvement in speaker similarity over the human baseline, with a Real-Time Factor (RTF) of 0.43, while maintaining strong reasoning and dialogue capabilities. Our code and models are publicly available at https://github.com/FlashLabs-AI-Corp/FlashLabs-Chroma and https://huggingface.co/FlashLabs/Chroma-4B .

Method: Created WenetSpeech-Wu corpus (8,000 hours of diverse speech), developed WenetSpeech-Wu-Bench benchmark covering multiple tasks (ASR, translation, TTS, etc.), and released pretrained models.

Result: Established competitive performance across multiple speech processing tasks, empirically validated dataset effectiveness, and created comprehensive Wu dialect speech processing ecosystem.

Conclusion: These contributions provide foundation for Wu dialect speech processing research; all resources (datasets, benchmarks, models) are open-sourced to support future dialectal speech intelligence work.

Abstract: Speech processing for low-resource dialects remains a fundamental challenge in developing inclusive and robust speech technologies. Despite its linguistic significance and large speaker population, the Wu dialect of Chinese has long been hindered by the lack of large-scale speech data, standardized evaluation benchmarks, and publicly available models. In this work, we present WenetSpeech-Wu, the first large-scale, multi-dimensionally annotated open-source speech corpus for the Wu dialect, comprising approximately 8,000 hours of diverse speech data. Building upon this dataset, we introduce WenetSpeech-Wu-Bench, the first standardized and publicly accessible benchmark for systematic evaluation of Wu dialect speech processing, covering automatic speech recognition (ASR), Wu-to-Mandarin translation, speaker attribute prediction, speech emotion recognition, text-to-speech (TTS) synthesis, and instruction-following TTS (instruct TTS). Furthermore, we release a suite of strong open-source models trained on WenetSpeech-Wu, establishing competitive performance across multiple tasks and empirically validating the effectiveness of the proposed dataset. Together, these contributions lay the foundation for a comprehensive Wu dialect speech processing ecosystem, and we open-source proposed datasets, benchmarks, and models to support future research on dialectal speech intelligence.

Method: SuperEar uses acoustic metamaterials to enhance faint signals, cover the full speech frequency range with compact design, and reduce noise/distortion. It’s implemented with low-cost 3D-printed parts and off-the-shelf hardware.

Result: SuperEar can recover phone call audio with over 80% success rate at distances up to 4.6 meters, more than doubling the range of previous approaches.

Conclusion: SuperEar demonstrates a new class of privacy threats enabled by metamaterial technology that requires attention, showing that practical acoustic eavesdropping in real outdoor scenarios is now feasible.

Abstract: Acoustic eavesdropping is a privacy risk, but existing attacks rarely work in real outdoor situations where people make phone calls on the move. We present SuperEar, the first portable system that uses acoustic metamaterials to reliably capture conversations in these scenarios. We show that the threat is real as a practical prototype can be implemented to enhance faint signals, cover the full range of speech with a compact design, and reduce noise and distortion to produce clear audio. We show that SuperEar can be implemented from low-cost 3D-printed parts and off-the-shelf hardware. Experimental results show that SuperEar can recover phone call audio with a success rate of over 80% at distances of up to 4.6 m - more than twice the range of previous approaches. Our findings highlight a new class of privacy threats enabled by metamaterial technology that requires attention.

Method: Introduces SonicBench with controllable generation toolbox to construct stimuli for two paradigms: recognition (absolute judgment) and comparison (relative judgment). Evaluates 12 core physical attributes across five perceptual dimensions. Uses linear probing analysis to examine audio encoder capabilities.

Result: LALMs show substantial deficiency in foundational auditory understanding - most perform near random guessing. Unlike humans, they fail to show expected advantage on comparison tasks. Explicit reasoning yields minimal gains. However, frozen audio encoders successfully capture physical cues (≥60% accuracy), indicating bottleneck is in alignment/decoding stages.

Conclusion: The primary limitation of LALMs for physical audio perception lies not in audio encoding but in the alignment and decoding stages where models fail to leverage sensory signals they’ve already captured. This reveals a critical gap in current audio language models’ fundamental auditory understanding.

Abstract: Large Audio Language Models (LALMs) excel at semantic and paralinguistic tasks, yet their ability to perceive the fundamental physical attributes of audio such as pitch, loudness, and spatial location remains under-explored. To bridge this gap, we introduce SonicBench, a psychophysically grounded benchmark that systematically evaluates 12 core physical attributes across five perceptual dimensions. Unlike previous datasets, SonicBench uses a controllable generation toolbox to construct stimuli for two complementary paradigms: recognition (absolute judgment) and comparison (relative judgment). This design allows us to probe not only sensory precision but also relational reasoning capabilities, a domain where humans typically exhibit greater proficiency. Our evaluation reveals a substantial deficiency in LALMs’ foundational auditory understanding; most models perform near random guessing and, contrary to human patterns, fail to show the expected advantage on comparison tasks. Furthermore, explicit reasoning yields minimal gains. However, our linear probing analysis demonstrates crucially that frozen audio encoders do successfully capture these physical cues (accuracy at least 60%), suggesting that the primary bottleneck lies in the alignment and decoding stages, where models fail to leverage the sensory signals they have already captured.

Method: Mixed-methods approach: 1) Review of existing standardization efforts, 2) Survey of 82 computational audiology community members, 3) Expert panel discussion with five experts at the 2024 Virtual Conference of Computational Audiology.

Result: While many are familiar with standardization concepts, few know about existing initiatives. 90% of respondents expressed willingness to follow or contribute to standardization efforts. Panel discussed relevant initiatives (OMOP, openEHR, Noah) and identified challenges (harmonization) and opportunities (alignment with other medical fields).

Conclusion: The study provides guidance for implementing interoperable data standards in audiology, highlighting community support, key issues to address, and suggesting future paths for standardization work in the field.

Abstract: Objective: This study addresses conceptual issues around data standardisation in audiology, and outlines steps toward achieving it. It reports a survey of the computational audiology community on their current understanding, needs, and preferences concerning data standards. Based on survey findings and a panel discussion, recommendations are made concerning moving forward with standardisation in audiology. Design: Mixed-methods: 1) review of existing standardisation efforts; 2) a survey of the computational audiology community; 3) expert panel discussion in a dedicated session at the 2024 Virtual Conference of Computational Audiology. Sample: Survey: 82 members of the global community; Panel discussion: five experts. Results: A prerequisite for any global audiology database are agreed data standards. Although many are familiar with the general idea, few know of existing initiatives, or have actively participated in them. Ninety percent of respondents expressed willingness to follow or contribute to standardisation efforts. The panel discussed relevant initiatives (e.g. OMOP, openEHR, Noah) and explored both challenges (around harmonisation) and opportunities (alignment with other medical fields and conversion among approaches). Conclusions: Combining conceptual discussion with stakeholder views, the study offers guidance for implementing interoperable data standards in audiology. It highlights community support, key issues to address, and suggests paths for future work.

Method: LIVI uses supervision from state-of-the-art transcription and text embedding models during training to learn effective representations, but removes the transcription step at inference to maintain efficiency.

Result: Achieves retrieval accuracy on par with or superior to harmonic-based systems while remaining lightweight and computationally efficient.

Conclusion: LIVI demonstrates that lyrics-based cover retrieval can balance accuracy and efficiency, challenging the dominance of complexity-heavy audio pipelines in version identification.

Abstract: Music Cover Retrieval, also known as Version Identification, aims to recognize distinct renditions of the same underlying musical work, a task central to catalog management, copyright enforcement, and music retrieval. State-of-the-art approaches have largely focused on harmonic and melodic features, employing increasingly complex audio pipelines designed to be invariant to musical attributes that often vary widely across covers. While effective, these methods demand substantial training time and computational resources. By contrast, lyrics constitute a strong invariant across covers, though their use has been limited by the difficulty of extracting them accurately and efficiently from polyphonic audio. Early methods relied on simple frameworks that limited downstream performance, while more recent systems deliver stronger results but require large models integrated within complex multimodal architectures. We introduce LIVI (Lyrics-Informed Version Identification), an approach that seeks to balance retrieval accuracy with computational efficiency. First, LIVI leverages supervision from state-of-the-art transcription and text embedding models during training to achieve retrieval accuracy on par with–or superior to–harmonic-based systems. Second, LIVI remains lightweight and efficient by removing the transcription step at inference, challenging the dominance of complexity-heavy pipelines.

Method: Proposes MARS with two stages: 1) Multi-modal retrieval to get candidate historical contexts with high acoustic/textual similarity to current utterance, 2) Multi-modal selection that calculates both acoustic and textual similarities and uses a near-ideal ranking method to select the best historical context.

Result: LLM-ASR trained on only 1.5K hours of data with MARS outperforms state-of-the-art top-ranking system trained on 179K hours of data on Interspeech 2025 Multilingual Conversational Speech Language Model Challenge dataset.

Conclusion: MARS effectively addresses context selection problem in conversational LLM-ASR by retrieving and selecting the most relevant historical context, significantly improving performance while reducing computational costs and training data requirements.

Abstract: Automatic Speech Recognition (ASR) aims to convert human speech content into corresponding text. In conversational scenarios, effectively utilizing context can enhance its accuracy. Large Language Models’ (LLMs) exceptional long-context understanding and reasoning abilities enable LLM-based ASR (LLM-ASR) to leverage historical context for recognizing conversational speech, which has a high degree of contextual relevance. However, existing conversational LLM-ASR methods use a fixed number of preceding utterances or the entire conversation history as context, resulting in significant ASR confusion and computational costs due to massive irrelevant and redundant information. This paper proposes a multi-modal retrieval-and-selection method named MARS that augments conversational LLM-ASR by enabling it to retrieve and select the most relevant acoustic and textual historical context for the current utterance. Specifically, multi-modal retrieval obtains a set of candidate historical contexts, each exhibiting high acoustic or textual similarity to the current utterance. Multi-modal selection calculates the acoustic and textual similarities for each retrieved candidate historical context and, by employing our proposed near-ideal ranking method to consider both similarities, selects the best historical context. Evaluations on the Interspeech 2025 Multilingual Conversational Speech Language Model Challenge dataset show that the LLM-ASR, when trained on only 1.5K hours of data and equipped with the MARS, outperforms the state-of-the-art top-ranking system trained on 179K hours of data.

Method: Introduces three families of analytic bijections (cubic rational, sinh, cubic polynomial) as drop-in replacements in coupling flows. Also develops radial flows - a novel architecture using direct parametrization that transforms radial coordinate while preserving angular direction.

Result: New bijections match or exceed spline performance in coupling flows. Radial flows show exceptional training stability, geometric interpretability, and on radially structured targets achieve comparable quality to coupling flows with 1000× fewer parameters. Outperform affine baselines on φ⁴ lattice field theory physics problems.

Conclusion: The proposed analytic bijections combine favorable properties of prior approaches (smoothness, global domain, analytic invertibility). Radial flows offer parameter efficiency and stability advantages, enabling problem-specific designs that address mode collapse in high-dimensional physics applications.

Abstract: A key challenge in designing normalizing flows is finding expressive scalar bijections that remain invertible with tractable Jacobians. Existing approaches face trade-offs: affine transformations are smooth and analytically invertible but lack expressivity; monotonic splines offer local control but are only piecewise smooth and act on bounded domains; residual flows achieve smoothness but need numerical inversion. We introduce three families of analytic bijections – cubic rational, sinh, and cubic polynomial – that are globally smooth ($C^\infty$), defined on all of $\mathbb{R}$, and analytically invertible in closed form, combining the favorable properties of all prior approaches. These bijections serve as drop-in replacements in coupling flows, matching or exceeding spline performance. Beyond coupling layers, we develop radial flows: a novel architecture using direct parametrization that transforms the radial coordinate while preserving angular direction. Radial flows exhibit exceptional training stability, produce geometrically interpretable transformations, and on targets with radial structure can achieve comparable quality to coupling flows with $1000\times$ fewer parameters. We provide comprehensive evaluation on 1D and 2D benchmarks, and demonstrate applicability to higher-dimensional physics problems through experiments on $φ^4$ lattice field theory, where our bijections outperform affine baselines and enable problem-specific designs that address mode collapse.

Method: Proposes UOWQ framework based on asymptotic analysis of KL-divergence generalization error. Formulates multi-source transfer learning as parameter estimation problem, provides closed-form solutions for single-source case and convex optimization for multi-source case. Includes practical algorithms for multi-source transfer and multi-task learning.

Result: Theoretical proof that using all available source samples is optimal with proper weight adjustment. Extensive experiments on DomainNet and Office-Home benchmarks show UOWQ consistently outperforms strong baselines.

Conclusion: UOWQ provides a unified theoretical framework for joint optimization of source weights and transfer quantities, effectively preventing negative transfer and improving transfer learning performance in data-scarce scenarios.

Abstract: Transfer learning plays a vital role in improving model performance in data-scarce scenarios. However, naive uniform transfer from multiple source tasks may result in negative transfer, highlighting the need to properly balance the contributions of heterogeneous sources. Moreover, existing transfer learning methods typically focus on optimizing either the source weights or the amount of transferred samples, while largely neglecting the joint consideration of the other. In this work, we propose a theoretical framework, Unified Optimization of Weights and Quantities (UOWQ), which formulates multi-source transfer learning as a parameter estimation problem grounded in an asymptotic analysis of a Kullback-Leibler divergence-based generalization error measure. The proposed framework jointly determines the optimal source weights and optimal transfer quantities for each source task. Firstly, we prove that using all available source samples is always optimal once the weights are properly adjusted, and we provide a theoretical explanation for this phenomenon. Moreover, to determine the optimal transfer weights, our analysis yields closed-form solutions in the single-source setting and develops a convex optimization-based numerical procedure for the multi-source case. Building on the theoretical results, we further propose practical algorithms for both multi-source transfer learning and multi-task learning settings. Extensive experiments on real-world benchmarks, including DomainNet and Office-Home, demonstrate that UOWQ consistently outperforms strong baselines. The results validate both the theoretical predictions and the practical effectiveness of our framework.

Method: Planner-guided, constrained-topology multi-agent framework where an LLM-powered planner uses a graph representation of team environment to orchestrate agent calls, generate context-aware prompts, and retroactively correct errors using downstream failures.

Result: 38% improvement over manually crafted workflows and 150% improvement over unplanned workflows on in-house dataset; reduced feature engineering cycles from 3 weeks to 1 day for recommendation models serving 120M+ users.

Conclusion: The framework successfully addresses real-world feature engineering challenges by enabling reliable, maintainable code generation aligned with team expectations through intelligent planning and human-in-the-loop intervention.

Abstract: Recent advances in code generation models have unlocked unprecedented opportunities for automating feature engineering, yet their adoption in real-world ML teams remains constrained by critical challenges: (i) the scarcity of datasets capturing the iterative and complex coding processes of production-level feature engineering, (ii) limited integration and personalization of widely used coding agents, such as CoPilot and Devin, with a team’s unique tools, codebases, workflows, and practices, and (iii) suboptimal human-AI collaboration due to poorly timed or insufficient feedback. We address these challenges with a planner-guided, constrained-topology multi-agent framework that generates code for repositories in a multi-step fashion. The LLM-powered planner leverages a team’s environment, represented as a graph, to orchestrate calls to available agents, generate context-aware prompts, and use downstream failures to retroactively correct upstream artifacts. It can request human intervention at critical steps, ensuring generated code is reliable, maintainable, and aligned with team expectations. On a novel in-house dataset, our approach achieves 38% and 150% improvement in the evaluation metric over manually crafted and unplanned workflows respectively. In practice, when building features for recommendation models serving over 120 million users, our approach has delivered real-world impact by reducing feature engineering cycles from three weeks to a single day.

Method: Systematic study of Tensor Network models (Matrix Product States and Tree Tensor Networks) using low-level jet constituent features, with post-training quantization for hardware-efficient FPGA implementation.

Result: Models achieve competitive performance compared to state-of-the-art deep learning classifiers, with sub-microsecond latency and efficient FPGA resource usage after quantization.

Conclusion: Tensor Network-based models demonstrate potential for fast, resource-efficient inference in low-latency environments like real-time trigger systems.

Abstract: We present a systematic study of Tensor Network (TN) models $\unicode{x2013}$ Matrix Product States (MPS) and Tree Tensor Networks (TTN) $\unicode{x2013}$ for real-time jet tagging in high-energy physics, with a focus on low-latency deployment on Field Programmable Gate Arrays (FPGAs). Motivated by the strict requirements of the HL-LHC Level-1 trigger system, we explore TNs as compact and interpretable alternatives to deep neural networks. Using low-level jet constituent features, our models achieve competitive performance compared to state-of-the-art deep learning classifiers. We investigate post-training quantization to enable hardware-efficient implementations without degrading classification performance or latency. The best-performing models are synthesized to estimate FPGA resource usage, latency, and memory occupancy, demonstrating sub-microsecond latency and supporting the feasibility of online deployment in real-time trigger systems. Overall, this study highlights the potential of TN-based models for fast and resource-efficient inference in low-latency environments.

Method: Introduces the Regenerative Logic-Core Protocol (RLCP), a dual-stream training framework that renders specific factual dependencies linearly undecodable via deep-layer gradient reversal. Applied to Qwen2.5-0.5B model to validate the digital metabolism hypothesis.

Result: The model achieves near-zero retention of targeted factual associations (Accuracy < 7%) while exhibiting changes consistent with emergent “structural crystallization.” On GSM8K, the metabolized model spontaneously adopts chain-of-thought scaffolding, compensating for loss of direct associative recall by shifting from O(1) recall to O(N) reasoning.

Conclusion: The findings provide a dynamic weight-level counterpart to architectural innovations like DeepSeek’s Engram, paving the way for modular “Neural CPU + Symbolic RAM” architectures. The causal mechanism underlying the behavioral shift requires further investigation.

Abstract: Large language models (LLMs) currently suffer from parameter entanglement, where general reasoning capabilities (logic) and specific factual knowledge (facts) exist in a superposition state within shared weights. This coupling leads to the “memory wall,” where computational capacity is squandered on simulating retrieval, often resulting in hallucinations. In this paper, we propose “digital metabolism,” a thermodynamic hypothesis suggesting that targeted forgetting is necessary for distilling a pure neural logic core. To validate this hypothesis, we introduce the Regenerative Logic-Core Protocol (RLCP), a dual-stream training framework that renders specific factual dependencies linearly undecodable via deep-layer gradient reversal. Applying RLCP to Qwen2.5-0.5B, we observe a distinct phase transition: the model achieves near-zero retention of targeted factual associations (Accuracy < 7%) while exhibiting changes consistent with an emergent “structural crystallization” effect. Empirical analysis on GSM8K reveals that the “metabolized” model spontaneously adopts chain-of-thought (CoT) scaffolding, which we interpret as compensating for the loss of direct associative recall (shifting from $O(1)$ recall to $O(N)$ reasoning). While the causal mechanism underlying this behavioral shift requires further investigation, our findings provide a dynamic weight-level counterpart to architectural innovations like DeepSeek’s Engram, paving the way for modular “Neural CPU + Symbolic RAM” architectures.

Method: 1) Generalize VLP for nonlinear approximations using value-centric approach; 2) Optimize VLP for small-batch GEMMs with asymmetric inputs; 3) Design Mugi architecture to support full LLM workloads with these innovations.

Result: Mugi achieves up to 45× throughput and 668× energy efficiency for nonlinear softmax operations, 2.07× throughput and 3.11× energy efficiency for LLMs, while reducing operational carbon by 1.45× and embodied carbon by 1.48×.

Conclusion: VLP can significantly benefit LLMs through generalized nonlinear approximations and optimized small-batch GEMMs, with the Mugi architecture demonstrating substantial improvements in performance, efficiency, and sustainability.

Abstract: Value level parallelism (VLP) has been proposed to improve the efficiency of large-batch, low-precision general matrix multiply (GEMM) between symmetric activations and weights. In transformer based large language models (LLMs), there exist more sophisticated operations beyond activation-weight GEMM. In this paper, we explore how VLP benefits LLMs. First, we generalize VLP for nonlinear approximations, outperforming existing nonlinear approximations in end-to-end LLM accuracy, performance, and efficiency. Our VLP approximation follows a value-centric approach, where important values are assigned with greater accuracy. Second, we optimize VLP for small-batch GEMMs with asymmetric inputs efficiently, which leverages timely LLM optimizations, including weight-only quantization, key-value (KV) cache quantization, and group query attention. Finally, we design a new VLP architecture, Mugi, to encapsulate the innovations above and support full LLM workloads, while providing better performance, efficiency and sustainability. Our experimental results show that Mugi can offer significant improvements on throughput and energy efficiency, up to $45\times$ and $668\times$ for nonlinear softmax operations, and $2.07\times$ and $3.11\times$ for LLMs, and also decrease operational carbon for LLM operation by $1.45\times$ and embodied carbon by $1.48\times$.

Method: Proposes M-LCB, a computationally efficient UCB-style meta-algorithm that builds confidence intervals directly from realized losses without additional optimization. It selects one predictor among K adaptive experts each round while updating at most M of them under fixed training budget.

Result: If each expert achieves internal regret Õ(T^α), then M-LCB ensures overall regret bounded by Õ(√(KT/M) + (K/M)^{1-α}T^α). This is the first result establishing regret guarantees when multiple adaptive experts are trained simultaneously under per-round budget constraints.

Conclusion: M-LCB extends classical bandit paradigm to coordinate stateful, self-learning experts under limited resources, with applications to parametric models trained online and experts that are themselves multi-armed bandit algorithms.

Abstract: In many modern applications, a system must dynamically choose between several adaptive learning algorithms that are trained online. Examples include model selection in streaming environments, switching between trading strategies in finance, and orchestrating multiple contextual bandit or reinforcement learning agents. At each round, a learner must select one predictor among $K$ adaptive experts to make a prediction, while being able to update at most $M \le K$ of them under a fixed training budget. We address this problem in the \emph{stochastic setting} and introduce \algname{M-LCB}, a computationally efficient UCB-style meta-algorithm that provides \emph{anytime regret guarantees}. Its confidence intervals are built directly from realized losses, require no additional optimization, and seamlessly reflect the convergence properties of the underlying experts. If each expert achieves internal regret $\tilde O(T^α)$, then \algname{M-LCB} ensures overall regret bounded by $\tilde O!\Bigl(\sqrt{\tfrac{KT}{M}} ;+; (K/M)^{1-α},T^α\Bigr)$. To our knowledge, this is the first result establishing regret guarantees when multiple adaptive experts are trained simultaneously under per-round budget constraints. We illustrate the framework with two representative cases: (i) parametric models trained online with stochastic losses, and (ii) experts that are themselves multi-armed bandit algorithms. These examples highlight how \algname{M-LCB} extends the classical bandit paradigm to the more realistic scenario of coordinating stateful, self-learning experts under limited resources.

Method: U-Net architecture configured as image segmentation task, trained on synthetic datasets of human preferences (identifying longest topological member or most complex structural connection).

Result: Model successfully predicts plausible modification regions, generalizes across diverse TO problems, and enables 39% improvement in linear buckling load with only 15 sec additional design time.

Conclusion: AI co-pilot reduces iterative trials in human-in-the-loop TO, making optimization more efficient while maintaining user control and improving manufacturability/performance.

Abstract: Inverse design tools such as Topology Optimization (TO) can achieve new levels of improvement for high-performance engineered structures. However, widespread use is hindered by high computational times and a black-box nature that inhibits user interaction. Human-in-the-loop TO approaches are emerging that integrate human intuition into the design generation process. However, these rely on the time-consuming bottleneck of iterative region selection for design modifications. To reduce the number of iterative trials, this contribution presents an AI co-pilot that uses machine learning to predict the user’s preferred regions. The prediction model is configured as an image segmentation task with a U-Net architecture. It is trained on synthetic datasets where human preferences either identify the longest topological member or the most complex structural connection. The model successfully predicts plausible regions for modification and presents them to the user as AI recommendations. The human preference model demonstrates generalization across diverse and non-standard TO problems and exhibits emergent behavior outside the single-region selection training data. Demonstration examples show that the new human-in-the-loop TO approach that integrates the AI co-pilot can improve manufacturability or improve the linear buckling load by 39% while only increasing the total design time by 15 sec compared to conventional simplistic TO.

Method: The authors introduce the forecast accuracy and coherence (AC) score, a new metric that measures both multi-horizon accuracy and stability. The score allows user-specified weights to balance accuracy and consistency requirements. They implement it as a differentiable objective function for training seasonal ARIMA models.

Result: When evaluated on the M4 Hourly benchmark dataset, AC-optimized models achieve a 75% reduction in forecast volatility for the same target timestamps while maintaining comparable or improved point forecast accuracy compared to traditional maximum likelihood estimation.

Conclusion: The forecast AC score provides a better way to measure and optimize forecast quality by accounting for both accuracy and temporal consistency, leading to more stable and reliable forecasts without sacrificing accuracy.

Abstract: Traditional time series forecasting methods optimize for accuracy alone. This objective neglects temporal consistency, in other words, how consistently a model predicts the same future event as the forecast origin changes. We introduce the forecast accuracy and coherence score (forecast AC score for short) for measuring the quality of probabilistic multi-horizon forecasts in a way that accounts for both multi-horizon accuracy and stability. Our score additionally provides for user-specified weights to balance accuracy and consistency requirements. As an example application, we implement the score as a differentiable objective function for training seasonal ARIMA models and evaluate it on the M4 Hourly benchmark dataset. Results demonstrate substantial improvements over traditional maximum likelihood estimation. Our AC-optimized models achieve a 75% reduction in forecast volatility for the same target timestamps while maintaining comparable or improved point forecast accuracy.

Method: Formulate invariance at backward adjoint/optimization geometry level. Replace Euclidean transpose with Unit-Consistent (UC) adjoint to derive UC gauge-consistent steepest descent and backpropagation.

Result: Develops operator-level recipe applicable uniformly across network components and optimizer state. Provides gauge-consistent optimization framework complementary to prior rescaling-invariant methods.

Conclusion: UC adjoint enables gauge-invariant optimization for positively homogeneous networks, addressing parameterization dependence in standard gradient descent through consistent backward geometry.

