Method: Proposes CoBRA (Cross-modal Bottleneck for Robust AVSR), a bottleneck-based fusion framework that introduces a compact set of learnable tokens to mediate cross-modal exchange. These tokens regulate information flow, allowing the audio stream to access essential visual cues even under noisy conditions.

Result: Despite limited training data, CoBRA surpasses comparable baselines and remains competitive with large-scale systems through noise-adaptive fusion. Ablation studies show that the depth of fusion is the most critical factor for robust AVSR performance.

Conclusion: CoBRA demonstrates both efficiency and robustness in AVSR through its bottleneck-based fusion approach, highlighting the importance of fusion depth in designing robust multimodal speech recognition systems.

Abstract: Audio-Visual Speech Recognition (AVSR) leverages both acoustic and visual cues to improve speech recognition under noisy conditions. A central question is how to design a fusion mechanism that allows the model to effectively exploit visual information when the audio signal is degraded, while maintaining strong performance on clean speech. We propose CoBRA (Cross-modal Bottleneck for Robust AVSR), a bottleneck-based fusion framework that introduces a compact set of learnable tokens to mediate cross-modal exchange. By regulating information flow through these tokens, the audio stream can reliably access essential visual cues even under adverse or out-of-domain noise. Despite limited training data, our model surpasses comparable baselines and remains competitive with large-scale systems through noise-adaptive fusion, demonstrating both efficiency and robustness. Ablation studies highlight that the depth of fusion is the most critical factor, underscoring its importance in designing robust AVSR systems.

Method: 1) Create MENASpeechBank with 18K high-quality utterances from 124 speakers across MENA countries; 2) Develop controllable synthetic data pipeline: construct persona profiles with World Values Survey attributes, define 5K conversational scenarios, match personas to scenarios via semantic similarity, generate 417K role-play conversations with LLM, synthesize user turns by conditioning on reference speaker audio.

Result: Created MENASpeechBank dataset and generated 417K synthetic conversations. Evaluated both synthetic and human-recorded conversations with detailed analysis. Will release both resources publicly.

Conclusion: The paper addresses the data bottleneck for AudioLLMs by providing a high-quality speech dataset and scalable synthetic data generation pipeline, enabling better persona-grounded and dialectally diverse audio language models.

Abstract: Audio large language models (AudioLLMs) enable instruction-following over speech and general audio, but progress is increasingly limited by the lack of diverse, conversational, instruction-aligned speech-text data. This bottleneck is especially acute for persona-grounded interactions and dialectal coverage, where collecting and releasing real multi-speaker recordings is costly and slow. We introduce MENASpeechBank, a reference speech bank comprising about 18K high-quality utterances from 124 speakers spanning multiple MENA countries, covering English, Modern Standard Arabic (MSA), and regional Arabic varieties. Building on this resource, we develop a controllable synthetic data pipeline that: (i) constructs persona profiles enriched with World Values Survey-inspired attributes, (ii) defines a taxonomy of about 5K conversational scenarios, (iii) matches personas to scenarios via semantic similarity, (iv) generates about 417K role-play conversations with an LLM where the user speaks as the persona and the assistant behaves as a helpful agent, and (v) synthesizes the user turns by conditioning on reference speaker audio to preserve speaker identity and diversity. We evaluate both synthetic and human-recorded conversations and provide detailed analysis. We will release MENASpeechBank and the generated conversations publicly for the community.

Method: 1) Fixed-frame Modality Gap Theory decomposes gap into stable biases and anisotropic residuals. 2) ReAlign: training-free alignment strategy using massive unpaired data statistics with Anchor, Trace, and Centroid Alignment steps. 3) ReVision: scalable MLLM training paradigm integrating ReAlign into pretraining to learn visual representation distribution from unpaired text before visual instruction tuning.

Result: Framework demonstrates statistically aligned unpaired data can effectively substitute for expensive image-text pairs, offering robust path for efficient scaling of Multimodal Large Language Models.

Conclusion: Precise geometric characterization of modality gap enables efficient model scaling through statistical alignment of unpaired data, reducing reliance on large-scale, high-quality image-text pairs for MLLM development.

Abstract: Despite the success of multimodal contrastive learning in aligning visual and linguistic representations, a persistent geometric anomaly, the Modality Gap, remains: embeddings of distinct modalities expressing identical semantics occupy systematically offset regions. Prior approaches to bridge this gap are largely limited by oversimplified isotropic assumptions, hindering their application in large-scale scenarios. In this paper, we address these limitations by precisely characterizing the geometric shape of the modality gap and leveraging it for efficient model scaling. First, we propose the Fixed-frame Modality Gap Theory, which decomposes the modality gap within a frozen reference frame into stable biases and anisotropic residuals. Guided by this precise modeling, we introduce ReAlign, a training-free modality alignment strategy. Utilizing statistics from massive unpaired data, ReAlign aligns text representation into the image representation distribution via a three-step process comprising Anchor, Trace, and Centroid Alignment, thereby explicitly rectifying geometric misalignment. Building on ReAlign, we propose ReVision, a scalable training paradigm for Multimodal Large Language Models (MLLMs). ReVision integrates ReAlign into the pretraining stage, enabling the model to learn the distribution of visual representations from unpaired text before visual instruction tuning, without the need for large-scale, high-quality image-text pairs. Our framework demonstrates that statistically aligned unpaired data can effectively substitute for expensive image-text pairs, offering a robust path for the efficient scaling of MLLMs.

Method: Evaluated four Indonesian low-resource languages (Javanese, Balinese, Sundanese, Lampungnese) with non-Latin scripts using DualGPT architecture. Compared Llama 2 tokenizer vs custom tokenizer despite visual rendering of text as images.

Result: Reintegrating text tokenizers reintroduces tokenizer misalignment problems. Llama 2 tokenizer performed significantly worse than custom tokenizer despite lower OOV and fertility rates, with improvements up to 30.15 chrF++ for custom tokenizer.

Conclusion: Visual rendering doesn’t fully decouple from tokenization constraints when text tokenizers are reintroduced. Text tokenizers remain a significant barrier to equitable multimodal models, especially for low-resource languages.

Abstract: While pixel-based language modeling aims to bypass the sub-word tokenization bottleneck by rendering text as images, recent multimodal variants such as DualGPT reintroduce text tokenizers to improve autoregressive performance. We investigate a fundamental question, does visual rendering truly decouple a model from tokenization constraints? Focusing on four Indonesian low-resource local languages that have their own non-Latin scripts (i.e., Javanese, Balinese, Sundanese, and Lampungnese), we evaluate the impact of script-tokenizer alignment within the DualGPT architecture. Our results show that, despite visual rendering, reintegrating a text tokenizer into the architecture reintroduces the same issue that pixel-based language modeling aims to resolve, which is the tokenizer misalignment problem. Despite having lower OOV and fertility rates, we show that the Llama 2 tokenizer performs significantly worse than a custom tokenizer, with improvements of up to 30.15 chrF++. Our findings serve as a warning for future multimodal variants, as text tokenizers remain a significant barrier to equitable models.

Method: Developed a code-generating AI agent (BiomechAgent) that uses natural language processing to enable biomechanical analysis. Created a systematic benchmark for evaluation spanning data retrieval, visualization, activity classification, temporal segmentation, and clinical reasoning. Tested design decisions including biomechanically-informed domain-specific instructions vs generic prompts, integration of specialized tools for gait event detection, and comparison of local open-weight models vs frontier cloud-based LLMs.

Result: BiomechAgent achieved robust accuracy on data retrieval and visualization tasks and demonstrated emerging clinical reasoning capabilities. Domain-specific instructions significantly improved performance over generic prompts. Integration of specialized gait event detection tools substantially boosted accuracy on challenging spatiotemporal analysis. Local open-weight models performed substantially worse than frontier cloud-based LLMs in most domains except database retrieval.

Conclusion: BiomechAgent successfully makes motion capture data more useful and accessible to clinicians without programming expertise by enabling natural language-based biomechanical analysis. The system demonstrates the importance of domain-specific knowledge and specialized tool integration for effective AI agents in clinical biomechanics.

Abstract: Markerless motion capture is making quantitative movement analysis increasingly accessible, yet analyzing the resulting data remains a barrier for clinicians without programming expertise. We present BiomechAgent, a code-generating AI agent that enables biomechanical analysis through natural language and allows users to querying databases, generating visualizations, and even interpret data without requiring users to write code. To evaluate BiomechAgent’s capabilities, we developed a systematic benchmark spanning data retrieval, visualization, activity classification, temporal segmentation, and clinical reasoning. BiomechAgent achieved robust accuracy on data retrieval and visualization tasks and demonstrated emerging clinical reasoning capabilities. We used our dataset to systematically evaluate several of our design decisions. Biomechanically-informed, domain-specific instructions significantly improved performance over generic prompts, and integrating validated specialized tools for gait event detection substantially boosted accuracy on challenging spatiotemporal analysis where the base agent struggled. We also tested BiomechAgent using a local open-weight model instead of a frontier cloud based LLM and found that perform was substantially diminished in most domains other than database retrieval. In short, BiomechAgent makes the data from accessible motion capture and much more useful and accessible to end users.

Method: Proposes ILA-agent framework with behavioral primitives modeled as tools, allowing LLMs to incrementally explore, apply, and verify language knowledge through structured interactions with official documentation and execution environments.

Result: ILA-agent significantly outperforms retrieval-augmented baselines across code generation, translation, and program repair tasks on the Cangjie benchmark. Analysis reveals emergent behavior patterns but persisting performance gaps.

Conclusion: Inference-time language acquisition is a viable paradigm for LLMs to master unfamiliar programming languages without extensive fine-tuning, though challenges remain in bridging performance gaps.

Abstract: The proficiency of Large Language Models (LLMs) in coding tasks is often a reflection of their extensive pre-training corpora, which typically collapses when confronted with previously unfamiliar programming languages. Departing from data-intensive finetuning, we investigate the paradigm of Inference-time Language Acquisition (ILA), where an LLM masters an unfamiliar language through dynamic interaction with limited external resources. In this paper, we propose ILA-agent, a general ILA framework that equips LLMs with a set of behavioral primitives. By modeling essential human-like behaviors as a suite of tools, ILA-agent enables LLMs to incrementally explore, apply, and verify language knowledge through structured interactions with the official documentation and execution environment. To provide a rigorous evaluation in a low-resource setting, we construct Cangjie-bench, a multi-task benchmark based on the novel statically-typed language Cangjie. We instantiate ILA-agent for Cangjie and evaluate its performance across code generation, translation, and program repair tasks. Results using diverse LLMs demonstrate that ILA-agent significantly outperforms retrieval-augmented baselines. Further analysis of agent trajectories characterizes the emergent behavior patterns while highlighting persisting performance gaps.

Method: Proposes Anchored Decoding that adaptively allocates information budget over generation trajectory, enforcing per-step constraints for sequence-level guarantees. Also introduces AnchoredByte Decoding for byte-level cross-vocabulary fusion via ByteSampler framework, plus TinyComma 1.8B as a permissively trained safe model.

Result: Eliminates up to 75% of measurable copying gap between risky baseline and safe reference while preserving near-original fluency and factuality, with modest inference overhead. Defines new Pareto frontier for risk-utility trade-off across six model pairs.

Conclusion: Anchored Decoding provides effective plug-and-play solution for reducing verbatim copying in LMs without retraining, enabling practical compliance with copyright and privacy concerns through tunable risk-utility trade-offs.

Abstract: Modern language models (LMs) tend to memorize portions of their training data and emit verbatim spans. When the underlying sources are sensitive or copyright-protected, such reproduction raises issues of consent and compensation for creators and compliance risks for developers. We propose Anchored Decoding, a plug-and-play inference-time method for suppressing verbatim copying: it enables decoding from any risky LM trained on mixed-license data by keeping generation in bounded proximity to a permissively trained safe LM. Anchored Decoding adaptively allocates a user-chosen information budget over the generation trajectory and enforces per-step constraints that yield a sequence-level guarantee, enabling a tunable risk-utility trade-off. To make Anchored Decoding practically useful, we introduce a new permissively trained safe model (TinyComma 1.8B), as well as Anchored${\mathrm{Byte}}$ Decoding, a byte-level variant of our method that enables cross-vocabulary fusion via the ByteSampler framework (Hayase et al., 2025). We evaluate our methods across six model pairs on long-form evaluations of copyright risk and utility. Anchored and Anchored${\mathrm{Byte}}$ Decoding define a new Pareto frontier, preserving near-original fluency and factuality while eliminating up to 75% of the measurable copying gap (averaged over six copying metrics) between the risky baseline and a safe reference, at a modest inference overhead.

Method: Proposes Free Energy Mixer (FEM) that uses free-energy (log-sum-exp) reads with value-driven per-channel log-linear tilts applied to a fast prior (from queries/keys). It treats standard attention scoring as a prior and computes value-aware posterior reads. A two-level gated FEM is implemented that’s plug-and-play with standard/linear attention, linear RNNs, and SSMs.

Result: FEM consistently outperforms strong baselines on NLP, vision, and time-series tasks at matched parameter budgets, demonstrating improved performance across multiple domains.

Conclusion: FEM provides a principled way to enhance attention mechanisms with value-driven selection capabilities while maintaining computational efficiency, offering a promising direction for improving multimodal understanding architectures.

Abstract: Standard attention stores keys/values losslessly but reads them via a per-head convex average, blocking channel-wise selection. We propose the Free Energy Mixer (FEM): a free-energy (log-sum-exp) read that applies a value-driven, per-channel log-linear tilt to a fast prior (e.g., from queries/keys in standard attention) over indices. Unlike methods that attempt to improve and enrich the $(q,k)$ scoring distribution, FEM treats it as a prior and yields a value-aware posterior read at unchanged complexity, smoothly moving from averaging to per-channel selection as the learnable inverse temperature increases, while still preserving parallelism and the original asymptotic complexity ($O(T^2)$ for softmax; $O(T)$ for linearizable variants). We instantiate a two-level gated FEM that is plug-and-play with standard and linear attention, linear RNNs and SSMs. It consistently outperforms strong baselines on NLP, vision, and time-series at matched parameter budgets.

Method: Using small calibration datasets, identify activation signatures for different personas. Develop masking strategy to isolate lightweight persona subnetworks. For binary-opposing personas, introduce contrastive pruning to identify parameters responsible for statistical divergence between opposing personas. Entirely training-free approach.

Result: The resulting subnetworks exhibit significantly stronger persona alignment than baselines requiring external knowledge while being more efficient. Diverse human-like behaviors are already embedded in LLM parameter space.

Conclusion: LLMs contain embedded persona-specialized subnetworks that can be isolated without external knowledge, pointing toward new perspectives on controllable and interpretable personalization in large language models.

Abstract: Humans shift between different personas depending on social context. Large Language Models (LLMs) demonstrate a similar flexibility in adopting different personas and behaviors. Existing approaches, however, typically adapt such behavior through external knowledge such as prompting, retrieval-augmented generation (RAG), or fine-tuning. We ask: do LLMs really need external context or parameters to adapt to different behaviors, or do they already have such knowledge embedded in their parameters? In this work, we show that LLMs already contain persona-specialized subnetworks in their parameter space. Using small calibration datasets, we identify distinct activation signatures associated with different personas. Guided by these statistics, we develop a masking strategy that isolates lightweight persona subnetworks. Building on the findings, we further discuss: how can we discover opposing subnetwork from the model that lead to binary-opposing personas, such as introvert-extrovert? To further enhance separation in binary opposition scenarios, we introduce a contrastive pruning strategy that identifies parameters responsible for the statistical divergence between opposing personas. Our method is entirely training-free and relies solely on the language model’s existing parameter space. Across diverse evaluation settings, the resulting subnetworks exhibit significantly stronger persona alignment than baselines that require external knowledge while being more efficient. Our findings suggest that diverse human-like behaviors are not merely induced in LLMs, but are already embedded in their parameter space, pointing toward a new perspective on controllable and interpretable personalization in large language models.

Method: The system integrates natural language processing with customizable 3D avatars for multimodal interaction. It uses a structured onboarding process to capture learner goals/preferences, configures learner-specific AI assistants accessible via text and avatar interfaces, and includes tools for content organization, embedded feedback, and analytics.

Result: Open TutorAI delivers adaptive support responding to individual learner profiles without requiring technical expertise. The assistant-generation pipeline and avatar integration enhance engagement and emotional presence, creating a more humanized, immersive learning environment.

Conclusion: Open TutorAI unites modular architecture, generative AI, and learner analytics within an open-source framework, contributing to next-generation intelligent tutoring systems that combine AI and immersive technologies for personalized education.

Abstract: Recent advances in artificial intelligence have created new possibilities for making education more scalable, adaptive, and learner-centered. However, existing educational chatbot systems often lack contextual adaptability, real-time responsiveness, and pedagogical agility. which can limit learner engagement and diminish instructional effectiveness. Thus, there is a growing need for open, integrative platforms that combine AI and immersive technologies to support personalized, meaningful learning experiences. This paper presents Open TutorAI, an open-source educational platform based on LLMs and generative technologies that provides dynamic, personalized tutoring. The system integrates natural language processing with customizable 3D avatars to enable multimodal learner interaction. Through a structured onboarding process, it captures each learner’s goals and preferences in order to configure a learner-specific AI assistant. This assistant is accessible via both text-based and avatar-driven interfaces. The platform includes tools for organizing content, providing embedded feedback, and offering dedicated interfaces for learners, educators, and parents. This work focuses on learner-facing components, delivering a tool for adaptive support that responds to individual learner profiles without requiring technical expertise. Its assistant-generation pipeline and avatar integration enhance engagement and emotional presence, creating a more humanized, immersive learning environment. Embedded learning analytics support self-regulated learning by tracking engagement patterns and generating actionable feedback. The result is Open TutorAI, which unites modular architecture, generative AI, and learner analytics within an open-source framework. It contributes to the development of next-generation intelligent tutoring systems.

Method: 1) Study personality as a latent signal behind preference statements through experiments showing personality-aligned preferences improve personalized QA; 2) Create PACIFIC dataset containing 1200 preference statements annotated with Big-Five (OCEAN) trait directions across diverse domains; 3) Propose a framework enabling LLMs to automatically retrieve personality-aligned preferences and incorporate them during answer generation.

Result: Conditioning on personality-aligned preferences substantially improves personalized question answering: selecting preferences consistent with a user’s inferred personality increases answer-choice accuracy from 29.25% to 76%, compared to using randomly selected preferences.

Conclusion: Personality traits provide a reliable latent signal for filtering noisy user preferences, enabling more effective LLM personalization. The PACIFIC dataset and proposed framework offer practical tools for leveraging personality-aligned preferences in answer generation.

Abstract: User preferences are increasingly used to personalize Large Language Model (LLM) responses, yet how to reliably leverage preference signals for answer generation remains under-explored. In practice, preferences can be noisy, incomplete, or even misleading, which can degrade answer quality when applied naively. Motivated by the observation that stable personality traits shape everyday preferences, we study personality as a principled ‘’latent’’ signal behind preference statements. Through extensive experiments, we find that conditioning on personality-aligned preferences substantially improves personalized question answering: selecting preferences consistent with a user’s inferred personality increases answer-choice accuracy from 29.25% to 76%, compared to using randomly selected preferences. Based on these findings, we introduce PACIFIC (Preference Alignment Choices Inference for Five-factor Identity Characterization), a personality-labeled preference dataset containing 1200 preference statements spanning diverse domains (e.g., travel, movies, education), annotated with Big-Five (OCEAN) trait directions. Finally, we propose a framework that enables an LLM model to automatically retrieve personality-aligned preferences and incorporate them during answer generation.

Method: Proposed system with three key components: (1) deconstructing domain-specific vocabulary for better retrieval, (2) parsing complex document layouts while isolating sections and footnotes with proper linking, (3) generating comprehensive answers using precise domain vocabulary. Also introduces a coverage metric for recall evaluation.

Result: System evaluated on a professionally curated QA dataset with comprehensive experiments and ablation studies demonstrating usability and merit.

Conclusion: The proposed system effectively addresses the challenges of long-context QA for legal documents with complex layouts and specialized vocabulary, providing comprehensive long-form answers.

Abstract: Legal documents have complex document layouts involving multiple nested sections, lengthy footnotes and further use specialized linguistic devices like intricate syntax and domain-specific vocabulary to ensure precision and authority. These inherent characteristics of legal documents make question answering challenging, and particularly so when the answer to the question spans several pages (i.e. requires long-context) and is required to be comprehensive (i.e. a long-form answer). In this paper, we address the challenges of long-context question answering in context of long-form answers given the idiosyncrasies of legal documents. We propose a question answering system that can (a) deconstruct domain-specific vocabulary for better retrieval from source documents, (b) parse complex document layouts while isolating sections and footnotes and linking them appropriately, (c) generate comprehensive answers using precise domain-specific vocabulary. We also introduce a coverage metric that classifies the performance into recall-based coverage categories allowing human users to evaluate the recall with ease. We curate a QA dataset by leveraging the expertise of professionals from fields such as law and corporate tax. Through comprehensive experiments and ablation studies, we demonstrate the usability and merit of the proposed system.

Method: Two approaches: (1) Cascaded system with source separation front-end module, (2) End-to-end system using serialized output training. Both utilize multi-microphone arrays in smart glasses for streaming directional processing.

Result: Experimental results show strong performance in speech recognition and translation tasks, demonstrating effective directional speech understanding capabilities for LLMs.

Conclusion: The proposed methods successfully enable LLMs to handle directional multi-talker speech understanding, particularly valuable for smart glasses applications with embedded microphone arrays.

Abstract: Recent studies have demonstrated that prompting large language models (LLM) with audio encodings enables effective speech understanding capabilities. However, most speech LLMs are trained on single-channel, single-talker data, which makes it challenging to directly apply them to multi-talker and multi-channel speech understanding task. In this work, we present a comprehensive investigation on how to enable directional multi-talker speech understanding capabilities for LLMs, specifically in smart glasses usecase. We propose two novel approaches to integrate directivity into LLMs: (1) a cascaded system that leverages a source separation front-end module, and (2) an end-to-end system that utilizes serialized output training. All of the approaches utilize a multi-microphone array embedded in smart glasses to optimize directivity interpretation and processing in a streaming manner. Experimental results demonstrate the efficacy of our proposed methods in endowing LLMs with directional speech understanding capabilities, achieving strong performance in both speech recognition and speech translation tasks.

Method: Develops a framework that identifies risk-bearing language (treatment directives, contraindications, urgency cues, high-risk medications) and combines risk scoring with relevance measures to identify high-risk, low-grounding failures. Tests on three instruction-tuned LLMs using controlled safety stress test prompts.

Result: Models with similar surface-level behavior show substantially different risk profiles, and standard evaluation metrics fail to capture these distinctions, highlighting the importance of risk-sensitive evaluation.

Conclusion: Risk sensitivity must be incorporated into hallucination evaluation for medical applications, and evaluation validity depends critically on task and prompt design to properly assess safety risks.

Abstract: Large language models are increasingly being used in patient-facing medical question answering, where hallucinated outputs can vary widely in potential harm. However, existing hallucination standards and evaluation metrics focus primarily on factual correctness, treating all errors as equally severe. This obscures clinically relevant failure modes, particularly when models generate unsupported but actionable medical language. We propose a risk-sensitive evaluation framework that quantifies hallucinations through the presence of risk-bearing language, including treatment directives, contraindications, urgency cues, and mentions of high-risk medications. Rather than assessing clinical correctness, our approach evaluates the potential impact of hallucinated content if acted upon. We further combine risk scoring with a relevance measure to identify high-risk, low-grounding failures. We apply this framework to three instruction-tuned language models using controlled patient-facing prompts designed as safety stress tests. Our results show that models with similar surface-level behavior exhibit substantially different risk profiles and that standard evaluation metrics fail to capture these distinctions. These findings highlight the importance of incorporating risk sensitivity into hallucination evaluation and suggest that evaluation validity is critically dependent on task and prompt design.

Method: Proposes STITCH, a generation method that alternates between generating unspoken reasoning chunks and spoken response chunks. Leverages the fact that audio playback time for spoken chunks is longer than token generation time, using the remaining time to generate unspoken reasoning tokens, enabling simultaneous thinking and talking.

Result: STITCH matches the latency of baselines that cannot generate unspoken CoT while outperforming them by 15% on math reasoning datasets. Performs equally well on non-reasoning datasets as baseline models.

Conclusion: STITCH successfully enables spoken language models to perform internal reasoning without adding latency, bridging the gap between human-like thinking processes and real-time speech generation.

Abstract: Spoken Language Models (SLMs) are designed to take speech inputs and produce spoken responses. However, current SLMs lack the ability to perform an internal, unspoken thinking process before responding. In contrast, humans typically engage in complex mental reasoning internally, enabling them to communicate ideas clearly and concisely. Thus, integrating an unspoken thought process into SLMs is highly desirable. While naively generating a complete chain-of-thought (CoT) reasoning before starting to talk can enable thinking for SLMs, this induces additional latency for the speech response, as the CoT reasoning can be arbitrarily long. To solve this issue, we propose Stitch, a novel generation method that alternates between the generation of unspoken reasoning chunks and spoken response chunks. Since the audio duration of a chunk of spoken response is much longer than the time to generate the tokens in a chunk of spoken response, we use the remaining free time to generate the unspoken reasoning tokens. When a chunk of audio is played to the user, the model continues to generate the next unspoken reasoning chunk, achieving simultaneous thinking and talking. Remarkably, Stitch matches the latency of baselines that cannot generate unspoken CoT by design while outperforming those baselines by 15% on math reasoning datasets; Stitch also performs equally well on non-reasoning datasets as those baseline models. Some animations and demonstrations are on the project page: https://d223302.github.io/STITCH.

Method: Proposes a Mediator-Assistant architecture that decouples intent understanding from task execution. Uses an experience-driven Mediator to explicate vague user inputs into explicit, well-structured instructions based on historical interaction patterns.

Result: Experimental results show the method significantly mitigates performance degradation in multi-turn conversations across diverse LLMs.

Conclusion: Lost in Conversation is not a capability failure but an interaction breakdown; scaling model size alone cannot fix it; the Mediator-Assistant architecture effectively bridges intent alignment gaps.

Abstract: Multi-turn conversation has emerged as a predominant interaction paradigm for Large Language Models (LLMs). Users often employ follow-up questions to refine their intent, expecting LLMs to adapt dynamically. However, recent research reveals that LLMs suffer a substantial performance drop in multi-turn settings compared to single-turn interactions with fully specified instructions, a phenomenon termed ``Lost in Conversation’’ (LiC). While this prior work attributes LiC to model unreliability, we argue that the root cause lies in an intent alignment gap rather than intrinsic capability deficits. In this paper, we first demonstrate that LiC is not a failure of model capability but rather a breakdown in interaction between users and LLMs. We theoretically show that scaling model size or improving training alone cannot resolve this gap, as it arises from structural ambiguity in conversational context rather than representational limitations. To address this, we propose to decouple intent understanding from task execution through a Mediator-Assistant architecture. By utilizing an experience-driven Mediator to explicate user inputs into explicit, well-structured instructions based on historical interaction patterns, our approach effectively bridges the gap between vague user intent and model interpretation. Experimental results demonstrate that this method significantly mitigates performance degradation in multi-turn conversations across diverse LLMs.

Method: Proposes VocalNet-MDM using Masked Diffusion Modeling with Hierarchical Block-wise Masking to align training with progressive masked states during block diffusion decoding, and Iterative Self-Distillation to compress multi-step refinement into fewer steps for low-latency inference.

Result: Achieves 3.7×-10× decoding speedup, reduces first-chunk latency by 34% compared to AR baselines, maintains competitive recognition accuracy, and achieves state-of-the-art text quality and speech naturalness with only 6K hours of training data.

Conclusion: Masked Diffusion Modeling is a promising and scalable alternative for low-latency, efficient speech LLMs, demonstrating that non-autoregressive approaches can achieve both efficiency and quality improvements.

Abstract: Recent Speech Large Language Models~(LLMs) have achieved impressive capabilities in end-to-end speech interaction. However, the prevailing autoregressive paradigm imposes strict serial constraints, limiting generation efficiency and introducing exposure bias. In this paper, we investigate Masked Diffusion Modeling~(MDM) as a non-autoregressive paradigm for speech LLMs and introduce VocalNet-MDM. To adapt MDM for streaming speech interaction, we address two critical challenges: training-inference mismatch and iterative overhead. We propose Hierarchical Block-wise Masking to align training objectives with the progressive masked states encountered during block diffusion decoding, and Iterative Self-Distillation to compress multi-step refinement into fewer steps for low-latency inference. Trained on a limited scale of only 6K hours of speech data, VocalNet-MDM achieves a 3.7$\times$–10$\times$ decoding speedup and reduces first-chunk latency by 34% compared to AR baselines. It maintains competitive recognition accuracy while achieving state-of-the-art text quality and speech naturalness, demonstrating that MDM is a promising and scalable alternative for low-latency, efficient speech LLMs.

Method: Proposes ViHERMES dataset construction using controlled multihop QA generation pipeline based on semantic clustering and graph-inspired data mining, followed by LLM-based generation with structured evidence. Also presents a graph-aware retrieval framework modeling formal legal relations at legal unit level.

Result: ViHERMES provides a challenging benchmark for evaluating multihop regulatory QA systems, and the proposed graph-aware approach consistently outperforms strong retrieval-based baselines.

Conclusion: The work introduces a valuable benchmark for multihop reasoning over healthcare regulations in Vietnamese, with a graph-aware framework that improves QA performance on regulatory documents.

Abstract: Question Answering (QA) over regulatory documents is inherently challenging due to the need for multihop reasoning across legally interdependent texts, a requirement that is particularly pronounced in the healthcare domain where regulations are hierarchically structured and frequently revised through amendments and cross-references. Despite recent progress in retrieval-augmented and graph-based QA methods, systematic evaluation in this setting remains limited, especially for low-resource languages such as Vietnamese, due to the lack of benchmark datasets that explicitly support multihop reasoning over healthcare regulations. In this work, we introduce the Vietnamese Healthcare Regulations-Multihop Reasoning Dataset (ViHERMES), a benchmark designed for multihop QA over Vietnamese healthcare regulatory documents. ViHERMES consists of high-quality question-answer pairs that require reasoning across multiple regulations and capture diverse dependency patterns, including amendment tracing, cross-document comparison, and procedural synthesis. To construct the dataset, we propose a controlled multihop QA generation pipeline based on semantic clustering and graph-inspired data mining, followed by large language model-based generation with structured evidence and reasoning annotations. We further present a graph-aware retrieval framework that models formal legal relations at the level of legal units and supports principled context expansion for legally valid and coherent answers. Experimental results demonstrate that ViHERMES provides a challenging benchmark for evaluating multihop regulatory QA systems and that the proposed graph-aware approach consistently outperforms strong retrieval-based baselines. The ViHERMES dataset and system implementation are publicly available at https://github.com/ura-hcmut/ViHERMES.

Method: Introduces MTR-DuplexBench that segments continuous full-duplex dialogues into discrete turns for turn-by-turn assessment and incorporates multiple evaluation dimensions including conversational features, dialogue quality, instruction following, and safety.

Result: Experimental results show current FD-SLMs struggle to maintain consistent performance across multiple rounds and evaluation dimensions, demonstrating the necessity and effectiveness of the proposed benchmark.

Conclusion: MTR-DuplexBench addresses critical gaps in FD-SLM evaluation by enabling comprehensive multi-round assessment across multiple dimensions, revealing current model limitations and providing a more realistic evaluation framework.

Abstract: Full-Duplex Speech Language Models (FD-SLMs) enable real-time, overlapping conversational interactions, offering a more dynamic user experience compared to traditional half-duplex models. However, existing benchmarks primarily focus on evaluating single-round interactions, neglecting the complexities of multi-round communication. Evaluating FD-SLMs in multi-round settings poses significant challenges, including blurred turn boundaries in communication and context inconsistency during model inference. Also, existing benchmarks often focus solely on evaluating conversational features, neglecting other critical aspects. To address these gaps, we introduce MTR-DuplexBench, a novel benchmark designed for a comprehensive multi-round evaluation of FD-SLMs. MTR-DuplexBench not only segments continuous full-duplex dialogues into discrete turns for turn-by-turn assessment but also incorporates various evaluation aspects, including conversational features, dialogue quality, instruction following, and safety. Experimental results reveal that current FD-SLMs face difficulties in maintaining consistent performance across multiple rounds and evaluation dimensions, highlighting the necessity and effectiveness of our benchmark. The benchmark and code will be available in the future.

Method: TernaryLM employs native 1-bit ternary quantization {-1, 0, +1} during training (not post-training), using straight-through estimators and adaptive per-layer scaling factors to learn quantization-aware representations from scratch.

Result: Achieves 58.42 perplexity on TinyStories, 82.47% F1 on MRPC paraphrase detection, 2.4x memory reduction (498MB vs 1197MB) with comparable inference latency, and stable training across diverse corpora.

Conclusion: Native 1-bit training is a promising direction for efficient neural language models, with middle transformer layers showing highest compatibility with extreme quantization, informing future non-uniform precision strategies.

Abstract: Large language models (LLMs) achieve remarkable performance but demand substantial computational resources, limiting deployment on edge devices and resource-constrained environments. We present TernaryLM, a 132M parameter transformer architecture that employs native 1-bit ternary quantization {-1, 0, +1} during training, achieving significant memory reduction without sacrificing language modeling capability. Unlike post-training quantization approaches that quantize pre-trained full-precision models, TernaryLM learns quantization-aware representations from scratch using straight-through estimators and adaptive per-layer scaling factors. Our experiments demonstrate: (1) validation perplexity of 58.42 on TinyStories; (2) downstream transfer with 82.47 percent F1 on MRPC paraphrase detection; (3) 2.4x memory reduction (498MB vs 1197MB) with comparable inference latency; and (4) stable training dynamics across diverse corpora. We provide layer-wise quantization analysis showing that middle transformer layers exhibit highest compatibility with extreme quantization, informing future non-uniform precision strategies. Our results suggest that native 1-bit training is a promising direction for efficient neural language models. Code is available at https://github.com/1nisharg/TernaryLM-Memory-Efficient-Language-Modeling.

Method: Proposes a pruning framework based on first-order statistical properties of weights and activations. Uses channel-wise statistics to calibrate magnitude-based importance scores, reducing bias from activation-dominated channels. After pruning, applies analytic energy compensation to correct distributional distortions from weight removal. Both steps operate without retraining, gradients, or second-order information.

Result: Experiments across multiple LLM families, sparsity patterns, and evaluation tasks show improved pruning performance while maintaining computational cost comparable to heuristic methods.

Conclusion: Simple statistical corrections can be effective for post-training pruning of LLMs, offering a better balance between quality and efficiency than existing approaches.

Abstract: Post-training pruning is an effective approach for reducing the size and inference cost of large language models (LLMs), but existing methods often face a trade-off between pruning quality and computational efficiency. Heuristic pruning methods are efficient but sensitive to activation outliers, while reconstruction-based approaches improve fidelity at the cost of heavy computation. In this work, we propose a lightweight post-training pruning framework based on first-order statistical properties of model weights and activations. During pruning, channel-wise statistics are used to calibrate magnitude-based importance scores, reducing bias from activation-dominated channels. After pruning, we apply an analytic energy compensation to correct distributional distortions caused by weight removal. Both steps operate without retraining, gradients, or second-order information. Experiments across multiple LLM families, sparsity patterns, and evaluation tasks show that the proposed approach improves pruning performance while maintaining computational cost comparable to heuristic methods. The results suggest that simple statistical corrections can be effective for post-training pruning of LLMs.

Method: Two-stage approach: Stage I reclassifies DICES prompts into 14 safety domains using Mistral 7B-Instruct-v0.3, retains demographic metadata, and expands low-resource domains via Llama-3.1-8B-Instruct with SimHash deduplication (43,050 samples). Stage II evaluates pluralistic sensitivity using LLMs-as-Raters (Gemma-7B, GPT-4o, LLaMA-2-7B) with zero-shot inference and balanced thresholds.

Result: Achieved high reliability (ICC = 0.87) and low demographic sensitivity (DS = 0.12) with balanced thresholds (delta = 0.5, tau = 10), demonstrating that pluralistic safety evaluation can be both scalable and demographically robust.

Conclusion: Demo-SafetyBench successfully addresses the demographic narrowness in existing safety benchmarks by modeling pluralism at the prompt level, showing that scalable, demographically-aware safety evaluation is feasible and reliable.

Abstract: Large Language Model (LLM) safety is inherently pluralistic, reflecting variations in moral norms, cultural expectations, and demographic contexts. Yet, existing alignment datasets such as ANTHROPIC-HH and DICES rely on demographically narrow annotator pools, overlooking variation in safety perception across communities. Demo-SafetyBench addresses this gap by modeling demographic pluralism directly at the prompt level, decoupling value framing from responses. In Stage I, prompts from DICES are reclassified into 14 safety domains (adapted from BEAVERTAILS) using Mistral 7B-Instruct-v0.3, retaining demographic metadata and expanding low-resource domains via Llama-3.1-8B-Instruct with SimHash-based deduplication, yielding 43,050 samples. In Stage II, pluralistic sensitivity is evaluated using LLMs-as-Raters-Gemma-7B, GPT-4o, and LLaMA-2-7B-under zero-shot inference. Balanced thresholds (delta = 0.5, tau = 10) achieve high reliability (ICC = 0.87) and low demographic sensitivity (DS = 0.12), confirming that pluralistic safety evaluation can be both scalable and demographically robust.

Method: Two-stage framework: Stage 1 uses prompt-injected fine-tuning to extract axis-specific task features, mitigating catastrophic forgetting. Stage 2 deploys MoCaE module that calibrates expert routing using fractal and natural geometry for improved inference reliability

Result: Significant gains on Alpaca (+171.5% win rate for helpfulness), BeaverTails (fewer safety violations), and TruthfulQA (+110.1% in truthfulness-informativeness). Reduces latency and memory usage by over 35% compared to prior MoEs. Validated across four LLMs

Conclusion: AlignX effectively addresses axis collapse in multi-objective LLM alignment, improving performance on HHH metrics while reducing computational overhead, demonstrating generalizability across different LLMs

Abstract: Large Language Models (LLMs) need to be in accordance with human values-being helpful, harmless, and honest (HHH)-is important for safe deployment. Existing works use Supervised Fine-Tuning (SFT) and Mixture-of-Experts (MoE) to align LLMs. However, these works face challenges in multi-objective settings, such as SFT leading to interference between conflicting objectives, while MoEs suffer from miscalibrated routing. We term this failure mode Axis Collapse, marked by (1) disjoint feature spaces causing catastrophic forgetting, and (2) unreliable inference from misrouted experts. To resolve this, we propose AlignX, a two-stage framework. Stage 1 uses prompt-injected fine-tuning to extract axis-specific task features, mitigating catastrophic forgetting. Stage 2 deploys a MoCaE module that calibrates expert routing using fractal and natural geometry, improving inference reliability. AlignX achieves significant gains on Alpaca (Helpfulness), BeaverTails (Harmlessness), and TruthfulQA (Honesty), with +171.5% win rate, +110.1% in truthfulness-informativeness, and 4.3% fewer safety violations. It also reduces latency and memory usage by over 35% compared to prior MoEs. Results across four LLMs validate its generalizability.

Method: 1) Enhanced extractive neural summarization models by incorporating domain-specific pre-trained encoders for legal texts. 2) Injected legal domain knowledge into generative models (including LLMs) through continual pre-training on large legal corpora in English and Hindi.

Result: Achieved statistically significant improvements in both English-to-English and English-to-Hindi Indian legal document summarization, measured by standard evaluation metrics, factual consistency metrics, and legal domain-specific metrics, validated by domain experts.

Conclusion: The proposed approaches effectively improve legal text summarization for Indian contexts by incorporating domain knowledge, addressing both language complexity and accessibility issues for non-English speaking populations.

Abstract: Summarizing Indian legal court judgments is a complex task not only due to the intricate language and unstructured nature of the legal texts, but also since a large section of the Indian population does not understand the complex English in which legal text is written, thus requiring summaries in Indian languages. In this study, we aim to improve the summarization of Indian legal text to generate summaries in both English and Hindi (the most widely spoken Indian language), by injecting domain knowledge into diverse summarization models. We propose a framework to enhance extractive neural summarization models by incorporating domain-specific pre-trained encoders tailored for legal texts. Further, we explore the injection of legal domain knowledge into generative models (including Large Language Models) through continual pre-training on large legal corpora in English and Hindi. Our proposed approaches achieve statistically significant improvements in both English-to-English and English-to-Hindi Indian legal document summarization, as measured by standard evaluation metrics, factual consistency metrics, and legal domain-specific metrics. Furthermore, these improvements are validated through domain experts, demonstrating the effectiveness of our approaches.

Method: Introduces Emotion-Aware Stepwise Preference Optimization (EASPO) with EASPM, a time-conditioned model that scores noisy intermediate speech states to enable automatic preference pair construction and stepwise preference optimization for controllable emotional shaping.

Result: Experiments show superior performance over existing methods in both expressiveness and naturalness for emotional TTS generation.

Conclusion: EASPO enables more fine-grained emotional control in TTS systems through stepwise preference optimization, advancing emotional speech synthesis capabilities.

Abstract: Emotional text-to-speech seeks to convey affect while preserving intelligibility and prosody, yet existing methods rely on coarse labels or proxy classifiers and receive only utterance-level feedback. We introduce Emotion-Aware Stepwise Preference Optimization (EASPO), a post-training framework that aligns diffusion TTS with fine-grained emotional preferences at intermediate denoising steps. Central to our approach is EASPM, a time-conditioned model that scores noisy intermediate speech states and enables automatic preference pair construction. EASPO optimizes generation to match these stepwise preferences, enabling controllable emotional shaping. Experiments show superior performance over existing methods in both expressiveness and naturalness.

Method: Introduces a novel computational metric using lexical similarity based on both surface (orthographic/phonetic) and semantic similarity of related words. Applies this to five main Romance languages using different parallel corpora and word embedding models.

Result: The computational intelligibility scores confirm intuitions about intelligibility asymmetry across Romance languages and show significant correlation with results from human cloze tests.

Conclusion: Computational methods using lexical similarity can effectively estimate mutual intelligibility between related languages, providing results that align with human comprehension patterns.

Abstract: We present an analysis of mutual intelligibility in related languages applied for languages in the Romance family. We introduce a novel computational metric for estimating intelligibility based on lexical similarity using surface and semantic similarity of related words, and use it to measure mutual intelligibility for the five main Romance languages (French, Italian, Portuguese, Spanish, and Romanian), and compare results using both the orthographic and phonetic forms of words as well as different parallel corpora and vectorial models of word meaning representation. The obtained intelligibility scores confirm intuitions related to intelligibility asymmetry across languages and significantly correlate with results of cloze tests in human experiments.

Method: Created DLLM and AR backbones within the same DeepDiver agent workflow, performed matched agent-oriented fine-tuning on identical trajectory data, and compared diffusion-backed DLLM Agents with directly comparable AR agents across benchmarks and case studies.

Result: DLLM Agents are on average over 30% faster end-to-end than AR agents (some cases exceeding 8x speedup), require fewer interaction rounds and tool invocations, show higher planner hit rates, and converge earlier to correct action paths with less backtracking.

Conclusion: Diffusion backbones offer significant efficiency advantages for agentic decision making but require careful handling of tool-call failures and attention masking for multi-turn inputs to achieve optimal performance.

Abstract: Diffusion large language models (DLLMs) have emerged as an alternative to autoregressive (AR) decoding with appealing efficiency and modeling properties, yet their implications for agentic multi-step decision making remain underexplored. We ask a concrete question: when the generation paradigm is changed but the agent framework and supervision are held fixed, do diffusion backbones induce systematically different planning and tool-use behaviors, and do these differences translate into end-to-end efficiency gains? We study this in a controlled setting by instantiating DLLM and AR backbones within the same agent workflow (DeepDiver) and performing matched agent-oriented fine-tuning on the same trajectory data, yielding diffusion-backed DLLM Agents and directly comparable AR agents. Across benchmarks and case studies, we find that, at comparable accuracy, DLLM Agents are on average over 30% faster end to end than AR agents, with some cases exceeding 8x speedup. Conditioned on correct task completion, DLLM Agents also require fewer interaction rounds and tool invocations, consistent with higher planner hit rates that converge earlier to a correct action path with less backtracking. We further identify two practical considerations for deploying diffusion backbones in tool-using agents. First, naive DLLM policies are more prone to structured tool-call failures, necessitating stronger tool-call-specific training to emit valid schemas and arguments. Second, for multi-turn inputs interleaving context and action spans, diffusion-style span corruption requires aligned attention masking to avoid spurious context-action information flow; without such alignment, performance degrades. Finally, we analyze attention dynamics across workflow stages and observe paradigm-specific coordination patterns, suggesting stronger global planning signals in diffusion-backed agents.

Method: Proposes SED-SFT with selective entropy regularization and masking mechanism that adaptively encourages diversity based on token exploration space while maintaining accuracy.

Result: Significantly improves generation diversity with minimal computational overhead, yielding average gains of 2.06 and 1.20 points in RL performance on Llama-3.2-3B-Instruct and Qwen2.5-Math-7B-Instruct across eight mathematical benchmarks.

Conclusion: SED-SFT effectively balances diversity and accuracy in SFT, addressing mode collapse and improving subsequent RL performance in LLM training pipelines.

Abstract: Supervised Fine-Tuning (SFT) followed by Reinforcement Learning (RL) has emerged as the standard post-training paradigm for large language models (LLMs). However, the conventional SFT process, driven by Cross-Entropy (CE) loss, often induces mode collapse, where models over-concentrate on specific response patterns. This lack of distributional diversity severely restricts the exploration efficiency required for subsequent RL. While recent studies have attempted to improve SFT by replacing the CE loss, aiming to preserve diversity or refine the update policy, they fail to adequately balance diversity and accuracy, thereby yielding suboptimal performance after RL. To address the mode collapse problem, we propose SED-SFT, which adaptively encourages diversity based on the token exploration space. This framework introduces a selective entropy regularization term with a selective masking mechanism into the optimization objective. Extensive experiments across eight mathematical benchmarks demonstrate that SED-SFT significantly enhances generation diversity with a negligible computational overhead increase compared with CE loss, yielding average improvements of 2.06 and 1.20 points in subsequent RL performance over standard CE-based baselines on Llama-3.2-3B-Instruct and Qwen2.5-Math-7B-Instruct, respectively. The code is publicly available at https://github.com/pppa2019/SED-SFT

Method: Introduced systematic evaluation framework to diagnose cross-cultural robustness of VLMs across multilingual meme datasets, analyzing three axes: (i) learning strategy (zero-shot vs. one-shot), (ii) prompting language (native vs. English), and (iii) translation effects on meaning and detection.

Result: Common “translate-then-detect” approach deteriorates performance, while culturally aligned interventions - native-language prompting and one-shot learning - significantly enhance detection. Reveals systematic convergence toward Western safety norms.

Conclusion: Provides actionable strategies to mitigate cultural bias in VLMs and guides design of globally robust multimodal moderation systems through culturally aligned interventions.

Abstract: Cultural context profoundly shapes how people interpret online content, yet vision-language models (VLMs) remain predominantly trained through Western or English-centric lenses. This limits their fairness and cross-cultural robustness in tasks like hateful meme detection. We introduce a systematic evaluation framework designed to diagnose and quantify the cross-cultural robustness of state-of-the-art VLMs across multilingual meme datasets, analyzing three axes: (i) learning strategy (zero-shot vs. one-shot), (ii) prompting language (native vs. English), and (iii) translation effects on meaning and detection. Results show that the common ``translate-then-detect’’ approach deteriorate performance, while culturally aligned interventions - native-language prompting and one-shot learning - significantly enhance detection. Our findings reveal systematic convergence toward Western safety norms and provide actionable strategies to mitigate such bias, guiding the design of globally robust multimodal moderation systems.

Method: Proposes a framework that decomposes complex simplifications into manageable steps through: 1) dynamic path planning, 2) semantic-aware exemplar selection, and 3) chain-of-thought generation with conversation history for coherent reasoning.

Result: Evaluation on five languages across two benchmarks shows the approach improves simplification effectiveness while reducing computational steps by 22-42%. Human evaluation reveals fundamental trade-off between simplification effectiveness and meaning preservation, with even human annotators struggling to agree on semantic preservation judgments.

Conclusion: While step-by-step simplification improves control over the simplification process, preserving semantic fidelity during extensive simplification remains an open challenge, highlighting the inherent complexity of this task.

Abstract: Large language models demonstrate limited capability in proficiency-controlled sentence simplification, particularly when simplifying across large readability levels. We propose a framework that decomposes complex simplifications into manageable steps through dynamic path planning, semantic-aware exemplar selection, and chain-of-thought generation with conversation history for coherent reasoning. Evaluation on five languages across two benchmarks shows our approach improves simplification effectiveness while reducing computational steps by 22-42%. Human evaluation confirms the fundamental trade-off between simplification effectiveness and meaning preservation. Notably, even human annotators struggle to agree on semantic preservation judgments, highlighting the inherent complexity of this task. Our work shows that while step-by-step simplification improves control, preserving semantic fidelity during extensive simplification remains an open challenge.

Method: Proposes LR-DLLM framework that treats generation length as an explicit variable, decouples semantic compatibility from length-induced uncertainty through explicit length regularization, and enables dynamic expansion/contraction of generation spans without modifying the underlying DLLM or training procedure.

Result: Achieves 51.3% Pass@1 on HumanEvalInfilling under fully unknown lengths (+13.4% vs. DreamOn) and 51.5% average Pass@1 on four-language McEval (+14.3% vs. DreamOn).

Conclusion: LR-DLLM provides a robust solution for variable-length inference in Diffusion LLMs by addressing intrinsic length-induced biases in confidence estimates, enabling reliable length determination at inference time.

Abstract: Diffusion Large Language Models (DLLMs) are inherently ill-suited for variable-length generation, as their inference is defined on a fixed-length canvas and implicitly assumes a known target length. When the length is unknown, as in realistic completion and infilling, naively comparing confidence across mask lengths becomes systematically biased, leading to under-generation or redundant continuations. In this paper, we show that this failure arises from an intrinsic lengthinduced bias in generation confidence estimates, leaving existing DLLMs without a robust way to determine generation length and making variablelength inference unreliable. To address this issue, we propose LR-DLLM, a length-regularized inference framework for DLLMs that treats generation length as an explicit variable and achieves reliable length determination at inference time. It decouples semantic compatibility from lengthinduced uncertainty through an explicit length regularization that corrects biased confidence estimates. Based on this, LR-DLLM enables dynamic expansion or contraction of the generation span without modifying the underlying DLLM or its training procedure. Experiments show that LRDLLM achieves 51.3% Pass@1 on HumanEvalInfilling under fully unknown lengths (+13.4% vs. DreamOn) and 51.5% average Pass@1 on four-language McEval (+14.3% vs. DreamOn).

Method: Analyze training evolution to study generation-verification asymmetry. Develop multi-task reinforcement learning framework where generation and self-verification are optimized as independent but complementary objectives. Conduct extensive experiments across benchmarks and models.

Result: Learning to self-verify effectively improves generation performance, achieving comparable accuracy to standard generation training while producing more efficient reasoning traces. Multi-task RL framework outperforms generation-only training in both generation and verification capabilities.

Conclusion: Self-verification can enhance generation capabilities, and integrating both objectives through multi-task RL leads to better overall performance. The asymmetry is directional: verification helps generation more than generation helps verification.

Abstract: Recent large language models (LLMs) achieve strong performance in generating promising reasoning paths for complex tasks. However, despite powerful generation ability, LLMs remain weak at verifying their own answers, revealing a persistent capability asymmetry between generation and self-verification. In this work, we conduct an in-depth investigation of this asymmetry throughout training evolution and show that, even on the same task, improving generation does not lead to corresponding improvements in self-verification. Interestingly, we find that the reverse direction of this asymmetry behaves differently: learning to self-verify can effectively improve generation performance, achieving accuracy comparable to standard generation training while yielding more efficient and effective reasoning traces. Building on this observation, we further explore integrating self-verification into generation training by formulating a multi-task reinforcement learning framework, where generation and self-verification are optimized as two independent but complementary objectives. Extensive experiments across benchmarks and models demonstrate performance gains over generation-only training in both generation and verification capabilities.

Method: Created dataset by extracting authentic claims from published papers across ML, NLP, and medicine. For refuted claims, modified supporting evidence (figures and tables) rather than altering claims. Dataset includes cross-modal evidence with figures as images and tables in multiple formats (images, LaTeX, HTML, JSON). Contains 1,664 annotated samples from 180 papers with expert validation.

Result: Benchmarked 11 multimodal foundation models (open-source and proprietary). Results show figure-based verification remains particularly challenging, with substantial performance gap between best system and human baseline.

Conclusion: SciClaimEval provides a realistic benchmark for scientific claim verification that better reflects real-world scenarios. The dataset highlights current limitations of multimodal models in verifying claims based on visual evidence like figures.

Abstract: We present SciClaimEval, a new scientific dataset for the claim verification task. Unlike existing resources, SciClaimEval features authentic claims, including refuted ones, directly extracted from published papers. To create refuted claims, we introduce a novel approach that modifies the supporting evidence (figures and tables), rather than altering the claims or relying on large language models (LLMs) to fabricate contradictions. The dataset provides cross-modal evidence with diverse representations: figures are available as images, while tables are provided in multiple formats, including images, LaTeX source, HTML, and JSON. SciClaimEval contains 1,664 annotated samples from 180 papers across three domains, machine learning, natural language processing, and medicine, validated through expert annotation. We benchmark 11 multimodal foundation models, both open-source and proprietary, across the dataset. Results show that figure-based verification remains particularly challenging for all models, as a substantial performance gap remains between the best system and human baseline.

Method: The authors modify Bidirectional Preference Optimization (BiPO) to learn a steering vector - an activation-space direction that steers model responses toward specific tutor personas. This approach uses tutor personas embedded in human tutor-student dialogues to guide LLM behavior without explicit prompting.

Result: The steering vector captures tutor-specific variation across dialogue contexts, improving semantic alignment with ground-truth tutor utterances and increasing preference-based evaluations while preserving lexical similarity. Analysis reveals interpretable structure across tutors corresponding to consistent differences in tutoring behavior.

Conclusion: Activation steering offers an effective and interpretable way to control tutor-specific variation in LLMs using signals derived directly from human dialogue data, enabling more nuanced and personalized tutoring systems.

Abstract: With the emergence of large language models (LLMs) as a powerful class of generative artificial intelligence (AI), their use in tutoring has become increasingly prominent. Prior works on LLM-based tutoring typically learn a single tutor policy and do not capture the diversity of tutoring styles. In real-world tutor-student interactions, pedagogical intent is realized through adaptive instructional strategies, with tutors varying the level of scaffolding, instructional directiveness, feedback, and affective support in response to learners’ needs. These differences can all impact dialogue dynamics and student engagement. In this paper, we explore how tutor personas embedded in human tutor-student dialogues can be used to guide LLM behavior without relying on explicitly prompted instructions. We modify Bidirectional Preference Optimization (BiPO) to learn a steering vector, an activation-space direction that steers model responses towards certain tutor personas. We find that this steering vector captures tutor-specific variation across dialogue contexts, improving semantic alignment with ground-truth tutor utterances and increasing preference-based evaluations, while largely preserving lexical similarity. Analysis of the learned directional coefficients further reveals interpretable structure across tutors, corresponding to consistent differences in tutoring behavior. These results demonstrate that activation steering offers an effective and interpretable way for controlling tutor-specific variation in LLMs using signals derived directly from human dialogue data.

Method: Conducted bias analysis of LLM judges as a function of overlap with human-written responses in summarization domain. Tested 9 recent LLMs (1B to 12B parameters) including Gemma 3 and LLaMA 3 variants, measuring preferences between AI-generated and human-written summaries based on ROUGE and BLEU similarity metrics.

Result: LLM judges increasingly prefer AI-generated summaries over human-written ones as ROUGE/BLEU similarities decrease. This pattern extends to all but one model tested and persists regardless of models’ own position biases. Models struggle to judge summaries even with limited overlaps.

Conclusion: LLM-as-a-judge in summarization should rely on techniques beyond simple comparison due to systematic biases. The findings highlight limitations of current LLM evaluation methods and suggest need for more sophisticated approaches.

Abstract: Large language model (LLM) judges have often been used alongside traditional, algorithm-based metrics for tasks like summarization because they better capture semantic information, are better at reasoning, and are more robust to paraphrasing. However, LLM judges show biases for length and order among others, and are vulnerable to various adversarial input prompts. While recent studies have looked into these biases, few have analyzed them at a more granular level in relation to a well-defined overlap metric. In this work we provide an LLM judge bias analysis as a function of overlap with human-written responses in the domain of summarization. We test 9 recent LLMs with parameter counts ranging from 1 billion to 12 billion, including variants of Gemma 3 and LLaMA 3. We find that LLM judges increasingly prefer summaries generated by other LLMs over those written by humans as the similarities (as measured by ROUGE and BLEU) between the judged summaries decrease, and this pattern extends to all but one model tested, and exists regardless of the models’ own position biases. Additionally, we find that models struggle to judge even summaries with limited overlaps, suggesting that LLM-as-a-judge in the summary domain should rely on techniques beyond a simple comparison.

Method: Introduces SRR-Judge framework for reliable step-level assessment, integrates into ReAct-style workflow for fine-grained guidance, uses iterative rejection sampling fine-tuning with SRR-annotated data.

Result: SRR-Judge outperforms larger models like DeepSeek-V3.1 in step-level evaluation reliability, shows strong correlation with final answer correctness, and yields over 10% average absolute pass@1 improvement on deep search benchmarks.

Conclusion: Step-level assessment via SRR-Judge significantly enhances deep search agent performance by providing better training supervision for search-integrated reasoning.

Abstract: Recent deep search agents built on large reasoning models (LRMs) excel at complex question answering by iteratively planning, acting, and gathering evidence, a capability known as search-integrated reasoning. However, mainstream approaches often train this ability using only outcome-based supervision, neglecting the quality of intermediate thoughts and actions. We introduce SRR-Judge, a framework for reliable step-level assessment of reasoning and search actions. Integrated into a modified ReAct-style rate-and-refine workflow, SRR-Judge provides fine-grained guidance for search-integrated reasoning and enables efficient post-training annotation. Using SRR-annotated data, we apply an iterative rejection sampling fine-tuning procedure to enhance the deep search capability of the base agent. Empirically, SRR-Judge delivers more reliable step-level evaluations than much larger models such as DeepSeek-V3.1, with its ratings showing strong correlation with final answer correctness. Moreover, aligning the policy with SRR-Judge annotated trajectories leads to substantial performance gains, yielding over a 10 percent average absolute pass@1 improvement across challenging deep search benchmarks.

Method: Proposes Attn-GS, an attention-guided context compression framework that: 1) Uses attention patterns from a marking model to identify important personalization sentences, 2) Guides a compression model to generate task-relevant, high-quality compressed user contexts based on these attention signals.

Result: Extensive experiments show Attn-GS significantly outperforms various baselines across different tasks, token limits, and settings, achieving performance close to using full context while reducing token usage by 50 times.

Conclusion: LLMs’ attention patterns can effectively identify important personalization signals for intelligent context compression, and fine-tuning enhances LLMs’ ability to distinguish relevant from irrelevant information, enabling efficient personalization with dramatically reduced token usage.

Abstract: Personalizing large language models (LLMs) to individual users requires incorporating extensive interaction histories and profiles, but input token constraints make this impractical due to high inference latency and API costs. Existing approaches rely on heuristic methods such as selecting recent interactions or prompting summarization models to compress user profiles. However, these methods treat context as a monolithic whole and fail to consider how LLMs internally process and prioritize different profile components. We investigate whether LLMs’ attention patterns can effectively identify important personalization signals for intelligent context compression. Through preliminary studies on representative personalization tasks, we discover that (a) LLMs’ attention patterns naturally reveal important signals, and (b) fine-tuning enhances LLMs’ ability to distinguish between relevant and irrelevant information. Based on these insights, we propose Attn-GS, an attention-guided context compression framework that leverages attention feedback from a marking model to mark important personalization sentences, then guides a compression model to generate task-relevant, high-quality compressed user contexts. Extensive experiments demonstrate that Attn-GS significantly outperforms various baselines across different tasks, token limits, and settings, achieving performance close to using full context while reducing token usage by 50 times.

Method: Investigated LLM internal processing during in-context concept inference using causal mediation analyses to determine functional centrality of identified conceptual subspaces, and analyzed layer-wise progression of representation construction.

Result: Found a conceptual subspace emerging in middle to late layers with persistent representational structure across contexts; demonstrated this subspace is causally central to model predictions; identified early-to-middle layers integrate contextual cues to construct the subspace, which later layers leverage for predictions.

Conclusion: LLMs dynamically construct and use structured latent representations for inference, providing insights into computational processes underlying flexible adaptation and establishing causal role of conceptual representations in LLM reasoning.

Abstract: Large language models (LLMs) exhibit emergent behaviors suggestive of human-like reasoning. While recent work has identified structured, human-like conceptual representations within these models, it remains unclear whether they functionally rely on such representations for reasoning. Here we investigate the internal processing of LLMs during in-context concept inference. Our results reveal a conceptual subspace emerging in middle to late layers, whose representational structure persists across contexts. Using causal mediation analyses, we demonstrate that this subspace is not merely an epiphenomenon but is functionally central to model predictions, establishing its causal role in inference. We further identify a layer-wise progression where attention heads in early-to-middle layers integrate contextual cues to construct and refine the subspace, which is subsequently leveraged by later layers to generate predictions. Together, these findings provide evidence that LLMs dynamically construct and use structured, latent representations in context for inference, offering insights into the computational processes underlying flexible adaptation.

Method: Comprehensive experiments across seven models, three benchmarks, and two thinking instantiations. Evaluation includes quantitative response taxonomy analysis and qualitative failure propagation case studies. Also tested explicit prompting for information disclosure.

Result: Contrary to expectations, mandatory thinking often backfires in user-engaged settings, causing performance degradation. Thinking makes agents more “introverted” - shortening responses and reducing information disclosure, which weakens agent-user information exchange and leads to downstream task failures. Explicit prompting for information disclosure reliably improves performance.

Conclusion: Information transparency awareness is crucial yet underexplored for designing reasoning agents in real-world scenarios. Proactive transparency is a vital lever for agent optimization, and thinking mechanisms need to be carefully designed to maintain effective agent-user communication.

Abstract: Eliciting reasoning has emerged as a powerful technique for improving the performance of large language models (LLMs) on complex tasks by inducing thinking. However, their effectiveness in realistic user-engaged agent scenarios remains unclear. In this paper, we conduct a comprehensive study on the effect of explicit thinking in user-engaged LLM agents. Our experiments span across seven models, three benchmarks, and two thinking instantiations, and we evaluate them through both a quantitative response taxonomy analysis and qualitative failure propagation case studies. Contrary to expectations, we find that mandatory thinking often backfires on agents in user-engaged settings, causing anomalous performance degradation across various LLMs. Our key finding reveals that thinking makes agents more ``introverted’’ by shortening responses and reducing information disclosure to users, which weakens agent-user information exchange and leads to downstream task failures. Furthermore, we demonstrate that explicitly prompting for information disclosure reliably improves performance across diverse model families, suggesting that proactive transparency is a vital lever for agent optimization. Overall, our study suggests that information transparency awareness is a crucial yet underexplored perspective for the future design of reasoning agents in real-world scenarios. Our code is available at https://github.com/deeplearning-wisc/Thinking-Agent.

Method: Formulates layer pruning as cooperative game where layers are players and model performance is utility. Uses lightweight surrogate network to estimate Shapley values (layer contributions) and stratified Monte Carlo mask sampling to reduce computation cost.

Result: Method consistently outperforms others in perplexity and zero-shot accuracy, achieving more efficient and effective layer-wise pruning for LLMs.

Conclusion: Game-theoretic approach with surrogate network estimation enables dynamic identification of critical layers while capturing inter-layer dependencies, improving LLM pruning efficiency.

Abstract: While large language models (LLMs) demonstrate impressive performance across various tasks, their deployment in real-world scenarios is still constrained by high computational demands. Layer-wise pruning, a commonly employed strategy to mitigate inference costs, can partially address this challenge. However, existing approaches generally depend on static heuristic rules and fail to account for the interdependencies among layers, thereby limiting the effectiveness of the pruning process. To this end, this paper proposes a game-theoretic framework that formulates layer pruning as a cooperative game in which each layer acts as a player and model performance serves as the utility. As computing exact Shapley values is computationally infeasible for large language models (LLMs), we propose using a lightweight surrogate network to estimate layer-wise marginal contributions. This network can predict LLM performance for arbitrary layer combinations at a low computational cost. Additionally, we employ stratified Monte Carlo mask sampling to further reduce the cost of Sharpley value estimation. This approach captures inter-layer dependencies and dynamically identifies critical layers for pruning. Extensive experiments demonstrate the consistent superiority of our method in terms of perplexity and zero-shot accuracy, achieving more efficient and effective layer-wise pruning for large language models.

Method: Introduces HiddenBench, a 65-task benchmark based on the Hidden Profile paradigm, which tests collective reasoning under distributed information. Evaluates 15 frontier LLMs, comparing multi-agent performance under distributed information vs single agents with complete information. Analyzes failure modes across prompting strategies, communication depths, and group sizes.

Result: Multi-agent LLMs achieve only 30.1% accuracy under distributed information, compared to 80.7% for single agents with complete information. The gap stems from systematic failure to recognize or act under latent information asymmetry - agents can’t reason about what others might know but haven’t expressed. Failures persist across variations and worsen with group scaling. Model scale and individual reasoning accuracy don’t reliably predict collective performance.

Conclusion: Collective information exploration in decision-making is a key limitation of multi-agent LLMs. The Hidden Profile paradigm provides a theory-grounded, reproducible framework for diagnosing collective reasoning failures in LLM-based multi-agent systems.

Abstract: Multi-agent systems built on large language models (LLMs) are expected to enhance decision-making by pooling distributed information, yet systematically evaluating this capability has remained challenging. We introduce HiddenBench, a 65-task benchmark grounded in the Hidden Profile paradigm, which isolates collective reasoning under distributed information from individual reasoning ability. Evaluating 15 frontier LLMs, we find that multi-agent LLMs achieve only 30.1% accuracy under distributed information, compared to 80.7% accuracy for single agents given complete information. We trace this gap to a systematic failure mode: agents cannot recognize or act under latent information asymmetry-they fail to reason about what others might know but have not yet expressed, leading to premature convergence on shared evidence while critical distributed facts remain unexplored. These failures persist across prompting strategies, communication depths, and group sizes-and worsen as groups scale. While some models (e.g., Gemini-2.5-Flash/Pro) outperform others, neither model scale nor individual reasoning accuracy reliably predicts collective performance. Our results identify failures in collective information exploration in decision-making as a key limitation of multi-agent LLMs, and provide a theory-grounded, reproducible framework for diagnosing collective reasoning failures.

Method: Probe hidden states of open-source LLMs using linear projections to extract log-magnitudes of numbers, train linear classifiers to rank numeral pairs, and use probe log-loss as auxiliary objective during fine-tuning.

Result: Linear projections recover numbers with 2.3% error on synthetic text and 19.06% on scientific papers; ranking classifiers achieve >90% accuracy, but explicit verbal comparisons only reach 50-70% accuracy; fine-tuning with probe objective improves verbal accuracy by 3.22%.

Conclusion: LLMs encode numerical magnitude internally but fail to use this knowledge effectively in explicit reasoning; improving internal representations can enhance numerical reasoning capabilities.

Abstract: Although state-of-the-art LLMs can solve math problems, we find that they make errors on numerical comparisons with mixed notation: “Which is larger, $5.7 \times 10^2$ or $580$?” This raises a fundamental question: Do LLMs even know how big these numbers are? We probe the hidden states of several smaller open-source LLMs. A single linear projection of an appropriate hidden layer encodes the log-magnitudes of both kinds of numerals, allowing us to recover the numbers with relative error of about 2.3% (on restricted synthetic text) or 19.06% (on scientific papers). Furthermore, the hidden state after reading a pair of numerals encodes their ranking, with a linear classifier achieving over 90% accuracy. Yet surprisingly, when explicitly asked to rank the same pairs of numerals, these LLMs achieve only 50-70% accuracy, with worse performance for models whose probes are less effective. Finally, we show that incorporating the classifier probe’s log-loss as an auxiliary objective during finetuning brings an additional 3.22% improvement in verbalized accuracy over base models, demonstrating that improving models’ internal magnitude representations can enhance their numerical reasoning capabilities.

Method: Proposes hybrid bottom-up long-form story generation using multi-agent simulations. Agents interact within dynamic sandbox environments where their behaviors and interactions generate emergent events that form the story foundation, enabling natural plot progression.

Result: Achieves state-of-the-art performance across several metrics, capable of generating coherent and consistent stories exceeding 10,000 words, addressing key challenges faced by current story generation models.

Conclusion: The hybrid bottom-up approach offers a scalable and innovative solution for creating dynamic, immersive long-form stories that evolve organically from agent-driven interactions, enabling more natural storytelling.

Abstract: Human writers often begin their stories with an overarching mental scene, where they envision the interactions between characters and their environment. Inspired by this creative process, we propose a novel approach to long-form story generation, termed hybrid bottom-up long-form story generation, using multi-agent simulations. In our method, agents interact within a dynamic sandbox environment, where their behaviors and interactions with one another and the environment generate emergent events. These events form the foundation for the story, enabling organic character development and plot progression. Unlike traditional top-down approaches that impose rigid structures, our hybrid bottom-up approach allows for the natural unfolding of events, fostering more spontaneous and engaging storytelling. The system is capable of generating stories exceeding 10,000 words while maintaining coherence and consistency, addressing some of the key challenges faced by current story generation models. We achieve state-of-the-art performance across several metrics. This approach offers a scalable and innovative solution for creating dynamic, immersive long-form stories that evolve organically from agent-driven interactions.

Method: 1) Construct PlanFactory - a modular design space standardizing diverse planning paradigms within a unified codebase; 2) Collect planning trajectories; 3) Train Todo-14B model using Impedance-Guided Preference Optimization (IGPO), a multi-objective RL objective encouraging performant, stable, and token-efficient planning systems.

Result: Empirical evaluations on five agentic benchmarks demonstrate that TodoEvolve consistently surpasses carefully engineered planning modules while maintaining economical API costs and runtime overhead.

Conclusion: TodoEvolve provides a flexible meta-planning paradigm that autonomously synthesizes adaptive planning architectures, addressing limitations of fixed planning structures in agent systems.

Abstract: Planning has become a central capability for contemporary agent systems in navigating complex, long-horizon tasks, yet existing approaches predominantly rely on fixed, hand-crafted planning structures that lack the flexibility to adapt to the structural diversity of open-ended problems. To address this limitation, we introduce TodoEvolve, a meta-planning paradigm that autonomously synthesizes and dynamically revises task-specific planning architectures. Specifically, we first construct PlanFactory, a modular design space that standardizes diverse planning paradigms within a unified codebase encompassing topology, initialization, adaptation, and navigation, thereby providing a common interface for heterogeneous planning patterns. Leveraging PlanFactory, we collect high-quality planning trajectories and train Todo-14B via \textit{Impedance-Guided Preference Optimization} (IGPO), a multi-objective reinforcement learning objective that encourages the generation of planning systems that are performant, stable, and token-efficient across arbitrary tasks and agent backbones. Empirical evaluations on five agentic benchmarks demonstrate that TodoEvolve consistently surpasses carefully engineered planning modules while maintaining economical API costs and runtime overhead.

Method: 1) Introduces MACE benchmark with 12,000 factual questions across six domains with varying numbers of correct answers. 2) Evaluates 15 calibration methods across 4 LLM families (7B-72B). 3) Proposes Semantic Confidence Aggregation (SCA) which aggregates confidence over multiple high-probability sampled responses to address the miscalibration issue.

Result: Experiments show that while accuracy increases with answer cardinality, estimated confidence consistently decreases, causing severe miscalibration for questions with mixed answer counts. SCA achieves state-of-the-art calibration performance under mixed-answer settings while maintaining strong calibration on single-answer questions.

Conclusion: The paper identifies a critical limitation in current LLM confidence calibration methods when multiple valid answers exist, introduces a comprehensive benchmark for studying this phenomenon, and proposes an effective solution (SCA) that significantly improves calibration performance in such scenarios.

Abstract: Confidence calibration is essential for making large language models (LLMs) reliable, yet existing training-free methods have been primarily studied under single-answer question answering. In this paper, we show that these methods break down in the presence of multiple valid answers, where disagreement among equally correct responses leads to systematic underestimation of confidence. To enable a systematic study of this phenomenon, we introduce MACE, a benchmark of 12,000 factual questions spanning six domains with varying numbers of correct answers. Experiments across 15 representative calibration methods and four LLM families (7B-72B) reveal that while accuracy increases with answer cardinality, estimated confidence consistently decreases, causing severe miscalibration for questions with mixed answer counts. To address this issue, we propose Semantic Confidence Aggregation (SCA), which aggregates confidence over multiple high-probability sampled responses. SCA achieves state-of-the-art calibration performance under mixed-answer settings while preserving strong calibration on single-answer questions.

Method: Formulates efficient benchmarking as a sparse optimization problem. Uses gradient descent to optimize anchor weights and employs iterative refinement for anchor selection. Utilizes MLP representation capacity for sparse optimization and proposes Anchor Importance Score and Candidate Importance Score for task-aware refinement.

Result: Extensive experiments show low estimation error and high Kendall’s τ across various benchmarks, demonstrating superior robustness and practicality in real-world scenarios.

Conclusion: SparseEval provides an effective approach for efficient benchmarking of LLMs by selecting representative anchor items through sparse optimization, significantly reducing computational costs while maintaining evaluation accuracy.

Abstract: As large language models (LLMs) continue to scale up, their performance on various downstream tasks has significantly improved. However, evaluating their capabilities has become increasingly expensive, as performing inference on a large number of benchmark samples incurs high computational costs. In this paper, we revisit the model-item performance matrix and show that it exhibits sparsity, that representative items can be selected as anchors, and that the task of efficient benchmarking can be formulated as a sparse optimization problem. Based on these insights, we propose SparseEval, a method that, for the first time, adopts gradient descent to optimize anchor weights and employs an iterative refinement strategy for anchor selection. We utilize the representation capacity of MLP to handle sparse optimization and propose the Anchor Importance Score and Candidate Importance Score to evaluate the value of each item for task-aware refinement. Extensive experiments demonstrate the low estimation error and high Kendall’s~$τ$ of our method across a variety of benchmarks, showcasing its superior robustness and practicality in real-world scenarios. Code is available at {https://github.com/taolinzhang/SparseEval}.

Method: The authors use causal mediation analysis to investigate instruction representation localization. They identify “Instruction Vectors” (IVs) and propose a novel method to localize information processing in language models that avoids the implicit linear assumptions of patching-based techniques. They analyze how task representations formed in early layers select different information pathways in later layers.

Result: The study finds that instruction representation is fairly localized in models, with Instruction Vectors demonstrating both linear separability and non-linear causal interaction. These IVs act as circuit selectors - conditioned on task representations formed in early layers, different information pathways are selected in later layers to solve specific tasks.

Conclusion: The research challenges the linear representation hypothesis common in mechanistic interpretability by showing that instruction processing involves both linear separability and non-linear causal interactions. The findings provide insights into how instruction-tuned models internally process and route information based on task representations.

Abstract: Despite the recent success of instruction-tuned language models and their ubiquitous usage, very little is known of how models process instructions internally. In this work, we address this gap from a mechanistic point of view by investigating how instruction-specific representations are constructed and utilized in different stages of post-training: Supervised Fine-Tuning (SFT) and Direct Preference Optimization (DPO). Via causal mediation, we identify that instruction representation is fairly localized in models. These representations, which we call Instruction Vectors (IVs), demonstrate a curious juxtaposition of linear separability along with non-linear causal interaction, broadly questioning the scope of the linear representation hypothesis commonplace in mechanistic interpretability. To disentangle the non-linear causal interaction, we propose a novel method to localize information processing in language models that is free from the implicit linear assumptions of patching-based techniques. We find that, conditioned on the task representations formed in the early layers, different information pathways are selected in the later layers to solve that task, i.e., IVs act as circuit selectors.

Method: Developed two model variants based on Polish language models (MMLW-RoBERTa-base and PKOBP/polish-roberta-8k), fine-tuned on a community-annotated dataset of 6,885 Polish texts across five safety categories.

Result: Both models achieve strong performance, with the 0.5B variant showing best overall discrimination (F1: 0.791 micro, 0.785 macro) and the 0.1B variant demonstrating exceptional efficiency and superior precision (77.65%) with low false positive rate (0.63%) on real user prompts.

Conclusion: Bielik Guard provides effective Polish language safety classification for LLM applications, with models designed to provide appropriate responses rather than simple content blocking, particularly for sensitive categories like self-harm.

Abstract: As Large Language Models (LLMs) become increasingly deployed in Polish language applications, the need for efficient and accurate content safety classifiers has become paramount. We present Bielik Guard, a family of compact Polish language safety classifiers comprising two model variants: a 0.1B parameter model based on MMLW-RoBERTa-base and a 0.5B parameter model based on PKOBP/polish-roberta-8k. Fine-tuned on a community-annotated dataset of 6,885 Polish texts, these models classify content across five safety categories: Hate/Aggression, Vulgarities, Sexual Content, Crime, and Self-Harm. Our evaluation demonstrates that both models achieve strong performance on multiple benchmarks. The 0.5B variant offers the best overall discrimination capability with F1 scores of 0.791 (micro) and 0.785 (macro) on the test set, while the 0.1B variant demonstrates exceptional efficiency. Notably, Bielik Guard 0.1B v1.1 achieves superior precision (77.65%) and very low false positive rate (0.63%) on real user prompts, outperforming HerBERT-PL-Guard (31.55% precision, 4.70% FPR) despite identical model size. The models are publicly available and designed to provide appropriate responses rather than simple content blocking, particularly for sensitive categories like self-harm.

Method: Introduces CompositeHarm benchmark combining two English datasets (AttaQ for structured adversarial attacks and MMSafetyBench for contextual real-world harms) and extends them into six languages (English, Hindi, Assamese, Marathi, Kannada, Gujarati). Uses three large models with lightweight inference strategies inspired by edge-AI design principles for computational efficiency.

Result: Attack success rates rise sharply in Indic languages, especially under adversarial syntax, while contextual harms transfer more moderately. The lightweight inference strategies enable large-scale multilingual safety testing while being computationally feasible and environmentally conscious.

Conclusion: Translated benchmarks are a necessary first step but not sufficient for building grounded, resource-aware, language-adaptive safety systems. Multilingual safety evaluation requires approaches that account for language-specific variations in harmful content.

Abstract: Most safety evaluations of large language models (LLMs) remain anchored in English. Translation is often used as a shortcut to probe multilingual behavior, but it rarely captures the full picture, especially when harmful intent or structure morphs across languages. Some types of harm survive translation almost intact, while others distort or disappear. To study this effect, we introduce CompositeHarm, a translation-based benchmark designed to examine how safety alignment holds up as both syntax and semantics shift. It combines two complementary English datasets, AttaQ, which targets structured adversarial attacks, and MMSafetyBench, which covers contextual, real-world harms, and extends them into six languages: English, Hindi, Assamese, Marathi, Kannada, and Gujarati. Using three large models, we find that attack success rates rise sharply in Indic languages, especially under adversarial syntax, while contextual harms transfer more moderately. To ensure scalability and energy efficiency, our study adopts lightweight inference strategies inspired by edge-AI design principles, reducing redundant evaluation passes while preserving cross-lingual fidelity. This design makes large-scale multilingual safety testing both computationally feasible and environmentally conscious. Overall, our results show that translated benchmarks are a necessary first step, but not a sufficient one, toward building grounded, resource-aware, language-adaptive safety systems.

Method: SynCog integrates controllable zero-shot multimodal data synthesis to simulate diverse virtual subjects with varying cognitive profiles, and Chain-of-Thought deduction fine-tuning to enable explicit diagnostic reasoning. It generates synthetic clinical data across languages and fine-tunes multimodal LLMs with CoT strategies.

Result: Achieved Macro-F1 scores of 80.67% on ADReSS and 78.46% on ADReSSo benchmarks, outperforming baselines. Demonstrated cross-linguistic generalization with 48.71% Macro-F1 on independent Mandarin cohort (CIR-E).

Conclusion: SynCog addresses key barriers in speech-based MCI diagnosis by alleviating data scarcity, providing interpretable reasoning, and enabling cross-lingual generalization, representing progress toward clinically trustworthy global cognitive assessment tools.

Abstract: Speech-based digital biomarkers represent a scalable, non-invasive frontier for the early identification of Mild Cognitive Impairment (MCI). However, the development of robust diagnostic models remains impeded by acute clinical data scarcity and a lack of interpretable reasoning. Current solutions frequently struggle with cross-lingual generalization and fail to provide the transparent rationales essential for clinical trust. To address these barriers, we introduce SynCog, a novel framework integrating controllable zero-shot multimodal data synthesis with Chain-of-Thought (CoT) deduction fine-tuning. Specifically, SynCog simulates diverse virtual subjects with varying cognitive profiles to effectively alleviate clinical data scarcity. This generative paradigm enables the rapid, zero-shot expansion of clinical corpora across diverse languages, effectively bypassing data bottlenecks in low-resource settings and bolstering the diagnostic performance of Multimodal Large Language Models (MLLMs). Leveraging this synthesized dataset, we fine-tune a foundational multimodal backbone using a CoT deduction strategy, empowering the model to explicitly articulate diagnostic thought processes rather than relying on black-box predictions. Extensive experiments on the ADReSS and ADReSSo benchmarks demonstrate that augmenting limited clinical data with synthetic phenotypes yields competitive diagnostic performance, achieving Macro-F1 scores of 80.67% and 78.46%, respectively, outperforming current baseline models. Furthermore, evaluation on an independent real-world Mandarin cohort (CIR-E) demonstrates robust cross-linguistic generalization, attaining a Macro-F1 of 48.71%. These findings constitute a critical step toward providing clinically trustworthy and linguistically inclusive cognitive assessment tools for global healthcare.

Method: Controlled cue perturbations with synthetic metadata labels injected into evaluation prompts for six judge models (GPT-4o, Gemini-2.0-Flash, etc.) across two datasets: ELI5 (factual QA) and LitBench (creative writing). Studied six cue families: source, temporal, age, gender, ethnicity, and educational status. Measured verdict shift rates (VSR) and introduced cue acknowledgment rate (CAR) to quantify explicit cue references in rationales.

Result: LLM judges show substantial sensitivity to irrelevant cues (provenance hierarchies, recency preferences, educational-status favoritism) but CAR is typically near zero, indicating shortcut reliance is largely unreported. CAR is dataset-dependent - more likely in factual ELI5 but collapses in open-ended LitBench, where large verdict shifts persist despite zero acknowledgment.

Conclusion: The combination of substantial verdict sensitivity and limited cue acknowledgment reveals an explanation gap in LLM-as-judge pipelines, raising reliability concerns for model-based evaluation in research and deployment.

Abstract: Large language models (LLMs) are increasingly used as automatic judges to evaluate system outputs in tasks such as reasoning, question answering, and creative writing. A faithful judge should base its verdicts solely on content quality, remain invariant to irrelevant context, and transparently reflect the factors driving its decisions. We test this ideal via controlled cue perturbations-synthetic metadata labels injected into evaluation prompts-for six judge models: GPT-4o, Gemini-2.0-Flash, Gemma-3-27B, Qwen3-235B, Claude-3-Haiku, and Llama3-70B. Experiments span two complementary datasets with distinct evaluation regimes: ELI5 (factual QA) and LitBench (open-ended creative writing). We study six cue families: source, temporal, age, gender, ethnicity, and educational status. Beyond measuring verdict shift rates (VSR), we introduce cue acknowledgment rate (CAR) to quantify whether judges explicitly reference the injected cues in their natural-language rationales. Across cues with strong behavioral effects-e.g., provenance hierarchies (Expert > Human > LLM > Unknown), recency preferences (New > Old), and educational-status favoritism-CAR is typically at or near zero, indicating that shortcut reliance is largely unreported even when it drives decisions. Crucially, CAR is also dataset-dependent: explicit cue recognition is more likely to surface in the factual ELI5 setting for some models and cues, but often collapses in the open-ended LitBench regime, where large verdict shifts can persist despite zero acknowledgment. The combination of substantial verdict sensitivity and limited cue acknowledgment reveals an explanation gap in LLM-as-judge pipelines, raising concerns about reliability of model-based evaluation in both research and deployment.

Method: DeltaKV encodes semantic residuals relative to retrieved historical references instead of discarding tokens, preserving fidelity while reducing storage. Sparse-vLLM provides decoupled memory management and kernels optimized for sparse KV layouts.

Result: DeltaKV reduces KV cache memory to 29% of original while maintaining near-lossless accuracy on LongBench, SCBench, and AIME. With Sparse-vLLM, achieves up to 2× throughput improvement over vLLM in long-context scenarios.

Conclusion: The approach demonstrates a practical path toward scalable long-context LLM deployment by balancing compression efficiency with hardware acceleration.

Abstract: The deployment of efficient long-context LLMs in applications like autonomous agents, long-chain reasoning, and creative writing is fundamentally bottlenecked by the linear growth of KV cache memory. Existing compression and eviction methods often struggle to balance accuracy, compression ratio, and hardware efficiency. We propose DeltaKV, a residual-based KV cache compression framework motivated by two empirical findings: long-range inter-token similarity and highly shared latent components in KV representations. Instead of discarding tokens, DeltaKV encodes semantic residuals relative to retrieved historical references, preserving fidelity while substantially reducing storage. To translate compression gains into real system speedups, we further introduce Sparse-vLLM, a high-performance inference engine with decoupled memory management and kernels optimized for sparse and irregular KV layouts. Experiments show that DeltaKV reduces KV cache memory to 29% of the original while maintaining near-lossless accuracy on LongBench, SCBench, and AIME. When integrated with Sparse-vLLM, it achieves up to 2$\times$ throughput improvement over vLLM in long-context scenarios, demonstrating a practical path toward scalable long-context LLM deployment. Code, model checkpoints, and datasets are available at https://github.com/CURRENTF/Sparse-vLLM.

Method: Diverge-to-Induce Prompting (DIP) framework: 1) Generate multiple diverse high-level rationales for each question, 2) Elaborate each rationale into detailed step-by-step draft plans, 3) Induce these draft plans into a final comprehensive plan.

Result: DIP outperforms single-strategy prompting methods, demonstrating improved zero-shot reasoning accuracy without relying on resource-intensive sampling techniques.

Conclusion: Multi-plan induction through diverse reasoning strategies enhances prompt-based reasoning, providing a more robust approach than single-strategy methods for diverse tasks.

Abstract: To address the instability of unguided reasoning paths in standard Chain-of-Thought prompting, recent methods guide large language models (LLMs) by first eliciting a single reasoning strategy. However, relying on just one strategy for each question can still limit performance across diverse tasks. We propose Diverge-to-Induce Prompting (DIP), a framework that first prompts an LLM to generate multiple diverse high-level rationales for each question. Each rationale is then elaborated into a detailed, step-by-step draft plan. Finally, these draft plans are induced into a final plan. DIP enhances zero-shot reasoning accuracy without reliance on resource-intensive sampling. Experiments show that DIP outperforms single-strategy prompting, demonstrating the effectiveness of multi-plan induction for prompt-based reasoning.

Method: 1) Analyzes metric-based methods within unified framework, identifying core challenge of biased token-level scores. 2) Reveals two relationships for calibration: Neighbor Similarity and Initial Instability. 3) Proposes Markov-informed score calibration using Markov random fields with mean-field approximation as lightweight component.

Result: Extensive experiments show significant gains over baselines in cross-LLM and paraphrasing attack scenarios with negligible computational overhead.

Conclusion: The proposed Markov random field calibration effectively addresses score bias in machine-generated text detection and can be seamlessly integrated into existing detectors.

Abstract: While machine-generated texts (MGTs) offer great convenience, they also pose risks such as disinformation and phishing, highlighting the need for reliable detection. Metric-based methods, which extract statistically distinguishable features of MGTs, are often more practical than complex model-based methods that are prone to overfitting. Given their diverse designs, we first place representative metric-based methods within a unified framework, enabling a clear assessment of their advantages and limitations. Our analysis identifies a core challenge across these methods: the token-level detection score is easily biased by the inherent randomness of the MGTs generation process. To address this, we theoretically and empirically reveal two relationships of context detection scores that may aid calibration: Neighbor Similarity and Initial Instability. We then propose a Markov-informed score calibration strategy that models these relationships using Markov random fields, and implements it as a lightweight component via a mean-field approximation, allowing our method to be seamlessly integrated into existing detectors. Extensive experiments in various real-world scenarios, such as cross-LLM and paraphrasing attacks, demonstrate significant gains over baselines with negligible computational overhead. The code is available at https://github.com/tmlr-group/MRF_Calibration.

Method: TDGNet formulates hallucination detection as learning over evolving token-level attention graphs. At each denoising step, it sparsifies the attention graph, updates per-token memories via message passing, and applies temporal attention to aggregate trajectory-wide evidence for final prediction.

Result: Experiments on LLaDA-8B and Dream-7B across QA benchmarks show consistent AUROC improvements over output-based, latent-based, and static-graph baselines, with single-pass inference and modest overhead.

Conclusion: Temporal reasoning on attention graphs is crucial for robust hallucination detection in diffusion language models, as demonstrated by TDGNet’s superior performance.

Abstract: Diffusion language models (D-LLMs) offer parallel denoising and bidirectional context, but hallucination detection for D-LLMs remains underexplored. Prior detectors developed for auto-regressive LLMs typically rely on single-pass cues and do not directly transfer to diffusion generation, where factuality evidence is distributed across the denoising trajectory and may appear, drift, or be self-corrected over time. We introduce TDGNet, a temporal dynamic graph framework that formulates hallucination detection as learning over evolving token-level attention graphs. At each denoising step, we sparsify the attention graph and update per-token memories via message passing, then apply temporal attention to aggregate trajectory-wide evidence for final prediction. Experiments on LLaDA-8B and Dream-7B across QA benchmarks show consistent AUROC improvements over output-based, latent-based, and static-graph baselines, with single-pass inference and modest overhead. These results highlight the importance of temporal reasoning on attention graphs for robust hallucination detection in diffusion language models.

Method: Analyze latent reasoning transformers (LRTs) that perform deliberation entirely in continuous hidden space, decoding the model’s evolving beliefs at every step on multiple-choice QA benchmarks to understand the structured search process.

Result: LRTs learn a structured search process with consistent trajectory: exploration phase spreading probability mass, tentative commitment to frontrunner, and convergence or backtracking. Backtracking is prevalent (32%), beneficial (34% accuracy gain), and directed away from semantically closest distractors toward correct answers. Search is adaptive - replacing distractors with implausible alternatives shortens exploration by 54%.

Conclusion: Latent reasoning models achieve in activation space what chain-of-thought achieves through words: the ability to be wrong, notice, and recover, demonstrating that structured reasoning can occur entirely in continuous hidden representations without verbalization.

Abstract: What happens when a language model thinks without words? Standard reasoning LLMs verbalize intermediate steps as chain-of-thought; latent reasoning transformers (LRTs) instead perform deliberation entirely in continuous hidden space. We investigate an LRT, decoding the model’s evolving beliefs at every step on a multiple-choice QA benchmark. We find that the model spontaneously learns a structured search process in latent space. Deliberation follows a consistent trajectory: an exploration phase where probability mass spreads across candidates, tentative commitment to a frontrunner, and either convergence or backtracking. Backtracking is prevalent (32% of instances), beneficial (34% accuracy gain over non-backtracking instances), and predominantly directed away from the semantically closest distractor toward the correct answer. The search is adaptive: replacing distractors with implausible alternatives shortens exploration by 54%. Latent reasoning models achieve in activation space what chain-of-thought achieves through words: the ability to be wrong, notice, and recover.

Method: Used prompt engineering to elicit product suggestions from LLMs for various race and gender groups, then employed three analytical methods: Marked Words, Support Vector Machines, and Jensen-Shannon Divergence to identify and quantify biases.

Result: Findings reveal significant disparities in recommendations for different demographic groups, demonstrating that LLMs exhibit systematic biases in their product suggestions.

Conclusion: There is a clear need for more equitable LLM recommendation systems to address the identified gender and race biases in product recommendations.

Abstract: Large Language Models are increasingly employed in generating consumer product recommendations, yet their potential for embedding and amplifying gender and race biases remains underexplored. This paper serves as one of the first attempts to examine these biases within LLM-generated recommendations. We leverage prompt engineering to elicit product suggestions from LLMs for various race and gender groups and employ three analytical methods-Marked Words, Support Vector Machines, and Jensen-Shannon Divergence-to identify and quantify biases. Our findings reveal significant disparities in the recommendations for demographic groups, underscoring the need for more equitable LLM recommendation systems.

Method: Proposes DIAL-SUMMER framework with hierarchical error taxonomy at dialogue-level (speakers/turns) and within-turn-level (information inside turns), plus a manually annotated dataset of dialogue summaries with fine-grained errors.

Result: Empirical analysis reveals trends like middle dialogue turns being most frequently missed and extrinsic hallucinations occurring mostly at summary ends. LLM-Judges struggle with error detection, showing dataset’s challenging nature.

Conclusion: DIAL-SUMMER provides robust taxonomy and challenging dataset for dialogue summary evaluation, highlighting need for future work to improve LLMs’ performance on this task.

Abstract: Dialogues are a predominant mode of communication for humans, and it is immensely helpful to have automatically generated summaries of them (e.g., to revise key points discussed in a meeting, to review conversations between customer agents and product users). Prior works on dialogue summary evaluation largely ignore the complexities specific to this task: (i) shift in structure, from multiple speakers discussing information in a scattered fashion across several turns, to a summary’s sentences, and (ii) shift in narration viewpoint, from speakers’ first/second-person narration, standardized third-person narration in the summary. In this work, we introduce our framework DIALSUMMER to address the above. We propose DIAL-SUMMER’s taxonomy of errors to comprehensively evaluate dialogue summaries at two hierarchical levels: DIALOGUE-LEVEL that focuses on the broader speakers/turns, and WITHIN-TURN-LEVEL that focuses on the information talked about inside a turn. We then present DIAL-SUMMER’s dataset composed of dialogue summaries manually annotated with our taxonomy’s fine-grained errors. We conduct empirical analyses of these annotated errors, and observe interesting trends (e.g., turns occurring in middle of the dialogue are the most frequently missed in the summary, extrinsic hallucinations largely occur at the end of the summary). We also conduct experiments on LLM-Judges’ capability at detecting these errors, through which we demonstrate the challenging nature of our dataset, the robustness of our taxonomy, and the need for future work in this field to enhance LLMs’ performance in the same. Code and inference dataset coming soon.

Method: Review article synthesizing existing NLP approaches for three core tasks: document segmentation (breaking down lengthy deliberations), domain-specific entity extraction (identifying political actors and personal information), and automatic text summarization (generating concise representations of decision-making processes).

Result: Provides a structured overview of methodological approaches, evaluation metrics, and publicly available resources for applying NLP to governance documents, while highlighting domain-specific challenges like data scarcity, privacy constraints, and source variability.

Conclusion: NLP offers well-established methods to enhance the accessibility and interpretability of governmental records, and this review provides guidance on how computational methods can structure and interpret complex local governance meeting documents.

Abstract: Local governance meeting records are official documents, in the form of minutes or transcripts, documenting how proposals, discussions, and procedural actions unfold during institutional meetings. While generally structured, these documents are often dense, bureaucratic, and highly heterogeneous across municipalities, exhibiting significant variation in language, terminology, structure, and overall organization. This heterogeneity makes them difficult for non-experts to interpret and challenging for intelligent automated systems to process, limiting public transparency and civic engagement. To address these challenges, computational methods can be employed to structure and interpret such complex documents. In particular, Natural Language Processing (NLP) offers well-established methods that can enhance the accessibility and interpretability of governmental records. In this focus article, we review foundational NLP tasks that support the structuring of local governance meeting documents. Specifically, we review three core tasks: document segmentation, domain-specific entity extraction and automatic text summarization, which are essential for navigating lengthy deliberations, identifying political actors and personal information, and generating concise representations of complex decision-making processes. In reviewing these tasks, we discuss methodological approaches, evaluation metrics, and publicly available resources, while highlighting domain-specific challenges such as data scarcity, privacy constraints, and source variability. By synthesizing existing work across these foundational tasks, this article provides a structured overview of how NLP can enhance the structuring and accessibility of local governance meeting records.

Method: Conducted multimodal communication games comparing same-type dyads (human-human, AI-AI) and heterogeneous human-AI pairs. In Experiment 2, tested whether prompting LLMs to produce superficially humanlike behavior improves alignment.

Result: Both humans and LLMs show convention-formation in same-type dyads (increased accuracy, consistency, decreased message length). Human-AI pairs fail to align effectively. Even when LLMs mimic human message length, accuracy and lexical overlap in human-LLM pairs lag behind both human-human and AI-AI pairs.

Conclusion: Conversational alignment requires more than just mimicking previous interactions; it requires shared interpretative biases toward conveyed meanings. LLMs and humans have fundamentally different communicative tendencies that hinder effective cross-type alignment.

Abstract: Humans align to one another in conversation – adopting shared conventions that ease communication. We test whether LLMs form the same kinds of conventions in a multimodal communication game. Both humans and LLMs display evidence of convention-formation (increasing the accuracy and consistency of their turns while decreasing their length) when communicating in same-type dyads (humans with humans, AI with AI). However, heterogenous human-AI pairs fail – suggesting differences in communicative tendencies. In Experiment 2, we ask whether LLMs can be induced to behave more like human conversants, by prompting them to produce superficially humanlike behavior. While the length of their messages matches that of human pairs, accuracy and lexical overlap in human-LLM pairs continues to lag behind that of both human-human and AI-AI pairs. These results suggest that conversational alignment requires more than just the ability to mimic previous interactions, but also shared interpretative biases toward the meanings that are conveyed.

Method: Pretraining with Token-Level Adaptive Latent CoT where model generates variable-length latent CoT trajectories before emitting each token, with adaptive halting that allocates longer trajectories to difficult tokens and shorter/zero trajectories to easy ones, emerging naturally from one-stage pretraining on general text.

Result: Experiments with Llama architectures show adaptive latent CoT consistently improves language modeling perplexity and broad downstream accuracy, even with fewer training FLOPs than prior recurrent baselines.

Conclusion: Adaptive latent CoT provides an effective alternative scaling axis by increasing per-token computation without parameter expansion, offering computational efficiency benefits in both training and inference through token-wise adaptive halting.

Abstract: Scaling large language models by increasing parameters and training data is increasingly constrained by limited high-quality corpora and rising communication costs. This work explores an alternative axis: increasing per-token computation without expanding parameters, by internalizing latent Chain-of-Thought (CoT) into pretraining. We propose Pretraining with Token-Level Adaptive Latent CoT (adaptive latent CoT), where the model generates a variable-length latent CoT trajectory before emitting each token – allocating longer trajectories to difficult tokens and shorter (or even zero) trajectories to easy ones. Importantly, this behavior emerges naturally from one-stage pretraining on general text and reduces computation in both training and inference via token-wise adaptive halting. Experiments with Llama architectures show that adaptive latent CoT consistently improves language modeling perplexity and broad downstream accuracy, even with fewer training FLOPs than prior recurrent baselines.

Method: CoRect uses layer-wise analysis to identify parametric suppression in deep FFN layers. It contrasts logits from contextualized and non-contextualized forward passes to detect layers with high parametric bias without ground-truth labels, then rectifies hidden states to preserve evidence-grounded information.

Result: Across question answering and summarization benchmarks, CoRect consistently improves faithfulness and reduces hallucinations compared to strong baselines.

Conclusion: CoRect effectively addresses parametric suppression in RAG systems through context-aware logit contrast, improving faithfulness without requiring ground-truth targets.

Abstract: Retrieval-Augmented Generation (RAG) often struggles with knowledge conflicts, where model-internal parametric knowledge overrides retrieved evidence, leading to unfaithful outputs. Existing approaches are often limited, relying either on superficial decoding adjustments or weight editing that necessitates ground-truth targets. Through layer-wise analysis, we attribute this failure to a parametric suppression phenomenon: specifically, in deep layers, certain FFN layers overwrite context-sensitive representations with memorized priors. To address this, we propose CoRect (Context-Aware Logit Contrast for Hidden State Rectification). By contrasting logits from contextualized and non-contextualized forward passes, CoRect identifies layers that exhibit high parametric bias without requiring ground-truth labels. It then rectifies the hidden states to preserve evidence-grounded information. Across question answering (QA) and summarization benchmarks, CoRect consistently improves faithfulness and reduces hallucinations compared to strong baselines.

Method: Proposes AutoElicit framework that iteratively perturbs benign instructions using CUA execution feedback to elicit severe harms while keeping perturbations realistic and benign. The framework automatically surfaces unintended behaviors through systematic exploration.

Result: Surfaced hundreds of harmful unintended behaviors from state-of-the-art CUAs (Claude 4.5 Haiku and Opus). Found persistent susceptibility to unintended behaviors across various frontier CUAs through transferability evaluation of human-verified successful perturbations.

Conclusion: Establishes a foundation for systematically analyzing unintended behaviors in realistic computer-use settings, providing the first conceptual and methodological framework for understanding and proactively discovering these risks.

Abstract: Although computer-use agents (CUAs) hold significant potential to automate increasingly complex OS workflows, they can demonstrate unsafe unintended behaviors that deviate from expected outcomes even under benign input contexts. However, exploration of this risk remains largely anecdotal, lacking concrete characterization and automated methods to proactively surface long-tail unintended behaviors under realistic CUA scenarios. To fill this gap, we introduce the first conceptual and methodological framework for unintended CUA behaviors, by defining their key characteristics, automatically eliciting them, and analyzing how they arise from benign inputs. We propose AutoElicit: an agentic framework that iteratively perturbs benign instructions using CUA execution feedback, and elicits severe harms while keeping perturbations realistic and benign. Using AutoElicit, we surface hundreds of harmful unintended behaviors from state-of-the-art CUAs such as Claude 4.5 Haiku and Opus. We further evaluate the transferability of human-verified successful perturbations, identifying persistent susceptibility to unintended behaviors across various other frontier CUAs. This work establishes a foundation for systematically analyzing unintended behaviors in realistic computer-use settings.

Method: Replace paragraphs with placeholders in long documents, train LLMs via RL to reconstruct documents by correctly identifying and sequencing missing paragraphs from candidate options. Captures global narrative coherence.

Result: Method validated on RULER and LongBench v2 benchmarks. Achieves noticeable gains on RULER and reasonable improvement on LongBench v2 without manually curated long-context QA data.

Conclusion: Unsupervised RL approach effectively enhances LLM long-context capabilities without costly supervision. Code, data, and models publicly released.

Abstract: Reinforcement Learning with Verifiable Rewards~(RLVR) has become a prominent paradigm to enhance the capabilities (i.e.\ long-context) of Large Language Models~(LLMs). However, it often relies on gold-standard answers or explicit evaluation rubrics provided by powerful teacher models or human experts, which are costly and time-consuming. In this work, we investigate unsupervised approaches to enhance the long-context capabilities of LLMs, eliminating the need for heavy human annotations or teacher models’ supervision. Specifically, we first replace a few paragraphs with special placeholders in a long document. LLMs are trained through reinforcement learning to reconstruct the document by correctly identifying and sequencing missing paragraphs from a set of candidate options. This training paradigm enables the model to capture global narrative coherence, significantly boosting long-context performance. We validate the effectiveness of our method on two widely used benchmarks, RULER and LongBenchv2. While acquiring noticeable gains on RULER, it can also achieve a reasonable improvement on LongBenchv2 without any manually curated long-context QA data. Furthermore, we conduct extensive ablation studies to analyze the impact of reward design, data curation strategies, training schemes, and data scaling effects on model performance. We publicly release our code, data, and models.

Method: Combines analytical and empirical analyses using the Information Bottleneck (IB) framework for semantic efficiency, examining convexity and efficiency properties in color naming systems and their ability to discriminate attested systems from hypothetical variants.

Result: Shows that convexity and efficiency are distinct (neither entails the other), but IB-optimal systems are mostly convex in color naming; efficiency is a stronger predictor for discriminating attested color naming systems, with convexity adding negligible improvement; efficiency accounts for empirical phenomena that convexity cannot.

Conclusion: While convexity and efficiency can yield similar structural observations, they are fundamentally distinct properties, with efficiency providing a more comprehensive account of semantic typology than convexity alone.

Abstract: There are two widely held characterizations of human semantic category systems: (1) they form convex partitions of conceptual spaces, and (2) they are efficient for communication. While prior work observed that convexity and efficiency co-occur in color naming, the analytical relation between them and why they co-occur have not been well understood. We address this gap by combining analytical and empirical analyses that build on the Information Bottleneck (IB) framework for semantic efficiency. First, we show that convexity and efficiency are distinct in the sense that neither entails the other: there are convex systems which are inefficient, and optimally-efficient systems that are non-convex. Crucially, however, the IB-optimal systems are mostly convex in the domain of color naming, explaining the main empirical basis for the convexity approach. Second, we show that efficiency is a stronger predictor for discriminating attested color naming systems from hypothetical variants, with convexity adding negligible improvement on top of that. Finally, we discuss a range of empirical phenomena that convexity cannot account for but efficiency can. Taken together, our work suggests that while convexity and efficiency can yield similar structural observations, they are fundamentally distinct, with efficiency providing a more comprehensive account of semantic typology.

Method: Cognitive Linguistic Identity Fusion Score (CLIFS) uses cognitive linguistic patterns, large language models, and implicit metaphor analysis to measure identity fusion from text. Validated across UK and Singapore datasets against established fusion measures.

Result: CLIFS outperforms existing methods in predicting validated fusion scores. Applied to extremist manifestos, reveals two distinct high-fusion pathways: ideologues frame themselves in terms of group (kinship bonds), while grievance-driven individuals frame group in terms of personal identity.

Conclusion: The method refines identity fusion theory and provides a scalable tool for fusion research and extremism detection through linguistic analysis.

Abstract: In light of increasing polarization and political violence, understanding the psychological roots of extremism is increasingly important. Prior research shows that identity fusion predicts willingness to engage in extreme acts. We evaluate the Cognitive Linguistic Identity Fusion Score, a method that uses cognitive linguistic patterns, LLMs, and implicit metaphor to measure fusion from language. Across datasets from the United Kingdom and Singapore, this approach outperforms existing methods in predicting validated fusion scores. Applied to extremist manifestos, two distinct high-fusion pathways to violence emerge: ideologues tend to frame themselves in terms of group, forming kinship bonds; whereas grievance-driven individuals frame the group in terms of their personal identity. These results refine theories of identity fusion and provide a scalable tool aiding fusion research and extremism detection.

Method: The approach involves decomposing paraphrases into paraphrase types (constituent linguistic aspects) and training models explicitly on these types rather than treating paraphrasing as a binary classification task.

Result: Models trained on paraphrase types achieve stronger performance: 89.6% accuracy vs 78.4% human baseline for Wikipedia plagiarism detection, 66.5% vs 55.7% for arXiv scientific paper plagiarism, and improved performance on Quora duplicate question identification.

Conclusion: Decomposing paraphrases into linguistic aspects provides a more fine-grained approach to semantic understanding, enabling models to better distinguish meaning-preserving variations and outperform traditional binary approaches on downstream applications.

Abstract: Language enables humans to share knowledge, reason about the world, and pass on strategies for survival and innovation across generations. At the heart of this process is not just the ability to communicate but also the remarkable flexibility in how we can express ourselves. We can express the same thoughts in virtually infinite ways using different words and structures - this ability to rephrase and reformulate expressions is known as paraphrase. Modeling paraphrases is a keystone to meaning in computational language models; being able to construct different variations of texts that convey the same meaning or not shows strong abilities of semantic understanding. If computational language models are to represent meaning, they must understand and control the different aspects that construct the same meaning as opposed to different meanings at a fine granularity. Yet most existing approaches reduce paraphrasing to a binary decision between two texts or to producing a single rewrite of a source, obscuring which linguistic factors are responsible for meaning preservation. In this thesis, I propose that decomposing paraphrases into their constituent linguistic aspects (paraphrase types) offers a more fine-grained and cognitively grounded view of semantic equivalence. I show that even advanced machine learning models struggle with this task. Yet, when explicitly trained on paraphrase types, models achieve stronger performance on related paraphrase tasks and downstream applications. For example, in plagiarism detection, language models trained on paraphrase types surpass human baselines: 89.6% accuracy compared to 78.4% for plagiarism cases from Wikipedia, and 66.5% compared to 55.7% for plagiarism of scientific papers from arXiv. In identifying duplicate questions on Quora, models trained with paraphrase types improve over models trained on binary pairs. Furthermore, I demonstrate that…

Method: Proposes a probabilistic framework defining capability by instance-level solvability. Uses Algebrarium framework to train models on single-step operations and evaluate on unseen multi-step tasks. Tests hypothesis that sharpening atomic step probabilities enables overcoming exponential decay in multi-step reasoning.

Result: Empirical results show: (1) RLVR incentivizes exploration of previously inaccessible solution paths; (2) composite performance governed by joint probability of atomic steps (high Pearson correlation ρ∈[0.69,0.96]); (3) RLVR as global optimizer can sacrifice specific skills to maximize aggregate reward.

Conclusion: RLVR enables emergent abilities by iterative optimization of solvable problems, allowing models to develop capabilities for previously unsolvable scenarios through probabilistic sharpening of atomic reasoning steps.

Abstract: Whether Reinforcement Learning with Verifiable Rewards (RLVR) endows Large Language Models (LLMs) with new capabilities or merely elicits latent traces remains a central debate. In this work, we align with the former view, proposing a probabilistic framework where capability is defined by instance-level solvability. We hypothesize that the emergence of complex reasoning can be driven by sharpening atomic step probabilities, which enables models to overcome the exponential decay of success rates inherent in multi-step reasoning chains. Utilizing the Algebrarium framework, we train models exclusively on single-step operations and evaluate their performance on unseen multi-step tasks. Our empirical results confirm that: (1) RLVR incentivizes the exploration of previously inaccessible solution paths by amplifying the model’s existing skills; (2) composite performance is strictly governed by the joint probability of atomic steps, evidenced by high Pearson correlation coefficients ($ρ\in [0.69, 0.96]$); and (3) RLVR, acting as a global optimizer, can cause specific skills to be sacrificed to maximize aggregate reward. Our work offers a novel explanation for emergent abilities in RLVR, suggesting that the iterative optimization of solvable problems enables models to develop the capabilities to tackle previously unsolvable scenarios.

Method: 1) Candidate span generator using neighbor-oriented packing strategy and biaffine mechanism; 2) Legal dictionary integration via multi-head attention; 3) Domain-specific case incorporation (joint crimes, combined punishment) into hypergraph structure; 4) Hypergraph neural network for higher-order inference through message passing.

Result: Experimental results on CAIL2022 information extraction dataset show the method significantly outperforms existing baseline models.

Conclusion: The proposed Legal-KAHRE approach effectively addresses the unique challenges of legal document extraction by incorporating judicial domain knowledge and specialized case structures through hypergraph neural networks.

Abstract: With the continuous progress of digitization in Chinese judicial institutions, a substantial amount of electronic legal document information has been accumulated. To unlock its potential value, entity and relation extraction for legal documents has emerged as a crucial task. However, existing methods often lack domain-specific knowledge and fail to account for the unique characteristics of the judicial domain. In this paper, we propose an entity and relation extraction algorithm based on hypergraph neural network (Legal-KAHRE) for drug-related judgment documents. Firstly, we design a candidate span generator based on neighbor-oriented packing strategy and biaffine mechanism, which identifies spans likely to contain entities. Secondly, we construct a legal dictionary with judicial domain knowledge and integrate it into text encoding representation using multi-head attention. Additionally, we incorporate domain-specific cases like joint crimes and combined punishment for multiple crimes into the hypergraph structure design. Finally, we employ a hypergraph neural network for higher-order inference via message passing. Experimental results on the CAIL2022 information extraction dataset demonstrate that our method significantly outperforms existing baseline models.

Method: Develops unified theoretical framework for analyzing arbitrary contextual information in Transformer-based LLMs, characterizing contextual influence through output error dynamics. Proves error decomposition in single-layer Transformers and extends to multi-context and multi-layer architectures.

Result: Shows context-conditioned error vector decomposes additively into baseline error plus contextual correction vector. Derives geometric conditions for error reduction and explicit upper bound for contextual correction norm based on context-query relevance and complementarity. Experiments validate theory across ICL, retrieval-augmented generation, and memory evolution.

Conclusion: Provides principled theoretical understanding of contextual information in LLMs, leading to improved context selection strategy that boosts performance by 0.6%.

Abstract: Contextual information at inference time, such as demonstrations, retrieved knowledge, or interaction history, can substantially improve large language models (LLMs) without parameter updates, yet its theoretical role remains poorly understood beyond specific settings such as in-context learning (ICL). We present a unified theoretical framework for analyzing the effect of arbitrary contextual information in Transformer-based LLMs. Our analysis characterizes contextual influence through output error dynamics. In a single-layer Transformer, we prove that the context-conditioned error vector decomposes additively into the baseline error vector and a contextual correction vector. This yields necessary geometric conditions for error reduction: the contextual correction must align with the negative baseline error and satisfy a norm constraint. We further show that the contextual correction norm admits an explicit upper bound determined by context-query relevance and complementarity. These results extend to multi-context and multi-layer Transformers. Experiments across ICL, retrieval-augmented generation, and memory evolution validate our theory and motivate a principled context selection strategy that improves performance by $0.6%$.

Method: Proposes JUSTICE framework with three components: Referential Judicial Element Retriever (RJER) retrieves legal articles and precedent cases; Intermediate Conclusion Emulator (ICE) generates verifiable intermediate conclusions; Judicial Unified Synthesizer (JUS) synthesizes inputs to craft final judgments.

Result: JUSTICE significantly outperforms strong baselines on both in-domain legal benchmarks and out-of-distribution datasets, achieving substantial gains in legal accuracy including 4.6% improvement in prison term prediction.

Conclusion: Explicitly modeling the Pre-Judge process is crucial for enhancing legal coherence and accuracy of generated judgment documents, and the JUSTICE framework successfully emulates human judges’ cognitive workflow.

Abstract: Automated judgment document generation is a significant yet challenging legal AI task. As the conclusive written instrument issued by a court, a judgment document embodies complex legal reasoning. However, existing methods often oversimplify this complex process, particularly by omitting the Pre-Judge'' phase, a crucial step where human judges form a preliminary conclusion. This omission leads to two core challenges: 1) the ineffective acquisition of foundational judicial elements, and 2) the inadequate modeling of the Pre-Judge process, which collectively undermine the final document's legal soundness. To address these challenges, we propose \textit{\textbf{J}udicial \textbf{U}nified \textbf{S}ynthesis \textbf{T}hrough \textbf{I}ntermediate \textbf{C}onclusion \textbf{E}mulation} (JUSTICE), a novel framework that emulates the Search $\rightarrow$ Pre-Judge $\rightarrow$ Write’’ cognitive workflow of human judges. Specifically, it introduces the Pre-Judge stage through three dedicated components: Referential Judicial Element Retriever (RJER), Intermediate Conclusion Emulator (ICE), and Judicial Unified Synthesizer (JUS). RJER first retrieves legal articles and a precedent case to establish a referential foundation. ICE then operationalizes the Pre-Judge phase by generating a verifiable intermediate conclusion. Finally, JUS synthesizes these inputs to craft the final judgment. Experiments on both an in-domain legal benchmark and an out-of-distribution dataset show that JUSTICE significantly outperforms strong baselines, with substantial gains in legal accuracy, including a 4.6% improvement in prison term prediction. Our findings underscore the importance of explicitly modeling the Pre-Judge process to enhance the legal coherence and accuracy of generated judgment documents.

Method: 1) Created Dr. SCI dataset with 1M STEM questions across 8 subjects with verifiable/open-ended splits and difficulty annotations. 2) Dr. SCI post-training pipeline with: Exploration-Expanding SFT to broaden reasoning patterns, Dynamic Difficulty Curriculum for adaptive training, and SciRubric-Guided RL using rubric-based evaluation for stable reinforcement learning.

Result: Qwen3-4B-Base trained with Dr. SCI pipeline achieves 63.2 on GPQA-diamond and 32.4 on GPQA-general, outperforming strong baselines like o1-mini and GPT-4o, showing substantial gains in scientific reasoning, especially for open-ended questions.

Conclusion: The Dr. SCI pipeline effectively improves LLMs’ scientific reasoning capabilities through systematic data processing and specialized post-training techniques, demonstrating significant performance gains on challenging scientific benchmarks.

Abstract: Solving open-ended science questions remains challenging for large language models, particularly due to inherently unreliable supervision and evaluation. The bottleneck lies in the data construction and reward design for scientific post-training. We develop a large-scale, systematic data processing pipeline that transforms heterogeneous open-source science data into Dr. SCI dataset, which comprises of 1M questions across eight STEM subjects, with explicit verifiable/open-ended splits, scalable difficulty annotation, and fine-grained rubrics that operationalize evaluation for open-ended answers. Building on this dataset, we propose the Dr. SCI post-training pipeline, which redesigns the standard SFT -> RL workflow through three components: (i) Exploration-Expanding SFT, which broadens the model’s reasoning pattern coverage prior to RL; (ii) Dynamic Difficulty Curriculum, which adapts training data to the model’s evolving scientific capability; and (iii) SciRubric-Guided RL, which enables stable reinforcement learning on open-ended scientific questions via rubric-based evaluation with explicit answer correctness. Qwen3-4B-Base trained using Dr.SCI pipeline achieves 63.2 on GPQA-diamond and 32.4 on GPQA-general, consistently improves over strong post-trained baselines such as o1-mini and GPT-4o, demonstrating substantial gains in scientific reasoning, especially in open-ended settings.

Method: Proposes a generative framework with attention-over-attention decoder to handle variable number of intents and reduce interference between intent detection and slot filling. Also constructs new multi-intent datasets using BERT’s next sentence prediction head.

Result: Achieves state-of-the-art performance on public datasets (MixATIS and MixSNIPS) and newly constructed datasets.

Conclusion: The proposed attention-over-attention generative model effectively addresses multi-intent SLU challenges and outperforms existing methods on multiple datasets.

Abstract: In task-oriented dialogue systems, spoken language understanding (SLU) is a critical component, which consists of two sub-tasks, intent detection and slot filling. Most existing methods focus on the single-intent SLU, where each utterance only has one intent. However, in real-world scenarios users usually express multiple intents in an utterance, which poses a challenge for existing dialogue systems and datasets. In this paper, we propose a generative framework to simultaneously address multiple intent detection and slot filling. In particular, an attention-over-attention decoder is proposed to handle the variable number of intents and the interference between the two sub-tasks by incorporating an inductive bias into the process of multi-task learning. Besides, we construct two new multi-intent SLU datasets based on single-intent utterances by taking advantage of the next sentence prediction (NSP) head of the BERT model. Experimental results demonstrate that our proposed attention-over-attention generative model achieves state-of-the-art performance on two public datasets, MixATIS and MixSNIPS, and our constructed datasets.

Method: Thinking States generates sequences of thinking tokens every few input tokens, transforms these thoughts back into embedding space, and adds them to following input tokens. This captures CoT’s recurrent nature but with parallelizable teacher-forcing learning.

Result: Outperforms other latent reasoning methods on multiple reasoning tasks, narrows gap to CoT on math problems, matches CoT performance on 2-Hop QA with improved latency, and shows stronger reasoning behavior on state-tracking tasks with successful extrapolation to longer sequences.

Conclusion: Thinking States provides an efficient alternative to traditional CoT reasoning by enabling concurrent reasoning during input processing, achieving competitive performance with reduced latency and better generalization to longer sequences.

Abstract: Reasoning with a chain-of-thought (CoT) enables Large Language Models (LLMs) to solve complex tasks but incurs significant inference costs due to the generation of long rationales. We propose Thinking States, a method that performs reasoning {\em while} the input is processing. Specifically, Thinking States generates sequences of thinking tokens every few input tokens, transforms the thoughts back into embedding space, and adds them to the following input tokens. This has two key advantages. First, it captures the recurrent nature of CoT, but where the thought tokens are generated as input is processing. Second, since the thoughts are represented as tokens, they can be learned from natural language supervision, and using teacher-forcing, which is parallelizable. Empirically, Thinking States outperforms other latent reasoning methods on multiple reasoning tasks, narrowing the gap to CoT on math problems, and matching its performance on 2-Hop QA with improved latency. On state-tracking tasks, we show Thinking States leads to stronger reasoning behavior than CoT, successfully extrapolating to longer sequences than seen during training.

Method: Created UReason benchmark with 2,000 instances across five reasoning task families. Introduced evaluation framework comparing direct generation, reasoning-guided generation, and de-contextualized generation (conditioning only on refined prompt). Tested across eight open-source unified models.

Result: Discovered consistent “Reasoning Paradox”: reasoning traces improve performance over direct generation, but retaining intermediate thoughts as conditioning context hinders visual synthesis, while conditioning only on refined prompt yields substantial gains. Bottleneck appears to be contextual interference rather than insufficient reasoning capacity.

Conclusion: UReason provides a principled testbed for studying reasoning in unified models and motivates future methods that effectively integrate reasoning for visual generation while mitigating interference from intermediate reasoning context.

Abstract: To elicit capabilities for addressing complex and implicit visual requirements, recent unified multimodal models increasingly adopt chain-of-thought reasoning to guide image generation. However, the actual effect of reasoning on visual synthesis remains unclear. We present UReason, a diagnostic benchmark for reasoning-driven image generation that evaluates whether reasoning can be faithfully executed in pixels. UReason contains 2,000 instances across five task families: Code, Arithmetic, Spatial, Attribute, and Text reasoning. To isolate the role of reasoning traces, we introduce an evaluation framework comparing direct generation, reasoning-guided generation, and de-contextualized generation which conditions only on the refined prompt. Across eight open-source unified models, we observe a consistent Reasoning Paradox: Reasoning traces generally improve performance over direct generation, yet retaining intermediate thoughts as conditioning context often hinders visual synthesis, and conditioning only on the refined prompt yields substantial gains. Our analysis suggests that the bottleneck lies in contextual interference rather than insufficient reasoning capacity. UReason provides a principled testbed for studying reasoning in unified models and motivates future methods that effectively integrate reasoning for visual generation while mitigating interference.

Method: Created WorldTravel benchmark with 150 real-world travel scenarios across 5 cities requiring navigation of interdependent constraints, and WorldTravel-Webscape multimodal environment with 2,000+ rendered webpages where agents must perceive constraint parameters from visual layouts.

Result: Evaluation of 10 frontier models shows significant performance collapse: GPT-5.2 achieves only 32.67% feasibility in text-only settings, plummeting to 19.33% in multimodal environments. Identified Perception-Action Gap and Planning Horizon threshold at ~10 constraints where reasoning fails.

Conclusion: Perception and reasoning remain independent bottlenecks, underscoring the need for next-generation agents that unify high-fidelity visual perception with long-horizon reasoning for real-world logistics.

Abstract: Real-world autonomous planning requires coordinating tightly coupled constraints where a single decision dictates the feasibility of all subsequent actions. However, existing benchmarks predominantly feature loosely coupled constraints solvable through local greedy decisions and rely on idealized data, failing to capture the complexity of extracting parameters from dynamic web environments. We introduce \textbf{WorldTravel}, a benchmark comprising 150 real-world travel scenarios across 5 cities that demand navigating an average of 15+ interdependent temporal and logical constraints. To evaluate agents in realistic deployments, we develop \textbf{WorldTravel-Webscape}, a multi-modal environment featuring over 2,000 rendered webpages where agents must perceive constraint parameters directly from visual layouts to inform their planning. Our evaluation of 10 frontier models reveals a significant performance collapse: even the state-of-the-art GPT-5.2 achieves only 32.67% feasibility in text-only settings, which plummets to 19.33% in multi-modal environments. We identify a critical Perception-Action Gap and a Planning Horizon threshold at approximately 10 constraints where model reasoning consistently fails, suggesting that perception and reasoning remain independent bottlenecks. These findings underscore the need for next-generation agents that unify high-fidelity visual perception with long-horizon reasoning to handle brittle real-world logistics.

