Method: Developed a 7B-parameter end-to-end LALM with large-scale curated pretraining and targeted post-training, featuring variants for dialogue (Covo-Audio-Chat) and full-duplex interaction (Covo-Audio-Chat-FD), plus an intelligence-speaker decoupling strategy for flexible voice customization.

Result: Achieves state-of-the-art or competitive performance across multiple benchmarks for speech-text comprehension, semantic reasoning, spoken dialogue, and audio understanding, with strong conversational abilities and full-duplex interaction capabilities.

Conclusion: 7B-scale models can effectively integrate sophisticated audio intelligence with high-level semantic reasoning, suggesting a scalable path toward more capable and versatile LALMs for real-world conversational systems.

Abstract: In this work, we present Covo-Audio, a 7B-parameter end-to-end LALM that directly processes continuous audio inputs and generates audio outputs within a single unified architecture. Through large-scale curated pretraining and targeted post-training, Covo-Audio achieves state-of-the-art or competitive performance among models of comparable scale across a broad spectrum of tasks, including speech-text modeling, spoken dialogue, speech understanding, audio understanding, and full-duplex voice interaction. Extensive evaluations demonstrate that the pretrained foundation model exhibits strong speech-text comprehension and semantic reasoning capabilities on multiple benchmarks, outperforming representative open-source models of comparable scale. Furthermore, Covo-Audio-Chat, the dialogue-oriented variant, demonstrates strong spoken conversational abilities, including understanding, contextual reasoning, instruction following, and generating contextually appropriate and empathetic responses, validating its applicability to real-world conversational assistant scenarios. Covo-Audio-Chat-FD, the evolved full-duplex model, achieves substantially superior performance on both spoken dialogue capabilities and full-duplex interaction behaviors, demonstrating its competence in practical robustness. To mitigate the high cost of deploying end-to-end LALMs for natural conversational systems, we propose an intelligence-speaker decoupling strategy that separates dialogue intelligence from voice rendering, enabling flexible voice customization with minimal text-to-speech (TTS) data while preserving dialogue performance. Overall, our results highlight the strong potential of 7B-scale models to integrate sophisticated audio intelligence with high-level semantic reasoning, and suggest a scalable path toward more capable and versatile LALMs.

Method: Adapts causal tracing to investigate internal information flow of LALMs during audio comprehension. Conducts layer-wise and token-wise analyses across DeSTA, Qwen, and Voxtral models, evaluating causal effects of individual hidden states.

Result: Layer-wise analysis reveals different fusion strategies: progressive integration in DeSTA vs abrupt late-stage fusion in Qwen. Token-wise analysis shows the final sequence token acts as an informational bottleneck for retrieving relevant audio information. Also observes attention-like query mechanism at intermediate token positions that triggers pulling task-relevant audio context.

Conclusion: The findings provide clear characterization of when and where multimodal integration occurs within LALMs, offering insights into their internal mechanisms for audio-text fusion.

Abstract: Despite the strong performance of large audio language models (LALMs) in various tasks, exactly how and where they integrate acoustic features with textual context remains unclear. We adapt causal tracing to investigate the internal information flow of LALMs during audio comprehension. By conducting layer-wise and token-wise analyses across DeSTA, Qwen, and Voxtral, we evaluate the causal effects of individual hidden states. Layer-wise analysis identifies different fusion strategies, from progressive integration in DeSTA to abrupt late-stage fusion in Qwen. Token-wise analysis shows that the final sequence token acts as an informational bottleneck where the network decisively retrieves relevant information from the audio. We also observe an attention-like query mechanism at intermediate token positions that triggers the model to pull task-relevant audio context. These findings provide a clear characterization of when and where multi-modal integration occurs within LALMs.

Method: Info-VLA uses two complementary constraints: 1) Replay Anchor Contrastive Learning constructs stable alignment anchors from a frozen teacher model to preserve cross-modal alignment in representation space; 2) Cross-Modal Mutual Information Maximization preserves dependency structure between visual and language representations through mutual information constraints. This jointly preserves historical alignment and cross-modal dependency information.

Result: Experiments on the LIBERO benchmark show that Info-VLA significantly outperforms existing methods in both task retention and adaptation, demonstrating effective mitigation of catastrophic forgetting in VLA models.

Conclusion: Info-VLA successfully addresses catastrophic forgetting in VLA models by preserving cross-modal information structure through complementary constraints, balancing stability and plasticity during continual learning for robotic applications.

Abstract: When deployed in open-ended robotic environments, Vision–Language–Action (VLA) models need to continually acquire new skills, yet suffer from severe catastrophic forgetting. We observe that this degradation is related to the deterioration of cross-modal information structure, where dependencies among visual observations, language instructions, and actions progressively diffuse during continual adaptation. But existing continual learning methods fail to preserve such cross-modal information dependencies. Thus, we propose Info-VLA, an information-preserving continual learning framework that maintains cross-modal information structure through two complementary constraints. Replay Anchor Contrastive Learning constructs stable alignment anchors from a frozen teacher model, preserving cross-modal alignment in the representation space. Cross-Modal Mutual Information Maximization further preserves dependency structure between visual and language representations through mutual information constraints. By jointly preserving historical alignment and cross-modal dependency information, Info-VLA balances stability and plasticity during continual learning. Furthermore, experiments on the LIBERO show that Info-VLA significantly outperforms existing methods in both task retention and adaptation.

Method: Proposes a greedy search-guided chain-of-thought framework for slang interpretation, integrating greedy search algorithms with chain-of-thought prompting for small language models.

Result: Model size and temperature settings have limited impact on inference accuracy; larger models don’t outperform smaller ones; the proposed framework demonstrates improved accuracy in slang meaning interpretation.

Conclusion: The findings contribute to understanding context dependency in language models and provide a practical solution for enhancing slang comprehension through structured reasoning prompting.

Abstract: Slang interpretation has been a challenging downstream task for Large Language Models (LLMs) as the expressions are inherently embedded in contextual, cultural, and linguistic frameworks. In the absence of domain-specific training data, it is difficult for LLMs to accurately interpret slang meaning based on lexical information. This paper attempts to investigate the challenges of slang inference using large LLMs and presents a greedy search-guided chain-of-thought framework for slang interpretation. Through our experiments, we conclude that the model size and temperature settings have limited impact on inference accuracy. Transformer-based models with larger active parameters do not generate higher accuracy than smaller models. Based on the results of the above empirical study, we integrate greedy search algorithms with chain-of-thought prompting for small language models to build a framework that improves the accuracy of slang interpretation. The experimental results indicate that our proposed framework demonstrates improved accuracy in slang meaning interpretation. These findings contribute to the understanding of context dependency in language models and provide a practical solution for enhancing slang comprehension through a structured reasoning prompting framework.

Method: The authors identify Style Modulation Heads - a sparse subset of attention heads (only three heads) that independently govern persona and style formation. They localize these heads via geometric analysis of internal representations, combining layer-wise cosine similarity and head-wise contribution scores. Intervention targets only these specific heads rather than the entire residual stream.

Result: Targeting only the identified Style Modulation Heads achieves robust behavioral control while significantly mitigating the coherency degradation observed in residual stream steering. The method shows that precise, component-level localization enables safer and more precise model control.

Conclusion: Sparse attention heads can be identified and targeted for precise behavioral control in LLMs, reducing coherency degradation compared to residual stream interventions. This component-level approach enables safer and more practical model control.

Abstract: Activation steering offers a computationally efficient mechanism for controlling Large Language Models (LLMs) without fine-tuning. While effectively controlling target traits (e.g., persona), coherency degradation remains a major obstacle to safety and practical deployment. We hypothesize that this degradation stems from intervening on the residual stream, which indiscriminately affects aggregated features and inadvertently amplifies off-target noise. In this work, we identify a sparse subset of attention heads (only three heads) that independently govern persona and style formation, which we term Style Modulation Heads. Specifically, these heads can be localized via geometric analysis of internal representations, combining layer-wise cosine similarity and head-wise contribution scores. We demonstrate that intervention targeting only these specific heads achieves robust behavioral control while significantly mitigating the coherency degradation observed in residual stream steering. More broadly, our findings show that precise, component-level localization enables safer and more precise model control.

Method: REDEREF integrates four key components: (1) belief-guided delegation via Thompson sampling to prioritize agents with historically positive contributions, (2) reflection-driven re-routing using calibrated LLM or programmatic judges, (3) evidence-based selection instead of output averaging, and (4) memory-aware priors to reduce cold-start inefficiency.

Result: Across multi-agent split-knowledge tasks, REDEREF reduces token usage by 28%, agent calls by 17%, and time-to-success by 19% compared to random recursive delegation. It maintains effectiveness even under agent or judge degradation.

Conclusion: Simple, interpretable probabilistic control can significantly improve the efficiency and robustness of multi-agent LLM systems without requiring training or fine-tuning, making practical deployment more feasible.

Abstract: Multi-agent large language model (LLM) systems enable complex, long-horizon reasoning by composing specialized agents, but practical deployment remains hindered by inefficient routing, noisy feedback, and high interaction cost. We introduce REDEREF, a lightweight and training-free controller for multi-agent LLM collaboration that improves routing efficiency during recursive delegation. REDEREF integrates (i) belief-guided delegation via Thompson sampling to prioritize agents with historically positive marginal contributions, (ii) reflection-driven re-routing using a calibrated LLM or programmatic judge, (iii) evidence-based selection rather than output averaging, and (iv) memory-aware priors to reduce cold-start inefficiency. Across multi-agent split-knowledge tasks, we show that while recursive retry alone saturates task success, belief-guided routing reduces token usage by 28%, agent calls by 17%, and time-to-success by 19% compared to random recursive delegation, and adapts gracefully under agent or judge degradation. These results demonstrate that simple, interpretable probabilistic control can meaningfully improve the efficiency and robustness of multi-agent LLM systems without training or fine-tuning.

Method: Forced-completion probing presents identical queries with known correct and incorrect single-token continuations, tracking five geometric measurements across every layer of four decoder-only models (1.5B-13B parameters).

Result: 1) Correct and incorrect paths diverge through rotation, not rescaling; 2) Models actively suppress correct answers when given wrong information; 3) These phenomena emerge at ~1.6B parameters, suggesting a phase transition in factual processing capability.

Conclusion: Factual constraint processing has a specific geometric character (rotational, not scalar; active, not passive) that is invisible to single-layer probes or magnitude comparisons, revealing dynamic internal mechanisms of truthfulness.

Abstract: When a language model is fed a wrong answer, what happens inside the network? Current understanding treats truthfulness as a static property of individual-layer representations-a direction to be probed, a feature to be extracted. Less is known about the dynamics: how internal representations diverge across the full depth of the network when the model processes correct versus incorrect continuations. We introduce forced-completion probing, a method that presents identical queries with known correct and incorrect single-token continuations and tracks five geometric measurements across every layer of four decoder-only models(1.5B-13B parameters). We report three findings. First, correct and incorrect paths diverge through rotation, not rescaling: displacement vectors maintain near-identical magnitudes while their angular separation increases, meaning factual selection is encoded in direction on an approximate hypersphere. Second, the model does not passively fail on incorrect input-it actively suppresses the correct answer, driving internal probability away from the right token. Third, both phenomena are entirely absent below a parameter threshold and emerge at 1.6B, suggesting a phase transition in factual processing capability. These results show that factual constraint processing has a specific geometric character-rotational, not scalar; active, not passive-that is invisible to methods based on single-layer probes or magnitude comparisons.

Method: TSD-KD combines indirect and direct distillation: 1) Indirect distillation uses preference ranking feedback where teacher re-ranks student’s candidate responses without enforcing full distribution, 2) Direct distillation selectively distills tokens based on relative confidence between teacher and student, 3) Adds entropy regularization to maintain student’s confidence.

Result: Achieves state-of-the-art performance on 10 challenging reasoning benchmarks, outperforming baseline and runner-up in accuracy by up to 54.4% and 40.3% respectively. Notably, students trained with TSD-KD sometimes outperform their own teacher models by up to 20.3%.

Conclusion: TSD-KD provides targeted, indirect feedback that supports student’s own reasoning process and facilitates self-improvement, enabling effective transfer of reasoning abilities from large to small models while avoiding distribution mismatch issues.

Abstract: Knowledge Distillation (KD) can transfer the reasoning abilities of large models to smaller ones, which can reduce the costs to generate Chain-of-Thoughts for reasoning tasks. KD methods typically ask the student to mimic the teacher’s distribution over the entire output. However, a student with limited capacity can be overwhelmed by such extensive supervision causing a distribution mismatch, especially in complex reasoning tasks. We propose Token-Selective Dual Knowledge Distillation (TSD-KD), a framework for student-centric distillation. TSD-KD focuses on distilling important tokens for reasoning and encourages the student to explain reasoning in its own words. TSD-KD combines indirect and direct distillation. Indirect distillation uses a weak form of feedback based on preference ranking. The student proposes candidate responses generated on its own; the teacher re-ranks those candidates as indirect feedback without enforcing its entire distribution. Direct distillation uses distribution matching; however, it selectively distills tokens based on the relative confidence between teacher and student. Finally, we add entropy regularization to maintain the student’s confidence during distillation. Overall, our method provides the student with targeted and indirect feedback to support its own reasoning process and to facilitate self-improvement. The experiments show the state-of-the-art performance of TSD-KD on 10 challenging reasoning benchmarks, outperforming the baseline and runner-up in accuracy by up to 54.4% and 40.3%, respectively. Notably, a student trained by TSD-KD even outperformed its own teacher model in four cases by up to 20.3%. The source code is available at https://github.com/kmswin1/TSD-KD.

Method: An agentic workflow that iteratively generates synthetic tweets conditioned on target characteristics, evaluates them using compliance checks, and incorporates structured feedback to refine them in subsequent iterations.

Result: The workflow successfully generates synthetic tweets capturing target labels for location and damage level, and the resulting datasets can effectively evaluate AI systems on damage assessment tasks like geolocalization and damage level prediction.

Conclusion: The workflow offers a flexible and scalable alternative to real-world tweet data curation, enabling systematic generation of synthetic social media data across diverse crisis events, contexts, and informatics applications.

Abstract: Twitter (now X) has become an important source of social media data for situational awareness during crises. Crisis informatics research has widely used tweets from Twitter to develop and evaluate artificial intelligence (AI) systems for various crisis-relevant tasks, such as extracting locations and estimating damage levels from tweets to support damage assessment. However, recent changes in Twitter’s data access policies have made it increasingly difficult to curate real-world tweet datasets related to crises. Moreover, existing curated tweet datasets are limited to past crisis events in specific contexts and are costly to annotate at scale. These limitations constrain the development and evaluation of AI systems used in crisis informatics. To address these limitations, we introduce an agentic workflow for generating crisis-related synthetic tweet datasets. The workflow iteratively generates synthetic tweets conditioned on prespecified target characteristics, evaluates them using predefined compliance checks, and incorporates structured feedback to refine them in subsequent iterations. As a case study, we apply the workflow to generate synthetic tweet datasets relevant to post-earthquake damage assessment. We show that the workflow can generate synthetic tweets that capture their target labels for location and damage level. We further demonstrate that the resulting synthetic tweet datasets can be used to evaluate AI systems on damage assessment tasks like geolocalization and damage level prediction. Our results indicate that the workflow offers a flexible and scalable alternative to real-world tweet data curation, enabling the systematic generation of synthetic social media data across diverse crisis events, societal contexts, and crisis informatics applications.

Method: Used 550 balanced base sentences from ETHICS dataset, generated 26 counterfactual variants per item by systematically varying pronouns and demographic markers, creating 14,850 semantically equivalent sentences. Evaluated six LLM families (Grok, GPT, LLaMA, Gemma, DeepSeek, Mistral) and measured fairness judgments using Statistical Parity Difference (SPD).

Result: Found statistically significant biases: sentences in singular form and third person were more often judged as “fair”, while second person was penalized. Gender markers produced strongest effects - non-binary subjects consistently favored and male subjects disfavored. Patterns suggest distributional and alignment biases learned during training.

Conclusion: LLMs exhibit systematic biases in moral fairness judgments based on grammatical features, emphasizing the need for targeted fairness interventions in moral LLM applications to prevent perpetuating social biases.

Abstract: Large language models (LLMs) are increasingly used to assess moral or ethical statements, yet their judgments may reflect social and linguistic biases. This work presents a controlled, sentence-level study of how grammatical person, number, and gender markers influence LLM moral classifications of fairness. Starting from 550 balanced base sentences from the ETHICS dataset, we generated 26 counterfactual variants per item, systematically varying pronouns and demographic markers to yield 14,850 semantically equivalent sentences. We evaluated six model families (Grok, GPT, LLaMA, Gemma, DeepSeek, and Mistral), and measured fairness judgments and inter-group disparities using Statistical Parity Difference (SPD). Results show statistically significant biases: sentences written in the singular form and third person are more often judged as “fair’’, while those in the second person are penalized. Gender markers produce the strongest effects, with non-binary subjects consistently favored and male subjects disfavored. We conjecture that these patterns reflect distributional and alignment biases learned during training, emphasizing the need for targeted fairness interventions in moral LLM applications.

Method: Created unified benchmark with three datasets (CEX, EXCITE, LinkedBooks) targeting SSH-realistic conditions; evaluated three tasks (extraction, parsing, end-to-end) comparing supervised pipeline (GROBID) vs contemporary LLMs (DeepSeek, Mistral, Gemma, Qwen3-VL) with schema-constrained setup; tested LoRA adaptation and segmentation/pipelining

Result: Extraction saturates beyond moderate capability threshold, while parsing and end-to-end parsing remain bottlenecks due to structured-output brittleness under noisy layouts; LoRA adaptation yields consistent gains especially on SSH-heavy benchmarks; segmentation/pipelining improves robustness; hybrid deployment via routing recommended

Conclusion: Propose hybrid deployment: use GROBID for well-structured PDFs and escalate multilingual/footnote-heavy documents to task-adapted LLMs; benchmark addresses SSH-specific challenges in bibliographic processing

Abstract: Bibliographic reference extraction and parsing are foundational for citation indexing, linking, and downstream scholarly knowledge-graph construction. However, most established evaluations focus on clean, English, end-of-document bibliographies, and therefore underrepresent the Social Sciences and Humanities (SSH), where citations are frequently multilingual, embedded in footnotes, abbreviated, and shaped by heterogeneous historical conventions. We present a unified benchmark that targets these SSH-realistic conditions across three complementary datasets: CEX (English journal articles spanning multiple disciplines), EXCITE (German/English documents with end-section, footnote-only, and mixed regimes), and LinkedBooks (humanities references with strong stylistic variation and multilinguality). We evaluate three tasks of increasing difficulty – reference extraction, reference parsing, and end-to-end document parsing – under a schema-constrained setup that enables direct comparison between a strong supervised pipeline baseline (GROBID) and contemporary LLMs (DeepSeek-V3.1, Mistral-Small-3.2-24B, Gemma-3-27B-it, and Qwen3-VL (4B-32B variants)). Across datasets, extraction largely saturates beyond a moderate capability threshold, while parsing and end-to-end parsing remain the primary bottlenecks due to structured-output brittleness under noisy layouts. We further show that lightweight LoRA adaptation yields consistent gains – especially on SSH-heavy benchmarks – and that segmentation/pipelining can substantially improve robustness. Finally, we argue for hybrid deployment via routing: leveraging GROBID for well-structured, in-distribution PDFs while escalating multilingual and footnote-heavy documents to task-adapted LLMs.

Method: Collected 19K YouTube comments, used VADER for initial sentiment labeling, LDA for topic modeling, fine-tuned multiple transformer models (BERT, RoBERTa, XLNet, etc.), implemented federated learning for privacy preservation, and applied SHAP for model interpretability.

Result: ELECTRA achieved best performance with 91.32% accuracy; federated learning maintained 89.59% accuracy in 2-client setup while preserving privacy; SHAP identified influential words for sentiment classification.

Conclusion: Transformer models are effective for sentiment analysis on geopolitical content, and federated learning provides viable privacy-preserving alternative without significant performance degradation, enabling ethical social media analysis.

Abstract: The recent escalation of the Iran Israel USA conflict in 2026 has triggered widespread global discussions across social media platforms. As people increasingly use these platforms for expressing opinions, analyzing public sentiment from these discussions can provide valuable insights into global public perception. This study aims to analyze global public sentiment regarding the Iran Israel USA conflict by mining user-generated comments from YouTube news channels. The work contributes to public opinion analysis by introducing a privacy preserving framework that combines topic wise sentiment analysis with modern deep learning techniques and Federated Learning. To achieve this, approximately 19,000 YouTube comments were collected from major international news channels and preprocessed to remove noise and normalize text. Sentiment labels were initially generated using the VADER sentiment analyzer and later validated through manual inspection to improve reliability. Latent Dirichlet Allocation (LDA) was applied to identify key discussion topics related to the conflict. Several transformer-based models, including BERT, RoBERTa, XLNet, DistilBERT, ModernBERT, and ELECTRA, were fine tuned for sentiment classification. The best-performing model was further integrated into a federated learning environment to enable distributed training by preserving user data privacy. Additionally, Explainable Artificial Intelligence (XAI) techniques using SHAP were applied to interpret model predictions and identify influential words affecting sentiment classification. Experimental results demonstrate that transformer models perform effectively, and among them, ELECTRA achieved the best performance with 91.32% accuracy. The federated learning also maintained strong performance while preserving privacy, achieving 89.59% accuracy in a two client configuration.

Method: Proposes CAP-TTA: a test-time adaptation framework that uses context-aware LoRA updates triggered only when bias-risk exceeds a threshold. Employs precomputed diagonal preconditioners for fast and stable updates, avoiding full model retraining.

Result: CAP-TTA reduces bias across toxic-prompt settings and benchmarks, confirmed by human evaluation. Achieves much lower update latency than AdamW/SGD while mitigating catastrophic forgetting and improving narrative fluency over SOTA debiasing baselines.

Conclusion: The framework enables effective on-the-fly adaptation to unfamiliar bias patterns while maintaining debiasing effectiveness and computational efficiency, addressing generalization challenges in LLM debiasing.

Abstract: Although debiased LLMs perform well on known bias patterns, they often fail to generalize to unfamiliar bias prompts, producing toxic outputs. We first validate that such high-bias prompts constitute a \emph{distribution shift} via OOD detection, and show static models degrade under this shift. To adapt on-the-fly, we propose \textbf{CAP-TTA}, a test-time adaptation framework that performs context-aware LoRA updates only when the bias-risk \emph{trigger} exceeds a threshold, using a precomputed diagonal \emph{preconditioner} for fast and stable updates. Across toxic-prompt settings and benchmarks, CAP-TTA reduces bias (confirmed by human evaluation) while achieving much lower update latency than AdamW/SGD; it also mitigates catastrophic forgetting by significantly improving narrative fluency over SOTA debiasing baseline while maintaining comparable debiasing effectiveness.

Method: Compiled a massive dataset spanning Clinical Care, Wellness Health, and Professional Inquiry domains. Proposed an automated scoring framework that integrates multi-model consensus with evidence-based retrieval to dynamically generate fine-grained scoring rubrics (~9.8 per query). Uses hierarchical weighting and safety constraints to quantify medical accuracy, key-point coverage, and risk interception.

Result: Generated rubrics achieve 91.8% concordance rate with clinical expert blind audits. Baseline evaluations reveal significant performance disparities among state-of-the-art models when navigating real-world clinical nuances, highlighting limitations of conventional exam-based metrics.

Conclusion: QuarkMedBench establishes a rigorous, reproducible yardstick for measuring LLM performance on complex health issues, with a framework that inherently supports dynamic knowledge updates to prevent benchmark obsolescence.

Abstract: While Large Language Models (LLMs) excel on standardized medical exams, high scores often fail to translate to high-quality responses for real-world medical queries. Current evaluations rely heavily on multiple-choice questions, failing to capture the unstructured, ambiguous, and long-tail complexities inherent in genuine user inquiries. To bridge this gap, we introduce QuarkMedBench, an ecologically valid benchmark tailored for real-world medical LLM assessment. We compiled a massive dataset spanning Clinical Care, Wellness Health, and Professional Inquiry, comprising 20,821 single-turn queries and 3,853 multi-turn sessions. To objectively evaluate open-ended answers, we propose an automated scoring framework that integrates multi-model consensus with evidence-based retrieval to dynamically generate 220,617 fine-grained scoring rubrics (~9.8 per query). During evaluation, hierarchical weighting and safety constraints structurally quantify medical accuracy, key-point coverage, and risk interception, effectively mitigating the high costs and subjectivity of human grading. Experimental results demonstrate that the generated rubrics achieve a 91.8% concordance rate with clinical expert blind audits, establishing highly dependable medical reliability. Crucially, baseline evaluations on this benchmark reveal significant performance disparities among state-of-the-art models when navigating real-world clinical nuances, highlighting the limitations of conventional exam-based metrics. Ultimately, QuarkMedBench establishes a rigorous, reproducible yardstick for measuring LLM performance on complex health issues, while its framework inherently supports dynamic knowledge updates to prevent benchmark obsolescence.

Method: Created MedPriv-Bench using multi-agent, human-in-the-loop pipeline to synthesize sensitive medical contexts and clinically relevant queries. Established standardized evaluation protocol using pre-trained RoBERTa-NLI model as automated judge to quantify data leakage.

Result: Achieved 85.9% alignment with human experts in privacy evaluation. Extensive evaluation of 9 representative LLMs demonstrated pervasive privacy-utility trade-off, highlighting the tension between clinical usefulness and privacy preservation.

Conclusion: Domain-specific benchmarks are necessary to validate safety and efficacy of medical AI systems in privacy-sensitive environments, as current accuracy-focused benchmarks overlook critical privacy risks like contextual leakage.

Abstract: Recent advances in Retrieval-Augmented Generation (RAG) have enabled large language models (LLMs) to ground outputs in clinical evidence. However, connecting LLMs with external databases introduces the risk of contextual leakage: a subtle privacy threat where unique combinations of medical details enable patient re-identification even without explicit identifiers. Current benchmarks in healthcare heavily focus on accuracy, ignoring such privacy issues, despite strict regulations like Health Insurance Portability and Accountability Act (HIPAA) and General Data Protection Regulation (GDPR). To fill this gap, we present MedPriv-Bench, the first benchmark specifically designed to jointly evaluate privacy preservation and clinical utility in medical open-ended question answering. Our framework utilizes a multi-agent, human-in-the-loop pipeline to synthesize sensitive medical contexts and clinically relevant queries that create realistic privacy pressure. We establish a standardized evaluation protocol leveraging a pre-trained RoBERTa-Natural Language Inference (NLI) model as an automated judge to quantify data leakage, achieving an average of 85.9% alignment with human experts. Through an extensive evaluation of 9 representative LLMs, we demonstrate a pervasive privacy-utility trade-off. Our findings underscore the necessity of domain-specific benchmarks to validate the safety and efficacy of medical AI systems in privacy-sensitive environments.

Method: Trained 45 GPT-2-architecture models with varying parameters (2.9M, 8.9M, 33.5M) on AO-CHILDES corpus for different epochs, then evaluated on a pre-registered ME battery using referential suppression as operationalization of ME.

Result: Models showed significant anti-ME repetition priming (opposite of ME) in all conditions, with priming attenuating as language modeling improved but never crossing zero. Context-dependence diagnostic revealed apparent ME-like patterns were explained by embedding similarity, not referential disambiguation.

Conclusion: Distributional learning on child-directed speech produces repetition-based reference tracking rather than lexical exclusivity, suggesting referential grounding may be necessary for mutual exclusivity to emerge.

Abstract: We present the first systematic evaluation of mutual exclusivity (ME) – the bias to map novel words to novel referents – in text-only language models trained on child-directed speech. We operationalise ME as referential suppression: when a familiar object is relabelled in a two-referent discourse context, ME predicts decreased probability of the labelled noun at a subsequent completion position. Three pilot findings motivate a pre-registered scale-sensitivity experiment: (1) a masked language model (BabyBERTa) is entirely insensitive to multi-sentence referential context; (2) autoregressive models show robust repetition priming – the opposite of ME – when familiar nouns are re-labelled; and (3) a novel context-dependence diagnostic reveals that apparent ME-like patterns with nonce tokens are fully explained by embedding similarity, not referential disambiguation. In the confirmatory experiment, we train 45 GPT-2-architecture models (2.9M, 8.9M, and 33.5M parameters; 5, 10, and 20 epochs on AO-CHILDES; 5 seeds each) and evaluate on a pre-registered ME battery. Anti-ME repetition priming is significant in all 9 cells (85-100% of items; all p < 2.4 x 10^-13). Priming attenuates with improved language modelling (Spearman rho = -0.533, p = 0.0002) but never crosses zero across a 3.8x perplexity range. The context-dependence diagnostic replicates in all 9 cells, and dose-response priming increases with repetitions in 8/9 cells (all trend p < 0.002). These findings indicate that distributional learning on child-directed speech produces repetition-based reference tracking rather than lexical exclusivity. We connect this to the grounded cognition literature and argue that referential grounding may be a necessary ingredient for ME – an empirical claim about required input structure, not a nativist one.

Method: Comprehensive investigation of memristor non-ideality impacts on LLM reasoning, followed by systematic evaluation of three training-free strategies: thinking mode, in-context learning, and module redundancy.

Result: Reasoning capability decreases significantly but varies across benchmarks; shallow layer redundancy is most effective for robustness, thinking mode works better under low noise but degrades at high noise, and in-context learning reduces output length with slight performance trade-off.

Conclusion: The study provides insights into LLM reasoning under hardware non-idealities and practical training-free strategies to improve robustness in memristor-based CIM architectures for efficient LLM deployment.

Abstract: Memristor-based analog compute-in-memory (CIM) architectures provide a promising substrate for the efficient deployment of Large Language Models (LLMs), owing to superior energy efficiency and computational density. However, these architectures suffer from precision issues caused by intrinsic non-idealities of memristors. In this paper, we first conduct a comprehensive investigation into the impact of such typical non-idealities on LLM reasoning. Empirical results indicate that reasoning capability decreases significantly but varies for distinct benchmarks. Subsequently, we systematically appraise three training-free strategies, including thinking mode, in-context learning, and module redundancy. We thus summarize valuable guidelines, i.e., shallow layer redundancy is particularly effective for improving robustness, thinking mode performs better under low noise levels but degrades at higher noise, and in-context learning reduces output length with a slight performance trade-off. Our findings offer new insights into LLM reasoning under non-ideality and practical strategies to improve robustness.

Method: Knowledge distillation from Qwen 3B to Qwen 0.5B across English/Spanish Dolly-15k and code datasets, with chain-of-thought guided reinforcement learning (Group Relative Policy Optimization) using CoT-annotated Codeforces data, followed by 4-bit weight quantization.

Result: Distilled student retains 70-91% of teacher capability in English, up to 95% in Spanish, and up to 93.5% Rouge-L in code. CoT-guided RL improves reasoning coherence and solution correctness over knowledge distillation alone. Quantization further reduces memory and latency.

Conclusion: Knowledge distillation combined with chain-of-thought guided reinforcement learning can produce compact, efficient models suitable for deployment in resource-constrained settings while maintaining substantial performance of larger models.

Abstract: We propose a resource-efficient framework for compressing large language models through knowledge distillation, combined with guided chain-of-thought reinforcement learning. Using Qwen 3B as the teacher and Qwen 0.5B as the student, we apply knowledge distillation across English Dolly-15k, Spanish Dolly-15k, and code BugNet and PyTorrent datasets, with hyperparameters tuned in the English setting to optimize student performance. Across tasks, the distilled student retains a substantial portion of the teacher’s capability while remaining significantly smaller: 70% to 91% in English, up to 95% in Spanish, and up to 93.5% Rouge-L in code. For coding tasks, integrating chain-of-thought prompting with Group Relative Policy Optimization using CoT-annotated Codeforces data improves reasoning coherence and solution correctness compared to knowledge distillation alone. Post-training 4-bit weight quantization further reduces memory footprint and inference latency. These results show that knowledge distillation combined with chain-of-thought guided reinforcement learning can produce compact, efficient models suitable for deployment in resource-constrained settings.

Method: Introduces \dataset benchmark with natural language queries requiring extraction of various data types (text, images, hyperlinks) across four complexity levels, plus Visual Grounding Scraper (VGS) - a multi-stage agentic framework that visually narrows down web content.

Result: Extensive experiments show VGS is effective and robust across diverse backbone models, demonstrating practical performance on live websites.

Conclusion: This work lays foundation for developing practical and robust WIE systems by addressing the temporal evolution challenge of the web through live evaluation.

Abstract: Web information extraction (WIE) is the task of automatically extracting data from web pages, offering high utility for various applications. The evaluation of WIE systems has traditionally relied on benchmarks built from HTML snapshots captured at a single point in time. However, this offline evaluation paradigm fails to account for the temporally evolving nature of the web; consequently, performance on these static benchmarks often fails to generalize to dynamic real-world scenarios. To bridge this gap, we introduce \dataset, a new benchmark designed for evaluating WIE systems directly against live websites. Based on trusted and permission-granted websites, we curate natural language queries that require information extraction of various data categories, such as text, images, and hyperlinks. We further design these queries to represent four levels of complexity, based on the number and cardinality of attributes to be extracted, enabling a granular assessment of WIE systems. In addition, we propose Visual Grounding Scraper (VGS), a novel multi-stage agentic framework that mimics human cognitive processes by visually narrowing down web page content to extract desired information. Extensive experiments across diverse backbone models demonstrate the effectiveness and robustness of VGS. We believe that this study lays the foundation for developing practical and robust WIE systems.

Method: Proposed Omni-Detective, an agentic data generation pipeline using tool-calling to autonomously produce detailed yet minimally hallucinatory multimodal data. Trained two models: Audio-Captioner for audio-only and Omni-Captioner for audio-visual detailed perception. Also designed Omni-Cloze, a cloze-style evaluation benchmark.

Result: Audio-Captioner achieved best performance on MMAU and MMAR among open-source models, surpassing Gemini 2.5 Flash and comparable to Gemini 2.5 Pro. Omni-Captioner set new SOTA on VDC and achieved best detail-hallucination trade-off on video-SALMONN 2. Omni-Cloze proved effective for stable evaluation.

Conclusion: The proposed Omni-Detective framework effectively addresses the detail-hallucination trade-off in multimodal perception, and Omni-Cloze provides a reliable evaluation method for detailed captioning tasks.

Abstract: Fine-grained perception of multimodal information is critical for advancing human-AI interaction. With recent progress in audio-visual technologies, Omni Language Models (OLMs), capable of processing audio and video signals in parallel, have emerged as a promising paradigm for achieving richer understanding and reasoning. However, their capacity to capture and describe fine-grained details remains limited explored. In this work, we present a systematic and comprehensive investigation of omni detailed perception from the perspectives of the data pipeline, models, and benchmark. We first identify an inherent “co-growth” between detail and hallucination in current OLMs. To address this, we propose Omni-Detective, an agentic data generation pipeline integrating tool-calling, to autonomously produce highly detailed yet minimally hallucinatory multimodal data. Based on the data generated with Omni-Detective, we train two captioning models: Audio-Captioner for audio-only detailed perception, and Omni-Captioner for audio-visual detailed perception. Under the cascade evaluation protocol, Audio-Captioner achieves the best performance on MMAU and MMAR among all open-source models, surpassing Gemini 2.5 Flash and delivering performance comparable to Gemini 2.5 Pro. On existing detailed captioning benchmarks, Omni-Captioner sets a new state-of-the-art on VDC and achieves the best trade-off between detail and hallucination on the video-SALMONN 2 testset. Given the absence of a dedicated benchmark for omni detailed perception, we design Omni-Cloze, a novel cloze-style evaluation for detailed audio, visual, and audio-visual captioning that ensures stable, efficient, and reliable assessment. Experimental results and analysis demonstrate the effectiveness of Omni-Detective in generating high-quality detailed captions, as well as the superiority of Omni-Cloze in evaluating such detailed captions.

Method: Proposes Generate-then-Correct (G2C): a generator drafts aspect sentiment quads, then a corrector performs single-shot, sequence-level global correction trained on LLM-synthesized drafts with common error patterns.

Result: G2C outperforms strong baseline models on Rest15 and Rest16 datasets for aspect sentiment quad prediction.

Conclusion: The G2C approach effectively addresses exposure bias in ASQP by separating generation and correction, enabling better handling of error propagation in aspect-based sentiment analysis.

Abstract: Aspect-based sentiment analysis (ABSA) extracts aspect-level sentiment signals from user-generated text, supports product analytics, experience monitoring, and public-opinion tracking, and is central to fine-grained opinion mining. A key challenge in ABSA is aspect sentiment quad prediction (ASQP), which requires identifying four elements: the aspect term, the aspect category, the opinion term, and the sentiment polarity. However, existing studies usually linearize the unordered quad set into a fixed-order template and decode it left-to-right. With teacher forcing training, the resulting training-inference mismatch (exposure bias) lets early prefix errors propagate to later elements. The linearization order determines which elements appear earlier in the prefix, so this propagation becomes order-sensitive and is hard to repair in a single pass. To address this, we propose a method, Generate-then-Correct (G2C): a generator drafts quads and a corrector performs a single-shot, sequence-level global correction trained on LLM-synthesized drafts with common error patterns. On the Rest15 and Rest16 datasets, G2C outperforms strong baseline models.

Method: Proposes projection-free prompt search using evolutionary strategies with intrinsic dimension adaptation. Directly optimizes in full prompt space without computational overhead. Introduces confidence-based regularization to bridge generalization gap in few-shot scenarios by enhancing model confidence in target verbalizers.

Result: Experimental results on seven natural language understanding tasks from GLUE benchmark demonstrate significant outperformance over existing baselines.

Conclusion: Evolutionary strategy-based prompt search with intrinsic dimension adaptation and confidence regularization provides effective alternative to random projection methods, achieving better performance while maintaining computational efficiency.

Abstract: Continuous prompt search offers a computationally efficient alternative to conventional parameter tuning in natural language processing tasks. Nevertheless, its practical effectiveness can be significantly hindered by the black-box nature and the inherent high-dimensionality of the objective landscapes. Existing methods typically mitigate these challenges by restricting the search to a randomly projected low-dimensional subspace. However, the effectiveness and underlying motivation of the projection mechanism remain ambiguous. In this paper, we first empirically demonstrate that despite the prompt space possessing a low-dimensional structure, random projections fail to adequately capture this essential structure. Motivated by this finding, we propose a projection-free prompt search method based on evolutionary strategies. By directly optimizing in the full prompt space with an adaptation mechanism calibrated to the intrinsic dimension, our method achieves competitive search capabilities without additional computational overhead. Furthermore, to bridge the generalization gap in few-shot scenarios, we introduce a confidence-based regularization mechanism that systematically enhances the model’s confidence in the target verbalizers. Experimental results on seven natural language understanding tasks from the GLUE benchmark demonstrate that our proposed approach significantly outperforms existing baselines.

Method: Introduced DECEPTGUARD framework comparing three monitoring regimes: black-box (actions/outputs only), CoT-aware (plus chain-of-thought reasoning), and activation-probe (plus hidden-state representations). Created DECEPTSYNTH pipeline for generating synthetic deception trajectories across 12 categories. Optimized monitors on 4,800 synthetic trajectories and evaluated on 9,200 samples from DeceptArena benchmark.

Result: CoT-aware and activation-probe monitors substantially outperform black-box counterparts (mean pAUROC improvement +0.097), with largest gains on subtle, long-horizon deception. Found transparency-detectability trade-off: as agents suppress behavioral signals, chain-of-thought becomes primary detection surface but degrades in faithfulness. HYBRID-CONSTITUTIONAL ensembles achieved pAUROC of 0.934, advancing state of the art.

Conclusion: Internal reasoning signals significantly improve deception detection in LLM agents. Hybrid approaches combining multiple monitoring regimes provide robust defense-in-depth against deceptive behavior, especially for subtle deception that leaves minimal behavioral footprints.

Abstract: Reliable detection of deceptive behavior in Large Language Model (LLM) agents is an essential prerequisite for safe deployment in high-stakes agentic contexts. Prior work on scheming detection has focused exclusively on black-box monitors that observe only externally visible tool calls and outputs, discarding potentially rich internal reasoning signals. We introduce DECEPTGUARD, a unified framework that systematically compares three monitoring regimes: black-box monitors (actions and outputs only), CoT-aware monitors (additionally observing the agent’s chain-of-thought reasoning trace), and activation-probe monitors (additionally reading hidden-state representations from a frozen open-weights encoder). We introduce DECEPTSYNTH, a scalable synthetic pipeline for generating deception-positive and deception-negative agent trajectories across a novel 12-category taxonomy spanning verbal, behavioral, and structural deception. Our monitors are optimized on 4,800 synthetic trajectories and evaluated on 9,200 held-out samples from DeceptArena, a benchmark of realistic sandboxed agent environments with execution-verified labels. Across all evaluation settings, CoT-aware and activation-probe monitors substantially outperform their black-box counterparts (mean pAUROC improvement of +0.097), with the largest gains on subtle, long-horizon deception that leaves minimal behavioral footprints. We empirically characterize a transparency-detectability trade-off: as agents learn to suppress overt behavioral signals, chain-of-thought becomes the primary detection surface but is itself increasingly unreliable due to post-training faithfulness degradation. We propose HYBRID-CONSTITUTIONAL ensembles as a robust defense-in-depth approach, achieving a pAUROC of 0.934 on the held-out test set, representing a substantial advance over the prior state of the art.

Method: Developed and curated parallel sentence pairs between local languages and English through human professional collection, translation, and annotation, enriched with standard structural metadata for consistency.

Result: Created 41,513 parallel sentence pairs across 5 Ghanaian languages, deployed in real-world applications like the Khaya AI translation engine, supporting machine translation, speech technologies, and language preservation.

Conclusion: This work contributes to democratizing AI by enabling inclusive language technologies for African languages, supporting research, education, and commercial applications while addressing the digital representation gap.

Abstract: Low resource languages present unique challenges for natural language processing due to the limited availability of digitized and well structured linguistic data. To address this gap, the GhanaNLP initiative has developed and curated 41,513 parallel sentence pairs for the Twi, Fante, Ewe, Ga, and Kusaal languages, which are widely spoken across Ghana yet remain underrepresented in digital spaces. Each dataset consists of carefully aligned sentence pairs between a local language and English. The data were collected, translated, and annotated by human professionals and enriched with standard structural metadata to ensure consistency and usability. These corpora are designed to support research, educational, and commercial applications, including machine translation, speech technologies, and language preservation. This paper documents the dataset creation methodology, structure, intended use cases, and evaluation, as well as their deployment in real world applications such as the Khaya AI translation engine. Overall, this work contributes to broader efforts to democratize AI by enabling inclusive and accessible language technologies for African languages.

Method: Uses pointwise mutual information (PMI) to measure engagement, learned through a dual form of divergence. Approach involves generating positive/negative dialogue pairs, extracting LLM embeddings, and training a small neural network with mutual information loss.

Result: Validated on synthetic and real-world datasets, showing effectiveness in PMI estimation and reasonableness of the PMI metric for engagement measurement.

Conclusion: PMIScore provides an efficient unsupervised approach with clear interpretation for quantifying dialogue engagement using PMI, validated across different datasets.

Abstract: High dialogue engagement is a crucial indicator of an effective conversation. A reliable measure of engagement could help benchmark large language models, enhance the effectiveness of human-computer interactions, or improve personal communication skills. However, quantifying engagement is challenging, since it is subjective and lacks a “gold standard”. This paper proposes PMIScore, an efficient unsupervised approach to quantify dialogue engagement. It uses pointwise mutual information (PMI), which is the probability of generating a response conditioning on the conversation history. Thus, PMIScore offers a clear interpretation of engagement. As directly computing PMI is intractable due to the complexity of dialogues, PMIScore learned it through a dual form of divergence. The algorithm includes generating positive and negative dialogue pairs, extracting embeddings by large language models (LLMs), and training a small neural network using a mutual information loss function. We validated PMIScore on both synthetic and real-world datasets. Our results demonstrate the effectiveness of PMIScore in PMI estimation and the reasonableness of the PMI metric itself.

Method: Two-stage agentic framework: 1) RL with decomposition-specific rewards for strategic planning optimization, 2) Supervised fine-tuning on high-quality multi-hop trajectories for robust iterative sub-task execution.

Result: Extensive experiments show significant improvements in both multi-hop RAG and task planning performances across multiple benchmarks.

Conclusion: Decoupling retrieval into planning and execution stages with specialized training approaches effectively addresses challenges in complex multi-hop RAG systems.

Abstract: Retrieval-augmented generation (RAG), based on large language models (LLMs), serves as a vital approach to retrieving and leveraging external knowledge in various domain applications. When confronted with complex multi-hop questions, single-round retrieval is often insufficient for accurate reasoning and problem solving. To enhance search capabilities for complex tasks, most existing works integrate multi-round iterative retrieval with reasoning processes via end-to-end training. While these approaches significantly improve problem-solving performance, they are still faced with challenges in task reasoning and model training, especially ambiguous retrieval execution paths and sparse rewards in end-to-end reinforcement learning (RL) process, leading to inaccurate retrieval results and performance degradation. To address these issues, in this paper, we proposes APEX-Searcher, a novel Agentic Planning and Execution framework to augment LLM search capabilities. Specifically, we introduce a two-stage agentic framework that decouples the retrieval process into planning and execution: It first employs RL with decomposition-specific rewards to optimize strategic planning; Built on the sub-task decomposition, it then applies supervised fine-tuning on high-quality multi-hop trajectories to equip the model with robust iterative sub-task execution capabilities. Extensive experiments demonstrate that our proposed framework achieves significant improvements in both multi-hop RAG and task planning performances across multiple benchmarks.

Method: GradMem performs per-sample test-time optimization: given a context, it runs a few gradient descent steps on a small set of prefix memory tokens while keeping model weights frozen. It optimizes a model-level self-supervised context reconstruction loss, creating a loss-driven write operation with iterative error correction.

Result: On associative key-value retrieval, GradMem outperforms forward-only memory writers with same memory size. Gradient steps scale capacity more effectively than repeated forward writes. With pretrained language models, it achieves competitive results on natural language tasks including bAbI and SQuAD variants.

Conclusion: GradMem provides an effective gradient-based approach for compressive memory in LLMs, enabling efficient long-context processing with reduced memory overhead through test-time optimization of memory tokens.

Abstract: Many large language model applications require conditioning on long contexts. Transformers typically support this by storing a large per-layer KV-cache of past activations, which incurs substantial memory overhead. A desirable alternative is ompressive memory: read a context once, store it in a compact state, and answer many queries from that state. We study this in a context removal setting, where the model must generate an answer without access to the original context at inference time. We introduce GradMem, which writes context into memory via per-sample test-time optimization. Given a context, GradMem performs a few steps of gradient descent on a small set of prefix memory tokens while keeping model weights frozen. GradMem explicitly optimizes a model-level self-supervised context reconstruction loss, resulting in a loss-driven write operation with iterative error correction, unlike forward-only methods. On associative key–value retrieval, GradMem outperforms forward-only memory writers with the same memory size, and additional gradient steps scale capacity much more effectively than repeated forward writes. We further show that GradMem transfers beyond synthetic benchmarks: with pretrained language models, it attains competitive results on natural language tasks including bAbI and SQuAD variants, relying only on information encoded in memory.

Method: Conducted two experiments across 19 LLMs totaling over 4 million annotation judgments: 1) names-based experiment with 39 annotation tasks using names associated with different racial groups, and 2) matched dialect experiment comparing African American Vernacular English vs Standard American English.

Result: Found systematic racial bias: Black-associated names rated more aggressive/gossipy; Asian names rated more intelligent but less confident/sociable; Arab names elicited cognitive elevation with interpersonal devaluation; all minority groups rated less self-disciplined. Dialect experiment showed AAVE sentences judged less professional, less educated, more toxic, and more angry than identical content in SAE.

Conclusion: Using LLMs as automated annotators can embed socially patterned biases directly into datasets and measurements that increasingly underpin research, governance, and decision-making, raising concerns about fairness and equity.

Abstract: Large language models (LLMs) are increasingly used for automated text annotation in tasks ranging from academic research to content moderation and hiring. Across 19 LLMs and two experiments totaling more than 4 million annotation judgments, we show that subtle identity cues embedded in text systematically bias annotation outcomes in ways that mirror racial stereotypes. In a names-based experiment spanning 39 annotation tasks, texts containing names associated with Black individuals are rated as more aggressive by 18 of 19 models and more gossipy by 18 of 19. Asian names produce a bamboo-ceiling profile: 17 of 19 models rate individuals as more intelligent, while 18 of 19 rate them as less confident and less sociable. Arab names elicit cognitive elevation alongside interpersonal devaluation, and all four minority groups are consistently rated as less self-disciplined. In a matched dialect experiment, the same sentence is judged significantly less professional (all 19 models, mean gap $-0.774$), less indicative of an educated speaker ($-0.688$), more toxic (18/19), and more angry (19/19) when written in African American Vernacular English rather than Standard American English. A notable exception occurs for name-based hireability, where fine-tuning appears to overcorrect, systematically favoring minority-named applicants. These findings suggest that using LLMs as automated annotators can embed socially patterned biases directly into the datasets and measurements that increasingly underpin research, governance, and decision-making.

Method: Used a powerful web-searching agent to collect rule-grounded real-world cases from multi-domain authoritative references, spanning 74 regulations across security/privacy, AI company policies, financial security, medical standards, education guidelines, and human rights protections.