Abstract: Deep neural networks constructed from linear maps and positively homogeneous nonlinearities (e.g., ReLU) possess a fundamental gauge symmetry: the network function is invariant to node-wise diagonal rescalings. However, standard gradient descent is not equivariant to this symmetry, causing optimization trajectories to depend heavily on arbitrary parameterizations. Prior work has proposed rescaling-invariant optimization schemes for positively homogeneous networks (e.g., path-based or path-space updates). Our contribution is complementary: we formulate the invariance requirement at the level of the backward adjoint/optimization geometry, which provides a simple, operator-level recipe that can be applied uniformly across network components and optimizer state. By replacing the Euclidean transpose with a Unit-Consistent (UC) adjoint, we derive UC gauge-consistent steepest descent and backprogation.

Method: Introduces Action Shapley as an agnostic metric for unbiased training data selection, with a randomized dynamic algorithm to overcome exponential complexity of traditional Shapley value computations.

Result: The algorithm achieves >80% computational efficiency improvement over exponential-time methods. Action Shapley-based data selection consistently outperforms ad-hoc selection across five real-world case studies.

Conclusion: Action Shapley provides an effective, efficient solution for training data selection in world models, addressing both computational complexity and selection bias issues in data-constrained RL scenarios.

Abstract: Numerous offline and model-based reinforcement learning systems incorporate world models to emulate the inherent environments. A world model is particularly important in scenarios where direct interactions with the real environment is costly, dangerous, or impractical. The efficacy and interpretability of such world models are notably contingent upon the quality of the underlying training data. In this context, we introduce Action Shapley as an agnostic metric for the judicious and unbiased selection of training data. To facilitate the computation of Action Shapley, we present a randomized dynamic algorithm specifically designed to mitigate the exponential complexity inherent in traditional Shapley value computations. Through empirical validation across five data-constrained real-world case studies, the algorithm demonstrates a computational efficiency improvement exceeding 80% in comparison to conventional exponential time computations. Furthermore, our Action Shapley-based training data selection policy consistently outperforms ad-hoc training data selection.

Method: Proposes a Curriculum Reinforcement Learning framework with: 1) data-driven marine simulation environment using real vessel movement data enhanced with Diffusion Model for dynamic conditions, 2) machine learning-based fuel consumption prediction module, 3) image-based environment representation, 4) lightweight policy-based CRL agent with comprehensive reward mechanism covering safety, emissions, timeliness, and goal completion.

Result: The framework effectively handles complex tasks progressively while ensuring stable and efficient learning in continuous action spaces. Validated in the Indian Ocean sea area, demonstrating efficacy in enabling sustainable and safe vessel navigation.

Conclusion: The proposed CRL framework successfully addresses the limitations of traditional vessel navigation by providing an autonomous, emission-aware solution that optimizes both environmental sustainability and navigational safety through realistic simulation and comprehensive learning objectives.

Abstract: Sustainability is becoming increasingly critical in the maritime transport, encompassing both environmental and social impacts, such as Greenhouse Gas (GHG) emissions and navigational safety. Traditional vessel navigation heavily relies on human experience, often lacking autonomy and emission awareness, and is prone to human errors that may compromise safety. In this paper, we propose a Curriculum Reinforcement Learning (CRL) framework integrated with a realistic, data-driven marine simulation environment and a machine learning-based fuel consumption prediction module. The simulation environment is constructed using real-world vessel movement data and enhanced with a Diffusion Model to simulate dynamic maritime conditions. Vessel fuel consumption is estimated using historical operational data and learning-based regression. The surrounding environment is represented as image-based inputs to capture spatial complexity. We design a lightweight, policy-based CRL agent with a comprehensive reward mechanism that considers safety, emissions, timeliness, and goal completion. This framework effectively handles complex tasks progressively while ensuring stable and efficient learning in continuous action spaces. We validate the proposed approach in a sea area of the Indian Ocean, demonstrating its efficacy in enabling sustainable and safe vessel navigation.

Method: FAConvLSTM uses factorized gate computations with 1×1 bottlenecks and shared depthwise spatial mixing, multi-scale dilated depthwise branches with squeeze-and-excitation, peephole connections, lightweight axial spatial attention applied sparsely, and a subspace head with temporal self-attention and seasonal positional encoding.

Result: Experiments show FAConvLSTM produces more stable, interpretable, and robust latent representations than standard ConvLSTM while significantly reducing computational overhead on multivariate spatiotemporal climate data.

Conclusion: FAConvLSTM effectively addresses ConvLSTM2D’s limitations by improving efficiency, spatial expressiveness, and physical interpretability for Earth observation data through factorized attention and multi-scale processing.

Abstract: Learning physically meaningful spatiotemporal representations from high-resolution multivariate Earth observation data is challenging due to strong local dynamics, long-range teleconnections, multi-scale interactions, and nonstationarity. While ConvLSTM2D is a commonly used baseline, its dense convolutional gating incurs high computational cost and its strictly local receptive fields limit the modeling of long-range spatial structure and disentangled climate dynamics. To address these limitations, we propose FAConvLSTM, a Factorized-Attention ConvLSTM layer designed as a drop-in replacement for ConvLSTM2D that simultaneously improves efficiency, spatial expressiveness, and physical interpretability. FAConvLSTM factorizes recurrent gate computations using lightweight [1 times 1] bottlenecks and shared depthwise spatial mixing, substantially reducing channel complexity while preserving recurrent dynamics. Multi-scale dilated depthwise branches and squeeze-and-excitation recalibration enable efficient modeling of interacting physical processes across spatial scales, while peephole connections enhance temporal precision. To capture teleconnection-scale dependencies without incurring global attention cost, FAConvLSTM incorporates a lightweight axial spatial attention mechanism applied sparsely in time. A dedicated subspace head further produces compact per timestep embeddings refined through temporal self-attention with fixed seasonal positional encoding. Experiments on multivariate spatiotemporal climate data shows superiority demonstrating that FAConvLSTM yields more stable, interpretable, and robust latent representations than standard ConvLSTM, while significantly reducing computational overhead.

Method: HOSL strategically integrates zeroth-order optimization on client side (for memory-efficient gradient estimation without backpropagation) with first-order optimization on server side (for fast convergence). This hybrid approach eliminates client-side activation storage while maintaining optimization effectiveness.

Result: HOSL reduces client GPU memory by up to 3.7× compared to first-order methods while achieving accuracy within 0.20%-4.23% of this baseline. It outperforms zeroth-order baselines by up to 15.55%. Theoretical analysis shows convergence rate depends on client-side model dimension rather than full model dimension.

Conclusion: HOSL effectively addresses the trade-off between memory efficiency and optimization effectiveness in split learning, enabling memory-efficient training on edge devices while maintaining competitive performance through strategic hybrid-order optimization.

Abstract: Split learning (SL) enables collaborative training of large language models (LLMs) between resource-constrained edge devices and compute-rich servers by partitioning model computation across the network boundary. However, existing SL systems predominantly rely on first-order (FO) optimization, which requires clients to store intermediate quantities such as activations for backpropagation. This results in substantial memory overhead, largely negating benefits of model partitioning. In contrast, zeroth-order (ZO) optimization eliminates backpropagation and significantly reduces memory usage, but often suffers from slow convergence and degraded performance. In this work, we propose HOSL, a novel Hybrid-Order Split Learning framework that addresses this fundamental trade-off between memory efficiency and optimization effectiveness by strategically integrating ZO optimization on the client side with FO optimization on the server side. By employing memory-efficient ZO gradient estimation at the client, HOSL eliminates backpropagation and activation storage, reducing client memory consumption. Meanwhile, server-side FO optimization ensures fast convergence and competitive performance. Theoretically, we show that HOSL achieves a $\mathcal{O}(\sqrt{d_c/TQ})$ rate, which depends on client-side model dimension $d_c$ rather than the full model dimension $d$, demonstrating that convergence improves as more computation is offloaded to the server. Extensive experiments on OPT models (125M and 1.3B parameters) across 6 tasks demonstrate that HOSL reduces client GPU memory by up to 3.7$\times$ compared to the FO method while achieving accuracy within 0.20%-4.23% of this baseline. Furthermore, HOSL outperforms the ZO baseline by up to 15.55%, validating the effectiveness of our hybrid strategy for memory-efficient training on edge devices.

Method: Develops a multivariate Long Short-Term Memory (LSTM)-based network that uses real-world historical data from both local and neighboring areas. The model captures long-term dependencies and interactions between different renewable energy sources to enhance predictive accuracy.

Result: The proposed forecasting approach demonstrates improved performance in case studies, resulting in lower CO2 emissions and more reliable electric load supply compared to traditional methods.

Conclusion: The multivariate LSTM-based forecasting model effectively addresses the limitations of traditional methods by better capturing complex temporal patterns in renewable energy generation, leading to more accurate predictions that support reduced emissions and enhanced power system reliability.

Abstract: The increasing integration of renewable energy sources (RESs) into modern power systems presents significant opportunities but also notable challenges, primarily due to the inherent variability of RES generation. Accurate forecasting of RES generation is crucial for maintaining the reliability, stability, and economic efficiency of power system operations. Traditional approaches, such as deterministic methods and stochastic programming, frequently depend on representative scenarios generated through clustering techniques like K-means. However, these methods may fail to fully capture the complex temporal dependencies and non-linear patterns within RES data. This paper introduces a multivariate Long Short-Term Memory (LSTM)-based network designed to forecast RESs generation using their real-world historical data. The proposed model effectively captures long-term dependencies and interactions between different RESs, utilizing historical data from both local and neighboring areas to enhance predictive accuracy. In the case study, we showed that the proposed forecasting approach results in lower CO2 emissions, and a more reliable supply of electric loads.

Method: Analyzed SGD learning dynamics through numerical experiments and a tractable physical model, revealing a transient exploratory phase and effective potential reshaping by SGD noise.

Result: SGD noise reshapes the loss landscape into an effective potential favoring flat solutions, with a transient freezing mechanism where growing energy barriers eventually trap dynamics in a single basin. Increased noise delays freezing and enhances convergence to flatter minima.

Conclusion: Provides a unified physical framework linking learning dynamics, loss-landscape geometry, and generalization, offering principles for designing more effective optimization algorithms.

Abstract: Stochastic gradient descent (SGD) is central to deep learning, yet the dynamical origin of its preference for flatter, more generalizable solutions remains unclear. Here, by analyzing SGD learning dynamics, we identify a nonequilibrium mechanism governing solution selection. Numerical experiments reveal a transient exploratory phase in which SGD trajectories repeatedly escape sharp valleys and transition toward flatter regions of the loss landscape. By using a tractable physical model, we show that the SGD noise reshapes the landscape into an effective potential that favors flat solutions. Crucially, we uncover a transient freezing mechanism: as training proceeds, growing energy barriers suppress inter-valley transitions and ultimately trap the dynamics within a single basin. Increasing the SGD noise strength delays this freezing, which enhances convergence to flatter minima. Together, these results provide a unified physical framework linking learning dynamics, loss-landscape geometry, and generalization, and suggest principles for the design of more effective optimization algorithms.

Method: Proposes MGF-RL framework that couples first-order meta-learning with evolutionary strategies to learn transferable initialization from historical outage experiences. Enables scalable policy search without gradient computation while accommodating nonlinear, constrained distribution-system dynamics.

Result: Outperforms standard RL, MAML-based meta-RL, and model predictive control across reliability, restoration speed, and adaptation efficiency under renewable forecast errors. Generalizes to unseen outages and renewable patterns with substantially fewer fine-tuning episodes than conventional RL.

Conclusion: MGF-RL provides an effective solution for real-time load restoration in renewable-rich distribution grids, with theoretical sublinear regret bounds relating adaptation efficiency to task similarity and environmental variation.

Abstract: Restoring critical loads after extreme events demands adaptive control to maintain distribution-grid resilience, yet uncertainty in renewable generation, limited dispatchable resources, and nonlinear dynamics make effective restoration difficult. Reinforcement learning (RL) can optimize sequential decisions under uncertainty, but standard RL often generalizes poorly and requires extensive retraining for new outage configurations or generation patterns. We propose a meta-guided gradient-free RL (MGF-RL) framework that learns a transferable initialization from historical outage experiences and rapidly adapts to unseen scenarios with minimal task-specific tuning. MGF-RL couples first-order meta-learning with evolutionary strategies, enabling scalable policy search without gradient computation while accommodating nonlinear, constrained distribution-system dynamics. Experiments on IEEE 13-bus and IEEE 123-bus test systems show that MGF-RL outperforms standard RL, MAML-based meta-RL, and model predictive control across reliability, restoration speed, and adaptation efficiency under renewable forecast errors. MGF-RL generalizes to unseen outages and renewable patterns while requiring substantially fewer fine-tuning episodes than conventional RL. We also provide sublinear regret bounds that relate adaptation efficiency to task similarity and environmental variation, supporting the empirical gains and motivating MGF-RL for real-time load restoration in renewable-rich distribution grids.

Method: Proposed reasoning distillation where a large teacher model provides structured symbolic reasoning tags alongside fix-type labels. Trained CodeT5-based student model under label-only and reasoning-distilled settings on IntroClass benchmark.

Result: Reasoning supervision consistently improves macro averaged performance, especially on less frequent bug categories. Correct reasoning traces strongly correlate with correct predictions but don’t fully determine them.

Conclusion: Symbolic reasoning distillation is a practical way to improve interpretability and robustness in lightweight program repair models.

Abstract: We study whether lightweight symbolic reasoning supervision can improve fix type classification in compact automated program repair models. Small code models are attractive for resource-constrained settings, but they typically produce only a single prediction, making it unclear whether they learn meaningful program structure or rely on shallow correlations. We propose a reasoning distillation approach in which a large teacher model provides structured symbolic reasoning tags alongside fix-type labels. These tags capture high-level causal properties of bugs without relying on free-form explanations. We train a CodeT5-based student model under label-only and reasoning-distilled settings on the IntroClass benchmark. Reasoning supervision consistently improves macro averaged performance, particularly on less frequent bug categories, without increasing model size or complexity. We further analyze the relationship between reasoning accuracy and fix-type prediction, showing that correct reasoning traces strongly correlate with correct predictions, while not fully determining them. Our results suggest that symbolic reasoning distillation is a practical way to improve interpretability and robustness in lightweight program repair models.

Method: Provides a self-contained theoretical analysis of constant metric scaling on arbitrary Riemannian manifolds, systematically categorizing which mathematical objects transform under scaling and which remain invariant, with specific focus on computational implications.

Result: Shows that while norms, distances, volume elements, and gradient magnitudes scale with the metric, fundamental geometric structures like the Levi-Civita connection, geodesics, exponential/logarithmic maps, and parallel transport remain invariant under constant rescaling.

Conclusion: Constant metric scaling can be safely introduced in Riemannian computation as a global step size parameter without altering the underlying geometric structures, providing practical guidance for implementing Riemannian optimization algorithms while maintaining geometric integrity.

Abstract: Constant rescaling of a Riemannian metric appears in many computational settings, often through a global scale parameter that is introduced either explicitly or implicitly. Although this operation is elementary, its consequences are not always made clear in practice and may be confused with changes in curvature, manifold structure, or coordinate representation. In this note we provide a short, self-contained account of constant metric scaling on arbitrary Riemannian manifolds. We distinguish between quantities that change under such a scaling, including norms, distances, volume elements, and gradient magnitudes, and geometric objects that remain invariant, such as the Levi–Civita connection, geodesics, exponential and logarithmic maps, and parallel transport. We also discuss implications for Riemannian optimization, where constant metric scaling can often be interpreted as a global rescaling of step sizes rather than a modification of the underlying geometry. The goal of this note is purely expository and is intended to clarify how a global metric scale parameter can be introduced in Riemannian computation without altering the geometric structures on which these methods rely.

Method: Conducts a thorough comparative review and analysis of threat models, attack methods, target domains (vision, multimodal, graphs, federated learning), and available defenses in contrastive learning.

Result: Summarizes recent advancements, identifies specific vulnerabilities inherent to contrastive learning, and highlights challenges in securing these systems against malicious attacks.

Conclusion: The findings have significant implications for secure deployment in industrial and distributed environments, with identified research directions needed to address the security challenges in contrastive learning.

Abstract: Contrastive learning has become a leading self- supervised approach to representation learning across domains, including vision, multimodal settings, graphs, and federated learning. However, recent studies have shown that contrastive learning is susceptible to backdoor and data poisoning attacks. In these attacks, adversaries can manipulate pretraining data or model updates to insert hidden malicious behavior. This paper offers a thorough and comparative review of backdoor attacks in contrastive learning. It analyzes threat models, attack methods, target domains, and available defenses. We summarize recent advancements in this area, underline the specific vulnerabilities inherent to contrastive learning, and discuss the challenges and future research directions. Our findings have significant implications for the secure deployment of systems in industrial and distributed environments.

Method: Proposes a self-reflection framework that integrates with existing interpretable graph learning methods. When a method produces node/edge importance scores, the framework feeds these predictions back into the original method for a second round of evaluation. Also develops a fine-tuning training method based on this feedback mechanism.

Result: The self-reflection technique effectively enhances interpretability on Spurious-Motif datasets with strong spurious correlations, improving performance over existing methods that struggle with these challenging benchmarks.

Conclusion: Self-reflection techniques from large language models can be successfully adapted to improve interpretable graph learning, particularly for challenging datasets with spurious correlations. The proposed framework provides an effective approach to enhance model performance on difficult benchmarks.

Abstract: Interpretable graph learning has recently emerged as a popular research topic in machine learning. The goal is to identify the important nodes and edges of an input graph that are crucial for performing a specific graph reasoning task. A number of studies have been conducted in this area, and various benchmark datasets have been proposed to facilitate evaluation. Among them, one of the most challenging is the Spurious-Motif benchmark, introduced at ICLR 2022. The datasets in this synthetic benchmark are deliberately designed to include spurious correlations, making it particularly difficult for models to distinguish truly relevant structures from misleading patterns. As a result, existing methods exhibit significantly worse performance on this benchmark compared to others. In this paper, we focus on improving interpretability on the challenging Spurious-Motif datasets. We demonstrate that the self-reflection technique, commonly used in large language models to tackle complex tasks, can also be effectively adapted to enhance interpretability in datasets with strong spurious correlations. Specifically, we propose a self-reflection framework that can be integrated with existing interpretable graph learning methods. When such a method produces importance scores for each node and edge, our framework feeds these predictions back into the original method to perform a second round of evaluation. This iterative process mirrors how large language models employ self-reflective prompting to reassess their previous outputs. We further analyze the reasons behind this improvement from the perspective of graph representation learning, which motivates us to propose a fine-tuning training method based on this feedback mechanism.

Method: Adds an adapter network that pre-processes input images before feeding them to the classifier. The adapter is trained using a novel quantile loss that corrects distribution shift by matching high-dimensional geometric quantiles rather than modifying classifier weights.

Result: Theoretical proof shows that minimizing quantile loss can learn the optimal adapter under suitable conditions. Experimental validation on CIFAR10-C, CIFAR100-C, and TinyImageNet-C datasets with both convolutional and transformer networks demonstrates effectiveness.

Conclusion: Proposes a novel architecture-agnostic TTA approach that uses an adapter network with quantile loss, offering theoretical guarantees and practical effectiveness across different architectures and datasets.

Abstract: Test-time adaptation (TTA) refers to adapting a classifier for the test data when the probability distribution of the test data slightly differs from that of the training data of the model. To the best of our knowledge, most of the existing TTA approaches modify the weights of the classifier relying heavily on the architecture. It is unclear as to how these approaches are extendable to generic architectures. In this article, we propose an architecture-agnostic approach to TTA by adding an adapter network pre-processing the input images suitable to the classifier. This adapter is trained using the proposed quantile loss. Unlike existing approaches, we correct for the distribution shift by matching high-dimensional geometric quantiles. We prove theoretically that under suitable conditions minimizing quantile loss can learn the optimal adapter. We validate our approach on CIFAR10-C, CIFAR100-C and TinyImageNet-C by training both classic convolutional and transformer networks on CIFAR10, CIFAR100 and TinyImageNet datasets.

Method: A two-stage framework: 1) General AVP identification using panoramic feature space (10 descriptors) with hierarchical fusion architecture combining CNN, BiLSTM, self-attention, and adaptive gating; 2) Functional subtype prediction with transfer learning and OHEM-driven contrastive learning enhanced by BLOSUM62 to sharpen decision boundaries.

Result: First stage achieved accuracy of 0.9531 and MCC of 0.9064, outperforming SOTA methods. Second stage successfully classified 6 viral families and 8 specific viruses under small-sample conditions. Web interface available for accessibility.

Conclusion: AVP-Pro provides a powerful, interpretable tool for high-throughput screening of antiviral drugs, addressing key limitations in existing methods through innovative feature fusion and contrastive learning approaches.

Abstract: The accurate identification of antiviral peptides (AVPs) is crucial for novel drug development. However, existing methods still have limitations in capturing complex sequence dependencies and distinguishing confusing samples with high similarity. To address these challenges, we propose AVP-Pro, a novel two-stage predictive framework that integrates adaptive feature fusion and contrastive learning. To comprehensively capture the physicochemical properties and deep-seated patterns of peptide sequences, we constructed a panoramic feature space encompassing 10 distinct descriptors and designed a hierarchical fusion architecture. This architecture integrates self-attention and adaptive gating mechanisms to dynamically modulate the weights of local motifs extracted by CNNs and global dependencies captured by BiLSTMs based on sequence context. Targeting the blurred decision boundary caused by the high similarity between positive and negative sample sequences, we adopted an Online Hard Example Mining (OHEM)-driven contrastive learning strategy enhanced by BLOSUM62. This approach significantly sharpened the model’s discriminative power. Model evaluation results show that in the first stage of general AVP identification, the model achieved an accuracy of 0.9531 and an MCC of 0.9064, outperforming existing state-of-the-art (SOTA) methods. In the second stage of functional subtype prediction, combined with a transfer learning strategy, the model realized accurate classification of 6 viral families and 8 specific viruses under small-sample conditions. AVP-Pro provides a powerful and interpretable new tool for the high-throughput screening of antiviral drugs. To further enhance accessibility for users, we have developed a user-friendly web interface, which is available at https://wwwy1031-avp-pro.hf.space.

Method: Proposes a self-augmented strategy where the model’s own predictions are partially masked and fed back for iterative refinement. Designs a self-augmented mixture-of-experts model with multiple expert networks that iteratively and collaboratively estimate QoS values through inter-expert communication.

Result: Experiments on benchmark datasets show the method outperforms existing baselines and achieves competitive results.

Conclusion: The iterative augmentation process naturally aligns with mixture-of-experts architecture by enabling inter-expert communication, effectively addressing the sparsity challenge in QoS prediction.

Abstract: Quality of Service (QoS) prediction is one of the most fundamental problems in service computing and personalized recommendation. In the problem, there is a set of users and services, each associated with a set of descriptive features. Interactions between users and services produce feedback values, typically represented as numerical QoS metrics such as response time or availability. Given the observed feedback for a subset of user-service pairs, the goal is to predict the QoS values for the remaining pairs. A key challenge in QoS prediction is the inherent sparsity of user-service interactions, as only a small subset of feedback values is typically observed. To address this, we propose a self-augmented strategy that leverages a model’s own predictions for iterative refinement. In particular, we partially mask the predicted values and feed them back into the model to predict again. Building on this idea, we design a self-augmented mixture-of-experts model, where multiple expert networks iteratively and collaboratively estimate QoS values. We find that the iterative augmentation process naturally aligns with the MoE architecture by enabling inter-expert communication: in the second round, each expert receives the first-round predictions and refines its output accordingly. Experiments on benchmark datasets show that our method outperforms existing baselines and achieves competitive results.

Method: Developed OpFML (Operational Forecasting with Machine Learning), a configurable and adaptable pipeline that can serve machine learning models for periodic forecasting. The system is demonstrated through application to daily Fire Danger Index forecasting.

Result: Created a working pipeline (OpFML) that can be utilized for operational forecasting tasks, specifically demonstrated for wildfire danger assessment. The pipeline includes various features for serving ML models in forecasting applications.

Conclusion: OpFML provides a practical solution for deploying machine learning models in operational forecasting systems, addressing the limitations of conventional methods in wildfire danger assessment and potentially other climate/earth science forecasting applications.

Abstract: Machine learning is finding its application in a multitude of areas in science and research, and Climate and Earth Sciences is no exception to this trend. Operational forecasting systems based on data-driven approaches and machine learning methods deploy models for periodic forecasting. Wildfire danger assessment using machine learning has garnered significant interest in the last decade, as conventional methods often overestimate the risk of wildfires. In this work, we present the code OpFML: Operational Forecasting with Machine Learning. OpFML is a configurable and adaptable pipeline that can be utilized to serve a machine learning model for periodic forecasting. We further demonstrate the capabilities of the pipeline through its application to daily Fire Danger Index forecasting and outline its various features.

Method: Used Path Patching, Logit Lens, JSD analysis, and Neural Differential Equations to investigate the phenomenon, uncovering a hidden Anchor-Adapter circuit that facilitates memorization shortcuts.

Result: Identified a Functional Anchor in middle layers (L18-20) that triggers retrieval of memorized solutions, followed by Structural Adapters in later layers (L21+) that transform representations. Demonstrated bidirectional causal steering by scaling specific MLP keys.

Conclusion: Provides a mechanistic roadmap for identifying and mitigating data contamination in RLVR-tuned models, revealing how spurious rewards can trigger memorization shortcuts rather than genuine reasoning.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) is highly effective for enhancing LLM reasoning, yet recent evidence shows models like Qwen 2.5 achieve significant gains even with spurious or incorrect rewards. We investigate this phenomenon and identify a “Perplexity Paradox”: spurious RLVR triggers a divergence where answer-token perplexity drops while prompt-side coherence degrades, suggesting the model is bypassing reasoning in favor of memorization. Using Path Patching, Logit Lens, JSD analysis, and Neural Differential Equations, we uncover a hidden Anchor-Adapter circuit that facilitates this shortcut. We localize a Functional Anchor in the middle layers (L18-20) that triggers the retrieval of memorized solutions, followed by Structural Adapters in later layers (L21+) that transform representations to accommodate the shortcut signal. Finally, we demonstrate that scaling specific MLP keys within this circuit allows for bidirectional causal steering-artificially amplifying or suppressing contamination-driven performance. Our results provide a mechanistic roadmap for identifying and mitigating data contamination in RLVR-tuned models. Code is available at https://github.com/idwts/How-RLVR-Activates-Memorization-Shortcuts.

Method: The model has two key components: 1) Cross-view inconsistency perception module inspired by hippocampus scene conflict monitoring, which captures subtle discrepancies across multiple transaction views to detect camouflaged fraud. 2) Novelty-aware hypergraph learning module inspired by CA1 region’s match-mismatch novelty detection, which measures feature deviations from neighborhood expectations and adaptively reweights messages to enhance sensitivity to rare fraud patterns.