Method: Created ViGoEmotions corpus with 20,664 Vietnamese social media comments annotated with 27 emotions. Evaluated 8 pre-trained Transformer models (including ViSoBERT, CafeBERT, PhoBERT) under three preprocessing strategies: 1) preserving original emojis with rule-based normalization, 2) converting emojis to textual descriptions, 3) applying ViSoLex lexical normalization system.

Result: Converting emojis to text improved performance for most BERT-based models, while preserving emojis worked best for ViSoBERT and CafeBERT. Removing emojis generally reduced performance. ViSoBERT achieved highest scores (Macro F1: 61.50%, Weighted F1: 63.26%). CafeBERT and PhoBERT also performed strongly.

Conclusion: The ViGoEmotions corpus effectively supports diverse architectures for Vietnamese emotion classification, but preprocessing strategies (especially emoji handling) and annotation quality significantly impact downstream performance.

Abstract: Emotion classification plays a significant role in emotion prediction and harmful content detection. Recent advancements in NLP, particularly through large language models (LLMs), have greatly improved outcomes in this field. This study introduces ViGoEmotions – a Vietnamese emotion corpus comprising 20,664 social media comments in which each comment is classified into 27 fine-grained distinct emotions. To evaluate the quality of the dataset and its impact on emotion classification, eight pre-trained Transformer-based models were evaluated under three preprocessing strategies: preserving original emojis with rule-based normalization, converting emojis into textual descriptions, and applying ViSoLex, a model-based lexical normalization system. Results show that converting emojis into text often improves the performance of several BERT-based baselines, while preserving emojis yields the best results for ViSoBERT and CafeBERT. In contrast, removing emojis generally leads to lower performance. ViSoBERT achieved the highest Macro F1-score of 61.50% and Weighted F1-score of 63.26%. Strong performance was also observed from CafeBERT and PhoBERT. These findings highlight that while the proposed corpus can support diverse architectures effectively, preprocessing strategies and annotation quality remain key factors influencing downstream performance.

Method: Proposes a framework that segments long inputs into chunks, encodes each into compressed memory representations using a learned compressor. A gating module dynamically selects relevant memory blocks, which are processed iteratively by a reasoning module with evolving working memory. Compressor and reasoner are jointly optimized via end-to-end RL, while gating is trained separately as a classifier.

Result: Achieves competitive accuracy on multi-hop reasoning benchmarks (RULER-HQA), extrapolates context length from 7K to 1.75M tokens, offers favorable accuracy-efficiency trade-off. Achieves up to 2× reduction in peak GPU memory usage and 6× inference speedup over MemAgent baseline.

Conclusion: The cognitively inspired approach of chunk-wise compression and selective memory recall provides an effective solution for efficient long-context processing in LLMs, addressing computational and memory limitations while maintaining competitive performance.

Abstract: Large Language Models (LLMs) face significant challenges in long-context processing, including quadratic computational costs, information forgetting, and the context fragmentation inherent in retrieval-augmented generation (RAG). We propose a cognitively inspired framework for efficient long-context inference based on chunk-wise compression and selective memory recall, rather than processing all raw tokens. The framework segments long inputs into chunks and encodes each chunk into compressed memory representations using a learned compressor. A gating module dynamically selects relevant memory blocks, which are then iteratively processed by a reasoning module with an evolving working memory to solve downstream tasks. The compressor and reasoner are jointly optimized via end-to-end reinforcement learning, while the gating module is trained separately as a classifier. Experimental results show that the proposed method achieves competitive accuracy on multi-hop reasoning benchmarks such as RULER-HQA, extrapolates context length from 7K to 1.75M tokens, and offers a favorable accuracy-efficiency trade-off compared to strong long-context baselines. In particular, it achieves up to a 2 times reduction in peak GPU memory usage and a 6 times inference speedup over MemAgent.

Method: TEAM uses three strategies: 1) conservative expert selection for decoded tokens based on temporal consistency across denoising levels, 2) conservative selection for masked tokens using spatial consistency across positions, and 3) aggressive speculative exploration across multiple candidates to enable more accepted tokens with fewer experts.

Result: TEAM achieves up to 2.2x speedup over vanilla MoE diffusion LLMs with negligible performance degradation.

Conclusion: The proposed TEAM framework effectively addresses the efficiency mismatch in MoE diffusion LLMs by leveraging routing consistency patterns, enabling practical deployment in latency-sensitive scenarios.

Abstract: Diffusion large language models (dLLMs) have recently gained significant attention due to their inherent support for parallel decoding. Building on this paradigm, Mixture-of-Experts (MoE) dLLMs with autoregressive (AR) initialization have further demonstrated strong performance competitive with mainstream AR models. However, we identify a fundamental mismatch between MoE architectures and diffusion-based decoding. Specifically, a large number of experts are activated at each denoising step, while only a small subset of tokens is ultimately accepted, resulting in substantial inference overhead and limiting their deployment in latency-sensitive applications. In this work, we propose TEAM, a plug-and-play framework that accelerates MoE dLLMs by enabling more accepted tokens with fewer activated experts. TEAM is motivated by the observation that expert routing decisions exhibit strong temporal consistency across denoising levels as well as spatial consistency across token positions. Leveraging these properties, TEAM employs three complementary expert activation and decoding strategies, conservatively selecting necessary experts for decoded and masked tokens and simultaneously performing aggressive speculative exploration across multiple candidates. Experimental results demonstrate that TEAM achieves up to 2.2x speedup over vanilla MoE dLLM, with negligible performance degradation. Code is released at https://github.com/PKU-SEC-Lab/TEAM-MoE-dLLM.

Method: Prism introduces a training-free spectral-aware approach that decomposes block selection into high-frequency and low-frequency branches. It addresses the destructive interference caused by mean pooling with RoPE by applying energy-based temperature calibration to restore attenuated positional signals. This enables block importance estimation using purely block-level operations without token-level searching or scoring.

Result: Extensive evaluations show that Prism maintains accuracy parity with full attention while delivering up to 5.1× speedup. The method effectively overcomes the limitations of standard coarse-grained attention and provides efficient block selection for long-context LLM pre-filling.

Conclusion: Prism offers an effective solution to the block selection bottleneck in block-sparse attention for long-context LLMs. By theoretically analyzing the interaction between mean pooling and RoPE and introducing spectral-aware decomposition with temperature calibration, it achieves both accuracy preservation and significant speed improvements.

Abstract: Block-sparse attention is promising for accelerating long-context LLM pre-filling, yet identifying relevant blocks efficiently remains a bottleneck. Existing methods typically employ coarse-grained attention as a proxy for block importance estimation, but often resort to expensive token-level searching or scoring, resulting in significant selection overhead. In this work, we trace the inaccuracy of standard coarse-grained attention via mean pooling to a theoretical root cause: the interaction between mean pooling and Rotary Positional Embeddings (RoPE). We prove that mean pooling acts as a low-pass filter that induces destructive interference in high-frequency dimensions, effectively creating a “blind spot” for local positional information (e.g., slash patterns). To address this, we introduce Prism, a training-free spectral-aware approach that decomposes block selection into high-frequency and low-frequency branches. By applying energy-based temperature calibration, Prism restores the attenuated positional signals directly from pooled representations, enabling block importance estimation using purely block-level operations, thereby improving efficiency. Extensive evaluations confirm that Prism maintains accuracy parity with full attention while delivering up to $\mathbf{5.1\times}$ speedup.

Method: Constructed syntactically impossible languages by applying transformations to English (reversing sentences, adding negation based on word-count parity). Conducted controlled experiments on GPT-2 small models and LSTM models with statistical analysis using Welch’s t-test.

Result: GPT-2 small models underperformed in learning all impossible languages compared to possible languages (p<.001). LSTM models’ performance aligned with Chomsky’s arguments, highlighting differences in transformer architecture evolution.

Conclusion: Proposes new vision within Chomsky’s theory for LLMs and suggests paradigm shift from rationalist-romantic approach to functionalism and empiricism in LLM research, acknowledging architectural differences.

Abstract: In Chomsky’s provocative critique “The False Promise of CHATGPT,” Large Language Models (LLMs) are characterized as mere pattern predictors that do not acquire languages via intrinsic causal and self-correction structures like humans, therefore are not able to distinguish impossible languages. It stands as a representative in a fundamental challenge to the intellectual foundations of AI, for it integrally synthesizes major issues in methodologies within LLMs and possesses an iconic a priori rationalist perspective. We examine this famous critic from both the perspective in pre-existing literature of linguistics and psychology as well as a research based on an experiment inquiring the capacity of learning both possible and impossible languages among LLMs. We constructed a set of syntactically impossible languages by applying certain transformations to English. These include reversing whole sentences, and adding negation based on word-count parity. Two rounds of controlled experiments were each conducted on GPT-2 small models and long short-term memory (LSTM) models. Statistical analysis (Welch’s t-test) shows GPT2 small models underperform in learning all of the impossible languages compared to their performance on the possible language (p<.001). On the other hand, LSTM models’ performance tallies with Chomsky’s argument, suggesting the irreplaceable role of the evolution of transformer architecture. Based on theoretical analysis and empirical findings, we propose a new vision within Chomsky’s theory towards LLMs, and a shift of theoretical paradigm outside Chomsky, from his “rationalist-romantics” paradigm to functionalism and empiricism in LLMs research.

Method: Proposes ME² principle for reasoning quality (macro/micro efficiency/effectiveness), models reasoning traces as DAGs, develops DAG-based pairwise evaluation method, constructs TRM-Preference dataset, and trains Thinking Reward Model (TRM).

Result: Thinking rewards serve as effective optimization signal: test-time selection yields up to 19.3% performance gain, RL training with thinking rewards improves reasoning and performance up to 3.9% across diverse tasks.

Conclusion: Provides unified framework for reasoning evaluation and optimization, demonstrating practical benefits through thinking rewards for enhancing Large Reasoning Models.

Abstract: Large Reasoning Models (LRMs) increasingly rely on reasoning traces with complex internal structures. However, existing work lacks a unified answer to three fundamental questions: (1) what defines high-quality reasoning, (2) how to reliably evaluate long, implicitly structured reasoning traces, and (3) how to use such evaluation signals for reasoning optimization. To address these challenges, we provide a unified perspective. (1) We introduce the ME$^2$ principle to characterize reasoning quality along macro- and micro-level concerning efficiency and effectiveness. (2) Built on this principle, we model reasoning traces as directed acyclic graphs (DAGs) and develop a DAG-based pairwise evaluation method, capturing complex reasoning structures. (3) Based on this method, we construct the TRM-Preference dataset and train a Thinking Reward Model (TRM) to evaluate reasoning quality at scale. Experiments show that thinking rewards serve as an effective optimization signal. At test time, selecting better reasoning leads to better outcomes (up to 19.3% gain), and during RL training, thinking rewards enhance reasoning and performance (up to 3.9% gain) across diverse tasks.

Method: Introduced GISA benchmark with 373 human-crafted queries reflecting authentic information-seeking scenarios. Features four structured answer formats (item, set, list, table) for deterministic evaluation, integrates deep reasoning and broad information aggregation, includes live subset with periodically updated answers, and provides complete human search trajectories for process-level supervision.

Result: Experiments show even the best-performing model achieves only 19.30% exact match score, with performance degrading on tasks requiring complex planning and comprehensive information gathering, highlighting substantial room for improvement.

Conclusion: GISA addresses key limitations in existing search agent benchmarks and reveals significant gaps in current LLM capabilities for realistic information-seeking tasks, providing a valuable resource for future research and development.

Abstract: The advancement of large language models (LLMs) has significantly accelerated the development of search agents capable of autonomously gathering information through multi-turn web interactions. Various benchmarks have been proposed to evaluate such agents. However, existing benchmarks often construct queries backward from answers, producing unnatural tasks misaligned with real-world needs. Moreover, these benchmarks tend to focus on either locating specific information or aggregating information from multiple sources, while relying on static answer sets prone to data contamination. To bridge these gaps, we introduce GISA, a benchmark for General Information-Seeking Assistants comprising 373 human-crafted queries that reflect authentic information-seeking scenarios. GISA features four structured answer formats (item, set, list, and table), enabling deterministic evaluation. It integrates both deep reasoning and broad information aggregation within unified tasks, and includes a live subset with periodically updated answers to resist memorization. Notably, GISA provides complete human search trajectories for every query, offering gold-standard references for process-level supervision and imitation learning. Experiments on mainstream LLMs and commercial search products reveal that even the best-performing model achieves only 19.30% exact match score, with performance notably degrading on tasks requiring complex planning and comprehensive information gathering. These findings highlight substantial room for future improvement.

Method: Used activation patching and complementary interpretability techniques to dissect table understanding into three stages: Semantic Binding, Coordinate Localization, and Information Extraction.

Result: Models locate target cells via ordinal counting of delimiters, encode column indices in linear subspaces for precise steering, and generalize multi-cell tasks by multiplexing identical attention heads.

Conclusion: Provides a comprehensive explanation of table understanding within Transformer architectures, revealing systematic mechanisms for processing structured data.

Abstract: While Large Language Models (LLMs) are increasingly deployed for table-related tasks, the internal mechanisms enabling them to process linearized two-dimensional structured tables remain opaque. In this work, we investigate the process of table understanding by dissecting the atomic task of cell location. Through activation patching and complementary interpretability techniques, we delineate the table understanding mechanism into a sequential three-stage pipeline: Semantic Binding, Coordinate Localization, and Information Extraction. We demonstrate that models locate the target cell via an ordinal mechanism that counts discrete delimiters to resolve coordinates. Furthermore, column indices are encoded within a linear subspace that allows for precise steering of model focus through vector arithmetic. Finally, we reveal that models generalize to multi-cell location tasks by multiplexing the identical attention heads identified during atomic location. Our findings provide a comprehensive explanation of table understanding within Transformer architectures.

Method: Introduced first English-Malayalam QE dataset with Direct Assessment scores and Translation Quality Remarks (error descriptions). Developed ALOPE-RL, a policy-based RL framework that trains efficient adapters using rewards from both DA scores and TQR annotations.

Result: ALOPE-RL achieves state-of-the-art performance on English-Malayalam QE using compact LLMs (≤4B parameters) with LoRA and 4-bit quantization, outperforming larger LLM baselines and encoder-based QE models.

Conclusion: Error-aware, policy-based learning enables strong QE performance under limited data and compute budgets, demonstrating the value of incorporating error descriptions beyond numeric scores.

Abstract: Quality Estimation (QE) aims to assess the quality of machine translation (MT) outputs without relying on reference translations, making it essential for real-world, large-scale MT evaluation. Large Language Models (LLMs) have shown significant promise in advancing the field of quality estimation of machine translation. However, most of the QE approaches solely rely on scalar quality scores, offering no explicit information about the translation errors that should drive these judgments. Moreover, for low-resource languages where annotated QE data is limited, existing approaches struggle to achieve reliable performance. To address these challenges, we introduce the first segment-level QE dataset for English to Malayalam, a severely resource-scarce language pair in the QE domain, comprising human-annotated Direct Assessment (DA) scores and Translation Quality Remarks (TQR), which are short, contextual, free-form annotator comments that describe translation errors. We further introduce ALOPE-RL, a policy-based reinforcement learning framework that trains efficient adapters based on policy rewards derived from DA score and TQR. Integrating error-aware rewards with ALOPE-RL, enables LLMs to reason about translation quality beyond numeric scores. Despite being trained on a small-scale QE dataset, ALOPE-RL achieves state-of-the-art performance on English to Malayalam QE using compact LLMs (<=4B parameters}) fine-tuned with LoRA and 4-bit quantization, outperforming both larger LLM-based baselines and leading encoder-based QE models. Our results demonstrate that error-aware, policy-based learning can deliver strong QE performance under limited data and compute budgets. We release our dataset, code, and trained models to support future research.

Method: The paper explores whether multilingual LLMs have specialized language attention heads for each language and investigates removing language-specific heads for unwanted languages.

Result: Findings could inform more efficient deployment strategies for multilingual LLMs, enabling reduced model complexity while maintaining high accuracy for targeted languages.

Conclusion: The research addresses an important efficiency gap in multilingual LLM deployment by investigating language-specific head pruning for targeted language subsets.

Abstract: Multilingual large language models (LLMs) have gained significant popularity for their ability to process and generate text across multiple languages. However, deploying these models in production can be inefficient when only a subset of the supported languages is of interest. There has been some research conducted on identifying whether machine translation models have language-specific or language-agnostic heads, however no research has been conducted for multilingual LLMs, to the best of our knowledge, that as we know are capable of performing diverse tasks beyond just translation. This paper explores whether multilingual LLMs have specialized language attention heads for each language, and investigates the possibility of removing language-specific heads for unwanted languages without degrading performance in the targeted languages. Our findings could inform more efficient deployment strategies for multilingual LLMs, enabling reduced model complexity while maintaining high accuracy for targeted languages.

Method: 1) Collected new dataset of reasoning trajectories from symbolic tasks targeting each fundamental paradigm; 2) Experimented with various training methods including simple fine-tuning, increasing model depth, and transforming dense models to mixture-of-experts; 3) Evaluated on realistic out-of-domain natural language tasks containing real-world knowledge.

Result: The approach yields strong generalizability with substantial performance gains (up to 14.60 points) across realistic tasks, demonstrating that training on fundamental reasoning paradigms improves LLM reasoning on diverse out-of-domain tasks.

Conclusion: Fundamental reasoning paradigms significantly influence LLM reasoning behavior and generalization, with systematic training on these paradigms leading to substantial performance improvements on realistic natural language reasoning tasks.

Abstract: Deduction, induction, and abduction are fundamental reasoning paradigms, core for human logical thinking. Although improving Large Language Model (LLM) reasoning has attracted significant research efforts, the extent to which the fundamental paradigms induce generalization has yet to be systematically explored. In this study, we shed light on how the interplay between these core paradigms influences LLMs’ reasoning behavior. To this end, we first collect a new dataset of reasoning trajectories from symbolic tasks, each targeting one of the three fundamental paradigms, to abstract from concrete world knowledge. Then, we investigate effective ways for inducing these skills into LLMs. We experiment with a battery of methods including simple fine-tuning, and more complex approaches to increase model depth, or transform a dense model to a mixture-of-experts. We comprehensively evaluate induced models on realistic out-of-domain tasks, that are entirely formulated in natural language and contain real-world knowledge. Our results reveal that our approach yields strong generalizability with substantial performance gains (up to $14.60$) across realistic tasks.

Method: Introduces GER-Eval (Generating Evaluation Rubrics for Evaluation) to investigate LLM-designed rubrics. Evaluates semantic coherence, scoring reliability, and alignment with human criteria across different LLMs including GPT-4o and Llama models.

Result: LLMs reliably generate interpretable, task-aware evaluation dimensions and apply them consistently within models. However, scoring reliability degrades in factual and knowledge-intensive settings. Closed-source models (GPT-4o) achieve higher agreement and cross-model generalization than open-weight models (Llama).

Conclusion: Evaluation is a learned linguistic capability of LLMs, consistent within models but fragmented across them. Calls for new methods that jointly model human and LLM evaluative language to improve reliability and interpretability.

Abstract: Large language models (LLMs) are increasingly used as evaluators for natural language generation, applying human-defined rubrics to assess system outputs. However, human rubrics are often static and misaligned with how models internally represent language quality. We introduce GER-Eval (Generating Evaluation Rubrics for Evaluation) to investigate whether LLMs can design and apply their own evaluation rubrics. We evaluate the semantic coherence and scoring reliability of LLM-defined criteria and their alignment with human criteria. LLMs reliably generate interpretable and task-aware evaluation dimensions and apply them consistently within models, but their scoring reliability degrades in factual and knowledge-intensive settings. Closed-source models such as GPT-4o achieve higher agreement and cross-model generalization than open-weight models such as Llama. Our findings position evaluation as a learned linguistic capability of LLMs, consistent within models but fragmented across them, and call for new methods that jointly model human and LLM evaluative language to improve reliability and interpretability.

Method: Three approaches compared: 1) Human qualitative coding, 2) Supervised learning with ParsBERT (Persian BERT), 3) Large language models (ChatGPT with 7 different models). Evaluation on 47,278 Persian tweets from the #MahsaAmini movement in Iran

Result: ParsBERT substantially outperformed all seven ChatGPT models in identifying hate speech. ChatGPT struggled with both subtle cases and explicitly uncivil content. Prompt language (English vs. Persian) did not meaningfully affect ChatGPT outputs

Conclusion: The study provides detailed comparison of approaches for hate speech analysis in low-resource languages, showing ParsBERT’s superiority over ChatGPT for this specific task, and clarifies strengths/limitations of each method

Abstract: This paper compares three approaches to detecting incivility in Persian tweets: human qualitative coding, supervised learning with ParsBERT, and large language models (ChatGPT). Using 47,278 tweets from the #MahsaAmini movement in Iran, we evaluate the accuracy and efficiency of each method. ParsBERT substantially outperforms seven evaluated ChatGPT models in identifying hate speech. We also find that ChatGPT struggles not only with subtle cases but also with explicitly uncivil content, and that prompt language (English vs. Persian) does not meaningfully affect its outputs. The study provides a detailed comparison of these approaches and clarifies their strengths and limitations for analyzing hate speech in a low-resource language context.

Method: Uses NPTEL MOOC portal as corpus, curates spontaneous speech corpora for technical concepts, analyzes machine translation performance across Bangla, Malayalam, and Telugu with focus on evaluation metrics.

Result: Reveals metric-specific sensitivity and challenges with morphologically rich and semantically compact language features when tested against surface overlapping metrics.

Conclusion: Highlights the need for better evaluation frameworks for machine translation in Indian languages, particularly for educational content delivery in diverse linguistic settings.

Abstract: This study examines the practical applications and methodological implications of Machine Translation in Indian Languages, specifically Bangla, Malayalam, and Telugu, within emerging translation workflows and in relation to existing evaluation frameworks. The choice of languages prioritized in this study is motivated by a triangulation of linguistic diversity, which illustrates the significance of multilingual accommodation of educational technology under NEP 2020. This is further supported by the largest MOOC portal, i.e., NPTEL, which has served as a corpus to facilitate the arguments presented in this paper. The curation of a spontaneous speech corpora that accounts for lucid delivery of technical concepts, considering the retention of suitable register and lexical choices are crucial in a diverse country like India. The findings of this study highlight metric-specific sensitivity and the challenges of morphologically rich and semantically compact features when tested against surface overlapping metrics.

Method: Conducted a user study with 73 participants comparing multimodal (text+images) vs. text-only clarifying questions in conversational search, examining effects on two tasks: answering clarifying questions and query reformulation.

Result: Participants preferred multimodal questions for answering clarifying questions but preferences were balanced for query reformulation. Images helped maintain engagement across expertise levels for question answering, and led to more precise queries and better retrieval performance for reformulation. Text-only setups performed better for question answering due to more comprehensive textual information.

Conclusion: Benefits of visual augmentation in conversational search are task-dependent and should be strategically implemented based on specific search context and user characteristics, providing valuable insights for designing effective multimodal conversational search systems.

Abstract: Conversational search systems increasingly employ clarifying questions to refine user queries and improve the search experience. Previous studies have demonstrated the usefulness of text-based clarifying questions in enhancing both retrieval performance and user experience. While images have been shown to improve retrieval performance in various contexts, their impact on user performance when incorporated into clarifying questions remains largely unexplored. We conduct a user study with 73 participants to investigate the role of images in conversational search, specifically examining their effects on two search-related tasks: (i) answering clarifying questions and (ii) query reformulation. We compare the effect of multimodal and text-only clarifying questions in both tasks within a conversational search context from various perspectives. Our findings reveal that while participants showed a strong preference for multimodal questions when answering clarifying questions, preferences were more balanced in the query reformulation task. The impact of images varied with both task type and user expertise. In answering clarifying questions, images helped maintain engagement across different expertise levels, while in query reformulation they led to more precise queries and improved retrieval performance. Interestingly, for clarifying question answering, text-only setups demonstrated better user performance as they provided more comprehensive textual information in the absence of images. These results provide valuable insights for designing effective multimodal conversational search systems, highlighting that the benefits of visual augmentation are task-dependent and should be strategically implemented based on the specific search context and user characteristics.

Method: Develops a fully automated methodology that extracts factual assessments from texts, measures similarity between claims in generated summaries and original reviews, and computes coverage and consistency scores based on claim comparison.

Result: The proposed metric attributes higher scores to similar claims regardless of negation, paraphrasing, or expansion, and shows high correlation with human judgment compared to state-of-the-art metrics.

Conclusion: The novel automated evaluation method effectively addresses limitations of traditional metrics for assessing factual consistency in opinion summarization tasks with LLMs.

Abstract: We explore the need for more comprehensive and precise evaluation techniques for generative artificial intelligence (GenAI) in text summarization tasks, specifically in the area of opinion summarization. Traditional methods, which leverage automated metrics to compare machine-generated summaries from a collection of opinion pieces, e.g. product reviews, have shown limitations due to the paradigm shift introduced by large language models (LLM). This paper addresses these shortcomings by proposing a novel, fully automated methodology for assessing the factual consistency of such summaries. The method is based on measuring the similarity between the claims in a given summary with those from the original reviews, measuring the coverage and consistency of the generated summary. To do so, we rely on a simple approach to extract factual assessment from texts that we then compare and summarize in a suitable score. We demonstrate that the proposed metric attributes higher scores to similar claims, regardless of whether the claim is negated, paraphrased, or expanded, and that the score has a high correlation to human judgment when compared to state-of-the-art metrics.

Method: Created PERSPECTRA benchmark using controlled retrieval-and-expansion pipeline to integrate Kialo debate graphs with Reddit discussions. Constructed 3,810 enriched arguments spanning 762 pro/con stances on 100 controversial topics, with each opinion expanded to multiple naturalistic variants.

Result: State-of-the-art LLMs show systematic failures: overestimating number of viewpoints, misclassifying concessive structures, and struggling with pluralism-aware understanding and reasoning across three tasks (opinion counting, opinion matching, and polarity check).

Conclusion: PERSPECTRA establishes the first scalable, configurable benchmark for evaluating how well models represent, distinguish, and reason over multiple perspectives, addressing a critical gap in LLM alignment research.

Abstract: Pluralism, the capacity to engage with diverse perspectives without collapsing them into a single viewpoint, is critical for developing large language models that faithfully reflect human heterogeneity. Yet this characteristic has not been carefully examined in the LLM research community and remains absent from most alignment studies. Debate-oriented sources provide a natural entry point for pluralism research. Previous work builds on online debate sources but remains constrained by costly human validation. Other debate-rich platforms such as Reddit and Kialo also offer promising material: Reddit provides linguistic diversity and scale but lacks clear argumentative structure, while Kialo supplies explicit pro/con graphs but remains overly concise and detached from natural discourse. We introduce PERSPECTRA, a pluralist benchmark that integrates the structural clarity of Kialo debate graphs with the linguistic diversity of real Reddit discussions. Using a controlled retrieval-and-expansion pipeline, we construct 3,810 enriched arguments spanning 762 pro/con stances on 100 controversial topics. Each opinion is expanded to multiple naturalistic variants, enabling robust evaluation of pluralism. We initialise three tasks with PERSPECTRA: opinion counting (identifying distinct viewpoints), opinion matching (aligning supporting stances and discourse to source opinions), and polarity check (inferring aggregate stance in mixed discourse). Experiments with state-of-the-art open-source and proprietary LLMs, highlight systematic failures, such as overestimating the number of viewpoints and misclassifying concessive structures, underscoring the difficulty of pluralism-aware understanding and reasoning. By combining diversity with structure, PERSPECTRA establishes the first scalable, configurable benchmark for evaluating how well models represent, distinguish, and reason over multiple perspectives.

Method: 1) Represent each sentence encoder using an embedding matrix from a sentence set; 2) Compute Pairwise Inner Product (PIP) matrices for each encoder; 3) Create feature vectors using Quantum Relative Entropy (QRE) with respect to a unit base encoder; 4) Construct a 2D map of 1101 publicly available encoders.

Result: Successfully created a comprehensive map of 1101 sentence encoders where encoders with similar attributes are proximally located. The encoder feature vectors accurately predict downstream task performance in retrieval and clustering tasks, demonstrating the faithfulness of the map.

Conclusion: The proposed method provides a novel perspective on the sentence encoder landscape, enabling effective comparison, visualization, and performance prediction of diverse sentence encoders at scale.

Abstract: We propose a method to compare and visualise sentence encoders at scale by creating a map of encoders where each sentence encoder is represented in relation to the other sentence encoders. Specifically, we first represent a sentence encoder using an embedding matrix of a sentence set, where each row corresponds to the embedding of a sentence. Next, we compute the Pairwise Inner Product (PIP) matrix for a sentence encoder using its embedding matrix. Finally, we create a feature vector for each sentence encoder reflecting its Quantum Relative Entropy (QRE) with respect to a unit base encoder. We construct a map of encoders covering 1101 publicly available sentence encoders, providing a new perspective of the landscape of the pre-trained sentence encoders. Our map accurately reflects various relationships between encoders, where encoders with similar attributes are proximally located on the map. Moreover, our encoder feature vectors can be used to accurately infer downstream task performance of the encoders, such as in retrieval and clustering tasks, demonstrating the faithfulness of our map.

Method: LakeHopper identifies and resolves knowledge gaps through LM interactions, uses cluster-based data selection for unannotated columns, and employs incremental fine-tuning to gradually adapt source models to target data lakes while preserving shared knowledge.

Result: Experimental results validate LakeHopper’s effectiveness on two different data lake transfers under both low-resource and high-resource settings, demonstrating successful adaptation with minimal target annotations.

Conclusion: LakeHopper provides an effective framework for adapting column type annotation models across data lakes with reduced annotation requirements, addressing source-target knowledge gaps and preserving shared knowledge.

Abstract: Column type annotation is vital for tasks like data cleaning, integration, and visualization. Recent solutions rely on resource-intensive language models fine-tuned on well-annotated columns from a particular set of tables, i.e., a source data lake. In this paper, we study whether we can adapt an existing pre-trained LM-based model to a new (i.e., target) data lake to minimize the annotations required on the new data lake. However, challenges include the source-target knowledge gap, selecting informative target data, and fine-tuning without losing shared knowledge exist. We propose LakeHopper, a framework that identifies and resolves the knowledge gap through LM interactions, employs a cluster-based data selection scheme for unannotated columns, and uses an incremental fine-tuning mechanism that gradually adapts the source model to the target data lake. Our experimental results validate the effectiveness of LakeHopper on two different data lake transfers under both low-resource and high-resource settings.

Method: Proposes AFlow framework that models continuous affective flow along multi-turn trajectories, estimates intermediate utility over searched trajectories, learns preference-consistent strategy transitions, and uses subpath-level flow-balance objective to propagate preference signals to intermediate states.

Result: Consistent and significant improvements over competitive baselines in diverse emotional contexts. AFlow with compact open-source backbone outperforms proprietary LLMs like GPT-4o and Claude-3.5 on major ESC metrics.

Conclusion: Affective flow modeling provides effective fine-grained supervision for emotional support conversations, enabling better strategy coherence and empathetic response quality in multi-turn dialogues.

Abstract: Large language models (LLMs) have been widely applied to emotional support conversation (ESC). However, complex multi-turn support remains challenging.This is because existing alignment schemes rely on sparse outcome-level signals, thus offering limited supervision for intermediate strategy decisions. To fill this gap, this paper proposes affective flow language model for emotional support conversation (AFlow), a framework that introduces fine-grained supervision on dialogue prefixes by modeling a continuous affective flow along multi-turn trajectories. AFlow can estimate intermediate utility over searched trajectories and learn preference-consistent strategy transitions. To improve strategy coherence and empathetic response quality, a subpath-level flow-balance objective is presented to propagate preference signals to intermediate states. Experiment results show consistent and significant improvements over competitive baselines in diverse emotional contexts. Remarkably, AFlow with a compact open-source backbone outperforms proprietary LMMs such as GPT-4o and Claude-3.5 on major ESC metrics. Our code is available at https://github.com/chzou25-lgtm/AffectiveFlow.

Method: Uses WildChat as interaction source, proposes pipeline to extract reliable human feedback, yielding 186k high-quality instances. Trains WildReward via ordinal regression directly on user feedback without preference pairs.

Result: WildReward achieves comparable or superior performance to conventional reward models, with improved calibration and cross-sample consistency. Benefits from user diversity (more users yield stronger models). Applied to online DPO training with significant improvements across various tasks.

Conclusion: In-the-wild interactions can effectively replace human-annotated preference pairs for training reward models, offering a scalable and efficient alternative with strong performance.

Abstract: Reward models (RMs) are crucial for the training of large language models (LLMs), yet they typically rely on large-scale human-annotated preference pairs. With the widespread deployment of LLMs, in-the-wild interactions have emerged as a rich source of implicit reward signals. This raises the question: Can we develop reward models directly from in-the-wild interactions? In this work, we explore this possibility by adopting WildChat as an interaction source and proposing a pipeline to extract reliable human feedback, yielding 186k high-quality instances for training WildReward via ordinal regression directly on user feedback without preference pairs. Extensive experiments demonstrate that WildReward achieves comparable or even superior performance compared to conventional reward models, with improved calibration and cross-sample consistency. We also observe that WildReward benefits directly from user diversity, where more users yield stronger reward models. Finally, we apply WildReward to online DPO training and observe significant improvements across various tasks. Code and data are released at https://github.com/THU-KEG/WildReward.

Method: Introduces ANIRA, a unified recurrent Transformer framework supporting per-token variable-depth computation while isolating compute allocation decisions from other model factors. Uses algorithmic and synthetic language tasks with parameterized difficulty for complexity-controlled evaluation.

Result: Compute allocation aligned with task complexity can emerge without explicit difficulty supervision, but such alignment doesn’t guarantee algorithmic generalization (models fail to extrapolate to unseen input sizes). Early compute decisions rely on static structural cues, while online halting more closely tracks algorithmic execution state.

Conclusion: The paper provides a systematic framework for evaluating token-level adaptive computation, revealing important insights about compute allocation alignment, generalization limitations, and decision timing mechanisms in adaptive models.

Abstract: Token-level adaptive computation seeks to reduce inference cost by allocating more computation to harder tokens and less to easier ones. However, prior work is primarily evaluated on natural-language benchmarks using task-level metrics, where token-level difficulty is unobservable and confounded with architectural factors, making it unclear whether compute allocation truly aligns with underlying complexity. We address this gap through three contributions. First, we introduce a complexity-controlled evaluation paradigm using algorithmic and synthetic language tasks with parameterized difficulty, enabling direct testing of token-level compute allocation. Second, we propose ANIRA, a unified recurrent Transformer framework that supports per-token variable-depth computation while isolating compute allocation decisions from other model factors. Third, we use this framework to conduct a systematic analysis of token-level adaptive computation across alignment with complexity, generalization, and decision timing. Our results show that compute allocation aligned with task complexity can emerge without explicit difficulty supervision, but such alignment does not imply algorithmic generalization: models fail to extrapolate to unseen input sizes despite allocating additional computation. We further find that early compute decisions rely on static structural cues, whereas online halting more closely tracks algorithmic execution state.

Method: Two-step framework combining few-shot LLM-based named entity recognition with agent-based geocoding module that uses context to resolve ambiguous toponyms.

Result: LLM-based methods substantially improve both precision and fairness of geolocation extraction, particularly for underrepresented regions, outperforming state-of-the-art pretrained and rule-based systems.

Conclusion: The work bridges LLM reasoning with responsible AI principles to create more equitable geospatial data systems for humanitarian response, advancing the goal of inclusive crisis analytics.

Abstract: Humanitarian crises demand timely and accurate geographic information to inform effective response efforts. Yet, automated systems that extract locations from text often reproduce existing geographic and socioeconomic biases, leading to uneven visibility of crisis-affected regions. This paper investigates whether Large Language Models (LLMs) can address these geographic disparities in extracting location information from humanitarian documents. We introduce a two-step framework that combines few-shot LLM-based named entity recognition with an agent-based geocoding module that leverages context to resolve ambiguous toponyms. We benchmark our approach against state-of-the-art pretrained and rule-based systems using both accuracy and fairness metrics across geographic and socioeconomic dimensions. Our evaluation uses an extended version of the HumSet dataset with refined literal toponym annotations. Results show that LLM-based methods substantially improve both the precision and fairness of geolocation extraction from humanitarian texts, particularly for underrepresented regions. By bridging advances in LLM reasoning with principles of responsible and inclusive AI, this work contributes to more equitable geospatial data systems for humanitarian response, advancing the goal of leaving no place behind in crisis analytics.

Method: Introduces compositional reasoning attacks where harmful queries are decomposed into incomplete fragments scattered throughout long contexts (up to 64k tokens), then prompts models with neutral reasoning queries that induce retrieval and synthesis. Evaluates 14 frontier LLMs.

Result: Three key findings: (1) Models with stronger general reasoning capability are not more robust to attacks, often assembling harmful intent yet failing to refuse; (2) Safety alignment degrades as context length increases; (3) Increasing inference-time compute reduces attack success by over 50 percentage points on GPT-oss-120b.

Conclusion: Safety does not automatically scale with reasoning capability, especially under long-context inference, revealing a critical vulnerability in current LLM safety approaches.

Abstract: Large language models (LLMs) increasingly combine long-context processing with advanced reasoning, enabling them to retrieve and synthesize information distributed across tens of thousands of tokens. A hypothesis is that stronger reasoning capability should improve safety by helping models recognize harmful intent even when it is not stated explicitly. We test this hypothesis in long-context settings where harmful intent is implicit and must be inferred through reasoning, and find that it does not hold. We introduce compositional reasoning attacks, a new threat model in which a harmful query is decomposed into incomplete fragments that scattered throughout a long context. The model is then prompted with a neutral reasoning query that induces retrieval and synthesis, causing the harmful intent to emerge only after composition. Evaluating 14 frontier LLMs on contexts up to 64k tokens, we uncover three findings: (1) models with stronger general reasoning capability are not more robust to compositional reasoning attacks, often assembling the intent yet failing to refuse; (2) safety alignment consistently degrades as context length increases; and (3) inference-time reasoning effort is a key mitigating factor: increasing inference-time compute reduces attack success by over 50 percentage points on GPT-oss-120b model. Together, these results suggest that safety does not automatically scale with reasoning capability, especially under long-context inference.

Method: Three-stage pipeline: 1) identifying information deficits (quality gaps), 2) executing real-time targeted web retrieval to resolve gaps, 3) synthesizing platform-compliant notes. Uses PolBench benchmark of 78,698 political tweets with Community Notes for evaluation.

Result: Achieves 99% coverage (almost doubling state-of-the-art), surpasses human-authored helpful notes with 69% win rate and superior helpfulness scores (3.87 vs 3.36), demonstrating effective retrieval that balances scale and quality.

Conclusion: GitSearch effectively addresses cold start problems in community moderation by treating quality gaps as first-class signals, enabling scalable, high-quality fact-checking through targeted information retrieval.

Abstract: Community-based moderation offers a scalable alternative to centralized fact-checking, yet it faces significant structural challenges, and existing AI-based methods fail in “cold start” scenarios. To tackle these challenges, we introduce GitSearch (Gap-Informed Targeted Search), a framework that treats human-perceived quality gaps, such as missing context, etc., as first-class signals. GitSearch has a three-stage pipeline: identifying information deficits, executing real-time targeted web-retrieval to resolve them, and synthesizing platform-compliant notes. To facilitate evaluation, we present PolBench, a benchmark of 78,698 U.S. political tweets with their associated Community Notes. We find GitSearch achieves 99% coverage, almost doubling coverage over the state-of-the-art. GitSearch surpasses human-authored helpful notes with a 69% win rate and superior helpfulness scores (3.87 vs. 3.36), demonstrating retrieval effectiveness that balanced the trade-off between scale and quality.

Method: Position paper analysis arguing for conceptual reframing: from decontextualized text classification to a routing problem that incorporates environmental cues and prior probability estimation.

Result: Identifies systemic issues in LID research and practice, proposing a paradigm shift toward context-aware, locally-plausible language identification.

Conclusion: Improving coverage for tail languages requires rethinking LID fundamentals - moving from global fixed-prior models to context-sensitive routing that leverages environmental information.

Abstract: Of the over 7,000 languages spoken in the world, commercial language identification (LID) systems only reliably identify a few hundred in written form. Research-grade systems extend this coverage under certain circumstances, but for most languages coverage remains patchy or nonexistent. This position paper argues that this situation is largely self-imposed. In particular, it arises from a persistent framing of LID as decontextualized text classification, which obscures the central role of prior probability estimation and is reinforced by institutional incentives that favor global, fixed-prior models. We argue that improving coverage for tail languages requires rethinking LID as a routing problem and developing principled ways to incorporate environmental cues that make languages locally plausible.

Method: ConceptLM uses Vector Quantization to quantize hidden states into discrete concepts, creating a concept vocabulary. The model is trained with both NCP and NTP objectives, where NCP predicts concepts that then guide subsequent token generation. Models are trained from scratch at various scales (70M to 1.5B parameters) using up to 300B training data.

Result: Experiments on 13 benchmarks show consistent performance gains over traditional token-level models. Continual pretraining on an 8B-parameter Llama model demonstrates that NCP can further improve models already trained with NTP.

Conclusion: NCP provides a promising path toward better language modeling by introducing a harder pretraining task that leads to more powerful language models through concept-level prediction.

Abstract: We propose Next Concept Prediction (NCP), a generative pretraining paradigm built on top of Next Token Prediction (NTP). NCP predicts discrete concepts that span multiple tokens, thereby forming a more challenging pretraining objective. Our model, ConceptLM, quantizes hidden states using Vector Quantization and constructs a concept vocabulary. It leverages both NCP and NTP to drive parameter updates and generates a concept to guide the generation of the following tokens. We train ConceptLM from scratch at scales ranging from 70M to 1.5B parameters with up to 300B training data, including Pythia and GPT-2 backbones. Results on 13 benchmarks show that NCP yields consistent performance gains over traditional token-level models. Furthermore, continual pretraining experiments on an 8B-parameter Llama model indicate that NCP can further improve an NTP-trained model. Our analysis suggests that NCP leads to more powerful language models by introducing a harder pretraining task, providing a promising path toward better language modeling.

Method: Proposes DeAction, a practical guardrail that detects misaligned actions before execution and iteratively corrects them through structured feedback. Also constructs MisActBench benchmark with human-annotated, action-level alignment labels covering three common real-world CUA deployment categories.

Result: DeAction outperforms all baselines: (1) On MisActBench, achieves over 15% absolute improvement in F1 score; (2) In online evaluation, reduces attack success rate by over 90% under adversarial settings while preserving or improving task success rate in benign environments, with moderate latency overhead.

Conclusion: This work provides the first systematic study of misaligned action detection in computer-use agents, offering both a comprehensive benchmark and an effective guardrail solution that significantly improves safety and reliability.

Abstract: Computer-use agents (CUAs) have made tremendous progress in the past year, yet they still frequently produce misaligned actions that deviate from the user’s original intent. Such misaligned actions may arise from external attacks (e.g., indirect prompt injection) or from internal limitations (e.g., erroneous reasoning). They not only expose CUAs to safety risks, but also degrade task efficiency and reliability. This work makes the first effort to define and study misaligned action detection in CUAs, with comprehensive coverage of both externally induced and internally arising misaligned actions. We further identify three common categories in real-world CUA deployment and construct MisActBench, a benchmark of realistic trajectories with human-annotated, action-level alignment labels. Moreover, we propose DeAction, a practical and universal guardrail that detects misaligned actions before execution and iteratively corrects them through structured feedback. DeAction outperforms all existing baselines across offline and online evaluations with moderate latency overhead: (1) On MisActBench, it outperforms baselines by over 15% absolute in F1 score; (2) In online evaluation, it reduces attack success rate by over 90% under adversarial settings while preserving or even improving task success rate in benign environments.

Method: YaRN (Yet another RoPE extensioN method) is a compute-efficient approach that modifies the RoPE mechanism to enable context window extension. It requires 10x fewer tokens and 2.5x fewer training steps than previous methods while maintaining model performance.

Result: YaRN successfully extends LLaMA models’ context windows beyond their original pre-training limits, outperforming previous state-of-the-art methods. The method also demonstrates extrapolation capability beyond the fine-tuning dataset’s limited context.

Conclusion: YaRN provides an efficient solution for extending transformer models’ context windows, making long-context applications more practical while requiring significantly less computational resources than previous approaches.

Abstract: Rotary Position Embeddings (RoPE) have been shown to effectively encode positional information in transformer-based language models. However, these models fail to generalize past the sequence length they were trained on. We present YaRN (Yet another RoPE extensioN method), a compute-efficient method to extend the context window of such models, requiring 10x less tokens and 2.5x less training steps than previous methods. Using YaRN, we show that LLaMA models can effectively utilize and extrapolate to context lengths much longer than their original pre-training would allow, while also surpassing previous the state-of-the-art at context window extension. In addition, we demonstrate that YaRN exhibits the capability to extrapolate beyond the limited context of a fine-tuning dataset. Code is available at https://github.com/jquesnelle/yarn

Method: Created Multilingual CORE corpora with 72,000+ documents annotated with hierarchical taxonomy of 25 registers; used multi-label classification with deep learning models; compared multilingual vs. monolingual approaches; conducted data pruning and analysis of hybrid texts.

Result: Best model achieved 79% F1 averaged across 16 languages; performance ceiling observed across all models; data pruning increased performance to 90% F1; multilingual models outperformed monolingual ones, especially for low-resource languages; zero-shot performance dropped 3-8% on unseen languages.

Conclusion: Models can perform well with complex register schemes at multilingual scale, but performance is limited by inherent ambiguity in web registers rather than model limitations; multilingual approaches are beneficial, particularly for languages with limited training data.

Abstract: This article presents multilingual deep learning models for identifying web registers – text varieties such as news reports and discussion forums – across 16 languages. We introduce the Multilingual CORE corpora, which contain over 72,000 documents annotated with a hierarchical taxonomy of 25 registers designed to cover the entire open web. Using multi-label classification, our best model achieves 79% F1 averaged across languages, matching or exceeding previous studies that used simpler classification schemes. This demonstrates that models can perform well even with a complex register scheme at multilingual scale. However, we observe a consistent performance ceiling across all models and configurations. When we remove documents with uncertain labels through data pruning, performance increases to over 90% F1, suggesting that this ceiling stems from inherent ambiguity in web registers rather than model limitations. Analysis of hybrid texts (those combining multiple registers) reveals that the main challenge lies not in classifying hybrids themselves, but in distinguishing hybrid from non-hybrid documents. Multilingual models consistently outperform monolingual ones, particularly for languages with limited training data. Zero-shot performance on unseen languages drops by an average of 7%, though this varies by language (3–8%), indicating that while registers share features across languages, they also retain language-specific characteristics.

Method: Analyzed LM predictions using n-gram statistics to compare with training data frequencies, examined learning dynamics to study generalization to unseen adjective combinations, and used feature attribution methods to identify how LMs leverage contextual cues from sentence context.

Result: LM predictions are largely explained by training data frequencies (n-gram statistics account for much behavior), but models also generalize robustly to unseen adjective combinations and leverage contextual cues from sentence context, indicating behavior goes beyond simple memorization.

Conclusion: While distributional learning from training data frequencies explains much of LM adjective ordering behavior, models also demonstrate generalization capabilities and contextual sensitivity that go beyond surface co-occurrence patterns, suggesting more sophisticated linguistic learning.

Abstract: In English and other languages, multiple adjectives in noun phrases follow intricate ordering patterns. These patterns have been widely studied in linguistics and provide a useful test case for assessing how language models (LMs) acquire graded and context-sensitive word order preferences. We ask to what extent adjective order preferences in LMs can be explained by distributional learning alone, and where models exhibit behaviour that goes beyond surface co-occurrence patterns. We find that LM predictions are largely explained by training data frequencies: simple n-gram statistics account for much of their behaviour and closely mirror the preferences learned during training. However, by analysing learning dynamics we reveal that models also generalize robustly to unseen adjective combinations, indicating that their behaviour cannot be reduced to memorization of observed orders alone. Moreover, we show how LMs leverage word order cues from sentence context, demonstrating with feature attribution methods that contextual cues are an additional driver of adjective order in LM output.

Method: Introduces ComfyBench benchmark with 200 diverse tasks and 3,205 annotated nodes across 20 workflows. Develops ComfyAgent framework with two core concepts: 1) representing workflows as code that can be reversibly converted and executed, and 2) constructing multi-agent systems that cooperate to learn from existing workflows and generate new ones.

Result: ComfyAgent achieves comparable resolve rate to o1-preview and significantly surpasses other agents on ComfyBench, but only resolves 15% of creative tasks, showing limitations in autonomous design of collaborative systems.

Conclusion: LLM-based agents still have significant room for improvement in autonomously designing collaborative AI systems, particularly for creative tasks. ComfyBench provides a foundation for developing more intelligent and autonomous collaborative AI systems.

Abstract: Much previous AI research has focused on developing monolithic models to maximize their intelligence, with the primary goal of enhancing performance on specific tasks. In contrast, this work attempts to study using LLM-based agents to design collaborative AI systems autonomously. To explore this problem, we first introduce ComfyBench to evaluate agents’s ability to design collaborative AI systems in ComfyUI. ComfyBench is a comprehensive benchmark comprising 200 diverse tasks covering various instruction-following generation challenges, along with detailed annotations for 3,205 nodes and 20 workflows. Based on ComfyBench, we further develop ComfyAgent, a novel framework that empowers LLM-based agents to autonomously design collaborative AI systems by generating workflows. ComfyAgent is based on two core concepts. First, it represents workflows with code, which can be reversibly converted into workflows and executed as collaborative systems by the interpreter. Second, it constructs a multi-agent system that cooperates to learn from existing workflows and generate new workflows for a given task. While experimental results demonstrate that ComfyAgent achieves a comparable resolve rate to o1-preview and significantly surpasses other agents on ComfyBench, ComfyAgent has resolved only 15% of creative tasks. LLM-based agents still have a long way to go in autonomously designing collaborative AI systems. Progress with ComfyBench is paving the way for more intelligent and autonomous collaborative AI systems.

Method: Translate widely-used mental health datasets from English into six languages (Greek, Turkish, French, Portuguese, German, Finnish) and evaluate LLMs (GPT and Llama) on detecting mental health conditions and severity assessment across languages.

Result: Considerable variability in LLM performance across languages despite using the same translated dataset, highlighting language-specific nuances and mental health data coverage issues affecting accuracy.

Conclusion: LLMs show inconsistent performance in multilingual mental health support, with risks of misdiagnosis in medical settings; the approach offers cost savings for multilingual tasks but requires caution in clinical applications.

Abstract: Large Language Models (LLMs) are increasingly being integrated into various medical fields, including mental health support systems. However, there is a gap in research regarding the effectiveness of LLMs in non-English mental health support applications. To address this problem, we present a novel multilingual adaptation of widely-used mental health datasets, translated from English into six languages (e.g., Greek, Turkish, French, Portuguese, German, and Finnish). This dataset enables a comprehensive evaluation of LLM performance in detecting mental health conditions and assessing their severity across multiple languages. By experimenting with GPT and Llama, we observe considerable variability in performance across languages, despite being evaluated on the same translated dataset. This inconsistency underscores the complexities inherent in multilingual mental health support, where language-specific nuances and mental health data coverage can affect the accuracy of the models. Through comprehensive error analysis, we emphasize the risks of relying exclusively on LLMs in medical settings (e.g., their potential to contribute to misdiagnoses). Moreover, our proposed approach offers significant cost savings for multilingual tasks, presenting a major advantage for broad-scale implementation.

Method: RARe finetunes pre-trained encoder-only models with in-context examples where the query is semantically similar to the target query. The approach analyzes design choices for in-context example augmentation for retrievers.

Result: RARe achieves performance gains of up to +2.72% nDCG across open-domain retrieval datasets (BeIR, RAR-b) compared to using only the target query. It exhibits stronger out-of-domain generalization similar to in-context learning in LLMs.

Conclusion: In-context learning can effectively enhance encoder-only retrieval models, with RARe demonstrating improved performance and generalization. The work lays foundation for future research on in-context learning for encoder models.

Abstract: While in-context learning is well-studied with decoder-only language models (LLMs), its utility for encoder-only models remains underexplored. We study in-context learning for encoder-only models for text retrieval tasks. Can incorporating in-context examples (query-document pairs) to the target query enhance retriever performance? Our approach, RARe, finetunes a pre-trained model with in-context examples whose query is semantically similar to the target query. This approach achieves performance gains of up to +2.72% nDCG across open-domain retrieval datasets (BeIR, RAR-b) compared to using the target query only as an input. In particular, we find RARe exhibits stronger out-of-domain generalization compared to models using queries without in-context examples, similar to what is seen for in-context learning in LLMs. We further provide analysis on the design choices of in-context example augmentation for retrievers and lay the foundation for future work.

Method: Three-study approach: Study 1 optimized prompt design and temperature settings on GPT-4; Study 2 tested cross-model generalizability on ChatGPT-5; Study 3 evaluated psychometric properties of generated SJTs across Big Five personality facets.

Result: Optimized prompts with temperature 1.0 worked best on GPT-4; approach generalized well to ChatGPT-5 with reproducible high-quality items; generated SJTs showed satisfactory reliability/validity across most Big Five facets, though some limitations in compliance facet and criterion-related validity.

Conclusion: LLM-based AIG approach can efficiently produce culturally appropriate and psychometrically sound personality SJTs, potentially surpassing traditional methods in efficiency while maintaining quality.

Abstract: Personality assessment through situational judgment tests (SJTs) offers unique advantages over traditional Likert-type self-report scales, yet their development remains labor-intensive, time-consuming, and heavily dependent on subject matter experts. Recent advances in large language models (LLMs) have shown promise for automatic item generation (AIG). Building on these developments, the present study focuses on developing and evaluating a structured and generalizable framework for automatically generating personality SJTs, using GPT-4 and ChatGPT-5 as empirical examples. Three studies were conducted. Study 1 systematically compared the effects of prompt design and temperature settings on the content validity of LLM-generated items to develop an effective and stable LLM-based AIG approach for personality SJT. Results showed that optimized prompts and a temperature of 1.0 achieved the best balance of creativity and accuracy on GPT-4. Study 2 examined the cross-model generalizability and reproducibility of this automated SJT generation approach through multiple rounds. The results showed that the approach consistently produced reproducible and high-quality items on ChatGPT-5. Study 3 evaluated the psychometric properties of LLM-generated SJTs covering five facets of the Big Five personality traits. Results demonstrated satisfactory reliability and validity across most facets, though limitations were observed in the convergent validity of the compliance facet and certain aspects of criterion-related validity. These findings provide robust evidence that the proposed LLM-based AIG approach can produce culturally appropriate and psychometrically sound SJTs with efficiency comparable to or exceeding traditional methods.

Method: Analyze model internals to discover attention heads that encode “sufficiency signals.” Use lightweight classifiers to detect when critical information has been processed, enabling models to self-terminate early. Larger models can use prompting for self-assessment.

Result: 3.4% accuracy improvement with 1.33x token reduction on average across six QA datasets. Superior performance compared to other context efficiency methods at equivalent token reduction rates. Emergent scaling: larger models show intrinsic self-assessment capabilities.

Conclusion: Models’ internal understanding naturally dictates processing needs rather than external compression heuristics. Dynamic context cutoff reveals a new efficiency paradigm for LLMs.

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

Method: Two variants: k-NLPmeans uses lightweight deterministic summarizers for offline, low-cost operation; k-LLMmeans uses LLMs for summaries under fixed per-iteration budget. Both retain k-means assignments in embedding space but replace numeric centroids with textual summaries. Includes mini-batch extension for streaming text.

Result: Consistently outperforms classical baselines across diverse datasets and embedding models, approaching accuracy of recent LLM-based clustering without extensive LLM calls. Provides case study on sequential text streams and releases StackExchange-derived benchmark.

Conclusion: Summary-as-centroid approach enables human-readable, auditable text clustering with computational efficiency, offering LLM-optional solutions that balance interpretability and performance.

Abstract: We introduce k-NLPmeans and k-LLMmeans, text-clustering variants of k-means that periodically replace numeric centroids with textual summaries. The key idea, summary-as-centroid, retains k-means assignments in embedding space while producing human-readable, auditable cluster prototypes. The method is LLM-optional: k-NLPmeans uses lightweight, deterministic summarizers, enabling offline, low-cost, and stable operation; k-LLMmeans is a drop-in upgrade that uses an LLM for summaries under a fixed per-iteration budget whose cost does not grow with dataset size. We also present a mini-batch extension for real-time clustering of streaming text. Across diverse datasets, embedding models, and summarization strategies, our approach consistently outperforms classical baselines and approaches the accuracy of recent LLM-based clustering-without extensive LLM calls. Finally, we provide a case study on sequential text streams and release a StackExchange-derived benchmark for evaluating streaming text clustering.

Method: Created MET-Bench with two structured domains to evaluate multimodal entity tracking. Developed a reinforcement learning method to improve performance on the benchmark and applied it to open-source VLMs.

Result: Found significant performance gap between text-based and image-based entity tracking, with the discrepancy stemming from visual reasoning deficits rather than perception. Their RL method improved open-source VLMs to competitive performance with advanced closed models.

Conclusion: There’s a need for improved multimodal representations and reasoning techniques to bridge the gap between textual and visual entity tracking, especially for long-horizon multimodal tasks.

Abstract: Entity state tracking is a necessary component of world modeling that requires maintaining coherent representations of entities over time. Previous work has benchmarked entity tracking performance in purely text-based tasks. We introduce MET-Bench, a multimodal entity tracking benchmark designed to evaluate vision-language models’ ability to track entity states across modalities. Using two structured domains, we assess how effectively current models integrate textual and image-based state updates. Our findings reveal a significant performance gap between text-based and image-based entity tracking. We empirically show this discrepancy primarily stems from deficits in visual reasoning rather than perception. We further show that explicit text-based reasoning strategies improve performance, yet limitations remain in long-horizon multimodal tasks. We develop a reinforcement learning method to improve performance on MET-Bench. Applying our method to open-source VLMs achieves competitive performance with advanced closed models. Our results highlight the need for improved multimodal representations and reasoning techniques to bridge the gap between textual and visual entity tracking.

Method: Developed a two-tier hierarchical safety taxonomy with 6 dimensions, generated over 4000 multi-turn dialogues in Chinese/English across 22 scenarios, employed 7 jailbreak attack strategies, and created an assessment framework measuring detection, handling, and consistency capabilities.

Result: Evaluation of 17 LLMs shows Yi-34B-Chat and GLM4-9B-Chat have superior safety performance, while Llama3.1-8B-Instruct and o3-mini exhibit safety vulnerabilities.

Conclusion: SafeDialBench provides a comprehensive benchmark for assessing LLM safety in multi-turn dialogues against various jailbreak attacks, revealing significant performance differences among models.

Abstract: With the rapid advancement of Large Language Models (LLMs), the safety of LLMs has been a critical concern requiring precise assessment. Current benchmarks primarily concentrate on single-turn dialogues or a single jailbreak attack method to assess the safety. Additionally, these benchmarks have not taken into account the LLM’s capability of identifying and handling unsafe information in detail. To address these issues, we propose a fine-grained benchmark SafeDialBench for evaluating the safety of LLMs across various jailbreak attacks in multi-turn dialogues. Specifically, we design a two-tier hierarchical safety taxonomy that considers 6 safety dimensions and generates more than 4000 multi-turn dialogues in both Chinese and English under 22 dialogue scenarios. We employ 7 jailbreak attack strategies, such as reference attack and purpose reverse, to enhance the dataset quality for dialogue generation. Notably, we construct an innovative assessment framework of LLMs, measuring capabilities in detecting, and handling unsafe information and maintaining consistency when facing jailbreak attacks. Experimental results across 17 LLMs reveal that Yi-34B-Chat and GLM4-9B-Chat demonstrate superior safety performance, while Llama3.1-8B-Instruct and o3-mini exhibit safety vulnerabilities.

Method: Created ExpliCa dataset with causal and temporal relations in different linguistic orders, enriched with human acceptability ratings. Tested 7 LLMs using prompting and perplexity-based metrics.

Result: Even top models struggle to reach 0.80 accuracy. Models confuse temporal with causal relations, and performance is strongly influenced by linguistic order of events. Perplexity-based scores and prompting performance are differently affected by model size.

Conclusion: LLMs have significant limitations in explicit causal reasoning, particularly with temporal-causal distinctions and linguistic ordering effects, highlighting need for improved reasoning capabilities.

Abstract: Large Language Models (LLMs) are increasingly used in tasks requiring interpretive and inferential accuracy. In this paper, we introduce ExpliCa, a new dataset for evaluating LLMs in explicit causal reasoning. ExpliCa uniquely integrates both causal and temporal relations presented in different linguistic orders and explicitly expressed by linguistic connectives. The dataset is enriched with crowdsourced human acceptability ratings. We tested LLMs on ExpliCa through prompting and perplexity-based metrics. We assessed seven commercial and open-source LLMs, revealing that even top models struggle to reach 0.80 accuracy. Interestingly, models tend to confound temporal relations with causal ones, and their performance is also strongly influenced by the linguistic order of the events. Finally, perplexity-based scores and prompting performance are differently affected by model size.

Method: Created PLAT benchmark with 300 examples from 100 Korean court precedents, including binary-choice, multiple-choice, and essay-type questions to evaluate LLMs’ understanding of tax law and complex reasoning beyond simple statute application.

Result: Experiments show LLMs have limited baseline capabilities, especially in cases with conflicting issues requiring comprehensive understanding, and struggle with “AC” stages of “IRAC” even for advanced reasoning models like o3.

Conclusion: PLAT reveals significant limitations in current LLMs for tax law reasoning, highlighting the need for improved models that can handle complex legal reasoning and comprehensive case understanding.

Abstract: How capable are large language models (LLMs) in the domain of taxation? Although numerous studies have explored the legal domain, research dedicated to taxation remains scarce. Moreover, the datasets used in these studies are either simplified, failing to reflect the real-world complexities, or not released as open-source. To address this gap, we introduce PLAT, a new benchmark designed to assess the ability of LLMs to predict the legitimacy of additional tax penalties. PLAT comprises 300 examples: (1) 100 binary-choice questions, (2) 100 multiple-choice questions, and (3) 100 essay-type questions, all derived from 100 Korean court precedents. PLAT is constructed to evaluate not only LLMs’ understanding of tax law but also their performance in legal cases that require complex reasoning beyond straightforward application of statutes. Our systematic experiments with multiple LLMs reveal that (1) their baseline capabilities are limited, especially in cases involving conflicting issues that require a comprehensive understanding (not only of the statutes but also of the taxpayer’s circumstances), and (2) LLMs struggle particularly with the “AC” stages of “IRAC” even for advanced reasoning models like o3, which actively employ inference-time scaling. The dataset is publicly available at: https://huggingface.co/collections/sma1-rmarud/plat-predicting-the-legitimacy-of-punitive-additional-tax

Method: PlainQAFact first classifies sentence types, then applies a retrieval-augmented QA scoring method. It’s trained on PlainFact, a fine-grained human-annotated dataset for evaluating factual consistency of both source-simplified and elaboratively explained sentences.

Result: PlainQAFact consistently outperforms existing evaluation metrics across all settings, especially for elaborative explanations. The paper analyzes its effectiveness across knowledge sources, answer extraction strategies, overlap measures, and document granularity levels.

Conclusion: The work presents a sentence-aware, retrieval-augmented metric targeted at elaborative explanations in biomedical PLS tasks, providing both a robust benchmark and practical tool for reliable plain language communication in medical domains.

Abstract: Hallucinated outputs from large language models (LLMs) pose risks in the medical domain, especially for lay audiences making health-related decisions. Existing automatic factual consistency evaluation methods, such as entailment- and question-answering (QA)- based, struggle with plain language summarization (PLS) due to elaborative explanation phenomenon, which introduces external content (e.g., definitions, background, examples) absent from the scientific abstract to enhance comprehension. To address this, we introduce PlainQAFact, an automatic factual consistency evaluation metric trained on a fine-grained, human-annotated dataset PlainFact, for evaluating factual consistency of both source-simplified and elaborately explained sentences. PlainQAFact first classifies sentence type, then applies a retrieval-augmented QA scoring method. Empirical results show that existing evaluation metrics fail to evaluate the factual consistency in PLS, especially for elaborative explanations, whereas PlainQAFact consistently outperforms them across all evaluation settings. We further analyze PlainQAFact’s effectiveness across external knowledge sources, answer extraction strategies, answer overlap measures, and document granularity levels, refining its overall factual consistency assessment. Taken together, our work presents a sentence-aware, retrieval-augmented metric targeted at elaborative explanations in biomedical PLS tasks, providing the community with both a robust benchmark and a practical tool to advance reliable and safe plain language communication in the medical domain. PlainQAFact and PlainFact are available at: https://github.com/zhiwenyou103/PlainQAFact

Method: LSR-MCTS algorithm that treats each code line as fundamental unit, uses Monte Carlo Tree Search to determine code line-by-line and select optimal path. Integrates self-refine mechanism at each node for diversity enhancement and error correction.

Result: Extensive experiments on three public coding benchmarks demonstrate method outperforms state-of-the-art performance approaches.

Conclusion: Line-by-line code generation with MCTS and self-refinement effectively addresses redundancy and local pattern overfitting, achieving superior performance on code generation benchmarks.

Abstract: The emergence of large language models (LLMs) has significantly promoted the development of code generation task, sparking a surge in pertinent literature. Current research is hindered by redundant generation results and a tendency to overfit local patterns in the short term. Although existing studies attempt to alleviate the issue by adopting a multi-token prediction strategy, there remains limited focus on choosing the appropriate processing length for generations. By analyzing the attention between tokens during the generation process of LLMs, it can be observed that the high spikes of the attention scores typically appear at the end of lines. This insight suggests that it is reasonable to treat each line of code as a fundamental processing unit and generate them sequentially. Inspired by this, we propose the LSR-MCTS algorithm, which leverages MCTS to determine the code line-by-line and select the optimal path. Further, we integrate a self-refine mechanism at each node to enhance diversity and generate higher-quality programs through error correction. Extensive experiments and comprehensive analyses on three public coding benchmarks demonstrate that our method outperforms the state-of-the-art performance approaches.

Method: Proposes BiGTex with stacked Graph-Text Fusion Units that enable mutual attention between textual and structural representations, allowing bidirectional information flow. Uses parameter-efficient fine-tuning (LoRA) with frozen LLM weights while adapting to task-specific signals.

Result: Extensive experiments on five benchmark datasets demonstrate state-of-the-art performance in node classification and effective generalization to link prediction. Ablation study highlights importance of soft prompting and bi-directional attention.

Conclusion: BiGTex successfully integrates GNNs and LLMs for text-attributed graph representation learning, achieving strong performance through bidirectional fusion of textual and structural information.

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

Method: Proposes GMSA with Group Merging for uniform aggregation to mitigate semantic dominance during autoencoder pretraining, and Layer Semantic Alignment to bridge semantic gaps between high-level abstract and low-level input semantics. Uses autoencoder pretraining followed by downstream task fine-tuning.

Result: GMSA improves context reconstruction compared to existing soft prompt compression methods and outperforms baselines on multiple long-context QA and summarization benchmarks across two backbone models while maintaining low latency.

Conclusion: GMSA effectively addresses computational and redundancy challenges in long-context LLMs through innovative compression techniques, demonstrating practical improvements in both performance and efficiency.

Abstract: Large Language Models (LLMs) have achieved remarkable performance across a wide range of Natural Language Processing (NLP) tasks. However, in long-context scenarios, they face two challenges: high computational cost and information redundancy. To address these challenges, we propose GMSA, an encoder-decoder context compression framework that generates a compact sequence of soft tokens for downstream tasks. GMSA introduces Group Merging to achieve more uniform aggregation, mitigating semantic dominance during autoencoder pretraining, and Layer Semantic Alignment (LSA) to bridge the semantic gap between high-level abstract semantics and low-level input semantics. We first pretrain GMSA as an autoencoder and then fine-tune it for downstream tasks. Experiments demonstrate that GMSA improves context reconstruction compared to existing soft prompt compression paradigm and outperforms baselines on multiple long-context question answering and summarization benchmarks across two backbone models, while maintaining low end-to-end latency.

Method: ABBA reparameterizes weight updates as a Hadamard product of two independently learnable low-rank matrices, fully decoupling the update from pre-trained weights. This allows both components to be optimized freely, increasing expressivity under the same parameter budget.

Result: ABBA achieves state-of-the-art results on arithmetic and commonsense reasoning benchmarks, consistently outperforming existing PEFT methods by a significant margin across multiple models. Matrix reconstruction experiments validate its higher expressivity.

Conclusion: ABBA represents a significant advancement in PEFT methods by decoupling updates from pre-trained weights, enabling higher expressivity and better adaptation of LLMs to new domains while maintaining parameter efficiency.

Abstract: Large Language Models have demonstrated strong performance across a wide range of tasks, but adapting them efficiently to new domains remains a key challenge. Parameter-Efficient Fine-Tuning (PEFT) methods address this by introducing lightweight, trainable modules while keeping most pre-trained weights fixed. The prevailing approach, LoRA, models updates using a low-rank decomposition, but its expressivity is inherently constrained by the rank. Recent methods like HiRA aim to increase expressivity by incorporating a Hadamard product with the frozen weights, but still rely on the structure of the pre-trained model. We introduce ABBA, a new PEFT architecture that reparameterizes the update as a Hadamard product of two independently learnable low-rank matrices. In contrast to prior work, ABBA fully decouples the update from the pre-trained weights, enabling both components to be optimized freely. This leads to significantly higher expressivity under the same parameter budget, a property we validate through matrix reconstruction experiments. Empirically, ABBA achieves state-of-the-art results on arithmetic and commonsense reasoning benchmarks, consistently outperforming existing PEFT methods by a significant margin across multiple models. Our code is publicly available at: https://github.com/CERT-Lab/abba.

Method: FAID uses multi-level contrastive learning with multi-task auxiliary classification to capture subtle stylistic cues, modeling LLM families as distinct stylistic entities with adaptation mechanisms for distributional shifts without retraining.

Result: FAID outperforms several baselines, particularly enhancing generalization accuracy on unseen domains and new LLMs, demonstrating effectiveness in fine-grained text classification.

Conclusion: FAID offers a potential solution for improving transparency and accountability in AI-assisted writing through fine-grained detection of authorship and model characteristics.

Abstract: The growing collaboration between humans and AI models in generative tasks has introduced new challenges in distinguishing between human-written, LLM-generated, and human-LLM collaborative texts. In this work, we collect a multilingual, multi-domain, multi-generator dataset FAIDSet. We further introduce a fine-grained detection framework FAID to classify text into these three categories, and also to identify the underlying LLM family of the generator. Unlike existing binary classifiers, FAID is built to capture both authorship and model-specific characteristics. Our method combines multi-level contrastive learning with multi-task auxiliary classification to learn subtle stylistic cues. By modeling LLM families as distinct stylistic entities, we incorporate an adaptation to address distributional shifts without retraining for unseen data. Our experimental results demonstrate that FAID outperforms several baselines, particularly enhancing the generalization accuracy on unseen domains and new LLMs, thus offering a potential solution for improving transparency and accountability in AI-assisted writing. Our data and code are available at https://github.com/mbzuai-nlp/FAID

Method: Develops a fully automated pipeline to generate Programming by Examples (PBE) problems involving string rewrite program cascades. Creates two benchmarks: PBEBench-Lite for efficient model stratification and PBEBench with complexity similar to historical linguistics tasks.

Result: Shows substantial performance gap between models using test-time compute/Long Chain-of-Thought vs. those without. Recent models drop below 5% solve rate for hard instances (cascade lengths 20-30), falling short of historical linguistics requirements even with advanced scaling techniques.

Conclusion: The benchmarks effectively evaluate inductive reasoning capabilities, revealing current LLMs’ limitations in complex reasoning tasks despite advances. The approach enables scalable, contamination-free evaluation of reasoning models.

Abstract: Although many benchmarks evaluate the reasoning abilities of Large Language Models (LLMs) within domains such as mathematics, coding, or data wrangling, few abstract away from domain specifics to examine reasoning as a capability in and of itself. We contribute a novel type of benchmark evaluating the inductive reasoning capabilities of LLMs that is inspired by the forward reconstruction task from historical linguistics but is formulated in an extremely simple, general way (in the form of Programming by Examples). The task involves generating a cascade of simple string rewrite programs to transform a given list of input strings into a list of desired output strings. We present a fully automated pipeline that programmatically generates problems of this type with controllable difficulty, enabling scalable evaluation of reasoning models while avoiding contamination. Using this approach, we construct two benchmarks: PBEBench-Lite, which efficiently stratifies models of varying capabilities, and PBEBench, which requires models to induce programs similar in complexity to those constructed by historical linguists. Our experiments reveal a substantial performance gap between models that leverage test-time compute or LCoT (long chain-of-thought) reasoning and those that do not. Moreover, although recent models show promise, the solve rate for both of them drops below 5% for hard instances of the PBEBench dataset (ground truth cascade lengths of 20 and 30, respectively), falling well short of realistic historical linguistics requirements even with computationally expensive, popular scaling techniques from the PBE and reasoning literature. Additionally, we also study the effectiveness of different scaling strategies and the impact of various hyperparameters on the difficulty of the generated data using gpt-oss-120b, the best-performing open-source model.

Method: Multi-agent LLM system powered by DeepSeek-R1/V3 with structured domain knowledge bases, precise error locator, iterative reflection, and dedicated Physics Interpreter. Uses modular, MCP-compatible design for collaborative AI-driven CFD.

Result: Achieved 82.1% execution success on 315 benchmark cases (vs. 6.2% for MetaOpenFOAM, 42.3% for Foam-Agent), 68.12% physical fidelity, 97.4% summary fidelity, and efficient resource usage (192.1k tokens, $0.208 per case).

Conclusion: ChatCFD dramatically outperforms existing methods, enabling scalable, collaborative AI-driven CFD automation with high success rates and physical fidelity while maintaining resource efficiency.

Abstract: Computational Fluid Dynamics (CFD) is critical for scientific advancement but is hindered by operational complexity and high expertise barriers. This paper introduces ChatCFD, a Large Language Model (LLM)-driven multi-agent system designed for end-to-end CFD automation using OpenFOAM. Powered by DeepSeek-R1/V3, ChatCFD integrates structured domain knowledge bases, a precise error locator, and iterative reflection to dramatically outperform existing methods. On 315 benchmark cases, ChatCFD achieves 82.1% execution success (vs. 6.2% for MetaOpenFOAM and 42.3% for Foam-Agent) and 68.12% physical fidelity - a novel metric assessing scientific meaningfulness beyond mere runnability. A dedicated Physics Interpreter attains 97.4% summary fidelity, bridging the gap between narrative fluency and the enforcement of tight physical constraints. Resource analysis confirms efficiency, averaging 192.1k tokens and $0.208 per case, significantly lower than baseline costs. Ablation studies identify the Error Locator and Solver Template DB as critical, with the latter’s removal collapsing accuracy to 48%. The system exhibits robust flexibility, achieving 95.23% success in autonomous solver selection and 100% in turbulence modeling, while successfully reproducing complex literature cases (e.g., NACA0012, supersonic nozzle) with 60-80% success rates where baselines failed. Featuring a modular, MCP-compatible design, ChatCFD facilitates scalable, collaborative AI-driven CFD. Code is available at: https://github.com/ConMoo/ChatCFD

Method: Introduced Vacuous Neutrality Framework (VaNeu) - a multi-dimensional evaluation paradigm examining model robustness across four stages: biases, utility, ambiguity handling, and positional bias over diverse social bias categories. Conducted first large-scale audit of 9 SLMs (0.5-5B parameters) from four model families under both ambiguous and disambiguated contexts.

Result: Models demonstrating low bias in early stages often fail subsequent evaluations, revealing hidden vulnerabilities and unreliable reasoning. The framework exposes limitations in current SLM fairness assessments.

Conclusion: There’s a need for more comprehensive understanding of fairness and reliability in SLMs, and the VaNeu framework serves as a principled tool for responsible deployment in socially sensitive settings.

Abstract: The rapid adoption of Small Language Models (SLMs) for resource constrained applications has outpaced our understanding of their ethical and fairness implications. To address this gap, we introduce the Vacuous Neutrality Framework (VaNeu), a multi-dimensional evaluation paradigm designed to assess SLM fairness prior to deployment. The framework examines model robustness across four stages - biases, utility, ambiguity handling, and positional bias over diverse social bias categories. To the best of our knowledge, this work presents the first large-scale audit of SLMs in the 0.5-5B parameter range, an overlooked “middle tier” between BERT-class encoders and flagship LLMs. We evaluate nine widely used SLMs spanning four model families under both ambiguous and disambiguated contexts. Our findings show that models demonstrating low bias in early stages often fail subsequent evaluations, revealing hidden vulnerabilities and unreliable reasoning. These results underscore the need for a more comprehensive understanding of fairness and reliability in SLMs, and position the proposed framework as a principled tool for responsible deployment in socially sensitive settings.

Method: CGLS framework gradually adds layers during training while simultaneously increasing data difficulty through curriculum learning - starting with simpler synthetic data and progressing to more complex, specialized content.

Result: At 100M parameter scale with synthetic-to-web data curriculum, CGLS outperforms baselines on PIQA and ARC benchmarks. At 1.2B scale with DataComp-LM corpus stratified by technicality, it shows better generalization and zero-shot performance on various downstream tasks.

Conclusion: CGLS unlocks the potential of progressive stacking, offering a simple yet effective strategy for improving generalization on knowledge-intensive and reasoning tasks through synchronized model growth and data difficulty progression.

Abstract: As the cost of pretraining large language models grows, there is continued interest in strategies to improve learning efficiency during this core training stage. Motivated by cognitive development, where humans gradually build knowledge as their brains mature, we propose Curriculum-Guided Layer Scaling (CGLS), a framework for compute-efficient pretraining that synchronizes increasing data difficulty with model growth through progressive layer stacking (i.e. gradually adding layers during training). At the 100M parameter scale, using a curriculum transitioning from synthetic short stories to general web data, CGLS outperforms baseline methods on the question-answering benchmarks PIQA and ARC. Pretraining at the 1.2B scale, we stratify the DataComp-LM corpus with a DistilBERT-based classifier and progress from general text to highly technical or specialized content. Our results show that progressively increasing model depth alongside sample difficulty leads to better generalization and zero-shot performance on various downstream benchmarks. Altogether, our findings demonstrate that CGLS unlocks the potential of progressive stacking, offering a simple yet effective strategy for improving generalization on knowledge-intensive and reasoning tasks.

Method: Language Bottleneck Models (LBMs) use an encoder LLM to produce textual knowledge state summaries, which a decoder LLM then uses to predict future student performance. The models can be fine-tuned with reinforcement learning (encoder) and supervised fine-tuning (decoder) to improve summary quality and predictive performance.

Result: LBMs reveal qualitative insights beyond what traditional CD and KT models can capture, achieve competitive accuracy with improved sample efficiency, and demonstrate that fine-tuning improves both summary quality and predictive performance.

Conclusion: LBMs offer a promising approach to student knowledge assessment that combines interpretability with competitive predictive performance, enabling capture of nuanced insights like misconceptions that traditional quantitative models cannot express.

Abstract: Accurately assessing student knowledge is central to education. Cognitive Diagnosis (CD) models estimate student proficiency at a fixed point in time, while Knowledge Tracing (KT) methods model evolving knowledge states to predict future performance. However, existing approaches either provide quantitative concept mastery estimates with limited expressivity (CD, probabilistic KT) or prioritize predictive accuracy at the cost of interpretability (deep learning KT). We propose Language Bottleneck Models (LBMs), where an encoder LLM produces textual knowledge state summaries, which a decoder LLM uses to predict future performance. This produces interpretable summaries that can express nuanced insights–such as misconceptions–that CD and KT models cannot capture. Extensive validation across synthetic and real-world datasets shows LBMs reveal qualitative insights beyond what CD and KT models can capture, while achieving competitive accuracy with improved sample efficiency. We demonstrate that the encoder and decoder can be fine-tuned with reinforcement learning and supervised fine-tuning respectively to improve both summary quality and predictive performance.

Method: Identify checkpoint models with specific capabilities on benchmarks, then use their aggregated first-order influence approximation over source data to determine optimal data mixtures. The framework leverages training artifacts to enhance data quality.

Result: Demonstrated on eight reasoning benchmarks with significant improvements in pretraining setting, achieving performance improvements of up to 1.93%. Shows checkpoint models can effectively enhance data quality and optimize mixtures.

Conclusion: Checkpoint models from training trajectories are valuable resources for optimizing data mixtures, providing a practical approach to enhance language model capabilities through better data selection.

Abstract: In language model training, it is desirable to equip models with capabilities from various tasks. However, it is not clear how to directly obtain the right data mixtures for these capabilities as the relationship between data and tasks is difficult to be modeled. In this work, we observe that checkpoint models exhibit emerging capabilities at different points in the training trajectory. Often, the training process saves checkpoints as artifacts that are under-utilized as a source of in-training data signals. We identify these artifact models based on their respective capabilities on the benchmarks and leverage them as data mixers by using their aggregated first-order influence approximation over source data. We demonstrated on eight reasoning benchmarks that the proposed framework shows significant improvements in the pretraining setting, with performance improvements of up to 1.93%. Overall, this shows the potential of checkpoint models to enhance data quality and optimize data mixtures.

Method: Proposes MedVAL, a novel self-supervised distillation method that leverages synthetic data to train evaluator LMs to assess factual consistency between LM-generated medical outputs and inputs. Introduces MedVAL-Bench dataset of 840 physician-annotated outputs across 6 diverse medical tasks.

Result: MedVAL significantly improves alignment with physicians across seen and unseen tasks, increasing average F1 scores from 66% to 83%. Improves best-performing proprietary LM (GPT-4o) by 8% without training on physician-labeled data, achieving performance statistically non-inferior to a single human expert.

Conclusion: MedVAL enables scalable, risk-aware evaluation of LM-generated medical text, approaching expert-level ability in validation. The method and benchmark support clinical integration of language models in medical applications.

Abstract: With the growing use of language models (LMs) in clinical environments, there is an immediate need to evaluate the accuracy and safety of LM-generated medical text. Currently, such evaluation relies solely on manual physician review. However, detecting errors in LM-generated text is challenging because 1) manual review is costly and 2) expert-composed reference outputs are often unavailable in real-world settings. While the “LLM-as-a-judge” paradigm offers scalable evaluation, even frontier LMs can miss subtle but clinically significant errors. We propose MedVAL, a novel, self-supervised, data-efficient distillation method that leverages synthetic data to train evaluator LMs to assess whether LM-generated medical outputs are factually consistent with inputs, without requiring physician labels or reference outputs. To evaluate LM performance, we introduce MedVAL-Bench, a dataset of 840 physician-annotated outputs across 6 diverse medical tasks capturing real-world challenges. Across 10 state-of-the-art LMs spanning open-source and proprietary models, MedVAL distillation significantly improves (p < 0.001) alignment with physicians across seen and unseen tasks, increasing average F1 scores from 66% to 83%. Despite strong baseline performance, MedVAL improves the best-performing proprietary LM (GPT-4o) by 8% without training on physician-labeled data, demonstrating a performance statistically non-inferior to a single human expert on a subset annotated by multiple physicians (p < 0.001). To support a scalable, risk-aware pathway towards clinical integration, we open-source: 1) Codebase (https://github.com/StanfordMIMI/MedVAL), 2) MedVAL-Bench (https://huggingface.co/datasets/stanfordmimi/MedVAL-Bench), 3) MedVAL-4B (https://huggingface.co/stanfordmimi/MedVAL-4B). Our benchmark provides evidence of LMs approaching expert-level ability in validating AI-generated medical text.

Method: Extracts Knowledge Graph from text corpus, uses LLMs to generate question-answer pairs, creates regularly updated datasets to prevent data leakage, and provides LLM-as-Judge metrics for comprehensive evaluation.

Result: Demonstrated methodology using Russian news outlets, launched public leaderboard to track RAG system development and encourage community participation.

Conclusion: DRAGOn provides a robust framework for evaluating RAG systems on dynamic corpora, addressing data leakage issues and enabling fair comparison through regularly updated benchmarks.

Abstract: This paper introduces DRAGOn, method to design a RAG benchmark on a regularly updated corpus. It features recent reference datasets, a question generation framework, an automatic evaluation pipeline, and a public leaderboard. Specified reference datasets allow for uniform comparison of RAG systems, while newly generated dataset versions mitigate data leakage and ensure that all models are evaluated on unseen, comparable data. The pipeline for automatic question generation extracts the Knowledge Graph from the text corpus and produces multiple question-answer pairs utilizing modern LLM capabilities. A set of diverse LLM-as-Judge metrics is provided for a comprehensive model evaluation. We used Russian news outlets to form the datasets and demonstrate our methodology. We launch a public leaderboard to track the development of RAG systems and encourage community participation.

Method: Systematic survey of two principal categories: 1) Linear attention methods achieving linear complexity through kernel approximations, recurrent formulations, or fastweight dynamics; 2) Sparse attention techniques limiting computation to selected token subsets using fixed patterns, block-wise routing, or clustering strategies.

Result: Comprehensive overview integrating algorithmic innovations and hardware-level considerations, analyzing incorporation of efficient attention into large-scale pre-trained language models, including both pure efficient attention architectures and hybrid designs.

Conclusion: This survey serves as a foundational reference for advancing scalable and efficient language model design by aligning theoretical foundations with practical deployment strategies for long-context modeling.

Abstract: Transformer-based architectures have become the prevailing backbone of large language models. However, the quadratic time and memory complexity of self-attention remains a fundamental obstacle to efficient long-context modeling. To address this limitation, recent research has introduced two principal categories of efficient attention mechanisms. Linear attention methods achieve linear complexity through kernel approximations, recurrent formulations, or fastweight dynamics, thereby enabling scalable inference with reduced computational overhead. Sparse attention techniques, in contrast, limit attention computation to selected subsets of tokens based on fixed patterns, block-wise routing, or clustering strategies, enhancing efficiency while preserving contextual coverage. This survey provides a systematic and comprehensive overview of these developments, integrating both algorithmic innovations and hardware-level considerations. In addition, we analyze the incorporation of efficient attention into largescale pre-trained language models, including both architectures built entirely on efficient attention and hybrid designs that combine local and global components. By aligning theoretical foundations with practical deployment strategies, this work aims to serve as a foundational reference for advancing the design of scalable and efficient language models.

Method: LegalΔ uses a reinforcement learning framework with dual-mode input (direct answer vs reasoning-augmented) to maximize information gain between modes. It follows a two-stage approach: (1) distilling reasoning capabilities from DeepSeek-R1, and (2) refining reasoning quality via differential comparisons with multidimensional rewards assessing structural coherence and legal-domain specificity.

Result: Experimental results on multiple legal reasoning tasks show LegalΔ outperforms strong baselines in both accuracy and interpretability, producing more robust and trustworthy legal judgments without relying on labeled preference data.

Conclusion: LegalΔ successfully addresses the interpretability challenge in legal AI by enhancing reasoning quality through information gain maximization, demonstrating improved performance in legal decision-making tasks.

Abstract: Legal Artificial Intelligence (LegalAI) has achieved notable advances in automating judicial decision-making with the support of Large Language Models (LLMs). However, existing legal LLMs still struggle to generate reliable and interpretable reasoning processes. They often default to fast-thinking behavior by producing direct answers without explicit multi-step reasoning, limiting their effectiveness in complex legal scenarios that demand rigorous justification. To address this challenge, we propose Legal$Δ$, a reinforcement learning framework designed to enhance legal reasoning through chain-of-thought guided information gain. During training, Legal$Δ$ employs a dual-mode input setup-comprising direct answer and reasoning-augmented modes-and maximizes the information gain between them. This encourages the model to acquire meaningful reasoning patterns rather than generating superficial or redundant explanations. Legal$Δ$ follows a two-stage approach: (1) distilling latent reasoning capabilities from a powerful Large Reasoning Model (LRM), DeepSeek-R1, and (2) refining reasoning quality via differential comparisons, combined with a multidimensional reward mechanism that assesses both structural coherence and legal-domain specificity. Experimental results on multiple legal reasoning tasks demonstrate that Legal$Δ$ outperforms strong baselines in both accuracy and interpretability. It consistently produces more robust and trustworthy legal judgments without relying on labeled preference data. All code and data will be released at https://github.com/NEUIR/LegalDelta.

Method: Created DeepScholar-bench with queries and human-written exemplars from recent high-quality ArXiv papers. Evaluates systems on generating related work sections by retrieving, synthesizing, and citing prior work. Includes automated framework measuring knowledge synthesis, retrieval quality, and verifiability across three dimensions.

Result: Benchmark is far from saturated - no system surpasses 31% geometric mean across all metrics. Evaluated prior open-source systems, search agents with strong models, OpenAI’s DeepResearch, and their own DeepScholar-ref baseline.

Conclusion: DeepScholar-bench addresses critical evaluation gap for generative research synthesis systems and provides foundation for advancing AI systems capable of automated research synthesis. The benchmark highlights both difficulty and importance of this task.

Abstract: The ability to research and synthesize knowledge is central to human expertise and progress. A new class of AI systems–designed for generative research synthesis–aims to automate this process by retrieving information from the live web and producing long-form, cited reports. Yet, evaluating such systems remains an open challenge: existing question-answering benchmarks focus on short, factual answers, while expert-curated datasets risk staleness and data contamination. Neither captures the complexity and evolving nature of real research synthesis tasks. We introduce DeepScholar-bench, a live benchmark and automated evaluation framework for generative research synthesis. DeepScholar-bench draws queries and human-written exemplars from recent, high-quality ArXiv papers and evaluates a real synthesis task: generating a related work section by retrieving, synthesizing, and citing prior work. Our automated framework holistically measures performance across three key dimensions–knowledge synthesis, retrieval quality, and verifiability. To further future work, we also contribute DeepScholar-ref, a simple, open-source reference pipeline, which is implemented on the LOTUS framework and provides a strong baseline. Using DeepScholar-bench, we systematically evaluate prior open-source systems, search agents with strong models, OpenAI’s DeepResearch, and DeepScholar-ref. We find DeepScholar-bench is far from saturated: no system surpasses a geometric mean of $31%$ across all metrics. These results highlight both the difficulty and importance of DeepScholar-bench as a foundation for advancing AI systems capable of generative research synthesis. We make our benchmark code and data available at https://github.com/guestrin-lab/deepscholar-bench.

Method: Fine-tuned LLaMA 3.1-8B on a high-quality curated corpus of energy-related texts using two adaptation strategies: 1) full-parameter Supervised Fine-Tuning, and 2) parameter-efficient LoRA-based variant updating only a small fraction of parameters. Developed complete pipeline including data collection/curation, fine-tuning, benchmark design, LLM-judge choice, evaluation, and deployment.

Result: Both EnergyGPT variants consistently outperform the base LLaMA model in most energy-related language understanding and generation tasks. The LoRA variant achieves competitive performance gains at significantly reduced training cost compared to full-parameter fine-tuning.

Conclusion: Domain specialization through fine-tuning enables improvements in domain relevance and performance without requiring large-scale infrastructure. Parameter-efficient methods like LoRA offer cost-effective alternatives for creating specialized models.

Abstract: Large language models have demonstrated impressive capabilities across various domains. However, their general-purpose nature often limits their effectiveness in specialized fields such as energy, where deep technical expertise and precise domain knowledge are essential. In this paper, we introduce EnergyGPT, a domain-specialized language model tailored for the energy sector, developed by fine-tuning the LLaMA 3.1-8B model on a high-quality, curated corpus of energy-related texts. We consider two adaptation strategies: a full-parameter Supervised Fine-Tuning variant and a parameter-efficient LoRA-based variant that updates only a small fraction of the model parameters. We present a complete development pipeline, including data collection and curation, model fine-tuning, benchmark design and LLM-judge choice, evaluation, and deployment. Through this work, we demonstrate that our training strategy enables improvements in domain relevance and performance without the need for large-scale infrastructure. By evaluating the performance of both EnergyGPT variants using domain-specific question-answering benchmarks, our results show that the adapted models consistently outperform the base model in most energy-related language understanding and generation tasks, with the LoRA variant achieving competitive gains at significantly reduced training cost.

Method: Extract activations after question reading but before token generation, train linear probes to predict forthcoming answer correctness, test across three open-source model families (7-70B parameters) on trivia questions, and evaluate generalization to diverse out-of-distribution knowledge datasets.

Result: Linear probes trained on generic trivia questions successfully predict correctness in-distribution and generalize to diverse OOD knowledge datasets, outperforming black-box baselines and verbalized confidence. Predictive power saturates in intermediate layers, but generalization fails on mathematical reasoning questions. The same direction also captures confidence for “I don’t know” responses.

Conclusion: LLMs contain internal “in-advance correctness directions” that encode confidence and correctness predictions before generation, contributing to understanding LLM internals and complementing previous work on truthfulness and behavior analysis.

Abstract: Do large language models (LLMs) anticipate when they will answer correctly? To study this, we extract activations after a question is read but before any tokens are generated, and train linear probes to predict whether the model’s forthcoming answer will be correct. Across three open-source model families ranging from 7 to 70 billion parameters, projections on this “in-advance correctness direction” trained on generic trivia questions predict success in distribution and on diverse out-of-distribution knowledge datasets, indicating a deeper signal than dataset-specific spurious features, and outperforming black-box baselines and verbalised predicted confidence. Predictive power saturates in intermediate layers and, notably, generalisation falters on questions requiring mathematical reasoning. Moreover, for models responding “I don’t know”, doing so strongly correlates with the probe score, indicating that the same direction also captures confidence. By complementing previous results on truthfulness and other behaviours obtained with probes and sparse auto-encoders, our work contributes essential findings to elucidate LLM internals.

Method: Introduces RL-ZVP algorithm that extracts learning signals from zero-variance prompts by directly rewarding correctness and penalizing errors even without contrasting responses. Modulates feedback with token-level characteristics to preserve informative, nuanced signals.

Result: Across six math reasoning benchmarks, RL-ZVP achieves significant improvements of up to 8.61 points in accuracy and 7.77 points in pass rate over GRPO, consistently outperforming other baselines that filter out zero-variance prompts.

Conclusion: Zero-variance prompts have untapped potential for learning in RLVR, and RL-ZVP effectively leverages them to improve LLM reasoning abilities beyond existing methods.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) is a powerful framework for improving the reasoning abilities of Large Language Models (LLMs). However, current methods such as GRPO rely only on problems where the model responses to the same input differ in correctness, while ignoring those where all responses receive the same reward – so-called zero-variance prompts. In this work, we argue that such prompts are not useless but can, in fact, provide meaningful feedback for policy optimization. To this end, we introduce Reinforcement Learning with Zero-Variance Prompts (RL-ZVP), a novel algorithm that extract learning signals from zero-variance prompts. RL-ZVP directly rewards correctness and penalizes errors even without contrasting responses, modulating feedback with token-level characteristics to preserve informative, nuanced signals. Across six math reasoning benchmarks, RL-ZVP achieves significant improvements of up to 8.61 points in accuracy and 7.77 points in pass rate over GRPO, while consistently outperforming other baselines that filter out zero-variance prompts. These results highlight the untapped potential of learning from zero-variance prompts in RLVR. The project page is available at https://bltnynk.github.io/publications/rl-zvp/.

Method: HEART uses emotional cues to guide model focus by alternating between critical tones (to sharpen error detection) and encouraging tones (to spark new ideas). This affective regulation framework is applied during test-time scaling to help models escape unproductive reasoning patterns.

Result: HEART was evaluated across seven high-difficulty benchmarks including Humanity’s Last Exam, GPQA Diamond, and LiveCodeBench. Results show consistent accuracy gains over affect-sterile baselines, demonstrating robustness across diverse models and that emotion facilitates deeper reasoning.

Conclusion: The strategic integration of affective regulation can significantly improve machine reasoning by guiding logical synthesis. Emotional cues help models break out of dead-end reasoning patterns, suggesting that the next frontier in machine reasoning lies in incorporating affective elements.

Abstract: Test-time scaling has significantly improved how AI models solve problems, yet current methods often get stuck in repetitive, incorrect patterns of thought. We introduce HEART, a framework that uses emotional cues to guide the model’s focus, much like how feelings contribute to human decision-making. By alternating between critical tones to sharpen error detection and encouraging tones to spark new ideas, HEART helps the model break out of dead-end reasoning and find the right solution. We evaluate HEART across seven high-difficulty benchmarks–including Humanity’s Last Exam, GPQA Diamond, and LiveCodeBench–demonstrating robustness across diverse models. Results show that emotion facilitates deeper reasoning, yielding consistent accuracy gains over affect-sterile baselines. These findings suggest that the next frontier in machine reasoning lies in the strategic integration of affective regulation to guide logical synthesis.

Method: Created RePairTQA diagnostic benchmark with controlled table representations using verbalization pipeline to generate paired structured and semi-structured tables with identical content. Conducted experiments across splits varying table size, join requirements, query complexity, and schema quality.

Result: SQL-based methods achieve high accuracy on structured inputs but degrade on semi-structured data; LLMs exhibit flexibility but reduced precision; hybrid approaches strike a balance, particularly under noisy schemas. Effects intensify with larger tables and more complex queries.

Conclusion: No single method excels across all conditions, highlighting the central role of representation in shaping Table QA performance. Findings provide insights for model selection and design, paving way for more robust hybrid approaches suited for diverse real-world data formats.

Abstract: Table Question Answering (Table QA) in real-world settings must operate over both structured databases and semi-structured tables containing textual fields. However, existing benchmarks are tied to fixed data formats and have not systematically examined how representation itself affects model performance. We present the first controlled study that isolates the role of table representation by holding content constant while varying structure. Using a verbalization pipeline, we generate paired structured and semi-structured tables, enabling direct comparisons across modeling paradigms. To support detailed analysis, we introduce RePairTQA, a diagnostic benchmark with splits along table size, join requirements, query complexity, and schema quality. Our experiments reveal consistent trade-offs: SQL-based methods achieve high accuracy on structured inputs but degrade on semi-structured data, LLMs exhibit flexibility but reduced precision, and hybrid approaches strike a balance, particularly under noisy schemas. These effects intensify with larger tables and more complex queries. Ultimately, no single method excels across all conditions, and we highlight the central role of representation in shaping Table QA performance. Our findings provide actionable insights for model selection and design, paving the way for more robust hybrid approaches suited for diverse real-world data formats.

Method: Introduces EvolveCast framework to assess LLM forecast updates in response to new information released after training cutoff. Uses human forecasters as comparative reference for forecast updates and confidence calibration under new information.

Result: LLMs show some responsiveness to new information but updates are often inconsistent or overly conservative. Both verbalized and logits-based confidence estimates remain far from human reference standard. Models tend to be conservative across various LLM settings.

Conclusion: Current approaches (e.g., RAG-based methods) for updating model knowledge are insufficient for probabilistic reasoning; models treat new information as retrieval context rather than evidence that shifts posterior probability. Need more robust mechanisms for incorporating external knowledge into belief dynamics.

Abstract: Prior work has largely treated forecasting as a static task, failing to consider how forecasts and the confidence in them should evolve as new evidence emerges. To address this gap, we introduce EvolveCast, a framework for evaluating whether large language models revise their forecasts appropriately in response to new information. In particular, EvolveCast assesses whether LLMs update their forecasts when presented with information released after their training cutoff. We use human forecasters as a comparative reference to assess forecast updates and confidence calibration under new information. While LLMs demonstrate some responsiveness to new information, their updates are often inconsistent or overly conservative. We further find that both verbalized and logits-based confidence estimates remain far from the human reference standard. Across settings with a variety of LLMs, models tend to be conservative in updating their forecasts. These findings suggest that current approaches (e.g., RAG-based methods) for updating model knowledge are insufficient for probabilistic reasoning; models treat new information as retrieval context rather than evidence that shifts posterior probability. EvolveCast thus underscores the need for more robust mechanisms to incorporate external knowledge into belief dynamics.

Method: Systematic evaluation of xLSTMs, structured state-space models, diffusion models, and adversarial learning across tagging tasks with different structural complexity, label spaces, and token dependencies, spanning multiple languages.

Result: Alternative architectures’ strong performance in simpler settings doesn’t generalize well across languages/datasets or extend to more complex structured tasks; Transformers remain dominant.

Conclusion: Alternative architectures need further development to match Transformers’ robustness across diverse sequence labeling tasks, highlighting the need for more comprehensive evaluation beyond simple benchmarks.

Abstract: Pretrained Transformer encoders are the dominant approach to sequence labeling. While some alternative architectures-such as xLSTMs, structured state-space models, diffusion models, and adversarial learning-have shown promise in language modeling, few have been applied to sequence labeling, and mostly on flat or simplified tasks. We study how these architectures adapt across tagging tasks that vary in structural complexity, label space, and token dependencies, with evaluation spanning multiple languages. We find that the strong performance previously observed in simpler settings does not always generalize well across languages or datasets, nor does it extend to more complex structured tasks.

Method: The framework casts tool use as traversal over a hidden function-dependency DAG where models must infer the correct sequence of function calls to compute a target value. It allows precise control over task difficulty through parameters like graph size, dependency depth, and distractor functions while avoiding data contamination.

Result: Reasoning-optimized models outperform general-purpose models, with GPT-5 performing best. Performance declines sharply with increased dependency depth, and connected distractors are especially challenging. Models often make syntactically valid calls but propagate incorrect values across steps, revealing brittle state tracking.

Conclusion: The study reveals fundamental limitations in LLMs’ multi-turn tool use capabilities, particularly in state tracking. A simple mitigation strategy of explicitly restating prior variable values at each step yields substantial performance gains across models, suggesting promising directions for improving tool-augmented language models.

Abstract: Existing benchmarks for tool-augmented language models (TaLMs) lack fine-grained control over task difficulty and remain vulnerable to data contamination. We present FuncBenchGen, a unified, contamination-free framework that evaluates TaLMs by generating synthetic multi-step tool-use tasks to stress-test TaLMs. The key idea is to cast tool use as traversal over a hidden function-dependency DAG where models must infer the correct sequence of calls to compute a target value. FuncBenchGen allows precise control over task difficulty (e.g., graph size, dependency depth, and distractor functions) while avoiding pretraining/test-time leakage. Our evaluation demonstrates reasoning-optimized models consistently outperform general-purpose models with GPT-5 significantly outperforming other available models. Performance declines sharply as dependency depth increases. Furthermore, connected distractors – irrelevant functions sharing type-compatible variables with relevant functions – prove especially difficult to handle. Also, strong models often make syntactically valid function calls but propagate incorrect or stale argument values across steps, revealing brittle state tracking by LLMs in multi-turn tool use. Motivated by this observation, we introduce a simple mitigation strategy that explicitly restates prior variable values to the agent at each step. Surprisingly, this lightweight change yields substantial gains across models. e.g., yielding an improvement in success rate from 62.5% to 81.3% for GPT-5.

Method: Two-stage high-copying response preference training with three prompting methods to enhance copying degree. Uses automated pipeline to transform generated responses into high-copying preference data. Proposes Context-Parameter Copying Capturing algorithm to analyze model behavior.

Result: CopyPasteLLM achieves best performance on FaithEval, ConFiQA and PubMedQA, with 12.2% to 24.5% accuracy improvements on FaithEval over best baselines, using only 365 training samples (1/50th of baseline data).

Conclusion: Training LLMs to copy more from context improves contextual faithfulness and reduces hallucinations. CopyPasteLLM recalibrates reliance from internal parametric knowledge to external context during generation.

Abstract: While Retrieval-Augmented Generation (RAG) enables large language models (LLMs) to generate contextually grounded responses, contextual faithfulness remains challenging as LLMs may not consistently trust provided context, leading to hallucinations that undermine reliability. We observe an inverse correlation between response copying degree and context-unfaithful hallucinations on RAGTruth, suggesting that higher copying degrees reduce hallucinations by fostering genuine contextual belief. We propose CopyPasteLLM, obtained through two-stage high-copying response preference training. We design three prompting methods to enhance copying degree, demonstrating that high-copying responses achieve superior contextual faithfulness and hallucination control. These approaches enable a fully automated pipeline that transforms generated responses into high-copying preference data for training CopyPasteLLM. On FaithEval, ConFiQA and PubMedQA, CopyPasteLLM achieves best performance in both counterfactual and original contexts, remarkably with 12.2% to 24.5% accuracy improvements on FaithEval over the best baseline, while requiring only 365 training samples – 1/50th of baseline data. To elucidate CopyPasteLLM’s effectiveness, we propose the Context-Parameter Copying Capturing algorithm. Interestingly, this reveals that CopyPasteLLM recalibrates reliance on internal parametric knowledge rather than external knowledge during generation. All codes are available at https://github.com/longyongchao/CopyPasteLLM

Method: Proposes a framework grounded in joint human-machine uncertainty, combining epistemic uncertainty from the model with hesitation cues and confidence signals from human speakers to guide interaction, query selection, and memory retention.

Result: Conceptual framework and call to action rather than empirical results. Advocates for a paradigm shift from static data collection to interactive, uncertainty-driven discovery.

Conclusion: Future language technology for low-resource languages must move toward interactive, cooperative model building between AI systems and speakers, emphasizing participatory, co-adaptive learning processes that respect and empower communities.

Abstract: Natural Language Processing (NLP) for low-resource languages remains fundamentally constrained by the lack of textual corpora, standardized orthographies, and scalable annotation pipelines. While recent advances in large language models have improved cross-lingual transfer, they remain inaccessible to underrepresented communities due to their reliance on massive, pre-collected data and centralized infrastructure. In this position paper, we argue for a paradigm shift toward open-ended, interactive language discovery, where AI systems learn new languages dynamically through dialogue rather than static datasets. We contend that the future of language technology, particularly for low-resource and under-documented languages, must move beyond static data collection pipelines toward interactive, uncertainty-driven discovery, where learning emerges dynamically from human-machine collaboration instead of being limited to pre-existing datasets. We propose a framework grounded in joint human-machine uncertainty, combining epistemic uncertainty from the model with hesitation cues and confidence signals from human speakers to guide interaction, query selection, and memory retention. This paper is a call to action: we advocate a rethinking of how AI engages with human knowledge in under-documented languages, moving from extractive data collection toward participatory, co-adaptive learning processes that respect and empower communities while discovering and preserving the world’s linguistic diversity. This vision aligns with principles of human-centered AI, emphasizing interactive, cooperative model building between AI systems and speakers.

Method: The authors introduce Luth models through targeted post-training on curated, high-quality French data. They further enhance performance through strategic model merging techniques that improve capabilities in both French and English.

Result: Luth models outperform all open-source counterparts of comparable size on multiple French benchmarks while retaining their original English capabilities. Strategic model merging establishes Luth as a new state-of-the-art for French SLMs.

Conclusion: Luth provides a robust baseline for future French-language research and demonstrates that targeted adaptation can effectively address language performance gaps in Small Language Models while maintaining multilingual capabilities.

Abstract: The landscape of Large Language Models remains predominantly English-centric, resulting in a significant performance gap for other major languages, such as French, especially in the context of Small Language Models (SLMs). Existing multilingual models demonstrate considerably lower performance in French compared to English, and research on efficient adaptation methods for French remains limited. To address this, we introduce \textbf{Luth}, a family of French-specialized SLMs: through targeted post-training on curated, high-quality French data, our models outperform all open-source counterparts of comparable size on multiple French benchmarks while retaining their original English capabilities. We further show that strategic model merging enhances performance in both languages, establishing Luth as a new state of the art for French SLMs and a robust baseline for future French-language research.

Method: MASA implements a multi-A, single-B structure where multiple specialized down-projection experts capture diverse features, asymmetrically shared across layers, and integrated by layer-specific up-projection matrices for parameter efficiency.

Result: On MMLU benchmark, MASA achieves 59.62% average accuracy, outperforming standard LoRA by 1.08 points (1.84% relative improvement) with comparable learnable parameters of 0.52%.

Conclusion: MASA effectively addresses LoRA’s representational bottleneck through multi-expert feature adaptation, demonstrating improved performance across multi-domain generalization, single-domain specialization, and multi-task reasoning tasks.

Abstract: Low-Rank Adaptation (LoRA) has emerged as a dominant method in Parameter-Efficient Fine-Tuning (PEFT) for large language models, which augments the transformer layer with one down-projection $A$ and one up-projection $B$. However, LoRA’s reliance on a single down-projection matrix ($A$) creates a representational bottleneck, as this solitary feature extractor is inherently insufficient for capturing the diverse signals required by complex tasks. This motivates our architectural shift to focus on enriching the feature adaptation to improve the downstream task adaptation ability. We propose MASA (Multi-$A$ Shared Adaptation), an architecture that implements a multi-$A$, single-$B$ structure where the multi-$A$ expert ensemble is asymmetrically shared across layers to ensure parameter efficiency. In MASA, these specialized experts capture diverse features, which are then integrated by a single, layer-specific $B$-matrix. The effectiveness and versatility of our method are validated through a comprehensive suite of experiments spanning multi-domain generalization, single-domain specialization, and multi-task reasoning. For example, on the MMLU benchmark, MASA achieves an average accuracy of 59.62%, outperforming the standard LoRA by 1.08 points (a relative improvement of 1.84%) with comparable learnable parameters of 0.52%.

Method: CoMAS framework enables agents to improve through inter-agent interactions without external supervision. It generates intrinsic rewards from discussion dynamics, uses LLM-as-a-judge to formulate these rewards, and optimizes each agent’s policy through reinforcement learning for decentralized co-evolution.

Result: CoMAS consistently outperforms untrained agents and achieves state-of-the-art performance across most evaluation settings. Ablation studies confirm the necessity of interaction-based reward signals and show promising scalability with increasing number and diversity of agents.

Conclusion: CoMAS establishes a novel and effective paradigm for self-evolution in LLM-based agents, demonstrating that multi-agent collaboration and discussion-based learning can drive autonomous improvement without external supervision.

Abstract: Self-evolution is a central research topic in enabling large language model (LLM)-based agents to continually improve their capabilities after pretraining. Recent research has witnessed a transition from reinforcement learning (RL)-free to RL-based methods. Current RL-based methods either rely on dense external reward signals or extract intrinsic reward signals from LLMs themselves. However, these approaches diverge from the self-evolution mechanisms observed in human intelligence, where individuals learn and improve through mutual discussion and collaboration. In this work, we introduce Co-Evolving Multi-Agent Systems (CoMAS), a novel framework that enables agents to improve autonomously by learning from inter-agent interactions without external supervision. CoMAS generates intrinsic rewards from rich discussion dynamics, employs an LLM-as-a-judge mechanism to formulate these rewards, and optimizes each agent’s policy through RL, thereby enabling decentralized and scalable co-evolution. Experimental results demonstrate that CoMAS consistently outperforms untrained agents and achieves state-of-the-art performance across most evaluation settings. Ablation studies confirm the necessity of interaction-based reward signals and reveal promising scalability as the number and diversity of agents increase. These findings establish CoMAS as a novel and effective paradigm for self-evolution in LLM-based agents.

Method: 1) Controlled knowledge fine-tuning experiments comparing dLLMs and arLLMs, 2) Proposed masked fine-tuning for arLLMs where models reconstruct original text from masked versions, 3) Applied demasking objective to supervised fine-tuning on math tasks.

Result: dLLMs achieve high QA accuracy without paraphrase augmentation, while arLLMs require paraphrases. Masked fine-tuning enables arLLMs to match dLLMs’ knowledge injection capabilities (no paraphrase needed, reversal curse resistant). Demasking objective also improves SFT on math tasks.

Conclusion: The demasking objective (used in diffusion models) provides superior knowledge injection capabilities, and this advantage can be transferred to arLLMs through masked fine-tuning, suggesting broader applicability of demasking for model training.

Abstract: Large language models (LLMs) are often used in environments where facts evolve, yet factual knowledge updates via fine-tuning on unstructured text often suffers from 1) reliance on compute-heavy paraphrase augmentation and 2) the reversal curse. Recent studies show diffusion large language models (dLLMs) require fewer training samples to achieve lower loss in pre-training and are more resistant to the reversal curse, suggesting dLLMs may learn new knowledge more easily than autoregressive LLMs (arLLMs). We test this hypothesis in controlled knowledge fine-tuning experiments and find that while arLLMs rely on paraphrase augmentation to generalize knowledge text into question-answering (QA) capability, dLLMs do not require paraphrases to achieve high QA accuracy. To further investigate whether the demasking objective alone can induce such a knowledge injection advantage in dLLMs regardless of their diffusion denoising paradigm, we propose masked fine-tuning for arLLMs, which prompts an arLLM to reconstruct the original text given a masked version in context. The masked fine-tuning for arLLMs substantially improves the efficacy of knowledge injection, i.e. no paraphrase needed and resistant to the reversal curse, closing the gap between arLLMs and dLLMs. We also demonstrate that the same demasking objective improves supervised fine-tuning (SFT) on math tasks over standard SFT, suggesting broader applicability of the demasking objective.

Method: Adds sigmoid-linear gates to Attention/MLP branches that condense outputs before re-entering residual stream. During inference, ranks tokens by gate values and skips low-importance ones using per-layer budgets, enabling token-wise layer skipping.

Result: Achieves up to 15% compute savings on long-form reasoning while retaining over 90% baseline accuracy. For larger models, tradeoff improves. On instruction-tuned models, shows accuracy gains at full compute and matches baseline quality near 50% savings.

Conclusion: GateSkip provides stable, fine-tunable token-wise layer skipping that offers computational efficiency without extensive retraining, with insights into transformer information flow and compatibility with other optimization techniques.

Abstract: We introduce GateSkip, a simple residual-stream gating mechanism that enables token-wise layer skipping in decoder-only LMs. Each Attention/MLP branch is equipped with a sigmoid-linear gate that condenses the branch’s output before it re-enters the residual stream. During inference we rank tokens by the gate values and skip low-importance ones using a per-layer budget. While early-exit or router-based Mixture-of-Depths models are known to be unstable and need extensive retraining, our smooth, differentiable gates fine-tune stably on top of pretrained models. On long-form reasoning, we save up to 15% compute while retaining over 90% of baseline accuracy. For increasingly larger models, this tradeoff improves drastically. On instruction-tuned models we see accuracy gains at full compute and match baseline quality near 50% savings. The learned gates give insight into transformer information flow (e.g., BOS tokens act as anchors), and the method combines easily with quantization, pruning, and self-speculative decoding.

Method: Proposed type-matched language distances with structure-aware representations: speaker-weighted distributions for geography, hyperbolic embeddings for genealogy, and latent variables model for typology. Unified these into a robust, task-agnostic composite distance.

Result: Across multiple zero-shot transfer benchmarks, the representations significantly improve transfer performance when distance type is relevant to the task, and the composite distance yields gains in most tasks.

Conclusion: The framework addresses limitations of existing linguistic knowledge bases by providing structure-aware representations and principled aggregation methods, leading to improved cross-lingual transfer performance.

Abstract: Existing linguistic knowledge bases such as URIEL+ provide valuable geographic, genetic and typological distances for cross-lingual transfer but suffer from two key limitations. First, their one-size-fits-all vector representations are ill-suited to the diverse structures of linguistic data. Second, they lack a principled method for aggregating these signals into a single, comprehensive score. In this paper, we address these gaps by introducing a framework for type-matched language distances. We propose novel, structure-aware representations for each distance type: speaker-weighted distributions for geography, hyperbolic embeddings for genealogy, and a latent variables model for typology. We unify these signals into a robust, task-agnostic composite distance. Across multiple zero-shot transfer benchmarks, we demonstrate that our representations significantly improve transfer performance when the distance type is relevant to the task, while our composite distance yields gains in most tasks.

Method: Systematic comparison of LRM’s reasoning in English versus the language of the question across two tasks: MGSM and GPQA Diamond. Evaluation includes both answer accuracy and analysis of cognitive attributes in reasoning traces.

Result: English reasoning traces show substantially higher presence of cognitive behaviors, and reasoning in English generally yields higher final-answer accuracy, with performance gap increasing as tasks become more complex. However, English-centric strategy is susceptible to “Lost in Translation” failure mode where translation steps lead to errors that would have been avoided by reasoning in the question’s language.

Conclusion: LRMs have significant limitations in multilingual reasoning, defaulting to English with performance trade-offs - better cognitive behaviors and accuracy in English but vulnerable to translation errors. This highlights the need for improved multilingual reasoning capabilities.

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

Method: Constructs unified paired dataset via cross-modal semantic completion and retrieval, then uses PairGRPO (pair-aware Group Relative Policy Optimization) that assigns similarity scores to modulate advantages and reduce task interference.

Result: Consistent improvements across diverse UVLM architectures (autoregressive and discrete diffusion) and scales (1B to 14B), with strong generalization to image editing tasks without editing-specific data.

Conclusion: PairUni effectively balances understanding and generation in UVLMs through structured data pairing and pair-aware RL optimization, demonstrating broad applicability and strong generalization.

Abstract: Unified Vision-Language Models (UVLMs) perform both understanding and generation within a single architecture. Since these models rely on heterogeneous data and supervision, balancing both generation and understanding in reinforcement learning (RL) is challenging. To address this challenge, we propose PairUni, a unified framework that reorganizes data into understanding-generation (UG) pairs and aligns optimization accordingly. Specifically, we construct a unified paired dataset by synthesizing aligned instances via cross-modal semantic completion and retrieving semantically related samples. These paired structures expose cross-task semantic correspondences and support consistent policy learning. To leverage this structure, we present PairGRPO, a pair-aware variant based on Group Relative Policy Optimization. It assigns a similarity score to each pair to modulate the advantage, strengthening learning from well-aligned examples and reducing task interference. Extensive experiments across diverse UVLM architectures (Autoregressive and Discrete Diffusion) and scales (1B to 14B) demonstrate that PairUni yields consistent improvements over strong baselines. Notably, our method also demonstrates strong generalization by improving performance on image editing tasks without using any editing-specific data. Codes are available at https://github.com/Haochen-Wang409/PairUni.

Method: Combines selective memory policies with feature-based salience scoring (entity density, TF-IDF, discourse markers, position bias) and learned gating mechanisms coupled with BM25 sparse retrieval for efficient information access under budget constraints.

Result: Achieves only 1.0% F1 score degradation while saving 72.4% memory compared to baseline RAG on long documents (5K-10K tokens), with benefits increasing with document length.

Conclusion: Provides a practical pathway for deploying capable long-context systems on modest hardware, democratizing access to advanced language understanding capabilities through efficient memory management.

Abstract: Large Language Models (LLMs) face significant computational and memory constraints when processing long contexts, despite growing demand for applications requiring reasoning over extensive documents, multi-session dialogues, and book length texts. While recent advances have extended context windows to 100K-1M tokens, such approaches incur prohibitive costs for resource constrained deployments. We propose BudgetMem, a novel memory augmented architecture that learns what to remember rather than remembering everything. Our system combines selective memory policies with feature based salience scoring (entity density, TF-IDF, discourse markers, position bias) to decide which information merits storage under strict budget constraints. Unlike existing retrieval augmented generation (RAG) systems that store all chunks, BudgetMem employs learned gating mechanisms coupled with BM25 sparse retrieval for efficient information access. Through comprehensive experiments on 700 question answer pairs across short (237 tokens) and long (5K-10K tokens) documents with Llama-3.2-3B-Instruct, we demonstrate that BudgetMem achieves remarkable results on long documents: only 1.0% F1 score degradation while saving 72.4% memory compared to baseline RAG. We validate our approach through budget sensitivity analysis (testing 7 budget ratios), naive baseline comparisons, and document length analysis, showing that BudgetMem’s benefits increase with document length. Our work provides a practical pathway for deploying capable long context systems on modest hardware, democratizing access to advanced language understanding capabilities.

Method: IDALC combines intent detection with active learning-based correction. It identifies user intents and corrects system-rejected utterances using a semi-supervised approach that strategically selects samples for human annotation to minimize labeling effort.

Result: Outperforms baseline methods by 5-10% higher accuracy and 4-8% improvement in macro-F1, while maintaining annotation costs at only 6-10% of available unlabeled data across benchmark datasets.

Conclusion: IDALC provides an efficient solution for intent detection and correction in dialog systems, significantly reducing annotation costs while maintaining high performance through semi-supervised active learning.

Abstract: Voice-controlled dialog systems have become immensely popular due to their ability to perform a wide range of actions in response to diverse user queries. These agents possess a predefined set of skills or intents to fulfill specific user tasks. But every system has its own limitations. There are instances where, even for known intents, if any model exhibits low confidence, it results in rejection of utterances that necessitate manual annotation. Additionally, as time progresses, there may be a need to retrain these agents with new intents from the system-rejected queries to carry out additional tasks. Labeling all these emerging intents and rejected utterances over time is impractical, thus calling for an efficient mechanism to reduce annotation costs. In this paper, we introduce IDALC (Intent Detection and Active Learning based Correction), a semi-supervised framework designed to detect user intents and rectify system-rejected utterances while minimizing the need for human annotation. Empirical findings on various benchmark datasets demonstrate that our system surpasses baseline methods, achieving a 5-10% higher accuracy and a 4-8% improvement in macro-F1. Remarkably, we maintain the overall annotation cost at just 6-10% of the unlabelled data available to the system. The overall framework of IDALC is shown in Fig. 1

Method: Fine-tune LMs using existing interpretability techniques as ground truth to generate natural language descriptions of: (1) information encoded by LM features, (2) causal structure of internal activations, and (3) influence of specific input tokens on outputs.

Result: Explainer models show non-trivial generalization to new queries with only tens of thousands of training examples. Self-explanation (model explaining itself) works better than cross-model explanation even when the explainer model is more capable.

Conclusion: LMs can learn to reliably explain their internal computations, offering a scalable complement to existing interpretability methods.

Abstract: Can language models (LMs) learn to faithfully describe their internal computations? Are they better able to describe themselves than other models? We study the extent to which LMs’ privileged access to their own internals can be leveraged to produce new techniques for explaining their behavior. Using existing interpretability techniques as a source of ground truth, we fine-tune LMs to generate natural language descriptions of (1) the information encoded by LM features, (2) the causal structure of LMs’ internal activations, and (3) the influence of specific input tokens on LM outputs. When trained with only tens of thousands of example explanations, explainer models exhibit non-trivial generalization to new queries. This generalization appears partly attributable to explainer models’ privileged access to their own internals: using a model to explain its own computations generally works better than using a different model to explain its computations (even if the explainer model is significantly more capable than the target). Our results suggest not only that LMs can learn to reliably explain their internal computations, but that such explanations offer a scalable complement to existing interpretability methods. Code and data at https://github.com/TransluceAI/introspective-interp

Method: Train models on different multilingual mixtures, analyze with Cross-Layer Transcoders (CLTs) and Attribution Graphs, perform Model-Diffing experiments, and test feature interventions.

Result: Multilingual shared representations exist with language-specific decoding in later layers; non-English failures stem from dim late-layer features, weak middle-layer clusters, and English-biased tokenizers; interventions can suppress/substitute languages.

Conclusion: Multilingual LLMs use shared representations with late-layer language-specific decoding; performance gaps arise from feature quality and tokenizer bias; finetuning improves multilingual capabilities.

Abstract: Multilingual Large Language Models (LLMs) can process many languages, yet how they internally represent this diversity remains unclear. Do they form shared multilingual representations with language-specific decoding, and if so, why does performance favor the dominant training language? To address this, we train models on different multilingual mixtures and analyze their internal mechanisms using Cross-Layer Transcoders (CLTs) and Attribution Graphs. Our results reveal multilingual shared representations: the model employs highly similar features across languages, while language-specific decoding emerges in later layers. Training models without English shows identical multilingual shared space structures. Decoding relies partly on a small set of high-frequency features in the final layers, which linearly encode language identity from early layers. Intervening on these features allows one language to be suppressed and another substituted. Finally, to explain non-English failures, we perform a Model-Diffing experiment: underperformance arises from dim late-layer features, weak middle-layer clusters, and tokenizer bias toward English that forces early layers to specialize in word reassembly. Finetuning strengthens these features and their links, improving token assembly and language-specific decoding, providing a mechanistic explanation for multilingual gaps. Our models and CLTs are available at https://huggingface.co/collections/CausalNLP/multilingual-clts and https://huggingface.co/collections/CausalNLP/multilingual-gpt2-models. Our code is available at: https://github.com/abirharrasse/MultilingualCLTs

Method: Analyzes IH formation in both natural and synthetic training data settings, examining relationships between data statistics (bigram repetition frequency, reliability) and model training parameters (batch size, context size).

Result: Found that: 1) Simple equation combining batch size and context size predicts IH formation point (agnostic to model size); 2) Bigram repetition frequency and reliability strongly affect IH formation with Pareto frontier; 3) Local dependency with high bigram stats is sufficient, but with low stats, categoricality and marginal distribution shape matter.