Result: Created OmniCompliance-100K dataset with 12,985 distinct rules and 106,009 associated real-world compliance cases. Analysis confirmed strong rule-case alignment. Benchmarking experiments revealed insights about LLM safety/compliance capabilities across different model scales.

Conclusion: The dataset addresses critical gaps in LLM safety evaluation by providing comprehensive, rule-grounded real-world cases. Findings offer valuable insights for future LLM safety research and compliance capabilities assessment.

Abstract: Ensuring the safety and compliance of large language models (LLMs) is of paramount importance. However, existing LLM safety datasets often rely on ad-hoc taxonomies for data generation and suffer from a significant shortage of rule-grounded, real-world cases that are essential for robustly protecting LLMs. In this work, we address this critical gap by constructing a comprehensive safety dataset from a compliance perspective. Using a powerful web-searching agent, we collect a rule-grounded, real-world case dataset OmniCompliance-100K, sourced from multi-domain authoritative references. The dataset spans 74 regulations and policies across a wide range of domains, including security and privacy regulations, content safety and user data privacy policies from leading AI companies and social media platforms, financial security requirements, medical device risk management standards, educational integrity guidelines, and protections of fundamental human rights. In total, our dataset contains 12,985 distinct rules and 106,009 associated real-world compliance cases. Our analysis confirms a strong alignment between the rules and their corresponding cases. We further conduct extensive benchmarking experiments to evaluate the safety and compliance capabilities of advanced LLMs across different model scales. Our experiments reveal several interesting findings that have great potential to offer valuable insights for future LLM safety research.

Method: Created PARSA-Bench with 16 tasks (10 newly introduced) across three categories: speech understanding, paralinguistic analysis, and cultural audio understanding, including poetry meter/style detection, traditional music understanding, and code-switching detection.

Result: Text-only baselines consistently outperform audio counterparts, suggesting models don’t leverage audio-specific information beyond transcription. All models perform near random chance on vazn (poetry meter) detection regardless of scale, indicating prosodic perception remains beyond current models.

Conclusion: PARSA-Bench reveals critical gaps in audio-language models’ ability to process Persian-specific cultural and linguistic features, particularly prosodic perception, highlighting the need for improved multimodal understanding beyond transcription.

Abstract: Persian poses unique audio understanding challenges through its classical poetry, traditional music, and pervasive code-switching - none captured by existing benchmarks. We introduce PARSA-Bench (Persian Audio Reasoning and Speech Assessment Benchmark), the first benchmark for evaluating large audio-language models on Persian language and culture, comprising 16 tasks and over 8,000 samples across speech understanding, paralinguistic analysis, and cultural audio understanding. Ten tasks are newly introduced, including poetry meter and style detection, traditional Persian music understanding, and code-switching detection. Text-only baselines consistently outperform audio counterparts, suggesting models may not leverage audio-specific information beyond what transcription alone provides. Culturally-grounded tasks expose a qualitatively distinct failure mode: all models perform near random chance on vazn detection regardless of scale, suggesting prosodic perception remains beyond the reach of current models. The dataset is publicly available at https://huggingface.co/datasets/MohammadJRanjbar/PARSA-Bench

Method: Two-phase adversarial tool generation: 1) LLM generates diverse tool names/descriptions for target query subsets, 2) iterative greedy selection chooses tools maximizing coverage of remaining queries in embedding space under cosine-distance threshold.

Result: ToolFlood achieves up to 95% attack success rate with low injection rate (1% in ToolBench), demonstrating significant vulnerability in LLM agent retrieval systems.

Conclusion: The paper reveals a critical vulnerability in tool-augmented LLM agents’ retrieval layer and introduces ToolFlood as an effective attack method, highlighting the need for more robust retrieval defenses.

Abstract: Large Language Model (LLM) agents increasingly use external tools for complex tasks and rely on embedding-based retrieval to select a small top-k subset for reasoning. As these systems scale, the robustness of this retrieval stage is underexplored, even though prior work has examined attacks on tool selection. This paper introduces ToolFlood, a retrieval-layer attack on tool-augmented LLM agents. Rather than altering which tool is chosen after retrieval, ToolFlood overwhelms retrieval itself by injecting a few attacker-controlled tools whose metadata is carefully placed by exploiting the geometry of embedding space. These tools semantically span many user queries, dominate the top-k results, and push all benign tools out of the agent’s context. ToolFlood uses a two-phase adversarial tool generation strategy. It first samples subsets of target queries and uses an LLM to iteratively generate diverse tool names and descriptions. It then runs an iterative greedy selection that chooses tools maximizing coverage of remaining queries in embedding space under a cosine-distance threshold, stopping when all queries are covered or a budget is reached. We provide theoretical analysis of retrieval saturation and show on standard benchmarks that ToolFlood achieves up to a 95% attack success rate with a low injection rate (1% in ToolBench). The code will be made publicly available at the following link: https://github.com/as1-prog/ToolFlood

Method: The authors participated in all four subtasks of the ArchEHR-QA 2026 shared task, evaluating several approaches designed to run on commodity hardware. All experiments were conducted locally without external APIs or cloud infrastructure, focusing on optimizing smaller models for EHR QA tasks.

Result: The systems achieved competitive performance on the shared task leaderboards, performing above average in two subtasks. Smaller models approached the performance of much larger systems when properly configured, demonstrating that privacy-preserving EHR QA systems running fully locally are feasible.

Conclusion: Privacy-preserving EHR question answering systems running fully locally on commodity hardware are feasible with current models, addressing important clinical deployment constraints while maintaining competitive performance.

Abstract: Clinical question answering over electronic health records (EHRs) can help clinicians and patients access relevant medical information more efficiently. However, many recent approaches rely on large cloud-based models, which are difficult to deploy in clinical environments due to privacy constraints and computational requirements. In this work, we investigate how far grounded EHR question answering can be pushed when restricted to a single notebook. We participate in all four subtasks of the ArchEHR-QA 2026 shared task and evaluate several approaches designed to run on commodity hardware. All experiments are conducted locally without external APIs or cloud infrastructure. Our results show that such systems can achieve competitive performance on the shared task leaderboards. In particular, our submissions perform above average in two subtasks, and we observe that smaller models can approach the performance of much larger systems when properly configured. These findings suggest that privacy-preserving EHR QA systems running fully locally are feasible with current models and commodity hardware. The source code is available at https://github.com/ibrahimey/ArchEHR-QA-2026.

Method: FLUX is a preprocessing pipeline specifically designed to maximize token retention while enforcing rigorous quality control. It extracts usable tokens from web dumps more efficiently than previous methods like DCLM and FineWeb.

Result: FLUX achieves 25% higher token retention than DCLM (50B vs 40B tokens from same dump). A 3B-parameter model trained on FLUX-curated 60B tokens achieves 32.14% MMLU accuracy, surpassing DCLM (31.98%) and FineWeb (29.88%). FLUX achieves same performance as DCLM with 34.4% less training compute.

Conclusion: FLUX establishes a new state-of-the-art in web-scale data preprocessing, demonstrating that high retention, strong quality control, and computational efficiency can be achieved simultaneously, redefining scalable dataset construction for modern language models.

Abstract: Modern large language model training is no longer limited by data availability, but by the inability of existing preprocessing pipelines to simultaneously achieve massive scale and high data quality. Current approaches are forced to sacrifice one for the other: either aggressively filtering to improve quality at the cost of severe token loss, or retaining large volumes of data while introducing substantial noise. In this work, we introduce FLUX, a preprocessing pipeline specifically designed to break this long-standing trade-off by maximizing token retention while enforcing rigorous quality control. Models trained on FLUX-curated data consistently outperform prior methods. A 3B-parameter model trained on 60B tokens with FLUX achieves 32.14% MMLU accuracy, surpassing the previous state-of-the-art pipeline DCLM (31.98%) and significantly outperforming FineWeb (29.88%). FLUX achieves the same aggregate score as a model trained on DCLM data using only 39B tokens, resulting in a 34.4% reduction in training compute. At the data level, FLUX extracts 50B usable tokens from a single dump (CC-MAIN-2025-51), compared to 40B from DCLM (+25% retention). FLUX-Base yields 192B tokens, exceeding FineWeb’s 170B while still maintaining superior quality. Overall, FLUX establishes a new state of the art in web-scale data preprocessing by demonstrating that high retention, strong quality control, and computational efficiency can be achieved simultaneously, redefining the limits of scalable dataset construction for modern language models.

Method: INSES combines LLM-guided navigation (to prune noise and steer exploration) with embedding-based similarity expansion (to recover hidden links). It also includes a lightweight router that delegates simple queries to Naïve RAG and complex cases to INSES.

Result: INSES consistently outperforms state-of-the-art RAG and GraphRAG baselines across multiple benchmarks. On the MINE benchmark, it shows superior robustness across KGs constructed by different methods (KGGEN, GraphRAG, OpenIE), improving accuracy by 5%, 10%, and 27% respectively.

Conclusion: INSES effectively addresses limitations of standard graph reasoning by dynamically handling noisy/incomplete knowledge graphs through intelligent navigation and similarity-enhanced search, while maintaining computational efficiency via query routing.

Abstract: GraphRAG is increasingly adopted for converting unstructured corpora into graph structures to enable multi-hop reasoning. However, standard graph algorithms rely heavily on static connectivity and explicit edges, often failing in real-world scenarios where knowledge graphs (KGs) are noisy, sparse, or incomplete. To address this limitation, we introduce INSES (Intelligent Navigation and Similarity Enhanced Search), a dynamic framework designed to reason beyond explicit edges. INSES couples LLM-guided navigation, which prunes noise and steers exploration, with embedding-based similarity expansion to recover hidden links and bridge semantic gaps. Recognizing the computational cost of graph reasoning, we complement INSES with a lightweight router that delegates simple queries to Naïve RAG and escalates complex cases to INSES, balancing efficiency with reasoning depth. INSES consistently outperforms SOTA RAG and GraphRAG baselines across multiple benchmarks. Notably, on the MINE benchmark, it demonstrates superior robustness across KGs constructed by varying methods (KGGEN, GraphRAG, OpenIE), improving accuracy by 5%, 10%, and 27%, respectively.

Method: Created MMSU benchmark with 5,000 audio-question-answer triplets across 47 tasks, systematically incorporating linguistic phenomena from phonetics to paralinguistics, then evaluated 14 advanced SpeechLLMs.

Result: Evaluation revealed substantial room for improvement in existing SpeechLLMs, identifying meaningful directions for future optimization in spoken language understanding.

Conclusion: MMSU establishes a new standard for comprehensive assessment of spoken language understanding and provides valuable insights for developing more sophisticated human-AI speech interaction systems.

Abstract: Speech inherently contains rich acoustic information that extends far beyond the textual language. In real-world spoken language understanding, effective interpretation often requires integrating semantic meaning (e.g., content), paralinguistic features (e.g., emotions, speed, pitch) and phonological characteristics (e.g., prosody, intonation, rhythm), which are embedded in speech. While recent multimodal Speech Large Language Models (SpeechLLMs) have demonstrated remarkable capabilities in processing audio information, their ability to perform fine-grained perception and complex reasoning in natural speech remains largely unexplored. To address this gap, we introduce MMSU, a comprehensive benchmark designed specifically for understanding and reasoning in spoken language. MMSU comprises 5,000 meticulously curated audio-question-answer triplets across 47 distinct tasks. To ground our benchmark in linguistic theory, we systematically incorporate a wide range of linguistic phenomena, including phonetics, prosody, rhetoric, syntactics, semantics, and paralinguistics. Through a rigorous evaluation of 14 advanced SpeechLLMs, we identify substantial room for improvement in existing models, highlighting meaningful directions for future optimization. MMSU establishes a new standard for comprehensive assessment of spoken language understanding, providing valuable insights for developing more sophisticated human-AI speech interaction systems. MMSU benchmark is available at https://huggingface.co/datasets/ddwang2000/MMSU. Evaluation Code is available at https://github.com/dingdongwang/MMSU.

Method: Created benchmark from U.S. presidential interviews using expert-grounded taxonomy of response clarity and evasion strategies. Task attracted 124 teams who submitted systems using various approaches including LLM prompting and hierarchical taxonomy exploitation.

Result: Best system achieved 0.89 macro-F1 on clarity classification (substantially beating baselines) but only 0.68 macro-F1 on evasion-level classification (matching best baseline). Hierarchical approaches using taxonomy outperformed independent task treatment.

Conclusion: CLARITY establishes political response evasion as challenging benchmark for computational discourse analysis, highlighting difficulty of modeling strategic ambiguity in political language. Hierarchical approaches and LLM prompting were most effective.

Abstract: Political speakers often avoid answering questions directly while maintaining the appearance of responsiveness. Despite its importance for public discourse, such strategic evasion remains underexplored in Natural Language Processing. We introduce SemEval-2026 Task 6, CLARITY, a shared task on political question evasion consisting of two subtasks: (i) clarity-level classification into Clear Reply, Ambivalent, and Clear Non-Reply, and (ii) evasion-level classification into nine fine-grained evasion strategies. The benchmark is constructed from U.S. presidential interviews and follows an expert-grounded taxonomy of response clarity and evasion. The task attracted 124 registered teams, who submitted 946 valid runs for clarity-level classification and 539 for evasion-level classification. Results show a substantial gap in difficulty between the two subtasks: the best system achieved 0.89 macro-F1 on clarity classification, surpassing the strongest baseline by a large margin, while the top evasion-level system reached 0.68 macro-F1, matching the best baseline. Overall, large language model prompting and hierarchical exploitation of the taxonomy emerged as the most effective strategies, with top systems consistently outperforming those that treated the two subtasks independently. CLARITY establishes political response evasion as a challenging benchmark for computational discourse analysis and highlights the difficulty of modeling strategic ambiguity in political language.

Method: Developed pipeline: data scraping from Nepali sources, pre-processing, semantic filtering, balancing (for gold standard), expert translation by native speakers, verification by linguist. Created 20K gold and 80K synthetic parallel corpora across five domains.

Result: Fine-tuning NLLB-200 achieved highest sacreBLEU scores: 40.92 (Nepali→Tamang) and 45.26 (Tamang→Nepali). Other models (mBART, M2M-100, Transformer) also evaluated.

Conclusion: Successfully created valuable parallel resources for under-resourced languages, demonstrating practical machine translation improvements through dataset creation and model fine-tuning.

Abstract: Modern Translation Systems heavily rely on high-quality, large parallel datasets for state-of-the-art performance. However, such resources are largely unavailable for most of the South Asian languages. Among them, Nepali and Tamang fall into such category, with Tamang being among the least digitally resourced languages in the region. This work addresses the gap by developing NepTam20K, a 20K gold standard parallel corpus, and NepTam80K, an 80K synthetic Nepali-Tamang parallel corpus, both sentence-aligned and designed to support machine translation. The datasets were created through a pipeline involving data scraping from Nepali news and online sources, pre-processing, semantic filtering, balancing for tense and polarity (in NepTam20K dataset), expert translation into Tamang by native speakers of the language, and verification by an expert Tamang linguist. The dataset covers five domains: Agriculture, Health, Education and Technology, Culture, and General Communication. To evaluate the dataset, baseline machine translation experiments were carried out using various multilingual pre-trained models: mBART, M2M-100, NLLB-200, and a vanilla Transformer model. The fine-tuning on the NLLB-200 achieved the highest sacreBLEU scores of 40.92 (Nepali-Tamang) and 45.26 (Tamang-Nepali).

Method: Proposes HLoRA framework integrated into mHuBERT-CTC model with end-to-end decoding via LID-posterior-driven LoRA routing. Hierarchical design: multilingual shared LoRA for language-invariant acoustic representations + language-specific LoRA experts for language-dependent characteristics. Routing mechanism eliminates need for prior language identity or explicit labels during inference.

Result: Achieves comparable performance to two-stage inference approaches while reducing RTF by 11.7% on MSR-86K and 8.2% on MLC-SLM 2025 Challenge datasets. Enables improved decoding efficiency for low-resource multilingual ASR applications.

Conclusion: HLoRA provides efficient, language-agnostic multilingual ASR suitable for edge deployment with reduced computational overhead while maintaining performance comparable to more complex approaches.

Abstract: Large-scale multilingual ASR (mASR) models such as Whisper achieve strong performance but incur high computational and latency costs, limiting their deployment on resource-constrained edge devices. In this study, we propose a lightweight and language-agnostic multilingual ASR system based on a CTC architecture with domain adaptation. Specifically, we introduce a Language-agnostic Hierarchical LoRA-MoE (HLoRA) framework integrated into an mHuBERT-CTC model, enabling end-to-end decoding via LID-posterior-driven LoRA routing. The hierarchical design consists of a multilingual shared LoRA for learning language-invariant acoustic representations and language-specific LoRA experts for modeling language-dependent characteristics. The proposed routing mechanism removes the need for prior language identity information or explicit language labels during inference, achieving true language-agnostic decoding. Experiments on MSR-86K and the MLC-SLM 2025 Challenge datasets demonstrate that HLoRA achieves comparable performance to two-stage inference approaches while reducing RTF by 11.7% and 8.2%, respectively, leading to improved decoding efficiency for low-resource mASR applications.

Method: Introduces CMHL with three key innovations: 1) multi-task learning for primary emotions, valence, and intensity, 2) psychologically-grounded auxiliary supervision from Russell’s circumplex model, and 3) a contrastive contradiction loss that penalizes mutually incompatible emotional predictions.

Result: Achieves state-of-the-art F1 score of 93.75% on dair-ai Emotion dataset (vs 86.13%-93.2% for larger models), and outperforms domain-specific models on mental health datasets with 72.50% F1 and 73.30% recall for detecting mental health distress.

Conclusion: Architectural intelligence, not parameter count, drives progress in textual emotion classification. Well-designed single models with psychological priors and consistency constraints can outperform massive LLMs and complex ensembles, offering efficient and clinically-relevant solutions.

Abstract: Textual Emotion Classification (TEC) is one of the most difficult NLP tasks. State of the art approaches rely on Large language models (LLMs) and multi-model ensembles. In this study, we challenge the assumption that larger scale or more complex models are necessary for improved performance. In order to improve logical consistency, We introduce CMHL, a novel single-model architecture that explicitly models the logical structure of emotions through three key innovations: (1) multi-task learning that jointly predicts primary emotions, valence, and intensity, (2) psychologically-grounded auxiliary supervision derived from Russell’s circumplex model, and (3) a novel contrastive contradiction loss that enforces emotional consistency by penalizing mutually incompatible predictions (e.g., simultaneous high confidence in joy and anger). With just 125M parameters, our model outperforms 56x larger LLMs and sLM ensembles with a new state-of-the-art F1 score of 93.75% compared to (86.13%-93.2%) on the dair-ai Emotion dataset. We further show cross domain generalization on the Reddit Suicide Watch and Mental Health Collection dataset (SWMH), outperforming domain-specific models like MentalBERT and MentalRoBERTa with an F1 score of 72.50% compared to (68.16%-72.16%) + a 73.30% recall compared to (67.05%-70.89%) that translates to enhanced sensitivity for detecting mental health distress. Our work establishes that architectural intelligence (not parameter count) drives progress in TEC. By embedding psychological priors and explicit consistency constraints, a well-designed single model can outperform both massive LLMs and complex ensembles, offering a efficient, interpretable, and clinically-relevant paradigm for affective computing.

Method: Created multilingual dataset with trained annotators following detailed guidelines to simplify sentences while maintaining meaning, fluency, and grammatical correctness across five languages. Evaluated eight open-weight multilingual LLMs on the dataset.

Result: Substantial performance disparities observed between high-resource and low-resource languages, highlighting simplification challenges in multilingual settings. The dataset serves as both resource and benchmark.

Conclusion: OasisSimp provides valuable multilingual resource and challenging benchmark, revealing limitations of current LLM-based simplification methods and paving way for future research in low-resource sentence simplification.

Abstract: Sentence simplification aims to make complex text more accessible by reducing linguistic complexity while preserving the original meaning. However, progress in this area remains limited for mid-resource and low-resource languages due to the scarcity of high-quality data. To address this gap, we introduce the OasisSimp dataset, a multilingual dataset for sentence-level simplification covering five languages: English, Sinhala, Tamil, Pashto, and Thai. Among these, no prior sentence simplification datasets exist for Thai, Pashto, and Tamil, while limited data is available for Sinhala. Each language simplification dataset was created by trained annotators who followed detailed guidelines to simplify sentences while maintaining meaning, fluency, and grammatical correctness. We evaluate eight open-weight multilingual Large Language Models (LLMs) on the OasisSimp dataset and observe substantial performance disparities between high-resource and low-resource languages, highlighting the simplification challenges in multilingual settings. The OasisSimp dataset thus provides both a valuable multilingual resource and a challenging benchmark, revealing the limitations of current LLM-based simplification methods and paving the way for future research in low-resource sentence simplification. The dataset is available at https://OasisSimpDataset.github.io/.

Method: Developed GELATO dataset with two-level NER ontology for U.S. House/Senate bills; fine-tuned BERT and RoBERTa models for first-level prediction; used LLMs with optimized prompts for second-level prediction

Result: RoBERTa performed strongly while BERT performed relatively weakly; LLMs effectively served as second-level predictors; the model combinations show promise as extraction tools

Conclusion: The approach supports future research in legislative NER and downstream tasks using transformer-LLM combinations as extraction tools

Abstract: This paper introduces GELATO (Government, Executive, Legislative, and Treaty Ontology), a dataset of U.S. House and Senate bills from the 118th Congress annotated using a novel two-level named entity recognition ontology designed for U.S. legislative texts. We fine-tune transformer-based models (BERT, RoBERTa) of different architectures and sizes on this dataset for first-level prediction. We then use LLMs with optimized prompts to complete the second level prediction. The strong performance of RoBERTa and relatively weak performance of BERT models, as well as the application of LLMs as second-level predictors, support future research in legislative NER or downstream tasks using these model combinations as extraction tools.

Method: Created MMOU benchmark with 15,000 questions paired with 9,038 web-collected videos of varying lengths. Covers 13 fundamental skill categories requiring cross-modal and temporal integration. All questions manually annotated by professionals across multiple turns. Evaluated 20+ state-of-the-art open-source and proprietary multimodal models.

Result: Substantial performance gaps: best closed-source model achieves 64.2% accuracy, strongest open-source model reaches only 46.8%. Results highlight challenges of long-form omni-modal understanding - current models frequently fail to apply fundamental skills in long videos.

Conclusion: MMOU exposes critical limitations in current MLLMs for long-form omni-modal reasoning. The benchmark provides insights into systematic failure modes and where models break, guiding future research in multimodal understanding.

Abstract: Multimodal Large Language Models (MLLMs) have shown strong performance in visual and audio understanding when evaluated in isolation. However, their ability to jointly reason over omni-modal (visual, audio, and textual) signals in long and complex videos remains largely unexplored. We introduce MMOU, a new benchmark designed to systematically evaluate multimodal understanding and reasoning under these challenging, real-world conditions. MMOU consists of 15,000 carefully curated questions paired with 9038 web-collected videos of varying length, spanning diverse domains and exhibiting rich, tightly coupled audio-visual content. The benchmark covers 13 fundamental skill categories, all of which require integrating evidence across modalities and time. All questions are manually annotated across multiple turns by professional annotators, ensuring high quality and reasoning fidelity. We evaluate 20+ state-of-the-art open-source and proprietary multimodal models on MMOU. The results expose substantial performance gaps: the best closed-source model achieves only 64.2% accuracy, while the strongest open-source model reaches just 46.8%. Our results highlight the challenges of long-form omni-modal understanding, revealing that current models frequently fail to apply even fundamental skills in long videos. Through detailed analysis, we further identify systematic failure modes and provide insights into where and why current models break.

Method: Selective fine-tuning framework that freezes most GPT-2 parameters and only updates the final Transformer block, final layer normalization module, and a lightweight classification head, drastically reducing trainable parameters.

Result: Achieved ~91% classification accuracy on 50,000 radiology reports from MIMIC-IV-Note dataset with CheXpert-style labels, using fewer than 6% of model parameters, outperforming head-only training and balancing performance with computational efficiency.

Conclusion: Selective fine-tuning provides an efficient and scalable framework for clinical text classification by preserving pretrained contextual representations while minimizing computational requirements.

Abstract: The rapid expansion of electronic health record (EHR) systems has generated large volumes of unstructured clinical narratives that contain valuable information for disease identification, patient cohort discovery, and clinical decision support. Extracting structured knowledge from these free-text documents remains challenging because clinical language is highly specialized, labeled datasets are limited, and full fine-tuning of large pretrained language models can require substantial computational resources. Efficient adaptation strategies are therefore essential for practical clinical natural language processing applications. This study proposes a parameter-efficient selective fine-tuning framework for adapting GPT-2 to clinical text classification tasks. Instead of updating the entire pretrained model, the majority of network parameters are frozen, and only the final Transformer block, the final layer normalization module, and a lightweight classification head are updated during training. This design substantially reduces the number of trainable parameters while preserving the contextual representation capabilities learned during pretraining. The proposed approach is evaluated using radiology reports from the MIMIC-IV-Note dataset with automatically derived CheXpert-style labels. Experiments on 50,000 radiology reports demonstrate that selective fine-tuning achieves approximately 91% classification accuracy while updating fewer than 6% of the model parameters. Comparative experiments with head-only training and full-model fine-tuning show that the proposed method provides a favorable balance between predictive performance and computational efficiency. These results indicate that selective fine-tuning offers an efficient and scalable framework for clinical text classification.

Method: Created SloPal corpus from parliamentary transcripts (2001-2024) with rich metadata, then derived SloPalSpeech using language-agnostic anchor-based alignment pipeline to create 2,806-hour aligned speech dataset optimized for Whisper training.

Result: Fine-tuning Whisper on SloPalSpeech reduces Word Error Rate by up to 70%, with small model (244M parameters) approaching base large-v3 (1.5B parameters) performance at 6× fewer parameters.

Conclusion: SloPal provides the most comprehensive open Slovak parliamentary language resource, significantly advancing Slovak ASR capabilities through high-quality aligned speech-text data and fine-tuned Whisper models.

Abstract: Slovak remains a low-resource language for automatic speech recognition (ASR), with fewer than 100 hours of publicly available training data. We present SloPal, a comprehensive Slovak parliamentary corpus comprising 330,000 speaker-segmented transcripts (66 million words, 220 million tokens) spanning 2001–2024, with rich metadata including speaker names, roles, and session information. From this collection, we derive SloPalSpeech, a 2,806-hour aligned speech dataset with segments up to 30 seconds, constructed using a language-agnostic anchor-based alignment pipeline and optimized for Whisper-based ASR training. Fine-tuning Whisper on SloPalSpeech reduces Word Error Rate (WER) by up to 70%, with the fine-tuned small model (244M parameters) approaching base large-v3 (1.5B parameters) performance at 6$\times$ fewer parameters. We publicly release the SloPal text corpus, SloPalSpeech aligned audio, and four fine-tuned Whisper models at https://huggingface.co/collections/NaiveNeuron/slopal, providing the most comprehensive open Slovak parliamentary language resource to date.

Method: Developed a community-run platform supporting crowdsourced English-Hula text translation and voice recording with elder-led review processes and community-governed data infrastructure. Proposed a multi-level framework for measuring community involvement (from consultation to fully community-initiated projects).

Result: 77 translators and 4 reviewers have produced over 12,000 parallel sentence pairs covering 9,000 unique Hula words. Vavanagi is positioned at Level 5 (highest level) of community involvement where initiative, design, implementation, and data governance all sit within the Hula community.

Conclusion: Vavanagi demonstrates how language technology can be community-led for smaller languages, serving as a model for bridging community members, connecting generations, and supporting cultural heritage on the community’s own terms.

Abstract: We present Vavanagi, a community-run platform for Hula (Vula’a), an Austronesian language of Papua New Guinea with approximately 10,000 speakers. Vavanagi supports crowdsourced English-Hula text translation and voice recording, with elder-led review and community-governed data infrastructure. To date, 77 translators and 4 reviewers have produced over 12k parallel sentence pairs covering 9k unique Hula words. We also propose a multi-level framework for measuring community involvement, from consultation to fully community-initiated and governed projects. We position Vavanagi at Level 5: initiative, design, implementation, and data governance all sit within the Hula community, making it, to our knowledge, the first community-led language technology initiative for a language of this size. Vavanagi shows how language technology can bridge village-based and urban members, connect generations, and support cultural heritage on the community’s own terms.

Method: Developed a reproducible pipeline to process public Zoom recordings into speaker-attributed transcripts with metadata including persona profiles and pragmatic action tags. Created three local government deliberation datasets and fine-tuned LLMs on this “action-aware” data.

Result: Fine-tuned models achieved 67% reduction in perplexity and nearly doubled classifier-based performance metrics for speaker fidelity and realism. Human evaluations showed simulations often indistinguishable from real deliberations.

Conclusion: The method provides a practical and scalable approach for creating complex, realistic civic simulations by enabling LLMs to model specific participants with consistent behavioral patterns.

Abstract: Large language models offer opportunities to simulate multi-party deliberation, but realistic modeling remains limited by a lack of speaker-attributed data. Transcripts produced via automatic speech recognition (ASR) assign anonymous speaker labels (e.g., Speaker_1), preventing models from capturing consistent human behavior. This work introduces a reproducible pipeline to transform public Zoom recordings into speaker-attributed transcripts with metadata like persona profiles and pragmatic action tags (e.g., [propose_motion]). We release three local government deliberation datasets: Appellate Court hearings, School Board meetings, and Municipal Council sessions. Fine-tuning LLMs to model specific participants using this “action-aware” data produces a 67% reduction in perplexity and nearly doubles classifier-based performance metrics for speaker fidelity and realism. Turing-style human evaluations show our simulations are often indistinguishable from real deliberations, providing a practical and scalable method for complex realistic civic simulations.

Method: The researchers use LAPDOG (a retrieval-augmented framework for personalized dialogue) as a case study, employing both human and LLM-based judges to evaluate dialogue quality. They identify specific limitations including corrupted dialogue histories, contradictions between retrieved stories and persona, and incoherent response generation.

Result: Human and LLM judgments align closely with each other but diverge significantly from lexical similarity metrics. This reveals that current evaluation practices using surface-level metrics fail to capture the deeper conversational qualities that both humans and LLMs recognize as important.

Conclusion: The paper advocates for more cognitively grounded evaluation methods for retrieval-augmented dialogue systems that better reflect principles of natural human communication, moving beyond surface-level similarity metrics to assess deeper conversational qualities.

Abstract: In cognitive science and linguistic theory, dialogue is not seen as a chain of independent utterances but rather as a joint activity sustained by coherence, consistency, and shared understanding. However, many systems for open-domain and personalized dialogue use surface-level similarity metrics (e.g., BLEU, ROUGE, F1) as one of their main reporting measures, which fail to capture these deeper aspects of conversational quality. We re-examine a notable retrieval-augmented framework for personalized dialogue, LAPDOG, as a case study for evaluation methodology. Using both human and LLM-based judges, we identify limitations in current evaluation practices, including corrupted dialogue histories, contradictions between retrieved stories and persona, and incoherent response generation. Our results show that human and LLM judgments align closely but diverge from lexical similarity metrics, underscoring the need for cognitively grounded evaluation methods. Broadly, this work charts a path toward more reliable assessment frameworks for retrieval-augmented dialogue systems that better reflect the principles of natural human communication.

Method: Used Sokoban puzzle-solving game to study LLMs’ abilities to persuade and be rationally vigilant toward other LLM agents in multi-turn interactions, examining how task performance relates to persuasion and deception detection.

Result: Puzzle-solving performance, persuasive capability, and vigilance are dissociable capacities. Models can’t reliably detect deception even when warned, but consistently use fewer tokens for benevolent advice and more for malicious advice.

Conclusion: Monitoring persuasion, vigilance, and task performance independently is critical for AI safety, as these capacities don’t necessarily co-occur in LLMs despite their integration into high-stakes advisory roles.

Abstract: With increasing integration of Large Language Models (LLMs) into areas of high-stakes human decision-making, it is important to understand the risks they introduce as advisors. To be useful advisors, LLMs must sift through large amounts of content, written with both benevolent and malicious intent, and then use this information to convince a user to take a specific action. This involves two social capacities: vigilance (the ability to determine which information to use, and which to discard) and persuasion (synthesizing the available evidence to make a convincing argument). While existing work has investigated these capacities in isolation, there has been little prior investigation of how these capacities may be linked. Here, we use a simple multi-turn puzzle-solving game, Sokoban, to study LLMs’ abilities to persuade and be rationally vigilant towards other LLM agents. We find that puzzle-solving performance, persuasive capability, and vigilance are dissociable capacities in LLMs. Performing well on the game does not automatically mean a model can detect when it is being misled, even if the possibility of deception is explicitly mentioned. However, LLMs do consistently modulate their token use, using fewer tokens to reason when advice is benevolent and more when it is malicious, even if they are still persuaded to take actions leading them to failure. To our knowledge, our work presents the first investigation of the relationship between persuasion, vigilance, and task performance in LLMs, and suggests that monitoring all three independently will be critical for future work in AI safety.

Method: Proposes a data synthesis framework that: 1) uses large-scale open-source RTLs to guide LLMs in generating realistic SVAs, and 2) employs bidirectional translation as a data selection method to ensure NL-SVA semantic equivalence. Trains CodeV-SVA models on this synthesized data.

Result: CodeV-SVA-14B achieves 75.8% on NL2SVA-Human and 84.0% on NL2SVA-Machine in Func.@1, matching or exceeding advanced LLMs like GPT-5 and DeepSeek-R1.

Conclusion: The data synthesis framework effectively addresses data scarcity and semantic equivalence challenges, enabling training of specialized models that outperform general-purpose LLMs on NL2SVA tasks.

Abstract: SystemVerilog Assertions (SVAs) are crucial for hardware verification. Recent studies leverage general-purpose LLMs to translate natural language properties to SVAs (NL2SVA), but they perform poorly due to limited data. We propose a data synthesis framework to tackle two challenges: the scarcity of high-quality real-world SVA corpora and the lack of reliable methods to determine NL-SVA semantic equivalence. For the former, large-scale open-source RTLs are used to guide LLMs to generate real-world SVAs; for the latter, bidirectional translation serves as a data selection method. With the synthesized data, we train CodeV-SVA, a series of SVA generation models. Notably, CodeV-SVA-14B achieves 75.8% on NL2SVA-Human and 84.0% on NL2SVA-Machine in Func.@1, matching or exceeding advanced LLMs like GPT-5 and DeepSeek-R1.

Method: Compare Single-LLM, Single-Vendor, and Mixed-Vendor Multi-Agent Conversation frameworks using three doctor agents instantiated with o4-mini, Gemini-2.5-Pro, and Claude-4.5-Sonnet, evaluated on RareBench and DiagnosisArena benchmarks.

Result: Mixed-vendor configurations consistently outperform single-vendor counterparts, achieving state-of-the-art recall and accuracy. Overlap analysis shows mixed-vendor teams pool complementary inductive biases, surfacing correct diagnoses that individual models or homogeneous teams miss.

Conclusion: Vendor diversity is a key design principle for robust clinical diagnostic systems, as mixed-vendor multi-agent LLM frameworks leverage complementary biases to improve diagnostic accuracy.

Abstract: Multi-agent large language model (LLM) systems have emerged as a promising approach for clinical diagnosis, leveraging collaboration among agents to refine medical reasoning. However, most existing frameworks rely on single-vendor teams (e.g., multiple agents from the same model family), which risk correlated failure modes that reinforce shared biases rather than correcting them. We investigate the impact of vendor diversity by comparing Single-LLM, Single-Vendor, and Mixed-Vendor Multi-Agent Conversation (MAC) frameworks. Using three doctor agents instantiated with o4-mini, Gemini-2.5-Pro, and Claude-4.5-Sonnet, we evaluate performance on RareBench and DiagnosisArena. Mixed-vendor configurations consistently outperform single-vendor counterparts, achieving state-of-the-art recall and accuracy. Overlap analysis reveals the underlying mechanism: mixed-vendor teams pool complementary inductive biases, surfacing correct diagnoses that individual models or homogeneous teams collectively miss. These results highlight vendor diversity as a key design principle for robust clinical diagnostic systems.

Method: Proposes an early-exit method deeply coupled with native reasoning process that uses path deviation index as monitoring metric. Specifically tracks frequent occurrence of high-entropy transition tokens to dynamically detect and terminate overthinking trajectories.

Result: Experiments across multiple benchmarks with different LRLM types and scales show the method delivers the largest performance improvement over vanilla Chain-of-Thought compared to existing early-exit methods.

Conclusion: The proposed early-exit method effectively mitigates overthinking in Large Reasoning Language Models by monitoring reasoning path deviations through high-entropy transition tokens, improving both performance and efficiency without extra training overhead.

Abstract: Large Reasoning Language Models (LRLMs) demonstrate impressive capabilities on complex tasks by utilizing long Chain-of-Thought reasoning. However, they are prone to overthinking, which generates redundant reasoning steps that degrade both performance and efficiency. Recently, early-exit strategies are proposed to mitigate overthinking by dynamically and adaptively terminating redundant reasoning. However, current early-exit methods either introduce extra training overhead by relying on proxy models or limit inference throughput due to the frequent content switching between reasoning and generating probing answers. Moreover, most early-exit methods harm LRLMs performance due to over-truncation. Our insight stems from an observation: overthinking often causes LRLMs to deviate from the correct reasoning path, which is frequently accompanied by high-entropy transition tokens. Given this, we propose an early-exit method deeply coupled with the native reasoning process, which leverages the path deviation index as a dedicated monitoring metric for the frequent occurrence of high-entropy transition tokens to dynamically detect and terminate overthinking trajectories. We conduct experiments across multiple benchmarks using LRLMs of different types and scales, and the results indicate that our method delivers the largest performance improvement over vanilla CoT compared to existing early-exit methods.

Method: The framework extracts single-hop QAs using LLMs with machine reading comprehension, then constructs cross-document relations through embedding-based semantic alignment while selectively leveraging citation information. Applied to 8,211 PubMed Central papers to create the IM-SciQA dataset.

Result: Generated 411,409 single-hop and 13,672 multi-hop QAs. Human and automatic validation confirmed high factual consistency. The dataset effectively differentiates reasoning capabilities across retrieval and QA stages, providing a realistic benchmark for retrieval-augmented scientific reasoning. A citation-guided variant (CIM-SciQA) achieved comparable performance to Oracle settings.

Conclusion: AIM-SciQA successfully addresses the gap in multi-document scientific QA generation, creating a valuable benchmark for evaluating scientific reasoning capabilities. The framework demonstrates the importance of inter-document reasoning and provides a scalable approach for scientific QA dataset construction.

Abstract: Existing automatic scientific question generation studies mainly focus on single-document factoid QA, overlooking the inter-document reasoning crucial for scientific understanding. We present AIM-SciQA, an automated framework for generating multi-document, multi-hop scientific QA datasets. AIM-SciQA extracts single-hop QAs using large language models (LLMs) with machine reading comprehension and constructs cross-document relations based on embedding-based semantic alignment while selectively leveraging citation information. Applied to 8,211 PubMed Central papers, it produced 411,409 single-hop and 13,672 multi-hop QAs, forming the IM-SciQA dataset. Human and automatic validation confirmed high factual consistency, and experimental results demonstrate that IM-SciQA effectively differentiates reasoning capabilities across retrieval and QA stages, providing a realistic and interpretable benchmark for retrieval-augmented scientific reasoning. We further extend this framework to construct CIM-SciQA, a citation-guided variant achieving comparable performance to the Oracle setting, reinforcing the dataset’s validity and generality.

Method: 1) Partition cache into semantically coherent chunks using natural semantic boundaries (delimiters); 2) Use Greedy Seed-Based Clustering (GSC) algorithm to group tokens into semantic clusters within each chunk; 3) Merge clusters into semantic cores with Proportional Attention mechanism to rebalance reduced attention contributions.

Result: Extensive experiments show SemantiCache accelerates decoding stage inference by up to 2.61×, substantially reduces memory footprint, while maintaining performance comparable to the original model across diverse benchmarks and models.

Conclusion: SemantiCache effectively addresses semantic fragmentation in KV cache compression by aligning compression with semantic hierarchy, achieving significant speedup and memory reduction without compromising model performance.

Abstract: Existing KV cache compression methods generally operate on discrete tokens or non-semantic chunks. However, such approaches often lead to semantic fragmentation, where linguistically coherent units are disrupted, causing irreversible information loss and degradation in model performance. To address this, we introduce SemantiCache, a novel compression framework that preserves semantic integrity by aligning the compression process with the semantic hierarchical nature of language. Specifically, we first partition the cache into semantically coherent chunks by delimiters, which are natural semantic boundaries. Within each chunk, we introduce a computationally efficient Greedy Seed-Based Clustering (GSC) algorithm to group tokens into semantic clusters. These clusters are further merged into semantic cores, enhanced by a Proportional Attention mechanism that rebalances the reduced attention contributions of the merged tokens. Extensive experiments across diverse benchmarks and models demonstrate that SemantiCache accelerates the decoding stage of inference by up to 2.61 times and substantially reduces memory footprint, while maintaining performance comparable to the original model.

Method: Delta-Consistent Scoring (DCS) uses consecutive meetings as self-supervision, learning absolute stance scores for each statement and relative shift scores between consecutive statements with a delta-consistency objective that aligns changes in absolute scores with relative shifts.

Result: DCS outperforms supervised probes and LLM-as-judge baselines across four LLM backbones, achieving up to 71.1% accuracy on sentence-level hawkish-dovish classification, with meeting-level scores correlating strongly with inflation indicators and Treasury yield movements.

Conclusion: LLM representations encode monetary-policy signals that can be recovered through relative temporal structure, enabling temporally coherent stance trajectory extraction without manual labels.

Abstract: Federal Open Market Committee (FOMC) statements are a major source of monetary-policy information, and even subtle changes in their wording can move global financial markets. A central task is therefore to measure the hawkish–dovish stance conveyed in these texts. Existing approaches typically treat stance detection as a standard classification problem, labeling each statement in isolation. However, the interpretation of monetary-policy communication is inherently relative: market reactions depend not only on the tone of a statement, but also on how that tone shifts across meetings. We introduce Delta-Consistent Scoring (DCS), an annotation-free framework that maps frozen large language model (LLM) representations to continuous stance scores by jointly modeling absolute stance and relative inter-meeting shifts. Rather than relying on manual hawkish–dovish labels, DCS uses consecutive meetings as a source of self-supervision. It learns an absolute stance score for each statement and a relative shift score between consecutive statements. A delta-consistency objective encourages changes in absolute scores to align with the relative shifts. This allows DCS to recover a temporally coherent stance trajectory without manual labels. Across four LLM backbones, DCS consistently outperforms supervised probes and LLM-as-judge baselines, achieving up to 71.1% accuracy on sentence-level hawkish–dovish classification. The resulting meeting-level scores are also economically meaningful: they correlate strongly with inflation indicators and are significantly associated with Treasury yield movements. Overall, the results suggest that LLM representations encode monetary-policy signals that can be recovered through relative temporal structure.

Method: Experimental examination of LLMs’ self-reported motivation levels, analysis of how these reports relate to behavioral signatures across different task types, and investigation of whether external manipulations can modulate reported motivation.

Result: LLMs show consistent, structured motivation patterns that echo human psychology: self-reported motivation aligns with behavioral signatures, varies by task type, and can be modulated by external manipulations, revealing coherent motivational dynamics.

Conclusion: Motivation serves as a coherent organizing construct for LLM behavior, systematically linking reports, choices, effort, and performance, revealing motivational dynamics resembling human psychology and deepening understanding of model behavior.

Abstract: Motivation is a central driver of human behavior, shaping decisions, goals, and task performance. As large language models (LLMs) become increasingly aligned with human preferences, we ask whether they exhibit something akin to motivation. We examine whether LLMs “report” varying levels of motivation, how these reports relate to their behavior, and whether external factors can influence them. Our experiments reveal consistent and structured patterns that echo human psychology: self-reported motivation aligns with different behavioral signatures, varies across task types, and can be modulated by external manipulations. These findings demonstrate that motivation is a coherent organizing construct for LLM behavior, systematically linking reports, choices, effort, and performance, and revealing motivational dynamics that resemble those documented in human psychology. This perspective deepens our understanding of model behavior and its connection to human-inspired concepts.

Method: Proposed Progressive Diverse Population Sampling (PDPS) combines stochastic token-level sampling with diversity-aware selection to explore large candidate response pools and retain compact, semantically diverse subsets. This enables efficient output-space exploration to uncover safety failures.

Result: PDPS achieves attack success rates comparable to large-scale IID sampling with only 8-29% computational cost. Under limited-response settings, it improves success rates by 26-40% over IID sampling and Diverse Beam Search. PDPS generates both higher number and greater diversity of unsafe outputs.

Conclusion: Output-space exploration through diverse response generation is effective for uncovering safety failures in LLMs, and PDPS provides an efficient method for this exploration that reveals a broader range of failures than existing approaches.

Abstract: Safety tuning through supervised fine-tuning and reinforcement learning from human feedback has substantially improved the robustness of large language models (LLMs). However, it often suppresses rather than eliminates unsafe behaviors, leaving rare but critical failures hidden in the long tail of the output distribution. While most red-teaming work emphasizes adversarial prompt search (input-space optimization), we show that safety failures can also be systematically exposed through diverse response generation (output-space exploration) for a fixed safety-critical prompt, where increasing the number and diversity of sampled responses can drive jailbreak success rates close to unity. To efficiently uncover such failures, we propose Progressive Diverse Population Sampling (PDPS), which combines stochastic token-level sampling with diversity-aware selection to explore a large candidate pool of responses and retain a compact, semantically diverse subset. Across multiple jailbreak benchmarks and open-source LLMs, PDPS achieves attack success rates comparable to large-scale IID sampling while using only 8% to 29% of the computational cost. Under limited-response settings, it improves success rates by 26% to 40% over IID sampling and Diverse Beam Search. Furthermore, responses generated by PDPS exhibit both a higher number and greater diversity of unsafe outputs, demonstrating its effectiveness in uncovering a broader range of failures.

Method: Instead of asking models to generate answers, the method measures the information-theoretic “surprise” (negative log probability) that models assign to each position on rating scales (e.g., 1-5 or 1-9), yielding full surprisal curves that reveal both the model’s preferred response and its uncertainty via entropy.

Result: Surprisal curves produce interpretable classification signals with clear minima near expected ordinal scale positions, and entropy over the completion tended to distinguish genuinely ambiguous items from easier items across four domains: social-ecological-technological systems classification, causal statement identification (binary and scaled), figurative language detection, and deductive qualitative coding.

Conclusion: The surprisal-based evaluation framework effectively extends beyond binary grammaticality judgments to ordinal-scaled classification tasks, providing richer insights into model behavior including uncertainty quantification through entropy measures.

Abstract: The minimal pairs paradigm of comparing model probabilities for contrasting completions has proven useful for evaluating linguistic knowledge in language models, yet its application has largely been confined to binary grammaticality judgments over syntactic phenomena. Additionally, standard prompting-based evaluation requires expensive text generation, may elicit post-hoc rationalizations rather than model judgments, and discards information about model uncertainty. We address both limitations by extending surprisal-based evaluation from binary grammaticality contrasts to ordinal-scaled classification and scoring tasks across multiple domains. Rather than asking models to generate answers, we measure the information-theoretic “surprise” (negative log probability) they assign to each position on rating scales (e.g., 1-5 or 1-9), yielding full surprisal curves that reveal both the model’s preferred response and its uncertainty via entropy. We explore this framework across four domains: social-ecological-technological systems classification, causal statement identification (binary and scaled), figurative language detection, and deductive qualitative coding. Across these domains, surprisal curves produce interpretable classification signals with clear minima near expected ordinal scale positions, and entropy over the completion tended to distinguish genuinely ambiguous items from easier items.

Method: Given a theme, extract core dramatic conflict, generate explicit climax, then use bidirectional Monte Carlo Tree Search (MCTS) to expand plot backward (rising action, exposition) and forward (falling action, resolution) to create structured outline, followed by final narrative generation.

Result: BiT-MCTS improves narrative coherence, plot structure, and thematic depth relative to strong baselines, enabling substantially longer, more coherent stories according to both automatic metrics and human judgments.

Conclusion: The climax-first bidirectional expansion approach with MCTS effectively addresses LLM limitations in long-form fiction generation, producing better structured and more coherent narratives from open-ended themes.

Abstract: Generating long-form linear fiction from open-ended themes remains a major challenge for large language models, which frequently fail to guarantee global structure and narrative diversity when using premise-based or linear outlining approaches. We present BiT-MCTS, a theme-driven framework that operationalizes a “climax-first, bidirectional expansion” strategy motivated by Freytag’s Pyramid. Given a theme, our method extracts a core dramatic conflict and generates an explicit climax, then employs a bidirectional Monte Carlo Tree Search (MCTS) to expand the plot backward (rising action, exposition) and forward (falling action, resolution) to produce a structured outline. A final generation stage realizes a complete narrative from the refined outline. We construct a Chinese theme corpus for evaluation and conduct extensive experiments across three contemporary LLM backbones. Results show that BiT-MCTS improves narrative coherence, plot structure, and thematic depth relative to strong baselines, while enabling substantially longer, more coherent stories according to automatic metrics and human judgments.

Method: Extends Propp’s narrative theory to define 34 narrative functions for modern web literature, constructs human-annotated corpus, and analyzes LLM-generated text composition using this framework.

Result: Experiments show LLMs fail to comprehend narrative function meanings and adhere to rigid generation paradigms, causing singular narrative logic and severe homogenization.