Result: Extensive experiments on six web-based financial fraud datasets show HIMVH achieves average improvements of 6.42% in AUC, 9.74% in F1, and 39.14% in AP over 15 state-of-the-art models.

Conclusion: HIMVH effectively addresses key challenges in web finance fraud detection by drawing inspiration from hippocampal mechanisms, demonstrating superior performance in detecting both camouflaged and rare fraudulent behaviors in long-tailed financial data.

Abstract: Online financial services constitute an essential component of contemporary web ecosystems, yet their openness introduces substantial exposure to fraud that harms vulnerable users and weakens trust in digital finance. Such threats have become a significant web harm that erodes societal fairness and affects the well being of online communities. However, existing detection methods based on graph neural networks (GNNs) struggle with two persistent challenges: (1) fraud camouflage, where malicious transactions mimic benign behaviors to evade detection, and (2) long-tailed data distributions, which obscure rare but critical fraudulent cases. To fill these gaps, we propose HIMVH, a Hippocampus-Inspired Multi-View Hypergraph learning model for web finance fraud detection. Specifically, drawing inspiration from the scene conflict monitoring role of the hippocampus, we design a cross-view inconsistency perception module that captures subtle discrepancies and behavioral heterogeneity across multiple transaction views. This module enables the model to identify subtle cross-view conflicts for detecting online camouflaged fraudulent behaviors. Furthermore, inspired by the match-mismatch novelty detection mechanism of the CA1 region, we introduce a novelty-aware hypergraph learning module that measures feature deviations from neighborhood expectations and adaptively reweights messages, thereby enhancing sensitivity to online rare fraud patterns in the long-tailed settings. Extensive experiments on six web-based financial fraud datasets demonstrate that HIMVH achieves 6.42% improvement in AUC, 9.74% in F1 and 39.14% in AP on average over 15 SOTA models.

Method: Proposes Soft-BCT with soft (probabilistic) context space splits instead of hard splits. Develops a learning algorithm based on variational inference for parameter estimation.

Result: On real-world datasets, Soft-BCT demonstrates comparable or superior performance to previous BCT models with hard splits.

Conclusion: Soft-BCT provides a more flexible Bayesian context tree framework for real-valued time series through probabilistic context splits, achieving competitive or better performance than deterministic alternatives.

Abstract: This paper proposes the soft Bayesian context tree model (Soft-BCT), which is a novel BCT model for real-valued time series. The Soft-BCT considers soft (probabilistic) splits of the context space, instead of hard (deterministic) splits of the context space as in the previous BCT for real-valued time series. A learning algorithm of the Soft-BCT is proposed based on the variational inference. For some real-world datasets, the Soft-BCT demonstrates almost the same or superior performance to the previous BCT.

Method: Two-stage approach: 1) Identify low-dimensional task-specific subspace by analyzing principal gradient directions; 2) Project full gradients onto subspace, add DP noise, map perturbed gradients back to original parameter space for model updates.

Result: Experiments show DP-SFT enhances accuracy and stability under DP constraints, accelerates convergence, and achieves substantial gains over DP fine-tuning baselines on multiple datasets.

Conclusion: DP-SFT effectively reduces noise magnitude while preserving formal DP guarantees by focusing noise injection on task-specific subspaces, addressing the performance degradation problem in DP fine-tuning.

Abstract: Fine-tuning large language models on downstream tasks is crucial for realizing their cross-domain potential but often relies on sensitive data, raising privacy concerns. Differential privacy (DP) offers rigorous privacy guarantees and has been widely adopted in fine-tuning; however, naively injecting noise across the high-dimensional parameter space creates perturbations with large norms, degrading performance and destabilizing training. To address this issue, we propose DP-SFT, a two-stage subspace fine-tuning method that substantially reduces noise magnitude while preserving formal DP guarantees. Our intuition is that, during fine-tuning, significant parameter updates lie within a low-dimensional, task-specific subspace, while other directions change minimally. Hence, we only inject DP noise into this subspace to protect privacy without perturbing irrelevant parameters. In phase one, we identify the subspace by analyzing principal gradient directions to capture task-specific update signals. In phase two, we project full gradients onto this subspace, add DP noise, and map the perturbed gradients back to the original parameter space for model updates, markedly lowering noise impact. Experiments on multiple datasets demonstrate that DP-SFT enhances accuracy and stability under rigorous DP constraints, accelerates convergence, and achieves substantial gains over DP fine-tuning baselines.

Method: Proposes constraint-set generation instead of pairwise constraints to reduce LLM queries. Develops constrained clustering algorithm with confidence threshold and penalty mechanism to handle potentially inaccurate LLM-generated constraints.

Result: Evaluated on five text datasets, achieves comparable clustering accuracy to state-of-the-art methods while reducing LLM queries by more than 20 times. Also improves query efficiency and constraint accuracy.

Conclusion: The proposed constraint-set generation approach significantly reduces resource consumption while maintaining clustering quality, making LLM-based constrained clustering more practical and efficient.

Abstract: Clustering is a fundamental tool that has garnered significant interest across a wide range of applications including text analysis. To improve clustering accuracy, many researchers have incorporated background knowledge, typically in the form of must-link and cannot-link constraints, to guide the clustering process. With the recent advent of large language models (LLMs), there is growing interest in improving clustering quality through LLM-based automatic constraint generation. In this paper, we propose a novel constraint-generation approach that reduces resource consumption by generating constraint sets rather than using traditional pairwise constraints. This approach improves both query efficiency and constraint accuracy compared to state-of-the-art methods. We further introduce a constrained clustering algorithm tailored to the characteristics of LLM-generated constraints. Our method incorporates a confidence threshold and a penalty mechanism to address potentially inaccurate constraints. We evaluate our approach on five text datasets, considering both the cost of constraint generation and the overall clustering performance. The results show that our method achieves clustering accuracy comparable to the state-of-the-art algorithms while reducing the number of LLM queries by more than 20 times.

Method: Spectral and Spatial Neural Operator (S2NO) integrates Laplacian eigenfunction encoding for global behavior and spatial convolutions for local behavior on irregular domains, combined with evolutionary algorithms for voxel-level optimization.

Result: Enables high-fidelity morphing prediction on complex geometries including irregular-boundary shapes, porous structures, and thin-walled structures, with super-resolution material distribution design capability.

Conclusion: S2NO significantly improves efficiency and capability of programming complex shape morphing, expanding design diversity and complexity for advanced applications.

Abstract: Shape-morphing soft materials can enable diverse target morphologies through voxel-level material distribution design, offering significant potential for various applications. Despite progress in basic shape-morphing design with simple geometries, achieving advanced applications such as conformal implant deployment or aerodynamic morphing requires accurate and diverse morphing designs on complex geometries, which remains challenging. Here, we present a Spectral and Spatial Neural Operator (S2NO), which enables high-fidelity morphing prediction on complex geometries. S2NO effectively captures global and local morphing behaviours on irregular computational domains by integrating Laplacian eigenfunction encoding and spatial convolutions. Combining S2NO with evolutionary algorithms enables voxel-level optimisation of material distributions for shape morphing programming on various complex geometries, including irregular-boundary shapes, porous structures, and thin-walled structures. Furthermore, the neural operator’s discretisation-invariant property enables super-resolution material distribution design, further expanding the diversity and complexity of morphing design. These advancements significantly improve the efficiency and capability of programming complex shape morphing.

Method: Proposed Federated Survival Learning framework with Bayesian Differential Privacy (FSL-BDP). This approach models time-to-default trajectories without centralizing sensitive borrower data. The framework provides Bayesian (data-dependent) differential privacy guarantees while enabling multiple financial institutions to jointly learn risk dynamics through federated learning.

Result: Experiments on three real-world credit datasets (LendingClub, SBA, Bondora) show federation fundamentally changes privacy mechanism effectiveness. While classical DP performs better than Bayesian DP in centralized settings, Bayesian DP benefits substantially more from federation (+7.0% vs +1.4%), achieving near parity with non-private performance and outperforming classical DP for most participating clients. This ranking reversal reveals that privacy mechanism selection should be evaluated in target deployment architecture rather than centralized benchmarks.

Conclusion: The findings provide actionable guidance for practitioners designing privacy-preserving decision support systems in regulated, multi-institutional environments. The proposed FSL-BDP framework addresses both limitations of traditional credit risk models while complying with data protection regulations, enabling effective cross-institution learning without data sharing.

Abstract: Credit risk models are a critical decision-support tool for financial institutions, yet tightening data-protection rules (e.g., GDPR, CCPA) increasingly prohibit cross-border sharing of borrower data, even as these models benefit from cross-institution learning. Traditional default prediction suffers from two limitations: binary classification ignores default timing, treating early defaulters (high loss) equivalently to late defaulters (low loss), and centralized training violates emerging regulatory constraints. We propose a Federated Survival Learning framework with Bayesian Differential Privacy (FSL-BDP) that models time-to-default trajectories without centralizing sensitive data. The framework provides Bayesian (data-dependent) differential privacy (DP) guarantees while enabling institutions to jointly learn risk dynamics. Experiments on three real-world credit datasets (LendingClub, SBA, Bondora) show that federation fundamentally alters the relative effectiveness of privacy mechanisms. While classical DP performs better than Bayesian DP in centralized settings, the latter benefits substantially more from federation (+7.0% vs +1.4%), achieving near parity of non-private performance and outperforming classical DP in the majority of participating clients. This ranking reversal yields a key decision-support insight: privacy mechanism selection should be evaluated in the target deployment architecture, rather than centralized benchmarks. These findings provide actionable guidance for practitioners designing privacy-preserving decision support systems in regulated, multi-institutional environments.

Method: 1) Context graph encoding chemical knowledge linking functional groups, molecules, and properties; 2) Learnable atom masking strategy to disentangle causal substructures from confounding ones; 3) Distribution intervener applying backdoor adjustment by combining causal substructures with chemically grounded confounders.

Result: Superior accuracy and sample efficiency in few-shot tasks across diverse molecular datasets, with strong generalizability to unseen properties. Discovered causal substructures align well with chemical knowledge about functional groups.

Conclusion: CaMol effectively addresses few-shot molecular property prediction by leveraging causal inference to identify chemically meaningful substructures, improving both performance and interpretability.

Abstract: Molecular property prediction is becoming one of the major applications of graph learning in Web-based services, e.g., online protein structure prediction and drug discovery. A key challenge arises in few-shot scenarios, where only a few labeled molecules are available for predicting unseen properties. Recently, several studies have used in-context learning to capture relationships among molecules and properties, but they face two limitations in: (1) exploiting prior knowledge of functional groups that are causally linked to properties and (2) identifying key substructures directly correlated with properties. We propose CaMol, a context-aware graph causality inference framework, to address these challenges by using a causal inference perspective, assuming that each molecule consists of a latent causal structure that determines a specific property. First, we introduce a context graph that encodes chemical knowledge by linking functional groups, molecules, and properties to guide the discovery of causal substructures. Second, we propose a learnable atom masking strategy to disentangle causal substructures from confounding ones. Third, we introduce a distribution intervener that applies backdoor adjustment by combining causal substructures with chemically grounded confounders, disentangling causal effects from real-world chemical variations. Experiments on diverse molecular datasets showed that CaMol achieved superior accuracy and sample efficiency in few-shot tasks, showing its generalizability to unseen properties. Also, the discovered causal substructures were strongly aligned with chemical knowledge about functional groups, supporting the model interpretability.

Method: Two approaches: 1) Supervised classification pipeline using labeled normal/faulty data, and 2) Unsupervised anomaly detection using autoencoders trained only on healthy engine data to flag deviations.

Result: Supervised models perform well when labeled failure data is available, while autoencoders effectively detect faults without requiring fault labels, making them suitable for scarce failure data scenarios.

Conclusion: Unsupervised learning via autoencoders offers a viable solution for early fault detection in aerospace, highlighting trade-offs between accuracy, data availability, and deployment feasibility.

Abstract: Unplanned engine failures in helicopters can lead to severe operational disruptions, safety hazards, and costly repairs. To mitigate these risks, this study compares two predictive maintenance strategies for helicopter engines: a supervised classification pipeline and an unsupervised anomaly detection approach based on autoencoders (AEs). The supervised method relies on labelled examples of both normal and faulty behaviour, while the unsupervised approach learns a model of normal operation using only healthy engine data, flagging deviations as potential faults. Both methods are evaluated on a real-world dataset comprising labelled snapshots of helicopter engine telemetry. While supervised models demonstrate strong performance when annotated failures are available, the AE achieves effective detection without requiring fault labels, making it particularly well suited for settings where failure data are scarce or incomplete. The comparison highlights the practical trade-offs between accuracy, data availability, and deployment feasibility, and underscores the potential of unsupervised learning as a viable solution for early fault detection in aerospace applications.

Method: Propose two algorithms inspired by the classic Lanczos method: 1) Lanczos Iteration - a near-linear time global algorithm with complexity Õ(√κ m), and 2) Lanczos Push - a local algorithm with complexity Õ(κ^{2.75}) independent of graph size.

Result: Lanczos Iteration achieves √κ speedup over previous power iteration-based global methods, while Lanczos Push shows κ^{0.25} improvement over state-of-the-art random walk-based local algorithms. Both outperform existing methods in extensive experiments on eight real-world datasets.

Conclusion: The proposed Lanczos-based algorithms provide efficient solutions for resistance distance computation on large graphs, addressing the limitations of existing methods and enabling better performance in graph analysis applications.

Abstract: The computation of resistance distance is pivotal in a wide range of graph analysis applications, including graph clustering, link prediction, and graph neural networks. Despite its foundational importance, efficient algorithms for computing resistance distances on large graphs are still lacking. Existing state-of-the-art (SOTA) methods, including power iteration-based algorithms and random walk-based local approaches, often struggle with slow convergence rates, particularly when the condition number of the graph Laplacian matrix, denoted by $κ$, is large. To tackle this challenge, we propose two novel and efficient algorithms inspired by the classic Lanczos method: Lanczos Iteration and Lanczos Push, both designed to reduce dependence on $κ$. Among them, Lanczos Iteration is a near-linear time global algorithm, whereas Lanczos Push is a local algorithm with a time complexity independent of the size of the graph. More specifically, we prove that the time complexity of Lanczos Iteration is $\tilde{O}(\sqrtκ m)$ ($m$ is the number of edges of the graph and $\tilde{O}$ means the complexity omitting the $\log$ terms) which achieves a speedup of $\sqrtκ$ compared to previous power iteration-based global methods. For Lanczos Push, we demonstrate that its time complexity is $\tilde{O}(κ^{2.75})$ under certain mild and frequently established assumptions, which represents a significant improvement of $κ^{0.25}$ over the SOTA random walk-based local algorithms. We validate our algorithms through extensive experiments on eight real-world datasets of varying sizes and statistical properties, demonstrating that Lanczos Iteration and Lanczos Push significantly outperform SOTA methods in terms of both efficiency and accuracy.

Method: The tutorial analyzes existing clustering approaches through the lens of abstraction-representation trade-off: classical K-means (high abstraction, simple representation), subspace clustering (learning separate latent spaces for clustering-relevant vs. other information), and deep clustering methods (using centroid-based and density-based losses to enforce abstraction).

Result: The analysis reveals that increasing representational expressiveness (as in deep learning) requires explicit enforcement of abstraction in objective functions to ensure proper clustering rather than just representation learning. Subspace clustering approaches help by learning separate latent spaces for clustering-relevant and other information.

Conclusion: Future clustering methods need to more adaptively balance abstraction and representation to improve performance, energy efficiency, and interpretability. The human brain’s ability to find this balance naturally suggests there’s significant room for improvement in automated clustering approaches.

Abstract: How to find a natural grouping of a large real data set? Clustering requires a balance between abstraction and representation. To identify clusters, we need to abstract from superfluous details of individual objects. But we also need a rich representation that emphasizes the key features shared by groups of objects that distinguish them from other groups of objects. Each clustering algorithm implements a different trade-off between abstraction and representation. Classical K-means implements a high level of abstraction - details are simply averaged out - combined with a very simple representation - all clusters are Gaussians in the original data space. We will see how approaches to subspace and deep clustering support high-dimensional and complex data by allowing richer representations. However, with increasing representational expressiveness comes the need to explicitly enforce abstraction in the objective function to ensure that the resulting method performs clustering and not just representation learning. We will see how current deep clustering methods define and enforce abstraction through centroid-based and density-based clustering losses. Balancing the conflicting goals of abstraction and representation is challenging. Ideas from subspace clustering help by learning one latent space for the information that is relevant to clustering and another latent space to capture all other information in the data. The tutorial ends with an outlook on future research in clustering. Future methods will more adaptively balance abstraction and representation to improve performance, energy efficiency and interpretability. By automatically finding the sweet spot between abstraction and representation, the human brain is very good at clustering and other related tasks such as single-shot learning. So, there is still much room for improvement.

Method: Combines Gaussian mixture model-based pseudo-labeling with mean teacher framework for stability, adds consistency losses for robustness in continual SF-UniDA setting.

Result: GMM-COMET consistently improves upon source-only model across all evaluated scenarios, establishing first strong baseline for continual SF-UniDA.

Conclusion: The method successfully addresses continual source-free universal domain adaptation, providing a foundation for future research in sequential multi-domain adaptation without source data access.

Abstract: Unsupervised domain adaptation tackles the problem that domain shifts between training and test data impair the performance of neural networks in many real-world applications. Thereby, in realistic scenarios, the source data may no longer be available during adaptation, and the label space of the target domain may differ from the source label space. This setting, known as source-free universal domain adaptation (SF-UniDA), has recently gained attention, but all existing approaches only assume a single domain shift from source to target. In this work, we present the first study on continual SF-UniDA, where the model must adapt sequentially to a stream of multiple different unlabeled target domains. Building upon our previous methods for online SF-UniDA, we combine their key ideas by integrating Gaussian mixture model-based pseudo-labeling within a mean teacher framework for improved stability over long adaptation sequences. Additionally, we introduce consistency losses for further robustness. The resulting method GMM-COMET provides a strong first baseline for continual SF-UniDA and is the only approach in our experiments to consistently improve upon the source-only model across all evaluated scenarios. Our code is available at https://github.com/pascalschlachter/GMM-COMET.

Method: Two unsupervised autoencoder schemes: 1) feed-forward model analyzing individual sensor snapshots, and 2) LSTM model capturing short temporal windows. Both trained exclusively on healthy data from 52 sensor channels at minute-level logging.

Result: Models achieve high reliability in fault detection despite being trained only on healthy data, evaluated on separate dataset containing seven annotated fault intervals.

Conclusion: Unsupervised autoencoder approaches can effectively detect early faults in hydraulic pumps using only healthy operational data, providing a practical solution for industrial applications where fault samples are scarce or unavailable.

Abstract: Unplanned failures in industrial hydraulic pumps can halt production and incur substantial costs. We explore two unsupervised autoencoder (AE) schemes for early fault detection: a feed-forward model that analyses individual sensor snapshots and a Long Short-Term Memory (LSTM) model that captures short temporal windows. Both networks are trained only on healthy data drawn from a minute-level log of 52 sensor channels; evaluation uses a separate set that contains seven annotated fault intervals. Despite the absence of fault samples during training, the models achieve high reliability.

Method: Uses a dual-path VQ-VAE to disentangle trend and seasonal components, encodes sequences into discrete tokens at multiple temporal resolutions, performs autoregressive generation in coarse-to-fine manner, and employs guidance-based reconstruction where coarse seasonal signals guide fine-grained pattern reconstruction.

Result: Outperforms existing methods on six datasets, produces higher-quality time series, achieves strong performance with significantly reduced parameter count, and exhibits superior capability in generating high-quality long-term sequences.

Conclusion: The proposed structure-disentangled multiscale generation framework effectively addresses time series complexity through hierarchical encoding, component disentanglement, and guided reconstruction, offering a promising solution for time series generative modeling.

Abstract: Generative modeling offers a promising solution to data scarcity and privacy challenges in time series analysis. However, the structural complexity of time series, characterized by multi-scale temporal patterns and heterogeneous components, remains insufficiently addressed. In this work, we propose a structure-disentangled multiscale generation framework for time series. Our approach encodes sequences into discrete tokens at multiple temporal resolutions and performs autoregressive generation in a coarse-to-fine manner, thereby preserving hierarchical dependencies. To tackle structural heterogeneity, we introduce a dual-path VQ-VAE that disentangles trend and seasonal components, enabling the learning of semantically consistent latent representations. Additionally, we present a guidance-based reconstruction strategy, where coarse seasonal signals are utilized as priors to guide the reconstruction of fine-grained seasonal patterns. Experiments on six datasets show that our approach produces higher-quality time series than existing methods. Notably, our model achieves strong performance with a significantly reduced parameter count and exhibits superior capability in generating high-quality long-term sequences. Our implementation is available at https://anonymous.4open.science/r/TimeMAR-BC5B.

Method: FAQ inputs original calibration samples into a larger LLM from the same family to regenerate high-fidelity data with Chain-of-Thought reasoning, then uses group competition under expert guidance to select best samples, followed by re-normalization.

Result: Experiments on multiple model series including Qwen3-8B show FAQ reduces accuracy loss by up to 28.5% compared to baseline with original calibration data.

Conclusion: FAQ demonstrates powerful potential for improving PTQ accuracy by leveraging family knowledge to generate representative calibration data, addressing the core bottleneck of calibration data representativeness.

Abstract: Although post-training quantization (PTQ) provides an efficient numerical compression scheme for deploying large language models (LLMs) on resource-constrained devices, the representativeness and universality of calibration data remain a core bottleneck in determining the accuracy of quantization parameters. Traditional PTQ methods typically rely on limited samples, making it difficult to capture the activation distribution during the inference phase, leading to biases in quantization parameters. To address this, we propose \textbf{FAQ} (Family-Aware Quantization), a calibration data regeneration framework that leverages prior knowledge from LLMs of the same family to generate high-fidelity calibration samples. Specifically, FAQ first inputs the original calibration samples into a larger LLM from the same family as the target model, regenerating a series of high-fidelity calibration data using a highly consistent knowledge system. Subsequently, this data, carrying Chain-of-Thought reasoning and conforming to the expected activation distribution, undergoes group competition under expert guidance to select the best samples, which are then re-normalized to enhance the effectiveness of standard PTQ. Experiments on multiple model series, including Qwen3-8B, show that FAQ reduces accuracy loss by up to 28.5% compared to the baseline with original calibration data, demonstrating its powerful potential and contribution.

Method: Proposes Selective Dual-module Federated LoRA (SDFLoRA) which decomposes each client’s LoRA adapter into: 1) a global module that captures transferable knowledge (selectively aligned and aggregated across clients), and 2) a local module that preserves client-specific adaptations (remains private). This design supports privacy-aware optimization by injecting differential privacy noise exclusively into the global module.

Result: Experiments on GLUE benchmarks demonstrate that SDFLoRA outperforms representative federated LoRA baselines and achieves a better utility-privacy trade-off.

Conclusion: SDFLoRA effectively addresses rank heterogeneity in federated LLM learning while enabling better personalization and privacy protection through its dual-module design with selective global aggregation and private local modules.

Abstract: Federated learning (FL) for large language models (LLMs) has attracted increasing attention as a way to enable privacy-preserving adaptation over distributed data. Parameter-efficient methods such as LoRA are widely adopted to reduce communication and memory costs. Despite these advances, practical FL deployments often exhibit rank heterogeneity, since different clients may use different low-rank configurations. This makes direct aggregation of LoRA updates biased and unstable. Existing solutions typically enforce unified ranks or align heterogeneous updates into a shared subspace, which over-constrains client-specific semantics, limits personalization, and provides weak protection of local client information under differential privacy noise. To address this issue, we propose Selective Dual-module Federated LoRA (SDFLoRA), which decomposes each client adapter into a global module that captures transferable knowledge and a local module that preserves client-specific adaptations. The global module is selectively aligned and aggregated across clients, while local modules remain private. This design enables robust learning under rank heterogeneity and supports privacy-aware optimization by injecting differential privacy noise exclusively into the global module. Experiments on GLUE benchmarks demonstrate that SDFLoRA outperforms representative federated LoRA baselines and achieves a better utility-privacy trade-off.

Method: Integrates Mathematical Artificial Data (MAD) method to synthesize boundary data from fundamental solutions, learns boundary-to-boundary mappings via MAD-BNO, and recovers interior solutions through boundary integral formulations.

Result: Achieves comparable or better accuracy than existing neural operators for 2D Laplace, Poisson, and Helmholtz equations while significantly reducing training time; extensible to 3D problems and complex geometries.

Conclusion: MAD-BNO provides an efficient, fully data-driven operator learning framework that uses only boundary data, enabling accurate solution recovery for various PDEs with reduced computational cost.

Abstract: For linear partial differential equations with known fundamental solutions, this work introduces a novel operator learning framework that relies exclusively on domain boundary data, including solution values and normal derivatives, rather than full-domain sampling. By integrating the previously developed Mathematical Artificial Data (MAD) method, which enforces physical consistency, all training data are synthesized directly from the fundamental solutions of the target problems, resulting in a fully data-driven pipeline without the need for external measurements or numerical simulations. We refer to this approach as the Mathematical Artificial Data Boundary Neural Operator (MAD-BNO), which learns boundary-to-boundary mappings using MAD-generated Dirichlet-Neumann data pairs. Once trained, the interior solution at arbitrary locations can be efficiently recovered through boundary integral formulations, supporting Dirichlet, Neumann, and mixed boundary conditions as well as general source terms. The proposed method is validated on benchmark operator learning tasks for two-dimensional Laplace, Poisson, and Helmholtz equations, where it achieves accuracy comparable to or better than existing neural operator approaches while significantly reducing training time. The framework is naturally extensible to three-dimensional problems and complex geometries.

Method: Parametric Skill Transfer (PaST) extracts domain-agnostic Skill Vectors from source domains and linearly injects knowledge manipulation skills into target models after lightweight SFT on new data.

Result: PaST outperforms SFT baselines by up to 9.9 points on SQuAD, achieves 8.0-point gain on LooGLE long-context QA, and improves ToolBench success rates by +10.3 points with consistent cross-domain transferability.

Conclusion: PaST provides an efficient framework for knowledge adaptation that separates factual updates from reasoning skill transfer, enabling scalable and effective knowledge incorporation in LLMs across diverse domains.