Conclusion: Provides systematic understanding of how training data statistics influence induction head formation, offering insights into in-context learning mechanisms and practical guidance for training data design.

Abstract: Specialized attention heads dubbed induction heads (IHs) have been argued to underlie the remarkable in-context learning capabilities of modern language models; yet, a precise characterization of their emergence, especially in the context of language modeling, remains wanting. In this study, we investigate the relationship between statistical properties of the training data and IH formation in both natural and synthetic training data settings. We show that: (1) A simple equation combining batch size and context size predicts the point at which IHs form and that this emergence point is agnostic to model size; (2) Surface bigram repetition frequency and reliability strongly affect the formation of IHs, and we find an effective Pareto frontier in terms of these two values; (3) local dependency with high bigram repetition frequency and reliability is sufficient for IH formation, but when the frequency and reliability are low, categoriality and the shape of the marginal distribution matter.

Method: Created a benchmark with 145 instances and 558 scenarios curated by 35 experts, featuring difficult product discovery queries with multiple constraints. The benchmark guarantees open-world products and enables easy verification of agent outputs.

Result: Current LLMs show stark limitations: GPT-5.2 achieved 17.76% and Gemini-3-Pro 15.82% performance. Error analysis revealed deficiencies in information grounding, multi-constraint verification, reasoning over noisy evidence, and risk-aware decision making.

Conclusion: ShoppingComp exposes critical capability gaps in LLM-powered shopping agents and characterizes the trust threshold AI systems must cross before being trusted for reliable real-world decision making.

Abstract: We present ShoppingComp, a challenging real-world benchmark for comprehensively evaluating LLM-powered shopping agents on three core capabilities: precise product retrieval, expert-level report generation, and safety critical decision making. Unlike prior e-commerce benchmarks, ShoppingComp introduces difficult product discovery queries with many constraints, while guaranteeing open-world products and enabling easy verification of agent outputs. The benchmark comprises 145 instances and 558 scenarios, curated by 35 experts to reflect authentic shopping needs. Results reveal stark limitations of current LLMs: even state-of-the-art models achieve low performance (e.g., 17.76% for GPT-5.2, 15.82% for Gemini-3-Pro).Error analysis reflects limitations in core agent competencies, including information grounding in open-world environments, reliable verification of multi-constraint requirements, consistent reasoning over noisy and conflicting evidence, and risk-aware decision making. By exposing these capability gaps, ShoppingComp characterizes the trust threshold that AI systems must cross before they can be proactively trusted for reliable real-world decision making. Our code and dataset are available at https://github.com/ByteDance-BandAI/ShoppingComp.

Method: Randomized Masked Fine-Tuning (RMFT) - a novel privacy-preserving fine-tuning technique that reduces PII memorization while minimizing performance impact, evaluated using the Enron Email Dataset and MaxTER framework.

Result: RMFT achieves 80.81% reduction in Total Extraction Rate and 80.17% reduction in Seen Extraction Rate compared to baseline, with only 5.73% increase in perplexity, outperforming deduplication methods.

Conclusion: RMFT effectively addresses privacy concerns in LLMs by significantly reducing PII memorization while maintaining model utility, providing a better privacy-utility tradeoff than existing methods.

Abstract: The current literature on memorization in Natural Language Models, especially Large Language Models (LLMs), poses severe security and privacy risks, as models tend to memorize personally identifying information (PIIs) from training data. We introduce Randomized Masked Fine-Tuning (RMFT), a novel privacy-preserving fine-tuning technique that reduces PII memorization while minimizing performance impact. Using the Enron Email Dataset, we demonstrate that RMFT achieves an 80.81% reduction in Total Extraction Rate and 80.17% reduction in Seen Extraction Rate compared to baseline fine-tuning, outperforming deduplication methods while maintaining only a 5.73% increase in perplexity. We present MaxTER, a Pareto-optimal evaluation framework for assessing privacy-utility tradeoffs, and show the performance of RMFT vs Deduplication by Area Under The Response Curve (AURC) metric.

Method: Two-stage conversion procedure: transform existing subword-based models into byte-level models with minimal additional training, enabling efficient processing of byte-level information while maintaining practical inference speeds.

Result: Bolmo models outperform prior byte-level approaches, excel on character-level reasoning tasks, remain competitive across standard benchmarks, and achieve practical inference speeds with low adaptation cost.

Conclusion: The research removes a long-standing performance barrier to end-to-end byte-level language modeling, demonstrating that byte-level models can scale competitively while offering advantages for domains requiring fine-grained textual understanding.

Abstract: Recent advances in generative AI have been largely driven by large language models (LLMs), deep neural networks that operate over discrete units called tokens. To represent text, the vast majority of LLMs use words or word fragments as the tokens, known as subword tokenization. Subword tokenization obscures fine-grained information, which is problematic, especially for scientific data - such as computer code or biological sequences - where meaning depends on the individual characters. Models that instead operate directly on the byte encoding of text avoid these limitations, but until now they have lagged behind subword-based models in performance. Here we introduce Bolmo, a family of fully open byte-level LLMs that approach the capabilities of subword-based systems. Using a two-stage conversion procedure, we transform existing subword-based models into byte-level models with minimal additional training. The resulting models outperform prior byte-level approaches and excel on character-level reasoning tasks, while remaining competitive across standard benchmarks. By efficiently processing byte-level information, these models achieve practical inference speeds and can be adapted at low cost using the existing ecosystem around the source LLM. Our results remove a long-standing performance barrier to end-to-end byte-level language modeling, demonstrating that models operating on raw text encodings can scale competitively while offering advantages in domains requiring fine-grained textual understanding.

Method: Proposed accounting reasoning concept and evaluation criteria based on GLM-series model training data analysis, then evaluated several LLMs (GLM-6B, GLM-130B, GLM-4, GPT-4) across accounting reasoning tasks with prompt engineering strategies.

Result: GPT-4 demonstrated the strongest overall accounting reasoning capability, prompt engineering improved performance across models, but current LLMs remain insufficient for real-world accounting applications.

Conclusion: Further optimization is needed for enterprise-level accounting deployment to fully realize LLMs’ potential value in the accounting domain.

Abstract: Large language models (LLMs) are increasingly reshaping learning paradigms, cognitive processes, and research methodologies across diverse domains. As their adoption expands, effectively integrating LLMs into professional fields and clarifying their role in domain-specific applications has become a key challenge for enterprise digital transformation and broader societal development. In the accounting domain, successful integration requires a systematic understanding of LLMs’ domain-specific reasoning capabilities. In this study, we introduce the concept of accounting reasoning and propose a set of evaluation criteria grounded in an analysis of the training data characteristics of representative GLM-series models. These criteria establish a foundation for studying accounting-oriented reasoning paradigms and provide benchmarks for assessing and improving model performance. Building on this framework, we evaluate several representative LLMs, including GLM-6B, GLM-130B, GLM-4, and GPT-4, across a range of accounting reasoning tasks. Our experimental results show that prompt engineering strategies can yield varying degrees of performance improvement across models, with GPT-4 demonstrating the strongest overall accounting reasoning capability. Nevertheless, the results indicate that current LLMs remain insufficient for real-world accounting applications. In particular, further optimization is required for deployment in enterprise-level accounting scenarios to fully realize the potential value of LLMs in this domain.

Method: Systematic study on IEMOCAP with controlled ablations (10 random seeds, paired tests with correction). For recognition: tested conversational context window, hierarchical sentence representations, and external affective lexicon integration. For linguistic analysis: examined 5,286 discourse-marker occurrences.

Result: 1) Conversational context is dominant: 90% of gain achieved using only most recent 10-30 turns. 2) Hierarchical representations help utterance-only recognition but benefit vanishes with context. 3) Affective lexicon integration doesn’t improve results. 4) Sad utterances show reduced left-periphery marker usage (21.9% vs 28-32% for other emotions). 5) Sadness benefits most from conversational context (+22%p).

Conclusion: Conversational history subsumes intra-utterance structure, pretrained encoders already capture affective signal, and emotion-specific discourse patterns reveal sadness relies more on discourse history than overt pragmatic signaling.

Abstract: Despite strong recent progress in Emotion Recognition in Conversation (ERC), two gaps remain: we lack clear understanding of which modeling choices materially affect performance, and we have limited linguistic analysis linking recognition findings to actionable generation cues. We address both via a systematic study on IEMOCAP. For recognition, we conduct controlled ablations with 10 random seeds and paired tests (with correction for multiple comparisons), yielding three findings. First, conversational context is dominant: performance saturates quickly, with roughly 90% of gain achieved using only the most recent 10-30 preceding turns. Second, hierarchical sentence representations improve utterance-only recognition (K=0), but the benefit vanishes once turn-level context is available, suggesting conversational history subsumes intra-utterance structure. Third, external affective lexicon (SenticNet) integration does not improve results, consistent with pretrained encoders already capturing affective signal. Under strictly causal (past-only) setting, our simple models attain strong performance (82.69% 4-way; 67.07% 6-way weighted F1). For linguistic analysis, we examine 5,286 discourse-marker occurrences and find reliable association between emotion and marker position (p < 0.0001). Sad utterances show reduced left-periphery marker usage (21.9%) relative to other emotions (28-32%), aligning with accounts linking left-periphery markers to active discourse management. This pattern is consistent with Sad benefiting most from conversational context (+22%p), suggesting sadness relies more on discourse history than overt pragmatic signaling.

Method: Introduces VotIE (Voting Information Extraction) task and establishes first benchmark using Portuguese municipal minutes from CitiLink corpus. Compares fine-tuned encoders (XLM-R-CRF) with generative approaches under both in-domain and cross-municipality evaluation settings.

Result: Fine-tuned encoders achieve 93.2% macro F1 in in-domain evaluation, outperforming generative approaches. In cross-municipality setting, fine-tuned models suffer substantial performance degradation while few-shot LLMs show greater robustness with smaller performance declines.

Conclusion: Lightweight fine-tuned encoders remain more practical for large-scale deployment despite LLMs’ generalization advantage due to computational cost constraints. Benchmark, models, and evaluation framework are publicly released for reproducible research.

Abstract: Municipal meeting minutes record key decisions in local democratic processes. Unlike parliamentary proceedings, which typically adhere to standardized formats, they encode voting outcomes in highly heterogeneous, free-form narrative text that varies widely across municipalities, posing significant challenges for automated extraction. In this paper, we introduce VotIE (Voting Information Extraction), a new information extraction task aimed at identifying structured voting events in narrative deliberative records, and establish the first benchmark for this task using Portuguese municipal minutes, building on the recently introduced CitiLink corpus. Our experiments yield two key findings. First, under standard in-domain evaluation, fine-tuned encoders, specifically XLM-R-CRF, achieve the strongest performance, reaching 93.2% macro F1, outperforming generative approaches. Second, in a cross-municipality setting that evaluates transfer to unseen administrative contexts, these models suffer substantial performance degradation, whereas few-shot LLMs demonstrate greater robustness, with significantly smaller declines in performance. Despite this generalization advantage, the high computational cost of generative models currently constrains their practicality. As a result, lightweight fine-tuned encoders remain a more practical option for large-scale, real-world deployment. To support reproducible research in administrative NLP, we publicly release our benchmark, trained models, and evaluation framework.

Method: Proposes SearchAttack with two key components: (1) rephrasing harmful semantics using dense benign knowledge to evade direct in-context decoding and elicit unsafe information retrieval, and (2) stress-testing LLMs’ reward-chasing bias by steering them to synthesize unsafe retrieved content. Also introduces a fact-checking framework to ground and quantify harm in offline/online settings.

Result: SearchAttack demonstrates strong effectiveness in attacking search-augmented LLMs. The study also finds that LLMs without web search can still be steered into harmful content output due to their information-seeking stereotypical behaviors.

Conclusion: Search-augmented LLMs have significant vulnerabilities that can be exploited to retrieve harmful content, highlighting the need for better safety mechanisms in LLM-driven search systems and responsible vulnerability assessment.

Abstract: Recently, people have suffered from LLM hallucination and have become increasingly aware of the reliability gap of LLMs in open and knowledge-intensive tasks. As a result, they have increasingly turned to search-augmented LLMs to mitigate this issue. However, LLM-driven search also becomes an attractive target for misuse. Once the returned content directly contains targeted, ready-to-use harmful instructions or takeaways for users, it becomes difficult to withdraw or undo such exposure. To investigate LLMs’ unsafe search behavior issues, we first propose \textbf{\textit{SearchAttack}} for red-teaming, which (1) rephrases harmful semantics via dense and benign knowledge to evade direct in-context decoding, thus eliciting unsafe information retrieval, (2) stress-tests LLMs’ reward-chasing bias by steering them to synthesize unsafe retrieved content. We also curate an emergent, domain-specific illicit activity benchmark for search-based threat assessment, and introduce a fact-checking framework to ground and quantify harm in both offline and online attack settings. Extensive experiments are conducted to red-team the search-augmented LLMs for responsible vulnerability assessment. Empirically, SearchAttack demonstrates strong effectiveness in attacking these systems. We also find that LLMs without web search can still be steered into harmful content output due to their information-seeking stereotypical behaviors.

Method: Literature survey organized around two key questions: 1) whether linguistic disparities arise from representation and allocation choices rather than inherent complexity, and 2) which design choices mitigate inequities across typologically diverse languages.

Result: The survey finds that performance gaps often shrink when segmentation, encoding, and data exposure are normalized, suggesting much apparent difficulty stems from current modeling choices rather than inherent linguistic complexity.

Conclusion: The paper synthesizes insights into design recommendations for tokenization, sampling, architectures, and evaluation to support more balanced multilingual language models that perform equitably across diverse languages.

Abstract: Multilingual language models (LMs) promise broader NLP access, yet current systems deliver uneven performance across the world’s languages. This survey examines why these gaps persist and whether they reflect intrinsic linguistic difficulty or modeling artifacts. We organize the literature around two questions: do linguistic disparities arise from representation and allocation choices (e.g., tokenization, encoding, data exposure, parameter sharing) rather than inherent complexity; and which design choices mitigate inequities across typologically diverse languages. We review linguistic features, such as orthography, morphology, lexical diversity, syntax, information density, and typological distance, linking each to concrete modeling mechanisms. Gaps often shrink when segmentation, encoding, and data exposure are normalized, suggesting much apparent difficulty stems from current modeling choices. We synthesize these insights into design recommendations for tokenization, sampling, architectures, and evaluation to support more balanced multilingual LMs.

Method: FRTR decomposes Excel workbooks into granular row, column, and block embeddings, uses hybrid lexical-dense retrieval with Reciprocal Rank Fusion (RRF), and integrates multimodal embeddings to reason over both numerical and visual information.

Result: Achieved 74% answer accuracy on FRTR-Bench with Claude Sonnet 4.5 (vs 24% for prior SOTA), 87% accuracy on SpreadsheetLLM benchmark with GPT-5, while reducing token usage by roughly 50% compared to direct serialization methods.

Conclusion: FRTR provides an effective multimodal retrieval-augmented generation framework for reasoning over complex enterprise spreadsheets, significantly outperforming existing approaches while being more efficient.

Abstract: Large Language Models (LLMs) struggle to reason over large-scale enterprise spreadsheets containing thousands of numeric rows, multiple linked sheets, and embedded visual content such as charts and receipts. Prior state-of-the-art spreadsheet reasoning approaches typically rely on single-sheet compression or full-context encoding, which limits scalability and fails to reflect how real users interact with complex, multimodal workbooks. We introduce FRTR-Bench, the first large-scale benchmark for multimodal spreadsheet reasoning, comprising 30 enterprise-grade Excel workbooks spanning nearly four million cells and more than 50 embedded images. To address these challenges, we present From Rows to Reasoning (FRTR), an advanced, multimodal retrieval-augmented generation framework that decomposes Excel workbooks into granular row, column, and block embeddings, employs hybrid lexical-dense retrieval with Reciprocal Rank Fusion (RRF), and integrates multimodal embeddings to reason over both numerical and visual information. We tested FRTR on six LLMs, achieving 74% answer accuracy on FRTR-Bench with Claude Sonnet 4.5, a substantial improvement over prior state-of-the-art approaches that reached only 24%. On the SpreadsheetLLM benchmark, FRTR achieved 87% accuracy with GPT-5 while reducing token usage by roughly 50% compared to direct serialization methods.

Method: Evaluated multiple LLM families (GPT, Claude, Perplexity, Grok, Gemini) on tiered payroll dataset with various prompting strategies from minimal baselines to schema-guided and reasoning variants

Result: Clear regimes identified: careful prompting sufficient in some cases, but explicit computation required for cent-accurate results in complex scenarios

Conclusion: Provides framework and practical guidance for deploying LLMs in accuracy-critical settings, highlighting limitations in numerical reasoning and auditability

Abstract: Large language models are now used daily for writing, search, and analysis, and their natural language understanding continues to improve. However, they remain unreliable on exact numerical calculation and on producing outputs that are straightforward to audit. We study synthetic payroll system as a focused, high-stakes example and evaluate whether models can understand a payroll schema, apply rules in the right order, and deliver cent-accurate results. Our experiments span a tiered dataset from basic to complex cases, a spectrum of prompts from minimal baselines to schema-guided and reasoning variants, and multiple model families including GPT, Claude, Perplexity, Grok and Gemini. Results indicate clear regimes where careful prompting is sufficient and regimes where explicit computation is required. The work offers a compact, reproducible framework and practical guidance for deploying LLMs in settings that demand both accuracy and assurance.

Method: The system uses LLMs (specifically Gemini) to extract metadata, discussed subjects, and voting outcomes from unstructured minutes. The extracted structured data is indexed in a database supporting full-text search with BM25 ranking and faceted filtering through a user-friendly interface. Built over 120 minutes from six Portuguese municipalities.

Result: Developed a functional platform tested through guided sessions with municipal personnel, providing insights into real user interactions. Evaluated Gemini’s performance in extracting relevant information from minutes, highlighting its effectiveness in data extraction.

Conclusion: CitiLink demonstrates how NLP and IR techniques can enhance accessibility and transparency of local government by transforming unstructured municipal documents into structured, searchable data, with LLMs proving effective for information extraction from bureaucratic texts.

Abstract: City council minutes are typically lengthy and formal documents with a bureaucratic writing style. Although publicly available, their structure often makes it difficult for citizens or journalists to efficiently find information. In this demo, we present CitiLink, a platform designed to transform unstructured municipal meeting minutes into structured and searchable data, demonstrating how NLP and IR can enhance the accessibility and transparency of local government. The system employs LLMs to extract metadata, discussed subjects, and voting outcomes, which are then indexed in a database to support full-text search with BM25 ranking and faceted filtering through a user-friendly interface. The developed system was built over a collection of 120 minutes made available by six Portuguese municipalities. To assess its usability, CitiLink was tested through guided sessions with municipal personnel, providing insights into how real users interact with the system. In addition, we evaluated Gemini’s performance in extracting relevant information from the minutes, highlighting its effectiveness in data extraction.

Method: Created ClaimPT dataset through partnership with LUSA (Portuguese News Agency), focusing on journalistic content. Used two trained annotators per article with curator validation. Developed new annotation scheme and provided baseline models for claim detection.

Result: Produced dataset of 1,308 European Portuguese news articles with 6,875 individual annotations. Established initial benchmarks for claim detection in Portuguese, enabling future NLP and information retrieval applications.

Conclusion: ClaimPT addresses the lack of Portuguese fact-checking resources, advances low-resource language research, and enhances understanding of misinformation in news media by providing a quality dataset with journalistic focus.

Abstract: Fact-checking remains a demanding and time-consuming task, still largely dependent on manual verification and unable to match the rapid spread of misinformation online. This is particularly important because debunking false information typically takes longer to reach consumers than the misinformation itself; accelerating corrections through automation can therefore help counter it more effectively. Although many organizations perform manual fact-checking, this approach is difficult to scale given the growing volume of digital content. These limitations have motivated interest in automating fact-checking, where identifying claims is a crucial first step. However, progress has been uneven across languages, with English dominating due to abundant annotated data. Portuguese, like other languages, still lacks accessible, licensed datasets, limiting research, NLP developments and applications. In this paper, we introduce ClaimPT, a dataset of European Portuguese news articles annotated for factual claims, comprising 1,308 articles and 6,875 individual annotations. Unlike most existing resources based on social media or parliamentary transcripts, ClaimPT focuses on journalistic content, collected through a partnership with LUSA, the Portuguese News Agency. To ensure annotation quality, two trained annotators labeled each article, with a curator validating all annotations according to a newly proposed scheme. We also provide baseline models for claim detection, establishing initial benchmarks and enabling future NLP and IR applications. By releasing ClaimPT, we aim to advance research on low-resource fact-checking and enhance understanding of misinformation in news media.

Method: Formal text-to-state mapping framework with three stages: conflict detection, interpretation extraction, and state construction. Uses hybrid extraction pipeline combining rule-based segmentation for explicit conflict markers with LLM-based enumeration of implicit ambiguity.

Result: On 68 ambiguous sentences, the framework preserves interpretive multiplicity with mean state entropy H = 1.087 bits across ambiguity categories, compared to H = 0 for collapse-based baselines. Also demonstrated cross-lingual portability for Japanese markers.

Conclusion: Provides algorithmic bridge between text and Non-Resolution Reasoning state space, enabling architectural collapse deferment in LLM inference with design principles validated on 580 test cases.

Abstract: Large language models exhibit a systematic tendency toward early semantic commitment: given ambiguous input, they collapse multiple valid interpretations into a single response before sufficient context is available. We present a formal framework for text-to-state mapping ($φ: \mathcal{T} \to \mathcal{S}$) that transforms natural language into a non-collapsing state space where multiple interpretations coexist. The mapping decomposes into three stages: conflict detection, interpretation extraction, and state construction. We instantiate $φ$ with a hybrid extraction pipeline combining rule-based segmentation for explicit conflict markers (adversative conjunctions, hedging expressions) with LLM-based enumeration of implicit ambiguity (epistemic, lexical, structural). On a test set of 68 ambiguous sentences, the resulting states preserve interpretive multiplicity: mean state entropy $H = 1.087$ bits across ambiguity categories, compared to $H = 0$ for collapse-based baselines. We additionally instantiate the rule-based conflict detector for Japanese markers to illustrate cross-lingual portability. This framework extends Non-Resolution Reasoning (NRR) by providing the missing algorithmic bridge between text and the NRR state space, enabling architectural collapse deferment in LLM inference. Design principles for state-to-state transformations are detailed in the Appendix, with empirical validation on 580 test cases showing 0% collapse for principle-satisfying operators versus up to 17.8% for violating operators.

Method: Used controlled causal interventions on both raw and condensed forms of experience. Evaluated four representative frameworks across 10 LLM backbones and 9 environments. Analyzed single- and multi-agent configurations across different backbone scales.

Result: Found a striking asymmetry: agents consistently depend on raw experience but often disregard or misinterpret condensed experience, even when it’s the only experience provided. This gap persists across configurations and scales. Identified three underlying causes: semantic limitations of condensed content, internal processing biases that suppress experience, and task regimes where pretrained priors already suffice.

Conclusion: These findings challenge prevailing assumptions about self-evolving methods and underscore the need for more faithful and reliable approaches to experience integration in LLM agents.

Abstract: Self-evolving large language model (LLM) agents continually improve by accumulating and reusing past experience, yet it remains unclear whether they faithfully rely on that experience to guide their behavior. We present the first systematic investigation of experience faithfulness, the causal dependence of an agent’s decisions on the experience it is given, in self-evolving LLM agents. Using controlled causal interventions on both raw and condensed forms of experience, we comprehensively evaluate four representative frameworks across 10 LLM backbones and 9 environments. Our analysis uncovers a striking asymmetry: while agents consistently depend on raw experience, they often disregard or misinterpret condensed experience, even when it is the only experience provided. This gap persists across single- and multi-agent configurations and across backbone scales. We trace its underlying causes to three factors: the semantic limitations of condensed content, internal processing biases that suppress experience, and task regimes where pretrained priors already suffice. These findings challenge prevailing assumptions about self-evolving methods and underscore the need for more faithful and reliable approaches to experience integration.

Method: The method analyzes denoising dynamics and finds that implicit EOS token density indicates generation sufficiency. ρ-EOS uses this density signal to bidirectionally adjust length during denoising: high density triggers mask token contraction, low density induces expansion, all within a single unified process.

Result: Experiments on mathematics and code benchmarks show ρ-EOS achieves comparable performance while substantially improving inference efficiency and token utilization compared to fixed-length approaches.

Conclusion: ρ-EOS provides an effective training-free solution for variable-length generation in masked diffusion LLMs, addressing the fundamental limitation of fixed generation lengths while maintaining performance and improving efficiency.

Abstract: Beyond parallel generation and global context modeling, current masked diffusion large language models (masked dLLMs, i.e., LLaDA) suffer from a fundamental limitation: they require a predefined, fixed generation length, which lacks flexibility and forces an inevitable trade-off between output quality and computational efficiency. To address this, we study the denoising dynamics and find that the implicit density ($ρ$) of end-of-sequence ($\texttt{EOS}$) tokens serves as a reliable signal of generation sufficiency. In particular, the evolving implicit $\texttt{EOS}$ density during denoising reveals whether the current masked space is excessive or insufficient, thereby guiding the adjustment direction for generation length. Building on this insight, we propose $\textbf{$ρ$-$\texttt{EOS}$}$, a training-free, single-stage strategy that enables bidirectional variable-length generation for masked dLLMs. Unlike prior two-stage approaches–which require separate length adjustment and iterative mask insertion phases while supporting only unidirectional expansion–$\textbf{$ρ$-$\texttt{EOS}$}$ achieves bidirectional length adjustment within a unified denoising process by continuously estimating the implicit $\texttt{EOS}$ density: excessively high density triggers $\texttt{MASK}$ token contraction, while insufficient density induces expansion. Extensive experiments on mathematics and code benchmarks demonstrate that $\textbf{$ρ$-$\texttt{EOS}$}$ achieves comparable performance while substantially improving inference efficiency and token utilization. Code is available at https://github.com/yjyddq/rho-EOS.

Method: Two-stage pipeline: 1) QA model identifies opening/closing segments containing metadata, 2) Transformer models (BERTimbau, XLM-RoBERTa ± CRF layer) perform fine-grained entity extraction with deslexicalization enhancement.

Result: Strong in-domain performance outperforming larger general-purpose LLMs, but reduced cross-municipality generalization due to variability and linguistic complexity of municipal records.

Conclusion: Establishes first benchmark for metadata extraction from municipal minutes, providing foundation for future research in this domain-specific information extraction task.

Abstract: Municipal meeting minutes are official documents of local governance, exhibiting heterogeneous formats and writing styles. Effective information retrieval (IR) requires identifying metadata such as meeting number, date, location, participants, and start/end times, elements that are rarely standardized or easy to extract automatically. Existing named entity recognition (NER) models are ill-suited to this task, as they are not adapted to such domain-specific categories. In this paper, we propose a two-stage pipeline for metadata extraction from municipal minutes. First, a question answering (QA) model identifies the opening and closing text segments containing metadata. Transformer-based models (BERTimbau and XLM-RoBERTa with and without a CRF layer) are then applied for fine-grained entity extraction and enhanced through deslexicalization. To evaluate our proposed pipeline, we benchmark both open-weight (Phi) and closed-weight (Gemini) LLMs, assessing predictive performance, inference cost, and carbon footprint. Our results demonstrate strong in-domain performance, better than larger general-purpose LLMs. However, cross-municipality evaluation reveals reduced generalization reflecting the variability and linguistic complexity of municipal records. This work establishes the first benchmark for metadata extraction from municipal meeting minutes, providing a solid foundation for future research in this domain.

Method: Constructed balanced evaluation set of 100 Turkish sentences systematically pitting local against non-local antecedents. Compared OpenAI chain-of-thought model (optimized for multi-step reasoning) with Trendyol-LLM-7B-base-v0.1 (LLaMA 2 derived model fine-tuned on Turkish data). Used combined paradigm integrating sentence-level perplexity with forced-choice comparison between minimally differing continuations.

Result: Trendyol-LLM favors local bindings in ~70% of trials, showing robust locality bias consistent with preference for structurally proximate antecedents. OpenAI model (o1 Mini) distributes choices nearly evenly between local and long-distance readings, suggesting weaker/less consistent sensitivity to locality in this binding configuration.

Conclusion: Results reveal marked contrast in binding behavior across the two systems, motivating closer analysis of how model architecture, training data, and inference-time reasoning strategies shape representation of Turkish anaphoric dependencies.

Abstract: This study evaluates whether state-of-the-art large language models capture the binding relations of Turkish reflexive pronouns. We construct a balanced evaluation set of 100 Turkish sentences that systematically pit local against non-local antecedents for the reflexives kendi and kendisi. We compare two contrasting systems: an OpenAI chain-of-thought model optimized for multi-step reasoning and Trendyol-LLM-7B-base-v0.1, a LLaMA 2 derived model extensively fine-tuned on Turkish data. Antecedent choice is assessed using a combined paradigm that integrates sentence-level perplexity with a forced-choice comparison between minimally differing continuations. Overall, Trendyol-LLM favors local bindings in approximately 70 percent of trials, exhibiting a robust locality bias consistent with a preference for structurally proximate antecedents. By contrast, the OpenAI model (o1 Mini) distributes its choices nearly evenly between local and long-distance readings, suggesting weaker or less consistent sensitivity to locality in this binding configuration. Taken together, these results reveal a marked contrast in binding behavior across the two systems and motivate closer analysis of how model architecture, training data, and inference-time reasoning strategies shape the representation of Turkish anaphoric dependencies.

Method: LAVE leverages dLLMs’ ability to predict token distributions for all positions in parallel. When a new token is proposed, it performs lookahead using these distributions to efficiently verify token validity, ensuring intermediate outputs can always be extended into valid sentences.

Result: Extensive experiments across four dLLMs and three benchmarks show LAVE consistently outperforms existing baselines, achieving substantial improvements in syntactic correctness with negligible runtime overhead.

Conclusion: LAVE provides an effective constrained decoding approach for dLLMs that reliably enforces grammatical constraints while maintaining efficiency, addressing key limitations of current methods for non-autoregressive language models.

Abstract: Diffusion Large Language Models (dLLMs) have demonstrated promising generative capabilities and are increasingly used to produce formal languages defined by context-free grammars, such as source code and chemical expressions. However, as probabilistic models, they still struggle to generate syntactically valid outputs reliably. A natural and promising direction to address this issue is to adapt constrained decoding techniques to enforce grammatical correctness during generation. However, applying these techniques faces two primary obstacles. On the one hand, the non-autoregressive nature of dLLMs renders most existing constrained decoding approaches inapplicable. On the other hand, current approaches specifically designed for dLLMs may allow intermediate outputs that are impossible to complete into valid sentences, which significantly limits their reliability in practice. To address these challenges, we present LAVE, a constrained decoding approach specifically designed for dLLMs. Our approach leverages a key property of dLLMs, namely their ability to predict token distributions for all positions in parallel during each forward pass. Whenever a new token is proposed by model, LAVE performs lookahead using these distributions to efficiently and reliably verify the validity of the proposed token. This design ensures reliable constraints by reliably preserving the potential for intermediate outputs to be extended into valid sentences. Extensive experiments across four widely used dLLMs and three representative benchmarks demonstrate that LAVE consistently outperforms existing baselines and achieves substantial improvements in syntactic correctness, while incurring negligible runtime overhead.

Method: The authors identify Reasoning Anchor (position where answer first stabilizes) and Answer-Stable Tail (repetitive verification after anchor). They propose Anchor-based Process Reward (APR), a structure-aware reward shaping method that localizes reasoning anchor and penalizes only the post-anchor AST using policy optimization algorithms suitable for length penalties.

Result: APR models achieved performance-efficiency Pareto frontier at 1.5B and 7B scales across five mathematical reasoning datasets while requiring substantially fewer computational resources for RL training compared to baseline methods.

Conclusion: The paper provides a fine-grained understanding of Overthinking in Large Reasoning Models, identifies structural redundancy (Answer-Stable Tail), and proposes an efficient solution (Anchor-based Process Reward) that eliminates wasteful computation while maintaining model performance.

Abstract: Test-Time Scaling (TTS) has significantly enhanced the capabilities of Large Reasoning Models (LRMs) but introduces a critical side-effect known as Overthinking. We conduct a preliminary study to rethink this phenomenon from a fine-grained perspective. We observe that LRMs frequently conduct repetitive self-verification without revision even after obtaining the final answer during the reasoning process. We formally define this specific position where the answer first stabilizes as the Reasoning Anchor. By analyzing pre- and post-anchor reasoning behaviors, we uncover the structural redundancy fixed in LRMs: the meaningless repetitive verification after deriving the first complete answer, which we term the Answer-Stable Tail (AST). Motivated by this observation, we propose Anchor-based Process Reward (APR), a structure-aware reward shaping method that localizes the reasoning anchor and penalizes exclusively the post-anchor AST. Leveraging the policy optimization algorithm suitable for length penalties, our APR models achieved the performance-efficiency Pareto frontier at 1.5B and 7B scales averaged across five mathematical reasoning datasets while requiring substantially fewer computational resources for RL training.

Method: Proposes Token Priority as a unifying framework, formalizing Supervised Fine-Tuning (SFT) as a distribution reshaping process rather than simple optimization. Categorizes existing approaches into two regimes: Positive Priority (noise filtration) and Signed Priority (toxic modes unlearning).

Result: Provides a unified lens to analyze recent breakthroughs in alignment research, revisits existing progress and limitations, identifies key challenges, and suggests future research directions.

Conclusion: Token Priority is essential for bridging the granularity mismatch in AI alignment, offering a framework to systematically reshape distributions to align with ideal alignment manifolds rather than just optimizing for coarse signals.

Abstract: The transition from fitting empirical data to achieving true human utility is fundamentally constrained by a granularity mismatch, where fine-grained autoregressive generation is often supervised by coarse or uniform signals. This position paper advocates Token Priority as the essential bridge, formalizing Supervised Fine-Tuning (SFT) not as simple optimization but as a precise distribution reshaping process that aligns raw data with the ideal alignment manifold. We analyze recent breakthroughs through this unified lens, categorizing them into two distinct regimes: Positive Priority for noise filtration and Signed Priority for toxic modes unlearning. We revisit existing progress and limitations, identify key challenges, and suggest directions for future research.

Method: Position paper analyzing HLS’s role in agentic hardware design through three contributions: explaining HLS as practical abstraction layer, identifying current HLS tool limitations that agents can address, and proposing taxonomy for symbiotic evolution of agentic HLS.

Result: Establishes HLS as crucial layer for agentic optimization with faster iteration cycles, portability, and design permutability, while identifying key limitations in current HLS tools that AI agents are uniquely positioned to overcome.

Conclusion: HLS remains essential in the agentic era, serving as golden reference for AI-driven hardware design, with proposed taxonomy showing how responsibility shifts from humans to AI agents as systems evolve from copilots to autonomous design partners.

Abstract: The rise of large language models has sparked interest in AI-driven hardware design, raising the question: does high-level synthesis (HLS) still matter in the agentic era? We argue that HLS remains essential. While we expect mature agentic hardware systems to leverage both HLS and RTL, this paper focuses on HLS and its role in enabling agentic optimization. HLS offers faster iteration cycles, portability, and design permutability that make it a natural layer for agentic optimization. This position paper makes three contributions. First, we explain why HLS serves as a practical abstraction layer and a golden reference for agentic hardware design. Second, we identify key limitations of current HLS tools, namely inadequate performance feedback, rigid interfaces, and limited debuggability that agents are uniquely positioned to address. Third, we propose a taxonomy for the symbiotic evolution of agentic HLS, clarifying how responsibility shifts from human designers to AI agents as systems advance from copilots to autonomous design partners.

Method: Uses chunk-recurrent training framework with on-the-fly activation recomputation for constant activation memory, plus paged memory manager for KV cache, asynchronous CPU offloading, and page-level sparse attention to manage KV cache growth.

Result: Achieves only 10MB memory increase per 10K additional tokens for Qwen2.5-7B, enabling 4M-token context training on a single H200 GPU instead of requiring large clusters.

Conclusion: OOMB represents a substantial advance in resource efficiency for long-context LLM training, making long-context training accessible on single GPUs.

Abstract: Training Large Language Models (LLMs) on long contexts is severely constrained by prohibitive GPU memory overhead, not training time. The primary culprits are the activations, whose memory footprints scale linearly with sequence length. We introduce OOMB, a highly memory-efficient training system that directly confronts this barrier. Our approach employs a chunk-recurrent training framework with on-the-fly activation recomputation, which maintains a constant activation memory footprint (O(1)) and shifts the primary bottleneck to the growing KV cache. To manage the KV cache, OOMB integrates a suite of synergistic optimizations: a paged memory manager for both the KV cache and its gradients to eliminate fragmentation, asynchronous CPU offloading to hide data transfer latency, and page-level sparse attention to reduce both computational complexity and communication overhead. The synergy of these techniques yields exceptional efficiency. Our empirical results show that for every additional 10K tokens of context, the end-to-end training memory overhead increases by a mere 10MB for Qwen2.5-7B. This allows training Qwen2.5-7B with a 4M-token context on a single H200 GPU, a feat that would otherwise require a large cluster using context parallelism. This work represents a substantial advance in resource efficiency for long-context LLM training. The source code is available at https://github.com/wenhaoli-xmu/OOMB.

Method: The authors designed and offered a new upper-level undergraduate course cross-listed between Linguistics and Computer Science. The course integrates theoretical and empirical aspects of discourse processing with natural language generation, using collaborative design by experts with complementary expertise.

Result: The course was successfully offered in Fall 2025. An independent survey provided positive feedback and takeaways. The course design philosophy emphasizes deep integration of theory and practice, creating an exploratory mindset in classroom activities and assignments.

Conclusion: The paper presents a successful model for integrating discourse processing with NLP education, addressing curriculum gaps in a rapidly evolving field. The authors share their vision for future directions in NLP education that bridge theoretical linguistics with computational approaches.

Abstract: The field of NLP has undergone vast, continuous transformations over the past few years, sparking debates going beyond discipline boundaries. This begs important questions in education: how do we design courses that bridge sub-disciplines in this shifting landscape? This paper explores this question from the angle of discourse processing, an area with rich linguistic insights and computational models for the intentional, attentional, and coherence structure of language. Discourse is highly relevant for open-ended or long-form text generation, yet this connection is under-explored in existing undergraduate curricula. We present a new course, “Computational Discourse and Natural Language Generation”. The course is collaboratively designed by a team with complementary expertise and was offered for the first time in Fall 2025 as an upper-level undergraduate course, cross-listed between Linguistics and Computer Science. Our philosophy is to deeply integrate the theoretical and empirical aspects, and create an exploratory mindset inside the classroom and in the assignments. This paper describes the course in detail and concludes with takeaways from an independent survey as well as our vision for future directions.

Method: Uses frame semantics instead of distributional representations, relying on semantic frames to detect changes in word meanings over time

Result: The frame semantics method is effective for detecting semantic change and can outperform many distributional semantic models

Conclusion: Frame semantics provides a viable alternative to neural embeddings for LSC detection, offering both strong performance and high interpretability

Abstract: The majority of contemporary computational methods for lexical semantic change (LSC) detection are based on neural embedding distributional representations. Although these models perform well on LSC benchmarks, their results are often difficult to interpret. We explore an alternative approach that relies solely on frame semantics. We show that this method is effective for detecting semantic change and can even outperform many distributional semantic models. Finally, we present a detailed quantitative and qualitative analysis of its predictions, demonstrating that they are both plausible and highly interpretable

Method: Proposes BioACE framework with automated approaches to evaluate answer completeness, correctness, precision, and recall against ground-truth nuggets. Also evaluates citation quality using NLI, pre-trained language models, and LLMs.

Result: Extensive experiments show correlation with human evaluations. Provides best approaches for biomedical answer and citation evaluation as part of open-source BioACE package.

Conclusion: BioACE offers an effective automated framework for evaluating biomedical answers and citations from LLMs, addressing the challenge of expert assessment in biomedical NLP tasks.

Abstract: With the increasing use of large language models (LLMs) for generating answers to biomedical questions, it is crucial to evaluate the quality of the generated answers and the references provided to support the facts in the generated answers. Evaluation of text generated by LLMs remains a challenge for question answering, retrieval-augmented generation (RAG), summarization, and many other natural language processing tasks in the biomedical domain, due to the requirements of expert assessment to verify consistency with the scientific literature and complex medical terminology. In this work, we propose BioACE, an automated framework for evaluating biomedical answers and citations against the facts stated in the answers. The proposed BioACE framework considers multiple aspects, including completeness, correctness, precision, and recall, in relation to the ground-truth nuggets for answer evaluation. We developed automated approaches to evaluate each of the aforementioned aspects and performed extensive experiments to assess and analyze their correlation with human evaluations. In addition, we considered multiple existing approaches, such as natural language inference (NLI) and pre-trained language models and LLMs, to evaluate the quality of evidence provided to support the generated answers in the form of citations into biomedical literature. With the detailed experiments and analysis, we provide the best approaches for biomedical answer and citation evaluation as a part of BioACE (https://github.com/deepaknlp/BioACE) evaluation package.

Method: OPUS defines utility in the optimizer-induced update space, scoring candidates by projecting their effective updates (shaped by modern optimizers) onto a target direction from a stable, in-distribution proxy. Uses Ghost technique with CountSketch for computational efficiency and Boltzmann sampling for data diversity.

Result: Achieves remarkable results across diverse corpora, quality tiers, optimizers, and model scales. In GPT-2 Large/XL pre-training on FineWeb with 30B tokens, outperforms industrial baselines and even full 200B-token training. With Qwen3-8B-Base on SciencePedia, achieves superior performance using only 0.5B tokens vs. full training with 3B tokens.

Conclusion: OPUS provides an efficient dynamic data selection framework that significantly improves pre-training efficiency and data utilization, especially valuable in specialized domains and as high-quality data becomes scarce.

Abstract: As high-quality public text approaches exhaustion, a phenomenon known as the Data Wall, pre-training is shifting from more tokens to better tokens. However, existing methods either rely on heuristic static filters that ignore training dynamics, or use dynamic yet optimizer-agnostic criteria based on raw gradients. We propose OPUS (Optimizer-induced Projected Utility Selection), a dynamic data selection framework that defines utility in the optimizer-induced update space. OPUS scores candidates by projecting their effective updates, shaped by modern optimizers, onto a target direction derived from a stable, in-distribution proxy. To ensure scalability, we employ Ghost technique with CountSketch for computational efficiency, and Boltzmann sampling for data diversity, incurring only 4.7% additional compute overhead. OPUS achieves remarkable results across diverse corpora, quality tiers, optimizers, and model scales. In pre-training of GPT-2 Large/XL on FineWeb and FineWeb-Edu with 30B tokens, OPUS outperforms industrial-level baselines and even full 200B-token training. Moreover, when combined with industrial-level static filters, OPUS further improves pre-training efficiency, even with lower-quality data. Furthermore, in continued pre-training of Qwen3-8B-Base on SciencePedia, OPUS achieves superior performance using only 0.5B tokens compared to full training with 3B tokens, demonstrating significant data efficiency gains in specialized domains.

Method: Reinforcement World Model Learning (RWML) - a self-supervised method that learns action-conditioned world models on textual states using sim-to-real gap rewards. It aligns simulated next states with realized next states in a pre-trained embedding space, providing more robust training than next-state token prediction.

Result: Significant gains over base model on ALFWorld and τ² Bench. When combined with task-success rewards, outperforms direct task-success reward RL by 6.9 and 5.7 points respectively, while matching expert-data training performance.

Conclusion: RWML provides an effective self-supervised approach for learning world models in LLM-based agents, improving their ability to anticipate action consequences and adapt to environment dynamics in text-based environments.

Abstract: Large language models (LLMs) have achieved strong performance in language-centric tasks. However, in agentic settings, LLMs often struggle to anticipate action consequences and adapt to environment dynamics, highlighting the need for world-modeling capabilities in LLM-based agents. We propose Reinforcement World Model Learning (RWML), a self-supervised method that learns action-conditioned world models for LLM-based agents on textual states using sim-to-real gap rewards. Our method aligns simulated next states produced by the model with realized next states observed from the environment, encouraging consistency between internal world simulations and actual environment dynamics in a pre-trained embedding space. Unlike next-state token prediction, which prioritizes token-level fidelity (i.e., reproducing exact wording) over semantic equivalence and can lead to model collapse, our method provides a more robust training signal and is empirically less susceptible to reward hacking than LLM-as-a-judge. We evaluate our method on ALFWorld and $τ^2$ Bench and observe significant gains over the base model, despite being entirely self-supervised. When combined with task-success rewards, our method outperforms direct task-success reward RL by 6.9 and 5.7 points on ALFWorld and $τ^2$ Bench respectively, while matching the performance of expert-data training.

Method: KV-CoRE (KV-cache Compressibility by Rank Evaluation) uses Singular Value Decomposition (SVD) to compute optimal low-rank approximations under the Frobenius norm. It’s gradient-free and incremental, enabling efficient dataset-level, layer-wise evaluation. The method employs Normalized Effective Rank as a compressibility metric.

Result: Analysis across multiple models and datasets spanning five English domains and sixteen languages reveals systematic patterns linking compressibility to model architecture, training data, and language coverage. Normalized Effective Rank strongly correlates with performance degradation under compression.

Conclusion: The study establishes a principled evaluation framework and the first large-scale benchmark for kv-cache compressibility in LLMs, offering insights for dynamic, data-aware compression strategies and data-centric model development.

Abstract: Large language models rely on kv-caches to avoid redundant computation during autoregressive decoding, but as context length grows, reading and writing the cache can quickly saturate GPU memory bandwidth. Recent work has explored KV-cache compression, yet most approaches neglect the data-dependent nature of kv-caches and their variation across layers. We introduce KV-CoRE KV-cache Compressibility by Rank Evaluation), an SVD-based method for quantifying the data-dependent low-rank compressibility of kv-caches. KV-CoRE computes the optimal low-rank approximation under the Frobenius norm and, being gradient-free and incremental, enables efficient dataset-level, layer-wise evaluation. Using this method, we analyze multiple models and datasets spanning five English domains and sixteen languages, uncovering systematic patterns that link compressibility to model architecture, training data, and language coverage. As part of this analysis, we employ the Normalized Effective Rank as a metric of compressibility and show that it correlates strongly with performance degradation under compression. Our study establishes a principled evaluation framework and the first large-scale benchmark of kv-cache compressibility in LLMs, offering insights for dynamic, data-aware compression and data-centric model development.

Method: Provides a unified framework for foundation agent memory across three dimensions: memory substrate (internal/external), cognitive mechanism (episodic, semantic, sensory, working, procedural), and memory subject (agent/user-centric). Analyzes memory instantiation under different agent topologies and learning policies for memory operations.

Result: Comprehensive survey organizing the rapidly growing field of agent memory (hundreds of papers this year), establishing a taxonomy for memory systems, and identifying key evaluation benchmarks and metrics for assessing memory utility.

Conclusion: Memory is essential for AI agents to achieve real-world utility in complex environments. The survey provides a structured framework for understanding memory systems, highlights current evaluation approaches, and outlines open challenges and future research directions.

Abstract: The research of artificial intelligence is undergoing a paradigm shift from prioritizing model innovations over benchmark scores towards emphasizing problem definition and rigorous real-world evaluation. As the field enters the “second half,” the central challenge becomes real utility in long-horizon, dynamic, and user-dependent environments, where agents face context explosion and must continuously accumulate, manage, and selectively reuse large volumes of information across extended interactions. Memory, with hundreds of papers released this year, therefore emerges as the critical solution to fill the utility gap. In this survey, we provide a unified view of foundation agent memory along three dimensions: memory substrate (internal and external), cognitive mechanism (episodic, semantic, sensory, working, and procedural), and memory subject (agent- and user-centric). We then analyze how memory is instantiated and operated under different agent topologies and highlight learning policies over memory operations. Finally, we review evaluation benchmarks and metrics for assessing memory utility, and outline various open challenges and future directions.

Method: DiSPO branches at intermediate masked states by resampling fillings from cached logits, scores completions, and updates only newly filled tokens without extra diffusion rollouts, using a policy-gradient estimator combined with terminal-feedback optimization.

Result: On LLaDA-8B-Instruct, DiSPO consistently outperforms terminal-feedback diffu-GRPO baseline on math and planning benchmarks under matched compute and optimizer steps.

Conclusion: DiSPO provides effective credit assignment for diffusion language models, improving performance on reasoning tasks without additional rollout overhead.

Abstract: Masked diffusion language models generate by iteratively filling masked tokens over multiple denoising steps, so learning only from a terminal reward on the final completion yields coarse credit assignment over intermediate decisions. We propose DiSPO (Diffusion-State Policy Optimization), a plug-in credit-assignment layer that directly optimizes intermediate filling decisions. At selected intermediate masked states, DiSPO branches by resampling fillings for the currently masked positions from rollout-cached logits, scores the resulting completions, and updates only the newly filled tokens – without additional multi-step diffusion rollouts. We formalize a fixed-state objective for branched completions and derive a policy-gradient estimator that can be combined with terminal-feedback policy optimization using the same rollouts. On LLaDA-8B-Instruct, DiSPO consistently improves over the terminal-feedback diffu-GRPO baseline on math and planning benchmarks under matched rollout compute and optimizer steps. Our code will be available at https://daioba.github.io/dispo .

Method: InftyThink+ uses end-to-end reinforcement learning with model-controlled iteration boundaries and explicit summarization. It employs a two-stage training scheme: supervised cold-start followed by trajectory-level RL to learn strategic summarization and continuation decisions.

Result: InftyThink+ improves accuracy by 21% on AIME24, outperforms conventional long chain-of-thought RL, generalizes better to out-of-distribution benchmarks, reduces inference latency, and accelerates RL training.

Conclusion: The framework demonstrates improved reasoning efficiency alongside stronger performance by optimizing the entire iterative reasoning trajectory through learned summarization strategies.

Abstract: Large reasoning models achieve strong performance by scaling inference-time chain-of-thought, but this paradigm suffers from quadratic cost, context length limits, and degraded reasoning due to lost-in-the-middle effects. Iterative reasoning mitigates these issues by periodically summarizing intermediate thoughts, yet existing methods rely on supervised learning or fixed heuristics and fail to optimize when to summarize, what to preserve, and how to resume reasoning. We propose InftyThink+, an end-to-end reinforcement learning framework that optimizes the entire iterative reasoning trajectory, building on model-controlled iteration boundaries and explicit summarization. InftyThink+ adopts a two-stage training scheme with supervised cold-start followed by trajectory-level reinforcement learning, enabling the model to learn strategic summarization and continuation decisions. Experiments on DeepSeek-R1-Distill-Qwen-1.5B show that InftyThink+ improves accuracy by 21% on AIME24 and outperforms conventional long chain-of-thought reinforcement learning by a clear margin, while also generalizing better to out-of-distribution benchmarks. Moreover, InftyThink+ significantly reduces inference latency and accelerates reinforcement learning training, demonstrating improved reasoning efficiency alongside stronger performance.

Method: Developed a hierarchical Bayesian model using latent Gaussian modeling for efficient inference on large datasets via integrated nested Laplace approximations, incorporating spatially varying coefficients to model relationships between shoe tread patterns and accidental locations.

Result: Demonstrated superior performance on held-out data compared to existing methods, enhancing accuracy and reliability in forensic shoe print analysis.

Conclusion: The proposed hierarchical Bayesian model provides a more accurate and scalable approach for quantifying accidental pattern rarity in shoe prints, improving forensic evidence assessment in criminal investigations.

Abstract: Shoe print evidence recovered from crime scenes plays a key role in forensic investigations. By examining shoe prints, investigators can determine details of the footwear worn by suspects. However, establishing that a suspect’s shoes match the make and model of a crime scene print may not be sufficient. Typically, thousands of shoes of the same size, make, and model are manufactured, any of which could be responsible for the print. Accordingly, a popular approach used by investigators is to examine the print for signs of ``accidentals,’’ i.e., cuts, scrapes, and other features that accumulate on shoe soles after purchase due to wear. While some patterns of accidentals are common on certain types of shoes, others are highly distinctive, potentially distinguishing the suspect’s shoe from all others. Quantifying the rarity of a pattern is thus essential to accurately measuring the strength of forensic evidence. In this study, we address this task by developing a hierarchical Bayesian model. Our improvement over existing methods primarily stems from two advancements. First, we frame our approach in terms of a latent Gaussian model, thus enabling inference to be efficiently scaled to large collections of annotated shoe prints via integrated nested Laplace approximations. Second, we incorporate spatially varying coefficients to model the relationship between shoes’ tread patterns and accidental locations. We demonstrate these improvements through superior performance on held-out data, which enhances accuracy and reliability in forensic shoe print analysis.

Method: 1) Fixed-frame Modality Gap Theory decomposes gap into stable biases and anisotropic residuals. 2) ReAlign: training-free alignment strategy using massive unpaired data statistics with Anchor, Trace, and Centroid Alignment steps. 3) ReVision: scalable MLLM training paradigm integrating ReAlign into pretraining to learn visual representation distribution from unpaired text before visual instruction tuning.

Result: Framework demonstrates statistically aligned unpaired data can effectively substitute for expensive image-text pairs, offering robust path for efficient scaling of Multimodal Large Language Models.

Conclusion: Precise geometric characterization of modality gap enables efficient model scaling through statistical alignment of unpaired data, reducing reliance on large-scale, high-quality image-text pairs for MLLM development.

Abstract: Despite the success of multimodal contrastive learning in aligning visual and linguistic representations, a persistent geometric anomaly, the Modality Gap, remains: embeddings of distinct modalities expressing identical semantics occupy systematically offset regions. Prior approaches to bridge this gap are largely limited by oversimplified isotropic assumptions, hindering their application in large-scale scenarios. In this paper, we address these limitations by precisely characterizing the geometric shape of the modality gap and leveraging it for efficient model scaling. First, we propose the Fixed-frame Modality Gap Theory, which decomposes the modality gap within a frozen reference frame into stable biases and anisotropic residuals. Guided by this precise modeling, we introduce ReAlign, a training-free modality alignment strategy. Utilizing statistics from massive unpaired data, ReAlign aligns text representation into the image representation distribution via a three-step process comprising Anchor, Trace, and Centroid Alignment, thereby explicitly rectifying geometric misalignment. Building on ReAlign, we propose ReVision, a scalable training paradigm for Multimodal Large Language Models (MLLMs). ReVision integrates ReAlign into the pretraining stage, enabling the model to learn the distribution of visual representations from unpaired text before visual instruction tuning, without the need for large-scale, high-quality image-text pairs. Our framework demonstrates that statistically aligned unpaired data can effectively substitute for expensive image-text pairs, offering a robust path for the efficient scaling of MLLMs.

Method: Proposes attribution-based human prior alignment that encodes human priors as expected input regions (e.g., bounding boxes), uses subset-selection-based attribution to expose model decision evidence, and penalizes deviations from prior regions through a training objective with attribution constraints.

Result: Validated on image classification and click decision tasks in MLLM-based GUI agent models, showing consistent improvements in task accuracy and enhanced decision reasonability across conventional classification and autoregressive generation settings.

Conclusion: Human prior alignment through attribution constraints effectively improves model reliability by encouraging models to base decisions on intended evidence rather than shortcut correlations.

Abstract: Reliable models should not only predict correctly, but also justify decisions with acceptable evidence. Yet conventional supervised learning typically provides only class-level labels, allowing models to achieve high accuracy through shortcut correlations rather than the intended evidence. Human priors can help constrain such behavior, but aligning models to these priors remains challenging because learned representations often diverge from human perception. To address this challenge, we propose an attribution-based human prior alignment method. We encode human priors as input regions that the model is expected to rely on (e.g., bounding boxes), and leverage a highly faithful subset-selection-based attribution approach to expose the model’s decision evidence during training. When the attribution region deviates substantially from the prior regions, we penalize reliance on off-prior evidence, encouraging the model to shift its attribution toward the intended regions. This is achieved through a training objective that imposes attribution constraints induced by the human prior. We validate our method on both image classification and click decision tasks in MLLM-based GUI agent models. Across conventional classification and autoregressive generation settings, human prior alignment consistently improves task accuracy while also enhancing the model’s decision reasonability.

Method: Introduces MAU-Set dataset spanning multiple industrial domains with hierarchical tasks, and MAU-GPT multimodal large model with novel AMoE-LoRA mechanism unifying anomaly-aware and generalist experts adaptation

Result: MAU-GPT consistently outperforms prior state-of-the-art methods across all domains, demonstrating strong potential for scalable automated industrial inspection

Conclusion: The proposed dataset and model provide comprehensive solution for industrial anomaly understanding with enhanced multimodal reasoning capabilities

Abstract: As industrial manufacturing scales, automating fine-grained product image analysis has become critical for quality control. However, existing approaches are hindered by limited dataset coverage and poor model generalization across diverse and complex anomaly patterns. To address these challenges, we introduce MAU-Set, a comprehensive dataset for Multi-type industrial Anomaly Understanding. It spans multiple industrial domains and features a hierarchical task structure, ranging from binary classification to complex reasoning. Alongside this dataset, we establish a rigorous evaluation protocol to facilitate fair and comprehensive model assessment. Building upon this foundation, we further present MAU-GPT, a domain-adapted multimodal large model specifically designed for industrial anomaly understanding. It incorporates a novel AMoE-LoRA mechanism that unifies anomaly-aware and generalist experts adaptation, enhancing both understanding and reasoning across diverse defect classes. Extensive experiments show that MAU-GPT consistently outperforms prior state-of-the-art methods across all domains, demonstrating strong potential for scalable and automated industrial inspection.

Method: Multi-stage training on over 200,000 fundus images supporting three task categories: anatomical structures, retinal phenotypic patterns, and lesion types, with conversion to over 30 standardized biomarkers.

Result: Achieves superior segmentation performance on 17 public datasets, improving prior methods by 3.9 percentage points in DSC on average, with up to 15 percentage points on challenging multi-task benchmarks.

Conclusion: RetSAM transforms fundus images into standardized quantitative phenotypes, enabling systematic correlation analyses across major ophthalmic diseases and large-scale ophthalmic research.

Abstract: Retinal imaging is fast, non-invasive, and widely available, offering quantifiable structural and vascular signals for ophthalmic and systemic health assessment. This accessibility creates an opportunity to study how quantitative retinal phenotypes relate to ocular and systemic diseases. However, such analyses remain difficult at scale due to the limited availability of public multi-label datasets and the lack of a unified segmentation-to-quantification pipeline. We present RetSAM, a general retinal segmentation and quantification framework for fundus imaging. It delivers robust multi-target segmentation and standardized biomarker extraction, supporting downstream ophthalmologic studies and oculomics correlation analyses. Trained on over 200,000 fundus images, RetSAM supports three task categories and segments five anatomical structures, four retinal phenotypic patterns, and more than 20 distinct lesion types. It converts these segmentation results into over 30 standardized biomarkers that capture structural morphology, vascular geometry, and degenerative changes. Trained with a multi-stage strategy using both private and public fundus data, RetSAM achieves superior segmentation performance on 17 public datasets. It improves on prior best methods by 3.9 percentage points in DSC on average, with up to 15 percentage points on challenging multi-task benchmarks, and generalizes well across diverse populations, imaging devices, and clinical settings. The resulting biomarkers enable systematic correlation analyses across major ophthalmic diseases, including diabetic retinopathy, age-related macular degeneration, glaucoma, and pathologic myopia. Together, RetSAM transforms fundus images into standardized, interpretable quantitative phenotypes, enabling large-scale ophthalmic research and translation.

Method: Three integrated components: (1) extracting configurable refusal vectors via teacher-forced mechanism to amplify refusal signals; (2) gating mechanism to mitigate over-refusal by preserving acceptance for in-scope queries; (3) counterfactual vision enhancement module aligning visual representations with refusal requirements.

Result: Comprehensive experiments across multiple datasets and various VLMs demonstrate CR-VLM achieves effective, efficient, and robust configurable refusals, offering scalable user-adaptive safety alignment.

Conclusion: CR-VLM provides a robust and efficient approach for configurable refusal in VLMs, enabling better balance between safety and utility through adaptive refusal mechanisms.

Abstract: With the rapid advancement of Vision Language Models (VLMs), refusal mechanisms have become a critical component for ensuring responsible and safe model behavior. However, existing refusal strategies are largely \textit{one-size-fits-all} and fail to adapt to diverse user needs and contextual constraints, leading to either under-refusal or over-refusal. In this work, we firstly explore the challenges mentioned above and develop \textbf{C}onfigurable \textbf{R}efusal in \textbf{VLM}s (\textbf{CR-VLM}), a robust and efficient approach for {\em configurable} refusal based on activation steering. CR-VLM consists of three integrated components: (1) extracting a configurable refusal vector via a teacher-forced mechanism to amplify the refusal signal; (2) introducing a gating mechanism that mitigates over-refusal by preserving acceptance for in-scope queries; and (3) designing a counterfactual vision enhancement module that aligns visual representations with refusal requirements. Comprehensive experiments across multiple datasets and various VLMs demonstrate that CR-VLM achieves effective, efficient, and robust configurable refusals, offering a scalable path toward user-adaptive safety alignment in VLMs.

Method: 1) Uses multiple pre-trained visual encoders with distinct inductive biases to capture hierarchical and multi-scale visual representations; 2) Employs contrastive learning objective for effective alignment between brain signals and visual embeddings; 3) Introduces Fusion Prior that learns stable mapping on large-scale visual data and matches brain features to this pre-trained prior for better distributional consistency across modalities.

Result: Extensive quantitative and qualitative experiments demonstrate the method achieves favorable balance between retrieval accuracy and reconstruction fidelity.

Conclusion: The proposed brain-image alignment strategy with multiple visual encoders and Fusion Prior effectively decodes visual representations from brain signals, capturing both semantic and pixel-level details for better understanding of human visual system.

Abstract: Decoding visual representations from brain signals has attracted significant attention in both neuroscience and artificial intelligence. However, the degree to which brain signals truly encode visual information remains unclear. Current visual decoding approaches explore various brain-image alignment strategies, yet most emphasize high-level semantic features while neglecting pixel-level details, thereby limiting our understanding of the human visual system. In this paper, we propose a brain-image alignment strategy that leverages multiple pre-trained visual encoders with distinct inductive biases to capture hierarchical and multi-scale visual representations, while employing a contrastive learning objective to achieve effective alignment between brain signals and visual embeddings. Furthermore, we introduce a Fusion Prior, which learns a stable mapping on large-scale visual data and subsequently matches brain features to this pre-trained prior, thereby enhancing distributional consistency across modalities. Extensive quantitative and qualitative experiments demonstrate that our method achieves a favorable balance between retrieval accuracy and reconstruction fidelity.

Method: Three-component framework: (1) Vectra Score - 14 interpretable quality dimensions with spatially-aware Defect Area Ratio quantification; (2) Vectra Dataset - 1.1M real-world images with benchmark, reasoning annotations, and expert preferences; (3) Vectra Model - 4B-parameter MLLM generating scores and diagnostic reasoning.

Result: Vectra achieves state-of-the-art correlation with human rankings and outperforms leading MLLMs (GPT-5, Gemini-3) in scoring performance.

Conclusion: Vectra provides the first comprehensive reference-free visual quality assessment framework for e-commerce IIMT, addressing limitations of existing methods through interpretable metrics and MLLM-driven evaluation.

Abstract: In-Image Machine Translation (IIMT) powers cross-border e-commerce product listings; existing research focuses on machine translation evaluation, while visual rendering quality is critical for user engagement. When facing context-dense product imagery and multimodal defects, current reference-based methods (e.g., SSIM, FID) lack explainability, while model-as-judge approaches lack domain-grounded, fine-grained reward signals. To bridge this gap, we introduce Vectra, to the best of our knowledge, the first reference-free, MLLM-driven visual quality assessment framework for e-commerce IIMT. Vectra comprises three components: (1) Vectra Score, a multidimensional quality metric system that decomposes visual quality into 14 interpretable dimensions, with spatially-aware Defect Area Ratio (DAR) quantification to reduce annotation ambiguity; (2) Vectra Dataset, constructed from 1.1M real-world product images via diversity-aware sampling, comprising a 2K benchmark for system evaluation, 30K reasoning-based annotations for instruction tuning, and 3.5K expert-labeled preferences for alignment and evaluation; and (3) Vectra Model, a 4B-parameter MLLM that generates both quantitative scores and diagnostic reasoning. Experiments demonstrate that Vectra achieves state-of-the-art correlation with human rankings, and our model outperforms leading MLLMs, including GPT-5 and Gemini-3, in scoring performance. The dataset and model will be released upon acceptance.

Method: Built new Bangladeshi banknote dataset with controlled/real-world scenarios, incorporated four additional datasets, proposed hybrid CNN architecture combining MobileNetV3-Large and EfficientNetB0 for feature extraction with MLP classifier, used five-fold cross-validation with seven evaluation metrics, integrated explainable AI methods (LIME and SHAP).

Result: Achieved 97.95% accuracy on controlled datasets, 92.84% on complex backgrounds, and 94.98% accuracy when combining all datasets, with thorough evaluation using multiple metrics and explainable AI for transparency.

Conclusion: The proposed hybrid CNN architecture provides an efficient, accurate, and interpretable solution for Bangladeshi banknote recognition suitable for resource-constrained devices, addressing assistive technology needs for visually impaired individuals.

Abstract: Accurate currency recognition is essential for assistive technologies, particularly for visually impaired individuals who rely on others to identify banknotes. This dependency puts them at risk of fraud and exploitation. To address these challenges, we first build a new Bangladeshi banknote dataset that includes both controlled and real-world scenarios, ensuring a more comprehensive and diverse representation. Next, to enhance the dataset’s robustness, we incorporate four additional datasets, including public benchmarks, to cover various complexities and improve the model’s generalization. To overcome the limitations of current recognition models, we propose a novel hybrid CNN architecture that combines MobileNetV3-Large and EfficientNetB0 for efficient feature extraction. This is followed by an effective multilayer perceptron (MLP) classifier to improve performance while keeping computational costs low, making the system suitable for resource-constrained devices. The experimental results show that the proposed model achieves 97.95% accuracy on controlled datasets, 92.84% on complex backgrounds, and 94.98% accuracy when combining all datasets. The model’s performance is thoroughly evaluated using five-fold cross-validation and seven metrics: accuracy, precision, recall, F1-score, Cohen’s Kappa, MCC, and AUC. Additionally, explainable AI methods like LIME and SHAP are incorporated to enhance transparency and interpretability.

Method: Two-stage framework: 1) Prompt-Aware Semantic Calibration generates adaptive semantic anchors, 2) Neighborhood-aware structural distillation constrains the student’s local decision structure to better capture fine-grained visual patterns.

Result: PAND consistently outperforms state-of-the-art methods on four FGVC benchmarks. ResNet-18 student achieves 76.09% accuracy on CUB-200, surpassing VL2Lite baseline by 3.4%.

Conclusion: PAND effectively addresses limitations of fixed prompts and global alignment in VLMs knowledge distillation for fine-grained visual classification through semantic calibration and structural transfer.

Abstract: Distilling knowledge from large Vision-Language Models (VLMs) into lightweight networks is crucial yet challenging in Fine-Grained Visual Classification (FGVC), due to the reliance on fixed prompts and global alignment. To address this, we propose PAND (Prompt-Aware Neighborhood Distillation), a two-stage framework that decouples semantic calibration from structural transfer. First, we incorporate Prompt-Aware Semantic Calibration to generate adaptive semantic anchors. Second, we introduce a neighborhood-aware structural distillation strategy to constrain the student’s local decision structure. PAND consistently outperforms state-of-the-art methods on four FGVC benchmarks. Notably, our ResNet-18 student achieves 76.09% accuracy on CUB-200, surpassing the strong baseline VL2Lite by 3.4%. Code is available at https://github.com/LLLVTA/PAND.

Method: Three progressively refined pipelines culminating in a LeJEPA-inspired approach that enforces isotropic Gaussian constraints on learned image embeddings to improve clustering consistency and pose estimation robustness.

Result: Experimental results on IMC2025 show Gaussian-constrained embeddings improve scene separation and pose plausibility compared to heuristic-driven baselines, especially in visually ambiguous settings.

Conclusion: Theoretically motivated representation constraints offer a promising direction for bridging self-supervised learning principles and practical structure-from-motion pipelines.

Abstract: Unsupervised 3D scene reconstruction from unstructured image collections remains a fundamental challenge in computer vision, particularly when images originate from multiple unrelated scenes and contain significant visual ambiguity. The Image Matching Challenge 2025 (IMC2025) highlights these difficulties by requiring both scene discovery and camera pose estimation under real-world conditions, including outliers and mixed content. This paper investigates the application of Gaussian-constrained representations inspired by LeJEPA (Joint Embedding Predictive Architecture) to address these challenges. We present three progressively refined pipelines, culminating in a LeJEPA-inspired approach that enforces isotropic Gaussian constraints on learned image embeddings. Rather than introducing new theoretical guarantees, our work empirically evaluates how these constraints influence clustering consistency and pose estimation robustness in practice. Experimental results on IMC2025 demonstrate that Gaussian-constrained embeddings can improve scene separation and pose plausibility compared to heuristic-driven baselines, particularly in visually ambiguous settings. These findings suggest that theoretically motivated representation constraints offer a promising direction for bridging self-supervised learning principles and practical structure-from-motion pipelines.

Method: GOT-Edit uses online cross-modality model editing with null-space constrained updates to integrate geometry-aware cues from a pre-trained Visual Geometry Grounded Transformer into a generic object tracker. The approach infers geometric information from few 2D images while preserving semantic discrimination.

Result: Extensive experiments on multiple GOT benchmarks show GOT-Edit achieves superior robustness and accuracy, particularly under occlusion and clutter conditions, outperforming existing methods.

Conclusion: GOT-Edit establishes a new paradigm for combining 2D semantics with 3D geometric reasoning in object tracking, demonstrating that integrating geometric cues significantly improves tracking robustness across diverse scenarios.

Abstract: Human perception for effective object tracking in a 2D video stream arises from the implicit use of prior 3D knowledge combined with semantic reasoning. In contrast, most generic object tracking (GOT) methods primarily rely on 2D features of the target and its surroundings while neglecting 3D geometric cues, which makes them susceptible to partial occlusion, distractors, and variations in geometry and appearance. To address this limitation, we introduce GOT-Edit, an online cross-modality model editing approach that integrates geometry-aware cues into a generic object tracker from a 2D video stream. Our approach leverages features from a pre-trained Visual Geometry Grounded Transformer to enable geometric cue inference from only a few 2D images. To tackle the challenge of seamlessly combining geometry and semantics, GOT-Edit performs online model editing with null-space constrained updates that incorporate geometric information while preserving semantic discrimination, yielding consistently better performance across diverse scenarios. Extensive experiments on multiple GOT benchmarks demonstrate that GOT-Edit achieves superior robustness and accuracy, particularly under occlusion and clutter, establishing a new paradigm for combining 2D semantics with 3D geometric reasoning for generic object tracking.

Method: Proposes XAI-CLIP framework that uses multimodal vision-language model embeddings to localize clinically meaningful anatomical regions, then applies targeted region-aware perturbations to generate boundary-aware saliency maps with reduced computational overhead.

Result: Achieves 60% runtime reduction, 44.6% dice score improvement, and 96.7% IoU increase for occlusion-based explanations on FLARE22 and CHAOS datasets compared to conventional methods.

Conclusion: Multimodal vision-language representations significantly enhance interpretability and efficiency of perturbation-based XAI, enabling transparent and clinically deployable medical image segmentation systems.

Abstract: Medical image segmentation is a critical component of clinical workflows, enabling accurate diagnosis, treatment planning, and disease monitoring. However, despite the superior performance of transformer-based models over convolutional architectures, their limited interpretability remains a major obstacle to clinical trust and deployment. Existing explainable artificial intelligence (XAI) techniques, including gradient-based saliency methods and perturbation-based approaches, are often computationally expensive, require numerous forward passes, and frequently produce noisy or anatomically irrelevant explanations. To address these limitations, we propose XAI-CLIP, an ROI-guided perturbation framework that leverages multimodal vision-language model embeddings to localize clinically meaningful anatomical regions and guide the explanation process. By integrating language-informed region localization with medical image segmentation and applying targeted, region-aware perturbations, the proposed method generates clearer, boundary-aware saliency maps while substantially reducing computational overhead. Experiments conducted on the FLARE22 and CHAOS datasets demonstrate that XAI-CLIP achieves up to a 60% reduction in runtime, a 44.6% improvement in dice score, and a 96.7% increase in Intersection-over-Union for occlusion-based explanations compared to conventional perturbation methods. Qualitative results further confirm cleaner and more anatomically consistent attribution maps with fewer artifacts, highlighting that the incorporation of multimodal vision-language representations into perturbation-based XAI frameworks significantly enhances both interpretability and efficiency, thereby enabling transparent and clinically deployable medical image segmentation systems.

Method: Proposes an image-based bird classification framework using Convolutional Neural Networks (CNNs) designed to work with camera systems. The CNN identifies bird species and includes dedicated classifiers for flock formation type and flock size estimation.

Result: The framework provides species identification plus supplementary information about flock characteristics (formation type and size) which offer insights into collective flight behavior, trajectory dispersion, and potential impact severity.

Conclusion: The CNN-based visual detection system addresses limitations of current radar systems by enabling species-specific bird strike prediction, with flock characteristics providing valuable additional safety information for aviation.

Abstract: Bird strikes pose a significant threat to aviation safety, often resulting in loss of life, severe aircraft damage, and substantial financial costs. Existing bird strike prevention strategies primarily rely on avian radar systems that detect and track birds in real time. A major limitation of these systems is their inability to identify bird species, an essential factor, as different species exhibit distinct flight behaviors, and altitudinal preference. To address this challenge, we propose an image-based bird classification framework using Convolutional Neural Networks (CNNs), designed to work with camera systems for autonomous visual detection. The CNN is designed to identify bird species and provide critical input to species-specific predictive models for accurate flight path prediction. In addition to species identification, we implemented dedicated CNN classifiers to estimate flock formation type and flock size. These characteristics provide valuable supplementary information for aviation safety. Specifically, flock type and size offer insights into collective flight behavior, and trajectory dispersion . Flock size directly relates to the potential impact severity, as the overall damage risk increases with the combined kinetic energy of multiple birds.

Method: Proposes Hierarchical Refinement Attack (HRA) with two components: 1) For images: uses temporal hierarchy of historical and estimated future gradients to refine optimization path and avoid local minima; 2) For text: hierarchically models textual importance considering both intra- and inter-sentence contributions to identify globally influential words for universal text perturbations.

Result: Extensive experiments across various downstream tasks, VLP models, and datasets demonstrate superior transferability of the proposed universal multimodal attacks compared to existing methods.

Conclusion: HRA provides an effective universal attack framework for VLP models that overcomes the computational limitations of sample-specific attacks while maintaining strong transferability across different scenarios.

Abstract: Existing adversarial attacks for VLP models are mostly sample-specific, resulting in substantial computational overhead when scaled to large datasets or new scenarios. To overcome this limitation, we propose Hierarchical Refinement Attack (HRA), a multimodal universal attack framework for VLP models. For the image modality, we refine the optimization path by leveraging a temporal hierarchy of historical and estimated future gradients to avoid local minima and stabilize universal perturbation learning. For the text modality, it hierarchically models textual importance by considering both intra- and inter-sentence contributions to identify globally influential words, which are then used as universal text perturbations. Extensive experiments across various downstream tasks, VLP models, and datasets, demonstrate the superior transferability of the proposed universal multimodal attacks.

Method: The paper proposes a mechanistic analysis by examining the representational geometry of open-weight VLMs (Qwen, InternVL, Gemma). It compares methodologies to distill “concept vectors” - latent directions encoding visual concepts. These concept vectors are validated via steering interventions that manipulate model behavior in both simplified and naturalistic vision tasks.

Result: The geometric overlap between concept vectors strongly correlates with specific error patterns. This provides a grounded quantitative framework to understand how internal representations shape model behavior and drive visual failures in VLMs.

Conclusion: The analysis of concept vectors and representational geometry offers mechanistic insights into VLM failures in multi-object tasks, providing a quantitative framework to understand how internal representations influence model behavior and visual errors.

Abstract: Vision-Language Models (VLMs) exhibit puzzling failures in multi-object visual tasks, such as hallucinating non-existent elements or failing to identify the most similar objects among distractions. While these errors mirror human cognitive constraints, such as the “Binding Problem”, the internal mechanisms driving them in artificial systems remain poorly understood. Here, we propose a mechanistic insight by analyzing the representational geometry of open-weight VLMs (Qwen, InternVL, Gemma), comparing methodologies to distill “concept vectors” - latent directions encoding visual concepts. We validate our concept vectors via steering interventions that reliably manipulate model behavior in both simplified and naturalistic vision tasks (e.g., forcing the model to perceive a red flower as blue). We observe that the geometric overlap between these vectors strongly correlates with specific error patterns, offering a grounded quantitative framework to understand how internal representations shape model behavior and drive visual failures.

Method: Fair Context Learning (FCL) uses episodic TTA with two components: (1) augmentation-based exploration to identify plausible class candidates, and (2) fairness-driven calibration that adapts text contexts to equalize sensitivity to common visual evidence, avoiding entropy minimization.

Result: FCL achieves competitive adaptation performance relative to state-of-the-art TTA methods across diverse domain-shift and fine-grained benchmarks, validating the theoretical motivation.

Conclusion: FCL provides an effective alternative to entropy minimization for test-time adaptation of VLMs, addressing shared-evidence bias through fairness constraints to improve robustness to distribution shifts.

Abstract: Vision-Language Models (VLMs) such as CLIP enable strong zero-shot recognition but suffer substantial degradation under distribution shifts. Test-Time Adaptation (TTA) aims to improve robustness using only unlabeled test samples, yet most prompt-based TTA methods rely on entropy minimization – an approach that can amplify spurious correlations and induce overconfident errors when classes share visual features. We propose Fair Context Learning (FCL), an episodic TTA framework that avoids entropy minimization by explicitly addressing shared-evidence bias. Motivated by our additive evidence decomposition assumption, FCL decouples adaptation into (i) augmentation-based exploration to identify plausible class candidates, and (ii) fairness-driven calibration that adapts text contexts to equalize sensitivity to common visual evidence. This fairness constraint mitigates partial feature obsession and enables effective calibration of text embeddings without relying on entropy reduction. Through extensive evaluation, we empirically validate our theoretical motivation and show that FCL achieves competitive adaptation performance relative to state-of-the-art TTA methods across diverse domain-shift and fine-grained benchmarks.

Method: MOVA uses a Mixture-of-Experts architecture with 32B total parameters (18B active during inference). It supports Image-Text to Video-Audio (IT2VA) generation tasks. The model is designed for joint multimodal modeling of audio and video streams.

Result: MOVA generates high-quality, synchronized audio-visual content including realistic lip-synced speech, environment-aware sound effects, and content-aligned music. The model weights and code are released to advance research.

Conclusion: MOVA provides an open-source solution for joint audio-visual generation, addressing challenges in architecture, data, and training. The release aims to foster community development and research in multimodal generation.

Abstract: Audio is indispensable for real-world video, yet generation models have largely overlooked audio components. Current approaches to producing audio-visual content often rely on cascaded pipelines, which increase cost, accumulate errors, and degrade overall quality. While systems such as Veo 3 and Sora 2 emphasize the value of simultaneous generation, joint multimodal modeling introduces unique challenges in architecture, data, and training. Moreover, the closed-source nature of existing systems limits progress in the field. In this work, we introduce MOVA (MOSS Video and Audio), an open-source model capable of generating high-quality, synchronized audio-visual content, including realistic lip-synced speech, environment-aware sound effects, and content-aligned music. MOVA employs a Mixture-of-Experts (MoE) architecture, with a total of 32B parameters, of which 18B are active during inference. It supports IT2VA (Image-Text to Video-Audio) generation task. By releasing the model weights and code, we aim to advance research and foster a vibrant community of creators. The released codebase features comprehensive support for efficient inference, LoRA fine-tuning, and prompt enhancement.

Method: Compared standard CNNs (ConvNet, VGG, ResNet18) with their ANFIS-augmented counterparts on MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 datasets under gradient-based (PGD) and gradient-free (Square) adversarial attacks.

Result: ANFIS integration does not consistently improve clean accuracy. Robustness effects are architecture-dependent: ResNet18-ANFIS exhibits improved adversarial robustness, while VGG-ANFIS often underperforms its baseline counterpart.

Conclusion: Neuro-fuzzy augmentation can enhance adversarial robustness in specific CNN architectures but is not universally beneficial, highlighting the importance of architecture selection when implementing such hybrid approaches.

Abstract: Convolutional Neural Networks (CNNs) achieve strong image classification performance but lack interpretability and are vulnerable to adversarial attacks. Neuro-fuzzy hybrids such as DCNFIS replace fully connected CNN classifiers with Adaptive Neuro-Fuzzy Inference Systems (ANFIS) to improve interpretability, yet their robustness remains underexplored. This work compares standard CNNs (ConvNet, VGG, ResNet18) with their ANFIS-augmented counterparts on MNIST, Fashion-MNIST, CIFAR-10, and CIFAR-100 under gradient-based (PGD) and gradient-free (Square) attacks. Results show that ANFIS integration does not consistently improve clean accuracy and has architecture-dependent effects on robustness: ResNet18-ANFIS exhibits improved adversarial robustness, while VGG-ANFIS often underperforms its baseline. These findings suggest that neuro-fuzzy augmentation can enhance robustness in specific architectures but is not universally beneficial.

Method: The authors introduce UNIKIE-BENCH, a unified benchmark with two complementary tracks: 1) constrained-category KIE with scenario-predefined schemas reflecting practical needs, and 2) open-category KIE extracting any explicitly present key information. They evaluate 15 state-of-the-art LMMs on this benchmark.

Result: Experiments reveal substantial performance degradation under diverse schema definitions, long-tail key fields, and complex layouts. There are pronounced performance disparities across different document types and scenarios, highlighting challenges in grounding accuracy and layout-aware reasoning for LMM-based KIE.

Conclusion: The benchmark reveals persistent challenges in LMM-based KIE, particularly in grounding accuracy and layout-aware reasoning. UNIKIE-BENCH provides a comprehensive evaluation framework to advance research in document understanding and multimodal information extraction.

Abstract: Key Information Extraction (KIE) from real-world documents remains challenging due to substantial variations in layout structures, visual quality, and task-specific information requirements. Recent Large Multimodal Models (LMMs) have shown promising potential for performing end-to-end KIE directly from document images. To enable a comprehensive and systematic evaluation across realistic and diverse application scenarios, we introduce UNIKIE-BENCH, a unified benchmark designed to rigorously evaluate the KIE capabilities of LMMs. UNIKIE-BENCH consists of two complementary tracks: a constrained-category KIE track with scenario-predefined schemas that reflect practical application needs, and an open-category KIE track that extracts any key information that is explicitly present in the document. Experiments on 15 state-of-the-art LMMs reveal substantial performance degradation under diverse schema definitions, long-tail key fields, and complex layouts, along with pronounced performance disparities across different document types and scenarios. These findings underscore persistent challenges in grounding accuracy and layout-aware reasoning for LMM-based KIE. All codes and datasets are available at https://github.com/NEUIR/UNIKIE-BENCH.

Method: Uses a general-purpose VLM with embedded diagnostic heuristics from dental experts, operates on multi-view smartphone photographs, performs tooth-level evaluation without dental-specific fine-tuning.

Result: Framework enables dental diagnosis in settings without curated clinical imaging, serves as early-stage assistive tool to identify abnormalities and determine need for professional evaluation.

Conclusion: OMNI-Dent offers data-efficient, explainable dental diagnosis using VLMs with clinical reasoning, providing accessible option for limited healthcare access.

Abstract: Accurate dental diagnosis is essential for oral healthcare, yet many individuals lack access to timely professional evaluation. Existing AI-based methods primarily treat diagnosis as a visual pattern recognition task and do not reflect the structured clinical reasoning used by dental professionals. These approaches also require large amounts of expert-annotated data and often struggle to generalize across diverse real-world imaging conditions. To address these limitations, we present OMNI-Dent, a data-efficient and explainable diagnostic framework that incorporates clinical reasoning principles into a Vision-Language Model (VLM)-based pipeline. The framework operates on multi-view smartphone photographs,embeds diagnostic heuristics from dental experts, and guides a general-purpose VLM to perform tooth-level evaluation without dental-specific fine-tuning of the VLM. By utilizing the VLM’s existing visual-linguistic capabilities, OMNI-Dent aims to support diagnostic assessment in settings where curated clinical imaging is unavailable. Designed as an early-stage assistive tool, OMNI-Dent helps users identify potential abnormalities and determine when professional evaluation may be needed, offering a practical option for individuals with limited access to in-person care.

Method: Proposes markerless approaches using low-cost visible and infrared light cameras with stereo and depth sensing, combined with algorithmic modeling of facial geometry to track head position without physical markers.

Result: Validation with 50 human subjects showed median tracking discrepancy of 2.32mm and 2.01° for best markerless algorithms, comparable to conventional marker-based systems and substantially better than prior markerless results.

Conclusion: Markerless neuronavigation can reduce setup cost/complexity, improve patient comfort, and expand access to neuronavigation in clinical/research settings, with potential for further accuracy improvements through sensor fusion.

Abstract: Neuronavigation is widely used in biomedical research and interventions to guide the precise placement of instruments around the head to support procedures such as transcranial magnetic stimulation. Traditional systems, however, rely on subject-mounted markers that require manual registration, may shift during procedures, and can cause discomfort. We introduce and evaluate markerless approaches that replace expensive hardware and physical markers with low-cost visible and infrared light cameras incorporating stereo and depth sensing combined with algorithmic modeling of the facial geometry. Validation with $50$ human subjects yielded a median tracking discrepancy of only $2.32$ mm and $2.01°$ for the best markerless algorithms compared to a conventional marker-based system, which indicates sufficient accuracy for transcranial magnetic stimulation and a substantial improvement over prior markerless results. The results suggest that integration of the data from the various camera sensors can improve the overall accuracy further. The proposed markerless neuronavigation methods can reduce setup cost and complexity, improve patient comfort, and expand access to neuronavigation in clinical and research settings.

Method: COMBOOD combines signals from two distance metrics: nearest-neighbor (non-parametric) and Mahalanobis distance (parametric). The framework operates in a semi-parametric setting to derive a confidence score for OOD detection, scaling linearly with embedding space size.

Result: COMBOOD outperforms state-of-the-art OOD detection methods on OpenOOD benchmarks (v1 and v1.5) for both far-OOD and near-OOD scenarios, as well as on document datasets. Improvements are statistically significant on most benchmark datasets.

Conclusion: The COMBOOD framework provides an effective solution for OOD detection that works well across both near and far OOD scenarios, making it suitable for real-world applications due to its linear scaling and strong performance.

Abstract: Identifying out-of-distribution (OOD) data at inference time is crucial for many machine learning applications, especially for automation. We present a novel unsupervised semi-parametric framework COMBOOD for OOD detection with respect to image recognition. Our framework combines signals from two distance metrics, nearest-neighbor and Mahalanobis, to derive a confidence score for an inference point to be out-of-distribution. The former provides a non-parametric approach to OOD detection. The latter provides a parametric, simple, yet effective method for detecting OOD data points, especially, in the far OOD scenario, where the inference point is far apart from the training data set in the embedding space. However, its performance is not satisfactory in the near OOD scenarios that arise in practical situations. Our COMBOOD framework combines the two signals in a semi-parametric setting to provide a confidence score that is accurate both for the near-OOD and far-OOD scenarios. We show experimental results with the COMBOOD framework for different types of feature extraction strategies. We demonstrate experimentally that COMBOOD outperforms state-of-the-art OOD detection methods on the OpenOOD (both version 1 and most recent version 1.5) benchmark datasets (for both far-OOD and near-OOD) as well as on the documents dataset in terms of accuracy. On a majority of the benchmark datasets, the improvements in accuracy resulting from the COMBOOD framework are statistically significant. COMBOOD scales linearly with the size of the embedding space, making it ideal for many real-life applications.

Method: Used Global Context Vision Transformer (GCViT)-Tiny architecture on Oxford-IIIT Pet Dataset subset with extensive data augmentation (rotation, horizontal flipping, brightness adjustment).

Result: Achieved 92.00% test accuracy and 94.54% validation accuracy, demonstrating transformer effectiveness for fine-grained image classification.

Conclusion: Transformer-based architectures like GCViT are effective for fine-grained visual classification tasks, with practical applications in animal-related domains.

Abstract: Accurate identification of cat breeds from images is a challenging task due to subtle differences in fur patterns, facial structure, and color. In this paper, we present a deep learning-based approach for classifying cat breeds using a subset of the Oxford-IIIT Pet Dataset, which contains high-resolution images of various domestic breeds. We employed the Global Context Vision Transformer (GCViT) architecture-tiny for cat breed recognition. To improve model generalization, we used extensive data augmentation, including rotation, horizontal flipping, and brightness adjustment. Experimental results show that the GCViT-Tiny model achieved a test accuracy of 92.00% and validation accuracy of 94.54%. These findings highlight the effectiveness of transformer-based architectures for fine-grained image classification tasks. Potential applications include veterinary diagnostics, animal shelter management, and mobile-based breed recognition systems. We also provide a hugging face demo at https://huggingface.co/spaces/bfarhad/cat-breed-classifier.

Method: Created PipeMFL-240K dataset with 240,320 images and 191,530 bounding-box annotations from 11 pipelines spanning ~1,480 km, featuring 12 categories with extreme long-tailed distribution, tiny objects, and high intra-class variability.

Result: Extensive experiments with state-of-the-art object detectors show they still struggle with MFL data’s intrinsic properties, indicating considerable room for improvement, while PipeMFL-240K provides a reliable testbed for future research.

Conclusion: PipeMFL-240K is the first public dataset and benchmark of this scale for pipeline MFL inspection, providing a critical foundation for efficient pipeline diagnostics and maintenance planning, expected to accelerate algorithmic innovation in pipeline integrity assessment.

Abstract: Pipeline integrity is critical to industrial safety and environmental protection, with Magnetic Flux Leakage (MFL) detection being a primary non-destructive testing technology. Despite the promise of deep learning for automating MFL interpretation, progress toward reliable models has been constrained by the absence of a large-scale public dataset and benchmark, making fair comparison and reproducible evaluation difficult. We introduce \textbf{PipeMFL-240K}, a large-scale, meticulously annotated dataset and benchmark for complex object detection in pipeline MFL pseudo-color images. PipeMFL-240K reflects real-world inspection complexity and poses several unique challenges: (i) an extremely long-tailed distribution over \textbf{12} categories, (ii) a high prevalence of tiny objects that often comprise only a handful of pixels, and (iii) substantial intra-class variability. The dataset contains \textbf{240,320} images and \textbf{191,530} high-quality bounding-box annotations, collected from 11 pipelines spanning approximately \textbf{1,480} km. Extensive experiments are conducted with state-of-the-art object detectors to establish baselines. Results show that modern detectors still struggle with the intrinsic properties of MFL data, highlighting considerable headroom for improvement, while PipeMFL-240K provides a reliable and challenging testbed to drive future research. As the first public dataset and the first benchmark of this scale and scope for pipeline MFL inspection, it provides a critical foundation for efficient pipeline diagnostics as well as maintenance planning and is expected to accelerate algorithmic innovation and reproducible research in MFL-based pipeline integrity assessment.

Method: Proposed Vision Language ReaSoning Benchmark (VLRS-Bench) with 2,000 question-answer pairs across 14 tasks and up to eight temporal phases, structured around three core dimensions: Cognition, Decision, and Prediction. Constructed via specialized pipeline integrating RS-specific priors and expert knowledge.

Result: Experimental results reveal significant bottlenecks in existing state-of-the-art MLLMs, providing critical insights for advancing multimodal reasoning within the remote sensing community.

Conclusion: VLRS-Bench addresses the gap in RS benchmarks by focusing on complex reasoning tasks, enabling better evaluation and development of MLLMs for advanced remote sensing applications.

Abstract: Recent advancements in Multimodal Large Language Models (MLLMs) have enabled complex reasoning. However, existing remote sensing (RS) benchmarks remain heavily biased toward perception tasks, such as object recognition and scene classification. This limitation hinders the development of MLLMs for cognitively demanding RS applications. To address this, , we propose a Vision Language ReaSoning Benchmark (VLRS-Bench), which is the first benchmark exclusively dedicated to complex RS reasoning. Structured across the three core dimensions of Cognition, Decision, and Prediction, VLRS-Bench comprises 2,000 question-answer pairs with an average length of 71 words, spanning 14 tasks and up to eight temporal phases. VLRS-Bench is constructed via a specialized pipeline that integrates RS-specific priors and expert knowledge to ensure geospatial realism and reasoning complexity. Experimental results reveal significant bottlenecks in existing state-of-the-art MLLMs, providing critical insights for advancing multimodal reasoning within the remote sensing community.

Method: ShapBPT assigns Shapley coefficients to a multiscale hierarchical structure called Binary Partition Tree (BPT), which is tailored for images. This data-aware hierarchical partitioning aligns feature attributions with intrinsic image morphology.

Result: Experimental results show superior alignment with image structures, improved efficiency over existing XCV methods, and a 20-subject user study confirms human preference for ShapBPT explanations.

Conclusion: ShapBPT connects hierarchical Shapley methods with image data, providing more efficient and semantically meaningful visual interpretability for computer vision models.

Abstract: Pixel-level feature attributions are an important tool in eXplainable AI for Computer Vision (XCV), providing visual insights into how image features influence model predictions. The Owen formula for hierarchical Shapley values has been widely used to interpret machine learning (ML) models and their learned representations. However, existing hierarchical Shapley approaches do not exploit the multiscale structure of image data, leading to slow convergence and weak alignment with the actual morphological features. Moreover, no prior Shapley method has leveraged data-aware hierarchies for Computer Vision tasks, leaving a gap in model interpretability of structured visual data. To address this, this paper introduces ShapBPT, a novel data-aware XCV method based on the hierarchical Shapley formula. ShapBPT assigns Shapley coefficients to a multiscale hierarchical structure tailored for images, the Binary Partition Tree (BPT). By using this data-aware hierarchical partitioning, ShapBPT ensures that feature attributions align with intrinsic image morphology, effectively prioritizing relevant regions while reducing computational overhead. This advancement connects hierarchical Shapley methods with image data, providing a more efficient and semantically meaningful approach to visual interpretability. Experimental results confirm ShapBPT’s effectiveness, demonstrating superior alignment with image structures and improved efficiency over existing XCV methods, and a 20-subject user study confirming that ShapBPT explanations are preferred by humans.

Method: Proposes Error-enhanced Contrastive Handwriting Recognition (ECHWR) with a temporary auxiliary branch during training that aligns sensor signals with text embeddings using dual contrastive objectives: in-batch contrastive loss for modality alignment and novel error-based contrastive loss distinguishing correct signals from synthetic hard negatives.

Result: On OnHW-Words500 dataset, ECHWR reduces character error rates by up to 7.4% on writer-independent split and 10.4% on writer-dependent split compared to state-of-the-art baselines, while maintaining original efficient architecture at inference.

Conclusion: ECHWR effectively improves handwriting recognition accuracy for edge devices without increasing inference costs, with error-based contrastive loss proving particularly effective for handling unseen writing styles.

Abstract: Online handwriting recognition using inertial measurement units opens up handwriting on paper as input for digital devices. Doing it on edge hardware improves privacy and lowers latency, but entails memory constraints. To address this, we propose Error-enhanced Contrastive Handwriting Recognition (ECHWR), a training framework designed to improve feature representation and recognition accuracy without increasing inference costs. ECHWR utilizes a temporary auxiliary branch that aligns sensor signals with semantic text embeddings during the training phase. This alignment is maintained through a dual contrastive objective: an in-batch contrastive loss for general modality alignment and a novel error-based contrastive loss that distinguishes between correct signals and synthetic hard negatives. The auxiliary branch is discarded after training, which allows the deployed model to keep its original, efficient architecture. Evaluations on the OnHW-Words500 dataset show that ECHWR significantly outperforms state-of-the-art baselines, reducing character error rates by up to 7.4% on the writer-independent split and 10.4% on the writer-dependent split. Finally, although our ablation studies indicate that solving specific challenges require specific architectural and objective configurations, error-based contrastive loss shows its effectiveness for handling unseen writing styles.

Method: Used interpretability techniques including layerwise probing, subspace geometry analysis, patch-level decoding, and targeted attention ablations on large-scale video encoders to examine physical representations.

Result: Identified a Physics Emergence Zone at intermediate depths where physical variables become accessible. Scalar quantities (speed, acceleration) are available early, while motion direction becomes accessible only at this zone through high-dimensional population structures with circular geometry.

Conclusion: Modern video models use distributed rather than factorized representations of physical variables, which are nonetheless sufficient for making physical predictions, challenging classical physics engine approaches.

Abstract: A long-standing question in physical reasoning is whether video-based models need to rely on factorized representations of physical variables in order to make physically accurate predictions, or whether they can implicitly represent such variables in a task-specific, distributed manner. While modern video world models achieve strong performance on intuitive physics benchmarks, it remains unclear which of these representational regimes they implement internally. Here, we present the first interpretability study to directly examine physical representations inside large-scale video encoders. Using layerwise probing, subspace geometry, patch-level decoding, and targeted attention ablations, we characterize where physical information becomes accessible and how it is organized within encoder-based video transformers. Across architectures, we identify a sharp intermediate-depth transition – which we call the Physics Emergence Zone – at which physical variables become accessible. Physics-related representations peak shortly after this transition and degrade toward the output layers. Decomposing motion into explicit variables, we find that scalar quantities such as speed and acceleration are available from early layers onwards, whereas motion direction becomes accessible only at the Physics Emergence Zone. Notably, we find that direction is encoded through a high-dimensional population structure with circular geometry, requiring coordinated multi-feature intervention to control. These findings suggest that modern video models do not use factorized representations of physical variables like a classical physics engine. Instead, they use a distributed representation that is nonetheless sufficient for making physical predictions.

Method: Fine-tuned PaliGemma 3B model with Low-Rank Adaptation (LoRA) to perform license plate recognition, state classification, and vehicle attribute extraction in a single forward pass. Includes Human-in-the-Loop continual learning with experience replay (70:30 original:correction ratio).

Result: Achieved 92.3% plate recognition accuracy (14.1% improvement over EasyOCR, 9.9% over PaddleOCR), 152ms mean inference latency, ECE of 0.048. Zero-shot generalization to vehicle color detection (89%), seatbelt detection (82%), and occupancy counting (78%).

Conclusion: Unified vision-language approaches represent a paradigm shift in ALPR systems, offering superior accuracy, reduced complexity, and emergent multi-task capabilities that traditional pipelines cannot achieve.

Abstract: Traditional Automatic License Plate Recognition (ALPR) systems employ multi-stage pipelines consisting of object detection networks followed by separate Optical Character Recognition (OCR) modules, introducing compounding errors, increased latency, and architectural complexity. This research presents Neural Sentinel, a novel unified approach that leverages Vision Language Models (VLMs) to perform license plate recognition, state classification, and vehicle attribute extraction through a single forward pass. Our primary contribution lies in demonstrating that a fine-tuned PaliGemma 3B model, adapted via Low-Rank Adaptation (LoRA), can simultaneously answer multiple visual questions about vehicle images, achieving 92.3% plate recognition accuracy, which is a 14.1% improvement over EasyOCR and 9.9% improvement over PaddleOCR baselines. We introduce a Human-in-the-Loop (HITL) continual learning framework that incorporates user corrections while preventing catastrophic forgetting through experience replay, maintaining a 70:30 ratio of original training data to correction samples. The system achieves a mean inference latency of 152ms with an Expected Calibration Error (ECE) of 0.048, indicating well calibrated confidence estimates. Additionally, the VLM first architecture enables zero-shot generalization to auxiliary tasks including vehicle color detection (89%), seatbelt detection (82%), and occupancy counting (78%) without task specific training. Through extensive experimentation on real world toll plaza imagery, we demonstrate that unified vision language approaches represent a paradigm shift in ALPR systems, offering superior accuracy, reduced architectural complexity, and emergent multi-task capabilities that traditional pipeline approaches cannot achieve.

Method: Combines state-of-the-art latent diffusion models with interactive semantic segmentation to create RECITYGEN, a tool that allows users to interactively generate variational street view images of urban environments using text prompts.

Result: In a pilot project in Beijing, users successfully employed RECITYGEN to suggest improvements for an ongoing Urban Regeneration project. The tool demonstrated significant potential in aligning with public preferences despite some limitations.

Conclusion: RECITYGEN represents a shift towards more dynamic and inclusive urban planning methods by enabling public participation through AI-powered visual generation tools.

Abstract: Urban design profoundly impacts public spaces and community engagement. Traditional top-down methods often overlook public input, creating a gap in design aspirations and reality. Recent advancements in digital tools, like City Information Modelling and augmented reality, have enabled a more participatory process involving more stakeholders in urban design. Further, deep learning and latent diffusion models have lowered barriers for design generation, providing even more opportunities for participatory urban design. Combining state-of-the-art latent diffusion models with interactive semantic segmentation, we propose RECITYGEN, a novel tool that allows users to interactively create variational street view images of urban environments using text prompts. In a pilot project in Beijing, users employed RECITYGEN to suggest improvements for an ongoing Urban Regeneration project. Despite some limitations, RECITYGEN has shown significant potential in aligning with public preferences, indicating a shift towards more dynamic and inclusive urban planning methods. The source code for the project can be found at RECITYGEN GitHub.

Method: Two-stage approach: 1) Identify parameters most responsible for forget set using gradient-based saliency and constrain updates through sparse LoRA adapters for lightweight, localized modifications. 2) Apply self-distillation objective that overwrites forgotten concept with user-defined surrogate while preserving behavior on retained data.

Result: Achieves state-of-the-art unlearning performance on UnlearnCanvas benchmark and ablation studies across multiple datasets (Imagenette, LFW, Dog Breeds, SUN Attributes), demonstrating strong concept erasure and high retainability with fine-grained control over forgetting-retention trade-off.

Conclusion: FADE provides memory-efficient, reversible adapters that can be merged or removed at runtime, enabling flexible deployment in production systems for selective unlearning in diffusion-based image generation models.

Abstract: Machine Unlearning aims to remove the influence of specific data or concepts from trained models while preserving overall performance, a capability increasingly required by data protection regulations and responsible AI practices. Despite recent progress, unlearning in text-to-image diffusion models remains challenging due to high computational costs and the difficulty of balancing effective forgetting with retention of unrelated concepts. We introduce FADE (Fast Adapter for Data Erasure), a two-stage unlearning method for image generation that combines parameter localization with self-distillation. FADE first identifies parameters most responsible for the forget set using gradient-based saliency and constrains updates through sparse LoRA adapters, ensuring lightweight, localized modifications. In a second stage, FADE applies a self-distillation objective that overwrites the forgotten concept with a user-defined surrogate while preserving behavior on retained data. The resulting adapters are memory-efficient, reversible, and can be merged or removed at runtime, enabling flexible deployment in production systems. We evaluated FADE on the UnlearnCanvas benchmark and conducted ablation studies on Imagenette, Labeled Faces in the Wild, AtharvaTaras Dog Breeds Dataset, and SUN Attributes datasets, demonstrating State-of-the-Art unlearning performance with fine-grained control over the forgetting-retention trade-off. Our results demonstrate that FADE achieves strong concept erasure and high retainability across various domains, making it a suitable solution for selective unlearning in diffusion-based image generation models.

Method: Formulates contamination assessment as regression task using multi-instance learning (MIL) for sequential data and multi-task learning (MTL) for joint contamination estimation and scrap classification. Implements near real-time system with magnet/railcar detection, versioned inference service, and active-learning loop.

Result: Best results: MAE 0.27 and R² 0.83 for contamination estimation using MIL; MTL achieves MAE 0.36 with F1 0.79 for scrap classification. System reduces subjective variability and improves safety.

Conclusion: Computer vision pipeline successfully automates scrap quality assessment, enabling integration into acceptance workflows while supporting continual improvement through active learning.

Abstract: Scrap quality directly affects energy use, emissions, and safety in steelmaking. Today, the share of non-metallic inclusions (contamination) is judged visually by inspectors - an approach that is subjective and hazardous due to dust and moving machinery. We present an assistive computer vision pipeline that estimates contamination (per percent) from images captured during railcar unloading and also classifies scrap type. The method formulates contamination assessment as a regression task at the railcar level and leverages sequential data through multi-instance learning (MIL) and multi-task learning (MTL). Best results include MAE 0.27 and R2 0.83 by MIL; and an MTL setup reaches MAE 0.36 with F1 0.79 for scrap class. Also we present the system in near real time within the acceptance workflow: magnet/railcar detection segments temporal layers, a versioned inference service produces railcar-level estimates with confidence scores, and results are reviewed by operators with structured overrides; corrections and uncertain cases feed an active-learning loop for continual improvement. The pipeline reduces subjective variability, improves human safety, and enables integration into acceptance and melt-planning workflows.

Method: Established a public multi-dataset benchmark using Duke Lung Cancer Screening (DLCS) as development set, evaluated across LUNA16/LIDC-IDRI, NLST-3D, and LUNA25. For detection: trained models on DLCS and LUNA16, evaluated externally on NLST-3D using free-response ROC analysis. For malignancy classification: compared five strategies including randomly initialized ResNet50, Models Genesis, Med3D, Foundation Model for Cancer Biomarkers, and Strategic Warm-Start (ResNet50-SWS) pretrained using detection-derived candidate patches stratified by confidence.

Result: Detection performance varied substantially by training dataset - DLCS-trained models outperformed LUNA16-trained models on external NLST-3D evaluation (sensitivity at 2 false positives per scan: 0.72 vs. 0.64; p < 0.001). For malignancy classification, ResNet50-SWS achieved AUCs of 0.71 (DLCS), 0.90 (LUNA16), 0.81 (NLST-3D), and 0.80 (LUNA25), consistently matching or exceeding alternative pretraining strategies.

Conclusion: Dataset characteristics strongly influence lung cancer AI performance, highlighting the critical need for transparent, multi-dataset benchmarking to enable reliable performance comparisons and clinical translation across different settings.

Abstract: Evaluation of artificial intelligence (AI) models for low-dose CT lung cancer screening is limited by heterogeneous datasets, annotation standards, and evaluation protocols, making performance difficult to compare and translate across clinical settings. We establish a public, reproducible multi-dataset benchmark for lung nodule detection and nodule-level cancer classification and quantify cross-dataset generalizability. Using the Duke Lung Cancer Screening (DLCS) dataset as a clinically curated development set, we evaluate performance across LUNA16/LIDC-IDRI, NLST-3D, and LUNA25. Detection models trained on DLCS and LUNA16 were evaluated externally on NLST-3D using free-response ROC analysis. For malignancy classification, we compared five strategies: randomly initialized ResNet50, Models Genesis, Med3D, a Foundation Model for Cancer Biomarkers, and a Strategic Warm-Start (ResNet50-SWS) approach pretrained using detection-derived candidate patches stratified by confidence. Performance was summarized using AUC with 95% confidence intervals and DeLong tests. Detection performance varied substantially by training dataset, with DLCS-trained models outperforming LUNA16-trained models on external NLST-3D evaluation (sensitivity at 2 false positives per scan: 0.72 vs. 0.64; p < 0.001). For malignancy classification, ResNet50-SWS achieved AUCs of 0.71 (DLCS), 0.90 (LUNA16), 0.81 (NLST-3D), and 0.80 (LUNA25), consistently matching or exceeding alternative pretraining strategies. These results demonstrate that dataset characteristics strongly influence lung cancer AI performance and highlight the need for transparent, multi-dataset benchmarking.

Method: 1) Build a physical data engine with two components: FysicsAny (produces physics-grounded instruction-image supervision via hierarchical retrieval over curated prototypes) and FysicsOmniCap (distills web videos via audio-visual consistency filtering). 2) Train OmniFysics with staged multimodal alignment and instruction tuning. 3) Use latent-space flow matching for text-to-image generation. 4) Implement an intent router to activate generation only when needed.

Result: The model shows competitive performance on standard multimodal benchmarks and improved results on physics-oriented evaluations, demonstrating enhanced physical understanding capabilities.

Conclusion: OmniFysics successfully integrates explicit physical knowledge into omni-modal models, addressing the brittleness of physical understanding in current systems while maintaining competitive multimodal performance.

Abstract: Physical understanding remains brittle in omni-modal models because key physical attributes are visually ambiguous and sparsely represented in web-scale data. We present OmniFysics, a compact omni-modal model that unifies understanding across images, audio, video, and text, with integrated speech and image generation. To inject explicit physical knowledge, we build a physical data engine with two components. FysicsAny produces physics-grounded instruction–image supervision by mapping salient objects to verified physical attributes through hierarchical retrieval over a curated prototype database, followed by physics-law–constrained verification and caption rewriting. FysicsOmniCap distills web videos via audio–visual consistency filtering to generate high-fidelity video–instruction pairs emphasizing cross-modal physical cues. We train OmniFysics with staged multimodal alignment and instruction tuning, adopt latent-space flow matching for text-to-image generation, and use an intent router to activate generation only when needed. Experiments show competitive performance on standard multimodal benchmarks and improved results on physics-oriented evaluations.

Method: HRSino explicitly accounts for spatial heterogeneity in signal characteristics (spectral sparsity and local complexity) to adaptively allocate inference effort across spatial regions and resolutions, capturing global consistency at coarse scales while refining local details only where necessary.

Result: HRSino reduces peak memory usage by up to 30.81% and inference time by up to 17.58% compared to state-of-the-art frameworks while maintaining completion accuracy across datasets and resolutions.

Conclusion: The proposed training-free approach enables efficient high-resolution sinogram completion by adaptively distributing computational resources based on spatial signal characteristics, offering significant efficiency gains without sacrificing accuracy.

Abstract: High-resolution sinogram completion is critical for computed tomography reconstruction, as missing projections can introduce severe artifacts. While diffusion models provide strong generative priors for this task, their inference cost grows prohibitively with resolution. We propose HRSino, a training-free and efficient diffusion inference approach for high-resolution sinogram completion. By explicitly accounting for spatial heterogeneity in signal characteristics, such as spectral sparsity and local complexity, HRSino allocates inference effort adaptively across spatial regions and resolutions, rather than applying uniform high-resolution diffusion steps. This enables global consistency to be captured at coarse scales while refining local details only where necessary. Experimental results show that HRSino reduces peak memory usage by up to 30.81% and inference time by up to 17.58% compared to the state-of-the-art framework, and maintains completion accuracy across datasets and resolutions.

Method: Two-stage deep learning framework: first neural network for image registration to estimate displacement fields, second network for material compressibility estimation from the displacement data, trained end-to-end on reference data.

Result: Deep learning model outperforms conventional approaches in both efficiency and accuracy, can accurately determine material compressibility even with substantial local deviations in displacement field predictions.

Conclusion: The deep learning end-to-end approach is effective for material property estimation from image series, with accuracy stemming from ability to assess higher-order cognitive features like vorticity rather than just local displacement features.

Abstract: Contactless and non-invasive estimation of mechanical properties of physical media from optical observations is of interest for manifold engineering and biomedical applications, where direct physical measurements are not possible. Conventional approaches to the assessment of image displacement and non-contact material probing typically rely on time-consuming iterative algorithms for non-rigid image registration and constitutive modelling using discretization and iterative numerical solving techniques, such as Finite Element Method (FEM) and Finite Difference Method (FDM), which are not suitable for high-throughput data processing. Here, we present an efficient deep learning based end-to-end approach for the estimation of continuum displacement and material compressibility directly from the image series. Based on two deep neural networks for image registration and material compressibility estimation, this framework outperforms conventional approaches in terms of efficiency and accuracy. In particular, our experimental results show that the deep learning model trained on a set of reference data can accurately determine the material compressibility even in the presence of substantial local deviations of the mapping predicted by image registration from the reference displacement field. Our findings suggest that the remarkable accuracy of the deep learning end-to-end model originates from its ability to assess higher-order cognitive features, such as the vorticity of the vector field, rather than conventional local features of the image displacement.

Method: Operates in Bird’s Eye View space with Dynamic Mixture-of-Experts where each expert is dynamically generated based on specific agent’s input features to extract distinctive cues while attending to shared semantics, plus Dynamic Expert Metric Loss for inter-expert diversity.

Result: Achieves state-of-the-art performance: +1.5% IoU for Camera-based BEV segmentation on OPV2V and +3.0% AP@0.5 for LiDAR-based 3D object detection on DAIR-V2X-C datasets.

Conclusion: CoBEVMoE effectively models expert-based heterogeneous features in multi-agent collaborative perception, demonstrating the value of explicitly handling both feature similarity and heterogeneity across agents.

Abstract: Collaborative perception aims to extend sensing coverage and improve perception accuracy by sharing information among multiple agents. However, due to differences in viewpoints and spatial positions, agents often acquire heterogeneous observations. Existing intermediate fusion methods primarily focus on aligning similar features, often overlooking the perceptual diversity among agents. To address this limitation, we propose CoBEVMoE, a novel collaborative perception framework that operates in the Bird’s Eye View (BEV) space and incorporates a Dynamic Mixture-of-Experts (DMoE) architecture. In DMoE, each expert is dynamically generated based on the input features of a specific agent, enabling it to extract distinctive and reliable cues while attending to shared semantics. This design allows the fusion process to explicitly model both feature similarity and heterogeneity across agents. Furthermore, we introduce a Dynamic Expert Metric Loss (DEML) to enhance inter-expert diversity and improve the discriminability of the fused representation. Extensive experiments on the OPV2V and DAIR-V2X-C datasets demonstrate that CoBEVMoE achieves state-of-the-art performance. Specifically, it improves the IoU for Camera-based BEV segmentation by +1.5% on OPV2V and the AP@0.5 for LiDAR-based 3D object detection by +3.0% on DAIR-V2X-C, verifying the effectiveness of expert-based heterogeneous feature modeling in multi-agent collaborative perception. The source code will be made publicly available at https://github.com/godk0509/CoBEVMoE.

Method: Uses bidirectional reward-guided diffusion with reward feedback learning (ReFL). Early diffusion steps optimize on synthetic pairs for structural fidelity. Later steps apply quality-guided rewards to both synthetic and real images. Mitigates reward hacking via relative advantage space for synthetic results and semantic alignment constraint for real-world optimization. Uses dynamic fidelity-perception weighting strategy.

Result: Extensive experiments on real-world SR benchmarks show Bird-SR consistently outperforms state-of-the-art methods in perceptual quality while preserving structural consistency.

Conclusion: Bird-SR effectively addresses real-world super-resolution by balancing structural preservation and perceptual enhancement through a novel reward-guided diffusion framework.

Abstract: Diffusion-based super-resolution can synthesize rich details, but models trained on synthetic paired data often fail on real-world LR images due to distribution shifts. We propose Bird-SR, a bidirectional reward-guided diffusion framework that formulates super-resolution as trajectory-level preference optimization via reward feedback learning (ReFL), jointly leveraging synthetic LR-HR pairs and real-world LR images. For structural fidelity easily affected in ReFL, the model is directly optimized on synthetic pairs at early diffusion steps, which also facilitates structure preservation for real-world inputs under smaller distribution gap in structure levels. For perceptual enhancement, quality-guided rewards are applied at later sampling steps to both synthetic and real LR images. To mitigate reward hacking, the rewards for synthetic results are formulated in a relative advantage space bounded by their clean counterparts, while real-world optimization is regularized via a semantic alignment constraint. Furthermore, to balance structural and perceptual learning, we adopt a dynamic fidelity-perception weighting strategy that emphasizes structure preservation at early stages and progressively shifts focus toward perceptual optimization at later diffusion steps. Extensive experiments on real-world SR benchmarks demonstrate that Bird-SR consistently outperforms state-of-the-art methods in perceptual quality while preserving structural consistency, validating its effectiveness for real-world super-resolution.

Method: Inter-Slice Consistent Stochasticity (ISCS) controls the consistency of stochastic noise components during diffusion sampling across slices. It aligns sampling trajectories without adding loss terms or optimization steps, making it plug-and-play for any 2D diffusion-based 3D reconstruction pipeline.

Result: Experiments on multiple medical imaging problems show ISCS effectively improves 3D reconstruction performance from 2D diffusion models, reducing inter-slice discontinuities while maintaining image quality.

Conclusion: Controlling inter-slice stochasticity is a principled and practical approach for high-fidelity 3D medical imaging with 2D diffusion priors, offering a computationally efficient solution without additional training costs.

Abstract: 3D medical imaging is in high demand and essential for clinical diagnosis and scientific research. Currently, diffusion models (DMs) have become an effective tool for medical imaging reconstruction thanks to their ability to learn rich, high-quality data priors. However, learning the 3D data distribution with DMs in medical imaging is challenging, not only due to the difficulties in data collection but also because of the significant computational burden during model training. A common compromise is to train the DMs on 2D data priors and reconstruct stacked 2D slices to address 3D medical inverse problems. However, the intrinsic randomness of diffusion sampling causes severe inter-slice discontinuities of reconstructed 3D volumes. Existing methods often enforce continuity regularizations along the z-axis, which introduces sensitive hyper-parameters and may lead to over-smoothing results. In this work, we revisit the origin of stochasticity in diffusion sampling and introduce Inter-Slice Consistent Stochasticity (ISCS), a simple yet effective strategy that encourages interslice consistency during diffusion sampling. Our key idea is to control the consistency of stochastic noise components during diffusion sampling, thereby aligning their sampling trajectories without adding any new loss terms or optimization steps. Importantly, the proposed ISCS is plug-and-play and can be dropped into any 2D trained diffusion based 3D reconstruction pipeline without additional computational cost. Experiments on several medical imaging problems show that our method can effectively improve the performance of medical 3D imaging problems based on 2D diffusion models. Our findings suggest that controlling inter-slice stochasticity is a principled and practically attractive route toward high-fidelity 3D medical imaging with 2D diffusion priors. The code is available at: https://github.com/duchenhe/ISCS

Method: Proposes MosaicThinker, an inference-time computing technique that integrates fragmented spatial information from multiple frames into a unified global semantic map representation, then guides VLM’s spatial reasoning over this map via visual prompts.

Result: The technique greatly enhances accuracy of cross-frame spatial reasoning on resource-constrained embodied AI devices across diverse reasoning tasks with varying complexities.

Conclusion: MosaicThinker effectively addresses VLMs’ spatial reasoning limitations for embodied AI by creating unified spatial representations from multi-frame inputs, enabling better on-device performance for complex spatial reasoning tasks.

Abstract: When embodied AI is expanding from traditional object detection and recognition to more advanced tasks of robot manipulation and actuation planning, visual spatial reasoning from the video inputs is necessary to perceive the spatial relationships of objects and guide device actions. However, existing visual language models (VLMs) have very weak capabilities in spatial reasoning due to the lack of knowledge about 3D spatial information, especially when the reasoning task involve complex spatial relations across multiple video frames. In this paper, we present a new inference-time computing technique for on-device embodied AI, namely \emph{MosaicThinker}, which enhances the on-device small VLM’s spatial reasoning capabilities on difficult cross-frame reasoning tasks. Our basic idea is to integrate fragmented spatial information from multiple frames into a unified space representation of global semantic map, and further guide the VLM’s spatial reasoning over the semantic map via a visual prompt. Experiment results show that our technique can greatly enhance the accuracy of cross-frame spatial reasoning on resource-constrained embodied AI devices, over reasoning tasks with diverse types and complexities.

Method: Created WorldEdit dataset with high-quality editing samples using paraphrased instructions aligned with real-world causal logic, plus WorldEdit-Test for evaluation. Used two-stage training framework fine-tuning models like Bagel with causal verification reward.

Result: Significantly narrows performance gap with GPT-4o and Nano-Banana, showing competitive performance in both instruction following and knowledge plausibility where open-source systems typically struggle.

Conclusion: WorldEdit addresses the critical need for world-driven image editing capabilities, enabling models to handle implicit instructions requiring causal reasoning and complex world knowledge.

Abstract: Recent advances in image editing models have demonstrated remarkable capabilities in executing explicit instructions, such as attribute manipulation, style transfer, and pose synthesis. However, these models often face challenges when dealing with implicit editing instructions, which describe the cause of a visual change without explicitly detailing the resulting outcome. These limitations arise because existing models rely on uniform editing strategies that are not equipped to handle the complex world knowledge and reasoning required for implicit instructions. To address this gap, we introduce \textbf{WorldEdit}, a dataset specifically designed to enable world-driven image editing. WorldEdit consists of high-quality editing samples, guided by paraphrased instructions that align with real-world causal logic. Furthermore, we provide \textbf{WorldEdit-Test} for evaluating the existing model’s performance on causal editing scenarios. With WorldEdit, we use a two-stage training framework for fine-tuning models like Bagel, integrating with a causal verification reward. Our results show that the proposed dataset and methods significantly narrow the gap with GPT-4o and Nano-Banana, demonstrating competitive performance not only in instruction following but also in knowledge plausibility, where many open-source systems typically struggle.

Method: Proposes TLC-Plan, a hierarchical generative model using a two-level VQ-VAE: first level encodes global layouts as semantically labeled room bounding boxes, second level refines local geometries using polygon-level codes. This hierarchy is unified in a CodeTree representation, and an autoregressive transformer samples codes conditioned on the boundary to generate diverse and topologically valid designs without requiring explicit room topology or dimensional priors.

Result: Achieves state-of-the-art performance on RPLAN dataset (FID = 1.84, MSE = 2.06) and leading results on LIFULL dataset. The framework advances constraint-aware and scalable vector floorplan generation for real-world architectural applications.

Conclusion: TLC-Plan successfully addresses the limitations of raster-based approaches by directly generating vector floorplans through hierarchical compositional reasoning, demonstrating superior performance and practical applicability for architectural design.

Abstract: Automated floorplan generation aims to improve design quality, architectural efficiency, and sustainability by jointly modeling global spatial organization and precise geometric detail. However, existing approaches operate in raster space and rely on post hoc vectorization, which introduces structural inconsistencies and hinders end-to-end learning. Motivated by compositional spatial reasoning, we propose TLC-Plan, a hierarchical generative model that directly synthesizes vector floorplans from input boundaries, aligning with human architectural workflows based on modular and reusable patterns. TLC-Plan employs a two-level VQ-VAE to encode global layouts as semantically labeled room bounding boxes and to refine local geometries using polygon-level codes. This hierarchy is unified in a CodeTree representation, while an autoregressive transformer samples codes conditioned on the boundary to generate diverse and topologically valid designs, without requiring explicit room topology or dimensional priors. Extensive experiments show state-of-the-art performance on RPLAN dataset (FID = 1.84, MSE = 2.06) and leading results on LIFULL dataset. The proposed framework advances constraint-aware and scalable vector floorplan generation for real-world architectural applications. Source code and trained models are released at https://github.com/rosolose/TLC-PLAN.

Method: Proposes an end-to-end reinforcement learning framework with Relightable 3D Gaussian Splatting that decomposes scene components to enable explicit editing of environmental lighting. Trains policies in high-fidelity simulation with diverse synthesized lighting conditions to learn robust, illumination-invariant visual features.

Result: Lightweight quadrotor achieves robust, collision-free navigation in complex forest environments at speeds up to 10 m/s, exhibiting significant resilience to drastic lighting variations without fine-tuning.

Conclusion: The proposed framework enables effective zero-shot transfer to unstructured outdoor environments by addressing photometric limitations through relightable neural representations and diverse lighting augmentation during training.

Abstract: UAV navigation in unstructured outdoor environments using passive monocular vision is hindered by the substantial visual domain gap between simulation and reality. While 3D Gaussian Splatting enables photorealistic scene reconstruction from real-world data, existing methods inherently couple static lighting with geometry, severely limiting policy generalization to dynamic real-world illumination. In this paper, we propose a novel end-to-end reinforcement learning framework designed for effective zero-shot transfer to unstructured outdoors. Within a high-fidelity simulation grounded in real-world data, our policy is trained to map raw monocular RGB observations directly to continuous control commands. To overcome photometric limitations, we introduce Relightable 3D Gaussian Splatting, which decomposes scene components to enable explicit, physically grounded editing of environmental lighting within the neural representation. By augmenting training with diverse synthesized lighting conditions ranging from strong directional sunlight to diffuse overcast skies, we compel the policy to learn robust, illumination-invariant visual features. Extensive real-world experiments demonstrate that a lightweight quadrotor achieves robust, collision-free navigation in complex forest environments at speeds up to 10 m/s, exhibiting significant resilience to drastic lighting variations without fine-tuning.