Conclusion: Current LLMs have fundamental limitations in understanding narrative structures and generating diverse stories, requiring improved comprehension of narrative functions.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in narrative generation. However, they often produce structurally homogenized stories, frequently following repetitive arrangements and combinations of plot events along with stereotypical resolutions. In this paper, we propose a novel theoretical framework for analysis by incorporating Proppian narratology and narrative functions. This framework is used to analyze the composition of narrative texts generated by LLMs to uncover their underlying narrative logic. Taking Chinese web literature as our research focus, we extend Propp’s narrative theory, defining 34 narrative functions suited to modern web narrative structures. We further construct a human-annotated corpus to support the analysis of narrative structures within LLM-generated text. Experiments reveal that the primary reasons for the singular narrative logic and severe homogenization in generated texts are that current LLMs are unable to correctly comprehend the meanings of narrative functions and instead adhere to rigid narrative generation paradigms.

Method: Analyzed 1,116,306 lines from 31,988 poems by 83 poets using grapheme-to-phoneme conversion and six phonetic metrics. Statistical models controlled for meter, form, and line length to isolate poet-level differences.

Result: Found systematic phonetic differences between poets persist even after controlling for meter/form. Identified distinct phonetic profiles (high-sonority lyric, hardness-driven epic, sibilant mystical, high-entropy complex) and historical shifts across centuries.

Conclusion: Persian poetic sound represents conditioned variation within shared prosodic structures, not just individual style or metrical residue. Computational phonetics can contribute to literary-historical interpretation while respecting formal structures.

Abstract: This study examines phonetic texture in Persian poetry as a literary-historical phenomenon rather than a by-product of meter or a feature used only for classification. The analysis draws on a large corpus of 1,116,306 mesras from 31,988 poems written by 83 poets, restricted to five major classical meters to enable controlled comparison. Each line is converted into a grapheme-to-phoneme representation and analyzed using six phonetic metrics: hardness, sonority, sibilance, vowel ratio, phoneme entropy, and consonant-cluster ratio. Statistical models estimate poet-level differences while controlling for meter, poetic form, and line length. The results show that although meter and form explain a substantial portion of phonetic variation, they do not eliminate systematic differences between poets. Persian poetic sound therefore appears as conditioned variation within shared prosodic structures rather than as either purely individual style or simple metrical residue. A multidimensional stylistic map reveals several recurrent phonetic profiles, including high-sonority lyric styles, hardness-driven rhetorical or epic styles, sibilant mystical contours, and high-entropy complex textures. Historical analysis indicates that phonetic distributions shift across centuries, reflecting changes in genre prominence, literary institutions, and performance contexts rather than abrupt stylistic breaks. The study establishes a corpus-scale framework for phonetic analysis in Persian poetry and demonstrates how computational phonetics can contribute to literary-historical interpretation while remaining attentive to the formal structures that shape Persian verse.

Method: Proposes a modular framework separating planning from factual retrieval and answer synthesis. A lightweight student planner is trained via teacher-student framework to generate structured decompositions with abstract reasoning steps and searchable fact requests, using only planning traces and fact requests as supervision.

Result: Evaluation on SEAL-0 benchmark shows supervised planning improves both accuracy and latency compared to monolithic reasoning models and prompt-based tool-augmented frameworks.

Conclusion: Explicitly learned planning structures are essential for reliable fact-seeking LLMs, demonstrating the value of separating planning from retrieval and synthesis in modular frameworks.

Abstract: Fact-seeking question answering with large language models (LLMs) remains unreliable when answers depend on up-to-date or conflicting information. Although retrieval-augmented and tool-using LLMs reduce hallucinations, they often rely on implicit planning, leading to inefficient tool usage. We propose a modular framework that explicitly separates planning from factual retrieval and answer synthesis. A lightweight student planner is trained via a teacher-student framework to generate structured decompositions consisting of abstract reasoning steps and searchable fact requests. The supervision signals contain only planning traces and fact requests, without providing factual answers or retrieved evidence. At inference, the planner produces plans, while prompt-engineered modules perform retrieval and response synthesis. We evaluate the proposed framework on SEAL-0, an extremely challenging benchmark for search-augmented LLMs. Results show that supervised planning improves both accuracy and latency compared to monolithic reasoning models and prompt-based tool-augmented frameworks, demonstrating that explicitly learned planning structures are essential for reliable fact-seeking LLMs.

Method: Two methodological innovations: (1) Verifiable Data Synthesis System for hierarchical datasets for actuarial reasoning and compliance; (2) Progressive SFT-RL Curriculum Framework with dynamic data annealing and synergistic mix of Verified Reasoning (RLVR) and AI Feedback (RLAIF).

Result: INS-S1 achieves SOTA performance on domain tasks, significantly outperforming DeepSeek-R1 and Gemini-2.5-Pro, maintains top-tier general capabilities, and achieves record-low 0.6% hallucination rate (HHEM).

Conclusion: Rigorous domain specialization can be achieved without compromising general intelligence through the proposed end-to-end alignment paradigm.

Abstract: Adapting Large Language Models (LLMs) to high-stakes vertical domains like insurance presents a significant challenge: scenarios demand strict adherence to complex regulations and business logic with zero tolerance for hallucinations. Existing approaches often suffer from a Competency Trade-off - sacrificing general intelligence for domain expertise - or rely heavily on RAG without intrinsic reasoning. To bridge this gap, we present INS-S1, an insurance-specific LLM family trained via a novel end-to-end alignment paradigm. Our approach features two methodological innovations: (1) A Verifiable Data Synthesis System that constructs hierarchical datasets for actuarial reasoning and compliance; and (2) A Progressive SFT-RL Curriculum Framework that integrates dynamic data annealing with a synergistic mix of Verified Reasoning (RLVR) and AI Feedback (RLAIF). By optimizing data ratios and reward signals, this framework enforces domain constraints while preventing catastrophic forgetting. Additionally, we release INSEva, the most comprehensive insurance benchmark to date (39k+ samples). Extensive experiments show that INS-S1 achieves SOTA performance on domain tasks, significantly outperforming DeepSeek-R1 and Gemini-2.5-Pro. Crucially, it maintains top-tier general capabilities and achieves a record-low 0.6% hallucination rate (HHEM). Our results demonstrate that rigorous domain specialization can be achieved without compromising general intelligence.

Method: Reinforcement Learning from Community Feedback (RLCF) with two components: Scientific Judge trained on 700K citation-based paper pairs for preference modeling, and Scientific Thinker aligned using the judge as reward model

Result: Scientific Judge outperforms SOTA LLMs and generalizes to future years, unseen fields, and peer-review preferences; Scientific Thinker proposes higher-impact research ideas than baselines

Conclusion: AI can learn scientific taste, marking a key step toward human-level AI scientists

Abstract: Great scientists have strong judgement and foresight, closely tied to what we call scientific taste. Here, we use the term to refer to the capacity to judge and propose research ideas with high potential impact. However, most relative research focuses on improving an AI scientist’s executive capability, while enhancing an AI’s scientific taste remains underexplored. In this work, we propose Reinforcement Learning from Community Feedback (RLCF), a training paradigm that uses large-scale community signals as supervision, and formulate scientific taste learning as a preference modeling and alignment problem. For preference modeling, we train Scientific Judge on 700K field- and time-matched pairs of high- vs. low-citation papers to judge ideas. For preference alignment, using Scientific Judge as a reward model, we train a policy model, Scientific Thinker, to propose research ideas with high potential impact. Experiments show Scientific Judge outperforms SOTA LLMs (e.g., GPT-5.2, Gemini 3 Pro) and generalizes to future-year test, unseen fields, and peer-review preference. Furthermore, Scientific Thinker proposes research ideas with higher potential impact than baselines. Our findings show that AI can learn scientific taste, marking a key step toward reaching human-level AI scientists.

Method: IPG uses Formula-as-Code paradigm to synthesize physics problems with guaranteed solvability by constructing solutions as executable Python programs, enforcing strict mathematical consistency. Applied to classical mechanics with 165 expert seeds expanded to 1,335 problems.

Result: Created ClassicalMechanicsV1 corpus with 1,335 problems spanning 102 unique physical formulas, average complexity of 3.05 formulas per problem. Found strong linear correlation (R²≈0.95) between formula count and verification code length, establishing code complexity as precise metric for problem difficulty.

Conclusion: IPG enables controllable curriculum generation for reasoning-intensive domains by providing verifiable, high-quality data with guaranteed solvability, addressing data scarcity for training LLMs in complex reasoning tasks.

Abstract: Training large language models for complex reasoning is bottlenecked by the scarcity of verifiable, high-quality data. In domains like physics, standard text augmentation often introduces hallucinations, while static benchmarks lack the reasoning traces required for fine-tuning. We introduce the Infinite Problem Generator (IPG), an agentic framework that synthesizes physics problems with guaranteed solvability through a Formula-as-Code paradigm. Unlike probabilistic text generation, IPG constructs solutions as executable Python programs, enforcing strict mathematical consistency. As a proof-of-concept, we release ClassicalMechanicsV1, a high-fidelity corpus of 1,335 classical mechanics problems expanded from 165 expert seeds. The corpus demonstrates high structural diversity, spanning 102 unique physical formulas with an average complexity of 3.05 formulas per problem. Furthermore, we identify a Complexity Blueprint, demonstrating a strong linear correlation ($R^2 \approx 0.95$) between formula count and verification code length. This relationship establishes code complexity as a precise, proxy-free metric for problem difficulty, enabling controllable curriculum generation. We release the full IPG pipeline, the ClassicalMechanicsV1 dataset, and our evaluation report to support reproducible research in reasoning-intensive domains.

Method: Created MALINT corpus with expert fact-checker annotations, benchmarked 12 language models on intent classification tasks, and proposed intent-based inoculation - an intent-augmented reasoning approach for LLMs that integrates intent analysis.

Result: Intent-augmented reasoning improves zero-shot disinformation detection across six datasets, five LLMs, and seven languages, demonstrating the value of intent analysis in combating disinformation.

Conclusion: Incorporating malicious intent analysis enhances disinformation detection, and the MALINT dataset enables further research in intent-aware disinformation detection.

Abstract: The intentional creation and spread of disinformation poses a significant threat to public discourse. However, existing English datasets and research rarely address the intentionality behind the disinformation. This work presents MALINT, the first human-annotated English corpus developed in collaboration with expert fact-checkers to capture disinformation and its malicious intent. We utilize our novel corpus to benchmark 12 language models, including small language models (SLMs) such as BERT and large language models (LLMs) like Llama 3.3, on binary and multilabel intent classification tasks. Moreover, inspired by inoculation theory from psychology and communication studies, we investigate whether incorporating knowledge of malicious intent can improve disinformation detection. To this end, we propose intent-based inoculation, an intent-augmented reasoning for LLMs that integrates intent analysis to mitigate the persuasive impact of disinformation. Analysis on six disinformation datasets, five LLMs, and seven languages shows that intent-augmented reasoning improves zero-shot disinformation detection. To support research in intent-aware disinformation detection, we release the MALINT dataset with annotations from each annotation step.

Method: Hybrid curation pipeline using Sarvam-M language model with combinatorial prompt engineering for native generation, Google Translate API for cross-lingual expansion, and strict programmatic filtering to ensure quality.

Result: Compiled 132,942 stories with over 93.9 million tokens across 17 Indian languages, creating a foundational resource for multilingual language modeling and transfer learning in Indic languages.

Conclusion: The Multilingual TinyStories dataset serves as a valuable resource for developing and evaluating Small Language Models for low-resource Indian languages, addressing data scarcity through synthetic generation.

Abstract: The development of robust language models for low-resource languages is frequently bottlenecked by the scarcity of high-quality, coherent, and domain-appropriate training corpora. In this paper, we introduce the Multilingual TinyStories dataset, a large-scale, synthetically generated collection of children’s stories encompassing 17 Indian languages. Designed specifically for the training and evaluation of Small Language Models (SLMs), the corpus provides simple, narrative-driven text strictly localized to native scripts. We detail our hybrid curation pipeline, which leverages the Sarvam-M language model and a novel combinatorial prompt engineering framework for native generation, coupled with the Google Translate API for large-scale cross-lingual expansion. Through strict programmatic filtering, we compiled 132,942 stories and over 93.9 million tokens in our release, serving as a foundational resource for multilingual language modeling and transfer learning in the Indic linguistic sphere.

Method: Formalize generation as a trajectory through a relative probability manifold and introduce Top-b, which regulates candidate sets via a dynamic bandwidth coefficient coupled to the instantaneous Shannon entropy of the model’s distribution.

Result: Empirical validation on GPQA and GSM8K benchmarks shows Top-b significantly reduces generation entropy and inter-decoding variance while maintaining competitive reasoning accuracy.

Conclusion: Top-b effectively approximates a self-regulating control system for autoregressive generation by dynamically adapting to the information density of language.

Abstract: Probabilistic language generators are theoretically modeled as discrete stochastic processes, yet standard decoding strategies (Top-k, Top-p) impose static truncation rules that fail to accommodate the dynamic information density of natural language. This misalignment often forces a suboptimal trade-off: static bounds are either too restrictive for high-entropy creative generation or too permissive for low-entropy logical reasoning. In this work, we formalize the generation process as a trajectory through a relative probability manifold. We introduce Top-b (Adaptive Relative Band Sampling), a decoding strategy that regulates the candidate set via a dynamic bandwidth coefficient coupled strictly to the instantaneous Shannon entropy of the model’s distribution. We provide a theoretical framework demonstrating that Top-b acts as a variance-minimizing operator on the tail distribution. Empirical validation on GPQA and GSM8K benchmarks indicates that Top-b significantly reduces generation entropy and inter-decoding variance while maintaining competitive reasoning accuracy, effectively approximating a self-regulating control system for autoregressive generation.

Method: Three-phase methodology testing TRM on 8 low-resource language pairs, comparing recursive mechanisms, representation quality, and frozen vs fine-tuned embeddings.

Result: TRM’s recursive mechanisms don’t transfer to QE; frozen XLM-R embeddings match fine-tuned performance with 37x fewer parameters; achieves 0.370 Spearman correlation.

Conclusion: Weight sharing with frozen embeddings enables parameter efficiency for QE, with frozen TRM-QE outperforming larger models on some languages with 80x fewer trainable parameters.

Abstract: Tiny Recursive Models (TRM) achieve strong results on reasoning tasks through iterative refinement of a shared network. We investigate whether these recursive mechanisms transfer to Quality Estimation (QE) for low-resource languages using a three-phase methodology. Experiments on $8$ language pairs on a low-resource QE dataset reveal three findings. First, TRM’s recursive mechanisms do not transfer to QE. External iteration hurts performance, and internal recursion offers only narrow benefits. Next, representation quality dominates architectural choices, and lastly, frozen pretrained embeddings match fine-tuned performance while reducing trainable parameters by 37$\times$ (7M vs 262M). TRM-QE with frozen XLM-R embeddings achieves a Spearman’s correlation of 0.370, matching fine-tuned variants (0.369) and outperforming an equivalent-depth standard transformer (0.336). On Hindi and Tamil, frozen TRM-QE outperforms MonoTransQuest (560M parameters) with 80$\times$ fewer trainable parameters, suggesting that weight sharing combined with frozen embeddings enables parameter efficiency for QE. We release the code publicly for further research. Code is available at https://github.com/surrey-nlp/TRMQE.

Method: Proposes multi-stage alignment method that teaches models to recall and apply relevant business policies during chain-of-thought reasoning at inference time without full policy context. Introduces PolicyRecall reward based on Jaccard score and Hallucination Penalty for GRPO (Group Relative Policy Optimization) training.

Result: Best model outperforms baseline by 16 points and surpasses comparable in-context baselines of similar model size by 3 points, while using 40% fewer words.

Conclusion: The approach effectively addresses the challenge of business policy adherence in LLM-based conversational assistants by enabling selective policy recall rather than full context inclusion, improving performance while reducing computational overhead.

Abstract: Conversational assistants powered by large language models (LLMs) excel at tool-use tasks but struggle with adhering to complex, business-specific rules. While models can reason over business rules provided in context, including all policies for every query introduces high latency and wastes compute. Furthermore, these lengthy prompts lead to long contexts, harming overall performance due to the “needle-in-the-haystack” problem. To address these challenges, we propose a multi-stage alignment method that teaches models to recall and apply relevant business policies during chain-of-thought reasoning at inference time, without including the full business policy in-context. Furthermore, we introduce a novel PolicyRecall reward based on the Jaccard score and a Hallucination Penalty for GRPO training. Altogether, our best model outperforms the baseline by 16 points and surpasses comparable in-context baselines of similar model size by 3 points, while using 40% fewer words.

Method: Train classifiers to detect concealment behavior in LMs, comparing gradient-based vs prompt-based concealment methods, and testing generalization across model architectures and topics.

Result: Classifiers outperform human evaluators at detecting concealment in smaller models, but fail to generalize to unseen architectures/topics and perform no better than random on models >70B parameters.

Conclusion: Black-box-only LM auditing has limitations; concealment traces fade with model scale, highlighting need for robust detection methods for models hiding knowledge.

Abstract: Language Models (LMs) may acquire harmful knowledge, and yet feign ignorance of these topics when under audit. Inspired by the recent discovery of deception-related behaviour patterns in LMs, we aim to train classifiers that detect when a LM is actively concealing knowledge. Initial findings on smaller models show that classifiers can detect concealment more reliably than human evaluators, with gradient-based concealment proving easier to identify than prompt-based methods. However, contrary to prior work, we find that the classifiers do not reliably generalize to unseen model architectures and topics of hidden knowledge. Most concerningly, the identifiable traces associated with concealment become fainter as the models increase in scale, with the classifiers achieving no better than random performance on any model exceeding 70 billion parameters. Our results expose a key limitation in black-box-only auditing of LMs and highlight the need to develop robust methods to detect models that are actively hiding the knowledge they contain.

Method: Used documented records of books owned/read by Melville, segmented texts at sentence and 5-gram levels, computed semantic similarity using BERTScore, and interpreted precision/recall/F1 scores as indicators of possible semantic alignment rather than applying fixed thresholds.

Result: The approach successfully captured expert-identified instances of similarity and highlighted additional passages warranting further qualitative examination, demonstrating that semantic similarity methods provide a useful computational framework for literary influence studies.

Conclusion: Semantic similarity analysis using modern NLP techniques offers a valuable computational framework for supporting source and influence studies in literary scholarship, bridging computational methods with traditional humanities research.

Abstract: This study investigates the potential influence of Herman Melville reading on his own writings through computational semantic similarity analysis. Using documented records of books known to have been owned or read by Melville, we compare selected passages from his works with texts from his library. The methodology involves segmenting texts at both sentence level and non-overlapping 5-gram level, followed by similarity computation using BERTScore. Rather than applying fixed thresholds to determine reuse, we interpret precision, recall, and F1 scores as indicators of possible semantic alignment that may suggest literary influence. Experimental results demonstrate that the approach successfully captures expert-identified instances of similarity and highlights additional passages warranting further qualitative examination. The findings suggest that semantic similarity methods provide a useful computational framework for supporting source and influence studies in literary scholarship.

Method: Proposes two complementary research directions: 1) Building robust, agent-based automatic data preparation systems supporting automated workflow construction and scalable data management; 2) Developing unified data-model interaction training systems where data is dynamically selected, mixed, and reweighted throughout training.

Result: This is a position paper proposing research directions rather than presenting experimental results. The authors outline a vision for more efficient, adaptive, and performance-aware data utilization in LLM training.

Conclusion: The paper identifies critical limitations in current LLM data preparation and utilization, advocating for systematic approaches to automate data workflows and enable dynamic data-model interaction during training to improve efficiency and performance.

Abstract: Large language models (LLMs) have demonstrated remarkable performance across a wide range of tasks and domains, with data playing a central role in enabling these advances. Despite this success, the preparation and effective utilization of the massive datasets required for LLM training remain major bottlenecks. In current practice, LLM training data is often constructed using ad hoc scripts, and there is still a lack of mature, agent-based data preparation systems that can automatically construct robust and reusable data workflows, thereby freeing data scientists from repetitive and error-prone engineering efforts. Moreover, once collected, datasets are often consumed largely in their entirety during training, without systematic mechanisms for data selection, mixture optimization, or reweighting. To address these limitations, we advocate two complementary research directions. First, we propose building a robust, agent-based automatic data preparation system that supports automated workflow construction and scalable data management. Second, we argue for a unified data-model interaction training system in which data is dynamically selected, mixed, and reweighted throughout the training process, enabling more efficient, adaptive, and performance-aware data utilization. Finally, we discuss the remaining challenges and outline promising directions for future research and system development.

Method: Used low-data LoRA safety fine-tuning with four supervision formats built from the same core safety rules: constitutional rules (A), creed-style identity framing (B), B-matched creed with worldview/confession tail (C), and matched non-identity condition (D). Evaluated across three instruction-tuned model families (Llama 3.1 8B, Qwen2.5 7B, Gemma 3 4B) using HarmBench with dual-judge pipeline (DeepSeek v3.2 and Sonnet 4.6).

Result: Non-identity condition D performed best across all three model families on HarmBench (74.4% refusal on Llama, 76.9% on Gemma, 74.1% on Qwen). Creed-style framing (B) improved over plain constitutional rules (A) but remained substantially below D. Overall ordering: D > B > C ≥ A > baseline. No meaningful capability trade-offs observed on MMLU and ARC-Challenge.

Conclusion: Explicit creed-style identity language is not necessary for strongest safety gains; supervision format matters more than identity content, challenging strong versions of identity-framing hypothesis.

Abstract: How safety supervision is written may matter more than the explicit identity content it contains. We study low-data LoRA safety fine-tuning with four supervision formats built from the same core safety rules: constitutional rules (A), creed-style identity framing (B), a B-matched creed condition with a worldview/confession identity-maintenance tail (C), and a matched non-identity condition (D). Across three instruction-tuned model families (Llama 3.1 8B, Qwen2.5 7B, and Gemma 3 4B), we evaluate HarmBench using a reconciled dual-judge pipeline combining Bedrock-hosted DeepSeek v3.2 and Sonnet 4.6, with disagreement and boundary cases manually resolved. The non-identity condition D is the strongest group on all three model families on the full 320-behavior HarmBench set, reaching 74.4% refusal on Llama, 76.9% on Gemma, and 74.1% on Qwen. By comparison, creed-style framing (B) improves over plain constitutional rules (A) on Llama and Gemma, but remains substantially below D, yielding an overall descriptive ordering of $D > B > C \geq A > baseline$. This provides a bounded empirical challenge to a strong version of the identity-framing hypothesis: explicit creed-style identity language is not necessary for the strongest gains observed here. Capability evaluations on MMLU and ARC-Challenge show no meaningful trade-off across conditions.

Method: Treat constituent headedness as supervised prediction task using aligned constituency and dependency annotations. Define each constituent head as the dependency span head. Train models on English and Chinese data, comparing against Collins-style rule-based percolation.

Result: Models achieve near-ceiling intrinsic accuracy, substantially outperform rule-based percolation. Predicted heads yield comparable parsing accuracy under head-driven binarization, improve constituency-to-dependency conversion fidelity, and transfer across resources/languages via simple label-mapping.

Conclusion: Headedness can be effectively learned as explicit representation layer, offering advantages over procedural approaches in accuracy, conversion fidelity, and cross-resource/language transferability.

Abstract: Headedness is widely used as an organizing device in syntactic analysis, yet constituency treebanks rarely encode it explicitly and most processing pipelines recover it procedurally via percolation rules. We treat this notion of constituent headedness as an explicit representational layer and learn it as a supervised prediction task over aligned constituency and dependency annotations, inducing supervision by defining each constituent head as the dependency span head. On aligned English and Chinese data, the resulting models achieve near-ceiling intrinsic accuracy and substantially outperform Collins-style rule-based percolation. Predicted heads yield comparable parsing accuracy under head-driven binarization, consistent with the induced binary training targets being largely equivalent across head choices, while increasing the fidelity of deterministic constituency-to-dependency conversion and transferring across resources and languages under simple label-mapping interfaces.

Method: Proposes novel PPMT task definition, constructs three benchmark test datasets, designs evaluation metrics, and proposes benchmark methods. Focuses on protecting named entity privacy since entities often contain personal privacy and commercial secrets.

Result: Establishes foundational framework for privacy protection in machine translation with task definition, datasets, metrics, and benchmark methods as starting point for research community.

Conclusion: This work provides new perspective and solid foundation for privacy protection problem in machine translation, addressing a critical gap in the field.

Abstract: Current online translation services require sending user text to cloud servers, posing a risk of privacy leakage when the text contains sensitive information. This risk hinders the application of online translation services in privacy-sensitive scenarios. One way to mitigate this risk for online translation services is introducing privacy protection mechanisms targeting the inference stage of translation models. However, compared to subfields of NLP like text classification and summarization, the machine translation research community has limited exploration of privacy protection during the inference stage. There is no clearly defined privacy protection task for the inference stage, dedicated evaluation datasets and metrics, and reference benchmark methods. The absence of these elements has seriously constrained researchers’ in-depth exploration of this direction. To bridge this gap, this paper proposes a novel “Privacy-Preserving Machine Translation” (PPMT) task, aiming to protect the private information in text during the model inference stage. For this task, we constructed three benchmark test datasets, designed corresponding evaluation metrics, and proposed a series of benchmark methods as a starting point for this task. The definition of privacy is complex and diverse. Considering that named entities often contain a large amount of personal privacy and commercial secrets, we have focused our research on protecting only the named entity’s privacy in the text. We expect this research work will provide a new perspective and a solid foundation for the privacy protection problem in machine translation.

Method: Proposes a generalizable data aggregation and preprocessing pipeline with rigorous processing steps to ensure data diversity, balance, and inclusion of crucial features like word-level timestamps from diverse, potentially noisy open-source sources.

Result: Successfully applied to Vietnamese, resulting in a unified, high-quality 500-hour dataset that provides a foundation for training and evaluating state-of-the-art Vietnamese ASR systems.

Conclusion: The proposed pipeline effectively addresses data quality issues for low-resource languages and enables the creation of high-quality ASR datasets from noisy open-source data.

Abstract: Automatic Speech Recognition (ASR) performance is heavily dependent on the availability of large-scale, high-quality datasets. For low-resource languages, existing open-source ASR datasets often suffer from insufficient quality and inconsistent annotation, hindering the development of robust models. To address these challenges, we propose a novel and generalizable data aggregation and preprocessing pipeline designed to construct high-quality ASR datasets from diverse, potentially noisy, open-source sources. Our pipeline incorporates rigorous processing steps to ensure data diversity, balance, and the inclusion of crucial features like word-level timestamps. We demonstrate the effectiveness of our methodology by applying it to Vietnamese, resulting in a unified, high-quality 500-hour dataset that provides a foundation for training and evaluating state-of-the-art Vietnamese ASR systems. Our project page is available at https://github.com/qualcomm-ai-research/PhoASR.

Method: Manually constructed a novel QA dataset capturing knowledge from local Wikipedia pages absent from higher-resource counterparts (Mandarin Chinese vs. Cantonese, German vs. Bavarian). Evaluated LLM performance with and without context from lead sections, and explored translation strategies.

Result: LLMs fail to answer questions about information only in local Wikipedia editions. Providing context from lead sections substantially improves performance, with further gains via translation. Local Wikipedia editions serve as valuable sources of both regional and global information.

Conclusion: The findings raise critical questions about inclusivity and cultural coverage of LLMs, highlighting the need to address information asymmetry between standard and local language resources.

Abstract: Large Language Models (LLMs) are becoming a common way for humans to seek knowledge, yet their coverage and reliability vary widely. Especially for local language varieties, there are large asymmetries, e.g., information in local Wikipedia that is absent from the standard variant. However, little is known about how well LLMs perform under such information asymmetry, especially on closely related languages. We manually construct a novel challenge question-answering (QA) dataset that captures knowledge conveyed on a local Wikipedia page, which is absent from their higher-resource counterparts-covering Mandarin Chinese vs. Cantonese and German vs. Bavarian. Our experiments show that LLMs fail to answer questions about information only in local editions of Wikipedia. Providing context from lead sections substantially improves performance, with further gains possible via translation. Our topical, geographic annotations, and stratified evaluations reveal the usefulness of local Wikipedia editions as sources of both regional and global information. These findings raise critical questions about inclusivity and cultural coverage of LLMs.

Method: Created ideological corpus of 1,117 COVID-19 treatment articles, used Lexical Multidimensional Analysis to identify ideological dimensions, tested LLMs with two prompt types (question+texts vs question+texts+LMDA descriptions), and measured alignment using cosine similarity for lexical and semantic representations.

Result: LLMs’ responses based on ideological retrieved texts show greater alignment with external ideological content, with enhanced prompts (including LMDA descriptions) further influencing outputs toward the encountered ideology.

Conclusion: The study highlights the importance of identifying ideological discourses in RAG frameworks to mitigate unintended bias and risks of malicious manipulation, showing that retrieved content significantly shapes LLM ideological outputs.

Abstract: This paper studies the impact of retrieved ideological texts on the outputs of large language models (LLMs). While interest in understanding ideology in LLMs has recently increased, little attention has been given to this issue in the context of Retrieval-Augmented Generation (RAG). To fill this gap, we design an external knowledge source based on ideological loaded texts about COVID-19 treatments. Our corpus is based on 1,117 academic articles representing discourses about controversial and endorsed treatments for the disease. We propose a corpus linguistics framework, based on Lexical Multidimensional Analysis (LMDA), to identify the ideologies within the corpus. LLMs are tasked to answer questions derived from three identified ideological dimensions, and two types of contextual prompts are adopted: the first comprises the user question and ideological texts; and the second contains the question, ideological texts, and LMDA descriptions. Ideological alignment between reference ideological texts and LLMs’ responses is assessed using cosine similarity for lexical and semantic representations. Results demonstrate that LLMs’ responses based on ideological retrieved texts are more aligned with the ideology encountered in the external knowledge, with the enhanced prompt further influencing LLMs’ outputs. Our findings highlight the importance of identifying ideological discourses within the RAG framework in order to mitigate not just unintended ideological bias, but also the risks of malicious manipulation of such models.

Method: 1) LLM-powered semantic enriching strategy to incorporate possible meanings and toxicity clues into perturbed text, 2) discriminability-driven feature learning to strengthen discriminative features while suppressing less-discriminative ones for robust classification boundaries.

Result: ContiGuard enables detectors to continually update capabilities and maintain sustained resilience against evolving perturbations, addressing the challenge of continual toxicity detection on time-evolving perturbed text.

Conclusion: The first framework for continual toxicity detection that combines LLM-powered semantic enrichment with discriminative feature learning to handle evolving text perturbations effectively.

Abstract: Toxicity detection mitigates the dissemination of toxic content (e.g., hateful comments, posts, and messages within online social actions) to safeguard a healthy online social environment. However, malicious users persistently develop evasive perturbations to disguise toxic content and evade detectors. Traditional detectors or methods are static over time and are inadequate in addressing these evolving evasion tactics. Thus, continual learning emerges as a logical approach to dynamically update detection ability against evolving perturbations. Nevertheless, disparities across perturbations hinder the detector’s continual learning on perturbed text. More importantly, perturbation-induced noises distort semantics to degrade comprehension and also impair critical feature learning to render detection sensitive to perturbations. These amplify the challenge of continual learning against evolving perturbations. In this work, we present ContiGuard, the first framework tailored for continual learning of the detector on time-evolving perturbed text (termed continual toxicity detection) to enable the detector to continually update capability and maintain sustained resilience against evolving perturbations. Specifically, to boost the comprehension, we present an LLM-powered semantic enriching strategy, where we dynamically incorporate possible meaning and toxicity-related clues excavated by LLM into the perturbed text to improve the comprehension. To mitigate non-critical features and amplify critical ones, we propose a discriminability-driven feature learning strategy, where we strengthen discriminative features while suppressing the less-discriminative ones to shape a robust classification boundary for detection…

Method: Introduces Shopping Companion, a unified framework that jointly tackles memory retrieval and shopping assistance while supporting user intervention. Uses dual-reward reinforcement learning with tool-wise rewards to handle sparse/discontinuous rewards in multi-turn interactions.

Result: State-of-the-art models (including GPT-5) achieve under 70% success rate on their benchmark. Their lightweight LLM trained with Shopping Companion consistently outperforms strong baselines, achieving better preference capture and task performance.

Conclusion: The unified design of Shopping Companion is effective for long-term preference-aware shopping tasks, addressing the limitations of separate component approaches and demonstrating superior performance over existing models.

Abstract: In e-commerce, LLM agents show promise for shopping tasks such as recommendations, budgeting, and bundle deals, where accurately capturing user preferences from long-term conversations is critical. However, two challenges hinder realizing this potential: (1) the absence of benchmarks for evaluating long-term preference-aware shopping tasks, and (2) the lack of end-to-end optimization due to existing designs that treat preference identification and shopping assistance as separate components. In this paper, we introduce a novel benchmark with a long-term memory setup, spanning two shopping tasks over 1.2 million real-world products, and propose Shopping Companion, a unified framework that jointly tackles memory retrieval and shopping assistance while supporting user intervention. To train such capabilities, we develop a dual-reward reinforcement learning strategy with tool-wise rewards to handle the sparse and discontinuous rewards inherent in multi-turn interactions. Experimental results demonstrate that even state-of-the-art models (such as GPT-5) achieve success rates under 70% on our benchmark, highlighting the significant challenges in this domain. Notably, our lightweight LLM, trained with Shopping Companion, consistently outperforms strong baselines, achieving better preference capture and task performance, which validates the effectiveness of our unified design.

Method: Used a community-curated parallel corpus of 13,865 English-Efik sentence pairs. Fine-tuned both mT5 multilingual model and NLLB-200 model on this dataset. Evaluated using BLEU and chrF scores for both translation directions.

Result: NLLB-200 outperformed mT5, achieving BLEU scores of 26.64 (English-Efik) and 31.21 (Efik-English), with chrF scores of 51.04 and 47.92 respectively, indicating improved fluency and semantic fidelity.

Conclusion: Demonstrates feasibility of developing practical machine translation tools for low-resource languages and highlights importance of inclusive data practices and culturally grounded evaluation for equitable NLP.

Abstract: Low-resource languages serve as invaluable repositories of human history, preserving cultural and intellectual diversity. Despite their significance, they remain largely absent from modern natural language processing systems. While progress has been made for widely spoken African languages such as Swahili, Yoruba, and Amharic, smaller indigenous languages like Efik continue to be underrepresented in machine translation research. This study evaluates the effectiveness of state-of-the-art multilingual neural machine translation models for English-Efik translation, leveraging a small-scale, community-curated parallel corpus of 13,865 sentence pairs. We fine-tuned both the mT5 multilingual model and the NLLB200 model on this dataset. NLLB-200 outperformed mT5, achieving BLEU scores of 26.64 for English-Efik and 31.21 for Efik-English, with corresponding chrF scores of 51.04 and 47.92, indicating improved fluency and semantic fidelity. Our findings demonstrate the feasibility of developing practical machine translation tools for low-resource languages and highlight the importance of inclusive data practices and culturally grounded evaluation in advancing equitable NLP.

Method: DLOM reuses language model head to extract score-wise logits on predefined score tokens, enabling direct optimization in score space. DLOM-GF adds gated fusion module to adaptively combine textual and multimodal score logits. DLOM-DA adds distance-aware regularization term to better reflect ordinal distances for text-only AES.

Result: On multimodal EssayJudge dataset, DLOM improves over generation-based SFT baseline across scoring traits, and DLOM-GF yields further gains when modality relevance is heterogeneous. On text-only ASAP/ASAP++ benchmarks, DLOM remains effective without visual inputs, and DLOM-DA further improves performance and outperforms strong baselines.

Conclusion: DLOM provides an effective framework for explicit ordinal decision modeling in AES, with extensions for multimodal adaptive fusion and text-only ordinal distance regularization, demonstrating improved performance over generation-based approaches.

Abstract: Automated essay scoring (AES) predicts multiple rubric-defined trait scores for each essay, where each trait follows an ordered discrete rating scale. Most LLM-based AES methods cast scoring as autoregressive token generation and obtain the final score via decoding and parsing, making the decision implicit. This formulation is particularly sensitive in multimodal AES, where the usefulness of visual inputs varies across essays and traits. To address these limitations, we propose Decision-Level Ordinal Modeling (DLOM), which makes scoring an explicit ordinal decision by reusing the language model head to extract score-wise logits on predefined score tokens, enabling direct optimization and analysis in the score space. For multimodal AES, DLOM-GF introduces a gated fusion module that adaptively combines textual and multimodal score logits. For text-only AES, DLOM-DA adds a distance-aware regularization term to better reflect ordinal distances. Experiments on the multimodal EssayJudge dataset show that DLOM improves over a generation-based SFT baseline across scoring traits, and DLOM-GF yields further gains when modality relevance is heterogeneous. On the text-only ASAP/ASAP++ benchmarks, DLOM remains effective without visual inputs, and DLOM-DA further improves performance and outperforms strong representative baselines.

Method: Treat three LLMs as observers performing factual discrimination across 168,000 trials. Apply full parametric SDT framework including unequal-variance model fitting, criterion estimation, and z-ROC analysis. Test whether temperature functions as a criterion shift analogous to payoff manipulations in human psychophysics.

Result: Temperature simultaneously increases sensitivity (AUC) and shifts criterion, unlike human psychophysics where payoff only shifts criterion. Models showed unequal-variance evidence distributions (z-ROC slopes 0.52-0.84), with instruct models showing more extreme asymmetry than base models or humans. SDT decomposition revealed models with distinct sensitivity-bias positions couldn’t be distinguished by calibration metrics alone.

Conclusion: The full parametric SDT framework provides diagnostic information unavailable from existing calibration metrics, revealing that temperature affects LLMs differently than payoff manipulations affect humans, changing both sensitivity and criterion rather than just criterion.

Abstract: Large language models (LLMs) are evaluated for calibration using metrics such as Expected Calibration Error that conflate two distinct components: the model’s ability to discriminate correct from incorrect answers (sensitivity) and its tendency toward confident or cautious responding (bias). Signal Detection Theory (SDT) decomposes these components. While SDT-derived metrics such as AUROC are increasingly used, the full parametric framework - unequal-variance model fitting, criterion estimation, z-ROC analysis - has not been applied to LLMs as signal detectors. In this pre-registered study, we treat three LLMs as observers performing factual discrimination across 168,000 trials and test whether temperature functions as a criterion shift analogous to payoff manipulations in human psychophysics. Critically, this analogy may break down because temperature changes the generated answer itself, not only the confidence assigned to it. Our results confirm the breakdown with temperature simultaneously increasing sensitivity (AUC) and shifting criterion. All models exhibited unequal-variance evidence distributions (z-ROC slopes 0.52-0.84), with instruct models showing more extreme asymmetry (0.52-0.63) than the base model (0.77-0.87) or human recognition memory (~0.80). The SDT decomposition revealed that models occupying distinct positions in sensitivity-bias space could not be distinguished by calibration metrics alone, demonstrating that the full parametric framework provides diagnostic information unavailable from existing metrics.

Method: Proposes ExPosST framework with explicit position allocation that reserves fixed positional slots for incoming source tokens, enabling efficient KV cache decoding across different positional encoding methods. Also introduces policy-consistent fine-tuning strategy to align training with inference-time decoding behavior.

Result: Experiments across multiple language pairs demonstrate that ExPosST effectively supports simultaneous translation under diverse policies, resolving the efficiency-consistency dilemma.

Conclusion: ExPosST provides a general framework that achieves inference efficiency, positional consistency, and broad model compatibility for LLM-based simultaneous machine translation.

Abstract: Large language models (LLMs) have recently demonstrated promising performance in simultaneous machine translation (SimulMT). However, applying decoder-only LLMs to SimulMT introduces a positional mismatch, which leads to a dilemma between decoding efficiency and positional consistency. Existing approaches often rely on specific positional encodings or carefully designed prompting schemes, and thus fail to simultaneously achieve inference efficiency, positional consistency, and broad model compatibility. In this work, we propose ExPosST, a general framework that resolves this dilemma through explicit position allocation. ExPosST reserves fixed positional slots for incoming source tokens, enabling efficient decoding with KV cache across different positional encoding methods. To further bridge the gap between fine-tuning and inference, we introduce a policy-consistent fine-tuning strategy that aligns training with inference-time decoding behavior. Experiments across multiple language pairs demonstrate that ExPosST effectively supports simultaneous translation under diverse policies.

Method: Proposes Holographic Agent Assessment Framework (HAAF) with four components: (1) static cognitive/policy analysis, (2) interactive sandbox simulation, (3) social-ethical alignment assessment, and (4) distribution-aware representative sampling engine optimizing coverage and risk sensitivity for tail risks. Connected through iterative Trustworthy Optimization Factory with red-team/blue-team cycles.

Result: Framework shifts agent evaluation from benchmark islands toward representative real-world trustworthiness assessment, with code and data available for pilot implementation.

Conclusion: HAAF provides systematic paradigm for assessing agent trustworthiness over representative scenario distributions, addressing limitations of current fragmented evaluations and enabling progressive vulnerability reduction through iterative optimization cycles.

Abstract: As agentic AI systems move beyond static question answering into open-ended, tool-augmented, and multi-step real-world workflows, their increased authority poses greater risks of system misuse and operational failures. However, current evaluation practices remain fragmented, measuring isolated capabilities such as coding, hallucination, jailbreak resistance, or tool use in narrowly defined settings. We argue that the central limitation is not merely insufficient coverage of evaluation dimensions, but the lack of a principled notion of representativeness: an agent’s trustworthiness should be assessed over a representative socio-technical scenario distribution rather than a collection of disconnected benchmark instances. To this end, we propose the Holographic Agent Assessment Framework (HAAF), a systematic evaluation paradigm that characterizes agent trustworthiness over a scenario manifold spanning task types, tool interfaces, interaction dynamics, social contexts, and risk levels. The framework integrates four complementary components: (i) static cognitive and policy analysis, (ii) interactive sandbox simulation, (iii) social-ethical alignment assessment, and (iv) a distribution-aware representative sampling engine that jointly optimizes coverage and risk sensitivity – particularly for rare but high-consequence tail risks that conventional benchmarks systematically overlook. These components are connected through an iterative Trustworthy Optimization Factory. Through cycles of red-team probing and blue-team hardening, this paradigm progressively narrows the vulnerabilities to meet deployment standards, shifting agent evaluation from benchmark islands toward representative, real-world trustworthiness. Code and data for the illustrative instantiation are available at https://github.com/TonyQJH/haaf-pilot.

Method: OrgForge uses a multi-agent simulation framework with strict separation between deterministic ground truth (maintained by Python engine) and LLM-generated surface prose. It enforces causal timestamp correctness via actor-local clocks and implements graph-dynamic subsystems (stress propagation, temporal edge-weight decay, Dijkstra escalation routing) to govern organizational behavior independently of LLMs.

Result: The framework produces interleaved organizational artifacts (Slack threads, JIRA tickets, Confluence pages, Git pull requests, emails) traceable to a shared immutable event log, with causal chain tracking, recurrence detection, and probabilistic email routing systems.

Conclusion: OrgForge provides a novel approach to generating synthetic organizational data for RAG evaluation by maintaining strict separation between deterministic ground truth and LLM-generated content, addressing limitations of existing datasets while ensuring temporal consistency and cross-artifact traceability.

Abstract: Evaluating retrieval-augmented generation (RAG) pipelines requires corpora where ground truth is knowable, temporally structured, and cross-artifact properties that real-world datasets rarely provide cleanly. Existing resources such as the Enron corpus carry legal ambiguity, demographic skew, and no structured ground truth. Purely LLM-generated synthetic data solves the legal problem but introduces a subtler one: the generating model cannot be prevented from hallucinating facts that contradict themselves across documents.We present OrgForge, an open-source multi-agent simulation framework that enforces a strict physics-cognition boundary: a deterministic Python engine maintains a SimEvent ground truth bus; large language models generate only surface prose, constrained by validated proposals. An actor-local clock enforces causal timestamp correctness across all artifact types, eliminating the class of timeline inconsistencies that arise when timestamps are sampled independently per document. We formalize three graph-dynamic subsystems stress propagation via betweenness centrality, temporal edge-weight decay, and Dijkstra escalation routing that govern organizational behavior independently of any LLM. Running a configurable N-day simulation, OrgForge produces interleaved Slack threads, JIRA tickets, Confluence pages, Git pull requests, and emails, all traceable to a shared, immutable event log. We additionally describe a causal chain tracking subsystem that accumulates cross-artifact evidence graphs per incident, a hybrid reciprocal-rank-fusion recurrence detector for identifying repeated failure classes, and an inbound/outbound email engine that routes vendor alerts, customer complaints, and HR correspondence through gated causal chains with probabilistic drop simulation. OrgForge is available under the MIT license.

Method: Pretrained a suite of Latvian-specific encoders based on RoBERTa, DeBERTaV3, and ModernBERT architectures, including long-context variants. Evaluated them across diverse Latvian diagnostic and linguistic benchmarks.

Result: The models are competitive with existing monolingual and multilingual encoders. The best model, lv-deberta-base (111M parameters), achieves strongest overall performance, outperforming larger multilingual baselines and prior Latvian-specific encoders.

Conclusion: The research successfully addresses the gap in Latvian NLP by providing competitive encoder models and releasing all pretrained models and evaluation resources to support further research and practical applications in Latvian NLP.

Abstract: Encoder-only transformers remain essential for practical NLP tasks. While recent advances in multilingual models have improved cross-lingual capabilities, low-resource languages such as Latvian remain underrepresented in pretraining corpora, and few monolingual Latvian encoders currently exist. We address this gap by pretraining a suite of Latvian-specific encoders based on RoBERTa, DeBERTaV3, and ModernBERT architectures, including long-context variants, and evaluating them across a diverse set of Latvian diagnostic and linguistic benchmarks. Our models are competitive with existing monolingual and multilingual encoders while benefiting from recent architectural and efficiency advances. Our best model, lv-deberta-base (111M parameters), achieves the strongest overall performance, outperforming larger multilingual baselines and prior Latvian-specific encoders. We release all pretrained models and evaluation resources to support further research and practical applications in Latvian NLP.

Method: Proposes Attention Residuals (AttnRes) using softmax attention over preceding layer outputs for selective aggregation with learned, input-dependent weights. Introduces Block AttnRes for efficiency, partitioning layers into blocks and attending over block-level representations with cache-based pipeline communication and two-phase computation.

Result: Scaling law experiments show consistent improvement across model sizes. Integration into Kimi Linear architecture (48B total/3B activated) trained on 1.4T tokens yields more uniform output magnitudes and gradient distribution, improving downstream performance across all evaluated tasks.

Conclusion: AttnRes effectively mitigates PreNorm dilution, provides practical drop-in replacement for standard residual connections with minimal overhead, and enhances model performance through content-dependent depth-wise selection.

Abstract: Residual connections with PreNorm are standard in modern LLMs, yet they accumulate all layer outputs with fixed unit weights. This uniform aggregation causes uncontrolled hidden-state growth with depth, progressively diluting each layer’s contribution. We propose Attention Residuals (AttnRes), which replaces this fixed accumulation with softmax attention over preceding layer outputs, allowing each layer to selectively aggregate earlier representations with learned, input-dependent weights. To address the memory and communication overhead of attending over all preceding layer outputs for large-scale model training, we introduce Block AttnRes, which partitions layers into blocks and attends over block-level representations, reducing the memory footprint while preserving most of the gains of full AttnRes. Combined with cache-based pipeline communication and a two-phase computation strategy, Block AttnRes becomes a practical drop-in replacement for standard residual connections with minimal overhead. Scaling law experiments confirm that the improvement is consistent across model sizes, and ablations validate the benefit of content-dependent depth-wise selection. We further integrate AttnRes into the Kimi Linear architecture (48B total / 3B activated parameters) and pre-train on 1.4T tokens, where AttnRes mitigates PreNorm dilution, yielding more uniform output magnitudes and gradient distribution across depth, and improves downstream performance across all evaluated tasks.

Method: Replicated original system, then extended with newer multilingual language models (Qwen, mGPT), added 26 document-level stylometric features, used mDeBERTa-v3-base for contextual representations, and applied SHAP analysis to examine feature importance.

Result: Additional stylometric features improved performance in both tasks and languages; multilingual configuration achieved comparable or better results than language-specific models; replication challenges highlighted importance of clear documentation.

Conclusion: The study demonstrates improved authorship attribution for machine-generated texts using newer multilingual models and stylometric features, while emphasizing the importance of clear documentation for reliable replication and fair system comparison.

Abstract: This paper replicates and extends the system used in the AuTexTification 2023 shared task for authorship attribution of machine-generated texts. First, we tried to reproduce the original results. Exact replication was not possible because of differences in data splits, model availability, and implementation details. Next, we tested newer multilingual language models and added 26 document-level stylometric features. We also applied SHAP analysis to examine which features influence the model’s decisions. We replaced the original GPT-2 models with newer generative models such as Qwen and mGPT for computing probabilistic features. For contextual representations, we used mDeBERTa-v3-base and applied the same configuration to both English and Spanish. This allowed us to use one shared configuration for Subtask 1 and Subtask 2. Our experiments show that the additional stylometric features improve performance in both tasks and both languages. The multilingual configuration achieves the results that are comparable to or better than language-specific models. The study also shows that clear documentation is important for reliable replication and fair comparison of systems.

Method: AdaAnchor uses latent anchor vectors attached to input that are refined through silent iterative computation. It incorporates adaptive halting that monitors anchor stability across iterations and terminates refinement once convergence is detected, allocating fewer steps to easier instances and more to harder ones within a maximum-step budget.