Abstract: Large Language Models (LLMs) face the “knowledge cutoff” challenge, where their frozen parametric memory prevents direct internalization of new information. While Supervised Fine-Tuning (SFT) is commonly used to update model knowledge, it often updates factual content without reliably improving the model’s ability to use the newly incorporated information for question answering or decision-making. Reinforcement Learning (RL) is essential for acquiring reasoning skills; however, its high computational cost makes it impractical for efficient online adaptation. We empirically observe that the parameter updates induced by SFT and RL are nearly orthogonal. Based on this observation, we propose Parametric Skill Transfer (PaST), a framework that supports modular skill transfer for efficient and effective knowledge adaptation. By extracting a domain-agnostic Skill Vector from a source domain, we can linearly inject knowledge manipulation skills into a target model after it has undergone lightweight SFT on new data. Experiments on knowledge-incorporation QA (SQuAD, LooGLE) and agentic tool-use benchmarks (ToolBench) demonstrate the effectiveness of our method. On SQuAD, PaST outperforms the state-of-the-art self-editing SFT baseline by up to 9.9 points. PaST further scales to long-context QA on LooGLE with an 8.0-point absolute accuracy gain, and improves zero-shot ToolBench success rates by +10.3 points on average with consistent gains across tool categories, indicating strong scalability and cross-domain transferability of the Skill Vector.

Method: Proposes Latent Dynamics Graph Convolutional Network (LD-GCN), a purely data-driven, encoder-free architecture that learns global low-dimensional representations of dynamical systems conditioned on inputs/parameters. Temporal evolution is modeled in latent space via time-stepping for extrapolation, with trajectories decoded onto geometrically parameterized domains using GNNs.

Result: The framework enables interpretable analysis of reduced dynamics and supports zero-shot prediction through latent interpolation. Mathematically validated via universal approximation theorem and numerically tested on complex computational mechanics problems including Navier-Stokes bifurcation detection.

Conclusion: LD-GCN successfully addresses the gap in combining geometric inductive biases with interpretable latent dynamics for nonlinear MOR of parameterized PDEs, offering enhanced interpretability and prediction capabilities while maintaining mathematical rigor.

Abstract: Graph Neural Networks (GNNs) are emerging as powerful tools for nonlinear Model Order Reduction (MOR) of time-dependent parameterized Partial Differential Equations (PDEs). However, existing methodologies struggle to combine geometric inductive biases with interpretable latent behavior, overlooking dynamics-driven features or disregarding spatial information. In this work, we address this gap by introducing Latent Dynamics Graph Convolutional Network (LD-GCN), a purely data-driven, encoder-free architecture that learns a global, low-dimensional representation of dynamical systems conditioned on external inputs and parameters. The temporal evolution is modeled in the latent space and advanced through time-stepping, allowing for time-extrapolation, and the trajectories are consistently decoded onto geometrically parameterized domains using a GNN. Our framework enhances interpretability by enabling the analysis of the reduced dynamics and supporting zero-shot prediction through latent interpolation. The methodology is mathematically validated via a universal approximation theorem for encoder-free architectures, and numerically tested on complex computational mechanics problems involving physical and geometric parameters, including the detection of bifurcating phenomena for Navier-Stokes equations. Code availability: https://github.com/lorenzotomada/ld-gcn-rom

Method: Proposed a new agnostic boosting algorithm that achieves near-optimal sample complexity while running in polynomial time relative to sample size (with other parameters fixed).

Result: First agnostic boosting algorithm with both near-optimal sample complexity and polynomial-time execution, solving the efficiency problem of previous exponential-time solutions.

Conclusion: This work bridges the gap between theoretical sample complexity bounds and practical implementation by providing an efficient algorithm for agnostic boosting.

Abstract: Boosting is a powerful method that turns weak learners, which perform only slightly better than random guessing, into strong learners with high accuracy. While boosting is well understood in the classic setting, it is less so in the agnostic case, where no assumptions are made about the data. Indeed, only recently was the sample complexity of agnostic boosting nearly settled arXiv:2503.09384, but the known algorithm achieving this bound has exponential running time. In this work, we propose the first agnostic boosting algorithm with near-optimal sample complexity, running in time polynomial in the sample size when considering the other parameters of the problem fixed.

Method: Targeted urinary metabolomics from 98 participants (52 ADHD, 46 controls) analyzed using Closest Resemblance classifier with feature selection.

Result: CR model outperformed RF and KNN with AUC >0.97 using 14 metabolites; identified dopamine 4-sulfate, N-acetylaspartylglutamic acid, citrulline linked to dopaminergic and amino acid pathways.

Conclusion: Interpretable ML + metabolomics provides translational framework for objective ADHD diagnostics with transparent decision boundaries suitable for point-of-care platforms.

Abstract: Attention Deficit Hyperactivity Disorder (ADHD) is a prevalent neurodevelopmental disorder with limited objective diagnostic tools, highlighting the urgent need for objective, biology-based diagnostic frameworks in precision psychiatry. We integrate urinary metabolomics with an interpretable machine learning framework to identify biochemical signatures associated with ADHD. Targeted metabolomic profiles from 52 ADHD and 46 control participants were analyzed using a Closest Resemblance (CR) classifier with embedded feature selection. The CR model outperformed Random Forest and K-Nearest Neighbor classifiers, achieving an AUC > 0.97 based on a reduced panel of 14 metabolites. These metabolites including dopamine 4-sulfate, N-acetylaspartylglutamic acid, and citrulline map to dopaminergic neurotransmission and amino acid metabolism pathways, offering mechanistic insight into ADHD pathophysiology. The CR classifier’s transparent decision boundaries and low computational cost support integration into targeted metabolomic assays and future point of care diagnostic platforms. Overall, this work demonstrates a translational framework combining metabolomics and interpretable machine learning to advance objective, biologically informed diagnostic strategies for ADHD.

Method: Two-fold approach: 1) Semantic splitting criterion where LLM evaluates candidate partitions based on coherence over labeled/unlabeled data; 2) One-time in-context inference for leaf node stabilization, where LLM distills decision paths into deterministic predictions.

Result: FORESTLLM achieves state-of-the-art performance across diverse few-shot classification and regression benchmarks.

Conclusion: The framework successfully unifies structural inductive biases of decision forests with semantic reasoning of LLMs, creating lightweight, interpretable models that eliminate LLM inference at test time while improving few-shot learning.

Abstract: Tabular data high-stakes critical decision-making in domains such as finance, healthcare, and scientific discovery. Yet, learning effectively from tabular data in few-shot settings, where labeled examples are scarce, remains a fundamental challenge. Traditional tree-based methods often falter in these regimes due to their reliance on statistical purity metrics, which become unstable and prone to overfitting with limited supervision. At the same time, direct applications of large language models (LLMs) often overlook its inherent structure, leading to suboptimal performance. To overcome these limitations, we propose FORESTLLM, a novel framework that unifies the structural inductive biases of decision forests with the semantic reasoning capabilities of LLMs. Crucially, FORESTLLM leverages the LLM only during training, treating it as an offline model designer that encodes rich, contextual knowledge into a lightweight, interpretable forest model, eliminating the need for LLM inference at test time. Our method is two-fold. First, we introduce a semantic splitting criterion in which the LLM evaluates candidate partitions based on their coherence over both labeled and unlabeled data, enabling the induction of more robust and generalizable tree structures under few-shot supervision. Second, we propose a one-time in-context inference mechanism for leaf node stabilization, where the LLM distills the decision path and its supporting examples into a concise, deterministic prediction, replacing noisy empirical estimates with semantically informed outputs. Across a diverse suite of few-shot classification and regression benchmarks, FORESTLLM achieves state-of-the-art performance.

Method: The authors propose SPREAD (Semantic-Preserving REtrieval-Augmented Diffusion), a novel framework that introduces a query-relevance-guided denoising strategy. This approach actively guides the denoising trajectory to ensure generation remains anchored to the query’s semantics and effectively suppresses semantic drift throughout the iterative denoising process.

Result: Experimental results demonstrate that SPREAD significantly enhances the precision of generated answers within the RAG framework and effectively mitigates Response Semantic Drift. The framework shows that DLMs coupled with RAG have promising potential with stronger dependency on contextual information.

Conclusion: SPREAD successfully addresses the Response Semantic Drift problem in Diffusion Language Models within RAG frameworks by introducing semantic-preserving denoising strategies, enabling more precise and semantically-aligned generation while leveraging the benefits of retrieval-augmented approaches.

Abstract: Diffusion Language Models (DLMs) have recently demonstrated remarkable capabilities in natural language processing tasks. However, the potential of Retrieval-Augmented Generation (RAG), which shows great successes for enhancing large language models (LLMs), has not been well explored, due to the fundamental difference between LLM and DLM decoding. To fill this critical gap, we systematically test the performance of DLMs within the RAG framework. Our findings reveal that DLMs coupled with RAG show promising potentials with stronger dependency on contextual information, but suffer from limited generation precision. We identify a key underlying issue: Response Semantic Drift (RSD), where the generated answer progressively deviates from the query’s original semantics, leading to low precision content. We trace this problem to the denoising strategies in DLMs, which fail to maintain semantic alignment with the query throughout the iterative denoising process. To address this, we propose Semantic-Preserving REtrieval-Augmented Diffusion (SPREAD), a novel framework that introduces a query-relevance-guided denoising strategy. By actively guiding the denoising trajectory, SPREAD ensures the generation remains anchored to the query’s semantics and effectively suppresses drift. Experimental results demonstrate that SPREAD significantly enhances the precision and effectively mitigates RSD of generated answers within the RAG framework.

Method: FEATHer uses: 1) multiscale frequency decomposition into pathways, 2) shared Dense Temporal Kernel with projection-depthwise convolution-projection (no recurrence/attention), 3) frequency-aware branch gating for adaptive fusion, and 4) Sparse Period Kernel for seasonality capture via period-wise downsampling.

Result: Achieves best ranking across 8 benchmarks with 60 first-place results and average rank of 2.05, outperforming baselines while maintaining compact architecture (as few as 400 parameters).

Conclusion: Reliable long-range forecasting is achievable on constrained edge hardware, offering practical direction for industrial real-time inference with ultra-lightweight models.

Abstract: Time-series forecasting is fundamental in industrial domains like manufacturing and smart factories. As systems evolve toward automation, models must operate on edge devices (e.g., PLCs, microcontrollers) with strict constraints on latency and memory, limiting parameters to a few thousand. Conventional deep architectures are often impractical here. We propose the Fourier-Efficient Adaptive Temporal Hierarchy Forecaster (FEATHer) for accurate long-term forecasting under severe limits. FEATHer introduces: (i) ultra-lightweight multiscale decomposition into frequency pathways; (ii) a shared Dense Temporal Kernel using projection-depthwise convolution-projection without recurrence or attention; (iii) frequency-aware branch gating that adaptively fuses representations based on spectral characteristics; and (iv) a Sparse Period Kernel reconstructing outputs via period-wise downsampling to capture seasonality. FEATHer maintains a compact architecture (as few as 400 parameters) while outperforming baselines. Across eight benchmarks, it achieves the best ranking, recording 60 first-place results with an average rank of 2.05. These results demonstrate that reliable long-range forecasting is achievable on constrained edge hardware, offering a practical direction for industrial real-time inference.

Method: Uses offline reinforcement learning with a dataset of state transitions collected from arbitrary policies before training. Combines online application-agnostic performance data (heartbeats) and hardware performance counters in a gray-box approach to ensure scientific objectives are met with limited performance degradation.

Result: The offline-trained agent substantially reduces energy consumption at tolerable performance degradation cost when evaluated on various compute-bound and memory-bound benchmarks, controlling power through Intel’s Running Average Power Limit on a live system.

Conclusion: Offline RL provides a viable alternative to online RL for designing autonomous CPU power controllers, effectively improving energy efficiency while avoiding the practical challenges of online training on live systems.

Abstract: Energy efficiency has become an integral aspect of modern computing infrastructure design, impacting the performance, cost, scalability, and durability of production systems. The incorporation of power actuation and sensing capabilities in CPU designs is indicative of this, enabling the deployment of system software that can actively monitor and adjust energy consumption and performance at runtime. While reinforcement learning (RL) would seem ideal for the design of such energy efficiency control systems, online training presents challenges ranging from the lack of proper models for setting up an adequate simulated environment, to perturbation (noise) and reliability issues, if training is deployed on a live system. In this paper we discuss the use of offline reinforcement learning as an alternative approach for the design of an autonomous CPU power controller, with the goal of improving the energy efficiency of parallel applications at runtime without unduly impacting their performance. Offline RL sidesteps the issues incurred by online RL training by leveraging a dataset of state transitions collected from arbitrary policies prior to training. Our methodology applies offline RL to a gray-box approach to energy efficiency, combining online application-agnostic performance data (e.g., heartbeats) and hardware performance counters to ensure that the scientific objectives are met with limited performance degradation. Evaluating our method on a variety of compute-bound and memory-bound benchmarks and controlling power on a live system through Intel’s Running Average Power Limit, we demonstrate that such an offline-trained agent can substantially reduce energy consumption at a tolerable performance degradation cost.

Method: Uses two autoencoders (one for parameter space, one for observation space) connected by learned mappings between their latent spaces, enabling surrogate regularized inversion and optimization in low-dimensional latent spaces.

Result: Produces more accurate reconstructions compared to paired autoencoders alone and end-to-end encoder-decoders of same architecture, especially with data inconsistencies. Demonstrated on medical tomography and geophysical seismic-waveform inversion.

Conclusion: The framework is broadly applicable to various inverse problems in scientific and engineering applications, offering flexible handling of data inconsistencies while maintaining physical model consistency.

Abstract: This work describes a novel data-driven latent space inference framework built on paired autoencoders to handle observational inconsistencies when solving inverse problems. Our approach uses two autoencoders, one for the parameter space and one for the observation space, connected by learned mappings between the autoencoders’ latent spaces. These mappings enable a surrogate for regularized inversion and optimization in low-dimensional, informative latent spaces. Our flexible framework can work with partial, noisy, or out-of-distribution data, all while maintaining consistency with the underlying physical models. The paired autoencoders enable reconstruction of corrupted data, and then use the reconstructed data for parameter estimation, which produces more accurate reconstructions compared to paired autoencoders alone and end-to-end encoder-decoders of the same architecture, especially in scenarios with data inconsistencies. We demonstrate our approaches on two imaging examples in medical tomography and geophysical seismic-waveform inversion, but the described approaches are broadly applicable to a variety of inverse problems in scientific and engineering applications.

Method: Uses a multiscale graph-based diffusion architecture trained on CFD simulations. The model interprets classifier-free guidance as a learned posterior reconstruction mechanism: unconditional branch learns geometry-aware flow prior, while sensor-conditioned branch injects observational constraints during sampling.

Result: GenDA reduces relative root-mean-square error (RRMSE) by 25-57% and increases structural similarity index (SSIM) by 23-33% compared to supervised GNN baselines and classical reduced-order data assimilation methods.

Conclusion: The framework provides a scalable path toward generative, geometry-aware data assimilation for environmental monitoring in complex urban domains, enabling obstacle-aware reconstruction and generalization across unseen geometries, wind directions, and mesh resolutions without retraining.

Abstract: Urban wind flow reconstruction is essential for assessing air quality, heat dispersion, and pedestrian comfort, yet remains challenging when only sparse sensor data are available. We propose GenDA, a generative data assimilation framework that reconstructs high-resolution wind fields on unstructured meshes from limited observations. The model employs a multiscale graph-based diffusion architecture trained on computational fluid dynamics (CFD) simulations and interprets classifier-free guidance as a learned posterior reconstruction mechanism: the unconditional branch learns a geometry-aware flow prior, while the sensor-conditioned branch injects observational constraints during sampling. This formulation enables obstacle-aware reconstruction and generalization across unseen geometries, wind directions, and mesh resolutions without retraining. We consider both sparse fixed sensors and trajectory-based observations using the same reconstruction procedure. When evaluated against supervised graph neural network (GNN) baselines and classical reduced-order data assimilation methods, GenDA reduces the relative root-mean-square error (RRMSE) by 25-57% and increases the structural similarity index (SSIM) by 23-33% across the tested meshes. Experiments are conducted on Reynolds-averaged Navier-Stokes (RANS) simulations of a real urban neighbourhood in Bristol, United Kingdom, at a characteristic Reynolds number of $\mathrm{Re}\approx2\times10^{7}$, featuring complex building geometry and irregular terrain. The proposed framework provides a scalable path toward generative, geometry-aware data assimilation for environmental monitoring in complex domains.

Method: Introduces Diffusion Value Function (DVF) that assigns value components by diffusing rewards over influence graphs with temporal discounting and spatial attenuation. Proposes DA2C algorithm and LD-GNN for decentralized learning under communication costs.

Result: DVF is well-defined, admits Bellman fixed point, and decomposes global discounted value. DA2C consistently outperforms local and global critic baselines across firefighting and distributed computation tasks, improving average reward by up to 11%.

Conclusion: DVF provides a principled, scalable approach to credit assignment in graph-based MARL, enabling effective decentralized learning in systems with structured local interactions.

Abstract: Credit assignment is a core challenge in multi-agent reinforcement learning (MARL), especially in large-scale systems with structured, local interactions. Graph-based Markov decision processes (GMDPs) capture such settings via an influence graph, but standard critics are poorly aligned with this structure: global value functions provide weak per-agent learning signals, while existing local constructions can be difficult to estimate and ill-behaved in infinite-horizon settings. We introduce the Diffusion Value Function (DVF), a factored value function for GMDPs that assigns to each agent a value component by diffusing rewards over the influence graph with temporal discounting and spatial attenuation. We show that DVF is well-defined, admits a Bellman fixed point, and decomposes the global discounted value via an averaging property. DVF can be used as a drop-in critic in standard RL algorithms and estimated scalably with graph neural networks. Building on DVF, we propose Diffusion A2C (DA2C) and a sparse message-passing actor, Learned DropEdge GNN (LD-GNN), for learning decentralised algorithms under communication costs. Across the firefighting benchmark and three distributed computation tasks (vector graph colouring and two transmit power optimisation problems), DA2C consistently outperforms local and global critic baselines, improving average reward by up to 11%.

Method: Proposed new probe architectures to handle long-context distribution shift, evaluated in cyber-offensive domain with various production-relevant shifts (multi-turn conversations, static jailbreaks, adaptive red teaming). Combined architecture choice with diverse training distributions, and paired probes with prompted classifiers.

Result: Multimax addresses context length but requires combination of architecture choice and diverse training for broad generalization. Pairing probes with prompted classifiers achieves optimal accuracy at low computational cost. Findings enabled successful deployment in Gemini. AlphaEvolve shows early positive results for automating probe architecture search and adaptive red teaming.

Conclusion: Effective misuse mitigation requires probes that generalize across distribution shifts, particularly to long contexts. Combining appropriate architectures, diverse training, and computational efficiency enables practical deployment. Automation of AI safety research through tools like AlphaEvolve is already feasible.

Abstract: Frontier language model capabilities are improving rapidly. We thus need stronger mitigations against bad actors misusing increasingly powerful systems. Prior work has shown that activation probes may be a promising misuse mitigation technique, but we identify a key remaining challenge: probes fail to generalize under important production distribution shifts. In particular, we find that the shift from short-context to long-context inputs is difficult for existing probe architectures. We propose several new probe architecture that handle this long-context distribution shift. We evaluate these probes in the cyber-offensive domain, testing their robustness against various production-relevant shifts, including multi-turn conversations, static jailbreaks, and adaptive red teaming. Our results demonstrate that while multimax addresses context length, a combination of architecture choice and training on diverse distributions is required for broad generalization. Additionally, we show that pairing probes with prompted classifiers achieves optimal accuracy at a low cost due to the computational efficiency of probes. These findings have informed the successful deployment of misuse mitigation probes in user-facing instances of Gemini, Google’s frontier language model. Finally, we find early positive results using AlphaEvolve to automate improvements in both probe architecture search and adaptive red teaming, showing that automating some AI safety research is already possible.

Method: Consolidated multiple publicly available T1D datasets into a unified resource requiring both CGM data and insulin pump dosing records, with auxiliary information retained when available. Created processing pipelines for standardized format conversion.

Result: MetaboNet comprises 3135 subjects and 1228 patient-years of overlapping CGM and insulin data, substantially larger than existing standalone datasets. Available as public subset for immediate download and DUA-restricted subset via application.

Conclusion: The consolidated dataset covers broad glycemic profiles and demographics, enabling more generalizable algorithmic performance than individual datasets, with clear access pathways for both unrestricted and DUA-governed components.

Abstract: Progress in Type 1 Diabetes (T1D) algorithm development is limited by the fragmentation and lack of standardization across existing T1D management datasets. Current datasets differ substantially in structure and are time-consuming to access and process, which impedes data integration and reduces the comparability and generalizability of algorithmic developments. This work aims to establish a unified and accessible data resource for T1D algorithm development. Multiple publicly available T1D datasets were consolidated into a unified resource, termed the MetaboNet dataset. Inclusion required the availability of both continuous glucose monitoring (CGM) data and corresponding insulin pump dosing records. Additionally, auxiliary information such as reported carbohydrate intake and physical activity was retained when present. The MetaboNet dataset comprises 3135 subjects and 1228 patient-years of overlapping CGM and insulin data, making it substantially larger than existing standalone benchmark datasets. The resource is distributed as a fully public subset available for immediate download at https://metabo-net.org/ , and with a Data Use Agreement (DUA)-restricted subset accessible through their respective application processes. For the datasets in the latter subset, processing pipelines are provided to automatically convert the data into the standardized MetaboNet format. A consolidated public dataset for T1D research is presented, and the access pathways for both its unrestricted and DUA-governed components are described. The resulting dataset covers a broad range of glycemic profiles and demographics and thus can yield more generalizable algorithmic performance than individual datasets.

Method: Developed stress-testing framework probing FNOs across 5 PDE families (dispersive, elliptic, multi-scale fluid, financial, chaotic) with controlled tests: parameter shifts, boundary condition changes, resolution extrapolation with spectral analysis, and iterative rollouts.

Result: Distribution shifts in parameters or boundary conditions inflate errors by more than 10x; resolution changes concentrate error in high-frequency modes; input perturbations generally don’t amplify error except worst-case scenarios like localized Poisson perturbations.

Conclusion: The study provides a comparative failure-mode atlas and actionable insights for improving robustness in operator learning, revealing specific vulnerabilities that need addressing for reliable PDE solution learning.

Abstract: Fourier Neural Operators (FNOs) have shown strong performance in learning solution maps of partial differential equations (PDEs), but their robustness under distribution shifts, long-horizon rollouts, and structural perturbations remains poorly understood. We present a systematic stress-testing framework that probes failure modes of FNOs across five qualitatively different PDE families: dispersive, elliptic, multi-scale fluid, financial, and chaotic systems. Rather than optimizing in-distribution accuracy, we design controlled stress tests–including parameter shifts, boundary or terminal condition changes, resolution extrapolation with spectral analysis, and iterative rollouts–to expose vulnerabilities such as spectral bias, compounding integration errors, and overfitting to restricted boundary regimes. Our large-scale evaluation (1{,}000 trained models) reveals that distribution shifts in parameters or boundary conditions can inflate errors by more than an order of magnitude, while resolution changes primarily concentrate error in high-frequency modes. Input perturbations generally do not amplify error, though worst-case scenarios (e.g., localized Poisson perturbations) remain challenging. These findings provide a comparative failure-mode atlas and actionable insights for improving robustness in operator learning.

Method: Proposes two novel architectures: Deep Differentiable Logic Gate Networks (LGNs) and Lookup Table Networks (LUTNs). Uses MIT-BIH arrhythmia dataset with inter-patient paradigm. Introduces novel preprocessing method, rate coding for LGNs/LUTNs, and a training method for LUTs using Boolean multiplexer equations. Benchmarks on FPGA with power/performance measurements.

Result: Achieves 94.28% accuracy and jκ index of 0.683 on four-class classification. Models use only 2.89k-6.17k FLOPs (3-6 orders magnitude less than SOTA). FPGA implementation requires 2000-2990 LUTs and consumes 5-7 mW (50-70 pJ per inference). Performance significantly exceeds previous LGN results.

Conclusion: LGNs and LUTNs are highly effective for ECG arrhythmia classification with exceptional energy efficiency and computational efficiency, making them suitable for deployment in ultra-low-power medical devices like heart implants and wearables, even for patients not included in training.

Abstract: Deep Differentiable Logic Gate Networks (LGNs) and Lookup Table Networks (LUTNs) are demonstrated to be suitable for the automatic classification of electrocardiograms (ECGs) using the inter-patient paradigm. The methods are benchmarked using the MIT-BIH arrhythmia data set, achieving up to 94.28% accuracy and a $jκ$ index of 0.683 on a four-class classification problem. Our models use between 2.89k and 6.17k FLOPs, including preprocessing and readout, which is three to six orders of magnitude less compared to SOTA methods. A novel preprocessing method is utilized that attains superior performance compared to existing methods for both the mixed-patient and inter-patient paradigms. In addition, a novel method for training the Lookup Tables (LUTs) in LUTNs is devised that uses the Boolean equation of a multiplexer (MUX). Additionally, rate coding was utilized for the first time in these LGNs and LUTNs, enhancing the performance of LGNs. Furthermore, it is the first time that LGNs and LUTNs have been benchmarked on the MIT-BIH arrhythmia dataset using the inter-patient paradigm. Using an Artix 7 FPGA, between 2000 and 2990 LUTs were needed, and between 5 to 7 mW (i.e. 50 pJ to 70 pJ per inference) was estimated for running these models. The performance in terms of both accuracy and $jκ$-index is significantly higher compared to previous LGN results. These positive results suggest that one can utilize LGNs and LUTNs for the detection of arrhythmias at extremely low power and high speeds in heart implants or wearable devices, even for patients not included in the training set.

Method: Evaluated various ensemble aggregation rules including Deep Ensembles and Monte Carlo Dropout across image datasets (CIFAR-10, FFHQ) and tabular data using random forests. Investigated the link between score estimation and image quality, and provided theoretical analysis of score model composition.

Result: Ensembling scores improves score-matching loss and model likelihood but fails to consistently enhance perceptual quality metrics like FID on image datasets. For tabular data, one aggregation strategy outperforms others.

Conclusion: While ensembling diffusion models improves likelihood-based metrics, it doesn’t reliably improve perceptual quality, highlighting a discrepancy between score estimation quality and generative performance. Theoretical insights explain score model composition techniques including guidance.