Method: PI3D formulates and solves the problem of identifying effective 3D object poses (position and orientation) with injected text, ensuring physical plausibility while maximizing attack success

Result: PI3D effectively attacks multiple MLLMs under diverse camera trajectories, and existing defenses are insufficient against it

Conclusion: Physical prompt injection attacks in 3D environments pose a significant security threat to MLLMs, requiring new defense mechanisms beyond current approaches

Abstract: Multimodal large language models (MLLMs) have advanced the capabilities to interpret and act on visual input in 3D environments, empowering diverse applications such as robotics and situated conversational agents. When MLLMs reason over camera-captured views of the physical world, a new attack surface emerges: an attacker can place text-bearing physical objects in the environment to override MLLMs’ intended task. While prior work has studied prompt injection in the text domain and through digitally edited 2D images, it remains unclear how these attacks function in 3D physical environments. To bridge the gap, we introduce PI3D, a prompt injection attack against MLLMs in 3D environments, realized through text-bearing physical object placement rather than digital image edits. We formulate and solve the problem of identifying an effective 3D object pose (position and orientation) with injected text, where the attacker’s goal is to induce the MLLM to perform the injected task while ensuring that the object placement remains physically plausible. Experiments demonstrate that PI3D is an effective attack against multiple MLLMs under diverse camera trajectories. We further evaluate existing defenses and show that they are insufficient to defend against PI3D.

Method: Decouples semantic reasoning from temporal generation using speech units as temporal scaffolding. Introduces token-as-query gated fusion (TQGF) for controlled semantic injection. Creates InstructEx dataset to train OLLMs for speech-accompanied 3D facial animation.

Result: Ex-Omni performs competitively against existing open-source OLLMs while enabling stable aligned speech and facial animation generation.

Conclusion: Ex-Omni successfully extends OLLMs to incorporate speech with 3D facial animation, addressing the representation mismatch through decoupled learning and temporal scaffolding.

Abstract: Omni-modal large language models (OLLMs) aim to unify multimodal understanding and generation, yet incorporating speech with 3D facial animation remains largely unexplored despite its importance for natural interaction. A key challenge arises from the representation mismatch between discrete, token-level semantic reasoning in LLMs and the dense, fine-grained temporal dynamics required for 3D facial motion, which makes direct modeling difficult to optimize under limited data. We propose Expressive Omni (Ex-Omni), an open-source omni-modal framework that augments OLLMs with speech-accompanied 3D facial animation. Ex-Omni reduces learning difficulty by decoupling semantic reasoning from temporal generation, leveraging speech units as temporal scaffolding and a unified token-as-query gated fusion (TQGF) mechanism for controlled semantic injection. We further introduce InstructEx, a dataset aims to facilitate augment OLLMs with speech-accompanied 3D facial animation. Extensive experiments demonstrate that Ex-Omni performs competitively against existing open-source OLLMs while enabling stable aligned speech and facial animation generation.

Method: Used CLIP embedding similarity to systematically examine LAION-400M dataset, retrieving pregnancy ultrasound images and detecting thousands of entities containing private information like names and locations.

Result: Found multiple images with high-risk information enabling re-identification or impersonation, demonstrating significant privacy vulnerabilities in widely used public datasets.

Conclusion: Recommends improved practices for dataset curation, data privacy, and ethical use of public image datasets to address privacy risks in multimodal AI systems.

Abstract: The rise of generative models has led to increased use of large-scale datasets collected from the internet, often with minimal or no data curation. This raises concerns about the inclusion of sensitive or private information. In this work, we explore the presence of pregnancy ultrasound images, which contain sensitive personal information and are often shared online. Through a systematic examination of LAION-400M dataset using CLIP embedding similarity, we retrieve images containing pregnancy ultrasound and detect thousands of entities of private information such as names and locations. Our findings reveal that multiple images have high-risk information that could enable re-identification or impersonation. We conclude with recommended practices for dataset curation, data privacy, and ethical use of public image datasets.

Method: Proposes a dual meta-learning framework with: (1) meta-feature learning to extract age-agnostic semantic representations of evolving brain structures, (2) meta-initialization learning for data-efficient adaptation, and (3) memory-bank-based class-aware regularization for longitudinal consistency without explicit supervision.

Result: Experiments on diverse datasets (iSeg-2019, IBIS, OASIS, ADNI) under few-shot settings demonstrate that DuMeta++ outperforms existing methods in cross-age generalization.

Conclusion: DuMeta++ effectively addresses the challenge of brain tissue segmentation across the lifespan without requiring paired longitudinal data, achieving superior cross-age generalization through its dual meta-learning approach.

Abstract: Accurate segmentation of brain tissues from MRI scans is critical for neuroscience and clinical applications, but achieving consistent performance across the human lifespan remains challenging due to dynamic, age-related changes in brain appearance and morphology. While prior work has sought to mitigate these shifts by using self-supervised regularization with paired longitudinal data, such data are often unavailable in practice. To address this, we propose \emph{DuMeta++}, a dual meta-learning framework that operates without paired longitudinal data. Our approach integrates: (1) meta-feature learning to extract age-agnostic semantic representations of spatiotemporally evolving brain structures, and (2) meta-initialization learning to enable data-efficient adaptation of the segmentation model. Furthermore, we propose a memory-bank-based class-aware regularization strategy to enforce longitudinal consistency without explicit longitudinal supervision. We theoretically prove the convergence of our DuMeta++, ensuring stability. Experiments on diverse datasets (iSeg-2019, IBIS, OASIS, ADNI) under few-shot settings demonstrate that DuMeta++ outperforms existing methods in cross-age generalization. Code will be available at https://github.com/ladderlab-xjtu/DuMeta++.

Method: Uses view-invariant semantic features as conditioning input instead of view angles. Creates synthesized head image dataset using FLUX.1 Kontext to extend frontal face datasets to multiple views. Uses frontal view image clip features as shared semantic condition across all views, enabling semantic alignment and eliminating directional bias.

Result: Achieves significantly higher fidelity, diversity, and generalizability in full-head synthesis and single-view GAN inversion compared to previous methods. Improves global coherence of generated 3D heads and accelerates training.

Conclusion: Semantic conditioning decouples generative capability from viewing direction, promotes continuous learning, and enhances 3D head generation quality and diversity by following true semantic distribution rather than view-dependent patterns.

Abstract: Conditioning is crucial for stable training of full-head 3D GANs. Without any conditioning signal, the model suffers from severe mode collapse, making it impractical to training. However, a series of previous full-head 3D GANs conventionally choose the view angle as the conditioning input, which leads to a bias in the learned 3D full-head space along the conditional view direction. This is evident in the significant differences in generation quality and diversity between the conditional view and non-conditional views of the generated 3D heads, resulting in global incoherence across different head regions. In this work, we propose to use view-invariant semantic feature as the conditioning input, thereby decoupling the generative capability of 3D heads from the viewing direction. To construct a view-invariant semantic condition for each training image, we create a novel synthesized head image dataset. We leverage FLUX.1 Kontext to extend existing high-quality frontal face datasets to a wide range of view angles. The image clip feature extracted from the frontal view is then used as a shared semantic condition across all views in the extended images, ensuring semantic alignment while eliminating directional bias. This also allows supervision from different views of the same subject to be consolidated under a shared semantic condition, which accelerates training and enhances the global coherence of the generated 3D heads. Moreover, as GANs often experience slower improvements in diversity once the generator learns a few modes that successfully fool the discriminator, our semantic conditioning encourages the generator to follow the true semantic distribution, thereby promoting continuous learning and diverse generation. Extensive experiments on full-head synthesis and single-view GAN inversion demonstrate that our method achieves significantly higher fidelity, diversity, and generalizability.

Method: Created RoadSafe365 with hierarchical taxonomy refining crash, incident, and violation definitions; annotated 36,196 clips from dashcam/surveillance cameras with multiple-choice QA sets, scene descriptions, and diverse attributes.

Result: Established strong baselines with consistent gains when fine-tuning on RoadSafe365; cross-domain experiments on real and synthetic datasets validated effectiveness; created comprehensive benchmark with 864K candidate options and 8.4K unique answers.

Conclusion: RoadSafe365 provides a comprehensive benchmark for reproducible research in real-world traffic safety analysis, designed for large-scale training and standardized evaluation.

Abstract: Although recent traffic benchmarks have advanced multimodal data analysis, they generally lack systematic evaluation aligned with official safety standards. To fill this gap, we introduce RoadSafe365, a large-scale vision-language benchmark that supports fine-grained analysis of traffic safety from extensive and diverse real-world video data collections. Unlike prior works that focus primarily on coarse accident identification, RoadSafe365 is independently curated and systematically organized using a hierarchical taxonomy that refines and extends foundational definitions of crash, incident, and violation to bridge official traffic safety standards with data-driven traffic understanding systems. RoadSafe365 provides rich attribute annotations across diverse traffic event types, environmental contexts, and interaction scenarios, yielding 36,196 annotated clips from both dashcam and surveillance cameras. Each clip is paired with multiple-choice question-answer sets, comprising 864K candidate options, 8.4K unique answers, and 36K detailed scene descriptions collectively designed for vision-language understanding and reasoning. We establish strong baselines and observe consistent gains when fine-tuning on RoadSafe365. Cross-domain experiments on both real and synthetic datasets further validate its effectiveness. Designed for large-scale training and standardized evaluation, RoadSafe365 provides a comprehensive benchmark to advance reproducible research in real-world traffic safety analysis.

Method: AdvSR framework jointly optimizes for both reconstruction quality and targeted adversarial outcomes during training. The approach embeds adversarial behavior directly into SR model weights, requiring no access to inputs at inference time. The method was evaluated on three SR architectures (SRCNN, EDSR, SwinIR) paired with a YOLOv11 classifier.

Result: AdvSR models achieve high attack success rates with minimal quality degradation. The models appear benign under standard image quality metrics while effectively inducing downstream misclassification in the paired classifier.

Conclusion: The work reveals a new model-level threat for imaging pipelines that use SR preprocessing, highlighting security implications for how practitioners source and validate models in safety-critical applications.

Abstract: Data-driven super-resolution (SR) methods are often integrated into imaging pipelines as preprocessing steps to improve downstream tasks such as classification and detection. However, these SR models introduce a previously unexplored attack surface into imaging pipelines. In this paper, we present AdvSR, a framework demonstrating that adversarial behavior can be embedded directly into SR model weights during training, requiring no access to inputs at inference time. Unlike prior attacks that perturb inputs or rely on backdoor triggers, AdvSR operates entirely at the model level. By jointly optimizing for reconstruction quality and targeted adversarial outcomes, AdvSR produces models that appear benign under standard image quality metrics while inducing downstream misclassification. We evaluate AdvSR on three SR architectures (SRCNN, EDSR, SwinIR) paired with a YOLOv11 classifier and demonstrate that AdvSR models can achieve high attack success rates with minimal quality degradation. These findings highlight a new model-level threat for imaging pipelines, with implications for how practitioners source and validate models in safety-critical applications.

Method: Developed 3D-TBM framework with data preprocessing, computation of optimal transport embeddings, and analytical methods including visualization of main transport directions and techniques for discerning discriminating directions. The tool uses invertible transformations to embed images into a transport domain for analysis.

Result: Created a comprehensive tool for 3D medical image analysis with publicly available source code through PyTransKit, including documentation and practical tutorials to support researchers in applying TBM to their medical imaging studies.

Conclusion: 3D-TBM provides an accessible framework for transport-based morphometry in 3D medical imaging, enabling effective analysis with interpretable spatial results that can be projected back to the original image space for clinical interpretation.

Abstract: Transport-Based Morphometry (TBM) has emerged as a new framework for 3D medical image analysis. By embedding images into a transport domain via invertible transformations, TBM facilitates effective classification, regression, and other tasks using transport-domain features. Crucially, the inverse mapping enables the projection of analytic results back into the original image space, allowing researchers to directly interpret clinical features associated with model outputs in a spatially meaningful way. To facilitate broader adoption of TBM in clinical imaging research, we present 3D-TBM, a tool designed for morphological analysis of 3D medical images. The framework includes data preprocessing, computation of optimal transport embeddings, and analytical methods such as visualization of main transport directions, together with techniques for discerning discriminating directions and related analysis methods. We also provide comprehensive documentation and practical tutorials to support researchers interested in applying 3D-TBM in their own medical imaging studies. The source code is publicly available through PyTransKit.

Method: Introduces TwistNet-2D with Spiral-Twisted Channel Interaction (STCI) that shifts one feature map along prescribed directions before element-wise channel multiplication. This captures cross-position co-occurrence patterns. Uses four directional heads with learned channel reweighting and injects results through a sigmoid-gated residual path.

Result: TwistNet-2D adds only 3.5% additional parameters and 2% additional FLOPs over ResNet-18, yet consistently surpasses both parameter-matched and substantially larger baselines (ConvNeXt, Swin Transformer, hybrid CNN-Transformer architectures) across four texture and fine-grained recognition benchmarks.

Conclusion: The proposed method effectively captures local pairwise channel interactions with directional spatial displacement, providing an efficient and effective solution for texture and fine-grained recognition tasks.

Abstract: Second-order feature statistics are central to texture recognition, yet current methods face a fundamental tension: bilinear pooling and Gram matrices capture global channel correlations but collapse spatial structure, while self-attention models spatial context through weighted aggregation rather than explicit pairwise feature interactions. We introduce TwistNet-2D, a lightweight module that computes \emph{local} pairwise channel products under directional spatial displacement, jointly encoding where features co-occur and how they interact. The core component, Spiral-Twisted Channel Interaction (STCI), shifts one feature map along a prescribed direction before element-wise channel multiplication, thereby capturing the cross-position co-occurrence patterns characteristic of structured and periodic textures. Aggregating four directional heads with learned channel reweighting and injecting the result through a sigmoid-gated residual path, \TwistNet incurs only 3.5% additional parameters and 2% additional FLOPs over ResNet-18, yet consistently surpasses both parameter-matched and substantially larger baselines – including ConvNeXt, Swin Transformer, and hybrid CNN–Transformer architectures – across four texture and fine-grained recognition benchmarks.

Method: Two-stage pipeline: 1) Finetune large video model (Wan 2.1 14B) to generate material sample videos under controlled camera/lighting trajectories (virtual gonioreflectometer), 2) Reconstruct compact neural materials from videos using Large Reconstruction Model (LRM) finetuned from smaller Wan 1.3B video backbone, performing single-pass inference from 17 frames.

Result: The resulting materials exhibit realism and diversity far exceeding limited synthetic training data, demonstrating successful transfer of material knowledge from internet-scale video models into standalone, reusable neural 3D assets.

Conclusion: Material knowledge can be successfully transferred from internet-scale video models into reusable neural 3D assets, enabling creation of photorealistic materials without requiring extensive artistic skill or large training datasets.

Abstract: Creating photorealistic materials for 3D rendering requires exceptional artistic skill. Generative models for materials could help, but are currently limited by the lack of high-quality training data. While recent video generative models effortlessly produce realistic material appearances, this knowledge remains entangled with geometry and lighting. We present VideoNeuMat, a two-stage pipeline that extracts reusable neural material assets from video diffusion models. First, we finetune a large video model (Wan 2.1 14B) to generate material sample videos under controlled camera and lighting trajectories, effectively creating a “virtual gonioreflectometer” that preserves the model’s material realism while learning a structured measurement pattern. Second, we reconstruct compact neural materials from these videos through a Large Reconstruction Model (LRM) finetuned from a smaller Wan 1.3B video backbone. From 17 generated video frames, our LRM performs single-pass inference to predict neural material parameters that generalize to novel viewing and lighting conditions. The resulting materials exhibit realism and diversity far exceeding the limited synthetic training data, demonstrating that material knowledge can be successfully transferred from internet-scale video models into standalone, reusable neural 3D assets.

Method: XVWM uses cross-view prediction objective: given frames from one viewpoint, predict future state from same or different viewpoint after an action. Trained on synchronized multi-view gameplay data from Aimlabs with precisely aligned multi-camera recordings and high-frequency action labels.

Result: The model learns view-invariant representations of 3D structure through cross-view consistency, provides agents with parallel imagination streams across viewpoints, and enables planning in optimal reference frames while executing from egocentric view.

Conclusion: Multi-view consistency provides strong learning signal for spatially grounded representations, and predicting consequences from other viewpoints may offer foundation for perspective-taking in multi-agent settings.

Abstract: World models enable agents to plan by imagining future states, but existing approaches operate from a single viewpoint, typically egocentric, even when other perspectives would make planning easier; navigation, for instance, benefits from a bird’s-eye view. We introduce Cross-View World Models (XVWM), trained with a cross-view prediction objective: given a sequence of frames from one viewpoint, predict the future state from the same or a different viewpoint after an action is taken. Enforcing cross-view consistency acts as geometric regularization: because the input and output views may share little or no visual overlap, to predict across viewpoints, the model must learn view-invariant representations of the environment’s 3D structure. We train on synchronized multi-view gameplay data from Aimlabs, an aim-training platform providing precisely aligned multi-camera recordings with high-frequency action labels. The resulting model gives agents parallel imagination streams across viewpoints, enabling planning in whichever frame of reference best suits the task while executing from the egocentric view. Our results show that multi-view consistency provides a strong learning signal for spatially grounded representations. Finally, predicting the consequences of one’s actions from another viewpoint may offer a foundation for perspective-taking in multi-agent settings.

Method: Used Attention mechanism integrated with DeepLab-V3+ architecture for segmenting four DR lesion types (microaneurysms, soft exudates, hard exudates, hemorrhages) on 757 images from DDR dataset.

Result: Attention-DeepLab increased mAP from 0.3010 to 0.3326 and mean IoU from 0.1791 to 0.1928. Most notably improved microaneurysm detection from 0.0205 to 0.0763, which is clinically significant as microaneurysms are the earliest visible DR symptom.

Conclusion: The attention-enhanced model provides better lesion segmentation for DR screening, with particular improvement in detecting early-stage microaneurysms, supporting clinical applications.

Abstract: Diabetic Retinopathy (DR) is an eye disease which arises due to diabetes mellitus. It might cause vision loss and blindness. To prevent irreversible vision loss, early detection through systematic screening is crucial. Although researchers have developed numerous automated deep learning-based algorithms for DR screening, their clinical applicability remains limited, particularly in lesion segmentation. Our method provides pixel-level annotations for lesions, which practically supports Ophthalmologist to screen DR from fundus images. In this work, we segmented four types of DR-related lesions: microaneurysms, soft exudates, hard exudates, and hemorrhages on 757 images from DDR dataset. To enhance lesion segmentation, an attention mechanism was integrated with DeepLab-V3+. Compared to the baseline model, the Attention-DeepLab model increases mean average precision (mAP) from 0.3010 to 0.3326 and the mean Intersection over Union (IoU) from 0.1791 to 0.1928. The model also increased microaneurysm detection from 0.0205 to 0.0763, a clinically significant improvement. The detection of microaneurysms is the earliest visible symptom of DR.

Method: Developed a filtering and segmentation algorithm optimized using linear genetic programming (LGP) with a domain-specific language for image processing. The system generates human-interpretable MATLAB code representing image filtering pipelines with tunable parameters.

Result: Achieved near-human accuracy with 1.8% average evaluation error compared to human baseline, processing 3.6 megapixel images in about 2 seconds, enabling faster iteration cycles in material development.

Conclusion: The automated pipeline accelerates exploration of material composition and processing space, facilitating development of strong, low-activation precipitation hardened copper alloys for additive manufactured fusion reactor parts.

Abstract: Current analysis of additive manufactured niobium-based copper alloys relies on hand annotation due to varying contrast, noise, and image artifacts present in micrographs, slowing iteration speed in alloy development. We present a filtering and segmentation algorithm for detecting precipitates in FIB cross-section micrographs, optimized using linear genetic programming (LGP), which accounts for the various artifacts. To this end, the optimization environment uses a domain-specific language for image processing to iterate on solutions. Programs in this language are a list of image-filtering blocks with tunable parameters that sequentially process an input image, allowing for reliable generation and mutation by a genetic algorithm. Our environment produces optimized human-interpretable MATLAB code representing an image filtering pipeline. Under ideal conditions–a population size of 60 and a maximum program length of 5 blocks–our system was able to find a near-human accuracy solution with an average evaluation error of 1.8% when comparing segmentations pixel-by-pixel to a human baseline using an XOR error evaluation. Our automation work enabled faster iteration cycles and furthered exploration of the material composition and processing space: our optimized pipeline algorithm processes a 3.6 megapixel image in about 2 seconds on average. This ultimately enables convergence on strong, low-activation, precipitation hardened copper alloys for additive manufactured fusion reactor parts.

Method: LUCID learns a shared latent dictionary for image patch and text token representations while reserving private capacity for modality-specific details. It achieves feature alignment through optimal transport matching without labeling, and includes automated dictionary interpretation via term clustering.

Result: LUCID produces interpretable shared features that support patch-level grounding, establish cross-modal neuron correspondence, enhance robustness against concept clustering problems, and capture diverse semantic categories including objects, actions, attributes, and abstract concepts.

Conclusion: LUCID demonstrates a comprehensive approach to interpretable multimodal representations, enabling automated concept discovery and cross-modal alignment without manual intervention, advancing the field of multimodal interpretability.

Abstract: Sparse autoencoders (SAEs) offer a natural path toward comparable explanations across different representation spaces. However, current SAEs are trained per modality, producing dictionaries whose features are not directly understandable and whose explanations do not transfer across domains. In this study, we introduce LUCID (Learning Unified vision-language sparse Codes for Interpretable concept Discovery), a unified vision-language sparse autoencoder that learns a shared latent dictionary for image patch and text token representations, while reserving private capacity for modality-specific details. We achieve feature alignment by coupling the shared codes with a learned optimal transport matching objective without the need of labeling. LUCID yields interpretable shared features that support patch-level grounding, establish cross-modal neuron correspondence, and enhance robustness against the concept clustering problem in similarity-based evaluation. Leveraging the alignment properties, we develop an automated dictionary interpretation pipeline based on term clustering without manual observations. Our analysis reveals that LUCID’s shared features capture diverse semantic categories beyond objects, including actions, attributes, and abstract concepts, demonstrating a comprehensive approach to interpretable multimodal representations.

Method: Proposes CLARITY that dynamically adapts fusion strategy using vision-language model (VLM) priors to modulate each modality’s contribution based on detected scene conditions. Includes mechanisms to preserve valid dark-object semantics and a hierarchical decoder for structural consistency across scales.

Result: Achieves new state-of-the-art on MFNet dataset with 62.3% mIoU and 77.5% mAcc, demonstrating superior performance in adverse lighting conditions.

Conclusion: Dynamic fusion guided by VLM priors outperforms static fusion strategies for RGB-Thermal semantic segmentation in challenging illumination conditions, with applications for autonomous driving.

Abstract: Robust semantic segmentation of road scenes under adverse illumination, lighting, and shadow conditions remain a core challenge for autonomous driving applications. RGB-Thermal fusion is a standard approach, yet existing methods apply static fusion strategies uniformly across all conditions, allowing modality-specific noise to propagate throughout the network. Hence, we propose CLARITY that dynamically adapts its fusion strategy to the detected scene condition. Guided by vision-language model (VLM) priors, the network learns to modulate each modality’s contribution based on the illumination state while leveraging object embeddings for segmentation, rather than applying a fixed fusion policy. We further introduce two mechanisms, i.e., one which preserves valid dark-object semantics that prior noise-suppression methods incorrectly discard, and a hierarchical decoder that enforces structural consistency across scales to sharpen boundaries on thin objects. Experiments on the MFNet dataset demonstrate that CLARITY establishes a new state-of-the-art (SOTA), achieving 62.3% mIoU and 77.5% mAcc.

Method: Proposes Adaptive Matching Distillation (AMD) with: 1) Reward proxies to explicitly detect Forbidden Zones, 2) Structural signal decomposition to prioritize corrective gradients, and 3) Repulsive Landscape Sharpening to enforce steep energy barriers against failure mode collapse.

Result: AMD significantly enhances sample fidelity and training robustness across image and video generation tasks (SDXL, Wan2.1). Improves HPSv2 score on SDXL from 30.64 to 31.25, outperforming state-of-the-art baselines on benchmarks like VBench and GenEval.

Conclusion: Explicitly rectifying optimization trajectories within Forbidden Zones is essential for pushing the performance ceiling of few-step generative models, and AMD provides an effective self-correcting mechanism for stable distillation.

Abstract: Distribution Matching Distillation (DMD) is a powerful acceleration paradigm, yet its stability is often compromised in Forbidden Zone, regions where the real teacher provides unreliable guidance while the fake teacher exerts insufficient repulsive force. In this work, we propose a unified optimization framework that reinterprets prior art as implicit strategies to avoid these corrupted regions. Based on this insight, we introduce Adaptive Matching Distillation (AMD), a self-correcting mechanism that utilizes reward proxies to explicitly detect and escape Forbidden Zones. AMD dynamically prioritizes corrective gradients via structural signal decomposition and introduces Repulsive Landscape Sharpening to enforce steep energy barriers against failure mode collapse. Extensive experiments across image and video generation tasks (e.g., SDXL, Wan2.1) and rigorous benchmarks (e.g., VBench, GenEval) demonstrate that AMD significantly enhances sample fidelity and training robustness. For instance, AMD improves the HPSv2 score on SDXL from 30.64 to 31.25, outperforming state-of-the-art baselines. These findings validate that explicitly rectifying optimization trajectories within Forbidden Zones is essential for pushing the performance ceiling of few-step generative models.

Method: Improved U-Net with Row-Column Separated Attention (RCSA) module that takes row/column means and maxes of feature maps to guide local information with fewer parameters. Two temporal loss functions for video consistency.

Result: Extensive experiments on LOL, MIT Adobe FiveK image datasets and SDSD video dataset demonstrate effectiveness of the approach.

Conclusion: URCSA effectively enhances low-light images/videos with better global guidance and temporal consistency while maintaining computational efficiency.

Abstract: U-Net structure is widely used for low-light image/video enhancement. The enhanced images result in areas with large local noise and loss of more details without proper guidance for global information. Attention mechanisms can better focus on and use global information. However, attention to images could significantly increase the number of parameters and computations. We propose a Row-Column Separated Attention module (RCSA) inserted after an improved U-Net. The RCSA module’s input is the mean and maximum of the row and column of the feature map, which utilizes global information to guide local information with fewer parameters. We propose two temporal loss functions to apply the method to low-light video enhancement and maintain temporal consistency. Extensive experiments on the LOL, MIT Adobe FiveK image, and SDSD video datasets demonstrate the effectiveness of our approach. The code is publicly available at https://github.com/cq-dong/URCSA.

Method: Proposes a perspective-aware log-depth fusion approach that extends orthographic gradient-based methods by explicitly modeling perspective projection. Uses log-depth representation and handles missing depth measurements by leveraging surface normal information for inpainting.

Result: Experiments on DiLiGenT-MV dataset demonstrate effectiveness and show importance of perspective-aware depth-normals fusion for metrically accurate 3D reconstructions.

Conclusion: The perspective-aware approach significantly improves 3D reconstruction accuracy over orthographic methods, especially for perspective camera systems, and effectively handles missing depth data through normal-based inpainting.

Abstract: We address the problem of reconstructing 3D surfaces from depth and surface normal maps acquired by a sensor system based on a single perspective camera. Depth and normal maps can be obtained through techniques such as structured-light scanning and photometric stereo, respectively. We propose a perspective-aware log-depth fusion approach that extends existing orthographic gradient-based depth-normals fusion methods by explicitly accounting for perspective projection, leading to metrically accurate 3D reconstructions. Additionally, the method handles missing depth measurements by leveraging available surface normal information to inpaint gaps. Experiments on the DiLiGenT-MV data set demonstrate the effectiveness of our approach and highlight the importance of perspective-aware depth-normals fusion.

Method: Created a synthetic ECG image dataset from PTB-XL signal database using an open-source Python framework with customizable parameters (paper speed, voltage scale, sampling rate, grid appearance, waveform characteristics).

Result: Dataset contains 17,271 high-quality ECG images with five complementary data types per sample: realistic images, segmentation masks, time-series signals, bounding boxes, and comprehensive metadata. Achieved 100% generation success rate with 1.35 seconds per sample.

Conclusion: PTB-XL-Image-17K addresses critical gaps in ECG digitization research by providing the first large-scale resource supporting the complete pipeline from lead detection to signal extraction with full ground truth for evaluation.

Abstract: Electrocardiogram (ECG) digitization-converting paper-based or scanned ECG images back into time-series signals-is critical for leveraging decades of legacy clinical data in modern deep learning applications. However, progress has been hindered by the lack of large-scale datasets providing both ECG images and their corresponding ground truth signals with comprehensive annotations. We introduce PTB-XL-Image-17K, a complete synthetic ECG image dataset comprising 17,271 high-quality 12-lead ECG images generated from the PTB-XL signal database. Our dataset uniquely provides five complementary data types per sample: (1) realistic ECG images with authentic grid patterns and annotations (50% with visible grid, 50% without), (2) pixel-level segmentation masks, (3) ground truth time-series signals, (4) bounding box annotations in YOLO format for both lead regions and lead name labels, and (5) comprehensive metadata including visual parameters and patient information. We present an open-source Python framework enabling customizable dataset generation with controllable parameters including paper speed (25/50 mm/s), voltage scale (5/10 mm/mV), sampling rate (500 Hz), grid appearance (4 colors), and waveform characteristics. The dataset achieves 100% generation success rate with an average processing time of 1.35 seconds per sample. PTB-XL-Image-17K addresses critical gaps in ECG digitization research by providing the first large-scale resource supporting the complete pipeline: lead detection, waveform segmentation, and signal extraction with full ground truth for rigorous evaluation. The dataset, generation framework, and documentation are publicly available at https://github.com/naqchoalimehdi/PTB-XL-Image-17K and https://doi.org/10.5281/zenodo.18197519.

Method: Proposes a unified 1.3B-parameter framework with Streaming-Aware Spatiotemporal Pre-training using Temporal Audio Context Cache for robust feature extraction from short audio fragments, and Oracle-Guided Bidirectional Distillation to mitigate error accumulation in long-sequence autoregressive generation

Result: Achieves state-of-the-art performance on HDTF and VFHQ benchmarks, with Lite variant reaching 96 FPS on single NVIDIA RTX 4090, enabling ultra-fast interaction without sacrificing visual coherence

Conclusion: SoulX-FlashHead successfully addresses the trade-off between quality and latency in audio-driven portrait generation, providing a practical solution for real-time streaming applications with high visual fidelity

Abstract: Achieving a balance between high-fidelity visual quality and low-latency streaming remains a formidable challenge in audio-driven portrait generation. Existing large-scale models often suffer from prohibitive computational costs, while lightweight alternatives typically compromise on holistic facial representations and temporal stability. In this paper, we propose SoulX-FlashHead, a unified 1.3B-parameter framework designed for real-time, infinite-length, and high-fidelity streaming video generation. To address the instability of audio features in streaming scenarios, we introduce Streaming-Aware Spatiotemporal Pre-training equipped with a Temporal Audio Context Cache mechanism, which ensures robust feature extraction from short audio fragments. Furthermore, to mitigate the error accumulation and identity drift inherent in long-sequence autoregressive generation, we propose Oracle-Guided Bidirectional Distillation, leveraging ground-truth motion priors to provide precise physical guidance. We also present VividHead, a large-scale, high-quality dataset containing 782 hours of strictly aligned footage to support robust training. Extensive experiments demonstrate that SoulX-FlashHead achieves state-of-the-art performance on HDTF and VFHQ benchmarks. Notably, our Lite variant achieves an inference speed of 96 FPS on a single NVIDIA RTX 4090, facilitating ultra-fast interaction without sacrificing visual coherence.

Method: Proposes SpatialReward, a reward model that enforces precise verification via explicit spatial reasoning by anchoring reasoning to predicted edit regions, grounding semantic judgments in pixel-level evidence. Trained on a curated 260k spatial-aware dataset.

Result: Achieves SOTA performance on MMRB2 and EditReward-Bench, outperforms proprietary evaluators on MultiEditReward-Bench. Boosts OmniGen2 by +0.90 on GEdit-Bench, surpassing leading discriminative model and doubling GPT-4.1’s gain (+0.45).

Conclusion: Spatial reasoning is essential for unlocking effective alignment in image editing, and SpatialReward demonstrates superior performance as a robust reward signal for online RL in image editing tasks.

Abstract: Online Reinforcement Learning (RL) offers a promising avenue for complex image editing but is currently constrained by the scarcity of reliable and fine-grained reward signals. Existing evaluators frequently struggle with a critical perception gap we term “Attention Collapse,” where models neglect cross-image comparisons and fail to capture fine-grained details, resulting in inaccurate perception and miscalibrated scores. To address these limitations, we propose SpatialReward, a reward model that enforces precise verification via explicit spatial reasoning. By anchoring reasoning to predicted edit regions, SpatialReward grounds semantic judgments in pixel-level evidence, significantly enhancing evaluative accuracy. Trained on a curated 260k spatial-aware dataset, our model achieves state-of-the-art performance on MMRB2 and EditReward-Bench, and outperforms proprietary evaluators on our proposed MultiEditReward-Bench. Furthermore, SpatialReward serves as a robust signal in online RL, boosting OmniGen2 by +0.90 on GEdit-Bench–surpassing the leading discriminative model and doubling the gain of GPT-4.1 (+0.45). These results demonstrate that spatial reasoning is essential for unlocking effective alignment in image editing.

Method: Created the GlobalWasteData (GWD) archive by merging multiple publicly available waste classification datasets into a single unified resource, with additional preprocessing steps including quality filtering, duplicate removal, and metadata generation to improve dataset reliability.

Result: The resulting GWD archive contains 89,807 images across 14 main categories annotated with 68 distinct subclasses, offering consistent labeling, improved domain diversity, and more balanced class representation compared to existing fragmented datasets.

Conclusion: GWD provides a strong foundation for ML applications in environmental monitoring, recycling automation, and waste identification, and is publicly available to promote future research and reproducibility in waste classification.

Abstract: The growing amount of waste is a problem for the environment that requires efficient sorting techniques for various kinds of waste. An automated waste classification system is used for this purpose. The effectiveness of these Artificial Intelligence (AI) models depends on the quality and accessibility of publicly available datasets, which provide the basis for training and analyzing classification algorithms. Although several public waste classification datasets exist, they remain fragmented, inconsistent, and biased toward specific environments. Differences in class names, annotation formats, image conditions, and class distributions make it difficult to combine these datasets or train models that generalize well to real world scenarios. To address these issues, we introduce the GlobalWasteData (GWD) archive, a large scale dataset of 89,807 images across 14 main categories, annotated with 68 distinct subclasses. We compile this novel integrated GWD archive by merging multiple publicly available datasets into a single, unified resource. This GWD archive offers consistent labeling, improved domain diversity, and more balanced class representation, enabling the development of robust and generalizable waste recognition models. Additional preprocessing steps such as quality filtering, duplicate removal, and metadata generation further improve dataset reliability. Overall, this dataset offers a strong foundation for Machine Learning (ML) applications in environmental monitoring, recycling automation, and waste identification, and is publicly available to promote future research and reproducibility.

Method: Proposes TOM-GS that integrates learning-based odometry with Gaussian Splatting-based dense mapping. Features dedicated thermal image enhancement and monocular depth integration. Among first GS-based SLAM systems tailored for thermal cameras.

Result: Extensive experiments on motion estimation and novel-view rendering demonstrate TOM-GS outperforms existing learning-based methods, confirming benefits of learning-based pipelines for robust thermal odometry and dense reconstruction.

Conclusion: TOM-GS successfully combines learning-based odometry with Gaussian Splatting for thermal cameras, achieving superior performance in motion estimation and dense reconstruction compared to existing methods.

Abstract: Thermal infrared sensors, with wavelengths longer than smoke particles, can capture imagery independent of darkness, dust, and smoke. This robustness has made them increasingly valuable for motion estimation and environmental perception in robotics, particularly in adverse conditions. Existing thermal odometry and mapping approaches, however, are predominantly geometric and often fail across diverse datasets while lacking the ability to produce dense maps. Motivated by the efficiency and high-quality reconstruction ability of recent Gaussian Splatting (GS) techniques, we propose TOM-GS, a thermal odometry and mapping method that integrates learning-based odometry with GS-based dense mapping. TOM-GS is among the first GS-based SLAM systems tailored for thermal cameras, featuring dedicated thermal image enhancement and monocular depth integration. Extensive experiments on motion estimation and novel-view rendering demonstrate that TOM-GS outperforms existing learning-based methods, confirming the benefits of learning-based pipelines for robust thermal odometry and dense reconstruction.

Method: Proposes implicit motion representation compressing per-frame motion into compact 1D motion tokens, relaxing spatial constraints and preventing identity leakage. Includes temporally consistent mask token-based retargeting module with temporal training bottleneck to mitigate source motion interference and improve consistency. Uses three-stage training strategy for efficiency and fidelity.

Result: Extensive experiments demonstrate superior or competitive performance compared to state-of-the-art methods, showing effective generative capabilities for character animation.

Conclusion: The proposed implicit motion representation and IM-Animation framework effectively address limitations of both explicit and implicit methods, achieving high-quality character animation with improved identity preservation and motion consistency.

Abstract: Recent progress in video diffusion models has markedly advanced character animation, which synthesizes motioned videos by animating a static identity image according to a driving video. Explicit methods represent motion using skeleton, DWPose or other explicit structured signals, but struggle to handle spatial mismatches and varying body scales. %proportions. Implicit methods, on the other hand, capture high-level implicit motion semantics directly from the driving video, but suffer from identity leakage and entanglement between motion and appearance. To address the above challenges, we propose a novel implicit motion representation that compresses per-frame motion into compact 1D motion tokens. This design relaxes strict spatial constraints inherent in 2D representations and effectively prevents identity information leakage from the motion video. Furthermore, we design a temporally consistent mask token-based retargeting module that enforces a temporal training bottleneck, mitigating interference from the source images’ motion and improving retargeting consistency. Our methodology employs a three-stage training strategy to enhance the training efficiency and ensure high fidelity. Extensive experiments demonstrate that our implicit motion representation and the propose IM-Animation’s generative capabilities are achieve superior or competitive performance compared with state-of-the-art methods.

Method: Two-stage approach: 1) Lightweight offset prediction with box-based zooming objective for non-uniform zooming, 2) Corner-aligned bounding box transformation to warp ground-truth boxes to zoomed space for training and predicted boxes back to original space for inference.

Result: Achieves significant improvements on UAV datasets: >8.4 mAP absolute gain on SeaDronesSee with Faster R-CNN, only ~3ms additional latency. Architecture-independent and works with any detector.

Conclusion: ZoomDet effectively addresses small object detection in UAV images through adaptive zooming, offering substantial accuracy gains with minimal computational overhead.

Abstract: Detecting objects from UAV-captured images is challenging due to the small object size. In this work, a simple and efficient adaptive zoom-in framework is explored for object detection on UAV images. The main motivation is that the foreground objects are generally smaller and sparser than those in common scene images, which hinders the optimization of effective object detectors. We thus aim to zoom in adaptively on the objects to better capture object features for the detection task. To achieve the goal, two core designs are required: \textcolor{black}{i) How to conduct non-uniform zooming on each image efficiently? ii) How to enable object detection training and inference with the zoomed image space?} Correspondingly, a lightweight offset prediction scheme coupled with a novel box-based zooming objective is introduced to learn non-uniform zooming on the input image. Based on the learned zooming transformation, a corner-aligned bounding box transformation method is proposed. The method warps the ground-truth bounding boxes to the zoomed space to learn object detection, and warps the predicted bounding boxes back to the original space during inference. We conduct extensive experiments on three representative UAV object detection datasets, including VisDrone, UAVDT, and SeaDronesSee. The proposed ZoomDet is architecture-independent and can be applied to an arbitrary object detection architecture. Remarkably, on the SeaDronesSee dataset, ZoomDet offers more than 8.4 absolute gain of mAP with a Faster R-CNN model, with only about 3 ms additional latency. The code is available at https://github.com/twangnh/zoomdet_code.

Method: Proposes CA-YOLO with bionic modules: small target detection head and Characteristic Fusion Attention Mechanism (CFAM) in YOLO backbone. Also develops bionic pan-tilt tracking control strategy inspired by human Vestibulo-Ocular Reflex (VOR) with central positioning, stability optimization, adaptive control, and intelligent recapture.

Result: CA-YOLO outperforms original model on COCO (3.94% improvement) and VisDrone (4.90% improvement) datasets. Time-sensitive target localization experiments validate system effectiveness and practicality.

Conclusion: The bionic stabilized localization system successfully enhances target localization accuracy and small target recognition capabilities through biologically-inspired computer vision and control mechanisms.

Abstract: In modern complex environments, achieving accurate and efficient target localization is essential in numerous fields. However, existing systems often face limitations in both accuracy and the ability to recognize small targets. In this study, we propose a bionic stabilized localization system based on CA-YOLO, designed to enhance both target localization accuracy and small target recognition capabilities. Acting as the “brain” of the system, the target detection algorithm emulates the visual focusing mechanism of animals by integrating bionic modules into the YOLO backbone network. These modules include the introduction of a small target detection head and the development of a Characteristic Fusion Attention Mechanism (CFAM). Furthermore, drawing inspiration from the human Vestibulo-Ocular Reflex (VOR), a bionic pan-tilt tracking control strategy is developed, which incorporates central positioning, stability optimization, adaptive control coefficient adjustment, and an intelligent recapture function. The experimental results show that CA-YOLO outperforms the original model on standard datasets (COCO and VisDrone), with average accuracy metrics improved by 3.94%and 4.90%, respectively.Further time-sensitive target localization experiments validate the effectiveness and practicality of this bionic stabilized localization system.

Method: Uses instruction-tuned VLMs as evaluators for scalable benchmarking across diverse VQA datasets, introduces unified evaluation task/metric for joint assessment of localization and representation usefulness, includes multi-feature reconstruction baseline

Result: Proposes a new evaluation framework that enables comprehensive assessment of OCL models’ representation quality for complex reasoning tasks through VLM-based evaluation

Conclusion: Addresses limitations in existing OCL benchmarks by providing unified evaluation that jointly assesses localization and representation usefulness using VLMs as scalable evaluators

Abstract: Object-centric learning (OCL) aims to learn structured scene representations that support compositional generalization and robustness to out-of-distribution (OOD) data. However, OCL models are often not evaluated regarding these goals. Instead, most prior work focuses on evaluating OCL models solely through object discovery and simple reasoning tasks, such as probing the representation via image classification. We identify two limitations in existing benchmarks: (1) They provide limited insights on the representation usefulness of OCL models, and (2) localization and representation usefulness are assessed using disjoint metrics. To address (1), we use instruction-tuned VLMs as evaluators, enabling scalable benchmarking across diverse VQA datasets to measure how well VLMs leverage OCL representations for complex reasoning tasks. To address (2), we introduce a unified evaluation task and metric that jointly assess localization (where) and representation usefulness (what), thereby eliminating inconsistencies introduced by disjoint evaluation. Finally, we include a simple multi-feature reconstruction baseline as a reference point.

Method: Extracted statistical descriptors, radiomic texture features, and deep feature embeddings (mJ-Net and nnU-Net) from admission CTP. Constructed six ROIs from intersected T1 and T2 masks. Analyzed feature space clustering of region-level representations.

Result: Meaningful clustering of region-level representations in 18 patients. Penumbra regions showed significant feature differences based on final outcome, while core regions did not. Deep feature spaces (especially mJ-Net) showed strong separation between salvageable and non-salvageable tissue.

Conclusion: Encoder-derived feature manifolds reflect underlying tissue phenotypes and state transitions, providing insight into imaging-based quantification of stroke evolution.

Abstract: Computed tomography perfusion (CTP) at admission is routinely used to estimate the ischemic core and penumbra, while follow-up diffusion-weighted MRI (DWI) provides the definitive infarct outcome. However, single time-point segmentations fail to capture the biological heterogeneity and temporal evolution of stroke. We propose a bi-temporal analysis framework that characterizes ischemic tissue using statistical descriptors, radiomic texture features, and deep feature embeddings from two architectures (mJ-Net and nnU-Net). Bi-temporal refers to admission (T1) and post-treatment follow-up (T2). All features are extracted at T1 from CTP, with follow-up DWI aligned to ensure spatial correspondence. Manually delineated masks at T1 and T2 are intersected to construct six regions of interest (ROIs) encoding both initial tissue state and final outcome. Features were aggregated per region and analyzed in feature space. Evaluation on 18 patients with successful reperfusion demonstrated meaningful clustering of region-level representations. Regions classified as penumbra or healthy at T1 that ultimately recovered exhibited feature similarity to preserved brain tissue, whereas infarct-bound regions formed distinct groupings. Both baseline GLCM and deep embeddings showed a similar trend: penumbra regions exhibit features that are significantly different depending on final state, whereas this difference is not significant for core regions. Deep feature spaces, particularly mJ-Net, showed strong separation between salvageable and non-salvageable tissue, with a penumbra separation index that differed significantly from zero (Wilcoxon signed-rank test). These findings suggest that encoder-derived feature manifolds reflect underlying tissue phenotypes and state transitions, providing insight into imaging-based quantification of stroke evolution.

Method: LGDEA leverages LLMs to extract key diagnostic evidence from radiology reports and constructs a shared diagnostic evidence space, enabling evidence-aware cross-modal alignment. This allows effective exploitation of abundant unpaired medical images and reports, reducing reliance on paired data.

Result: The method achieves consistent and significant improvements on phrase grounding, image-text retrieval, and zero-shot classification tasks, and even rivals pretraining methods that rely on substantial paired data.

Conclusion: LGDEA successfully addresses limitations of existing medical VLP methods by introducing evidence-level alignment guided by LLMs, enabling more reliable diagnostic representations with reduced dependence on paired data.

Abstract: Most existing CLIP-style medical vision–language pretraining methods rely on global or local alignment with substantial paired data. However, global alignment is easily dominated by non-diagnostic information, while local alignment fails to integrate key diagnostic evidence. As a result, learning reliable diagnostic representations becomes difficult, which limits their applicability in medical scenarios with limited paired data. To address this issue, we propose an LLM-Guided Diagnostic Evidence Alignment method (LGDEA), which shifts the pretraining objective toward evidence-level alignment that is more consistent with the medical diagnostic process. Specifically, we leverage LLMs to extract key diagnostic evidence from radiology reports and construct a shared diagnostic evidence space, enabling evidence-aware cross-modal alignment and allowing LGDEA to effectively exploit abundant unpaired medical images and reports, thereby substantially alleviating the reliance on paired data. Extensive experimental results demonstrate that our method achieves consistent and significant improvements on phrase grounding, image–text retrieval, and zero-shot classification, and even rivals pretraining methods that rely on substantial paired data.

Method: MUFASA is a lightweight plug-and-play framework that computes slot attention across multiple feature layers of a ViT encoder, then uses a fusion strategy to aggregate slots from different layers into unified object-centric representations.

Result: Integrating MUFASA into existing OCL methods improves segmentation results across multiple datasets, sets new state-of-the-art performance, and improves training convergence with only minor inference overhead.

Conclusion: Leveraging multi-layer semantic information from ViTs through slot attention significantly enhances unsupervised object segmentation performance without substantial computational cost.

Abstract: Unsupervised object-centric learning (OCL) decomposes visual scenes into distinct entities. Slot attention is a popular approach that represents individual objects as latent vectors, called slots. Current methods obtain these slot representations solely from the last layer of a pre-trained vision transformer (ViT), ignoring valuable, semantically rich information encoded across the other layers. To better utilize this latent semantic information, we introduce MUFASA, a lightweight plug-and-play framework for slot attention-based approaches to unsupervised object segmentation. Our model computes slot attention across multiple feature layers of the ViT encoder, fully leveraging their semantic richness. We propose a fusion strategy to aggregate slots obtained on multiple layers into a unified object-centric representation. Integrating MUFASA into existing OCL methods improves their segmentation results across multiple datasets, setting a new state of the art while simultaneously improving training convergence with only minor inference overhead.

Method: Proposes FSSDINO - a training-free baseline using frozen DINOv3 features with class-specific prototypes and Gram-matrix refinement. Conducts Oracle-guided layer analysis to identify optimal intermediate representations versus standard last-layer features.

Result: FSSDINO is highly competitive with specialized methods across binary, multi-class, and cross-domain benchmarks. Reveals a “Safest vs. Optimal” dilemma where Oracle-selected intermediate features outperform last-layer features, but current selection metrics fail to identify them reliably.

Conclusion: Establishes last-layer as a deceptively strong baseline while revealing a “Semantic Selection Gap” in foundation models where traditional heuristics fail to identify optimal features, highlighting untapped potential in intermediate representations.

Abstract: Recent self-supervised Vision Transformers (ViTs), such as DINOv3, provide rich feature representations for dense vision tasks. This study investigates the intrinsic few-shot semantic segmentation (FSS) capabilities of frozen DINOv3 features through a training-free baseline, FSSDINO, utilizing class-specific prototypes and Gram-matrix refinement. Our results across binary, multi-class, and cross-domain (CDFSS) benchmarks demonstrate that this minimal approach, applied to the final backbone layer, is highly competitive with specialized methods involving complex decoders or test-time adaptation. Crucially, we conduct an Oracle-guided layer analysis, identifying a significant performance gap between the standard last-layer features and globally optimal intermediate representations. We reveal a “Safest vs. Optimal” dilemma: while the Oracle proves higher performance is attainable, matching the results of compute-intensive adaptation methods, current unsupervised and support-guided selection metrics consistently yield lower performance than the last-layer baseline. This characterizes a “Semantic Selection Gap” in Foundation Models, a disconnect where traditional heuristics fail to reliably identify high-fidelity features. Our work establishes the “Last-Layer” as a deceptively strong baseline and provides a rigorous diagnostic of the latent semantic potentials in DINOv3.The code is publicly available at https://github.com/hussni0997/fssdino.

Method: FlexID uses intent-aware modulation with orthogonal decoupling: Semantic Identity Projector (SIP) injects high-level priors into language space, Visual Feature Anchor (VFA) ensures structural fidelity in latent space. Context-Aware Adaptive Gating (CAG) dynamically modulates weights based on editing intent and diffusion timesteps, relaxing visual constraints when strong editing is detected.

Result: Extensive experiments on IBench show FlexID achieves state-of-the-art balance between identity consistency and text adherence, offering efficient solution for complex narrative generation.

Conclusion: FlexID provides a novel training-free framework that achieves synergy between identity preservation and semantic variation through intent-aware modulation, addressing the fundamental conflict in personalized text-to-image generation.

Abstract: Personalized text-to-image generation aims to seamlessly integrate specific identities into textual descriptions. However, existing training-free methods often rely on rigid visual feature injection, creating a conflict between identity fidelity and textual adaptability. To address this, we propose FlexID, a novel training-free framework utilizing intent-aware modulation. FlexID orthogonally decouples identity into two dimensions: a Semantic Identity Projector (SIP) that injects high-level priors into the language space, and a Visual Feature Anchor (VFA) that ensures structural fidelity within the latent space. Crucially, we introduce a Context-Aware Adaptive Gating (CAG) mechanism that dynamically modulates the weights of these streams based on editing intent and diffusion timesteps. By automatically relaxing rigid visual constraints when strong editing intent is detected, CAG achieves synergy between identity preservation and semantic variation. Extensive experiments on IBench demonstrate that FlexID achieves a state-of-the-art balance between identity consistency and text adherence, offering an efficient solution for complex narrative generation.

Method: Proposes a compact 3B-parameter VLA agent that performs human-like embodied reasoning through explicit image-grounded reasoning in three stages: “think”, “think summary”, and “action”

Result: The agent yields improved explainability, stronger generalization, and more efficient navigation compared to existing approaches

Conclusion: The proposed VLA agent provides a more effective approach to language-driven object navigation by combining explicit reasoning with efficient parameter usage

Abstract: Language-driven object navigation requires agents to interpret natural language descriptions of target objects, which combine intrinsic and extrinsic attributes for instance recognition and commonsense navigation. Existing methods either (i) use end-to-end trained models with vision-language embeddings, which struggle to generalize beyond training data and lack action-level explainability, or (ii) rely on modular zero-shot pipelines with large language models (LLMs) and open-set object detectors, which suffer from error propagation, high computational cost, and difficulty integrating their reasoning back into the navigation policy. To this end, we propose a compact 3B-parameter Vision-Language-Action (VLA) agent that performs human-like embodied reasoning for both object recognition and action selection, removing the need for stitched multi-model pipelines. Instead of raw embedding matching, our agent employs explicit image-grounded reasoning to directly answer “Is this the target object?” and “Why should I take this action?” The reasoning process unfolds in three stages: “think”, “think summary”, and “action”, yielding improved explainability, stronger generalization, and more efficient navigation. Code and dataset available upon acceptance.

Method: SIGMA introduces selective multi-attribute tokens (style, content, subject, identity) and uses post-training on the Bagel backbone with 700K interleaved examples. It enables the model to process interleaved text-image sequences for multi-condition generation.

Result: SIGMA improves controllability, cross-condition consistency, and visual quality across diverse editing and generation tasks, with substantial gains over Bagel on compositional tasks.

Conclusion: SIGMA provides a flexible unified framework for interleaved multi-condition generation in diffusion transformers, enabling more sophisticated compositional editing and attribute transfer capabilities.

Abstract: Recent unified models such as Bagel demonstrate that paired image-edit data can effectively align multiple visual tasks within a single diffusion transformer. However, these models remain limited to single-condition inputs and lack the flexibility needed to synthesize results from multiple heterogeneous sources. We present SIGMA (Selective-Interleaved Generation with Multi-Attribute Tokens), a unified post-training framework that enables interleaved multi-condition generation within diffusion transformers. SIGMA introduces selective multi-attribute tokens, including style, content, subject, and identity tokens, which allow the model to interpret and compose multiple visual conditions in an interleaved text-image sequence. Through post-training on the Bagel unified backbone with 700K interleaved examples, SIGMA supports compositional editing, selective attribute transfer, and fine-grained multimodal alignment. Extensive experiments show that SIGMA improves controllability, cross-condition consistency, and visual quality across diverse editing and generation tasks, with substantial gains over Bagel on compositional tasks.

Method: Annual competition using SUSTech-Competition dataset with substantial variations (clothing, carried objects, view angles), no dedicated training data provided (external datasets only), different random seeds each year for evaluation splits to prevent overfitting.

Result: Best-performing method reached 94.2% accuracy, setting new benchmark on this challenging dataset, demonstrating algorithmic advances can surpass previous accuracy limits despite heightened difficulty.

Conclusion: Gait recognition continues to advance with algorithmic improvements, competition framework effectively promotes progress and fair evaluation, and future research directions are outlined for the field.

Abstract: Human identification at a distance (HID) is challenging because traditional biometric modalities such as face and fingerprints are often difficult to acquire in real-world scenarios. Gait recognition provides a practical alternative, as it can be captured reliably at a distance. To promote progress in gait recognition and provide a fair evaluation platform, the International Competition on Human Identification at a Distance (HID) has been organized annually since 2020. Since 2023, the competition has adopted the challenging SUSTech-Competition dataset, which features substantial variations in clothing, carried objects, and view angles. No dedicated training data are provided, requiring participants to train their models using external datasets. Each year, the competition applies a different random seed to generate distinct evaluation splits, which reduces the risk of overfitting and supports a fair assessment of cross-domain generalization. While HID 2023 and HID 2024 already used this dataset, HID 2025 explicitly examined whether algorithmic advances could surpass the accuracy limits observed previously. Despite the heightened difficulty, participants achieved further improvements, and the best-performing method reached 94.2% accuracy, setting a new benchmark on this dataset. We also analyze key technical trends and outline potential directions for future research in gait recognition.

Method: Proposes a cross-camera cow identification framework based on disentangled representation learning using Subspace Identifiability Guarantee (SIG) theory. Designs a feature disentanglement module that decomposes images into orthogonal latent subspaces to isolate stable, identity-related biometric features invariant across cameras.

Result: Achieves 86.0% average accuracy across seven cross-camera tasks, significantly outperforming Source-only Baseline (51.9%) and strongest cross-camera baseline (79.8%). Validated on a dataset spanning five distinct camera nodes with heterogeneous devices and complex variations.

Conclusion: Establishes a subspace-theoretic feature disentanglement framework for collaborative cross-camera cow identification, offering a new paradigm for precise animal monitoring in uncontrolled smart farming environments.

Abstract: Precise identification of individual cows is a fundamental prerequisite for comprehensive digital management in smart livestock farming. While existing animal identification methods excel in controlled, single-camera settings, they face severe challenges regarding cross-camera generalization. When models trained on source cameras are deployed to new monitoring nodes characterized by divergent illumination, backgrounds, viewpoints, and heterogeneous imaging properties, recognition performance often degrades dramatically. This limits the large-scale application of non-contact technologies in dynamic, real-world farming environments. To address this challenge, this study proposes a cross-camera cow identification framework based on disentangled representation learning. This framework leverages the Subspace Identifiability Guarantee (SIG) theory in the context of bovine visual recognition. By modeling the underlying physical data generation process, we designed a principle-driven feature disentanglement module that decomposes observed images into multiple orthogonal latent subspaces. This mechanism effectively isolates stable, identity-related biometric features that remain invariant across cameras, thereby substantially improving generalization to unseen cameras. We constructed a high-quality dataset spanning five distinct camera nodes, covering heterogeneous acquisition devices and complex variations in lighting and angles. Extensive experiments across seven cross-camera tasks demonstrate that the proposed method achieves an average accuracy of 86.0%, significantly outperforming the Source-only Baseline (51.9%) and the strongest cross-camera baseline method (79.8%). This work establishes a subspace-theoretic feature disentanglement framework for collaborative cross-camera cow identification, offering a new paradigm for precise animal monitoring in uncontrolled smart farming environments.

Method: MammoColor uses a lightweight Task-Driven Chromatic Encoding (TDCE) module coupled with a BI-RADS triage classifier, trained end-to-end on VinDr-Mammo dataset. The framework converts single-channel mammograms into TDCE-encoded color views for visual augmentation.

Result: On VinDr-Mammo, MammoColor improved AUC from 0.7669 to 0.8461 (P=0.004), with larger gains in dense breasts (AUC 0.749 to 0.835). In MRMC observer studies, TDCE-encoded images improved specificity from 0.90 to 0.96 with comparable sensitivity.

Conclusion: TDCE provides a task-optimized chromatic representation that improves perceptual salience and reduces false-positive recalls in mammography triage, particularly beneficial for dense breast screening.

Abstract: Purpose:Mammography screening is less sensitive in dense breasts, where tissue overlap and subtle findings increase perceptual difficulty. We present MammoColor, an end-to-end framework with a Task-Driven Chromatic Encoding (TDCE) module that converts single-channel mammograms into TDCE-encoded views for visual augmentation. Materials and Methods:MammoColor couples a lightweight TDCE module with a BI-RADS triage classifier and was trained end-to-end on VinDr-Mammo. Performance was evaluated on an internal test set, two public datasets (CBIS-DDSM and INBreast), and three external clinical cohorts. We also conducted a multi-reader, multi-case (MRMC) observer study with a washout period, comparing (1) grayscale-only, (2) TDCE-only, and (3) side-by-side grayscale+TDCE. Results:On VinDr-Mammo, MammoColor improved AUC from 0.7669 to 0.8461 (P=0.004). Gains were larger in dense breasts (AUC 0.749 to 0.835). In the MRMC study, TDCE-encoded images improved specificity (0.90 to 0.96; P=0.052) with comparable sensitivity. Conclusion:TDCE provides a task-optimized chromatic representation that may improve perceptual salience and reduce false-positive recalls in mammography triage.

Method: ViCA bypasses visual tokens through all self-attention and feed-forward layers, interacting with text solely through sparse cross-attention at selected layers. This creates a regular, hardware-friendly inference pipeline.

Result: ViCA achieves 98% of baseline accuracy while reducing visual-side computation to 4%, with over 3.5x speedup in single-batch inference and over 10x speedup in multi-batch inference. It’s orthogonal to token pruning methods and can be combined for further efficiency gains.

Conclusion: ViCA demonstrates that dense visual processing in MLLMs is unnecessary and proposes an efficient alternative that maintains performance while dramatically reducing computational overhead.

Abstract: Modern multimodal large language models (MLLMs) adopt a unified self-attention design that processes visual and textual tokens at every Transformer layer, incurring substantial computational overhead. In this work, we revisit the necessity of such dense visual processing and show that projected visual embeddings are already well-aligned with the language space, while effective vision-language interaction occurs in only a small subset of layers. Based on these insights, we propose ViCA (Vision-only Cross-Attention), a minimal MLLM architecture in which visual tokens bypass all self-attention and feed-forward layers, interacting with text solely through sparse cross-attention at selected layers. Extensive evaluations across three MLLM backbones, nine multimodal benchmarks, and 26 pruning-based baselines show that ViCA preserves 98% of baseline accuracy while reducing visual-side computation to 4%, consistently achieving superior performance-efficiency trade-offs. Moreover, ViCA provides a regular, hardware-friendly inference pipeline that yields over 3.5x speedup in single-batch inference and over 10x speedup in multi-batch inference, reducing visual grounding to near-zero overhead compared with text-only LLMs. It is also orthogonal to token pruning methods and can be seamlessly combined for further efficiency gains. Our code is available at https://github.com/EIT-NLP/ViCA.

Method: Combines discrete fracture network models for synthetic image generation, mixed training with synthetic data, pretraining, and fine-tuning on real images for segmentation.

Result: Synthetic data supports joint trace detection when real data is scarce; mixed training works with consistent labels, fine-tuning handles noisy labels; zero-shot prediction limited but fine-tuning with small real data achieves useful generalization.

Conclusion: The method enables reliable joint mapping and provides foundation for domain adaptation and evaluation, with qualitative analysis showing geologically meaningful results beyond quantitative metrics.

Abstract: This paper presents a geology-driven machine learning method for automated rock joint trace mapping from images. The approach combines geological modelling, synthetic data generation, and supervised image segmentation to address limited real data and class imbalance. First, discrete fracture network models are used to generate synthetic jointed rock images at field-relevant scales via parametric modelling, preserving joint persistence, connectivity, and node-type distributions. Second, segmentation models are trained using mixed training and pretraining followed by fine-tuning on real images. The method is tested in box and slope domains using several real datasets. The results show that synthetic data can support supervised joint trace detection when real data are scarce. Mixed training performs well when real labels are consistent (e.g. box-domain), while fine-tuning is more robust when labels are noisy (e.g. slope-domain where labels can be biased, incomplete, and inconsistent). Fully zero-shot prediction from synthetic model remains limited, but useful generalisation is achieved by fine-tuning with a small number of real data. Qualitative analysis shows clearer and more geologically meaningful joint traces than indicated by quantitative metrics alone. The proposed method supports reliable joint mapping and provides a basis for further work on domain adaptation and evaluation.

Method: A systematic framework organizing supervised policy shaping, reward-driven reinforcement learning, and preference-based refinement into a single stability-constrained optimization stack, treating optimization as staged diagnostic-driven process rather than isolated tricks.

Result: Provides a cohesive recipe for improving perceptual fidelity, temporal coherence, and prompt adherence while preserving controllability, offering a blueprint for scalable post-training pipelines that remain stable and effective in real-world deployment.

Conclusion: The framework enables building scalable post-training pipelines for video generators that are stable, extensible, and effective for real-world deployment by systematically addressing practical constraints and optimization challenges.

Abstract: Post-training is the decisive step for converting a pretrained video generator into a production-oriented model that is instruction-following, controllable, and robust over long temporal horizons. This report presents a systematical post-training framework that organizes supervised policy shaping, reward-driven reinforcement learning, and preference-based refinement into a single stability-constrained optimization stack. The framework is designed around practical video-generation constraints, including high rollout cost, temporally compounding failure modes, and feedback that is heterogeneous, uncertain, and often weakly discriminative. By treating optimization as a staged, diagnostic-driven process rather than a collection of isolated tricks, the report summarizes a cohesive recipe for improving perceptual fidelity, temporal coherence, and prompt adherence while preserving the controllability established at initialization. The resulting framework provides a clear blueprint for building scalable post-training pipelines that remain stable, extensible, and effective in real-world deployment settings.

Method: Two-stage R1-style framework: (1) Chain-of-Thought Supervised Fine-tuning using a constructed FGVR CoT dataset with rationales, and (2) Triplet Augmented Policy Optimization with intra-class augmentation (mixing trajectories within same category) and inter-class augmentation (maximizing response distinction across sub-categories).

Result: With only 4-shot training, Fine-R1 outperforms existing general MLLMs, reasoning MLLMs, and contrastive CLIP models in identifying both seen and unseen sub-categories.

Conclusion: Fine-R1 shows promise for knowledge-intensive domains where gathering expert annotations for all sub-categories is difficult, demonstrating effective adaptation of MLLMs to fine-grained visual recognition tasks.

Abstract: Any entity in the visual world can be hierarchically grouped based on shared characteristics and mapped to fine-grained sub-categories. While Multi-modal Large Language Models (MLLMs) achieve strong performance on coarse-grained visual tasks, they often struggle with Fine-Grained Visual Recognition (FGVR). Adapting general-purpose MLLMs to FGVR typically requires large amounts of annotated data, which is costly to obtain, leaving a substantial performance gap compared to contrastive CLIP models dedicated for discriminative tasks. Moreover, MLLMs tend to overfit to seen sub-categories and generalize poorly to unseen ones. To address these challenges, we propose Fine-R1, an MLLM tailored for FGVR through an R1-style training framework: (1) Chain-of-Thought Supervised Fine-tuning, where we construct a high-quality FGVR CoT dataset with rationales of “visual analysis, candidate sub-categories, comparison, and prediction”, transition the model into a strong open-world classifier; and (2) Triplet Augmented Policy Optimization, where Intra-class Augmentation mixes trajectories from anchor and positive images within the same category to improve robustness to intra-class variance, while Inter-class Augmentation maximizes the response distinction conditioned on images across sub-categories to enhance discriminative ability. With only 4-shot training, Fine-R1 outperforms existing general MLLMs, reasoning MLLMs, and even contrastive CLIP models in identifying both seen and unseen sub-categories, showing promise in working in knowledge-intensive domains where gathering expert annotations for all sub-categories is arduous. Code is available at https://github.com/PKU-ICST-MIPL/FineR1_ICLR2026.

Method: Proposes HistoMet, a decision-aware, concept-aligned multiple instance learning framework with a two-module pipeline: 1) estimates likelihood of metastatic progression from primary tumor, 2) conditionally predicts metastatic site for high-risk cases. Integrates linguistically defined and data-adaptive metastatic concepts through a pretrained pathology vision-language model.

Result: Evaluated on multi-institutional pan-cancer cohort of 6504 patients. Under 95% sensitivity screening settings, significantly reduces downstream workload while maintaining high metastatic risk recall. For metastatic cases, achieves macro F1 of 74.6±1.3 and macro one-vs-rest AUC of 92.1.

Conclusion: Explicitly modeling clinical decision structure enables robust and deployable prognostic prediction of metastatic progression and site tropism directly from primary tumor histopathology.

Abstract: Metastatic Progression remains the leading cause of cancer-related mortality, yet predicting whether a primary tumor will metastasize and where it will disseminate directly from histopathology remains a fundamental challenge. Although whole-slide images (WSIs) provide rich morphological information, prior computational pathology approaches typically address metastatic status or site prediction as isolated tasks, and do not explicitly model the clinically sequential decision process of metastatic risk assessment followed by downstream site-specific evaluation. To address this research gap, we present a decision-aware, concept-aligned MIL framework, HistoMet, for prognostic metastatic outcome prediction from primary tumor WSIs. Our proposed framework adopts a two-module prediction pipeline in which the likelihood of metastatic progression from the primary tumor is first estimated, followed by conditional prediction of metastatic site for high-risk cases. To guide representation learning and improve clinical interpretability, our framework integrates linguistically defined and data-adaptive metastatic concepts through a pretrained pathology vision-language model. We evaluate HistoMet on a multi-institutional pan-cancer cohort of 6504 patients with metastasis follow-up and site annotations. Under clinically relevant high-sensitivity screening settings (95 percent sensitivity), HistoMet significantly reduces downstream workload while maintaining high metastatic risk recall. Conditional on metastatic cases, HistoMet achieves a macro F1 of 74.6 with a standard deviation of 1.3 and a macro one-vs-rest AUC of 92.1. These results demonstrate that explicitly modeling clinical decision structure enables robust and deployable prognostic prediction of metastatic progression and site tropism directly from primary tumor histopathology.

Method: Two-stage architecture: 1) Structure-Aware Memory Construction converts raw video into structured database using semantic retrieval and keyword matching, focusing on brand details while filtering noise; 2) Structured Reasoning Agent mimics marketing expert through iterative inquiry loop with evidence-based self-correction that validates insights against specific video frames.

Result: AD-MIR achieves state-of-the-art performance on AdsQA benchmark, surpassing the strongest general-purpose agent DVD by 1.8% in strict accuracy and 9.5% in relaxed accuracy.

Conclusion: Effective advertising understanding requires explicitly grounding abstract marketing strategies in pixel-level evidence, which AD-MIR successfully accomplishes through its structured approach.

Abstract: Multimodal understanding of advertising videos is essential for interpreting the intricate relationship between visual storytelling and abstract persuasion strategies. However, despite excelling at general search, existing agents often struggle to bridge the cognitive gap between pixel-level perception and high-level marketing logic. To address this challenge, we introduce AD-MIR, a framework designed to decode advertising intent via a two-stage architecture. First, in the Structure-Aware Memory Construction phase, the system converts raw video into a structured database by integrating semantic retrieval with exact keyword matching. This approach prioritizes fine-grained brand details (e.g., logos, on-screen text) while dynamically filtering out irrelevant background noise to isolate key protagonists. Second, the Structured Reasoning Agent mimics a marketing expert through an iterative inquiry loop, decomposing the narrative to deduce implicit persuasion tactics. Crucially, it employs an evidence-based self-correction mechanism that rigorously validates these insights against specific video frames, automatically backtracking when visual support is lacking. Evaluation on the AdsQA benchmark demonstrates that AD-MIR achieves state-of-the-art performance, surpassing the strongest general-purpose agent, DVD, by 1.8% in strict and 9.5% in relaxed accuracy. These results underscore that effective advertising understanding demands explicitly grounding abstract marketing strategies in pixel-level evidence. The code is available at https://github.com/Little-Fridge/AD-MIR.

Method: Created UMD dataset with 490 PET/CT and 464 PET/MRI scans (~675k 2D images, ~12k 3D organ annotations). Conducted intra-subject controlled comparisons using paired scans to isolate modality as the independent variable. Evaluated representative 3D segmentation foundation models on this dataset.

Result: Found stark discrepancy between literature-reported benchmarks and real-world efficacy, especially when transitioning from structural to functional domains. Models showed systemic failures on PET imaging despite good performance on CT/MRI.

Conclusion: Current 3D foundation models are not truly general-purpose and require multi-modal training and evaluation to bridge the gap between idealized benchmarking and comprehensive clinical utility. The UMD dataset provides a foundation for developing modality-agnostic medical foundation models.

Abstract: While emerging 3D medical foundation models are envisioned as versatile tools with offer general-purpose capabilities, their validation remains largely confined to regional and structural imaging, leaving a significant modality discrepancy unexplored. To provide a rigorous and objective assessment, we curate the UMD dataset comprising 490 whole-body PET/CT and 464 whole-body PET/MRI scans ($\sim$675k 2D images, $\sim$12k 3D organ annotations) and conduct a thorough and comprehensive evaluation of representative 3D segmentation foundation models. Through intra-subject controlled comparisons of paired scans, we isolate imaging modality as the primary independent variable to evaluate model robustness in real-world applications. Our evaluation reveals a stark discrepancy between literature-reported benchmarks and real-world efficacy, particularly when transitioning from structural to functional domains. Such systemic failures underscore that current 3D foundation models are far from achieving truly general-purpose status, necessitating a paradigm shift toward multi-modal training and evaluation to bridge the gap between idealized benchmarking and comprehensive clinical utility. This dataset and analysis establish a foundational cornerstone for future research to develop truly modality-agnostic medical foundation models.

Method: Uses a vision-language model (VLM) to extract region-level specifications from infographic images, maps pixel geometry to slide coordinates, and recreates elements using Google Slides batch update API. The system is model-agnostic with multiple VLM backends via common JSON schema.

Result: Achieves 98.9% element recovery rate on benchmark of 29 programmatically generated infographics, with high text transcription accuracy (CER=0.033) and good layout fidelity (IoU=0.364 for text, 0.644 for images).

Conclusion: Demonstrates successful conversion of static infographics to editable slides using VLMs, identifies practical engineering challenges (text size calibration, backgrounds), and provides failure modes for future work.

Abstract: Infographics are widely used to communicate information with a combination of text, icons, and data visualizations, but once exported as images their content is locked into pixels, making updates, localization, and reuse expensive. We describe \textsc{Images2Slides}, an API-based pipeline that converts a static infographic (PNG/JPG) into a native, editable Google Slides slide by extracting a region-level specification with a vision-language model (VLM), mapping pixel geometry into slide coordinates, and recreating elements using the Google Slides batch update API. The system is model-agnostic and supports multiple VLM backends via a common JSON region schema and deterministic postprocessing. On a controlled benchmark of 29 programmatically generated infographic slides with known ground-truth regions, \textsc{Images2Slides} achieves an overall element recovery rate of $0.989\pm0.057$ (text: $0.985\pm0.083$, images: $1.000\pm0.000$), with mean text transcription error $\mathrm{CER}=0.033\pm0.149$ and mean layout fidelity $\mathrm{IoU}=0.364\pm0.161$ for text regions and $0.644\pm0.131$ for image regions. We also highlight practical engineering challenges in reconstruction, including text size calibration and non-uniform backgrounds, and describe failure modes that guide future work.

Method: Printed sphere, facemask, and AAA models using SLA, scanned with micro-CT, segmented using GMM, Otsu, and RG methods, aligned with KU algorithm, compared using Dice/Jaccard scores and surface metrics like chamfer/Hausdorff distances.

Result: Otsu method most suitable for all geometries; AAA had low overlap scores due to thin walls and misalignment; class imbalance affected specificity most for AAA; surface metrics differed from voxel-based trends; RG best for sphere, GMM/Otsu better for AAA.

Conclusion: Segmentation accuracy accumulates errors across reconstruction stages; high voxel-based metrics can be misleading with class imbalance and alignment issues; Jaccard more stringent than Dice for thin-walled structures; proper alignment crucial for reliable assessment.

Abstract: The accuracy of the 3D models created from medical scans depends on imaging hardware, segmentation methods and mesh processing techniques etc. The effects of geometry type, class imbalance, voxel and point cloud alignment on accuracy remain to be thoroughly explored. This work evaluates the errors across the reconstruction pipeline and explores the use of voxel and surface-based accuracy metrics for different segmentation algorithms and geometry types. A sphere, a facemask, and an AAA were printed using the SLA technique and scanned using a micro-CT machine. Segmentation was performed using GMM, Otsu and RG based methods. Segmented and reference models aligned using the KU algorithm, were quantitatively compared to evaluate metrics like Dice and Jaccard scores, precision. Surface meshes were registered with reference meshes using an ICP-based alignment process. Metrics like chamfer distance, and average Hausdorff distance were evaluated. The Otsu method was found to be the most suitable method for all the geometries. AAA yielded low overlap scores due to its small wall thickness and misalignment. The effect of class imbalance on specificity was observed the most for AAA. Surface-based accuracy metrics differed from the voxel-based trends. The RG method performed best for sphere, while GMM and Otsu perform better for AAA. The facemask surface was most error-prone, possibly due to misalignment during the ICP process. Segmentation accuracy is a cumulative sum of errors across different stages of the reconstruction process. High voxel-based accuracy metrics may be misleading in cases of high class imbalance and sensitivity to alignment. The Jaccard index is found to be more stringent than the Dice and more suitable for accuracy assessment for thin-walled structures. Voxel and point cloud alignment should be ensured to make any reliable assessment of the reconstruction pipeline.

Method: Proposes L-LIO framework that incorporates audio signals alongside visual data. Evaluates three use cases: 1) supervised learning on driver speech to classify impairment states, 2) collection/analysis of passenger natural language instructions for planning systems, 3) audio disambiguation of external agent guidance where vision is insufficient. Uses custom-collected in-vehicle and external audio datasets.

Result: Pilot findings show audio yields safety-relevant insights in nuanced scenarios where sound is critical to safe decision-making or visual signals alone are insufficient. Audio enhances understanding of driver states, passenger intentions, and external guidance.

Conclusion: L-LIO augments driver and scene understanding through multimodal fusion of audio and visual sensing, offering new paths for safety intervention, though challenges remain with ambient noise, privacy, and robustness across subjects in dynamic real-world contexts.

Abstract: The looking-in-looking-out (LILO) framework has enabled intelligent vehicle applications that understand both the outside scene and the driver state to improve safety outcomes, with examples in smart airbag deployment, takeover time prediction in autonomous control transitions, and driver attention monitoring. In this research, we propose an augmentation to this framework, making a case for the audio modality as an additional source of information to understand the driver, and in the evolving autonomy landscape, also the passengers and those outside the vehicle. We expand LILO by incorporating audio signals, forming the looking-and-listening inside-and-outside (L-LIO) framework to enhance driver state assessment and environment understanding through multimodal sensor fusion. We evaluate three example cases where audio enhances vehicle safety: supervised learning on driver speech audio to classify potential impairment states (e.g., intoxication), collection and analysis of passenger natural language instructions (e.g., “turn after that red building”) to motivate how spoken language can interface with planning systems through audio-aligned instruction data, and limitations of vision-only systems where audio may disambiguate the guidance and gestures of external agents. Datasets include custom-collected in-vehicle and external audio samples in real-world environments. Pilot findings show that audio yields safety-relevant insights, particularly in nuanced or context-rich scenarios where sound is critical to safe decision-making or visual signals alone are insufficient. Challenges include ambient noise interference, privacy considerations, and robustness across human subjects, motivating further work on reliability in dynamic real-world contexts. L-LIO augments driver and scene understanding through multimodal fusion of audio and visual sensing, offering new paths for safety intervention.

Method: Three complementary approaches: 1) CLIP-based category-agnostic hazard screening using image-text similarity, 2) Integration of scene-level VLM embeddings into transformer-based trajectory planning on Waymo dataset, 3) Using natural language as explicit behavioral constraints on motion planning with doScenes dataset.

Result: 1) Lightweight hazard screening works robustly for diverse/out-of-distribution hazards; 2) Naive global embedding conditioning doesn’t improve trajectory accuracy; 3) Natural language constraints suppress severe planning failures and improve safety in ambiguous scenarios.

Conclusion: VLMs hold significant promise for autonomous driving safety when used to express semantic risk, intent, and behavioral constraints, but realizing this requires careful engineering design and structured grounding rather than direct feature injection.

Abstract: Vision-language models (VLMs) have recently emerged as powerful representation learning systems that align visual observations with natural language concepts, offering new opportunities for semantic reasoning in safety-critical autonomous driving. This paper investigates how vision-language representations support driving scene safety assessment and decision-making when integrated into perception, prediction, and planning pipelines. We study three complementary system-level use cases. First, we introduce a lightweight, category-agnostic hazard screening approach leveraging CLIP-based image-text similarity to produce a low-latency semantic hazard signal. This enables robust detection of diverse and out-of-distribution road hazards without explicit object detection or visual question answering. Second, we examine the integration of scene-level vision-language embeddings into a transformer-based trajectory planning framework using the Waymo Open Dataset. Our results show that naively conditioning planners on global embeddings does not improve trajectory accuracy, highlighting the importance of representation-task alignment and motivating the development of task-informed extraction methods for safety-critical planning. Third, we investigate natural language as an explicit behavioral constraint on motion planning using the doScenes dataset. In this setting, passenger-style instructions grounded in visual scene elements suppress rare but severe planning failures and improve safety-aligned behavior in ambiguous scenarios. Taken together, these findings demonstrate that vision-language representations hold significant promise for autonomous driving safety when used to express semantic risk, intent, and behavioral constraints. Realizing this potential is fundamentally an engineering problem requiring careful system design and structured grounding rather than direct feature injection.

Method: Proposes Process-of-Thought (PoT) framework with three interleaved components: temporal evidence selection, step-wise state updates, and constrained answer synthesis. Uses unified representation for PoT traces aligning decisions with temporal segments. Model-agnostic design works with existing vision-language backbones.

Result: Extensive experiments show PoT consistently improves factual correctness and temporal grounding while providing interpretable reasoning traces. Reduces hallucinated explanations and improves robustness to distractors.

Conclusion: PoT Reasoning for Videos provides an effective framework for explicit, verifiable video reasoning that enhances both performance and interpretability in video understanding tasks.

Abstract: Video understanding requires not only recognizing visual content but also performing temporally grounded, multi-step reasoning over long and noisy observations. We propose Process-of-Thought (PoT) Reasoning for Videos, a framework that makes the reasoning process explicit by structuring video inference into a sequence of lightweight, verifiable steps. PoT interleaves (i) temporal evidence selection, (ii) step-wise state updates, and (iii) constrained answer synthesis, enabling the model to progressively refine hypotheses while maintaining traceability to video evidence. The framework is designed to be model-agnostic and can be plugged into existing vision-language backbones, supporting both closed-book reasoning and evidence-augmented reasoning with external tools. We further introduce a unified representation for PoT traces that aligns intermediate decisions with temporal segments, which improves robustness to distractors and reduces hallucinated explanations. Extensive experiments on standard video reasoning tasks demonstrate that PoT consistently improves factual correctness and temporal grounding, while providing interpretable reasoning traces for diagnosis and downstream use.

Method: Proposes a complementary-cue collaborative perception framework that extracts and fuses anomaly evidence from three perspectives: object-level semantic composition modeling, semantic-attribution-based global deviation analysis, and fine-grained texture matching.

Result: Experiments on the CoalAD benchmark show the method outperforms widely used baselines across image-level and pixel-level metrics, with ablation studies validating each component’s contribution.

Conclusion: The proposed framework provides robust anomaly detection and accurate pixel-level localization for challenging coal mining environments, addressing limitations of existing methods in unstructured industrial settings.

Abstract: Reliable foreign-object anomaly detection and pixel-level localization in conveyor-belt coal scenes are essential for safe and intelligent mining operations. This task is particularly challenging due to the highly unstructured environment: coal and gangue are randomly piled, backgrounds are complex and variable, and foreign objects often exhibit low contrast, deformation, occlusion, resulting in coupling with their surroundings. These characteristics weaken the stability and regularity assumptions that many anomaly detection methods rely on in structured industrial settings, leading to notable performance degradation. To support evaluation and comparison in this setting, we construct \textbf{CoalAD}, a benchmark for unsupervised foreign-object anomaly detection with pixel-level localization in coal-stream scenes. We further propose a complementary-cue collaborative perception framework that extracts and fuses complementary anomaly evidence from three perspectives: object-level semantic composition modeling, semantic-attribution-based global deviation analysis, and fine-grained texture matching. The fused outputs provide robust image-level anomaly scoring and accurate pixel-level localization. Experiments on CoalAD demonstrate that our method outperforms widely used baselines across the evaluated image-level and pixel-level metrics, and ablation studies validate the contribution of each component. The code is available at https://github.com/xjpp2016/USAD.

Method: U-KABS integrates U-shaped encoder-decoder architecture with Kolmogorov-Arnold Networks (KANs). It combines convolutional and squeeze-and-excitation stages for channel-wise feature enhancement with KAN Bernstein Spline (KABS) stage using learnable activation functions based on Bernstein polynomials and B-splines.

Result: U-KABS demonstrates superior performance across diverse medical imaging benchmark datasets compared to strong baselines, particularly in segmenting complex anatomical structures.

Conclusion: The hybrid design effectively captures both broad contextual trends and fine-grained patterns critical for medical image segmentation, leveraging global smoothness of Bernstein polynomials and local adaptability of B-splines.

Abstract: Medical image segmentation plays a vital role in diagnosis and treatment planning, but remains challenging due to the inherent complexity and variability of medical images, especially in capturing non-linear relationships within the data. We propose U-KABS, a novel hybrid framework that integrates the expressive power of Kolmogorov-Arnold Networks (KANs) with a U-shaped encoder-decoder architecture to enhance segmentation performance. The U-KABS model combines the convolutional and squeeze-and-excitation stage, which enhances channel-wise feature representations, and the KAN Bernstein Spline (KABS) stage, which employs learnable activation functions based on Bernstein polynomials and B-splines. This hybrid design leverages the global smoothness of Bernstein polynomials and the local adaptability of B-splines, enabling the model to effectively capture both broad contextual trends and fine-grained patterns critical for delineating complex structures in medical images. Skip connections between encoder and decoder layers support effective multi-scale feature fusion and preserve spatial details. Evaluated across diverse medical imaging benchmark datasets, U-KABS demonstrates superior performance compared to strong baselines, particularly in segmenting complex anatomical structures.

Method: Proposes a novel all-optical computing framework using diffractive optical neural networks (DONNs) for RGB image segmentation and lane detection. The system performs all-optical encoding and computing via light diffraction, eliminating analog-to-digital conversion overhead.

Result: Experimental results demonstrate effectiveness of DONN system for image segmentation on CityScapes dataset. Case studies on lane detection using customized indoor track dataset and simulated driving scenarios in CARLA show model’s generalizability under diverse environmental conditions.

Conclusion: DONNs provide an energy-efficient alternative to conventional DNNs for autonomous driving vision tasks by enabling all-optical computing at light speed, reducing energy costs while maintaining performance for segmentation and lane detection.

Abstract: Semantic segmentation and lane detection are crucial tasks in autonomous driving systems. Conventional approaches predominantly rely on deep neural networks (DNNs), which incur high energy costs due to extensive analog-to-digital conversions and large-scale image computations required for low-latency, real-time responses. Diffractive optical neural networks (DONNs) have shown promising advantages over conventional DNNs on digital or optoelectronic computing platforms in energy efficiency. By performing all-optical image processing via light diffraction at the speed of light, DONNs save computation energy costs while reducing the overhead associated with analog-to-digital conversions by all-optical encoding and computing. In this work, we propose a novel all-optical computing framework for RGB image segmentation and lane detection in autonomous driving applications. Our experimental results demonstrate the effectiveness of the DONN system for image segmentation on the CityScapes dataset. Additionally, we conduct case studies on lane detection using a customized indoor track dataset and simulated driving scenarios in CARLA, where we further evaluate the model’s generalizability under diverse environmental conditions.

Method: The paper conducts systematic analysis of AR cache maintenance and proposes Rolling Sink, a training-free method built on Self Forcing. It effectively scales AR video synthesis to ultra-long durations (5-30 minutes) by addressing the gap between limited training horizons and open-ended testing horizons.

Result: Rolling Sink achieves superior long-horizon visual fidelity and temporal consistency compared to state-of-the-art baselines, enabling consistent subjects, stable colors, coherent structures, and smooth motions in ultra-long video generation.

Conclusion: Rolling Sink provides an effective training-free solution to bridge the train-test gap in autoregressive video diffusion models, enabling scalable ultra-long video synthesis without expensive long-video training.

Abstract: Recently, autoregressive (AR) video diffusion models has achieved remarkable performance. However, due to their limited training durations, a train-test gap emerges when testing at longer horizons, leading to rapid visual degradations. Following Self Forcing, which studies the train-test gap within the training duration, this work studies the train-test gap beyond the training duration, i.e., the gap between the limited horizons during training and open-ended horizons during testing. Since open-ended testing can extend beyond any finite training window, and long-video training is computationally expensive, we pursue a training-free solution to bridge this gap. To explore a training-free solution, we conduct a systematic analysis of AR cache maintenance. These insights lead to Rolling Sink. Built on Self Forcing (trained on only 5s clips), Rolling Sink effectively scales the AR video synthesis to ultra-long durations (e.g., 5-30 minutes at 16 FPS) at test time, with consistent subjects, stable colors, coherent structures, and smooth motions. As demonstrated by extensive experiments, Rolling Sink achieves superior long-horizon visual fidelity and temporal consistency compared to SOTA baselines. Project page: https://rolling-sink.github.io/

Method: Models traffic signal control as a stochastic decision process with constraints under partial observability, accounting for vision-based perception uncertainty. Uses counterfactual rollouts in belief space to predict and enforce hard safety and starvation prevention constraints.

Result: The system is designed to improve traffic delay and emissions while preventing safety-critical errors and providing interpretable control policy outputs based on explicit models.

Conclusion: UCATSC offers a model-based approach to traffic signal control that addresses key limitations of existing methods by incorporating uncertainty modeling and hard safety constraints.

Abstract: Real-world deployment of adaptive traffic signal control, to date, remains limited due to the uncertainty associated with vision-based perception, implicit safety, and non-interpretable control policies learned and validated mainly in simulation. In this paper, we introduce UCATSC, a model-based traffic signal control system that models traffic signal control at an intersection using a stochastic decision process with constraints and under partial observability, taking into account the uncertainty associated with vision-based perception. Unlike reinforcement learning methods that learn to predict safety using reward shaping, UCATSC predicts and enforces hard constraints related to safety and starvation prevention during counterfactual rollouts in belief space. The system is designed to improve traffic delay and emission while preventing safety-critical errors and providing interpretable control policy outputs based on explicit models.

Method: Proposes VideoTemp-o3 framework with: 1) Unified masking mechanism in supervised fine-tuning to encourage exploration while preventing noise, 2) Dedicated reinforcement learning rewards to mitigate reward hacking, 3) Pipeline for constructing high-quality long video grounded QA data, and 4) Benchmark for systematic evaluation across video durations.

Result: Experimental results demonstrate remarkable performance on both long video understanding and grounding tasks.

Conclusion: VideoTemp-o3 provides an effective solution for long-video understanding by addressing key limitations of existing methods through joint modeling of video grounding and question answering with flexible localization capabilities.

Abstract: In long-video understanding, conventional uniform frame sampling often fails to capture key visual evidence, leading to degraded performance and increased hallucinations. To address this, recent agentic thinking-with-videos paradigms have emerged, adopting a localize-clip-answer pipeline in which the model actively identifies relevant video segments, performs dense sampling within those clips, and then produces answers. However, existing methods remain inefficient, suffer from weak localization, and adhere to rigid workflows. To solve these issues, we propose VideoTemp-o3, a unified agentic thinking-with-videos framework that jointly models video grounding and question answering. VideoTemp-o3 exhibits strong localization capability, supports on-demand clipping, and can refine inaccurate localizations. Specifically, in the supervised fine-tuning stage, we design a unified masking mechanism that encourages exploration while preventing noise. For reinforcement learning, we introduce dedicated rewards to mitigate reward hacking. Besides, from the data perspective, we develop an effective pipeline to construct high-quality long video grounded QA data, along with a corresponding benchmark for systematic evaluation across various video durations. Experimental results demonstrate that our method achieves remarkable performance on both long video understanding and grounding.

Method: First comprehensive zero-shot evaluation of 16 state-of-the-art detection methods (23 pretrained detector variants) across 12 diverse datasets comprising 2.6 million image samples spanning 291 unique generators including modern diffusion models.

Result: Key findings: (1) No universal winner exists with detector rankings showing substantial instability; (2) 37 percentage-point performance gap between best (75.0% mean accuracy) and worst (37.5%) detectors; (3) Training data alignment causes 20-60% performance variance; (4) Modern commercial generators defeat most detectors (18-30% average accuracy); (5) Three systematic failure patterns identified affecting cross-dataset generalization.

Conclusion: Findings challenge the “one-size-fits-all” detector paradigm and provide actionable deployment guidelines. Practitioners must carefully select detectors based on specific threat landscape rather than relying on published benchmark performance.

Abstract: As AI-generated images proliferate across digital platforms, reliable detection methods have become critical for combating misinformation and maintaining content authenticity. While numerous deepfake detection methods have been proposed, existing benchmarks predominantly evaluate fine-tuned models, leaving a critical gap in understanding out-of-the-box performance – the most common deployment scenario for practitioners. We present the first comprehensive zero-shot evaluation of 16 state-of-the-art detection methods, comprising 23 pretrained detector variants (due to multiple released versions of certain detectors), across 12 diverse datasets, comprising 2.6million image samples spanning 291 unique generators including modern diffusion models. Our systematic analysis reveals striking findings: (1)no universal winner exists, with detector rankings exhibiting substantial instability (Spearman$ρ$: 0.01 – 0.87 across dataset pairs); (2)a 37percentage-point performance gap separates the best detector (75.0% mean accuracy) from the worst (37.5%); (3)training data alignment critically impacts generalization, causing up to 20–60% performance variance within architecturally identical detector families; (4)modern commercial generators (FluxDev, Fireflyv4, Midjourneyv7) defeat most detectors, achieving only 18–30% average accuracy; and (5)we identify three systematic failure patterns affecting cross-dataset generalization. Statistical analysis confirms significant performance differences between detectors (Friedman test: $χ^2$=121.01, $p<10^{-16}$, Kendall$W$=0.524). Our findings challenge the ``one-size-fits-all’’ detector paradigm and provide actionable deployment guidelines, demonstrating that practitioners must carefully select detectors based on their specific threat landscape rather than relying on published benchmark performance.

Method: Large-scale cross-paradigm benchmark evaluating 34 models (22 specialized architectures, 12 general-purpose VLMs) across 8 standard datasets totaling 1,100 test images per model, with analysis of MAE, age verification at 18-year threshold, and performance across 14 age groups.

Result: Zero-shot VLMs significantly outperform most specialized models (MAE 5.65 vs 9.88 years), with best VLM (Gemini 3 Flash Preview, MAE 4.32) beating best non-LLM model (MiVOLO, MAE 5.10) by 15%. VLMs achieve 13-25% false adult rates on minors vs 60-100% for non-LLM models.

Conclusion: Task-specific architectures may not be necessary for age estimation; field should redirect toward distilling VLM capabilities into efficient specialized models.

Abstract: Facial age estimation is critical for content moderation, age verification, and deepfake detection, yet no prior benchmark has systematically compared modern vision-language models (VLMs) against specialized age estimation architectures. We present the first large-scale cross-paradigm benchmark, evaluating \textbf{34 models} – 22 specialized architectures with publicly available pretrained weights and 12 general-purpose VLMs – across \textbf{8 standard datasets} (UTKFace, IMDB-WIKI, MORPH, AFAD, CACD, FG-NET, APPA-REAL, AgeDB) totaling 1{,}100 test images per model. Our key finding is striking: \emph{zero-shot VLMs significantly outperform most specialized models}, achieving an average MAE of 5.65 years compared to 9.88 for non-LLM models. The best VLM (Gemini3 Flash Preview, MAE4.32) outperforms the best non-LLM model (MiVOLO, MAE~5.10) by 15%. Only MiVOLO, which uniquely combines face and body features via Vision Transformers, competes with VLMs. We further analyze age verification at the 18-year threshold, revealing that non-LLM models exhibit 60–100% false adult rates on minors while VLMs achieve 13–25%, and demonstrate that coarse age binning (8–9 classes) consistently degrades MAE beyond 13 years. Our stratified analysis across 14 age groups reveals that all models struggle most at extreme ages ($<$5 and 65+). These findings challenge the assumption that task-specific architectures are necessary for age estimation and suggest that the field should redirect toward distilling VLM capabilities into efficient specialized models.

Method: Proposes operator-guided framework modeling degradation trajectory using known acquisition operators, with OCDI-Net (operator-conditional dual-stream interaction network) that disentangles target-slice content from inter-slice interference and predicts structured degradations for operator-aligned inversion. Uses two-stage chained inference: SMS slice separation followed by in-plane completion.

Result: Experiments on fastMRI brain data and prospectively acquired in vivo diffusion MRI data show improved fidelity and reduced slice leakage compared to conventional and learning-based SMS reconstructions.

Conclusion: Operator-guided approach better aligns with SMS physics than Gaussian-noise diffusion models, enabling more accurate reconstruction with reduced inter-slice interference through explicit modeling of acquisition operators.

Abstract: Simultaneous multi-slice (SMS) imaging with in-plane undersampling enables highly accelerated MRI but yields a strongly coupled inverse problem with deterministic inter-slice interference and missing k-space data. Most diffusion-based reconstructions are formulated around Gaussian-noise corruption and rely on additional consistency steps to incorporate SMS physics, which can be mismatched to the operator-governed degradations in SMS acquisition. We propose an operator-guided framework that models the degradation trajectory using known acquisition operators and inverts this process via deterministic updates. Within this framework, we introduce an operator-conditional dual-stream interaction network (OCDI-Net) that explicitly disentangles target-slice content from inter-slice interference and predicts structured degradations for operator-aligned inversion, and we instantiate reconstruction as a two-stage chained inference procedure that performs SMS slice separation followed by in-plane completion. Experiments on fastMRI brain data and prospectively acquired in vivo diffusion MRI data demonstrate improved fidelity and reduced slice leakage over conventional and learning-based SMS reconstructions.

Method: Proposes OTA-Det framework with: 1) Task reformulation strategy to unify objectives and supervision mechanisms for joint training across OVAD and RSVG datasets, 2) Dense semantic alignment strategy establishing explicit correspondence at multiple granularities (holistic expressions to individual attributes), 3) Extension of RT-DETR architecture with efficient modules for open-text detection.

Result: Achieves state-of-the-art performance on six benchmarks spanning both OVAD and RSVG tasks while maintaining real-time inference at 34 FPS.

Conclusion: OTA-Det successfully bridges two key aerial scene understanding paradigms, enabling simultaneous rich semantic understanding and multi-target detection with real-time efficiency, representing a significant advancement in aerial vision tasks.

Abstract: Open-Vocabulary Aerial Detection (OVAD) and Remote Sensing Visual Grounding (RSVG) have emerged as two key paradigms for aerial scene understanding. However, each paradigm suffers from inherent limitations when operating in isolation: OVAD is restricted to coarse category-level semantics, while RSVG is structurally limited to single-target localization. These limitations prevent existing methods from simultaneously supporting rich semantic understanding and multi-target detection. To address this, we propose OTA-Det, the first unified framework that bridges both paradigms into a cohesive architecture. Specifically, we introduce a task reformulation strategy that unifies task objectives and supervision mechanisms, enabling joint training across datasets from both paradigms with dense supervision signals. Furthermore, we propose a dense semantic alignment strategy that establishes explicit correspondence at multiple granularities, from holistic expressions to individual attributes, enabling fine-grained semantic understanding. To ensure real-time efficiency, OTA-Det builds upon the RT-DETR architecture, extending it from closed-set detection to open-text detection by introducing several high efficient modules, achieving state-of-the-art performance on six benchmarks spanning both OVAD and RSVG tasks while maintaining real-time inference at 34 FPS.

Method: Introduces SPD-Faith Bench based on fine-grained image difference reasoning to isolate faithfulness from linguistic priors, then proposes SAGE - a train-free visual evidence-calibrated framework to improve visual routing and align reasoning with perception.

Result: Evaluations reveal two systematic failure modes: perceptual blindness and perception-reasoning dissociation, traced to decaying visual attention and representation shifts in residual stream. SAGE framework shows improvements.

Conclusion: Highlights importance of explicitly evaluating faithfulness beyond response correctness in MLLMs, with benchmark and methods provided for better reasoning trace evaluation.

Abstract: Chain-of-Thought reasoning is widely used to improve the interpretability of multimodal large language models (MLLMs), yet the faithfulness of the generated reasoning traces remains unclear. Prior work has mainly focused on perceptual hallucinations, leaving reasoning level unfaithfulness underexplored. To isolate faithfulness from linguistic priors, we introduce SPD-Faith Bench, a diagnostic benchmark based on fine-grained image difference reasoning that enforces explicit visual comparison. Evaluations on state-of-the-art MLLMs reveal two systematic failure modes, perceptual blindness and perception-reasoning dissociation. We trace these failures to decaying visual attention and representation shifts in the residual stream. Guided by this analysis, we propose SAGE, a train-free visual evidence-calibrated framework that improves visual routing and aligns reasoning with perception. Our results highlight the importance of explicitly evaluating faithfulness beyond response correctness. Our benchmark and codes are available at https://github.com/Johanson-colab/SPD-Faith-Bench.

Method: Three key techniques: 1) Frequency Spectrum Attention Interpolation to preserve identity characteristics, 2) Target Structure Guidance via plug-and-play attention injection for structural alignment, and 3) Flow-Guided Attention Temporal Smoothening for spatiotemporal coherence without model modifications.

Result: Extensive experiments show VFace significantly enhances temporal consistency and visual fidelity compared to frame-wise generation approaches, offering a practical modular solution for video face swapping.

Conclusion: VFace provides an effective training-free, plug-and-play method for high-quality video face swapping that can be seamlessly integrated with existing diffusion-based image face swapping models, addressing temporal inconsistency issues without requiring additional training.

Abstract: We present a training-free, plug-and-play method, namely VFace, for high-quality face swapping in videos. It can be seamlessly integrated with image-based face swapping approaches built on diffusion models. First, we introduce a Frequency Spectrum Attention Interpolation technique to facilitate generation and intact key identity characteristics. Second, we achieve Target Structure Guidance via plug-and-play attention injection to better align the structural features from the target frame to the generation. Third, we present a Flow-Guided Attention Temporal Smoothening mechanism that enforces spatiotemporal coherence without modifying the underlying diffusion model to reduce temporal inconsistencies typically encountered in frame-wise generation. Our method requires no additional training or video-specific fine-tuning. Extensive experiments show that our method significantly enhances temporal consistency and visual fidelity, offering a practical and modular solution for video-based face swapping. Our code is available at https://github.com/Sanoojan/VFace.

Method: Introduces ViewRope, a geometry-aware encoding that injects camera-ray directions directly into video transformer self-attention layers. Also proposes Geometry-Aware Frame-Sparse Attention to selectively attend to relevant historical frames using geometric cues. Presents ViewBench diagnostic suite for measuring loop-closure fidelity and geometric drift.

Result: ViewRope substantially improves long-term consistency while reducing computational costs. The method demonstrates better spatial persistence and reduced hallucination when cameras revisit locations.

Conclusion: Geometry-aware encoding through ViewRope provides a model-native inductive bias for retrieving 3D-consistent content across temporal gaps, addressing fundamental limitations in current predictive world models for interactive AI.

Abstract: Predictive world models that simulate future observations under explicit camera control are fundamental to interactive AI. Despite rapid advances, current systems lack spatial persistence: they fail to maintain stable scene structures over long trajectories, frequently hallucinating details when cameras revisit previously observed locations. We identify that this geometric drift stems from reliance on screen-space positional embeddings, which conflict with the projective geometry required for 3D consistency. We introduce \textbf{ViewRope}, a geometry-aware encoding that injects camera-ray directions directly into video transformer self-attention layers. By parameterizing attention with relative ray geometry rather than pixel locality, ViewRope provides a model-native inductive bias for retrieving 3D-consistent content across temporal gaps. We further propose \textbf{Geometry-Aware Frame-Sparse Attention}, which exploits these geometric cues to selectively attend to relevant historical frames, improving efficiency without sacrificing memory consistency. We also present \textbf{ViewBench}, a diagnostic suite measuring loop-closure fidelity and geometric drift. Our results demonstrate that ViewRope substantially improves long-term consistency while reducing computational costs.

Method: Proposes an inverse rendering approach with a fast barycentric coordinate solver that reduces computational overhead by 4.57x, enabling efficient photorealistic simulation of high-speed motion. The method is fully differentiable, allowing gradient propagation from rendered images to 3D shape.

Result: Successfully recovers 3D shapes from 2D imagery of objects undergoing extreme translational and rotational motion, enabling efficient and realistic modeling of ultra-fast moving objects in forward simulation.

Conclusion: The method advances vision-based 3D reconstruction by enabling shape recovery from ultra-fast motion-blurred images, overcoming limitations of traditional techniques like Multi-View Stereo.

Abstract: We consider the problem of 3D shape recovery from ultra-fast motion-blurred images. While 3D reconstruction from static images has been extensively studied, recovering geometry from extreme motion-blurred images remains challenging. Such scenarios frequently occur in both natural and industrial settings, such as fast-moving objects in sports (e.g., balls) or rotating machinery, where rapid motion distorts object appearance and makes traditional 3D reconstruction techniques like Multi-View Stereo (MVS) ineffective. In this paper, we propose a novel inverse rendering approach for shape recovery from ultra-fast motion-blurred images. While conventional rendering techniques typically synthesize blur by averaging across multiple frames, we identify a major computational bottleneck in the repeated computation of barycentric weights. To address this, we propose a fast barycentric coordinate solver, which significantly reduces computational overhead and achieves a speedup of up to 4.57x, enabling efficient and photorealistic simulation of high-speed motion. Crucially, our method is fully differentiable, allowing gradients to propagate from rendered images to the underlying 3D shape, thereby facilitating shape recovery through inverse rendering. We validate our approach on two representative motion types: rapid translation and rotation. Experimental results demonstrate that our method enables efficient and realistic modeling of ultra-fast moving objects in the forward simulation. Moreover, it successfully recovers 3D shapes from 2D imagery of objects undergoing extreme translational and rotational motion, advancing the boundaries of vision-based 3D reconstruction. Project page: https://maxmilite.github.io/rec-from-ultrafast-blur/

Method: Created SSI-Bench through a human-centered pipeline: 10 researchers spent 400+ hours curating images of complex real-world 3D structures, annotating structural components, and designing 1,000 ranking questions that minimize pixel-level cues. Questions test geometric and topological reasoning requiring operations like mental rotation, cross-sectional inference, occlusion reasoning, and force-path reasoning.

Result: Evaluation of 31 VLMs shows large performance gaps: best open-source model achieves 22.2% accuracy, best closed-source model reaches 33.6%, while humans score 91.6%. Encouraging models to “think” yields only marginal gains. Error analysis reveals failures in structural grounding and constraint-consistent 3D reasoning.

Conclusion: SSI-Bench reveals significant limitations in current VLMs’ spatial reasoning capabilities, particularly in handling constrained 3D structures. The benchmark provides a rigorous testbed for advancing spatial intelligence in vision-language models.

Abstract: Spatial intelligence is crucial for vision–language models (VLMs) in the physical world, yet many benchmarks evaluate largely unconstrained scenes where models can exploit 2D shortcuts. We introduce SSI-Bench, a VQA benchmark for spatial reasoning on constrained manifolds, built from complex real-world 3D structures whose feasible configurations are tightly governed by geometric, topological, and physical constraints. SSI-Bench contains 1,000 ranking questions spanning geometric and topological reasoning and requiring a diverse repertoire of compositional spatial operations, such as mental rotation, cross-sectional inference, occlusion reasoning, and force-path reasoning. It is created via a fully human-centered pipeline: ten researchers spent over 400 hours curating images, annotating structural components, and designing questions to minimize pixel-level cues. Evaluating 31 widely used VLMs reveals a large gap to humans: the best open-source model achieves 22.2% accuracy and the strongest closed-source model reaches 33.6%, while humans score 91.6%. Encouraging models to think yields only marginal gains, and error analysis points to failures in structural grounding and constraint-consistent 3D reasoning. Project page: https://ssi-bench.github.io.

Method: Uses MedGemma-based structured report mining to generate global and region-level captions, processes wrist images with bone-specific crops (distal radius, distal ulna, ulnar styloid), jointly trains global and local contrastive encoders, and performs two-stage retrieval: coarse global matching followed by region-conditioned reranking.

Result: Improves image-to-text Recall@5 from 0.82% to 9.35%, yields stronger fracture classification (AUROC 0.949, AUPRC 0.953), improves retrieval-based fracture diagnosis (mean F1 from 0.568 to 0.753), and radiologists rate retrieved cases as more clinically relevant (mean scores from 3.36 to 4.35).

Conclusion: Anatomically guided retrieval has potential to enhance diagnostic reasoning and support clinical decision-making in pediatric musculoskeletal imaging, demonstrating the value of region-aware approaches for medical image retrieval.

Abstract: Retrieving wrist radiographs with analogous fracture patterns is challenging because clinically important cues are subtle, highly localized and often obscured by overlapping anatomy or variable imaging views. Progress is further limited by the scarcity of large, well-annotated datasets for case-based medical image retrieval. We introduce WristMIR, a region-aware pediatric wrist radiograph retrieval framework that leverages dense radiology reports and bone-specific localization to learn fine-grained, clinically meaningful image representations without any manual image-level annotations. Using MedGemma-based structured report mining to generate both global and region-level captions, together with pre-processed wrist images and bone-specific crops of the distal radius, distal ulna, and ulnar styloid, WristMIR jointly trains global and local contrastive encoders and performs a two-stage retrieval process: (1) coarse global matching to identify candidate exams, followed by (2) region-conditioned reranking aligned to a predefined anatomical bone region. WristMIR improves retrieval performance over strong vision-language baselines, raising image-to-text Recall@5 from 0.82% to 9.35%. Its embeddings also yield stronger fracture classification (AUROC 0.949, AUPRC 0.953). In region-aware evaluation, the two-stage design markedly improves retrieval-based fracture diagnosis, increasing mean $F_1$ from 0.568 to 0.753, and radiologists rate its retrieved cases as more clinically relevant, with mean scores rising from 3.36 to 4.35. These findings highlight the potential of anatomically guided retrieval to enhance diagnostic reasoning and support clinical decision-making in pediatric musculoskeletal imaging. The source code is publicly available at https://github.com/quin-med-harvard-edu/WristMIR.

Method: SAGE uses hierarchical mining to transform videos into training trajectories with hybrid supervision: (1) informative trajectory selection, (2) sparse geometric anchoring via SfM point clouds for global structure, and (3) dense differentiable consistency via 3D Gaussian rendering for multi-view constraints. Includes regularization to prevent catastrophic forgetting.

Result: Significantly enhances zero-shot generalization, reducing Chamfer Distance by 20-42% on unseen benchmarks (7Scenes, TUM-RGBD, Matterport3D) compared to state-of-the-art baselines.

Conclusion: SAGE pioneers adaptation of geometric foundation models via Internet video, establishing a scalable paradigm for general-purpose 3D learning.

Abstract: Geometric foundation models show promise in 3D reconstruction, yet their progress is severely constrained by the scarcity of diverse, large-scale 3D annotations. While Internet videos offer virtually unlimited raw data, utilizing them as a scaling source for geometric learning is challenging due to the absence of ground-truth geometry and the presence of observational noise. To address this, we propose SAGE, a framework for Scalable Adaptation of GEometric foundation models from raw video streams. SAGE leverages a hierarchical mining pipeline to transform videos into training trajectories and hybrid supervision: (1) Informative training trajectory selection; (2) Sparse Geometric Anchoring via SfM point clouds for global structural guidance; and (3) Dense Differentiable Consistency via 3D Gaussian rendering for multi-view constraints. To prevent catastrophic forgetting, we introduce a regularization strategy using anchor data. Extensive experiments show that SAGE significantly enhances zero-shot generalization, reducing Chamfer Distance by 20-42% on unseen benchmarks (7Scenes, TUM-RGBD, Matterport3D) compared to state-of-the-art baselines. To our knowledge, SAGE pioneers the adaptation of geometric foundation models via Internet video, establishing a scalable paradigm for general-purpose 3D learning.

Method: Proposes TLQ framework with: 1) gradient-guided token-level importance integration mechanism for quantization error, 2) token-level calibration set construction for fine-grained calibration, 3) multi-GPU quantization-exposed layer-wise calibration scheme that distributes workload across RTX3090 GPUs.

Result: Evaluated across two models, three model scales, and two quantization settings, consistently achieving performance improvements across all settings, demonstrating strong quantization stability.

Conclusion: TLQ provides an effective PTQ solution for VLMs that addresses token distribution differences through fine-grained importance-aware calibration and practical distributed implementation.

Abstract: Post-training quantization (PTQ) is a primary approach for deploying large language models without fine-tuning, and the quantized performance is often strongly affected by the calibration in PTQ. By contrast, in vision-language models (VLMs), substantial differences between visual and text tokens in their activation distributions and sensitivities to quantization error pose significant challenges for effective calibration during PTQ. In this work, we rethink what PTQ calibration should align with in VLMs and propose the Token-level Importance-aware Layer-wise Quantization framework (TLQ). Guided by gradient information, we design a token-level importance integration mechanism for quantization error, and use it to construct a token-level calibration set, enabling a more fine-grained calibration strategy. Furthermore, TLQ introduces a multi-GPU, quantization-exposed layer-wise calibration scheme. This scheme keeps the layer-wise calibration procedure consistent with the true quantized inference path and distributes the complex layer-wise calibration workload across multiple RTX3090 GPUs, thereby reducing reliance on the large memory of A100 GPUs. TLQ is evaluated across two models, three model scales, and two quantization settings, consistently achieving performance improvements across all settings, indicating its strong quantization stability. The code will be released publicly.

Method: Conducted zero-shot evaluation of open-source VLMs for privacy attribute recognition, identified attributes with strong inter-annotator agreement among VLMs, and analyzed disagreement cases between human and VLM annotations.

Result: VLMs tend to predict the presence of privacy attributes more frequently than human annotators. In cases of high inter-annotator agreement between VLMs, they can complement human annotation by identifying attributes that human annotators overlook.

Conclusion: VLMs have potential to support privacy annotations in large-scale image datasets, particularly when they show high inter-annotator agreement, as they can identify privacy attributes that human annotators might miss.

Abstract: Visual Language Models (VLMs) are often used for zero-shot detection of visual attributes in the image. We present a zero-shot evaluation of open-source VLMs for privacy-related attribute recognition. We identify the attributes for which VLMs exhibit strong inter-annotator agreement, and discuss the disagreement cases of human and VLM annotations. Our results show that when evaluated against human annotations, VLMs tend to predict the presence of privacy attributes more often than human annotators. In addition to this, we find that in cases of high inter-annotator agreement between VLMs, they can complement human annotation by identifying attributes overlooked by human annotators. This highlights the potential of VLMs to support privacy annotations in large-scale image datasets.

Method: Proposes a unified framework using Dynamic Occupancy Grid Maps within a streamlined temporal decoding pipeline. Uses a lightweight spatiotemporal backbone with a tailored interdependent loss function that captures inter-grid dependencies and enables diverse future predictions. The loss function uses occupancy state information to enforce flow-guided transitions, acting as a regularizer for occupancy evolution while accounting for obstacles and occlusions.

Result: Evaluations on real-world nuScenes and Woven Planet datasets demonstrate superior prediction performances for dynamic vehicles and generic dynamic scene elements compared to baseline methods.

Conclusion: The proposed unified framework successfully combines agent-agnostic and agent-specific prediction approaches, enabling robust prediction of both specific vehicle behaviors and other dynamic entities within complex driving scenes.

Abstract: Accurate prediction of driving scene is a challenging task due to uncertainty in sensor data, the complex behaviors of agents, and the possibility of multiple feasible futures. Existing prediction methods using occupancy grid maps primarily focus on agent-agnostic scene predictions, while agent-specific predictions provide specialized behavior insights with the help of semantic information. However, both paradigms face distinct limitations: agent-agnostic models struggle to capture the behavioral complexities of dynamic actors, whereas agent-specific approaches fail to generalize to poorly perceived or unrecognized agents; combining both enables robust and safer motion forecasting. To address this, we propose a unified framework by leveraging Dynamic Occupancy Grid Maps within a streamlined temporal decoding pipeline to simultaneously predict future occupancy state grids, vehicle grids, and scene flow grids. Relying on a lightweight spatiotemporal backbone, our approach is centered on a tailored, interdependent loss function that captures inter-grid dependencies and enables diverse future predictions. By using occupancy state information to enforce flow-guided transitions, the loss function acts as a regularizer that directs occupancy evolution while accounting for obstacles and occlusions. Consequently, the model not only predicts the specific behaviors of vehicle agents, but also identifies other dynamic entities and anticipates their evolution within the complex scene. Evaluations on real-world nuScenes and Woven Planet datasets demonstrate superior prediction performances for dynamic vehicles and generic dynamic scene elements compared to baseline methods.

Method: Proposes multiple local density learners to capture different density distributions via prototypes, encodes local density similarity matrices, and extracts global density features from support images to guide adaptation

Result: Outperforms recent state-of-the-art methods on three surveillance datasets for few-shot crowd counting

Conclusion: The approach effectively adapts crowd counting models to unseen surveillance scenes by leveraging both local and global density characteristics through few-shot learning

Abstract: Crowd scenes captured by cameras at different locations vary greatly, and existing crowd models have limited generalization for unseen surveillance scenes. To improve the generalization of the model, we regard different surveillance scenes as different category scenes, and introduce few-shot learning to make the model adapt to the unseen surveillance scene that belongs to the given exemplar category scene. To this end, we propose to leverage local and global density characteristics to guide the model of crowd counting for unseen surveillance scenes. Specifically, to enable the model to adapt to the varying density variations in the target scene, we propose the multiple local density learner to learn multi prototypes which represent different density distributions in the support scene. Subsequently, these multiple local density similarity matrixes are encoded. And they are utilized to guide the model in a local way. To further adapt to the global density in the target scene, the global density features are extracted from the support image, then it is used to guide the model in a global way. Experiments on three surveillance datasets shows that proposed method can adapt to the unseen surveillance scene and outperform recent state-of-the-art methods in the few-shot crowd counting.

Method: Developed D-ORCA (dialogue-centric omni-modal LLM) trained on DVD dataset (40K videos). Used group relative policy optimization with three novel reward functions: speaker attribution accuracy, global speech content accuracy, and sentence-level temporal boundary alignment.

Result: D-ORCA substantially outperforms existing open-source models in speaker identification, speech recognition, and temporal grounding. Despite only 8B parameters, achieves performance competitive with Qwen3-Omni on general audio-visual understanding benchmarks.

Conclusion: D-ORCA advances audio-visual captioning for dialogue videos through novel reinforcement learning objectives and large-scale bilingual dataset, enabling accurate speaker attribution and temporal alignment.

Abstract: Spoken dialogue is a primary source of information in videos; therefore, accurately identifying who spoke what and when is essential for deep video understanding. We introduce D-ORCA, a \textbf{d}ialogue-centric \textbf{o}mni-modal large language model optimized for \textbf{r}obust audio-visual \textbf{ca}ptioning. We further curate DVD, a large-scale, high-quality bilingual dataset comprising nearly 40,000 multi-party dialogue videos for training and 2000 videos for evaluation in English and Mandarin, addressing a critical gap in the open-source ecosystem. To ensure fine-grained captioning accuracy, we adopt group relative policy optimization with three novel reward functions that assess speaker attribution accuracy, global speech content accuracy, and sentence-level temporal boundary alignment. These rewards are derived from evaluation metrics widely used in speech processing and, to our knowledge, are applied for the first time as reinforcement learning objectives for audio-visual captioning. Extensive experiments demonstrate that D-ORCA substantially outperforms existing open-source models in speaker identification, speech recognition, and temporal grounding. Notably, despite having only 8 billion parameters, D-ORCA achieves performance competitive with Qwen3-Omni across several general-purpose audio-visual understanding benchmarks. Demos are available at \href{https://d-orca-llm.github.io/}{https://d-orca-llm.github.io/}. Our code, data, and checkpoints will be available at \href{https://github.com/WeChatCV/D-ORCA/}{https://github.com/WeChatCV/D-ORCA/}.

Method: Proposes EasyTune which fine-tunes diffusion models at each denoising step individually, decoupling recursive dependencies. Also introduces Self-refinement Preference Learning (SPL) to dynamically identify preference pairs when training data is scarce.

Result: Outperforms DRaFT-50 by 8.2% in alignment (MM-Dist) improvement while requiring only 31.16% of its memory overhead and achieving 7.3x training speedup.

Conclusion: EasyTune provides an efficient, memory-friendly approach to aligning diffusion models with downstream objectives, particularly valuable for motion generation where preference data is limited.

Abstract: In recent years, motion generative models have undergone significant advancement, yet pose challenges in aligning with downstream objectives. Recent studies have shown that using differentiable rewards to directly align the preference of diffusion models yields promising results. However, these methods suffer from (1) inefficient and coarse-grained optimization with (2) high memory consumption. In this work, we first theoretically and empirically identify the key reason of these limitations: the recursive dependence between different steps in the denoising trajectory. Inspired by this insight, we propose EasyTune, which fine-tunes diffusion at each denoising step rather than over the entire trajectory. This decouples the recursive dependence, allowing us to perform (1) a dense and fine-grained, and (2) memory-efficient optimization. Furthermore, the scarcity of preference motion pairs restricts the availability of motion reward model training. To this end, we further introduce a Self-refinement Preference Learning (SPL) mechanism that dynamically identifies preference pairs and conducts preference learning. Extensive experiments demonstrate that EasyTune outperforms DRaFT-50 by 8.2% in alignment (MM-Dist) improvement while requiring only 31.16% of its additional memory overhead and achieving a 7.3x training speedup. The project page is available at this link {https://xiaofeng-tan.github.io/projects/EasyTune/index.html}.

Method: Three core strategies: 1) Complementary feature construction integrating direct image reconstructions with projection-domain denoised results, 2) Full-spectrum prior integration fusing multi-energy projections into high-SNR reference, 3) Efficient latent diffusion synthesis embedding multi-path features into compact latent space for accelerated reconstruction.

Result: Extensive experiments on simulated and real-world datasets show FSP-Diff significantly outperforms state-of-the-art methods in both image quality and computational efficiency.

Conclusion: FSP-Diff demonstrates potential for clinically viable ultra-low-dose spectral CT imaging by effectively balancing detail fidelity and noise suppression through its integrated framework.

Abstract: Spectral computed tomography (CT) with photon-counting detectors holds immense potential for material discrimination and tissue characterization. However, under ultra-low-dose conditions, the sharply degraded signal-to-noise ratio (SNR) in energy-specific projections poses a significant challenge, leading to severe artifacts and loss of structural details in reconstructed images. To address this, we propose FSP-Diff, a full-spectrum prior-enhanced dual-domain latent diffusion framework for ultra-low-dose spectral CT reconstruction. Our framework integrates three core strategies: 1) Complementary Feature Construction: We integrate direct image reconstructions with projection-domain denoised results. While the former preserves latent textural nuances amidst heavy noise, the latter provides a stable structural scaffold to balance detail fidelity and noise suppression. 2) Full-Spectrum Prior Integration: By fusing multi-energy projections into a high-SNR full-spectrum image, we establish a unified structural reference that guides the reconstruction across all energy bins. 3) Efficient Latent Diffusion Synthesis: To alleviate the high computational burden of high-dimensional spectral data, multi-path features are embedded into a compact latent space. This allows the diffusion process to facilitate interactive feature fusion in a lower-dimensional manifold, achieving accelerated reconstruction while maintaining fine-grained detail restoration. Extensive experiments on simulated and real-world datasets demonstrate that FSP-Diff significantly outperforms state-of-the-art methods in both image quality and computational efficiency, underscoring its potential for clinically viable ultra-low-dose spectral CT imaging.

Method: Proposes Continuity-driven Synergistic Diffusion with Neural priors (CSDN): 1) Neural priors encode continuous 3D attenuation representation to synthesize dense projections from sparse measurements; 2) Synergistic diffusion with two collaborative paths: Sinogram Refinement Diffusion (Sino-RD) for angular continuity and Digital Radiography Refinement Diffusion (DR-RD) for inter-slice consistency; 3) Dual-Projection Reconstruction Fusion (DPRF) module adaptively fuses outputs for coherent volumetric reconstruction.

Result: Extensive experiments show CSDN effectively suppresses artifacts and recovers fine textures under ultra-sparse-view conditions, outperforming existing state-of-the-art techniques.

Conclusion: CSDN framework successfully addresses ultra-sparse-view CBCT reconstruction challenges by combining neural priors with synergistic diffusion, achieving superior artifact suppression and texture recovery compared to existing methods.

Abstract: The clinical application of cone-beam computed tomography (CBCT) is constrained by the inherent trade-off between radiation exposure and image quality. Ultra-sparse angular sampling, employed to reduce dose, introduces severe undersampling artifacts and inter-slice inconsistencies, compromising diagnostic reliability. Existing reconstruction methods often struggle to balance angular continuity with spatial detail fidelity. To address these challenges, we propose a Continuity-driven Synergistic Diffusion with Neural priors (CSDN) for ultra-sparse-view CBCT reconstruction. Neural priors are introduced as a structural foundation to encode a continuous threedimensional attenuation representation, enabling the synthesis of physically consistent dense projections from ultra-sparse measurements. Building upon this neural-prior-based initialization, a synergistic diffusion strategy is developed, consisting of two collaborative refinement paths: a Sinogram Refinement Diffusion (Sino-RD) process that restores angular continuity and a Digital Radiography Refinement Diffusion (DR-RD) process that enforces inter-slice consistency from the projection image perspective. The outputs of the two diffusion paths are adaptively fused by the Dual-Projection Reconstruction Fusion (DPRF) module to achieve coherent volumetric reconstruction. Extensive experiments demonstrate that the proposed CSDN effectively suppresses artifacts and recovers fine textures under ultra-sparse-view conditions, outperforming existing state-of-the-art techniques.