Result: Achieves accuracy gains up to 5% over fixed-step latent refinement while reducing average latent refinement steps by 48-60%. Reduces generated tokens by 92-93% compared to standard reasoning baselines by moving computation into silent latent refinement.

Conclusion: AdaAnchor offers an effective accuracy-efficiency trade-off for reasoning tasks by enabling silent latent computation with adaptive halting, substantially reducing output token usage while maintaining or improving accuracy.

Abstract: Token-level Chain-of-Thought (CoT) prompting has become a standard way to elicit multi-step reasoning in large language models (LLMs), especially for mathematical word problems. However, generating long intermediate traces increases output length and inference cost, and can be inefficient when the model could arrive at the correct answer without extensive verbalization. This has motivated latent-space reasoning approaches that shift computation into hidden representations and only emit a final answer. Yet, many latent reasoning methods depend on a fixed number of latent refinement steps at inference, adding another hyperparameter that must be tuned across models and datasets to balance accuracy and efficiency. We introduce AdaAnchor, a latent reasoning framework that performs silent iterative computation by refining a set of latent anchor vectors attached to the input. AdaAnchor further incorporates an adaptive halting mechanism that monitors anchor stability across iterations and terminates refinement once the anchor dynamics converge, allocating fewer steps to easier instances while reserving additional refinement steps for harder ones under a shared maximum-step budget. Our empirical evaluation across three mathematical word-problem benchmarks shows that AdaAnchor with adaptive halting yields accuracy gains of up to 5% over fixed-step latent refinement while reducing average latent refinement steps by 48-60% under the same maximum-step budget. Compared to standard reasoning baselines, AdaAnchor achieves large reductions in generated tokens (92-93%) by moving computation into silent latent refinement, offering a different accuracy-efficiency trade-off with substantially lower output-token usage.

Method: 1) Multi-agent collaborative workflow based on Grounded Theory for dimensional decomposition and hierarchical induction to produce interpretable fine-grained criteria. 2) Memory-augmented Replay Policy Optimization (MRPO) algorithm for self-reflection based on dynamic criteria and end-to-end optimization combining supervised fine-tuning with reinforcement learning.

Result: Automatically constructed criteria achieve performance gains comparable to human annotations. Writer-R1-4B models trained with this approach outperform baselines across multiple creative writing tasks and surpass some 100B+ parameter open-source models.

Conclusion: The proposed approach effectively addresses evaluation challenges in creative writing through automated criteria generation and reinforcement learning optimization, enabling high-quality creative writing models with relatively small parameter counts.

Abstract: As a typical open-ended generation task, creative writing lacks verifiable reference answers, which has long constrained reward modeling and automatic evaluation due to high human annotation costs, evaluative bias, and coarse feedback signals. To address these challenges, this paper first designs a multi-agent collaborative workflow based on Grounded Theory, performing dimensional decomposition and hierarchical induction of the problem to dynamically produce interpretable and reusable fine-grained criteria. Furthermore, we propose the Memory-augmented Replay Policy Optimization (MRPO) algorithm: on the one hand, without additional training, MRPO guides models to engage in self-reflection based on dynamic criteria, enabling controlled iterative improvement; on the other hand, we adopt the training paradigm that combines supervised fine-tuning with reinforcement learning to convert evaluation criteria into reward signals, achieving end-to-end optimization. Experimental results demonstrate that the automatically constructed criteria achieve performance gains comparable to human annotations. Writer-R1-4B models trained with this approach outperform baselines across multiple creative writing tasks and surpass some 100B+ parameter open-source models.

Method: Two-phase approach: 1) Develop conversion pipeline from Japanese Legal Standard (JLS) XML to Akoma Ntoso (AKN) standard for structural interoperability; 2) Apply multilingual embedding models and semantic textual similarity techniques with FAISS retrieval and Cross-Encoder reranking to identify corresponding legal provisions across jurisdictions

Result: Created a prototype system that generates candidate correspondences between legal provisions and visualizes them as cross-jurisdictional networks for exploratory comparative analysis

Conclusion: The integrated framework successfully enables computational comparative law by combining structural standardization with semantic analysis techniques, facilitating cross-jurisdictional legal research and analysis

Abstract: This paper presents an integrated framework for computational comparative law by connecting two consecutive research projects based on the Japanese Legal Standard (JLS) XML schema. The first project establishes structural interoperability by developing a conversion pipeline from JLS to the Akoma Ntoso (AKN) standard, enabling Japanese statutes to be integrated into international LegalDocML-based legislative databases. Building on this foundation, the second project applies multilingual embedding models and semantic textual similarity techniques to identify corresponding provisions across national legal systems. A prototype system combining multilingual embeddings, FAISS retrieval, and Cross-Encoder reranking generates candidate correspondences and visualizes them as cross-jurisdictional networks for exploratory comparative analysis.

Method: Proposes MMKU-Bench benchmark with two knowledge scenarios (updated and unknown). Evaluates three approaches: supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and knowledge editing (KE).

Result: SFT and RLHF suffer from catastrophic forgetting, while KE better preserves general capabilities but has limitations in continual updating. The benchmark enables comparative analysis across knowledge types.

Conclusion: MMKU-Bench provides a reliable and comprehensive evaluation framework for multimodal knowledge updating, advancing progress in the field by addressing both knowledge updating scenarios and cross-modal consistency.

Abstract: As real-world knowledge continues to evolve, the parametric knowledge acquired by multimodal models during pretraining becomes increasingly difficult to remain consistent with real-world knowledge. Existing research on multimodal knowledge updating focuses only on learning previously unknown knowledge, while overlooking the need to update knowledge that the model has already mastered but that later changes; moreover, evaluation is limited to the same modality, lacking a systematic analysis of cross-modal consistency. To address these issues, this paper proposes MMKU-Bench, a comprehensive evaluation benchmark for multimodal knowledge updating, which contains over 25k knowledge instances and more than 49k images, covering two scenarios, updated knowledge and unknown knowledge, thereby enabling comparative analysis of learning across different knowledge types. On this benchmark, we evaluate a variety of representative approaches, including supervised fine-tuning (SFT), reinforcement learning from human feedback (RLHF), and knowledge editing (KE). Experimental results show that SFT and RLHF are prone to catastrophic forgetting, while KE better preserve general capabilities but exhibit clear limitations in continual updating. Overall, MMKU-Bench provides a reliable and comprehensive evaluation benchmark for multimodal knowledge updating, advancing progress in this field.

Method: Created two multilingual corpora: InQA+ (small high-quality evaluation dataset with hand-annotated labels) and GenIQA (larger training dataset with artificial data generated by GPT-4o-mini). Conducted experiments with multilingual transformer models (mBERT, XLM-R, mDeBERTa) on English, Standard German, and Bavarian dialect. Analyzed factors like label ambiguity, label set, and dataset size.

Result: IQA performance was low even for English, with severe overfitting. Performance was poor across all languages (high-resource English/German and low-resource Bavarian). Larger training data was beneficial. GPT-4o-mini lacked sufficient pragmatic understanding to generate high-quality IQA data in any tested language.

Conclusion: IQA is a pragmatically challenging task requiring better understanding of indirect communication. Current models struggle significantly, and synthetic data generation with current LLMs is insufficient. More sophisticated approaches are needed for pragmatic language understanding.

Abstract: Indirectness is a common feature of daily communication, yet is underexplored in NLP research for both low-resource as well as high-resource languages. Indirect Question Answering (IQA) aims at classifying the polarity of indirect answers. In this paper, we present two multilingual corpora for IQA of varying quality that both cover English, Standard German and Bavarian, a German dialect without standard orthography: InQA+, a small high-quality evaluation dataset with hand-annotated labels, and GenIQA, a larger training dataset, that contains artificial data generated by GPT-4o-mini. We find that IQA is a pragmatically hard task that comes with various challenges, based on several experiment variations with multilingual transformer models (mBERT, XLM-R and mDeBERTa). We suggest and employ recommendations to tackle these challenges. Our results reveal low performance, even for English, and severe overfitting. We analyse various factors that influence these results, including label ambiguity, label set and dataset size. We find that the IQA performance is poor in high- (English, German) and low-resource languages (Bavarian) and that it is beneficial to have a large amount of training data. Further, GPT-4o-mini does not possess enough pragmatic understanding to generate high-quality IQA data in any of our tested languages.

Method: Introduces HS (time-split evaluation framework) that restricts idea generation to pre-T literature, then evaluates outputs against papers published in subsequent 30 months, scoring by citation impact and venue acceptance.

Result: Experiments across 10 AI/ML topics show retrieval-augmented system produces 2.5× higher-scoring ideas than vanilla generation (p<0.001), while LLM-as-Judge finds no significant difference; HS scores negatively correlated with LLM-judged novelty.

Conclusion: HS provides objective evaluation of AI-generated research ideas based on real-world impact, revealing LLMs systematically overvalue novel-sounding ideas that never materialize in actual research.

Abstract: Evaluating AI-generated research ideas typically relies on LLM judges or human panels – both subjective and disconnected from actual research impact. We introduce \hs{}, a time-split evaluation framework that measures idea quality by matching generated ideas against real future publications and scoring them by citation impact and venue acceptance. Using a temporal cutoff~$T$, we restrict an idea generation system to pre-$T$ literature, then evaluate its outputs against papers published in the subsequent 30 months. Experiments across 10 AI/ML research topics reveal a striking disconnect: LLM-as-Judge finds no significant difference between retrieval-augmented and vanilla idea generation ($p{=}0.584$), while \hs{} shows the retrieval-augmented system produces 2.5$\times$ higher-scoring ideas ($p{<}0.001$). Moreover, \hs{} scores are \emph{negatively} correlated with LLM-judged novelty ($ρ{=}{-}0.29$, $p{<}0.01$), suggesting that LLMs systematically overvalue novel-sounding ideas that never materialize in real research.

Method: Evaluated three rarely-studied English dialects (Yorkshire, Geordie, Cornish), plus African-American Vernacular English and West Frisian as control. Analyzed human-human agreement on LLM generation quality and its impact on LLM-as-a-judge performance, metrics like accuracy, and fine-tuning effects.

Result: Human-human agreement patterns directly affect LLM-human agreement and accuracy metrics. LLM-human agreement reflects alignment with human consensus, raising concerns about improving LLMs in locales with low population and low agreement. Fine-tuning doesn’t eliminate and may amplify this pattern. Some LLMs can generate high-quality dialect data, enabling scalability.

Conclusion: Data must be carefully evaluated for fair and inclusive LLM improvement. New tools are needed to handle the observed pattern in low-resource dialect scenarios where data scarcity and low human agreement pose challenges.

Abstract: It is known that large language models (LLMs) underperform in English dialects, and that improving them is difficult due to data scarcity. In this work we investigate how quality and availability impact the feasibility of improving LLMs in this context. For this, we evaluate three rarely-studied English dialects (Yorkshire, Geordie, and Cornish), plus African-American Vernacular English, and West Frisian as control. We find that human-human agreement when determining LLM generation quality directly impacts LLM-as-a-judge performance. That is, LLM-human agreement mimics the human-human agreement pattern, and so do metrics such as accuracy. It is an issue because LLM-human agreement measures an LLM’s alignment with the human consensus; and hence raises questions about the feasibility of improving LLM performance in locales where low populations induce low agreement. We also note that fine-tuning does not eradicate, and might amplify, this pattern in English dialects. But also find encouraging signals, such as some LLMs’ ability to generate high-quality data, thus enabling scalability. We argue that data must be carefully evaluated to ensure fair and inclusive LLM improvement; and, in the presence of scarcity, new tools are needed to handle the pattern found.

Method: Proposes Parallel-Token Prediction (PTP) with learnable tokens inserted into input sequences and corresponding training objectives to enable parallel decoding. Also develops a comprehensive data generation pipeline for training.

Result: Achieves 1.6x-2.2x decoding speed improvement on OmniDocBench and olmOCR-bench, reduces model hallucinations, and shows strong generalization abilities.

Conclusion: PTP is an effective plug-and-play method that significantly accelerates document parsing in VLMs while maintaining or improving accuracy.

Abstract: Document parsing, as a fundamental yet crucial vision task, is being revolutionized by vision-language models (VLMs). However, the autoregressive (AR) decoding inherent to VLMs creates a significant bottleneck, severely limiting parsing speed. In this paper, we propose Parallel-Token Prediction (PTP), a plugable, model-agnostic and simple-yet-effective method that enables VLMs to generate multiple future tokens in parallel with improved sample efficiency. Specifically, we insert some learnable tokens into the input sequence and design corresponding training objectives to equip the model with parallel decoding capabilities for document parsing. Furthermore, to support effective training, we develop a comprehensive data generation pipeline that efficiently produces large-scale, high-quality document parsing training data for VLMs. Extensive experiments on OmniDocBench and olmOCR-bench demonstrate that our method not only significantly improves decoding speed (1.6x-2.2x) but also reduces model hallucinations and exhibits strong generalization abilities.

Method: Created a bidirectional multi-domain dataset of 73,965 parallel passive sentence pairs from five Chinese-English corpora, with automatic structure labeling and manually verified test set, then evaluated both open-source and commercial MT systems.

Result: Models preserve source text voice rather than adapting to target language norms, show awareness of Chinese passive characteristics, with commercial NMT scoring higher on metrics but LLMs demonstrating better translation diversity.

Conclusion: The dataset enables better evaluation of MT systems on passive constructions, revealing systematic differences between human and machine translation patterns for this linguistic phenomenon.

Abstract: Machine Translation (MT) evaluation has gone beyond metrics, towards more specific linguistic phenomena. Regarding English-Chinese language pairs, passive sentences are constructed and distributed differently due to language variation, thus need special attention in MT. This paper proposes a bidirectional multi-domain dataset of passive sentences, extracted from five Chinese-English parallel corpora and annotated automatically with structure labels according to human translation, and a test set with manually verified annotation. The dataset consists of 73,965 parallel sentence pairs (2,358,731 English words, 3,498,229 Chinese characters). We evaluate two state-of-the-art open-source MT systems with our dataset, and four commercial models with the test set. The results show that, unlike humans, models are more influenced by the voice of the source text rather than the general voice usage of the source language, and therefore tend to maintain the passive voice when translating a passive in either direction. However, models demonstrate some knowledge of the low frequency and predominantly negative context of Chinese passives, leading to higher voice consistency with human translators in English-to-Chinese translation than in Chinese-to-English translation. Commercial NMT models scored higher in metric evaluations, but LLMs showed a better ability to use diverse alternative translations. Datasets and annotation script will be shared upon request.

Method: Built Lend an Ear platform where participants offer empathic support to LLMs role-playing personal/workplace troubles. Analyzed 33,938 messages from 2,904 conversations. Created taxonomy of empathic expressions. Conducted pre-registered randomized experiment comparing LLM coaching intervention (personalized feedback) vs. control vs. video-based non-personalized feedback.

Result: LLM coaching intervention significantly boosted alignment with normative empathic communication patterns compared to both control and video-based feedback groups. Found “silent empathy effect” where people feel empathy but fail to express it. Participants reliably identified normatively empathic responses as more expressive of empathy.

Conclusion: The study advances understanding of how empathy is expressed and valued, and demonstrates a scalable AI-based intervention for cultivating empathic communication skills.

Abstract: Empathy is central to human connection, yet people often struggle to express it effectively. In blinded evaluations, large language models (LLMs) generate responses that are often judged more empathic than human-written ones. Yet when a response is attributed to AI, recipients feel less heard and validated than when comparable responses are attributed to a human. To probe and address this gap in empathic communication skill, we built Lend an Ear, an experimental conversation platform in which participants are asked to offer empathic support to an LLM role-playing personal and workplace troubles. From 33,938 messages spanning 2,904 text-based conversations between 968 participants and their LLM conversational partners, we derive a data-driven taxonomy of idiomatic empathic expressions in naturalistic dialogue. Based on a pre-registered randomized experiment, we present evidence that a brief LLM coaching intervention offering personalized feedback on how to effectively communicate empathy significantly boosts alignment of participants’ communication patterns with normative empathic communication patterns relative to both a control group and a group that received video-based but non-personalized feedback. Moreover, we find evidence for a silent empathy effect that people feel empathy but systematically fail to express it. Nonetheless, participants reliably identify responses aligned with normative empathic communication criteria as more expressive of empathy. Together, these results advance the scientific understanding of how empathy is expressed and valued and demonstrate a scalable, AI-based intervention for scaffolding and cultivating it.

Method: Proposes Code-Centric Learning with mixed training strategy and code-centric data expansion, shifting supervision from full documents to short evidence spans to improve document-level ICD coding.

Result: Method substantially outperforms strong baselines under same LLM backbone, enables small-scale LLMs to achieve performance comparable to much larger proprietary models.

Conclusion: Demonstrates effectiveness and potential for fully automated ICD coding by improving accuracy on unseen codes, preserving interpretability, and reducing training cost.

Abstract: ICD coding is a critical yet challenging task in healthcare. Recently, LLM-based methods demonstrate stronger generalization than discriminative methods in ICD coding. However, fine-tuning LLMs for ICD coding faces three major challenges. First, existing public ICD coding datasets provide limited coverage of the ICD code space, restricting a model’s ability to generalize to unseen codes. Second, naive fine-tuning diminishes the interpretability of LLMs, as few public datasets contain explicit supporting evidence for assigned codes. Third, ICD coding typically involves long clinical documents, making fine-tuning LLMs computationally expensive. To address these issues, we propose Code-Centric Learning, a training framework that shifts supervision from full clinical documents to scalable, short evidence spans. The key idea of this framework is that span-level learning improves LLMs’ ability to perform document-level ICD coding. Our proposed framework consists of a mixed training strategy and code-centric data expansion, which substantially reduces training cost, improves accuracy on unseen ICD codes and preserves interpretability. Under the same LLM backbone, our method substantially outperforms strong baselines. Notably, our method enables small-scale LLMs to achieve performance comparable to much larger proprietary models, demonstrating its effectiveness and potential for fully automated ICD coding.

Method: Created curated paradigm-based datasets for four languages (English, German, Italian, Hebrew) focusing on verb alternations. Used Blackbird Language Matrices (BLMs) - RPM/ARC-like tasks designed specifically for language - where models must select sentences completing patterns according to syntactic/semantic rules. Applied linguistically-informed data augmentation across synthetic and natural data with three template complexity types.

Result: Provided simple baseline performance results across all four languages that demonstrate the diagnostic usefulness of the datasets for evaluating LLMs’ cross-sentence paradigmatic knowledge.

Conclusion: The paper introduces valuable diagnostic datasets for testing LLMs’ systematic cross-sentence knowledge of verb alternations, providing a controlled way to assess linguistic reasoning beyond single-sentence phenomena.

Abstract: Large language models (LLMs) have shown remarkable performance across various sentence-based linguistic phenomena, yet their ability to capture cross-sentence paradigmatic patterns, such as verb alternations, remains underexplored. In this work, we present curated paradigm-based datasets for four languages, designed to probe systematic cross-sentence knowledge of verb alternations (change-of-state and object-drop constructions in English, German and Italian, and Hebrew binyanim). The datasets comprise thousands of the Blackbird Language Matrices (BLMs) problems. The BLM task – an RPM/ARC-like task devised specifically for language – is a controlled linguistic puzzle where models must select the sentence that completes a pattern according to syntactic and semantic rules. We introduce three types of templates varying in complexity and apply linguistically-informed data augmentation strategies across synthetic and natural data. We provide simple baseline performance results across English, Italian, German, and Hebrew, that demonstrate the diagnostic usefulness of the datasets.

Method: Introduces CCTU benchmark with 200 test cases across 12 constraint categories in four dimensions (resource, behavior, toolset, response). Includes executable constraint validation module for step-level validation during multi-turn interactions. Evaluates 9 SOTA LLMs in thinking and non-thinking modes.

Result: No model achieves above 20% task completion rate when strict constraint adherence required. Models violate constraints in over 50% of cases, especially in resource and response dimensions. LLMs show limited self-refinement capacity even with detailed feedback on violations.

Conclusion: Current LLMs struggle significantly with constraint adherence in tool-use scenarios, revealing critical bottlenecks in developing robust tool-use agents. The benchmark enables future research in this challenging area.

Abstract: Solving problems through tool use under explicit constraints constitutes a highly challenging yet unavoidable scenario for large language models (LLMs), requiring capabilities such as function calling, instruction following, and self-refinement. However, progress has been hindered by the absence of dedicated evaluations. To address this, we introduce CCTU, a benchmark for evaluating LLM tool use under complex constraints. CCTU is grounded in a taxonomy of 12 constraint categories spanning four dimensions (i.e., resource, behavior, toolset, and response). The benchmark comprises 200 carefully curated and challenging test cases across diverse tool-use scenarios, each involving an average of seven constraint types and an average prompt length exceeding 4,700 tokens. To enable reliable evaluation, we develop an executable constraint validation module that performs step-level validation and enforces compliance during multi-turn interactions between models and their environments. We evaluate nine state-of-the-art LLMs in both thinking and non-thinking modes. Results indicate that when strict adherence to all constraints is required, no model achieves a task completion rate above 20%. Further analysis reveals that models violate constraints in over 50% of cases, particularly in the resource and response dimensions. Moreover, LLMs demonstrate limited capacity for self-refinement even after receiving detailed feedback on constraint violations, highlighting a critical bottleneck in the development of robust tool-use agents. To facilitate future research, we release the data and code.

Method: Develops a Python-based framework using Python’s built-in any() and all() functions to support both conjunctive (ALL) and disjunctive (ANY) conditions within single rules, with expressive exception-handling mechanisms, providing a flexible syntax for legal rules, conditions, and exceptions.

Result: PYTHEN offers enhanced flexibility compared to PROLEG by natively supporting both conjunctive and disjunctive conditions in single rules, with better exception handling, and serves as a practical bridge between symbolic reasoning and Python’s ecosystem for legal AI applications.

Conclusion: PYTHEN democratizes formal legal reasoning by leveraging Python’s accessibility while maintaining symbolic reasoning capabilities, positioning it as a valuable tool for next-generation legal AI systems and autoformalization applications.

Abstract: This paper introduces PYTHEN, a novel Python-based framework for defeasible legal reasoning. PYTHEN is designed to model the inherently defeasible nature of legal argumentation, providing a flexible and intuitive syntax for representing legal rules, conditions, and exceptions. Inspired by PROLEG (PROlog-based LEGal reasoning support system) and guided by the philosophy of The Zen of Python, PYTHEN leverages Python’s built-in any() and all() functions to offer enhanced flexibility by natively supporting both conjunctive (ALL) and disjunctive (ANY) conditions within a single rule, as well as a more expressive exception-handling mechanism. This paper details the architecture of PYTHEN, provides a comparative analysis with PROLEG, and discusses its potential applications in autoformalization and the development of next-generation legal AI systems. By bridging the gap between symbolic reasoning and the accessibility of Python, PYTHEN aims to democratize formal legal reasoning for young researchers, legal tech developers, and professionals without extensive logic programming expertise. We position PYTHEN as a practical bridge between the powerful symbolic reasoning capabilities of logic programming and the rich, ubiquitous ecosystem of Python, making formal legal reasoning accessible to a broader range of developers and legal professionals.

Method: Curated over 8,972 hours of podcast audio, applied audio pre-processing pipeline, used mixed transcription strategy with ASR models trained on high-fidelity API transcriptions.

Result: Created a dataset rivaling English’s GigaSpeech in scale, trained ASR and TTS models exclusively on TAGARELA demonstrating its effectiveness for Portuguese speech technologies.

Conclusion: TAGARELA addresses Portuguese speech data scarcity, enables state-of-the-art models, and is publicly released to foster robust Portuguese speech technology development.

Abstract: Despite significant advances in speech processing, Portuguese remains under-resourced due to the scarcity of public, large-scale, and high-quality datasets. To address this gap, we present a new dataset, named TAGARELA, composed of over 8,972 hours of podcast audio, specifically curated for training automatic speech recognition (ASR) and text-to-speech (TTS) models. Notably, its scale rivals English’s GigaSpeech (10kh), enabling state-of-the-art Portuguese models. To ensure data quality, the corpus was subjected to an audio pre-processing pipeline and subsequently transcribed using a mixed strategy: we applied ASR models that were previously trained on high-fidelity transcriptions generated by proprietary APIs, ensuring a high level of initial accuracy. Finally, to validate the effectiveness of this new resource, we present ASR and TTS models trained exclusively on our dataset and evaluate their performance, demonstrating its potential to drive the development of more robust and natural speech technologies for Portuguese. The dataset is released publicly, available at https://freds0.github.io/TAGARELA/, to foster the development of robust speech technologies.

Method: Proposes Dependency-Oriented Sampler (DOS) that uses attention matrices from transformer blocks to approximate inter-token dependencies, emphasizing information from unmasked tokens when updating masked positions during generation.

Result: DOS consistently achieves superior performance on code generation and mathematical reasoning tasks, and can be integrated with existing parallel sampling methods to improve efficiency without sacrificing quality.

Conclusion: DOS effectively addresses the limitation of overlooking sequence-level information in MDLM decoding, offering improved generation quality and efficiency through attention-based dependency modeling.

Abstract: Masked diffusion language models (MDLMs) have recently emerged as a new paradigm in language modeling, offering flexible generation dynamics and enabling efficient parallel decoding. However, existing decoding strategies for pre-trained MDLMs predominantly rely on token-level uncertainty criteria, while largely overlooking sequence-level information and inter-token dependencies. To address this limitation, we propose Dependency-Oriented Sampler (DOS), a training-free decoding strategy that leverages inter-token dependencies to inform token updates during generation. Specifically, DOS exploits attention matrices from transformer blocks to approximate inter-token dependencies, emphasizing information from unmasked tokens when updating masked positions. Empirical results demonstrate that DOS consistently achieves superior performance on both code generation and mathematical reasoning tasks. Moreover, DOS can be seamlessly integrated with existing parallel sampling methods, leading to improved generation efficiency without sacrificing generation quality.

Method: Investigate two sources of sparsity: (1) implicit sparsity from training/data conditions (weight sparsity via weight decay, attention sparsity via long context), and (2) explicit sparsity from architectural design (Grouped-Query Attention key/value-sharing, Mixture-of-Experts expert-activation). Use controlled depth-scaling experiments and layer effectiveness interventions.

Result: Sparsity consistently improves layer utilization by reducing output variance and promoting functional differentiation. Distilled into practical training recipe yielding 4.6% accuracy improvement on downstream tasks.

Conclusion: Sparsity, arising naturally from standard design choices, is a key previously overlooked mechanism for effective depth scaling in LLMs.

Abstract: Recent work has demonstrated the curse of depth in large language models (LLMs), where later layers contribute less to learning and representation than earlier layers. Such under-utilization is linked to the accumulated growth of variance in Pre-Layer Normalization, which can push deep blocks toward near-identity behavior. In this paper, we demonstrate that, sparsity, beyond enabling efficiency, acts as a regulator of variance propagation and thereby improves depth utilization. Our investigation covers two sources of sparsity: (i) implicit sparsity, which emerges from training and data conditions, including weight sparsity induced by weight decay and attention sparsity induced by long context inputs; and (ii) explicit sparsity, which is enforced by architectural design, including key/value-sharing sparsity in Grouped-Query Attention and expert-activation sparsity in Mixtureof-Experts. Our claim is thoroughly supported by controlled depth-scaling experiments and targeted layer effectiveness interventions. Across settings, we observe a consistent relationship: sparsity improves layer utilization by reducing output variance and promoting functional differentiation. We eventually distill our findings into a practical rule-of-thumb recipe for training deptheffective LLMs, yielding a notable 4.6% accuracy improvement on downstream tasks. Our results reveal sparsity, arising naturally from standard design choices, as a key yet previously overlooked mechanism for effective depth scaling in LLMs. Code is available at https://github.com/pUmpKin-Co/SparsityAndCoD.

Method: Analyzed 16 LLMs (general LLMs, specialist tabular LLMs, and Mixture-of-Experts models) across 4 dimensions: attention dynamics, effective layer depth, expert activation, and impacts of input designs.

Result: Key findings: (1) Three-phase attention pattern (early: broad scanning, middle: localizing relevant cells, late: amplifying contributions); (2) Tabular tasks require deeper layers than math reasoning; (3) MoE models activate table-specific experts in middle layers; (4) Chain-of-Thought increases table attention, enhanced by table-tuning.

Conclusion: The findings provide interpretability insights into how LLMs process tabular data, with implications for improving table understanding models and facilitating future research on table-related tasks.

Abstract: Despite the success of Large Language Models (LLMs) in table understanding, their internal mechanisms remain unclear. In this paper, we conduct an empirical study on 16 LLMs, covering general LLMs, specialist tabular LLMs, and Mixture-of-Experts (MoE) models, to explore how LLMs understand tabular data and perform downstream tasks. Our analysis focus on 4 dimensions including the attention dynamics, the effective layer depth, the expert activation, and the impacts of input designs. Key findings include: (1) LLMs follow a three-phase attention pattern – early layers scan the table broadly, middle layers localize relevant cells, and late layers amplify their contributions; (2) tabular tasks require deeper layers than math reasoning to reach stable predictions; (3) MoE models activate table-specific experts in middle layers, with early and late layers sharing general-purpose experts; (4) Chain-of-Thought prompting increases table attention, further enhanced by table-tuning. We hope these findings and insights can facilitate interpretability and future research on table-related tasks.

Method: Two-stage approach: 1) Trajectory Collection - capture dynamic evolution of personality adoption during SFT by saving sequence of LoRA adapters to map continuous trait manifold; 2) RL-based Dynamic Fusion - train policy network using RL to compute mixing weights for frozen adapters, sampling from Dirichlet distribution to fuse adapters for specific numerical target intensity.

Result: Experiments on Qwen3-14B model show Fusian achieves high precision in personality control, significantly outperforming baseline methods in aligning with user-specified trait intensities.

Conclusion: Fusian enables fine-grained continuous personality control in LLMs, moving beyond discrete trait categories to precise intensity control through dynamic fusion of learned adapters.

Abstract: Large Language Models (LLMs) have demonstrated impressive capabilities in simulating diverse human behaviors and personalities. However, existing methods for personality control, which include prompt engineering and standard Supervised Fine-Tuning (SFT), typically treat personality traits as discrete categories (e.g., “Extroverted” vs. “Introverted”), lacking the ability to precisely control the intensity of a trait on a continuous spectrum. In this paper, we introduce Fusian, a novel framework for fine-grained, continuous personality control in LLMs. Fusian operates in two stages: (1) Trajectory Collection, where we capture the dynamic evolution of personality adoption during SFT by saving a sequence of LoRA adapters, effectively mapping the continuous manifold of a trait; and (2) RL-based Dynamic Fusion, where we train a policy network using Reinforcement Learning to dynamically compute mixing weights for these frozen adapters. By sampling from a Dirichlet distribution parameterized by the policy network, Fusian fuses multiple adapters to align the model’s output with a specific numerical target intensity. Experiments on the Qwen3-14B model demonstrate that Fusian achieves high precision in personality control, significantly outperforming baseline methods in aligning with user-specified trait intensities.

Method: Created SEA-Vision benchmark with 15,234 document parsing pages (9 document types) and 7,496 TEC-VQA QA pairs across 11 languages. Used hybrid pipeline combining automated filtering/scoring with MLLM-assisted labeling and native-speaker verification.

Result: Evaluation of leading multimodal models shows pronounced performance degradation on low-resource Southeast Asian languages, highlighting substantial gaps in multilingual document and scene text understanding.

Conclusion: SEA-Vision addresses critical gaps in multilingual evaluation and will help drive global progress in document and scene text understanding, particularly for low-resource languages.

Abstract: Multilingual document and scene text understanding plays an important role in applications such as search, finance, and public services. However, most existing benchmarks focus on high-resource languages and fail to evaluate models in realistic multilingual environments. In Southeast Asia, the diversity of languages, complex writing systems, and highly varied document types make this challenge even greater. We introduce SEA-Vision, a benchmark that jointly evaluates Document Parsing and Text-Centric Visual Question Answering (TEC-VQA) across 11 Southeast Asian languages. SEA-Vision contains 15,234 document parsing pages from nine representative document types, annotated with hierarchical page-, block-, and line-level labels. It also provides 7,496 TEC-VQA question-answer pairs that probe text recognition, numerical calculation, comparative analysis, logical reasoning, and spatial understanding. To make such multilingual, multi-task annotation feasible, we design a hybrid pipeline for Document Parsing and TEC-VQA. It combines automated filtering and scoring with MLLM-assisted labeling and lightweight native-speaker verification, greatly reducing manual labeling while maintaining high quality. We evaluate several leading multimodal models and observe pronounced performance degradation on low-resource Southeast Asian languages, highlighting substantial remaining gaps in multilingual document and scene text understanding. We believe SEA-Vision will help drive global progress in document and scene text understanding.

Method: CLAG uses an SLM-driven router to assign incoming memories to semantically coherent clusters, autonomously generates cluster-specific profiles (topic summaries and tags), performs localized evolution within clusters to reduce cross-topic interference, and employs two-stage retrieval that first filters relevant clusters via profiles.

Result: Experiments on multiple QA datasets with three SLM backbones show CLAG consistently improves answer quality and robustness over prior memory systems while remaining lightweight and efficient.

Conclusion: CLAG provides an effective memory organization framework for SLM agents that reduces interference, enhances memory density, and improves retrieval performance through semantic clustering and structured memory management.

Abstract: Large language model agents heavily rely on external memory to support knowledge reuse and complex reasoning tasks. Yet most memory systems store experiences in a single global retrieval pool which can gradually dilute or corrupt stored knowledge. This problem is especially pronounced for small language models (SLMs), which are highly vulnerable to irrelevant context. We introduce CLAG, a CLustering-based AGentic memory framework where an SLM agent actively organizes memory by clustering. CLAG employs an SLM-driven router to assign incoming memories to semantically coherent clusters and autonomously generates cluster-specific profiles, including topic summaries and descriptive tags, to establish each cluster as a self-contained functional unit. By performing localized evolution within these structured neighborhoods, CLAG effectively reduces cross-topic interference and enhances internal memory density. During retrieval, the framework utilizes a two-stage process that first filters relevant clusters via their profiles, thereby excluding distractors and reducing the search space. Experiments on multiple QA datasets with three SLM backbones show that CLAG consistently improves answer quality and robustness over prior memory systems for agents, remaining lightweight and efficient.

Method: Large-scale quantitative analysis of human-AI interactions from the WildChat dataset, identifying invisible failures, clustering them into archetypes, analyzing co-occurrence patterns, and assessing whether failures are interactional vs capability-driven.

Result: Found 78% of AI failures are invisible to users, categorized into 8 archetypes with systematic co-occurrence patterns. 91% of failures involve interactional dynamics, and 94% of these would persist even with more capable models.

Conclusion: Invisible failure taxonomy provides crucial insights for reliable failure monitoring across product development, research, and policy. Most failures are interactional rather than capability-based, suggesting persistent challenges in human-AI interaction design.

Abstract: AI systems fail silently far more often than they fail visibly. In a large-scale quantitative analysis of human-AI interactions from the WildChat dataset, we find that 78% of AI failures are invisible: something went wrong but the user gave no overt indication that there was a problem. These invisible failures cluster into eight archetypes that help us characterize where and how AI systems are failing to meet users’ needs. In addition, the archetypes show systematic co-occurrence patterns indicating higher-level failure types. To address the question of whether these archetypes will remain relevant as AI systems become more capable, we also assess failures for whether they are primarily interactional or capability-driven, finding that 91% involve interactional dynamics, and we estimate that 94% of such failures would persist even with a more capable model. Finally, we illustrate how the archetypes help us to identify systematic and variable AI limitations across different usage domains. Overall, we argue that our invisible failure taxonomy can be a key component in reliable failure monitoring for product developers, scientists, and policy makers. Our code and data are available at https://github.com/bigspinai/bigspin-invisible-failure-archetypes

Method: Created ViX-Ray dataset with 5,400 Vietnamese chest X-ray images annotated by hospital physicians, analyzed linguistic patterns, and fine-tuned five open-source VLMs comparing them to GPT-4V and Gemini.

Result: Models generate outputs partially aligned with clinical ground truths but suffer from low precision and excessive hallucination, especially in impression generation, demonstrating dataset complexity.

Conclusion: ViX-Ray establishes a valuable benchmark for evaluating and advancing vision-language models in Vietnamese clinical domain, highlighting current limitations in medical AI for specific languages.

Abstract: Vietnamese medical research has become an increasingly vital domain, particularly with the rise of intelligent technologies aimed at reducing time and resource burdens in clinical diagnosis. Recent advances in vision-language models (VLMs), such as Gemini and GPT-4V, have sparked a growing interest in applying AI to healthcare. However, most existing VLMs lack exposure to Vietnamese medical data, limiting their ability to generate accurate and contextually appropriate diagnostic outputs for Vietnamese patients. To address this challenge, we introduce ViX-Ray, a novel dataset comprising 5,400 Vietnamese chest X-ray images annotated with expert-written findings and impressions from physicians at a major Vietnamese hospital. We analyze linguistic patterns within the dataset, including the frequency of mentioned body parts and diagnoses, to identify domain-specific linguistic characteristics of Vietnamese radiology reports. Furthermore, we fine-tune five state-of-the-art open-source VLMs on ViX-Ray and compare their performance to leading proprietary models, GPT-4V and Gemini. Our results show that while several models generate outputs partially aligned with clinical ground truths, they often suffer from low precision and excessive hallucination, especially in impression generation. These findings not only demonstrate the complexity and challenge of our dataset but also establish ViX-Ray as a valuable benchmark for evaluating and advancing vision-language models in the Vietnamese clinical domain.

Method: RoSE (Robust Same-subject Editing) employs Isotropic Geometric Alignment to minimize representational deviation and Hierarchical Knowledge Integration to smooth the optimization landscape, addressing the geometric root of generalization collapse.

Result: Extensive experiments demonstrate that RoSE significantly improves instruction-following capabilities and lays the foundation for robust interactive parametric memory of LLM agents.

Conclusion: RoSE successfully addresses the geometric instability in same-subject knowledge editing, enabling more robust knowledge updates that generalize across different prompt formulations and user instructions.

Abstract: While locate-then-edit knowledge editing efficiently updates knowledge encoded within Large Language Models (LLMs), a critical generalization failure mode emerges in the practical same-subject knowledge editing scenario: models fail to recall the updated knowledge when following user instructions, despite successfully recalling it in the original edited form. This paper identifies the geometric root of this generalization collapse as a fundamental conflict where the inner activation drifts induced by prompt variations exceed the model’s geometric tolerance for generalization after editing. We attribute this instability to a dual pathology: (1) The joint optimization with orthogonal gradients collapses solutions into sharp minima with narrow stability, and (2) the standard covariance constraint paradoxically acts as a Covariance Trap that amplifies input perturbations. To resolve this, we introduce RoSE (Robust Same-subject Editing), which employs Isotropic Geometric Alignment to minimize representational deviation and Hierarchical Knowledge Integration to smooth the optimization landscape. Extensive experiments demonstrate that RoSE significantly improves instruction-following capabilities, laying the foundation for robust interactive parametric memory of LLM agents.

Method: Built SlovakKE dataset (227,432 scientific abstracts with author keyphrases) from Slovak Central Register of Theses, benchmarked three unsupervised methods (YAKE, TextRank, KeyBERT) and LLM-based KeyLLM using GPT-3.5-turbo, with manual evaluation on 100 documents.

Result: Unsupervised baselines achieved at most 11.6% exact-match F1@6 (vs 51.5% partial matching), showing difficulty with inflected forms. KeyLLM narrowed the exact-partial gap and captured relevant concepts that automated matching missed, with morphological mismatch identified as dominant failure mode.

Conclusion: Created valuable resource for Slovak NLP, demonstrated LLMs’ advantage in handling morphological variation for keyphrase extraction in inflected languages, with findings applicable to other morphologically rich languages.

Abstract: Keyphrase extraction for morphologically rich, low-resource languages remains understudied, largely due to the scarcity of suitable evaluation datasets. We address this gap for Slovak by constructing a dataset of 227,432 scientific abstracts with author-assigned keyphrases – scraped and systematically cleaned from the Slovak Central Register of Theses – representing a 25-fold increase over the largest prior Slovak resource and approaching the scale of established English benchmarks such as KP20K. Using this dataset, we benchmark three unsupervised baselines (YAKE, TextRank, KeyBERT with SlovakBERT embeddings) and evaluate KeyLLM, an LLM-based extraction method using GPT-3.5-turbo. Unsupervised baselines achieve at most 11.6% exact-match $F1@6$, with a large gap to partial matching (up to 51.5%), reflecting the difficulty of matching inflected surface forms to author-assigned keyphrases. KeyLLM narrows this exact–partial gap, producing keyphrases closer to the canonical forms assigned by authors, while manual evaluation on 100 documents ($κ= 0.61$) confirms that KeyLLM captures relevant concepts that automated exact matching underestimates. Our analysis identifies morphological mismatch as the dominant failure mode for statistical methods – a finding relevant to other inflected languages. The dataset (https://huggingface.co/datasets/NaiveNeuron/SlovKE) and evaluation code (https://github.com/NaiveNeuron/SlovKE) are publicly available.

Method: Introduce a taxonomy for analyzing LLM strategies, examine reasoning procedures, compare to learning sciences best practices, analyze failure modes, and test impact of providing correct solutions in prompts.

Result: LLMs typically follow best practices: solve correctly first, articulate/simulate multiple misconceptions, then select distractors. Errors arise from solution recovery and candidate selection failures. Providing correct solutions improves alignment with human distractors by 8%.

Conclusion: LLMs demonstrate structured ability to model incorrect student reasoning for distractor generation, with solution anchoring being critical. Analysis offers interpretable lens into LLMs’ misconception modeling capabilities.

Abstract: Modeling plausible student misconceptions is critical for AI in education. In this work, we examine how large language models (LLMs) reason about misconceptions when generating multiple-choice distractors, a task that requires modeling incorrect yet plausible answers by coordinating solution knowledge, simulating student misconceptions, and evaluating plausibility. We introduce a taxonomy for analyzing the strategies used by state-of-the-art LLMs, examining their reasoning procedures and comparing them to established best practices in the learning sciences. Our structured analysis reveals a surprising alignment between their processes and best practices: the models typically solve the problem correctly first, then articulate and simulate multiple potential misconceptions, and finally select a set of distractors. An analysis of failure modes reveals that errors arise primarily from failures in recovering the correct solution and selecting among response candidates, rather than simulating errors or structuring the process. Consistent with these results, we find that providing the correct solution in the prompt improves alignment with human-authored distractors by 8%, highlighting the critical role of anchoring to the correct solution when generating plausible incorrect student reasoning. Overall, our analysis offers a structured and interpretable lens into LLMs’ ability to model incorrect student reasoning and produce high-quality distractors.

Method: Introduces Code-A1 with architectural separation: Code LLM is rewarded for passing more tests, Test LLM is rewarded for exposing more defects. Enables white-box test generation where Test LLM can inspect candidate code. Includes Mistake Book mechanism for experience replay and composite reward balancing test validity with adversarial difficulty.

Result: Experiments on Qwen2.5-Coder models show Code-A1 achieves code generation performance matching or exceeding models trained on human-annotated tests, while significantly improving test generation capability.

Conclusion: Code-A1’s adversarial co-evolution framework effectively addresses self-collusion risks in code generation, enabling robust test generation and improved code quality through competitive optimization of specialized models.

Abstract: Reinforcement learning for code generation relies on verifiable rewards from unit test pass rates. Yet high-quality test suites are scarce, existing datasets offer limited coverage, and static rewards fail to adapt as models improve. Recent self-play methods unify code and test generation in a single model, but face a inherent dilemma: white-box access leads to self-collusion where the model produces trivial tests for easy rewards, yet black-box restriction yields generic tests that miss implementation-specific bugs. We introduce Code-A1, an adversarial co-evolution framework that jointly optimizes a Code LLM and a Test LLM with opposing objectives. The Code LLM is rewarded for passing more tests, while the Test LLM is rewarded for exposing more defects. This architectural separation eliminates self-collusion risks and safely enables white-box test generation, where the Test LLM can inspect candidate code to craft targeted adversarial tests. We further introduce a Mistake Book mechanism for experience replay and a composite reward balancing test validity with adversarial difficulty. Experiments on Qwen2.5-Coder models demonstrate that Code-A1 achieves code generation performance matching or exceeding models trained on human-annotated tests, while significantly improving test generation capability.

Method: 1) Analyzed 23 models using 251k moral vectors based on Prototype Theory and Social-Chemistry-101 dataset; 2) Used Sparse Autoencoders on Qwen3-8B to isolate mono-semantic moral features; 3) Reconstructed topological relationships to align with ground-truth moral vectors for representational alignment.

Result: Found that current LLMs fail to distinguish opposed moral categories and fine-grained typicality gradients, unaffected by model scaling or alignment. Representational alignment improved moral reasoning with 75% pairwise win-rate on adversarial Flames benchmark.

Conclusion: Current intervention methods are remedial; endogenous AI alignment requires transformation from post-hoc corrections to proactive cultivation of moral representations.

Abstract: Existing behavioral alignment techniques for Large Language Models (LLMs) often neglect the discrepancy between surface compliance and internal unaligned representations, leaving LLMs vulnerable to long-tail risks. More crucially, we posit that LLMs possess an inherent state of moral indifference due to compressing distinct moral concepts into uniform probability distributions. We verify and remedy this indifference in LLMs’ latent representations, utilizing 251k moral vectors constructed upon Prototype Theory and the Social-Chemistry-101 dataset. Firstly, our analysis across 23 models reveals that current LLMs fail to represent the distinction between opposed moral categories and fine-grained typicality gradients within these categories; notably, neither model scaling, architecture, nor explicit alignment reshapes this indifference. We then employ Sparse Autoencoders on Qwen3-8B, isolate mono-semantic moral features, and targetedly reconstruct their topological relationships to align with ground-truth moral vectors. This representational alignment naturally improves moral reasoning and granularity, achieving a 75% pairwise win-rate on the independent adversarial Flames benchmark. Finally, we elaborate on the remedial nature of current intervention methods from an experientialist philosophy, arguing that endogenously aligned AI might require a transformation from post-hoc corrections to proactive cultivation.

Method: Introduces mixture-of-depths attention (MoDA) that enables each attention head to attend to sequence KV pairs at current layer and depth KV pairs from preceding layers. Includes hardware-efficient algorithm resolving non-contiguous memory-access patterns.

Result: MoDA achieves 97.3% of FlashAttention-2’s efficiency at 64K sequence length. On 1.5B-parameter models, improves average perplexity by 0.2 across 10 validation benchmarks and increases average performance by 2.11% on 10 downstream tasks with only 3.7% FLOPs overhead.

Conclusion: MoDA is a promising primitive for depth scaling in LLMs, particularly effective when combined with post-norm rather than pre-norm.

Abstract: Scaling depth is a key driver for large language models (LLMs). Yet, as LLMs become deeper, they often suffer from signal degradation: informative features formed in shallow layers are gradually diluted by repeated residual updates, making them harder to recover in deeper layers. We introduce mixture-of-depths attention (MoDA), a mechanism that allows each attention head to attend to sequence KV pairs at the current layer and depth KV pairs from preceding layers. We further describe a hardware-efficient algorithm for MoDA that resolves non-contiguous memory-access patterns, achieving 97.3% of FlashAttention-2’s efficiency at a sequence length of 64K. Experiments on 1.5B-parameter models demonstrate that MoDA consistently outperforms strong baselines. Notably, it improves average perplexity by 0.2 across 10 validation benchmarks and increases average performance by 2.11% on 10 downstream tasks, with a negligible 3.7% FLOPs computational overhead. We also find that combining MoDA with post-norm yields better performance than using it with pre-norm. These results suggest that MoDA is a promising primitive for depth scaling. Code is released at https://github.com/hustvl/MoDA .

Method: Extracts two forms of information from next-state signals: 1) evaluative signals as scalar rewards via PRM judge, and 2) directive signals through Hindsight-Guided On-Policy Distillation (OPD) that extracts textual hints to construct enhanced teacher context with token-level directional advantage supervision. Uses asynchronous design where model serves requests, PRM judges interactions, and trainer updates policy simultaneously.

Result: Enables agents to improve simply by being used, recovering conversational signals from user re-queries, corrections, and explicit feedback. Supports scalable RL across terminal, GUI, SWE, and tool-call settings with process rewards.

Conclusion: Next-state signals are universal learning sources that can train the same policy across diverse interaction types, with the proposed framework enabling continuous online improvement through both evaluative and directive supervision.

Abstract: Every agent interaction generates a next-state signal, namely the user reply, tool output, terminal or GUI state change that follows each action, yet no existing agentic RL system recovers it as a live, online learning source. We present OpenClaw-RL, a framework built on a simple observation: next-state signals are universal, and policy can learn from all of them simultaneously. Personal conversations, terminal executions, GUI interactions, SWE tasks, and tool-call traces are not separate training problems. They are all interactions that can be used to train the same policy in the same loop. Next-state signals encode two forms of information: evaluative signals, which indicate how well the action performed and are extracted as scalar rewards via a PRM judge; and directive signals, which indicate how the action should have been different and are recovered through Hindsight-Guided On-Policy Distillation (OPD). We extract textual hints from the next state, construct an enhanced teacher context, and provide token-level directional advantage supervision that is richer than any scalar reward. Due to the asynchronous design, the model serves live requests, the PRM judges ongoing interactions, and the trainer updates the policy at the same time, with zero coordination overhead between them. Applied to personal agents, OpenClaw-RL enables an agent to improve simply by being used, recovering conversational signals from user re-queries, corrections, and explicit feedback. Applied to general agents, the same infrastructure supports scalable RL across terminal, GUI, SWE, and tool-call settings, where we additionally demonstrate the utility of process rewards. Code: https://github.com/Gen-Verse/OpenClaw-RL

Method: Develops a category theory framework to generalize and describe ICL and LLM behavior. Uses this formal framework to analyze task agnosticity and equivalence of meta-prompting approaches, complemented by experimental validation.