Abstract: Diffusion models now generate high-quality, diverse samples, with an increasing focus on more powerful models. Although ensembling is a well-known way to improve supervised models, its application to unconditional score-based diffusion models remains largely unexplored. In this work we investigate whether it provides tangible benefits for generative modelling. We find that while ensembling the scores generally improves the score-matching loss and model likelihood, it fails to consistently enhance perceptual quality metrics such as FID on image datasets. We confirm this observation across a breadth of aggregation rules using Deep Ensembles, Monte Carlo Dropout, on CIFAR-10 and FFHQ. We attempt to explain this discrepancy by investigating possible explanations, such as the link between score estimation and image quality. We also look into tabular data through random forests, and find that one aggregation strategy outperforms the others. Finally, we provide theoretical insights into the summing of score models, which shed light not only on ensembling but also on several model composition techniques (e.g. guidance).

Method: Low-rank KV adaptation (LRKV) modifies multi-head attention by using shared full-rank KV projections across all heads, augmented with low-rank, head-specific residuals. This creates a continuous trade-off between complete sharing and fully independent attention while preserving token-level resolution.

Result: LRKV consistently outperforms standard attention, MQA/GQA, and MLA across large-scale pretraining: faster loss reduction, lower validation perplexity, stronger downstream task performance. At 2.5B scale, achieves better performance with half the KV cache and equivalent quality with 20-25% less training compute.

Conclusion: LRKV is a practical and effective attention mechanism for scaling Transformer pretraining under memory- and compute-constrained regimes, preserving functional head diversity better than aggressive KV-sharing mechanisms while significantly reducing computational requirements.

Abstract: Transformer pretraining is increasingly constrained by memory and compute requirements, with the key-value (KV) cache emerging as a dominant bottleneck during training and autoregressive decoding. We propose \textit{low-rank KV adaptation} (LRKV), a simple modification of multi-head attention that reduces KV cache memory by exploiting redundancy across attention heads while preserving full token-level resolution. Each layer uses a shared full-rank KV projection augmented with low-rank, head-specific residuals, yielding a continuous trade-off between complete sharing and fully independent attention. LRKV is a drop-in replacement for standard multi-head attention and directly subsumes query-sharing approaches such as multi-query and grouped-query attention, while remaining distinct from latent-compression methods such as multi-latent attention (MLA). Across large-scale pretraining experiments, LRKV consistently achieves faster loss reduction, lower validation perplexity, and stronger downstream task performance than standard attention, MQA/GQA, and MLA. At the 2.5B scale, LRKV outperforms standard attention while using roughly half the KV cache, and reaches equivalent model quality with up to \textbf{20-25% less training compute} when measured in cumulative FLOPs. To explain these gains, we analyze attention head structure in operator space and show that LRKV preserves nearly all functional head diversity relative to standard attention, whereas more aggressive KV-sharing mechanisms rely on compensatory query specialization. Together, these results establish LRKV as a practical and effective attention mechanism for scaling Transformer pretraining under memory- and compute-constrained regimes.

Method: The authors develop: (1) a hardware-aware Ising formulation that reduces scale imbalance between local fields and coupling terms to improve robustness to coefficient quantization; (2) a complete ES pipeline with stochastic rounding and iterative refinement to compensate for precision loss; (3) a decomposition strategy that partitions large ES problems into smaller Ising subproblems solvable on COBI hardware.

Result: On CNN/DailyMail dataset, the COBI-based pipeline achieves 3-4.5x runtime speedups compared to brute-force methods (comparable to software Tabu search), with two to three orders of magnitude energy reduction while maintaining competitive summary quality.

Conclusion: CMOS Ising solvers like COBI show strong potential for deploying real-time, low-energy text summarization on edge devices, offering significant energy savings while maintaining competitive performance.

Abstract: Extractive summarization (ES) aims to generate a concise summary by selecting a subset of sentences from a document while maximizing relevance and minimizing redundancy. Although modern ES systems achieve high accuracy using powerful neural models, their deployment typically relies on CPU or GPU infrastructures that are energy-intensive and poorly suited for real-time inference in resource-constrained environments. In this work, we explore the feasibility of implementing McDonald-style extractive summarization on a low-power CMOS coupled oscillator-based Ising machine (COBI) that supports integer-valued, all-to-all spin couplings. We first propose a hardware-aware Ising formulation that reduces the scale imbalance between local fields and coupling terms, thereby improving robustness to coefficient quantization: this method can be applied to any problem formulation that requires k of n variables to be chosen. We then develop a complete ES pipeline including (i) stochastic rounding and iterative refinement to compensate for precision loss, and (ii) a decomposition strategy that partitions a large ES problem into smaller Ising subproblems that can be efficiently solved on COBI and later combined. Experimental results on the CNN/DailyMail dataset show that our pipeline can produce high-quality summaries using only integer-coupled Ising hardware with limited precision. COBI achieves 3-4.5x runtime speedups compared to a brute-force method, which is comparable to software Tabu search, and two to three orders of magnitude reductions in energy, while maintaining competitive summary quality. These results highlight the potential of deploying CMOS Ising solvers for real-time, low-energy text summarization on edge devices.

Method: Proposed QUPID, a partitioned quantum neural network (PQNN) that uses quantum-enhanced feature representations. Extended to R-QUPID with differential privacy for enhanced robustness. The partitioning framework addresses scalability by distributing computational workloads efficiently.

Result: QUPID and R-QUPID significantly outperform traditional state-of-the-art ML models in anomaly detection across various scenarios. R-QUPID maintains performance even with differential privacy, demonstrating both improved detection capabilities and enhanced robustness.

Conclusion: Quantum ML with partitioned architectures like QUPID provides practical, scalable solutions for smart grid anomaly detection, offering superior performance and robustness compared to traditional ML approaches while addressing scalability challenges in quantum computing.

Abstract: Smart grid infrastructures have revolutionized energy distribution, but their day-to-day operations require robust anomaly detection methods to counter risks associated with cyber-physical threats and system faults potentially caused by natural disasters, equipment malfunctions, and cyber attacks. Conventional machine learning (ML) models are effective in several domains, yet they struggle to represent the complexities observed in smart grid systems. Furthermore, traditional ML models are highly susceptible to adversarial manipulations, making them increasingly unreliable for real-world deployment. Quantum ML (QML) provides a unique advantage, utilizing quantum-enhanced feature representations to model the intricacies of the high-dimensional nature of smart grid systems while demonstrating greater resilience to adversarial manipulation. In this work, we propose QUPID, a partitioned quantum neural network (PQNN) that outperforms traditional state-of-the-art ML models in anomaly detection. We extend our model to R-QUPID that even maintains its performance when including differential privacy (DP) for enhanced robustness. Moreover, our partitioning framework addresses a significant scalability problem in QML by efficiently distributing computational workloads, making quantum-enhanced anomaly detection practical in large-scale smart grid environments. Our experimental results across various scenarios exemplifies the efficacy of QUPID and R-QUPID to significantly improve anomaly detection capabilities and robustness compared to traditional ML approaches.

Method: Developed MSCR metric using test data augmentation with robustness distance ε calculated from dataset’s minimal class separation distance. This allows dataset-specific corruption robustness evaluation and comparison.

Result: MSCR effectively reflects different robustness levels in 2D and image classifiers. Unexpected optima found in robust accuracy with varying noise levels. Data augmentation for robustness training can slightly improve accuracy, challenging the inherent accuracy-robustness tradeoff assumption.

Conclusion: MSCR provides interpretable, comparable corruption robustness assessment. The accuracy-robustness tradeoff is not inherent, as simple data augmentation can improve both accuracy and robustness simultaneously.

Abstract: Robustness is a fundamental pillar of Machine Learning (ML) classifiers, substantially determining their reliability. Methods for assessing classifier robustness are therefore essential. In this work, we address the challenge of evaluating corruption robustness in a way that allows comparability and interpretability on a given dataset. We propose a test data augmentation method that uses a robustness distance $ε$ derived from the datasets minimal class separation distance. The resulting MSCR (minimal separation corruption robustness) metric allows a dataset-specific comparison of different classifiers with respect to their corruption robustness. The MSCR value is interpretable, as it represents the classifiers avoidable loss of accuracy due to statistical corruptions. On 2D and image data, we show that the metric reflects different levels of classifier robustness. Furthermore, we observe unexpected optima in classifiers robust accuracy through training and testing classifiers with different levels of noise. While researchers have frequently reported on a significant tradeoff on accuracy when training robust models, we strengthen the view that a tradeoff between accuracy and corruption robustness is not inherent. Our results indicate that robustness training through simple data augmentation can already slightly improve accuracy.

Method: SUNG uses VAE-based state-action visitation density estimator to quantify uncertainty. It implements optimistic exploration by selecting actions with high value and high uncertainty, and adaptive exploitation by applying conservative offline RL objectives to high-uncertainty samples and standard online RL objectives to low-uncertainty samples.

Result: SUNG achieves state-of-the-art online finetuning performance when combined with different offline RL methods across various D4RL benchmark environments and datasets.

Conclusion: SUNG provides a simple unified framework that effectively addresses offline-to-online RL challenges using uncertainty guidance, enabling smooth transition between offline and online learning stages with superior performance.

Abstract: Offline reinforcement learning (RL) provides a promising solution to learning an agent fully relying on a data-driven paradigm. However, constrained by the limited quality of the offline dataset, its performance is often sub-optimal. Therefore, it is desired to further finetune the agent via extra online interactions before deployment. Unfortunately, offline-to-online RL can be challenging due to two main challenges: constrained exploratory behavior and state-action distribution shift. In view of this, we propose a Simple Unified uNcertainty-Guided (SUNG) framework, which naturally unifies the solution to both challenges with the tool of uncertainty. Specifically, SUNG quantifies uncertainty via a VAE-based state-action visitation density estimator. To facilitate efficient exploration, SUNG presents a practical optimistic exploration strategy to select informative actions with both high value and high uncertainty. Moreover, SUNG develops an adaptive exploitation method by applying conservative offline RL objectives to high-uncertainty samples and standard online RL objectives to low-uncertainty samples to smoothly bridge offline and online stages. SUNG achieves state-of-the-art online finetuning performance when combined with different offline RL methods, across various environments and datasets in D4RL benchmark. Codes are made publicly available in https://github.com/guosyjlu/SUNG.

Method: Decouple the acting policy from the policy evaluated by the critic. This allows using greedier updates (like value improvement) for the critic’s policy while maintaining slow gradient-based updates for the parameterized acting policy. The approach is analyzed using generalized Policy Iteration in finite-horizon domains and implemented in off-policy actor-critic algorithms TD3 and SAC.

Result: Empirical results show that incorporating value-improvement into TD3 and SAC significantly improves or matches performance across different DeepMind continuous control environments, with negligible compute and implementation overhead.

Conclusion: Decoupling acting and critic policies allows better balance between greedification and stability, enabling more aggressive policy improvement while maintaining learning stability, resulting in improved performance in continuous control tasks.

Abstract: To learn approximately optimal acting policies for decision problems, modern Actor Critic algorithms rely on deep Neural Networks (DNNs) to parameterize the acting policy and greedification operators to iteratively improve it. The reliance on DNNs suggests an improvement that is gradient based, which is per step much less greedy than the improvement possible by greedier operators such as the greedy update used by Q-learning algorithms. On the other hand, slow changes to the policy can also be beneficial for the stability of the learning process, resulting in a tradeoff between greedification and stability. To better address this tradeoff, we propose to decouple the acting policy from the policy evaluated by the critic. This allows the agent to separately improve the critic’s policy (e.g. value improvement) with greedier updates while maintaining the slow gradient-based improvement to the parameterized acting policy. We investigate the convergence of this approach using the popular analysis scheme of generalized Policy Iteration in the finite-horizon domain. Empirically, incorporating value-improvement into the popular off-policy actor-critic algorithms TD3 and SAC significantly improves or matches performance over their respective baselines, across different environments from the DeepMind continuous control domain, with negligible compute and implementation cost.

Method: REGAD uses reinforcement learning with two novel components: (1) A tailored policy network with two-step actions to remove negative effect propagation step by step, and (2) A policy-in-the-loop mechanism to identify suitable edge removal strategies that control noise propagation and estimate updated structure to obtain reliable pseudo labels iteratively.

Result: Experiments on three real-world datasets demonstrate that REGAD outperforms all baselines under different noisy ratios.

Conclusion: The proposed REGAD framework effectively addresses the challenge of noisy labels in graph anomaly detection by strategically pruning edges around potentially mislabeled nodes and using reinforcement learning to guide the edge removal process, leading to superior performance compared to existing methods.

Abstract: Graph anomaly detection (GAD) is widely applied in many areas, such as financial fraud detection and social spammer detection. Anomalous nodes in the graph not only impact their own communities but also create a ripple effect on neighbors throughout the graph structure. Detecting anomalous nodes in complex graphs has been a challenging task. While existing GAD methods assume all labels are correct, real-world scenarios often involve inaccurate annotations. These noisy labels can severely degrade GAD performance because, with anomalies representing a minority class, even a small number of mislabeled instances can disproportionately interfere with detection models. Cutting edges to mitigate the negative effects of noisy labels is a good option; however, it has both positive and negative influences and also presents an issue of weak supervision. To perform effective GAD with noisy labels, we propose REinforced Graph Anomaly Detector (REGAD) by pruning the edges of candidate nodes potentially with mistaken labels. Moreover, we design the performance feedback based on strategically crafted confident labels to guide the cutting process, ensuring optimal results. Specifically, REGAD contains two novel components. (i) A tailored policy network, which involves two-step actions to remove negative effect propagation step by step. (ii) A policy-in-the-loop mechanism to identify suitable edge removal strategies that control the propagation of noise on the graph and estimate the updated structure to obtain reliable pseudo labels iteratively. Experiments on three real-world datasets demonstrate that REGAD outperforms all baselines under different noisy ratios.

Method: Proposes a novel framework that jointly optimizes both graph structure and model for fair unlearning. The method rewires the graph by removing redundant edges that hinder forgetting while preserving fairness through targeted edge augmentation. Also introduces a worst-case evaluation mechanism to assess robustness under challenging scenarios.

Result: Experiments on real-world datasets demonstrate that the approach achieves more effective and fair unlearning compared to existing baselines.

Conclusion: The proposed framework successfully addresses the fairness gap in graph unlearning by jointly optimizing graph structure and model, providing a more balanced approach that maintains fairness while achieving effective forgetting of sensitive information.

Abstract: With growing emphasis on privacy regulations, machine unlearning has become increasingly critical in real-world applications such as social networks and recommender systems, many of which are naturally represented as graphs. However, existing graph unlearning methods often modify nodes or edges indiscriminately, overlooking their impact on fairness. For instance, forgetting links between users of different genders may inadvertently exacerbate group disparities. To address this issue, we propose a novel framework that jointly optimizes both the graph structure and the model to achieve fair unlearning. Our method rewires the graph by removing redundant edges that hinder forgetting while preserving fairness through targeted edge augmentation. We further introduce a worst-case evaluation mechanism to assess robustness under challenging scenarios. Experiments on real-world datasets show that our approach achieves more effective and fair unlearning than existing baselines.

Method: Enhanced Chebyshev1KANs with wavelet-activated MLPs with learnable parameters and internal attention to preserve full-rank Jacobian. Added Residual Gradient Attention (RGA) mechanism to dynamically re-weight loss terms based on gradient norms and residual magnitudes.

Result: AC-PKAN consistently outperforms or matches state-of-the-art models like PINNsFormer across nine benchmark tasks in three domains, establishing it as effective for complex real-world engineering problems in zero-data or data-sparse regimes.

Conclusion: AC-PKAN successfully overcomes limitations of Chebyshev1KANs through internal and external attention mechanisms, enhancing weakly supervised PINNs and extending KANs’ expressive power for PDE solving.

Abstract: Kolmogorov-Arnold Networks (KANs) have recently shown promise for solving partial differential equations (PDEs). Yet their original formulation is computationally and memory intensive, motivating the introduction of Chebyshev Type-I-based KANs (Chebyshev1KANs). Although Chebyshev1KANs have outperformed the vanilla KANs architecture, our rigorous theoretical analysis reveals that they still suffer from rank collapse, ultimately limiting their expressive capacity. To overcome these limitations, we enhance Chebyshev1KANs by integrating wavelet-activated MLPs with learnable parameters and an internal attention mechanism. We prove that this design preserves a full-rank Jacobian and is capable of approximating solutions to PDEs of arbitrary order. Furthermore, to alleviate the loss instability and imbalance introduced by the Chebyshev polynomial basis, we externally incorporate a Residual Gradient Attention (RGA) mechanism that dynamically re-weights individual loss terms according to their gradient norms and residual magnitudes. By jointly leveraging internal and external attention, we present AC-PKAN, a novel architecture that constitutes an enhancement to weakly supervised Physics-Informed Neural Networks (PINNs) and extends the expressive power of KANs. Experimental results from nine benchmark tasks across three domains show that AC-PKAN consistently outperforms or matches state-of-the-art models such as PINNsFormer, establishing it as a highly effective tool for solving complex real-world engineering problems in zero-data or data-sparse regimes. The code will be made publicly available upon acceptance.

Method: DREAM-CL uses dynamic prototype rehearsal memory that clusters data based on learning behavior, applies smooth sorting to rank samples by training difficulty (compressing extremes and removing outliers), and selects more challenging samples as prototypes for rehearsal memory.

Result: DREAM-CL outperforms state-of-the-art continual learning methods on time-incremental, class-incremental, and lead-incremental scenarios using Chapman and PTB-XL ECG datasets.

Conclusion: The dynamic prototype selection approach with smooth sorting effectively retains knowledge across continual learning sessions for ECG arrhythmia detection, validated through ablation and sensitivity studies.

Abstract: Continual Learning (CL) methods aim to learn from a sequence of tasks while avoiding the challenge of forgetting previous knowledge. We present DREAM-CL, a novel CL method for ECG arrhythmia detection that introduces dynamic prototype rehearsal memory. DREAM-CL selects representative prototypes by clustering data based on learning behavior during each training session. Within each cluster, we apply a smooth sorting operation that ranks samples by training difficulty, compressing extreme values and removing outliers. The more challenging samples are then chosen as prototypes for the rehearsal memory, ensuring effective knowledge retention across sessions. We evaluate our method on time-incremental, class-incremental, and lead-incremental scenarios using two widely used ECG arrhythmia datasets, Chapman and PTB-XL. The results demonstrate that DREAM-CL outperforms the state-of-the-art in CL for ECG arrhythmia detection. Detailed ablation and sensitivity studies are performed to validate the different design choices of our method.

Method: Proposes MoLAN framework that performs modality-aware blocking (dividing each modality’s features into multiple blocks), then dynamically assigns distinct denoising strengths based on each block’s noise level and semantic relevance. Also introduces MoLAN+ as a specific multimodal sentiment analysis approach built on this framework.

Result: Experiments across five models and four datasets demonstrate broad effectiveness of MoLAN framework. MoLAN+ achieves state-of-the-art performance in multimodal sentiment analysis.

Conclusion: MoLAN provides a unified, flexible framework for fine-grained noise suppression in multimodal learning that preserves essential information while being easily integrated into various multimodal models.

Abstract: Multimodal Sentiment Analysis aims to integrate information from various modalities, such as audio, visual, and text, to make complementary predictions. However, it often struggles with irrelevant or misleading visual and auditory information. Most existing approaches typically treat the entire modality information (e.g., a whole image, audio segment, or text paragraph) as an independent unit for feature enhancement or denoising. They often suppress the redundant and noise information at the risk of losing critical information. To address this challenge, we propose MoLAN, a unified ModaLity-aware noise dynAmic editiNg framework. Specifically, MoLAN performs modality-aware blocking by dividing the features of each modality into multiple blocks. Each block is then dynamically assigned a distinct denoising strength based on its noise level and semantic relevance, enabling fine-grained noise suppression while preserving essential multimodal information. Notably, MoLAN is a unified and flexible framework that can be seamlessly integrated into a wide range of multimodal models. Building upon this framework, we further introduce MoLAN+, a new multimodal sentiment analysis approach. Experiments across five models and four datasets demonstrate the broad effectiveness of the MoLAN framework. Extensive evaluations show that MoLAN+ achieves the state-of-the-art performance. The code is publicly available at https://github.com/betterfly123/MoLAN-Framework.

Method: Proposes RCCDA: a dynamic model update policy that analytically characterizes model loss evolution under concept drift, integrates into Lyapunov drift-plus-penalty framework to produce lightweight greedy-optimal policy with provable update frequency/cost limits.

Result: Experimental results on four domain generalization datasets show RCCDA outperforms baseline methods in inference accuracy while adhering to strict resource constraints under various concept drift schedules.

Conclusion: RCCDA provides a unique solution for real-time ML deployments by offering theoretical guarantees on resource usage and performance while maintaining lightweight computational requirements.

Abstract: Machine learning (ML) algorithms deployed in real-world environments are often faced with the challenge of adapting models to concept drift, where the task data distributions are shifting over time. The problem becomes even more difficult when model performance must be maintained under adherence to strict resource constraints. Existing solutions often depend on drift-detection methods that produce high computational overhead for resource-constrained environments, and fail to provide strict guarantees on resource usage or theoretical performance assurances. To address these shortcomings, we propose RCCDA: a dynamic model update policy that optimizes ML training dynamics while ensuring compliance to predefined resource constraints, utilizing only past loss information and a tunable drift threshold. In developing our policy, we analytically characterize the evolution of model loss under concept drift with arbitrary training update decisions. Integrating these results into a Lyapunov drift-plus-penalty framework produces a lightweight greedy-optimal policy that provably limits update frequency and cost. Experimental results on four domain generalization datasets demonstrate that our policy outperforms baseline methods in inference accuracy while adhering to strict resource constraints under several schedules of concept drift, making our solution uniquely suited for real-time ML deployments.

Method: Model demand using Weibull distribution with Gamma prior for Thompson Sampling, then extend to general parametric distribution families. Use TS to dynamically adjust order quantities while analyzing frequentist and Bayesian regret bounds.

Result: Established optimal (up to logarithmic factors) frequentist regret bounds for TS without restrictive prior assumptions. TS automatically increases order quantities when past orders are small to gather more demand information, and accurately estimates parameters when orders are sufficiently large. TS outperforms online convex optimization, upper confidence bounds, and myopic Bayesian dynamic programming in simulations.

Conclusion: Thompson Sampling provides an effective solution for online learning with censored feedback, offering interpretable exploration-exploitation trade-offs and superior performance compared to existing approaches in inventory management problems.

Abstract: In this paper, we investigate the performance of Thompson Sampling (TS) for online learning with censored feedback, focusing primarily on the classic repeated newsvendor model–a foundational framework in inventory management–and demonstrating how our techniques can be naturally extended to a broader class of problems. We first model demand using a Weibull distribution and initialize TS with a Gamma prior to dynamically adjust order quantities. Our analysis establishes optimal (up to logarithmic factors) frequentist regret bounds for TS without imposing restrictive prior assumptions. More importantly, it yields novel and highly interpretable insights on how TS addresses the exploration-exploitation trade-off in the repeated newsvendor setting. Specifically, our results show that when past order quantities are sufficiently large to overcome censoring, TS accurately estimates the unknown demand parameters, leading to near-optimal ordering decisions. Conversely, when past orders are relatively small, TS automatically increases future order quantities to gather additional demand information. Then, we extend our analysis to general parametric distribution family and provide proof for Bayesian regret. Extensive numerical simulations further demonstrate that TS outperforms more conservative and widely-used approaches such as online convex optimization, upper confidence bounds, and myopic Bayesian dynamic programming.

Method: Introduces selective non-alignment by adding a “skip” operator to contrastive learning’s pull and push operations. SkipAlign selectively skips alignment (pulling) for low-confidence unlabeled samples, retaining only gentle repulsion against ID prototypes, transforming uncertain samples into pure repulsion signals.

Result: Extensive experiments show SkipAlign significantly outperforms state-of-the-art methods in detecting unseen OOD data without sacrificing ID classification accuracy, resulting in tighter ID clusters and naturally dispersed OOD features.

Conclusion: Selective non-alignment through the SkipAlign framework effectively addresses limitations of existing OSSL methods by better handling uncertain samples, improving both OOD detection and maintaining ID classification performance.

Abstract: Open-set semi-supervised learning (OSSL) leverages unlabeled data containing both in-distribution (ID) and unknown out-of-distribution (OOD) samples, aiming simultaneously to improve closed-set accuracy and detect novel OOD instances. Existing methods either discard valuable information from uncertain samples or force-align every unlabeled sample into one or a few synthetic “catch-all” representations, resulting in geometric collapse and overconfidence on only seen OODs. To address the limitations, we introduce selective non-alignment, adding a novel “skip” operator into conventional pull and push operations of contrastive learning. Our framework, SkipAlign, selectively skips alignment (pulling) for low-confidence unlabeled samples, retaining only gentle repulsion against ID prototypes. This approach transforms uncertain samples into a pure repulsion signal, resulting in tighter ID clusters and naturally dispersed OOD features. Extensive experiments demonstrate that SkipAlign significantly outperforms state-of-the-art methods in detecting unseen OOD data without sacrificing ID classification accuracy.

Method: Introduces dictator clients as a well-defined class of malicious participants, proposes concrete attack strategies, provides theoretical analysis of their impact on convergence, and explores complex scenarios with multiple dictator clients (collaboration, independence, betrayal).

Result: Theoretical analysis shows dictator clients can completely erase other clients’ contributions while preserving their own. Empirical evaluations on computer vision and NLP benchmarks support the theoretical findings about various attack scenarios.

Conclusion: Dictator clients represent a significant threat to federated learning systems, capable of dominating the training process. The analysis reveals complex dynamics when multiple dictator clients interact, highlighting the need for robust defense mechanisms in FL.

Abstract: Federated learning (FL) has emerged as a promising paradigm for decentralized model training, enabling multiple clients to collaboratively learn a shared model without exchanging their local data. However, the decentralized nature of FL also introduces vulnerabilities, as malicious clients can compromise or manipulate the training process. In this work, we introduce dictator clients, a novel, well-defined, and analytically tractable class of malicious participants capable of entirely erasing the contributions of all other clients from the server model, while preserving their own. We propose concrete attack strategies that empower such clients and systematically analyze their effects on the learning process. Furthermore, we explore complex scenarios involving multiple dictator clients, including cases where they collaborate, act independently, or form an alliance in order to ultimately betray one another. For each of these settings, we provide a theoretical analysis of their impact on the global model’s convergence. Our theoretical algorithms and findings about the complex scenarios including multiple dictator clients are further supported by empirical evaluations on both computer vision and natural language processing benchmarks.