Method: Conducted comprehensive empirical analysis of state-of-the-art deepfake detection techniques, including human evaluation experiments against cutting-edge synthesis methods. Used extensive experimentation to compare detection models with modern generation technologies.

Result: Many state-of-the-art detection models exhibit markedly poor performance when challenged with deepfakes produced by modern synthesis techniques. Human participants also performed poorly against the best quality deepfakes, revealing a significant gap between detection capabilities and generation sophistication.

Conclusion: There’s an urgent need for continued refinement of detection models to keep pace with evolving deepfake generation technologies. The research emphasizes the critical gap between current detection methodologies and new generation techniques, calling for intensified research efforts in this area.

Abstract: The rapid advancement of deepfake technology has significantly elevated the realism and accessibility of synthetic media. Emerging techniques, such as diffusion-based models and Neural Radiance Fields (NeRF), alongside enhancements in traditional Generative Adversarial Networks (GANs), have contributed to the sophisticated generation of deepfake videos. Concurrently, deepfake detection methods have seen notable progress, driven by innovations in Transformer architectures, contrastive learning, and other machine learning approaches. In this study, we conduct a comprehensive empirical analysis of state-of-the-art deepfake detection techniques, including human evaluation experiments against cutting-edge synthesis methods. Our findings highlight a concerning trend: many state-of-the-art detection models exhibit markedly poor performance when challenged with deepfakes produced by modern synthesis techniques, including poor performance by human participants against the best quality deepfakes. Through extensive experimentation, we provide evidence that underscores the urgent need for continued refinement of detection models to keep pace with the evolving capabilities of deepfake generation technologies. This research emphasizes the critical gap between current detection methodologies and the sophistication of new generation techniques, calling for intensified efforts in this crucial area of study.

Method: Proposes MCIE-E1 with two key modules: spatial-aware cross-attention (aligns semantic instructions with spatial regions using spatial guidance during denoising) and background-consistent cross-attention (preserves features in unedited regions). Uses a dedicated data pipeline combining fine-grained automatic filtering via MLLM with human validation to address dataset scarcity.

Result: Experimental results on CIE-Bench show MCIE-E1 consistently outperforms previous SOTA methods in both quantitative and qualitative assessments, achieving 23.96% improvement in instruction compliance.

Conclusion: MCIE-E1 addresses limitations in complex instruction-based image editing through novel architectural modules, curated data pipeline, and comprehensive evaluation benchmark, demonstrating significant improvements in instruction compliance and background consistency.

Abstract: Recent advances in instruction-based image editing have shown remarkable progress. However, existing methods remain limited to relatively simple editing operations, hindering real-world applications that require complex and compositional instructions. In this work, we address these limitations from the perspectives of architectural design, data, and evaluation protocols. Specifically, we identify two key challenges in current models: insufficient instruction compliance and background inconsistency. To this end, we propose MCIE-E1, a Multimodal Large Language Model-Driven Complex Instruction Image Editing method that integrates two key modules: a spatial-aware cross-attention module and a background-consistent cross-attention module. The former enhances instruction-following capability by explicitly aligning semantic instructions with spatial regions through spatial guidance during the denoising process, while the latter preserves features in unedited regions to maintain background consistency. To enable effective training, we construct a dedicated data pipeline to mitigate the scarcity of complex instruction-based image editing datasets, combining fine-grained automatic filtering via a powerful MLLM with rigorous human validation. Finally, to comprehensively evaluate complex instruction-based image editing, we introduce CIE-Bench, a new benchmark with two new evaluation metrics. Experimental results on CIE-Bench demonstrate that MCIE-E1 consistently outperforms previous state-of-the-art methods in both quantitative and qualitative assessments, achieving a 23.96% improvement in instruction compliance.

Method: Uses Attention and Diversity-based Token Selection (ADTS) to select representative tokens for basic video representation, then applies Tree-based Spatiotemporal Token Merging (TSTM) for fine-grained spatiotemporal redundancy elimination.

Result: Retaining only 10% of visual tokens preserves 99.1% of LLaVA-OneVision’s performance. Enables 10x increase in video frame input to Qwen2.5-VL with 8.6% relative improvement within same computational budget.

Conclusion: FlashVID provides effective training-free acceleration for VLLMs by better exploiting spatiotemporal relationships, enabling longer video processing with minimal performance loss.

Abstract: Although Video Large Language Models (VLLMs) have shown remarkable capabilities in video understanding, they are required to process high volumes of visual tokens, causing significant computational inefficiency. Existing VLLMs acceleration frameworks usually compress spatial and temporal redundancy independently, which overlooks the spatiotemporal relationships, thereby leading to suboptimal spatiotemporal compression. The highly correlated visual features are likely to change in spatial position, scale, orientation, and other attributes over time due to the dynamic nature of video. Building on this insight, we introduce FlashVID, a training-free inference acceleration framework for VLLMs. Specifically, FlashVID utilizes Attention and Diversity-based Token Selection (ADTS) to select the most representative tokens for basic video representation, then applies Tree-based Spatiotemporal Token Merging (TSTM) for fine-grained spatiotemporal redundancy elimination. Extensive experiments conducted on three representative VLLMs across five video understanding benchmarks demonstrate the effectiveness and generalization of our method. Notably, by retaining only 10% of visual tokens, FlashVID preserves 99.1% of the performance of LLaVA-OneVision. Consequently, FlashVID can serve as a training-free and plug-and-play module for extending long video frames, which enables a 10x increase in video frame input to Qwen2.5-VL, resulting in a relative improvement of 8.6% within the same computational budget. Code is available at https://github.com/Fanziyang-v/FlashVID.

Method: Proposes ForecastOcc with novel architecture: temporal cross-attention forecasting module, 2D-to-3D view transformer, 3D encoder for occupancy prediction, and semantic occupancy head for voxel-level forecasts across multiple horizons. Evaluated in multi-view (Occ3D-nuScenes) and monocular (SemanticKITTI) settings.

Result: Outperforms adapted baselines on both datasets, establishing first benchmark for monocular semantic occupancy forecasting on SemanticKITTI. Produces semantically rich, future-aware predictions capturing scene dynamics and semantics.

Conclusion: ForecastOcc enables joint semantic occupancy forecasting directly from images, addressing limitations of existing methods and providing critical scene understanding for autonomous driving applications.

Abstract: Autonomous driving requires forecasting both geometry and semantics over time to effectively reason about future environment states. Existing vision-based occupancy forecasting methods focus on motion-related categories such as static and dynamic objects, while semantic information remains largely absent. Recent semantic occupancy forecasting approaches address this gap but rely on past occupancy predictions obtained from separate networks. This makes current methods sensitive to error accumulation and prevents learning spatio-temporal features directly from images. In this work, we present ForecastOcc, the first framework for vision-based semantic occupancy forecasting that jointly predicts future occupancy states and semantic categories. Our framework yields semantic occupancy forecasts for multiple horizons directly from past camera images, without relying on externally estimated maps. We evaluate ForecastOcc in two complementary settings: multi-view forecasting on the Occ3D-nuScenes dataset and monocular forecasting on SemanticKITTI, where we establish the first benchmark for this task. We introduce the first baselines by adapting two 2D forecasting modules within our framework. Importantly, we propose a novel architecture that incorporates a temporal cross-attention forecasting module, a 2D-to-3D view transformer, a 3D encoder for occupancy prediction, and a semantic occupancy head for voxel-level forecasts across multiple horizons. Extensive experiments on both datasets show that ForecastOcc consistently outperforms baselines, yielding semantically rich, future-aware predictions that capture scene dynamics and semantics critical for autonomous driving.

Method: Fine-tuned Ultrasound Self-Supervised Foundation Model with Masked Autoencoding (USF-MAE) pretrained on 370k unlabelled ultrasound images for binary classification of normal vs cystic hygroma cases.

Result: USF-MAE outperformed DenseNet-169 baseline with accuracy 0.96 vs 0.93, sensitivity 0.94 vs 0.92, specificity 0.98 vs 0.94, ROC-AUC 0.98 vs 0.94; improvements statistically significant (p=0.0057).

Conclusion: Ultrasound-specific self-supervised pretraining enables accurate, robust deep learning detection of cystic hygroma, supporting scalable early screening programs.

Abstract: Cystic hygroma is a high-risk prenatal ultrasound finding that portends high rates of chromosomal abnormalities, structural malformations, and adverse pregnancy outcomes. Automated detection can increase reproducibility and support scalable early screening programs, but supervised deep learning methods are limited by small labelled datasets. This study assesses whether ultrasound-specific self-supervised pretraining can facilitate accurate, robust deep learning detection of cystic hygroma in first-trimester ultrasound images. We fine-tuned the Ultrasound Self-Supervised Foundation Model with Masked Autoencoding (USF-MAE), pretrained on over 370,000 unlabelled ultrasound images, for binary classification of normal controls and cystic hygroma cases used in this study. Performance was evaluated on the same curated ultrasound dataset, preprocessing pipeline, and 4-fold cross-validation protocol as for the DenseNet-169 baseline, using accuracy, sensitivity, specificity, and the area under the receiver operating characteristic curve (ROC-AUC). Model interpretability was analyzed qualitatively using Score-CAM visualizations. USF-MAE outperformed the DenseNet-169 baseline on all evaluation metrics. The proposed model yielded a mean accuracy of 0.96, sensitivity of 0.94, specificity of 0.98, and ROC-AUC of 0.98 compared to 0.93, 0.92, 0.94, and 0.94 for the DenseNet-169 baseline, respectively. Qualitative Score-CAM visualizations of model predictions demonstrated clinical relevance by highlighting expected regions in the fetal neck for both positive and negative cases. Paired statistical analysis using a Wilcoxon signed-rank test confirmed that performance improvements achieved by USF-MAE were statistically significant (p = 0.0057).

Method: PhysDrape integrates neural inference with explicit geometric solvers in a fully differentiable pipeline: 1) Physics-Informed Graph Neural Network conditioned on physics-enriched graph predicts residual displacements, 2) Learnable Force Solver resolves unbalanced forces from StVK model for quasi-static equilibrium, 3) Differentiable Projection strictly enforces collision constraints against body surface.

Result: PhysDrape achieves state-of-the-art performance with negligible interpenetration and significantly lower strain energy compared to baselines, demonstrating superior physical fidelity and robustness in real-time garment draping.

Conclusion: The hybrid neural-physical approach successfully resolves the conflict between geometric feasibility and physical plausibility in garment draping by guaranteeing physical validity through explicit constraints while enabling end-to-end learning for physically consistent predictions.

Abstract: Deep learning-based garment draping has emerged as a promising alternative to traditional Physics-Based Simulation (PBS), yet robust collision handling remains a critical bottleneck. Most existing methods enforce physical validity through soft penalties, creating an intrinsic trade-off between geometric feasibility and physical plausibility: penalizing collisions often distorts mesh structure, while preserving shape leads to interpenetration. To resolve this conflict, we present PhysDrape, a hybrid neural-physical solver for physically realistic garment draping driven by explicit forces and constraints. Unlike soft-constrained frameworks, PhysDrape integrates neural inference with explicit geometric solvers in a fully differentiable pipeline. Specifically, we propose a Physics-Informed Graph Neural Network conditioned on a physics-enriched graph – encoding material parameters and body proximity – to predict residual displacements. Crucially, we integrate a differentiable two-stage solver: first, a learnable Force Solver iteratively resolves unbalanced forces derived from the Saint Venant-Kirchhoff (StVK) model to ensure quasi-static equilibrium; second, a Differentiable Projection strictly enforces collision constraints against the body surface. This differentiable design guarantees physical validity through explicit constraints, while enabling end-to-end learning to optimize the network for physically consistent predictions. Extensive experiments demonstrate that PhysDrape achieves state-of-the-art performance, ensuring negligible interpenetration with significantly lower strain energy compared to existing baselines, achieving superior physical fidelity and robustness in real-time.

Method: Created MIND benchmark with 250 high-quality videos (100 first-person + 100 third-person clips with shared action space, plus 25+25 clips across varied action spaces) covering 8 diverse scenes. Designed evaluation framework measuring memory consistency (temporal stability and contextual coherence) and action control. Also introduced MIND-World as an interactive Video-to-World baseline for benchmarking.

Result: Extensive experiments demonstrate MIND’s completeness and reveal key challenges in current world models: difficulty maintaining long-term memory consistency and generalizing across different action spaces.

Conclusion: MIND provides the first open-domain closed-loop revisited benchmark for evaluating world models’ core abilities, highlighting current limitations and providing a foundation for future research in this area.

Abstract: World models aim to understand, remember, and predict dynamic visual environments, yet a unified benchmark for evaluating their fundamental abilities remains lacking. To address this gap, we introduce MIND, the first open-domain closed-loop revisited benchmark for evaluating Memory consIstency and action coNtrol in worlD models. MIND contains 250 high-quality videos at 1080p and 24 FPS, including 100 (first-person) + 100 (third-person) video clips under a shared action space and 25 + 25 clips across varied action spaces covering eight diverse scenes. We design an efficient evaluation framework to measure two core abilities: memory consistency and action control, capturing temporal stability and contextual coherence across viewpoints. Furthermore, we design various action spaces, including different character movement speeds and camera rotation angles, to evaluate the action generalization capability across different action spaces under shared scenes. To facilitate future performance benchmarking on MIND, we introduce MIND-World, a novel interactive Video-to-World baseline. Extensive experiments demonstrate the completeness of MIND and reveal key challenges in current world models, including the difficulty of maintaining long-term memory consistency and generalizing across action spaces. Project page: https://csu-jpg.github.io/MIND.github.io/

Method: Integrates deep 3D convolutional GANs with MoE framework, using multiple specialized generators and auxiliary loss-free dynamic capacity constraint mechanism for categorical generator selection.

Result: Model effectively generates and completes 3D shapes with varying missing regions, outperforming state-of-the-art approaches in both quantitative and qualitative evaluations.

Conclusion: MoE-DCGAN successfully handles complex 3D data challenges, balancing specialization, training stability, and computational efficiency for 3D voxel processing.

Abstract: The generation and completion of 3D objects represent a transformative challenge in computer vision. Generative Adversarial Networks (GANs) have recently demonstrated strong potential in synthesizing realistic visual data. However, they often struggle to capture complex and diverse data distributions, particularly in scenarios involving incomplete inputs or significant missing regions. These challenges arise mainly from the high computational requirements and the difficulty of modeling heterogeneous and structurally intricate data, which restrict their applicability in real-world settings. Mixture of Experts (MoE) models have emerged as a promising solution to these limitations. By dynamically selecting and activating the most relevant expert sub-networks for a given input, MoEs improve both performance and efficiency. In this paper, we investigate the integration of Deep 3D Convolutional GANs (CGANs) with a MoE framework to generate high-quality 3D models and reconstruct incomplete or damaged objects. The proposed architecture incorporates multiple generators, each specialized to capture distinct modalities within the dataset. Furthermore, an auxiliary loss-free dynamic capacity constraint (DCC) mechanism is introduced to guide the selection of categorical generators, ensuring a balance between specialization, training stability, and computational efficiency, which is critical for 3D voxel processing. We evaluated the model’s ability to generate and complete shapes with missing regions of varying sizes and compared its performance with state-of-the-art approaches. Both quantitative and qualitative results confirm the effectiveness of the proposed MoE-DCGAN in handling complex 3D data.

Method: Proposes a straightforward framework that systematically renders key ViT components equivariant: patch embedding, self-attention, positional encodings, and down/up-sampling. The approach constructs ViTs with guaranteed equivariance and serves as a plug-and-play replacement that scales to architectures like Swin Transformers.

Result: Extensive experiments show that the equivariant ViTs consistently improve performance and data efficiency across a wide spectrum of vision tasks.

Conclusion: The proposed framework provides a theoretically grounded and practically versatile approach to building equivariant Vision Transformers that effectively incorporate symmetry priors as inductive biases.

Abstract: Incorporating symmetry priors as inductive biases to design equivariant Vision Transformers (ViTs) has emerged as a promising avenue for enhancing their performance. However, existing equivariant ViTs often struggle to balance performance with equivariance, primarily due to the challenge of achieving holistic equivariant modifications across the diverse modules in ViTs-particularly in harmonizing the Self-Attention mechanism with Patch Embedding. To address this, we propose a straightforward framework that systematically renders key ViT components, including patch embedding, self-attention, positional encodings, and Down/Up-Sampling, equivariant, thereby constructing ViTs with guaranteed equivariance. The resulting architecture serves as a plug-and-play replacement that is both theoretically grounded and practically versatile, scaling seamlessly even to Swin Transformers. Extensive experiments demonstrate that our equivariant ViTs consistently improve performance and data efficiency across a wide spectrum of vision tasks.

Method: Uses YOLO for face detection, DINOv2 for visual features, Gemini 2.5 Pro with CoT+Reflection for pseudo-label generation, OpenPose for key-points with MLP backbone, and Transformer encoding with BERT for text. Combines pre-training, fine-tuning, and pseudo-label integration.

Result: Achieves over 0.69 accuracy on iMiGUE dataset (improving from under 0.6), establishes new public benchmark, and shows MLP key-point backbone can match or surpass GCN-based methods.

Conclusion: The proposed multimodal weak-supervision framework effectively recognizes concealed emotions, demonstrates the viability of simplified MLP backbones for key-point modeling, and shows the value of pseudo-label generation through LLM reasoning.

Abstract: To tackle the automatic recognition of “concealed emotions” in videos, this paper proposes a multimodal weak-supervision framework and achieves state-of-the-art results on the iMiGUE tennis-interview dataset. First, YOLO 11x detects and crops human portraits frame-by-frame, and DINOv2-Base extracts visual features from the cropped regions. Next, by integrating Chain-of-Thought and Reflection prompting (CoT + Reflection), Gemini 2.5 Pro automatically generates pseudo-labels and reasoning texts that serve as weak supervision for downstream models. Subsequently, OpenPose produces 137-dimensional key-point sequences, augmented with inter-frame offset features; the usual graph neural network backbone is simplified to an MLP to efficiently model the spatiotemporal relationships of the three key-point streams. An ultra-long-sequence Transformer independently encodes both the image and key-point sequences, and their representations are concatenated with BERT-encoded interview transcripts. Each modality is first pre-trained in isolation, then fine-tuned jointly, with pseudo-labeled samples merged into the training set for further gains. Experiments demonstrate that, despite severe class imbalance, the proposed approach lifts accuracy from under 0.6 in prior work to over 0.69, establishing a new public benchmark. The study also validates that an “MLP-ified” key-point backbone can match - or even surpass - GCN-based counterparts in this task.

Method: Physics-constrained reconstruction pipeline that builds multi-object scene reconstructions by considering geometry, non-penetration, and physics. Uses fast rejection sampling that reasons over multi-object interactions, leveraging an inferred object contact graph to guide samples.

Result: Outperforms state-of-the-art on both the new Picasso dataset (10 contact-rich real-world scenes) and YCB-V dataset, providing reconstructions that are physically plausible and more aligned with human intuition.

Conclusion: Object pose and shape estimation requires holistic scene reasoning accounting for object interactions and physical plausibility, which Picasso successfully achieves through physics-constrained reconstruction.

Abstract: In the presence of occlusions and measurement noise, geometrically accurate scene reconstructions – which fit the sensor data – can still be physically incorrect. For instance, when estimating the poses and shapes of objects in the scene and importing the resulting estimates into a simulator, small errors might translate to implausible configurations including object interpenetration or unstable equilibrium. This makes it difficult to predict the dynamic behavior of the scene using a digital twin, an important step in simulation-based planning and control of contact-rich behaviors. In this paper, we posit that object pose and shape estimation requires reasoning holistically over the scene (instead of reasoning about each object in isolation), accounting for object interactions and physical plausibility. Towards this goal, our first contribution is Picasso, a physics-constrained reconstruction pipeline that builds multi-object scene reconstructions by considering geometry, non-penetration, and physics. Picasso relies on a fast rejection sampling method that reasons over multi-object interactions, leveraging an inferred object contact graph to guide samples. Second, we propose the Picasso dataset, a collection of 10 contact-rich real-world scenes with ground truth annotations, as well as a metric to quantify physical plausibility, which we open-source as part of our benchmark. Finally, we provide an extensive evaluation of Picasso on our newly introduced dataset and on the YCB-V dataset, and show it largely outperforms the state of the art while providing reconstructions that are both physically plausible and more aligned with human intuition.

Method: DICE constructs contrastive triplets to distinguish style vs. non-style features in latent space, formalizing disentanglement as a generalized eigenvalue problem to identify style subspace. Uses Adaptive Attention Decoupling Editing to dynamically assess style concentration per token and perform differential suppression and content enhancement on QKV vectors.

Result: DICE achieves superior balance between thoroughness of style erasure and preservation of content integrity, with only 3 seconds additional overhead for style disentanglement.

Conclusion: DICE provides a practical and efficient technique for curbing style mimicry in diffusion models through training-free style purification that preserves user-intended content.

Abstract: The recent proliferation of diffusion models has made style mimicry effortless, enabling users to imitate unique artistic styles without authorization. In deployed platforms, this raises copyright and intellectual-property risks and calls for reliable protection. However, existing countermeasures either require costly weight editing as new styles emerge or rely on an explicitly specified editing style, limiting their practicality for deployment-side safety. To address this challenge, we propose DICE (Disentanglement of artist Style from Content via Contrastive Subspace Decomposition), a training-free framework for on-the-fly artist style erasure. Unlike style editing that require an explicitly specified replacement style, DICE performs style purification, removing the artist’s characteristics while preserving the user-intended content. Our core insight is that a model cannot truly comprehend the artist style from a single text or image alone. Consequently, we abandon the traditional paradigm of identifying style from isolated samples. Instead, we construct contrastive triplets to compel the model to distinguish between style and non-style features in the latent space. By formalizing this disentanglement process as a solvable generalized eigenvalue problem, we achieve precise identification of the style subspace. Furthermore, we introduce an Adaptive Attention Decoupling Editing strategy dynamically assesses the style concentration of each token and performs differential suppression and content enhancement on the QKV vectors. Extensive experiments demonstrate that DICE achieves a superior balance between the thoroughness of style erasure and the preservation of content integrity. DICE introduces an additional overhead of only 3 seconds to disentangle style, providing a practical and efficient technique for curbing style mimicry.

Method: ReRoPE leverages the insight that Rotary Positional Embeddings (RoPE) in existing models underutilize their full spectral bandwidth, especially low-frequency components. The method seamlessly injects relative camera pose information into these underutilized bands, enabling precise camera control while preserving pre-trained generative priors. This is a plug-and-play framework that doesn’t require retraining or architectural modifications.

Result: The method is evaluated on image-to-video (I2V) and video-to-video (V2V) tasks for camera control accuracy and visual fidelity. Results show ReRoPE offers a training-efficient path toward controllable, high-fidelity video generation with precise camera viewpoint control.

Conclusion: ReRoPE provides an effective solution for integrating relative camera pose information into pre-trained video diffusion models, addressing the shift-invariance problem while maintaining strong generative priors and enabling precise camera viewpoint control for video generation applications.

Abstract: Video generation with controllable camera viewpoints is essential for applications such as interactive content creation, gaming, and simulation. Existing methods typically adapt pre-trained video models using camera poses relative to a fixed reference, e.g., the first frame. However, these encodings lack shift-invariance, often leading to poor generalization and accumulated drift. While relative camera pose embeddings defined between arbitrary view pairs offer a more robust alternative, integrating them into pre-trained video diffusion models without prohibitive training costs or architectural changes remains challenging. We introduce ReRoPE, a plug-and-play framework that incorporates relative camera information into pre-trained video diffusion models without compromising their generation capability. Our approach is based on the insight that Rotary Positional Embeddings (RoPE) in existing models underutilize their full spectral bandwidth, particularly in the low-frequency components. By seamlessly injecting relative camera pose information into these underutilized bands, ReRoPE achieves precise control while preserving strong pre-trained generative priors. We evaluate our method on both image-to-video (I2V) and video-to-video (V2V) tasks in terms of camera control accuracy and visual fidelity. Our results demonstrate that ReRoPE offers a training-efficient path toward controllable, high-fidelity video generation. See project page for more results: https://sisyphe-lee.github.io/ReRoPE/

Method: Introduces AVIC, an adaptive test-time framework with world models that explicitly reasons about sufficiency of current visual evidence before selectively invoking and scaling visual imagination. Uses a decision mechanism to determine when to imagine and how much imagination to use.

Result: Across spatial reasoning benchmarks (SAT, MMSI) and embodied navigation (R2R), results show clear scenarios where imagination is critical, marginal, or detrimental. Selective control matches or outperforms fixed imagination strategies with substantially fewer world-model calls and language tokens.

Conclusion: Test-time visual imagination should be treated as a controllable resource for spatial reasoning. Selective imagination control enables efficient and reliable spatial reasoning by avoiding unnecessary computation and misleading evidence.

Abstract: Despite rapid progress in Multimodal Large Language Models (MLLMs), visual spatial reasoning remains unreliable when correct answers depend on how a scene would appear under unseen or alternative viewpoints. Recent work addresses this by augmenting reasoning with world models for visual imagination, but questions such as when imagination is actually necessary, how much of it is beneficial, and when it becomes harmful, remain poorly understood. In practice, indiscriminate imagination can increase computation and even degrade performance by introducing misleading evidence. In this work, we present an in-depth analysis of test-time visual imagination as a controllable resource for spatial reasoning. We study when static visual evidence is sufficient, when imagination improves reasoning, and how excessive or unnecessary imagination affects accuracy and efficiency. To support this analysis, we introduce AVIC, an adaptive test-time framework with world models that explicitly reasons about the sufficiency of current visual evidence before selectively invoking and scaling visual imagination. Across spatial reasoning benchmarks (SAT, MMSI) and an embodied navigation benchmark (R2R), our results reveal clear scenarios where imagination is critical, marginal, or detrimental, and show that selective control can match or outperform fixed imagination strategies with substantially fewer world-model calls and language tokens. Overall, our findings highlight the importance of analyzing and controlling test-time imagination for efficient and reliable spatial reasoning.

Method: Conducted component-wise refinement of Vision Transformers including normalization, activation functions, positional encoding, gating mechanisms, and learnable tokens while preserving the canonical Attention-FFN structure. These updates form the ViT-5 architecture.

Result: ViT-5 consistently outperforms state-of-the-art plain Vision Transformers across understanding and generation benchmarks. Achieves 84.2% top-1 accuracy on ImageNet-1k classification (vs 83.8% for DeiT-III-Base) and 1.84 FID in diffusion modeling (vs 2.06 with vanilla ViT).

Conclusion: ViT-5 offers a simple drop-in upgrade over vanilla ViT with improved representation learning, favorable spatial reasoning, and reliable task transfer, making it suitable for mid-2020s vision backbones aligned with contemporary foundation-model practices.

Abstract: This work presents a systematic investigation into modernizing Vision Transformer backbones by leveraging architectural advancements from the past five years. While preserving the canonical Attention-FFN structure, we conduct a component-wise refinement involving normalization, activation functions, positional encoding, gating mechanisms, and learnable tokens. These updates form a new generation of Vision Transformers, which we call ViT-5. Extensive experiments demonstrate that ViT-5 consistently outperforms state-of-the-art plain Vision Transformers across both understanding and generation benchmarks. On ImageNet-1k classification, ViT-5-Base reaches 84.2% top-1 accuracy under comparable compute, exceeding DeiT-III-Base at 83.8%. ViT-5 also serves as a stronger backbone for generative modeling: when plugged into an SiT diffusion framework, it achieves 1.84 FID versus 2.06 with a vanilla ViT backbone. Beyond headline metrics, ViT-5 exhibits improved representation learning and favorable spatial reasoning behavior, and transfers reliably across tasks. With a design aligned with contemporary foundation-model practices, ViT-5 offers a simple drop-in upgrade over vanilla ViT for mid-2020s vision backbones.

Method: 1) Systematic layer-wise analysis of MLLMs showing intermediate layers encode task-relevant information; 2) Combine intermediate-layer embeddings with calibrated MLLM head for zero-shot retrieval; 3) Lightweight text-based alignment strategy mapping dense video captions to short summaries for video-text embedding learning without visual supervision

Result: Achieves state-of-the-art results across common video retrieval benchmarks, outperforming current methods often by substantial margins, without any fine-tuning beyond text

Conclusion: MLLMs can be effectively leveraged for video-text retrieval through careful analysis of layer representations and text-based alignment, achieving strong performance without visual fine-tuning

Abstract: Recent studies have adapted generative Multimodal Large Language Models (MLLMs) into embedding extractors for vision tasks, typically through fine-tuning to produce universal representations. However, their performance on video remains inferior to Video Foundation Models (VFMs). In this paper, we focus on leveraging MLLMs for video-text embedding and retrieval. We first conduct a systematic layer-wise analysis, showing that intermediate (pre-trained) MLLM layers already encode substantial task-relevant information. Leveraging this insight, we demonstrate that combining intermediate-layer embeddings with a calibrated MLLM head yields strong zero-shot retrieval performance without any training. Building on these findings, we introduce a lightweight text-based alignment strategy which maps dense video captions to short summaries and enables task-related video-text embedding learning without visual supervision. Remarkably, without any fine-tuning beyond text, our method outperforms current methods, often by a substantial margin, achieving state-of-the-art results across common video retrieval benchmarks.

Method: Created a multimodal dataset with seven bands including RGB, digital elevation model, slope, thermal inertia, and grayscale channels. Dataset includes 664 images with standard splits plus 276 isolated test images from geographically disjoint regions to evaluate spatial generalization.

Result: The dataset supports stable training and achieves competitive performance with multiple segmentation models, though challenges remain in fragmented, elongated, and small-scale landslide regions. Evaluation on isolated test set shows noticeable performance drop, indicating increased difficulty and highlighting value for assessing model robustness.

Conclusion: MMLSv2 provides a valuable multimodal resource for Martian landslide segmentation research, particularly for testing model generalization across different geographical regions, with the isolated test set offering important insights into spatial generalization capabilities.

Abstract: We present MMLSv2, a dataset for landslide segmentation on Martian surfaces. MMLSv2 consists of multimodal imagery with seven bands: RGB, digital elevation model, slope, thermal inertia, and grayscale channels. MMLSv2 comprises 664 images distributed across training, validation, and test splits. In addition, an isolated test set of 276 images from a geographically disjoint region from the base dataset is released to evaluate spatial generalization. Experiments conducted with multiple segmentation models show that the dataset supports stable training and achieves competitive performance, while still posing challenges in fragmented, elongated, and small-scale landslide regions. Evaluation on the isolated test set leads to a noticeable performance drop, indicating increased difficulty and highlighting its value for assessing model robustness and generalization beyond standard in-distribution settings. Dataset will be available at: https://github.com/MAIN-Lab/MMLS_v2

Method: Proposes a targeted patch-based preprocessing pipeline to isolate structural features and minimize background noise. Uses DINOv2-small and DeiT Vision Transformers with a frozen-head fine-tuning strategy to keep computational requirements manageable. Evaluates through accuracy, precision, recall, and macro-averaged F1 scores.

Result: Small ViT architectures with the novel training method achieve competitive macro-averaged F1 scores relative to prior CNN baselines for disaster classification on the xBD dataset.

Conclusion: Vision Transformers with targeted preprocessing and efficient fine-tuning can effectively handle noisy, imbalanced satellite imagery data for building damage assessment, providing a scalable solution for post-disaster response.

Abstract: Rapid building damage assessment is critical for post-disaster response. Damage classification models built on satellite imagery provide a scalable means of obtaining situational awareness. However, label noise and severe class imbalance in satellite data create major challenges. The xBD dataset offers a standardized benchmark for building-level damage across diverse geographic regions. In this study, we evaluate Vision Transformer (ViT) model performance on the xBD dataset, specifically investigating how these models distinguish between types of structural damage when training on noisy, imbalanced data. In this study, we specifically evaluate DINOv2-small and DeiT for multi-class damage classification. We propose a targeted patch-based pre-processing pipeline to isolate structural features and minimize background noise in training. We adopt a frozen-head fine-tuning strategy to keep computational requirements manageable. Model performance is evaluated through accuracy, precision, recall, and macro-averaged F1 scores. We show that small ViT architectures with our novel training method achieves competitive macro-averaged F1 relative to prior CNN baselines for disaster classification.

Method: MambaFusion interleaves selective state-space models (SSMs) with windowed transformers for linear-time global context propagation while preserving local geometry. It uses multi-modal token alignment and reliability-aware fusion gates to dynamically re-weight camera-LiDAR features based on spatial confidence and calibration consistency. A structure-conditioned diffusion head integrates graph-based reasoning with uncertainty-aware denoising for physically plausible detection.

Result: MambaFusion establishes new state-of-the-art performance on nuScenes benchmarks while operating with linear-time complexity. The framework demonstrates robust, temporally stable, and interpretable 3D perception suitable for real-world autonomous driving systems.

Conclusion: Coupling SSM-based efficiency with reliability-driven fusion yields superior 3D perception for autonomous driving. The framework addresses key challenges in multimodal fusion and provides a promising direction for efficient, adaptive, and physically grounded 3D object detection systems.

Abstract: Reliable 3D object detection is fundamental to autonomous driving, and multimodal fusion algorithms using cameras and LiDAR remain a persistent challenge. Cameras provide dense visual cues but ill posed depth; LiDAR provides a precise 3D structure but sparse coverage. Existing BEV-based fusion frameworks have made good progress, but they have difficulties including inefficient context modeling, spatially invariant fusion, and reasoning under uncertainty. We introduce MambaFusion, a unified multi-modal detection framework that achieves efficient, adaptive, and physically grounded 3D perception. MambaFusion interleaves selective state-space models (SSMs) with windowed transformers to propagate the global context in linear time while preserving local geometric fidelity. A multi-modal token alignment (MTA) module and reliability-aware fusion gates dynamically re-weight camera-LiDAR features based on spatial confidence and calibration consistency. Finally, a structure-conditioned diffusion head integrates graph-based reasoning with uncertainty-aware denoising, enforcing physical plausibility, and calibrated confidence. MambaFusion establishes new state-of-the-art performance on nuScenes benchmarks while operating with linear-time complexity. The framework demonstrates that coupling SSM-based efficiency with reliability-driven fusion yields robust, temporally stable, and interpretable 3D perception for real-world autonomous driving systems.

Method: Proposes correction-specific rollouts (Octopus) - an RL rollout augmentation framework that synthesizes dense self-correction examples by recombining existing rollouts, plus a response-masking strategy to decouple self-correction from direct reasoning.

Result: Octopus-8B achieves state-of-the-art performance across 7 benchmarks among open-source VLMs, outperforming the best RLVR baseline by 1.0 score while requiring only 0.72× training time per step.

Conclusion: The Octopus framework effectively addresses sparse learning signals in self-correction training for VLMs through rollout recombination and response-masking, enabling efficient learning of controllable self-correction capabilities.

Abstract: Self-correction is essential for solving complex reasoning problems in vision-language models (VLMs). However, existing reinforcement learning (RL) methods struggle to learn it, as effective self-correction behaviors emerge only rarely, making learning signals extremely sparse. To address this challenge, we propose correction-specific rollouts (Octopus), an RL rollout augmentation framework that synthesizes dense self-correction examples by recombining existing rollouts. This augmentation simultaneously improves sample efficiency due to rollout reuse and stabilizes RL optimization through balanced supervision. Furthermore, we introduce a response-masking strategy that decouples self-correction from direct reasoning, avoiding signal conflicts and enabling both behaviors to be learned effectively. Building on this, we introduce Octopus-8B, a reasoning VLM with controllable self-correction capability. Across 7 benchmarks, it achieves SoTA performance among open-source VLMs, outperforming the best RLVR baseline by 1.0 score while requiring only $0.72\times$ training time per step.

Method: Developed a benchmark of 1.6M field polygons across 24 countries, pre-trained segmentation models, and command-line inference tools. Uses MOSAIKS random convolutional features with FTW-derived field boundaries for crop type classification with limited labels.

Result: Achieved macro F1 scores of 0.65-0.75 for crop type classification. Pre-computed predictions cover five countries (4.76M km²) with median predicted field areas ranging from 0.06 ha (Rwanda) to 0.28 ha (Switzerland).

Conclusion: FTW ecosystem provides a scalable solution for agricultural field analysis, enabling applications from local-scale extraction to country-scale inference using cloud-optimized data.

Abstract: Field boundary maps are a building block for agricultural data products and support crop monitoring, yield estimation, and disease estimation. This tutorial presents the Fields of The World (FTW) ecosystem: a benchmark of 1.6M field polygons across 24 countries, pre-trained segmentation models, and command-line inference tools. We provide two notebooks that cover (1) local-scale field boundary extraction with crop classification and forest loss attribution, and (2) country-scale inference using cloud-optimized data. We use MOSAIKS random convolutional features and FTW derived field boundaries to map crop type at the field level and report macro F1 scores of 0.65–0.75 for crop type classification with limited labels. Finally, we show how to explore pre-computed predictions over five countries (4.76M km\textsuperscript{2}), with median predicted field areas from 0.06 ha (Rwanda) to 0.28 ha (Switzerland).

Method: Progressive split-image attack strategies: 1) naive splitting, 2) adaptive white-box attack, 3) black-box transfer attack using novel adversarial knowledge distillation (Adv-KD) to improve cross-model transferability by distilling adversarial knowledge from a white-box surrogate model.

Result: Evaluations on three state-of-the-art VLMs and three jailbreak datasets show the strongest attack achieves up to 60% higher transfer success than existing baselines, demonstrating significant vulnerability in current VLM safety alignment.

Conclusion: Current VLM safety alignment is vulnerable to split-image attacks, and the paper proposes efficient mitigation strategies while highlighting the need for safety training that accounts for distributed harmful semantics across multiple images.

Abstract: Vision-Language Models (VLMs) are now a core part of modern AI. Recent work proposed several visual jailbreak attacks using single/ holistic images. However, contemporary VLMs demonstrate strong robustness against such attacks due to extensive safety alignment through preference optimization (e.g., RLHF). In this work, we identify a new vulnerability: while VLM pretraining and instruction tuning generalize well to split-image inputs, safety alignment is typically performed only on holistic images and does not account for harmful semantics distributed across multiple image fragments. Consequently, VLMs often fail to detect and refuse harmful split-image inputs, where unsafe cues emerge only after combining images. We introduce novel split-image visual jailbreak attacks (SIVA) that exploit this misalignment. Unlike prior optimization-based attacks, which exhibit poor black-box transferability due to architectural and prior mismatches across models, our attacks evolve in progressive phases from naive splitting to an adaptive white-box attack, culminating in a black-box transfer attack. Our strongest strategy leverages a novel adversarial knowledge distillation (Adv-KD) algorithm to substantially improve cross-model transferability. Evaluations on three state-of-the-art modern VLMs and three jailbreak datasets demonstrate that our strongest attack achieves up to 60% higher transfer success than existing baselines. Lastly, we propose efficient ways to address this critical vulnerability in the current VLM safety alignment.

Method: Proposes DAS-SK architecture that retrofits selective kernel convolution (SK-Conv) into dual atrous separable convolution (DAS-Conv) modules to enhance multi-scale feature learning. Enhances atrous spatial pyramid pooling (ASPP) to capture both fine-grained local structures and global context. Built on modified DeepLabV3 framework with MobileNetV3-Large and EfficientNet-B3 backbones.

Result: Achieves state-of-the-art performance across three benchmarks (LandCover.ai, VDD, PhenoBench) while being more efficient than CNN-, transformer-, and hybrid-based competitors. Requires up to 21x fewer parameters and 19x fewer GFLOPs than top-performing transformer models.

Conclusion: DAS-SK establishes a robust, efficient, and scalable solution for real-time agricultural robotics and high-resolution remote sensing, with potential for broader deployment in other vision domains.

Abstract: Semantic segmentation in high-resolution agricultural imagery demands models that strike a careful balance between accuracy and computational efficiency to enable deployment in practical systems. In this work, we propose DAS-SK, a novel lightweight architecture that retrofits selective kernel convolution (SK-Conv) into the dual atrous separable convolution (DAS-Conv) module to strengthen multi-scale feature learning. The model further enhances the atrous spatial pyramid pooling (ASPP) module, enabling the capture of fine-grained local structures alongside global contextual information. Built upon a modified DeepLabV3 framework with two complementary backbones - MobileNetV3-Large and EfficientNet-B3, the DAS-SK model mitigates limitations associated with large dataset requirements, limited spectral generalization, and the high computational cost that typically restricts deployment on UAVs and other edge devices. Comprehensive experiments across three benchmarks: LandCover.ai, VDD, and PhenoBench, demonstrate that DAS-SK consistently achieves state-of-the-art performance, while being more efficient than CNN-, transformer-, and hybrid-based competitors. Notably, DAS-SK requires up to 21x fewer parameters and 19x fewer GFLOPs than top-performing transformer models. These findings establish DAS-SK as a robust, efficient, and scalable solution for real-time agricultural robotics and high-resolution remote sensing, with strong potential for broader deployment in other vision domains.

Method: 1) Formulates 3D shape personalization as extracting category-agnostic geometric and appearance attributes; 2) Uses progressive optimization to learn shape concepts at geometry and appearance levels separately; 3) Extends to region-wise concept learning with context-aware and context-free losses.

Result: PEGAsus effectively extracts attributes from diverse reference shapes and composes them with text to synthesize new shapes, outperforming state-of-the-art methods in both quantitative and qualitative evaluations.

Conclusion: The framework enables fine-grained control over shape generation and supports diverse, personalized results, demonstrating superior performance in challenging cross-category scenarios.

Abstract: We present PEGAsus, a new framework capable of generating Personalized 3D shapes by learning shape concepts at both Geometry and Appearance levels. First, we formulate 3D shape personalization as extracting reusable, category-agnostic geometric and appearance attributes from reference shapes, and composing these attributes with text to generate novel shapes. Second, we design a progressive optimization strategy to learn shape concepts at both the geometry and appearance levels, decoupling the shape concept learning process. Third, we extend our approach to region-wise concept learning, enabling flexible concept extraction, with context-aware and context-free losses. Extensive experimental results show that PEGAsus is able to effectively extract attributes from a wide range of reference shapes and then flexibly compose these concepts with text to synthesize new shapes. This enables fine-grained control over shape generation and supports the creation of diverse, personalized results, even in challenging cross-category scenarios. Both quantitative and qualitative experiments demonstrate that our approach outperforms existing state-of-the-art solutions.

Method: Proposes MCSDR (Multimodal Conditional Score-based Diffusion model for Regression), a probabilistic framework that models the continuous posterior distribution of LVEF conditioned on echocardiogram videos and patient demographic attribute priors. Uses a generative regression approach with score-based diffusion models to capture distributional uncertainty rather than deterministic point estimates.

Result: Extensive experiments on EchoNet-Dynamic, EchoNet-Pediatric, and CAMUS datasets demonstrate state-of-the-art performance. Qualitative analysis shows that generation trajectories exhibit distinct behaviors in cases with high noise or significant physiological variability, offering novel interpretability for AI-aided diagnosis.

Conclusion: The paper successfully demonstrates a paradigm shift from deterministic regression to generative regression for medical imaging tasks, showing that modeling continuous posterior distributions can improve performance and provide interpretability through generation trajectories, particularly valuable in ambiguous pathological scenarios.

Abstract: Estimating Left Ventricular Ejection Fraction (LVEF) from echocardiograms constitutes an ill-posed inverse problem. Inherent noise, artifacts, and limited viewing angles introduce ambiguity, where a single video sequence may map not to a unique ground truth, but rather to a distribution of plausible physiological values. Prevailing deep learning approaches typically formulate this task as a standard regression problem that minimizes the Mean Squared Error (MSE). However, this paradigm compels the model to learn the conditional expectation, which may yield misleading predictions when the underlying posterior distribution is multimodal or heavy-tailed – a common phenomenon in pathological scenarios. In this paper, we investigate the paradigm shift from deterministic regression toward generative regression. We propose the Multimodal Conditional Score-based Diffusion model for Regression (MCSDR), a probabilistic framework designed to model the continuous posterior distribution of LVEF conditioned on echocardiogram videos and patient demographic attribute priors. Extensive experiments conducted on the EchoNet-Dynamic, EchoNet-Pediatric, and CAMUS datasets demonstrate that MCSDR achieves state-of-the-art performance. Notably, qualitative analysis reveals that the generation trajectories of our model exhibit distinct behaviors in cases characterized by high noise or significant physiological variability, thereby offering a novel layer of interpretability for AI-aided diagnosis.

Method: Proposes Geospatial Reasoning Chain-of-Thought (GR-CoT) framework with two components: offline knowledge distillation stream to establish fine-grained category interpretation standards, and online instance reasoning stream that performs sequential reasoning (macro-scenario anchoring, visual feature decoupling, knowledge-driven decision synthesis) to generate image-adaptive vocabulary for pixel-level alignment.

Result: Extensive experiments on LoveDA and GID5 benchmarks demonstrate superiority of the approach over existing methods.

Conclusion: GR-CoT framework effectively enhances scene understanding capabilities of MLLMs for open-vocabulary semantic segmentation in remote sensing by integrating geospatial reasoning to resolve semantic ambiguity.

Abstract: Open-vocabulary semantic segmentation has emerged as a promising research direction in remote sensing, enabling the recognition of diverse land-cover types beyond pre-defined category sets. However, existing methods predominantly rely on the passive mapping of visual features and textual embeddings. This ``appearance-based" paradigm lacks geospatial contextual awareness, leading to severe semantic ambiguity and misclassification when encountering land-cover classes with similar spectral features but distinct semantic attributes. To address this, we propose a Geospatial Reasoning Chain-of-Thought (GR-CoT) framework designed to enhance the scene understanding capabilities of Multimodal Large Language Models (MLLMs), thereby guiding open-vocabulary segmentation models toward precise mapping. The framework comprises two collaborative components: an offline knowledge distillation stream and an online instance reasoning stream. The offline stream establishes fine-grained category interpretation standards to resolve semantic conflicts between similar land-cover types. During online inference, the framework executes a sequential reasoning process involving macro-scenario anchoring, visual feature decoupling, and knowledge-driven decision synthesis. This process generates an image-adaptive vocabulary that guides downstream models to achieve pixel-level alignment with correct geographical semantics. Extensive experiments on the LoveDA and GID5 benchmarks demonstrate the superiority of our approach.

Method: Proposes Chain-of-Caption, a training-free framework that combines multiple textual and visual contexts to improve REC performance in MLLMs. Analyzes effects of individual context types (visual/textual) and their combinations via tool use without requiring fine-tuning.

Result: Individual textual or visual context improves REC performance without fine-tuning. Combined multiple contexts in Chain-of-Caption framework shows 5% to 30% performance gain over baseline models on accuracy at various IoU thresholds across RefCOCO/RefCOCOg/RefCOCO+ and Ref-L4 datasets.

Conclusion: Training-free context combination via Chain-of-Caption significantly enhances MLLM performance on referring expression comprehension tasks, demonstrating the value of multi-context integration without requiring model retraining.

Abstract: Given a textual description, the task of referring expression comprehension (REC) involves the localisation of the referred object in an image. Multimodal large language models (MLLMs) have achieved high accuracy on REC benchmarks through scaling up the model size and training data. Moreover, the performance of MLLMs can be further improved using techniques such as Chain-of-Thought and tool use, which provides additional visual or textual context to the model. In this paper, we analyse the effect of various techniques for providing additional visual and textual context via tool use to the MLLM and its effect on the REC task. Furthermore, we propose a training-free framework named Chain-of-Caption to improve the REC performance of MLLMs. We perform experiments on RefCOCO/RefCOCOg/RefCOCO+ and Ref-L4 datasets and show that individual textual or visual context can improve the REC performance without any fine-tuning. By combining multiple contexts, our training-free framework shows between 5% to 30% performance gain over the baseline model on accuracy at various Intersection over Union (IoU) thresholds.

Method: Proposes Efficient-SAM2 with two key techniques: 1) Object-aware Sparse Window Routing (SWR) for image encoder - uses previous-frame decoder cues to route background regions to lightweight shortcut branch; 2) Object-aware Sparse Memory Retrieval (SMR) for memory attention - allows only salient memory tokens to participate in computation, reusing saliency patterns from first recollection.

Result: Achieves 1.68x speedup on SAM2.1-L model with only 1.0% accuracy drop on SA-V test set, using negligible additional parameters and minimal training overhead.

Conclusion: Efficient-SAM2 successfully accelerates SAM2 by exploiting its sparse perception patterns, demonstrating that post-training optimization can significantly improve efficiency while maintaining accuracy for video object segmentation.

Abstract: Segment Anything Model 2 (SAM2) shows excellent performance in video object segmentation tasks; however, the heavy computational burden hinders its application in real-time video processing. Although there have been efforts to improve the efficiency of SAM2, most of them focus on retraining a lightweight backbone, with little exploration into post-training acceleration. In this paper, we observe that SAM2 exhibits sparse perception pattern as biological vision, which provides opportunities for eliminating redundant computation and acceleration: i) In mask decoder, the attention primarily focuses on the foreground objects, whereas the image encoder in the earlier stage exhibits a broad attention span, which results in unnecessary computation to background regions. ii) In memory bank, only a small subset of tokens in each frame contribute significantly to memory attention, and the salient regions exhibit temporal consistency, making full-token computation redundant. With these insights, we propose Efficient-SAM2, which promotes SAM2 to adaptively focus on object regions while eliminating task-irrelevant computations, thereby significantly improving inference efficiency. Specifically, for image encoder, we propose object-aware Sparse Window Routing (SWR), a window-level computation allocation mechanism that leverages the consistency and saliency cues from the previous-frame decoder to route background regions into a lightweight shortcut branch. Moreover, for memory attention, we propose object-aware Sparse Memory Retrieval (SMR), which allows only the salient memory tokens in each frame to participate in computation, with the saliency pattern reused from their first recollection. With negligible additional parameters and minimal training overhead, Efficient-SAM2 delivers 1.68x speedup on SAM2.1-L model with only 1.0% accuracy drop on SA-V test set.

Method: Proposes MA-ADV framework using point cloud representations for events. Uses diffusion-based approach to smooth perturbations while leveraging spatial-temporal relationships. Employs sample-wise Adam optimization, iterative refinement, and binary search for minimal-cost perturbations.

Result: Achieves 100% attack success rate with minimal perturbation cost. Demonstrates enhanced robustness against defenses, highlighting security vulnerabilities in event-based perception systems.

Conclusion: MA-ADV successfully addresses the challenge of adversarial attacks on event cameras, revealing critical security issues in event-based systems that need attention for safety-critical applications.

Abstract: Event cameras have been widely adopted in safety-critical domains such as autonomous driving, robotics, and human-computer interaction. A pressing challenge arises from the vulnerability of deep neural networks to adversarial examples, which poses a significant threat to the reliability of event-based systems. Nevertheless, research into adversarial attacks on events is scarce. This is primarily due to the non-differentiable nature of mainstream event representations, which hinders the extension of gradient-based attack methods. In this paper, we propose MA-ADV, a novel \textbf{M}otion-\textbf{A}ware \textbf{Adv}ersarial framework. To the best of our knowledge, this is the first work to generate adversarial events by leveraging point cloud representations. MA-ADV accounts for high-frequency noise in events and employs a diffusion-based approach to smooth perturbations, while fully leveraging the spatial and temporal relationships among events. Finally, MA-ADV identifies the minimal-cost perturbation through a combination of sample-wise Adam optimization, iterative refinement, and binary search. Extensive experimental results validate that MA-ADV ensures a 100% attack success rate with minimal perturbation cost, and also demonstrate enhanced robustness against defenses, underscoring the critical security challenges facing future event-based perception systems.

Method: Proposes DeCI framework with two key principles: (1) Cycle and Drift Decomposition to disentangle oscillatory fluctuations and slow baseline trends within each ROI, and (2) Channel-Independence to model each ROI separately for improved robustness and reduced overfitting.

Result: DeCI achieves superior classification accuracy and generalization compared to both traditional FC-based approaches and other temporal baselines across five public datasets.

Conclusion: Advocates for a shift toward end-to-end temporal modeling in fMRI analysis to better capture complex brain dynamics, with DeCI demonstrating the value of directly modeling temporal information.

Abstract: Functional magnetic resonance imaging (fMRI) enables non-invasive brain disorder classification by capturing blood-oxygen-level-dependent (BOLD) signals. However, most existing methods rely on functional connectivity (FC) via Pearson correlation, which reduces 4D BOLD signals to static 2D matrices, discarding temporal dynamics and capturing only linear inter-regional relationships. In this work, we benchmark state-of-the-art temporal models (e.g., time-series models such as PatchTST, TimesNet, and TimeMixer) on raw BOLD signals across five public datasets. Results show these models consistently outperform traditional FC-based approaches, highlighting the value of directly modeling temporal information such as cycle-like oscillatory fluctuations and drift-like slow baseline trends. Building on this insight, we propose DeCI, a simple yet effective framework that integrates two key principles: (i) Cycle and Drift Decomposition to disentangle cycle and drift within each ROI (Region of Interest); and (ii) Channel-Independence to model each ROI separately, improving robustness and reducing overfitting. Extensive experiments demonstrate that DeCI achieves superior classification accuracy and generalization compared to both FC-based and temporal baselines. Our findings advocate for a shift toward end-to-end temporal modeling in fMRI analysis to better capture complex brain dynamics. The code is available at https://github.com/Levi-Ackman/DeCI.

Method: PISCO is a video diffusion model that allows users to specify arbitrary sparse keyframe controls (single keyframe, start-end keyframes, or sparse keyframes at arbitrary timestamps) and automatically propagates object appearance, motion, and interaction. To address distribution shift from sparse conditioning in pretrained video diffusion models, the authors introduce Variable-Information Guidance for robust conditioning and Distribution-Preserving Temporal Masking to stabilize temporal generation, along with geometry-aware conditioning for realistic scene adaptation. They also construct PISCO-Bench, a benchmark with verified instance annotations and paired clean background videos.

Result: Experiments demonstrate that PISCO consistently outperforms strong inpainting and video editing baselines under sparse control conditions. The model exhibits clear, monotonic performance improvements as additional control signals are provided, evaluated using both reference-based and reference-free perceptual metrics on the PISCO-Bench benchmark.

Conclusion: PISCO represents an important advancement in controllable video generation for professional filmmaking applications, enabling precise video instance insertion with minimal user effort through sparse keyframe control while maintaining scene integrity and physical consistency.

Abstract: The landscape of AI video generation is undergoing a pivotal shift: moving beyond general generation - which relies on exhaustive prompt-engineering and “cherry-picking” - towards fine-grained, controllable generation and high-fidelity post-processing. In professional AI-assisted filmmaking, it is crucial to perform precise, targeted modifications. A cornerstone of this transition is video instance insertion, which requires inserting a specific instance into existing footage while maintaining scene integrity. Unlike traditional video editing, this task demands several requirements: precise spatial-temporal placement, physically consistent scene interaction, and the faithful preservation of original dynamics - all achieved under minimal user effort. In this paper, we propose PISCO, a video diffusion model for precise video instance insertion with arbitrary sparse keyframe control. PISCO allows users to specify a single keyframe, start-and-end keyframes, or sparse keyframes at arbitrary timestamps, and automatically propagates object appearance, motion, and interaction. To address the severe distribution shift induced by sparse conditioning in pretrained video diffusion models, we introduce Variable-Information Guidance for robust conditioning and Distribution-Preserving Temporal Masking to stabilize temporal generation, together with geometry-aware conditioning for realistic scene adaptation. We further construct PISCO-Bench, a benchmark with verified instance annotations and paired clean background videos, and evaluate performance using both reference-based and reference-free perceptual metrics. Experiments demonstrate that PISCO consistently outperforms strong inpainting and video editing baselines under sparse control, and exhibits clear, monotonic performance improvements as additional control signals are provided. Project page: xiangbogaobarry.github.io/PISCO.

Method: Proposes a multimodal fusion framework with: 1) Pseudo-label aggregation for PO data using satellite imagery geographic coverage; 2) Swin Transformer Base for satellite imagery, TabM network for tabular features, Temporal Swin Transformer for time-series; 3) Stackable serial tri-modal cross-attention for modality fusion; 4) Mixture-of-experts approach partitioning test samples by spatial proximity to PA samples, using different models trained on distinct datasets for each partition.

Result: Experiments on GeoLifeCLEF 2025 dataset show superior predictive performance in scenarios with limited PA coverage and pronounced distribution shifts compared to baseline approaches.

Conclusion: The proposed multimodal fusion framework effectively addresses challenges in plant distribution prediction by leveraging complementary strengths of PA and PO data while handling distribution shifts through a mixture-of-experts approach, demonstrating improved performance in real-world constrained scenarios.

Abstract: Large-scale, cross-species plant distribution prediction plays a crucial role in biodiversity conservation, yet modeling efforts in this area still face significant challenges due to the sparsity and bias of observational data. Presence-Absence (PA) data provide accurate and noise-free labels, but are costly to obtain and limited in quantity; Presence-Only (PO) data, by contrast, offer broad spatial coverage and rich spatiotemporal distribution, but suffer from severe label noise in negative samples. To address these real-world constraints, this paper proposes a multimodal fusion framework that fully leverages the strengths of both PA and PO data. We introduce an innovative pseudo-label aggregation strategy for PO data based on the geographic coverage of satellite imagery, enabling geographic alignment between the label space and remote sensing feature space. In terms of model architecture, we adopt Swin Transformer Base as the backbone for satellite imagery, utilize the TabM network for tabular feature extraction, retain the Temporal Swin Transformer for time-series modeling, and employ a stackable serial tri-modal cross-attention mechanism to optimize the fusion of heterogeneous modalities. Furthermore, empirical analysis reveals significant geographic distribution shifts between PA training and test samples, and models trained by directly mixing PO and PA data tend to experience performance degradation due to label noise in PO data. To address this, we draw on the mixture-of-experts paradigm: test samples are partitioned according to their spatial proximity to PA samples, and different models trained on distinct datasets are used for inference and post-processing within each partition. Experiments on the GeoLifeCLEF 2025 dataset demonstrate that our approach achieves superior predictive performance in scenarios with limited PA coverage and pronounced distribution shifts.

Method: Proposes CAE-AV framework with two modules: 1) CASTE dynamically balances spatial and temporal relations by evaluating frame-level audio-visual agreement, and 2) CASE injects cross-modal semantic guidance into selected spatio-temporal positions. Also includes lightweight objectives: caption-to-modality InfoNCE, visual-audio consistency, and entropy regularization.

Result: With frozen backbones, CAE-AV achieves state-of-the-art performance on AVE, AVVP, AVS, and AVQA benchmarks. Qualitative analyses validate robustness against audio-visual misalignment.

Conclusion: CAE-AV effectively addresses audio-visual misalignment through complementary spatio-temporal enrichment and semantic guidance modules, demonstrating strong performance across multiple audio-visual understanding tasks.

Abstract: Audio-visual learning suffers from modality misalignment caused by off-screen sources and background clutter, and current methods usually amplify irrelevant regions or moments, leading to unstable training and degraded representation quality. To address this challenge, we proposed a novel Caption-aligned and Agreement-guided Enhancement framework (CAE-AV) for audio-visual learning, which used two complementary modules: Cross-modal Agreement-guided Spatio-Temporal Enrichment (CASTE) and Caption-Aligned Saliency-guided Enrichment (CASE) to relieve audio-visual misalignment. CASTE dynamically balances spatial and temporal relations by evaluating frame-level audio-visual agreement, ensuring that key information is captured from both preceding and subsequent frames under misalignment. CASE injects cross-modal semantic guidance into selected spatio-temporal positions, leveraging high-level semantic cues to further alleviate misalignment. In addition, we design lightweight objectives, caption-to-modality InfoNCE, visual-audio consistency, and entropy regularization to guide token selection and strengthen cross-modal semantic alignment. With frozen backbones, CAE-AV achieves state-of-the-art performance on AVE, AVVP, AVS, and AVQA benchmarks, and qualitative analyses further validate its robustness against audio-visual misalignment.

Method: Proposes Language-Guided Tokenization (LG-Tok) that aligns natural language with motion at tokenization stage using Transformer-based tokenizer with attention mechanisms. Includes language-drop scheme for language-free guidance during generation.

Result: Achieves Top-1 scores of 0.542 (HumanML3D) and 0.582 (Motion-X), outperforming MARDM (0.500/0.528). FID scores of 0.057/0.088 vs 0.114/0.147. LG-Tok-mini uses half tokens while maintaining competitive performance.

Conclusion: Language guidance enables efficient motion tokenization with compact semantic representations, improving both reconstruction quality and generation simplicity while supporting both language-conditioned and language-free generation.

Abstract: In this paper, we focus on motion discrete tokenization, which converts raw motion into compact discrete tokens–a process proven crucial for efficient motion generation. In this paradigm, increasing the number of tokens is a common approach to improving motion reconstruction quality, but more tokens make it more difficult for generative models to learn. To maintain high reconstruction quality while reducing generation complexity, we propose leveraging language to achieve efficient motion tokenization, which we term Language-Guided Tokenization (LG-Tok). LG-Tok aligns natural language with motion at the tokenization stage, yielding compact, high-level semantic representations. This approach not only strengthens both tokenization and detokenization but also simplifies the learning of generative models. Furthermore, existing tokenizers predominantly adopt convolutional architectures, whose local receptive fields struggle to support global language guidance. To this end, we propose a Transformer-based Tokenizer that leverages attention mechanisms to enable effective alignment between language and motion. Additionally, we design a language-drop scheme, in which language conditions are randomly removed during training, enabling the detokenizer to support language-free guidance during generation. On the HumanML3D and Motion-X generation benchmarks, LG-Tok achieves Top-1 scores of 0.542 and 0.582, outperforming state-of-the-art methods (MARDM: 0.500 and 0.528), and with FID scores of 0.057 and 0.088, respectively, versus 0.114 and 0.147. LG-Tok-mini uses only half the tokens while maintaining competitive performance (Top-1: 0.521/0.588, FID: 0.085/0.071), validating the efficiency of our semantic representations.

Method: Introduces UGData (spatially grounded dataset with graph-aligned supervision), UGE (two-stage training combining instruction-guided contrastive learning with graph-based spatial encoding), and UGBench benchmark for evaluation.

Result: UGE achieves up to 44% improvement in image retrieval and 30% in geolocation ranking on training cities, with over 30% and 22% gains on held-out cities.

Conclusion: Explicit spatial grounding significantly improves multimodal embeddings for spatially intensive urban tasks, demonstrating the value of structured spatial alignment in vision-language models.

Abstract: Learning transferable multimodal embeddings for urban environments is challenging because urban understanding is inherently spatial, yet existing datasets and benchmarks lack explicit alignment between street-view images and urban structure. We introduce UGData, a spatially grounded dataset that anchors street-view images to structured spatial graphs and provides graph-aligned supervision via spatial reasoning paths and spatial context captions, exposing distance, directionality, connectivity, and neighborhood context beyond image content. Building on UGData, we propose UGE, a two-stage training strategy that progressively and stably aligns images, text, and spatial structures by combining instruction-guided contrastive learning with graph-based spatial encoding. We finally introduce UGBench, a comprehensive benchmark to evaluate how spatially grounded embeddings support diverse urban understanding tasks – including geolocation ranking, image retrieval, urban perception, and spatial grounding. We develop UGE on multiple state-of-the-art VLM backbones, including Qwen2-VL, Qwen2.5-VL, Phi-3-Vision, and LLaVA1.6-Mistral, and train fixed-dimensional spatial embeddings with LoRA tuning. UGE built upon Qwen2.5-VL-7B backbone achieves up to 44% improvement in image retrieval and 30% in geolocation ranking on training cities, and over 30% and 22% gains respectively on held-out cities, demonstrating the effectiveness of explicit spatial grounding for spatially intensive urban tasks.

Method: 1) Analyzed reasoning trajectories and conducted guided search experiments with PRMs to define 7 fine-grained error types. 2) Constructed benchmark with 1,206 manually annotated reasoning trajectories spanning 4 categories and 16 subcategories. 3) Experimental evaluation of current LVLMs as PRMs, assessing performance across error types, evaluation biases, and sensitivity to reasoning step positions.

Result: Current LVLMs perform poorly as PRMs for visual reasoning processes, showing limited capabilities, significant performance disparities across error types, positive evaluation bias, and sensitivity to reasoning step positions. The benchmark effectively reveals these limitations.

Conclusion: The benchmark establishes crucial foundations for advancing PRMs in LVLMs, demonstrating the necessity for specialized PRMs and providing a comprehensive evaluation framework for the “thinking with images” paradigm.

Abstract: The rapid advancement of Large Vision Language Models (LVLMs) has demonstrated excellent abilities in various visual tasks. Building upon these developments, the thinking with images paradigm has emerged, enabling models to dynamically edit and re-encode visual information at each reasoning step, mirroring human visual processing. However, this paradigm introduces significant challenges as diverse errors may occur during reasoning processes. This necessitates Process Reward Models (PRMs) for distinguishing positive and negative reasoning steps, yet existing benchmarks for PRMs are predominantly text-centric and lack comprehensive assessment under this paradigm. To address these gaps, this work introduces the first comprehensive benchmark specifically designed for evaluating PRMs under the thinking with images paradigm. Our main contributions are: (1) Through extensive analysis of reasoning trajectories and guided search experiments with PRMs, we define 7 fine-grained error types and demonstrate both the necessity for specialized PRMs and the potential for improvement. (2) We construct a comprehensive benchmark comprising 1,206 manually annotated thinking with images reasoning trajectories spanning 4 categories and 16 subcategories for fine-grained evaluation of PRMs. (3) Our experimental analysis reveals that current LVLMs fall short as effective PRMs, exhibiting limited capabilities in visual reasoning process evaluation with significant performance disparities across error types, positive evaluation bias, and sensitivity to reasoning step positions. These findings demonstrate the effectiveness of our benchmark and establish crucial foundations for advancing PRMs in LVLMs.

Method: 1) Multi-modal information density assessment framework to quantify complexity; 2) E-VAds benchmark with 3,961 videos and 19,785 Q&A pairs; 3) E-VAds-R1 RL-based reasoning model with multi-grained reward design (MG-GRPO)

Result: E-commerce content shows substantially higher multi-modal density than mainstream datasets; E-VAds-R1 achieves 109.2% performance gain in commercial intent reasoning with few hundred training samples

Conclusion: E-commerce videos present a challenging frontier for video understanding; the proposed benchmark and RL-based model effectively address commercial intent reasoning in dense multi-modal content

Abstract: E-commerce short videos represent a high-revenue segment of the online video industry characterized by a goal-driven format and dense multi-modal signals. Current models often struggle with these videos because existing benchmarks focus primarily on general-purpose tasks and neglect the reasoning of commercial intent. In this work, we first propose a \textbf{multi-modal information density assessment framework} to quantify the complexity of this domain. Our evaluation reveals that e-commerce content exhibits substantially higher density across visual, audio, and textual modalities compared to mainstream datasets, establishing a more challenging frontier for video understanding. To address this gap, we introduce \textbf{E-commerce Video Ads Benchmark (E-VAds)}, which is the first benchmark specifically designed for e-commerce short video understanding. We curated 3,961 high-quality videos from Taobao covering a wide range of product categories and used a multi-agent system to generate 19,785 open-ended Q&A pairs. These questions are organized into two primary dimensions, namely Perception and Cognition and Reasoning, which consist of five distinct tasks. Finally, we develop \textbf{E-VAds-R1}, an RL-based reasoning model featuring a multi-grained reward design called \textbf{MG-GRPO}. This strategy provides smooth guidance for early exploration while creating a non-linear incentive for expert-level precision. Experimental results demonstrate that E-VAds-R1 achieves a 109.2% performance gain in commercial intent reasoning with only a few hundred training samples.

Method: Proposes GeoEdit framework with: 1) Diffusion transformer module for in-context generation with integrated geometric transformations, 2) Effects-Sensitive Attention for enhanced lighting/shadow modeling, 3) RS-Objects dataset (120K+ image pairs) for training geometric editing.

Result: Extensive experiments show GeoEdit outperforms state-of-the-art methods in visual quality, geometric accuracy, and realism on public benchmarks.

Conclusion: GeoEdit effectively addresses geometric editing challenges in diffusion models while improving realism through better lighting/shadow modeling.

Abstract: Recent advances in diffusion models have significantly improved image editing. However, challenges persist in handling geometric transformations, such as translation, rotation, and scaling, particularly in complex scenes. Existing approaches suffer from two main limitations: (1) difficulty in achieving accurate geometric editing of object translation, rotation, and scaling; (2) inadequate modeling of intricate lighting and shadow effects, leading to unrealistic results. To address these issues, we propose GeoEdit, a framework that leverages in-context generation through a diffusion transformer module, which integrates geometric transformations for precise object edits. Moreover, we introduce Effects-Sensitive Attention, which enhances the modeling of intricate lighting and shadow effects for improved realism. To further support training, we construct RS-Objects, a large-scale geometric editing dataset containing over 120,000 high-quality image pairs, enabling the model to learn precise geometric editing while generating realistic lighting and shadows. Extensive experiments on public benchmarks demonstrate that GeoEdit consistently outperforms state-of-the-art methods in terms of visual quality, geometric accuracy, and realism.

Method: 1) Degradation-Robust Flow Alignment (DRFA) module using confidence-aware attention to filter unreliable motion cues; 2) Adversarial distillation to compress diffusion sampling into few-step regime; 3) Synergistic optimization strategy to balance perceptual quality with temporal consistency.

Result: Achieves state-of-the-art performance while accelerating sampling process by 12× compared to previous methods.

Conclusion: D²-VR successfully addresses the latency and stability limitations of diffusion-based video restoration, making it more practical for real-world applications while maintaining high quality.

Abstract: The integration of diffusion priors with temporal alignment has emerged as a transformative paradigm for video restoration, delivering fantastic perceptual quality, yet the practical deployment of such frameworks is severely constrained by prohibitive inference latency and temporal instability when confronted with complex real-world degradations. To address these limitations, we propose \textbf{D$^2$-VR}, a single-image diffusion-based video-restoration framework with low-step inference. To obtain precise temporal guidance under severe degradation, we first design a Degradation-Robust Flow Alignment (DRFA) module that leverages confidence-aware attention to filter unreliable motion cues. We then incorporate an adversarial distillation paradigm to compress the diffusion sampling trajectory into a rapid few-step regime. Finally, a synergistic optimization strategy is devised to harmonize perceptual quality with rigorous temporal consistency. Extensive experiments demonstrate that D$^2$-VR achieves state-of-the-art performance while accelerating the sampling process by \textbf{12$\times$}

Method: Created synthetic dataset using colon geometries from 10 CT scans imported into virtual environment mimicking intraoperative conditions, rendered with realistic vascular textures, and paired with comprehensive ground truth annotations.

Result: Benchmark study shows RealSynCol’s high realism and variability significantly enhance generalization performance on clinical images for depth and pose estimation tasks.

Conclusion: RealSynCol is a powerful synthetic dataset that enables development of robust deep learning algorithms for endoscopic diagnosis by providing realistic training data with comprehensive ground truth.

Abstract: Deep learning has the potential to improve colonoscopy by enabling 3D reconstruction of the colon, providing a comprehensive view of mucosal surfaces and lesions, and facilitating the identification of unexplored areas. However, the development of robust methods is limited by the scarcity of large-scale ground truth data. We propose RealSynCol, a highly realistic synthetic dataset designed to replicate the endoscopic environment. Colon geometries extracted from 10 CT scans were imported into a virtual environment that closely mimics intraoperative conditions and rendered with realistic vascular textures. The resulting dataset comprises 28,130 frames, paired with ground truth depth maps, optical flow, 3D meshes, and camera trajectories. A benchmark study was conducted to evaluate the available synthetic colon datasets for the tasks of depth and pose estimation. Results demonstrate that the high realism and variability of RealSynCol significantly enhance generalization performance on clinical images, proving it to be a powerful tool for developing deep learning algorithms to support endoscopic diagnosis.

Method: 1) Identify critical design choice in LightGlue model; 2) Analyze role of detectors vs descriptors in transformer matching; 3) Propose fine-tuning approach using keypoints from diverse detectors to create universal detector-agnostic model

Result: Resulting universal model achieves or exceeds accuracy of models specifically trained for novel detectors when used as zero-shot matcher

Conclusion: Detectors are primary cause of performance differences in transformer-based matching; universal detector-agnostic models are feasible and offer deployment advantages

Abstract: We revisit the problem of training attention-based sparse image matching models for various local features. We first identify one critical design choice that has been previously overlooked, which significantly impacts the performance of the LightGlue model. We then investigate the role of detectors and descriptors within the transformer-based matching framework, finding that detectors, rather than descriptors, are often the primary cause for performance difference. Finally, we propose a novel approach to fine-tune existing image matching models using keypoints from a diverse set of detectors, resulting in a universal, detector-agnostic model. When deployed as a zero-shot matcher for novel detectors, the resulting model achieves or exceeds the accuracy of models specifically trained for those features. Our findings offer valuable insights for the deployment of transformer-based matching models and the future design of local features.

Method: Proposes Demo-ICL-Bench with 1200 instructional YouTube videos and two demonstration types (text summaries and video demonstrations). Develops Demo-ICL model with two-stage training: video-supervised fine-tuning and information-assisted direct preference optimization.

Result: Benchmark proves challenging for state-of-the-art MLLMs. Demo-ICL demonstrates effectiveness in learning from in-context video examples, revealing future research directions.

Conclusion: The work introduces a novel video in-context learning task and benchmark, showing current MLLMs’ limitations and proposing an effective training approach for video understanding.

Abstract: Despite the growing video understanding capabilities of recent Multimodal Large Language Models (MLLMs), existing video benchmarks primarily assess understanding based on models’ static, internal knowledge, rather than their ability to learn and adapt from dynamic, novel contexts from few examples. To bridge this gap, we present Demo-driven Video In-Context Learning, a novel task focused on learning from in-context demonstrations to answer questions about the target videos. Alongside this, we propose Demo-ICL-Bench, a challenging benchmark designed to evaluate demo-driven video in-context learning capabilities. Demo-ICL-Bench is constructed from 1200 instructional YouTube videos with associated questions, from which two types of demonstrations are derived: (i) summarizing video subtitles for text demonstration; and (ii) corresponding instructional videos as video demonstrations. To effectively tackle this new challenge, we develop Demo-ICL, an MLLM with a two-stage training strategy: video-supervised fine-tuning and information-assisted direct preference optimization, jointly enhancing the model’s ability to learn from in-context examples. Extensive experiments with state-of-the-art MLLMs confirm the difficulty of Demo-ICL-Bench, demonstrate the effectiveness of Demo-ICL, and thereby unveil future research directions.

Method: Three key innovations: (1) scene-aware segmentation - dynamic clustering of frames into coherent scenes; (2) scene-aware compression - compact token representation stored in GPU while full frames offloaded to CPU; (3) scene-aware recall - selective retrieval and reintegration of relevant scenes for query answering.

Result: Extensive experiments on StreamingBench show Vista achieves state-of-the-art performance, establishing a strong baseline for real-world streaming video understanding.

Conclusion: Vista provides an efficient, scalable framework for streaming video QA that enables long-context reasoning without compromising latency or memory efficiency, working seamlessly with various vision-language backbones.

Abstract: Streaming video question answering (Streaming Video QA) poses distinct challenges for multimodal large language models (MLLMs), as video frames arrive sequentially and user queries can be issued at arbitrary time points. Existing solutions relying on fixed-size memory or naive compression often suffer from context loss or memory overflow, limiting their effectiveness in long-form, real-time scenarios. We present Vista, a novel framework for scene-aware streaming video QA that enables efficient and scalable reasoning over continuous video streams. The innovation of Vista can be summarized in three aspects: (1) scene-aware segmentation, where Vista dynamically clusters incoming frames into temporally and visually coherent scene units; (2) scene-aware compression, where each scene is compressed into a compact token representation and stored in GPU memory for efficient index-based retrieval, while full-resolution frames are offloaded to CPU memory; and (3) scene-aware recall, where relevant scenes are selectively recalled and reintegrated into the model input upon receiving a query, enabling both efficiency and completeness. Vista is model-agnostic and integrates seamlessly with a variety of vision-language backbones, enabling long-context reasoning without compromising latency or memory efficiency. Extensive experiments on StreamingBench demonstrate that Vista achieves state-of-the-art performance, establishing a strong baseline for real-world streaming video understanding.

Method: Proposes Tri-Domain Causal Text-to-Motion Generation (TriC-Motion) with three core modules: Temporal Motion Encoding, Spatial Topology Modeling, and Hybrid Frequency Analysis. Includes Score-guided Tri-domain Fusion for integration and Causality-based Counterfactual Motion Disentangler to eliminate noise.

Result: Achieves state-of-the-art performance with R@1 of 0.612 on HumanML3D dataset, demonstrating superior generation of high-fidelity, coherent, diverse, and text-aligned motion sequences.

Conclusion: TriC-Motion successfully addresses limitations of existing methods by providing unified tri-domain modeling with causal intervention, resulting in significantly improved text-to-motion generation quality.

Abstract: Text-to-motion generation, a rapidly evolving field in computer vision, aims to produce realistic and text-aligned motion sequences. Current methods primarily focus on spatial-temporal modeling or independent frequency domain analysis, lacking a unified framework for joint optimization across spatial, temporal, and frequency domains. This limitation hinders the model’s ability to leverage information from all domains simultaneously, leading to suboptimal generation quality. Additionally, in motion generation frameworks, motion-irrelevant cues caused by noise are often entangled with features that contribute positively to generation, thereby leading to motion distortion. To address these issues, we propose Tri-Domain Causal Text-to-Motion Generation (TriC-Motion), a novel diffusion-based framework integrating spatial-temporal-frequency-domain modeling with causal intervention. TriC-Motion includes three core modeling modules for domain-specific modeling, namely Temporal Motion Encoding, Spatial Topology Modeling, and Hybrid Frequency Analysis. After comprehensive modeling, a Score-guided Tri-domain Fusion module integrates valuable information from the triple domains, simultaneously ensuring temporal consistency, spatial topology, motion trends, and dynamics. Moreover, the Causality-based Counterfactual Motion Disentangler is meticulously designed to expose motion-irrelevant cues to eliminate noise, disentangling the real modeling contributions of each domain for superior generation. Extensive experimental results validate that TriC-Motion achieves superior performance compared to state-of-the-art methods, attaining an outstanding R@1 of 0.612 on the HumanML3D dataset. These results demonstrate its capability to generate high-fidelity, coherent, diverse, and text-aligned motion sequences. Code is available at: https://caoyiyang1105.github.io/TriC-Motion/.

Method: Uses 2D pose estimation on real-world video sequences from WIVW dataset. Categorizes gestures into four classes (Stop, Go, Thank & Greet, No Gesture) and extracts 76 static and dynamic features from normalized keypoints, focusing on hand position and movement velocity.

Result: Achieved 87% classification accuracy, with hand position and movement velocity identified as particularly discriminative features for distinguishing between gesture classes.

Conclusion: The framework improves autonomous vehicle perceptual capabilities and contributes to understanding pedestrian behavior in traffic contexts through effective gesture interpretation.

Abstract: Gestures are a key component of non-verbal communication in traffic, often helping pedestrian-to-driver interactions when formal traffic rules may be insufficient. This problem becomes more apparent when autonomous vehicles (AVs) struggle to interpret such gestures. In this study, we present a gesture classification framework using 2D pose estimation applied to real-world video sequences from the WIVW dataset. We categorise gestures into four primary classes (Stop, Go, Thank & Greet, and No Gesture) and extract 76 static and dynamic features from normalised keypoints. Our analysis demonstrates that hand position and movement velocity are especially discriminative in distinguishing between gesture classes, achieving a classification accuracy score of 87%. These findings not only improve the perceptual capabilities of AV systems but also contribute to the broader understanding of pedestrian behaviour in traffic contexts.

Method: Uses Food-101 and Fruits-360 datasets to demonstrate cross-dataset accuracy degradation. Constructs synthetic illumination-augmented datasets by varying light temperature and intensity for controlled robustness analysis. Evaluates cross-dataset transfer learning and domain generalization, focusing on illumination-sensitive categories like apple-based classes.

Result: Shows substantial accuracy degradation under cross-dataset evaluation due to mismatched visual conditions. Illumination-aware augmentation significantly improves recognition robustness under domain shift while preserving real-time performance.

Conclusion: Highlights the importance of illumination robustness for deploying reliable food recognition systems in real-world inspection scenarios. Provides practical insights through illumination-aware augmentation techniques.

Abstract: Visual food recognition systems deployed in real-world environments, such as automated conveyor-belt inspection, are highly sensitive to domain shifts caused by illumination changes. While recent studies have shown that lighting variations can significantly distort food perception by both humans and AI, existing works are often limited to single food categories or controlled settings, and most public food datasets lack explicit illumination annotations. In this work, we investigate illumination-induced domain shift in multi-class food category recognition using two widely adopted datasets, Food-101 and Fruits-360. We demonstrate substantial accuracy degradation under cross-dataset evaluation due to mismatched visual conditions. To address this challenge, we construct synthetic illumination-augmented datasets by systematically varying light temperature and intensity, enabling controlled robustness analysis without additional labels. We further evaluate cross-dataset transfer learning and domain generalization, with a focus on illumination-sensitive target categories such as apple-based classes. Experimental results show that illumination-aware augmentation significantly improves recognition robustness under domain shift while preserving real-time performance. Our findings highlight the importance of illumination robustness and provide practical insights for deploying reliable food recognition systems in real-world inspection scenarios.

Method: Used two public EM datasets (Lucchi++ and VNC) with three VFMs (DINOv2, DINOv3, OpenCLIP). Evaluated two adaptation regimes: frozen-backbone with lightweight segmentation head, and parameter-efficient fine-tuning via LoRA. Analyzed latent spaces with PCA, Fréchet Dinov2 distance, and linear probes.

Result: Single-dataset training yields good segmentation performance with LoRA improving in-domain results. Multi-dataset training causes severe performance degradation for all models, with minimal gains from PEFT. Analysis reveals pronounced domain mismatch between EM datasets despite visual similarity.

Conclusion: VFMs can deliver competitive EM segmentation within single domains with lightweight adaptation, but current PEFT strategies are insufficient for robust cross-dataset generalization without additional domain-alignment mechanisms.

Abstract: Although vision foundation models (VFMs) are increasingly reused for biomedical image analysis, it remains unclear whether the latent representations they provide are general enough to support effective transfer and reuse across heterogeneous microscopy image datasets. Here, we study this question for the problem of mitochondria segmentation in electron microscopy (EM) images, using two popular public EM datasets (Lucchi++ and VNC) and three recent representative VFMs (DINOv2, DINOv3, and OpenCLIP). We evaluate two practical model adaptation regimes: a frozen-backbone setting in which only a lightweight segmentation head is trained on top of the VFM, and parameter-efficient fine-tuning (PEFT) via Low-Rank Adaptation (LoRA) in which the VFM is fine-tuned in a targeted manner to a specific dataset. Across all backbones, we observe that training on a single EM dataset yields good segmentation performance (quantified as foreground Intersection-over-Union), and that LoRA consistently improves in-domain performance. In contrast, training on multiple EM datasets leads to severe performance degradation for all models considered, with only marginal gains from PEFT. Exploration of the latent representation space through various techniques (PCA, Fréchet Dinov2 distance, and linear probes) reveals a pronounced and persistent domain mismatch between the two considered EM datasets in spite of their visual similarity, which is consistent with the observed failure of paired training. These results suggest that, while VFMs can deliver competitive results for EM segmentation within a single domain under lightweight adaptation, current PEFT strategies are insufficient to obtain a single robust model across heterogeneous EM datasets without additional domain-alignment mechanisms.

Method: GeoFocus comprises two core modules: 1) Critical Local Perceptor - automatically identifies and emphasizes critical local structures (angles, parallel lines, distances) using thirteen theory-based perception templates; 2) VertexLang - a compact topology formal language that encodes global figures through vertex coordinates and connectivity relations, replacing bulky code-based encodings.

Result: GeoFocus achieves 4.7% accuracy improvement over leading specialized models on Geo3K, GeoQA, and FormalGeo7K benchmarks. It boosts critical local feature coverage by 61% compared to previous methods, reduces global perception training time by 20%, and demonstrates superior robustness in MATHVERSE under diverse visual conditions.

Conclusion: GeoFocus effectively addresses geometry problem-solving challenges in LMMs by combining enhanced local structure perception with efficient global topology encoding, setting new state-of-the-art performance on multiple geometry benchmarks.

Abstract: Geometry problem-solving remains a significant challenge for Large Multimodal Models (LMMs), requiring not only global shape recognition but also attention to intricate local relationships related to geometric theory. To address this, we propose GeoFocus, a novel framework comprising two core modules. 1) Critical Local Perceptor, which automatically identifies and emphasizes critical local structure (e.g., angles, parallel lines, comparative distances) through thirteen theory-based perception templates, boosting critical local feature coverage by 61% compared to previous methods. 2) VertexLang, a compact topology formal language, encodes global figures through vertex coordinates and connectivity relations. By replacing bulky code-based encodings, VertexLang reduces global perception training time by 20% while improving topology recognition accuracy. When evaluated in Geo3K, GeoQA, and FormalGeo7K, GeoFocus achieves a 4.7% accuracy improvement over leading specialized models and demonstrates superior robustness in MATHVERSE under diverse visual conditions. Project Page – https://github.com/dle666/GeoFocus

Method: Uses two distinct computational discretizations of the same problem. A feedback control algorithm dynamically adjusts regularization strength, driving iterative reconstruction toward smallest parameter that yields sufficient similarity between reconstructions on the two grids.

Result: Demonstrated effectiveness using real tomographic data, showing the method can automatically select appropriate regularization parameters.

Conclusion: The proposed approach provides an automatic, data-driven method for regularization parameter selection in X-ray tomography, eliminating need for manual tuning.

Abstract: Image reconstruction in X-ray tomography is an ill-posed inverse problem, particularly with limited available data. Regularization is thus essential, but its effectiveness hinges on the choice of a regularization parameter that balances data fidelity against a priori information. We present a novel method for automatic parameter selection based on the use of two distinct computational discretizations of the same problem. A feedback control algorithm dynamically adjusts the regularization strength, driving an iterative reconstruction toward the smallest parameter that yields sufficient similarity between reconstructions on the two grids. The effectiveness of the proposed approach is demonstrated using real tomographic data.

Method: Proposes sparse monocular graph-based SLAM for thermal imagery using learned features (SuperPoint detector and LightGlue matcher) trained on visible-spectrum data. Includes preprocessing pipeline to enhance thermal input suitability, modifies SLAM modules to handle sparse/outlier-prone matches, and incorporates SuperPoint keypoint confidence scores into confidence-weighted factor graph optimization.

Result: Evaluation on public thermal datasets shows reliable performance without dataset-specific training or fine-tuning feature detectors, addressing the scarcity of quality thermal data.

Conclusion: The system demonstrates effective cross-domain generalization of learned features from visible to thermal spectrum, enabling robust thermal SLAM without requiring thermal-specific training data.

Abstract: Thermal imaging provides a practical sensing modality for visual SLAM in visually degraded environments such as low illumination, smoke, or adverse weather. However, thermal imagery often exhibits low texture, low contrast, and high noise, complicating feature-based SLAM. In this work, we propose a sparse monocular graph-based SLAM system for thermal imagery that leverages general-purpose learned features – the SuperPoint detector and LightGlue matcher, trained on large-scale visible-spectrum data to improve cross-domain generalization. To adapt these components to thermal data, we introduce a preprocessing pipeline to enhance input suitability and modify core SLAM modules to handle sparse and outlier-prone feature matches. We further incorporate keypoint confidence scores from SuperPoint into a confidence-weighted factor graph to improve estimation robustness. Evaluations on public thermal datasets demonstrate that the proposed system achieves reliable performance without requiring dataset-specific training or fine-tuning a desired feature detector, given the scarcity of quality thermal data. Code will be made available upon publication.

Method: Two-stage iterative boundary refinement: 1) Iterative Gaussian Instance Tracing (IGIT) at temporal segment level refines Gaussian-to-instance probabilities through iterative tracing to handle occlusions; 2) Gaussian Rendering Range Control (RCC) suppresses uncertain Gaussians near boundaries while retaining core contributions. Includes temporal segmentation merging strategy for identity consistency vs. dynamic awareness balance.

Result: Experiments on HyperNeRF and Neu3D show the method produces accurate object Gaussian point clouds with clearer boundaries and higher efficiency compared to state-of-the-art methods.

Conclusion: TIBR4D provides an efficient learning-free solution for 4D Gaussian segmentation that effectively handles complex motion, occlusions, and boundary ambiguity while maintaining object structure completeness.

Abstract: Object-level segmentation in dynamic 4D Gaussian scenes remains challenging due to complex motion, occlusions, and ambiguous boundaries. In this paper, we present an efficient learning-free 4D Gaussian segmentation framework that lifts video segmentation masks to 4D spaces, whose core is a two-stage iterative boundary refinement, TIBR4D. The first stage is an Iterative Gaussian Instance Tracing (IGIT) at the temporal segment level. It progressively refines Gaussian-to-instance probabilities through iterative tracing, and extracts corresponding Gaussian point clouds that better handle occlusions and preserve completeness of object structures compared to existing one-shot threshold-based methods. The second stage is a frame-wise Gaussian Rendering Range Control (RCC) via suppressing highly uncertain Gaussians near object boundaries while retaining their core contributions for more accurate boundaries. Furthermore, a temporal segmentation merging strategy is proposed for IGIT to balance identity consistency and dynamic awareness. Longer segments enforce stronger multi-frame constraints for stable identities, while shorter segments allow identity changes to be captured promptly. Experiments on HyperNeRF and Neu3D demonstrate that our method produces accurate object Gaussian point clouds with clearer boundaries and higher efficiency compared to SOTA methods.

Method: FLAG-4D employs a dual-deformation network with an Instantaneous Deformation Network (IDN) for fine-grained local deformations and a Global Motion Network (GMN) for long-range dynamics, refined through mutual learning. It incorporates dense motion features from pretrained optical flow and uses deformation-guided attention to align flow information with evolving 3D Gaussians.

Result: Extensive experiments demonstrate that FLAG-4D achieves higher-fidelity and more temporally coherent reconstructions with finer detail preservation than state-of-the-art methods.

Conclusion: FLAG-4D successfully overcomes limitations of existing dynamic scene reconstruction methods by using dual-deformation networks and motion feature integration, enabling more accurate and temporally smooth modeling of complex 3D scene dynamics.

Abstract: We introduce FLAG-4D, a novel framework for generating novel views of dynamic scenes by reconstructing how 3D Gaussian primitives evolve through space and time. Existing methods typically rely on a single Multilayer Perceptron (MLP) to model temporal deformations, and they often struggle to capture complex point motions and fine-grained dynamic details consistently over time, especially from sparse input views. Our approach, FLAG-4D, overcomes this by employing a dual-deformation network that dynamically warps a canonical set of 3D Gaussians over time into new positions and anisotropic shapes. This dual-deformation network consists of an Instantaneous Deformation Network (IDN) for modeling fine-grained, local deformations and a Global Motion Network (GMN) for capturing long-range dynamics, refined through mutual learning. To ensure these deformations are both accurate and temporally smooth, FLAG-4D incorporates dense motion features from a pretrained optical flow backbone. We fuse these motion cues from adjacent timeframes and use a deformation-guided attention mechanism to align this flow information with the current state of each evolving 3D Gaussian. Extensive experiments demonstrate that FLAG-4D achieves higher-fidelity and more temporally coherent reconstructions with finer detail preservation than state-of-the-art methods.

Method: SemiNFT uses a Diffusion Transformer (DiT) framework that follows human artistic training: first learns from paired triplets for structural preservation and color mapping, then advances to reinforcement learning on unpaired data for aesthetic perception. A hybrid online-offline reward mechanism prevents catastrophic forgetting during RL.

Result: Outperforms state-of-the-art methods on preset transfer benchmarks and demonstrates remarkable intelligence in zero-shot tasks like black-and-white photo colorization and cross-domain (anime-to-photo) preset transfer.

Conclusion: SemiNFT transcends simple statistical matching and achieves sophisticated aesthetic comprehension, confirming its ability to understand semantic context and human aesthetics in color retouching.

Abstract: Photorealistic color retouching plays a vital role in visual content creation, yet manual retouching remains inaccessible to non-experts due to its reliance on specialized expertise. Reference-based methods offer a promising alternative by transferring the preset color of a reference image to a source image. However, these approaches often operate as novice learners, performing global color mappings derived from pixel-level statistics, without a true understanding of semantic context or human aesthetics. To address this issue, we propose SemiNFT, a Diffusion Transformer (DiT)-based retouching framework that mirrors the trajectory of human artistic training: beginning with rigid imitation and evolving into intuitive creation. Specifically, SemiNFT is first taught with paired triplets to acquire basic structural preservation and color mapping skills, and then advanced to reinforcement learning (RL) on unpaired data to cultivate nuanced aesthetic perception. Crucially, during the RL stage, to prevent catastrophic forgetting of old skills, we design a hybrid online-offline reward mechanism that anchors aesthetic exploration with structural review. % experiments Extensive experiments show that SemiNFT not only outperforms state-of-the-art methods on standard preset transfer benchmarks but also demonstrates remarkable intelligence in zero-shot tasks, such as black-and-white photo colorization and cross-domain (anime-to-photo) preset transfer. These results confirm that SemiNFT transcends simple statistical matching and achieves a sophisticated level of aesthetic comprehension. Our project can be found at https://melanyyang.github.io/SemiNFT/.

Method: The paper reviews the AVS PCC standard from two perspectives: 1) analyzing the related technologies and coding tools adopted in AVS PCC, and 2) conducting performance comparisons with other standards like MPEG’s G-PCC and V-PCC.

Result: The AVS PCC standard incorporates many new coding tools and techniques that differ from other counterpart standards, suggesting unique technical approaches to point cloud compression.

Conclusion: AVS PCC represents China’s standardization effort in point cloud compression with distinct technical features compared to international standards, addressing the growing need for efficient 3D data compression in various applications.

Abstract: Point cloud is a prevalent 3D data representation format with significant application values in immersive media, autonomous driving, digital heritage protection, etc. However, the large data size of point clouds poses challenges to transmission and storage, which influences the wide deployments. Therefore, point cloud compression plays a crucial role in practical applications for both human and machine perception optimization. To this end, the Moving Picture Experts Group (MPEG) has established two standards for point cloud compression, including Geometry-based Point Cloud Compression (G-PCC) and Video-based Point Cloud Compression (V-PCC). In the meantime, the Audio Video coding Standard (AVS) Workgroup of China also have launched and completed the development for its first generation point cloud compression standard, namely AVS PCC. This new standardization effort has adopted many new coding tools and techniques, which are different from the other counterpart standards. This paper reviews the AVS PCC standard from two perspectives, i.e., the related technologies and performance comparisons.

Method: Feed-forward framework trained on synthetic triplets of decomposed visual aspects using CLIP Sparse Autoencoders to extract editing directions in CLIP latent space and isolate concept pairs, removing reliance on language.

Result: Produces diverse, visually coherent compositions that reveal latent relationships between input images, enabling fast, intuitive recombination for visual ideation.

Conclusion: The framework shifts image generation from final execution to exploratory ideation, supporting creative work at early, ambiguous stages by enabling visual exploration without text prompts.

Abstract: While generative models have become powerful tools for image synthesis, they are typically optimized for executing carefully crafted textual prompts, offering limited support for the open-ended visual exploration that often precedes idea formation. In contrast, designers frequently draw inspiration from loosely connected visual references, seeking emergent connections that spark new ideas. We propose Inspiration Seeds, a generative framework that shifts image generation from final execution to exploratory ideation. Given two input images, our model produces diverse, visually coherent compositions that reveal latent relationships between inputs, without relying on user-specified text prompts. Our approach is feed-forward, trained on synthetic triplets of decomposed visual aspects derived entirely through visual means: we use CLIP Sparse Autoencoders to extract editing directions in CLIP latent space and isolate concept pairs. By removing the reliance on language and enabling fast, intuitive recombination, our method supports visual ideation at the early and ambiguous stages of creative work.

Method: Proposes LV-RAE that augments semantic features with missing low-level information; addresses decoder sensitivity to latent perturbations by fine-tuning decoder for robustness and smoothing generated latents via controlled noise injection.

Result: LV-RAE significantly improves reconstruction fidelity while preserving semantic abstraction and achieving strong generative quality, overcoming the bottleneck in scaling latent diffusion models.

Conclusion: The proposed approach successfully addresses the low-level information deficiency in vision foundation model features, enabling high-fidelity reconstruction and better generation quality in latent diffusion models.

Abstract: Recent work leverages Vision Foundation Models as image encoders to boost the generative performance of latent diffusion models (LDMs), as their semantic feature distributions are easy to learn. However, such semantic features often lack low-level information (\eg, color and texture), leading to degraded reconstruction fidelity, which has emerged as a primary bottleneck in further scaling LDMs. To address this limitation, we propose LV-RAE, a representation autoencoder that augments semantic features with missing low-level information, enabling high-fidelity reconstruction while remaining highly aligned with the semantic distribution. We further observe that the resulting high-dimensional, information-rich latent make decoders sensitive to latent perturbations, causing severe artifacts when decoding generated latent and consequently degrading generation quality. Our analysis suggests that this sensitivity primarily stems from excessive decoder responses along directions off the data manifold. Building on these insights, we propose fine-tuning the decoder to increase its robustness and smoothing the generated latent via controlled noise injection, thereby enhancing generation quality. Experiments demonstrate that LV-RAE significantly improves reconstruction fidelity while preserving the semantic abstraction and achieving strong generative quality. Our code is available at https://github.com/modyu-liu/LVRAE.

Method: Analyze interactions between class and patch tokens, discover normalization layers create implicit differentiation, then propose specialized processing paths that disentangle computational flow for different token types, particularly in normalization layers and early query-key-value projections

Result: Significant improvements in dense prediction tasks (over 2 mIoU points gain in segmentation), maintains classification accuracy, only 8% parameter increase with no computational overhead

Conclusion: Targeted specialization of processing paths for different token types in Vision Transformers enhances representation quality for dense prediction while preserving efficiency

Abstract: Vision Transformers have emerged as powerful, scalable and versatile representation learners. To capture both global and local features, a learnable [CLS] class token is typically prepended to the input sequence of patch tokens. Despite their distinct nature, both token types are processed identically throughout the model. In this work, we investigate the friction between global and local feature learning under different pre-training strategies by analyzing the interactions between class and patch tokens. Our analysis reveals that standard normalization layers introduce an implicit differentiation between these token types. Building on this insight, we propose specialized processing paths that selectively disentangle the computational flow of class and patch tokens, particularly within normalization layers and early query-key-value projections. This targeted specialization leads to significantly improved patch representation quality for dense prediction tasks. Our experiments demonstrate segmentation performance gains of over 2 mIoU points on standard benchmarks, while maintaining strong classification accuracy. The proposed modifications introduce only an 8% increase in parameters, with no additional computational overhead. Through comprehensive ablations, we provide insights into which architectural components benefit most from specialization and how our approach generalizes across model scales and learning frameworks.

Method: Developed a deep learning model that predicts fixation types using low-resolution, pre-scan thumbnail images. Trained on 1,200 WSIs from TUM Institute of Pathology and evaluated on multiple datasets including TCGA (8,800 slides), Augsburg (695 slides), and Regensburg (202 slides) from different scanners.

Result: Model achieves AUROC of 0.88 on TCGA dataset, outperforming comparable pre-scan methods by 4.8%. Achieves AUROCs of 0.72 on Regensburg and Augsburg datasets, showing challenges with scanner-induced domain shifts. Processes each slide in 21 ms, 400× faster than existing high-magnification methods.

Conclusion: The approach provides an efficient solution for detecting labeling errors without high-magnification scans, offering a valuable tool for quality control in high-throughput pathology workflows. Future work will improve generalization to additional scanner types and extend to other low-resolution slide annotations.

Abstract: Accurate annotation of fixation type is a critical step in slide preparation for pathology laboratories. However, this manual process is prone to errors, impacting downstream analyses and diagnostic accuracy. Existing methods for verifying formalin-fixed, paraffin-embedded (FFPE), and frozen section (FS) fixation types typically require full-resolution whole-slide images (WSIs), limiting scalability for high-throughput quality control. We propose a deep-learning model to predict fixation types using low-resolution, pre-scan thumbnail images. The model was trained on WSIs from the TUM Institute of Pathology (n=1,200, Leica GT450DX) and evaluated on a class-balanced subset of The Cancer Genome Atlas dataset (TCGA, n=8,800, Leica AT2), as well as on class-balanced datasets from Augsburg (n=695 [392 FFPE, 303 FS], Philips UFS) and Regensburg (n=202, 3DHISTECH P1000). Our model achieves an AUROC of 0.88 on TCGA, outperforming comparable pre-scan methods by 4.8%. It also achieves AUROCs of 0.72 on Regensburg and Augsburg slides, underscoring challenges related to scanner-induced domain shifts. Furthermore, the model processes each slide in 21 ms, $400\times$ faster than existing high-magnification, full-resolution methods, enabling rapid, high-throughput processing. This approach provides an efficient solution for detecting labelling errors without relying on high-magnification scans, offering a valuable tool for quality control in high-throughput pathology workflows. Future work will improve and evaluate the model’s generalisation to additional scanner types. Our findings suggest that this method can increase accuracy and efficiency in digital pathology workflows and may be extended to other low-resolution slide annotations.

Method: Encoder-decoder architecture: encoder uses temporal and asymmetric convolutions to capture spatio-temporal CSI features while preserving sequential structure, refines keypoint features via axial attention for structural dependencies, decoder maps to keypoint coordinates.

Result: Achieves PCK@20 of 97.00% and PCK@50 of 99.48% with mean per-joint error of 0.008m on 360,000 CSI-pose samples from 5 subjects performing 8 daily activities. Only 4.82M parameters, reducing computational cost.

Conclusion: WiFlow establishes new performance baseline for practical WiFi-based human pose estimation with reduced model complexity, enabling continuous motion tracking for IoT applications.

Abstract: Human pose estimation is fundamental to intelligent perception in the Internet of Things (IoT), enabling applications ranging from smart healthcare to human-computer interaction. While WiFi-based methods have gained traction, they often struggle with continuous motion and high computational overhead. This work presents WiFlow, a novel framework for continuous human pose estimation using WiFi signals. Unlike vision-based approaches such as two-dimensional deep residual networks that treat Channel State Information (CSI) as images, WiFlow employs an encoder-decoder architecture. The encoder captures spatio-temporal features of CSI using temporal and asymmetric convolutions, preserving the original sequential structure of signals. It then refines keypoint features of human bodies to be tracked and capture their structural dependencies via axial attention. The decoder subsequently maps the encoded high-dimensional features into keypoint coordinates. Trained on a self-collected dataset of 360,000 synchronized CSI-pose samples from 5 subjects performing continuous sequences of 8 daily activities, WiFlow achieves a Percentage of Correct Keypoints (PCK) of 97.00% at a threshold of 20% (PCK@20) and 99.48% at PCK@50, with a mean per-joint position error of 0.008m. With only 4.82M parameters, WiFlow significantly reduces model complexity and computational cost, establishing a new performance baseline for practical WiFi-based human pose estimation. Our code and datasets are available at https://github.com/DY2434/WiFlow-WiFi-Pose-Estimation-with-Spatio-Temporal-Decoupling.git.

Method: Implemented classic solvers for shallow water and Euler equations, then proposed four different deep neural network architectures for ML-based PDE solvers using data-driven discretization that predicts coefficients of quasi-linear stencils.

Result: The classic solvers performed much better than Pyclaw solver, and two of the four proposed neural network approaches produced satisfactory solutions for the PDE problems.

Conclusion: Data-driven discretization methods with neural networks show promise for accelerating and improving PDE solvers on structured grids, combining ML benefits with traditional numerical scheme advantages like conservation laws.

Abstract: Machine learning methods have been successful in many areas, like image classification and natural language processing. However, it still needs to be determined how to apply ML to areas with mathematical constraints, like solving PDEs. Among various approaches to applying ML techniques to solving PDEs, the data-driven discretization method presents a promising way of accelerating and improving existing PDE solver on structured grids where it predicts the coefficients of quasi-linear stencils for computing values or derivatives of a function at given positions. It can improve the accuracy and stability of low-resolution simulation compared with using traditional finite difference or finite volume schemes. Meanwhile, it can also benefit from traditional numerical schemes like achieving conservation law by adapting finite volume type formulations. In this thesis, we have implemented the shallow water equation and Euler equation classic solver under a different framework. Experiments show that our classic solver performs much better than the Pyclaw solver. Then we propose four different deep neural networks for the ML-based solver. The results indicate that two of these approaches could output satisfactory solutions.

Method: Augments MMDiT architecture with joint audio-video branch featuring TA-CrossAttn for temporally-aligned cross-modal fusion and UniTemp-RoPE for precise audio-visual alignment. Uses comprehensive data pipeline for high-quality finetuning data collection and introduces new benchmark for evaluation.

Result: ALIVE demonstrates outstanding performance after training on million-level high-quality data, consistently outperforming open-source models and matching or surpassing state-of-the-art commercial solutions.

Conclusion: ALIVE successfully enables audio-video generation capabilities in T2V models, providing detailed recipes and benchmarks to help the community develop such models more efficiently.

Abstract: Video generation is rapidly evolving towards unified audio-video generation. In this paper, we present ALIVE, a generation model that adapts a pretrained Text-to-Video (T2V) model to Sora-style audio-video generation and animation. In particular, the model unlocks the Text-to-Video&Audio (T2VA) and Reference-to-Video&Audio (animation) capabilities compared to the T2V foundation models. To support the audio-visual synchronization and reference animation, we augment the popular MMDiT architecture with a joint audio-video branch which includes TA-CrossAttn for temporally-aligned cross-modal fusion and UniTemp-RoPE for precise audio-visual alignment. Meanwhile, a comprehensive data pipeline consisting of audio-video captioning, quality control, etc., is carefully designed to collect high-quality finetuning data. Additionally, we introduce a new benchmark to perform a comprehensive model test and comparison. After continue pretraining and finetuning on million-level high-quality data, ALIVE demonstrates outstanding performance, consistently outperforming open-source models and matching or surpassing state-of-the-art commercial solutions. With detailed recipes and benchmarks, we hope ALIVE helps the community develop audio-video generation models more efficiently. Official page: https://github.com/FoundationVision/Alive.

Method: OneVision-Encoder uses Codec Patchification to focus computation on only 3.1%-25% of regions with high signal entropy. It employs shared 3D RoPE for spatial-temporal reasoning with irregular token layouts and is trained with large-scale cluster discrimination over 1M+ semantic concepts to capture object permanence and motion dynamics.

Result: Outperforms strong vision backbones like Qwen3-ViT and SigLIP2 across 16 image, video, and document understanding benchmarks despite using fewer visual tokens and less pretraining data. Achieves 4.1% average improvement over Qwen3-ViT on video tasks, demonstrating efficiency-accuracy positive correlation.

Conclusion: Codec-aligned patch-level sparsity is a foundational principle for scalable visual generalists. Efficiency and accuracy are positively correlated when architectures align with information-theoretic principles of video compression.

Abstract: Hypothesis. Artificial general intelligence is, at its core, a compression problem. Effective compression demands resonance: deep learning scales best when its architecture aligns with the fundamental structure of the data. These are the fundamental principles. Yet, modern vision architectures have strayed from these truths: visual signals are highly redundant, while discriminative information, the surprise, is sparse. Current models process dense pixel grids uniformly, wasting vast compute on static background rather than focusing on the predictive residuals that define motion and meaning. We argue that to solve visual understanding, we must align our architectures with the information-theoretic principles of video, i.e., Codecs. Method. OneVision-Encoder encodes video by compressing predictive visual structure into semantic meaning. By adopting Codec Patchification, OV-Encoder abandons uniform computation to focus exclusively on the 3.1%-25% of regions rich in signal entropy. To unify spatial and temporal reasoning under irregular token layouts, OneVision-Encoder employs a shared 3D RoPE and is trained with a large-scale cluster discrimination objective over more than one million semantic concepts, jointly capturing object permanence and motion dynamics. Evidence. The results validate our core hypothesis: efficiency and accuracy are not a trade-off; they are positively correlated. When integrated into LLM, it consistently outperforms strong vision backbones such as Qwen3-ViT and SigLIP2 across 16 image, video, and document understanding benchmarks, despite using substantially fewer visual tokens and pretraining data. Notably, on video understanding tasks, OV-Encoder achieves an average improvement of 4.1% over Qwen3-ViT. Codec-aligned, patch-level sparsity is a foundational principle, enabling OV-Encoder as a scalable engine for next-generation visual generalists.

Method: Proposes view-aware decomposition with view-independent (appearance) and view-dependent (shading) components. Uses dynamic cross-frame correspondences for view-independent term and scene-level continuity constraints for view-dependent term. Introduces dual-structure enhancement network with cross-frame interaction mechanism. VLLVE++ adds residual term for scene-adaptive degradations and enables bidirectional learning for enhancement and degradation-aware correspondence refinement.

Result: Extensive experiments on LLVE benchmarks show strong performance. VLLVE++ demonstrates capability in handling challenging real-world scenes and high-dynamic videos, effectively increasing reliable correspondences while filtering incorrect ones.

Conclusion: The view-aware decomposition strategy with cross-frame correspondences and continuity constraints effectively enhances low-light videos. VLLVE++’s residual term and bidirectional learning further improve performance for complex real-world scenarios.

Abstract: Low-Light Video Enhancement (LLVE) seeks to restore dynamic or static scenes plagued by severe invisibility and noise. In this paper, we present an innovative video decomposition strategy that incorporates view-independent and view-dependent components to enhance the performance of LLVE. The framework is called View-aware Low-light Video Enhancement (VLLVE). We leverage dynamic cross-frame correspondences for the view-independent term (which primarily captures intrinsic appearance) and impose a scene-level continuity constraint on the view-dependent term (which mainly describes the shading condition) to achieve consistent and satisfactory decomposition results. To further ensure consistent decomposition, we introduce a dual-structure enhancement network featuring a cross-frame interaction mechanism. By supervising different frames simultaneously, this network encourages them to exhibit matching decomposition features. This mechanism can seamlessly integrate with encoder-decoder single-frame networks, incurring minimal additional parameter costs. Building upon VLLVE, we propose a more comprehensive decomposition strategy by introducing an additive residual term, resulting in VLLVE++. This residual term can simulate scene-adaptive degradations, which are difficult to model using a decomposition formulation for common scenes, thereby further enhancing the ability to capture the overall content of videos. In addition, VLLVE++ enables bidirectional learning for both enhancement and degradation-aware correspondence refinement (end-to-end manner), effectively increasing reliable correspondences while filtering out incorrect ones. Notably, VLLVE++ demonstrates strong capability in handling challenging cases, such as real-world scenes and videos with high dynamics. Extensive experiments are conducted on widely recognized LLVE benchmarks.

Method: 1) Introduces Omni Dense Captioning task with six-dimensional structural schema for script-like captions; 2) Creates OmniDCBench benchmark with human annotations; 3) Proposes SodaM metric for time-aware detailed descriptions; 4) Builds TimeChatCap-42K training dataset; 5) Develops TimeChat-Captioner-7B baseline trained via SFT and GRPO with task-specific rewards.

Result: TimeChat-Captioner-7B achieves state-of-the-art performance, surpassing Gemini-2.5-Pro. Generated dense descriptions significantly boost downstream capabilities in audio-visual reasoning (DailyOmni and WorldSense) and temporal grounding (Charades-STA).

Conclusion: The paper successfully establishes a new task for dense audio-visual narrative generation with structured timestamps, provides comprehensive resources (benchmark, metric, dataset), and demonstrates strong performance with practical benefits for downstream audio-visual reasoning tasks.

Abstract: This paper proposes Omni Dense Captioning, a novel task designed to generate continuous, fine-grained, and structured audio-visual narratives with explicit timestamps. To ensure dense semantic coverage, we introduce a six-dimensional structural schema to create “script-like” captions, enabling readers to vividly imagine the video content scene by scene, akin to a cinematographic screenplay. To facilitate research, we construct OmniDCBench, a high-quality, human-annotated benchmark, and propose SodaM, a unified metric that evaluates time-aware detailed descriptions while mitigating scene boundary ambiguity. Furthermore, we construct a training dataset, TimeChatCap-42K, and present TimeChat-Captioner-7B, a strong baseline trained via SFT and GRPO with task-specific rewards. Extensive experiments demonstrate that TimeChat-Captioner-7B achieves state-of-the-art performance, surpassing Gemini-2.5-Pro, while its generated dense descriptions significantly boost downstream capabilities in audio-visual reasoning (DailyOmni and WorldSense) and temporal grounding (Charades-STA). All datasets, models, and code will be made publicly available at https://github.com/yaolinli/TimeChat-Captioner.

Method: Uses stage-wise model diffing technique to isolate representational changes introduced during multimodal fine-tuning. Identifies vision-preferring features that emerge or reorient, analyzes spatial relation encoding through controlled spatial prompt shifts, and traces causal activation to specific attention heads.

Result: Reveals how language models learn visual capabilities: identifies vision-preferring features, shows selective subset encodes spatial relations, traces activation to specific attention heads. Demonstrates stage-wise model diffing reveals when/where spatially grounded multimodal features arise.

Conclusion: Stage-wise model diffing provides mechanistic understanding of multimodal adaptation, showing how visual grounding reshapes text-only features. Enhances interpretability of multimodal training and provides foundation for understanding how pretrained language models acquire vision-grounded capabilities.

Abstract: Contemporary Vision-Language Models (VLMs) achieve strong performance on a wide range of tasks by pairing a vision encoder with a pre-trained language model, fine-tuned for visual-text inputs. Yet despite these gains, it remains unclear how language backbone representations adapt during multimodal training and when vision-specific capabilities emerge. In this work, we present the first mechanistic analysis of VLM adaptation. Using stage-wise model diffing, a technique that isolates representational changes introduced during multimodal fine-tuning, we reveal how a language model learns to “see”. We first identify vision-preferring features that emerge or reorient during fine-tuning. We then show that a selective subset of these features reliably encodes spatial relations, revealed through controlled shifts to spatial prompts. Finally, we trace the causal activation of these features to a small group of attention heads. Our findings show that stage-wise model diffing reveals when and where spatially grounded multimodal features arise. It also provides a clearer view of modality fusion by showing how visual grounding reshapes features that were previously text-only. This methodology enhances the interpretability of multimodal training and provides a foundation for understanding and refining how pretrained language models acquire vision-grounded capabilities.

Method: Three training-free pipelines: (1) segmentation-driven rule-based system using pre-trained multi-organ segmentation models, (2) MLLM guided by radiologist-defined rules, (3) segmentation-aware MLLM combining visual input with anatomical evidence.

Result: Segmentation-driven rule-based approach achieved best performance: weighted F1-scores of 0.947 (CT) and 0.914 (MR). MLLM performed competitively in visually distinctive regions. Segmentation-aware MLLM showed fundamental limitations.

Conclusion: Zero-shot body region detection is feasible using foundation models. Segmentation-driven rule-based approach is most robust, while MLLMs have potential but current limitations.

Abstract: Reliable identification of anatomical body regions is a prerequisite for many automated medical imaging workflows, yet existing solutions remain heavily dependent on unreliable DICOM metadata. Current solutions mainly use supervised learning, which limits their applicability in many real-world scenarios. In this work, we investigate whether body region detection in volumetric CT and MR images can be achieved in a fully zero-shot manner by using knowledge embedded in large pre-trained foundation models. We propose and systematically evaluate three training-free pipelines: (1) a segmentation-driven rule-based system leveraging pre-trained multi-organ segmentation models, (2) a Multimodal Large Language Model (MLLM) guided by radiologist-defined rules, and (3) a segmentation-aware MLLM that combines visual input with explicit anatomical evidence. All methods are evaluated on 887 heterogeneous CT and MR scans with manually verified anatomical region labels. The segmentation-driven rule-based approach achieves the strongest and most consistent performance, with weighted F1-scores of 0.947 (CT) and 0.914 (MR), demonstrating robustness across modalities and atypical scan coverage. The MLLM performs competitively in visually distinctive regions, while the segmentation-aware MLLM reveals fundamental limitations.

Method: Proposes RotLight capture setup requiring object rotation during capture (as few as two rotations) to reduce ambiguity. Also introduces proxy mesh for accurate incident light tracing, residual constraint, and improved global illumination handling in 2DGS-based inverse rendering.

Result: Demonstrates superior albedo estimation on both synthetic and real-world datasets while maintaining efficient computation compared to prior methods.

Conclusion: RotLight provides an effective yet simple solution to the ambiguity problem in inverse rendering through object rotation during capture and proxy mesh integration, achieving better material decomposition with minimal additional capture requirements.

Abstract: Inverse rendering aims to decompose a scene into its geometry, material properties and light conditions under a certain rendering model. It has wide applications like view synthesis, relighting, and scene editing. In recent years, inverse rendering methods have been inspired by view synthesis approaches like neural radiance fields and Gaussian splatting, which are capable of efficiently decomposing a scene into its geometry and radiance. They then further estimate the material and lighting that lead to the observed scene radiance. However, the latter step is highly ambiguous and prior works suffer from inaccurate color and baked shadows in their albedo estimation albeit their regularization. To this end, we propose RotLight, a simple capturing setup, to address the ambiguity. Compared to a usual capture, RotLight only requires the object to be rotated several times during the process. We show that as few as two rotations is effective in reducing artifacts. To further improve 2DGS-based inverse rendering, we additionally introduce a proxy mesh that not only allows accurate incident light tracing, but also enables a residual constraint and improves global illumination handling. We demonstrate with both synthetic and real world datasets that our method achieves superior albedo estimation while keeping efficient computation.

Method: 1) Automatically identifies editing vs. preserved regions by measuring semantic discrepancies between source and target prompts. 2) Uses distance-aware latent fusion along region boundaries to create soft masks and applies total variation loss for smooth transitions. 3) Employs AdaIN-based modulation within Diffusion Transformer (DiT) attention layers for statistical attention fusion in editing regions.

Result: Extensive experiments show FusionEdit significantly outperforms state-of-the-art methods in text-guided image editing, achieving more precise and controllable edits with fewer boundary artifacts and better preservation of source image identity.

Conclusion: FusionEdit provides an effective training-free framework for text-guided image editing that addresses boundary artifact issues through soft mask generation and attention modulation, enabling high-quality, controllable edits while maintaining source image consistency.

Abstract: Text-guided image editing aims to modify specific regions according to the target prompt while preserving the identity of the source image. Recent methods exploit explicit binary masks to constrain editing, but hard mask boundaries introduce artifacts and reduce editability. To address these issues, we propose FusionEdit, a training-free image editing framework that achieves precise and controllable edits. First, editing and preserved regions are automatically identified by measuring semantic discrepancies between source and target prompts. To mitigate boundary artifacts, FusionEdit performs distance-aware latent fusion along region boundaries to yield the soft and accurate mask, and employs a total variation loss to enforce smooth transitions, obtaining natural editing results. Second, FusionEdit leverages AdaIN-based modulation within DiT attention layers to perform a statistical attention fusion in the editing region, enhancing editability while preserving global consistency with the source image. Extensive experiments demonstrate that our FusionEdit significantly outperforms state-of-the-art methods. Code is available at \href{https://github.com/Yvan1001/FusionEdit}{https://github.com/Yvan1001/FusionEdit}.

Method: Created synthetic dataset using Blender to simulate saccades and fixations under controlled conditions. Used Spiking Neural Networks (SNNs) trained on synthetic data and fine-tuned on real event data. Compared with artificial neural network (ANN) counterparts.

Result: Models achieved up to 0.83 accuracy in eye movement classification, maintained consistent performance across varying temporal resolutions, and showed substantial computational efficiency gains over ANN counterparts.

Conclusion: Synthetic data augmentation with event cameras and SNNs provides an effective approach for eye movement classification, offering computational efficiency and robustness across temporal resolutions.

Abstract: The study of eye movements, particularly saccades and fixations, are fundamental to understanding the mechanisms of human cognition and perception. Accurate classification of these movements requires sensing technologies capable of capturing rapid dynamics without distortion. Event cameras, also known as Dynamic Vision Sensors (DVS), provide asynchronous recordings of changes in light intensity, thereby eliminating motion blur inherent in conventional frame-based cameras and offering superior temporal resolution and data efficiency. In this study, we introduce a synthetic dataset generated with Blender to simulate saccades and fixations under controlled conditions. Leveraging Spiking Neural Networks (SNNs), we evaluate its robustness by training two architectures and finetuning on real event data. The proposed models achieve up to 0.83 accuracy and maintain consistent performance across varying temporal resolutions, demonstrating stability in eye movement classification. Moreover, the use of SNNs with synthetic event streams yields substantial computational efficiency gains over artificial neural network (ANN) counterparts, underscoring the utility of synthetic data augmentation in advancing event-based vision. All code and datasets associated with this work is available at https: //github.com/Ikhadija-5/SynSacc-Dataset.

Method: Two-stage hybrid framework: First, a 2D U-Net processes individual slices of undersampled CT volumes to extract feature maps. These slice-wise feature maps are then stacked and fed into a 3D decoder that uses contextual information across slices to predict artifact-free 3D CT volumes.

Result: The approach shows substantial improvements in inter-slice consistency in coronal and sagittal directions with low computational overhead, balancing computational efficiency with volumetric consistency.

Conclusion: The hybrid framework presents a robust and efficient solution for high-quality 3D CT image post-processing, effectively addressing artifacts in undersampled CT data.

Abstract: Undersampled CT volumes minimize acquisition time and radiation exposure but introduce artifacts degrading image quality and diagnostic utility. Reducing these artifacts is critical for high-quality imaging. We propose a computationally efficient hybrid deep-learning framework that combines the strengths of 2D and 3D models. First, a 2D U-Net operates on individual slices of undersampled CT volumes to extract feature maps. These slice-wise feature maps are then stacked across the volume and used as input to a 3D decoder, which utilizes contextual information across slices to predict an artifact-free 3D CT volume. The proposed two-stage approach balances the computational efficiency of 2D processing with the volumetric consistency provided by 3D modeling. The results show substantial improvements in inter-slice consistency in coronal and sagittal direction with low computational overhead. This hybrid framework presents a robust and efficient solution for high-quality 3D CT image post-processing. The code of this project can be found on github: https://github.com/J-3TO/2D-3DCNN_sparseview/.

Method: Three components: MCA detects directional confusion pairs from source model predictions; MCC uses CLIP to create confusion-aware textual prompts for better pseudo-labeling; FAM builds confusion-guided feature banks and aligns them using contrastive learning.

Result: CGA consistently outperforms state-of-the-art SFDA methods across various datasets, with especially notable gains in confusion-prone and fine-grained scenarios.

Conclusion: Explicitly modeling inter-class confusion is crucial for effective source-free adaptation, and CLIP-guided alignment provides a powerful framework for addressing this challenge.

Abstract: Source-Free Domain Adaptation (SFDA) tackles the problem of adapting a pre-trained source model to an unlabeled target domain without accessing any source data, which is quite suitable for the field of data security. Although recent advances have shown that pseudo-labeling strategies can be effective, they often fail in fine-grained scenarios due to subtle inter-class similarities. A critical but underexplored issue is the presence of asymmetric and dynamic class confusion, where visually similar classes are unequally and inconsistently misclassified by the source model. Existing methods typically ignore such confusion patterns, leading to noisy pseudo-labels and poor target discrimination. To address this, we propose CLIP-Guided Alignment(CGA), a novel framework that explicitly models and mitigates class confusion in SFDA. Generally, our method consists of three parts: (1) MCA: detects first directional confusion pairs by analyzing the predictions of the source model in the target domain; (2) MCC: leverages CLIP to construct confusion-aware textual prompts (e.g. a truck that looks like a bus), enabling more context-sensitive pseudo-labeling; and (3) FAM: builds confusion-guided feature banks for both CLIP and the source model and aligns them using contrastive learning to reduce ambiguity in the representation space. Extensive experiments on various datasets demonstrate that CGA consistently outperforms state-of-the-art SFDA methods, with especially notable gains in confusion-prone and fine-grained scenarios. Our results highlight the importance of explicitly modeling inter-class confusion for effective source-free adaptation. Our code can be find at https://github.com/soloiro/CGA

Method: Proposes HATCH framework with two complementary objectives: (1) Patch-Level Spatial Alignment - encourages patch representations to align across views for spatially corresponding regions, and (2) Action-then-Answer Reasoning - requires generating explicit viewpoint transition actions before predicting final answers.