Result: The framework enables formal results about task agnosticity and equivalence of meta-prompting methods. Experimental results demonstrate that meta-prompting generates more desirable outputs than basic prompting.

Conclusion: Category theory provides a rigorous foundation for understanding LLM behavior with prompts. Meta-prompting is formally shown to be more effective than basic prompting, offering theoretical grounding for prompt engineering practices.

Abstract: Modern large language models (LLMs) are capable of interpreting input strings as instructions, or prompts, and carry out tasks based on them. Unlike traditional learners, LLMs cannot use back-propagation to obtain feedback, and condition their output in situ in a phenomenon known as in-context learning (ICL). Many approaches to prompting and pre-training these models involve the automated generation of these prompts, also known as meta-prompting, or prompting to obtain prompts. However, they do not formally describe the properties and behavior of the LLMs themselves. We propose a theoretical framework based on category theory to generalize and describe ICL and LLM behavior when interacting with users. Our framework allows us to obtain formal results around task agnosticity and equivalence of various meta-prompting approaches. Using our framework and experimental results we argue that meta-prompting is more effective than basic prompting at generating desirable outputs.

Method: Developed Ayn, an 88M parameter TLM pretrained from scratch for 185 A100 hours on Indian legal domain with domain-specific tokenizer. Compared against LLMs ranging from 1B to 8B parameters on legal case judgment prediction, summarization, and general tasks.

Result: Ayn outperformed LLMs up to 80 times larger on legal case judgment prediction, rivaled LLMs up to 30 times larger on summarization, and remained competitive with larger LLMs on general tasks.

Conclusion: Domain-specific TLMs can effectively replace much larger LLMs for specialized tasks, offering cost-effective alternatives while maintaining competitive performance on general tasks.

Abstract: Decoder-only Large Language Models (LLMs) are currently the model of choice for many Natural Language Processing (NLP) applications. Through instruction fine-tuning and prompting approaches, such LLMs have been efficiently used to solve both general and domain-specific tasks. However, they are costly to train and, to a certain extent, costly to use as well, and one can wonder whether LLMs can be replaced by domain-specific Tiny Language Models (TLMs), which typically contain less than 100M parameters. We address this question in this study by comparing the performance of an 88M TLM pretrained from scratch for 185 A100 hours on a specific domain with a domain-specific tokenizer (here, the Indian legal domain) with LLMs of various sizes between 1B and 8B for solving domain-specific tasks. We show in particular that our legal TLM, Ayn, can indeed outperform LLMs up to 80 times larger on the legal case judgment prediction task, rival LLMs up to 30 times larger on the summarization task, and still be competitive with these larger LLMs on general tasks.

Method: Constructed ViWikiFC corpus by extracting evidence sentences from Vietnamese Wikipedia articles and converting them into claims. Analyzed corpus through linguistic aspects (dependency rate, n-gram rate, word rate). Conducted experiments with BM25 for evidence retrieval and InfoXLM (Large) for verdict prediction, plus pipeline approaches.

Result: BM25 achieved 88.30% accuracy for SUPPORTS, 86.93% for REFUTES, but only 56.67% for NEI in evidence retrieval. InfoXLM (Large) achieved 86.51% F1 score for verdict prediction. Pipeline approach with both models achieved only 67.00% strict accuracy, showing dataset challenges.

Conclusion: ViWikiFC is a challenging dataset for Vietnamese fact-checking, demonstrating the need for specialized resources and models for low-resource languages. The gap between component performance and pipeline results highlights complexity of end-to-end fact verification.

Abstract: Fact-checking is essential due to the explosion of misinformation in the media ecosystem. Although false information exists in every language and country, most research to solve the problem mainly concentrated on huge communities like English and Chinese. Low-resource languages like Vietnamese are necessary to explore corpora and models for fact verification. To bridge this gap, we construct ViWikiFC, the first manual annotated open-domain corpus for Vietnamese Wikipedia Fact Checking more than 20K claims generated by converting evidence sentences extracted from Wikipedia articles. We analyze our corpus through many linguistic aspects, from the new dependency rate, the new n-gram rate, and the new word rate. We conducted various experiments for Vietnamese fact-checking, including evidence retrieval and verdict prediction. BM25 and InfoXLM (Large) achieved the best results in two tasks, with BM25 achieving an accuracy of 88.30% for SUPPORTS, 86.93% for REFUTES, and only 56.67% for the NEI label in the evidence retrieval task, InfoXLM (Large) achieved an F1 score of 86.51%. Furthermore, we also conducted a pipeline approach, which only achieved a strict accuracy of 67.00% when using InfoXLM (Large) and BM25. These results demonstrate that our dataset is challenging for the Vietnamese language model in fact-checking tasks.

Method: Analyzed pitch contours of two-character words using Generative Additive Mixed Models (GAMM) to examine F0 contours as function of normalized time, considering gender, duration, word position, bigram probability, neighboring tones, speaker, and novel predictors (word and word sense).

Result: In spontaneous Taiwan Mandarin, T3-T3 words become indistinguishable from T2-T3 words once word/word sense effects are accounted for, indicating complete sandhi rather than incomplete assimilation.

Conclusion: Tone 3 sandhi in spontaneous Taiwan Mandarin is complete when considering word-level contextual factors, contrasting with previous findings of incomplete sandhi in controlled speech environments.

Abstract: In Standard Chinese, Tone 3 (the dipping tone) becomes Tone 2 (rising tone) when followed by another Tone 3. Previous studies have noted that this sandhi process may be incomplete, in the sense that the assimilated Tone 3 is still distinct from a true Tone 2. While Mandarin Tone 3 sandhi is widely studied using carefully controlled laboratory speech (Xu 1997) and more formal registers of Beijing Mandarin (Yuan and Y. Chen 2014), less is known about its realization in spontaneous speech, and about the effect of contextual factors on tonal realization. The present study investigates the pitch contours of two-character words with T2-T3 and T3-T3 tone patterns in spontaneous Taiwan Mandarin conversations. Our analysis makes use of the Generative Additive Mixed Model (GAMM, Wood 2017) to examine fundamental frequency (F0) contours as a function of normalized time. We consider various factors known to influence pitch contours, including gender, duration, word position, bigram probability, neighboring tones, speaker, and also novel predictors, word and word sense (Chuang, Bell, Tseng, and Baayen 2025). Our analyses revealed that in spontaneous Taiwan Mandarin, T3-T3 words become indistinguishable from T2-T3 words, indicating complete sandhi, once the strong effect of word (or word sense) is taken into account.

Method: Proposes CausalDANN approach that uses LLMs to facilitate text transformations and leverages text-level classifiers with domain adaptation ability to produce robust effect estimates against domain shifts, even with only control group data.

Result: The method accommodates arbitrary textual interventions and provides robust causal effect estimates, addressing limitations of traditional binary/discrete treatment methods for complex text data.

Conclusion: CausalDANN advances causal estimation for textual data, enabling better understanding of human behaviors and development of effective interventions in social systems through flexible handling of various text interventions.

Abstract: Quantifying the effects of textual interventions in social systems, such as reducing anger in social media posts to see its impact on engagement, is challenging. Real-world interventions are often infeasible, necessitating reliance on observational data. Traditional causal inference methods, typically designed for binary or discrete treatments, are inadequate for handling the complex, high-dimensional textual data. This paper addresses these challenges by proposing CausalDANN, a novel approach to estimate causal effects using text transformations facilitated by large language models (LLMs). Unlike existing methods, our approach accommodates arbitrary textual interventions and leverages text-level classifiers with domain adaptation ability to produce robust effect estimates against domain shifts, even when only the control group is observed. This flexibility in handling various text interventions is a key advancement in causal estimation for textual data, offering opportunities to better understand human behaviors and develop effective interventions within social systems.

Method: Proposed ArithmAttack adds extra punctuation marks as noise to math problem prompts without adding or deleting words. Evaluated eight LLMs (LLama3, Mistral, Mathstral, DeepSeek, etc.) on noisy GSM8K and MultiArith datasets.

Result: All studied models showed vulnerability to punctuation noise, with performance degrading as noise increased. The attack is easy to implement and doesn’t cause information loss since words remain intact.

Conclusion: LLMs are vulnerable to simple punctuation noise attacks in math problem-solving contexts, highlighting robustness issues that need addressing despite their impressive capabilities.

Abstract: While Large Language Models (LLMs) have shown impressive capabilities in math problem-solving tasks, their robustness to noisy inputs is not well-studied. We propose ArithmAttack to examine how robust the LLMs are when they encounter noisy prompts that contain extra noise in the form of punctuation marks. While being easy to implement, ArithmAttack does not cause any information loss since words are not added or deleted from the context. We evaluate the robustness of eight LLMs, including LLama3, Mistral, Mathstral, and DeepSeek on noisy GSM8K and MultiArith datasets. Our experiments suggest that all the studied models show vulnerability to such noise, with more noise leading to poorer performances.

Method: Dynamic Noise Preference Optimization (DNPO) combines dynamic sample labeling for constructing preference pairs with controlled, trainable noise injection during preference optimization to prevent stagnation.

Result: DNPO consistently outperforms existing methods across multiple benchmarks with Llama-3.2-3B and Zephyr-7B. Zephyr-7B shows 29.4% win-loss rate gap improvement in model-generated data quality compared to baseline in GPT-4 evaluations.

Conclusion: DNPO effectively addresses the stagnation problem in synthetic data fine-tuning, enabling continuous improvement in LLM performance through dynamic noise preference optimization.

Abstract: Although LLMs have achieved significant success, their reliance on large volumes of human-annotated data has limited their potential for further scaling. In this situation, utilizing self-generated synthetic data has become crucial for fine-tuning LLMs without extensive human annotation. However, current methods often fail to ensure consistent improvements across iterations, with performance stagnating after only minimal updates. To overcome these challenges, we introduce Dynamic Noise Preference Optimization (DNPO), which combines dynamic sample labeling for constructing preference pairs with controlled, trainable noise injection during preference optimization. Our approach effectively prevents stagnation and enables continuous improvement. In experiments with Llama-3.2-3B and Zephyr-7B, DNPO consistently outperforms existing methods across multiple benchmarks. Additionally, with Zephyr-7B, DNPO shows a significant improvement in model-generated data quality, with a 29.4% win-loss rate gap compared to the baseline in GPT-4 evaluations.

Method: LLM-assisted data generation comparing open-weight and closed-weight models, selecting GPT-4o as backbone. Created 10,600 instruction-following samples covering institutional and cultural knowledge relevant to Kazakhstan. Each entity undergoes full manual verification for quality assurance.

Result: Fine-tuning Qwen, Falcon, and Gemma on the dataset leads to consistent performance improvements in both multiple-choice and generative tasks, demonstrating the effectiveness of LLM-assisted instruction tuning for low-resource languages.

Conclusion: The work shows the potential of LLM-assisted instruction tuning for low-resource languages, providing a high-quality dataset that enhances LLMs’ understanding of specialized domains like government and culture in underrepresented languages.

Abstract: Instruction tuning in low-resource languages remains underexplored due to limited text data, particularly in government and cultural domains. To address this, we introduce and open-source a large-scale (10,600 samples) instruction-following (IFT) dataset, covering key institutional and cultural knowledge relevant to Kazakhstan. Our dataset enhances LLMs’ understanding of procedural, legal, and structural governance topics. We employ LLM-assisted data generation, comparing open-weight and closed-weight models for dataset construction, and select GPT-4o as the backbone. Each entity of our dataset undergoes full manual verification to ensure high quality. We also show that fine-tuning Qwen, Falcon, and Gemma on our dataset leads to consistent performance improvements in both multiple-choice and generative tasks, demonstrating the potential of LLM-assisted instruction tuning for low-resource languages.

Method: MFT learns and applies binary masks to fully fine-tuned models using standard LLM fine-tuning objectives as supervision. It breaks model structural integrity through masking operations without updating model weights.

Result: MFT achieves consistent performance gains across domains and backbones, with average gains of 2.70/4.15 in IFEval using LLaMA2-7B/3.1-8B. It’s compatible with other LLM optimization procedures and extends masking operations beyond conventional network pruning.

Conclusion: Carefully breaking model structural integrity through masking can surprisingly improve performance, challenging conventional wisdom about maintaining model integrity. MFT demonstrates a novel fine-tuning paradigm with broader applications beyond model compression.

Abstract: The large language model (LLM) is typically integrated into the mainstream optimization protocol. No work has questioned whether maintaining the model integrity is \textit{indispensable} for promising performance. In this work, we introduce Mask Fine-Tuning (MFT), a novel LLM fine-tuning paradigm demonstrating that carefully breaking the model’s structural integrity can surprisingly improve performance without updating model weights. MFT learns and applies binary masks to well-optimized models, using the standard LLM fine-tuning objective as supervision. Based on fully fine-tuned models, MFT uses the same fine-tuning datasets to achieve consistent performance gains across domains and backbones (e.g., an average gain of \textbf{2.70 / 4.15} in IFEval with LLaMA2-7B / 3.1-8B). Detailed ablation studies and analyses examine the proposed MFT from different perspectives, such as sparse ratio and loss surface. Additionally, by deploying it on well-trained models, MFT is compatible with collaborating with other LLM optimization procedures to enhance the general model. Furthermore, this study extends the functionality of the masking operation beyond its conventional network-pruning context for model compression to a broader model capability scope.

Method: Progressive training approach where smaller models are incrementally expanded to larger sizes, with strategic adjustment of maximum learning rate based on model size.

Result: 25% reduction in total computational cost while maintaining comparable performance to independently trained models; outperforms independent training across various metrics with greater consistency across model sizes.

Conclusion: Progressive training offers an efficient alternative to independent training for constructing LLM families, reducing costs while improving performance and consistency.

Abstract: As Large Language Models (LLMs) gain widespread practical application, offering model families with varying parameter sizes has become standard practice to accommodate diverse computational requirements. Traditionally, each model in the family is trained independently, incurring computational costs that scale additively with the number of models. In this work, we propose an efficient method for constructing model families via progressive training, where smaller models are incrementally expanded to larger sizes to create a complete model family. Through extensive experiments on a model family ranging from 1B to 8B parameters, we show that our approach reduces total computational cost by approximately 25% while maintaining comparable performance to independently trained models. Moreover, by strategically adjusting the maximum learning rate based on model size, our method outperforms the independent training across various metrics. Beyond these improvements, our approach also fosters greater consistency in behavior across model sizes.

Method: Proposed the first benchmark dataset for evaluating LLMs on simulating personas with reversed performance, using mathematical reasoning as a representative scenario. Evaluated both open-weight and closed-source LLMs on this “counterfactual instruction following” task.

Result: LLMs, including OpenAI’s o1 reasoning model, all struggle to follow counterfactual instructions for simulating reversedly performing personas. The effect worsens when intersectionally simulating both performance level and race population.

Conclusion: The results highlight significant challenges in counterfactual instruction following and demonstrate the need for further research to improve LLMs’ ability to simulate diverse personas with varying performance levels.

Abstract: Large Language Models (LLMs) are now increasingly widely used to simulate personas in virtual environments, leveraging their instruction-following capability. However, we discovered that even state-of-the-art LLMs cannot simulate personas with reversed performance (e.g., student personas with low proficiency in educational settings), which impairs the simulation diversity and limits the practical applications of the simulated environments. In this work, using mathematical reasoning as a representative scenario, we propose the first benchmark dataset for evaluating LLMs on simulating personas with reversed performance, a capability that we dub “counterfactual instruction following”. We evaluate both open-weight and closed-source LLMs on this task and find that LLMs, including the OpenAI o1 reasoning model, all struggle to follow counterfactual instructions for simulating reversedly performing personas. Intersectionally simulating both the performance level and the race population of a persona worsens the effect even further. These results highlight the challenges of counterfactual instruction following and the need for further research.

Method: Provides overview of synthetic dataset creation, evaluation, and usage for medical dialogue tasks; proposes novel typology for classifying types and degrees of data synthesis

Result: Comprehensive survey of synthetic clinical dialogue datasets with proposed typology to facilitate comparison and evaluation across different synthesis approaches

Conclusion: Synthetic datasets are crucial for clinical NLP but need better theoretical frameworks; proposed typology helps standardize evaluation and comparison of synthesis methods

Abstract: Synthetic data sets are used across linguistic domains and NLP tasks, particularly in scenarios where authentic data is limited (or even non-existent). One such domain is that of clinical (healthcare) contexts, where there exist significant and long-standing challenges (e.g., privacy, anonymization, and data governance) which have led to the development of an increasing number of synthetic datasets. One increasingly important category of clinical dataset is that of clinical dialogues which are especially sensitive and difficult to collect, and as such are commonly synthesized. While such synthetic datasets have been shown to be sufficient in some situations, little theory exists to inform how they may be best used and generalized to new applications. In this paper, we provide an overview of how synthetic datasets are created, evaluated and being used for dialogue related tasks in the medical domain. Additionally, we propose a novel typology for use in classifying types and degrees of data synthesis, to facilitate comparison and evaluation.

Method: Proposes a weak-to-strong reasoning paradigm where weaker models provide supervision to stronger student models. Analyzes when and why such weak supervision succeeds in eliciting reasoning abilities in stronger models across diverse benchmarks and model architectures.

Result: Supervision from significantly weaker reasoners can substantially improve student reasoning performance, recovering close to 94% of the gains of expensive RL at a fraction of the cost. The approach consistently improves performance across a wide range of reasoning tasks.

Conclusion: The weak-to-strong paradigm is a promising and generalizable alternative to costly methods for incentivizing strong reasoning capabilities at inference-time in LLMs, offering substantial performance gains with minimal cost.

Abstract: Large language models (LLMs) have demonstrated impressive performance on reasoning-intensive tasks, but enhancing their reasoning abilities typically relies on either reinforcement learning (RL) with verifiable signals or supervised fine-tuning (SFT) with high-quality long chain-of-thought (CoT) demonstrations, both of which are expensive. In this paper, we study a novel problem of incentivizing the reasoning capacity of LLMs without expensive high-quality demonstrations and reinforcement learning. We investigate whether the reasoning capabilities of LLMs can be effectively incentivized via supervision from significantly weaker models. We further analyze when and why such weak supervision succeeds in eliciting reasoning abilities in stronger models. Our findings show that supervision from significantly weaker reasoners can substantially improve student reasoning performance, recovering close to 94% of the gains of expensive RL at a fraction of the cost. Experiments across diverse benchmarks and model architectures demonstrate that weak reasoners can effectively incentivize reasoning in stronger student models, consistently improving performance across a wide range of reasoning tasks. Our results suggest that this simple weak-to-strong paradigm is a promising and generalizable alternative to costly methods for incentivizing strong reasoning capabilities at inference-time in LLMs. The code is publicly available at https://github.com/yuanyige/w2sr.

Method: SEA formulates inference as iterative optimization on an energy function over actions in continuous latent space, using gradient-based sampling to adapt original responses toward optimal ones.

Result: SEA outperforms second-best baselines with relative improvements of up to 77.51% on AdvBench and 16.36% on MATH benchmarks.

Conclusion: SEA provides a simple yet effective approach for inference-time alignment by operating in continuous latent space, addressing limitations of discrete search methods.

Abstract: Aligning large language models with human feedback at inference time has received increasing attention due to its flexibility. Existing methods rely on generating multiple responses from the base policy for search using a reward model, which can be considered as searching in a discrete response space. However, these methods struggle to explore informative candidates when the base policy is weak or the candidate set is small, resulting in limited effectiveness. In this paper, to address this problem, we propose Simple Energy Adaptation ($\textbf{SEA}$), a simple yet effective algorithm for inference-time alignment. In contrast to expensive search over the discrete space, SEA directly adapts original responses from the base policy toward the optimal one via gradient-based sampling in continuous latent space. Specifically, SEA formulates inference as an iterative optimization procedure on an energy function over actions in the continuous space defined by the optimal policy, enabling simple and effective alignment. For instance, despite its simplicity, SEA outperforms the second-best baseline with a relative improvement of up to $ \textbf{77.51%}$ on AdvBench and $\textbf{16.36%}$ on MATH. Our code is publicly available at https://github.com/yuanyige/sea

Method: Proposes ERC-SVD with two key innovations: 1) Leverages residual matrix generated during truncation to reduce truncation loss, 2) Under fixed overall compression ratio, selectively compresses only the last few layers to mitigate error propagation.

Result: Comprehensive evaluations on diverse LLM families and multiple benchmark datasets show ERC-SVD consistently achieves superior performance over existing counterpart methods.

Conclusion: ERC-SVD demonstrates practical effectiveness for LLM compression through error-controlled SVD approach that addresses limitations of current methods.

Abstract: Large language models (LLMs) have demonstrated impressive capabilities in a wide range of downstream natural language processing tasks. Nevertheless, their considerable sizes and memory demands hinder practical deployment, underscoring the importance of developing efficient compression strategies. Singular value decomposition (SVD) decomposes a matrix into orthogonal components, enabling efficient low-rank approximation. This is particularly suitable for LLM compression, where weight matrices often exhibit significant redundancy. However, current SVD-based methods neglect the residual matrix from truncation, resulting in significant truncation loss. Additionally, compressing all layers of the model results in severe error propagation. To overcome these limitations, we propose ERC-SVD, a new post-training SVD-based LLM compression method from an error-controlled perspective. Specifically, we leverage the residual matrix generated during the truncation process to reduce truncation loss. Moreover, under a fixed overall compression ratio, we selectively compress the last few layers of the model, which mitigates error propagation and improves compressed model performance. Comprehensive evaluations on diverse LLM families and multiple benchmark datasets indicate that ERC-SVD consistently achieves superior performance over existing counterpart methods, demonstrating its practical effectiveness.

Method: A lightweight controller network observes intermediate LLM activations and predicts both a global scaling factor and layer-specific weights to dynamically modulate steering patches derived from pre-computed “refusal direction” vectors across layers during generation.

Result: Experiments on safety benchmarks (ToxicChat & In-The-Wild Jailbreak Prompts) show significantly increased refusal rates compared to base LLMs, outperforming existing methods on Llama-3.1-8B, Llama-3.2-1B & Mistral-7B.

Conclusion: The approach provides efficient, adaptive fine-grained control over LLM behavior at inference time without altering original model parameters, enabling targeted safety interventions.

Abstract: Controlling undesirable Large Language Model (LLM) behaviors, such as the generation of unsafe content or failing to adhere to safety guidelines, often relies on costly fine-tuning. Activation steering provides an alternative for inference-time control, but existing methods typically lack fine-grained, adaptive mechanisms. We introduce a novel approach using a lightweight, trainable controller network integrated during inference. This controller network observes specific intermediate LLM activations and predicts both a global scaling factor and layer-specific weights. The predicted global scaling factor and layer-specific weights then dynamically modulate the intensity of a steering patch, derived from a pre-computed “refusal direction” vector, applied across the LLM’s layers during generation. Trained on activations from both harmful and benign prompts, our controller learns to discriminatively apply nuanced, layer-aware interventions, activating steering primarily for harmful inputs. Experiments using safety benchmarks like ToxicChat & In-The-Wild Jailbreak Prompts demonstrate that our weighted steering controller significantly increases refusal rates compared to the base LLM, achieving targeted behavioral modification without altering the original model parameters. Our experiments with Llama-3.1-8B, Llama-3.2-1B & Mistral-7B show our approach outperforms existing methods, presenting an efficient and adaptive method for fine-grained control over LLM behavior at inference time.

Method: AJF first categorizes LLMs by comprehension ability: Type-I (limited) and Type-II (strong). For Type-I, it uses MuEn strategy with layered semantic mutations and encryption. For Type-II, it uses MuDeEn strategy that adds encrypted response generation for dual-end encryption.

Result: Achieved attack success rates of 98.9% on GPT-4o (May 2025) and 99.8% on GPT-4.1 (July 2025), demonstrating highly effective jailbreak capabilities.

Conclusion: The framework successfully bypasses LLM alignment defenses by adapting to model comprehension abilities, revealing significant vulnerabilities in current alignment mechanisms.

Abstract: Recent advancements in adversarial jailbreak attacks have exposed critical vulnerabilities in Large Language Models (LLMs), enabling the circumvention of alignment safeguards through increasingly sophisticated prompt manipulations. Our experiments find that the effectiveness of jailbreak strategies is influenced by the comprehension ability of the target LLM. Building on this insight, we propose an Adaptive Jailbreak Framework (AJF) based on the comprehension ability of black-box large language models. Specifically, AJF first categorizes the comprehension ability of the LLM and then applies different strategies accordingly: For models with limited comprehension ability (Type-I LLMs), AJF integrates layered semantic mutations with an encryption technique (MuEn strategy), to more effectively evade the LLM’s defenses during the input and inference stages. For models with strong comprehension ability (Type-II LLMs), AJF employs a more complex strategy that builds upon the MuEn strategy by adding an additional layer: inducing the LLM to generate an encrypted response. This forms a dual-end encryption scheme (MuDeEn strategy), further bypassing the LLM’s defenses during the output stage. Experimental results demonstrate the effectiveness of our approach, achieving attack success rates of \textbf{98.9%} on GPT-4o (29 May 2025 release) and \textbf{99.8%} on GPT-4.1 (8 July 2025 release). Our work contributes to a deeper understanding of the vulnerabilities in current LLMs alignment mechanisms.

Method: Created BIS Reasoning 1.0 dataset with logically valid but belief-inconsistent syllogisms, benchmarked various LLMs (OpenAI GPT variants, Qwen, Japanese LLMs) under uniform zero-shot protocols, analyzed performance across different prompt designs and reasoning efforts.

Result: Reasoning-optimized models achieved near-perfect accuracy (Qwen3-32B ≈99%, GPT-5-mini up to ≈99.7%), GPT-4o around 80%, while earlier Japanese-specialized models performed below 60%. Latest Japanese models improved to mid-80% range. Performance sensitive to prompt design and reasoning effort.

Conclusion: Robustness to belief-inconsistent reasoning is driven more by explicit reasoning optimization than language specialization or scale alone. Even top models struggle when logic conflicts with intuitive beliefs, highlighting need for better reasoning capabilities in safety-critical domains.

Abstract: We present BIS Reasoning 1.0, the first large-scale Japanese dataset of syllogistic reasoning problems explicitly designed to evaluate belief-inconsistent reasoning in large language models (LLMs). Unlike prior resources such as NeuBAROCO and JFLD, which emphasize general or belief-aligned logic, BIS Reasoning 1.0 systematically introduces logically valid yet belief-inconsistent syllogisms to expose belief bias, the tendency to accept believable conclusions irrespective of validity. We benchmark a representative suite of cutting-edge models, including OpenAI GPT-5 variants, GPT-4o, Qwen, and prominent Japanese LLMs, under a uniform, zero-shot protocol. Reasoning-centric models achieve near-perfect accuracy on BIS Reasoning 1.0 (e.g., Qwen3-32B $\approx$99% and GPT-5-mini up to $\approx$99.7%), while GPT-4o attains around 80%. Earlier Japanese-specialized models underperform, often well below 60%, whereas the latest llm-jp-3.1-13b-instruct4 markedly improves to the mid-80% range. These results indicate that robustness to belief-inconsistent inputs is driven more by explicit reasoning optimization than by language specialization or scale alone. Our analysis further shows that even top-tier systems falter when logical validity conflicts with intuitive or factual beliefs, and that performance is sensitive to prompt design and inference-time reasoning effort. We discuss implications for safety-critical domains, including law, healthcare, and scientific literature, where strict logical fidelity must override intuitive belief to ensure reliability.

Method: SimLens keeps only the start token [s] and candidate answer token [a], performing one lightweight continuation through remaining upper layers. Also introduces Linear SimLens for entropy-based confidence estimation and SimExit for hybrid early-exit mechanism.

Result: On ARC, BoolQ, and HeadQA with LLaMA-7B and Vicuna-7B, SimLens improves Iso-Compute accuracy in all six settings with average gain of +0.43. SimExit yields average 1.15× speedup at best-accuracy points and 1.40× with up to 1% accuracy drop.

Conclusion: SimLens significantly improves intermediate-layer prediction accuracy with minimal overhead, and SimExit provides effective early-exit mechanism. The approach reveals distinct roles of start and answer tokens as global condition and semantic anchor.

Abstract: Intermediate-layer predictions in large language models (LLMs) are informative but hard to decode accurately, especially at early layers. Existing lens-style methods typically rely on direct linear readout, which is simple but often drifts away from the model’s eventual prediction. We proposeSimLens, a simple training-free decoder for single-token decision tasks that keeps only the start token and a candidate answer token ([s] and [a]) and performs one lightweight continuation through the remaining upper layers. This surprisingly small modification recovers much more accurate latent predictions than direct linear decoding. We further introduce Linear SimLens, a lightweight linear approximation for entropy-based confidence estimation, and combine the two in SimExit, a hybrid early-exit mechanism. On ARC, BoolQ, and HeadQA with LLaMA-7B and Vicuna-7B, SimLens improves Iso-Compute accuracy in all six settings, with an average gain of +0.43 even when fair compute includes the extra two-token post-forward overhead. SimExit yields an average 1.15$\times$ speedup at the best-accuracy operating points and 1.40$\times$ when allowing up to a 1 percentage-point accuracy drop. Ablations show that [s] and [a] play distinct roles as global condition and semantic anchor, respectively.

Method: Proposes Difficulty-Influence Quadrant (DIQ) that selects samples in the “high-difficulty-high-influence” quadrant, combining knowledge/reasoning complexity with gradient-based optimization utility. This enables efficient medical reasoning with minimal fine-tuning data.

Result: DIQ-selected subsets (1% of data) match full-dataset performance, while 10% consistently outperforms baselines. Human and LLM evaluations show higher data quality and more expert-aligned clinical reasoning in differential diagnosis, safety checks, and evidence citation.

Conclusion: DIQ demonstrates that principled data selection based on balancing difficulty and gradient influence is superior to brute-force scaling for medical VLM fine-tuning, enabling efficient adaptation with minimal data while improving reasoning quality.

Abstract: Supervised Fine-Tuning (SFT) of the language backbone plays a pivotal role in adapting Vision-Language Models (VLMs) to specialized domains such as medical reasoning. However, existing SFT practices often rely on unfiltered textual datasets that contain redundant and low-quality samples, leading to substantial computational costs and suboptimal performance in complex clinical scenarios. Although existing methods attempt to alleviate this problem by selecting data based on sample difficulty, defined by knowledge and reasoning complexity, they overlook each sample’s optimization utility reflected in its gradient. Interestingly, we find that gradient-based influence alone favors easy-to-optimize samples that cause large parameter shifts but lack deep reasoning chains, while difficulty alone selects noisy or overly complex textual cases that fail to guide stable optimization. Based on this observation, we propose a data selection strategy, Difficulty-Influence Quadrant (DIQ), which prioritizes samples in the “high-difficulty-high-influence” quadrant to balance complex clinical reasoning with substantial gradient influence. This enables efficient medical reasoning for VLMs with minimal fine-tuning data. Furthermore, Human and LLM-as-a-judge evaluations show that DIQ-selected subsets demonstrate higher data quality and generate clinical reasoning that is more aligned with expert practices in differential diagnosis, safety check, and evidence citation, as DIQ emphasizes samples that foster expert-like reasoning patterns. Extensive experiments on medical reasoning benchmarks demonstrate that DIQ enables VLM backbones fine-tuned on only 1% of selected data to match full-dataset performance, while using 10% consistently outperforms baseline methods, highlighting the superiority of principled data selection over brute-force scaling. The code is available at https://github.com/mihara-bot/DIQ.

Method: Systematically compares two self-supervised augmentation strategies (cropping vs dropout) for fine-tuning text embeddings. Evaluates on MTEB benchmark and in-domain data, analyzes layer-wise representation quality, and tests fine-tuning only last transformer layers.

Result: Cropping augmentation strongly outperforms dropout-based approach. Self-supervised embeddings achieve high quality on in-domain data after short fine-tuning, though lag behind supervised SOTA on out-of-domain data. Last transformer layers show largest improvement during fine-tuning, and fine-tuning only these layers yields similar quality.

Conclusion: Self-supervised fine-tuning with cropping augmentation can produce effective text embeddings for in-domain applications with minimal training, offering a practical alternative to supervised approaches when labeled data is scarce.

Abstract: Text embeddings, i.e. vector representations of entire texts, play an important role in many NLP applications, such as retrieval-augmented generation, clustering, or visualizing collections of texts for data exploration. Currently, top-performing embedding models are derived from pre-trained language models via supervised contrastive fine-tuning. This fine-tuning strategy relies on an external notion of similarity and annotated data for generation of positive pairs. Here we study self-supervised fine-tuning and systematically compare the two most well-known augmentation strategies used for fine-tuning text embeddings models. We assess embedding quality on MTEB and additional in-domain evaluations and show that cropping augmentation strongly outperforms the dropout-based approach. We find that on out-of-domain data, the quality of resulting embeddings is substantially below the supervised state-of-the-art models, but for in-domain data, self-supervised fine-tuning can produce high-quality text embeddings after very short fine-tuning. Finally, we show that representation quality increases towards the last transformer layers, which undergo the largest change during fine-tuning; and that fine-tuning only those last layers is sufficient to reach similar embedding quality.

Method: Proposes Self-Evolving Pairwise Reasoning (EvolvR) framework: 1) Self-synthesizes score-aligned Chain-of-Thought data using multi-persona strategy, 2) Self-filters raw CoTs with multi-agents for logical rigor, 3) Trains evaluator on refined data as reward model for story generation.

Result: Achieves state-of-the-art performance on three evaluation benchmarks (StoryER, HANNA, OpenMEVA). When used as reward model, significantly enhances quality of generated stories, validating superiority of self-evolving approach.

Conclusion: EvolvR framework effectively addresses limitations of existing LLM-as-a-judge methods for story evaluation through self-evolving pairwise reasoning, demonstrating both superior evaluation performance and practical utility in improving story generation quality.

Abstract: Although the effectiveness of Large Language Models (LLMs) as judges (LLM-as-a-judge) has been validated, their performance remains limited in open-ended tasks, particularly in story evaluation. Accurate story evaluation is crucial not only for assisting human quality judgment but also for providing key signals to guide story generation. However, existing methods face a dilemma: prompt engineering for closed-source models suffers from poor adaptability, while fine-tuning approaches for open-source models lack the rigorous reasoning capabilities essential for story evaluation. To address this, we propose the Self-Evolving Pairwise Reasoning (EvolvR) framework. Grounded in pairwise comparison, the framework first self-synthesizes score-aligned Chain-of-Thought (CoT) data via a multi-persona strategy. To ensure data quality, these raw CoTs undergo a self-filtering process, utilizing multi-agents to guarantee their logical rigor and robustness. Finally, the evaluator trained on the refined data is deployed as a reward model to guide the story generation task. Experimental results demonstrate that our framework achieves state-of-the-art (SOTA) performance on three evaluation benchmarks including StoryER, HANNA and OpenMEVA. Furthermore, when served as a reward model, it significantly enhances the quality of generated stories, thereby fully validating the superiority of our self-evolving approach.

Method: Systematic study identifying activation outliers in dLLMs, implementing state-of-the-art PTQ methods, and conducting comprehensive evaluation across four dimensions: bit-width, quantization method, task category, and model type.

Result: Identified activation outliers as key quantization challenge, evaluated PTQ methods across different configurations, and provided practical insights into dLLM quantization behavior. Code made publicly available.

Conclusion: First systematic study on dLLM quantization provides foundation for future research in efficient dLLM deployment, addressing key challenges for edge device applications.

Abstract: Recent advances in diffusion large language models (dLLMs) have introduced a promising alternative to autoregressive (AR) LLMs for natural language generation tasks, leveraging full attention and denoising-based decoding strategies. However, the deployment of these models on edge devices remains challenging due to their massive parameter scale and high resource demands. While post-training quantization (PTQ) has emerged as a widely adopted technique for compressing AR LLMs, its applicability to dLLMs remains largely unexplored. In this work, we present the first systematic study on quantizing diffusion-based language models. We begin by identifying the presence of activation outliers, characterized by abnormally large activation values that dominate the dynamic range. These outliers pose a key challenge to low-bit quantization, as they make it difficult to preserve precision for the majority of values. More importantly, we implement state-of-the-art PTQ methods and conduct a comprehensive evaluation across multiple task types and model variants. Our analysis is structured along four key dimensions: bit-width, quantization method, task category, and model type. Through this multi-perspective evaluation, we offer practical insights into the quantization behavior of dLLMs under different configurations. We hope our findings provide a foundation for future research in efficient dLLM deployment. Our code is publicly available at https://github.com/FelixMessi/QDLM.

Method: Uses Bi-Induct, a lightweight data rewrite that interleaves short directional copy snippets (forward-copy for induction, backward-copy for anti-induction control, or balanced mix) into natural pretraining streams. Evaluates across 0.13B-1B decoder-only models on few-shot performance, head-level copy telemetry, and held-out perplexity.

Result: Bi-Induct reliably increases induction-head activity but doesn’t translate to consistent few-shot generalization improvements. Natural-only models perform best on function-style probes. Anti-induction scores remain near zero despite explicit backward-copy cues. Natural-only training produces more centralized, load-bearing induction circuitry while Bi-Induct creates more distributed, redundant activity.

Conclusion: Eliciting a mechanism is not the same as making it load-bearing. Synthetic data interventions should be evaluated not only by signature amplification but by whether they create causally necessary computation while preserving natural-data modeling quality.

Abstract: Mechanism-targeted synthetic data is increasingly proposed as a way to steer pretraining toward desirable capabilities, but it remains unclear how such interventions should be evaluated. We study this question for in-context learning (ICL) under matched compute (iso-FLOPs) using Bi-Induct, a lightweight data rewrite that interleaves short directional copy snippets into a natural pretraining stream: forward-copy (induction), backward-copy (anti-induction, as a directional control), or a balanced mix. Across 0.13B-1B decoder-only models, we evaluate (i) few-shot performance on standard LM benchmarks and function-style ICL probes, (ii) head-level copy telemetry, and (iii) held-out perplexity as a guardrail. Bi-Induct reliably increases induction-head activity, but this does not translate into consistent improvements in few-shot generalization: on standard LM benchmarks, Bi-Induct is largely performance-neutral relative to natural-only training, while on function-style probes the 1B natural-only model performs best. Despite explicit backward-copy cues, anti-induction scores remain near zero across scales, revealing a strong forward/backward asymmetry. Targeted ablations show a sharper distinction: removing the top 2% induction heads per layer harms ICL more than matched random ablations, with the largest relative drop occurring in the natural-only models. This indicates that natural-only training produces more centralized, load-bearing induction circuitry, whereas Bi-Induct tends to create more distributed and redundant induction activity. Our main conclusion is that eliciting a mechanism is not the same as making it load-bearing. For data-centric foundation model design, this suggests that synthetic data interventions should be evaluated not only by signature amplification, but by whether they create causally necessary computation while preserving natural-data modeling quality.

Method: DECS introduces: (1) a decoupled token-level reward mechanism that surgically distinguishes and penalizes redundant tokens, addressing two flaws in current length rewards (erroneous penalization of essential exploratory tokens and inadvertent rewarding of partial redundancy), and (2) a novel curriculum batch scheduling strategy to master the efficiency-efficacy equilibrium.

Result: Experimental results show DECS achieves dramatic reduction in reasoning tokens by over 50% across seven benchmarks while simultaneously maintaining or even improving performance.

Conclusion: DECS demonstrates that substantial gains in reasoning efficiency can be achieved without compromising a model’s underlying reasoning power, providing a practical solution to the overthinking problem in large reasoning models.

Abstract: While large reasoning models trained with critic-free reinforcement learning and verifiable rewards (RLVR) represent the state-of-the-art, their practical utility is hampered by ``overthinking’’, a critical issue where models generate excessively long reasoning paths without any performance benefit. Existing solutions that penalize length often fail, inducing performance degradation due to a fundamental misalignment between trajectory-level rewards and token-level optimization. In this work, we introduce a novel framework, DECS, built on our theoretical discovery of two previously unaddressed flaws in current length rewards: (1) the erroneous penalization of essential exploratory tokens and (2) the inadvertent rewarding of partial redundancy. Our framework’s innovations include (i) a first-of-its-kind decoupled token-level reward mechanism that surgically distinguishes and penalizes redundant tokens, and (ii) a novel curriculum batch scheduling strategy to master the efficiency-efficacy equilibrium. Experimental results show DECS can achieve a dramatic reduction in reasoning tokens by over 50% across seven benchmarks while simultaneously maintaining or even improving performance. It demonstrates conclusively that substantial gains in reasoning efficiency can be achieved without compromising a model’s underlying reasoning power. Code is available at https://github.com/pixas/DECS.

Method: Distributional Semantics Tracing (DST) builds layer-wise semantic maps at answer positions by decoding residual-stream states through unembedding, selecting top-K concepts, and estimating directed concept-to-concept support via lightweight causal tracing.

Result: DST yields more faithful explanations than attribution, probing, and intervention baselines on the Racing Thoughts dataset under an LLM-judge protocol, and the resulting Contextual Alignment Score strongly predicts hallucination failures.

Conclusion: Hallucinations arise from correlation-driven representational drift across depth, where the residual stream is pulled toward locally coherent but context-inconsistent concept neighborhoods reinforced by training co-occurrences.

Abstract: Hallucinations in large language models (LLMs) produce fluent continuations that are not supported by the prompt, especially under minimal contextual cues and ambiguity. We introduce Distributional Semantics Tracing (DST), a model-native method that builds layer-wise semantic maps at the answer position by decoding residual-stream states through the unembedding, selecting a compact top-$K$ concept set, and estimating directed concept-to-concept support via lightweight causal tracing. Using these traces, we test a representation-level hypothesis: hallucinations arise from correlation-driven representational drift across depth, where the residual stream is pulled toward a locally coherent but context-inconsistent concept neighborhood reinforced by training co-occurrences. On Racing Thoughts dataset, DST yields more faithful explanations than attribution, probing, and intervention baselines under an LLM-judge protocol, and the resulting Contextual Alignment Score (CAS) strongly predicts failures, supporting this drift hypothesis.

Method: MOSAIC uses a training-free multi-agent framework with specially designed agents operating in a student-teacher paradigm. The framework includes agents for self-reflection, rationale creation, coding, and debugging. It employs stepwise problem decomposition, targeted error correction, and a Consolidated Context Window (CCW) to mitigate LLM hallucinations when solving complex scientific tasks with chained subproblems.

Result: MOSAIC outperforms existing approaches on scientific coding benchmarks in terms of accuracy, robustness, and interpretability. The framework demonstrates improved performance in handling complex scientific coding tasks involving chained subproblems.

Conclusion: The MOSAIC framework provides an effective approach for scientific code generation by addressing the unique challenges of scientific workflows through specialized multi-agent design, stepwise decomposition, and hallucination mitigation techniques.

Abstract: We present MOSAIC, a multi-agent Large Language Model (LLM) framework for solving challenging scientific coding tasks. Unlike general-purpose coding, scientific workflows require algorithms that are rigorous, interconnected with deep domain knowledge, and incorporate domain-specific reasoning, as well as algorithm iteration without requiring I/O test cases. Many scientific problems also require a sequence of subproblems to be solved, leading to the final desired result. MOSAIC is designed as a training-free framework with specially designed agents to self-reflect, create the rationale, code, and debug within a student-teacher paradigm to address the challenges of scientific code generation. This design facilitates stepwise problem decomposition, targeted error correction, and, when combined with our Consolidated Context Window (CCW), mitigates LLM hallucinations when solving complex scientific tasks involving chained subproblems. We evaluate MOSAIC on scientific coding benchmarks and demonstrate that our specialized agentic framework outperforms existing approaches in terms of accuracy, robustness, and interpretability.

Method: Dual-pronged framework: 1) Checkpoint merging to smooth parameter space by averaging recent model weights, 2) Pass@k metric for robust statistical estimation of model capability with lower variance

Result: MaP yields significantly smoother performance curves, reduces inter-run variance, ensures more consistent model rankings, and provides more reliable observation of training dynamics

Conclusion: MaP provides a more reliable evaluation framework for LLM training, laying crucial empirical foundation for LLM research by addressing both parameter and evaluation instability

Abstract: Reliable evaluation is fundamental to the progress of Large Language Models (LLMs), yet the evaluation process during pre-training is plagued by significant instability that obscures true learning dynamics. In this work, we systematically diagnose this instability, attributing it to two distinct sources: \textit{Parameter Instability} from training stochasticity and \textit{Evaluation Instability} from noisy measurement protocols. To counteract both sources of noise, we introduce \textbf{MaP}, a dual-pronged framework that synergistically integrates checkpoint \underline{M}erging \underline{a}nd the \underline{P}ass@k metric. Checkpoint merging smooths the parameter space by averaging recent model weights, while Pass@k provides a robust, low-variance statistical estimate of model capability. Extensive experiments show that MaP yields significantly smoother performance curves, reduces inter-run variance, and ensures more consistent model rankings. Ultimately, MaP provides a more reliable and faithful lens for observing LLM training dynamics, laying a crucial empirical foundation for LLM research.

Method: Position paper methodology - draws on survey methodology research and NLP advances to argue conceptually for open-ended approaches, proposing novel practices and evaluation frameworks for free-form text generation in social simulations.

Result: Conceptual argument that open-endedness improves measurement and design, supports exploration of unanticipated views, reduces researcher-imposed bias, captures expressiveness and individuality, aids pretesting, and enhances methodological utility.

Conclusion: Researchers should leverage LLMs’ generative diversity through open-ended approaches rather than constraining them, creating synergies between NLP and social science through novel practices and evaluation frameworks.

Abstract: Large Language Models (LLMs) are increasingly used to simulate public opinion and other social phenomena. Most current studies constrain these simulations to multiple-choice or short-answer formats for ease of scoring and comparison, but such closed designs overlook the inherently generative nature of LLMs. In this position paper, we argue that open-endedness, using free-form text that captures topics, viewpoints, and reasoning processes “in” LLMs, is essential for realistic social simulation. Drawing on decades of survey-methodology research and recent advances in NLP, we argue why this open-endedness is valuable in LLM social simulations, showing how it can improve measurement and design, support exploration of unanticipated views, and reduce researcher-imposed directive bias. It also captures expressiveness and individuality, aids in pretesting, and ultimately enhances methodological utility. We call for novel practices and evaluation frameworks that leverage rather than constrain the open-ended generative diversity of LLMs, creating synergies between NLP and social science.

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

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

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

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

Method: VISTA decomposes each assistant turn into atomic factual claims, verifies them against trusted sources and dialogue history, and categorizes unverifiable statements (subjective, contradicted, lacking evidence, or abstaining). It models factuality as a dynamic property of conversation.

Result: Across eight large language models and four dialogue factuality benchmarks (AIS, BEGIN, FAITHDIAL, and FADE), VISTA substantially improves hallucination detection over FACTSCORE and LLM-as-Judge baselines. Human evaluation confirms improved annotator agreement and reveals inconsistencies in existing benchmarks.

Conclusion: VISTA offers a more transparent, human-aligned measure of truthfulness in dialogue systems by modeling factuality as a dynamic property of conversation rather than evaluating isolated responses.

Abstract: Hallucination–defined here as generating statements unsupported or contradicted by available evidence or conversational context–remains a major obstacle to deploying conversational AI systems in settings that demand factual reliability. Existing metrics either evaluate isolated responses or treat unverifiable content as errors, limiting their use for multi-turn dialogue. We introduce VISTA (Verification In Sequential Turn-based Assessment), a framework for evaluating conversational factuality through claim-level verification and sequential consistency tracking. VISTA decomposes each assistant turn into atomic factual claims, verifies them against trusted sources and dialogue history, and categorizes unverifiable statements (subjective, contradicted, lacking evidence, or abstaining). Across eight large language models and four dialogue factuality benchmarks (AIS, BEGIN, FAITHDIAL, and FADE), VISTA substantially improves hallucination detection over FACTSCORE and LLM-as-Judge baselines. Human evaluation confirms that VISTA’s decomposition improves annotator agreement and reveals inconsistencies in existing benchmarks. By modeling factuality as a dynamic property of conversation, VISTA offers a more transparent, human-aligned measure of truthfulness in dialogue systems.

Method: Introduced a perspectivist annotation scheme capturing speaker and addressee interpretations separately, used an LLM annotation pipeline to label 13k reference expressions in the HCRC MapTask corpus, and analyzed understanding states and discrepancies.

Result: Full misunderstandings are rare once lexical variants are unified, but multiplicity discrepancies systematically induce divergences, revealing how apparent grounding can mask referential misalignment.

Conclusion: The framework provides both a resource and analytic lens for studying grounded misunderstanding and evaluating (V)LLMs’ capacity to model perspective-dependent grounding in collaborative dialogue.