Method: Combines KLD-weighted Clustered Federated Learning to handle data heterogeneity and multi-domain datasets, with Heterogeneous U-Shaped split learning to address device heterogeneity under strict data privacy constraints where no labels or raw data (real or synthetic) are shared between nodes.

Result: Achieves average 10% boost in classification metrics (up to 60% in multi-domain non-IID settings), 1.1x-3x higher image generation scores for MNIST family datasets, and 2x-70x lower FID scores for higher resolution datasets.

Conclusion: The proposed approach successfully enables decentralized GAN training on distributed, underutilized devices while maintaining strict data privacy, addressing both data and device heterogeneity challenges in real-world settings.

Abstract: Federated Learning has gained attention for its ability to enable multiple nodes to collaboratively train machine learning models without sharing raw data. At the same time, Generative AI – particularly Generative Adversarial Networks (GANs) – have achieved remarkable success across a wide range of domains, such as healthcare, security, and Image Generation. However, training generative models typically requires large datasets and significant computational resources, which are often unavailable in real-world settings. Acquiring such resources can be costly and inefficient, especially when many underutilized devices – such as IoT devices and edge devices – with varying capabilities remain idle. Moreover, obtaining large datasets is challenging due to privacy concerns and copyright restrictions, as most devices are unwilling to share their data. To address these challenges, we propose a novel approach for decentralized GAN training that enables utilizing distributed data and underutilized, low-capability devices while not sharing data in its raw form. Our approach is designed to tackle key challenges in decentralized environments, combining KLD-weighted Clustered Federated Learning to address the issues of data heterogeneity and multi-domain datasets, with Heterogeneous U-Shaped split learning to tackle the challenge of device heterogeneity under strict data sharing constraints – ensuring that no labels or raw data, whether real or synthetic, are ever shared between nodes. Experiments show that our approach demonstrates significant improvements across key metrics, where it achieves an average 10% boost in classification metrics (up to 60% in multi-domain non-IID settings), 1.1x – 3x higher image generation scores for the MNIST family datasets, and 2x – 70x lower FID scores for higher resolution datasets. Find our code at https://distributed-gen-ai.github.io/huscf-gan.github.io/.

Method: ProteinGuide provides a unified statistical framework for conditioning various protein generative models (Masked Language Models, auto-regressive models, diffusion/flow matching) on auxiliary information without retraining.

Result: Successfully designed proteins with specified properties, optimized conflicting properties, and achieved higher editing efficiency in wet lab experiments than 7 rounds of directed evolution using only 2,000 variants.

Conclusion: ProteinGuide enables efficient, on-the-fly conditioning of protein generative models on experimental data, significantly accelerating protein engineering compared to traditional methods.

Abstract: Sequence generative models are transforming protein engineering. However, no principled framework exists for conditioning these models on auxiliary information, such as experimental data, without additional training of a generative model. Herein, we present ProteinGuide, a method for such “on-the-fly” conditioning, amenable to a broad class of protein generative models including Masked Language Models (e.g. ESM3), any-order auto-regressive models (e.g. ProteinMPNN) as well as diffusion and flow matching models (e.g. MultiFlow). ProteinGuide stems from our unifying view of these model classes under a single statistical framework. As proof of principle, we perform several in silico experiments. We first guide pre-trained generative models to design proteins with user-specified properties, such as higher stability or activity. Next, we design for optimizing two desired properties that are in tension with each other. Finally, we apply our method in the wet lab, using ProteinGuide to increase the editing activity of an adenine base editor in vivo with data from only a single pooled library of 2,000 variants. We find that a single round of ProteinGuide achieves a higher editing efficiency than was previously achieved using seven rounds of directed evolution.

Method: PHAR transforms numeric feature attributions from instance-wise explainers into structured rules with human-readable intervals. Includes rule fusion step using weighted selection and lasso-based refinement to consolidate rule sets, balancing coverage, confidence, and simplicity. Also introduces visualization techniques for specificity-generalization trade-offs.

Result: PHAR performs comparably to native rule-based methods like Anchor, scales efficiently to long time series sequences, achieves broader instance coverage, resolves conflicting explanations from Rashomon phenomenon, and provides coherent domain-adaptable insights.

Conclusion: PHAR improves interpretability, decision transparency, and practical applicability for time series classification by providing concise, human-readable rules aligned with model predictions, as demonstrated on UCR/UEA Time Series Classification Archive.

Abstract: Explaining machine learning (ML) models for time series (TS) classification remains challenging due to the difficulty of interpreting raw time series and the high dimensionality of the input space. We introduce PHAR–Post-hoc Attribution Rules–a unified framework that transforms numeric feature attributions from post-hoc, instance-wise explainers (e.g. LIME, SHAP) into structured, human-readable rules. These rules define human-readable intervals that indicate where and when decision-relevant segments occur and can enhance model transparency by localizing threshold-based conditions on the raw series. PHAR performs comparably to native rule-based methods, such as Anchor, while scaling more efficiently to long TS sequences and achieving broader instance coverage. A dedicated rule fusion step consolidates rule sets using strategies like weighted selection and lasso-based refinement, balancing key quality metrics: coverage, confidence, and simplicity. This fusion ensures each instance receives a concise and unambiguous rule, improving both explanation fidelity and consistency. We further introduce visualization techniques to illustrate specificity-generalization trade-offs in the derived rules. PHAR resolves conflicting and overlapping explanations–a common effect of the Rashomon phenomenon–into coherent, domain-adaptable insights. Comprehensive experiments on UCR/UEA Time Series Classification Archive demonstrate that PHAR may improve interpretability, decision transparency, and practical applicability for TS classification tasks by providing concise, human-readable rules aligned with model predictions.

Method: Developed ThinkEval framework that builds and employs specialized knowledge graphs to analyze causal structure of facts before and after editing. Created KnowGIC benchmark dataset comprising multi-step reasoning paths to precisely measure complex knowledge transformation effects.

Result: Evaluated five editing techniques (AlphaEdit, RECT, ROME, MEMIT, PRUNE) across multiple LLMs. Results show these techniques struggle to balance indirect fact suppression with preservation of related knowledge, compromising the contextual integrity of a model’s knowledge.

Conclusion: ThinkEval provides a systematic framework for evaluating model-editing techniques, revealing that current methods fail to adequately handle indirect knowledge leakage and ripple effects, highlighting the need for more sophisticated editing approaches that maintain contextual knowledge integrity.

Abstract: Robust model-editing techniques are essential for deploying large language models (LLMs) in practical applications, as they enable cost-effective ways to deal with challenges such as privacy breaches, bias mitigation and misinformation spread. For example, an LLM-based healthcare assistance may need to update out-dated or incorrect knowledge to prevent harmful recommendations. However, many editing techniques focus on isolated facts, which critically fail to prevent indirect knowledge leakage – the unintended reconstruction of edited-out information through persistent causal links and contextual relationships. To assist users in selecting the right editing technique, we develop and present ThinkEval, a framework to systematically quantify indirect knowledge leakage and ripple effects in model-editing. ThinkEval builds and employs specialized knowledge graphs to analyze the causal structure of facts before and after editing. To support this approach, we present KnowGIC, a benchmark dataset comprising multi-step reasoning paths that precisely measure these complex knowledge transformation effects. We evaluate five editing techniques: AlphaEdit, RECT, ROME, MEMIT, and PRUNE across multiple LLMs. Our results show that these techniques struggle to balance indirect fact suppression with the preservation of related knowledge, compromising the contextual integrity of a model’s knowledge. Our dataset is available at: https://github.com/manitbaser/KnowGIC.

Method: Systematic comparison using HiQ dataset (U.S., 2000-2022) with multiple evaluation protocols: statistical baselines, fully supervised LSTM, and Sundial foundation model in zero-shot setting (no task-specific tuning).

Result: Sundial in zero-shot setting outperforms fully trained LSTM when input context window covers more than one or two full seasonal cycles. General-purpose foundation model surpasses specialized supervised models without any task-specific tuning.

Conclusion: Pretrained time series foundation models have strong potential as effective plug-and-play forecasters in agricultural and environmental applications, demonstrating zero-shot forecasting capability that can outperform specialized supervised models.

Abstract: This work investigates the zero-shot forecasting capability of time series foundation models for Leaf Area Index (LAI) forecasting in agricultural monitoring. Using the HiQ dataset (U.S., 2000-2022), we systematically compare statistical baselines, a fully supervised LSTM, and the Sundial foundation model under multiple evaluation protocols. We find that Sundial, in the zero-shot setting, can outperform a fully trained LSTM provided that the input context window is sufficiently long-specifically, when covering more than one or two full seasonal cycles. We show that a general-purpose foundation model can surpass specialized supervised models on remote-sensing time series prediction without any task-specific tuning. These results highlight the strong potential of pretrained time series foundation models to serve as effective plug-and-play forecasters in agricultural and environmental applications.

Method: End-to-end physics-informed hierarchical network that reconstructs scattering quantities via hierarchical operator design inspired by near-far field decomposition. Incorporates physics-guided graph neural network to capture electromagnetic coupling among mesh elements, with governing equations embedded as residual constraints.

Result: Achieves EM-solver-level RCS accuracy and 3D object reconstruction with orders-of-magnitude speedups, generalizes well to unseen geometries under limited training data.

Conclusion: U-PINet bridges the gap between high-fidelity but expensive CEM solvers and efficient but physically inconsistent data-driven approaches, enabling accurate, efficient, and physically consistent RCS prediction.

Abstract: Conventional computational electromagnetics (CEM) solvers can deliver high fidelity radar cross section (RCS) signatures by first solving the induced surface currents on 3-dimensional (3D) targets and then evaluating the scattered fields via radiation integrals. However, their computational cost becomes prohibitive for repeated queries and large-scale 3D scenarios. Recent purely data-driven networks improve efficiency, yet they often bypass this scattering mechanism, which may compromise physical consistency and generalization. To bridge this gap, in this paper, we propose U-PINet, a fully end-to-end, physics-informed hierarchical network for efficient RCS prediction via 3D electromagnetic scattering reconstruction. Once the scattering quantities are reconstructed, scattered fields and RCS can be evaluated for arbitrary observation directions via the radiation integral. U-PINet explicitly learns physics-consistent intermediate scattering representations by modeling local electromagnetic coupling and long-range radiation effects through a hierarchical operator design inspired by near-far field decomposition in fast solvers. A physics-guided graph neural network is incorporated to capture self- and mutual-coupling among mesh elements of complex targets, enabling physically interpretable intermediate representations. By embedding governing equations as residual constraints, U-PINet enables accurate object reconstruction of scattering quantities and consequently reliable RCS prediction across observation directions, while significantly reducing runtime. Extensive numerical experiments demonstrate that U-PINet achieves EM-solver-level RCS accuracy and 3D object reconstruction with orders-of-magnitude speedups, and generalizes well to unseen geometries under limited training data.

Method: Proposed a physiological-model-based neural network (PMB-NN) framework that embeds physiological constraints from a simplified human movement physiological model into neural network training. The model uses VO2 data during daily physical activities (resting, cycling, running) and was trained/tested on individual datasets from 12 participants.

Result: PMB-NN achieved median R² score of 0.8 and RMSE of 8.3 bpm. It performed on par with benchmark neural network models while significantly outperforming traditional physiological models (p=0.002). The framework also successfully identified personalized physiological parameters.

Conclusion: The PMB-NN framework enables personalized, real-time cardiac monitoring during daily activities by combining physiological principles with neural network accuracy. This approach offers potential for early heart failure detection through individualized heart rate tracking.

Abstract: Heart failure (HF) poses a significant global health challenge, with early detection offering opportunities for improved outcomes. Abnormalities in heart rate (HR), particularly during daily activities, may serve as early indicators of HF risk. However, existing HR monitoring tools for HF detection are limited by their reliability on population-based averages. The estimation of individualized HR serves as a dynamic digital twin, enabling precise tracking of cardiac health biomarkers. Current HR estimation methods, categorized into physiologically-driven and purely data-driven models, struggle with efficiency and interpretability. This study introduces a novel physiological-model-based neural network (PMB-NN) framework for HR estimation based on oxygen uptake (VO2) data during daily physical activities. The framework was trained and tested on individual datasets from 12 participants engaged in activities including resting, cycling, and running. By embedding physiological constraints, which were derived from our proposed simplified human movement physiological model (PM), into the neural network training process, the PMB-NN model adheres to human physiological principles while achieving high estimation accuracy, with a median R$^2$ score of 0.8 and an RMSE of 8.3 bpm. Comparative statistical analysis demonstrates that the PMB-NN achieves performance on par with the benchmark neural network model while significantly outperforming traditional physiological model (p=0.002). In addition, our PMB-NN is adept at identifying personalized parameters of the PM, enabling the PM to generate reasonable HR estimation. The proposed framework with a precise VO2 estimation system derived from body movements enables the future possibilities of personalized and real-time cardiac monitoring during daily life physical activities.

Method: Proposes an enhanced autoencoder framework that incorporates clustering loss (based on soft silhouette score) and contrastive loss. Uses EfficientNetV2 to generate pseudo-labels for unlabeled data. Jointly optimizes reconstruction, clustering, and contrastive objectives to group similar data points and separate distinct clusters in latent space. Applies UMAP to latent representation for 2D projections, evaluated using silhouette score. Tests multiple autoencoder types.

Result: Experiments on two scientific ensemble datasets (channel structures in soil from MCMC, and droplet-on-film impact dynamics) show that models with clustering or contrastive loss marginally outperform baseline approaches in extracting meaningful features.

Conclusion: The enhanced autoencoder framework with clustering and contrastive losses improves visualization and interpretability of high-dimensional scientific ensemble datasets, though improvements over baselines are marginal.

Abstract: Analyzing and visualizing scientific ensemble datasets with high dimensionality and complexity poses significant challenges. Dimensionality reduction techniques and autoencoders are powerful tools for extracting features, but they often struggle with such high-dimensional data. This paper presents an enhanced autoencoder framework that incorporates a clustering loss, based on the soft silhouette score, alongside a contrastive loss to improve the visualization and interpretability of ensemble datasets. First, EfficientNetV2 is used to generate pseudo-labels for the unlabeled portions of the scientific ensemble datasets. By jointly optimizing the reconstruction, clustering, and contrastive objectives, our method encourages similar data points to group together while separating distinct clusters in the latent space. UMAP is subsequently applied to this latent representation to produce 2D projections, which are evaluated using the silhouette score. Multiple types of autoencoders are evaluated and compared based on their ability to extract meaningful features. Experiments on two scientific ensemble datasets - channel structures in soil derived from Markov chain Monte Carlo, and droplet-on-film impact dynamics - show that models incorporating clustering or contrastive loss marginally outperform the baseline approaches.

Method: PasoDoble adversarially trains two models from the same base: a Proposer generates challenging questions with ground-truth answers, enriched with pre-training dataset knowledge; a Solver attempts to solve them. The Proposer is rewarded for valid questions that push the Solver’s limits, while the Solver is rewarded for correct solutions. An optional offline paradigm decouples updates for stability.

Result: Experimental results show that PasoDoble can improve the reasoning performance of LLMs. The framework operates without supervision during training.

Conclusion: PasoDoble successfully adapts dual-play adversarial training to LLMs, addressing reward hacking and stability issues while reducing dependency on external supervision, demonstrating improved reasoning capabilities.

Abstract: Large Language Models (LLMs) have achieved remarkable progress through Reinforcement Learning with Verifiable Rewards (RLVR), yet still rely heavily on external supervision (e.g., curated labels). Adversarial learning, particularly through self-play, offers a promising alternative that enables models to iteratively learn from themselves - thus reducing reliance on external supervision. Dual-play extends adversarial learning by assigning specialized roles to two models and training them against each other, fostering sustained competition and mutual evolution. Despite its promise, adapting dual-play training to LLMs remains limited, largely due to their susceptibility to reward hacking and training instability. In this paper, we introduce PasoDoble, a novel LLM dual-play framework. PasoDoble adversarially trains two models initialized from the same base model: a Proposer, which generates challenging questions with ground-truth answers, and a Solver, which attempts to solve them. We enrich the Proposer with knowledge from a pre-training dataset to ensure the questions’ quality and diversity. To avoid reward hacking, the Proposer is rewarded for producing only valid questions that push the Solver’s limit, while the Solver is rewarded for solving them correctly, and both are updated jointly. To further enhance training stability, we introduce an optional offline paradigm that decouples Proposer and Solver updates, alternately updating each for several steps while holding the other fixed. Notably, PasoDoble operates without supervision during training. Experimental results show that PasoDoble can improve the reasoning performance of LLMs. Our project page is available at https://hcy123902.github.io/PasoDoble.

Method: The paper presents a primer/review approach, examining technical foundations of FWPs, their computational characteristics, and establishing connections to transformers and state space models through theoretical analysis and comparison.

Result: The review establishes FWPs as a family of RNNs with 2D matrix-form hidden states that serve as short-term memory storage, where fast weights dynamically change over time via a programmer network, showing connections to modern architectures and biological learning.

Conclusion: FWPs represent an important architectural innovation bridging conventional RNNs with modern sequence models, offering insights into both artificial intelligence design and potential biological parallels in synaptic plasticity mechanisms.

Abstract: Recent advances in artificial neural networks for machine learning, and language modeling in particular, have established a family of recurrent neural network (RNN) architectures that, unlike conventional RNNs with vector-form hidden states, use two-dimensional (2D) matrix-form hidden states. Such 2D-state RNNs, known as Fast Weight Programmers (FWPs), can be interpreted as a neural network whose synaptic weights (called fast weights) dynamically change over time as a function of input observations, and serve as short-term memory storage; corresponding synaptic weight modifications are controlled or programmed by another network (the programmer) whose parameters are trained (e.g., by gradient descent). In this Primer, we review the technical foundations of FWPs, their computational characteristics, and their connections to transformers and state space models. We also discuss connections between FWPs and models of synaptic plasticity in the brain, suggesting a convergence of natural and artificial intelligence.

Method: Three core components: 1) Unified preprocessing pipeline for lightcurve refinement; 2) LoRA-finetuned LLM backbone enhanced with first-order difference augmentation, flare statistics, and historical records for multimodal fusion; 3) Physics-informed loss with minimum rising rate prior added to cross-entropy loss.

Result: Extensive experiments on Kepler and TESS datasets show StellarF achieves state-of-the-art performance across key metrics, setting new benchmarks for flare forecasting.

Conclusion: StellarF bridges general AI with astrophysics, offering a practical, physically interpretable paradigm for transient event forecasting in time-domain astronomy.

Abstract: Stellar flare forecasting represents a critical frontier in astrophysics, offering profound insights into stellar activity mechanisms and exoplanetary habitability assessments. Yet the inherent unpredictability of flare activity, rooted in stellar diversity and evolutionary stages, underpins the field’s core challenges: (1) sparse, incomplete, noisy lightcurve data from traditional observations; (2) ineffective multi-scale flare evolution capture via single representations; (3) poor physical interpretability in data-driven models lacking physics-informed priors. To address these challenges, we propose StellarF, a physics-informed framework synergizing general Al with astrophysical domain knowledge via three core components: a unified preprocessing pipeline for lightcurve refinement (missing-value imputation, temporal patch partitioning, adaptive sample filtering); a Low-Rank Adaptation (LoRA)-finetuned large language model (LLM) backbone enhanced by first-order difference augmentation, flare statistical information, and flare historical record modules for multimodal fusion instead of only simple representations; and a novel physics-informed loss embedding a minimum rising rate prior, appended to the cross-entropy loss, to align with flare physics. Extensive experiments on Kepler and TESS datasets show StellarF achieves state-of-the-art performance across key metrics, setting new benchmarks for flare forecasting. This work bridges general AI with astrophysics, offering a practical, physically interpretable paradigm for transient event forecasting in time-domain astronomy.

Method: DALI integrates within Dreamer architecture, trains a self-supervised encoder to predict forward dynamics, generates actionable latent context representations that condition the world model and policy, bridging perception and control.

Result: DALI achieves significant gains over context-unaware baselines, often surpasses context-aware baselines in extrapolation tasks, enables zero-shot generalization to unseen contextual variations, and demonstrates counterfactual consistency in latent space.

Conclusion: DALI provides an effective framework for learning latent context representations that enable robust generalization to unseen environmental conditions without requiring explicit context variables, making it practical for real-world RL applications.

Abstract: Real-world reinforcement learning demands adaptation to unseen environmental conditions without costly retraining. Contextual Markov Decision Processes (cMDP) model this challenge, but existing methods often require explicit context variables (e.g., friction, gravity), limiting their use when contexts are latent or hard to measure. We introduce Dynamics-Aligned Latent Imagination (DALI), a framework integrated within the Dreamer architecture that infers latent context representations from agent-environment interactions. By training a self-supervised encoder to predict forward dynamics, DALI generates actionable representations conditioning the world model and policy, bridging perception and control. We theoretically prove this encoder is essential for efficient context inference and robust generalization. DALI’s latent space enables counterfactual consistency: Perturbing a gravity-encoding dimension alters imagined rollouts in physically plausible ways. On challenging cMDP benchmarks, DALI achieves significant gains over context-unaware baselines, often surpassing context-aware baselines in extrapolation tasks, enabling zero-shot generalization to unseen contextual variations.

Method: Proposes DCCD-CONF framework that alternates between optimizing graph structure and estimating confounder distribution by maximizing log-likelihood of interventional data. Handles nonlinear cyclic graphs with unmeasured confounders.

Result: Outperforms state-of-the-art methods in both causal graph recovery and confounder identification on synthetic data and real-world gene perturbation datasets.

Conclusion: DCCD-CONF provides an effective solution for causal discovery in complex real-world systems with cycles and unmeasured confounders, supported by both empirical results and theoretical consistency guarantees.

Abstract: Understanding causal relationships between variables is fundamental across scientific disciplines. Most causal discovery algorithms rely on two key assumptions: (i) all variables are observed, and (ii) the underlying causal graph is acyclic. While these assumptions simplify theoretical analysis, they are often violated in real-world systems, such as biological networks. Existing methods that account for confounders either assume linearity or struggle with scalability. To address these limitations, we propose DCCD-CONF, a novel framework for differentiable learning of nonlinear cyclic causal graphs in the presence of unmeasured confounders using interventional data. Our approach alternates between optimizing the graph structure and estimating the confounder distribution by maximizing the log-likelihood of the data. Through experiments on synthetic data and real-world gene perturbation datasets, we show that DCCD-CONF outperforms state-of-the-art methods in both causal graph recovery and confounder identification. Additionally, we also provide consistency guarantees for our framework, reinforcing its theoretical soundness.

Method: Introduces Bayesian-inspired posterior proxy by injecting stochasticity directly into normalization layers (avoiding full Bayesian NN training). During generation, samples a single realization from the random distribution and holds it fixed for temporal consistency. Creates a population of “minds” with diverse behaviors while maintaining general competence.

Result: Experiments on zero-shot generation and Reinforcement Learning with Verifiable Rewards (RLVR) show B-Trans effectively leverages stochastic model diversity, yielding superior response diversity while achieving better task performance compared to deterministic baselines.

Conclusion: B-Trans successfully creates transformer populations that combine behavioral diversity with maintained competence, demonstrating the value of population-level approaches inspired by human intelligence emergence.

Abstract: Despite their scale and success, modern transformers are usually trained as single-minded systems: optimization produces a deterministic set of parameters, representing a single functional hypothesis about the data. Motivated by the analogy to human populations, in which population-level intelligence emerges from diverse individual behaviors, we propose Population Bayesian Transformers (B-Trans), which enable sampling diverse yet coherent transformer large language model instances (hereafter referred to as a ‘mind’) from a single pre-trained LLM. B-Trans introduces a Bayesian-inspired posterior proxy by injecting stochasticity directly into normalization layers, avoiding the prohibitive cost of training full Bayesian neural networks. Sampling from this proxy yields a population of minds with diverse behaviors while maintaining general competence. During the generation of each response, we sample a single realization from the random distribution and hold it fixed, ensuring temporal consistency and reasoning coherence. Experiments on zero-shot generation and Reinforcement Learning with Verifiable Rewards (RLVR) demonstrate that B-Trans effectively leverages the stochastic model diversity, yielding superior response diversity while achieving better task performance compared to deterministic baselines.

Method: BLIP uses a scalable variational Bayesian framework built on adaptive Variational Dropout, which is architecture-agnostic and integrates seamlessly with (equivariant) message-passing architectures, providing uncertainty estimates with minimal computational overhead.

Result: BLIP demonstrates improved predictive accuracy over standard MLIPs, delivers trustworthy uncertainty estimates especially in data-scarce or heavy out-of-distribution regimes, and fine-tuning pretrained MLIPs with BLIP yields consistent performance gains with calibrated uncertainties.

Conclusion: BLIP addresses critical limitations of MLIPs by providing scalable Bayesian uncertainty estimation, enabling more reliable simulation-based chemistry with better handling of data-scarce and out-of-distribution scenarios.

Abstract: Machine Learning Interatomic Potentials (MLIPs) are becoming a central tool in simulation-based chemistry. However, like most deep learning models, MLIPs struggle to make accurate predictions on out-of-distribution data or when trained in a data-scarce regime, both common scenarios in simulation-based chemistry. Moreover, MLIPs do not provide uncertainty estimates by construction, which are fundamental to guide active learning pipelines and to ensure the accuracy of simulation results compared to quantum calculations. To address this shortcoming, we propose BLIPs: Bayesian Learned Interatomic Potentials. BLIP is a scalable, architecture-agnostic variational Bayesian framework for training or fine-tuning MLIPs, built on an adaptive version of Variational Dropout. BLIP delivers well-calibrated uncertainty estimates and minimal computational overhead for energy and forces prediction at inference time, while integrating seamlessly with (equivariant) message-passing architectures. Empirical results on simulation-based computational chemistry tasks demonstrate improved predictive accuracy with respect to standard MLIPs, and trustworthy uncertainty estimates, especially in data-scarse or heavy out-of-distribution regimes. Moreover, fine-tuning pretrained MLIPs with BLIP yields consistent performance gains and calibrated uncertainties.

Method: Uses controlled experiments with unambiguous graph concepts, extracts features as (1) class-conditional centroids at graph level and (2) linear-probe directions at node level, then analyzes their geometry with basis-invariant diagnostics across GCN/GIN/GAT architectures.