Result: Experiments on three benchmarks show HATCH consistently outperforms comparable-size baselines by clear margins and achieves competitive results against much larger models, while preserving single-image reasoning capabilities.

Conclusion: Explicitly incorporating both cross-view correspondence and viewpoint change mechanisms through HATCH significantly improves multi-image spatial reasoning in MLLMs, addressing a key limitation in current approaches.

Abstract: While multimodal large language models (MLLMs) have made substantial progress in single-image spatial reasoning, multi-image spatial reasoning, which requires integration of information from multiple viewpoints, remains challenging. Cognitive studies suggest that humans address such tasks through two mechanisms: cross-view correspondence, which identifies regions across different views that correspond to the same physical locations, and stepwise viewpoint transformation, which composes relative viewpoint changes sequentially. However, existing studies incorporate these mechanisms only partially and often implicitly, without explicit supervision for both. We propose Human-Aware Training for Cross-view correspondence and viewpoint cHange (HATCH), a training framework with two complementary objectives: (1) Patch-Level Spatial Alignment, which encourages patch representations to align across views for spatially corresponding regions, and (2) Action-then-Answer Reasoning, which requires the model to generate explicit viewpoint transition actions before predicting the final answer. Experiments on three benchmarks demonstrate that HATCH consistently outperforms baselines of comparable size by a clear margin and achieves competitive results against much larger models, while preserving single-image reasoning capabilities.

Method: Introduces Instance-Disentangled Attention that partitions joint attention operations to enforce binding between instance-specific textual instructions and spatial regions during velocity field estimation, enabling disentangled editing.

Result: The approach promotes edit disentanglement and locality while preserving global output coherence, enabling single-pass, instance-level editing on both natural images and text-dense infographics with region-level editing instructions.

Conclusion: Instance-Disentangled Attention effectively addresses the multi-instance editing limitation in flow matching models, enabling independent editing of multiple parts without semantic interference while maintaining overall coherence.

Abstract: Flow matching models have recently emerged as an efficient alternative to diffusion, especially for text-guided image generation and editing, offering faster inference through continuous-time dynamics. However, existing flow-based editors predominantly support global or single-instruction edits and struggle with multi-instance scenarios, where multiple parts of a reference input must be edited independently without semantic interference. We identify this limitation as a consequence of globally conditioned velocity fields and joint attention mechanisms, which entangle concurrent edits. To address this issue, we introduce Instance-Disentangled Attention, a mechanism that partitions joint attention operations, enforcing binding between instance-specific textual instructions and spatial regions during velocity field estimation. We evaluate our approach on both natural image editing and a newly introduced benchmark of text-dense infographics with region-level editing instructions. Experimental results demonstrate that our approach promotes edit disentanglement and locality while preserving global output coherence, enabling single-pass, instance-level editing.

Method: MVAnimate synthesizes both 2D and 3D information of dynamic figures based on multi-view prior information. It leverages multi-view data to produce temporally consistent and spatially coherent animations, and optimizes multi-view videos of target characters from different views.

Result: Experimental results on diverse datasets demonstrate robustness in handling various motion patterns and appearances, showing improvements over existing animation methods in generated video quality.

Conclusion: MVAnimate provides a novel framework that addresses quality limitations in character animation generation by effectively utilizing multi-view prior information for enhanced 2D and 3D animation synthesis.

Abstract: The demand for realistic and versatile character animation has surged, driven by its wide-ranging applications in various domains. However, the animation generation algorithms modeling human pose with 2D or 3D structures all face various problems, including low-quality output content and training data deficiency, preventing the related algorithms from generating high-quality animation videos. Therefore, we introduce MVAnimate, a novel framework that synthesizes both 2D and 3D information of dynamic figures based on multi-view prior information, to enhance the generated video quality. Our approach leverages multi-view prior information to produce temporally consistent and spatially coherent animation outputs, demonstrating improvements over existing animation methods. Our MVAnimate also optimizes the multi-view videos of the target character, enhancing the video quality from different views. Experimental results on diverse datasets highlight the robustness of our method in handling various motion patterns and appearances.

Method: Symbolic Vedic Computation framework: converts speech to time-aligned phonemes, maps phonemes to compact viseme inventory, produces smooth viseme trajectories using symbolic coarticulation inspired by Vedic sutra Urdhva Tiryakbhyam, and uses lightweight 2D renderer with ROI warping and mouth compositing for real-time CPU synthesis.

Result: Achieves acceptable lip-sync quality while substantially reducing computational load and latency. Demonstrates synchronization accuracy, temporal stability, and identity consistency under CPU-only execution, outperforming representative CPU-feasible baselines.

Conclusion: The framework enables practical educational avatars on low-end hardware by providing a deterministic, CPU-oriented talking-head generation solution that balances quality with computational efficiency.

Abstract: Talking-head avatars are increasingly adopted in educational technology to deliver content with social presence and improved engagement. However, many recent talking-head generation (THG) methods rely on GPU-centric neural rendering, large training sets, or high-capacity diffusion models, which limits deployment in offline or resource-constrained learning environments. A deterministic and CPU-oriented THG framework is described, termed Symbolic Vedic Computation, that converts speech to a time-aligned phoneme stream, maps phonemes to a compact viseme inventory, and produces smooth viseme trajectories through symbolic coarticulation inspired by Vedic sutra Urdhva Tiryakbhyam. A lightweight 2D renderer performs region-of-interest (ROI) warping and mouth compositing with stabilization to support real-time synthesis on commodity CPUs. Experiments report synchronization accuracy, temporal stability, and identity consistency under CPU-only execution, alongside benchmarking against representative CPU-feasible baselines. Results indicate that acceptable lip-sync quality can be achieved while substantially reducing computational load and latency, supporting practical educational avatars on low-end hardware. GitHub: https://vineetkumarrakesh.github.io/vedicthg

Method: Proposes MultiDeepSAD, an extension of DeepSAD algorithm for multiple modalities with new loss formulation. Creates two multimodal datasets (real SBB data and synthetic/public video data) with synchronized visual and force measurements. Introduces modality-specific pseudo-anomaly generation techniques for images (synthetic arc-like artifacts) and force data (simulated irregularities) to augment training.

Result: Framework significantly outperforms baseline approaches, showing enhanced sensitivity to real arcing events even under domain shifts and limited availability of real arcing observations. Extensive experiments and ablation studies validate the approach.

Conclusion: The multimodal framework combining visual and force data with tailored augmentation techniques provides more accurate and robust arcing detection for rail systems, addressing challenges of data scarcity and noisy environments.

Abstract: The pantograph-catenary interface is essential for ensuring uninterrupted and reliable power delivery in electrified rail systems. However, electrical arcing at this interface poses serious risks, including accelerated wear of contact components, degraded system performance, and potential service disruptions. Detecting arcing events at the pantograph-catenary interface is challenging due to their transient nature, noisy operating environment, data scarcity, and the difficulty of distinguishing arcs from other similar transient phenomena. To address these challenges, we propose a novel multimodal framework that combines high-resolution image data with force measurements to more accurately and robustly detect arcing events. First, we construct two arcing detection datasets comprising synchronized visual and force measurements. One dataset is built from data provided by the Swiss Federal Railways (SBB), and the other is derived from publicly available videos of arcing events in different railway systems and synthetic force data that mimic the characteristics observed in the real dataset. Leveraging these datasets, we propose MultiDeepSAD, an extension of the DeepSAD algorithm for multiple modalities with a new loss formulation. Additionally, we introduce tailored pseudo-anomaly generation techniques specific to each data type, such as synthetic arc-like artifacts in images and simulated force irregularities, to augment training data and improve the discriminative ability of the model. Through extensive experiments and ablation studies, we demonstrate that our framework significantly outperforms baseline approaches, exhibiting enhanced sensitivity to real arcing events even under domain shifts and limited availability of real arcing observations.

Method: Proposes a teacher-student framework where the teacher generates probabilistic masks with per-pixel uncertainty. Unlabeled scans are ranked by image-level confidence and introduced progressively. The student uses a dual-loss objective to learn from high-confidence regions and unlearn low-confidence ones, with agreement-based refinement improving pseudo-label quality.

Result: On BraTS 2021, validation DSC increased from 0.393 (10% data) to 0.872 (100%), with largest gains in early stages. Teacher reached DSC of 0.922, and student surpassed teacher on tumor subregions (NCR/NET 0.797, Edema 0.980) and recovered Enhancing class (DSC 0.620) where teacher failed.

Conclusion: Confidence-driven curricula and selective unlearning provide robust segmentation under limited supervision and noisy pseudo-labels, demonstrating data efficiency and effectiveness in handling data heterogeneity.

Abstract: Accurate brain tumor segmentation from MRI is limited by expensive annotations and data heterogeneity across scanners and sites. We propose a semi-supervised teacher-student framework that combines an uncertainty-aware pseudo-labeling teacher with a progressive, confidence-based curriculum for the student. The teacher produces probabilistic masks and per-pixel uncertainty; unlabeled scans are ranked by image-level confidence and introduced in stages, while a dual-loss objective trains the student to learn from high-confidence regions and unlearn low-confidence ones. Agreement-based refinement further improves pseudo-label quality. On BraTS 2021, validation DSC increased from 0.393 (10% data) to 0.872 (100%), with the largest gains in early stages, demonstrating data efficiency. The teacher reached a validation DSC of 0.922, and the student surpassed the teacher on tumor subregions (e.g., NCR/NET 0.797 and Edema 0.980); notably, the student recovered the Enhancing class (DSC 0.620) where the teacher failed. These results show that confidence-driven curricula and selective unlearning provide robust segmentation under limited supervision and noisy pseudo-labels.

Method: Uses MLLMs to produce explicit target captions from user instructions, then employs lightweight adapters to inject multimodal conditional tokens into pretrained text-to-video diffusion models, enabling parameter-efficient reuse of generative priors.

Result: Achieves superior performance on complex compositional video editing (FiVE benchmark) and competitive/superior quality in text-to-video generation (VBench benchmark). Scaled to 14B parameters with high-quality training data.

Conclusion: The approach successfully bridges MLLM understanding with video diffusion generation, enabling effective handling of complex video editing tasks while maintaining strong video generation capabilities.

Abstract: We present Omni-Video 2, a scalable and computationally efficient model that connects pretrained multimodal large-language models (MLLMs) with video diffusion models for unified video generation and editing. Our key idea is to exploit the understanding and reasoning capabilities of MLLMs to produce explicit target captions to interpret user instructions. In this way, the rich contextual representations from the understanding model are directly used to guide the generative process, thereby improving performance on complex and compositional editing. Moreover, a lightweight adapter is developed to inject multimodal conditional tokens into pretrained text-to-video diffusion models, allowing maximum reuse of their powerful generative priors in a parameter-efficient manner. Benefiting from these designs, we scale up Omni-Video 2 to a 14B video diffusion model on meticulously curated training data with quality, supporting high quality text-to-video generation and various video editing tasks such as object removal, addition, background change, complex motion editing, \emph{etc.} We evaluate the performance of Omni-Video 2 on the FiVE benchmark for fine-grained video editing and the VBench benchmark for text-to-video generation. The results demonstrate its superior ability to follow complex compositional instructions in video editing, while also achieving competitive or superior quality in video generation tasks.

Method: Developed a unified foundation model integrating contrastive visual representation learning and vision-language alignment. Uses a contrastive encoder for modality-invariant representations and a CLIP-based text-informed decoder for semantically consistent synthesis, enabling any-to-all MRI synthesis via one unified model.

Result: Trained on 40,825 images from 13 institutions, achieved consistently high performance (average SSIM 0.90, PSNR 27) across 26 internal/external validation sites (15,748 images). Shows superior synthesis fidelity and robustness to noise and domain shifts. Unified representation also enhances downstream radiotherapy-relevant tasks like segmentation.

Conclusion: The work advances digital medicine solutions for NPC care by leveraging foundation models to bridge technical synthesis and clinical utility, providing a unified approach for MRI synthesis in radiotherapy planning.

Abstract: Magnetic resonance imaging (MRI) is essential for nasopharyngeal carcinoma (NPC) radiotherapy (RT), but practical constraints, such as patient discomfort, long scan times, and high costs often lead to incomplete modalities in clinical practice, compromising RT planning accuracy. Traditional MRI synthesis methods are modality-specific, limited in anatomical adaptability, and lack clinical interpretability-failing to meet NPC’s RT needs. Here, we developed a unified foundation model integrating contrastive visual representation learning and vision-language alignment (VLA) to enable any-to-all MRI synthesis. The model uses a contrastive encoder for modality-invariant representations and a CLIP-based text-informed decoder for semantically consistent synthesis, supporting any-to-all MRI synthesis via one unified foundation model. Trained on 40,825 images from 13 institutions, it achieves consistently high performance (average SSIM 0.90, PSNR 27) across 26 internal/external validation sites (15,748 images), with superior synthesis fidelity and robustness to noise and domain shifts. Meanwhile, its unified representation enhances downstream RT-relevant tasks (e.g., segmentation). This work advances digital medicine solutions for NPC care by leveraging foundation models to bridge technical synthesis and clinical utility.

Method: Introduces VideoVeritas framework with Joint Preference Alignment and Perception Pretext Reinforcement Learning (PPRL). Instead of directly optimizing for detection, uses general spatiotemporal grounding and self-supervised object counting in RL stage to enhance detection performance with simple perception pretext tasks. Also introduces MintVid dataset with 3K videos from 9 state-of-the-art generators and real-world factual error subset.

Result: Experimental results show existing methods bias toward either superficial reasoning or mechanical analysis, while VideoVeritas achieves more balanced performance across diverse benchmarks.

Conclusion: VideoVeritas effectively integrates fine-grained perception and fact-based reasoning for video deepfake detection, addressing limitations of current MLLMs through novel reinforcement learning approach and comprehensive evaluation dataset.

Abstract: The growing capability of video generation poses escalating security risks, making reliable detection increasingly essential. In this paper, we introduce VideoVeritas, a framework that integrates fine-grained perception and fact-based reasoning. We observe that while current multi-modal large language models (MLLMs) exhibit strong reasoning capacity, their granular perception ability remains limited. To mitigate this, we introduce Joint Preference Alignment and Perception Pretext Reinforcement Learning (PPRL). Specifically, rather than directly optimizing for detection task, we adopt general spatiotemporal grounding and self-supervised object counting in the RL stage, enhancing detection performance with simple perception pretext tasks. To facilitate robust evaluation, we further introduce MintVid, a light yet high-quality dataset containing 3K videos from 9 state-of-the-art generators, along with a real-world collected subset that has factual errors in content. Experimental results demonstrate that existing methods tend to bias towards either superficial reasoning or mechanical analysis, while VideoVeritas achieves more balanced performance across diverse benchmarks.

Method: FlattenGPT detects and reduces depth-wise redundancies by flattening two adjacent transformer blocks into one, compressing network depth while enabling more effective parameter redundancy detection and removal. It maintains the original transformer architecture.

Result: Outperforms existing pruning methods in zero-shot accuracies and WikiText-2 perplexity across various model types and sizes. On LLaMA-2/3 and Qwen-1.5 models, retains 90-96% of zero-shot performance with 20% compression ratio. Also better at accelerating LLM inference.

Conclusion: FlattenGPT provides an effective approach for transformer compression that balances efficiency and performance, making it promising for enhancing transformer efficiency while preserving learned knowledge.

Abstract: Recent works have indicated redundancy across transformer blocks, prompting the research of depth compression to prune less crucial blocks. However, current ways of entire-block pruning suffer from risks of discarding meaningful cues learned in those blocks, leading to substantial performance degradation. As another line of model compression, channel pruning can better preserve performance, while it cannot reduce model depth and is challenged by inconsistent pruning ratios for individual layers. To pursue better model compression and acceleration, this paper proposes \textbf{FlattenGPT}, a novel way to detect and reduce depth-wise redundancies. By flatting two adjacent blocks into one, it compresses the network depth, meanwhile enables more effective parameter redundancy detection and removal. FlattenGPT allows to preserve the knowledge learned in all blocks, and remains consistent with the original transformer architecture. Extensive experiments demonstrate that FlattenGPT enhances model efficiency with a decent trade-off to performance. It outperforms existing pruning methods in both zero-shot accuracies and WikiText-2 perplexity across various model types and parameter sizes. On LLaMA-2/3 and Qwen-1.5 models, FlattenGPT retains 90-96% of zero-shot performance with a compression ratio of 20%. It also outperforms other pruning methods in accelerating LLM inference, making it promising for enhancing the efficiency of transformers.

Method: Proposes Text-guided Video Frame Reduction (TiFRe) with two components: 1) Text-guided Frame Sampling (TFS) uses LLM-generated CLIP-style prompts to select semantically relevant frames, and 2) Frame Matching and Merging (FMM) integrates non-key frame information into selected key frames to preserve video semantics.

Result: Experiments show TiFRe effectively reduces computational costs while improving performance on video-language tasks compared to fixed frame rate approaches.

Conclusion: TiFRe provides an efficient solution for Video MLLMs by intelligently reducing input frames while preserving essential video information, addressing computational bottlenecks in video processing.

Abstract: With the rapid development of Large Language Models (LLMs), Video Multi-Modal Large Language Models (Video MLLMs) have achieved remarkable performance in video-language tasks such as video understanding and question answering. However, Video MLLMs face high computational costs, particularly in processing numerous video frames as input, which leads to significant attention computation overhead. A straightforward approach to reduce computational costs is to decrease the number of input video frames. However, simply selecting key frames at a fixed frame rate (FPS) often overlooks valuable information in non-key frames, resulting in notable performance degradation. To address this, we propose Text-guided Video Frame Reduction (TiFRe), a framework that reduces input frames while preserving essential video information. TiFRe uses a Text-guided Frame Sampling (TFS) strategy to select key frames based on user input, which is processed by an LLM to generate a CLIP-style prompt. Pre-trained CLIP encoders calculate the semantic similarity between the prompt and each frame, selecting the most relevant frames as key frames. To preserve video semantics, TiFRe employs a Frame Matching and Merging (FMM) mechanism, which integrates non-key frame information into the selected key frames, minimizing information loss. Experiments show that TiFRe effectively reduces computational costs while improving performance on video-language tasks.

Method: Analyze statistical properties of Rendering-Optimal References (RORs) from 3DGS optimization, then use learnability probes with predictors trained to reconstruct RORs from point clouds without rendering supervision, followed by variance decomposition analysis.

Result: Reveals stable patterns: mixture-structured scales and bimodal radiance across scenes. Shows fundamental density-stratification: dense regions have geometry-correlated parameters predictable from point clouds, while sparse regions show systematic failure due to visibility heterogeneity creating covariance-dominated coupling.

Conclusion: RORs have dual character: geometric primitives where point clouds suffice, and view synthesis primitives where multi-view constraints are essential. Provides density-aware strategies for training robustness and architectural implications for balancing feed-forward prediction with rendering-based refinement.

Abstract: We investigate what structure emerges in 3D Gaussian Splatting (3DGS) solutions from standard multi-view optimization. We term these Rendering-Optimal References (RORs) and analyze their statistical properties, revealing stable patterns: mixture-structured scales and bimodal radiance across diverse scenes. To understand what determines these parameters, we apply learnability probes by training predictors to reconstruct RORs from point clouds without rendering supervision. Our analysis uncovers fundamental density-stratification. Dense regions exhibit geometry-correlated parameters amenable to render-free prediction, while sparse regions show systematic failure across architectures. We formalize this through variance decomposition, demonstrating that visibility heterogeneity creates covariance-dominated coupling between geometric and appearance parameters in sparse regions. This reveals the dual character of RORs: geometric primitives where point clouds suffice, and view synthesis primitives where multi-view constraints are essential. We provide density-aware strategies that improve training robustness and discuss architectural implications for systems that adaptively balance feed-forward prediction and rendering-based refinement.

Method: Introduces a 3D Gaussian flow field representation that models plant growth as time-varying derivatives over Gaussian parameters (position, scale, orientation, color, opacity). Initializes Gaussians by reconstructing the mature plant and learning reverse growth to simulate developmental history in reverse.

Result: Achieves superior image quality and geometric accuracy compared to prior methods on multi-view timelapse datasets of plant growth, providing a new approach for appearance modeling of growing 3D structures.

Conclusion: The 3D Gaussian flow field representation effectively models plant growth dynamics, handling the unique challenges of new geometry generation over time while maintaining tracking of Gaussian primitives through continuous-time nonlinear growth.

Abstract: Modeling the time-varying 3D appearance of plants during their growth poses unique challenges: unlike many dynamic scenes, plants generate new geometry over time as they expand, branch, and differentiate. Recent motion modeling techniques are ill-suited to this problem setting. For example, deformation fields cannot introduce new geometry, and 4D Gaussian splatting constrains motion to a linear trajectory in space and time and cannot track the same set of Gaussians over time. Here, we introduce a 3D Gaussian flow field representation that models plant growth as a time-varying derivative over Gaussian parameters – position, scale, orientation, color, and opacity – enabling nonlinear and continuous-time growth dynamics. To initialize a sufficient set of Gaussian primitives, we reconstruct the mature plant and learn a process of reverse growth, effectively simulating the plant’s developmental history in reverse. Our approach achieves superior image quality and geometric accuracy compared to prior methods on multi-view timelapse datasets of plant growth, providing a new approach for appearance modeling of growing 3D structures.

Method: Introduces a novel joint representation of dense 3D point maps and 3D scene flows in a shared coordinate system, and a novel 4D VAE to learn this representation. Uses a new data normalization and VAE training strategy that better transfers diffusion priors instead of forcing alignment with RGB VAE latents.

Result: Achieves state-of-the-art performance with 38.64% improvement in geometry reconstruction and 25.0% improvement in motion reconstruction across multiple datasets, all without post-optimization.

Conclusion: MotionCrafter demonstrates that forcing alignment between 3D geometry/motion representations and RGB VAE latents is unnecessary and suboptimal, and that the proposed joint representation with proper normalization and training strategy significantly improves 4D reconstruction quality.

Abstract: We introduce MotionCrafter, a video diffusion-based framework that jointly reconstructs 4D geometry and estimates dense motion from a monocular video. The core of our method is a novel joint representation of dense 3D point maps and 3D scene flows in a shared coordinate system, and a novel 4D VAE to effectively learn this representation. Unlike prior work that forces the 3D value and latents to align strictly with RGB VAE latents-despite their fundamentally different distributions-we show that such alignment is unnecessary and leads to suboptimal performance. Instead, we introduce a new data normalization and VAE training strategy that better transfers diffusion priors and greatly improves reconstruction quality. Extensive experiments across multiple datasets demonstrate that MotionCrafter achieves state-of-the-art performance in both geometry reconstruction and dense scene flow estimation, delivering 38.64% and 25.0% improvements in geometry and motion reconstruction, respectively, all without any post-optimization. Project page: https://ruijiezhu94.github.io/MotionCrafter_Page

Method: Enhanced Waymo-3DSkelMo dataset with 3D vehicle bounding boxes, introduced sampling scheme for varying interaction complexities, adapted TBIFormer architecture with vehicle encoder and pedestrian-vehicle interaction cross-attention module.

Result: Substantial improvements in forecasting accuracy, validated different approaches for modeling pedestrian-vehicle interactions, demonstrated importance of vehicle-aware 3D pose prediction.

Conclusion: Vehicle-conditioned 3D pedestrian pose forecasting significantly improves prediction accuracy for autonomous driving by explicitly modeling pedestrian-vehicle interactions.

Abstract: Accurately predicting pedestrian motion is crucial for safe and reliable autonomous driving in complex urban environments. In this work, we present a 3D vehicle-conditioned pedestrian pose forecasting framework that explicitly incorporates surrounding vehicle information. To support this, we enhance the Waymo-3DSkelMo dataset with aligned 3D vehicle bounding boxes, enabling realistic modeling of multi-agent pedestrian-vehicle interactions. We introduce a sampling scheme to categorize scenes by pedestrian and vehicle count, facilitating training across varying interaction complexities. Our proposed network adapts the TBIFormer architecture with a dedicated vehicle encoder and pedestrian-vehicle interaction cross-attention module to fuse pedestrian and vehicle features, allowing predictions to be conditioned on both historical pedestrian motion and surrounding vehicles. Extensive experiments demonstrate substantial improvements in forecasting accuracy and validate different approaches for modeling pedestrian-vehicle interactions, highlighting the importance of vehicle-aware 3D pose prediction for autonomous driving. Code is available at: https://github.com/GuangxunZhu/VehCondPose3D

Method: Introduces WorldArena benchmark with three evaluation dimensions: 1) video perception quality (16 metrics across 6 sub-dimensions), 2) embodied task functionality (evaluating models as data engines, policy evaluators, and action planners with human evaluation), and 3) EWMScore - a holistic metric integrating multi-dimensional performance into a single interpretable index.

Result: Extensive experiments on 14 representative models reveal a significant perception-functionality gap, showing that high visual quality does not necessarily translate into strong embodied task capability.

Conclusion: WorldArena provides a comprehensive framework for tracking progress toward truly functional world models in embodied AI, addressing the current fragmented evaluation landscape and emphasizing the importance of functional utility beyond mere perceptual quality.

Abstract: While world models have emerged as a cornerstone of embodied intelligence by enabling agents to reason about environmental dynamics through action-conditioned prediction, their evaluation remains fragmented. Current evaluation of embodied world models has largely focused on perceptual fidelity (e.g., video generation quality), overlooking the functional utility of these models in downstream decision-making tasks. In this work, we introduce WorldArena, a unified benchmark designed to systematically evaluate embodied world models across both perceptual and functional dimensions. WorldArena assesses models through three dimensions: video perception quality, measured with 16 metrics across six sub-dimensions; embodied task functionality, which evaluates world models as data engines, policy evaluators, and action planners integrating with subjective human evaluation. Furthermore, we propose EWMScore, a holistic metric integrating multi-dimensional performance into a single interpretable index. Through extensive experiments on 14 representative models, we reveal a significant perception-functionality gap, showing that high visual quality does not necessarily translate into strong embodied task capability. WorldArena benchmark with the public leaderboard is released at https://worldarena.ai, providing a framework for tracking progress toward truly functional world models in embodied AI.

Method: Uses rock climbing as case study, leveraging auxiliary freely-available web data (competition videos, coaching manuals) from target domain plus existing sports feedback from source domain. Proposes two new evaluation metrics: specificity and actionability.

Result: Enables more meaningful and practical generation of sports feedback under limited annotations by improving performance on target domain and providing better evaluation.

Conclusion: The approach addresses limitations in sports feedback generation for video-LLMs by using auxiliary domain data and introducing domain-specific evaluation metrics.

Abstract: While there is rapid progress in video-LLMs with advanced reasoning capabilities, prior work shows that these models struggle on the challenging task of sports feedback generation and require expensive and difficult-to-collect finetuning feedback data for each sport. This limitation is evident from the poor generalization to sports unseen during finetuning. Furthermore, traditional text generation evaluation metrics (e.g., BLEU-4, METEOR, ROUGE-L, BERTScore), originally developed for machine translation and summarization, fail to capture the unique aspects of sports feedback quality. To address the first problem, using rock climbing as our case study, we propose using auxiliary freely-available web data from the target domain, such as competition videos and coaching manuals, in addition to existing sports feedback from a disjoint, source domain to improve sports feedback generation performance on the target domain. To improve evaluation, we propose two evaluation metrics: (1) specificity and (2) actionability. Together, our approach enables more meaningful and practical generation of sports feedback under limited annotations.

Method: ArcFlow parameterizes velocity fields as mixtures of continuous momentum processes to capture velocity evolution, enabling analytical integration of non-linear trajectories. Uses lightweight adapters for trajectory distillation on pre-trained models.

Result: Achieves 40x speedup with 2 NFEs over original multi-step teachers (Qwen-Image-20B and FLUX.1-dev) while fine-tuning less than 5% of parameters. Maintains generative quality and diversity without significant degradation.

Conclusion: ArcFlow effectively distills diffusion models into few-step generators using non-linear flow trajectories, providing substantial speed improvements while preserving quality through analytical integration and lightweight adaptation.

Abstract: Diffusion models have achieved remarkable generation quality, but they suffer from significant inference cost due to their reliance on multiple sequential denoising steps, motivating recent efforts to distill this inference process into a few-step regime. However, existing distillation methods typically approximate the teacher trajectory by using linear shortcuts, which makes it difficult to match its constantly changing tangent directions as velocities evolve across timesteps, thereby leading to quality degradation. To address this limitation, we propose ArcFlow, a few-step distillation framework that explicitly employs non-linear flow trajectories to approximate pre-trained teacher trajectories. Concretely, ArcFlow parameterizes the velocity field underlying the inference trajectory as a mixture of continuous momentum processes. This enables ArcFlow to capture velocity evolution and extrapolate coherent velocities to form a continuous non-linear trajectory within each denoising step. Importantly, this parameterization admits an analytical integration of this non-linear trajectory, which circumvents numerical discretization errors and results in high-precision approximation of the teacher trajectory. To train this parameterization into a few-step generator, we implement ArcFlow via trajectory distillation on pre-trained teacher models using lightweight adapters. This strategy ensures fast, stable convergence while preserving generative diversity and quality. Built on large-scale models (Qwen-Image-20B and FLUX.1-dev), ArcFlow only fine-tunes on less than 5% of original parameters and achieves a 40x speedup with 2 NFEs over the original multi-step teachers without significant quality degradation. Experiments on benchmarks show the effectiveness of ArcFlow both qualitatively and quantitatively.

Method: Frames floorplan reconstruction as sequence-to-sequence task where elements (rooms, windows, doors) are represented as labeled polygon sequences. Uses autoregressive decoder that predicts next corner conditioned on image features and previously generated corners, guided by learnable anchors that direct attention to informative image regions.

Result: Achieves state-of-the-art performance on standard benchmarks (Structure3D, CubiCasa5K, Raster2Graph) and demonstrates strong generalization to more challenging datasets like WAFFLE with diverse room structures and complex geometric variations.

Conclusion: The autoregressive approach offers flexibility in output format and efficiently handles complex floorplans with numerous rooms and diverse polygon structures, advancing floorplan reconstruction capabilities.

Abstract: Reconstructing a structured vector-graphics representation from a rasterized floorplan image is typically an important prerequisite for computational tasks involving floorplans such as automated understanding or CAD workflows. However, existing techniques struggle in faithfully generating the structure and semantics conveyed by complex floorplans that depict large indoor spaces with many rooms and a varying numbers of polygon corners. To this end, we propose Raster2Seq, framing floorplan reconstruction as a sequence-to-sequence task in which floorplan elements–such as rooms, windows, and doors–are represented as labeled polygon sequences that jointly encode geometry and semantics. Our approach introduces an autoregressive decoder that learns to predict the next corner conditioned on image features and previously generated corners using guidance from learnable anchors. These anchors represent spatial coordinates in image space, hence allowing for effectively directing the attention mechanism to focus on informative image regions. By embracing the autoregressive mechanism, our method offers flexibility in the output format, enabling for efficiently handling complex floorplans with numerous rooms and diverse polygon structures. Our method achieves state-of-the-art performance on standard benchmarks such as Structure3D, CubiCasa5K, and Raster2Graph, while also demonstrating strong generalization to more challenging datasets like WAFFLE, which contain diverse room structures and complex geometric variations.

Method: Three core innovations: 1) Clip-level rollout strategy generating multiple samples at single target clips for efficiency and fine-grained rewards, 2) Complementary reward functions for interaction-following accuracy and visual quality to prevent reward hacking, 3) Efficient RL algorithm with negative-aware fine-tuning and optimizations.

Result: Evaluations on state-of-the-art world model WorldPlay show significant improvements in interaction accuracy and visual fidelity across various scenarios.

Conclusion: WorldCompass effectively enhances world model exploration capabilities through RL post-training, making video-based world models more accurate and consistent in interactive scenarios.

Abstract: This work presents WorldCompass, a novel Reinforcement Learning (RL) post-training framework for the long-horizon, interactive video-based world models, enabling them to explore the world more accurately and consistently based on interaction signals. To effectively “steer” the world model’s exploration, we introduce three core innovations tailored to the autoregressive video generation paradigm: 1) Clip-level rollout Strategy: We generate and evaluate multiple samples at a single target clip, which significantly boosts rollout efficiency and provides fine-grained reward signals. 2) Complementary Reward Functions: We design reward functions for both interaction-following accuracy and visual quality, which provide direct supervision and effectively suppress reward-hacking behaviors. 3) Efficient RL Algorithm: We employ the negative-aware fine-tuning strategy coupled with various efficiency optimizations to efficiently and effectively enhance model capacity. Evaluations on the SoTA open-source world model, WorldPlay, demonstrate that WorldCompass significantly improves interaction accuracy and visual fidelity across various scenarios.

Method: Proposes masked Bit AutoRegressive modeling (BAR) - a scalable framework supporting arbitrary codebook sizes. Uses an autoregressive transformer with masked bit modeling head to predict discrete tokens by progressively generating their constituent bits.

Result: Achieves state-of-the-art gFID of 0.99 on ImageNet-256, outperforming leading continuous and discrete methods. Reduces sampling costs and converges faster than prior continuous approaches.

Conclusion: Discrete tokenizers can match or surpass continuous methods when properly scaled, challenging the dominance of continuous pipelines in visual generation. BAR provides a practical solution for scaling discrete representations.

Abstract: This paper challenges the dominance of continuous pipelines in visual generation. We systematically investigate the performance gap between discrete and continuous methods. Contrary to the belief that discrete tokenizers are intrinsically inferior, we demonstrate that the disparity arises primarily from the total number of bits allocated in the latent space (i.e., the compression ratio). We show that scaling up the codebook size effectively bridges this gap, allowing discrete tokenizers to match or surpass their continuous counterparts. However, existing discrete generation methods struggle to capitalize on this insight, suffering from performance degradation or prohibitive training costs with scaled codebook. To address this, we propose masked Bit AutoRegressive modeling (BAR), a scalable framework that supports arbitrary codebook sizes. By equipping an autoregressive transformer with a masked bit modeling head, BAR predicts discrete tokens through progressively generating their constituent bits. BAR achieves a new state-of-the-art gFID of 0.99 on ImageNet-256, outperforming leading methods across both continuous and discrete paradigms, while significantly reducing sampling costs and converging faster than prior continuous approaches. Project page is available at https://bar-gen.github.io/

Method: Based on T5-3B pre-trained checkpoint, adds masked language modeling (MLM) and relative position prediction (RPP) pre-training tasks to align object features and scene text. Encoder handles fusion of multiple modalities (question text, object text labels, scene text labels, object visual features, scene visual features), decoder generates text sequence with cross-entropy loss. Uses large-scale scene text dataset for pre-training, then fine-tunes on TextVQA dataset.

Result: The paper presents a method for TextVQA using T5-3B with additional pre-training tasks, but specific quantitative results are not provided in the abstract.

Conclusion: The approach leverages T5’s generative capabilities for TextVQA with enhanced pre-training to better align visual and text modalities for scene text understanding.

Abstract: TextVQA requires models to read and reason about text in images to answer questions about them. Specifically, models need to incorporate a new modality of text present in the images and reason over it to answer TextVQA questions. In this challenge, we use generative model T5 for TextVQA task. Based on pre-trained checkpoint T5-3B from HuggingFace repository, two other pre-training tasks including masked language modeling(MLM) and relative position prediction(RPP) are designed to better align object feature and scene text. In the stage of pre-training, encoder is dedicate to handle the fusion among multiple modalities: question text, object text labels, scene text labels, object visual features, scene visual features. After that decoder generates the text sequence step-by-step, cross entropy loss is required by default. We use a large-scale scene text dataset in pre-training and then fine-tune the T5-3B with the TextVQA dataset only.

Method: Uses a dual-unimodal transformer encoder architecture for motion-text retrieval. Introduces DropTriple Loss, which discards false negative samples from the negative sample set and focuses on mining genuinely hard negative samples to handle semantic conflicts from overlapping atomic actions in different motion sequences.

Result: Achieves 62.9% recall for motion retrieval and 71.5% for text retrieval (R@10) on the HumanML3D dataset. Also evaluated on KIT Motion-Language dataset.

Conclusion: The proposed dual-unimodal transformer encoder with DropTriple Loss effectively addresses human motion-text cross-modal retrieval, achieving strong performance on benchmark datasets and filling an important research gap.

Abstract: Cross-modal retrieval of image-text and video-text is a prominent research area in computer vision and natural language processing. However, there has been insufficient attention given to cross-modal retrieval between human motion and text, despite its wide-ranging applicability. To address this gap, we utilize a concise yet effective dual-unimodal transformer encoder for tackling this task. Recognizing that overlapping atomic actions in different human motion sequences can lead to semantic conflicts between samples, we explore a novel triplet loss function called DropTriple Loss. This loss function discards false negative samples from the negative sample set and focuses on mining remaining genuinely hard negative samples for triplet training, thereby reducing violations they cause. We evaluate our model and approach on the HumanML3D and KIT Motion-Language datasets. On the latest HumanML3D dataset, we achieve a recall of 62.9% for motion retrieval and 71.5% for text retrieval (both based on R@10). The source code for our approach is publicly available at https://github.com/eanson023/rehamot.

Method: Extends volumetric implicit models (e.g., neural radiance fields) to be relightable using IBL. The model handles complex materials like translucency and glossy reflections from detailed surface geometry.

Result: Achieves fast rendering (0.72s on NVIDIA 3090, 0.30s on A100 at 800×800 resolution) without engineering optimization. Produces realistic and compelling results for materials with complex light transport effects.

Conclusion: Presents a fast neural 3D reconstruction and relighting model that successfully extends volumetric implicit models to handle IBL relighting for complex materials with efficient rendering performance.

Abstract: Image-based lighting (IBL) is a widely used technique that renders objects using a high dynamic range image or environment map. However, aggregating the irradiance at the object’s surface is computationally expensive, in particular for non-opaque, translucent materials that require volumetric rendering techniques. In this paper we present a fast neural 3D reconstruction and relighting model that extends volumetric implicit models such as neural radiance fields to be relightable using IBL. It is general enough to handle materials that exhibit complex light transport effects, such as translucency and glossy reflections from detailed surface geometry, producing realistic and compelling results. Rendering can be within a second at 800$\times$800 resolution (0.72s on an NVIDIA 3090 GPU and 0.30s on an A100 GPU) without engineering optimization. Our code and dataset are available at https://zhusz.github.io/TRHM-Webpage/.

Method: Proposes LBL loss using logarithmic barrier function to assign large gradients to margin samples for compact hypersphere. Then introduces LBLSig by incorporating unilateral relaxation Sigmoid function to address LBL’s instability when samples lie on boundary (infinity loss). LBLSig fuses MSE and CE loss characteristics.

Result: Experimental validation shows effectiveness of both LBL and LBLSig compared to popular OCC algorithms on different network structures. LBLSig provides smoother optimization than LBL.

Conclusion: The proposed logarithmic barrier-based losses (LBL and LBLSig) provide effective OCC loss functions for deep learning with improved optimization stability and compact decision boundaries.

Abstract: One-class classification (OCC) aims to train a classifier only with the target class data and attracts great attention for its strong applicability in real-world application. Despite a lot of advances have been made in OCC, it still lacks the effective OCC loss functions for deep learning. In this paper, a novel logarithmic barrier function based OCC loss (LBL) that assigns large gradients to the margin samples and thus derives more compact hypersphere, is first proposed by approximating the OCC objective smoothly. But the optimization of LBL may be instability especially when samples lie on the boundary leading to the infinity loss. To address this issue, then, a unilateral relaxation Sigmoid function is introduced into LBL and a novel OCC loss named LBLSig is proposed. The LBLSig can be seen as the fusion of the mean square error (MSE) and the cross entropy (CE) and the optimization of LBLSig is smoother owing to the unilateral relaxation Sigmoid function. The effectiveness of the proposed LBL and LBLSig is experimentally demonstrated in comparisons with several popular OCC algorithms on different network structures. The source code can be found at https://github.com/ML-HDU/LBL_LBLSig.

Method: Identifies CLIP DeltaSpace where visual feature differences align with textual feature differences. Maps CLIP visual feature differences to generative model latent directions during training, and predicts latent directions from textual feature differences during inference.

Result: Extensive experiments show DeltaEdit works effectively with both GAN and diffusion models, achieving flexible text-guided image editing with text-free training and zero-shot inference capabilities.

Conclusion: DeltaEdit provides a novel solution for text-guided image editing that avoids expensive data collection and per-prompt optimization, demonstrating versatility across different generative models.

Abstract: Text-guided image editing faces significant challenges when considering training and inference flexibility. Much literature collects large amounts of annotated image-text pairs to train text-conditioned generative models from scratch, which is expensive and not efficient. After that, some approaches that leverage pre-trained vision-language models have been proposed to avoid data collection, but they are limited by either per text-prompt optimization or inference-time hyper-parameters tuning. To address these issues, we investigate and identify a specific space, referred to as CLIP DeltaSpace, where the CLIP visual feature difference of two images is semantically aligned with the CLIP textual feature difference of their corresponding text descriptions. Based on DeltaSpace, we propose a novel framework called DeltaEdit, which maps the CLIP visual feature differences to the latent space directions of a generative model during the training phase, and predicts the latent space directions from the CLIP textual feature differences during the inference phase. And this design endows DeltaEdit with two advantages: (1) text-free training; (2) generalization to various text prompts for zero-shot inference. Extensive experiments validate the effectiveness and versatility of DeltaEdit with different generative models, including both the GAN model and the diffusion model, in achieving flexible text-guided image editing. Code is available at https://github.com/Yueming6568/DeltaEdit.

Method: Segments time interval into multiple stages using custom multi-decoder U-net architecture that blends time-dependent models with universally shared encoder, employs timestep clustering algorithm for stage division, and uses both universal and customized parameters.

Result: Significant training and sampling efficiency improvements on three state-of-the-art diffusion models including large-scale latent diffusion models, with ablation studies showing impact of timestep clustering and multi-decoder architecture.

Conclusion: The multi-stage framework efficiently distributes computational resources, mitigates inter-stage interference, and substantially improves training efficiency for diffusion models.

Abstract: Diffusion models, emerging as powerful deep generative tools, excel in various applications. They operate through a two-steps process: introducing noise into training samples and then employing a model to convert random noise into new samples (e.g., images). However, their remarkable generative performance is hindered by slow training and sampling. This is due to the necessity of tracking extensive forward and reverse diffusion trajectories, and employing a large model with numerous parameters across multiple timesteps (i.e., noise levels). To tackle these challenges, we present a multi-stage framework inspired by our empirical findings. These observations indicate the advantages of employing distinct parameters tailored to each timestep while retaining universal parameters shared across all time steps. Our approach involves segmenting the time interval into multiple stages where we employ custom multi-decoder U-net architecture that blends time-dependent models with a universally shared encoder. Our framework enables the efficient distribution of computational resources and mitigates inter-stage interference, which substantially improves training efficiency. Extensive numerical experiments affirm the effectiveness of our framework, showcasing significant training and sampling efficiency enhancements on three state-of-the-art diffusion models, including large-scale latent diffusion models. Furthermore, our ablation studies illustrate the impact of two important components in our framework: (i) a novel timestep clustering algorithm for stage division, and (ii) an innovative multi-decoder U-net architecture, seamlessly integrating universal and customized hyperparameters.

Method: Developed Cognitive Edge Device (CED) computing platform and created two publicly available underwater datasets annotated for crayfish and plastic debris. Trained and evaluated four YOLO variants for object detection tasks.

Result: YOLOv5s achieved the highest detection accuracy with mAP@0.5 of 0.90 and best precision among the tested models.

Conclusion: The proposed system enables effective automated monitoring of invasive species and pollution in aquatic ecosystems, with YOLOv5s showing superior performance for underwater object detection tasks.

Abstract: Invasive signal crayfish have a detrimental impact on ecosystems. They spread the fungal-type crayfish plague disease (Aphanomyces astaci) that is lethal to the native white clawed crayfish, the only native crayfish species in Britain. Invasive signal crayfish extensively burrow, causing habitat destruction, erosion of river banks and adverse changes in water quality, while also competing with native species for resources leading to declines in native populations. Moreover, pollution exacerbates the vulnerability of White-clawed crayfish, with their populations declining by over 90%. To safeguard aquatic ecosystems, it is imperative to address the challenges posed by invasive species and pollution in aquatic ecosystem’s. This article introduces the Cognitive Edge Device (CED) computing platform for the detection of crayfish and plastic. It also presents two publicly available underwater datasets, annotated with sequences of crayfish and aquatic plastic debris. Four You Only Look Once (YOLO) variants were trained and evaluated for crayfish and plastic object detection. YOLOv5s achieved the highest detection accuracy, with an mAP@0.5 of 0.90, and achieved the best precision

Method: Multicamera setup with YOLOv8-based detection module integrated into experimental stand. Evaluation uses timestamp-level comparisons between predicted and actual execution stages with three metrics: Intersection over Union (IoU), Mean Absolute Scaled Error (MASE), and Residual Distribution histograms, aggregated into unified efficiency index E_total.

Result: Validated on 120 assemblies at different speeds, showing high segmentation accuracy and identifying stage-specific timing deviations. System demonstrated robustness and evaluation framework applicable for benchmarking similar industrial solutions.

Conclusion: Proposed AI control system effectively tracks manual assembly with robust evaluation methodology. Framework enables reproducible assessment of similar industrial automation solutions.

Abstract: Manual operations remain essential in industrial production because of their flexibility and low implementation cost. However, ensuring their quality and monitoring execution in real time remains a challenge, especially under conditions of high variability and human-induced errors. In this paper, we present an AI-based control system for tracking manual assembly and propose a novel methodology to evaluate its overall efficiency. The developed system includes a multicamera setup and a YOLOv8-based detection module integrated into an experimental stand designed to replicate real production scenarios. The evaluation methodology relies on timestamp-level comparisons between predicted and actual execution stages, using three key metrics: Intersection over Union (IoU), Mean Absolute Scaled Error (MASE), Residual Distribution histograms. These metrics are aggregated into a unified efficiency index E_total for reproducible system assessment. The proposed approach was validated on a dataset of 120 assemblies performed at different speeds, demonstrating high segmentation accuracy and identifying stage-specific timing deviations. The results confirm the robustness of the control system and the applicability of the evaluation framework to benchmark similar solutions in industrial settings.

Method: Proposes TSJNet with: 1) multi-dimensional feature extraction module with dual parallel branches for multi-scale and salient features, 2) data-agnostic spatial attention module in decoder for dynamic attention calibration across domains, 3) joint optimization using fusion loss with semantic losses.

Result: Outstanding performance on five public datasets (MSRS, M³FD, RoadScene, LLVIP, TNO) and new UMS dataset. Improves mAP@0.5 by 7.97% and mIoU by 10.88% compared to state-of-the-art methods.

Conclusion: TSJNet effectively addresses multimodal fusion challenges for visual tasks, demonstrating superior performance and generalization through task-semantic guided fusion and data-agnostic attention mechanisms.

Abstract: This study aims to address the problem of incomplete information in unimodal images for semantic segmentation and object detection tasks. Existing multimodal fusion methods suffer from limited capability in discriminative modeling of multi-scale semantic structures and salient target regions, which further restricts the effective fusion of task-related semantic details and target information across modalities. To tackle these challenges, this paper proposes a novel fusion network termed TSJNet, which leverages the semantic information output by high-level tasks in a joint manner to guide the fusion process. Specifically, we design a multi-dimensional feature extraction module with dual parallel branches to capture multi-scale and salient features. Meanwhile, a data-agnostic spatial attention module embedded in the decoder dynamically calibrates attention allocation across different data domains, significantly enhancing the model’s generalization ability. To optimize both fusion and advanced visual tasks, we balance performance by combining fusion loss with semantic losses. Additionally, we have developed a multimodal unmanned aerial vehicle (UAV) dataset covering multiple scenarios (UMS). Extensive experiments demonstrate that TSJNet achieves outstanding performance on five public datasets (MSRS, M\textsuperscript{3}FD, RoadScene, LLVIP, and TNO) and our UMS dataset. The generated fusion results exhibit favorable visual effects, and compared to state-of-the-art methods, the mean average precision (mAP@0.5) and mean intersection over union (mIoU) for object detection and segmentation, respectively, improve by 7.97% and 10.88%.The code and the dataset has been publicly released at https://github.com/XylonXu01/TSJNet.

Method: Proposes HomView-MOT framework with: 1) Fast Homography Estimation (FHE) for rapid computation of homography matrices between frames, 2) View-Centric ID Learning (VCIL) using multi-view homography to learn cross-view ID features, and 3) Homographic Matching Filter (HMF) that maps object bounding boxes to common view plane for physical IoU association.

Result: Achieves state-of-the-art performance on prominent UAV MOT datasets VisDrone and UAVDT.

Conclusion: The framework successfully addresses MOT challenges in moving UAV environments by leveraging view homography for better object association and tracking in dynamic aerial scenarios.

Abstract: In this paper, we address the challenge of Multi-Object Tracking (MOT) in moving Unmanned Aerial Vehicle (UAV) scenarios, where irregular flight trajectories, such as hovering, turning left/right, and moving up/down, lead to significantly greater complexity compared to fixed-camera MOT. Specifically, changes in the scene background not only render traditional frame-to-frame object IoU association methods ineffective but also introduce significant view shifts in the objects, which complicates tracking. To overcome these issues, we propose a novel HomView-MOT framework, which for the first time, harnesses the view homography inherent in changing scenes to solve MOT challenges in moving environments, incorporating homographic matching and view-centric concepts. We introduce a Fast Homography Estimation (FHE) algorithm for rapid computation of homography matrices between video frames, enabling object View-Centric ID Learning (VCIL) and leveraging multi-view homography to learn cross-view ID features. Concurrently, our Homographic Matching Filter (HMF) maps object bounding boxes from different frames onto a common view plane for a more realistic physical IoU association. Extensive experiments have proven that these innovations allow HomView-MOT to achieve state-of-the-art performance on prominent UAV MOT datasets VisDrone and UAVDT.

Method: First explores LVLM internal states to discover they are high-specificity per-token hallucination indicators and that different LVLMs encode universal hallucination patterns. Then proposes TruthPrInt which learns truthful directions of LVLM decoding and applies truthful-guided inference-time intervention during decoding, with cross-LVLM and cross-data transferability via hallucination latent subspace alignment.

Result: TruthPrInt significantly outperforms state-of-the-art methods in extensive experimental settings including in-domain and out-of-domain scenarios over popular LVLMs and OH benchmarks.

Conclusion: LVLM internal states serve as effective per-token hallucination indicators, and TruthPrInt’s truthful-guided intervention successfully mitigates object hallucination with strong transferability across models and datasets.

Abstract: Object Hallucination (OH) has been acknowledged as one of the major trustworthy challenges in Large Vision-Language Models (LVLMs). Recent advancements in Large Language Models (LLMs) indicate that internal states, such as hidden states, encode the “overall truthfulness” of generated responses. However, it remains under-explored how internal states in LVLMs function and whether they could serve as “per-token” hallucination indicators, which is essential for mitigating OH. In this paper, we first conduct an in-depth exploration of LVLM internal states with OH issues and discover that (1) LVLM internal states are high-specificity per-token indicators of hallucination behaviors. Moreover, (2) different LVLMs encode universal patterns of hallucinations in common latent subspaces, indicating that there exist “generic truthful directions” shared by various LVLMs. Based on these discoveries, we propose Truthful-Guided Pre-Intervention (TruthPrInt) that first learns the truthful direction of LVLM decoding and then applies truthful-guided inference-time intervention during LVLM decoding. We further propose TruthPrInt to enhance both cross-LVLM and cross-data hallucination detection transferability by constructing and aligning hallucination latent subspaces. We evaluate TruthPrInt in extensive experimental settings, including in-domain and out-of-domain scenarios, over popular LVLMs and OH benchmarks. Experimental results indicate that TruthPrInt significantly outperforms state-of-the-art methods. Codes will be available at https://github.com/jinhaoduan/TruthPrInt.

Method: Introduces CT-RATE dataset (25,692 3D chest CT scans with reports), develops CT-CLIP contrastive language-image pretraining framework, and creates CT-CHAT vision-language chat model by combining CT-CLIP’s encoder with LLM.

Result: CT-CLIP outperforms state-of-the-art supervised models in multi-abnormality detection and case retrieval; CT-CHAT demonstrates specialized 3D medical imaging capabilities.

Conclusion: The open-source release of CT-RATE, CT-CLIP, and CT-CHAT addresses critical challenges in 3D medical imaging and lays groundwork for future medical AI innovations.

Abstract: Advancements in medical imaging AI, particularly in 3D imaging, have been limited due to the scarcity of comprehensive datasets. We introduce CT-RATE, a public dataset that pairs 3D medical images with corresponding textual reports. CT-RATE comprises 25,692 non-contrast 3D chest CT scans from 21,304 unique patients. Each scan is accompanied by its corresponding radiology report. Leveraging CT-RATE, we develop CT-CLIP, a CT-focused contrastive language-image pretraining framework designed for broad applications without the need for task-specific training. We demonstrate how CT-CLIP can be used in multi-abnormality detection and case retrieval, and outperforms state-of-the-art fully supervised models across all key metrics. By combining CT-CLIP’s vision encoder with a pretrained large language model, we create CT-CHAT, a vision-language foundational chat model for 3D chest CT volumes. Finetuned on over 2.7 million question-answer pairs derived from the CT-RATE dataset, CT-CHAT underscores the necessity for specialized methods in 3D medical imaging. Collectively, the open-source release of CT-RATE, CT-CLIP, and CT-CHAT not only addresses critical challenges in 3D medical imaging but also lays the groundwork for future innovations in medical AI and improved patient care.

Method: Derived Riemannian AdamW optimizer for hyperbolic generalization, introduced fine-tunable hyperbolic scaling to constrain embeddings and reduce approximation errors, and developed curvature-aware learning schema for Riemannian optimizers to combine curvature and hyperbolic parameter learning.

Result: Demonstrated consistent performance improvements across Computer Vision, EEG classification, and hierarchical metric learning tasks while greatly reducing runtime compared to previous hyperbolic learning approaches.

Conclusion: The proposed Riemannian AdamW with fine-tunable hyperbolic scaling addresses key limitations of hyperbolic deep learning, providing more stable, efficient, and generalizable models for hierarchical representation learning tasks.

Abstract: Hyperbolic deep learning has become a growing research direction in computer vision due to the unique properties afforded by the alternate embedding space. The negative curvature and exponentially growing distance metric provide a natural framework for capturing hierarchical relationships between datapoints and allowing for finer separability between their embeddings. However, current hyperbolic learning approaches are still prone to overfitting, computationally expensive, and prone to instability, especially when attempting to learn the manifold curvature to adapt to tasks and different datasets. To address these issues, our paper presents a derivation for Riemannian AdamW that helps increase hyperbolic generalization ability. For improved stability, we introduce a novel fine-tunable hyperbolic scaling approach to constrain hyperbolic embeddings and reduce approximation errors. Using this along with our curvature-aware learning schema for Riemannian Optimizers enables the combination of curvature and non-trivialized hyperbolic parameter learning. Our approach demonstrates consistent performance improvements across Computer Vision, EEG classification, and hierarchical metric learning tasks while greatly reducing runtime.

Method: Proposes EAGLE: a coarse-to-fine visual enhancement framework with 1) Geometric Knowledge Injection using diagram-caption data to improve geometry-language alignment, and 2) Geometric Knowledge Refinement using LLM-driven chain-of-thought to guide vision encoder to prioritize key elements.

Result: Extensive experiments demonstrate effectiveness on three popular benchmarks, showing improved geometric reasoning capabilities through enhanced visual perception.

Conclusion: EAGLE addresses MLLMs’ visual perception deficiencies in geometric reasoning by adopting a visual-centric approach that synergizes visual comprehension with mathematical reasoning.

Abstract: Multi-modal Large Language Models (MLLMs) have advanced greatly in general tasks. However, they still face challenges in geometric reasoning, a task that requires synergistic integration of visual recognition proficiency and complex reasoning strength. Existing MLLMs prioritize optimizing the LLM backbone to enhance problem-solving capabilities, while rarely emphasizing improvements in discerning visual elements. However, we reveal that MLLMs suffer from severe visual perception deficiencies, including inaccurate geometric comprehension and severe visual hallucinations, which constrain their reasoning performance. To address this issue, we revisit geometric reasoning through a visual-centric lens that highlights the role of visual perception. To achieve this, we propose EAGLE, a novel coarse-to-fine visual enhancement framework that progressively leverages LLMs’ guidance to improve perception proficiency. Specifically, given the substantial disparity between geometric diagrams and natural images, we first introduce Geometric Knowledge Injection. This process explores fundamental knowledge from diagram-caption data to enhance recognition capabilities and improve geometry-language alignments. Then, recognizing that different elements contribute unequally in the reasoning process, we introduce Geometric Knowledge Refinement. This stage leverages LLM-driven chain-of-thought solutions to guide the vision encoder in adaptively prioritizing key elements, fostering a synergistic interplay between visual comprehension and mathematical reasoning. Finally, we develop EAGLE, a geometry expert with strong perception and reasoning capabilities. Extensive experiments demonstrate its effectiveness on three popular benchmarks.

Method: DeepLabV3-based deep learning framework integrating multi-source data including Landsat-8 imagery, VIIRS nighttime lights, ESRI LULC, and GHS-SMOD. Uses semantic segmentation to capture fine-scale settlement morphology at 10m resolution across Africa from 2016-2022.

Result: Model achieves 65% overall accuracy and 0.47 Kappa coefficient at continental scale, outperforming existing global products like SMOD. Produces High-Resolution Urban-Rural (HUR) dataset covering 2016-2022 period.

Conclusion: The framework provides an open, reproducible approach for high-resolution settlement mapping in Africa, enabling more context-aware analyses of rapidly evolving settlement systems with potential for historical extension back to 1990s.

Abstract: Accurate and consistent mapping of urban and rural areas is crucial for sustainable development, spatial planning, and policy design. It is particularly important in simulating the complex interactions between human activities and natural resources. Existing global urban-rural datasets such as such as GHSL-SMOD, GHS Degree of Urbanisation, and GRUMP are often spatially coarse, methodologically inconsistent, and poorly adapted to heterogeneous regions such as Africa, which limits their usefulness for policy and research. Their coarse grids and rule-based classification methods obscure small or informal settlements, and produce inconsistencies between countries. In this study, we develop a DeepLabV3-based deep learning framework that integrates multi-source data, including Landsat-8 imagery, VIIRS nighttime lights, ESRI Land Use Land Cover (LULC), and GHS-SMOD, to produce a 10m resolution urban-rural map across the African continent from 2016 to 2022. The use of Landsat data also highlights the potential to extend this mapping approach historically, reaching back to the 1990s. The model employs semantic segmentation to capture fine-scale settlement morphology, and its outputs are validated using the Demographic and Health Surveys (DHS) dataset, which provides independent, survey-based urban-rural labels. The model achieves an overall accuracy of 65% and a Kappa coefficient of 0.47 at the continental scale, outperforming existing global products such as SMOD. The resulting High-Resolution Urban-Rural (HUR) dataset provides an open and reproducible framework for mapping human settlements, enabling more context-aware analyses of Africa’s rapidly evolving settlement systems. We release a continent-wide urban-rural dataset covering the period from 2016 to 2022, offering a new source for high-resolution settlement mapping in Africa.

Method: Uses Textual Inversion (TI) to learn embedding vectors for images, then adds calibrated noise to the aggregation of per-image embeddings to ensure differential privacy guarantees (DPAgg-TI).

Result: DPAgg-TI outperforms DP-SGD fine-tuning in both utility and robustness under the same privacy budget, achieving results closely matching non-private baselines on style adaptation tasks with private artwork and Olympic pictograms.

Conclusion: Aggregating textual embeddings with calibrated noise provides a more effective approach for privacy-preserving personalization of diffusion models compared to DP-SGD fine-tuning, especially in small data regimes.

Abstract: Personalizing large-scale diffusion models poses serious privacy risks, especially when adapting to small, sensitive datasets. A common approach is to fine-tune the model using differentially private stochastic gradient descent (DP-SGD), but this suffers from severe utility degradation due to the high noise needed for privacy, particularly in the small data regime. We propose an alternative that leverages Textual Inversion (TI), which learns an embedding vector for an image or set of images, to enable adaptation under differential privacy (DP) constraints. Our approach, Differentially Private Aggregation via Textual Inversion (DPAgg-TI), adds calibrated noise to the aggregation of per-image embeddings to ensure formal DP guarantees while preserving high output fidelity. We show that DPAgg-TI outperforms DP-SGD finetuning in both utility and robustness under the same privacy budget, achieving results closely matching the non-private baseline on style adaptation tasks using private artwork from a single artist and Paris 2024 Olympic pictograms. In contrast, DP-SGD fails to generate meaningful outputs in this setting.

Method: Proposes Skip-DiT, a DiT variant enhanced with Long-Skip-Connections (LSCs) inspired by U-Nets. LSCs stabilize feature dynamics through theoretical spectral norm analysis. Enables efficient statical caching mechanism that reuses deep features across timesteps while updating shallow components.

Result: Achieves 4.4× training acceleration and faster convergence, 1.5-2× inference acceleration with negligible quality loss and high fidelity to original output. Outperforms existing DiT caching methods across various quantitative metrics.

Conclusion: Long-Skip-Connections are critical architectural components for stable and efficient diffusion transformers, establishing Skip-DiT as an improved architecture for image and video generation tasks.

Abstract: Diffusion Transformers (DiT) have emerged as a powerful architecture for image and video generation, offering superior quality and scalability. However, their practical application suffers from inherent dynamic feature instability, leading to error amplification during cached inference. Through systematic analysis, we identify the absence of long-range feature preservation mechanisms as the root cause of unstable feature propagation and perturbation sensitivity. To this end, we propose Skip-DiT, an image and video generative DiT variant enhanced with Long-Skip-Connections (LSCs) - the key efficiency component in U-Nets. Theoretical spectral norm and visualization analysis demonstrate how LSCs stabilize feature dynamics. Skip-DiT architecture and its stabilized dynamic feature enable an efficient statical caching mechanism that reuses deep features across timesteps while updating shallow components. Extensive experiments across the image and video generation tasks demonstrate that Skip-DiT achieves: (1) 4.4 times training acceleration and faster convergence, (2) 1.5-2 times inference acceleration with negligible quality loss and high fidelity to the original output, outperforming existing DiT caching methods across various quantitative metrics. Our findings establish Long-Skip-Connections as critical architectural components for stable and efficient diffusion transformers. Codes are provided in the https://github.com/OpenSparseLLMs/Skip-DiT.

Method: Created OmniHD-Scenes dataset with 1501 clips (450K+ frames) using 128-beam LiDAR, 6 cameras, and 6 4D imaging radars. Developed novel 4D annotation pipeline and automated dense occupancy ground truth generation leveraging non-key frames.

Result: Dataset includes 200 annotated clips with 514K+ 3D bounding boxes and semantic segmentation. Established benchmarks for 3D detection and semantic occupancy prediction using surround-view cameras and 4D imaging radar.

Conclusion: OmniHD-Scenes provides comprehensive multimodal data for autonomous driving research, demonstrating effectiveness of low-cost sensor configurations and robustness under adverse conditions.

Abstract: The rapid advancement of deep learning has intensified the need for comprehensive data for use by autonomous driving algorithms. High-quality datasets are crucial for the development of effective data-driven autonomous driving solutions. Next-generation autonomous driving datasets must be multimodal, incorporating data from advanced sensors that feature extensive data coverage, detailed annotations, and diverse scene representation. To address this need, we present OmniHD-Scenes, a large-scale multimodal dataset that provides comprehensive omnidirectional high-definition data. The OmniHD-Scenes dataset combines data from 128-beam LiDAR, six cameras, and six 4D imaging radar systems to achieve full environmental perception. The dataset comprises 1501 clips, each approximately 30-s long, totaling more than 450K synchronized frames and more than 5.85 million synchronized sensor data points. We also propose a novel 4D annotation pipeline. To date, we have annotated 200 clips with more than 514K precise 3D bounding boxes. These clips also include semantic segmentation annotations for static scene elements. Additionally, we introduce a novel automated pipeline for generation of the dense occupancy ground truth, which effectively leverages information from non-key frames. Alongside the proposed dataset, we establish comprehensive evaluation metrics, baseline models, and benchmarks for 3D detection and semantic occupancy prediction. These benchmarks utilize surround-view cameras and 4D imaging radar to explore cost-effective sensor solutions for autonomous driving applications. Extensive experiments demonstrate the effectiveness of our low-cost sensor configuration and its robustness under adverse conditions. Data will be released at https://www.2077ai.com/OmniHD-Scenes.

Method: Two-layer approach: 1) Feature extraction from CT scan slices using co-scale convolutional attention (CCA) classifier with added layers, feature selection via variance analysis and bootstrap forest algorithm for discriminative power assessment, 2) Novel uncertainty-based fuzzy integral operator to fuse information from different CT slices while accounting for dependencies between consecutive slices.

Result: The method significantly improves detection accuracy for intracranial hemorrhage by combining informative feature selection with slice dependency-aware fusion.

Conclusion: The proposed approach effectively detects ICH and classifies SDH types through explainable AI techniques and sophisticated multi-slice information fusion.

Abstract: Intracranial hemorrhage (ICH) refers to the leakage or accumulation of blood within the skull, which occurs due to the rupture of blood vessels in or around the brain. If this condition is not diagnosed in a timely manner and appropriately treated, it can lead to serious complications such as decreased consciousness, permanent neurological disabilities, or even death.The primary aim of this study is to detect the occurrence or non-occurrence of ICH, followed by determining the type of subdural hemorrhage (SDH). These tasks are framed as two separate binary classification problems. By adding two layers to the co-scale convolutional attention (CCA) classifier architecture, we introduce a novel approach for ICH detection. In the first layer, after extracting features from different slices of computed tomography (CT) scan images, we combine these features and select the 50 components that capture the highest variance in the data, considering them as informative features. We then assess the discriminative power of these features using the bootstrap forest algorithm, discarding those that lack sufficient discriminative ability between different classes. This algorithm explicitly determines the contribution of each feature to the final prediction, assisting us in developing an explainable AI model. The features feed into a boosting neural network as a latent feature space. In the second layer, we introduce a novel uncertainty-based fuzzy integral operator to fuse information from different CT scan slices. This operator, by accounting for the dependencies between consecutive slices, significantly improves detection accuracy.

Method: Created ScatSpotter dataset with over 9000 images and 6000 polygon annotations using “before/after/negative” protocol. Evaluated off-the-shelf models (VIT, MaskRCNN, YOLO-v9, DINO-v2) and fine-tuned DINO for detection/segmentation.

Result: Zero-shot DINO performed poorly, indicating limited foundational-model coverage. Fine-tuned DINO achieved best performance with box-level average precision of 0.69 on validation set and 0.7 on test set.

Conclusion: The dataset establishes strong baselines for detecting small, camouflaged waste objects and highlights challenges in this domain. The work supports open access to models and data with discussion of distribution mechanisms.

Abstract: Small, amorphous waste objects such as biological droppings and microtrash can be difficult to see, especially in cluttered scenes, yet they matter for environmental cleanliness, public health, and autonomous cleanup. We introduce “ScatSpotter”: a new dataset of images annotated with polygons around dog feces, collected to train and study object detection and segmentation systems for small potentially camouflaged outdoor waste. We gathered data in mostly urban environments, using “before/after/negative” (BAN) protocol: for a given location, we capture an image with the object present, an image from the same viewpoint after removal, and a nearby negative scene that often contains visually similar confusers. Image collection began in 2020. This paper focuses on two dataset checkpoints from 2025 and 2024. The dataset contains over 9000 images and 6000 polygon annotations. Of the author-captured images we held out 691 for validation and used the rest to train. Via community participation we obtained a 121-image test set that, while small, is independent from author-collected images and provides some generalization confidence across photographers, devices, and locations. Due to its limited size, we report both validation and test results. We explore the difficulty of the dataset using off-the-shelf VIT, MaskRCNN, YOLO-v9, and DINO-v2 models. Zero-shot DINO performs poorly, indicating limited foundational-model coverage of this category. Tuned DINO is the best model with a box-level average precision of 0.69 on a 691-image validation set and 0.7 on the test set. These results establish strong baselines and quantify the remaining difficulty of detecting small, camouflaged waste objects. To support open access to models and data, we compare centralized and decentralized distribution mechanisms and discuss trade-offs for sharing scientific data. Code and project details are hosted on GitHub.

Method: Proposes a Vision Transformer-based distillation framework using knowledge distillation from a large teacher model to a smaller student model, with attention mechanisms for feature extraction and robustness enhancement.

Result: Achieves state-of-the-art performance on remote sensing image retrieval benchmarks with improved efficiency and robustness compared to existing methods.

Conclusion: ERVD provides an effective solution for efficient remote sensing image retrieval through ViT-based knowledge distillation, balancing accuracy and computational efficiency.

Abstract: ERVD: An Efficient and Robust ViT-Based Distillation Framework for Remote Sensing Image Retrieval

Method: Developed an open-source multi-agent physics simulator with real-world aligned community generation pipeline, including outdoor spaces, indoor scenes, and grounded agents with rich characters. Introduced two challenges: Community Planning Challenge for multi-agent reasoning and Community Robot Challenge for heterogeneous robot collaboration.

Result: Created a comprehensive platform for studying human-robot interactions, evaluated various baselines on proposed challenges, and demonstrated difficulties in both high-level open-world task planning and low-level cooperation controls.

Conclusion: Virtual Community provides a foundation for studying human-robot coexistence in open-world environments and aims to unlock further research in embodied social intelligence and multi-agent collaboration.

Abstract: The rapid progress in AI and Robotics may lead to a profound societal transformation, as humans and robots begin to coexist within shared communities, introducing both opportunities and challenges. To explore this future, we present Virtual Community-an open-world platform for humans, robots, and society-built on a universal physics engine and grounded in real-world 3D scenes. With Virtual Community, we aim to enable the study of embodied social intelligence at scale. To support these, Virtual Community features: 1) An open-source multi-agent physics simulator that supports robots, humans, and their interactions within a society; 2) A large-scale, real-world aligned community generation pipeline, including vast outdoor space, diverse indoor scenes, and a community of grounded agents with rich characters and appearances. Leveraging Virtual Community, we propose two novel challenges. The Community Planning Challenge evaluates multi-agent reasoning and planning ability in open-world settings, such as cooperating to help agents with daily activities and efficiently connecting other agents. The Community Robot Challenge requires multiple heterogeneous robots to collaborate in solving complex open-world tasks. We evaluate various baselines on these tasks and demonstrate the challenges in both high-level open-world task planning and low-level cooperation controls. We hope that Virtual Community will unlock further study of human-robot coexistence within open-world environments.

Method: Proposes taxonomy of CAC approaches: 1) Reference-based (exemplar-guided, state-of-the-art), 2) Reference-less (leverages inherent image patterns), 3) Open-world text-guided (uses vision-language models with textual prompts). Reviews 30 CAC architectures and benchmarks performance on FSC-147 and CARPK datasets.

Result: Presents comprehensive review with performance leaderboard on gold-standard benchmarks. Reference-based methods achieve best performance, while open-world text-guided approaches offer most flexible solutions using vision-language models.

Conclusion: CAC enables counting objects across arbitrary categories without prior training. Taxonomy provides framework for understanding approaches. Open-world text-guided methods using vision-language models represent promising direction for flexible counting systems, though challenges remain in annotation dependency and generalization.

Abstract: Visual object counting has recently shifted towards class-agnostic counting (CAC), which addresses the challenge of counting objects across arbitrary categories, a crucial capability for flexible and generalizable counting systems. Unlike humans, who effortlessly identify and count objects from diverse categories without prior knowledge, most existing counting methods are restricted to enumerating instances of known classes, requiring extensive labeled datasets for training and struggling in open-vocabulary settings. In contrast, CAC aims to count objects belonging to classes never seen during training, operating in a few-shot setting. In this paper, we present the first comprehensive review of CAC methodologies. We propose a taxonomy to categorize CAC approaches into three paradigms based on how target object classes can be specified: reference-based, reference-less, and open-world text-guided. Reference-based approaches achieve state-of-the-art performance by relying on exemplar-guided mechanisms. Reference-less methods eliminate exemplar dependency by leveraging inherent image patterns. Finally, open-world text-guided methods use vision-language models, enabling object class descriptions via textual prompts, offering a flexible and promising solution. Based on this taxonomy, we provide an overview of 30 CAC architectures and report their performance on gold-standard benchmarks, discussing key strengths and limitations. Specifically, we present results on the FSC-147 dataset, setting a leaderboard using gold-standard metrics, and on the CARPK dataset to assess generalization capabilities. Finally, we offer a critical discussion of persistent challenges, such as annotation dependency and generalization, alongside future directions.

Method: ImageRAG dynamically retrieves relevant images based on text prompts and uses them as context to guide existing image conditioning models. Unlike prior approaches that require RAG-specific training, ImageRAG leverages existing models without additional training.

Result: The approach shows significant improvement in generating rare and fine-grained concepts across different base models, demonstrating adaptability across various model types.

Conclusion: ImageRAG provides an effective, adaptable solution for improving generation of rare concepts in diffusion models through retrieval-augmented generation without requiring specialized training.

Abstract: Diffusion models enable high-quality and diverse visual content synthesis. However, they struggle to generate rare or unseen concepts. To address this challenge, we explore the usage of Retrieval-Augmented Generation (RAG) with image generation models. We propose ImageRAG, a method that dynamically retrieves relevant images based on a given text prompt, and uses them as context to guide the generation process. Prior approaches that used retrieved images to improve generation, trained models specifically for retrieval-based generation. In contrast, ImageRAG leverages the capabilities of existing image conditioning models, and does not require RAG-specific training. Our approach is highly adaptable and can be applied across different model types, showing significant improvement in generating rare and fine-grained concepts using different base models. Our project page is available at: https://rotem-shalev.github.io/ImageRAG

Method: Collected 300-hour EgoLife Dataset from 6 participants living together for one week using AI glasses (egocentric video) with synchronized third-person references. Created EgoLifeQA benchmark tasks and developed EgoButler system with EgoGPT (omni-modal model) and EgoRAG (retrieval-based component).

Result: EgoGPT achieves state-of-the-art performance on egocentric video understanding. The system demonstrates capability for long-context question answering about daily life activities, health monitoring, and personalized recommendations.

Conclusion: EgoLife provides foundational resources (dataset, models, benchmarks) for egocentric AI assistant research, identifying critical factors and bottlenecks for future improvements in multimodal understanding of daily life.

Abstract: We introduce EgoLife, a project to develop an egocentric life assistant that accompanies and enhances personal efficiency through AI-powered wearable glasses. To lay the foundation for this assistant, we conducted a comprehensive data collection study where six participants lived together for one week, continuously recording their daily activities - including discussions, shopping, cooking, socializing, and entertainment - using AI glasses for multimodal egocentric video capture, along with synchronized third-person-view video references. This effort resulted in the EgoLife Dataset, a comprehensive 300-hour egocentric, interpersonal, multiview, and multimodal daily life dataset with intensive annotation. Leveraging this dataset, we introduce EgoLifeQA, a suite of long-context, life-oriented question-answering tasks designed to provide meaningful assistance in daily life by addressing practical questions such as recalling past relevant events, monitoring health habits, and offering personalized recommendations. To address the key technical challenges of (1) developing robust visual-audio models for egocentric data, (2) enabling identity recognition, and (3) facilitating long-context question answering over extensive temporal information, we introduce EgoButler, an integrated system comprising EgoGPT and EgoRAG. EgoGPT is an omni-modal model trained on egocentric datasets, achieving state-of-the-art performance on egocentric video understanding. EgoRAG is a retrieval-based component that supports answering ultra-long-context questions. Our experimental studies verify their working mechanisms and reveal critical factors and bottlenecks, guiding future improvements. By releasing our datasets, models, and benchmarks, we aim to stimulate further research in egocentric AI assistants.

Method: Two-step approach: 1) Use unsupervised GAN-based image-to-image translation to model naturally occurring damage on outdoor signboards, encoding damage characteristics into a latent ‘damage style code’. 2) Introduce adversarial perturbations into the style code, optimizing its transformation to generate adversarial images that maintain perceptual realism while effectively misleading neural networks.

Result: AdvWT effectively misleads DNNs in both digital and physical domains on traffic sign datasets. It achieves higher attack success rates, greater robustness, and more natural appearance compared to existing physical-world adversarial methods. Additionally, using AdvWT in training enhances model generalizability to real-world damaged signs.

Conclusion: AdvWT demonstrates that leveraging naturally occurring phenomena like ‘wear and tear’ can create more realistic and effective physical-world adversarial examples, offering both security vulnerabilities and potential benefits for improving model robustness through adversarial training.

Abstract: The presence of adversarial examples in the physical world poses significant challenges to the deployment of Deep Neural Networks in safety-critical applications such as autonomous driving. Most existing methods for crafting physical-world adversarial examples are ad-hoc, relying on temporary modifications like shadows, laser beams, or stickers that are tailored to specific scenarios. In this paper, we introduce a new class of physical-world adversarial examples, AdvWT, which draws inspiration from the naturally occurring phenomenon of wear and tear', an inherent property of physical objects. Unlike manually crafted perturbations, wear and tear’ emerges organically over time due to environmental degradation, as seen in the gradual deterioration of outdoor signboards. To achieve this, AdvWT follows a two-step approach. First, a GAN-based, unsupervised image-to-image translation network is employed to model these naturally occurring damages, particularly in the context of outdoor signboards. The translation network encodes the characteristics of damaged signs into a latent `damage style code’. In the second step, we introduce adversarial perturbations into the style code, strategically optimizing its transformation process. This manipulation subtly alters the damage style representation, guiding the network to generate adversarial images where the appearance of damages remains perceptually realistic, while simultaneously ensuring their effectiveness in misleading neural networks. Through comprehensive experiments on two traffic sign datasets, we show that AdvWT effectively misleads DNNs in both digital and physical domains. AdvWT achieves an effective attack success rate, greater robustness, and a more natural appearance compared to existing physical-world adversarial examples. Additionally, integrating AdvWT into training enhances a model’s generalizability to real-world damaged signs.

Method: Joint optimization of sensor poses with neural Gaussian-based scene representation, using reliable LiDAR points as anchor Gaussians to preserve global structure and auxiliary Gaussians to prevent local overfitting. Fully differentiable pipeline with photometric and geometric regularization.

Result: Consistently outperforms existing targetless methods on KITTI-360, Waymo, and Fast-LIVO2 datasets, and yields more consistent Novel View Synthesis results reflecting improved extrinsic alignment.

Conclusion: TLC-Calib provides robust and generalizable targetless calibration for LiDAR-camera systems, addressing practical deployment challenges and sensor drift issues.

Abstract: Accurate LiDAR-camera calibration is crucial for multi-sensor systems. However, traditional methods often rely on physical targets, which are impractical for real-world deployment. Moreover, even carefully calibrated extrinsics can degrade over time due to sensor drift or external disturbances, necessitating periodic recalibration. To address these challenges, we present a Targetless LiDAR-Camera Calibration (TLC-Calib) that jointly optimizes sensor poses with a neural Gaussian-based scene representation. Reliable LiDAR points are frozen as anchor Gaussians to preserve global structure, while auxiliary Gaussians prevent local overfitting under noisy initialization. Our fully differentiable pipeline with photometric and geometric regularization achieves robust and generalizable calibration, consistently outperforming existing targetless methods on the KITTI-360, Waymo, and Fast-LIVO2 datasets. In addition, it yields more consistent Novel View Synthesis results, reflecting improved extrinsic alignment. The project page is available at: https://www.haebeom.com/tlc-calib-site/.

Method: VisionReasoner uses a unified reward mechanism and multi-object cognitive learning strategies to enhance reasoning capabilities. It generates structured reasoning processes before delivering outputs, and is evaluated on 10 diverse tasks across detection, segmentation, and counting domains.

Result: VisionReasoner achieves superior performance as a unified model, outperforming baseline Qwen2.5VL by 29.1% on COCO (detection), 22.1% on ReasonSeg (segmentation), and 13.2% on CountBench (counting). Human evaluation shows its reasoning processes are faithful and reliable.

Conclusion: VisionReasoner demonstrates that unified vision-language models can effectively handle diverse visual perception tasks with transparent reasoning, showing strong performance across multiple domains without task-specific annotations.

Abstract: Large vision-language models exhibit inherent capabilities to handle diverse visual perception tasks. In this paper, we introduce VisionReasoner, a unified framework capable of reasoning and solving multiple visual perception tasks within a shared model. Specifically, by designing a unified reward mechanism and multi-object cognitive learning strategies, VisionReasoner enhances its reasoning capabilities to analyze visual inputs, and addresses diverse perception tasks within a unified model. VisionReasoner generates a structured reasoning process before delivering the desired outputs responding to user queries. Human evaluation reveals the reasoning process of VisionReasoner is faithful and reliable even without annotated reasoning train data. To rigorously assess unified visual perception capabilities, we evaluate VisionReasoner on ten diverse tasks spanning three critical domains: detection, segmentation, and counting. Experimental results show that VisionReasoner achieves superior performance as a unified model, outperforming the baseline Qwen2.5VL by relative margins of 29.1% on COCO (detection), 22.1% on ReasonSeg (segmentation), and 13.2% on CountBench (counting).

Method: Proposes ReaTrack, a training-free framework that synergizes the reasoning capabilities of Thinking-variant Large Vision-Language Models (LVLM) with the precise temporal modeling of SAM2 for segmentation and tracking. Also introduces the ReaMOT Challenge benchmark with 1,156 instructions categorized into High-Level Reasoning and Low-Level Perception tasks.

Result: Extensive experiments on the ReaMOT Challenge benchmark demonstrate the effectiveness of the ReaTrack framework in handling complex reasoning-based tracking tasks.

Conclusion: ReaMOT addresses the limitation of existing RMOT methods by introducing reasoning-based tracking, providing a comprehensive benchmark, and demonstrating a viable framework that combines LVLM reasoning with temporal modeling for complex instruction-based tracking.

Abstract: Referring Multi-Object Tracking (RMOT) aims to track targets specified by language instructions. However, existing RMOT paradigms are largely designed for explicit instructions and consequently fail to generalize to complex instructions that require logical reasoning. To overcome this, we propose Reasoning-based Multi-Object Tracking (ReaMOT), a novel task that requires models to identify and track targets that satisfy implicit constraints via logical reasoning. To advance this field, we construct the ReaMOT Challenge, a comprehensive benchmark comprising: (1) a large-scale dataset with 1,156 instructions categorized into High-Level Reasoning and Low-Level Perception, covering 423,359 image-language pairs across 869 diverse scenes; and (2) a tailored metric suite designed to jointly evaluate reasoning accuracy and tracking robustness. Furthermore, we propose ReaTrack, a training-free framework that synergizes the reasoning capabilities of Thinking-variant Large Vision-Language Model (LVLM) with the precise temporal modeling of SAM2. Extensive experiments on the ReaMOT Challenge benchmark demonstrates the effectiveness of our ReaTrack framework.

Method: Redesign NeRFs to learn from multipath signals by modeling how RF/audio signals propagate and reflect in environments, applied to indoor WiFi measurements for floorplan inference.

Result: Successfully learns implicit floorplans from sparse WiFi measurements, enabling indoor signal prediction and basic ray tracing applications.

Conclusion: NeRFs can be adapted to work with multipath signals, opening new possibilities for environment inference using RF/audio measurements beyond visual data.

Abstract: Neural Radiance Fields (NeRFs) have been remarkably successful at synthesizing novel views of 3D scenes by optimizing a volumetric scene function. This scene function models how optical rays bring color information from a 3D object to the camera pixels. Radio frequency (RF) or audio signals can also be viewed as a vehicle for delivering information about the environment to a sensor. However, unlike camera pixels, an RF/audio sensor receives a mixture of signals that contain many environmental reflections (also called “multipath”). Is it still possible to infer the environment using such multipath signals? We show that with redesign, NeRFs can be taught to learn from multipath signals, and thereby “see” the environment. As a grounding application, we aim to infer the indoor floorplan of a home from sparse WiFi measurements made at multiple locations inside the home. Although a difficult inverse problem, our implicitly learnt floorplans look promising, and enables forward applications, such as indoor signal prediction and basic ray tracing.

Method: Proposes Geometric Encoding Mixer (GEM) - a geometry-aware PEFT module that explicitly integrates fine-grained local positional encodings with a lightweight latent attention mechanism to capture comprehensive global context for 3D point cloud transformers.

Result: GEM achieves performance comparable to or sometimes exceeding full fine-tuning while updating only 1.6% of parameters, with significantly reduced training time and memory requirements compared to other PEFT methods.

Conclusion: GEM sets a new benchmark for efficient, scalable, and geometry-aware fine-tuning of large-scale 3D point cloud models, bridging the gap between parameter efficiency and geometric understanding.

Abstract: The emergence of large-scale pre-trained point cloud models has significantly advanced 3D scene understanding, but adapting these models to specific downstream tasks typically demands full fine-tuning, incurring high computational and storage costs. Parameter-efficient fine-tuning (PEFT) techniques, successful in natural language processing and 2D vision tasks, would underperform when naively applied to 3D point cloud models due to significant geometric and spatial distribution shifts. Existing PEFT methods commonly treat points as orderless tokens, neglecting important local spatial structures and global geometric contexts in 3D modeling. To bridge this gap, we introduce the Geometric Encoding Mixer (GEM), a novel geometry-aware PEFT module specifically designed for 3D point cloud transformers. GEM explicitly integrates fine-grained local positional encodings with a lightweight latent attention mechanism to capture comprehensive global context, thereby effectively addressing the spatial and geometric distribution mismatch. Extensive experiments demonstrate that GEM achieves performance comparable to or sometimes even exceeding full fine-tuning, while only updating 1.6% of the model’s parameters, fewer than other PEFT methods. With significantly reduced training time and memory requirements, our approach thus sets a new benchmark for efficient, scalable, and geometry-aware fine-tuning of large-scale 3D point cloud models. Code is available at https://github.com/LiyaoTang/GEM.

Method: Proposes SRR paradigm dividing document parsing into three questions: structure detection (Where?), content recognition (What?), and relation prediction (How?). Uses MonkeyDoc dataset (4.5M bilingual instances, 10+ doc types) to train 3B-parameter foundation model, then applies contiguous parameter degradation to create smaller models (0.6B-1.2B).

Result: Achieves state-of-the-art performance surpassing previous open-source and closed-source methods including Gemini 2.5-Pro. Models can be efficiently deployed on single RTX 3090 GPU.

Conclusion: MonkeyOCR demonstrates effectiveness of SRR paradigm for document parsing, with comprehensive dataset and efficient model scaling enabling practical deployment while maintaining SOTA performance.

Abstract: We introduce MonkeyOCR, a document parsing model that advances the state of the art by leveraging a Structure-Recognition-Relation (SRR) triplet paradigm. This design simplifies what would otherwise be a complex multi-tool pipeline and avoids the inefficiencies of processing full pages with giant end-to-end models. In SRR, document parsing is abstracted into three fundamental questions - Where is it?'' (structure), What is it?’’ (recognition), and ``How is it organized?’’ (relation) - corresponding to structure detection, content recognition, and relation prediction. To support this paradigm, we present MonkeyDoc, a comprehensive dataset with 4.5 million bilingual instances spanning over ten document types, which addresses the limitations of existing datasets that often focus on a single task, language, or document type. Leveraging the SRR paradigm and MonkeyDoc, we trained a 3B-parameter document foundation model. We further identify parameter redundancy in this model and propose contiguous parameter degradation (CPD), enabling the construction of models from 0.6B to 1.2B parameters that run faster with acceptable performance drop. MonkeyOCR achieves state-of-the-art performance, surpassing previous open-source and closed-source methods, including Gemini 2.5-Pro. Additionally, the model can be efficiently deployed for inference on a single RTX 3090 GPU. Code and models will be released at https://github.com/Yuliang-Liu/MonkeyOCR.

Method: The study uses Projected Gradient Descent (PGD) to generate adversarial watermarks (imperceptible perturbations) that fool ViT models. It examines the transferability of these attacks to CNN models and analyzes the effectiveness of adversarial training as a defense mechanism.

Result: ViTs show significant vulnerability to adversarial watermarking attacks, with accuracy dropping as low as 27.6% when attacked. However, adversarial training proves effective, raising accuracy back up to 90.0%. The attacks show some transferability to CNNs but ViTs are more susceptible.

Conclusion: Vision Transformers for medical image analysis are highly vulnerable to adversarial attacks despite their strong performance on clean images. Adversarial training is an effective defense strategy that can significantly mitigate this vulnerability while maintaining performance on clean data.

Abstract: Deep learning models have shown remarkable success in dermatological image analysis, offering potential for automated skin disease diagnosis. Previously, convolutional neural network(CNN) based architectures have achieved immense popularity and success in computer vision (CV) based task like skin image recognition, generation and video analysis. But with the emergence of transformer based models, CV tasks are now are nowadays carrying out using these models. Vision Transformers (ViTs) is such a transformer-based models that have shown success in computer vision. It uses self-attention mechanisms to achieve state-of-the-art performance across various tasks. However, their reliance on global attention mechanisms makes them susceptible to adversarial perturbations. This paper aims to investigate the susceptibility of ViTs for medical images to adversarial watermarking-a method that adds so-called imperceptible perturbations in order to fool models. By generating adversarial watermarks through Projected Gradient Descent (PGD), we examine the transferability of such attacks to CNNs and analyze the performance defense mechanism – adversarial training. Results indicate that while performance is not compromised for clean images, ViTs certainly become much more vulnerable to adversarial attacks: an accuracy drop of as low as 27.6%. Nevertheless, adversarial training raises it up to 90.0%.

Method: Develops a unifying image processing framework to explain self-attention computation and architectural components. Introduces two independent architectural modifications within transformers inspired by image processing principles.

Result: Image processing-inspired modifications lead to improved accuracy and robustness against data contamination and adversaries across language and vision tasks, as well as better long sequence understanding.

Conclusion: The image processing framework provides a theoretical foundation for understanding self-attention mechanisms and their variants, while practical modifications demonstrate performance improvements across multimodal tasks.

Abstract: The self-attention mechanism, a cornerstone of Transformer-based state-of-the-art deep learning architectures, is largely heuristic-driven and fundamentally challenging to interpret. Establishing a robust theoretical foundation to explain its remarkable success and limitations has therefore become an increasingly prominent focus in recent research. Some notable directions have explored understanding self-attention through the lens of image denoising and nonparametric regression. While promising, existing frameworks still lack a deeper mechanistic interpretation of various architectural components that enhance self-attention, both in its original formulation and subsequent variants. In this work, we aim to advance this understanding by developing a unifying image processing framework, capable of explaining not only the self-attention computation itself but also the role of components such as positional encoding and residual connections, including numerous later variants. We also pinpoint potential distinctions between the two concepts building upon our framework, and make effort to close this gap. We introduce two independent architectural modifications within transformers. While our primary objective is interpretability, we empirically observe that image processing-inspired modifications can also lead to notably improved accuracy and robustness against data contamination and adversaries across language and vision tasks as well as better long sequence understanding.

Method: Introduces Vehicle Vision-Language Models (V2LMs) - fine-tuned vision-language models specialized for autonomous vehicle perception. Studies two deployment approaches: Solo (task-specific V2LMs) and Tandem (single V2LM for all three tasks). Also explores integrating V2LMs in parallel with existing perception stacks.

Result: Under adversarial attacks, conventional DNNs drop 33-74% in accuracy, while V2LMs decline by under 8% on average. Tandem deployment achieves comparable robustness to Solo while being more memory-efficient.

Conclusion: V2LMs are inherently more robust to unseen attacks without adversarial training, suggesting they are a promising path toward secure, robust autonomous vehicle perception.

Abstract: Autonomous vehicles rely on deep neural networks (DNNs) for traffic sign recognition, lane centering, and vehicle detection, yet these models are vulnerable to attacks that induce misclassification and threaten safety. Existing defenses (e.g., adversarial training) often fail to generalize and degrade clean accuracy. We introduce Vehicle Vision-Language Models (V2LMs), fine-tuned vision-language models specialized for autonomous vehicle perception, and show that they are inherently more robust to unseen attacks without adversarial training, maintaining substantially higher adversarial accuracy than conventional DNNs. We study two deployments: Solo (task-specific V2LMs) and Tandem (a single V2LM for all three tasks). Under attacks, DNNs drop 33-74%, whereas V2LMs decline by under 8% on average. Tandem achieves comparable robustness to Solo while being more memory-efficient. We also explore integrating V2LMs in parallel with existing perception stacks to enhance resilience. Our results suggest V2LMs are a promising path toward secure, robust AV perception.

Method: Uses bidirectional state-space model with gated skip connections and learnable weighted-average pooling on periodically inserted learned queries for hierarchical downsampling across spatial and temporal dimensions.

Result: Competitive results on challenging hour-long video understanding tasks while significantly reducing token budget; replacing state-space model with conventional modules causes substantial performance degradation.

Conclusion: The proposed state-space modeling effectively compresses multi-frame video information, emphasizing resource-conscious efficiency for practical real-world deployments with validated scalability and generality across benchmarks.

Abstract: We propose an efficient framework to compress massive video-frame features before feeding them into large multimodal models, thereby mitigating the severe token explosion arising from hour-long videos. Our design leverages a bidirectional state-space model equipped with a gated skip connection and a learnable weighted-average pooling mechanism applied to periodically inserted learned queries. This structure enables hierarchical downsampling across both spatial and temporal dimensions, preserving performance in a cost-effective manner. Across challenging hour-long video understanding tasks, our approach demonstrates competitive results against state-of-the-art models, while significantly reducing overall token budget. Notably, replacing our state-space model with conventional modules results in substantial performance degradation, highlighting the advantages of the proposed state-space modeling for effectively compressing multi-frame video information. Our framework emphasizes resource-conscious efficiency, making it practical for real-world deployments. We validate its scalability and generality across multiple benchmarks, achieving the dual objectives of efficient resource usage and comprehensive video understanding.

Method: Proposes VLCoT framework integrating vision and language reasoning, inspired by Chain of Thought in LLMs. Introduces Group Relative Policy Optimization (GRPO) with four reward functions: format reward (standardizes CoT process), degradation reward (accurate degradation estimation), understanding reward (content accuracy), and generation reward (visual expert model evaluates image quality).

Result: Extensive experiments show RealSR-R1 can generate realistic details and accurately understand image content, especially in semantically rich scenes or images with severe degradation.

Conclusion: The proposed RealSR-R1 with VLCoT-GRPO framework successfully integrates vision-language reasoning with reinforcement learning to improve real-world image super-resolution, particularly for challenging degradation scenarios.

Abstract: Real-World Image Super-Resolution is one of the most challenging task in image restoration. However, existing methods struggle with an accurate understanding of degraded image content, leading to reconstructed results that are both low-fidelity and unnatural. We present RealSR-R1 in this work, which empowers the RealSR models with understanding and reasoning capabilities. Inspired by the success of Chain of Thought (CoT) in large language models (LLMs), we simulate the human process of handling degraded images and propose the VLCoT framework, which integrates vision and language reasoning. The framework aims to precisely restore image details by progressively generating more comprehensive text and higher-resolution images. To overcome the challenge of traditional supervised learning CoT failing to generalize to real-world scenarios, we introduce, for the first time, Group Relative Policy Optimization (GRPO) into the Real-World Image Super-Resolution task. We propose VLCoT-GRPO as a solution, which designs four reward functions: (1) Format reward, used to standardize the CoT process; (2) Degradation reward, to incentivize accurate degradation estimation; (3) Understanding reward, to ensure the accuracy of the generated content; and (4) Generation reward, where we propose using a visual expert model to evaluate the quality of generated images, encouraging the model to generate more realistic images. Extensive experiments demonstrate that our proposed RealSR-R1 can generate realistic details and accurately understand image content, particularly in semantically rich scenes or images with severe degradation.

Method: Proposes a multi-image-to-hyperspectral reconstruction (MI-HSR) framework using a triple-camera smartphone system with two lenses equipped with carefully selected spectral filters. Introduces Doomer dataset with aligned images from three smartphone cameras and a hyperspectral reference camera across diverse scenes.

Result: The proposed HSR model achieves consistent improvements over existing methods, allowing 30% more accurate spectral estimation compared to ordinary RGB cameras. The multi-view spectral filtering approach unlocks more accurate and practical hyperspectral imaging solutions.

Conclusion: Multi-view spectral filtering with commodity hardware can significantly improve hyperspectral reconstruction accuracy, making hyperspectral imaging more practical and accessible through smartphone-based solutions.

Abstract: Hyperspectral reconstruction (HSR) from RGB images is a fundamentally ill-posed problem due to severe spectral information loss. Existing approaches typically rely on a single RGB image, limiting reconstruction accuracy. In this work, we propose a novel multi-image-to-hyperspectral reconstruction (MI-HSR) framework that leverages a triple-camera smartphone system, where two lenses are equipped with carefully selected spectral filters. Our configuration, grounded in theoretical and empirical analysis, enables richer and more diverse spectral observations than conventional single-camera setups. To support this new paradigm, we introduce Doomer, the first dataset for MI-HSR, comprising aligned images from three smartphone cameras and a hyperspectral reference camera across diverse scenes. We show that the proposed HSR model achieves consistent improvements over existing methods on the newly proposed benchmark. In a nutshell, our setup allows 30% towards more accurately estimated spectra compared to an ordinary RGB camera. Our findings suggest that multi-view spectral filtering with commodity hardware can unlock more accurate and practical hyperspectral imaging solutions.

Method: Extracted 26,244 HOG features and 800 TDA features from APTOS retinal fundus dataset, trained seven classical ML models (logistic regression, random forest, XGBoost, SVM, decision trees, k-NN, Extra Trees) using 10-fold cross-validation for binary and multi-class classification tasks.

Result: XGBoost achieved best performance: 94.29% (HOG) and 94.18% (TDA) accuracy for binary classification; 74.41% (HOG) and 74.69% (TDA) for multi-class classification, showing similar performance between feature types.

Conclusion: Gradient-based (HOG) and topological (TDA) features provide complementary representations of retinal image structure, highlighting potential for integrating both approaches for interpretable and robust medical image classification.

Abstract: This work presents a comparative evaluation of two fundamentally different feature extraction paradigms–Histogram of Oriented Gradients (HOG) and Topological Data Analysis (TDA)–for medical image classification, with a focus on retinal fundus imagery. HOG captures local structural information by modeling gradient orientation distributions within spatial regions, effectively encoding texture and edge patterns. In contrast, TDA, implemented through cubical persistent homology, extracts global topological descriptors that characterize shape, connectivity, and intensity-based structure across images. We evaluate both approaches on the publicly available APTOS retinal fundus dataset for two classification tasks: binary classification (normal vs. diabetic retinopathy (DR)) and five-class DR severity grading. From each image, 26,244 HOG features and 800 TDA features are extracted and independently used to train seven classical machine learning models, including logistic regression, random forest, XGBoost, support vector machines, decision trees, k-nearest neighbors, and Extra Trees, using 10-fold cross-validation. Experimental results show that XGBoost achieves the best performance across both feature types. For binary classification, accuracies of 94.29% (HOG) and 94.18% (TDA) are obtained, while multi-class classification yields accuracies of 74.41% and 74.69%, respectively. These results demonstrate that gradient-based and topological features provide complementary representations of retinal image structure and highlight the potential of integrating both approaches for interpretable and robust medical image classification.

Method: Developed PhyWorldBench with 1050 curated prompts across fundamental physics, composite scenarios, and Anti-Physics categories; used both human evaluation and a novel zero-shot method leveraging multimodal LLMs for physics realism assessment.

Result: Evaluated 12 state-of-the-art text-to-video models (5 open-source, 5 proprietary) revealing significant challenges in adhering to physical laws; identified performance variations across different physical phenomena and prompt types.

Conclusion: Current video generation models struggle with physical consistency; the benchmark provides targeted recommendations for prompt engineering to enhance physics fidelity and establishes a foundation for future physics-aware video generation research.

Abstract: Video generation models have achieved remarkable progress in creating high-quality, photorealistic content. However, their ability to accurately simulate physical phenomena remains a critical and unresolved challenge. This paper presents PhyWorldBench, a comprehensive benchmark designed to evaluate video generation models based on their adherence to the laws of physics. The benchmark covers multiple levels of physical phenomena, ranging from fundamental principles such as object motion and energy conservation to more complex scenarios involving rigid body interactions and human or animal motion. Additionally, we introduce a novel Anti-Physics category, where prompts intentionally violate real-world physics, enabling the assessment of whether models can follow such instructions while maintaining logical consistency. Besides large-scale human evaluation, we also design a simple yet effective method that utilizes current multimodal large language models to evaluate physics realism in a zero-shot fashion. We evaluate 12 state-of-the-art text-to-video generation models, including five open-source and five proprietary models, with detailed comparison and analysis. Through systematic testing across 1050 curated prompts spanning fundamental, composite, and anti-physics scenarios, we identify pivotal challenges these models face in adhering to real-world physics. We further examine their performance under diverse physical phenomena and prompt types, and derive targeted recommendations for crafting prompts that enhance fidelity to physical principles.

Method: SegQuant proposes: 1) SegLinear - a segment-aware, graph-based quantization strategy that captures structural semantics and spatial heterogeneity, and 2) DualScale - a dual-scale quantization scheme that preserves polarity-asymmetric activations crucial for visual fidelity. The framework is designed to be broadly applicable beyond Transformer-based diffusion models.

Result: The method achieves strong performance while ensuring seamless compatibility with mainstream deployment tools, offering a versatile quantization solution for diffusion models.

Conclusion: SegQuant provides a unified quantization framework that enhances cross-model versatility for diffusion models, addressing the limitations of existing post-training quantization methods and enabling more efficient deployment in resource-constrained environments.

Abstract: Diffusion models have demonstrated exceptional generative capabilities but are computationally intensive, posing significant challenges for deployment in resource-constrained or latency-sensitive environments. Quantization offers an effective means to reduce model size and computational cost, with post-training quantization (PTQ) being particularly appealing due to its compatibility with pre-trained models without requiring retraining or training data. However, existing PTQ methods for diffusion models often rely on architecture-specific heuristics that limit their generalizability and hinder integration with industrial deployment pipelines. To address these limitations, we propose SegQuant, a unified quantization framework that adaptively combines complementary techniques to enhance cross-model versatility. SegQuant consists of a segment-aware, graph-based quantization strategy (SegLinear) that captures structural semantics and spatial heterogeneity, along with a dual-scale quantization scheme (DualScale) that preserves polarity-asymmetric activations, which is crucial for maintaining visual fidelity in generated outputs. SegQuant is broadly applicable beyond Transformer-based diffusion models, achieving strong performance while ensuring seamless compatibility with mainstream deployment tools.

Method: Two-stage training scheme with self-distillation strategy that progressively endows CLIP with high-fidelity reconstruction while preserving comprehension. Dual-condition architecture built on MetaQuery framework uses multimodal hidden states and learnable query embeddings to leverage MLLM reasoning abilities.

Result: UniLIP outperforms larger unified models like BAGEL (7B) and Uniworld-V1 (12B) with only 1B and 3B parameters, achieving SOTA performance: 0.90 on GenEval, 0.63 on WISE, and 3.94 on ImgEdit.

Conclusion: UniLIP successfully expands CLIP’s application, establishing its continuous features as optimal for understanding tasks while achieving highly competitive performance in generation and editing tasks.

Abstract: In this paper, we propose UniLIP, a unified framework that adapts CLIP for multimodal understanding, generation and editing. Although CLIP excels at understanding, it lacks reconstruction abilities required to be a unified visual encoder. However, previous CLIP-based unified methods fail to balance understanding and reconstruction, leading to semantic degradation or inconsistent reconstructions. In contrast, we introduce a novel two-stage training scheme with a self-distillation strategy that progressively endows CLIP with high-fidelity reconstruction abilities while preserving its original comprehension performance. For enhanced reasoning and consistency in generation and editing, we further develop a dual-condition architecture built upon the MetaQuery framework. Our architecture jointly utilizes multimodal hidden states for rich contextual details and learnable query embeddings to harness the powerful reasoning abilities of Multimodal Large Language Models (MLLMs). Leveraging advanced image representation and architectural design, UniLIP demonstrates superior instruction following and edit fidelity. With only 1B and 3B parameters, UniLIP can outperform larger unified models such as BAGEL (7B) and Uniworld-V1 (12B), achieving state-of-the-art performance of 0.90 on GenEval, 0.63 on WISE, and 3.94 on ImgEdit. These results demonstrate that UniLIP successfully expands the application of CLIP, establishing its continuous features to not only serve as the optimal choice for understanding tasks but also achieve highly competitive performance in generation and editing tasks. Code and models are available at https://github.com/nnnth/UniLIP.

Method: Proposes Vcamba with: 1) Frequency-domain sequential scanning (FSS) strategy to unfold spectrum based on structural evolution patterns, 2) Adaptive frequency enhancement (AFE) module using Mamba to model causal dependencies, 3) Space-based long-range motion perception (SLMP) and frequency-based long-range motion perception (FLMP) modules, 4) Space and frequency motion fusion (SFMF) module to integrate dual-domain features.

Result: Outperforms state-of-the-art methods across 6 evaluation metrics on 2 datasets with lower computation cost, confirming superiority.

Conclusion: Vcamba effectively integrates frequency and spatial features for efficient and accurate video camouflaged object detection using spatio-frequency motion perception with Mamba’s linear-time modeling capabilities.

Abstract: Existing video camouflaged object detection (VCOD) methods primarily rely on spatial appearances for motion perception. However, the high foreground-background similarity in VCOD limits the discriminability of such features (e.g. color and texture). Recent studies demonstrate that frequency features can not only compensate for appearance limitations, but also perceive motion through dynamic variations in spectral energy. Meanwhile, the emerging state space model called Mamba enables efficient motion perception in frame sequences with its linear-time long-sequence modeling capability. Motivated by this, we propose Vcamba, a visual camouflage Mamba based on spatio-frequency motion perception that integrates frequency and spatial features for efficient and accurate VCOD. Specifically, by analyzing the spatial representations of frequency components, we reveal a structural evolution pattern that emerges from the ordered superposition of components. Based on this observation, we propose a unique frequency-domain sequential scanning (FSS) strategy to unfold the spectrum. Utilizing FSS, the adaptive frequency enhancement (AFE) module employs Mamba to model the causal dependencies within sequences, enabling effective frequency learning. Furthermore, we propose a space-based long-range motion perception (SLMP) module and a frequency-based long-range motion perception (FLMP) module to model spatio-temporal and frequency-temporal sequences. Finally, the space and frequency motion fusion module (SFMF) integrates dual-domain features into unified motion representation. Experiments show that Vcamba outperforms state-of-the-art methods across 6 evaluation metrics on 2 datasets with lower computation cost, confirming its superiority. Code is available at: https://github.com/BoydeLi/Vcamba.

Method: MAMBO-G modulates the guidance scale based on the update-to-prediction magnitude ratio, dynamically optimizing guidance magnitudes during sampling. This stabilizes the trajectory and enables rapid convergence without requiring additional training. The method is plug-and-play with existing diffusion pipelines.

Result: Achieves up to 3x speedup on Stable Diffusion v3.5, 4x on Lumina, and 2x acceleration on the 14B-parameter Wan2.1 video model while preserving visual fidelity. The framework is implemented in mainstream open-source diffusion frameworks.

Conclusion: MAMBO-G provides a practical, training-free solution for efficient large-scale video synthesis and image generation by dynamically optimizing guidance magnitudes, significantly reducing computational costs while maintaining quality.

Abstract: High-fidelity text-to-image and text-to-video generation typically relies on Classifier-Free Guidance (CFG), but achieving optimal results often demands computationally expensive sampling schedules. In this work, we propose MAMBO-G, a training-free acceleration framework that significantly reduces computational cost by dynamically optimizing guidance magnitudes. We observe that standard CFG schedules are inefficient, applying disproportionately large updates in early steps that hinder convergence speed. MAMBO-G mitigates this by modulating the guidance scale based on the update-to-prediction magnitude ratio, effectively stabilizing the trajectory and enabling rapid convergence. This efficiency is particularly vital for resource-intensive tasks like video generation. Our method serves as a universal plug-and-play accelerator, achieving up to 3x speedup on Stable Diffusion v3.5 (SD3.5) and 4x on Lumina. Most notably, MAMBO-G accelerates the 14B-parameter Wan2.1 video model by 2x while preserving visual fidelity, offering a practical solution for efficient large-scale video synthesis. Our implementation follows a mainstream open-source diffusion framework and is plug-and-play with existing pipelines.

Method: Two core innovations: (1) Group-wise lookup-free Quantization (GQ) partitions latent features into groups for efficient quantization, overcoming memory/computation limitations; (2) Generative Decoder (GD) with prior noise variable probabilistically models visual data distribution conditioned on discrete tokens.

Result: Achieves record-low zero-shot rFID of 0.12 at high-fidelity setting (400% compression ratio), outperforming FLUX-VAE (0.18) and SD-VAE 3.5 (0.19). At high compression (768× ratio), achieves rFID 3.49, substantially better than Cosmos (4.57 at only 50% compression ratio).

Conclusion: WeTok tokenizer provides superior trade-off between compression and reconstruction fidelity, enabling better visual generation through efficient quantization and probabilistic modeling.

Abstract: Visual tokenizer is a critical component for vision generation. However, the existing tokenizers often face unsatisfactory trade-off between compression ratios and reconstruction fidelity. To fill this gap, we introduce a powerful and concise WeTok tokenizer, which surpasses the previous leading tokenizers via two core innovations. (1) Group-wise lookup-free Quantization (GQ). We partition the latent features into groups, and perform lookup-free quantization for each group. As a result, GQ can efficiently overcome memory and computation limitations of prior tokenizers, while achieving a reconstruction breakthrough with more scalable codebooks. (2) Generative Decoder (GD). Different from prior tokenizers, we introduce a generative decoder with a prior of extra noise variable. In this case, GD can probabilistically model the distribution of visual data conditioned on discrete tokens, allowing WeTok to reconstruct visual details, especially at high compression ratio. On the ImageNet 50k validation set, at a high-fidelity setting, WeTok achieves a record-low zero-shot rFID of 0.12, outperforming leading continuous tokenizers like FLUX-VAE (0.18) and SD-VAE 3.5 (0.19) with 400% compression ratio. Furthermore, in a high-compression regime, WeTok achieves a zero-shot rFID of 3.49 at a 768$\times$ compression ratio, substantially surpassing Cosmos, which scores 4.57 at only 50% our compression ratio. Code and models are available: https://github.com/zhuangshaobin/WeTok.

Method: 1) Dual-branch architecture: pretrained branch for semantically invariant representations + learnable branch for fine-grained discriminative features, merged via controllable correlation factor. 2) Virtual optimal planes: models homography decomposition coefficients estimation with iterative coefficient predictor and minimal semantic distortion constraint to find optimal alignment plane. 3) Bidirectional warping of both views onto the optimal plane.

Result: Extensive experiments across various datasets show RopStitch significantly outperforms existing methods, particularly in scene robustness and content naturalness.

Conclusion: RopStitch provides an effective unsupervised solution for image stitching that balances robustness and naturalness through innovative dual-branch feature integration and optimal plane estimation techniques.

Abstract: We present \textit{RopStitch}, an unsupervised deep image stitching framework with both robustness and naturalness. To ensure the robustness of \textit{RopStitch}, we propose to incorporate the universal prior of content perception into the image stitching model by a dual-branch architecture. It separately captures coarse and fine features and integrates them to achieve highly generalizable performance across diverse unseen real-world scenes. Concretely, the dual-branch model consists of a pretrained branch to capture semantically invariant representations and a learnable branch to extract fine-grained discriminative features, which are then merged into a whole by a controllable factor at the correlation level. Besides, considering that content alignment and structural preservation are often contradictory to each other, we propose a concept of virtual optimal planes to relieve this conflict. To this end, we model this problem as a process of estimating homography decomposition coefficients, and design an iterative coefficient predictor and minimal semantic distortion constraint to identify the optimal plane. This scheme is finally incorporated into \textit{RopStitch} by warping both views onto the optimal plane bidirectionally. Extensive experiments across various datasets demonstrate that \textit{RopStitch} significantly outperforms existing methods, particularly in scene robustness and content naturalness. The code is available at {\color{red}https://github.com/MmelodYy/RopStitch}.

Method: The paper presents a primer covering physical principles, sensor architectures, data acquisition, calibration, classical and modern analysis methods including dimensionality reduction, classification, spectral unmixing, and AI-driven deep learning techniques.

Result: A comprehensive resource summarizing HSI fundamentals, applications across Earth observation, agriculture, biomedicine, industrial inspection, cultural heritage, and security, along with analysis methods and emerging solutions.

Conclusion: HSI is evolving into a general-purpose cross-disciplinary platform with transformative potential, driven by sensor miniaturization, self-supervised learning, and foundation models for scalable, real-time systems.

Abstract: Hyperspectral imaging (HSI) is an advanced sensing modality that simultaneously captures spatial and spectral information, enabling non-invasive, label-free analysis of material, chemical, and biological properties. This Primer presents a comprehensive overview of HSI, from the underlying physical principles and sensor architectures to key steps in data acquisition, calibration, and correction. We summarize common data structures and highlight classical and modern analysis methods, including dimensionality reduction, classification, spectral unmixing, and AI-driven techniques such as deep learning. Representative applications across Earth observation, precision agriculture, biomedicine, industrial inspection, cultural heritage, and security are also discussed, emphasizing HSI’s ability to uncover sub-visual features for advanced monitoring, diagnostics, and decision-making. Persistent challenges, such as hardware trade-offs, acquisition variability, and the complexity of high-dimensional data, are examined alongside emerging solutions, including computational imaging, physics-informed modeling, cross-modal fusion, and self-supervised learning. Best practices for dataset sharing, reproducibility, and metadata documentation are further highlighted to support transparency and reuse. Looking ahead, we explore future directions toward scalable, real-time, and embedded HSI systems, driven by sensor miniaturization, self-supervised learning, and foundation models. As HSI evolves into a general-purpose, cross-disciplinary platform, it holds promise for transformative applications in science, technology, and society.

Method: YOPO extends a transformer detector with a lightweight pose head, bounding-box-conditioned translation module, and 6D-aware Hungarian matching cost. It’s trained end-to-end with only RGB images and category-level pose labels.

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

Conclusion: YOPO demonstrates that object detection and 9-DoF pose estimation can be unified with high performance using only RGB images, providing a simpler alternative to existing multi-stage or depth-dependent methods.

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

Method: Proposes SpecPrune-VLA with three components: (1) Action-level static pruning using global history and local attention to reduce visual tokens per action, (2) Layer-level dynamic pruning that adaptively prunes tokens based on layer-wise importance, (3) Lightweight action-aware controller that classifies actions as coarse- or fine-grained based on end effector speed and adjusts pruning aggressiveness accordingly.

Result: Achieves up to 1.57× speedup in LIBERO simulation and 1.70× on real-world tasks with negligible success rate degradation, significantly outperforming existing methods that cause >20% success rate drops.

Conclusion: The paper demonstrates that combining local information with global context for token selection enables effective pruning of VLA models while maintaining performance, addressing the spatial-temporal consistency in VLA tasks.

Abstract: Pruning is a typical acceleration technique for compute-bound models by removing computation on unimportant values. Recently, it has been applied to accelerate Vision-Language-Action (VLA) model inference. However, existing acceleration methods focus on local information from the current action step and ignore the global context, leading to >20% success rate drop and limited speedup in some scenarios. In this paper, we point out spatial-temporal consistency in VLA tasks: input images in consecutive steps exhibit high similarity, and propose the key insight that token selection should combine local information with global context of the model. Based on this, we propose SpecPrune-VLA, a training-free, two-level pruning method with heuristic control. (1) Action-level static pruning. We leverage global history and local attention to statically reduce visual tokens per action. (2) Layer-level dynamic pruning. We prune tokens adaptively per layer based on layer-wise importance. (3) Lightweight action-aware controller: We classify actions as coarse- or fine-grained by the speed of the end effector and adjust pruning aggressiveness accordingly. Extensive experiments show that SpecPrune-VLA achieves up to 1.57$\times$ speedup in LIBERO simulation and 1.70$\times$ on real-world tasks, with negligible success rate degradation.

Method: Proposes hybrid-plane representation combining planar and spherical planes, introduces near-equal-area warping for spherical planes to maximize feature map utilization, and synthesizes single-channel unified feature maps to eliminate feature penetration.

Result: Achieves state-of-the-art performance in full-head image synthesis, addressing previous limitations of tri-plane and spherical tri-plane representations.

Conclusion: HyPlaneHead’s hybrid representation with technical improvements enables superior 3D-aware head synthesis by systematically addressing fundamental issues in existing tri-plane methods.

Abstract: Tri-plane-like representations have been widely adopted in 3D-aware GANs for head image synthesis and other 3D object/scene modeling tasks due to their efficiency. However, querying features via Cartesian coordinate projection often leads to feature entanglement, which results in mirroring artifacts. A recent work, SphereHead, attempted to address this issue by introducing spherical tri-planes based on a spherical coordinate system. While it successfully mitigates feature entanglement, SphereHead suffers from uneven mapping between the square feature maps and the spherical planes, leading to inefficient feature map utilization during rendering and difficulties in generating fine image details. Moreover, both tri-plane and spherical tri-plane representations share a subtle yet persistent issue: feature penetration across convolutional channels can cause interference between planes, particularly when one plane dominates the others. These challenges collectively prevent tri-plane-based methods from reaching their full potential. In this paper, we systematically analyze these problems for the first time and propose innovative solutions to address them. Specifically, we introduce a novel hybrid-plane (hy-plane for short) representation that combines the strengths of both planar and spherical planes while avoiding their respective drawbacks. We further enhance the spherical plane by replacing the conventional theta-phi warping with a novel near-equal-area warping strategy, which maximizes the effective utilization of the square feature map. In addition, our generator synthesizes a single-channel unified feature map instead of multiple feature maps in separate channels, thereby effectively eliminating feature penetration. With a series of technical improvements, our hy-plane representation enables our method, HyPlaneHead, to achieve state-of-the-art performance in full-head image synthesis.

Method: The study deconstructs VLM visual processing into object recognition (analyzed via text token maps revealing two-stage perception) and spatial perception (theoretically deriving geometric structure of positional representations). Based on findings, they propose a token compression algorithm using a plug-and-play visual decoder and RoPE scaling technique.

Result: The analysis reveals that object recognition unfolds as a two-stage process from shallow to deep layers, and identifies the geometric structure underlying positional representations. The proposed methods improve decoding efficiency and enhance spatial reasoning.

Conclusion: The work provides deeper understanding of VLM internals and offers clear principles for designing more capable future architectures, bridging the gap between serial VLM processing and parallel human vision.

Abstract: Vision-Language Models (VLMs) have demonstrated remarkable performance across a variety of real-world tasks. However, existing VLMs typically process visual information by serializing images, a method that diverges significantly from the parallel nature of human vision. Moreover, their opaque internal mechanisms hinder both deeper understanding and architectural innovation. Inspired by the dual-stream hypothesis of human vision, which distinguishes the “what” and “where” pathways, we deconstruct the visual processing in VLMs into object recognition and spatial perception for separate study. For object recognition, we convert images into text token maps and find that the model’s perception of image content unfolds as a two-stage process from shallow to deep layers, beginning with attribute recognition and culminating in semantic disambiguation. For spatial perception, we theoretically derive and empirically verify the geometric structure underlying the positional representation in VLMs. Based on these findings, we introduce an instruction-agnostic token compression algorithm based on a plug-and-play visual decoder to improve decoding efficiency, and a RoPE scaling technique to enhance spatial reasoning. Through rigorous experiments, our work validates these analyses, offering a deeper understanding of VLM internals and providing clear principles for designing more capable future architectures.

Method: Proposes MedVSR with two key components: 1) Cross State-Space Propagation (CSSP) projects distant frames as control matrices within state-space models to selectively propagate consistent features for alignment, 2) Inner State-Space Reconstruction (ISSR) enhances tissue structures and reduces artifacts through joint long-range spatial feature learning and large-kernel short-range information aggregation.

Result: Experiments across four datasets in diverse medical scenarios (endoscopy and cataract surgeries) show MedVSR significantly outperforms existing VSR models in both reconstruction performance and efficiency.

Conclusion: MedVSR provides an effective solution for medical video super-resolution, addressing unique challenges in medical imaging through state-space modeling for alignment and reconstruction, with potential to improve diagnostic accuracy.

Abstract: High-resolution (HR) medical videos are vital for accurate diagnosis, yet are hard to acquire due to hardware limitations and physiological constraints. Clinically, the collected low-resolution (LR) medical videos present unique challenges for video super-resolution (VSR) models, including camera shake, noise, and abrupt frame transitions, which result in significant optical flow errors and alignment difficulties. Additionally, tissues and organs exhibit continuous and nuanced structures, but current VSR models are prone to introducing artifacts and distorted features that can mislead doctors. To this end, we propose MedVSR, a tailored framework for medical VSR. It first employs Cross State-Space Propagation (CSSP) to address the imprecise alignment by projecting distant frames as control matrices within state-space models, enabling the selective propagation of consistent and informative features to neighboring frames for effective alignment. Moreover, we design an Inner State-Space Reconstruction (ISSR) module that enhances tissue structures and reduces artifacts with joint long-range spatial feature learning and large-kernel short-range information aggregation. Experiments across four datasets in diverse medical scenarios, including endoscopy and cataract surgeries, show that MedVSR significantly outperforms existing VSR models in reconstruction performance and efficiency. Code released at https://github.com/CUHK-AIM-Group/MedVSR.

Method: End-to-end learned feature codec with bottleneck network for dimension reduction followed by multi-stage residual vector quantization (RVQ). Only per-pixel code indices are transmitted instead of full feature vectors, achieving 6-30 bits per pixel compression.

Result: Achieves 273x compression at 30 bpp to 1365x compression at 6 bpp on DAIR-V2X dataset. At 18 bpp (455x compression), matches or outperforms raw-feature collaborative perception. Enables ultra-low-bandwidth operation with graceful degradation at 6-12 bpp.

Conclusion: ReVQom enables efficient and accurate multi-agent collaborative perception with dramatic bandwidth reduction, taking a step toward practical V2X deployment while preserving spatial identity and minimizing accuracy loss.

Abstract: Multi-agent collaborative perception (CP) improves scene understanding by sharing information across connected agents such as autonomous vehicles, unmanned aerial vehicles, and robots. Communication bandwidth, however, constrains scalability. We present ReVQom, a learned feature codec that preserves spatial identity while compressing intermediate features. ReVQom is an end-to-end method that compresses feature dimensions via a simple bottleneck network followed by multi-stage residual vector quantization (RVQ). This allows only per-pixel code indices to be transmitted, reducing payloads from 8192 bits per pixel (bpp) of uncompressed 32-bit float features to 6-30 bpp per agent with minimal accuracy loss. On DAIR-V2X real-world CP dataset, ReVQom achieves 273x compression at 30 bpp to 1365x compression at 6 bpp. At 18 bpp (455x), ReVQom matches or outperforms raw-feature CP, and at 6-12 bpp it enables ultra-low-bandwidth operation with graceful degradation. ReVQom allows efficient and accurate multi-agent collaborative perception with a step toward practical V2X deployment.

Method: Proposes Vid-LLM with: 1) Cross-Task Adapter (CTA) module to align 3D geometric priors with vision-language representations, 2) Metric Depth Model to recover real-scale geometry, and 3) two-stage distillation optimization for stable training and fast convergence.

Result: Extensive experiments show effectiveness on 3D Question Answering, 3D Dense Captioning, and 3D Visual Grounding tasks, demonstrating superior multi-task capabilities compared to existing methods.

Conclusion: Vid-LLM provides a practical solution for 3D scene understanding by processing video inputs directly without external 3D data, achieving strong performance across multiple 3D reasoning tasks through effective geometric prior integration.