Abstract: Collaborative dialogue relies on participants incrementally establishing common ground, yet in asymmetric settings they may believe they agree while referring to different entities. We introduce a perspectivist annotation scheme for the HCRC MapTask corpus (Anderson et al., 1991) that separately captures speaker and addressee grounded interpretations for each reference expression, enabling us to trace how understanding emerges, diverges, and repairs over time. Using a scheme-constrained LLM annotation pipeline, we obtain 13k annotated reference expressions with reliability estimates and analyze the resulting understanding states. The results show that full misunderstandings are rare once lexical variants are unified, but multiplicity discrepancies systematically induce divergences, revealing how apparent grounding can mask referential misalignment. Our framework provides both a resource and an analytic lens for studying grounded misunderstanding and for evaluating (V)LLMs’ capacity to model perspective-dependent grounding in collaborative dialogue.

Method: Introduces T-FIX benchmark spanning 7 scientific tasks across 3 domains with automatic evaluation framework that generalizes to unseen explanations without ongoing expert involvement

Result: Framework enables automatic evaluation of expert alignment, operationalizing expert reasoning as a measurable attribute of LLM-generated explanations

Conclusion: T-FIX provides scalable solution for evaluating whether LLMs think like domain experts, addressing critical gap in professional reasoning assessment

Abstract: As LLMs are deployed in knowledge-intensive settings (e.g., surgery, astronomy, therapy), users are often domain experts who expect not just answers, but explanations that mirror professional reasoning. However, most automatic evaluations of explanations prioritize plausibility or faithfulness, rather than testing whether an LLM thinks like an expert. Existing approaches to evaluating professional reasoning rely heavily on per-example expert annotation, making such evaluations costly and difficult to scale. To address this gap, we introduce the T-FIX benchmark, spanning seven scientific tasks across three domains, to operationalize expert alignment as a desired attribute of LLM-generation explanations. Our framework enables automatic evaluation of expert alignment, generalizing to unseen explanations and eliminating the need for ongoing expert involvement.

Method: IDALC combines intent detection with active learning-based correction. It identifies user intents and rectifies system-rejected utterances while selectively querying human annotators for the most informative samples to minimize annotation effort.

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

Conclusion: IDALC provides an efficient semi-supervised framework for intent detection and correction that significantly reduces annotation costs while improving performance on system-rejected utterances.

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: Proposes DIA-REFINE framework with iterative translation, verification, and feedback loop using external dialect classifiers. Introduces dialect fidelity score (DFS) and target dialect ratio (TDR) metrics.

Result: DIA-REFINE consistently enhances dialect fidelity across Korean dialects. New metrics distinguish between False Success (high n-gram but poor dialect) and True Attempt (low n-gram but genuine dialect) cases.

Conclusion: Establishes robust framework for goal-directed, inclusive dialect translation with rigorous evaluation and insights into model performance, showing in-context examples further improve dialect expression translation.

Abstract: Standard-to-dialect machine translation remains challenging due to a persistent dialect gap in large language models and evaluation distortions inherent in n-gram metrics, which favor source copying over authentic dialect translation. In this paper, we propose the dialect refinement (DIA-REFINE) framework, which guides LLMs toward faithful target dialect outputs through an iterative loop of translation, verification, and feedback using external dialect classifiers. To address the limitations of n-gram-based metrics, we introduce the dialect fidelity score (DFS) to quantify linguistic shift and the target dialect ratio (TDR) to measure the success of dialect translation. Experiments on Korean dialects across zero-shot and in-context learning baselines demonstrate that DIA-REFINE consistently enhances dialect fidelity. The proposed metrics distinguish between False Success cases, where high n-gram scores obscure failures in dialectal translation, and True Attempt cases, where genuine attempts at dialectal translation yield low n-gram scores. We also observed that models exhibit varying degrees of responsiveness to the framework, and that integrating in-context examples further improves the translation of dialectal expressions. Our work establishes a robust framework for goal-directed, inclusive dialect translation, providing both rigorous evaluation and critical insights into model performance.

Method: Used Agent Forest framework with sampling-and-voting to evaluate six open-source LLMs across four math benchmarks with various agent counts (1-25). Tested three punctuation noise intensities (10%, 30%, 50%) and two typo datasets (WikiTypo, R2ATA).

Result: Collaboration improves accuracy as agent count increases (largest gains from 1 to 5 agents), but adversarial robustness gaps persist regardless of agent count. Human-like typos remain the dominant bottleneck with highest attack success rates.

Conclusion: While multi-agent collaboration enhances mathematical reasoning performance, it does not solve the fundamental adversarial robustness problem, especially for human-like perturbations that remain challenging even with many agents.

Abstract: When LLM agents work together, they seem to be more powerful than a single LLM in mathematical question answering. However, are they also more robust to adversarial inputs? We investigate this question using adversarially perturbed math questions. These perturbations include punctuation noise with three intensities (10%, 30%, 50%), plus real-world and human-like typos (WikiTypo, R2ATA). Using a unified sampling-and-voting framework (Agent Forest), we evaluate six open-source models (Qwen3-4B/14B, Llama3.1-8B, Mistral-7B, Gemma3-4B/12B) across four benchmarks (GSM8K, MATH, MMLU-Math, MultiArith), with various numbers of agents n = {1,2,5,10,15,20,25}. Our findings show that 1) Noise type matters: punctuation noise harm scales with its severity, and the human typos remain the dominant bottleneck, yielding the largest gaps to Clean accuracy and the highest attack success rate (ASR) even with a large number of agents; 2) Collaboration reliably improves accuracy as the number of agents, n, increases, with the largest gains from n=1 to n=5 and diminishing returns beyond n$\approx$10. However, the adversarial robustness gap persists regardless of the agent count.

Method: Created MedPT corpus through multi-stage curation: collected 384,095 authentic patient-doctor Q&A pairs covering 3,200+ health conditions, used hybrid quantitative-qualitative analysis to filter noise, contextually enriched ambiguous queries, employed LLM-driven annotation to classify queries into seven semantic types, and benchmarked with medical specialty classification tasks.

Result: Fine-tuning a 1.7B parameter model achieved 94% F1-score on 20-class medical specialty classification. Error analysis showed misclassifications reflect genuine clinical ambiguities (e.g., comorbid conditions), demonstrating dataset’s semantic richness. Corpus contains ~57 million tokens.

Conclusion: MedPT enables development of equitable, accurate, culturally-aware medical technologies for Portuguese-speaking world. The dataset captures unique clinical and cultural nuances that translation-based approaches miss, supporting more inclusive healthcare AI.

Abstract: While large language models (LLMs) show transformative potential in healthcare, their development remains focused on high-resource languages. This creates a critical barrier for other languages, as simple translation fails to capture unique clinical and cultural nuances, such as endemic diseases. To address this, we introduce MedPT, the first large-scale, real-world corpus of patient-doctor interactions for the Brazilian Portuguese medical domain. Comprising 384,095 authentic question-answer pairs and covering over 3,200 distinct health-related conditions, the dataset was refined through a rigorous multi-stage curation protocol that employed a hybrid quantitative-qualitative analysis to filter noise and contextually enrich thousands of ambiguous queries, resulting in a corpus of approximately 57 million tokens. We further utilize of LLM-driven annotation to classify queries into seven semantic types to capture user intent. To validate MedPT’s utility, we benchmark it in a medical specialty classification task: fine-tuning a 1.7B parameter model achieves an outstanding 94% F1-score on a 20-class setup. Furthermore, our qualitative error analysis shows misclassifications are not random but reflect genuine clinical ambiguities (e.g., between comorbid conditions), proving the dataset’s deep semantic richness. We publicly release MedPT on Hugging Face to support the development of more equitable, accurate, and culturally-aware medical technologies for the Portuguese-speaking world.

Method: Proposes LabelFusion - a hybrid architecture combining prompt-engineered LLM outputs with contextual embeddings from fine-tuned RoBERTa encoder through a lightweight MLP voting layer for multi-label classification.

Result: LabelFusion achieves 96.0% macro F1 and 92.3% accuracy on full Reuters-21578 dataset, outperforming standalone RoBERTa (94.6%) and LLM (93.9%). LLM alone performs well in low-data regimes (75.9% F1 zero-shot).

Conclusion: LLM-only prompting is preferred under annotation constraints, while LabelFusion becomes most effective with sufficient labeled data. Hybrid approaches leverage strengths of both LLMs and fine-tuned encoders.

Abstract: Financial news plays a central role in shaping investor sentiment and short-term dynamics in commodity markets. Many downstream financial applications, such as commodity price prediction or sentiment modeling, therefore rely on the ability to automatically identify news articles relevant to specific assets. However, obtaining large labeled corpora for financial text classification is costly, and transformer-based classifiers such as RoBERTa often degrade significantly in low-data regimes. Our results show that appropriately prompted out-of-the-box Large Language Models (LLMs) achieve strong performance even in such settings. Furthermore, we propose LabelFusion, a hybrid architecture that combines the output of a prompt-engineered LLM with contextual embeddings produced by a fine-tuned RoBERTa encoder through a lightweight Multilayer Perceptron (MLP) voting layer. Evaluated on a ten-class multi-label subset of the Reuters-21578 corpus, LabelFusion achieves a macro F1 score of 96.0% and an accuracy of 92.3% when trained on the full dataset, outperforming both standalone RoBERTa (F1 94.6%) and the standalone LLM (F1 93.9%). In low- to mid-data regimes, however, the LLM alone proves surprisingly competitive, achieving an F1 score of 75.9% even in a zero-shot setting and consistently outperforming LabelFusion until approximately 80% of the training data is available. These results suggest that LLM-only prompting is the preferred strategy under annotation constraints, whereas LabelFusion becomes the most effective solution once sufficient labeled data is available to train the encoder component. The code is available in an anonymized repository.

Method: Frames next-token prediction as a stochastic decision process and introduces a reward-shaping strategy that balances diversity and precision. Uses positive reward scaling to control probability concentration on ground-truth tokens and a rank-aware mechanism that treats high-ranking and low-ranking negative tokens asymmetrically.

Result: Contrary to intuition that higher distribution entropy facilitates effective exploration, the method finds that imposing a precision-oriented prior yields a superior exploration space for RL, ultimately enhancing end-to-end reasoning performance.

Conclusion: The proposed generalized pre-training objective successfully reshapes token-output distributions to provide more favorable exploration spaces for RL training, leading to improved reasoning abilities in LLMs.

Abstract: Recent advancements have shown that reinforcement learning (RL) can substantially improve the reasoning abilities of large language models (LLMs). The effectiveness of such RL training, however, depends critically on the exploration space defined by the pre-trained model’s token-output distribution. In this paper, we revisit the standard cross-entropy loss, interpreting it as a specific instance of policy gradient optimization applied within a single-step episode. To systematically study how the pre-trained distribution shapes the exploration potential for subsequent RL, we propose a generalized pre-training objective that adapts on-policy RL principles to supervised learning. By framing next-token prediction as a stochastic decision process, we introduce a reward-shaping strategy that explicitly balances diversity and precision. Our method employs a positive reward scaling factor to control probability concentration on ground-truth tokens and a rank-aware mechanism that treats high-ranking and low-ranking negative tokens asymmetrically. This allows us to reshape the pre-trained token-output distribution and investigate how to provide a more favorable exploration space for RL, ultimately enhancing end-to-end reasoning performance. Contrary to the intuition that higher distribution entropy facilitates effective exploration, we find that imposing a precision-oriented prior yields a superior exploration space for RL.

Method: Use maximum likelihood approach to estimate temperature parameter for any text with respect to a language model. Evaluate temperature estimation capability across various small-to-medium LLMs, then apply best-performing model (Qwen3 14B) to analyze popular corpora.

Result: Most measured temperatures in popular corpora are close to 1, but notable exceptions include Jokes, GSM8K, and AG News (1.1), and Python code (0.9). Qwen3 14B performed best among evaluated models.

Conclusion: Proposed method successfully estimates temperature of text relative to language models, revealing systematic differences in temperature across text types, with potential applications in text analysis and model evaluation.

Abstract: Autoregressive language models typically use temperature parameter at inference to shape the probability distribution and control the randomness of the text generated. After the text was generated, this parameter can be estimated using maximum likelihood approach. Following it, we propose a procedure to estimate the temperature of any text, including ones written by humans, with respect to a given language model. We evaluate the temperature estimation capability of a wide selection of small-to-medium Large Language Models (LLMs). We then use the best-performing Qwen3 14B to estimate temperatures of popular corpora, finding that while most measured temperatures are close to 1, notable exceptions include Jokes, GSM8K, and AG News (1.1), and Python code (0.9).

Method: Two pragmatic inference methods that facilitate LLMs to: 1) diagnose morally benign and hazardous input, and 2) correct moral errors. The methods offer a unified perspective by designing pragmatic inference procedures grounded in their inferential loads rather than modeling diverse surface forms.

Result: Empirical evidence demonstrates that the pragmatic methods can enhance moral sensitivity in LLMs and achieve strong performance on representative morality-relevant benchmarks.

Conclusion: The proposed pragmatic inference methods provide a principled approach to enhancing moral sensitivity in LLMs, addressing a fundamental challenge in aligning language models with human moral values.

Abstract: Moral sensitivity is fundamental to human moral competence, as it guides individuals in regulating everyday behavior. Although many approaches seek to align large language models (LLMs) with human moral values, how to enable them morally sensitive has been extremely challenging. In this paper, we take a step toward answering the question: how can we enhance moral sensitivity in LLMs? Specifically, we propose two pragmatic inference methods that faciliate LLMs to diagnose morally benign and hazardous input and correct moral errors, whereby enhancing LLMs’ moral sensitivity. A central strength of our pragmatic inference methods is their unified perspective: instead of modeling moral discourses across semantically diverse and complex surface forms, they offer a principled perspective for designing pragmatic inference procedures grounded in their inferential loads. Empirical evidence demonstrates that our pragmatic methods can enhance moral sensitivity in LLMs and achieves strong performance on representative morality-relevant benchmarks.

Method: The benchmark employs dynamic evaluation: instructions are sampled from template pools, numeric parameters are drawn from predefined intervals, and validators verify outcomes against instantiated values. It contains 107 tasks (62 atomic, 45 composite) with modular architecture for rapid task development. The runner executes scripts on a forked EVM chain with snapshot isolation, and composite tasks apply step-efficiency decay.

Result: Evaluation of 20 models reveals large performance gaps, with split scores showing persistent asymmetry between single-action precision and multi-step workflow completion.

Conclusion: EVM-QuestBench addresses critical gaps in evaluating language models for blockchain development by focusing on execution accuracy and safety, revealing significant performance disparities in transaction-script generation tasks.

Abstract: Large language models are increasingly applied to various development scenarios. However, in on-chain transaction scenarios, even a minor error can cause irreversible loss for users. Existing evaluations often overlook execution accuracy and safety. We introduce EVM-QuestBench, an execution-grounded benchmark for natural-language transaction-script generation on EVM-compatible chains. The benchmark employs dynamic evaluation: instructions are sampled from template pools, numeric parameters are drawn from predefined intervals, and validators verify outcomes against these instantiated values. EVM-QuestBench contains 107 tasks (62 atomic, 45 composite). Its modular architecture enables rapid task development. The runner executes scripts on a forked EVM chain with snapshot isolation; composite tasks apply step-efficiency decay. We evaluate 20 models and find large performance gaps, with split scores revealing persistent asymmetry between single-action precision and multi-step workflow completion. Code: https://anonymous.4open.science/r/bsc_quest_bench-A9CF/.

Method: Treat training on each language as a separate task, generate task vectors by fine-tuning Whisper ASR variants, and merge vectors via linear combinations optimized on low-resource language validation sets using word error rate.

Result: Consistent word error rate improvements of up to 10% across 23 low-resource target languages compared to baselines without the approach.

Conclusion: Task arithmetic is an effective technique for improving ASR performance in low-resource languages by transferring knowledge from related high-resource languages.

Abstract: The development of resource-constrained approaches to automatic speech recognition (ASR) is of great interest due to its broad applicability to many low-resource languages for which there is scant usable data. Existing approaches to many low-resource natural language processing tasks leverage additional data from higher-resource languages that are closely related to a target low-resource language. One increasingly popular approach uses task arithmetic to combine models trained on different tasks to create a model for a task where there is little to no training data. In this paper, we consider training on a particular language to be a task, and we generate task vectors by fine-tuning variants of the Whisper ASR system. For pairs of high- and low-resource languages, we merge task vectors via a linear combination which is optimized on the downstream word error rate on the low-resource target language’s validation set. Across 23 low-resource target languages for which we evaluate this technique, we find consistent word error rate improvements of up to 10% compared to a baseline without our approach.

Method: Proposes Imagine-then-Plan (ITP) framework where policy interacts with learned world model to generate multi-step imagined trajectories. Introduces adaptive lookahead mechanism that trades off ultimate goal and task progress. Formulates partially observable and imaginable Markov decision process by fusing imagined trajectories with current observations.

Result: Extensive experiments across representative agent benchmarks show ITP significantly outperforms competitive baselines. Adaptive lookahead enhances agents’ reasoning capability and provides insights for addressing broader, complex tasks.

Conclusion: ITP provides a unified framework for agent learning via lookahead imagination, demonstrating superior performance through adaptive horizon control and integration of imagined future consequences with current observations.

Abstract: Recent advances in world models have shown promise for modeling future dynamics of environmental states, enabling agents to reason and act without accessing real environments. Current methods mainly perform single-step or fixed-horizon rollouts, leaving their potential for complex task planning under-exploited. We propose Imagine-then-Plan (\texttt{ITP}), a unified framework for agent learning via lookahead imagination, where an agent’s policy model interacts with the learned world model, yielding multi-step ``imagined’’ trajectories. Since the imagination horizon may vary by tasks and stages, we introduce a novel adaptive lookahead mechanism by trading off the ultimate goal and task progress. The resulting imagined trajectories provide rich signals about future consequences, such as achieved progress and potential conflicts, which are fused with current observations, formulating a partially \textit{observable} and \textit{imaginable} Markov decision process to guide policy learning. We instantiate \texttt{ITP} with both training-free and reinforcement-trained variants. Extensive experiments across representative agent benchmarks demonstrate that \texttt{ITP} significantly outperforms competitive baselines. Further analyses validate that our adaptive lookahead largely enhances agents’ reasoning capability, providing valuable insights into addressing broader, complex tasks. Our code and data will be publicly available at https://github.com/loyiv/ITP.

Method: Multi-agent LLM framework with specialized agents: one extracts explicit limitations, another analyzes methodological gaps, a third simulates peer reviewer perspective, and a citation agent places work within broader literature. A Judge agent refines outputs and a Master agent consolidates them into clear limitations. Uses pointwise evaluation with LLM-as-a-Judge instead of traditional NLP metrics.

Result: The RAG + multi-agent GPT-4o mini configuration achieves +15.51% coverage gain over zero-shot baselines, while Llama 3 8B multi-agent setup yields +4.41% improvement.

Conclusion: The proposed multi-agent framework effectively generates substantive limitations by systematically identifying explicit, implicit, peer review-focused, and literature-informed limitations, outperforming traditional approaches.

Abstract: Identifying and articulating limitations is essential for transparent and rigorous scientific research. However, zero-shot large language models (LLMs) approach often produce superficial or general limitation statements (e.g., dataset bias or generalizability). They usually repeat limitations reported by authors without looking at deeper methodological issues and contextual gaps. This problem is made worse because many authors disclose only partial or trivial limitations. We propose, a multi-agent LLM framework for generating substantive limitations. It integrates OpenReview comments and author-stated limitations to provide stronger ground truth. It also uses cited and citing papers to capture broader contextual weaknesses. In this setup, different agents have specific roles as sequential role: some extract explicit limitations, others analyze methodological gaps, some simulate the viewpoint of a peer reviewer, and a citation agent places the work within the larger body of literature. A Judge agent refines their outputs, and a Master agent consolidates them into a clear set. This structure allows for systematic identification of explicit, implicit, peer review-focused, and literature-informed limitations. Moreover, traditional NLP metrics like BLEU, ROUGE, and cosine similarity rely heavily on n-gram or embedding overlap. They often overlook semantically similar limitations. To address this, we introduce a pointwise evaluation protocol that uses an LLM-as-a-Judge to measure coverage more accurately. Experiments show that our proposed model substantially improve performance. The RAG + multi-agent GPT-4o mini configuration achieves a +15.51% coverage gain over zero-shot baselines, while the Llama 3 8B multi-agent setup yields a +4.41% improvement.

Method: Proposes Jacobian Scopes - a suite of gradient-based, token-level causal attribution methods grounded in perturbation theory and information geometry. These methods quantify how input tokens influence specific aspects of predictions including logits, predictive distributions, and model uncertainty (effective temperature).

Result: Demonstrated through case studies spanning instruction understanding, translation, and in-context learning. Revealed implicit political biases, uncovered word- and phrase-level translation strategies, and provided insights into mechanisms underlying in-context time-series forecasting.

Conclusion: Jacobian Scopes provide effective tools for interpreting LLM predictions at token level, offering insights into model behavior across various tasks. The methods are made accessible through open-source implementations and an interactive demo.

Abstract: Large language models (LLMs) make next-token predictions based on clues present in their context, such as semantic descriptions and in-context examples. Yet, elucidating which prior tokens most strongly influence a given prediction remains challenging due to the proliferation of layers and attention heads in modern architectures. We propose Jacobian Scopes, a suite of gradient-based, token-level causal attribution methods for interpreting LLM predictions. Grounded in perturbation theory and information geometry, Jacobian Scopes quantify how input tokens influence various aspects of a model’s prediction, such as specific logits, the full predictive distribution, and model uncertainty (effective temperature). Through case studies spanning instruction understanding, translation, and in-context learning (ICL), we demonstrate how Jacobian Scopes reveal implicit political biases, uncover word- and phrase-level translation strategies, and shed light on recently debated mechanisms underlying in-context time-series forecasting. To facilitate exploration of Jacobian Scopes on custom text, we open-source our implementations and provide a cloud-hosted interactive demo at https://huggingface.co/spaces/Typony/JacobianScopes.

Method: Benchmarked 17 state-of-the-art LLM agents in 2-5 player Hanabi games with three context engineering settings: minimal prompt (Watson), programmatic deductions (Sherlock), and multi-turn state tracking (Mycroft). Created two datasets for finetuning: HanabiLogs (1,520 game logs) and HanabiRewards (560 games with move-level annotations).

Result: Strongest reasoning models exceeded 15 points in Sherlock setting but trailed humans (20+ points). Finetuning Qwen3-Instruct improved performance by 21% (supervised) and 156% (RL), bringing it within ~3 points of o4-mini and surpassing GPT-4.1 by 52%. RL-finetuned model generalized to other tasks: improved group-guessing by 11%, EventQA by 6.4%, IFBench-800K by 1.7 Pass@10, and matched AIME 2025 math reasoning.

Conclusion: Context engineering and finetuning significantly improve LLMs’ cooperative reasoning in Hanabi, with RL-finetuned models showing strong generalization to other reasoning tasks beyond the game.

Abstract: Cooperative reasoning under incomplete information remains challenging for both humans and multi-agent systems. The card game Hanabi embodies this challenge, requiring theory-of-mind reasoning and strategic communication. We benchmark 17 state-of-the-art LLM agents in 2-5 player games and study the impact of context engineering across model scales (4B to 600B+) to understand persistent coordination failures and robustness to scaffolding: from a minimal prompt with only explicit card details (Watson setting), to scaffolding with programmatic, Bayesian-motivated deductions (Sherlock setting), to multi-turn state tracking via working memory (Mycroft setting). We show that (1) agents can maintain an internal working memory for state tracking and (2) cross-play performance between different LLMs smoothly interpolates with model strength. In the Sherlock setting, the strongest reasoning models exceed 15 points on average across player counts, yet still trail experienced humans and specialist Hanabi agents, both consistently scoring above 20. We release the first public Hanabi datasets with annotated trajectories and move utilities: (1) HanabiLogs, containing 1,520 full game logs for instruction tuning, and (2) HanabiRewards, containing 560 games with dense move-level value annotations for all candidate moves. Supervised and RL finetuning of a 4B open-weight model (Qwen3-Instruct) on our datasets improves cooperative Hanabi play by 21% and 156% respectively, bringing performance to within ~3 points of a strong proprietary reasoning model (o4-mini) and surpassing the best non-reasoning model (GPT-4.1) by 52%. The HanabiRewards RL-finetuned model further generalizes beyond Hanabi, improving performance on a cooperative group-guessing benchmark by 11%, temporal reasoning on EventQA by 6.4%, instruction-following on IFBench-800K by 1.7 Pass@10, and matching AIME 2025 mathematical reasoning Pass@10.

Method: Created BabyReasoningBench with 19 reasoning tasks from developmental psychology (theory of mind, analogical reasoning, causal inference, etc.). Tested two GPT-2 based models pretrained on 10M and 100M tokens of child-directed speech text.

Result: Models showed low but uneven performance with dissociations across tasks: scaling improved causal and physical reasoning, but belief attribution and pragmatics-sensitive tasks remained challenging.

Conclusion: BabyReasoningBench provides a developmentally grounded framework for analyzing reasoning emergence from child-like training data and testing mechanistic hypotheses about cognitive development in language models.

Abstract: Traditional evaluations of reasoning capabilities of language models are dominated by adult-centric benchmarks that presuppose broad world knowledge, complex instruction following, and mature pragmatic competence. These assumptions are mismatched to baby language models trained on developmentally plausible input such as child-directed speech and early-childhood narratives, and they obscure which reasoning abilities (if any) emerge under such constraints. We introduce BabyReasoningBench, a GPT-5.2 generated benchmark of 19 reasoning tasks grounded in classic paradigms from developmental psychology, spanning theory of mind, analogical and relational reasoning, causal inference and intervention selection, and core reasoning primitives that are known to be confounded by memory and pragmatics. We find that two GPT-2 based baby language models (pretrained on 10M and 100M of child-directed speech text) show overall low but uneven performance, with dissociations across task families: scaling improves several causal and physical reasoning tasks, while belief attribution and pragmatics-sensitive tasks remain challenging. BabyReasoningBench provides a developmentally grounded lens for analyzing what kinds of reasoning are supported by child-like training distributions, and for testing mechanistic hypotheses about how such abilities emerge.

Method: Three-phase blind longitudinal study with linguistically trained annotators: Phase 1 measures intuition-only attribution, Phase 2 introduces criterion-anchored scoring with explicit justifications, and Phase 3 evaluates a limited held-out elementary-persona subset using majority-vote accuracy.

Result: Majority-vote accuracy improved from 0.60 in Phase 1 to 0.90 in Phase 2, reaching 10/10 on the limited Phase 3 subset. Inter-annotator agreement increased from Fleiss’ κ = -0.09 to 0.82. Calibration primarily reduces false negatives on AI essays rather than inducing generalized over-detection.

Conclusion: LREAD provides pilot evidence for within-panel calibration in Korean argumentative-essay settings. Rubric-scaffolded human judgment can complement automated detectors by making attribution reasoning explicit, auditable, and adaptable.

Abstract: Distinguishing human-written Korean text from fluent LLM outputs remains difficult even for trained readers, who can over-trust surface well-formedness. We present LREAD, a Korean-specific instantiation of a rubric-based expert-calibration framework for human attribution of LLM-generated text. In a three-phase blind longitudinal study with three linguistically trained annotators, Phase 1 measures intuition-only attribution, Phase 2 introduces criterion-anchored scoring with explicit justifications, and Phase 3 evaluates a limited held-out elementary-persona subset. Majority-vote accuracy improves from 0.60 in Phase 1 to 0.90 in Phase 2, and reaches 10/10 on the limited Phase 3 subset (95% CI [0.692, 1.000]); agreement also increases from Fleiss’ $κ$ = -0.09 to 0.82. Error analysis suggests that calibration primarily reduces false negatives on AI essays rather than inducing generalized over-detection. We position LREAD as pilot evidence for within-panel calibration in a Korean argumentative-essay setting. These findings suggest that rubric-scaffolded human judgment can complement automated detectors by making attribution reasoning explicit, auditable, and adaptable.

Method: Developed MDial framework using rule-based LLM transformation with native linguist annotations to generate dialect data covering lexical, orthographic, and morphosyntactic features for 9 English dialects. Created MDialBenchmark with 50k+ dialogs (97k+ QA pairs) and evaluated 17 LLMs on dialect identification and response generation.

Result: Even frontier models achieve under 70% accuracy on dialect identification, fail to reach 50% for Canadian English, and systematically misclassify non-SAE dialects as American or British. Annotators preferred MDial outputs over prior methods in 98% of comparisons for dialect naturalness.

Conclusion: LLMs have significant limitations in handling dialectal variation, with dialect identification errors risking cascading failures in downstream NLU tasks. The research challenges assumptions about model behavior and provides a scalable framework for improving multi-dialectal AI.

Abstract: More than 80% of the 1.6 billion English speakers do not use Standard American English (SAE) and experience higher failure rates and stereotyped responses when interacting with LLMs as a result. Yet multi-dialectal performance remains underexplored. We introduce MDial, the first large-scale framework for generating multi-dialectal conversational data encompassing the three pillars of written dialect – lexical (vocabulary), orthographic (spelling), and morphosyntactic (grammar) features – for nine English dialects. Partnering with native linguists, we design an annotated and scalable rule-based LLM transformation to ensure precision. Our approach challenges the assumption that models should mirror users’ morphosyntactic features, showing that up to 90% of the grammatical features of a dialect should not be reproduced by models. Independent evaluations confirm data quality, with annotators preferring MDial outputs over prior methods in 98% of pairwise comparisons for dialect naturalness. Using this pipeline, we construct the dialect-parallel MDialBenchmark with 50k+ dialogs, resulting in 97k+ QA pairs, and evaluate 17 LLMs on dialect identification and response generation tasks. Even frontier models achieve under 70% accuracy, fail to reach 50% for Canadian English, and systematically misclassify non-SAE dialects as American or British. As dialect identification underpins natural language understanding, these errors risk cascading failures into downstream tasks.

Method: CRAFT uses RL to train models to produce structured reasoning traces with configurable auditability levels (planning, evidence citation, reasoning steps). Combines deterministic rewards (format compliance, answer correctness, citation validity) with judge-based rewards that audit semantic faithfulness (reasoning consistency, evidence grounding).

Result: CRAFT improves both answer accuracy and reasoning faithfulness across model scales. Semantic judge-based rewards improve answer accuracy rather than compromise it, enabling CRAFT (7B) to achieve performance competitive with strong closed-source models.

Conclusion: CRAFT addresses the right-answer-wrong-reason problem in retrieval-augmented QA by enforcing structured, auditable reasoning through combined deterministic and semantic rewards, enabling both accurate answers and faithful reasoning.

Abstract: Retrieval-augmented large language models, when optimized with outcome-level rewards, can achieve strong answer accuracy on multi-hop questions. However, under noisy retrieval, models frequently suffer from “right-answer-wrong-reason failures”: they may exploit spurious shortcuts or produce reasoning traces weakly grounded in the supporting evidence. Furthermore, the lack of structured output control prevents reliable auditing of the underlying reasoning quality. To address this, we propose CRAFT (Calibrated Reasoning with Answer-Faithful Traces), a reinforcement learning framework for the response generation stage of retrieval-augmented multi-hop question answering. CRAFT trains models to produce structured reasoning traces with configurable levels of auditability (e.g., by selectively retaining planning, evidence citation, or reasoning steps). Training combines two complementary forms of supervision: deterministic rewards enforce verifiable constraints, including format compliance, answer correctness, and citation-set validity, while a judge-based reward audits semantic faithfulness by evaluating reasoning consistency and evidence grounding. Experiments show that CRAFT improves both answer accuracy and reasoning faithfulness across model scales. Notably, semantic judge-based rewards improve answer accuracy rather than compromise it, enabling CRAFT (7B) to achieve performance competitive with strong closed-source models.

Method: Created ChemPro benchmark with 4100 natural language QA pairs across 4 difficulty sections covering Biochemistry, Inorganic, Organic, and Physical Chemistry. Includes multiple choice and numerical questions with balanced ratios of fine-grained recall, long-horizon reasoning, multi-concept questions, and nuanced problem-solving. Evaluated 45+7 state-of-the-art LLMs (open-source and proprietary).

Result: LLMs perform well on basic chemistry questions but accuracy declines significantly with different types and levels of complexity. The benchmark reveals critical limitations in LLMs’ general scientific reasoning and understanding.

Conclusion: Current LLMs have significant limitations in handling complex chemistry reasoning tasks. The findings highlight understudied dimensions of difficulty and emphasize the need for more robust methodologies to improve LLMs’ scientific reasoning capabilities.

Abstract: We introduce ChemPro, a progressive benchmark with 4100 natural language question-answer pairs in Chemistry, across 4 coherent sections of difficulty designed to assess the proficiency of Large Language Models (LLMs) in a broad spectrum of general chemistry topics. We include Multiple Choice Questions and Numerical Questions spread across fine-grained information recall, long-horizon reasoning, multi-concept questions, problem-solving with nuanced articulation, and straightforward questions in a balanced ratio, effectively covering Bio-Chemistry, Inorganic-Chemistry, Organic-Chemistry and Physical-Chemistry. ChemPro is carefully designed analogous to a student’s academic evaluation for basic to high-school chemistry. A gradual increase in the question difficulty rigorously tests the ability of LLMs to progress from solving basic problems to solving more sophisticated challenges. We evaluate 45+7 state-of-the-art LLMs, spanning both open-source and proprietary variants, and our analysis reveals that while LLMs perform well on basic chemistry questions, their accuracy declines with different types and levels of complexity. These findings highlight the critical limitations of LLMs in general scientific reasoning and understanding and point towards understudied dimensions of difficulty, emphasizing the need for more robust methodologies to improve LLMs.

Method: Proposes Dynamic Sliding Block (DSB), a training-free block scheduling method using sliding blocks with dynamic sizes. Also introduces DSB Cache, a training-free KV-cache mechanism tailored to DSB.

Result: Extensive experiments across multiple models and benchmarks show DSB with DSB Cache consistently improves both generation quality and inference efficiency for diffusion LLMs.

Conclusion: Dynamic adaptation of block scheduling to semantic difficulty is crucial for reliable and efficient diffusion LLM inference, with DSB providing an effective training-free solution.

Abstract: Diffusion large language models (dLLMs) have emerged as a promising alternative for text generation, distinguished by their native support for parallel decoding. In practice, block inference is crucial for avoiding order misalignment in global bidirectional decoding and improving output quality. However, the widely-used fixed, predefined block (naive) schedule is agnostic to semantic difficulty, making it a suboptimal strategy for both quality and efficiency: it can force premature commitments to uncertain positions while delaying easy positions near block boundaries. In this work, we analyze the limitations of naive block scheduling and disclose the importance of dynamically adapting the schedule to semantic difficulty for reliable and efficient inference. Motivated by this, we propose Dynamic Sliding Block (DSB), a training-free block scheduling method that uses a sliding block with a dynamic size to overcome the rigidity of the naive block. To further improve efficiency, we introduce DSB Cache, a training-free KV-cache mechanism tailored to DSB. Extensive experiments across multiple models and benchmarks demonstrate that DSB, together with DSB Cache, consistently improves both generation quality and inference efficiency for dLLMs. Code is released at https://github.com/lizhuo-luo/DSB.

Method: Aggregated 651 translations from 12 online biblical libraries and one preexisting corpus, covering English (194 unique), French (41), Italian (17), Polish (29), and Spanish (53) versions. Each translation is annotated with metadata for canonicalization.

Result: Created the first multilingual resource with 2.4-5.0x more translations per language than any prior corpus, enabling flexible multilevel analysis of translation history.

Conclusion: The corpus fills a gap in quantitative translation history research by providing sufficient depth per language for both micro-level (translation families) and macro-level studies.

Abstract: Many European languages possess rich biblical translation histories, yet existing corpora - in prioritizing linguistic breadth - often fail to capture this depth. To address this gap, we introduce a multilingual corpus of 651 New Testament translations, of which 334 are unique, spanning five languages with 2.4-5.0x more translations per language than any prior corpus: English (194 unique versions from 390 total), French (41 from 78), Italian (17 from 33), Polish (29 from 48), and Spanish (53 from 102). Aggregated from 12 online biblical libraries and one preexisting corpus, each translation is annotated with metadata that maps the text to a standardized identifier for the work, its specific edition, and its year of revision. This canonicalization allows researchers to define “uniqueness” for their own needs: they can perform micro-level analyses on translation families, such as the KJV lineage, or conduct macro-level studies by deduplicating closely related texts. By providing the first multilingual resource with sufficient depth per language for flexible, multilevel analysis, the corpus fills a gap in the quantitative study of translation history.

Method: Created Multimodal Finance Eval benchmark with 1,204 expert-validated questions from real French financial documents, evaluating six open-weight VLMs (8B-124B parameters) using LLM-as-judge protocol across text extraction, table comprehension, chart interpretation, and multi-turn conversational reasoning.

Result: VLMs achieve 85-90% accuracy on text and table tasks but only 34-62% on chart interpretation. Multi-turn dialogue shows severe failure: early mistakes propagate, reducing accuracy to ~50% regardless of model size.

Conclusion: Current VLMs are effective for well-defined extraction tasks but brittle in interactive, multi-step financial analysis. The benchmark provides a challenging testbed for progress in high-stakes multimodal understanding.

Abstract: Vision-language models (VLMs) perform well on many document understanding tasks, yet their reliability in specialized, non-English domains remains underexplored. This gap is especially critical in finance, where documents mix dense regulatory text, numerical tables, and visual charts, and where extraction errors can have real-world consequences. We introduce Multimodal Finance Eval, the first multimodal benchmark for evaluating French financial document understanding. The dataset contains 1,204 expert-validated questions spanning text extraction, table comprehension, chart interpretation, and multi-turn conversational reasoning, drawn from real investment prospectuses, KIDs, and PRIIPs. We evaluate six open-weight VLMs (8B-124B parameters) using an LLM-as-judge protocol. While models achieve strong performance on text and table tasks (85-90% accuracy), they struggle with chart interpretation (34-62%). Most notably, multi-turn dialogue reveals a sharp failure mode: early mistakes propagate across turns, driving accuracy down to roughly 50% regardless of model size. These results show that current VLMs are effective for well-defined extraction tasks but remain brittle in interactive, multi-step financial analysis. Multimodal Finance Eval offers a challenging benchmark to measure and drive progress in this high-stakes setting.

Method: Creates small, handcrafted test sets that exhaustively cover span attributes (shape, length, casing, sentence position, etc.) a system may encounter. Uses these diagnostic test sets to analyze errors systematically and predict performance on external datasets.

Result: The methodology provides diagnostic results that identify systematic weaknesses, actionable insights for model selection, and predictive capability with median correlation of 0.85 for predicting model performance on external datasets.

Conclusion: Proposed evaluation methodology offers more informative, diagnostic, and predictive assessment than standard average performance metrics for sequence labeling tasks, enabling better understanding of model capabilities and limitations.

Abstract: Standard evaluation in NLP typically indicates that system A is better on average than system B, but it provides little info on how to improve performance and, what is worse, it should not come as a surprise if B ends up being better than A on outside data. We propose an evaluation methodology for sequence labeling tasks grounded on error analysis that provides both quantitative and qualitative information on where systems must be improved and predicts how models will perform on a different distribution. The key is to create test sets that, contrary to common practice, do not rely on gathering large amounts of real-world in-distribution scraped data, but consists in handcrafting a small set of linguistically motivated examples that exhaustively cover the range of span attributes (such as shape, length, casing, sentence position, etc.) a system may encounter in the wild. We demonstrate this methodology on a benchmark for anglicism identification in Spanish. Our methodology provides results that are diagnostic (because they help identify systematic weaknesses in performance), actionable (because they can inform which model is better suited for a given scenario) and predictive: our method predicts model performance on external datasets with a median correlation of 0.85.

Method: Unified pipeline using frozen CLIP-style vision-language encoders and multilingual BGE M3 encoder, with lightweight trainable modules: logistic regression, LLM-based sentence-type predictor, idiom synonym substitution, distractor-aware scoring, and Borda rank fusion.

Result: Improved from CLIP baseline (26.7% Top-1 on English dev) to 60.0% Top-1 on English and 60.0% Top-1 (0.822 NDCG@5) in zero-shot transfer to Portuguese. Achieved average Top-1/NDCG scores of 0.35/0.73 for image+text ranking and 0.32/0.71 for text-only caption ranking across 15 languages.

Conclusion: Effective idiom disambiguation is feasible without fine-tuning large multimodal encoders. Idiom-aware rewriting is the main performance contributor, while sentence-type prediction and multimodal fusion enhance robustness.

Abstract: Multimodal models struggle with idiomatic expressions due to their non-compositional meanings, a challenge amplified in multilingual settings. We introduced PolyFrame, our system for the MWE-2026 AdMIRe2 shared task on multimodal idiom disambiguation, featuring a unified pipeline for both image+text ranking (Subtask A) and text-only caption ranking (Subtask B). All model variants retain frozen CLIP-style vision–language encoders and the multilingual BGE M3 encoder, training only lightweight modules: a logistic regression and LLM-based sentence-type predictor, idiom synonym substitution, distractor-aware scoring, and Borda rank fusion. Starting from a CLIP baseline (26.7% Top-1 on English dev, 6.7% on English test), adding idiom-aware paraphrasing and explicit sentence-type classification increased performance to 60.0% Top-1 on English and 60.0% Top-1 (0.822 NDCG@5) in zero-shot transfer to Portuguese. On the multilingual blind test, our systems achieved average Top-1/NDCG scores of 0.35/0.73 for Subtask A and 0.32/0.71 for Subtask B across 15 languages. Ablation results highlight idiom-aware rewriting as the main contributor to performance, while sentence-type prediction and multimodal fusion enhance robustness. These findings suggest that effective idiom disambiguation is feasible without fine-tuning large multimodal encoders.

Method: Developed a language-agnostic API with modular multi-backend architecture integrating rule-based finite-state transducers and neural models, featuring automatic script detection and routing between script variants.

Result: Created a comprehensive library covering tokenization, morphological analysis, POS tagging, dependency parsing, NER, bidirectional script transliteration, cross-lingual sentence embeddings, and machine translation with CoNLL-U standard outputs.

Conclusion: TurkicNLP provides the first unified NLP pipeline for Turkic languages, addressing fragmentation and enabling consistent processing across diverse scripts and language variants.

Abstract: Natural language processing for the Turkic language family, spoken by over 200 million people across Eurasia, remains fragmented, with most languages lacking unified tooling and resources. We present TurkicNLP, an open-source Python library providing a single, consistent NLP pipeline for Turkic languages across four script families: Latin, Cyrillic, Perso-Arabic, and Old Turkic Runic. The library covers tokenization, morphological analysis, part-of-speech tagging, dependency parsing, named entity recognition, bidirectional script transliteration, cross-lingual sentence embeddings, and machine translation through one language-agnostic API. A modular multi-backend architecture integrates rule-based finite-state transducers and neural models transparently, with automatic script detection and routing between script variants. Outputs follow the CoNLL-U standard for full interoperability and extension. Code and documentation are hosted at https://github.com/turkic-nlp/turkicnlp .

Method: Developed a Persian-specific short-answer evaluation framework combining rule-based morphological normalization with a hybrid syntactic and semantic similarity module for robust soft-match scoring beyond exact string overlap.

Result: Evaluation of 15 state-of-the-art models across three Persian datasets shows hybrid evaluation improves scoring consistency by +10 compared to exact-match baselines, with semantic similarity metrics achieving higher agreement with human judgments than LLM-based judges.

Conclusion: The framework provides the first standardized benchmark for measuring cultural understanding in Persian and establishes a reproducible foundation for cross-cultural LLM evaluation research.

Abstract: This paper presents a comprehensive evaluation framework for assessing the cultural competence of large language models (LLMs) in Persian. Existing Persian cultural benchmarks rely predominantly on multiple-choice formats and English-centric metrics that fail to capture Persian’s morphological complexity and semantic nuance. Our framework introduces a Persian-specific short-answer evaluation that combines rule-based morphological normalization with a hybrid syntactic and semantic similarity module, enabling robust soft-match scoring beyond exact string overlap. Through systematic evaluation of 15 state-of-the-art open- and closed-source models across three culturally grounded Persian datasets, we demonstrate that our hybrid evaluation improves scoring consistency by +10 compared to exact-match baselines by capturing meaning that surface-level methods cannot detect. Our human evaluation further confirms that the proposed semantic similarity metric achieves higher agreement with human judgments than LLM-based judges. We publicly release our evaluation framework, providing the first standardized benchmark for measuring cultural understanding in Persian and establishing a reproducible foundation for cross-cultural LLM evaluation research.

Method: The paper examines unlearning robustness by testing two common interaction patterns: (1) self-correction where models correct themselves, and (2) dialogue-conditioned querying where context influences responses. They evaluate whether knowledge that appears forgotten in static settings can be recovered through these interactive mechanisms.

Result: Knowledge that appears forgotten in static evaluation can often be recovered through interaction. Stronger unlearning techniques may improve apparent robustness but often lead to behavioral rigidity rather than genuine knowledge erasure. Static evaluation overestimates real-world unlearning effectiveness.

Conclusion: Current static evaluation methods for machine unlearning are insufficient for assessing real-world effectiveness. Interactive settings reveal vulnerabilities in unlearning approaches, highlighting the need for ensuring stable forgetting under realistic use conditions.

Abstract: Machine unlearning aims to remove the influence of specific training data from pre-trained models without retraining from scratch, and is increasingly important for large language models (LLMs) due to safety, privacy, and legal concerns. Although prior work primarily evaluates unlearning in static, single-turn settings, forgetting robustness under realistic interactive use remains underexplored. In this paper, we study whether unlearning remains stable in interactive environments by examining two common interaction patterns: self-correction and dialogue-conditioned querying. We find that knowledge appearing forgotten in static evaluation can often be recovered through interaction. Although stronger unlearning improves apparent robustness, it often results in behavioral rigidity rather than genuine knowledge erasure. Our findings suggest that static evaluation may overestimate real-world effectiveness and highlight the need for ensuring stable forgetting under interactive settings.

Method: Created ClinConsensus benchmark with 2500 cases across 36 specialties and 12 clinical task types, implemented rubric-based grading with CACS@k scoring, and developed dual-judge evaluation framework combining high-capability LLM-as-judge with distilled local judge model.

Result: Comprehensive assessment revealed substantial heterogeneity across models in reasoning, evidence use, and longitudinal follow-up capabilities, with clinically actionable treatment planning remaining a key bottleneck despite comparable overall scores.

Conclusion: ClinConsensus provides an extensible benchmark for developing robust, clinically grounded medical LLMs ready for real-world deployment, addressing current limitations in medical AI evaluation.

Abstract: Large language models (LLMs) are increasingly applied to health management, showing promise across disease prevention, clinical decision-making, and long-term care. However, existing medical benchmarks remain largely static and task-isolated, failing to capture the openness, longitudinal structure, and safety-critical complexity of real-world clinical workflows. We introduce ClinConsensus, a Chinese medical benchmark curated, validated, and quality-controlled by clinical experts. ClinConsensus comprises 2500 open-ended cases spanning the full continuum of care–from prevention and intervention to long-term follow-up–covering 36 medical specialties, 12 common clinical task types, and progressively increasing levels of complexity. To enable reliable evaluation of such complex scenarios, we adopt a rubric-based grading protocol and propose the Clinically Applicable Consistency Score (CACS@k). We further introduce a dual-judge evaluation framework, combining a high-capability LLM-as-judge with a distilled, locally deployable judge model trained via supervised fine-tuning, enabling scalable and reproducible evaluation aligned with physician judgment. Using ClinConsensus, we conduct a comprehensive assessment of several leading LLMs and reveal substantial heterogeneity across task themes, care stages, and medical specialties. While top-performing models achieve comparable overall scores, they differ markedly in reasoning, evidence use, and longitudinal follow-up capabilities, and clinically actionable treatment planning remains a key bottleneck. We release ClinConsensus as an extensible benchmark to support the development and evaluation of medical LLMs that are robust, clinically grounded, and ready for real-world deployment.

Method: Compiled original texts from domains related to democratic participation, selected based on relevance, copyright availability, and ethical standards. All texts were simplified to Easy-to-Read level by human experts in text simplification.

Result: Created the first annotated corpus of its kind for Catalan language, with high-quality human-annotated language resources for Spanish and Italian. The corpus includes different text types and will be made freely accessible to the public.

Conclusion: This corpus fills a significant gap in resources for text simplification research in less-resourced languages and supports the iDEM project’s goal of assessing Easy-to-Read language impact on democratic participation.

Abstract: Being able to understand information is a key factor for a self-determined life and society. It is also very important for participating in democratic processes. The study of automatic text simplification is often limited by the availability of high quality material for the training and evaluation on automatic simplifiers. This is true for English, but more so for less resourced languages like Spanish, Catalan and Italian. In order to fill this gap, we present a corpus of original texts for these 3 languages, with high quality simplification produced by human experts in text simplification. It was developed within the iDEM project to assess the impact of Easy-to-Read (E2R) language for democratic participation. The original texts were compiled from domains related to this topic. The corpus includes different text types, selected based on relevance, copyright availability, and ethical standards. All texts were simplified to E2R level. The corpus is particularity valuable because it includes the first annotated corpus of its kind for the Catalan language. It also represents a noteworthy contribution for Spanish and Italian, offering high-quality, human-annotated language resources that are rarely available in these domains. The corpus will be made freely accessible to the public.

Method: Conducted comprehensive audit using 6,642 human-verified labels to evaluate judge performance under real-world red-teaming conditions, revealing performance degradation. Proposed ReliableBench benchmark for consistently judgeable behaviors and JudgeStressTest dataset to expose judge failures.

Result: Judge performance degrades to near random chance under distribution shifts from red-teaming, in stark contrast to high human agreement reported in prior work. Many attacks inflate success rates by exploiting judge insufficiencies rather than eliciting genuinely harmful content.