Result: Increasing width produces phase patterns in overlap; topology imprints overlap onto node-level features that pooling partially remixes into task-aligned graph axes; sharper pooling increases axis alignment and reduces channel sharing; shallow models can settle into metastable low-rank embeddings.

Conclusion: The results connect representational geometry with concrete design choices (width, pooling, final-layer activations) and suggest practical approaches for building more interpretable GNNs.

Abstract: Interpreting graph neural networks (GNNs) is difficult because message passing mixes signals and internal channels rarely align with human concepts. We study superposition, the sharing of directions by multiple features, directly in the latent space of GNNs. Using controlled experiments with unambiguous graph concepts, we extract features as (i) class-conditional centroids at the graph level and (ii) linear-probe directions at the node level, and then analyze their geometry with simple basis-invariant diagnostics. Across GCN/GIN/GAT we find: increasing width produces a phase pattern in overlap; topology imprints overlap onto node-level features that pooling partially remixes into task-aligned graph axes; sharper pooling increases axis alignment and reduces channel sharing; and shallow models can settle into metastable low-rank embeddings. These results connect representational geometry with concrete design choices (width, pooling, and final-layer activations) and suggest practical approaches for more interpretable GNNs.

Method: Novel trial & error approach using vanilla decoder-only Transformer (GPT-2) without external tools. Combines imitation learning of Sudoku rules with explicit Depth-First Search (DFS) exploration involving informed guessing and backtracking. Uses depth-1 guessing strategy to minimize guesses until solution.

Result: Achieves state-of-the-art 99% accuracy on Sudoku puzzles, outperforming prior neuro-symbolic approaches. Shows empirically that almost all Sudoku puzzles can be solved using puzzle rules with at most one guess.

Conclusion: The method successfully addresses LLMs’ limitations in combinatorial reasoning through a trial & error approach with DFS exploration. Provides rigorous analysis connecting the setup to Min-Sum Set Cover, demonstrating effective solving of NP-class problems using standard Transformer architectures without custom tools.

Abstract: Despite their proficiency in various language tasks, Large Language Models (LLMs) struggle with combinatorial problems like Satisfiability, Traveling Salesman Problem, or even basic arithmetic. We address this gap through a novel trial & error approach for solving problems in the class NP, where candidate solutions are iteratively generated and efficiently validated using verifiers. We focus on the paradigmatic task of Sudoku and achieve state-of-the-art accuracy (99%) compared to prior neuro-symbolic approaches. Unlike prior work that used custom architectures, our method employs a vanilla decoder-only Transformer (GPT-2) without external tools or function calling. Our method integrates imitation learning of simple Sudoku rules with an explicit Depth-First Search (DFS) exploration strategy involving informed guessing and backtracking. Moving beyond imitation learning, we seek to minimize the number of guesses until reaching a solution. This is achieved using depth-1 guessing, showing empirically that almost all Sudoku can be solved using the puzzle’s rules with at most one guess. We provide a rigorous analysis of this setup formalizing its connection to a contextual variant of Min-Sum Set Cover, a well-studied problem in algorithms and stochastic optimization.

Method: Proposes TetriServe with step-level sequence parallelism that dynamically adjusts parallelism per request based on deadlines. Uses round-based scheduling: (1) discretizes time into fixed rounds for tractable deadline-aware scheduling, (2) adapts parallelism at step level to minimize GPU hour consumption, and (3) jointly packs requests to minimize late completions.

Result: Extensive evaluation on state-of-the-art DiT models shows TetriServe achieves up to 32% higher SLO attainment compared to existing solutions without degrading image quality.

Conclusion: TetriServe’s step-level sequence parallelism and round-based scheduling effectively address the challenge of serving DiT models under strict SLOs for heterogeneous workloads, significantly improving SLO attainment while maintaining image quality.

Abstract: Diffusion Transformer (DiT) models excel at generating high-quality images through iterative denoising steps, but serving them under strict Service Level Objectives (SLOs) is challenging due to their high computational cost, particularly at large resolutions. Existing serving systems use fixed degree sequence parallelism, which is inefficient for heterogeneous workloads with mixed resolutions and deadlines, leading to poor GPU utilization and low SLO attainment. In this paper, we propose step-level sequence parallelism to dynamically adjust the degree of parallelism of individual requests according to their deadlines. We present TetriServe, a DiT serving system that implements this strategy for highly efficient image generation. Specifically, TetriServe introduces a novel round-based scheduling mechanism that improves SLO attainment: (1) discretizing time into fixed rounds to make deadline-aware scheduling tractable, (2) adapting parallelism at the step level and minimize GPU hour consumption, and (3) jointly packing requests to minimize late completions. Extensive evaluation on state-of-the-art DiT models shows that TetriServe achieves up to 32% higher SLO attainment compared to existing solutions without degrading image quality.

Method: Leverages Hankel singular value analysis and eigenvalue stability properties to identify high-influence dimensions during training. Applies balanced truncation to compress Linear Time-Invariant SSMs (like Linear Recurrent Units) while preserving performance-critical structure.

Result: In-training reduction significantly accelerates optimization while preserving expressivity. Compressed models retain task-critical structure that models trained directly at smaller dimensions lose. SSMs that begin large and shrink during training achieve computational efficiency with higher performance.

Conclusion: CompreSSM demonstrates that applying control theory principles during SSM training enables effective model compression, maintaining performance while reducing computational costs. The approach works for LTI SSMs and can be extended to selective models.

Abstract: State Space Models (SSMs), developed to tackle long sequence modeling tasks efficiently, offer both parallelizable training and fast inference. At their core are recurrent dynamical systems that maintain a hidden state, with update costs scaling with the state dimension. A key design challenge is striking the right balance between maximizing expressivity and limiting this computational burden. Control theory, and more specifically Hankel singular value analysis, provides a potent framework for the measure of energy for each state, as well as the balanced truncation of the original system down to a smaller representation with performance guarantees. Leveraging the eigenvalue stability properties of Hankel matrices, we apply this lens to SSMs \emph{during training}, where only dimensions of high influence are identified and preserved. Our approach, \textsc{CompreSSM}, applies to Linear Time-Invariant SSMs such as Linear Recurrent Units, but is also extendable to selective models. Experiments show that in-training reduction significantly accelerates optimization while preserving expressivity, with compressed models retaining task-critical structure lost by models trained directly at smaller dimension. In other words, SSMs that begin large and shrink during training achieve computational efficiency while maintaining higher performance. Project code is available at github.com/camail-official/compressm.

Method: Propose multi-task neural diffusion processes with a task encoder that extracts low-dimensional representations from context observations. This task representation conditions the diffusion model, allowing information sharing across tasks while preserving input-size agnosticity and equivariance properties of neural diffusion processes.

Result: Empirical results show improved point prediction accuracy and better-calibrated predictive uncertainty compared to single-task neural diffusion processes and Gaussian process baselines. Validated on real wind farm data for wind power prediction, demonstrating effective few-shot adaptation in challenging real-world multi-task regression.

Conclusion: The framework retains the expressiveness and scalability of neural diffusion processes while enabling efficient transfer to unseen tasks, with practical applications in high-impact domains like wind farm management where reliable uncertainty quantification supports operational decision-making.

Abstract: Neural diffusion processes provide a scalable, non-Gaussian approach to modelling distributions over functions, but existing formulations are limited to single-task inference and do not capture dependencies across related tasks. In many multi-task regression settings, jointly modelling correlated functions and enabling task-aware conditioning is crucial for improving predictive performance and uncertainty calibration, particularly in low-data regimes. We propose multi-task neural diffusion processes, an extension that incorporates a task encoder to enable task-conditioned probabilistic regression and few-shot adaptation across related functions. The task encoder extracts a low-dimensional representation from context observations and conditions the diffusion model on this representation, allowing information sharing across tasks while preserving input-size agnosticity and the equivariance properties of neural diffusion processes. The resulting framework retains the expressiveness and scalability of neural diffusion processes while enabling efficient transfer to unseen tasks. Empirical results demonstrate improved point prediction accuracy and better-calibrated predictive uncertainty compared to single-task neural diffusion processes and Gaussian process baselines. We validate the approach on real wind farm data appropriate for wind power prediction. In this high-impact application, reliable uncertainty quantification directly supports operational decision-making in wind farm management, illustrating effective few-shot adaptation in a challenging real-world multi-task regression setting.

Method: Two experimental approaches tested: 1) Direct communication via a one-word “cheap talk” channel in a 4-player Stag Hunt game, and 2) Curriculum learning through progressively complex games in an Iterated Public Goods Game with Punishment. Qualitative analysis examined agent learning patterns.

Result: Communication was highly effective - cheap talk increased cooperation from 0% to 48.3%. Curriculum learning backfired - pedagogical curriculum reduced agent payoffs by 27.4% and induced “learned pessimism” when emphasizing defection-equilibrium games.

Conclusion: For coordination problems, simple communication protocols are more reliable than experience-based training. Curriculum design for social dilemmas requires careful attention to strategic lessons in game sequences, as optimizing for short-term rationality can undermine alignment goals.

Abstract: Eliciting cooperation in multi-agent LLM systems is critical for AI alignment. We investigate two approaches: direct communication and curriculum learning. In a 4-player Stag Hunt, a one-word “cheap talk” channel increases cooperation from 0% to 48.3%, demonstrating communication as a robust coordination mechanism. In contrast, we find that curriculum learning is highly sensitive to design choices: our pedagogical curriculum through progressively complex games reduced agent payoffs by 27.4% in an Iterated Public Goods Game with Punishment, demonstrating that optimizing for short-term rationality can actively undermine alignment goals. Qualitative analysis reveals that curricula emphasizing defection-equilibrium games can induce “learned pessimism” in agents. These findings suggest that for coordination problems, simple communication protocols may be more reliable than experience-based training, and that curriculum design for social dilemmas requires careful attention to the strategic lessons embedded in game sequences.

Method: Analyze collections of distributions on {0,1}^ℕ via their mean vectors (expected values in each coordinate). Prove separability of mean vectors is sufficient for UME-learnability, construct counterexamples showing non-separable but UME-learnable collections, and prove closure under countable unions.

Result: Separability of mean vectors is sufficient but not necessary for UME-learnability. Countable unions of UME-learnable collections remain UME-learnable, resolving Cohen et al. (2025) conjecture.

Conclusion: The paper establishes UME-learnability as a fundamental property for uniform mean estimation, showing separability provides one sufficient condition but not the only path to uniform estimability, and proving closure properties that enable broader applications.

Abstract: We characterize conditions under which collections of distributions on ${0,1}^\mathbb{N}$ admit uniform estimation of their mean. Prior work from Vapnik and Chervonenkis (1971) has focused on uniform convergence using the empirical mean estimator, leading to the principle known as $P-$ Glivenko-Cantelli. We extend this framework by moving beyond the empirical mean estimator and introducing Uniform Mean Estimability, also called UME-learnability, which captures when a collection permits uniform mean estimation by any arbitrary estimator. We work on the space created by the mean vectors of the collection of distributions. For each distribution, the mean vector records the expected value in each coordinate. We show that separability of the mean vectors is a sufficient condition for UME-learnability. However, we show that separability of the mean vectors is not necessary for UME-learnability by constructing a collection of distributions whose mean vectors are non-separable yet UME-learnable using techniques fundamentally different from those used in our separability-based analysis. Finally, we establish that countable unions of UME-learnable collections are also UME-learnable, solving the conjecture posed in Cohen et al. (2025).

Method: Introduces Feature Error Measure (FEM) to quantify learned feature quality, analyzes training dynamics of adaptive feature models (linear regression, single/multiple index models) using over-parameterized sequence models as a simplification framework.

Result: Shows that FEM consistently decreases during training across various adaptive feature models, providing empirical evidence supporting the validity and potential of the adaptive feature program approach.

Conclusion: The decreasing FEM during training across multiple models suggests the adaptive feature program is a promising framework for analyzing neural network feature learning, potentially advancing theoretical understanding of neural network advantages.

Abstract: Theoretically exploring the advantages of neural networks might be one of the most challenging problems in the AI era. An adaptive feature program has recently been proposed to analyze feature learning, the characteristic property of neural networks, in a more abstract way. Motivated by the celebrated Le Cam equivalence, we advocate the over-parameterized sequence models to further simplify the analysis of the training dynamics of adaptive feature program and present several pieces of supporting evidence for the adaptive feature program. More precisely, after having introduced the feature error measure (FEM) to characterize the quality of the learned feature, we show that the FEM is decreasing during the training process of several concrete adaptive feature models including linear regression, single/multiple index models, etc. We believe that this hints at the potential successes of the adaptive feature program.

Method: Separates inference and training deployment with improved data loader, creating periodically asynchronous framework; uses unified tri-model architecture in training phase and shared-prompt attention mask to reduce repetitive computation.

Result: Achieves significant end-to-end training efficiency improvements on NPU platforms while maintaining algorithm accuracy equivalent to synchronous methods (both on-policy strategies).

Conclusion: The proposed framework enables demand-driven, independent, and elastic scaling of inference and training components, showing potential for widespread application in RL systems.

Abstract: Since the introduction of the GRPO algorithm, reinforcement learning (RL) has attracted increasing attention, with growing efforts to reproduce and apply it. However, training efficiency remains a critical challenge. In mainstream RL frameworks, inference and training are typically deployed on the same devices. While this approach reduces costs through resource consolidation, its synchronous execution imposes a computational coupling that prevents concurrent inference and training. In this study, we are returning to the strategy of separating inference and training deployment, and by introducing improvements in the data loader, we transform the conventional synchronous architecture into a periodically asynchronous framework, which allows for demand-driven, independent, and elastic scaling of each component, while the accuracy of the algorithm remains completely equivalent to the synchronization method, with both belonging to the on-policy strategy. It is worth emphasizing that we apply a unified tri-model architecture in the training phase, and we also proposed a shared-prompt attention mask to reduce repetitive computation. In practice, our approach consistently delivers significant end-to-end training efficiency improvements on NPU platforms, indicating its potential for widespread application.

Method: Proposes DemoTuner with two key components: 1) Structured chain-of-thought prompts using LLMs to extract condition-aware tuning hints from documents, and 2) HA-DDPGfD algorithm (hint-aware demonstration reinforcement learning) that integrates mined hints into RL agent training via demonstration reinforcement learning.

Result: Achieves performance gains up to 44.01% for MySQL and 39.95% for PostgreSQL over default configurations. Reduces execution time by up to 10.03% compared to baseline methods while consuming least online tuning cost. Shows superior adaptability to unknown workloads.

Conclusion: DemoTuner successfully leverages textual documents to improve RL-based DBMS tuning, introducing the first demonstration reinforcement learning approach for this domain. The framework effectively mines and integrates tuning hints to accelerate convergence and enhance performance.

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 that DemoTuner achieves performance gains of up to 44.01% for MySQL and 39.95% for PostgreSQL over default configurations. Compared with three representative baseline methods, DemoTuner is able to further reduce the execution time by up to 10.03%, while always consuming the least online tuning cost. Additionally, DemoTuner also exhibits superior adaptability to application scenarios with unknown workloads.

Method: Two-stage cascaded approach: 1) Lightweight neural-operator functional autoencoder infers coarse-scale approximation as explicit intermediate variable; 2) Conditional diffusion model refines details using mask-cascade training for robustness. Enforces measurement consistency via manifold-constrained gradients in Bayesian posterior framework.

Result: Substantially alleviates ill-posedness, enabling accurate and stable reconstructions even under extreme sparsity with diverse sensing patterns.

Conclusion: Cas-Sensing provides a general probabilistic paradigm for multi-scale field reconstruction that explicitly handles uncertainty and decomposes ill-posed problems into better-conditioned subproblems through cascaded inference.

Abstract: Reconstructing full fields from extremely sparse and random measurements constitutes a fundamentally ill-posed inverse problem, in which deterministic end-to-end mappings often break down due to intrinsic non-uniqueness and uncertainty. Rather than treating sparse reconstruction as a regression task, we recast it as a hierarchical probabilistic inference problem, where uncertainty is explicitly represented, structured, and progressively resolved. From this perspective, we propose Cascaded Sensing (Cas-Sensing) as a general reconstruction paradigm for multi-scale physical fields under extreme data sparsity. Central to this paradigm is the introduction of an explicit intermediate representation that decomposes the original ill-posed problem into two substantially better-conditioned subproblems. First, a lightweight neural-operator-based functional autoencoder infers a coarse-scale approximation of the target field from sparse observations acting as an explicit intermediate variable. Rather than modeling multiple scales jointly, this intermediate estimate is deterministically fixed and subsequently used as the sole conditioning input to a conditional diffusion model that generates refined-scale details, yielding a cascaded inference structure with clearly separated reconstruction responsibilities. To ensure robustness under diverse sensing patterns, the diffusion model is trained using a mask-cascade strategy, which exposes it to a distribution of imperfect conditioning structures induced by extreme sparsity. During inference, measurement consistency is enforced through manifold-constrained gradients within a Bayesian posterior framework, ensuring fidelity to sparse observations while preserving data manifold coherence. This cascaded probabilistic formulation substantially alleviates ill-posedness, enabling accurate and stable reconstructions even under extreme sparsity.

Method: Interpret graphs as walk-summable Gaussian graphical models, compute transformations via Gaussian Belief Propagation to aggregate information from both direct and indirect neighbors without explicit enumeration of multi-hop paths. Different precision matrix constructions induce distinct propagation biases.

Result: Graph Linear Transformations achieve competitive or superior performance compared to both local message-passing GNNs and dynamic neighborhood aggregation models across homophilic and heterophilic benchmark datasets.

Conclusion: GLT provides an efficient alternative to traditional GNN architectures by enabling direct and indirect feature mixing through a single well-defined linear operator, offering interpretable propagation biases and strong performance across diverse graph types.

Abstract: Graph neural networks (GNN) typically rely on localized message passing, requiring increasing depth to capture long range dependencies. In this work, we introduce Graph Linear Transformations, a linear transformation that realizes direct and indirect feature mixing on graphs through a single, well-defined linear operator derived from the graph structure. By interpreting graphs as walk-summable Gaussian graphical models, we compute these transformations via Gaussian Belief Propagation, enabling each node to aggregate information from both direct and indirect neighbors without explicit enumeration of multi-hop paths. We show that different constructions of the underlying precision matrix induce distinct and interpretable propagation biases, ranging from selective edge-level interactions to uniform structural smoothing, and that Graph Linear Transformations can achieve competitive or superior performance compared to both local message-passing GNNs and dynamic neighborhood aggregation models across homophilic and heterophilic benchmark datasets.

Method: Theoretical analysis shows global error is linearly bounded by local reconstruction errors for Lipschitz networks. Proposes Subset Fidelity metric to quantify component importance via local reconstruction behavior. Introduces ModHiFi algorithm that uses Subset Fidelity for model modification without training data or loss function.

Result: ModHiFi-P achieves 11% speedup over SOTA on ImageNet models with competitive language model performance. ModHiFi-U achieves complete unlearning on CIFAR-10 without fine-tuning and competitive performance on Swin Transformers.

Conclusion: ModHiFi provides effective model modification without requiring training data or loss function access, addressing key limitations in open weight model adaptation for pruning and unlearning tasks.

Abstract: Open weight models, which are ubiquitous, rarely provide access to their training data or loss function. This makes modifying such models for tasks such as pruning or unlearning, which are constrained by this unavailability, an active area of research. Existing techniques typically require gradients or ground-truth labels, rendering them infeasible in settings with limited computational resources. In this work, we investigate the fundamental question of identifying components that are critical to the model’s predictive performance, without access to either gradients or the loss function, and with only distributional access such as synthetic data. We theoretically demonstrate that the global error is linearly bounded by local reconstruction errors for Lipschitz-continuous networks such as CNNs and well-trained Transformers (which, contrary to existing literature, we find exhibit Lipschitz continuity). This motivates using the locally reconstructive behavior of component subsets to quantify their global importance, via a metric that we term Subset Fidelity. In the uncorrelated features setting, selecting individual components based on their Subset Fidelity scores is optimal, which we utilize to propose ModHiFi, an algorithm for model modification that requires neither training data nor access to a loss function. ModHiFi-P, for structured pruning, achieves an 11% speedup over the current state of the art on ImageNet models and competitive performance on language models. ModHiFi-U, for classwise unlearning, achieves complete unlearning on CIFAR-10 without fine-tuning and demonstrates competitive performance on Swin Transformers.

Method: Integrates EAGLE-I outage records (2014-2024) with weather, socioeconomic, infrastructure, and seasonal event data. Evaluates four machine learning models: Random Forest, Graph Neural Network, Adaptive Boosting, and Long Short-Term Memory networks.

Result: Experimental validation on Michigan counties shows LSTM network achieves the highest accuracy among all tested models for predicting power outages during extreme conditions.

Conclusion: The proposed learning-based framework effectively predicts power outages from extreme events, with LSTM performing best, and incorporating social/demographic indicators improves understanding of community vulnerability and outage risk.

Abstract: This paper presents a novel learning based framework for predicting power outages caused by extreme events. The proposed approach targets low probability high consequence outage scenarios and leverages a comprehensive set of features derived from publicly available data sources. We integrate EAGLE-I outage records from 2014 to 2024 with weather, socioeconomic, infrastructure, and seasonal event data. Incorporating social and demographic indicators reveals patterns of community vulnerability and improves understanding of outage risk during extreme conditions. Four machine learning models are evaluated including Random Forest (RF), Graph Neural Network (GNN), Adaptive Boosting (AdaBoost), and Long Short Term Memory (LSTM). Experimental validation is performed on a large scale dataset covering counties in the lower peninsula of Michigan. Among all models tested, the LSTM network achieves higher accuracy.

Method: Propose entropy-based retraining framework grounded in nonequilibrium statistical physics. Interpret drift as probability flow via Fokker-Planck equation, quantify model-data mismatch using relative entropy, and implement entropy-triggered retraining using EWMA control statistic on streaming kernel density estimator of KL divergence.

Result: In synthetic, financial, and web-traffic domains, entropy-based retraining achieves comparable predictive performance to frequent retraining while reducing retraining frequency by 1-2 orders of magnitude. However, in biomedical ECG setting, it underperforms maximum-frequency baseline due to limitations with complex label-conditional drift.

Conclusion: Entropy-based retraining provides effective cost-performance tradeoff for many nonstationary environments, but has limitations for complex label-conditional drift scenarios where feature-space entropy monitoring may be insufficient.

Abstract: Machine learning models deployed in nonstationary environments inevitably experience performance degradation due to data drift. While numerous drift detection heuristics exist, most lack a dynamical interpretation and provide limited guidance on how retraining decisions should be balanced against operational cost. In this work, we propose an entropy-based retraining framework grounded in nonequilibrium statistical physics. Interpreting drift as probability flow governed by a Fokker-Planck equation, we quantify model-data mismatch using relative entropy and show that its time derivative admits an entropy-balance decomposition featuring a nonnegative entropy production term driven by probability currents. Guided by this theory, we implement an entropy-triggered retraining policy using an exponentially weighted moving-average (EWMA) control statistic applied to a streaming kernel density estimator of the Kullback-Leibler divergence. We evaluate this approach across multiple nonstationary data streams. In synthetic, financial, and web-traffic domains, entropy-based retraining achieves predictive performance comparable to frequent retraining while reducing retraining frequency by one to two orders of magnitude. However, in a challenging biomedical ECG setting, the entropy-based trigger underperforms the maximum-frequency baseline, highlighting limitations of feature-space entropy monitoring under complex label-conditional drift.

Method: BSZO applies Kalman filtering to combine finite-difference measurements across multiple perturbation directions within a subspace, treating each measurement as a noisy observation. It builds a posterior distribution over the subspace-projected gradient through Bayesian inference with residual-based adaptive mechanism to handle noise variations.

Result: Theoretical analysis shows BSZO improves convergence rate by factor k/γ compared to standard ZO methods. Experiments on RoBERTa, Mistral, and OPT models show BSZO outperforms baselines across tasks, achieving up to 6.67% absolute average improvement on OPT-13B while remaining robust under fp16/bf16 precision with memory usage close to inference-only baselines (1.00×-1.08× of MeZO).

Conclusion: BSZO provides an effective Bayesian subspace approach for zeroth-order optimization that significantly improves performance and robustness for LLM fine-tuning while maintaining low memory overhead.

Abstract: Fine-tuning large language models (LLMs) with zeroth-order (ZO) optimization reduces memory by approximating gradients through function evaluations. However, existing methods essentially perform updates in a one-dimensional space, and suffer from collapse or substantial performance degradation under low-precision training. We introduce BSZO, an adaptive \textbf{B}ayesian \textbf{S}ubspace \textbf{Z}eroth-Order \textbf{O}ptimizer, which applies Kalman filtering to combine finite-difference information across multiple perturbation directions within a subspace. By treating each finite-difference measurement as a noisy observation, BSZO builds a posterior distribution over the subspace-projected gradient and updates it through Bayesian inference, with a residual-based adaptive mechanism to adapt to noise variations. Theoretical analysis shows that BSZO improves the convergence rate by a factor of $k/γ$ compared to standard ZO methods. Experiments on RoBERTa, Mistral, and OPT models show that BSZO outperforms the baselines across various tasks, achieving up to 6.67% absolute average improvement on OPT-13B while remaining robust under fp16/bf16 precision and keeping memory usage close to inference-only baselines (1.00$\times$–1.08$\times$ of MeZO).

Method: The method extends existing sLID (spatial LID) technique with three key enhancements: (1) kinematic enhancement by incorporating velocity into sLID computation, (2) spatial fusion using Bayesian estimation to aggregate sLID values across neighborhoods, (3) temporal modeling with tLID that learns long-term dynamics from time series data, integrated into a unified stLID framework.

Result: Extensive experiments show that stLID consistently outperforms existing methods in both failure detection precision and lead-time for identifying landslide failure zones.

Conclusion: The proposed stLID framework effectively addresses the limitations of existing methods by jointly incorporating spatial and temporal information, enabling more accurate and early detection of complex landslides and multiple successive failures in distinct slope areas.