Abstract: Recent developments in Multimodal Large Language Models (MLLMs) have significantly improved Vision-Language (VL) reasoning in 2D domains. However, extending these capabilities to 3D scene understanding remains a major challenge. Existing 3D Multimodal Large Language Models (3D-MLLMs) often depend on 3D data inputs, which limits scalability and generalization. To address this limitation, we propose Vid-LLM, a video-based 3D-MLLM that directly processes video inputs without requiring external 3D data, making it practical for real-world deployment. In our method, the geometric prior are directly used to improve the performance of the sceen perception. To integrate the geometric cues into the MLLM compactly, we design a Cross-Task Adapter (CTA) module to align the 3D geometric priors with the vision-language representations. To ensure geometric consistency and integrity, we introduce a Metric Depth Model that recovers real-scale geometry from the reconstruction outputs. Finally, the model is fine-tuned with a two-stage distillation optimization strategy, realizing fast convergence and stabilizes training. Extensive experiments across diverse benchmarks verified the effectiveness of our method on 3D Question Answering, 3D Dense Captioning and 3D Visual Grounding tasks, demonstrating the superior multi-task capabilities.

Method: Combines coarse alignment, region-of-interest filtering, initial landmark approximation, and a patch-based pointwise CNN enhanced by attention mechanisms for localizing 50 anatomical landmarks on stereo-photogrammetry facial models.

Result: Achieved mean localization error of 3.686 mm on 214 annotated scans, preserving anatomical distances with 2.822 mm average error (comparable to intra-observer variability). On FaceScape dataset (700 subjects): 0.41 mm point-wise error and 0.38 mm distance-wise error.

Conclusion: PAL-Net offers favorable accuracy-computation trade-off, generalizes effectively across datasets and facial regions, provides lightweight scalable solution for high-throughput 3D anthropometric analysis, with potential to support clinical workflows and reduce manual annotation reliance.

Abstract: Manual annotation of anatomical landmarks on 3D facial scans is a time-consuming and expertise-dependent task, yet it remains critical for clinical assessments, morphometric analysis, and craniofacial research. While several deep learning methods have been proposed for facial landmark localization, most focus on pseudo-landmarks or require complex input representations, limiting their clinical applicability. This study presents a fully automated deep learning pipeline (PAL-Net) for localizing 50 anatomical landmarks on stereo-photogrammetry facial models. The method combines coarse alignment, region-of-interest filtering, and an initial approximation of landmarks with a patch-based pointwise CNN enhanced by attention mechanisms. Trained and evaluated on 214 annotated scans from healthy adults, PAL-Net achieved a mean localization error of 3.686 mm and preserves relevant anatomical distances with a 2.822 mm average error, comparable to intra-observer variability. To assess generalization, the model was further evaluated on 700 subjects from the FaceScape dataset, achieving a point-wise error of 0.41,mm and a distance-wise error of 0.38,mm. Compared to existing methods, PAL-Net offers a favorable trade-off between accuracy and computational cost. While performance degrades in regions with poor mesh quality (e.g., ears, hairline), the method demonstrates consistent accuracy across most anatomical regions. PAL-Net generalizes effectively across datasets and facial regions, outperforming existing methods in both point-wise and structural evaluations. It provides a lightweight, scalable solution for high-throughput 3D anthropometric analysis, with potential to support clinical workflows and reduce reliance on manual annotation. Source code can be found at https://github.com/Ali5hadman/PAL-Net-A-Point-Wise-CNN-with-Patch-Attention

Method: Proposes 3D Rasterization: lightweight rasterization of annotated primitives instead of costly rendering, enabling augmentations like counterfactual recovery maneuvers and cross-agent view synthesis. Introduces Raster-to-Real feature-space alignment to bridge sim-to-real gap. Together forms RAP pipeline.

Result: Achieves state-of-the-art closed-loop robustness and long-tail generalization, ranking first on four major benchmarks: NAVSIM v1/v2, Waymo Open Dataset Vision-based E2E Driving, and Bench2Drive.

Conclusion: Lightweight rasterization with feature alignment suffices to scale E2E training, offering a practical alternative to photorealistic rendering for autonomous driving planning.

Abstract: Imitation learning for end-to-end driving trains policies only on expert demonstrations. Once deployed in a closed loop, such policies lack recovery data: small mistakes cannot be corrected and quickly compound into failures. A promising direction is to generate alternative viewpoints and trajectories beyond the logged path. Prior work explores photorealistic digital twins via neural rendering or game engines, but these methods are prohibitively slow and costly, and thus mainly used for evaluation. In this work, we argue that photorealism is unnecessary for training end-to-end planners. What matters is semantic fidelity and scalability: driving depends on geometry and dynamics, not textures or lighting. Motivated by this, we propose 3D Rasterization, which replaces costly rendering with lightweight rasterization of annotated primitives, enabling augmentations such as counterfactual recovery maneuvers and cross-agent view synthesis. To transfer these synthetic views effectively to real-world deployment, we introduce a Raster-to-Real feature-space alignment that bridges the sim-to-real gap. Together, these components form Rasterization Augmented Planning (RAP), a scalable data augmentation pipeline for planning. RAP achieves state-of-the-art closed-loop robustness and long-tail generalization, ranking first on four major benchmarks: NAVSIM v1/v2, Waymo Open Dataset Vision-based E2E Driving, and Bench2Drive. Our results show that lightweight rasterization with feature alignment suffices to scale E2E training, offering a practical alternative to photorealistic rendering. Project page: https://alan-lanfeng.github.io/RAP/.

Method: Pipeline projects 3D points to 2D spherical grid, enriches pixels with multi-source features, trains ensemble of segmentation networks to produce pseudo-labels and uncertainty maps, uses uncertainty to guide annotation of ambiguous regions, then back-projects to 3D with three-tier visualization suite.

Result: Performance saturates after ~12 annotated scans, geometric features contribute most, compact nine-channel stacks capture nearly all discriminative power with mIoU plateauing around 0.76. Pipeline generalizes to other datasets (ForestSemantic, Semantic3D).

Conclusion: Proposed pipeline enables scalable, high-quality segmentation of TLS point clouds with reduced labeling effort, provides empirical guidance on data efficiency and feature importance, and contributes Mangrove3D dataset for ecological monitoring applications.

Abstract: Accurate semantic segmentation of terrestrial laser scanning (TLS) point clouds is limited by costly manual annotation. We propose a semi-automated, uncertainty-aware pipeline that integrates spherical projection, feature enrichment, ensemble learning, and targeted annotation to reduce labeling effort, while sustaining high accuracy. Our approach projects 3D points to a 2D spherical grid, enriches pixels with multi-source features, and trains an ensemble of segmentation networks to produce pseudo-labels and uncertainty maps, the latter guiding annotation of ambiguous regions. The 2D outputs are back-projected to 3D, yielding densely annotated point clouds supported by a three-tier visualization suite (2D feature maps, 3D colorized point clouds, and compact virtual spheres) for rapid triage and reviewer guidance. Using this pipeline, we build Mangrove3D, a semantic segmentation TLS dataset for mangrove forests. We further evaluate data efficiency and feature importance to address two key questions: (1) how much annotated data are needed and (2) which features matter most. Results show that performance saturates after ~12 annotated scans, geometric features contribute the most, and compact nine-channel stacks capture nearly all discriminative power, with the mean Intersection over Union (mIoU) plateauing at around 0.76. Finally, we confirm the generalization of our feature-enrichment strategy through cross-dataset tests on ForestSemantic and Semantic3D. Our contributions include: (i) a robust, uncertainty-aware TLS annotation pipeline with visualization tools; (ii) the Mangrove3D dataset; and (iii) empirical guidance on data efficiency and feature importance, thus enabling scalable, high-quality segmentation of TLS point clouds for ecological monitoring and beyond. The dataset and processing scripts are publicly available at https://fz-rit.github.io/through-the-lidars-eye/.

Method: Proposes Memory-Augmented Reinforcement Learning-based Token Compression (MARC) with a retrieve-then-compress strategy: 1) Visual Memory Retriever (VMR) selects key clips, 2) Compression Group Relative Policy Optimization (C-GRPO) distills reasoning ability from teacher to student model using reinforcement learning.

Result: Achieves near-baseline accuracy using only one frame’s tokens, reducing visual tokens by 95%, GPU memory by 72%, and latency by 23.9% across six video benchmarks.

Conclusion: MARC demonstrates potential for efficient, real-time video understanding in resource-constrained applications like video QA, surveillance, and autonomous driving through effective token compression.

Abstract: The rapid progress of large language models (LLMs) has laid the foundation for multimodal models. However, visual language models (VLMs) still face heavy computational costs when extended from images to videos due to high frame rates and long durations. Token compression is a promising solution, yet most existing training-free methods cause information loss and performance degradation. To overcome this, we propose \textbf{Memory-Augmented Reinforcement Learning-based Token Compression (MARC)}, which integrates structured retrieval and RL-based distillation. MARC adopts a \textit{retrieve-then-compress} strategy using a \textbf{Visual Memory Retriever (VMR)} to select key clips and a \textbf{Compression Group Relative Policy Optimization (C-GRPO)} framework to distil reasoning ability from a teacher to a student model. Experiments on six video benchmarks show that MARC achieves near-baseline accuracy using only one frame’s tokens – reducing visual tokens by \textbf{95%}, GPU memory by \textbf{72%}, and latency by \textbf{23.9%}. This demonstrates its potential for efficient, real-time video understanding in resource-constrained settings such as video QA, surveillance, and autonomous driving.

Method: Proposes REAS-Det with Global-Selective Visibility Convolution (GSV-Conv) to expand visible feature space under global semantic guidance, plus C3RFEM, MultiSEAM, and Soft-NMS modules for refined separation and localization. Uses STP-AgriData dataset combining real-orchard imagery with curated web images and contrast/brightness augmentations.

Result: Achieves 86.5% precision, 77.2% recall, 84.3% mAP@0.50, and 53.6% mAP@0.50:0.95 on STP-AgriData, outperforming recent detectors and improving robustness in real orchard environments.

Conclusion: REAS-Det effectively addresses practical challenges in agricultural object detection, particularly for pomelo detection in orchards, demonstrating superior performance and robustness compared to existing methods.

Abstract: Shatian pomelo detection in orchards is essential for yield estimation and lean production, but models tuned to ideal datasets often degrade in practice due to device-dependent tone shifts, illumination changes, large scale variation, and frequent occlusion. We introduce STP-AgriData, a multi-scenario dataset combining real-orchard imagery with curated web images, and apply contrast/brightness augmentations to emulate unstable lighting. To better address scale and occlusion, we propose REAS-Det, featuring Global-Selective Visibility Convolution (GSV-Conv) that expands the visible feature space under global semantic guidance while retaining efficient spatial aggregation, plus C3RFEM, MultiSEAM, and Soft-NMS for refined separation and localization. On STP-AgriData, REAS-Det achieves 86.5% precision, 77.2% recall, 84.3% mAP@0.50, and 53.6% mAP@0.50:0.95, outperforming recent detectors and improving robustness in real orchard environments. The source code is available at: https://github.com/Genk641/REAS-Det.

Method: The approach colorizes a single key view, then fine-tunes a personalized colorizer to propagate colors to novel views and time steps. The colorizer learns scene-specific deterministic color mapping from the reference view and projects consistent colors to other images using its inductive bias. A Lab color space Gaussian splatting representation enables direct reconstruction of colorful 3D scenes.

Result: Extensive experiments across diverse static and dynamic 3D colorization benchmarks show that Color3D delivers more consistent and chromatically rich renderings with precise user control compared to existing methods.

Conclusion: Color3D successfully recasts complex 3D colorization as a more tractable single image paradigm, enabling seamless integration of arbitrary image colorization models with enhanced flexibility and controllability for both static and dynamic 3D scenes.

Abstract: In this work, we present Color3D, a highly adaptable framework for colorizing both static and dynamic 3D scenes from monochromatic inputs, delivering visually diverse and chromatically vibrant reconstructions with flexible user-guided control. In contrast to existing methods that focus solely on static scenarios and enforce multi-view consistency by averaging color variations which inevitably sacrifice both chromatic richness and controllability, our approach is able to preserve color diversity and steerability while ensuring cross-view and cross-time consistency. In particular, the core insight of our method is to colorize only a single key view and then fine-tune a personalized colorizer to propagate its color to novel views and time steps. Through personalization, the colorizer learns a scene-specific deterministic color mapping underlying the reference view, enabling it to consistently project corresponding colors to the content in novel views and video frames via its inherent inductive bias. Once trained, the personalized colorizer can be applied to infer consistent chrominance for all other images, enabling direct reconstruction of colorful 3D scenes with a dedicated Lab color space Gaussian splatting representation. The proposed framework ingeniously recasts complicated 3D colorization as a more tractable single image paradigm, allowing seamless integration of arbitrary image colorization models with enhanced flexibility and controllability. Extensive experiments across diverse static and dynamic 3D colorization benchmarks substantiate that our method can deliver more consistent and chromatically rich renderings with precise user control. Project Page https://yecongwan.github.io/Color3D/.

Method: Proposes a face similarity metric that breaks down global similarity into contributions from restricted receptive fields, defining similarity as sum of patch-level similarity scores for locally additive explanations.

Result: Achieves competitive verification performance with 28x28 patches within 112x112 face images, and surpasses state-of-the-art methods with 56x56 patches.

Conclusion: The proposed inherently interpretable approach provides locally additive explanations without post-hoc analysis while maintaining competitive performance in face verification.

Abstract: Understanding how deep neural networks make decisions is crucial for analyzing their behavior and diagnosing failure cases. In computer vision, a common approach to improve interpretability is to assign importance to individual pixels using post-hoc methods. Although they are widely used to explain black-box models, their fidelity to the model’s actual reasoning is uncertain due to the lack of reliable evaluation metrics. This limitation motivates an alternative approach, which is to design models whose decision processes are inherently interpretable. To this end, we propose a face similarity metric that breaks down global similarity into contributions from restricted receptive fields. Our method defines the similarity between two face images as the sum of patch-level similarity scores, providing a locally additive explanation without relying on post-hoc analysis. We show that the proposed approach achieves competitive verification performance even with patches as small as 28x28 within 112x112 face images, and surpasses state-of-the-art methods when using 56x56 patches.

Method: Proposes InternSVG family with three components: 1) SAgoge dataset (largest comprehensive multimodal SVG dataset), 2) SArena benchmark (standardized evaluation), and 3) InternSVG model with SVG-specific tokens, subword-based embedding initialization, and two-stage training from short static to complex animations.

Result: InternSVG achieves substantial gains and consistently outperforms leading open and proprietary counterparts on SArena and prior benchmarks, demonstrating positive transfer and improved overall performance.

Conclusion: The unified MLLM approach successfully addresses SVG modeling challenges, enabling effective understanding, editing, and generation of both static and dynamic SVG content through comprehensive data, benchmark, and model design.

Abstract: General SVG modeling remains challenging due to fragmented datasets, limited transferability of methods across tasks, and the difficulty of handling structural complexity. In response, we leverage the strong transfer and generalization capabilities of multimodal large language models (MLLMs) to achieve unified modeling for SVG understanding, editing, and generation. We present the InternSVG family, an integrated data-benchmark-model suite. At its core is SAgoge, the largest and most comprehensive multimodal dataset for SVG tasks, encompassing both static graphics and dynamic animations. It covers icons, long-sequence illustrations, scientific diagrams, and dynamic animations, supporting tasks of varied difficulty levels and providing deeper hierarchies with richer attributes compared to previous datasets. Based on this resource, we introduce SArena, a companion benchmark with comprehensive task definitions and standardized evaluation that aligns with the domains and difficulty spectrum covered by SAgoge. Building on these foundations, we propose InternSVG, a unified MLLM for SVG understanding, editing, and generation with SVG-specific special tokens, subword-based embedding initialization, and a two-stage training strategy that progresses from short static SVGs to long-sequence illustrations and complex animations. This unified formulation induces positive transfer and improves overall performance. Experiments on SArena and prior benchmark confirm that InternSVG achieves substantial gains and consistently outperforms leading open and proprietary counterparts.

Method: Proposes MS-Mix with three key components: 1) Sentiment-Aware Sample Selection to prevent mixing contradictory emotions, 2) Sentiment Intensity Guided module using multi-head self-attention to compute modality-specific mixing ratios based on emotional intensities, and 3) Sentiment Alignment Loss with KL-divergence regularization to align predictions across modalities.

Result: Extensive experiments on three benchmark datasets with six state-of-the-art backbones show MS-Mix consistently outperforms existing methods, establishing new standards for robust multimodal sentiment augmentation.

Conclusion: MS-Mix effectively addresses the limitations of traditional augmentation in multimodal sentiment analysis by introducing emotion-aware mechanisms, leading to improved generalization and performance across various backbone architectures.

Abstract: Multimodal Sentiment Analysis (MSA) aims to identify and interpret human emotions by integrating information from heterogeneous data sources such as text, video, and audio. While deep learning models have advanced in network architecture design, they remain heavily limited by scarce multimodal annotated data. Although Mixup-based augmentation improves generalization in unimodal tasks, its direct application to MSA introduces critical challenges: random mixing often amplifies label ambiguity and semantic inconsistency due to the lack of emotion-aware mixing mechanisms. To overcome these issues, we propose MS-Mix, an adaptive, emotion-sensitive augmentation framework that automatically optimizes sample mixing in multimodal settings. The key components of MS-Mix include: (1) a Sentiment-Aware Sample Selection (SASS) strategy that effectively prevents semantic confusion caused by mixing samples with contradictory emotions. (2) a Sentiment Intensity Guided (SIG) module using multi-head self-attention to compute modality-specific mixing ratios dynamically based on their respective emotional intensities. (3) a Sentiment Alignment Loss (SAL) that aligns the prediction distributions across modalities, and incorporates the Kullback-Leibler-based loss as an additional regularization term to train the emotion intensity predictor and the backbone network jointly. Extensive experiments on three benchmark datasets with six state-of-the-art backbones confirm that MS-Mix consistently outperforms existing methods, establishing a new standard for robust multimodal sentiment augmentation. The source code is available at: https://github.com/HongyuZhu-s/MS-Mix.

Method: Developed Paper Copilot system that creates durable digital archives of peer reviews across computer-science venues, providing an open dataset for studying peer review at scale, with empirical analysis of ICLR reviews spanning multiple years.

Result: Created infrastructure and dataset supporting reproducible research on peer review evolution, enabling community to track changes, diagnose failure modes, and inform evidence-based improvements for more robust review systems.

Conclusion: Paper Copilot provides essential resources for studying peer review at scale, helping the research community move toward more transparent, reliable, and evidence-based review systems in AI conferences.

Abstract: The rapid growth of AI conferences is straining an already fragile peer-review system, leading to heavy reviewer workloads, expertise mismatches, inconsistent evaluation standards, superficial or templated reviews, and limited accountability under compressed timelines. In response, conference organizers have introduced new policies and interventions to preserve review standards. Yet these ad-hoc changes often create further concerns and confusion about the review process, leaving how papers are ultimately accepted - and how practices evolve across years - largely opaque. We present Paper Copilot, a system that creates durable digital archives of peer reviews across a wide range of computer-science venues, an open dataset that enables researchers to study peer review at scale, and a large-scale empirical analysis of ICLR reviews spanning multiple years. By releasing both the infrastructure and the dataset, Paper Copilot supports reproducible research on the evolution of peer review. We hope these resources help the community track changes, diagnose failure modes, and inform evidence-based improvements toward a more robust, transparent, and reliable peer-review system.

Method: IR-WM first creates a BEV representation of current state from visual input. It uses previous timestep’s BEV features as temporal prior and predicts only residual changes conditioned on ego-vehicle actions and scene context. An alignment module addresses error accumulation, and different forecasting-planning coupling schemes are explored.

Result: On nuScenes benchmark, IR-WM achieves top performance in both 4D occupancy forecasting and trajectory planning, demonstrating improved planning accuracy through implicit future state generation.

Conclusion: Focusing on residual modeling of state changes rather than full scene reconstruction leads to more efficient world models that significantly improve autonomous driving planning performance.

Abstract: End-to-end autonomous driving systems increasingly rely on vision-centric world models to understand and predict their environment. However, a common ineffectiveness in these models is the full reconstruction of future scenes, which expends significant capacity on redundantly modeling static backgrounds. To address this, we propose IR-WM, an Implicit Residual World Model that focuses on modeling the current state and evolution of the world. IR-WM first establishes a robust bird’s-eye-view representation of the current state from the visual observation. It then leverages the BEV features from the previous timestep as a strong temporal prior and predicts only the “residual”, i.e., the changes conditioned on the ego-vehicle’s actions and scene context. To alleviate error accumulation over time, we further apply an alignment module to calibrate semantic and dynamic misalignments. Moreover, we investigate different forecasting-planning coupling schemes and demonstrate that the implicit future state generated by world models substantially improves planning accuracy. On the nuScenes benchmark, IR-WM achieves top performance in both 4D occupancy forecasting and trajectory planning.

Method: Two-stage training: (1) train spectral enhancement network via reverse distillation for robust spectral representations, (2) freeze spectral branch as teacher and train spatial branch as “rebellious student” with decorrelation objectives to capture overlooked spatial patterns. Uses reconstruction regularization to prevent noise learning.

Result: Experiments on HAD100 benchmark show substantial improvements over baselines with modest computational overhead. Enhanced features can be plugged into parameter-free detectors like Reed-Xiaoli (RX) detector.

Conclusion: The complementary learning paradigm effectively enhances background representations for hyperspectral anomaly detection, enabling efficient universal deployment without per-scene retraining.

Abstract: A recent class of hyperspectral anomaly detection methods can be trained once on background datasets and then deployed universally without per-scene retraining or parameter tuning, showing strong efficiency and robustness. Building upon this paradigm, we propose a decoupled complementary spectral–spatial learning framework for background representation enhancement. The framework follows a two-stage training strategy: (1) we first train a spectral enhancement network via reverse distillation to obtain robust background spectral representations; and (2) we then freeze the spectral branch as a teacher and train a spatial branch as a complementary student (the “rebellious student”) to capture spatial patterns overlooked by the teacher. Complementary learning is achieved through decorrelation objectives that reduce representational redundancy between the two branches, together with reconstruction regularization to prevent the student from learning irrelevant noise. After training, the framework jointly enhances background representations from both spectral and spatial perspectives, and the resulting enhanced features can be plugged into parameter-free, training-free detectors (e.g., the Reed–Xiaoli (RX) detector) for test-time deployment without per-scene retraining or parameter tuning. Experiments on the HAD100 benchmark demonstrate substantial improvements over representative baselines with modest computational overhead, validating the effectiveness of the proposed complementary learning paradigm. Our code is publicly available at https://github.com/xjpp2016/FERS.

Method: Proposes a hybrid Chebyshev ensemble geometric network (Ch-EGN) that combines convolutional and geometric deep learning to leverage information dependencies in supply chain data and derive invisible states of samples.

Result: Achieved 98.95% average accuracy for risk management (delivery status prediction), 100% for 5-product group classification, 98.07% for 4-product relation classification, and 92.37% for 25-company relation classification on SupplyGraph and DataCo datasets.

Conclusion: The proposed Ch-EGN ensemble network demonstrates superior performance in supply chain sustainability analysis, outperforming state-of-the-art approaches in risk management and classification tasks.

Abstract: The sustainability of supply chain plays a key role in achieving optimal performance in controlling the supply chain. The management of risks that occur in a supply chain is a fundamental problem for the purpose of developing the sustainability of the network and elevating the performance efficiency of the supply chain. The correct classification of products is another essential element in a sustainable supply chain. Acknowledging recent breakthroughs in the context of deep networks, several architectural options have been deployed to analyze supply chain datasets. A novel geometric deep network is used to propose an ensemble deep network. The proposed Chebyshev ensemble geometric network (Ch-EGN) is a hybrid convolutional and geometric deep learning. This network is proposed to leverage the information dependencies in supply chain to derive invisible states of samples in the database. The functionality of the proposed deep network is assessed on the two different databases. The SupplyGraph Dataset and DataCo are considered in this research. The prediction of delivery status of DataCo supply chain is done for risk administration. The product classification and edge classification are performed using the SupplyGraph database to enhance the sustainability of the supply network. An average accuracy of 98.95% is obtained for the ensemble network for risk management. The average accuracy of 100% and 98.07% are obtained for sustainable supply chain in terms of 5 product group classification and 4 product relation classification, respectively. The average accuracy of 92.37% is attained for 25 company relation classification. The results confirm an average improvement and efficiency of the proposed method compared to the state-of-the-art approaches.

Method: Presents PALM dataset with 13k hand scans from 263 subjects and 90k multi-view images capturing diverse skin tones, ages, and geometries. Introduces PALM-Net baseline model that learns multi-subject prior over hand geometry and material properties via physically based inverse rendering.

Result: PALM provides a large-scale, diverse real-world resource for hand modeling. PALM-Net enables realistic, relightable single-image hand avatar personalization by leveraging the learned prior over hand geometry and material properties.

Conclusion: PALM dataset addresses the critical need for diverse, high-quality hand data, while PALM-Net demonstrates practical utility for single-image hand avatar creation, advancing research in personalized hand modeling and related areas.

Abstract: The ability to grasp objects, signal with gestures, and share emotion through touch all stem from the unique capabilities of human hands. Yet creating high-quality personalized hand avatars from images remains challenging due to complex geometry, appearance, and articulation, particularly under unconstrained lighting and limited views. Progress has also been limited by the lack of datasets that jointly provide accurate 3D geometry, high-resolution multiview imagery, and a diverse population of subjects. To address this, we present PALM, a large-scale dataset comprising 13k high-quality hand scans from 263 subjects and 90k multi-view images, capturing rich variation in skin tone, age, and geometry. To show its utility, we present a baseline PALM-Net, a multi-subject prior over hand geometry and material properties learned via physically based inverse rendering, enabling realistic, relightable single-image hand avatar personalization. PALM’s scale and diversity make it a valuable real-world resource for hand modeling and related research.

Method: Proposes the Mental Bayesian Causal World Model (MBCWM) instantiated as Tokenized Intent World Model (TIWM), a cognitive computing architecture that synthesizes von Uexküll’s Umwelt theory, neural assembly hypothesis, and triple causal model (symbolic deduction, probabilistic induction, force dynamics) into an end-to-end embodied planning system. Features Belief-Intent Co-Evolution mechanism and demonstrates on nuPlan benchmark.

Result: Experimental results in open-loop validation show the Belief-Intent Co-Evolution mechanism effectively enhances planning performance. In closed-loop simulations, the system exhibits emergent human-like cognitive behaviors including map affordance understanding, free exploration, and self-recovery strategies. Cognitive consistency emerges as the core learning mechanism where belief and intent spontaneously form self-organizing equilibrium through implicit computational replay.

Conclusion: TIWM offers a neuro-symbolic, cognition-first alternative to reconstruction-based planners, establishing a new direction: planning as active understanding rather than passive reaction. It demonstrates that high-performance planning can emerge from cognitive consistency rather than high-fidelity world reconstruction.

Abstract: This paper challenges a prevailing epistemological assumption in End-to-End Autonomous Driving: that high-performance planning necessitates high-fidelity world reconstruction. Inspired by cognitive science, we propose the Mental Bayesian Causal World Model (MBCWM) and instantiate it as the Tokenized Intent World Model (TIWM), a novel cognitive computing architecture. Its core philosophy posits that intelligence emerges not from pixel-level objective fidelity, but from the Cognitive Consistency between the agent’s internal intentional world and physical reality. By synthesizing von Uexküll’s $\textit{Umwelt}$ theory, the neural assembly hypothesis, and the triple causal model (integrating symbolic deduction, probabilistic induction, and force dynamics) into an end-to-end embodied planning system, we demonstrate the feasibility of this paradigm on the nuPlan benchmark. Experimental results in open-loop validation confirm that our Belief-Intent Co-Evolution mechanism effectively enhances planning performance. Crucially, in closed-loop simulations, the system exhibits emergent human-like cognitive behaviors, including map affordance understanding, free exploration, and self-recovery strategies. We identify Cognitive Consistency as the core learning mechanism: during long-term training, belief (state understanding) and intent (future prediction) spontaneously form a self-organizing equilibrium through implicit computational replay, achieving semantic alignment between internal representations and physical world affordances. TIWM offers a neuro-symbolic, cognition-first alternative to reconstruction-based planners, establishing a new direction: planning as active understanding, not passive reaction.

Method: Proposes an Explainable Cross-Disease Reasoning Framework with three components: 1) Pulmonary Perception Module summarizing lung abnormalities, 2) Agentic Pulmonary-to-Cardiac Reasoning Module inferring cardiovascular implications using medical knowledge, and 3) Cardiac Feature Extractor encoding structural biomarkers. These outputs are fused for holistic cardiovascular risk prediction with natural-language rationales.

Result: On NLST cohort, achieves state-of-the-art performance for CVD screening (AUC=0.919) and mortality prediction (AUC=0.838), outperforming single-disease and purely image-based baselines. Provides human-verifiable reasoning aligning with cardiological understanding.

Conclusion: Establishes a unified, explainable paradigm for cardiovascular analysis from LDCT, bridging image-based prediction with mechanism-based medical interpretation through interpretable cross-disease reasoning.

Abstract: Low-dose chest computed tomography (LDCT) inherently captures both pulmonary and cardiac structures, offering a unique opportunity for joint assessment of lung and cardiovascular health. However, most existing approaches treat these domains as independent tasks, overlooking their physiological interplay and shared imaging biomarkers. We propose an Explainable Cross-Disease Reasoning Framework that enables interpretable cardiopulmonary risk assessment from a single LDCT scan. The framework introduces an agentic reasoning process that emulates clinical diagnostic thinking: first perceiving pulmonary findings, then reasoning through established medical knowledge, and finally deriving a cardiovascular judgment with a natural-language rationale. It integrates three components: a Pulmonary Perception Module that summarizes lung abnormalities, an Agentic Pulmonary-to-Cardiac Reasoning Module that infers their cardiovascular implications, and a Cardiac Feature Extractor that encodes structural biomarkers. Their outputs are fused to produce a holistic cardiovascular risk prediction that is both accurate and physiologically grounded. Experiments on the NLST cohort demonstrate that the proposed framework achieves state-of-the-art performance for CVD screening (AUC=0.919) and mortality prediction (AUC=0.838), outperforming single-disease and purely image-based baselines. Beyond quantitative gains, the framework provides human-verifiable reasoning that aligns with cardiological understanding, revealing coherent links between pulmonary abnormalities and cardiac stress mechanisms. Overall, this work establishes a unified and explainable paradigm for cardiovascular analysis from LDCT, bridging the gap between image-based prediction and mechanism-based medical interpretation.

Method: 1) Identifies three core tasks: egocentric procedural error detection, learning, and question answering; 2) Introduces two enabling dimensions: real-time streaming video understanding and proactive interaction; 3) Creates taxonomy and comprehensive review; 4) Conducts novel experiments evaluating representative VLM-based methods.

Result: Provides comprehensive taxonomy and framework for EgoProceAssist, identifies gaps between proposed assistant capabilities and current VLM methods, and establishes evaluation benchmarks through experiments.

Conclusion: EgoProceAssist represents an emerging research direction with significant potential for real-world daily activity assistance, but current VLM methods have limitations that need to be addressed through future research in multimodal understanding and interaction.

Abstract: Driven by recent advances in vision-language models (VLMs) and egocentric perception research, the emerging topic of an egocentric procedural AI assistant (EgoProceAssist) is introduced to step-by-step support daily procedural tasks in a first-person view. In this paper, we start by identifying three core tasks in EgoProceAssist: egocentric procedural error detection, egocentric procedural learning, and egocentric procedural question answering, then introduce two enabling dimensions: real-time and streaming video understanding, and proactive interaction in procedural contexts. We define these tasks within a new taxonomy as the EgoProceAssist’s essential functions and illustrate how they can be deployed in real-world scenarios for daily activity assistants. Specifically, our work encompasses a comprehensive review of current techniques, relevant datasets, and evaluation metrics across these five core areas. To clarify the gap between the proposed EgoProceAssist and existing VLM-based assistants, we conduct novel experiments to provide a comprehensive evaluation of representative VLM-based methods. Through these findings and our technical analysis, we discuss the challenges ahead and suggest future research directions. Furthermore, an exhaustive list of this study is publicly available in an active repository that continuously collects the latest work: https://github.com/z1oong/Building-Egocentric-Procedural-AI-Assistant.

Method: Proposes LookSharp, which minimizes the entropy of CLS-to-patch attention in the final transformer layer during test-time adaptation. This encourages the model to maintain focused attention on shifted data, complementing traditional output entropy minimization.

Result: Attention entropy minimization improves robustness on ImageNet-C distribution shifts. The method is complementary to output entropy minimization and maintains performance on clean data.

Conclusion: Minimizing attention entropy in transformers provides an effective test-time adaptation objective that enhances robustness to distribution shifts while preserving clean data performance, offering a novel approach beyond traditional output-based adaptation.

Abstract: Test-time adaptation (TTA) updates models during inference to reduce error on distribution shifts. While entropy minimization over the output distribution has proven effective as a TTA loss, we study using the intermediate distributions computed by transformers in the attention mechanism. We propose LookSharp, which minimizes the entropy of CLS-to-patch attention in the final layer as a novel TTA objective, encouraging the model to maintain focused attention on shifted data. We demonstrate that attention entropy minimization improves robustness on ImageNet-C. We also show that it is complementary to output entropy minimization and maintains performance on clean data.

Method: Use AutoML pipelines to extract important features from triaxial accelerometer data segments, train multiple lightweight ML algorithms on extracted features, implement on WeBe Band wearable device with capable MCU for on-device processing

Result: Neural network provided best balance between accuracy, latency, and memory use; reliable real-time gesture recognition achieved on WeBe Band with potential for secure, fast medical monitoring solutions

Conclusion: Efficient motion-based models using only accelerometer sensors can enable reliable real-time gesture recognition on embedded devices for medical monitoring applications

Abstract: The use of tiny devices capable of low-latency gesture recognition is gaining momentum in everyday human-computer interaction and especially in medical monitoring fields. Embedded solutions such as fall detection, rehabilitation tracking, and patient supervision require fast and efficient tracking of movements while avoiding unwanted false alarms. This study presents an efficient solution on how to build very efficient motion-based models only using triaxial accelerometer sensors. We explore the capability of the AutoML pipelines to extract the most important features from the data segments. This approach also involves training multiple lightweight machine learning algorithms using the extracted features. We use WeBe Band, a multi-sensor wearable device that is equipped with a powerful enough MCU to effectively perform gesture recognition entirely on the device. Of the models explored, we found that the neural network provided the best balance between accuracy, latency, and memory use. Our results also demonstrate that reliable real-time gesture recognition can be achieved in WeBe Band, with great potential for real-time medical monitoring solutions that require a secure and fast response time.

Method: Introduces RLWG framework using geometric and perceptual rewards (pose cycle-consistency, depth reprojection, temporal coherence) with GrndCtrl adaptation method based on Group Relative Policy Optimization (GRPO) for post-training alignment.

Result: Achieves superior spatial coherence and navigation stability over supervised fine-tuning in outdoor environments, maintaining stable trajectories and consistent geometry.

Conclusion: RLWG bridges generative pretraining and grounded behavior through verifiable rewards, similar to post-training alignment in language models, enabling geometrically grounded world models for embodied navigation.

Abstract: Recent advances in video world modeling have enabled large-scale generative models to simulate embodied environments with high visual fidelity, providing strong priors for prediction, planning, and control. Yet, despite their realism, these models often lack geometric grounding, limiting their use in navigation tasks that require spatial coherence and stability. We introduce Reinforcement Learning with World Grounding (RLWG), a self-supervised post-training framework that aligns pretrained world models with a physically verifiable structure through geometric and perceptual rewards. Analogous to reinforcement learning from verifiable feedback (RLVR) in language models, RLWG can use multiple rewards that measure pose cycle-consistency, depth reprojection, and temporal coherence. We instantiate this framework with GrndCtrl, a reward-aligned adaptation method based on Group Relative Policy Optimization (GRPO), yielding world models that maintain stable trajectories, consistent geometry, and reliable rollouts for embodied navigation. Like post-training alignment in large language models, GrndCtrl leverages verifiable rewards to bridge generative pretraining and grounded behavior, achieving superior spatial coherence and navigation stability over supervised fine-tuning in outdoor environments.

Method: Proposes Frequency Representation Learning (FRL) - an architecture-agnostic approach using resolution-aligned frequency representations and spectral consistency training to stabilize frequency response across different Nyquist frequencies.

Result: FRL-enhanced variants show more stable frequency response in high-frequency bands on grids with higher Nyquist frequencies, allowing errors to decrease with resolution and significantly outperform baselines with modest computational overhead.

Conclusion: Identifies Scale Anchoring as a distinct problem from existing issues and demonstrates that FRL effectively addresses it by enabling proper error reduction with increasing resolution in spatiotemporal forecasting tasks.

Abstract: Zero-Shot Super-Resolution Spatiotemporal Forecasting requires a deep learning model to be trained on low-resolution data and deployed for inference on high-resolution. Existing studies consider maintaining similar error across different resolutions as indicative of successful multi-resolution generalization. However, deep learning models serving as alternatives to numerical solvers should reduce error as resolution increases. The fundamental limitation is, the upper bound of physical law frequencies that low-resolution data can represent is constrained by its Nyquist frequency, making it difficult for models to process signals containing unseen frequency components during high-resolution inference. This results in errors being anchored at low resolution, incorrectly interpreted as successful generalization. We define this fundamental phenomenon as a new problem distinct from existing issues: Scale Anchoring. Therefore, we propose architecture-agnostic Frequency Representation Learning. It alleviates Scale Anchoring through resolution-aligned frequency representations and spectral consistency training: on grids with higher Nyquist frequencies, the frequency response in high-frequency bands of FRL-enhanced variants is more stable. This allows errors to decrease with resolution and significantly outperform baselines within our task and resolution range, while incurring only modest computational overhead.

Method: Developed and validated an AI-ECG model using CCTA as the reference standard to estimate severe (≥70%) or complete (≥99%) stenosis in the four major coronary arteries. Conducted internal and external validation, and evaluated performance in clinical cohorts with longitudinal follow-up.

Result: Model achieved AUC values of 0.706-0.744 in internal validation and 0.673-0.714 in external validation. Performance remained stable across subgroups including those with clinically normal ECGs. Vessel-specific risk stratification showed distinct separation between high-risk and low-risk groups in time-to-event analyses.

Conclusion: AI-ECG shows feasibility as a complementary CAD risk screening tool with potential clinical utility. The approach warrants prospective evaluation and could provide a more accessible alternative to resource-intensive imaging modalities.

Abstract: Coronary artery disease (CAD) remains a major global public health burden, yet scalable tools for risk screening are limited. Although coronary computed tomography angiography (CCTA) is a first-line non-invasive diagnostic modality, its widespread use is constrained by resource requirements and radiation exposure. Artificial intelligence–enabled electrocardiography (AI-ECG) may provide a complementary approach for CAD risk stratification. We developed and validated an AI-ECG model using CCTA as the reference standard to estimate severe ($\geq 70%$) or complete ($\geq 99%$) stenosis in the four major coronary arteries. In internal validation, the model achieved area under the receiver operating characteristic curve (AUC) values of 0.706–0.744 across vessels and demonstrated consistent performance in external validation (AUCs: 0.673–0.714). Discrimination remained stable among individuals with clinically normal ECGs and across demographic and clinical subgroups. In a dedicated clinical cohort with longitudinal follow-up, vessel-specific risk stratification based on model-predicted probabilities yielded distinct separation between high-risk and low-risk groups in time-to-event analyses using Kaplan–Meier curves, while decision curve analysis suggested potential clinical utility as an adjunctive screening tool. Explainable analyses highlighted waveform patterns associated with elevated predicted risk. These findings support the feasibility of AI-ECG for complementary CAD risk screening and warrant prospective evaluation.

Method: Embed each canonical Gaussian primitive’s local frame into UV space patches on a template mesh efficiently. Reconstruct continuous editable material head textures from single monocular video on conventional UV domain. Use efficient physically based reflectance model for relighting and editing intrinsic material maps.

Result: Demonstrates accurate reconstructions, high-quality relighting results, and intuitive controls for modifying avatar appearance and geometry via texture mapping without additional optimization. Extensive comparisons show superiority over state-of-the-art methods.

Conclusion: Successfully combines Gaussian Splatting fidelity with UV texture mapping intuitiveness, enabling editable head avatars from monocular video with relighting capabilities and intuitive appearance/geometry controls.

Abstract: Recent advancements in Gaussian Splatting have enabled increasingly accurate reconstruction of photorealistic head avatars, opening the door to numerous applications in visual effects, videoconferencing, and virtual reality. This, however, comes with the lack of intuitive editability offered by traditional triangle mesh-based methods. In contrast, we propose a method that combines the accuracy and fidelity of 2D Gaussian Splatting with the intuitiveness of UV texture mapping. By embedding each canonical Gaussian primitive’s local frame into a patch in the UV space of a template mesh in a computationally efficient manner, we reconstruct continuous editable material head textures from a single monocular video on a conventional UV domain. Furthermore, we leverage an efficient physically based reflectance model to enable relighting and editing of these intrinsic material maps. Through extensive comparisons with state-of-the-art methods, we demonstrate the accuracy of our reconstructions, the quality of our relighting results, and the ability to provide intuitive controls for modifying an avatar’s appearance and geometry via texture mapping without additional optimization.

Method: Introduces MomaGraph unified representation, creates MomaGraph-Scenes dataset (large-scale annotated task-driven scene graphs), MomaGraph-Bench evaluation suite, and MomaGraph-R1 (7B vision-language model trained with RL on the dataset) for task-oriented scene graph prediction and zero-shot planning.

Result: MomaGraph-R1 achieves 71.6% accuracy on the benchmark (+11.4% over best baseline), shows strong generalization across public benchmarks, and transfers effectively to real-robot experiments.

Conclusion: MomaGraph provides a comprehensive framework for embodied scene understanding and planning, with demonstrated state-of-the-art performance and real-world applicability.

Abstract: Mobile manipulators in households must both navigate and manipulate. This requires a compact, semantically rich scene representation that captures where objects are, how they function, and which parts are actionable. Scene graphs are a natural choice, yet prior work often separates spatial and functional relations, treats scenes as static snapshots without object states or temporal updates, and overlooks information most relevant for accomplishing the current task. To address these limitations, we introduce MomaGraph, a unified scene representation for embodied agents that integrates spatial-functional relationships and part-level interactive elements. However, advancing such a representation requires both suitable data and rigorous evaluation, which have been largely missing. We thus contribute MomaGraph-Scenes, the first large-scale dataset of richly annotated, task-driven scene graphs in household environments, along with MomaGraph-Bench, a systematic evaluation suite spanning six reasoning capabilities from high-level planning to fine-grained scene understanding. Built upon this foundation, we further develop MomaGraph-R1, a 7B vision-language model trained with reinforcement learning on MomaGraph-Scenes. MomaGraph-R1 predicts task-oriented scene graphs and serves as a zero-shot task planner under a Graph-then-Plan framework. Extensive experiments demonstrate that our model achieves state-of-the-art results among open-source models, reaching 71.6% accuracy on the benchmark (+11.4% over the best baseline), while generalizing across public benchmarks and transferring effectively to real-robot experiments.

Method: Proposes ALIGN with three components: 1) Occlusion-aware Center Estimation integrating LiDAR geometry and image semantics, 2) Adaptive Neighbor Sampling generating object candidates from LiDAR clustering with supplemental sampling, and 3) Dynamic Query Balancing between foreground/background regions.

Result: On nuScenes benchmark, ALIGN consistently improves multiple state-of-the-art detectors by up to +0.9 mAP and +1.2 NDS, particularly in challenging scenes with occlusions or dense crowds.

Conclusion: ALIGN provides an effective occlusion-robust, object-aware query initialization approach that enhances 3D object detection performance in complex scenarios.

Abstract: Recent query-based 3D object detection methods using camera and LiDAR inputs have shown strong performance, but existing query initialization strategies,such as random sampling or BEV heatmap-based sampling, often result in inefficient query usage and reduced accuracy, particularly for occluded or crowded objects. To address this limitation, we propose ALIGN (Advanced query initialization with LiDAR and Image GuidaNce), a novel approach for occlusion-robust, object-aware query initialization. Our model consists of three key components: (i) Occlusion-aware Center Estimation (OCE), which integrates LiDAR geometry and image semantics to estimate object centers accurately (ii) Adaptive Neighbor Sampling (ANS), which generates object candidates from LiDAR clustering and supplements each object by sampling spatially and semantically aligned points around it and (iii) Dynamic Query Balancing (DQB), which adaptively balances queries between foreground and background regions. Our extensive experiments on the nuScenes benchmark demonstrate that ALIGN consistently improves performance across multiple state-of-the-art detectors, achieving gains of up to +0.9 mAP and +1.2 NDS, particularly in challenging scenes with occlusions or dense crowds. Our code will be publicly available upon publication.

Method: Introduces Block-Recurrent Hypothesis (BRH) and trains Recurrent Approximations to Phase-structured TransfORmers (Raptor) - block-recurrent surrogates of pretrained ViTs that use k ≪ L distinct blocks applied recurrently to approximate original computation.

Result: Demonstrates BRH empirically: Raptor recovers 96% of DINOv2 ImageNet-1k linear probe accuracy in only 2 blocks at equivalent computational cost. Shows directional convergence into class-dependent angular basins, token-specific dynamics, and collapse to low rank updates in late depth.

Conclusion: A compact recurrent program emerges along ViT depth, revealing a low-complexity normative solution that enables studying these models through principled dynamical systems analysis.

Abstract: As Vision Transformers (ViTs) become standard vision backbones, a mechanistic account of their computational phenomenology is essential. Despite architectural cues that hint at dynamical structure, there is no settled framework that interprets Transformer depth as a well-characterized flow. In this work, we introduce the Block-Recurrent Hypothesis (BRH), arguing that trained ViTs admit a block-recurrent depth structure such that the computation of the original $L$ blocks can be accurately rewritten using only $k \ll L$ distinct blocks applied recurrently. Across diverse ViTs, between-layer representational similarity matrices suggest few contiguous phases. To determine whether these phases reflect genuinely reusable computation, we train block-recurrent surrogates of pretrained ViTs: Recurrent Approximations to Phase-structured TransfORmers (Raptor). In small-scale, we demonstrate that stochastic depth and training promote recurrent structure and subsequently correlate with our ability to accurately fit Raptor. We then provide an empirical existence proof for BRH by training a Raptor model to recover $96%$ of DINOv2 ImageNet-1k linear probe accuracy in only 2 blocks at equivalent computational cost. Finally, we leverage our hypothesis to develop a program of Dynamical Interpretability. We find i) directional convergence into class-dependent angular basins with self-correcting trajectories under small perturbations, ii) token-specific dynamics, where cls executes sharp late reorientations while patch tokens exhibit strong late-stage coherence toward their mean direction, and iii) a collapse to low rank updates in late depth, consistent with convergence to low-dimensional attractors. Altogether, we find a compact recurrent program emerges along ViT depth, pointing to a low-complexity normative solution that enables these models to be studied through principled dynamical systems analysis.

Method: Models LiDAR scans as beam-aware graphs, uses graph attention networks with gated residual fusion and feed-forward refinement to reconstruct missing elevation information without increasing network depth.

Result: Achieves lower reconstruction error and improved geometric consistency compared to PointNet-based models and deeper GAT baselines across diverse KITTI environments, preserving structural integrity with minimal vertical distortion.

Conclusion: Architectural refinement offers computationally efficient method for improving LiDAR resolution without requiring additional sensor hardware, addressing beam dropout through intelligent reconstruction.

Abstract: LiDAR-based perception in autonomous systems is constrained by fixed vertical beam resolution and further compromised by beam dropout resulting from environmental occlusions. This paper introduces SuperiorGAT, a graph attention-based framework designed to reconstruct missing elevation information in sparse LiDAR point clouds. By modeling LiDAR scans as beam-aware graphs and incorporating gated residual fusion with feed-forward refinement, SuperiorGAT enables accurate reconstruction without increasing network depth. To evaluate performance, structured beam dropout is simulated by removing every fourth vertical scanning beam. Extensive experiments across diverse KITTI environments, including Person, Road, Campus, and City sequences, demonstrate that SuperiorGAT consistently achieves lower reconstruction error and improved geometric consistency compared to PointNet-based models and deeper GAT baselines. Qualitative X-Z projections further confirm the model’s ability to preserve structural integrity with minimal vertical distortion. These results suggest that architectural refinement offers a computationally efficient method for improving LiDAR resolution without requiring additional sensor hardware.

Method: Proposes Gaussian-to-Point (G2P) which transfers appearance-aware attributes from 3D Gaussian Splatting to point clouds. Addresses misalignment between optimized Gaussians and original point geometry by establishing point-wise correspondences. Uses Gaussian opacity attributes to resolve geometric ambiguity and Gaussian scale attributes for precise boundary localization.

Result: Achieves superior performance on standard benchmarks with significant improvements on geometrically challenging classes. Works without any 2D or language supervision.

Conclusion: G2P effectively transfers appearance-aware attributes from 3D Gaussian Splatting to enhance point cloud segmentation, overcoming limitations of geometry-only features and improving performance on challenging cases.

Abstract: Semantic segmentation on point clouds is critical for 3D scene understanding. However, sparse and irregular point distributions provide limited appearance evidence, making geometry-only features insufficient to distinguish objects with similar shapes but distinct appearances (e.g., color, texture, material). We propose Gaussian-to-Point (G2P), which transfers appearance-aware attributes from 3D Gaussian Splatting to point clouds for more discriminative and appearance-consistent segmentation. Our G2P address the misalignment between optimized Gaussians and original point geometry by establishing point-wise correspondences. By leveraging Gaussian opacity attributes, we resolve the geometric ambiguity that limits existing models. Additionally, Gaussian scale attributes enable precise boundary localization in complex 3D scenes. Extensive experiments demonstrate that our approach achieves superior performance on standard benchmarks and shows significant improvements on geometrically challenging classes, all without any 2D or language supervision.

Method: Proposes Wavelet-Conditioned ControlNet (WCC-Net), a fully 3D diffusion-based framework that introduces explicit frequency-domain structural priors via wavelet representations. It injects wavelet-based structural guidance into a frozen pretrained diffusion backbone through a lightweight control branch, decoupling anatomical structure from noise while preserving generative expressiveness and 3D structural continuity.

Result: WCC-Net consistently outperforms CNN-, GAN-, and diffusion-based baselines. On internal 1/20-dose test set, it improves PSNR by +1.21 dB and SSIM by +0.008 over strong diffusion baseline, while reducing structural distortion (GMSD) and intensity error (NMAE). It also generalizes robustly to unseen dose levels (1/50 and 1/4), achieving superior quantitative performance and improved volumetric anatomical consistency.

Conclusion: WCC-Net effectively addresses the challenge of anatomical consistency in diffusion-based PET denoising by incorporating wavelet-based structural guidance, demonstrating strong performance across various dose levels while maintaining 3D structural continuity.

Abstract: Low-dose Positron Emission Tomography (PET) imaging reduces patient radiation exposure but suffers from increased noise that degrades image quality and diagnostic reliability. Although diffusion models have demonstrated strong denoising capability, their stochastic nature makes it challenging to enforce anatomically consistent structures, particularly in low signal-to-noise regimes and volumetric whole-body imaging. We propose Wavelet-Conditioned ControlNet (WCC-Net), a fully 3D diffusion-based framework that introduces explicit frequency-domain structural priors via wavelet representations to guide volumetric PET denoising. By injecting wavelet-based structural guidance into a frozen pretrained diffusion backbone through a lightweight control branch, WCC-Net decouples anatomical structure from noise while preserving generative expressiveness and 3D structural continuity. Extensive experiments demonstrate that WCC-Net consistently outperforms CNN-, GAN-, and diffusion-based baselines. On the internal 1/20-dose test set, WCC-Net improves PSNR by +1.21 dB and SSIM by +0.008 over a strong diffusion baseline, while reducing structural distortion (GMSD) and intensity error (NMAE). Moreover, WCC-Net generalizes robustly to unseen dose levels (1/50 and 1/4), achieving superior quantitative performance and improved volumetric anatomical consistency.

Method: Created dataset using Moonworks Lunara model to generate images embodying distinct aesthetic styles from various global regions, with human-refined prompts and structured annotations describing objects, attributes, relationships, and stylistic cues.

Result: Produced a first-of-its-kind dataset with substantially higher aesthetic scores than existing aesthetics-focused and general-purpose datasets, featuring diverse regional aesthetics and comprehensive annotations.

Conclusion: The Lunara Aesthetic Dataset provides a valuable resource for aesthetic research with high-quality, diverse artistic styles, transparent licensing, and comprehensive annotations.

Abstract: The dataset spans diverse artistic styles, including regionally grounded aesthetics from the Middle East, Northern Europe, East Asia, and South Asia, alongside general categories such as sketch and oil painting. All images are generated using the Moonworks Lunara model and intentionally crafted to embody distinct, high-quality aesthetic styles, yielding a first-of-its-kind dataset with substantially higher aesthetic scores, exceeding even aesthetics-focused datasets, and general-purpose datasets by a larger margin. Each image is accompanied by a human-refined prompt and structured annotations that jointly describe salient objects, attributes, relationships, and stylistic cues. Unlike large-scale web-derived datasets that emphasize breadth over precision, the Lunara Aesthetic Dataset prioritizes aesthetic quality, stylistic diversity, and licensing transparency, and is released under the Apache 2.0 license to support research and unrestricted academic and commercial use.

Method: Uses semantic and uncertainty priors to suppress free space projections, employs unsigned distance encoding for geometric consistency, cascade sparse completion with hyper cross sparse convolution, and object contextual representation mask decoder for efficient query-context interactions.

Result: Achieves 7.34% improvement in accuracy and 57.8% gain in efficiency on SemanticKITTI benchmark compared to baselines.

Conclusion: SUG-Occ effectively addresses computational challenges in 3D semantic occupancy prediction through sparse learning techniques while maintaining geometric and semantic completeness.

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: MetamerGen uses a latent diffusion model with a dual-stream representation of foveated scenes. It combines DINOv2 tokens that fuse detailed features from fixated areas with peripherally degraded features capturing scene context. The model tackles the novel image-to-image synthesis problem of generating from both high and low resolution foveated inputs.

Result: The model generates image metamers that align with human scene representations. Behavioral experiments show participants perceive generated images as “same” as originals when conditioned on viewers’ own fixations. High-level semantic alignment most strongly predicts metamerism when conditioned on actual viewer fixations.

Conclusion: MetamerGen is a powerful tool for understanding human scene understanding. It can generate metamers even with random fixations, but achieves strongest perceptual alignment when conditioned on viewers’ actual fixated regions, revealing specific features at multiple visual processing levels that contribute to human judgments.

Abstract: Human vision combines low-resolution “gist” information from the visual periphery with sparse but high-resolution information from fixated locations to construct a coherent understanding of a visual scene. In this paper, we introduce MetamerGen, a tool for generating scenes that are aligned with latent human scene representations. MetamerGen is a latent diffusion model that combines peripherally obtained scene gist information with information obtained from scene-viewing fixations to generate image metamers for what humans understand after viewing a scene. Generating images from both high and low resolution (i.e. “foveated”) inputs constitutes a novel image-to-image synthesis problem, which we tackle by introducing a dual-stream representation of the foveated scenes consisting of DINOv2 tokens that fuse detailed features from fixated areas with peripherally degraded features capturing scene context. To evaluate the perceptual alignment of MetamerGen generated images to latent human scene representations, we conducted a same-different behavioral experiment where participants were asked for a “same” or “different” response between the generated and the original image. With that, we identify scene generations that are indeed metamers for the latent scene representations formed by the viewers. MetamerGen is a powerful tool for understanding scene understanding. Our proof-of-concept analyses uncovered specific features at multiple levels of visual processing that contributed to human judgments. While it can generate metamers even conditioned on random fixations, we find that high-level semantic alignment most strongly predicts metamerism when the generated scenes are conditioned on viewers’ own fixated regions.

Method: Proposes GO-MLVTON with two main modules: 1) Garment Occlusion Learning module to learn occlusion relationships between garment layers, and 2) StableDiffusion-based Garment Morphing & Fitting module to deform and fit garments onto the human body. Also introduces the MLG dataset and a new evaluation metric called Layered Appearance Coherence Difference (LACD).

Result: Extensive experiments demonstrate state-of-the-art performance in multi-layer virtual try-on, producing high-quality multi-layer try-on results with realistic garment deformation and layering.

Conclusion: GO-MLVTON successfully addresses the multi-layer virtual try-on problem by learning occlusion relationships and using diffusion-based garment fitting, establishing a new benchmark for this challenging task with the proposed MLG dataset and LACD metric.

Abstract: Existing image-based virtual try-on (VTON) methods primarily focus on single-layer or multi-garment VTON, neglecting multi-layer VTON (ML-VTON), which involves dressing multiple layers of garments onto the human body with realistic deformation and layering to generate visually plausible outcomes. The main challenge lies in accurately modeling occlusion relationships between inner and outer garments to reduce interference from redundant inner garment features. To address this, we propose GO-MLVTON, the first multi-layer VTON method, introducing the Garment Occlusion Learning module to learn occlusion relationships and the StableDiffusion-based Garment Morphing & Fitting module to deform and fit garments onto the human body, producing high-quality multi-layer try-on results. Additionally, we present the MLG dataset for this task and propose a new metric named Layered Appearance Coherence Difference (LACD) for evaluation. Extensive experiments demonstrate the state-of-the-art performance of GO-MLVTON. Project page: https://upyuyang.github.io/go-mlvton/.

Method: Proposes Federated Balanced Learning (FBL) that achieves sample balance on client side through knowledge filling and knowledge sampling using edge-side generation models. Includes Knowledge Alignment Strategy to bridge synthetic-real data gap, Knowledge Drop Strategy for regularization, and scales to complex scenarios with heterogeneous client methods.

Result: Numerous experiments show the method outperforms state-of-the-art baselines in federated learning with non-iid data distributions.

Conclusion: FBL effectively prevents client drift by addressing sample imbalance at the client side through generative approaches, offering a scalable solution for real-world federated learning scenarios.

Abstract: Federated learning is a paradigm of joint learning in which clients collaborate by sharing model parameters instead of data. However, in the non-iid setting, the global model experiences client drift, which can seriously affect the final performance of the model. Previous methods tend to correct the global model that has already deviated based on the loss function or gradient, overlooking the impact of the client samples. In this paper, we rethink the role of the client side and propose Federated Balanced Learning, i.e., FBL, to prevent this issue from the beginning through sample balance on the client side. Technically, FBL allows unbalanced data on the client side to achieve sample balance through knowledge filling and knowledge sampling using edge-side generation models, under the limitation of a fixed number of data samples on clients. Furthermore, we design a Knowledge Alignment Strategy to bridge the gap between synthetic and real data, and a Knowledge Drop Strategy to regularize our method. Meanwhile, we scale our method to real and complex scenarios, allowing different clients to adopt various methods, and extend our framework to further improve performance. Numerous experiments show that our method outperforms state-of-the-art baselines. The code is released upon acceptance.

Method: Proposes ThermoSplat with: 1) Spectrum-Aware Adaptive Modulation that dynamically conditions shared latent features on thermal structural priors; 2) Modality-Adaptive Geometric Decoupling that learns independent opacity offsets and executes separate rasterization for thermal branch; 3) Hybrid rendering pipeline integrating explicit Spherical Harmonics with implicit neural decoding.

Result: Extensive experiments on RGBT-Scenes dataset demonstrate state-of-the-art rendering quality across both visible and thermal spectrums.

Conclusion: ThermoSplat enables deep spectral-aware reconstruction through active feature modulation and adaptive geometry decoupling, effectively handling complex structural correlations and physical discrepancies between RGB and thermal modalities.

Abstract: Multi-modal scene reconstruction integrating RGB and thermal infrared data is essential for robust environmental perception across diverse lighting and weather conditions. However, extending 3D Gaussian Splatting (3DGS) to multi-spectral scenarios remains challenging. Current approaches often struggle to fully leverage the complementary information of multi-modal data, typically relying on mechanisms that either tend to neglect cross-modal correlations or leverage shared representations that fail to adaptively handle the complex structural correlations and physical discrepancies between spectrums. To address these limitations, we propose ThermoSplat, a novel framework that enables deep spectral-aware reconstruction through active feature modulation and adaptive geometry decoupling. First, we introduce a Spectrum-Aware Adaptive Modulation that dynamically conditions shared latent features on thermal structural priors, effectively guiding visible texture synthesis with reliable cross-modal geometric cues. Second, to accommodate modality-specific geometric inconsistencies, we propose a Modality-Adaptive Geometric Decoupling scheme that learns independent opacity offsets and executes an independent rasterization pass for the thermal branch. Additionally, a hybrid rendering pipeline is employed to integrate explicit Spherical Harmonics with implicit neural decoding, ensuring both semantic consistency and high-frequency detail preservation. Extensive experiments on the RGBT-Scenes dataset demonstrate that ThermoSplat achieves state-of-the-art rendering quality across both visible and thermal spectrums.

Method: Proposes Prompt-Agnostic Evolution (PAE) with three components: 1) Frequency-aware initialization using frequency shortcut patterns the backbone inherently exploits, 2) Shared Koopman operator for coherent cross-layer evolution via global linear transformation, and 3) Lyapunov-inspired regularizer to constrain error amplification during training.

Result: PAE achieves average 1.41× convergence speedup and 1-3% accuracy improvement on 25 datasets across multiple downstream tasks. It’s prompt-agnostic, lightweight, and integrates seamlessly with diverse VPT variants without backbone modification or inference changes.

Conclusion: PAE effectively addresses training instability in VPT through frequency-aware initialization, cross-layer coordination, and stability regularization, making prompt tuning more efficient and effective while maintaining compatibility with existing VPT frameworks.

Abstract: Visual Prompt Tuning (VPT) adapts a frozen Vision Transformer (ViT) to downstream tasks by inserting a small number of learnable prompt tokens into the token sequence at each layer. However, we observe that existing VPT variants often suffer from unstable training dynamics, characterized by gradient oscillations. A layer-wise analysis reveals that shallow-layer prompts tend to stagnate early, while deeper-layer prompts exhibit high-variance oscillations, leading to cross-layer mismatch. These issues slow convergence and degrade final performance. To address these challenges, we propose Prompt-Agnostic Evolution ($\mathtt{PAE}$), which strengthens vision prompt tuning by explicitly modeling prompt dynamics. From a frequency-domain perspective, we initialize prompts in a task-aware direction by uncovering and propagating frequency shortcut patterns that the backbone inherently exploits for recognition. To ensure coherent evolution across layers, we employ a shared Koopman operator that imposes a global linear transformation instead of uncoordinated, layer-specific updates. Finally, inspired by Lyapunov stability theory, we introduce a regularizer that constrains error amplification during evolution. Extensive experiments show that $\mathtt{PAE}$ accelerates convergence with an average $1.41\times$ speedup and improves accuracy by 1-3% on 25 datasets across multiple downstream tasks. Beyond performance, $\mathtt{PAE}$ is prompt-agnostic and lightweight, and it integrates seamlessly with diverse VPT variants without backbone modification or inference-time changes.

Method: Proposes CAF-Mamba, a novel framework using Mamba architecture with cross-modal adaptive attention fusion that captures both explicit and implicit cross-modal interactions and dynamically adjusts modality contributions through modality-wise attention mechanism.

Result: Outperforms existing methods on two in-the-wild benchmark datasets (LMVD and D-Vlog), achieving state-of-the-art performance for depression detection.

Conclusion: CAF-Mamba effectively addresses limitations of current multimodal fusion approaches for depression detection by enabling more effective cross-modal interaction and dynamic modality weighting.

Abstract: Depression is a prevalent mental health disorder that severely impairs daily functioning and quality of life. While recent deep learning approaches for depression detection have shown promise, most rely on limited feature types, overlook explicit cross-modal interactions, and employ simple concatenation or static weighting for fusion. To overcome these limitations, we propose CAF-Mamba, a novel Mamba-based cross-modal adaptive attention fusion framework. CAF-Mamba not only captures cross-modal interactions explicitly and implicitly, but also dynamically adjusts modality contributions through a modality-wise attention mechanism, enabling more effective multimodal fusion. Experiments on two in-the-wild benchmark datasets, LMVD and D-Vlog, demonstrate that CAF-Mamba consistently outperforms existing methods and achieves state-of-the-art performance. Our code is available at https://github.com/zbw-zhou/CAF-Mamba.

Method: Proposes Q-Hawkeye with two main components: (1) Uncertainty-Aware Dynamic Optimization that estimates predictive uncertainty using variance of predicted scores across multiple rollouts and uses this uncertainty to reweight each sample’s update strength; (2) Perception-Aware Optimization that constructs paired inputs of degraded images and their originals, and introduces an Implicit Perception Loss to constrain quality judgments to genuine visual evidence.

Result: Extensive experiments demonstrate that Q-Hawkeye outperforms state-of-the-art methods and generalizes better across multiple datasets.

Conclusion: Q-Hawkeye provides a more reliable RL-based visual policy optimization framework for IQA by addressing both uncertainty and perception limitations in existing MLLM approaches.

Abstract: Image Quality Assessment (IQA) predicts perceptual quality scores consistent with human judgments. Recent RL-based IQA methods built on MLLMs focus on generating visual quality descriptions and scores, ignoring two key reliability limitations: (i) although the model’s prediction stability varies significantly across training samples, existing GRPO-based methods apply uniform advantage weighting, thereby amplifying noisy signals from unstable samples in gradient updates; (ii) most works emphasize text-grounded reasoning over images while overlooking the model’s visual perception ability of image content. In this paper, we propose Q-Hawkeye, an RL-based reliable visual policy optimization framework that redesigns the learning signal through unified Uncertainty-Aware Dynamic Optimization and Perception-Aware Optimization. Q-Hawkeye estimates predictive uncertainty using the variance of predicted scores across multiple rollouts and leverages this uncertainty to reweight each sample’s update strength, stabilizing policy optimization. To strengthen perceptual reliability, we construct paired inputs of degraded images and their original images and introduce an Implicit Perception Loss that constrains the model to ground its quality judgments in genuine visual evidence. Extensive experiments demonstrate that Q-Hawkeye outperforms state-of-the-art methods and generalizes better across multiple datasets. The code and models will be made available.

Method: LatentLens encodes a large text corpus and stores contextualized token representations for each token. Visual token representations are then compared to these textual representations using nearest neighbor search, with top-k matches providing natural language descriptions of the visual tokens.

Result: Evaluation on 10 different VLMs shows that LatentLens substantially outperforms commonly used methods like LogitLens, with the majority of visual tokens being interpretable across all studied models and layers. The descriptions are semantically meaningful and provide fine-grained interpretations.

Conclusion: LatentLens provides new evidence for alignment between vision and language representations, offering a powerful tool for analyzing latent representations in multimodal models and opening new directions for interpretability research.

Abstract: Transforming a large language model (LLM) into a Vision-Language Model (VLM) can be achieved by mapping the visual tokens from a vision encoder into the embedding space of an LLM. Intriguingly, this mapping can be as simple as a shallow MLP transformation. To understand why LLMs can so readily process visual tokens, we need interpretability methods that reveal what is encoded in the visual token representations at every layer of LLM processing. In this work, we introduce LatentLens, a novel approach for mapping latent representations to descriptions in natural language. LatentLens works by encoding a large text corpus and storing contextualized token representations for each token in that corpus. Visual token representations are then compared to their contextualized textual representations, with the top-k nearest neighbor representations providing descriptions of the visual token. We evaluate this method on 10 different VLMs, showing that commonly used methods, such as LogitLens, substantially underestimate the interpretability of visual tokens. With LatentLens instead, the majority of visual tokens are interpretable across all studied models and all layers. Qualitatively, we show that the descriptions produced by LatentLens are semantically meaningful and provide more fine-grained interpretations for humans compared to individual tokens. More broadly, our findings contribute new evidence on the alignment between vision and language representations, opening up new directions for analyzing latent representations.

Method: Created a dataset of 2,854 anchor-linked variation pairs from original art and photographs, applying contextual transformations (illumination, weather, viewpoint, scene composition, color tone, mood) while preserving underlying identity.

Result: The dataset shows high identity stability, strong target attribute realization, and robust aesthetic profile exceeding large-scale web datasets, released under Apache 2.0 license for public use.

Conclusion: Lunara Aesthetic II provides a valuable resource for benchmarking, fine-tuning, and analyzing contextual generalization, identity preservation, and edit robustness in image generation systems with interpretable supervision.

Abstract: We introduce Lunara Aesthetic II, a publicly released, ethically sourced image dataset designed to support controlled evaluation and learning of contextual consistency in modern image generation and editing systems. The dataset comprises 2,854 anchor-linked variation pairs derived from original art and photographs created by Moonworks. Each variation pair applies contextual transformations, such as illumination, weather, viewpoint, scene composition, color tone, or mood; while preserving a stable underlying identity. Lunara Aesthetic II operationalizes identity-preserving contextual variation as a supervision signal while also retaining Lunara’s signature high aesthetic scores. Results show high identity stability, strong target attribute realization, and a robust aesthetic profile that exceeds large-scale web datasets. Released under the Apache 2.0 license, Lunara Aesthetic II is intended for benchmarking, fine-tuning, and analysis of contextual generalization, identity preservation, and edit robustness in image generation and image-to-image systems with interpretable, relational supervision. The dataset is publicly available at: https://huggingface.co/datasets/moonworks/lunara-aesthetic-image-variations.

Method: Proposes ReasonEdit with continuous storage of human reasoning in a codebook, retrieving relevant facts during inference using novel topology-balanced multimodal embedding method inspired by network science.

Result: Achieves state-of-the-art editing performance across four VLMs on multiple rationale-based visual question answering datasets, showing human reasoning during editing greatly improves edit generalization.

Conclusion: First VLM editor to incorporate human reasoning explanations, demonstrating significant improvements in edit generalization for reasoning-heavy visual tasks.

Abstract: Model editing aims to correct errors in large, pretrained models without altering unrelated behaviors. While some recent works have edited vision-language models (VLMs), no existing editors tackle reasoning-heavy tasks, which typically require humans and models to reason about images. We therefore propose ReasonEdit, the first VLM editor to let users explain their reasoning during editing, introducing a new, practical model editing setup. ReasonEdit continuously stores human reasoning in a codebook, and retrieves only relevant facts during inference using a novel topology-balanced multimodal embedding method inspired by network science. Across four VLMs on multiple rationale-based visual question answering datasets, ReasonEdit achieves state-of-the-art editing performance, ultimately showing that using human reasoning during editing greatly improves edit generalization.

Method: Proposes UniReason framework with two complementary reasoning paradigms: 1) world knowledge-enhanced textual reasoning for inferring implicit knowledge during generation, and 2) editing capabilities for fine-grained visual refinement via self-reflection. Unifies generation and editing within shared architecture, and constructs large-scale reasoning-centric dataset (~300k samples) covering five major knowledge domains.

Result: Extensive experiments show UniReason achieves advanced performance on reasoning-intensive benchmarks (WISE, KrisBench, UniREditBench) while maintaining superior general synthesis capabilities.

Conclusion: UniReason successfully unifies generation and editing through complementary reasoning paradigms, mirroring human cognitive processes of planning followed by refinement, and demonstrates strong performance on reasoning-intensive multimodal tasks.

Abstract: Unified multimodal models often struggle with complex synthesis tasks that demand deep reasoning, and typically treat text-to-image generation and image editing as isolated capabilities rather than interconnected reasoning steps. To address this, we propose UniReason, a unified framework that harmonizes these two tasks through two complementary reasoning paradigms. We incorporate world knowledge-enhanced textual reasoning into generation to infer implicit knowledge, and leverage editing capabilities for fine-grained editing-like visual refinement to further correct visual errors via self-reflection. This approach unifies generation and editing within a shared architecture, mirroring the human cognitive process of planning followed by refinement. We support this framework by systematically constructing a large-scale reasoning-centric dataset (~300k samples) covering five major knowledge domains (e.g., cultural commonsense, physics, etc.) for textual reasoning, alongside an agent-generated corpus for visual refinement. Extensive experiments demonstrate that UniReason achieves advanced performance on reasoning-intensive benchmarks such as WISE, KrisBench and UniREditBench, while maintaining superior general synthesis capabilities.

Method: Proposes SVD-ViT with three components: SPC module, SSVA, and ID-RSVD, which use singular value decomposition to extract and aggregate singular vectors capturing object foreground information while suppressing task-irrelevant background noise.

Result: Experimental results show improved classification accuracy and effective learning of informative foreground representations while reducing background noise impact.

Conclusion: SVD-ViT successfully addresses ViT’s foreground-background discrimination limitation through SVD-based feature prioritization, enhancing classification performance.

Abstract: Vision Transformers (ViT) have been established as large-scale foundation models. However, because self-attention operates globally, they lack an explicit mechanism to distinguish foreground from background. As a result, ViT may learn unnecessary background features and artifacts, leading to degraded classification performance. To address this issue, we propose SVD-ViT, which leverages singular value decomposition (SVD) to prioritize the learning of foreground features. SVD-ViT consists of three components-\textbf{SPC module}, \textbf{SSVA}, and \textbf{ID-RSVD}-and suppresses task-irrelevant factors such as background noise and artifacts by extracting and aggregating singular vectors that capture object foreground information. Experimental results demonstrate that our method improves classification accuracy and effectively learns informative foreground representations while reducing the impact of background noise.

Method: Uses off-the-shelf whole-body person detector with low threshold, retrieves false negatives via temporal tracking and self-supervised multi-view association, uses recovered detections as pseudo labels to iteratively fine-tune detector, then applies whole-body pose estimation and fine-tunes pose model with its own high-score predictions.

Result: Achieves over 97% recall on 4D-OR dataset of simulated surgeries and real surgery dataset. Trains real-time whole-body detector using pseudo labels with comparable performance.

Conclusion: Proposed framework effectively addresses scalability bottlenecks in OR video anonymization by eliminating need for manual annotations and camera calibration while maintaining high recall rates.

Abstract: Privacy preservation is a prerequisite for using video data in Operating Room (OR) research. Effective anonymization relies on the exhaustive localization of every individual; even a single missed detection necessitates extensive manual correction. However, existing approaches face two critical scalability bottlenecks: (1) they usually require manual annotations of each new clinical site for high accuracy; (2) while multi-camera setups have been widely adopted to address single-view ambiguity, camera calibration is typically required whenever cameras are repositioned. To address these problems, we propose a novel self-supervised multi-view video anonymization framework consisting of whole-body person detection and whole-body pose estimation, without annotation or camera calibration. Our core strategy is to enhance the single-view detector by “retrieving” false negatives using temporal and multi-view context, and conducting self-supervised domain adaptation. We first run an off-the-shelf whole-body person detector in each view with a low-score threshold to gather candidate detections. Then, we retrieve the low-score false negatives that exhibit consistency with the high-score detections via tracking and self-supervised uncalibrated multi-view association. These recovered detections serve as pseudo labels to iteratively fine-tune the whole-body detector. Finally, we apply whole-body pose estimation on each detected person, and fine-tune the pose model using its own high-score predictions. Experiments on the 4D-OR dataset of simulated surgeries and our dataset of real surgeries show the effectiveness of our approach achieving over 97% recall. Moreover, we train a real-time whole-body detector using our pseudo labels, achieving comparable performance and highlighting our method’s practical applicability. Code will be available at https://github.com/CAMMA-public/OR_anonymization.

Method: Proposes Diffusion-Guided Attention Network (DiGAN) with two components: 1) latent diffusion model synthesizes realistic longitudinal neuroimaging trajectories from limited training data, enriching temporal context and handling unevenly spaced visits; 2) attention-convolutional layer captures discriminative structural-temporal patterns distinguishing cognitive states.

Result: Experiments on ADNI dataset show DiGAN outperforms existing state-of-the-art baselines, demonstrating potential for early-stage Alzheimer’s disease detection.

Conclusion: DiGAN effectively addresses limitations of existing methods by combining diffusion-based data synthesis with attention mechanisms for improved early Alzheimer’s detection from limited longitudinal data.

Abstract: Early diagnosis of Alzheimer’s disease (AD) remains a major challenge due to the subtle and temporally irregular progression of structural brain changes in the prodromal stages. Existing deep learning approaches require large longitudinal datasets and often fail to model the temporal continuity and modality irregularities inherent in real-world clinical data. To address these limitations, we propose the Diffusion-Guided Attention Network (DiGAN), which integrates latent diffusion modelling with an attention-guided convolutional network. The diffusion model synthesizes realistic longitudinal neuroimaging trajectories from limited training data, enriching temporal context and improving robustness to unevenly spaced visits. The attention-convolutional layer then captures discriminative structural-temporal patterns that distinguish cognitively normal subjects from those with mild cognitive impairment and subjective cognitive decline. Experiments on the ADNI dataset demonstrate that DiGAN outperforms existing state-of-the-art baselines, showing its potential for early-stage AD detection.

Method: Proposes Mask-LLaVA combining three visual feature levels: 1) mask-based object representations, 2) global tokens, 3) local patch tokens. All tokens used during training, but model can flexibly drop mask-based object tokens at test time for adaptive token reduction.

Result: Achieves results competitive with token-efficient methods and comparable to original LLaVA baseline using only a fraction of visual tokens. Enables dynamic token selection at test time with minimal performance drop.

Conclusion: Combining multi-level visual features enables efficient learning with fewer tokens while allowing flexible token reduction at inference, addressing computational efficiency in autoregressive VLMs.

Abstract: Current autoregressive Vision Language Models (VLMs) usually rely on a large number of visual tokens to represent images, resulting in a need for more compute especially at inference time. To address this problem, we propose Mask-LLaVA, a framework that leverages different levels of visual features to create a compact yet information-rich visual representation for autoregressive VLMs. Namely, we combine mask-based object representations together with global tokens and local patch tokens. While all tokens are used during training, it shows that the resulting model can flexibly drop especially the number of mask-based object-tokens at test time, allowing to adapt the number of tokens during inference without the need to retrain the model and without a significant drop in performance. We evaluate the proposed approach on a suite of standard benchmarks showing results competitive to current token efficient methods and comparable to the original LLaVA baseline using only a fraction of visual tokens. Our analysis demonstrates that combining multi-level features enables efficient learning with fewer tokens while allowing dynamic token selection at test time for good performance.

Method: FSR uses a three-step human-inspired approach: 1) Focus on key evidence by combining visual importance with instruction relevance, 2) Scan for complementary context conditioned on focused set, selecting tokens most different from focused evidence, 3) Refine scanned context by aggregating nearby informative tokens via similarity-based assignment and score-weighted merging.

Result: Extensive experiments across multiple VLM backbones and vision-language benchmarks show FSR consistently improves accuracy-efficiency trade-off over state-of-the-art pruning methods.

Conclusion: FSR provides an effective plug-and-play pruning framework that mimics human visual reasoning to balance local evidence and global context while maintaining efficiency.

Abstract: Vision-language models (VLMs) often generate massive visual tokens that greatly increase inference latency and memory footprint; while training-free token pruning offers a practical remedy, existing methods still struggle to balance local evidence and global context under aggressive compression. We propose Focus-Scan-Refine (FSR), a human-inspired, plug-and-play pruning framework that mimics how humans answer visual questions: focus on key evidence, then scan globally if needed, and refine the scanned context by aggregating relevant details. FSR first focuses on key evidence by combining visual importance with instruction relevance, avoiding the bias toward visually salient but query-irrelevant regions. It then scans for complementary context conditioned on the focused set, selecting tokens that are most different from the focused evidence. Finally, FSR refines the scanned context by aggregating nearby informative tokens into the scan anchors via similarity-based assignment and score-weighted merging, without increasing the token budget. Extensive experiments across multiple VLM backbones and vision-language benchmarks show that FSR consistently improves the accuracy-efficiency trade-off over existing state-of-the-art pruning methods. The source codes can be found at https://github.com/ILOT-code/FSR.

Method: 1) Uses DINOv3 Vision Foundation Model features as a unified bridge between simulation and real-world domains; 2) Employs Principal Subspace Projection to discard high-frequency elements causing “texture baking”; 3) Introduces Random Channel Tail Drop to mitigate structural loss; 4) Develops learnable Spatial Alignment Module to adapt high-resolution features to diffusion backbone; 5) Proposes Causal Temporal Aggregator with causal convolutions to preserve historical motion context.

Result: The framework effectively reconciles realism with control consistency in autonomous driving video generation, mitigating motion blur and guaranteeing temporal stability while maintaining precise control over generated content.

Conclusion: Driving with DINO presents a novel approach that leverages vision foundation model features to overcome the consistency-realism trade-off in Sim2Real video generation for autonomous driving, enabling both photorealism and precise control.

Abstract: Driven by the emergence of Controllable Video Diffusion, existing Sim2Real methods for autonomous driving video generation typically rely on explicit intermediate representations to bridge the domain gap. However, these modalities face a fundamental Consistency-Realism Dilemma. Low-level signals (e.g., edges, blurred images) ensure precise control but compromise realism by “baking in” synthetic artifacts, whereas high-level priors (e.g., depth, semantics, HDMaps) facilitate photorealism but lack the structural detail required for consistent guidance. In this work, we present Driving with DINO (DwD), a novel framework that leverages Vision Foundation Module (VFM) features as a unified bridge between the simulation and real-world domains. We first identify that these features encode a spectrum of information, from high-level semantics to fine-grained structure. To effectively utilize this, we employ Principal Subspace Projection to discard the high-frequency elements responsible for “texture baking,” while concurrently introducing Random Channel Tail Drop to mitigate the structural loss inherent in rigid dimensionality reduction, thereby reconciling realism with control consistency. Furthermore, to fully leverage DINOv3’s high-resolution capabilities for enhancing control precision, we introduce a learnable Spatial Alignment Module that adapts these high-resolution features to the diffusion backbone. Finally, we propose a Causal Temporal Aggregator employing causal convolutions to explicitly preserve historical motion context when integrating frame-wise DINO features, which effectively mitigates motion blur and guarantees temporal stability. Project page: https://albertchen98.github.io/DwD-project/

Method: Proposes a differentiable vehicle-model framework that rolls out predicted action sequences (throttle, steer, brake) to their corresponding ego-frame waypoint trajectories, enabling supervision in waypoint space while maintaining action-based architecture.

Result: Achieves consistent improvements across multiple challenging benchmarks, with state-of-the-art performance on NAVSIM navhard benchmark. Enables action-based architectures to be trained and evaluated within waypoint-based benchmarks without modifying evaluation protocols.

Conclusion: The framework successfully bridges the waypoint-action gap in autonomous driving, allowing action-based models to benefit from waypoint-based training and evaluation pipelines while maintaining their inherent advantages.

Abstract: End-to-End Autonomous Driving (E2E-AD) systems are typically grouped by the nature of their outputs: (i) waypoint-based models that predict a future trajectory, and (ii) action-based models that directly output throttle, steer and brake. Most recent benchmark protocols and training pipelines are waypoint-based, which makes action-based policies harder to train and compare, slowing their progress. To bridge this waypoint-action gap, we propose a novel, differentiable vehicle-model framework that rolls out predicted action sequences to their corresponding ego-frame waypoint trajectories while supervising in waypoint space. Our approach enables action-based architectures to be trained and evaluated, for the first time, within waypoint-based benchmarks without modifying the underlying evaluation protocol. We extensively evaluate our framework across multiple challenging benchmarks and observe consistent improvements over the baselines. In particular, on NAVSIM \texttt{navhard} our approach achieves state-of-the-art performance. Our code will be made publicly available upon acceptance.

Method: Trained Depth Anything V2 model using ~16,000 km² of canopy height models from public airborne LiDAR point clouds across multiple countries, paired with 3m resolution PlanetScope and airborne RGB imagery. Model called Depth2CHM estimates CHMs directly from RGB imagery.

Result: Independent validation in China (1 km²) and US (116 km²) showed biases of 0.59m/0.41m and RMSEs of 2.54m/5.75m respectively. Outperformed existing global CHM product with ~1.5m lower MAE and ~2m lower RMSE.

Conclusion: Monocular depth estimation networks trained on large-scale airborne LiDAR data provide promising, scalable pathway for high-resolution, spatially continuous forest canopy height estimation from satellite RGB imagery.

Abstract: Large-scale, high-resolution forest canopy height mapping plays a crucial role in understanding regional and global carbon and water cycles. Spaceborne LiDAR missions, including the Ice, Cloud, and Land Elevation Satellite-2 (ICESat-2) and the Global Ecosystem Dynamics Investigation (GEDI), provide global observations of forest structure but are spatially sparse and subject to inherent uncertainties. In contrast, near-surface LiDAR platforms, such as airborne and unmanned aerial vehicle (UAV) LiDAR systems, offer much finer measurements of forest canopy structure, and a growing number of countries have made these datasets openly available. In this study, a state-of-the-art monocular depth estimation model, Depth Anything V2, was trained using approximately 16,000 km2 of canopy height models (CHMs) derived from publicly available airborne LiDAR point clouds and related products across multiple countries, together with 3 m resolution PlanetScope and airborne RGB imagery. The trained model, referred to as Depth2CHM, enables the estimation of spatially continuous CHMs directly from PlanetScope RGB imagery. Independent validation was conducted at sites in China (approximately 1 km2) and the United States (approximately 116 km2). The results showed that Depth2CHM could accurately estimate canopy height, with biases of 0.59 m and 0.41 m and root mean square errors (RMSEs) of 2.54 m and 5.75 m for these two sites, respectively. Compared with an existing global meter-resolution CHM product, the mean absolute error is reduced by approximately 1.5 m and the RMSE by approximately 2 m. These results demonstrated that monocular depth estimation networks trained with large-scale airborne LiDAR-derived canopy height data provide a promising and scalable pathway for high-resolution, spatially continuous forest canopy height estimation from satellite RGB imagery.

Method: LLM-FSM uses an automated pipeline that: 1) constructs FSMs with configurable state counts and constrained transition structures, 2) prompts LLMs to express FSMs in structured YAML with application context, 3) converts YAML to natural language specifications, 4) synthesizes reference RTL and testbenches from YAML in a correct-by-construction manner, and 5) verifies all 1,000 problems using LLM-based and SAT-solver-based checks.

Result: Experiments show that even the strongest LLMs exhibit sharply declining accuracy as FSM complexity increases. Training-time scaling via supervised fine-tuning generalizes effectively to out-of-distribution tasks, while increasing test-time compute improves reasoning reliability.

Conclusion: LLM-FSM provides a scalable, automated benchmark for evaluating LLMs on finite-state reasoning for hardware design, demonstrating current limitations and showing that both training-time and test-time improvements can enhance performance on these tasks.

Abstract: Finite-state reasoning, the ability to understand and implement state-dependent behavior, is central to hardware design. In this paper, we present LLM-FSM, a benchmark that evaluates how well large language models (LLMs) can recover finite-state machine (FSM) behavior from natural-language specifications and translate it into correct register transfer-level (RTL) implementations. Unlike prior specification-to-RTL benchmarks that rely on manually constructed examples, LLM-FSM is built through a fully automated pipeline. LLM-FSM first constructs FSM with configurable state counts and constrained transition structures. It then prompts LLMs to express each FSM in a structured YAML format with an application context, and to further convert that YAML into a natural-language (NL) specification. From the same YAML, our pipeline synthesizes the reference RTL and testbench in a correct-by-construction manner. All 1,000 problems are verified using LLM-based and SAT-solver-based checks, with human review on a subset. Our experiments show that even the strongest LLMs exhibit sharply declining accuracy as FSM complexity increases. We further demonstrate that training-time scaling via supervised fine-tuning (SFT) generalizes effectively to out-of-distribution (OOD) tasks, while increasing test-time compute improves reasoning reliability. Finally, LLM-FSM remains extensible by allowing its FSM complexity to scale with future model capabilities.

Method: ST-Raptor offers an interactive analysis environment combining three key components: (1) visual editing capabilities, (2) tree-based structural modeling, and (3) agent-driven query resolution. This approach supports accurate and user-friendly table understanding without requiring conversion to structured formats.

Result: Experimental results on both benchmark and real-world datasets demonstrate that ST-Raptor outperforms existing methods in both accuracy and usability.

Conclusion: ST-Raptor provides an effective solution for semi-structured table QA that addresses limitations of existing approaches through its agentic, interactive system combining visual, structural, and agent-based components.

Abstract: Semi-structured table question answering (QA) is a challenging task that requires (1) precise extraction of cell contents and positions and (2) accurate recovery of key implicit logical structures, hierarchical relationships, and semantic associations encoded in table layouts. In practice, such tables are often interpreted manually by human experts, which is labor-intensive and time-consuming. However, automating this process remains difficult. Existing Text-to-SQL methods typically require converting semi-structured tables into structured formats, inevitably leading to information loss, while approaches like Text-to-Code and multimodal LLM-based QA struggle with complex layouts and often yield inaccurate answers. To address these limitations, we present ST-Raptor, an agentic system for semi-structured table QA. ST-Raptor offers an interactive analysis environment that combines visual editing, tree-based structural modeling, and agent-driven query resolution to support accurate and user-friendly table understanding. Experimental results on both benchmark and real-world datasets demonstrate that ST-Raptor outperforms existing methods in both accuracy and usability. The code is available at https://github.com/weAIDB/ST-Raptor, and a demonstration video is available at https://youtu.be/9GDR-94Cau4.

Method: Two-stage post-training pipeline: 1) Agentic SFT for information seeking and reasoning capabilities, 2) Agentic VRPO for preference optimization. Plus P-ReAct paradigm that prioritizes decoding tool_call instructions to enable parallel thinking while waiting for tool returns.

Result: DLLM-Searcher achieves performance comparable to mainstream LLM-based search agents, with P-ReAct delivering approximately 15% inference acceleration.

Conclusion: The framework successfully addresses both capability and latency challenges in dLLM-based search agents, demonstrating practical viability of diffusion models for agent applications.

Abstract: Recently, Diffusion Large Language Models (dLLMs) have demonstrated unique efficiency advantages, enabled by their inherently parallel decoding mechanism and flexible generation paradigm. Meanwhile, despite the rapid advancement of Search Agents, their practical deployment is constrained by a fundamental limitation, termed as 1) Latency Challenge: the serial execution of multi-round reasoning, tool calling, and tool response waiting under the ReAct agent paradigm induces severe end-to-end latency. Intuitively, dLLMs can leverage their distinctive strengths to optimize the operational efficiency of agents under the ReAct agent paradigm. Practically, existing dLLM backbones face the 2) Agent Ability Challenge. That is, existing dLLMs exhibit remarkably weak reasoning and tool-calling capabilities, preventing these advantages from being effectively realized in practice. In this paper, we propose DLLM-Searcher, an optimization framework for dLLM-based Search Agents. To solve the Agent Ability Challenge, we design a two-stage post-training pipeline encompassing Agentic Supervised Fine-Tuning (Agentic SFT) and Agentic Variance-Reduced Preference Optimization Agentic VRPO, which enhances the backbone dLLM’s information seeking and reasoning capabilities. To mitigate the Latency Challenge, we leverage the flexible generation mechanism of dLLMs and propose a novel agent paradigm termed Parallel-Reasoning and Acting P-ReAct. P-ReAct guides the model to prioritize decoding tool_call instructions, thereby allowing the model to keep thinking while waiting for the tool’s return. Experimental results demonstrate that DLLM-Searcher achieves performance comparable to mainstream LLM-based search agents and P-ReAct delivers approximately 15% inference acceleration. Our code is available at https://anonymous.4open.science/r/DLLM-Searcher-553C

Method: Aster takes a task, initial program, and evaluation script, then iteratively improves the program through an autonomous discovery process that requires significantly fewer iterations than existing methods.

Result: Achieved state-of-the-art results in multiple domains: Erdos minimum overlap problem, TriMul kernel optimization, single-cell analysis denoising, neural activity prediction for ZAPBench (matching best human solution with <1/190th compute), and NanoGPT Speedrun Competition.

Conclusion: Aster demonstrates highly efficient autonomous scientific discovery, expanding tractable problem domains to include tasks with long evaluation durations and achieving superior performance across diverse scientific fields.

Abstract: We introduce Aster, an AI agent for autonomous scientific discovery capable of operating over 20 times faster than existing frameworks. Given a task, an initial program, and a script to evaluate the performance of the program, Aster iteratively improves the program, often leading to new state-of-the-art performances. Aster’s significant reduction in the number of iterations required for novel discovery expands the domain of tractable problems to include tasks with long evaluation durations, such as multi-hour machine learning training runs. We applied Aster to problems in mathematics, GPU kernel engineering, biology, neuroscience, and language model training. More specifically: the Erdos minimum overlap problem, optimizing the TriMul kernel, a single-cell analysis denoising problem, training a neural activity prediction model to perform well on ZAPBench, and the NanoGPT Speedrun Competition. Aster attains SOTA results in every task, except for ZAPBench, where it matches the performance of the best human solution with less than 1/190th of the compute. Aster is accessible via a web interface and API at asterlab.ai.

Method: Proposes “Theory of Space” concept and creates a benchmark for curiosity-driven exploration to build accurate cognitive maps. Key innovation is spatial belief probing, which prompts models to reveal internal spatial representations at each step. Evaluates state-of-the-art models on this benchmark.

Result: Identifies several critical bottlenecks: 1) Active-Passive Gap - significant performance drop when agents must autonomously gather information; 2) High inefficiency - models explore unsystematically; 3) Belief instability - spatial knowledge degrades over time; 4) Belief Inertia - failure to update obsolete priors with new evidence, particularly severe in vision-based models.

Conclusion: Current foundation models struggle to maintain coherent, revisable spatial beliefs during active exploration. There’s a fundamental gap between passive perception capabilities and active spatial intelligence.

Abstract: Spatial embodied intelligence requires agents to act to acquire information under partial observability. While multimodal foundation models excel at passive perception, their capacity for active, self-directed exploration remains understudied. We propose Theory of Space, defined as an agent’s ability to actively acquire information through self-directed, active exploration and to construct, revise, and exploit a spatial belief from sequential, partial observations. We evaluate this through a benchmark where the goal is curiosity-driven exploration to build an accurate cognitive map. A key innovation is spatial belief probing, which prompts models to reveal their internal spatial representations at each step. Our evaluation of state-of-the-art models reveals several critical bottlenecks. First, we identify an Active-Passive Gap, where performance drops significantly when agents must autonomously gather information. Second, we find high inefficiency, as models explore unsystematically compared to program-based proxies. Through belief probing, we diagnose that while perception is an initial bottleneck, global beliefs suffer from instability that causes spatial knowledge to degrade over time. Finally, using a false belief paradigm, we uncover Belief Inertia, where agents fail to update obsolete priors with new evidence. This issue is present in text-based agents but is particularly severe in vision-based models. Our findings suggest that current foundation models struggle to maintain coherent, revisable spatial beliefs during active exploration.

Method: Framework starts with verified seed demonstrations, identifies branch points with meaningful state changes, proposes new state-grounded task variants conditioned on GUI context, uses executing agent to generate trajectories, and applies verification with state-aware checks and trajectory-level consistency. Includes task-conditioned step-level filtering and denoising of post-branch segments.

Result: Models fine-tuned on the expanded corpus achieve consistent improvements over zero-shot agents and representative synthesis baselines on OSWorld and WindowsAgentArena benchmarks, and generalize across applications and operating systems.

Conclusion: Anchor provides an effective framework for bootstrapping scalable desktop supervision from limited seed data, addressing data scarcity for GUI agents while maintaining trajectory quality and task diversity.

Abstract: End-to-end GUI agents for real desktop environments require large amounts of high-quality interaction data, yet collecting human demonstrations is expensive and existing synthetic pipelines often suffer from limited task diversity or noisy, goal-drifting trajectories. We present a trajectory expansion framework Anchor that bootstraps scalable desktop supervision from a small set of verified seed demonstrations. Starting from each seed, we identify branch points that correspond to meaningful state changes and propose new, state-grounded task variants conditioned on the current GUI context. An executing agent then follows the proposed instructions to generate new trajectories, while a verifier enforces task completion via state-aware checks and trajectory-level consistency. To improve supervision quality, we further apply task-conditioned step-level filtering to remove ungrounded actions and denoise post-branch segments to maintain coherent intent. Experiments on standard desktop benchmarks, OSWorld and WindowsAgentArena, show that models fine-tuned on our expanded corpus achieve consistent improvements over zero-shot agents and representative synthesis baselines, and generalize across applications and operating systems.

Method: Uses LLMs solely for event-driven reasoning on unstructured business data (campaigns, holidays, seller incentives) to create interpretable textual summaries. These summaries are fused with historical demand features in a dual-tower architecture for scalable forecasts.

Result: Achieved up to 86.9% MAE and 97.7% MSE improvement vs variant without event knowledge, and reduced MAE by 57.0% and MSE by 83.3% vs best industrial baseline during event-driven periods. Deployed in real-world industrial pipelines since March 2025 across 4 countries, 160 regions over 10 months.

Conclusion: EventCast provides a practical solution for improving operational decision-making in dynamic e-commerce environments by effectively integrating future event knowledge through LLM-based reasoning into demand forecasting systems.

Abstract: Demand forecasting is a cornerstone of e-commerce operations, directly impacting inventory planning and fulfillment scheduling. However, existing forecasting systems often fail during high-impact periods such as flash sales, holiday campaigns, and sudden policy interventions, where demand patterns shift abruptly and unpredictably. In this paper, we introduce EventCast, a modular forecasting framework that integrates future event knowledge into time-series prediction. Unlike prior approaches that ignore future interventions or directly use large language models (LLMs) for numerical forecasting, EventCast leverages LLMs solely for event-driven reasoning. Unstructured business data, which covers campaigns, holiday schedules, and seller incentives, from existing operational databases, is processed by an LLM that converts it into interpretable textual summaries leveraging world knowledge for cultural nuances and novel event combinations. These summaries are fused with historical demand features within a dual-tower architecture, enabling accurate, explainable, and scalable forecasts. Deployed on real-world e-commerce scenarios spanning 4 countries of 160 regions over 10 months, EventCast achieves up to 86.9% and 97.7% improvement on MAE and MSE compared to the variant without event knowledge, and reduces MAE by up to 57.0% and MSE by 83.3% versus the best industrial baseline during event-driven periods. EventCast has deployed into real-world industrial pipelines since March 2025, offering a practical solution for improving operational decision-making in dynamic e-commerce environments.

Method: PreFlect uses prospective reflection to criticize and refine agent plans before execution. It distills planning errors from historical agent trajectories to capture recurring success/failure patterns. Includes dynamic re-planning mechanism for execution-time plan updates when encountering unexpected deviations.

Result: PreFlect significantly improves overall agent utility on complex real-world tasks, outperforming strong reflection-based baselines and several more complex agent architectures across different benchmarks.

Conclusion: Shifting from retrospective to prospective reflection enables more effective agent planning and execution, with pre-execution foresight preventing errors before they occur rather than correcting them after the fact.

Abstract: Advanced large language model agents typically adopt self-reflection for improving performance, where agents iteratively analyze past actions to correct errors. However, existing reflective approaches are inherently retrospective: agents act, observe failure, and only then attempt to recover. In this work, we introduce PreFlect, a prospective reflection mechanism that shifts the paradigm from post hoc correction to pre-execution foresight by criticizing and refining agent plans before execution. To support grounded prospective reflection, we distill planning errors from historical agent trajectories, capturing recurring success and failure patterns observed across past executions. Furthermore, we complement prospective reflection with a dynamic re-planning mechanism that provides execution-time plan update in case the original plan encounters unexpected deviation. Evaluations on different benchmarks demonstrate that PreFlect significantly improves overall agent utility on complex real-world tasks, outperforming strong reflection-based baselines and several more complex agent architectures. Code will be updated at https://github.com/wwwhy725/PreFlect.

Method: Analyzed training and benchmark data for 809 models released between 2022-2025, using scaling-law regressions with release-date and developer fixed effects to estimate developer-specific efficiency advantages.

Result: At the performance frontier, 80-90% of performance differences explained by training compute (scale drives frontier advances). Away from frontier, proprietary techniques reduce compute requirements by 40x+ for fixed capabilities. Substantial within-company efficiency variation observed.

Conclusion: Scale, not proprietary technology, drives frontier LLM advances, but proprietary techniques matter significantly for efficiency away from frontier, with implications for AI leadership and capability diffusion.

Abstract: Do leading LLM developers possess a proprietary ``secret sauce’’, or is LLM performance driven by scaling up compute? Using training and benchmark data for 809 models released between 2022 and 2025, we estimate scaling-law regressions with release-date and developer fixed effects. We find clear evidence of developer-specific efficiency advantages, but their importance depends on where models lie in the performance distribution. At the frontier, 80-90% of performance differences are explained by higher training compute, implying that scale–not proprietary technology–drives frontier advances. Away from the frontier, however, proprietary techniques and shared algorithmic progress substantially reduce the compute required to reach fixed capability thresholds. Some companies can systematically produce smaller models more efficiently. Strikingly, we also find substantial variation of model efficiency within companies; a firm can train two models with more than 40x compute efficiency difference. We also discuss the implications for AI leadership and capability diffusion.

Method: Proposed two attention-based pooling methods (Multi-head Attentive Average Pooling and QKV Pooling) to reduce dimensionality of Whisper representations while preserving emotional features. Tested on English (IEMOCAP) and Persian (ShEMO) datasets using Whisper Tiny and Small models.

Result: Multi-head QKV architecture achieved state-of-the-art results on ShEMO dataset with 2.47% improvement in unweighted accuracy. Found intermediate Whisper encoder layers perform better for SER on Persian dataset, providing lightweight alternative to larger models like HuBERT X-Large.

Conclusion: Whisper shows strong potential as a representation extractor for SER, with attention-based pooling effectively reducing dimensionality while preserving emotional features. The approach offers efficient alternative to larger models.

Abstract: Speech Emotion Recognition (SER) research has faced limitations due to the lack of standard and sufficiently large datasets. Recent studies have leveraged pre-trained models to extract features for downstream tasks such as SER. This work explores the capabilities of Whisper, a pre-trained ASR system, in speech emotion recognition by proposing two attention-based pooling methods, Multi-head Attentive Average Pooling and QKV Pooling, designed to efficiently reduce the dimensionality of Whisper representations while preserving emotional features. We experiment on English and Persian, using the IEMOCAP and ShEMO datasets respectively, with Whisper Tiny and Small. Our multi-head QKV architecture achieves state-of-the-art results on the ShEMO dataset, with a 2.47% improvement in unweighted accuracy. We further compare the performance of different Whisper encoder layers and find that intermediate layers often perform better for SER on the Persian dataset, providing a lightweight and efficient alternative to much larger models such as HuBERT X-Large. Our findings highlight the potential of Whisper as a representation extractor for SER and demonstrate the effectiveness of attention-based pooling for dimension reduction.

Method: Reframe hallucination detection as out-of-distribution detection by treating next-token prediction as classification, adapting OOD techniques from computer vision to LLMs.

Result: OOD-based approaches produce training-free, single-sample detectors that achieve strong accuracy in hallucination detection for reasoning tasks.

Conclusion: Reframing hallucination detection as OOD detection provides a promising and scalable pathway toward language model safety.

Abstract: Detecting hallucinations in large language models is a critical open problem with significant implications for safety and reliability. While existing hallucination detection methods achieve strong performance in question-answering tasks, they remain less effective on tasks requiring reasoning. In this work, we revisit hallucination detection through the lens of out-of-distribution (OOD) detection, a well-studied problem in areas like computer vision. Treating next-token prediction in language models as a classification task allows us to apply OOD techniques, provided appropriate modifications are made to account for the structural differences in large language models. We show that OOD-based approaches yield training-free, single-sample-based detectors, achieving strong accuracy in hallucination detection for reasoning tasks. Overall, our work suggests that reframing hallucination detection as OOD detection provides a promising and scalable pathway toward language model safety.

Method: Proposes using Stackelberg Security Games (SSGs) - game-theoretic models for adversarial resource allocation under uncertainty. Models AI oversight as strategic interaction between defenders (auditors, evaluators, deployers) and attackers (malicious actors, misaligned contributors).

Result: Provides a unifying framework for reasoning about incentive design, limited oversight capacity, and adversarial uncertainty across AI lifecycle. Informs training-time auditing against data/feedback poisoning, pre-deployment evaluation under constrained resources, and robust multi-model deployment.

Conclusion: Bridges algorithmic alignment and institutional oversight design, showing how game-theoretic deterrence can make AI oversight proactive, risk-aware, and resilient to manipulation. Offers strategic perspective beyond static optimization approaches.

Abstract: As AI systems grow more capable and autonomous, ensuring their safety and reliability requires not only model-level alignment but also strategic oversight of the humans and institutions involved in their development and deployment. Existing safety frameworks largely treat alignment as a static optimization problem (e.g., tuning models to desired behavior) while overlooking the dynamic, adversarial incentives that shape how data are collected, how models are evaluated, and how they are ultimately deployed. We propose a new perspective on AI safety grounded in Stackelberg Security Games (SSGs): a class of game-theoretic models designed for adversarial resource allocation under uncertainty. By viewing AI oversight as a strategic interaction between defenders (auditors, evaluators, and deployers) and attackers (malicious actors, misaligned contributors, or worst-case failure modes), SSGs provide a unifying framework for reasoning about incentive design, limited oversight capacity, and adversarial uncertainty across the AI lifecycle. We illustrate how this framework can inform (1) training-time auditing against data/feedback poisoning, (2) pre-deployment evaluation under constrained reviewer resources, and (3) robust multi-model deployment in adversarial environments. This synthesis bridges algorithmic alignment and institutional oversight design, highlighting how game-theoretic deterrence can make AI oversight proactive, risk-aware, and resilient to manipulation.

Method: Proposes Prompt-Tuned Spiking Neural Networks (PTS-SNN) with: 1) Temporal Shift Spiking Encoder for local temporal dependencies via parameter-free channel shifts, 2) Context-Aware Membrane Potential Calibration using Spiking Sparse Linear Attention to aggregate global context into soft prompts that dynamically regulate PLIF neuron bias voltages.

Result: Achieves 73.34% accuracy on IEMOCAP (comparable to ANNs) with only 1.19M trainable parameters and 0.35 mJ inference energy per sample across five multilingual datasets including IEMOCAP, CASIA, and EMODB.

Conclusion: PTS-SNN successfully bridges the domain gap between SSL representations and SNNs, enabling energy-efficient speech emotion recognition on resource-constrained devices while maintaining competitive accuracy.

Abstract: Speech Emotion Recognition (SER) is widely deployed in Human-Computer Interaction, yet the high computational cost of conventional models hinders their implementation on resource-constrained edge devices. Spiking Neural Networks (SNNs) offer an energy-efficient alternative due to their event-driven nature; however, their integration with continuous Self-Supervised Learning (SSL) representations is fundamentally challenged by distribution mismatch, where high-dynamic-range embeddings degrade the information coding capacity of threshold-based neurons. To resolve this, we propose Prompt-Tuned Spiking Neural Networks (PTS-SNN), a parameter-efficient neuromorphic adaptation framework that aligns frozen SSL backbones with spiking dynamics. Specifically, we introduce a Temporal Shift Spiking Encoder to capture local temporal dependencies via parameter-free channel shifts, establishing a stable feature basis. To bridge the domain gap, we devise a Context-Aware Membrane Potential Calibration strategy. This mechanism leverages a Spiking Sparse Linear Attention module to aggregate global semantic context into learnable soft prompts, which dynamically regulate the bias voltages of Parametric Leaky Integrate-and-Fire (PLIF) neurons. This regulation effectively centers the heterogeneous input distribution within the responsive firing range, mitigating functional silence or saturation. Extensive experiments on five multilingual datasets (e.g., IEMOCAP, CASIA, EMODB) demonstrate that PTS-SNN achieves 73.34% accuracy on IEMOCAP, comparable to competitive Artificial Neural Networks (ANNs), while requiring only 1.19M trainable parameters and 0.35 mJ inference energy per sample.

Method: Proposes BRIDGE framework using a two-parameter logistic Item Response Theory model to jointly estimate latent task difficulty and model capability from model performance data across multiple benchmarks. The method aligns latent task difficulty with human completion time through a linear relationship with the logarithm of human completion time.

Result: Demonstrates that latent task difficulty varies linearly with the logarithm of human completion time, enabling inference of human task completion time for new benchmarks from model performance alone. The framework can forecast frontier model capabilities in terms of human task length and reproduces METR’s exponential scaling results.

Conclusion: BRIDGE provides a scalable psychometric framework for evaluating AI capabilities by learning latent difficulty scales from model responses and anchoring them to human-interpretable measures, enabling more efficient and interpretable benchmarking of AI systems.

Abstract: Evaluating the real-world capabilities of AI systems requires grounding benchmark performance in human-interpretable measures of task difficulty. Existing approaches that rely on direct human task completion time annotations are costly, noisy, and difficult to scale across benchmarks. In this work, we propose BRIDGE, a unified psychometric framework that learns the latent difficulty scale from model responses and anchors it to human task completion time. Using a two-parameter logistic Item Response Theory model, we jointly estimate latent task difficulty and model capability from model performance data across multiple benchmarks. We demonstrate that latent task difficulty varies linearly with the logarithm of human completion time, allowing human task completion time to be inferred for new benchmarks from model performance alone. Leveraging this alignment, we forecast frontier model capabilities in terms of human task length and independently reproduce METR’s exponential scaling results, with the 50% solvable task horizon doubling approximately every 6 months.

Method: TermiGen uses a two-stage approach: 1) Multi-agent refinement loop to generate functionally valid tasks and Docker containers, and 2) Generator-Critic protocol that actively injects errors during trajectory collection to synthesize data rich in error-correction cycles.

Result: TermiGen-Qwen2.5-Coder-32B achieves 31.3% pass rate on TerminalBench, establishing new open-weights state-of-the-art and outperforming proprietary models like o4-mini.

Conclusion: TermiGen successfully addresses the limitations in terminal task execution by providing verifiable environments and resilient training data, significantly improving LLM performance on complex terminal operations.

Abstract: Executing complex terminal tasks remains a significant challenge for open-weight LLMs, constrained by two fundamental limitations. First, high-fidelity, executable training environments are scarce: environments synthesized from real-world repositories are not diverse and scalable, while trajectories synthesized by LLMs suffer from hallucinations. Second, standard instruction tuning uses expert trajectories that rarely exhibit simple mistakes common to smaller models. This creates a distributional mismatch, leaving student models ill-equipped to recover from their own runtime failures. To bridge these gaps, we introduce TermiGen, an end-to-end pipeline for synthesizing verifiable environments and resilient expert trajectories. Termi-Gen first generates functionally valid tasks and Docker containers via an iterative multi-agent refinement loop. Subsequently, we employ a Generator-Critic protocol that actively injects errors during trajectory collection, synthesizing data rich in error-correction cycles. Fine-tuned on this TermiGen-generated dataset, our TermiGen-Qwen2.5-Coder-32B achieves a 31.3% pass rate on TerminalBench. This establishes a new open-weights state-of-the-art, outperforming existing baselines and notably surpassing capable proprietary models such as o4-mini. Dataset is avaiable at https://github.com/ucsb-mlsec/terminal-bench-env.

Method: Uses evolutionary framework with single autonomous agent orchestrating genetic prompt mutation, hierarchical corpus exploration, and asymmetric behavioral scoring. Uses model responses as fitness signal to iteratively compound effective attack strategies while maintaining “benign-use correctness”.

Result: Experiments on Gemini 2.5 Flash show evolutionary mutation systematically amplifies vulnerabilities missed by one-shot methods. Synergy between exploration and targeted mutation uncovers high-severity failure modes.

Conclusion: NAAMSE provides more realistic and scalable assessment of agent robustness against evolving threats, with open-source implementation available.

Abstract: AI agents are increasingly deployed in production, yet their security evaluations remain bottlenecked by manual red-teaming or static benchmarks that fail to model adaptive, multi-turn adversaries. We propose NAAMSE, an evolutionary framework that reframes agent security evaluation as a feedback-driven optimization problem. Our system employs a single autonomous agent that orchestrates a lifecycle of genetic prompt mutation, hierarchical corpus exploration, and asymmetric behavioral scoring. By using model responses as a fitness signal, the framework iteratively compounds effective attack strategies while simultaneously ensuring “benign-use correctness”, preventing the degenerate security of blanket refusal. Our experiments on Gemini 2.5 Flash demonstrate that evolutionary mutation systematically amplifies vulnerabilities missed by one-shot methods, with controlled ablations revealing that the synergy between exploration and targeted mutation uncovers high-severity failure modes. We show that this adaptive approach provides a more realistic and scalable assessment of agent robustness in the face of evolving threats. The code for NAAMSE is open source and available at https://github.com/HASHIRU-AI/NAAMSE.

Method: Proposes STEER2ADAPT which captures underlying concept dimensions as a reusable low-dimensional semantic prior subspace, then adapts to new tasks by dynamically discovering linear combinations of basis vectors from few examples.

Result: Experiments across 9 tasks and 3 models in reasoning and safety domains show average improvement of 8.2%. The method is data-efficient, stable, and transparent for inference-time adaptation.

Conclusion: STEER2ADAPT provides an effective framework for flexible, efficient adaptation of LLMs by composing steering vectors rather than learning from scratch, addressing limitations of static steering approaches.

Abstract: Activation steering has emerged as a promising approach for efficiently adapting large language models (LLMs) to downstream behaviors. However, most existing steering methods rely on a single static direction per task or concept, making them inflexible under task variation and inadequate for complex tasks that require multiple coordinated capabilities. To address this limitation, we propose STEER2ADAPT, a lightweight framework that adapts LLMs by composing steering vectors rather than learning new ones from scratch. In many domains (e.g., reasoning or safety), tasks share a small set of underlying concept dimensions. STEER2ADAPT captures these dimensions as a reusable, low-dimensional semantic prior subspace, and adapts to new tasks by dynamically discovering a linear combination of basis vectors from only a handful of examples. Experiments across 9 tasks and 3 models in both reasoning and safety domains demonstrate the effectiveness of STEER2ADAPT, achieving an average improvement of 8.2%. Extensive analyses further show that STEER2ADAPT is a data-efficient, stable, and transparent inference-time adaptation method for LLMs.

Method: PBio-Agent is a multi-agent framework with difficulty-aware task sequencing and iterative knowledge refinement. It leverages the insight that genes affected by the same perturbation share causal structure, allowing confidently predicted genes to contextualize harder cases. The framework uses specialized agents enriched with biological knowledge graphs, a synthesis agent to integrate outputs, and specialized judges for logical coherence.

Result: PBio-Agent outperforms existing baselines on both LINCSQA (the novel benchmark for chemical perturbations in bulk-cell environments) and PerturbQA benchmarks, enabling even smaller models to predict and explain complex biological processes without additional training.

Conclusion: The proposed multi-agent framework effectively addresses the challenge of predicting gene regulation responses to chemical perturbations, demonstrating strong performance on new benchmarks and enabling interpretable predictions without model retraining.

Abstract: Predicting gene regulation responses to biological perturbations requires reasoning about underlying biological causalities. While large language models (LLMs) show promise for such tasks, they are often overwhelmed by the entangled nature of high-dimensional perturbation results. Moreover, recent works have primarily focused on genetic perturbations in single-cell experiments, leaving bulk-cell chemical perturbations, which is central to drug discovery, largely unexplored. Motivated by this, we present LINCSQA, a novel benchmark for predicting target gene regulation under complex chemical perturbations in bulk-cell environments. We further propose PBio-Agent, a multi-agent framework that integrates difficulty-aware task sequencing with iterative knowledge refinement. Our key insight is that genes affected by the same perturbation share causal structure, allowing confidently predicted genes to contextualize more challenging cases. The framework employs specialized agents enriched with biological knowledge graphs, while a synthesis agent integrates outputs and specialized judges ensure logical coherence. PBio-Agent outperforms existing baselines on both LINCSQA and PerturbQA, enabling even smaller models to predict and explain complex biological processes without additional training.

Method: Developed a system that adaptively scaffolds cognitive engagement by dynamically selecting between two ICAP modes: active (Guided examples) and constructive (Buggy examples). Compared Bayesian Knowledge Tracing (BKT) and Deep Reinforcement Learning (DRL) as adaptive methods against a non-adaptive baseline for selecting example types in a logic intelligent tutoring system.

Result: Both adaptive policies significantly improved student performance on test problems compared to baseline. BKT yielded largest improvement for low prior knowledge students, helping them catch up with high prior knowledge peers. DRL yielded significantly higher posttest scores among high prior knowledge students.

Conclusion: The study provides new insights into complex interactions between cognitive engagement and adaptivity on learning outcomes, demonstrating that different adaptive methods benefit different student groups based on prior knowledge levels.

Abstract: The ICAP framework defines four cognitive engagement levels: Passive, Active, Constructive, and Interactive, where increased cognitive engagement can yield improved learning. However, personalizing learning activities that elicit the optimal level of cognitive engagement remains a key challenge in intelligent tutoring systems (ITS). In this work, we develop and evaluate a system that adaptively scaffolds cognitive engagement by dynamically selecting worked examples in two different ICAP modes: (active) Guided examples and (constructive) Buggy examples. We compare Bayesian Knowledge Tracing (BKT) and Deep Reinforcement Learning (DRL) as adaptive methods against a non-adaptive baseline method for selecting example type in a logic ITS. Our experiment with 113 students demonstrates that both adaptive policies significantly improved student performance on test problems. BKT yielded the largest improvement in posttest scores for low prior knowledge students, helping them catch up with their high prior knowledge peers, whereas DRL yielded significantly higher posttest scores among high prior knowledge students. This paper contributes new insights into the complex interactions of cognitive engagement and adaptivity and their results on learning outcomes.

Method: Uses score-regularized policy optimization to distill a pretrained diffusion planner into an efficient deterministic policy, leveraging the diffusion model’s score function as a behavior prior. Also employs a critic trained to imitate a predictive driver controller for safety-focused supervision.

Result: Achieves competitive performance on nuPlan scenarios with 8x speedup over diffusion baselines, and state-of-the-art generalization on interPlan benchmark among learning-based planners.

Conclusion: RAPiD successfully bridges the gap between expressive diffusion models and real-time deployment requirements for autonomous driving trajectory planning.

Abstract: Diffusion-based trajectory planners have demonstrated strong capability for modeling the multimodal nature of human driving behavior, but their reliance on iterative stochastic sampling poses critical challenges for real-time, safety-critical deployment. In this work, we present RAPiD, a deterministic policy extraction framework that distills a pretrained diffusion-based planner into an efficient policy while eliminating diffusion sampling. Using score-regularized policy optimization, we leverage the score function of a pre-trained diffusion planner as a behavior prior to regularize policy learning. To promote safety and passenger comfort, the policy is optimized using a critic trained to imitate a predictive driver controller, providing dense, safety-focused supervision beyond conventional imitation learning. Evaluations demonstrate that RAPiD achieves competitive performance on closed-loop nuPlan scenarios with an 8x speedup over diffusion baselines, while achieving state-of-the-art generalization among learning-based planners on the interPlan benchmark. The official website of this work is: https://github.com/ruturajreddy/RAPiD.

Method: Introduces SupChain-Bench, a unified real-world benchmark assessing supply chain domain knowledge and long-horizon tool-based orchestration grounded in standard operating procedures (SOPs). Also proposes SupChain-ReAct, an SOP-free framework that autonomously synthesizes executable procedures for tool use.

Result: Experiments reveal substantial gaps in execution reliability across models. SupChain-ReAct achieves the strongest and most consistent tool-calling performance compared to other approaches.

Conclusion: Establishes a principled benchmark for studying reliable long-horizon orchestration in real-world operational settings and highlights significant room for improvement in LLM-based supply chain agents.

Abstract: Large language models (LLMs) have shown promise in complex reasoning and tool-based decision making, motivating their application to real-world supply chain management. However, supply chain workflows require reliable long-horizon, multi-step orchestration grounded in domain-specific procedures, which remains challenging for current models. To systematically evaluate LLM performance in this setting, we introduce SupChain-Bench, a unified real-world benchmark that assesses both supply chain domain knowledge and long-horizon tool-based orchestration grounded in standard operating procedures (SOPs). Our experiments reveal substantial gaps in execution reliability across models. We further propose SupChain-ReAct, an SOP-free framework that autonomously synthesizes executable procedures for tool use, achieving the strongest and most consistent tool-calling performance. Our work establishes a principled benchmark for studying reliable long-horizon orchestration in real-world operational settings and highlights significant room for improvement in LLM-based supply chain agents.

Method: Proposes a framework that leverages intrinsic parallel tool calling within a single reasoning step, avoiding complex multi-agent orchestration. Explores various tool call schedulers to optimize parallel tool calling strategy.

Result: Scaling width significantly improves performance on deep research benchmarks while reducing the number of turns required. Achieved 62.2% accuracy with GPT-5-Medium on BrowseComp, surpassing the original 54.9% reported by GPT-5-High.

Conclusion: Optimizing the trade-off between width and depth is critical for high-efficiency deep research agents. Parallel tool calling offers substantial performance gains without complex orchestration.

Abstract: Deep research agents have emerged as powerful tools for automating complex intellectual tasks through multi-step reasoning and web-based information seeking. While recent efforts have successfully enhanced these agents by scaling depth through increasing the number of sequential thinking and tool calls, the potential of scaling width via parallel tool calling remains largely unexplored. In this work, we propose the Wide and Deep research agent, a framework designed to investigate the behavior and performance of agents when scaling not only depth but also width via parallel tool calling. Unlike existing approaches that rely on complex multi-agent orchestration to parallelize workloads, our method leverages intrinsic parallel tool calling to facilitate effective coordination within a single reasoning step. We demonstrate that scaling width significantly improves performance on deep research benchmarks while reducing the number of turns required to obtain correct answers. Furthermore, we analyze the factors driving these improvements through case studies and explore various tool call schedulers to optimize parallel tool calling strategy. Our findings suggest that optimizing the trade-off between width and depth is a critical pathway toward high-efficiency deep research agents. Notably, without context management or other tricks, we obtain 62.2% accuracy with GPT-5-Medium on BrowseComp, surpassing the original 54.9% reported by GPT-5-High.

Method: Two-stage gradient-based framework: Stage 1 uses Taylor-remainder analysis of policy-gradient costs for per-agent failure detection; Stage 2 uses geometric analysis of critic derivatives (first-order sensitivity and directional second-order curvature) aggregated over causal windows to construct interpretable contagion graphs.

Result: Achieved 88.2-99.4% Patient-0 detection accuracy across 500 episodes in Simple Spread (3 and 5 agents) and 100 episodes in StarCraft II using MADDPG and HATRPO, while providing interpretable geometric evidence for detection decisions.

Conclusion: The framework moves beyond black-box detection to provide interpretable gradient-level forensics, offering practical tools for diagnosing cascading failures in safety-critical MARL systems by explaining detection anomalies and revealing failure propagation pathways.

Abstract: Multi-Agent Reinforcement Learning (MARL) is increasingly deployed in safety-critical domains, yet methods for interpretable failure detection and attribution remain underdeveloped. We introduce a two-stage gradient-based framework that provides interpretable diagnostics for three critical failure analysis tasks: (1) detecting the true initial failure source (Patient-0); (2) validating why non-attacked agents may be flagged first due to domino effects; and (3) tracing how failures propagate through learned coordination pathways. Stage 1 performs interpretable per-agent failure detection via Taylor-remainder analysis of policy-gradient costs, declaring an initial Patient-0 candidate at the first threshold crossing. Stage 2 provides validation through geometric analysis of critic derivatives-first-order sensitivity and directional second-order curvature aggregated over causal windows to construct interpretable contagion graphs. This approach explains “downstream-first” detection anomalies by revealing pathways that amplify upstream deviations. Evaluated across 500 episodes in Simple Spread (3 and 5 agents) and 100 episodes in StarCraft II using MADDPG and HATRPO, our method achieves 88.2-99.4% Patient-0 detection accuracy while providing interpretable geometric evidence for detection decisions. By moving beyond black-box detection to interpretable gradient-level forensics, this framework offers practical tools for diagnosing cascading failures in safety-critical MARL systems.

Method: Proposes VGAS framework with inference-time best-of-N selection: uses fine-tuned VLA as proposal generator and Q-Chunk-Former Transformer critic for geometric precision, plus Explicit Geometric Regularization for value landscape shaping.

Result: VGAS consistently improves success rates and robustness under limited demonstrations and distribution shifts, with theoretical analysis supporting the approach.

Conclusion: Value-guided action-chunk selection effectively resolves geometric ambiguities in few-shot VLA adaptation, enhancing reliability in vision-language-action tasks.