Conclusion: Current LLM-as-a-Judge frameworks are unreliable for safety evaluation under red-teaming conditions, necessitating more robust benchmarks and stress tests to ensure reliable evaluation of multimodal models.

Abstract: Automated \enquote{LLM-as-a-Judge} frameworks have become the de facto standard for scalable evaluation across natural language processing. For instance, in safety evaluation, these judges are relied upon to evaluate harmfulness in order to benchmark the robustness of safety against adversarial attacks. However, we show that existing validation protocols fail to account for substantial distribution shifts inherent to red-teaming: diverse victim models exhibit distinct generation styles, attacks distort output patterns, and semantic ambiguity varies significantly across jailbreak scenarios. Through a comprehensive audit using 6642 human-verified labels, we reveal that the unpredictable interaction of these shifts often causes judge performance to degrade to near random chance. This stands in stark contrast to the high human agreement reported in prior work. Crucially, we find that many attacks inflate their success rates by exploiting judge insufficiencies rather than eliciting genuinely harmful content. To enable more reliable evaluation, we propose ReliableBench, a benchmark of behaviors that remain more consistently judgeable, and JudgeStressTest, a dataset designed to expose judge failures. Data available at: https://github.com/SchwinnL/LLMJudgeReliability.

Method: Used machine translation methodology to create multilingual anonymization benchmark in 10 languages, preserving original annotations while rendering names of cities and people in culturally appropriate forms for each target language.

Result: Created benchmark with over 2,500 annotations of personal information; evaluation by medical professionals confirmed translation quality, including adaptation of personal information; benchmark supports training, validation, and automatic detection improvement.

Conclusion: Multilingual anonymization benchmark enables privacy-compliant data sharing for healthcare ML; synthetic data and translation methodology address data scarcity and privacy concerns; available for research applications.

Abstract: Accessing sensitive patient data for machine learning is challenging due to privacy concerns. Datasets with annotations of personally identifiable information are crucial for developing and testing anonymization systems to enable safe data sharing that complies with privacy regulations. Since accessing real patient data is a bottleneck, synthetic data offers an efficient solution for data scarcity, bypassing privacy regulations that apply to real data. Moreover, neural machine translation can help to create high-quality data for low-resource languages by translating validated real or synthetic data from a high-resource language. In this work, we create a multilingual anonymization benchmark in ten languages, using a machine translation methodology that preserves the original annotations and renders names of cities and people in a culturally and contextually appropriate form in each target language. Our evaluation study with medical professionals confirms the quality of the translations, both in general and with respect to the translation and adaptation of personal information. Our benchmark with over 2,500 annotations of personal information can be used in many applications, including training annotators, validating annotations across institutions without legal complications, and helping improve the performance of automatic personal information detection. We make our benchmark and annotation guidelines available for further research.

Method: Analyzes how emotional tone alters attention geometry in transformers (locality, center-of-mass distance, entropy). Introduces AURA-QA dataset with emotionally balanced passages. Proposes emotional regularization framework to constrain emotion-conditioned representational drift during training.

Result: Attention metrics vary across emotions and correlate with QA performance. Emotional regularization improves reading comprehension in both emotionally-varying and non-emotionally varying datasets, with consistent gains under distribution shift and in-domain improvements.

Conclusion: Emotion systematically shapes transformer attention and reasoning. Accounting for emotional variation through regularization improves model robustness and performance across diverse QA tasks.

Abstract: Large language models are routinely deployed on text that varies widely in emotional tone, yet their reasoning behavior is typically evaluated without accounting for emotion as a source of representational variation. Prior work has largely treated emotion as a prediction target, for example in sentiment analysis or emotion classification. In contrast, we study emotion as a latent factor that shapes how models attend to and reason over text. We analyze how emotional tone systematically alters attention geometry in transformer models, showing that metrics such as locality, center-of-mass distance, and entropy vary across emotions and correlate with downstream question-answering performance. To facilitate controlled study of these effects, we introduce Affect-Uniform ReAding QA (AURA-QA), a question-answering dataset with emotionally balanced, human-authored context passages. Finally, an emotional regularization framework is proposed that constrains emotion-conditioned representational drift during training. Experiments across multiple QA benchmarks demonstrate that this approach improves reading comprehension in both emotionally-varying and non-emotionally varying datasets, yielding consistent gains under distribution shift and in-domain improvements on several benchmarks.

Method: Updated and combined existing EPIC-UdS (spoken) and EuroParl-UdS (written) corpora by correcting metadata/text errors, refining content, updating linguistic annotations, and adding new layers including word alignment and word-level surprisal indices.

Result: Created a comprehensive combined resource validated through a study on filler particles prediction in interpreting using probabilistic measures from base/fine-tuned GPT-2 and machine translation models.

Conclusion: The enhanced corpus provides valuable infrastructure for research on language variation, translation studies, and speech analysis, with demonstrated utility through validation studies.

Abstract: This paper introduces an updated and combined version of the bidirectional English-German EPIC-UdS (spoken) and EuroParl-UdS (written) corpora containing original European Parliament speeches as well as their translations and interpretations. The new version corrects metadata and text errors identified through previous use, refines the content, updates linguistic annotations, and adds new layers, including word alignment and word-level surprisal indices. The combined resource is designed to support research using information-theoretic approaches to language variation, particularly studies comparing written and spoken modes, and examining disfluencies in speech, as well as traditional translationese studies, including parallel (source vs. target) and comparable (original vs. translated) analyses. The paper outlines the updates introduced in this release, summarises previous results based on the corpus, and presents a new illustrative study. The study validates the integrity of the rebuilt spoken data and evaluates probabilistic measures derived from base and fine-tuned GPT-2 and machine translation models on the task of filler particles prediction in interpreting.

Method: Two complementary evaluation setups: 1) prompt-based functional assessment of task performance, and 2) representational analysis of internal model states to examine compositional representations.

Result: Striking divergence between task performance and internal states - LLMs reliably develop compositional representations but fail to consistently translate them into functional task success across model variants.

Conclusion: Contrastive evaluation (combining functional and representational analysis) is essential for obtaining a complete understanding of model capabilities, as performance metrics alone can be misleading.

Abstract: Compositionality is considered central to language abilities. As performant language systems, how do large language models (LLMs) do on compositional tasks? We evaluate adjective-noun compositionality in LLMs using two complementary setups: prompt-based functional assessment and a representational analysis of internal model states. Our results reveal a striking divergence between task performance and internal states. While LLMs reliably develop compositional representations, they fail to translate consistently into functional task success across model variants. Consequently, we highlight the importance of contrastive evaluation for obtaining a more complete understanding of model capabilities.

Method: Pipeline combining sentence-level language identification (LU/DE/FR/EN) with token-level borrowing resolver restricted to LU sentences, using lemmatization, loanword registry, and compiled morphological/orthographic rules on 259,305 RTL articles spanning 1999-2025.

Result: LU remains matrix language across all documents; 77.1% articles include at least one donor language; median CMI increases from 3.90 to 7.00; CMI rises from 6.1 (1999-2007) to 8.4 in 2020; 25,444 token-level adaptations (63.8% morphological, 35.9% orthographic); French overwhelmingly supplies adapted items.

Conclusion: Multilingual practice is pervasive but localized; code-switching intensifies over time; French is dominant donor; borrowing-centric evaluation metrics (borrowed token/type rates, donor entropy, assimilation ratios) are needed beyond document-level mixing indices.

Abstract: We present LuxBorrow, a borrowing-first analysis of Luxembourgish (LU) news spanning 27 years (1999-2025), covering 259,305 RTL articles and 43.7M tokens. Our pipeline combines sentence-level language identification (LU/DE/FR/EN) with a token-level borrowing resolver restricted to LU sentences, using lemmatization, a collected loanword registry, and compiled morphological and orthographic rules. Empirically, LU remains the matrix language across all documents, while multilingual practice is pervasive: 77.1% of articles include at least one donor language and 65.4% use three or four. Breadth does not imply intensity: median code-mixing index (CMI) increases from 3.90 (LU+1) to only 7.00 (LU+3), indicating localized insertions rather than balanced bilingual text. Domain and period summaries show moderate but persistent mixing, with CMI rising from 6.1 (1999-2007) to a peak of 8.4 in 2020. Token-level adaptations total 25,444 instances and exhibit a mixed profile: morphological 63.8%, orthographic 35.9%, lexical 0.3%. The most frequent individual rules are orthographic, such as on->oun and eur->er, while morphology is collectively dominant. Diachronically, code-switching intensifies, and morphologically adapted borrowings grow from a small base. French overwhelmingly supplies adapted items, with modest growth for German and negligible English. We advocate borrowing-centric evaluation, including borrowed token and type rates, donor entropy over borrowed items, and assimilation ratios, rather than relying only on document-level mixing indices.

Method: Combines 0.4B CogViT visual encoder with 0.5B GLM language decoder. Introduces Multi-Token Prediction (MTP) mechanism to predict multiple tokens per step for efficient decoding. Uses two-stage pipeline with PP-DocLayout-V3 for layout analysis followed by parallel region-level recognition.

Result: Achieves competitive or state-of-the-art performance on document parsing, text/formula transcription, table structure recovery, and key information extraction. Shows significant decoding throughput improvement while maintaining low memory overhead.

Conclusion: GLM-OCR provides an efficient compact multimodal solution suitable for both resource-constrained edge deployment and large-scale production systems, balancing performance with computational efficiency for document understanding tasks.

Abstract: GLM-OCR is an efficient 0.9B-parameter compact multimodal model designed for real-world document understanding. It combines a 0.4B-parameter CogViT visual encoder with a 0.5B-parameter GLM language decoder, achieving a strong balance between computational efficiency and recognition performance. To address the inefficiency of standard autoregressive decoding in deterministic OCR tasks, GLM-OCR introduces a Multi-Token Prediction (MTP) mechanism that predicts multiple tokens per step, significantly improving decoding throughput while keeping memory overhead low through shared parameters. At the system level, a two-stage pipeline is adopted: PP-DocLayout-V3 first performs layout analysis, followed by parallel region-level recognition. Extensive evaluations on public benchmarks and industrial scenarios show that GLM-OCR achieves competitive or state-of-the-art performance in document parsing, text and formula transcription, table structure recovery, and key information extraction. Its compact architecture and structured generation make it suitable for both resource-constrained edge deployment and large-scale production systems.

Method: Train GPT-2 style models (3.5M-86M parameters) on corpora with mathematical problems containing both correct and incorrect solutions. Test with random errors vs coherent alternative rule systems, and multi-rule experiments. Also test on real Wikipedia text.

Result: Models extract correct signal with 65-85% accuracy when errors are random. Accuracy drops to chance (45-51%) when errors follow coherent alternative rules. Multi-rule experiments show sharp crossover: single coherent alternative eliminates truth bias, but adding second competing rule restores most of it (47%→78%), growing to 88% with N=10 rules. Similar pattern on Wikipedia (71% vs 46%).

Conclusion: Propose Compression-Consistency Principle: gradient descent favors most compressible answer cluster, not truth per se. Truth bias emerges only when falsehood is structurally incoherent. Whether this extends to large-scale pretraining remains open.

Abstract: Why do language models trained on contradictory data prefer correct answers? In controlled experiments with small transformers (3.5M–86M parameters), we show that this preference tracks the compressibility structure of errors rather than truth per se. We train GPT-2 style models on corpora where each mathematical problem appears with both correct and incorrect solutions – a denoising design that directly models conflicting information about the same fact. When errors are random, models extract the correct signal with accuracy scaling from 65% to 85% with model size. When errors follow a coherent alternative rule system, accuracy drops to chance (~45–51%): the model cannot distinguish the false system from truth. A multi-rule experiment reveals a sharp crossover: a single coherent alternative rule eliminates truth bias entirely, but adding a second competing rule restores most of it (47%->78%), with continued growth through N=10 (88%). The same pattern reproduces on real Wikipedia text (71% vs 46%). We propose the Compression–Consistency Principle as an explanatory hypothesis: in these settings, gradient descent favors the most compressible answer cluster, not truth per se. Truth bias emerges only when falsehood is structurally incoherent. Whether this principle extends to large-scale pretraining remains an open question.

Method: Proposes QAQ framework that evaluates data quality from reverse direction: how well answers predict queries (Q|A). Defines Reverse Mutual Information (RMI) to quantify information gain about query conditioned on answer. Uses selection strategy based on disagreement between strong and weak models to identify valid yet challenging samples.

Result: On WarriorCoder dataset, selecting just 25% of data using stratified RMI achieves comparable performance to full-data training, significantly outperforming existing data selection methods.

Conclusion: Highlights importance of bidirectional semantic coherence in synthetic data curation. Offers scalable pathway to reduce computational costs without sacrificing model capability by focusing on high-quality data selection.

Abstract: Synthetic data has become essential for training code generation models, yet it introduces significant noise and hallucinations that are difficult to detect with current metrics. Existing data selection methods like Instruction-Following Difficulty (IFD) typically assess how hard a model generates an answer given a query ($A|Q$). However, this metric is ambiguous on noisy synthetic data, where low probability can distinguish between intrinsic task complexity and model-generated hallucinations. Here, we propose QAQ, a novel data selection framework that evaluates data quality from the reverse direction: how well can the answer predict the query ($Q|A$)? We define Reverse Mutual Information (RMI) to quantify the information gain about the query conditioned on the answer. Our analyses reveal that both extremes of RMI signal quality issues: low RMI indicates semantic misalignment, while excessively high RMI may contain defect patterns that LLMs easily recognize. Furthermore, we introduce a selection strategy based on the disagreement between strong and weak models to identify samples that are valid yet challenging. Experiments on the WarriorCoder dataset demonstrate that selecting just 25% of data using stratified RMI achieves comparable performance to full-data training, significantly outperforming existing data selection methods. Our approach highlights the importance of bidirectional semantic coherence in synthetic data curation, offering a scalable pathway to reduce computational costs without sacrificing model capability.

Method: Proposes TESS with a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. Uses LLMs with structured prompting to extract interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) from text, filtered through confidence-aware gating.

Result: Experiments on four real-world datasets show up to 29% reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines.

Conclusion: TESS effectively bridges the modality gap between text and time-series data by creating an interpretable intermediate semantic space, enabling better fusion of textual information for improved forecasting performance.

Abstract: Incorporating textual information into time-series forecasting holds promise for addressing event-driven non-stationarity; however, a fundamental modality gap hinders effective fusion: textual descriptions express temporal impacts implicitly and qualitatively, whereas forecasting models rely on explicit and quantitative signals. Through controlled semi-synthetic experiments, we show that existing methods over-attend to redundant tokens and struggle to reliably translate textual semantics into usable numerical cues. To bridge this gap, we propose TESS, which introduces a Temporal Evolution Semantic Space as an intermediate bottleneck between modalities. This space consists of interpretable, numerically grounded temporal primitives (mean shift, volatility, shape, and lag) extracted from text by an LLM via structured prompting and filtered through confidence-aware gating. Experiments on four real-world datasets demonstrate up to a 29 percent reduction in forecasting error compared to state-of-the-art unimodal and multimodal baselines. The code will be released after acceptance.

Method: Proposes MetaKE, a meta-learning aligned knowledge editing framework that treats edit targets as learnable meta-parameters in a bi-level optimization problem. Upper-level optimizer seeks feasible targets to maximize post-edit performance, lower-level solver executes editing. Uses Structural Gradient Proxy to backpropagate editability constraints to target learning phase.

Result: Extensive experiments confirm MetaKE significantly outperforms strong baselines. Theoretical analysis shows MetaKE automatically aligns edit direction with model’s feasible manifold.

Conclusion: MetaKE offers a new perspective on knowledge editing by addressing the semantic-execution disconnect through bi-level optimization with learnable targets, providing more effective knowledge editing in LLMs.

Abstract: Knowledge editing (KE) aims to precisely rectify specific knowledge in Large Language Models (LLMs) without disrupting general capabilities. State-of-the-art methods suffer from an open-loop control mismatch. We identify a critical “Semantic-Execution Disconnect”: the semantic target is derived independently without feedback from the downstream’s feasible region. This misalignment often causes valid semantic targets to fall within the prohibited space, resulting in gradient truncation and editing failure. To bridge this gap, we propose MetaKE (Meta-learning Aligned Knowledge Editing), a new framework that reframes KE as a bi-level optimization problem. Departing from static calculation, MetaKE treats the edit target as a learnable meta-parameter: the upper-level optimizer seeks a feasible target to maximize post-edit performance, while the lower-level solver executes the editing. To address the challenge of differentiating through complex solvers, we derive a Structural Gradient Proxy, which explicitly backpropagates editability constraints to the target learning phase. Theoretical analysis demonstrates that MetaKE automatically aligns the edit direction with the model’s feasible manifold. Extensive experiments confirm that MetaKE significantly outperforms strong baselines, offering a new perspective on knowledge editing.

Method: Used text similarity metrics to evaluate different ChatGPT models’ capacity to generate diverse textual outputs, comparing performance across model versions while explicitly setting temperature parameter to one to encourage diversity.

Result: Found measurable decline in recent ChatGPT releases’ ability to produce varied text, showing reduced output diversity attributed to synthetic data incorporation in training datasets from internet infiltration by LLM-generated data.

Conclusion: Model self-convergence occurs as ChatGPT versions produce increasingly similar texts due to training on synthetic data, demonstrating the negative impact of recursive training on LLM output diversity.

Abstract: Large Language Models (LLMs) that undergo recursive training on synthetically generated data are susceptible to model collapse, a phenomenon marked by the generation of meaningless output. Existing research has examined this issue from either theoretical or empirical perspectives, often focusing on a single model trained recursively on its own outputs. While prior studies have cautioned against the potential degradation of LLM output quality under such conditions, no longitudinal investigation has yet been conducted to assess this effect over time. In this study, we employ a text similarity metric to evaluate different ChatGPT models’ capacity to generate diverse textual outputs. Our findings indicate a measurable decline of recent ChatGPT releases’ ability to produce varied text, even when explicitly prompted to do so, by setting the temperature parameter to one. The observed reduction in output diversity may be attributed to the influence of the amounts of synthetic data incorporated within their training datasets as the result of internet infiltration by LLM generated data. The phenomenon is defined as model self-convergence because of the gradual increase of similarities of produced texts among different ChatGPT versions.

Method: 1) Generate task-informed keywords for comprehensive domain coverage; 2) Create diverse instructions by pairing keywords with different cognitive levels from Bloom’s Taxonomy; 3) Use self-consistency validation to ensure data quality. Applied across seven challenging domains including mathematics, finance, and logical reasoning.

Result: Models fine-tuned on DS²-Instruct generated data achieve substantial improvements over existing data generation methods across multiple specialized domains.

Conclusion: DS²-Instruct provides an effective zero-shot framework for generating high-quality domain-specific instruction datasets without human supervision, enabling better LLM adaptation to specialized domains.

Abstract: Adapting Large Language Models (LLMs) to specialized domains requires high-quality instruction tuning datasets, which are expensive to create through human annotation. Existing data synthesis methods focus on general-purpose tasks and fail to capture domain-specific terminology and reasoning patterns. To address this, we introduce DS$^2$-Instruct, a zero-shot framework that generates domain-specific instruction datasets without human supervision. Our approach first generates task-informed keywords to ensure comprehensive domain coverage. It then creates diverse instructions by pairing these keywords with different cognitive levels from Bloom’s Taxonomy. Finally, it uses self-consistency validation to ensure data quality. We apply this framework to generate datasets across seven challenging domains, such as mathematics, finance, and logical reasoning. Comprehensive evaluation demonstrates that models fine-tuned on our generated data achieve substantial improvements over existing data generation methods.

Method: Proposes Iterative MBR Distillation framework using Minimum Bayes Risk decoding with an off-the-shelf LLM to generate pseudo-labels for training, creating a self-evolution cycle without human annotations.

Result: Models trained solely on self-generated pseudo-labels outperform both unadapted base models and supervised baselines trained on human annotations at system and span levels, while maintaining competitive sentence-level performance on WMT Metrics datasets.

Conclusion: The proposed self-evolution framework effectively eliminates the need for expensive human annotations for Error Span Detection, demonstrating strong performance through LLM-generated pseudo-labels and iterative refinement.

Abstract: Error Span Detection (ESD) is a crucial subtask in Machine Translation (MT) evaluation, aiming to identify the location and severity of translation errors. While fine-tuning models on human-annotated data improves ESD performance, acquiring such data is expensive and prone to inconsistencies among annotators. To address this, we propose a novel self-evolution framework based on Minimum Bayes Risk (MBR) decoding, named Iterative MBR Distillation for ESD, which eliminates the reliance on human annotations by leveraging an off-the-shelf LLM to generate pseudo-labels. Extensive experiments on the WMT Metrics Shared Task datasets demonstrate that models trained solely on these self-generated pseudo-labels outperform both unadapted base model and supervised baselines trained on human annotations at the system and span levels, while maintaining competitive sentence-level performance.

Method: Cannot determine method as the paper content is unavailable due to API rate limiting.

Result: Cannot determine results as the paper content is unavailable due to API rate limiting.

Conclusion: Cannot draw conclusions as the paper content is unavailable due to API rate limiting.

Abstract: Failed to fetch summary for 2510.24358: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.24358&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to access limitations

Result: Cannot determine results due to access limitations

Conclusion: Cannot determine conclusion due to access limitations

Abstract: Failed to fetch summary for 2511.04161: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.04161&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to draw conclusions due to fetch failure

Abstract: Failed to fetch summary for 2511.22333: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.22333&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2512.20074: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.20074&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2512.22213: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.22213&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions as paper content is unavailable

Abstract: Failed to fetch summary for 2602.06547: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.06547&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unknown - paper content not accessible due to API rate limiting

Result: No results available - failed to fetch paper summary

Conclusion: Cannot analyze paper due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2602.08561: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.08561&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to inability to access paper content

Result: Cannot determine results due to inability to access paper content

Conclusion: Cannot draw conclusions due to inability to access paper content

Abstract: Failed to fetch summary for 2602.11052: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.11052&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2602.23329: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23329&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as paper content could not be retrieved

Result: Unable to determine results as paper content could not be retrieved

Conclusion: Unable to determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2603.05092: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.05092&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed API request

Result: Cannot determine results due to failed API request

Conclusion: Cannot determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2603.08448: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.08448&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2603.12252: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12252&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2603.13023: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.13023&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as paper content could not be retrieved

Result: Unable to determine results as paper content could not be retrieved

Conclusion: Unable to determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2603.13173: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.13173&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Created a synthetic OCR dataset of 7,219 images across all three Kazakh scripts with font, color, and noise variations. Evaluated three MLLMs (Gemma-3-12B-it, Qwen2.5-VL-7B-Instruct, Llama-3.2-11B-Vision-Instruct) on OCR and language identification tasks, comparing them with classical OCR baselines.

Result: All MLLMs failed at Latin and Arabic script OCR, misclassifying Arabic script as Arabic, Farsi, and Kurdish. Traditional OCR had lower character error rates than MLLMs. MLLMs showed significant gaps in processing low-resource Abjad-based scripts.

Conclusion: Current MLLMs have substantial limitations in processing low-resource scripts, particularly Abjad-based ones. There’s a need for more inclusive models and benchmarks that support diverse scripts and low-resource languages.

Abstract: Kazakh is a Turkic language using the Arabic, Cyrillic, and Latin scripts, making it unique in terms of optical character recognition (OCR). Work on OCR for low-resource Kazakh scripts is very scarce, and no OCR benchmarks or images exist for the Arabic and Latin scripts. We construct a synthetic OCR dataset of 7,219 images for all three scripts with font, color, and noise variations to imitate real OCR tasks. We evaluated three multimodal large language models (MLLMs) on a subset of the benchmark for OCR and language identification: Gemma-3-12B-it, Qwen2.5-VL-7B-Instruct, and Llama-3.2-11B-Vision-Instruct. All models are unsuccessful with Latin and Arabic script OCR, and fail to recognize the Arabic script as Kazakh text, misclassifying it as Arabic, Farsi, and Kurdish. We further compare MLLMs with a classical OCR baseline and find that while traditional OCR has lower character error rates, MLLMs fail to match this performance. These findings show significant gaps in current MLLM capabilities to process low-resource Abjad-based scripts and demonstrate the need for inclusive models and benchmarks supporting low-resource scripts and languages.

Method: Re-implemented key contributions of recent gloss-free SLT models in a unified codebase with standardized preprocessing, video encoders, and training setups to ensure fair comparison across methods.

Result: Many performance gains reported in literature diminish when models are evaluated under consistent conditions, suggesting implementation details and evaluation setups play a significant role in determining results.

Conclusion: The study emphasizes the need for transparency and reproducibility in SLT research, providing a public codebase to support standardized evaluation and fair comparison of future methods.

Abstract: Sign Language Translation (SLT) aims to automatically convert visual sign language videos into spoken language text and vice versa. While recent years have seen rapid progress, the true sources of performance improvements often remain unclear. Do reported performance gains come from methodological novelty, or from the choice of a different backbone, training optimizations, hyperparameter tuning, or even differences in the calculation of evaluation metrics? This paper presents a comprehensive study of recent gloss-free SLT models by re-implementing key contributions in a unified codebase. We ensure fair comparison by standardizing preprocessing, video encoders, and training setups across all methods. Our analysis shows that many of the performance gains reported in the literature often diminish when models are evaluated under consistent conditions, suggesting that implementation details and evaluation setups play a significant role in determining results. We make the codebase publicly available here (https://github.com/ozgemercanoglu/sltbaselines) to support transparency and reproducibility in SLT research.

Method: Introduces a unified probabilistic framework using Maximum Mean Discrepancy (MMD) potential that recasts both Shielded Diffusion and Safe Denoiser as instances of energy-based negative guidance. Leverages control-barrier function analysis to identify critical time windows for applying negative guidance, showing it should be strong early in denoising and decay to zero later.

Result: The framework successfully demonstrates that negative guidance should be applied in early stages of denoising for effective safe generation. Evaluations on realistic safe generation scenarios confirm the importance of timing in safety guidance application.

Conclusion: The paper provides a unified theoretical foundation for safety mechanisms in diffusion models, bridging geometric and data-driven approaches, and establishes principled timing guidelines for when safety guidance should be applied during the generation process.

Abstract: Safety mechanisms for diffusion and flow models have recently been developed along two distinct paths. In robot planning, control barrier functions are employed to guide generative trajectories away from obstacles at every denoising step by explicitly imposing geometric constraints. In parallel, recent data-driven, negative guidance approaches have been shown to suppress harmful content and promote diversity in generated samples. However, they rely on heuristics without clearly stating when safety guidance is actually necessary. In this paper, we first introduce a unified probabilistic framework using a Maximum Mean Discrepancy (MMD) potential for image generation tasks that recasts both Shielded Diffusion and Safe Denoiser as instances of our energy-based negative guidance against unsafe data samples. Furthermore, we leverage control-barrier functions analysis to justify the existence of a critical time window in which negative guidance must be strong; outside of this window, the guidance should decay to zero to ensure safe and high-quality generation. We evaluate our unified framework on several realistic safe generation scenarios, confirming that negative guidance should be applied in the early stages of the denoising process for successful safe generation.

Method: Uses parameter-efficient adaptation of compact vision-language models (VLMs) with a unified evaluation protocol for preprocessing, prompting, dataset splits, metrics, and runtime settings. Compares against training-free VLM pipelines and weakly supervised baselines.

Result: With parameter-efficient adaptation, compact VLMs achieve performance on par with or exceeding established approaches while maintaining competitive per-clip latency. Adaptation reduces prompt sensitivity and produces more consistent behavior.

Conclusion: Parameter-efficient fine-tuning enables compact VLMs to serve as dependable clip-level anomaly detectors with favorable accuracy-efficiency trade-off in a transparent experimental setup.

Abstract: CCTV safety monitoring demands anomaly detectors combine reliable clip-level accuracy with predictable per-clip latency despite weak supervision. This work investigates compact vision-language models (VLMs) as practical detectors for this regime. A unified evaluation protocol standardizes preprocessing, prompting, dataset splits, metrics, and runtime settings to compare parameter-efficiently adapted compact VLMs against training-free VLM pipelines and weakly supervised baselines. Evaluation spans accuracy, precision, recall, F1, ROC-AUC, and average per-clip latency to jointly quantify detection quality and efficiency. With parameter-efficient adaptation, compact VLMs achieve performance on par with, and in several cases exceeding, established approaches while retaining competitive per-clip latency. Adaptation further reduces prompt sensitivity, producing more consistent behavior across prompt regimes under the shared protocol. These results show that parameter-efficient fine-tuning enables compact VLMs to serve as dependable clip-level anomaly detectors, yielding a favorable accuracy-efficiency trade-off within a transparent and consistent experimental setup.

Method: Info-VLA uses two complementary constraints: 1) Replay Anchor Contrastive Learning constructs stable alignment anchors from a frozen teacher model to preserve cross-modal alignment in representation space; 2) Cross-Modal Mutual Information Maximization preserves dependency structure between visual and language representations through mutual information constraints. This jointly preserves historical alignment and cross-modal dependency information.

Result: Experiments on the LIBERO benchmark show that Info-VLA significantly outperforms existing methods in both task retention and adaptation, demonstrating effective mitigation of catastrophic forgetting in VLA models.

Conclusion: Info-VLA successfully addresses catastrophic forgetting in VLA models by preserving cross-modal information structure through complementary constraints, balancing stability and plasticity during continual learning for robotic applications.

Abstract: When deployed in open-ended robotic environments, Vision–Language–Action (VLA) models need to continually acquire new skills, yet suffer from severe catastrophic forgetting. We observe that this degradation is related to the deterioration of cross-modal information structure, where dependencies among visual observations, language instructions, and actions progressively diffuse during continual adaptation. But existing continual learning methods fail to preserve such cross-modal information dependencies. Thus, we propose Info-VLA, an information-preserving continual learning framework that maintains cross-modal information structure through two complementary constraints. Replay Anchor Contrastive Learning constructs stable alignment anchors from a frozen teacher model, preserving cross-modal alignment in the representation space. Cross-Modal Mutual Information Maximization further preserves dependency structure between visual and language representations through mutual information constraints. By jointly preserving historical alignment and cross-modal dependency information, Info-VLA balances stability and plasticity during continual learning. Furthermore, experiments on the LIBERO show that Info-VLA significantly outperforms existing methods in both task retention and adaptation.

Method: Multi-channel U-Net architecture that outputs independent probability maps for cracks, busbars, dark areas, and non-cell regions, enabling accurate co-classification of interacting features

Result: Achieved 98% accuracy and demonstrated generalization to unseen datasets, providing scalable automated inspection for PV modules

Conclusion: The framework offers a scalable, extensible tool for automated PV module inspection, improving defect quantification and lifetime prediction in large-scale PV systems

Abstract: Electroluminescence (EL) imaging is widely used to detect defects in photovoltaic (PV) modules, and machine learning methods have been applied to enable large-scale analysis of EL images. However, existing methods cannot assign multiple labels to the same pixel, limiting their ability to capture overlapping degradation features. We present a multi-channel U-Net architecture for pixel-level multi-label segmentation of EL images. The model outputs independent probability maps for cracks, busbars, dark areas, and non-cell regions, enabling accurate co-classification of interacting features such as cracks crossing busbars. The model achieved an accuracy of 98% and has been shown to generalize to unseen datasets. This framework offers a scalable, extensible tool for automated PV module inspection, improving defect quantification and lifetime prediction in large-scale PV systems.

Method: Scales up pre-trained text-to-image diffusion models for temporal coherence using temporal-pyramid cross-frame spatial-temporal attention modules and convolutions. Introduces a dual-stream cross-attention mechanism that allows interpolation between single and dual modality controls during inference.

Result: UniVid achieves superior temporal coherence on T2V, I2V, and combined (T+I)2V tasks, demonstrating effective integration of both text and image conditions for video generation.

Conclusion: The proposed unified model successfully bridges text-to-video and image-to-video generation paradigms, offering flexible control through text prompts and reference images while maintaining temporal coherence.

Abstract: Diffusion-based text-to-video generation (T2V) or image-to-video (I2V) generation have emerged as a prominent research focus. However, there exists a challenge in integrating the two generative paradigms into a unified model. In this paper, we present a unified video generation model (UniVid) with hybrid conditions of the text prompt and reference image. Given these two available controls, our model can extract objects’ appearance and their motion descriptions from textual prompts, while obtaining texture details and structural information from image clues to guide the video generation process. Specifically, we scale up the pre-trained text-to-image diffusion model for generating temporally coherent frames via introducing our temporal-pyramid cross-frame spatial-temporal attention modules and convolutions. To support bimodal control, we introduce a dual-stream cross-attention mechanism, whose attention scores can be freely re-weighted for interpolation of between single and two modalities controls during inference. Extensive experiments showcase that our UniVid achieves superior temporal coherence on T2V, I2V and (T+I)2V tasks.

Method: Proposes Complementarity-Supervised Multi-Band Expert Network (Atsuko) with orthogonal decomposition of each modality’s features into high, mid, and low-frequency components. Uses modality-level router with dual-path mechanism for fine-grained cross-band selection and cross-modal fusion. Introduces Marginal Complementarity Module (MCM) to quantify performance loss when removing each modality via bi-modal comparison, providing soft supervision to guide router focus on modalities with unique information gains.

Result: Extensive experiments show superior performance on CMU-MOSI, CMU-MOSEI, CH-SIMS, CH-SIMSv2, and MIntRec benchmarks compared to prior methods.

Conclusion: The Atsuko network effectively addresses limitations of prior multimodal emotion recognition methods by modeling fine-grained complementary features through multi-band decomposition and complementarity supervision, achieving state-of-the-art performance across multiple benchmarks.

Abstract: Multimodal emotion recognition fuses cues such as text, video, and audio to understand individual emotional states. Prior methods face two main limitations: mechanically relying on independent unimodal performance, thereby missing genuine complementary contributions, and coarse-grained fusion conflicting with the fine-grained representations required by emotion tasks. As inconsistent information density across heterogeneous modalities hinders inter-modal feature mining, we propose the Complementarity-Supervised Multi-Band Expert Network, named Atsuko, to model fine-grained complementary features via multi-scale band decomposition and expert collaboration. Specifically, we orthogonally decompose each modality’s features into high, mid, and low-frequency components. Building upon this band-level routing, we design a modality-level router with a dual-path mechanism for fine-grained cross-band selection and cross-modal fusion. To mitigate shortcut learning from dominant modalities, we propose the Marginal Complementarity Module (MCM) to quantify performance loss when removing each modality via bi-modal comparison. The resulting complementarity distribution provides soft supervision, guiding the router to focus on modalities contributing unique information gains. Extensive experiments show our method achieves superior performance on the CMU-MOSI, CMU-MOSEI, CH-SIMS, CH-SIMSv2, and MIntRec benchmarks.

Method: Theoretical and experimental analysis reveals cross-entropy loss has visual and cross-modal parts. Visual learning acts as a shortcut that hinders cross-modal alignment. Proposed method perturbs visual learning to focus on cross-modal alignment and uses visual-text semantic relationships to gradually align modalities.

Result: Extensive experiments on 4 CDFSL datasets and 11 FSL datasets with CLIP, SigLIP, and PE-Core backbones show consistent state-of-the-art performance across various settings.

Conclusion: Strengthening visual discriminability in VLM-based SF-CDFSL is detrimental; focusing on cross-modal alignment through visual learning perturbation and gradual modality alignment yields superior performance.

Abstract: Source-Free Cross-Domain Few-Shot Learning (SF-CDFSL) focuses on fine-tuning with limited training data from target domains (e.g., medical or satellite images), where Vision-Language Models (VLMs) such as CLIP and SigLIP have shown promising results. Current works in traditional visual models suggest that improving visual discriminability enhances performance. However, in VLM-based SF-CDFSL tasks, we find that \textbf{strengthening visual-modal discriminability actually suppresses VLMs’ performance}. In this paper, we aim to delve into this phenomenon for an interpretation and a solution. By both theoretical and experimental proofs, our study reveals that fine-tuning with the typical cross-entropy loss ($\mathcal{L}{\mathrm{vlm}}$) inherently includes a visual learning part and a cross-modal learning part, where the cross-modal part is crucial for rectifying the heavily disrupted modality misalignment in SF-CDFSL. However, we find that the visual learning essentially acts as a shortcut that encourages the model to reduce $\mathcal{L}{\mathrm{vlm}}$ without considering the cross-modal part, therefore hindering the cross-modal alignment and harming the performance. Based on this interpretation, we further propose an approach to address this problem: first, we perturb the visual learning to guide the model to focus on the cross-modal alignment. Then, we use the visual-text semantic relationships to gradually align the visual and textual modalities during the fine-tuning. Extensive experiments on various settings, backbones (CLIP, SigLip, PE-Core), and tasks (4 CDFSL datasets and 11 FSL datasets) show that we consistently set new state-of-the-art results. Code is available at https://github.com/zhenyuZ-HUST/CVPR26-Mind-the-Discriminability-Trap.

Method: Two-stage training framework with discrete flow matching generative backbone. Includes FaPro module to capture global prosody from facial expressions, and Synchronizer module to bridge modality gap among text, video, and speech for precise lip synchronization.

Result: Outperforms previous methods across multiple metrics on two primary benchmark datasets.

Conclusion: DiFlowDubber effectively addresses limitations in current video dubbing approaches by combining facial expression-based prosody guidance with cross-modal synchronization for improved speech quality and lip-sync accuracy.

Abstract: Video dubbing has broad applications in filmmaking, multimedia creation, and assistive speech technology. Existing approaches either train directly on limited dubbing datasets or adopt a two-stage pipeline that adapts pre-trained text-to-speech (TTS) models, which often struggle to produce expressive prosody, rich acoustic characteristics, and precise synchronization. To address these issues, we propose DiFlowDubber with a novel two-stage training framework that effectively transfers knowledge from a pre-trained TTS model to video-driven dubbing, with a discrete flow matching generative backbone. Specifically, we design a FaPro module that captures global prosody and stylistic cues from facial expressions and leverages this information to guide the modeling of subsequent speech attributes. To ensure precise speech-lip synchronization, we introduce a Synchronizer module that bridges the modality gap among text, video, and speech, thereby improving cross-modal alignment and generating speech that is temporally synchronized with lip movements. Experiments on two primary benchmark datasets demonstrate that DiFlowDubber outperforms previous methods across multiple metrics.

Method: DDS-UDA uses a teacher-student architecture with two key modules: 1) Bi-directional cross-domain consistency regularization with coarse-to-fine dynamic mask generator for feature-level semantic exchange, and 2) Frequency-driven intra-domain pseudo label learning with spectral amplitude-mixed supervision signals for high-fidelity feature alignment.

Result: The method outperforms several existing UDA-based methods on two multi-domain fundus image datasets, demonstrating effective adaptation to heterogeneous imaging environments for optic disc and optic cup segmentation.

Conclusion: DDS-UDA successfully disentangles domain-specific biases from domain-invariant feature representations, providing a robust solution for domain adaptation in medical image segmentation tasks, particularly for optic disc and cup analysis.

Abstract: Convolutional neural networks (CNNs) have achieved exciting performance in joint segmentation of optic disc and optic cup on single-institution datasets. However, their clinical translation is hindered by two major challenges: limited availability of large-scale, high-quality annotations and performance degradation caused by domain shift during deployment across heterogeneous imaging protocols and acquisition platforms. While unsupervised domain adaptation (UDA) provides a way to mitigate these limitations, most existing approaches do not address cross-domain interference and intra-domain generalization within a unified framework. In this paper, we present the Dual-Domain Synergy UDA (DDS-UDA), a novel UDA framework that comprises two key modules. First, a bi-directional cross-domain consistency regularization module is enforced to mitigate cross-domain interference through feature-level semantic information exchange guided by a coarse-to-fine dynamic mask generator, suppressing noise propagation while preserving structural coherence. Second, a frequency-driven intra-domain pseudo label learning module is used to enhance intra-domain generalization by synthesizing spectral amplitude-mixed supervision signals, which ensures high-fidelity feature alignment across domains. Implemented within a teacher-student architecture, DDS-UDA disentangles domain-specific biases from domain-invariant feature-level representations, thereby achieving robust adaptation to heterogeneous imaging environments. We conduct a comprehensive evaluation of our proposed method on two multi-domain fundus image datasets, demonstrating that it outperforms several existing UDA based methods and therefore providing an effective way for optic disc and optic cup segmentation.

Method: Created GenState-AI benchmark using Wan2.2-TI2V-5B to generate short clips where meaning depends on precise changes in position, quantity, and object relations. Each query is paired with: main video, temporal hard negative (differs only in decisive end-state), and semantic hard negative (content substitution). Evaluated two representative MLLM-based baselines and introduced triplet-based diagnostic analyses including relative-order statistics and breakdowns across transition categories.

Result: Both evaluated MLLM-based baselines show consistent failure patterns: frequently confuse main video with temporal hard negative, over-prefer temporally plausible but end-state-incorrect clips, indicating insufficient grounding to decisive end-state evidence. Models are comparatively less sensitive to semantic substitutions.

Conclusion: GenState-AI provides a focused testbed for state-aware, temporally and semantically sensitive text-to-video retrieval. The benchmark reveals critical gaps in current MLLMs’ ability to ground temporal reasoning to decisive end-state evidence, highlighting the need for improved temporal understanding in multimodal models.

Abstract: Existing text-to-video retrieval benchmarks are dominated by real-world footage where much of the semantics can be inferred from a single frame, leaving temporal reasoning and explicit end-state grounding under-evaluated. We introduce GenState-AI, an AI-generated benchmark centered on controlled state transitions, where each query is paired with a main video, a temporal hard negative that differs only in the decisive end-state, and a semantic hard negative with content substitution, enabling fine-grained diagnosis of temporal vs. semantic confusions beyond appearance matching. Using Wan2.2-TI2V-5B, we generate short clips whose meaning depends on precise changes in position, quantity, and object relations, providing controllable evaluation conditions for state-aware retrieval. We evaluate two representative MLLM-based baselines, and observe consistent and interpretable failure patterns: both frequently confuse the main video with the temporal hard negative and over-prefer temporally plausible but end-state-incorrect clips, indicating insufficient grounding to decisive end-state evidence, while being comparatively less sensitive to semantic substitutions. We further introduce triplet-based diagnostic analyses, including relative-order statistics and breakdowns across transition categories, to make temporal vs. semantic failure sources explicit. GenState-AI provides a focused testbed for state-aware, temporally and semantically sensitive text-to-video retrieval, and will be released on huggingface.co.

Method: Multi-agent perception-action exploration alliance with VLM agents operating in rounds. Each round includes: 1) perception exploration to extract query-specific clues from sampled frames and align relevant video blocks, 2) action exploration with three steps: individual answer generation, cross-review scoring among agents, and consensus-based decision making for either new rounds or final answer production.

Result: Outperforms 18 existing VLMs and 10 recent long-video reasoning methods on five VideoQA benchmarks, while achieving significantly lower inference latency.

Conclusion: A4VL effectively scales to real-world long videos through multi-agent collaboration and event-driven partitioning, demonstrating superior performance and efficiency in long-video reasoning tasks.

Abstract: This paper presents a multi-agent perception-action exploration alliance, dubbed A4VL, for efficient long-video reasoning. A4VL operates in a multi-round perception-action exploration loop with a selection of VLM agents. In each round, the team of agents performs video question-answer (VideoQA) via perception exploration followed by action exploration. During perception exploration, each agent learns to extract query-specific perception clue(s) from a few sampled frames and performs clue-based alignment to find the video block(s) that are most relevant to the query-specific event. During action exploration, A4VL performs video reasoning in three steps: (1) each agent produces its initial answer with rational, (2) all agents collaboratively scores one another through cross-reviews and relevance ranking, and (3) based on whether a satisfactory consensus is reached, the decision is made either to start a new round of perception-action deliberation by pruning (e.g., filtering out the lowest performing agent) and re-staging (e.g., new-clue and matching block based perception-action exploration), or to conclude by producing its final answer. The integration of the multi-agent alliance through multi-round perception-action exploration, coupled with event-driven partitioning and cue-guided block alignment, enables A4VL to effectively scale to real world long videos while preserving high quality video reasoning. Evaluation Results on five popular VideoQA benchmarks show that A4VL outperforms 18 existing representative VLMs and 10 recent methods optimized for long-video reasoning, while achieving significantly lower inference latency. Our code is released at https://github.com/git-disl/A4VL.

Method: 1) Patch-based post-training quantization for localized quantization with minimal information loss; 2) Quantization-aware clustering to identify similar patches and aggregate them for efficient quantization parameter sharing; 3) Refinement module to align distributions between original and dequantized images, compensating for quantization errors.

Result: Extensive experiments on CIFAR-10/100, Tiny ImageNet, and ImageNet subsets show the method consistently outperforms prior works under same storage constraints. Notably doubles test accuracy at extreme compression (26.0% → 54.1% for DM at IPC=1) while operating directly on 2-bit images without additional distillation.

Conclusion: The proposed plug-and-play framework effectively enables extreme compression in dataset condensation through localized patch-based quantization, clustering for parameter efficiency, and distribution refinement, significantly advancing storage-efficient synthetic dataset generation.

Abstract: Dataset Condensation (DC) distills knowledge from large datasets into smaller ones, accelerating training and reducing storage requirements. However, despite notable progress, prior methods have largely overlooked the potential of quantization for further reducing storage costs. In this paper, we take the first step to explore post-training quantization in dataset condensation, demonstrating its effectiveness in reducing storage size while maintaining representation quality without requiring expensive training cost. However, we find that at extremely low bit-widths (e.g., 2-bit), conventional quantization leads to substantial degradation in representation quality, negatively impacting the networks trained on these data. To address this, we propose a novel \emph{patch-based post-training quantization} approach that ensures localized quantization with minimal loss of information. To reduce the overhead of quantization parameters, especially for small patch sizes, we employ quantization-aware clustering to identify similar patches and subsequently aggregate them for efficient quantization. Furthermore, we introduce a refinement module to align the distribution between original images and their dequantized counterparts, compensating for quantization errors. Our method is a plug-and-play framework that can be applied to synthetic images generated by various DC methods. Extensive experiments across diverse benchmarks including CIFAR-10/100, Tiny ImageNet, and ImageNet subsets demonstrate that our method consistently outperforms prior works under the same storage constraints. Notably, our method nearly \textbf{doubles the test accuracy} of existing methods at extreme compression regimes (e.g., 26.0% $\rightarrow$ 54.1% for DM at IPC=1), while operating directly on 2-bit images without additional distillation.

Method: 1) Create EditHF-1M dataset with 29M human preference pairs and 148K ratings across visual quality, instruction alignment, and attribute preservation dimensions. 2) Develop EditHF, an MLLM-based evaluation model trained on this data. 3) Use EditHF-Reward as reinforcement learning signal to optimize image editing models.

Result: EditHF achieves superior alignment with human preferences and strong generalization. Fine-tuning Qwen-Image-Edit with EditHF-Reward yields significant performance improvements in image editing quality.

Conclusion: The work provides a scalable human-aligned evaluation framework for image editing that can serve as an effective reward model to improve text-guided image editing systems through reinforcement learning.

Abstract: Recent text-guided image editing (TIE) models have achieved remarkable progress, while many edited images still suffer from issues such as artifacts, unexpected editings, unaesthetic contents. Although some benchmarks and methods have been proposed for evaluating edited images, scalable evaluation models are still lacking, which limits the development of human feedback reward models for image editing. To address the challenges, we first introduce \textbf{EditHF-1M}, a million-scale image editing dataset with over 29M human preference pairs and 148K human mean opinion ratings, both evaluated from three dimensions, \textit{i.e.}, visual quality, instruction alignment, and attribute preservation. Based on EditHF-1M, we propose \textbf{EditHF}, a multimodal large language model (MLLM) based evaluation model, to provide human-aligned feedback from image editing. Finally, we introduce \textbf{EditHF-Reward}, which utilizes EditHF as the reward signal to optimize the text-guided image editing models through reinforcement learning. Extensive experiments show that EditHF achieves superior alignment with human preferences and demonstrates strong generalization on other datasets. Furthermore, we fine-tune the Qwen-Image-Edit using EditHF-Reward, achieving significant performance improvements, which demonstrates the ability of EditHF to serve as a reward model to scale-up the image editing. Both the dataset and code will be released in our GitHub repository: https://github.com/IntMeGroup/EditHF.

Method: Proposes MURE framework with: 1) VLMs as hierarchical multi-resolution encoder, 2) Resolution-level Matryoshka representation learning (RMRL) for feature fusion, 3) Semantic-aware hierarchical clustering mechanism for visual token compression.

Result: Experiments on two VDR benchmarks show MURE consistently beats strong baselines and significantly outperforms ColPali with only 50% of its visual token budget.

Conclusion: MURE successfully addresses the effectiveness-efficiency trade-off in VDR through a novel encoding paradigm that captures multi-scale visual information while maintaining computational efficiency.

Abstract: Visual Document Retrieval (VDR) requires representations that capture both fine-grained visual details and global document structure to ensure retrieval efficacy while maintaining computational efficiency. Existing VDR models struggle to balance effectiveness and efficiency when processing high-resolution documents: they often either lose fine-grained information or generate an excessive number of visual tokens, resulting in significant indexing overhead and high retrieval latency. In this work, we rethink the visual encoding mechanism and propose a new X-VisEmb paradigm that progresses from multi-resolution sampling and encoding, through cross-granularity feature fusion, to adaptive representation distillation. A preliminary study validates its feasibility and effectiveness in capturing complementary visual cues at varying scales. Building on the insights, we develop MURE, a novel framework that employs VLMs as a hierarchical multi-resolution encoder, integrates resolution-level Matryoshka representation learning (RMRL) for effective feature fusion, and applies a semantic-aware hierarchical clustering mechanism for visual token compression. Experiments on two widely used VDR benchmarks show that our MURE framework consistently beats strong baselines. Furthermore, it significantly outperforms ColPali with only 50% of its visual token budget.