Abstract: Local Intrinsic Dimensionality (LID) has shown strong potential for identifying anomalies and outliers in high-dimensional data across a wide range of real-world applications, including landslide failure detection in granular media. Early and accurate identification of failure zones in landslide-prone areas is crucial for effective geohazard mitigation. While existing approaches typically rely on surface displacement data analyzed through statistical or machine learning techniques, they often fall short in capturing both the spatial correlations and temporal dynamics that are inherent in such data. To address this gap, we focus on ground-monitored landslides and introduce a novel approach that jointly incorporates spatial and temporal information, enabling the detection of complex landslides and including multiple successive failures occurring in distinct areas of the same slope. To be specific, our method builds upon an existing LID-based technique, known as sLID. We extend its capabilities in three key ways. (1) Kinematic enhancement: we incorporate velocity into the sLID computation to better capture short-term temporal dependencies and deformation rate relationships. (2) Spatial fusion: we apply Bayesian estimation to aggregate sLID values across spatial neighborhoods, effectively embedding spatial correlations into the LID scores. (3) Temporal modeling: we introduce a temporal variant, tLID, that learns long-term dynamics from time series data, providing a robust temporal representation of displacement behavior. Finally, we integrate both components into a unified framework, referred to as spatiotemporal LID (stLID), to identify samples that are anomalous in either or both dimensions. Extensive experiments show that stLID consistently outperforms existing methods in failure detection precision and lead-time.

Method: Two-stage RL framework: Stage 1 rewards local token swaps to repair syntax; Stage 2 provides chemistry-aware feedback from RDKit diagnostics to reduce valence, aromaticity, and connectivity issues. Uses GRU policy with PPO, evaluated in both pure RL (from random) and fine-tuning RL (from pretrained model).

Result: In pure RL, TSSR significantly improves syntactic validity, chemical validity, and novelty. In fine-tuning RL, it preserves drug-likeness and synthesizability while increasing validity and novelty. Token analysis shows syntax edits and chemistry fixes jointly reduce RDKit errors.

Conclusion: TSSR converts sparse terminal objectives into denser, interpretable rewards, improving both syntactic and chemical quality without reducing diversity. The framework is dataset-agnostic and adaptable to various RL approaches.

Abstract: The design of reliable, valid, and diverse molecules is fundamental to modern drug discovery, as improved molecular generation supports efficient exploration of the chemical space for potential drug candidates and reduces the cost of early design efforts. Despite these needs, current chemical language models that generate molecules as SMILES strings are vulnerable to compounding token errors: many samples are unparseable or chemically implausible, and hard constraints meant to prevent failure can restrict exploration. To address this gap, we introduce TSSR, a Two-Stage, Swap-Reward-driven reinforcement learning (RL) framework for character-level SMILES generation. Stage one rewards local token swaps that repair syntax, promoting transitions from invalid to parseable strings. Stage two provides chemistry-aware feedback from RDKit diagnostics, rewarding reductions in valence, aromaticity, and connectivity issues. The reward decomposes into interpretable terms (swap efficiency, error reduction, distance to validity), is model agnostic, and requires no task-specific labels or hand-crafted grammars. We evaluated TSSR on the MOSES benchmark using a GRU policy trained with PPO in both pure RL (P-RL) from random initialization and fine-tuning RL (F-RL) starting from a pretrained chemical language model, assessing 10,000 generated SMILES per run. In P-RL, TSSR significantly improves syntactic validity, chemical validity, and novelty. In F-RL, TSSR preserves drug-likeness and synthesizability while increasing validity and novelty. Token-level analysis shows that syntax edits and chemistry fixes act jointly to reduce RDKit detected errors. TSSR converts a sparse terminal objective into a denser and more interpretable reward, improving both syntactic and chemical quality without reducing diversity. TSSR is dataset-agnostic and can be adapted to various reinforcement learning approaches.

Method: 1) Theoretical analysis showing ℓ₁-regularized SAEs are intrinsically biased toward low-dimensional patterns. 2) Falsification-oriented evaluation combining causal token injection and LLM-guided falsification to test feature activation. 3) Testing across 20 configurations spanning multiple model families, layers, and reasoning datasets.

Result: 45-90% of contrastive features activate when associated tokens are injected into non-reasoning text. Remaining features can be activated by non-reasoning inputs and deactivated by reasoning inputs. No analyzed feature satisfied criteria for genuine reasoning behavior. Steering these features yields no benchmark performance improvements.

Conclusion: SAE features identified by current contrastive approaches primarily capture linguistic correlates of reasoning rather than underlying reasoning computations themselves.

Abstract: We investigate whether sparse autoencoders (SAEs) identify genuine reasoning features in large language models (LLMs). We first show through a simple theoretical analysis that $\ell_1$-regularized SAEs are intrinsically biased toward low-dimensional patterns, providing a mechanistic explanation for why shallow linguistic cues may be preferentially captured over distributed reasoning behaviors. Motivated by this bias, we introduce a falsification-oriented evaluation framework that combines causal token injection and LLM-guided falsification to test whether feature activation reflects reasoning processes or superficial linguistic correlates. Across 20 configurations spanning multiple model families, layers, and reasoning datasets, we find that features identified by contrastive methods are highly sensitive to token-level interventions, with 45% to 90% activating when a small number of associated tokens are injected into non-reasoning text. For the remaining features, LLM-guided falsification consistently produces non-reasoning inputs that activate the feature and reasoning inputs that do not, with no analyzed feature satisfying our criteria for genuine reasoning behavior. Steering these features yields no improvements in benchmark performance. Overall, our results suggest that SAE features identified by current contrastive approaches primarily capture linguistic correlates of reasoning rather than the underlying reasoning computations themselves. Code is available at https://github.com/GeorgeMLP/reasoning-probing.

Method: SPIKE regularizes PINNs with continuous-time Koopman operators to learn parsimonious dynamics representations. It enforces linear dynamics dz/dt = Az in a learned observable space, with L1 regularization on A to promote sparsity (PIKE without sparsity, SPIKE with L1). Uses matrix exponential integration for unconditional stability.

Result: Experiments across parabolic, hyperbolic, dispersive, and stiff PDEs (including Navier-Stokes) and chaotic ODEs (Lorenz) show consistent improvements in temporal extrapolation, spatial generalization, and long-term prediction accuracy.

Conclusion: SPIKE framework successfully addresses PINN overfitting by learning sparse Koopman representations, enabling better generalization and long-term predictions while maintaining stability for stiff systems.

Abstract: Physics-Informed Neural Networks (PINNs) provide a mesh-free approach for solving differential equations by embedding physical constraints into neural network training. However, PINNs tend to overfit within the training domain, leading to poor generalization when extrapolating beyond trained spatiotemporal regions. This work presents SPIKE (Sparse Physics-Informed Koopman-Enhanced), a framework that regularizes PINNs with continuous-time Koopman operators to learn parsimonious dynamics representations. By enforcing linear dynamics $dz/dt = Az$ in a learned observable space, both PIKE (without explicit sparsity) and SPIKE (with L1 regularization on $A$) learn sparse generator matrices, embodying the parsimony principle that complex dynamics admit low-dimensional structure. Experiments across parabolic, hyperbolic, dispersive, and stiff PDEs, including fluid dynamics (Navier-Stokes) and chaotic ODEs (Lorenz), demonstrate consistent improvements in temporal extrapolation, spatial generalization, and long-term prediction accuracy. The continuous-time formulation with matrix exponential integration provides unconditional stability for stiff systems while avoiding diagonal dominance issues inherent in discrete-time Koopman operators.

Method: Two-agent architecture: 1) Dimension Generating Agent (DGA) - fine-tuned LLM that uses explicit user data to generate implicit dimensions/knowledge gaps; 2) Response Generating Agent (RGA) - balances explicit and implicit dimensions with selective filtering (quality, diversity, task relevance) to create personalized proactive responses.

Result: ProPer improves quality scores and win rates across multiple domains, achieving up to 84% gains in single-turn evaluation and consistent dominance in multi-turn interactions, measured by coverage, initiative appropriateness, and intent alignment.

Conclusion: ProPer successfully addresses the limitations of reactive assistants by proactively identifying and addressing implicit user needs through a structured two-agent approach, demonstrating significant improvements in personalized assistance quality.

Abstract: Most language-based assistants follow a reactive ask-and-respond paradigm, requiring users to explicitly state their needs. As a result, relevant but unexpressed needs often go unmet. Existing proactive agents attempt to address this gap either by eliciting further clarification, preserving this burden, or by extrapolating future needs from context, often leading to unnecessary or mistimed interventions. We introduce ProPer, Proactivity-driven Personalized agents, a novel two-agent architecture consisting of a Dimension Generating Agent (DGA) and a Response Generating Agent (RGA). DGA, a fine-tuned LLM agent, leverages explicit user data to generate multiple implicit dimensions (latent aspects relevant to the user’s task but not considered by the user) or knowledge gaps. These dimensions are selectively filtered using a reranker based on quality, diversity, and task relevance. RGA then balances explicit and implicit dimensions to tailor personalized responses with timely and proactive interventions. We evaluate ProPer across multiple domains using a structured, gap-aware rubric that measures coverage, initiative appropriateness, and intent alignment. Our results show that ProPer improves quality scores and win rates across all domains, achieving up to 84% gains in single-turn evaluation and consistent dominance in multi-turn interactions.

Method: 1) Created novel NorthEast monsoon climate index from sea surface temperature data; 2) Used Deep Q-Network reinforcement learning to optimize calculation areas by selecting rectangles with highest correlation to seasonal rainfall; 3) Classified rainfall stations into 12 clusters to distinguish regional patterns; 4) Incorporated optimized index into LSTM models for monthly rainfall prediction.

Result: The optimized index significantly improves long-term monthly rainfall prediction skill in most cluster areas. Most importantly, it effectively reduces Root Mean Square Error for 12-month-ahead forecasts, demonstrating practical value for long-term climate prediction.

Conclusion: The reinforcement learning-optimized local climate index approach successfully enhances rainfall prediction accuracy in Thailand, providing a valuable framework for developing region-specific climate indices that outperform global indices for local forecasting needs.

Abstract: Climate prediction is a challenge due to the intricate spatiotemporal patterns within Earth systems. Global climate indices, such as the El Niño Southern Oscillation, are standard input features for long-term rainfall prediction. However, a significant gap persists regarding local-scale indices capable of improving predictive accuracy in specific regions of Thailand. This paper introduces a novel NorthEast monsoon climate index calculated from sea surface temperature to reflect the climatology of the boreal winter monsoon. To optimise the calculated areas used for this index, a Deep Q-Network reinforcement learning agent explores and selects the most effective rectangles based on their correlation with seasonal rainfall. Rainfall stations were classified into 12 distinct clusters to distinguish rainfall patterns between southern and upper Thailand. Experimental results show that incorporating the optimised index into Long Short-Term Memory models significantly improves long-term monthly rainfall prediction skill in most cluster areas. This approach effectively reduces the Root Mean Square Error for 12-month-ahead forecasts.

Method: Formulates trajectory planning as a Decentralized Partially Observable Semi-Markov Decision Process (Dec-POSMDP) and introduces an asynchronous multi-agent learning algorithm. Once trajectory policies are learned, bandwidth allocation is optimally solved based on local observations at each collection point.

Result: The proposed method demonstrates superiority over other learning-based and heuristic baselines in terms of both energy efficiency and mission completion time. The learned policies also exhibit robustness under varying environmental conditions.

Conclusion: The asynchronous multi-agent learning approach effectively addresses the practical challenge of action synchronization in cooperative UAV data collection, achieving better energy efficiency and mission completion while maintaining robustness to environmental variations.

Abstract: This paper addresses the joint optimization of trajectories and bandwidth allocation for multiple Unmanned Aerial Vehicles (UAVs) to enhance energy efficiency in the cooperative data collection problem. We focus on an important yet underestimated aspect of the system, where action synchronization across all UAVs is impossible. Since most existing learning-based solutions are not designed to learn in this asynchronous environment, we formulate the trajectory planning problem as a Decentralized Partially Observable Semi-Markov Decision Process and introduce an asynchronous multi-agent learning algorithm to learn UAVs’ cooperative policies. Once the UAVs’ trajectory policies are learned, the bandwidth allocation can be optimally solved based on local observations at each collection point. Comprehensive empirical results demonstrate the superiority of the proposed method over other learning-based and heuristic baselines in terms of both energy efficiency and mission completion time. Additionally, the learned policies exhibit robustness under varying environmental conditions.

Method: Used Qwen3 models (4B-32B) within an adapted Agentic-Reasoning framework, systematically testing combinations of model scale, explicit thinking strategies (no thinking, planner-only, full thinking), and tool use (search, code, mind-map).

Result: Tool augmentation provided the largest and most consistent performance gains - 4B models with tools outperformed 32B models without tool access. Explicit thinking showed mixed results: planner-only thinking improved decomposition and constraint tracking, while full thinking often degraded performance by causing tool orchestration issues.

Conclusion: Tool augmentation is more effective than model scaling for improving performance on complex reasoning tasks like GAIA, while explicit thinking strategies require careful configuration to avoid destabilizing tool orchestration and verification processes.

Abstract: This report studies whether small, tool-augmented agents can match or outperform larger monolithic models on the GAIA benchmark. Using Qwen3 models (4B-32B) within an adapted Agentic-Reasoning framework, we isolate the effects of model scale, explicit thinking (no thinking, planner-only, or full), and tool use (search, code, mind-map). Tool augmentation provides the largest and most consistent gains. Using tools, 4B models can outperform 32B models without tool access on GAIA in our experimental setup. In contrast, explicit thinking is highly configuration- and difficulty-dependent: planner-only thinking can improve decomposition and constraint tracking, while unrestricted full thinking often degrades performance by destabilizing tool orchestration, leading to skipped verification steps, excessive tool calls, non-termination, and output-format drift.

Method: EvidFuse uses a training-free multi-agent framework with two collaborating components: 1) Data-Augmented Analysis Agent with EDA knowledge and raw table access, and 2) Real-Time Evidence Construction Writer that plans outlines and drafts reports while issuing fine-grained analysis requests.

Result: EvidFuse achieves top rank in both LLM-as-a-judge and human evaluations on chart quality, chart-text alignment, and report-level usefulness.

Conclusion: The framework enables writing-time text-chart interleaved generation, allowing visual evidence to be constructed exactly when needed, directly constraining claims and enabling on-demand expansion of evidence space for deeper analysis.

Abstract: Data-driven reports communicate decision-relevant insights by tightly interleaving narrative text with charts grounded in underlying tables. However, current LLM-based systems typically generate narratives and visualizations in staged pipelines, following either a text-first-graph-second or a graph-first-text-second paradigm. These designs often lead to chart-text inconsistency and insight freezing, where the intermediate evidence space becomes fixed and the model can no longer retrieve or construct new visual evidence as the narrative evolves, resulting in shallow and predefined analysis. To address the limitations, we propose \textbf{EvidFuse}, a training-free multi-agent framework that enables writing-time text-chart interleaved generation for data-driven reports. EvidFuse decouples visualization analysis from long-form drafting via two collaborating components: a \textbf{Data-Augmented Analysis Agent}, equipped with Exploratory Data Analysis (EDA)-derived knowledge and access to raw tables, and a \textbf{Real-Time Evidence Construction Writer} that plans an outline and drafts the report while intermittently issuing fine-grained analysis requests. This design allows visual evidence to be constructed and incorporated exactly when the narrative requires it, directly constraining subsequent claims and enabling on-demand expansion of the evidence space. Experiments demonstrate that EvidFuse attains the top rank in both LLM-as-a-judge and human evaluations on chart quality, chart-text alignment, and report-level usefulness.

Method: DAGR-VQA integrates learnable register tokens directly into a convolutional backbone as global context carriers. This enables dynamic, HVS-inspired attention that produces temporally adaptive saliency maps. The model combines dynamic saliency maps with RGB inputs, capturing spatial data and analyzing it through a temporal transformer for perceptually consistent quality assessment.

Result: Comprehensive tests on LSVQ, KonVid-1k, LIVE-VQC, and YouTube-UGC datasets show highly competitive performance, surpassing most top baselines. The model achieves 387.7 FPS at 1080p, making it suitable for real-time applications like multimedia streaming systems. Ablation studies confirm that register tokens promote stable and temporally consistent attention mechanisms.

Conclusion: DAGR-VQA successfully integrates register tokens into a convolutional backbone for dynamic spatio-temporal saliency prediction, creating an efficient NR-VQA framework that outperforms existing methods while maintaining real-time computational performance for practical applications.

Abstract: No-reference video quality assessment (NR-VQA) estimates perceptual quality without a reference video, which is often challenging. While recent techniques leverage saliency or transformer attention, they merely address global context of the video signal by using static maps as auxiliary inputs rather than embedding context fundamentally within feature extraction of the video sequence. We present Dynamic Attention with Global Registers for Video Quality Assessment (DAGR-VQA), the first framework integrating register-token directly into a convolutional backbone for spatio-temporal, dynamic saliency prediction. By embedding learnable register tokens as global context carriers, our model enables dynamic, HVS-inspired attention, producing temporally adaptive saliency maps that track salient regions over time without explicit motion estimation. Our model integrates dynamic saliency maps with RGB inputs, capturing spatial data and analyzing it through a temporal transformer to deliver a perceptually consistent video quality assessment. Comprehensive tests conducted on the LSVQ, KonVid-1k, LIVE-VQC, and YouTube-UGC datasets show that the performance is highly competitive, surpassing the majority of top baselines. Research on ablation studies demonstrates that the integration of register tokens promotes the development of stable and temporally consistent attention mechanisms. Achieving an efficiency of 387.7 FPS at 1080p, DAGR-VQA demonstrates computational performance suitable for real-time applications like multimedia streaming systems.

Method: Two-stage approach: 1) Use Vector Quantized Generative Adversarial Network to capture HDRTV priors from real HDRTV content, 2) Match these priors to input SDRTV content to recover realistic HDRTV outputs. This transforms the problem from unreferenced prediction to referenced selection.

Result: Method evaluated on public datasets shows significant improvements in both objective and subjective metrics across real and synthetic datasets, demonstrating effectiveness of using HDRTV priors to constrain the solution space.

Conclusion: Introducing real HDRTV as reference priors significantly constrains the ill-posed SDRTV-to-HDRTV conversion problem, transforming it from prediction to selection and enhancing accuracy and reliability of the conversion process.

Abstract: The rise of HDR-WCG display devices has highlighted the need to convert SDRTV to HDRTV, as most video sources are still in SDR. Existing methods primarily focus on designing neural networks to learn a single-style mapping from SDRTV to HDRTV. However, the limited information in SDRTV and the diversity of styles in real-world conversions render this process an ill-posed problem, thereby constraining the performance and generalization of these methods. Inspired by generative approaches, we propose a novel method for SDRTV to HDRTV conversion guided by real HDRTV priors. Despite the limited information in SDRTV, introducing real HDRTV as reference priors significantly constrains the solution space of the originally high-dimensional ill-posed problem. This shift transforms the task from solving an unreferenced prediction problem to making a referenced selection, thereby markedly enhancing the accuracy and reliability of the conversion process. Specifically, our approach comprises two stages: the first stage employs a Vector Quantized Generative Adversarial Network to capture HDRTV priors, while the second stage matches these priors to the input SDRTV content to recover realistic HDRTV outputs. We evaluate our method on public datasets, demonstrating its effectiveness with significant improvements in both objective and subjective metrics across real and synthetic datasets.

Method: Provides a faithful implementation of CTS metric using skeleton-based matching framework, with optional preprocessing extensions to handle common prediction artifacts (small holes, edge noise). Extensions are disabled by default for strict comparability.

Result: Implementation supports PyTorch-based workflows and includes visualization tools for transparency. Code and archival resources will be made available on GitHub.

Conclusion: The paper presents a reliable implementation of CTS that enables proper evaluation of topological correctness in crack segmentation while maintaining compatibility with original metric definition.

Abstract: The Crack Topology Score (CTS) is a recently proposed metric that focuses on evaluating the topological correctness of crack segmentation outputs. While pixel-wise metrics such as IoU or F1-score fail to capture structural validity, CTS offers a skeleton-based matching framework to measure the preservation of connectivity. This paper presents a faithful implementation of the CTS metric, along with optional preprocessing extensions designed to handle common prediction artifacts (e.g., small holes and edge noise) found in deep learning outputs. All extensions are disabled by default to ensure strict comparability with the original definition. The implementation supports PyTorch-based workflows and includes visualization tools for transparency. Code and archival resources will be made available at https://github.com/SH-Joo/crack-topology-score.

Method: Used LIDC-IDRI dataset chest CT images, extracted ROI around pulmonary nodules, defined findings/questions based on morphological characteristics in database, constructed VQA dataset, and fine-tuned VQA model on it.

Result: Created VQA dataset with natural radiological descriptions, achieved high CIDEr score of 3.896 for generated findings, and obtained high agreement with reference findings based on morphological characteristics.

Conclusion: The proposed method effectively enables interactive diagnostic support that can present image findings according to physicians’ interests, as demonstrated by generated results and evaluation metrics.

Abstract: Interpretation of imaging findings based on morphological characteristics is important for diagnosing pulmonary nodules on chest computed tomography (CT) images. In this study, we constructed a visual question answering (VQA) dataset from structured data in an open dataset and investigated an image-finding generation method for chest CT images, with the aim of enabling interactive diagnostic support that presents findings based on questions that reflect physicians’ interests rather than fixed descriptions. In this study, chest CT images included in the Lung Image Database Consortium and Image Database Resource Initiative (LIDC-IDRI) datasets were used. Regions of interest surrounding the pulmonary nodules were extracted from these images, and image findings and questions were defined based on morphological characteristics recorded in the database. A dataset comprising pairs of cropped images, corresponding questions, and image findings was constructed, and the VQA model was fine-tuned on it. Language evaluation metrics such as BLEU were used to evaluate the generated image findings. The VQA dataset constructed using the proposed method contained image findings with natural expressions as radiological descriptions. In addition, the generated image findings showed a high CIDEr score of 3.896, and a high agreement with the reference findings was obtained through evaluation based on morphological characteristics. We constructed a VQA dataset for chest CT images using structured information on the morphological characteristics from the LIDC-IDRI dataset. Methods for generating image findings in response to these questions have also been investigated. Based on the generated results and evaluation metric scores, the proposed method was effective as an interactive diagnostic support system that can present image findings according to physicians’ interests.

Method: Used LIDC-IDRI dataset to create nodule image-text prompt pairs based on physician evaluations. Fine-tuned Stable Diffusion v1.5 and v2.0 models. Adjusted guidance scale (GS) parameter to control text fidelity during generation.

Result: SDv2 with GS=5 performed best in quantitative and subjective evaluations. Generated images showed high quality, diversity, and text consistency. No statistically significant differences between generated and real images in subjective evaluation.

Conclusion: The proposed LDM-based method successfully generates high-quality chest CT nodule images that capture specific medical features, providing a solution for data scarcity in medical imaging applications.

Abstract: Recently, computer-aided diagnosis systems have been developed to support diagnosis, but their performance depends heavily on the quality and quantity of training data. However, in clinical practice, it is difficult to collect the large amount of CT images for specific cases, such as small cell carcinoma with low epidemiological incidence or benign tumors that are difficult to distinguish from malignant ones. This leads to the challenge of data imbalance. In this study, to address this issue, we proposed a method to automatically generate chest CT nodule images that capture target features using latent diffusion models (LDM) and verified its effectiveness. Using the LIDC-IDRI dataset, we created pairs of nodule images and finding-based text prompts based on physician evaluations. For the image generation models, we used Stable Diffusion version 1.5 (SDv1) and 2.0 (SDv2), which are types of LDM. Each model was fine-tuned using the created dataset. During the generation process, we adjusted the guidance scale (GS), which indicates the fidelity to the input text. Both quantitative and subjective evaluations showed that SDv2 (GS = 5) achieved the best performance in terms of image quality, diversity, and text consistency. In the subjective evaluation, no statistically significant differences were observed between the generated images and real images, confirming that the quality was equivalent to real clinical images. We proposed a method for generating chest CT nodule images based on input text using LDM. Evaluation results demonstrated that the proposed method could generate high-quality images that successfully capture specific medical features.

Method: Hybrid CNN-LSTM framework where CNN extracts spatial features from epidemiological data and LSTM models temporal patterns, combined with hybrid Whale Optimization Algorithm (WOA) and Gray Wolf Optimization (GWO) for hyperparameter tuning of learning rates, batch sizes, and training epochs.

Result: Applied to COVID-19 case data from 24 countries across six continents, the method outperformed established benchmarks (ARIMA and standalone LSTM) with statistically significant gains in predictive accuracy, including reduced RMSE.

Conclusion: The framework demonstrates potential as a versatile method for forecasting epidemic trends, offering valuable insights for resource planning and decision-making in both historical contexts like COVID-19 and future outbreaks.

Abstract: Effective epidemic modeling is essential for managing public health crises, requiring robust methods to predict disease spread and optimize resource allocation. This study introduces a novel deep learning framework that advances time series forecasting for infectious diseases, with its application to COVID 19 data as a critical case study. Our hybrid approach integrates Convolutional Neural Networks (CNNs) and Long Short Term Memory (LSTM) models to capture spatial and temporal dynamics of disease transmission across diverse regions. The CNN extracts spatial features from raw epidemiological data, while the LSTM models temporal patterns, yielding precise and adaptable predictions. To maximize performance, we employ a hybrid optimization strategy combining the Whale Optimization Algorithm (WOA) and Gray Wolf Optimization (GWO) to fine tune hyperparameters, such as learning rates, batch sizes, and training epochs enhancing model efficiency and accuracy. Applied to COVID 19 case data from 24 countries across six continents, our method outperforms established benchmarks, including ARIMA and standalone LSTM models, with statistically significant gains in predictive accuracy (e.g., reduced RMSE). This framework demonstrates its potential as a versatile method for forecasting epidemic trends, offering insights for resource planning and decision making in both historical contexts, like the COVID 19 pandemic, and future outbreaks.

Method: Formulates the model using factor graphs with NUP (normal with unknown parameters) priors, where local parameters are piecewise constant and automatically adapted to each image. The computational approach involves iterations of conjugate-gradient steps and Gaussian message passing.

Result: The proposed model and algorithms are successfully demonstrated with applications to image denoising and contrast enhancement, showing the effectiveness of the adaptive piecewise smooth approach.

Conclusion: The piecewise smooth image model with adaptive local parameters using factor graphs and NUP priors provides an effective framework for image processing tasks like denoising and contrast enhancement, offering automatic adaptation to local image characteristics.

Abstract: We propose a novel piecewise smooth image model with piecewise constant local parameters that are automatically adapted to each image. Technically, the model is formulated in terms of factor graphs with NUP (normal with unknown parameters) priors, and the pertinent computations amount to iterations of conjugate-gradient steps and Gaussian message passing. The proposed model and algorithms are demonstrated with applications to denoising and contrast enhancement.