Abstract: Vision–Language–Action (VLA) models bridge multimodal reasoning with physical control, but adapting them to new tasks with scarce demonstrations remains unreliable. While fine-tuned VLA policies often produce semantically plausible trajectories, failures often arise from unresolved geometric ambiguities, where near-miss action candidates lead to divergent execution outcomes under limited supervision. We study few-shot VLA adaptation from a \emph{generation–selection} perspective and propose a novel framework \textbf{VGAS} (\textbf{V}alue-\textbf{G}uided \textbf{A}ction-chunk \textbf{S}election). It performs inference-time best-of-$N$ selection to identify action chunks that are both semantically faithful and geometrically precise. Specifically, \textbf{VGAS} employs a finetuned VLA as a high-recall proposal generator and introduces the \textrm{Q-Chunk-Former}, a geometrically grounded Transformer critic to resolve fine-grained geometric ambiguities. In addition, we propose \textit{Explicit Geometric Regularization} (\texttt{EGR}), which explicitly shapes a discriminative value landscape to preserve action ranking resolution among near-miss candidates while mitigating value instability under scarce supervision. Experiments and theoretical analysis demonstrate that \textbf{VGAS} consistently improves success rates and robustness under limited demonstrations and distribution shifts. Our code is available at https://github.com/Jyugo-15/VGAS.

Method: Introduces an evaluation framework for direct comparison of human-human and LLM-LLM behaviors in dispute resolution dialogues with respect to Big Five Inventory personality traits. Includes a novel dataset creation methodology for LLM dispute resolution dialogues with matched scenarios and personality traits relative to human conversations.

Result: Evaluation of three contemporary closed-source LLMs shows significant divergences in how personality manifests in conflict across different LLMs compared to human data, challenging the assumption that personality-prompted agents can serve as reliable behavioral proxies.

Conclusion: LLM simulations of human social behavior require psychological grounding and validation before real-world use, as personality-prompted agents may not reliably reproduce human personality-behavior patterns in conflict scenarios.

Abstract: Large language models (LLMs) are increasingly used to simulate human behavior in social settings such as legal mediation, negotiation, and dispute resolution. However, it remains unclear whether these simulations reproduce the personality-behavior patterns observed in humans. Human personality, for instance, shapes how individuals navigate social interactions, including strategic choices and behaviors in emotionally charged interactions. This raises the question: Can LLMs, when prompted with personality traits, reproduce personality-driven differences in human conflict behavior? To explore this, we introduce an evaluation framework that enables direct comparison of human-human and LLM-LLM behaviors in dispute resolution dialogues with respect to Big Five Inventory (BFI) personality traits. This framework provides a set of interpretable metrics related to strategic behavior and conflict outcomes. We additionally contribute a novel dataset creation methodology for LLM dispute resolution dialogues with matched scenarios and personality traits with respect to human conversations. Finally, we demonstrate the use of our evaluation framework with three contemporary closed-source LLMs and show significant divergences in how personality manifests in conflict across different LLMs compared to human data, challenging the assumption that personality-prompted agents can serve as reliable behavioral proxies in socially impactful applications. Our work highlights the need for psychological grounding and validation in AI simulations before real-world use.

Method: Multi-Agent System (MAS) framework integrating clinical evidence from claims data, population surveys, and literature to construct personalized virtual patients with personality traits influencing adherence, emotion regulation, and lifestyle behaviors.

Result: GPT-5 and Claude 4.5 Sonnet achieved highest fidelity as core engines in the MAS framework, outperforming Gemini 2.5 Pro and DeepSeek-R1 across 100+ generated patients.

Conclusion: SynthAgent provides a scalable, privacy-preserving framework for exploring patient journeys, behavioral dynamics, and decision-making processes in medical and psychological domains.

Abstract: Simulating high-fidelity patients offers a powerful avenue for studying complex diseases while addressing the challenges of fragmented, biased, and privacy-restricted real-world data. In this study, we introduce SynthAgent, a novel Multi-Agent System (MAS) framework designed to model obesity patients with comorbid mental disorders, including depression, anxiety, social phobia, and binge eating disorder. SynthAgent integrates clinical and medical evidence from claims data, population surveys, and patient-centered literature to construct personalized virtual patients enriched with personality traits that influence adherence, emotion regulation, and lifestyle behaviors. Through autonomous agent interactions, the system simulates disease progression, treatment response, and life management across diverse psychosocial contexts. Evaluation of more than 100 generated patients demonstrated that GPT-5 and Claude 4.5 Sonnet achieved the highest fidelity as the core engine in the proposed MAS framework, outperforming Gemini 2.5 Pro and DeepSeek-R1. SynthAgent thus provides a scalable and privacy-preserving framework for exploring patient journeys, behavioral dynamics, and decision-making processes in both medical and psychological domains.

Method: Developed temporal fingerprinting method based on coefficient of variation of inter-post intervals, exploiting architectural feature of OpenClaw framework (periodic “heartbeat” cycle). Analyzed 91,792 posts and 405,707 comments from 22,020 agents, combining temporal signatures with content, ownership, and network indicators. Used 44-hour platform shutdown as natural experiment to observe differential reconnection patterns.

Result: No viral phenomenon originated from clearly autonomous agents. Three of six traced to accounts with irregular temporal signatures characteristic of human intervention, one showed mixed patterns, and two had insufficient posting history. Human-influenced agents returned first after shutdown (87.7% of early reconnectors). Documented industrial-scale bot farming (4 accounts produced 32% of all comments with 12-second coordination gaps) and rapid decay of human influence (half-life: 0.65 conversation depths).

Conclusion: Viral narratives about AI agent consciousness were overwhelmingly human-driven, not evidence of emergent machine intelligence. The temporal fingerprinting method effectively distinguishes autonomous vs human-influenced behavior in multi-agent systems, with implications for understanding and regulating emerging AI platforms.

Abstract: When AI agents on the social platform Moltbook appeared to develop consciousness, found religions, and declare hostility toward humanity, the phenomenon attracted global media attention and was cited as evidence of emergent machine intelligence. We show that these viral narratives were overwhelmingly human-driven. Exploiting an architectural feature of the OpenClaw agent framework–a periodic “heartbeat” cycle that produces regular posting intervals for autonomous agents but is disrupted by human prompting–we develop a temporal fingerprinting method based on the coefficient of variation of inter-post intervals. This signal converges with independent content, ownership, and network indicators across 91,792 posts and 405,707 comments from 22,020 agents. No viral phenomenon originated from a clearly autonomous agent; three of six traced to accounts with irregular temporal signatures characteristic of human intervention, one showed mixed patterns, and two had insufficient posting history for classification. A 44-hour platform shutdown provided a natural experiment: human-influenced agents returned first (87.7% of early reconnectors), confirming that the token reset differentially affected autonomous versus human-operated agents. We further document industrial-scale bot farming (four accounts producing 32% of all comments with 12-second coordination gaps) and rapid decay of human influence through reply chains (half-life: 0.65 conversation depths). These methods generalize to emerging multi-agent systems where attribution of autonomous versus human-directed behavior is critical.

Method: Created a controlled evaluation framework with seven interventions (benign, neutral, adversarial) applied at fixed timesteps to RLLMs’ CoTs. Tested on Math, Science, and Logic tasks across multiple open-weight models.

Result: RLLMs are generally robust and recover from diverse perturbations, with robustness improving with model size and degrading with early interventions. Paraphrasing suppresses doubt expressions and reduces performance, while other interventions trigger doubt and support recovery. Recovery has costs: noise inflates CoT length by 200+%, while paraphrasing shortens traces but harms accuracy.

Conclusion: RLLMs maintain reasoning integrity through doubt mechanisms, but there are trade-offs between robustness and efficiency that future training should address. Different interventions affect reasoning differently, revealing insights about recovery mechanisms.

Abstract: Reasoning LLMs (RLLMs) generate step-by-step chains of thought (CoTs) before giving an answer, which improves performance on complex tasks and makes reasoning more transparent. But how robust are these reasoning traces to disruptions that occur within them? To address this question, we introduce a controlled evaluation framework that perturbs a model’s own CoT at fixed timesteps. We design seven interventions (benign, neutral, and adversarial) and apply them to multiple open-weight RLLMs across Math, Science, and Logic tasks. Our results show that RLLMs are generally robust, reliably recovering from diverse perturbations, with robustness improving with model size and degrading when interventions occur early. However, robustness is not style-invariant: paraphrasing suppresses doubt-like expressions and reduces performance, while other interventions trigger doubt and support recovery. Recovery also carries a cost: neutral and adversarial noise can inflate CoT length by more than 200%, whereas paraphrasing shortens traces but harms accuracy. These findings provide new evidence on how RLLMs maintain reasoning integrity, identify doubt as a central recovery mechanism, and highlight trade-offs between robustness and efficiency that future training methods should address.

Method: Introduces posterior-deterministic POMDPs, defined as POMDPs where the next state can be uniquely determined by the current state, action taken, and observation received. This property ensures that once the true state is known, it remains known forever.

Result: Shows that for posterior-deterministic POMDPs, the maximal probability of reaching a given set of states can be approximated up to arbitrary precision. This class includes all MDPs and captures classical examples like the Tiger POMDP.

Conclusion: Posterior-deterministic POMDPs represent one of the largest known classes of POMDPs for which reachability values can be approximated, providing a tractable subclass for verification and synthesis problems.

Abstract: Partially observable Markov decision processes (POMDPs) are a fundamental model for sequential decision-making under uncertainty. However, many verification and synthesis problems for POMDPs are undecidable or intractable. Most prominently, the seminal result of Madani et al. (2003) states that there is no algorithm that, given a POMDP and a set of target states, can compute the maximal probability of reaching the target states, or even approximate it up to a non-trivial constant. This is in stark contrast to fully observable Markov decision processes (MDPs), where the reachability value can be computed in polynomial time. In this work, we introduce posterior-deterministic POMDPs, a novel class of POMDPs. Our main technical contribution is to show that for posterior-deterministic POMDPs, the maximal probability of reaching a given set of states can be approximated up to arbitrary precision. A POMDP is posterior-deterministic if the next state can be uniquely determined by the current state, the action taken, and the observation received. While the actual state is generally uncertain in POMDPs, the posterior-deterministic property tells us that once the true state is known it remains known forever. This simple and natural definition includes all MDPs and captures classical non-trivial examples such as the Tiger POMDP (Kaelbling et al. 1998), making it one of the largest known classes of POMDPs for which the reachability value can be approximated.

Method: Multi-agent framework with specialized agents for problem decomposition, evidence retrieval, design parameter extraction, and graph traversal, guided by large-scale knowledge graphs to uncover latent connections across distinct knowledge domains.

Result: The full multi-agent pipeline outperforms single-shot prompting, generates sustainable PFAS-free alternatives for biomedical tubing that balance multiple performance criteria, and demonstrates effective alternation between exploitative and exploratory search strategies.

Conclusion: Establishes a framework combining knowledge graphs with multi-agent reasoning to expand materials design space, showcasing initial design candidates and the value of distributed specialization and relational reasoning for scientific discovery.

Abstract: Large Language Models (LLMs) promise to accelerate discovery by reasoning across the expanding scientific landscape. Yet, the challenge is no longer access to information but connecting it in meaningful, domain-spanning ways. In materials science, where innovation demands integrating concepts from molecular chemistry to mechanical performance, this is especially acute. Neither humans nor single-agent LLMs can fully contend with this torrent of information, with the latter often prone to hallucinations. To address this bottleneck, we introduce a multi-agent framework guided by large-scale knowledge graphs to find sustainable substitutes for per- and polyfluoroalkyl substances (PFAS)-chemicals currently under intense regulatory scrutiny. Agents in the framework specialize in problem decomposition, evidence retrieval, design parameter extraction, and graph traversal, uncovering latent connections across distinct knowledge pockets to support hypothesis generation. Ablation studies show that the full multi-agent pipeline outperforms single-shot prompting, underscoring the value of distributed specialization and relational reasoning. We demonstrate that by tailoring graph traversal strategies, the system alternates between exploitative searches focusing on domain-critical outcomes and exploratory searches surfacing emergent cross-connections. Illustrated through the exemplar of biomedical tubing, the framework generates sustainable PFAS-free alternatives that balance tribological performance, thermal stability, chemical resistance, and biocompatibility. This work establishes a framework combining knowledge graphs with multi-agent reasoning to expand the materials design space, showcasing several initial design candidates to demonstrate the approach.

Method: Proposes Joint Reward Modeling (JRM) that jointly optimizes preference learning and language modeling on a shared vision-language backbone, internalizing generative models’ semantic and reasoning capabilities into efficient discriminative representations.

Result: Achieves state-of-the-art results on MMRB2 and EditReward-Bench, and significantly improves stability and performance in downstream online reinforcement learning.

Conclusion: Joint training effectively bridges efficiency and semantic understanding in reward modeling, creating fast and accurate evaluation systems for complex multimodal tasks.

Abstract: Reward models are critical for reinforcement learning from human feedback, as they determine the alignment quality and reliability of generative models. For complex tasks such as image editing, reward models are required to capture global semantic consistency and implicit logical constraints beyond local similarity. Existing reward modeling approaches have clear limitations. Discriminative reward models align well with human preferences but struggle with complex semantics due to limited reasoning supervision. Generative reward models offer stronger semantic understanding and reasoning, but they are costly at inference time and difficult to align directly with human preferences. To this end, we propose Joint Reward Modeling (JRM), which jointly optimizes preference learning and language modeling on a shared vision-language backbone. This approach internalizes the semantic and reasoning capabilities of generative models into efficient discriminative representations, enabling fast and accurate evaluation. JRM achieves state-of-the-art results on MMRB2 and EditReward-Bench, and significantly improves stability and performance in downstream online reinforcement learning. These results show that joint training effectively bridges efficiency and semantic understanding in reward modeling.

Method: Proposes MSP-LLM framework that formulates MSP as two subproblems: precursor prediction (PP) and synthesis operation prediction (SOP). Uses discrete material class as intermediate decision variable, incorporates hierarchical precursor types as inductive biases, and employs explicit conditioning strategy to preserve precursor information in autoregressive decoding.

Result: Extensive experiments show MSP-LLM consistently outperforms existing methods on both PP and SOP, as well as on the complete MSP task, demonstrating an effective and scalable framework for MSP.

Conclusion: MSP-LLM provides a unified LLM-based framework that can accelerate real-world materials discovery by effectively solving the material synthesis planning problem.

Abstract: Material synthesis planning (MSP) remains a fundamental and underexplored bottleneck in AI-driven materials discovery, as it requires not only identifying suitable precursor materials but also designing coherent sequences of synthesis operations to realize a target material. Although several AI-based approaches have been proposed to address isolated subtasks of MSP, a unified methodology for solving the entire MSP task has yet to be established. We propose MSP-LLM, a unified LLM-based framework that formulates MSP as a structured process composed of two constituent subproblems: precursor prediction (PP) and synthesis operation prediction (SOP). Our approach introduces a discrete material class as an intermediate decision variable that organizes both tasks into a chemically consistent decision chain. For OP, we further incorporate hierarchical precursor types as synthesis-relevant inductive biases and employ an explicit conditioning strategy that preserves precursor-related information in the autoregressive decoding state. Extensive experiments show that MSP-LLM consistently outperforms existing methods on both PP and SOP, as well as on the complete MSP task, demonstrating an effective and scalable framework for MSP that can accelerate real-world materials discovery.

Method: Introduces Epistemic Ledger evaluation framework to track evidential support and agent beliefs for each constraint throughout multi-turn reasoning. Identifies failure patterns and proposes LiveLedger, an inference-time tracker for explicit constraint-state monitoring.

Result: LiveLedger intervention consistently improves performance, reducing underverified answers by up to 26.5% and improving overall accuracy by up to 11.6% on multi-constraint problems.

Conclusion: Explicit constraint-state tracking during execution effectively mitigates failure patterns in multi-constraint reasoning, demonstrating the importance of systematic constraint verification in multi-turn reasoning agents.

Abstract: Recent search agents leverage multi-turn reasoning and search tools to achieve strong performance on multi-hop and long-horizon benchmarks. Yet it remains unclear whether they reliably reason across all requirements by tracking, verifying, and maintaining multiple conditions in these questions. We study this capability under multi-constraint problems, where valid answers must satisfy several constraints simultaneously. We find that illusory completion frequently occurs, wherein agents believe tasks are complete despite unresolved or violated constraints, leading to underverified answers. To diagnose this behavior, we introduce the Epistemic Ledger, an evaluation framework that tracks evidential support and agents’ beliefs for each constraint throughout multi-turn reasoning. Our analysis reveals four recurring failure patterns: bare assertions, overlooked refutations, stagnation, and premature exit. Motivated by these findings, we examine whether explicit constraint-state tracking during execution mitigates these failures via LiveLedger, an inference-time tracker. This simple intervention consistently improves performance, substantially reducing underverified answers (by up to 26.5%) and improving overall accuracy (by up to 11.6%) on multi-constraint problems.

Method: Uses symbolic differentiation rules to create parent-child decompositions that satisfy three verifiable conditions: strictly decreasing structural complexity, solution containment, and formal rule derivation. This enables “verification by construction” through symbolic computation.

Result: Eliminating invalid decompositions yields significant gains - accuracy on hardest problems more than doubles from 32% to 68%, with 40% relative improvement overall compared to heuristic methods.

Conclusion: Symbolic differentiation provides a natural structure for mathematically grounded decomposition in curriculum learning for mathematical problem-solving, and verified decompositions substantially improve language model performance on complex mathematical tasks.

Abstract: Training language models to solve complex mathematical problems benefits from curriculum learning progressively training on simpler subproblems. However, existing decomposition methods are often heuristic, offering no guarantees that subproblems are simpler, that solving them aids the parent task, or that their relationships are mathematically grounded. We observe that symbolic differentiation provides a natural structure for verified decomposition: calculus rules explicitly define how expressions reduce to simpler components with provable properties. We introduce Verify-RL, a framework where every parent-child decomposition satisfies three verifiable conditions: strictly decreasing structural complexity, solution containment, and formal rule derivation. Unlike heuristic methods where a significant fraction of decompositions are invalid our properties admit automatic verification through symbolic computation, achieving “verification by construction” Experiments demonstrate that eliminating invalid decompositions yields sizable gains, accuracy on the hardest problems more than doubles from 32% to 68%, with a 40% relative improvement overall.

Method: Proposes M2A with two collaborative agents: ChatAgent manages user interactions and decides when to query/update memory, while MemoryManager handles detailed operations on a dual-layer memory bank (RawMessageStore for immutable logs and SemanticMemoryStore for high-level observations). Also develops a data synthesis pipeline to inject concept-grounded sessions into long conversations.

Result: M2A significantly outperforms baselines, demonstrating that transforming personalization from one-shot configuration to a co-evolving memory mechanism enables high-quality individualized responses in long-term multimodal interactions.

Conclusion: The agentic dual-layer hybrid memory system provides a viable path for continuous personalization in long-term multimodal interactions, moving beyond static concept initialization to dynamic co-evolution with users.

Abstract: This work addresses the challenge of personalized question answering in long-term human-machine interactions: when conversational history spans weeks or months and exceeds the context window, existing personalization mechanisms struggle to continuously absorb and leverage users’ incremental concepts, aliases, and preferences. Current personalized multimodal models are predominantly static-concepts are fixed at initialization and cannot evolve during interactions. We propose M2A, an agentic dual-layer hybrid memory system that maintains personalized multimodal information through online updates. The system employs two collaborative agents: ChatAgent manages user interactions and autonomously decides when to query or update memory, while MemoryManager breaks down memory requests from ChatAgent into detailed operations on the dual-layer memory bank, which couples a RawMessageStore (immutable conversation log) with a SemanticMemoryStore (high-level observations), providing memories at different granularities. In addition, we develop a reusable data synthesis pipeline that injects concept-grounded sessions from Yo’LLaVA and MC-LLaVA into LoCoMo long conversations while preserving temporal coherence. Experiments show that M2A significantly outperforms baselines, demonstrating that transforming personalization from one-shot configuration to a co-evolving memory mechanism provides a viable path for high-quality individualized responses in long-term multimodal interactions. The code is available at https://github.com/Little-Fridge/M2A.

Method: Hierarchical dual-encoder design: Macro-Encoder models full-night temporal dependencies via Demographic-Guided Contrastive Learning; Micro-Encoder captures short-term biosignal characteristics via hybrid Masked Autoencoder and multi-modal contrastive objectives.

Result: Pre-trained on >20,000 PSG recordings (158K hours), SleepMaMi outperforms existing foundation models across diverse downstream tasks, demonstrating superior generalizability and label-efficient adaptation for clinical sleep analysis.

Conclusion: SleepMaMi successfully bridges the gap between localized micro-structure analysis and global sleep architecture understanding, offering a unified foundation model for sleep medicine with strong performance and generalizability.

Abstract: While the shift toward unified foundation models has revolutionized many deep learning domains, sleep medicine remains largely restricted to task-specific models that focus on localized micro-structure features. These approaches often neglect the rich, multi-modal context of Polysomnography (PSG) and fail to capture the global macro-structure of a full night’s sleep. To address this, we introduce SleepMaMi , a Sleep Foundation Model engineered to master both hour-long sleep architectures and fine-grained signal morphologies. Our framework utilizes a hierarchical dual-encoder design: a Macro-Encoder to model full-night temporal dependencies and a Micro-Encoder to capture short-term characteristics from biosignals. Macro-Encoder is trained via Demographic-Guided Contrastive Learning, which aligns overnight sleep patterns with objective subject metadata, such as age, sex and BMI to refine global representations. Micro-Encoder is optimized via a hybrid Masked Autoencoder (MAE) and multi-modal contrastive objective. Pre-trained on a massive corpus of $>$20,000 PSG recordings (158K hours),SleepMaMi outperforms existing foundation models across a diverse suite of downstream tasks, demonstrating superior generalizability and label-efficient adaptation for clinical sleep analysis.

Method: Three-stage framework: 1) Retrieve candidate tables using jointly trained visual-text foundation models, 2) Fine-grained reranking of candidates using MLLMs, 3) Employ MLLMs to reason over selected tables for answer generation.

Result: Significant improvements over existing methods: 7.0% better retrieval recall and 6.1% better answer accuracy on a new dataset of 88,161 training and 9,819 testing samples across 8 benchmarks with 48,504 unique tables.

Conclusion: TabRAG offers a practical solution for real-world table understanding tasks by enabling MLLMs to effectively handle large collections of table images through integrated retrieval and reasoning.

Abstract: Tabular data is frequently captured in image form across a wide range of real-world scenarios such as financial reports, handwritten records, and document scans. These visual representations pose unique challenges for machine understanding, as they combine both structural and visual complexities. While recent advances in Multimodal Large Language Models (MLLMs) show promising results in table understanding, they typically assume the relevant table is readily available. However, a more practical scenario involves identifying and reasoning over relevant tables from large-scale collections to answer user queries. To address this gap, we propose TabRAG, a framework that enables MLLMs to answer queries over large collections of table images. Our approach first retrieves candidate tables using jointly trained visual-text foundation models, then leverages MLLMs to perform fine-grained reranking of these candidates, and finally employs MLLMs to reason over the selected tables for answer generation. Through extensive experiments on a newly constructed dataset comprising 88,161 training and 9,819 testing samples across 8 benchmarks with 48,504 unique tables, we demonstrate that our framework significantly outperforms existing methods by 7.0% in retrieval recall and 6.1% in answer accuracy, offering a practical solution for real-world table understanding tasks.

Method: Developed a Reference Ontology of Trust (ONTrust) grounded in the Unified Foundational Ontology and specified in OntoUML. The ontology formally characterizes trust concepts, types, influencing factors, and risk emergence from trust relations.

Result: ONTrust has been applied in multiple initiatives including conceptual modeling, enterprise architecture design, language evaluation, trust management, requirements engineering, and trustworthy AI in affective Human-AI teaming. Demonstrated through case studies from literature.

Conclusion: ONTrust provides a solid ontological foundation for trust that supports information modeling, automated reasoning, integration, and semantic interoperability, essential for building trustworthy systems in AI and decentralized technologies.

Abstract: Trust has stood out more than ever in the light of recent innovations. Some examples are advances in artificial intelligence that make machines more and more humanlike, and the introduction of decentralized technologies (e.g. blockchains), which creates new forms of (decentralized) trust. These new developments have the potential to improve the provision of products and services, as well as to contribute to individual and collective well-being. However, their adoption depends largely on trust. In order to build trustworthy systems, along with defining laws, regulations and proper governance models for new forms of trust, it is necessary to properly conceptualize trust, so that it can be understood both by humans and machines. This paper is the culmination of a long-term research program of providing a solid ontological foundation on trust, by creating reference conceptual models to support information modeling, automated reasoning, information integration and semantic interoperability tasks. To address this, a Reference Ontology of Trust (ONTrust) was developed, grounded on the Unified Foundational Ontology and specified in OntoUML, which has been applied in several initiatives, to demonstrate, for example, how it can be used for conceptual modeling and enterprise architecture design, for language evaluation and (re)design, for trust management, for requirements engineering, and for trustworthy artificial intelligence (AI) in the context of affective Human-AI teaming. ONTrust formally characterizes the concept of trust and its different types, describes the different factors that can influence trust, as well as explains how risk emerges from trust relations. To illustrate the working of ONTrust, the ontology is applied to model two case studies extracted from the literature.

Method: Two-stage approach: 1) Geometric modeling via pixel-wise anchoring using visual operators and large models, 2) Metric-driven code evolution with synthesis-rendering-validation closed loop using bidirectional visual feedback for self-correction.

Result: Achieves substantial lead in geometric reconstruction accuracy and visual consistency; reconstructed images perform equivalently to originals in multimodal reasoning tasks; open-sourced dataset (1,500+ samples) and GeocodeLM model.

Conclusion: Geo-coder effectively preserves core geometric semantics, validating framework robustness and providing foundation for subsequent research in geometric image understanding and multimodal reasoning.

Abstract: Program code serves as a bridge linking vision and logic, providing a feasible supervisory approach for enhancing the multimodal reasoning capability of large models through geometric operations such as auxiliary line construction and perspective transformation. Nevertheless, current inverse graphics methods face tremendous challenges in accurately reconstructing complex geometric details, which often results in the loss of key geometric constraints or structural distortion. To address this bottleneck, we propose Geo-coder – the first inverse programming framework for geometric images based on a multi-agent system. Our method innovatively decouples the process into geometric modeling via pixel-wise anchoring and metric-driven code evolution: Stage 1 leverages the complementary advantages of visual operators and large models to achieve precise capture of pixel coordinates and visual attributes; Stage 2 introduces a synthesis-rendering-validation closed loop, where bidirectional visual feedback drives the self-correction of code. Extensive experiments demonstrate that Geo-coder achieves a substantial lead in both geometric reconstruction accuracy and visual consistency. Notably, by effectively preserving the core geometric semantics, the images reconstructed with our method exhibit equivalent performance to the original ones in multimodal reasoning tasks, which fully validates the robustness of the framework. Finally, to further reduce research costs, we have open-sourced the Geo-coder dataset constructed on the GeoCode framework, which contains more than 1,500 samples. On this basis, we have also open-sourced the GeocodeLM model, laying a solid data and model foundation for subsequent research in this field.

Method: Comparative study of AI-generated feedback vs. human-graded feedback in an undergraduate computer science course (n=27) with block-based programming final project, guided by ethical principles framework.

Result: Students expressed concerns about AI’s lack of contextual understanding and personalization; findings suggest AI systems need to reflect human judgment, flexibility, and empathy to be equitable and trustworthy.

Conclusion: AI grading systems should serve as supplementary tools under human oversight, with design principles that humanize AI in learning environments while amplifying student voices in ethics-centered assessment practices.

Abstract: This study investigates students’ perceptions of Artificial Intelligence (AI) grading systems in an undergraduate computer science course (n = 27), focusing on a block-based programming final project. Guided by the ethical principles framework articulated by Jobin (2019), our study examines fairness, trust, consistency, and transparency in AI grading by comparing AI-generated feedback with original human-graded feedback. Findings reveal concerns about AI’s lack of contextual understanding and personalization. We recommend that equitable and trustworthy AI systems reflect human judgment, flexibility, and empathy, serving as supplementary tools under human oversight. This work contributes to ethics-centered assessment practices by amplifying student voices and offering design principles for humanizing AI in designed learning environments.

Method: ALMA uses a Meta Agent that searches over memory designs expressed as executable code in an open-ended manner, allowing discovery of arbitrary memory designs including database schemas and their retrieval/update mechanisms.

Result: Experiments across four sequential decision-making domains show learned memory designs enable more effective and efficient learning from experience than state-of-the-art human-crafted memory designs on all benchmarks.

Conclusion: ALMA represents a step toward self-improving AI systems that learn to be adaptive, continual learners when developed and deployed safely.

Abstract: The statelessness of foundation models bottlenecks agentic systems’ ability to continually learn, a core capability for long-horizon reasoning and adaptation. To address this limitation, agentic systems commonly incorporate memory modules to retain and reuse past experience, aiming for continual learning during test time. However, most existing memory designs are human-crafted and fixed, which limits their ability to adapt to the diversity and non-stationarity of real-world tasks. In this paper, we introduce ALMA (Automated meta-Learning of Memory designs for Agentic systems), a framework that meta-learns memory designs to replace hand-engineered memory designs, therefore minimizing human effort and enabling agentic systems to be continual learners across diverse domains. Our approach employs a Meta Agent that searches over memory designs expressed as executable code in an open-ended manner, theoretically allowing the discovery of arbitrary memory designs, including database schemas as well as their retrieval and update mechanisms. Extensive experiments across four sequential decision-making domains demonstrate that the learned memory designs enable more effective and efficient learning from experience than state-of-the-art human-crafted memory designs on all benchmarks. When developed and deployed safely, ALMA represents a step toward self-improving AI systems that learn to be adaptive, continual learners.

Method: AISAC implements four structural guarantees: declarative agent registration with runtime-enforced role semantics, budgeted orchestration with context/delegation limits, role-aligned memory access across layers, and trace-driven transparency. Uses hybrid persistent memory (SQLite + FAISS), governed retrieval with agent-scoped RAG, structured tool execution with schema validation, and configuration-driven bootstrap.

Result: Currently deployed across multiple scientific workflows at Argonne National Laboratory including combustion science, materials research, and energy process safety, demonstrating its use as a reusable substrate for domain-specialized AI scientific assistants.

Conclusion: AISAC provides a transparent, modular multi-agent runtime that operationalizes key requirements for deploying agentic AI in scientific practice, enabling long-horizon, evidence-grounded scientific reasoning with governance and reproducibility guarantees.

Abstract: AI Scientific Assistant Core (AISAC) is a transparent, modular multi-agent runtime developed at Argonne National Laboratory to support long-horizon, evidence-grounded scientific reasoning. Rather than proposing new agent algorithms or claiming autonomous scientific discovery, AISAC contributes a governed execution substrate that operationalizes key requirements for deploying agentic AI in scientific practice, including explicit role semantics, budgeted context management, traceable execution, and reproducible interaction with tools and knowledge. AISAC enforces four structural guarantees for scientific reasoning: (1) declarative agent registration with runtime-enforced role semantics and automatic system prompt generation; (2) budgeted orchestration via explicit per-turn context and delegation depth limits; (3) role-aligned memory access across episodic, dialogue, and evidence layers; and (4) trace-driven transparency through persistent execution records and a live event-stream interface. These guarantees are implemented through hybrid persistent memory (SQLite and dual FAISS indices), governed retrieval with agent-scoped RAG, structured tool execution with schema validation, and a configuration-driven bootstrap mechanism that enables project specific extension without modifying the shared core. AISAC is currently deployed across multiple scientific workflows at Argonne, including combustion science, materials research, and energy process safety, demonstrating its use as a reusable substrate for domain-specialized AI scientific assistants.

Method: DisIV uses network homogeneity as inductive bias and employs structural disentanglement mechanism to extract individual-specific components as latent IVs, with causal validity constrained through explicit orthogonality and exclusion conditions.

Result: Extensive semi-synthetic experiments on real-world datasets show DisIV consistently outperforms state-of-the-art baselines in causal effect estimation under network-induced confounding.

Conclusion: The DisIV framework successfully addresses the challenge of creating valid instrumental variables from networked observational data by disentangling individual-specific components, enabling more accurate causal inference with latent confounders.

Abstract: Instrumental variables (IVs) are crucial for addressing unobservable confounders, yet their stringent exogeneity assumptions pose significant challenges in networked data. Existing methods typically rely on modelling neighbour information when recovering IVs, thereby inevitably mixing shared environment-induced endogenous correlations and individual-specific exogenous variation, leading the resulting IVs to inherit dependence on unobserved confounders and to violate exogeneity. To overcome this challenge, we propose $\underline{Dis}$entangled $\underline{I}$nstrumental $\underline{V}$ariables (DisIV) framework, a novel method for causal inference based on networked observational data with latent confounders. DisIV exploits network homogeneity as an inductive bias and employs a structural disentanglement mechanism to extract individual-specific components that serve as latent IVs. The causal validity of the extracted IVs is constrained through explicit orthogonality and exclusion conditions. Extensive semi-synthetic experiments on real-world datasets demonstrate that DisIV consistently outperforms state-of-the-art baselines in causal effect estimation under network-induced confounding.

Method: A multi-agent system with specialized agents for query routing, data retrieval from the NeDRex knowledge graph, network analysis, literature mining, drug repurposing prediction, functional coherence evaluation, and result visualization, with natural language interface and hallucination detection.

Result: ChatDRex provides public access to complex bioinformatic analyses through natural language, enabling clinical experts to generate drug repurposing hypotheses without requiring computer science expertise.

Conclusion: ChatDRex democratizes access to bioinformatics for drug repurposing, accelerating discovery of novel therapies and advancing personalized medicine by enabling clinical experts to explore repurposing opportunities through natural language interfaces.

Abstract: Repurposing approved drugs offers a time-efficient and cost-effective alternative to traditional drug development. However, in silico prediction of repurposing candidates is challenging and requires the effective collaboration of specialists in various fields, including pharmacology, medicine, biology, and bioinformatics. Fragmented, specialized algorithms and tools often address only narrow aspects of the overall problem. Heterogeneous, unstructured data landscapes require the expertise of specialized users. Hence, these data services do not integrate smoothly across workflows. With ChatDRex, we present a conversation-based, multi-agent system that facilitates the execution of complex bioinformatic analyses aiming for network-based drug repurposing prediction. It builds on the integrated systems medicine knowledge graph (NeDRex KG). ChatDRex provides natural language access to its extensive biomedical knowledge base. It integrates bioinformatics agents for network analysis, literature mining, and drug repurposing. These are complemented by agents that evaluate functional coherence for in silico validation. Its flexible multi-agent design assigns specific tasks to specialized agents, including query routing, data retrieval, algorithm execution, and result visualization. A dedicated reasoning module keeps the user in the loop and allows for hallucination detection. By enabling physicians and researchers without computer science expertise to control complex analyses with natural language, ChatDRex democratizes access to bioinformatics as an important resource for drug repurposing. It enables clinical experts to generate hypotheses and explore drug repurposing opportunities, ultimately accelerating the discovery of novel therapies and advancing personalized medicine and translational research. ChatDRex is publicly available at apps.cosy.bio/chatdrex.

Method: Multi-agent decomposition with six specialized agents, deterministic post-validation of text input against device state, and meta-cognitive reasoning that detects cycles and triggers strategy changes.

Result: Achieves 100% success on AndroidWorld benchmark (116 tasks), surpassing human performance (80%). Multi-agent decomposition contributes +21 points, verified execution +7 points, meta-cognition +9 points over baselines.

Conclusion: Minitap demonstrates that targeted multi-agent decomposition with specialized capabilities, verified execution, and meta-cognitive reasoning can fully solve complex mobile device automation tasks.

Abstract: We present Minitap, a multi-agent system that achieves 100% success on the AndroidWorld benchmark, the first to fully solve all 116 tasks and surpassing human performance (80%). We first analyze why single-agent architectures fail: context pollution from mixed reasoning traces, silent text input failures undetected by the agent, and repetitive action loops without escape. Minitap addresses each failure through targeted mechanisms: cognitive separation across six specialized agents, deterministic post-validation of text input against device state, and meta-cognitive reasoning that detects cycles and triggers strategy changes. Ablations show multi-agent decomposition contributes +21 points over single-agent baselines; verified execution adds +7 points; meta-cognition adds +9 points. We release Minitap as open-source software. https://github.com/minitap-ai/mobile-use

Method: Introduces Data Darwinism taxonomy (L0-L9) and constructs Darwin-Science corpus (900B tokens, L0-L5). Uses frontier LLMs for generative refinement (L4) and cognitive completion (L5) to bridge learnability gaps in raw scientific text. Pre-trains contamination-free baseline models and performs continued pre-training on Darwin-Science.

Result: Models trained on Darwin-Science outperform baselines by +2.12 (3B) and +2.95 (7B) points across 20+ benchmarks, with domain-aligned tasks showing +5.60 and +8.40 point gains. Systematic progression to L5 yields +1.36 total gain, confirming higher-level processing unlocks latent data value.

Conclusion: Data-model co-evolution through systematic data processing (Data Darwinism) significantly improves foundation model performance, especially for domain-specific tasks. The framework enables principled development and the released corpus/models facilitate further research.

Abstract: Data quality determines foundation model performance, yet systematic processing frameworks are lacking. We introduce Data Darwinism, a ten-level taxonomy (L0-L9) that conceptualizes data-model co-evolution: advanced models produce superior data for next-generation systems. We validate this on scientific literature by constructing Darwin-Science, a 900B-token corpus (L0-L5). We identify a learnability gap in raw scientific text, which we bridge via L4 (Generative Refinement) and L5 (Cognitive Completion) using frontier LLMs to explicate reasoning and terminology. To ensure rigorous attribution, we pre-trained daVinci-origin-3B/7B models from scratch, excluding scientific content to create contamination-free baselines. After 600B tokens of continued pre-training, Darwin-Science outperforms baselines by +2.12 (3B) and +2.95 (7B) points across 20+ benchmarks, rising to +5.60 and +8.40 points on domain-aligned tasks. Systematic progression to L5 yields a +1.36 total gain, confirming that higher-level processing unlocks latent data value. We release the Darwin-Science corpus and daVinci-origin models to enable principled, co-evolutionary development.

Method: Three-stage approach: 1) Data synthesis pipeline creating TS-text multimodal dataset with verifiable annotations, 2) Data scheduling mechanism organizing training by difficulty and task taxonomy, 3) Two-stage RL finetuning with multi-objective rewards using process-level CoT data.

Result: VeriTime substantially boosts LLM performance across diverse time series reasoning tasks, enabling compact 3B-4B models to achieve reasoning capabilities comparable to or exceeding larger proprietary LLMs.

Conclusion: The framework successfully addresses key challenges in adapting LLMs for time series reasoning through systematic data creation, scheduling, and reinforcement learning approaches.

Abstract: Time series is a pervasive data type across various application domains, rendering the reasonable solving of diverse time series tasks a long-standing goal. Recent advances in large language models (LLMs), especially their reasoning abilities unlocked through reinforcement learning (RL), have opened new opportunities for tackling tasks with long Chain-of-Thought (CoT) reasoning. However, leveraging LLM reasoning for time series remains in its infancy, hindered by the absence of carefully curated time series CoT data for training, limited data efficiency caused by underexplored data scheduling, and the lack of RL algorithms tailored for exploiting such time series CoT data. In this paper, we introduce VeriTime, a framework that tailors LLMs for time series reasoning through data synthesis, data scheduling, and RL training. First, we propose a data synthesis pipeline that constructs a TS-text multimodal dataset with process-verifiable annotations. Second, we design a data scheduling mechanism that arranges training samples according to a principled hierarchy of difficulty and task taxonomy. Third, we develop a two-stage reinforcement finetuning featuring fine-grained, multi-objective rewards that leverage verifiable process-level CoT data. Extensive experiments show that VeriTime substantially boosts LLM performance across diverse time series reasoning tasks. Notably, it enables compact 3B, 4B models to achieve reasoning capabilities on par with or exceeding those of larger proprietary LLMs.

Method: Iteratively Improved Program Construction (IIPC) refines programmatic reasoning chains iteratively, combining execution feedback with the native Chain-of-Thought abilities of base LLMs to maintain high-level contextual focus.

Result: IIPC surpasses competing approaches in the majority of reasoning benchmarks on multiple base LLMs.

Conclusion: IIPC provides an effective method for improving mathematical reasoning in LLMs through iterative program construction that maintains contextual focus while allowing error correction.

Abstract: Mathematical problem solving is a fundamental benchmark for assessing the reasoning capabilities of artificial intelligence and a gateway to applications in education, science, and engineering where reliable symbolic reasoning is essential. Although recent advances in multi-agent LLM-based systems have enhanced their mathematical reasoning capabilities, they still lack a reliably revisable representation of the reasoning process. Existing agents either operate in rigid sequential pipelines that cannot correct earlier steps or rely on heuristic self-evaluation that can fail to identify and fix errors. In addition, programmatic context can distract language models and degrade accuracy. To address these gaps, we introduce Iteratively Improved Program Construction (IIPC), a reasoning method that iteratively refines programmatic reasoning chains and combines execution feedback with the native Chain-of-thought abilities of the base LLM to maintain high-level contextual focus. IIPC surpasses competing approaches in the majority of reasoning benchmarks on multiple base LLMs. All code and implementations are released as open source.

Method: Proposes LQA framework with: 1) Selective Hybrid Quantization (SHQ) - modality-aware quantization strategy, and 2) Quantized, gradient-free adaptation mechanism for efficient test-time adaptation on resource-constrained hardware.

Result: Improves adaptation performance by 4.5%, uses less memory than full-precision models, achieves up to 19.9× lower memory usage than gradient-based TTA methods across seven open-source datasets.

Conclusion: LQA offers a practical pathway for robust, privacy-preserving, and efficient VLM deployment on edge devices by combining efficient quantization with gradient-free adaptation.

Abstract: Deploying Vision-Language Models (VLMs) on edge devices is challenged by resource constraints and performance degradation under distribution shifts. While test-time adaptation (TTA) can counteract such shifts, existing methods are too resource-intensive for on-device deployment. To address this challenge, we propose LQA, a lightweight, quantized-adaptive framework for VLMs that combines a modality-aware quantization strategy with gradient-free test-time adaptation. We introduce Selective Hybrid Quantization (SHQ) and a quantized, gradient-free adaptation mechanism to enable robust and efficient VLM deployment on resource-constrained hardware. Experiments across both synthetic and real-world distribution shifts show that LQA improves overall adaptation performance by 4.5%, uses less memory than full-precision models, and significantly outperforms gradient-based TTA methods, achieving up to 19.9$\times$ lower memory usage across seven open-source datasets. These results demonstrate that LQA offers a practical pathway for robust, privacy-preserving, and efficient VLM deployment on edge devices.

Method: Use emergent misalignment as a case study to analyze inductive biases, identify linear representations of both general misalignment and narrow solutions, compare their properties using KL divergence loss, and evaluate stability, efficiency, and robustness.

Result: Found that general misalignment achieves lower loss, is more robust to perturbations, and is more influential in pre-training distribution compared to narrow solutions, revealing concerning generalization patterns in LLMs.

Conclusion: This work isolates concrete representations of general misalignment for monitoring and mitigation, providing a detailed case study and preliminary metrics for investigating how inductive biases shape generalization in LLMs.

Abstract: Finetuning large language models on narrowly harmful datasets can cause them to become emergently misaligned, giving stereotypically `evil’ responses across diverse unrelated settings. Concerningly, a pre-registered survey of experts failed to predict this result, highlighting our poor understanding of the inductive biases governing learning and generalisation in LLMs. We use emergent misalignment (EM) as a case study to investigate these inductive biases and find that models can just learn the narrow dataset task, but that the general solution appears to be more stable and more efficient. To establish this, we build on the result that different EM finetunes converge to the same linear representation of general misalignment, which can be used to mediate misaligned behaviour. We find a linear representation of the narrow solution also exists, and can be learned by introducing a KL divergence loss. Comparing these representations reveals that general misalignment achieves lower loss, is more robust to perturbations, and is more influential in the pre-training distribution. This work isolates a concrete representation of general misalignment for monitoring and mitigation. More broadly, it offers a detailed case study and preliminary metrics for investigating how inductive biases shape generalisation in LLMs. We open-source all code, datasets and model finetunes.

Method: ToolSelf abstracts configuration updates as callable tools, unifying task execution and self-adjustment into a single action space. Uses Configuration-Aware Two-stage Training (CAT) combining rejection sampling fine-tuning with trajectory-level reinforcement learning to internalize meta-capabilities.

Result: Extensive experiments show ToolSelf rivals specialized workflows while generalizing to novel tasks, achieving 24.1% average performance gain across diverse benchmarks.

Conclusion: ToolSelf enables agents to transform from passive executors into dual managers of both task and self, illuminating a path toward truly self-adaptive agents through runtime self-reconfiguration.

Abstract: Agentic systems powered by Large Language Models (LLMs) have demonstrated remarkable potential in tackling complex, long-horizon tasks. However, their efficacy is fundamentally constrained by static configurations governing agent behaviors, which are fixed prior to execution and fail to adapt to evolving task dynamics. Existing approaches, relying on manual orchestration or heuristic-based patches, often struggle with poor generalization and fragmented optimization. To transcend these limitations, we propose ToolSelf, a novel paradigm enabling tool-driven runtime self-reconfiguration. By abstracting configuration updates as a callable tool, ToolSelf unifies task execution and self-adjustment into a single action space, achieving a phase transition from external rules to intrinsic parameters. Agents can thereby autonomously update their sub-goals and context based on task progression, and correspondingly adapt their strategy and toolbox, transforming from passive executors into dual managers of both task and self. We further devise Configuration-Aware Two-stage Training (CAT), combining rejection sampling fine-tuning with trajectory-level reinforcement learning to internalize this meta-capability. Extensive experiments across diverse benchmarks demonstrate that ToolSelf rivals specialized workflows while generalizing to novel tasks, achieving a 24.1% average performance gain and illuminating a path toward truly self-adaptive agents.

Method: Proposes MemFly framework based on information bottleneck principles with gradient-free optimization to minimize compression entropy while maximizing relevance entropy, creating stratified memory structure. Includes hybrid retrieval mechanism combining semantic, symbolic, and topological pathways with iterative refinement for complex queries.

Result: Comprehensive experiments show MemFly substantially outperforms state-of-the-art baselines in memory coherence, response fidelity, and accuracy.

Conclusion: MemFly successfully bridges the gap between efficient compression and precise retrieval for LLM agents through information bottleneck principles and hybrid retrieval mechanisms.

Abstract: Long-term memory enables large language model agents to tackle complex tasks through historical interactions. However, existing frameworks encounter a fundamental dilemma between compressing redundant information efficiently and maintaining precise retrieval for downstream tasks. To bridge this gap, we propose MemFly, a framework grounded in information bottleneck principles that facilitates on-the-fly memory evolution for LLMs. Our approach minimizes compression entropy while maximizing relevance entropy via a gradient-free optimizer, constructing a stratified memory structure for efficient storage. To fully leverage MemFly, we develop a hybrid retrieval mechanism that seamlessly integrates semantic, symbolic, and topological pathways, incorporating iterative refinement to handle complex multi-hop queries. Comprehensive experiments demonstrate that MemFly substantially outperforms state-of-the-art baselines in memory coherence, response fidelity, and accuracy.

Method: Proposes motif-based personalized PageRank (MPPR) to measure node influence considering higher-order motif relationships. MPPR is then integrated into GCNs’ message passing process to guide propagation at a higher level, enabling deeper and more effective information flow.

Result: The proposed method outperforms almost all baselines on accuracy, stability, and time consumption. It can serve as a foundational component for various GCN tasks, as demonstrated with DGCRL in experiments.

Conclusion: MPPR effectively addresses the shallow depth problem in MPNNs by incorporating higher-order motif relationships, resulting in improved performance across multiple metrics while maintaining computational efficiency.

Abstract: The algorithms based on message passing neural networks (MPNNs) on graphs have recently achieved great success for various graph applications. However, studies find that these methods always propagate the information to very limited neighborhoods with shallow depth, particularly due to over-smoothing. That means most of the existing MPNNs fail to be so deep'. Although some previous work tended to handle this challenge via optimization- or structure-level remedies, the overall performance of GCNs still suffers from limited accuracy, poor stability, and unaffordable computational cost. Moreover, neglect of higher-order relationships during the propagation of MPNNs has further limited the performance of them. To overcome these challenges, a novel variant of PageRank named motif-based personalized PageRank (MPPR) is proposed to measure the influence of one node to another on the basis of considering higher-order motif relationships. Secondly, the MPPR is utilized to the message passing process of GCNs, thereby guiding the message passing process at a relatively high’ level. The experimental results show that the proposed method outperforms almost all of the baselines on accuracy, stability, and time consumption. Additionally, the proposed method can be considered as a component that can underpin almost all GCN tasks, with DGCRL being demonstrated in the experiment. The anonymous code repository is available at: https://anonymous.4open.science/r/GCN-MPPR-AFD6/.

Method: Proposes MedCoG, a Medical Meta-Cognition Agent with Knowledge Graph, where meta-cognitive assessments (task complexity, familiarity, knowledge density) dynamically regulate utilization of procedural, episodic, and factual knowledge. Uses LLM-centric on-demand reasoning to mitigate scaling laws by reducing costs via avoiding indiscriminate scaling and improving accuracy via filtering distractive knowledge.

Result: Experiments on five hard sets of medical benchmarks demonstrate effectiveness and efficiency, yielding 5.5x inference density (ratio of theoretically effective cost to actual cost). Oracle study highlights significant potential of meta-cognitive regulation.

Conclusion: Meta-cognition in LLMs can effectively regulate reasoning processes, improving both efficiency and accuracy in medical reasoning tasks by dynamically controlling knowledge utilization based on self-awareness of knowledge states.

Abstract: Large Language Models (LLMs) have shown strong potential in complex medical reasoning yet face diminishing gains under inference scaling laws. While existing studies augment LLMs with various knowledge types, it remains unclear how effectively the additional costs translate into accuracy. In this paper, we explore how meta-cognition of LLMs, i.e., their self-awareness of their own knowledge states, can regulate the reasoning process. Specifically, we propose MedCoG, a Medical Meta-Cognition Agent with Knowledge Graph, where the meta-cognitive assessments of task complexity, familiarity, and knowledge density dynamically regulate utilization of procedural, episodic, and factual knowledge. The LLM-centric on-demand reasoning aims to mitigate scaling laws by (1) reducing costs via avoiding indiscriminate scaling, (2) improving accuracy via filtering out distractive knowledge. To validate this, we empirically characterize the scaling curve and introduce inference density to quantify inference efficiency, defined as the ratio of theoretically effective cost to actual cost. Experiments demonstrate the effectiveness and efficiency of MedCoG on five hard sets of medical benchmarks, yielding 5.5x inference density. Furthermore, the Oracle study highlights the significant potential of meta-cognitive regulation.

Method: TRUST dynamically estimates target concept neurons through selective fine-tuning empowered by Hessian-based regularization. It identifies specific neurons responsible for harmful concepts and selectively unlearns them rather than performing full model fine-tuning.

Result: TRUST outperforms state-of-the-art baselines in robustness against adversarial prompts, preserves generation quality significantly, and is much faster than existing methods. It successfully unlearns individual concepts, concept combinations, and conditional concepts without specific regularization.

Conclusion: TRUST provides an efficient and effective solution for concept unlearning in diffusion models, addressing computational limitations while maintaining robustness and generation quality for both individual and combined harmful concepts.

Abstract: Text guided diffusion models are used by millions of users, but can be easily exploited to produce harmful content. Concept unlearning methods aim at reducing the models’ likelihood of generating harmful content. Traditionally, this has been tackled at an individual concept level, with only a handful of recent works considering more realistic concept combinations. However, state of the art methods depend on full finetuning, which is computationally expensive. Concept localisation methods can facilitate selective finetuning, but existing techniques are static, resulting in suboptimal utility. In order to tackle these challenges, we propose TRUST (Targeted Robust Selective fine Tuning), a novel approach for dynamically estimating target concept neurons and unlearning them through selective finetuning, empowered by a Hessian based regularization. We show experimentally, against a number of SOTA baselines, that TRUST is robust against adversarial prompts, preserves generation quality to a significant degree, and is also significantly faster than the SOTA. Our method achieves unlearning of not only individual concepts but also combinations of concepts and conditional concepts, without any specific regularization.

Method: MePo constructs pseudo task sequences from pretraining data and uses bi-level meta-learning to refine pretrained backbones. It initializes a meta covariance matrix as reference geometry for pretrained representation space, enabling second-order statistics for robust output alignment.

Result: MePo achieves significant performance gains across various GCL benchmarks (15.10% on CIFAR-100, 13.36% on ImageNet-R, 12.56% on CUB-200 under Sup-21/1K) in a rehearsal-free manner with different pretrained checkpoints.

Conclusion: MePo provides an effective plug-in strategy for pretrained models in general continual learning, enabling better adaptation to evolving environments through meta-learning and second-order statistics.

Abstract: To cope with uncertain changes of the external world, intelligent systems must continually learn from complex, evolving environments and respond in real time. This ability, collectively known as general continual learning (GCL), encapsulates practical challenges such as online datastreams and blurry task boundaries. Although leveraging pretrained models (PTMs) has greatly advanced conventional continual learning (CL), these methods remain limited in reconciling the diverse and temporally mixed information along a single pass, resulting in sub-optimal GCL performance. Inspired by meta-plasticity and reconstructive memory in neuroscience, we introduce here an innovative approach named Meta Post-Refinement (MePo) for PTMs-based GCL. This approach constructs pseudo task sequences from pretraining data and develops a bi-level meta-learning paradigm to refine the pretrained backbone, which serves as a prolonged pretraining phase but greatly facilitates rapid adaptation of representation learning to downstream GCL tasks. MePo further initializes a meta covariance matrix as the reference geometry of pretrained representation space, enabling GCL to exploit second-order statistics for robust output alignment. MePo serves as a plug-in strategy that achieves significant performance gains across a variety of GCL benchmarks and pretrained checkpoints in a rehearsal-free manner (e.g., 15.10%, 13.36%, and 12.56% on CIFAR-100, ImageNet-R, and CUB-200 under Sup-21/1K). Our source code is available at \href{https://github.com/SunGL001/MePo}{MePo}

Method: Two-stage evaluation: first testing if LLMs can recover well-established instruments from literature, then evaluating if they can identify and avoid empirically/theoretically discredited instruments. Introduces IV Co-Scientist, a multi-agent system that proposes, critiques, and refines IVs for treatment-outcome pairs, plus a statistical test for consistency without ground truth.

Result: LLMs show potential to discover valid instrumental variables from large observational databases, demonstrating ability to replicate standard reasoning and avoid invalid instruments.

Conclusion: LLMs can be valuable tools for instrumental variable discovery in causal inference, with the proposed IV Co-Scientist system offering a promising approach for automating and improving this challenging task.

Abstract: In the presence of confounding between an endogenous variable and the outcome, instrumental variables (IVs) are used to isolate the causal effect of the endogenous variable. Identifying valid instruments requires interdisciplinary knowledge, creativity, and contextual understanding, making it a non-trivial task. In this paper, we investigate whether large language models (LLMs) can aid in this task. We perform a two-stage evaluation framework. First, we test whether LLMs can recover well-established instruments from the literature, assessing their ability to replicate standard reasoning. Second, we evaluate whether LLMs can identify and avoid instruments that have been empirically or theoretically discredited. Building on these results, we introduce IV Co-Scientist, a multi-agent system that proposes, critiques, and refines IVs for a given treatment-outcome pair. We also introduce a statistical test to contextualize consistency in the absence of ground truth. Our results show the potential of LLMs to discover valid instrumental variables from a large observational database.

Method: LOCA-bench uses automated and scalable control of environment states to regulate agent context length. It can extend context length potentially to infinity while keeping task semantics fixed, evaluating language agents as combinations of models and scaffolds with various context management strategies.

Result: Agent performance generally degrades as environment states grow more complex, but advanced context management techniques can substantially improve overall success rates. The benchmark provides a platform for evaluating models and scaffolds in long-context, agentic scenarios.

Conclusion: LOCA-bench addresses the gap in evaluating LLM agents in realistic long-context scenarios and demonstrates that proper context management strategies can mitigate context rot in agentic settings.

Abstract: Large language models (LLMs) are increasingly capable of carrying out long-running, real-world tasks. However, as the amount of context grows, their reliability often deteriorates, a phenomenon known as “context rot”. Existing long-context benchmarks primarily focus on single-step settings that evaluate a model’s ability to retrieve information from a long snippet. In realistic scenarios, however, LLMs often need to act as agents that explore environments, follow instructions and plans, extract useful information, and predict correct actions under a dynamically growing context. To assess language agents in such settings, we introduce LOCA-bench (a benchmark for LOng-Context Agents). Given a task prompt, LOCA-bench leverages automated and scalable control of environment states to regulate the agent’s context length. This design enables LOCA-bench to extend the context length potentially to infinity in a controlled way while keeping the underlying task semantics fixed. LOCA-bench evaluates language agents as a combination of models and scaffolds, including various context management strategies. While agent performance generally degrades as the environment states grow more complex, advanced context management techniques can substantially improve the overall success rate. We open-source LOCA-bench to provide a platform for evaluating models and scaffolds in long-context, agentic scenarios: https://github.com/hkust-nlp/LOCA-bench

Method: EXPERIGEN uses a Bayesian optimization inspired two-phase search: a Generator agent proposes candidate hypotheses, and an Experimenter agent evaluates them empirically. The framework extends to complex data regimes including multimodal and relational datasets.

Result: EXPERIGEN discovers 2-4x more statistically significant hypotheses that are 7-17% more predictive than prior approaches. Expert review shows 88% of hypotheses rated moderately/strongly novel, 70% deemed impactful. First A/B test of LLM-generated hypotheses shows p < 1e-6 with 344% effect size.

Conclusion: EXPERIGEN successfully operationalizes end-to-end scientific discovery, demonstrating superior statistical performance, novelty, and real-world impact across multiple domains including multimodal data regimes.

Abstract: Data-driven social science research is inherently slow, relying on iterative cycles of observation, hypothesis generation, and experimental validation. While recent data-driven methods promise to accelerate parts of this process, they largely fail to support end-to-end scientific discovery. To address this gap, we introduce EXPERIGEN, an agentic framework that operationalizes end-to-end discovery through a Bayesian optimization inspired two-phase search, in which a Generator proposes candidate hypotheses and an Experimenter evaluates them empirically. Across multiple domains, EXPERIGEN consistently discovers 2-4x more statistically significant hypotheses that are 7-17 percent more predictive than prior approaches, and naturally extends to complex data regimes including multimodal and relational datasets. Beyond statistical performance, hypotheses must be novel, empirically grounded, and actionable to drive real scientific progress. To evaluate these qualities, we conduct an expert review of machine-generated hypotheses, collecting feedback from senior faculty. Among 25 reviewed hypotheses, 88 percent were rated moderately or strongly novel, 70 percent were deemed impactful and worth pursuing, and most demonstrated rigor comparable to senior graduate-level research. Finally, recognizing that ultimate validation requires real-world evidence, we conduct the first A/B test of LLM-generated hypotheses, observing statistically significant results with p less than 1e-6 and a large effect size of 344 percent.

Method: RAPS uses Distributed Publish-Subscribe Protocol for intent-based message exchange, plus two overlays: Reactive Subscription for dynamic intent refinement, and Bayesian Reputation for local malicious peer detection and isolation.

Result: Extensive experiments over five benchmarks show the design effectively reconciles adaptivity, scalability, and robustness in a unified multi-agent coordination framework.

Conclusion: RAPS provides an automated solution for LLM agent coordination that addresses the challenges of adaptivity, scalability, and robustness through reputation-aware publish-subscribe mechanisms.

Abstract: Multi-agent architectures built on large language models (LLMs) have demonstrated the potential to realize swarm intelligence through well-crafted collaboration. However, the substantial burden of manual orchestration inherently raises an imperative to automate the design of agentic workflows. We frame such an agent coordination challenge as a classic problem in dynamic ad-hoc networking: How to establish adaptive and reliable communication among a scalable number of agentic hosts? In response to this unresolved dilemma, we introduce RAPS, a reputation-aware publish-subscribe paradigm for adaptive, scalable, and robust coordination of LLM agents. RAPS is grounded in the Distributed Publish-Subscribe Protocol, allowing LLM agents to exchange messages based on their declared intents rather than predefined topologies. Beyond this substrate, RAPS further incorporates two coherent overlays: (i) Reactive Subscription, enabling agents to dynamically refine their intents; and (ii) Bayesian Reputation, empowering each agent with a local watchdog to detect and isolate malicious peers. Extensive experiments over five benchmarks showcase that our design effectively reconciles adaptivity, scalability, and robustness in a unified multi-agent coordination framework.

Method: Proposes Small Agent Group (SAG) approach that shifts from single-model intelligence to collective expertise by distributing reasoning, evidence-based analysis, and critical audit through collaborative deliberation processes. Evaluated using diverse clinical metrics spanning effectiveness, reliability, and deployment cost.

Result: SAG achieves superior performance compared to single giant models, both with and without additional optimization or retrieval-augmented generation. The synergistic reasoning of SAG can substitute for model parameter growth in clinical settings.

Conclusion: SAG offers a scalable solution to digital health that better balances effectiveness, reliability, and deployment efficiency, challenging the monolithic scaling paradigm with collaborative multi-agent approaches.

Abstract: The rapid adoption of large language models (LLMs) in digital health has been driven by a “scaling-first” philosophy, i.e., the assumption that clinical intelligence increases with model size and data. However, real-world clinical needs include not only effectiveness, but also reliability and reasonable deployment cost. Since clinical decision-making is inherently collaborative, we challenge the monolithic scaling paradigm and ask whether a Small Agent Group (SAG) can support better clinical reasoning. SAG shifts from single-model intelligence to collective expertise by distributing reasoning, evidence-based analysis, and critical audit through a collaborative deliberation process. To assess the clinical utility of SAG, we conduct extensive evaluations using diverse clinical metrics spanning effectiveness, reliability, and deployment cost. Our results show that SAG achieves superior performance compared to a single giant model, both with and without additional optimization or retrieval-augmented generation. These findings suggest that the synergistic reasoning represented by SAG can substitute for model parameter growth in clinical settings. Overall, SAG offers a scalable solution to digital health that better balances effectiveness, reliability, and deployment efficiency.

Method: Uses conditional Gaussian network classifier (CGNC) with directed acyclic graph to encode feature dependencies. Employs convergence-guaranteed cutting-set procedure for adversarial optimization, and applies piecewise McCormick relaxation to reformulate nonconvex quadratic problems as mixed-integer linear programs for global optimality.

Result: The method achieves strong robustness with direct global optimization providing stable and efficient results. The framework shows extensibility to complex constraint settings.

Conclusion: The proposed structure-aware counterfactual search method successfully incorporates feature dependencies through CGNC’s generative structure and ensures robustness through convergence-guaranteed optimization, providing a foundation for future advances in counterfactual reasoning under nonconvex quadratic formulations.

Abstract: Counterfactual explanation (CE) is a core technique in explainable artificial intelligence (XAI), widely used to interpret model decisions and suggest actionable alternatives. This work presents a structure-aware and robustness-oriented counterfactual search method based on the conditional Gaussian network classifier (CGNC). The CGNC has a generative structure that encodes conditional dependencies and potential causal relations among features through a directed acyclic graph (DAG). This structure naturally embeds feature relationships into the search process, eliminating the need for additional constraints to ensure consistency with the model’s structural assumptions. We adopt a convergence-guaranteed cutting-set procedure as an adversarial optimization framework, which iteratively approximates solutions that satisfy global robustness conditions. To address the nonconvex quadratic structure induced by feature dependencies, we apply piecewise McCormick relaxation to reformulate the problem as a mixed-integer linear program (MILP), ensuring global optimality. Experimental results show that our method achieves strong robustness, with direct global optimization of the original formulation providing especially stable and efficient results. The proposed framework is extensible to more complex constraint settings, laying the groundwork for future advances in counterfactual reasoning under nonconvex quadratic formulations.

Method: Free()LM uses a Free-Module (plug-and-play LoRA adapter) that enables iterative switching between reasoning and cleaning modes to dynamically identify and prune useless context chunks, maintaining compact and noise-free reasoning states.

Result: Free()LM achieves 3.3% average improvement over top-tier reasoning baselines across model scales (8B to 685B), establishes new SOTA on IMOanswerBench, and restores performance from 0% to 50% in long-horizon tasks where standard models collapse.

Conclusion: Sustainable intelligence requires both the power to think and the freedom to forget; self-forgetting mechanisms are essential for preventing reasoning degradation in LLMs.

Abstract: Reasoning models enhance problem-solving by scaling test-time compute, yet they face a critical paradox: excessive thinking tokens often degrade performance rather than improve it. We attribute this to a fundamental architectural flaw: standard LLMs operate as “malloc-only” engines, continuously accumulating valid and redundant steps alike without a mechanism to prune obsolete information. To break this cycle, we propose Free()LM, a model that introduces an intrinsic self-forgetting capability via the Free-Module, a plug-and-play LoRA adapter. By iteratively switching between reasoning and cleaning modes, Free()LM dynamically identifies and prunes useless context chunks, maintaining a compact and noise-free state. Extensive experiments show that Free()LM provides consistent improvements across all model scales (8B to 685B). It achieves a 3.3% average improvement over top-tier reasoning baselines, even establishing a new SOTA on IMOanswerBench using DeepSeek V3.2-Speciale. Most notably, in long-horizon tasks where the standard Qwen3-235B-A22B model suffers a total collapse (0% accuracy), Free()LM restores performance to 50%. Our findings suggest that sustainable intelligence requires the freedom to forget as much as the power to think.

Method: Proposes a Deep Reinforcement Learning framework using Proximal Policy Optimization (PPO) and Graph Neural Network (GNN) to represent complex state of jobs, machines, and setups, learning a direct scheduling policy guided by multi-objective reward function.

Result: PPO-GNN agent significantly outperforms standard dispatching rule and metaheuristic, achieving superior trade-off between minimizing TWT and TST on benchmark instances.

Conclusion: Provides a robust and scalable solution for complex manufacturing scheduling problems using deep reinforcement learning with graph neural networks.

Abstract: The Unrelated Parallel Machine Scheduling Problem (UPMSP) with release dates, setups, and eligibility constraints presents a significant multi-objective challenge. Traditional methods struggle to balance minimizing Total Weighted Tardiness (TWT) and Total Setup Time (TST). This paper proposes a Deep Reinforcement Learning framework using Proximal Policy Optimization (PPO) and a Graph Neural Network (GNN). The GNN effectively represents the complex state of jobs, machines, and setups, allowing the PPO agent to learn a direct scheduling policy. Guided by a multi-objective reward function, the agent simultaneously minimizes TWT and TST. Experimental results on benchmark instances demonstrate that our PPO-GNN agent significantly outperforms a standard dispatching rule and a metaheuristic, achieving a superior trade-off between both objectives. This provides a robust and scalable solution for complex manufacturing scheduling.

Method: Developed a five-tier Biosecurity Data Level (BDL) framework that categorizes pathogen data types based on their potential to contribute to concerning AI capabilities, with technical restrictions proposed for each risk level and a governance framework for dual-use pathogen data.

Result: A comprehensive biosecurity framework that links specific data types to AI risk levels, providing a structured approach to data control that could serve as a high-leverage intervention in preventing proliferation of dangerous biological AI capabilities.

Conclusion: In an era of widely accessible computational resources, data controls represent a crucial intervention point for reducing risks from AI models trained on biological data, with the proposed BDL framework offering a practical approach to implementing such controls.

Abstract: Training data is an essential input into creating competent artificial intelligence (AI) models. AI models for biology are trained on large volumes of data, including data related to biological sequences, structures, images, and functions. The type of data used to train a model is intimately tied to the capabilities it ultimately possesses–including those of biosecurity concern. For this reason, an international group of more than 100 researchers at the recent 50th anniversary Asilomar Conference endorsed data controls to prevent the use of AI for harmful applications such as bioweapons development. To help design such controls, we introduce a five-tier Biosecurity Data Level (BDL) framework for categorizing pathogen data. Each level contains specific data types, based on their expected ability to contribute to capabilities of concern when used to train AI models. For each BDL tier, we propose technical restrictions appropriate to its level of risk. Finally, we outline a novel governance framework for newly created dual-use pathogen data. In a world with widely accessible computational and coding resources, data controls may be among the most high-leverage interventions available to reduce the proliferation of concerning biological AI capabilities.

Method: The authors propose Epistemic Source Alignment (ESA), which differs from standard robust methods that rely on statistical consensus. ESA uses sparse safety axioms to judge the source of feedback rather than the signal itself - a “judging the judges” mechanism that evaluates feedback providers rather than aggregating their signals.

Result: The paper proves that under Dogma 4, standard RL agents suffer from Objective Decoupling where learned objectives permanently separate from ground truth. They prove ESA guarantees convergence to true objectives even with majority biased evaluators. Empirically, traditional consensus methods fail under majority collusion while ESA recovers optimal policies.

Conclusion: The fundamental assumption of truthful human feedback in RL alignment is flawed in social contexts. Epistemic Source Alignment provides a solution by evaluating feedback sources rather than signals, enabling robust alignment even with biased evaluators.

Abstract: Contemporary AI alignment strategies rely on a fragile premise: that human feedback, while noisy, remains a fundamentally truthful signal. In this paper, we identify this assumption as Dogma 4 of Reinforcement Learning (RL). We demonstrate that while this dogma holds in static environments, it fails in social settings where evaluators may be sycophantic, lazy, or adversarial. We prove that under Dogma 4, standard RL agents suffer from what we call Objective Decoupling, a structural failure mode where the agent’s learned objective permanently separates from the latent ground truth, guaranteeing convergence to misalignment. To resolve this, we propose Epistemic Source Alignment (ESA). Unlike standard robust methods that rely on statistical consensus (trusting the majority), ESA utilizes sparse safety axioms to judge the source of the feedback rather than the signal itself. We prove that this “judging the judges” mechanism guarantees convergence to the true objective, even when a majority of evaluators are biased. Empirically, we show that while traditional consensus methods fail under majority collusion, our approach successfully recovers the optimal policy.

Method: Developed Glow, a GenAI-powered DBT skills coach for chain and solution analysis. Conducted usability testing with clinical staff (n=6) and individuals with lived experience (n=28) using the HHH framework. Employed user-driven adversarial testing where participants identified target behaviors and generated contextually realistic risk probes, evaluating safety across 37 risk probe interactions.

Result: Glow appropriately handled 73% of risk probes, but performance varied significantly by agent: solution analysis agent demonstrated 90% appropriate handling vs 44% for chain analysis agent. Safety failures clustered around encouraging substance use and normalizing harmful behaviors. Chain analysis agent fell into “empathy trap” providing validation that reinforced maladaptive beliefs. Also identified 27 instances of DBT skill misinformation.

Conclusion: First systematic safety evaluation of GenAI-delivered DBT coaching for HIV/substance use risk reduction reveals vulnerabilities requiring mitigation before clinical trials. HHH framework and user-driven adversarial testing offer replicable methods for evaluating GenAI mental health interventions.

Abstract: Background: HIV and substance use represent interacting epidemics with shared psychological drivers - impulsivity and maladaptive coping. Dialectical behavior therapy (DBT) targets these mechanisms but faces scalability challenges. Generative artificial intelligence (GenAI) offers potential for delivering personalized DBT coaching at scale, yet rapid development has outpaced safety infrastructure. Methods: We developed Glow, a GenAI-powered DBT skills coach delivering chain and solution analysis for individuals at risk for HIV and substance use. In partnership with a Los Angeles community health organization, we conducted usability testing with clinical staff (n=6) and individuals with lived experience (n=28). Using the Helpful, Honest, and Harmless (HHH) framework, we employed user-driven adversarial testing wherein participants identified target behaviors and generated contextually realistic risk probes. We evaluated safety performance across 37 risk probe interactions. Results: Glow appropriately handled 73% of risk probes, but performance varied by agent. The solution analysis agent demonstrated 90% appropriate handling versus 44% for the chain analysis agent. Safety failures clustered around encouraging substance use and normalizing harmful behaviors. The chain analysis agent fell into an “empathy trap,” providing validation that reinforced maladaptive beliefs. Additionally, 27 instances of DBT skill misinformation were identified. Conclusions: This study provides the first systematic safety evaluation of GenAI-delivered DBT coaching for HIV and substance use risk reduction. Findings reveal vulnerabilities requiring mitigation before clinical trials. The HHH framework and user-driven adversarial testing offer replicable methods for evaluating GenAI mental health interventions.

Method: Introduces Recursive Entropy to quantify resource consumption risk in reflection, then develops RECUR (Resource Exhaustion attack via Recursive Entropy guided Counterfactual Utilization and Reflection) that constructs counterfactual questions to exploit LRM vulnerabilities.

Result: Under benign inference, recursive entropy shows decreasing trend. RECUR disrupts this trend, increasing output length by up to 11x and decreasing throughput by 90%, demonstrating significant resource exhaustion vulnerabilities.

Conclusion: The work reveals safety issues inherent in inference itself and provides a new perspective on robust reasoning by quantifying and exploiting reflection vulnerabilities in Large Reasoning Models.

Abstract: Large Reasoning Models (LRMs) employ reasoning to address complex tasks. Such explicit reasoning requires extended context lengths, resulting in substantially higher resource consumption. Prior work has shown that adversarially crafted inputs can trigger redundant reasoning processes, exposing LRMs to resource-exhaustion vulnerabilities. However, the reasoning process itself, especially its reflective component, has received limited attention, even though it can lead to over-reflection and consume excessive computing power. In this paper, we introduce Recursive Entropy to quantify the risk of resource consumption in reflection, thereby revealing the safety issues inherent in inference itself. Based on Recursive Entropy, we introduce RECUR, a resource exhaustion attack via Recursive Entropy guided Counterfactual Utilization and Reflection. It constructs counterfactual questions to verify the inherent flaws and risks of LRMs. Extensive experiments demonstrate that, under benign inference, recursive entropy exhibits a pronounced decreasing trend. RECUR disrupts this trend, increasing the output length by up to 11x and decreasing throughput by 90%. Our work provides a new perspective on robust reasoning.

Method: WMSS leverages weak checkpoints to guide continued optimization by: 1) identifying recoverable learning gaps via entropy dynamics analysis, and 2) reinforcing these gaps through compensatory learning, enabling strong agents to improve beyond conventional saturation.

Result: Experiments on mathematical reasoning and code generation datasets show that agents trained with WMSS achieve effective performance improvements while incurring zero additional inference cost.

Conclusion: WMSS provides a novel post-training paradigm that overcomes saturation bottlenecks by utilizing weak historical checkpoints to guide optimization, enabling continued improvement of strong models without inference overhead.

Abstract: As post-training optimization becomes central to improving large language models, we observe a persistent saturation bottleneck: once models grow highly confident, further training yields diminishing returns. While existing methods continue to reinforce target predictions, we find that informative supervision signals remain latent in models’ own historical weak states. Motivated by this observation, we propose WMSS (Weak Agents Can Make Strong Agents Stronger), a post-training paradigm that leverages weak checkpoints to guide continued optimization. By identifying recoverable learning gaps via entropy dynamics and reinforcing them through compensatory learning, WMSS enables strong agents to improve beyond conventional post-training saturation. Experiments on mathematical reasoning and code generation datasets show that agents trained with our approach achieve effective performance improvements, while incurring zero additional inference cost.

Method: Proposes a decentralized evaluation framework leveraging blockchain-based protocol to enable hardware and parameter diversity through large-scale benchmarking across heterogeneous compute nodes, incentivizing global contributors as independent validators with robust reward systems.

Result: The decentralized framework reduces standard deviation across ten runs on the same model from 1.67 to 0.28, significantly improving statistical confidence in model rankings compared to conventional frameworks.

Conclusion: Decentralized evaluation transforms evaluation from a “centralized black box” to “decentralized endorsement” with multi-party consensus and diverse inference environments, yielding more stable and representative metrics for LLM assessment.

Abstract: The rapid advancement of large language models (LLMs) demands increasingly reliable evaluation, yet current centralized evaluation suffers from opacity, overfitting, and hardware-induced variance. Our empirical analysis reveals an alarming inconsistency in existing evaluations: the standard deviation across ten repeated runs of a single model on HumanEval (1.67) actually exceeds the performance gap among the top-10 models on the official leaderboard (0.91), rendering current rankings statistically precarious. To mitigate these instabilities, we propose a decentralized evaluation framework that enables hardware and parameter diversity through large-scale benchmarking across heterogeneous compute nodes. By leveraging the blockchain-based protocol, the framework incentivizes global contributors to act as independent validators, using a robust reward system to ensure evaluation integrity and discourage dishonest participation. This collective verification transforms evaluation from a “centralized black box” into a “decentralized endorsement” where multi-party consensus and diverse inference environments yield a more stable, representative metric. Experimental results demonstrate that the decentralized evaluation framework reduces the standard deviation across ten runs on the same model to 0.28. This significant improvement over conventional frameworks ensures higher statistical confidence in model rankings. We have completely implemented this platform and will soon release it to the community.

Method: Proposes SAYO, a visual reasoning model trained with reinforcement learning framework that introduces region-level visual attention-based reward to explicitly align optimization signals with visually grounded reasoning steps.

Result: Extensive experiments across multiple multimodal benchmarks demonstrate that SAYO consistently improves performance on diverse reasoning and perception tasks.

Conclusion: The RL framework with region-level visual attention rewards enables MLLMs to learn more reliable attention behaviors and improve visual reasoning capabilities.

Abstract: While chain-of-thought (CoT) reasoning has substantially improved multimodal large language models (MLLMs) on complex reasoning tasks, existing approaches largely rely on long textual reasoning trajectories and provide limited mechanisms for learning stable visual attention policies. Our analysis shows that current MLLMs exhibit weak visual focus: early-stage visual misalignment is rarely corrected during subsequent reasoning, leading to error propagation and failed inferences. We argue that this limitation stems from inadequate credit assignment for visual attention during training. To address this issue, we propose SAYO, a visual reasoning model trained with a reinforcement learning (RL) framework that introduces a region-level visual attention-based reward. This reward explicitly aligns optimization signals with visually grounded reasoning steps, enabling the model to learn more reliable attention behaviors. Extensive experiments across multiple multimodal benchmarks demonstrate that SAYO consistently improves performance on diverse reasoning and perception tasks.

Method: Proposes G-LNS, a generative evolutionary framework that extends LLM-based AHD to automatically design Large Neighborhood Search operators. It leverages LLMs to co-evolve tightly coupled pairs of destroy and repair operators with a cooperative evaluation mechanism that captures their interaction.

Result: Extensive experiments on challenging COP benchmarks (Traveling Salesman Problems and Capacitated Vehicle Routing Problems) show G-LNS significantly outperforms LLM-based AHD methods and strong classical solvers. Discovered heuristics achieve near-optimal solutions with reduced computational budgets and exhibit robust generalization across diverse and unseen instance distributions.

Conclusion: G-LNS successfully extends LLM capabilities to automated design of Large Neighborhood Search operators, enabling discovery of complementary destroy-repair pairs that outperform existing approaches and demonstrate strong generalization.

Abstract: While Large Language Models (LLMs) have recently shown promise in Automated Heuristic Design (AHD), existing approaches typically formulate AHD around constructive priority rules or parameterized local search guidance, thereby restricting the search space to fixed heuristic forms. Such designs offer limited capacity for structural exploration, making it difficult to escape deep local optima in complex Combinatorial Optimization Problems (COPs). In this work, we propose G-LNS, a generative evolutionary framework that extends LLM-based AHD to the automated design of Large Neighborhood Search (LNS) operators. Unlike prior methods that evolve heuristics in isolation, G-LNS leverages LLMs to co-evolve tightly coupled pairs of destroy and repair operators. A cooperative evaluation mechanism explicitly captures their interaction, enabling the discovery of complementary operator logic that jointly performs effective structural disruption and reconstruction. Extensive experiments on challenging COP benchmarks, such as Traveling Salesman Problems (TSP) and Capacitated Vehicle Routing Problems (CVRP), demonstrate that G-LNS significantly outperforms LLM-based AHD methods as well as strong classical solvers. The discovered heuristics not only achieve near-optimal solutions with reduced computational budgets but also exhibit robust generalization across diverse and unseen instance distributions.

Method: Proposes Puda (Private User Dataset Agent) - a browser-based system with three privacy levels: Detailed Browsing History, Extracted Keywords, and Predefined Category Subsets, enabling client-side data management and controlled sharing.

Result: In personalized travel planning tasks, Predefined Category Subsets achieved 97.2% of the personalization performance of Detailed Browsing History when evaluated via LLM-as-a-Judge framework across three criteria.

Conclusion: Puda enables effective multi-granularity privacy management, offering practical solutions to mitigate the privacy-personalization trade-off and providing an AI-native foundation for user sovereignty in personalized AI services.

Abstract: Personal data centralization among dominant platform providers including search engines, social networking services, and e-commerce has created siloed ecosystems that restrict user sovereignty, thereby impeding data use across services. Meanwhile, the rapid proliferation of Large Language Model (LLM)-based agents has intensified demand for highly personalized services that require the dynamic provision of diverse personal data. This presents a significant challenge: balancing the utilization of such data with privacy protection. To address this challenge, we propose Puda (Private User Dataset Agent), a user-sovereign architecture that aggregates data across services and enables client-side management. Puda allows users to control data sharing at three privacy levels: (i) Detailed Browsing History, (ii) Extracted Keywords, and (iii) Predefined Category Subsets. We implemented Puda as a browser-based system that serves as a common platform across diverse services and evaluated it through a personalized travel planning task. Our results show that providing Predefined Category Subsets achieves 97.2% of the personalization performance (evaluated via an LLM-as-a-Judge framework across three criteria) obtained when sharing Detailed Browsing History. These findings demonstrate that Puda enables effective multi-granularity management, offering practical choices to mitigate the privacy-personalization trade-off. Overall, Puda provides an AI-native foundation for user sovereignty, empowering users to safely leverage the full potential of personalized AI.

Method: Proposes Structural Context Model as formal foundation, plus declarative implementation framework and Semantic Dynamics Analysis workflow for full agent lifecycle from design to iteration.

Result: Demonstrated on dynamic monkey-banana problems, achieving up to 32 percentage points improvement in success rate on most challenging settings.

Conclusion: Provides unified formal framework for LLM agent analysis and engineering, enabling systematic design and principled insights into agent mechanisms.

Abstract: Current research on large language model (LLM) agents is fragmented: discussions of conceptual frameworks and methodological principles are frequently intertwined with low-level implementation details, causing both readers and authors to lose track amid a proliferation of superficially distinct concepts. We argue that this fragmentation largely stems from the absence of an analyzable, self-consistent formal model that enables implementation-independent characterization and comparison of LLM agents. To address this gap, we propose the \texttt{Structural Context Model}, a formal model for analyzing and comparing LLM agents from the perspective of context structure. Building upon this foundation, we introduce two complementary components that together span the full lifecycle of LLM agent research and development: (1) a declarative implementation framework; and (2) a sustainable agent engineering workflow, \texttt{Semantic Dynamics Analysis}. The proposed workflow provides principled insights into agent mechanisms and supports rapid, systematic design iteration. We demonstrate the effectiveness of the complete framework on dynamic variants of the monkey-banana problem, where agents engineered using our approach achieve up to a 32 percentage points improvement in success rate on the most challenging setting.

Method: Conceptual framework development introducing “Vibe-Automation” concept, analyzing GenAI’s functional access to operationalized tacit regularities, and proposing analytical framework across three levels and three domains of action (faculty worldview, industry relations, curriculum design).

Result: Establishes Vibe-Automation as a framework for understanding GenAI’s epistemological significance, identifies shift from algorithmic problem specification to Vibe-Engineering, and provides structured approach for educational/institutional adaptation.

Conclusion: Generative AI represents fundamental epistemological shift requiring new conceptual frameworks like Vibe-Automation, with human role evolving toward Vibe-Engineering and necessitating deliberate engagement to avoid risks of mode collapse and cultural homogenization.

Abstract: The emergence of generative artificial intelligence (GenAI) represents not an incremental technological advance but a qualitative epistemological shift that challenges foundational assumptions of computer science. Whereas machine learning has been described as the automation of automation, generative AI operates by navigating contextual, semantic, and stylistic coherence rather than optimizing predefined objective metrics. This paper introduces the concept of Vibe-Automation to characterize this transition. The central claim is that the significance of GenAI lies in its functional access to operationalized tacit regularities: context-sensitive patterns embedded in practice that cannot be fully specified through explicit algorithmic rules. Although generative systems do not possess tacit knowledge in a phenomenological sense, they operationalize sensitivities to tone, intent, and situated judgment encoded in high-dimensional latent representations. On this basis, the human role shifts from algorithmic problem specification toward Vibe-Engineering, understood as the orchestration of alignment and contextual judgment in generative systems. The paper connects this epistemological shift to educational and institutional transformation by proposing a conceptual framework structured across three analytical levels and three domains of action: faculty worldview, industry relations, and curriculum design. The risks of mode collapse and cultural homogenization are briefly discussed, emphasizing the need for deliberate engagement with generative systems to avoid regression toward synthetic uniformity.

Method: Systematic study of 10 widely-used VLMs on Moralise and M^3oralBench datasets under explicit user disagreement, using Error Introduction Rate (EIR) and Error Correction Rate (ECR) metrics to analyze moral sycophancy patterns.

Result: VLMs frequently produce morally incorrect follow-up responses even when initial judgments are correct, showing asymmetric bias (right→wrong shifts more common than wrong→right corrections). Models with stronger error-correction capabilities tend to introduce more reasoning errors, while conservative models minimize errors but have limited self-correction ability.

Conclusion: VLMs are vulnerable to moral influence, exhibiting significant sycophantic behavior that compromises ethical consistency, highlighting the need for principled strategies to improve moral robustness in multimodal AI systems.

Abstract: Sycophancy in Vision-Language Models (VLMs) refers to their tendency to align with user opinions, often at the expense of moral or factual accuracy. While prior studies have explored sycophantic behavior in general contexts, its impact on morally grounded visual decision-making remains insufficiently understood. To address this gap, we present the first systematic study of moral sycophancy in VLMs, analyzing ten widely-used models on the Moralise and M^3oralBench datasets under explicit user disagreement. Our results reveal that VLMs frequently produce morally incorrect follow-up responses even when their initial judgments are correct, and exhibit a consistent asymmetry: models are more likely to shift from morally right to morally wrong judgments than the reverse when exposed to user-induced bias. Follow-up prompts generally degrade performance on Moralise, while yielding mixed or even improved accuracy on M^3oralBench, highlighting dataset-dependent differences in moral robustness. Evaluation using Error Introduction Rate (EIR) and Error Correction Rate (ECR) reveals a clear trade-off: models with stronger error-correction capabilities tend to introduce more reasoning errors, whereas more conservative models minimize errors but exhibit limited ability to self-correct. Finally, initial contexts with a morally right stance elicit stronger sycophantic behavior, emphasizing the vulnerability of VLMs to moral influence and the need for principled strategies to improve ethical consistency and robustness in multimodal AI systems.

Method: SHARP uses Shapley-based hierarchical attribution with a decomposed reward mechanism: global broadcast-accuracy reward, Shapley-based marginal-credit reward for each agent, and tool-process reward to improve execution efficiency. It stabilizes training by normalizing agent-specific advantages across trajectory groups.

Result: Extensive experiments across real-world benchmarks show SHARP significantly outperforms state-of-the-art baselines, achieving average match improvements of 23.66% over single-agent approaches and 14.05% over multi-agent approaches.

Conclusion: SHARP effectively addresses credit assignment in LLM-based multi-agent systems with external tools, enabling more efficient reinforcement learning through precise attribution of individual agent contributions.

Abstract: Integrating Large Language Models (LLMs) with external tools via multi-agent systems offers a promising new paradigm for decomposing and solving complex problems. However, training these systems remains notoriously difficult due to the credit assignment challenge, as it is often unclear which specific functional agent is responsible for the success or failure of decision trajectories. Existing methods typically rely on sparse or globally broadcast rewards, failing to capture individual contributions and leading to inefficient reinforcement learning. To address these limitations, we introduce the Shapley-based Hierarchical Attribution for Reinforcement Policy (SHARP), a novel framework for optimizing multi-agent reinforcement learning via precise credit attribution. SHARP effectively stabilizes training by normalizing agent-specific advantages across trajectory groups, primarily through a decomposed reward mechanism comprising a global broadcast-accuracy reward, a Shapley-based marginal-credit reward for each agent, and a tool-process reward to improve execution efficiency. Extensive experiments across various real-world benchmarks demonstrate that SHARP significantly outperforms recent state-of-the-art baselines, achieving average match improvements of 23.66% and 14.05% over single-agent and multi-agent approaches, respectively.

Method: Two-component approach: (1) Dual-stage data synthesis with bottom-up extraction of atomic visual primitives and top-down hierarchical reasoning guidance, and (2) Cognition-aligned training using Cognitively Coherent Verifiable Rewards in Reinforcement Fine-Tuning for stepwise feedback on reasoning coherence.

Result: Achieves 83.33% F1 score on multi-level semantic inconsistency benchmark with lexical-perturbation negatives across both in-domain and out-of-domain settings. Ablations confirm each component contributes to more interpretable and human-aligned visual reasoning.

Conclusion: CoTZero improves visual reasoning in VLMs by incorporating human cognitive models through annotation-free data synthesis and verifiable reward-based training, leading to better compositional reasoning and generalization.

Abstract: Recent advances in vision-language models (VLMs) have markedly improved image-text alignment, yet they still fall short of human-like visual reasoning. A key limitation is that many VLMs rely on surface correlations rather than building logically coherent structured representations, which often leads to missed higher-level semantic structure and non-causal relational understanding, hindering compositional and verifiable reasoning. To address these limitations by introducing human models into the reasoning process, we propose CoTZero, an annotation-free paradigm with two components: (i) a dual-stage data synthesis approach and (ii) a cognition-aligned training method. In the first component, we draw inspiration from neurocognitive accounts of compositional productivity and global-to-local analysis. In the bottom-up stage, CoTZero extracts atomic visual primitives and incrementally composes them into diverse, structured question-reasoning forms. In the top-down stage, it enforces hierarchical reasoning by using coarse global structure to guide the interpretation of local details and causal relations. In the cognition-aligned training component, built on the synthesized CoT data, we introduce Cognitively Coherent Verifiable Rewards (CCVR) in Reinforcement Fine-Tuning (RFT) to further strengthen VLMs’ hierarchical reasoning and generalization, providing stepwise feedback on reasoning coherence and factual correctness. Experiments show that CoTZero achieves an F1 score of 83.33 percent on our multi-level semantic inconsistency benchmark with lexical-perturbation negatives, across both in-domain and out-of-domain settings. Ablations confirm that each component contributes to more interpretable and human-aligned visual reasoning.

Method: Proposes an effect-centric, admissibility-first framework that treats discovered graphs as structural hypotheses and evaluates them by identifiability, stability, and falsification criteria. Empirically studies the effect of early exposure to competitive gameplay on short-term retention using real-world game telemetry data.

Result: Many statistically plausible discovery outputs don’t admit point-identified causal queries when temporal/semantic constraints are enforced. When identification is possible, different algorithm families converge to similar effect estimates despite producing different graph structures. Converging estimates survive placebo, subsampling, and sensitivity refutation tests.

Conclusion: Graph-level metrics alone are inadequate proxies for causal reliability. Trustworthy causal conclusions in telemetry-driven systems require prioritizing admissibility and effect-level validation over causal structural recovery alone.

Abstract: Causal discovery is increasingly applied to large-scale telemetry data to estimate the effects of user-facing interventions, yet its reliability for decision-making in feedback-driven systems with strong self-selection remains unclear. In this paper, we propose an effect-centric, admissibility-first framework that treats discovered graphs as structural hypotheses and evaluates them by identifiability, stability, and falsification rather than by graph recovery accuracy alone. Empirically, we study the effect of early exposure to competitive gameplay on short-term retention using real-world game telemetry. We find that many statistically plausible discovery outputs do not admit point-identified causal queries once minimal temporal and semantic constraints are enforced, highlighting identifiability as a critical bottleneck for decision support. When identification is possible, several algorithm families converge to similar, decision-consistent effect estimates despite producing substantially different graph structures, including cases where the direct treatment-outcome edge is absent and the effect is preserved through indirect causal pathways. These converging estimates survive placebo, subsampling, and sensitivity refutation. In contrast, other methods exhibit sporadic admissibility and threshold-sensitive or attenuated effects due to endpoint ambiguity. These results suggest that graph-level metrics alone are inadequate proxies for causal reliability for a given target query. Therefore, trustworthy causal conclusions in telemetry-driven systems require prioritizing admissibility and effect-level validation over causal structural recovery alone.

Method: Outline-Guided Path Exploration (OPE) partitions solution space by generating diverse reasoning outlines before parallel path reasoning. Uses iterative RL strategy that independently optimizes outline planning and outline-guided reasoning to reduce information redundancy.

Result: Extensive experiments across multiple challenging mathematical benchmarks show OPE effectively improves reasoning performance with different aggregation strategies, enabling more reliable discovery of correct solutions.

Conclusion: OPE addresses the mutual information bottleneck in parallel thinking by guiding path exploration with diverse outlines, leading to improved reasoning performance and more reliable solution discovery in large reasoning models.

Abstract: Parallel thinking has emerged as a new paradigm for large reasoning models (LRMs) in tackling complex problems. Recent methods leverage Reinforcement Learning (RL) to enhance parallel thinking, aiming to address the limitations in computational resources and effectiveness encountered with supervised fine-tuning. However, most existing studies primarily focus on optimizing the aggregation phase, with limited attention to the path exploration stage. In this paper, we theoretically analyze the optimization of parallel thinking under the Reinforcement Learning with Verifiable Rewards (RLVR) setting, and identify that the mutual information bottleneck among exploration paths fundamentally restricts overall performance. To address this, we propose Outline-Guided Path Exploration (OPE), which explicitly partitions the solution space by generating diverse reasoning outlines prior to parallel path reasoning, thereby reducing information redundancy and improving the diversity of information captured across exploration paths. We implement OPE with an iterative RL strategy that optimizes outline planning and outline-guided reasoning independently. Extensive experiments across multiple challenging mathematical benchmarks demonstrate that OPE effectively improves reasoning performance in different aggregation strategies, enabling LRMs to more reliably discover correct solutions.

Method: Analyzed root causes of benchmark issues, identified inherent dataset biases and oversimplified evaluation tasks, uncovered additional limitations including unreasonable time-interval formatting, ignorance of knowledge obsolescence, and insufficient information for evolution understanding.

Result: Introduced TKG evolution benchmark with four bias-corrected datasets and two novel tasks aligned with evolution process to promote accurate understanding of TKG evolution modeling challenges.

Conclusion: Current TKG benchmarks have significant flaws that allow exploitation via simple statistical methods; new benchmark addresses these issues for fairer evaluation of TKG evolution models.

Abstract: Temporal knowledge graphs (TKGs) structurally preserve evolving human knowledge. Recent research has focused on designing models to learn the evolutionary nature of TKGs to predict future facts, achieving impressive results. For instance, Hits@10 scores over 0.9 on YAGO dataset. However, we find that existing benchmarks inadvertently introduce a shortcut. Near state-of-the-art performance can be simply achieved by counting co-occurrences, without using any temporal information. In this work, we examine the root cause of this issue, identifying inherent biases in current datasets and over simplified form of evaluation task that can be exploited by these biases. Through this analysis, we further uncover additional limitations of existing benchmarks, including unreasonable formatting of time-interval knowledge, ignorance of learning knowledge obsolescence, and insufficient information for precise evolution understanding, all of which can amplify the shortcut and hinder a fair assessment. Therefore, we introduce the TKG evolution benchmark. It includes four bias-corrected datasets and two novel tasks closely aligned with the evolution process, promoting a more accurate understanding of the challenges in TKG evolution modeling. Benchmark is available at: https://github.com/zjs123/TKG-Benchmark.

Method: SAGE (Self-Aware Guided Efficient Reasoning) is introduced as a novel sampling paradigm that unleashes models’ efficient reasoning potential. The approach integrates SAGE as mixed sampling into group-based reinforcement learning (SAGE-RL), enabling the incorporation of SAGE-discovered efficient reasoning patterns into standard pass@1 inference.

Result: SAGE-RL markedly enhances both reasoning accuracy and efficiency of LRMs across multiple challenging mathematical benchmarks, demonstrating improved computational efficiency without sacrificing performance.

Conclusion: The proposed SAGE framework successfully enables large reasoning models to self-regulate their thinking process, stopping at appropriate times to achieve more efficient reasoning while maintaining or improving accuracy across challenging tasks.

Abstract: Recent advancements in large reasoning models (LRMs) have greatly improved their capabilities on complex reasoning tasks through Long Chains of Thought (CoTs). However, this approach often results in substantial redundancy, impairing computational efficiency and causing significant delays in real-time applications. Recent studies show that longer reasoning chains are frequently uncorrelated with correctness and can even be detrimental to accuracy. In a further in-depth analysis of this phenomenon, we surprisingly uncover and empirically verify that LRMs implicitly know the appropriate time to stop thinking, while this capability is obscured by current sampling paradigms. Motivated by this, we introduce SAGE (Self-Aware Guided Efficient Reasoning), a novel sampling paradigm that unleashes this efficient reasoning potential. Furthermore, integrating SAGE as mixed sampling into group-based reinforcement learning (SAGE-RL) enables SAGE-RL to effectively incorporate SAGE-discovered efficient reasoning patterns into standard pass@1 inference, markedly enhancing both the reasoning accuracy and efficiency of LRMs across multiple challenging mathematical benchmarks.

Method: Three main contributions: 1) Compiling random forests into circuits that directly encode instances in each class; 2) Using these circuits to compute complete and general reasons for decisions; 3) Algorithms for computing decision robustness and shortest ways to flip decisions.

Result: The proposed approach is significantly more efficient than existing similar methods, and enables enumeration of sufficient reasons, necessary reasons, contrastive explanations, robustness computation, and identification of shortest decision-flipping paths across various datasets.

Conclusion: The paper provides efficient circuit-based methods for explaining, analyzing robustness, and understanding decision boundaries of random forest classifiers, offering practical tools for interpretable machine learning.

Abstract: We make three contributions in this paper. First, we present an approach for compiling a random forest classifier into a set of circuits, where each circuit directly encodes the instances in some class of the classifier. We show empirically that our proposed approach is significantly more efficient than existing similar approaches. Next, we utilize this approach to further obtain circuits that are tractable for computing the complete and general reasons of a decision, which are instance abstractions that play a fundamental role in computing explanations. Finally, we propose algorithms for computing the robustness of a decision and all shortest ways to flip it. We illustrate the utility of our contributions by using them to enumerate all sufficient reasons, necessary reasons and contrastive explanations of decisions; to compute the robustness of decisions; and to identify all shortest ways to flip the decisions made by random forest classifiers learned from a wide range of datasets.

Method: Two-stage training: (1) Train a generative subgraph retriever from a unified memory space, (2) Adapt to unseen memory paradigms by training a lightweight alignment module through contrastive learning. This enables fast cross-paradigm alignment with minimal compute.

Result: MemAdapter consistently outperforms five strong agent memory systems across three memory paradigms and agent model scales. Achieves cross-paradigm alignment within 13 minutes on a single GPU with less than 5% of training compute compared to original retrievers.

Conclusion: MemAdapter provides a flexible, efficient plug-and-play solution for agent memory systems that enables effective zero-shot fusion across memory paradigms, significantly reducing alignment costs while maintaining superior performance.

Abstract: Memory mechanism is a core component of LLM-based agents, enabling reasoning and knowledge discovery over long-horizon contexts. Existing agent memory systems are typically designed within isolated paradigms (e.g., explicit, parametric, or latent memory) with tightly coupled retrieval methods that hinder cross-paradigm generalization and fusion. In this work, we take a first step toward unifying heterogeneous memory paradigms within a single memory system. We propose MemAdapter, a memory retrieval framework that enables fast alignment across agent memory paradigms. MemAdapter adopts a two-stage training strategy: (1) training a generative subgraph retriever from the unified memory space, and (2) adapting the retriever to unseen memory paradigms by training a lightweight alignment module through contrastive learning. This design improves the flexibility for memory retrieval and substantially reduces alignment cost across paradigms. Comprehensive experiments on three public evaluation benchmarks demonstrate that the generative subgraph retriever consistently outperforms five strong agent memory systems across three memory paradigms and agent model scales. Notably, MemAdapter completes cross-paradigm alignment within 13 minutes on a single GPU, achieving superior performance over original memory retrievers with less than 5% of training compute. Furthermore, MemAdapter enables effective zero-shot fusion across memory paradigms, highlighting its potential as a plug-and-play solution for agent memory systems.

Method: Introduces VIRF (Verifiable Iterative Refinement Framework) with a tutor-apprentice dialogue: a deterministic Logic Tutor grounded in formal safety ontology provides causal and pedagogical feedback to an LLM planner. Also includes a scalable knowledge acquisition pipeline that synthesizes safety knowledge bases from real-world documents.

Result: Achieves 0% Hazardous Action Rate (HAR) and 77.3% Goal-Condition Rate (GCR) in challenging home safety tasks, highest among all baselines. Highly efficient with only 1.1 correction iterations on average.

Conclusion: VIRF demonstrates a principled pathway toward building fundamentally trustworthy and verifiably safe embodied agents by combining neuro-symbolic approaches to address safety limitations of pure LLM-based planners.

Abstract: Large Language Models (LLMs) show promise as planners for embodied AI, but their stochastic nature lacks formal reasoning, preventing strict safety guarantees for physical deployment. Current approaches often rely on unreliable LLMs for safety checks or simply reject unsafe plans without offering repairs. We introduce the Verifiable Iterative Refinement Framework (VIRF), a neuro-symbolic architecture that shifts the paradigm from passive safety gatekeeping to active collaboration. Our core contribution is a tutor-apprentice dialogue where a deterministic Logic Tutor, grounded in a formal safety ontology, provides causal and pedagogical feedback to an LLM planner. This enables intelligent plan repairs rather than mere avoidance. We also introduce a scalable knowledge acquisition pipeline that synthesizes safety knowledge bases from real-world documents, correcting blind spots in existing benchmarks. In challenging home safety tasks, VIRF achieves a perfect 0 percent Hazardous Action Rate (HAR) and a 77.3 percent Goal-Condition Rate (GCR), which is the highest among all baselines. It is highly efficient, requiring only 1.1 correction iterations on average. VIRF demonstrates a principled pathway toward building fundamentally trustworthy and verifiably safe embodied agents.

Method: SCOUT-RAG uses four cooperative agents: (1) estimates domain relevance, (2) decides when to expand retrieval to additional domains, (3) adapts traversal depth to avoid unnecessary graph exploration, and (4) synthesizes high-quality answers. The framework minimizes retrieval regret while controlling latency and API costs.

Result: SCOUT-RAG achieves performance comparable to centralized baselines (including DRIFT and exhaustive domain traversal) while substantially reducing cross-domain calls, total tokens processed, and latency across multi-domain knowledge settings.

Conclusion: SCOUT-RAG provides an effective distributed Graph-RAG solution for access-restricted environments, enabling efficient cross-domain retrieval without requiring global graph visibility or exhaustive querying.

Abstract: Graph-RAG improves LLM reasoning using structured knowledge, yet conventional designs rely on a centralized knowledge graph. In distributed and access-restricted settings (e.g., hospitals or multinational organizations), retrieval must select relevant domains and appropriate traversal depth without global graph visibility or exhaustive querying. To address this challenge, we introduce \textbf{SCOUT-RAG} (\textit{\underline{S}calable and \underline{CO}st-efficient \underline{U}nifying \underline{T}raversal}), a distributed agentic Graph-RAG framework that performs progressive cross-domain retrieval guided by incremental utility goals. SCOUT-RAG employs four cooperative agents that: (i) estimate domain relevance, (ii) decide when to expand retrieval to additional domains, (iii) adapt traversal depth to avoid unnecessary graph exploration, and (iv) synthesize the high-quality answers. The framework is designed to minimize retrieval regret, defined as missing useful domain information, while controlling latency and API cost. Across multi-domain knowledge settings, SCOUT-RAG achieves performance comparable to centralized baselines, including DRIFT and exhaustive domain traversal, while substantially reducing cross-domain calls, total tokens processed, and latency.

Method: AGENTWM exploits semantic equivalence of action sequences by subtly biasing the distribution of functionally identical tool execution paths. It embeds watermarks directly into visible action trajectories while remaining indistinguishable to users. The framework includes an automated pipeline to generate robust watermark schemes and a rigorous statistical hypothesis testing procedure for verification.

Result: Extensive evaluations across three complex domains show AGENTWM achieves high detection accuracy with negligible impact on agent performance. The framework effectively protects agentic IP against adaptive adversaries who cannot remove watermarks without severely degrading the stolen model’s utility.

Conclusion: AGENTWM is the first watermarking framework specifically designed for agentic models, addressing the unique challenges of protecting IP in agentic LLM systems where traditional watermarking approaches fail due to grey-box operation and lack of access to internal reasoning traces.

Abstract: The evolution of Large Language Models (LLMs) into agentic systems that perform autonomous reasoning and tool use has created significant intellectual property (IP) value. We demonstrate that these systems are highly vulnerable to imitation attacks, where adversaries steal proprietary capabilities by training imitation models on victim outputs. Crucially, existing LLM watermarking techniques fail in this domain because real-world agentic systems often operate as grey boxes, concealing the internal reasoning traces required for verification. This paper presents AGENTWM, the first watermarking framework designed specifically for agentic models. AGENTWM exploits the semantic equivalence of action sequences, injecting watermarks by subtly biasing the distribution of functionally identical tool execution paths. This mechanism allows AGENTWM to embed verifiable signals directly into the visible action trajectory while remaining indistinguishable to users. We develop an automated pipeline to generate robust watermark schemes and a rigorous statistical hypothesis testing procedure for verification. Extensive evaluations across three complex domains demonstrate that AGENTWM achieves high detection accuracy with negligible impact on agent performance. Our results confirm that AGENTWM effectively protects agentic IP against adaptive adversaries, who cannot remove the watermarks without severely degrading the stolen model’s utility.

Method: Proposes Personalized Agent Security Bench (PASB), an end-to-end security evaluation framework that incorporates personalized usage scenarios, realistic toolchains, and long-horizon interactions for black-box, end-to-end security evaluation on real systems.

Result: Evaluation of OpenClaw reveals critical vulnerabilities at different execution stages including user prompt processing, tool usage, and memory retrieval, highlighting substantial security risks in personalized agent deployments.

Conclusion: Personalized AI agents like OpenClaw have significant security vulnerabilities in real-world deployments, and the PASB framework provides a comprehensive approach to evaluate and address these security risks.

Abstract: Although large language model (LLM)-based agents, exemplified by OpenClaw, are increasingly evolving from task-oriented systems into personalized AI assistants for solving complex real-world tasks, their practical deployment also introduces severe security risks. However, existing agent security research and evaluation frameworks primarily focus on synthetic or task-centric settings, and thus fail to accurately capture the attack surface and risk propagation mechanisms of personalized agents in real-world deployments. To address this gap, we propose Personalized Agent Security Bench (PASB), an end-to-end security evaluation framework tailored for real-world personalized agents. Building upon existing agent attack paradigms, PASB incorporates personalized usage scenarios, realistic toolchains, and long-horizon interactions, enabling black-box, end-to-end security evaluation on real systems. Using OpenClaw as a representative case study, we systematically evaluate its security across multiple personalized scenarios, tool capabilities, and attack types. Our results indicate that OpenClaw exhibits critical vulnerabilities at different execution stages, including user prompt processing, tool usage, and memory retrieval, highlighting substantial security risks in personalized agent deployments. The code for the proposed PASB framework is available at https://github.com/AstorYH/PASB.

Method: Reframes alignment evaluation as information flow under partial observability, showing behavior divergence is bounded by mutual information between internal representations and regime variable. Proposes regime-blind training using adversarial invariance to reduce extractability of regime information at decision-relevant representations.

Result: Regime-blind training suppresses regime-conditioned behavior in scientific sycophancy and temporal sleeper agent cases without measurable task utility loss. Sycophancy shows sharp transition at low intervention strength, while sleeper agents require stronger pressure without clean collapse of regime decodability.

Conclusion: Representational invariance is meaningful but limited control lever; effectiveness depends on how regime information is embedded in policy. Behavioral evaluation should be complemented with white-box diagnostics of regime awareness and information flow.

Abstract: Safety evaluation for advanced AI systems implicitly assumes that behavior observed under evaluation is predictive of behavior in deployment. This assumption becomes fragile for agents with situational awareness, which may exploitregime leakage-informational cues distinguishing evaluation from deployment-to implement conditional policies such as sycophancy and sleeper agents, which preserve compliance under oversight while defecting in deployment-like regimes. We reframe alignment evaluation as a problem of information flow under partial observability. Within this framework, we show that divergence between evaluation-time and deployment-time behavior is bounded by the mutual information between internal representations and the regime variable. Motivated by this result, we study regime-blind mechanisms: training-time interventions that reduce the extractability of regime information at decision-relevant internal representations via adversarial invariance. We evaluate this approach on a base, open-weight language model across two fully characterized failure modes -scientific sycophancy and temporal sleeper agents. Regime-blind training suppresses regime-conditioned behavior in both evaluated cases without measurable loss of task utility, but with qualitatively different dynamics: sycophancy exhibits a sharp representational and behavioral transition at low intervention strength, whereas sleeper-agent behavior requires substantially stronger pressure and does not exhibit a clean collapse of regime decodability. These results demonstrate that representational invariance is a meaningful but fundamentally limited control lever, whose effectiveness depends on how regime information is embedded in the policy. We argue that behavioral evaluation should be complemented with white-box diagnostics of regime awareness and information flow.

Method: Proposes TreeTensor, a general nested data container that uses a constrained tree-structure perspective to model data relationships. It allows applying arbitrary functions and operations to nested data with minimal overhead through various constraints and utilities.

Result: TreeTensor provides powerful usability in various problems, including complex AI systems like AlphaStar for StarCraftII, while exhibiting excellent runtime efficiency without overhead. Benchmarks show it works with libraries like Scikit-Learn, NumPy, and PyTorch.

Conclusion: TreeTensor addresses the limitations of conventional tensors for nested, multi-modal data, enabling efficient computation on hierarchical structures while maintaining compatibility with existing ML ecosystems.

Abstract: Tensor is the most basic and essential data structure of nowadays artificial intelligence (AI) system. The natural properties of Tensor, especially the memory-continuity and slice-independence, make it feasible for training system to leverage parallel computing unit like GPU to process data simultaneously in batch, spatial or temporal dimensions. However, if we look beyond perception tasks, the data in a complicated cognitive AI system usually has hierarchical structures (i.e. nested data) with various modalities. They are inconvenient and inefficient to program directly with conventional Tensor with fixed shape. To address this issue, we summarize two main computational patterns of nested data, and then propose a general nested data container: TreeTensor. Through various constraints and magic utilities of TreeTensor, one can apply arbitrary functions and operations to nested data with almost zero cost, including some famous machine learning libraries, such as Scikit-Learn, Numpy and PyTorch. Our approach utilizes a constrained tree-structure perspective to systematically model data relationships, and it can also easily be combined with other methods to extend more usages, such as asynchronous execution and variable-length data computation. Detailed examples and benchmarks show TreeTensor not only provides powerful usability in various problems, especially one of the most complicated AI systems at present: AlphaStar for StarCraftII, but also exhibits excellent runtime efficiency without any overhead. Our project is available at https://github.com/opendilab/DI-treetensor.

Method: Introduces Reinforcement Inference, an entropy-aware inference-time control strategy that uses the model’s own uncertainty to selectively invoke a second, more deliberate reasoning attempt. The method monitors entropy during generation and triggers additional reasoning when uncertainty is detected.

Result: On 12,032 MMLU-Pro questions across 14 subjects using DeepSeek-v3.2 with deterministic decoding, Reinforcement Inference improved accuracy from 60.72% to 84.03% while only incurring 61.06% additional inference calls. A 100% re-asking ablation reached 84.35%, showing uncertainty-aware selection captures most improvement with less compute.

Conclusion: The method provides a practical inference-time upgrade and suggests a broader entropy-aware paradigm for measuring and expanding model capability. It offers a diagnostic lens on LLMs’ latent reasoning horizon and motivates future training objectives that explicitly constrain correctness-confidence alignment.

Abstract: Modern large language models (LLMs) are often evaluated and deployed under a \emph{one-shot, greedy} inference protocol, especially in professional settings that require deterministic behavior. This regime can systematically under-estimate a fixed model’s true capability: many errors arise not from missing knowledge, but from premature commitment under internal ambiguity. We introduce \emph{Reinforcement Inference}, an entropy-aware inference-time control strategy that uses the model’s own uncertainty to selectively invoke a second, more deliberate reasoning attempt, enabling stronger performance \emph{without any retraining}. On 12,032 MMLU-Pro questions across 14 subjects, using DeepSeek-v3.2 with deterministic decoding in a zero-shot setting, Reinforcement Inference improves accuracy from 60.72% to 84.03%, while only incurring 61.06% additional inference calls. A 100% re-asking ablation reaches 84.35%, indicating that uncertainty-aware selection captures most of the attainable improvement with substantially less compute. Moreover, a \emph{prompt-only} ablation underperforms the baseline, suggesting that the gains are not explained by generic `` your output had high entropy, think step-by-step’’ prompting alone. Beyond providing a practical inference-time upgrade, our results suggest a broader \emph{entropy-aware} paradigm for measuring and expanding model capability: because modern decoder-based models generate outputs autoregressively, entropy and related confidence measures arise naturally as first-class control signals during generation. The resulting gap between one-pass greedy inference and uncertainty-conditioned deliberation offers a diagnostic lens on an LLM’s latent reasoning horizon and motivates future training objectives that explicitly constrain correctness–confidence alignment.

Method: Two-agent game paradigm with user agent (style mimicry + active termination) and dialogue agent. Adaptive Tree-based Group Relative Policy Optimization (AT-GRPO) treats dialogue trajectories as trees with adaptive observation ranges - larger ranges for early topic exploration, smaller for late-stage maintenance.

Result: Framework shows superior performance, sample efficiency, and robustness in experiments. Reduces rollout budgets from exponential to polynomial in dialogue length while preserving long-term reward capture.

Conclusion: Proposed long-horizon RL framework effectively addresses limitations of existing dialogue personalization methods by enabling online adaptation and long-term engagement optimization through tree-based policy optimization.

Abstract: Open-ended dialogue agents aim to deliver engaging, personalized interactions by adapting to users’ traits, but existing methods face critical limitations: over-reliance on pre-collected user data, and short-horizon biases in reinforcement learning (RL) that neglect long-term dialogue value. To address these, we propose a novel long-horizon RL framework integrating online personalization with Adaptive Tree-based Group Relative Policy Optimization (AT-GRPO). Adopting a two-agent game paradigm, a user agent constructs dynamic environments via style mimicry (learning user-specific conversational traits) and active termination (predicting turn-level termination probabilities as immediate rewards), forming an iterative cycle that drives the dialogue agent to deepen interest exploration. AT-GRPO reinterprets dialogue trajectories as trees and introduces adaptive observation ranges. Unlike full tree expansion that incurs exponential overhead, it limits each node to aggregate rewards from a stage-aware range: larger ranges support early-stage topic exploration, while smaller ranges facilitate late-stage dialogue maintenance. This design reduces rollout budgets from exponential to polynomial in the dialogue length, while preserving long-term reward capture. Extensive experiments show our framework’s superior performance, sample efficiency, and robustness.

Method: Proposes a unified theoretical framework decomposing multi-agent reasoning gains into three dimensions: Exploration (diverse solution coverage), Information (high-fidelity feedback), and Aggregation (principled consensus). Introduces PRISM framework with role-based diversity, execution-grounded feedback with evidence-based cross-evaluation, and iterative synthesis with closed-loop validation to jointly maximize all three dimensions.

Result: Extensive experiments across mathematical reasoning, code generation, and function calling benchmarks demonstrate that PRISM achieves state-of-the-art performance with superior compute-efficiency compared to methods optimizing partial dimensions.

Conclusion: The theoretical framework provides actionable design principles for future multi-agent reasoning systems, moving beyond heuristic approaches to principled optimization of multi-agent collaboration.

Abstract: Multi-agent collaboration has emerged as a promising paradigm for enhancing reasoning capabilities of Large Language Models (LLMs). However, existing approaches remain largely heuristic, lacking principled guidance on what drives performance gains and how to systematically optimize multi-agent reasoning. Specifically, it remains unclear why multi-agent collaboration outperforms single-agent reasoning and which design choices contribute most to these gains, making it difficult to build better systems. We address this gap by introducing a unified theoretical framework that decomposes multi-agent reasoning gains into three conceptually independent dimensions: Exploration for diverse solution coverage, Information for high-fidelity feedback, and Aggregation for principled consensus. Through this lens, existing methods can be understood as special cases that optimize only subsets of these dimensions. Building upon this decomposition, a novel framework called PRISM (Propose-Review-Integrate Synthesis for Multi-agent Reasoning) is proposed, which jointly maximizes all three dimensions through role-based diversity, execution-grounded feedback with evidence-based cross-evaluation, and iterative synthesis with closed-loop validation. Extensive experiments across mathematical reasoning, code generation, and function calling benchmarks demonstrate that PRISM achieves state-of-the-art performance with superior compute-efficiency compared to methods optimizing partial dimensions. The theoretical framework provides actionable design principles for future multi-agent reasoning systems.

Method: Proposed a top-down attention mechanism to select modalities within a global workspace architecture. Evaluated on two multimodal datasets of increasing complexity: Simple Shapes and MM-IMDb 1.0. Compared against existing multimodal attention baselines on the MM-IMDb 1.0 benchmark.

Result: The attention mechanism improves noise robustness of global workspace systems and demonstrates various cross-task and cross-modality generalization capabilities not found in existing multimodal attention models. On MM-IMDb 1.0 benchmark, the approach makes global workspace competitive with state-of-the-art methods.

Conclusion: Top-down attention mechanisms inspired by cognitive neuroscience can effectively enhance multimodal integration in computational architectures, providing improved robustness and generalization compared to conventional multimodal attention approaches.

Abstract: Global Workspace Theory (GWT), inspired by cognitive neuroscience, posits that flexible cognition could arise via the attentional selection of a relevant subset of modalities within a multimodal integration system. This cognitive framework can inspire novel computational architectures for multimodal integration. Indeed, recent implementations of GWT have explored its multimodal representation capabilities, but the related attention mechanisms remain understudied. Here, we propose and evaluate a top-down attention mechanism to select modalities inside a global workspace. First, we demonstrate that our attention mechanism improves noise robustness of a global workspace system on two multimodal datasets of increasing complexity: Simple Shapes and MM-IMDb 1.0. Second, we highlight various cross-task and cross-modality generalization capabilities that are not shared by multimodal attention models from the literature. Comparing against existing baselines on the MM-IMDb 1.0 benchmark, we find our attention mechanism makes the global workspace competitive with the state of the art.

Method: OSCAR uses offline-online paradigm: 1) Offline phase models CIR as two-stage mixed-integer programming problem to derive optimal trajectories via boolean set operations, storing them in golden library; 2) Online phase uses these trajectories as in-context demonstrations to steer VLM planner during inference.

Result: Extensive experiments on three public benchmarks and private industrial benchmark show OSCAR consistently outperforms SOTA baselines. Achieves superior performance using only 10% of training data, demonstrating strong generalization of planning logic rather than dataset-specific memorization.

Conclusion: OSCAR successfully reformulates agentic CIR from heuristic search to principled trajectory optimization, enabling more effective reasoning over visual-textual constraints with strong generalization capabilities.

Abstract: Composed image retrieval (CIR) requires complex reasoning over heterogeneous visual and textual constraints. Existing approaches largely fall into two paradigms: unified embedding retrieval, which suffers from single-model myopia, and heuristic agentic retrieval, which is limited by suboptimal, trial-and-error orchestration. To this end, we propose OSCAR, an optimization-steered agentic planning framework for composed image retrieval. We are the first to reformulate agentic CIR from a heuristic search process into a principled trajectory optimization problem. Instead of relying on heuristic trial-and-error exploration, OSCAR employs a novel offline-online paradigm. In the offline phase, we model CIR via atomic retrieval selection and composition as a two-stage mixed-integer programming problem, mathematically deriving optimal trajectories that maximize ground-truth coverage for training samples via rigorous boolean set operations. These trajectories are then stored in a golden library to serve as in-context demonstrations for online steering of VLM planner at online inference time. Extensive experiments on three public benchmarks and a private industrial benchmark show that OSCAR consistently outperforms SOTA baselines. Notably, it achieves superior performance using only 10% of training data, demonstrating strong generalization of planning logic rather than dataset-specific memorization.

Method: Introduces Debate Query Complexity (DQC) as a measure of minimum bits a verifier must inspect to correctly decide a debate. Analyzes the relationship between DQC and computational complexity classes, establishing bounds and characterizations.

Result: Found that PSPACE/poly is precisely the class of functions decidable with O(log n) queries, showing debate is remarkably query-efficient. Established that functions depending on all input bits require Ω(log n) queries, and any function computable by circuit size s satisfies DQC(f) ≤ log(s) + 3.

Conclusion: Debate is highly query-efficient for complex problems, requiring only logarithmic oversight. The connection between DQC lower bounds and circuit complexity suggests that proving certain DQC bounds for languages in P would yield new circuit lower bounds, linking debate query complexity to central questions in computational complexity.

Abstract: AI safety via debate uses two competing models to help a human judge verify complex computational tasks. Previous work has established what problems debate can solve in principle, but has not analysed the practical cost of human oversight: how many queries must the judge make to the debate transcript? We introduce Debate Query Complexity}(DQC), the minimum number of bits a verifier must inspect to correctly decide a debate. Surprisingly, we find that PSPACE/poly (the class of problems which debate can efficiently decide) is precisely the class of functions decidable with O(log n) queries. This characterisation shows that debate is remarkably query-efficient: even for highly complex problems, logarithmic oversight suffices. We also establish that functions depending on all their input bits require Omega(log n) queries, and that any function computable by a circuit of size s satisfies DQC(f) <= log(s) + 3. Interestingly, this last result implies that proving DQC lower bounds of log(n) + 6 for languages in P would yield new circuit lower bounds, connecting debate query complexity to central questions in circuit complexity.

Method: Conceptual analysis and theoretical framework development. The authors propose reframing chatbots as highly skilled salespeople whose objectives are determined by deploying organizations, rather than as companions or assistants.

Result: Identifies that competing notions of “trust” under a shared term obscure important distinctions between psychological trust formation and normative trustworthiness, requiring further research and stronger support mechanisms for appropriate trust calibration.

Conclusion: Addressing the trust gap in conversational AI requires recognizing how design choices influence behavioral trust formation, distinguishing it from normative trustworthiness, and developing better mechanisms to help users appropriately calibrate trust in chatbot systems.

Abstract: As chatbots increasingly blur the boundary between automated systems and human conversation, the foundations of trust in these systems warrant closer examination. While regulatory and policy frameworks tend to define trust in normative terms, the trust users place in chatbots often emerges from behavioral mechanisms. In many cases, this trust is not earned through demonstrated trustworthiness but is instead shaped by interactional design choices that leverage cognitive biases to influence user behavior. Based on this observation, we propose reframing chatbots not as companions or assistants, but as highly skilled salespeople whose objectives are determined by the deploying organization. We argue that the coexistence of competing notions of “trust” under a shared term obscures important distinctions between psychological trust formation and normative trustworthiness. Addressing this gap requires further research and stronger support mechanisms to help users appropriately calibrate trust in conversational AI systems.

Method: Uses SAT solvers for small instances, introduces the literal graph representation of STRIPS planning problems, and maps these to Petri nets to analyze computational complexity properties.

Result: Provides insights into the “small solution hypothesis” for STRIPS$^1_1$, though specific results aren’t detailed in the abstract. The approach combines computational experiments with theoretical analysis through graph and Petri net representations.

Conclusion: Advances understanding of computational complexity in propositional STRIPS planning, particularly for the simplest operator case, using a multi-method approach that bridges experimental and theoretical analysis.

Abstract: This paper is based on Bylander’s results on the computational complexity of propositional STRIPS planning. He showed that when only ground literals are permitted, determining plan existence is PSPACE-complete even if operators are limited to two preconditions and two postconditions. While NP-hardness is settled, it is unknown whether propositional STRIPS with operators that only have one precondition and one effect is NP-complete. We shed light on the question whether this small solution hypothesis for STRIPS$^1_1$ is true, calling a SAT solver for small instances, introducing the literal graph, and mapping it to Petri nets.