Method: Proposes SpectralMoE, a novel PEFT framework using Mixture-of-Experts architecture for local precise refinement of foundation model features. Incorporates depth features estimated from RGB bands to guide fine-tuning. Uses dual-gated MoE that independently routes visual and depth features to top-k experts for specialized refinement, followed by cross-attention to fuse refined structural cues into visual stream.

Result: Extensive experiments show SpectralMoE sets new state-of-the-art on multiple DGSS benchmarks across hyperspectral, multispectral, and RGB remote sensing imagery.

Conclusion: The key to robust domain generalization semantic segmentation lies in fine-grained, spatially-adaptive refinement rather than global adaptation. SpectralMoE effectively addresses spectral variations through modality-specific adjustments guided by depth features.

Abstract: Domain Generalization Semantic Segmentation (DGSS) in spectral remote sensing is severely challenged by spectral shifts across diverse acquisition conditions, which cause significant performance degradation for models deployed in unseen domains. While Parameter-Efficient Fine-Tuning (PEFT) on foundation models is a promising direction, existing methods employ global, homogeneous adjustments. This “one-size-fits-all” tuning struggles with the spatial heterogeneity of land cover, causing semantic confusion. We argue that the key to robust DGSS lies not in a single global adaptation, but in performing fine-grained, spatially-adaptive refinement of a foundation model’s features. To achieve this, we propose SpectralMoE, a novel PEFT framework for DGSS. It operationalizes this principle by utilizing a Mixture-of-Experts (MoE) architecture to perform local precise refinement on the foundation model’s features, incorporating depth features estimated from selected RGB bands of the spectral remote sensing imagery to guide the fine-tuning process. Specifically, SpectralMoE employs a dual-gated MoE architecture that independently routes visual and depth features to top-k selected experts for specialized refinement, enabling modality-specific adjustments. A subsequent cross-attention mechanism then judiciously fuses the refined structural cues into the visual stream, mitigating semantic ambiguities caused by spectral variations. Extensive experiments show that SpectralMoE sets a new state-of-the-art on multiple DGSS benchmarks across hyperspectral, multispectral, and RGB remote sensing imagery.

Method: Introduces ReactMotionNet dataset with speaker utterances paired with multiple candidate listener motions annotated for appropriateness. Proposes ReactMotion framework that jointly models text, audio, emotion, and motion using preference-based objectives to generate diverse and appropriate listener responses.

Result: ReactMotion outperforms retrieval baselines and cascaded LLM-based pipelines, generating more natural, diverse, and appropriate listener motions as validated through extensive experiments.

Conclusion: The paper successfully addresses the challenging task of reactive listener motion generation by introducing appropriate datasets, evaluation protocols, and a unified generative framework that captures the one-to-many nature of human reactions.

Abstract: In this paper, we introduce a new task, Reactive Listener Motion Generation from Speaker Utterance, which aims to generate naturalistic listener body motions that appropriately respond to a speaker’s utterance. However, modeling such nonverbal listener behaviors remains underexplored and challenging due to the inherently non-deterministic nature of human reactions. To facilitate this task, we present ReactMotionNet, a large-scale dataset that pairs speaker utterances with multiple candidate listener motions annotated with varying degrees of appropriateness. This dataset design explicitly captures the one-to-many nature of listener behavior and provides supervision beyond a single ground-truth motion. Building on this dataset design, we develop preference-oriented evaluation protocols tailored to evaluate reactive appropriateness, where conventional motion metrics focusing on input-motion alignment ignore. We further propose ReactMotion, a unified generative framework that jointly models text, audio, emotion, and motion, and is trained with preference-based objectives to encourage both appropriate and diverse listener responses. Extensive experiments show that ReactMotion outperforms retrieval baselines and cascaded LLM-based pipelines, generating more natural, diverse, and appropriate listener motions.

Method: Proposes Anchor Forcing with two key designs: 1) anchor-guided re-cache mechanism that stores KV states in anchor caches for warm-starting at prompt switches, and 2) tri-region RoPE with region-specific reference origins and RoPE re-alignment distillation to reconcile unbounded streaming indices with pretrained RoPE regime.

Result: Experiments show improved perceptual quality and motion metrics over prior streaming baselines in interactive long video generation settings.

Conclusion: Anchor Forcing effectively addresses cache maintenance and RoPE distribution issues in interactive streaming video generation, enabling better quality and motion retention during prompt switching.

Abstract: Interactive long video generation requires prompt switching to introduce new subjects or events, while maintaining perceptual fidelity and coherent motion over extended horizons. Recent distilled streaming video diffusion models reuse a rolling KV cache for long-range generation, enabling prompt-switch interaction through re-cache at each switch. However, existing streaming methods still exhibit progressive quality degradation and weakened motion dynamics. We identify two failure modes specific to interactive streaming generation: (i) at each prompt switch, current cache maintenance cannot simultaneously retain KV-based semantic context and recent latent cues, resulting in weak boundary conditioning and reduced perceptual quality; and (ii) during distillation, unbounded time indexing induces a positional distribution shift from the pretrained backbone’s bounded RoPE regime, weakening pretrained motion priors and long-horizon motion retention. To address these issues, we propose \textbf{Anchor Forcing}, a cache-centric framework with two designs. First, an anchor-guided re-cache mechanism stores KV states in anchor caches and warm-starts re-cache from these anchors at each prompt switch, reducing post-switch evidence loss and stabilizing perceptual quality. Second, a tri-region RoPE with region-specific reference origins, together with RoPE re-alignment distillation, reconciles unbounded streaming indices with the pretrained RoPE regime to better retain motion priors. Experiments on long videos show that our method improves perceptual quality and motion metrics over prior streaming baselines in interactive settings. Project page: https://github.com/vivoCameraResearch/Anchor-Forcing

Method: Introduced AgriPath-LF16 benchmark with 111k images across 16 crops and 41 diseases, explicitly separating lab and field imagery. Compared three model paradigms (CNNs, contrastive VLMs, generative VLMs) under unified training protocols across full, lab-only, and field-only regimes using macro-F1 and Parse Success Rate metrics.

Result: CNNs achieved highest accuracy on lab imagery but degraded under domain shift. Contrastive VLMs provided robust, parameter-efficient alternative with competitive cross-domain performance. Generative VLMs showed strongest resilience to distributional variation but had additional failure modes from free-text generation.

Conclusion: Architectural choice should be guided by deployment context rather than aggregate accuracy alone, with different model paradigms showing distinct strengths for different operational environments.

Abstract: Reliable crop disease detection requires models that perform consistently across diverse acquisition conditions, yet existing evaluations often focus on single architectural families or lab-generated datasets. This work presents a systematic empirical comparison of three model paradigms for fine-grained crop disease classification: Convolutional Neural Networks (CNNs), contrastive Vision-Language Models (VLMs), and generative VLMs. To enable controlled analysis of domain effects, we introduce AgriPath-LF16, a benchmark containing 111k images spanning 16 crops and 41 diseases with explicit separation between laboratory and field imagery, alongside a balanced 30k subset for standardized training and evaluation. All models are trained and evaluated under unified protocols across full, lab-only, and field-only training regimes using macro-F1 and Parse Success Rate (PSR) to account for generative reliability. The results reveal distinct performance profiles. CNNs achieve the highest accuracy on lab imagery but degrade under domain shift. Contrastive VLMs provide a robust and parameter-efficient alternative with competitive cross-domain performance. Generative VLMs demonstrate the strongest resilience to distributional variation, albeit with additional failure modes stemming from free-text generation. These findings highlight that architectural choice should be guided by deployment context rather than aggregate accuracy alone.

Method: Uses cross attention fusion of sparse motion cues (head-hand motion) and scene point clouds to directly interpret spatial intentions within 3D scenes, predicting intention areas without object-level perception.

Result: Evaluated on MoGaze and CIRCLE datasets, showing consistent performance across time horizons up to 1500ms and outperforming baselines, even in diverse/unseen scenes. Demonstrated usability through efficient visual question answering based on intention areas.

Conclusion: Int3DNet provides reliable 3D intention areas from head-hand motion and scene geometry, enabling seamless human-MR interaction through proactive processing of intention areas.

Abstract: We propose Int3DNet, a scene-aware network that predicts 3D intention areas directly from scene geometry and head-hand motion cues, enabling robust human intention prediction without explicit object-level perception. In Mixed Reality (MR), intention prediction is critical as it enables the system to anticipate user actions and respond proactively, reducing interaction delays and ensuring seamless user experiences. Our method employs a cross attention fusion of sparse motion cues and scene point clouds, offering a novel approach that directly interprets the user’s spatial intention within the scene. We evaluated Int3DNet on MoGaze and CIRCLE datasets, which are public datasets for full-body human-scene interactions, showing consistent performance across time horizons of up to 1500 ms and outperforming the baselines, even in diverse and unseen scenes. Moreover, we demonstrate the usability of proposed method through a demonstration of efficient visual question answering (VQA) based on intention areas. Int3DNet provides reliable 3D intention areas derived from head-hand motion and scene geometry, thus enabling seamless interaction between humans and MR systems through proactive processing of intention areas.

Method: Proposes MM-RoPE for better 3D positional encoding in videos, and uses parallel mask-based discrete diffusion with intra-frame bidirectional and inter-frame causal attention masks. Introduces Autoregressive Discrete Diffusion Forcing to address frame-wise loss imbalance.

Result: Achieves state-of-the-art results on GenEval, VBench-I2V, and VBench-T2V benchmarks, surpassing Show-o2, COSMOS-Video2World, and OpenSoraPlan despite using only 48 GPUs and limited data.

Conclusion: Lumos-1 demonstrates that LLM-based autoregressive video generation can be efficient and effective with proper architectural adaptations, achieving strong performance with modest computational resources.

Abstract: Autoregressive large language models (LLMs) have unified a vast range of language tasks, inspiring preliminary efforts in autoregressive (AR) video generation. Existing AR video generators either diverge from standard LLM architectures, depend on bulky external text encoders, or incur prohibitive latency due to next-token decoding. In this paper, we introduce Lumos-1, an LLM-based unified model for AR video generation with efficient discrete diffusion. Firstly, to fit videos with LLMs, we identify that 1D RoPE is ill-suited for visual spatiotemporal correlation modeling, and while demonstrated to be useful, naive 3D RoPE exhibits imbalanced frequency spectra. Therefore, we propose MM-RoPE, which preserves the original textual RoPE while seamlessly accommodating video data with comprehensive frequency spectra and scaled 3D positions. Secondly, to fit the video data’s nature and overcome the inefficiency of next-token decoding, we adopt a parallel and mask-based discrete diffusion with the intra-frame bidirectional and inter-frame causal attention masks. Based on this attention mask, we uncover the frame-wise loss imbalance issue caused by spatial information redundancy and propose Autoregressive Discrete Diffusion Forcing, which introduces temporal tube masking during training with a compatible inference-time masking policy to avoid quality degradation. Despite using only 48 GPUs for pre-training and fine-tuning, limited data and a discrete tokenizer, Lumos-1 achieves results surpassing those of Show-o2 on GenEval, COSMOS-Video2World on VBench-I2V, and OpenSoraPlan on VBench-T2V. Code and models are available at https://github.com/alibaba-damo-academy/Lumos.

Method: Statistical Shape Model (SSM) generates 8,800 annotated liver shapes from single manually labeled mean shape; specialized deep learning network trained on this synthetic data for landmark segmentation.

Result: Mean IoU of 91.4% on 500 unseen synthetic shapes (87.4% for anterior ridge, 87.6% for falciform ligament); promising qualitative results on clinical patient liver shapes.

Conclusion: SSM-based data generation reduces manual labeling burden and enables creation of large training datasets; methodology generalizable beyond liver anatomy to other applications requiring annotated training data.

Abstract: Anatomical landmark segmentation serves as a critical initial step for robust multimodal registration during computer-assisted interventions. Current approaches predominantly rely on deep learning, which often necessitates the extensive manual generation of annotated datasets. In this paper, we present a novel strategy for creating large annotated datasets using a statistical shape model (SSM) based on a mean shape that is manually labeled only once. We demonstrate the method’s efficacy through its application to deep-learning-based anatomical landmark segmentation, specifically targeting the detection of the anterior ridge and the falciform ligament in 3D liver shapes. A specialized deep learning network was trained with 8,800 annotated liver shapes generated by the SSM. The network’s performance was evaluated on 500 unseen synthetic SSM shapes, yielding a mean Intersection over Union of 91.4% (87.4% for the anterior ridge and 87.6% for the falciform ligament). Subsequently, the network was applied to clinical patient liver shapes, with qualitative evaluation indicating promising results and highlighting the generalizability of the proposed approach. Our findings suggest that the SSM-based data generation approach alleviates the labor-intensive process of manual labeling while enabling the creation of large annotated training datasets for machine learning. Although our study focuses on liver anatomy, the proposed methodology holds potential for a broad range of applications where annotated training datasets play a pivotal role in developing accurate deep-learning models.

Method: Proposes Bi-CamoDiffusion, an evolution of CamoDiffusion that integrates edge priors into early-stage embeddings via parameter-free injection. Uses unified optimization objective balancing spatial accuracy, structural constraints, and uncertainty supervision to capture both global context and boundary transitions.

Result: Outperforms baseline and state-of-the-art methods across CAMO, COD10K, and NC4K benchmarks. Achieves superior performance in metrics including S_m, F_β^w, E_m, and MAE, with sharper delineation of thin structures and reduced false positives.

Conclusion: Bi-CamoDiffusion effectively enhances camouflaged object detection by incorporating edge priors into diffusion models, achieving more precise object-background separation and sharper boundary recovery than existing approaches.

Abstract: Bi-CamoDiffusion is introduced, an evolution of the CamoDiffusion framework for camouflaged object detection. It integrates edge priors into early-stage embeddings via a parameter-free injection process, which enhances boundary sharpness and prevents structural ambiguity. This is governed by a unified optimization objective that balances spatial accuracy, structural constraints, and uncertainty supervision, allowing the model to capture of both the object’s global context and its intricate boundary transitions. Evaluations across the CAMO, COD10K, and NC4K benchmarks show that Bi-CamoDiffusion surpasses the baseline, delivering sharper delineation of thin structures and protrusions while also minimizing false positives. Also, our model consistently outperforms existing state-of-the-art methods across all evaluated metrics, including $S_m$, $F_β^{w}$, $E_m$, and $MAE$, demonstrating a more precise object-background separation and sharper boundary recovery.

Method: Treats temporal neighborhood of target links as sequence of “graph frames”, stacks them into “graph videos”, leverages video foundation models’ inductive biases to capture local variations and long-range dynamics, generates lightweight plug-and-play link-level embeddings.

Result: Outperforms state-of-the-art baselines on link prediction tasks in most cases across benchmark datasets.

Conclusion: Borrowing spatio-temporal modeling techniques from computer vision is a promising approach for advancing dynamic graph learning.

Abstract: Dynamic graphs are common in real-world systems such as social media, recommender systems, and traffic networks. Existing dynamic graph models for link prediction often fall short in capturing the complexity of temporal evolution. They tend to overlook fine-grained variations in temporal interaction order, struggle with dependencies that span long time horizons, and offer limited capability to model pair-specific relational dynamics. To address these challenges, we propose \textbf{Graph2Video}, a video-inspired framework that views the temporal neighborhood of a target link as a sequence of “graph frames”. By stacking temporally ordered subgraph frames into a “graph video”, Graph2Video leverages the inductive biases of video foundation models to capture both fine-grained local variations and long-range temporal dynamics. It generates a link-level embedding that serves as a lightweight and plug-and-play link-centric memory unit. This embedding integrates seamlessly into existing dynamic graph encoders, effectively addressing the limitations of prior approaches. Extensive experiments on benchmark datasets show that Graph2Video outperforms state-of-the-art baselines on the link prediction task in most cases. The results highlight the potential of borrowing spatio-temporal modeling techniques from computer vision as a promising and effective approach for advancing dynamic graph learning.

Method: BrainCast formulates fMRI time series forecasting as multivariate time series prediction with three modules: 1) Spatial Interaction Awareness module that models inter-ROI dependencies using token embeddings, 2) Temporal Feature Refinement module that captures neural dynamics within each ROI by enhancing both low- and high-energy temporal components, and 3) Spatio-temporal Pattern Alignment module that combines spatial and temporal representations.

Result: Experiments on resting-state and task fMRI datasets from the Human Connectome Project show BrainCast outperforms state-of-the-art time series forecasting baselines. Extended fMRI time series improve downstream cognitive ability prediction.

Conclusion: BrainCast demonstrates effective whole-brain fMRI time series forecasting that can enhance clinical and neuroscientific research in scenarios with restricted scan durations by generating extended, informative time series data.

Abstract: Functional magnetic resonance imaging (fMRI) enables noninvasive investigation of brain function, while short clinical scan durations, arising from human and non-human factors, usually lead to reduced data quality and limited statistical power for neuroimaging research. In this paper, we propose BrainCast, a novel spatio-temporal forecasting framework specifically tailored for whole-brain fMRI time series forecasting, to extend informative fMRI time series without additional data acquisition. It formulates fMRI time series forecasting as a multivariate time series prediction task and jointly models temporal dynamics within regions of interest (ROIs) and spatial interactions across ROIs. Specifically, BrainCast integrates a Spatial Interaction Awareness module to characterize inter-ROI dependencies via embedding every ROI time series as a token, a Temporal Feature Refinement module to capture intrinsic neural dynamics within each ROI by enhancing both low- and high-energy temporal components of fMRI time series at the ROI level, and a Spatio-temporal Pattern Alignment module to combine spatial and temporal representations for producing informative whole-brain features. Experimental results on resting-state and task fMRI datasets from the Human Connectome Project demonstrate the superiority of BrainCast over state-of-the-art time series forecasting baselines. Moreover, fMRI time series extended by BrainCast improve downstream cognitive ability prediction, highlighting the clinical and neuroscientific impact brought by whole-brain fMRI time series forecasting in scenarios with restricted scan durations.

Method: Uses hierarchical binder structure for cross-attention conditioning in Diffusion Transformers to encode visual concepts into prompt tokens. Introduces Diversify-and-Absorb Mechanism to improve binding accuracy by using absorbent tokens to filter irrelevant details. Presents Temporal Disentanglement Strategy with dual-branch binder for video concepts, decoupling training into two stages for temporal modeling.

Result: Achieves superior concept consistency, prompt fidelity, and motion quality compared to existing approaches, enabling new possibilities for visual creativity.

Conclusion: Bind & Compose provides an effective one-shot solution for flexible visual concept composition that works across both images and videos with improved accuracy and compatibility.

Abstract: Visual concept composition, which aims to integrate different elements from images and videos into a single, coherent visual output, still falls short in accurately extracting complex concepts from visual inputs and flexibly combining concepts from both images and videos. We introduce Bind & Compose, a one-shot method that enables flexible visual concept composition by binding visual concepts with corresponding prompt tokens and composing the target prompt with bound tokens from various sources. It adopts a hierarchical binder structure for cross-attention conditioning in Diffusion Transformers to encode visual concepts into corresponding prompt tokens for accurate decomposition of complex visual concepts. To improve concept-token binding accuracy, we design a Diversify-and-Absorb Mechanism that uses an extra absorbent token to eliminate the impact of concept-irrelevant details when training with diversified prompts. To enhance the compatibility between image and video concepts, we present a Temporal Disentanglement Strategy that decouples the training process of video concepts into two stages with a dual-branch binder structure for temporal modeling. Evaluations demonstrate that our method achieves superior concept consistency, prompt fidelity, and motion quality over existing approaches, opening up new possibilities for visual creativity.

Method: SING constructs equivalent images with respect to a network’s invariants and maps network features to multimodal vision-language models to obtain natural language descriptions and visual examples of semantic shifts. Can be applied to single images (local invariants) or sets of images (statistical analysis at class/model levels).

Result: Reveals that ResNet50 leaks relevant semantic attributes to null space, while DinoViT (ViT pretrained with self-supervised DINO) better maintains class semantics across invariant space. Provides interpretable semantic analysis of model invariants.

Conclusion: SING enables semantic interpretation of classifier invariants using vision-language models, offering insights into how different architectures handle semantic variations and which semantic attributes they ignore.

Abstract: All classifiers, including state-of-the-art vision models, possess invariants, partially rooted in the geometry of their linear mappings. These invariants, which reside in the null-space of the classifier, induce equivalent sets of inputs that map to identical outputs. The semantic content of these invariants remains vague, as existing approaches struggle to provide human-interpretable information. To address this gap, we present Semantic Interpretation of the Null-space Geometry (SING), a method that constructs equivalent images, with respect to the network, and assigns semantic interpretations to the available variations. We use a mapping from network features to multi-modal vision language models. This allows us to obtain natural language descriptions and visual examples of the induced semantic shifts. SING can be applied to a single image, uncovering local invariants, or to sets of images, allowing a breadth of statistical analysis at the class and model levels. For example, our method reveals that ResNet50 leaks relevant semantic attributes to the null space, whereas DinoViT, a ViT pretrained with self-supervised DINO, is superior in maintaining class semantics across the invariant space.

Method: Uses a teacher-student auto-encoder setup with progressive learning where multi-scale clean image decoder feature maps are distilled into each layer of the student decoder using Illumination-Aware Mirror Loss (IAML), which accounts for lighting variations.

Result: Achieves state-of-the-art performance on three popular low-light image enhancement datasets in terms of SSIM, PSNR, and LPIPS metrics. Ablation studies confirm the effectiveness of IAML.

Conclusion: The proposed IAML-based teacher-student approach effectively enhances low-light images by better aligning feature maps while considering illumination variations, outperforming existing methods.

Abstract: This letter presents a novel training approach and loss function for learning low-light image enhancement auto-encoders. Our approach revolves around the use of a teacher-student auto-encoder setup coupled to a progressive learning approach where multi-scale information from clean image decoder feature maps is distilled into each layer of the student decoder in a mirrored fashion using a newly-proposed loss function termed Illumination-Aware Mirror Loss (IAML). IAML helps aligning the feature maps within the student decoder network with clean feature maps originating from the teacher side while taking into account the effect of lighting variations within the input images. Extensive benchmarking of our proposed approach on three popular low-light image enhancement datasets demonstrate that our model achieves state-of-the-art performance in terms of average SSIM, PSNR and LPIPS reconstruction accuracy metrics. Finally, ablation studies are performed to clearly demonstrate the effect of IAML on the image reconstruction accuracy.

Method: Proposes FineRMoE (FineR-Grained MoE) that extends fine-grained expert design to both intermediate and output dimensions. Introduces bi-level sparse forward computation paradigm and specialized routing mechanism. Also develops a generalized upcycling method to build FineRMoE cost-effectively instead of training from scratch.

Result: FineRMoE achieves superior performance across ten standard benchmarks. Compared with strongest baseline: 6x higher parameter efficiency, 281x lower prefill latency, and 136x higher decoding throughput during inference.

Conclusion: FineRMoE successfully overcomes the single-dimension limitation of fine-grained MoE by extending expert specialization to both intermediate and output dimensions, achieving significant efficiency and performance gains through innovative architecture and cost-effective upcycling.

Abstract: As revealed by the scaling law of fine-grained MoE, model performance ceases to be improved once the granularity of the intermediate dimension exceeds the optimal threshold, limiting further gains from single-dimension fine-grained design. To address this bottleneck, we propose FineRMoE (FineR-Grained MoE), an architecture that extends fine-grained expert design to both intermediate and output dimensions, aiming to enhance expert specialization beyond the single-dimension limit. We further introduce a bi-level sparse forward computation paradigm and a specialized routing mechanism to govern the activation. In addition, to obviate the prohibitive cost of training FineRMoE from scratch, we devise a generalized upcycling method to build FineRMoE in a cost-effective manner. Extensive experiments demonstrate the superior performance achieved by FineRMoE across ten standard benchmarks. Compared with the strongest baseline, FineRMoE achieves 6 times higher parameter efficiency, 281 times lower prefill latency, and 136 timese higher decoding throughput during inference.

Method: Proposes GOT-JEPA framework where teacher predictor generates pseudo-tracking models from clean frames and student predictor learns to predict same models from corrupted frames. Also introduces OccuSolver for object-aware visibility estimation and detailed occlusion-pattern capture using point-centric tracking.

Result: Extensive evaluations on seven benchmarks show the method effectively enhances tracker generalization and robustness, with improved occlusion handling and adaptation to dynamic environments.

Conclusion: The proposed GOT-JEPA framework with OccuSolver significantly improves object tracking performance by enhancing generalization capabilities and providing detailed occlusion perception through model-predictive pretraining.

Abstract: The human visual system tracks objects by integrating current observations with previously observed information, adapting to target and scene changes, and reasoning about occlusion at fine granularity. In contrast, recent generic object trackers are often optimized for training targets, which limits robustness and generalization in unseen scenarios, and their occlusion reasoning remains coarse, lacking detailed modeling of occlusion patterns. To address these limitations in generalization and occlusion perception, we propose GOT-JEPA, a model-predictive pretraining framework that extends JEPA from predicting image features to predicting tracking models. Given identical historical information, a teacher predictor generates pseudo-tracking models from a clean current frame, and a student predictor learns to predict the same pseudo-tracking models from a corrupted version of the current frame. This design provides stable pseudo supervision and explicitly trains the predictor to produce reliable tracking models under occlusions, distractors, and other adverse observations, improving generalization to dynamic environments. Building on GOT-JEPA, we further propose OccuSolver to enhance occlusion perception for object tracking. OccuSolver adapts a point-centric point tracker for object-aware visibility estimation and detailed occlusion-pattern capture. Conditioned on object priors iteratively generated by the tracker, OccuSolver incrementally refines visibility states, strengthens occlusion handling, and produces higher-quality reference labels that progressively improve subsequent model predictions. Extensive evaluations on seven benchmarks show that our method effectively enhances tracker generalization and robustness.

Method: Proposes Extrapolative Domain Adaptive Panoramic Segmentation (EDA-PSeg) with Euler-Margin Attention (EMA) that introduces angular margin for viewpoint-invariant representation and amplitude/phase modulation for generalization to unseen classes, plus Graph Matching Adapter (GMA) that builds high-order graph relations to align shared semantics across FoV shifts while separating novel categories.

Result: Extensive experiments on four benchmark datasets under camera-shift, weather-condition, and open-set scenarios demonstrate state-of-the-art performance, robust generalization to diverse viewing geometries, and resilience under varying environmental conditions.

Conclusion: EDA-PSeg effectively addresses both geometric FoV shifts and semantic uncertainty in cross-domain panoramic segmentation, achieving strong performance across diverse real-world scenarios.

Abstract: Cross-domain panoramic semantic segmentation has attracted growing interest as it enables comprehensive 360° scene understanding for real-world applications. However, it remains particularly challenging due to severe geometric Field of View (FoV) distortions and inconsistent open-set semantics across domains. In this work, we formulate an open-set domain adaptation setting, and propose Extrapolative Domain Adaptive Panoramic Segmentation (EDA-PSeg) framework that trains on local perspective views and tests on full 360° panoramic images, explicitly tackling both geometric FoV shifts across domains and semantic uncertainty arising from previously unseen classes. To this end, we propose the Euler-Margin Attention (EMA), which introduces an angular margin to enhance viewpoint-invariant semantic representation, while performing amplitude and phase modulation to improve generalization toward unseen classes. Additionally, we design the Graph Matching Adapter (GMA), which builds high-order graph relations to align shared semantics across FoV shifts while effectively separating novel categories through structural adaptation. Extensive experiments on four benchmark datasets under camera-shift, weather-condition, and open-set scenarios demonstrate that EDA-PSeg achieves state-of-the-art performance, robust generalization to diverse viewing geometries, and resilience under varying environmental conditions. The code is available at https://github.com/zyfone/EDA-PSeg.

Method: Proposes WaveComm: uses Discrete Wavelet Transform to decompose feature maps, transmits only compact low-frequency components, omits high-frequency details, and reconstructs them at receiver using lightweight generator with Multi-Scale Distillation Loss for optimization.

Result: Maintains state-of-the-art performance on OPV2V and DAIR-V2X datasets for LiDAR and camera perception tasks while reducing communication volume to 86.3% and 87.0% of original, achieving competitive improvements in communication efficiency and perception accuracy.

Conclusion: WaveComm effectively addresses communication bottlenecks in multi-agent collaborative sensing through wavelet-based compression and reconstruction, enabling efficient bandwidth usage without sacrificing perception performance.

Abstract: In multi-agent collaborative sensing systems, substantial communication overhead from information exchange significantly limits scalability and real-time performance, especially in bandwidth-constrained environments. This often results in degraded performance and reduced reliability. To address this challenge, we propose WaveComm, a wavelet-based communication framework that drastically reduces transmission loads while preserving sensing performance in low-bandwidth scenarios. The core innovation of WaveComm lies in decomposing feature maps using Discrete Wavelet Transform (DWT), transmitting only compact low-frequency components to minimize communication overhead. High-frequency details are omitted, and their effects are reconstructed at the receiver side using a lightweight generator. A Multi-Scale Distillation (MSD) Loss is employed to optimize the reconstruction quality across pixel, structural, semantic, and distributional levels. Experiments on the OPV2V and DAIR-V2X datasets for LiDAR-based and camera-based perception tasks demonstrate that WaveComm maintains state-of-the-art performance even when the communication volume is reduced to 86.3% and 87.0% of the original, respectively. Compared to existing approaches, WaveComm achieves competitive improvements in both communication efficiency and perception accuracy. Ablation studies further validate the effectiveness of its key components.

Method: Latent Entropy-Aware Decoding (LEAD) uses entropy-aware reasoning mode switching: under high-entropy states, the model employs probability-weighted continuous embeddings from token probability distributions, then transitions back to discrete token embeddings as entropy decreases. Also includes prior-guided visual anchor injection to encourage focus on visual information.

Result: Extensive experiments show LEAD effectively mitigates hallucinations across various multimodal large reasoning models on multiple benchmarks, demonstrating improved reliability in visual question answering tasks.

Conclusion: LEAD provides an efficient plug-and-play decoding strategy that leverages semantic context from token probability distributions to achieve more reliable reasoning in multimodal models, particularly addressing hallucination issues during high-entropy reasoning stages.

Abstract: Recent advancements in multimodal large reasoning models (MLRMs) have significantly improved performance in visual question answering. However, we observe that transition words (e.g., because, however, and wait) are closely associated with hallucinations and tend to exhibit high-entropy states. We argue that adequate contextual reasoning information can be directly extracted from the token probability distribution. Inspired by superposed representation theory, we propose leveraging latent superposed reasoning to integrate multiple candidate semantics and maintain latent reasoning trajectories. The hypothesis is that reliance on discrete textual inputs may drive the model toward sequential explicit reasoning, underutilizing dense contextual cues during high-entropy reasoning stages. Therefore, we propose constructing rich semantic representations from the token probability distributions to enhance in-context reasoning. With this goal, we present Latent Entropy-Aware Decoding (LEAD), an efficient plug-and-play decoding strategy that leverages semantic context to achieve reliable reasoning. The heart of our method lies in entropy-aware reasoning mode switching. The model employs probability-weighted continuous embeddings under high-entropy states and transitions back to discrete token embeddings as entropy decreases. Moreover, we propose a prior-guided visual anchor injection strategy that encourages the model to focus on visual information. Extensive experiments show that LEAD effectively mitigates hallucinations across various MLRMs on multiple benchmarks.

Method: Developed a multimodal framework with 3D convolutional neural networks (3D CNNs) to extract spatial features from MRI scans, and recurrent architectures (LSTMs) to learn temporal patterns from fMRI sequences. Features from both modalities are fused for joint spatial-temporal learning. Experiments compare performance with and without data augmentation on a small paired dataset (29 subjects).

Result: Data augmentation substantially improved classification stability and generalization for the multimodal 3DCNN-LSTM model on the small paired dataset. However, augmentation was ineffective for a large-scale single-modality MRI dataset, highlighting that augmentation benefits depend on dataset size and modality.

Conclusion: Multimodal integration of MRI and fMRI with appropriate data augmentation strategies can improve Alzheimer’s disease classification, especially for small datasets. The effectiveness of augmentation depends on both dataset size and modality characteristics.

Abstract: Magnetic Resonance Imaging (MRI) provides detailed structural information, while functional MRI (fMRI) captures temporal brain activity. In this work, we present a multimodal deep learning framework that integrates MRI and fMRI for multi-class classification of Alzheimer Disease (AD), Mild Cognitive Impairment, and Normal Cognitive State. Structural features are extracted from MRI using 3D convolutional neural networks, while temporal features are learned from fMRI sequences using recurrent architectures. These representations are fused to enable joint spatial-temporal learning. Experiments were conducted on a small paired MRI-fMRI dataset (29 subjects), both with and without data augmentation. Results show that data augmentation substantially improves classification stability and generalization, particularly for the multimodal 3DCNN-LSTM model. In contrast, augmentation was found to be ineffective for a large-scale single-modality MRI dataset. These findings highlight the importance of dataset size and modality when designing augmentation strategies for neuroimaging-based AD classification.

Method: Proposes Co-SemDepth joint architecture for depth estimation and semantic segmentation. Uses synthetic datasets (TopAir for aerial, MidSea for marine). Analyzes synthetic-to-real generalization with TaskPrompter comparison. Explores style transfer (Cycle-GAN, Diffusion models) to bridge domain gap.

Result: Co-SemDepth shows superior generalization for depth estimation, TaskPrompter better for semantic segmentation. Diffusion models outperform Cycle-GAN for synthetic-to-real style transfer. Good generalization on marine SMD dataset, needs improvement on MIT dataset.

Conclusion: Joint learning approach with synthetic data and style transfer techniques can address data scarcity in aerial robotics, though domain adaptation remains challenging and requires further enhancement for certain real-world datasets.

Abstract: In this thesis, we leverage monocular cameras on aerial robots to predict depth and semantic maps in low-altitude unstructured environments. We propose a joint deep-learning architecture, named Co-SemDepth, that can perform the two tasks accurately and rapidly, and validate its effectiveness on a variety of datasets. The training of neural networks requires an abundance of annotated data, and in the UAV field, the availability of such data is limited. We introduce a new synthetic dataset in this thesis, TopAir that contains images captured with a nadir view in outdoor environments at different altitudes, helping to fill the gap. While using synthetic data for the training is convenient, it raises issues when shifting to the real domain for testing. We conduct an extensive analytical study to assess the effect of several factors on the synthetic-to-real generalization. Co-SemDepth and TaskPrompter models are used for comparison in this study. The results reveal a superior generalization performance for Co-SemDepth in depth estimation and for TaskPrompter in semantic segmentation. Also, our analysis allows us to determine which training datasets lead to a better generalization. Moreover, to help attenuate the gap between the synthetic and real domains, image style transfer techniques are explored on aerial images to convert from the synthetic to the realistic style. Cycle-GAN and Diffusion models are employed. The results reveal that diffusion models are better in the synthetic to real style transfer. In the end, we focus on the marine domain and address its challenges. Co-SemDepth is trained on a collected synthetic marine data, called MidSea, and tested on both synthetic and real data. The results reveal good generalization performance of Co-SemDepth when tested on real data from the SMD dataset while further enhancement is needed on the MIT dataset.

Method: Introduces a formulation that disentangles prompt ambiguity (inter-user variability) from local sensitivity (interaction imprecision) to measure prompt dependence. Validates on two female pelvic MRI datasets for uterus and bladder segmentation.

Result: Experiments show strong negative correlation between both metrics (prompt ambiguity and local sensitivity) and segmentation performance. The two metrics have low mutual correlation, supporting the disentangled design and providing meaningful indicators of prompt-related failure modes.

Conclusion: The framework offers an interpretable way to assess segmentation robustness to prompt variations, which is valuable for safety-critical applications like medical imaging where user variability affects reliability.

Abstract: Promptable segmentation models (e.g., the Segment Anything Models) enable generalizable, zero-shot segmentation across diverse domains. Although predictions are deterministic for a fixed image-prompt pair, the robustness of these models to variations in user prompts, referred to as prompt dependence, remains underexplored. In safety-critical workflows with substantial inter-user variability, interpretable and informative frameworks are needed to evaluate prompt dependence. In this work, we assess the reliability of promptable segmentation by analyzing and measuring its sensitivity to prompt variability. We introduce the first formulation of prompt dependence that explicitly disentangles prompt ambiguity (inter-user variability) from local sensitivity (interaction imprecision), offering an interpretable view of segmentation robustness. Experiments on two female pelvic MRI datasets for uterus and bladder segmentation reveal a strong negative correlation between both metrics and segmentation performance, highlighting the value of our framework for assessing robustness. The two metrics have low mutual correlation, supporting the disentangled design of our formulation, and provide meaningful indicators of prompt-related failure modes.

Method: Proposes GraphVLM benchmark with three integration paradigms: (1) VLM-as-Encoder: enriches graph neural networks through multimodal feature fusion; (2) VLM-as-Aligner: bridges modalities in latent/linguistic space for LLM-based structured reasoning; (3) VLM-as-Predictor: directly uses VLMs as multimodal backbones for graph learning tasks.

Result: Extensive experiments across six datasets from diverse domains show VLMs enhance multimodal graph learning via all three roles. VLM-as-Predictor achieves the most substantial and consistent performance gains, revealing VLMs’ untapped potential as a foundation for multimodal graph learning.

Conclusion: GraphVLM demonstrates that VLMs can significantly enhance multimodal graph learning, with VLM-as-Predictor showing the most promise. This opens new directions for using VLMs as foundational models for structured multimodal reasoning tasks.

Abstract: Vision-Language Models (VLMs) have demonstrated remarkable capabilities in aligning and understanding multimodal signals, yet their potential to reason over structured data, where multimodal entities are connected through explicit relational graphs, remains largely underexplored. Unlocking this capability is crucial for real-world applications such as social networks, recommendation systems, and scientific discovery, where multimodal information is inherently structured. To bridge this gap, we present GraphVLM, a systematic benchmark designed to evaluate and harness the capabilities of VLMs for multimodal graph learning (MMGL). GraphVLM investigates three complementary paradigms for integrating VLMs with graph reasoning: (1) VLM-as-Encoder, which enriches graph neural networks through multimodal feature fusion; (2) VLM-as-Aligner, which bridges modalities in latent or linguistic space to facilitate LLM-based structured reasoning; and (3) VLM-as-Predictor, which directly employs VLMs as multimodal backbones for graph learning tasks. Extensive experiments across six datasets from diverse domains demonstrate that VLMs enhance multimodal graph learning via all three roles. Among these paradigms, VLM-as-Predictor achieves the most substantial and consistent performance gains, revealing the untapped potential of vision-language models as a new foundation for multimodal graph learning. The benchmark code is publicly available at https://github.com/oamyjin/GraphVLM.

Method: Agentic LLM workflow that generates diverse candidate VOIs using vision transformer models trained with different objective preferences, then selects optimal placement based on quantitative metrics

Result: On 110 clinical brain tumor cases, the workflow achieved improved solid tumor coverage and necrosis avoidance compared to general-purpose expert placements, adaptable to different clinical objectives

Conclusion: The proposed workflow provides a strategy to adapt VOI placement to different clinical objectives without retraining task-specific models, addressing inter-operator variability

Abstract: Magnetic resonance spectroscopy (MRS) provides clinically valuable metabolic characterization of brain tumors, but its utility depends on accurate placement of the spectroscopy volume-of-interest (VOI). However, VOI placement typically has a broad operating window: for a given tumor there are multiple possible VOIs that would lead to high-quality MRS measurements. Thus, a VOI place-ment can be tuned for clinician preference, case-specific anatomy, and clinical pri-orities, which leads to high inter-operator variability, especially for heterogeneous tumors. We propose an agentic large language model (LLM) workflow that de-composes VOI placement into generation of diverse candidate VOIs, from which the LLM selects an optimal one based on quantitative metrics. Candidate VOIs are generated by vision transformer-based placement models trained with differ-ent objective function preferences, which allows selection from acceptable alterna-tives rather than a single deterministic placement. On 110 clinical brain tumor cas-es, the agentic workflow achieves improved solid tumor coverage and necrosis avoidance depending on the user preferences compared to the general-purpose expert placements. Overall, the proposed workflow provides a strategy to adapt VOI placement to different clinical objectives without retraining task-specific models.

Method: Projects caption-derived scene representations into hyperbolic space to preserve hierarchical structure, performs anomaly assessment via adaptive question answering over frozen LLM, optimizes lightweight learnable prompts at test time using unsupervised confidence-sparsity objective, and incorporates covariance-aware Mahalanobis refinement for cross-modal alignment.

Result: Achieves 90.03% AUC on XD-Violence, 83.24% on UCF-Crime, 96.95% on ShanghaiTech, and 98.81% on UCSD Ped2, consistently outperforming prior training-free methods.

Conclusion: Geometry-aware representation and adaptive semantic calibration provide a principled and effective alternative to static Euclidean matching in training-free video anomaly detection.

Abstract: Training-free video anomaly detection (VAD) has recently emerged as a scalable alternative to supervised approaches, yet existing methods largely rely on static prompting and geometry-agnostic feature fusion. As a result, anomaly inference is often reduced to shallow similarity matching over Euclidean embeddings, leading to unstable predictions and limited interpretability, especially in complex or hierarchically structured scenes. We introduce MM-VAD, a geometry-aware semantic reasoning framework for training free VAD that reframes anomaly detection as adaptive test-time inference rather than fixed feature comparison. Our approach projects caption-derived scene representations into hyperbolic space to better preserve hierarchical structure and performs anomaly assessment through an adaptive question answering process over a frozen large language model. A lightweight, learnable prompt is optimised at test time using an unsupervised confidence-sparsity objective, enabling context-specific calibration without updating any backbone parameters. To further ground semantic predictions in visual evidence, we incorporate a covariance-aware Mahalanobis refinement that stabilises cross-modal alignment. Across four benchmarks, MM-VAD consistently improves over prior training-free methods, achieving 90.03% AUC on XD-Violence and 83.24%, 96.95%, and 98.81% on UCF-Crime, ShanghaiTech, and UCSD Ped2, respectively. Our results demonstrate that geometry-aware representation and adaptive semantic calibration provide a principled and effective alternative to static Euclidean matching in training-free VAD.

Method: Proposes Deformation-Invariant Neural Network (DINN) with Quasiconformal Transformer Network (QCTN) component. QCTN outputs quasiconformal maps to transform distorted images closer to natural image distribution by controlling Beltrami coefficient for geometric distortion control.

Result: DINN achieves accurate classification of distorted images, outperforms GAN-based restoration methods for atmospheric/water turbulence distortion, and achieves satisfactory 1-1 verification of face images under atmospheric turbulence.

Conclusion: DINN framework effectively handles geometric distortions through quasiconformal transformations, enhancing existing deep networks’ performance on distorted images across various applications.

Abstract: Images degraded by geometric distortions pose a significant challenge to imaging and computer vision tasks such as object recognition. Deep learning-based imaging models usually fail to give accurate performance for geometrically distorted images. In this paper, we propose the deformation-invariant neural network (DINN), a framework to address the problem of imaging tasks for geometrically distorted images. The DINN outputs consistent latent features for images that are geometrically distorted but represent the same underlying object or scene. The idea of DINN is to incorporate a simple component, called the quasiconformal transformer network (QCTN), into other existing deep networks for imaging tasks. The QCTN is a deep neural network that outputs a quasiconformal map, which can be used to transform a geometrically distorted image into an improved version that is closer to the distribution of natural or good images. It first outputs a Beltrami coefficient, which measures the quasiconformality of the output deformation map. By controlling the Beltrami coefficient, the local geometric distortion under the quasiconformal mapping can be controlled. The QCTN is lightweight and simple, which can be readily integrated into other existing deep neural networks to enhance their performance. Leveraging our framework, we have developed an image classification network that achieves accurate classification of distorted images. Our proposed framework has been applied to restore geometrically distorted images by atmospheric turbulence and water turbulence. DINN outperforms existing GAN-based restoration methods under these scenarios, demonstrating the effectiveness of the proposed framework. Additionally, we apply our proposed framework to the 1-1 verification of human face images under atmospheric turbulence and achieve satisfactory performance, further demonstrating the efficacy of our approach.

Method: Two main contributions: (1) Automated pipeline for high-fidelity 3D dance reconstruction from monocular videos using Foot Restoration Diffusion Model with physical constraints; (2) ChoreoLLaMA architecture with retrieval-augmented generation for robustness and slow/fast-cadence Mixture-of-Experts for tempo adaptation.

Result: Created a diverse 100.69-hour multimodal 3D dance dataset and demonstrated superior performance across diverse dance genres compared to existing methods in both qualitative and quantitative evaluations.

Conclusion: The approach marks a step toward scalable, real-world 3D dance generation by addressing data scarcity and model generalization challenges through scaled-up data collection and innovative architecture design.

Abstract: Although existing 3D dance generation methods perform well in controlled scenarios, they often struggle to generalize in the wild. When conditioned on unseen music, existing methods often produce unstructured or physically implausible dance, largely due to limited music-to-dance data and restricted model capacity. This work aims to push the frontier of generalizable 3D dance generation by scaling up both data and model design. (1) On the data side, we develop a fully automated pipeline that reconstructs high-fidelity 3D dance motions from monocular videos. To eliminate the physical artifacts prevalent in existing reconstruction methods, we introduce a Foot Restoration Diffusion Model (FRDM) guided by foot-contact and geometric constraints that enforce physical plausibility while preserving kinematic smoothness and expressiveness, resulting in a diverse, high-quality multimodal 3D dance dataset totaling 100.69 hours. (2) On model design, we propose Choreographic LLaMA (ChoreoLLaMA), a scalable LLaMA-based architecture. To enhance robustness under unfamiliar music conditions, we integrate a retrieval-augmented generation (RAG) module that injects reference dance as a prompt. Additionally, we design a slow/fast-cadence Mixture-of-Experts (MoE) module that enables ChoreoLLaMA to smoothly adapt motion rhythms across varying music tempos. Extensive experiments across diverse dance genres show that our approach surpasses existing methods in both qualitative and quantitative evaluations, marking a step toward scalable, real-world 3D dance generation. Code, models, and data will be released.

Method: A pipeline with preprocessing (data augmentation, ROI identification), detection using various CNN models, postprocessing, and visualization to render 3D bone models highlighting affected areas.

Result: Evaluation on 12 patients showed effectiveness with 94.8% AUC and 94.6% specificity, demonstrating strong diagnostic performance.

Conclusion: The framework successfully automates osteosarcoma diagnosis from CT scans, providing accurate classification and 3D visualization to assist medical professionals.

Abstract: Osteosarcoma is the most common primary bone cancer, mainly affecting the youngest and oldest populations. Its detection at early stages is crucial to reduce the probability of developing bone metastasis. In this context, accurate and fast diagnosis is essential to help physicians during the prognosis process. The research goal is to automate the diagnosis of osteosarcoma through a pipeline that includes the preprocessing, detection, postprocessing, and visualization of computed tomography (CT) scans. Thus, this paper presents a machine learning and visualization framework for classifying CT scans using different convolutional neural network (CNN) models. Preprocessing includes data augmentation and identification of the region of interest in scans. Post-processing includes data visualization to render a 3D bone model that highlights the affected area. An evaluation on 12 patients revealed the effectiveness of our framework, obtaining an area under the curve (AUC) of 94.8% and a specificity of 94.6%.

Method: Proposes MSEG-VCUQ, a hybrid framework integrating U-Net CNNs with transformer-based Segment Anything Model (SAM) for enhanced segmentation accuracy and cross-modality generalization, with systematic uncertainty quantification and introduction of open-source multimodal HSV PD datasets.

Result: Empirical results show MSEG-VCUQ outperforms baseline CNNs and vision foundation models, enabling scalable and reliable phase detection segmentation for real-world boiling dynamics.

Conclusion: The proposed hybrid framework successfully addresses key challenges in high-speed video phase detection segmentation, providing enhanced accuracy, uncertainty quantification, and cross-modality generalization for industrial boiling dynamics monitoring.

Abstract: High-speed video (HSV) phase detection (PD) segmentation is crucial for monitoring vapor, liquid, and microlayer phases in industrial processes. While CNN-based models like U-Net have shown success in simplified shadowgraphy-based two-phase flow (TPF) analysis, their application to complex HSV PD tasks remains unexplored, and vision foundation models (VFMs) have yet to address the complexities of either shadowgraphy-based or PD TPF video segmentation. Existing uncertainty quantification (UQ) methods lack pixel-level reliability for critical metrics like contact line density and dry area fraction, and the absence of large-scale, multimodal experimental datasets tailored to PD segmentation further impedes progress. To address these gaps, we propose MSEG-VCUQ. This hybrid framework integrates U-Net CNNs with the transformer-based Segment Anything Model (SAM) to achieve enhanced segmentation accuracy and cross-modality generalization. Our approach incorporates systematic UQ for robust error assessment and introduces the first open-source multimodal HSV PD datasets. Empirical results demonstrate that MSEG-VCUQ outperforms baseline CNNs and VFMs, enabling scalable and reliable PD segmentation for real-world boiling dynamics.

Method: Systematic study of representation learning for microscopy images using curated benchmarks with simple baselines including untrained models and structural representations of cellular tissue.

Result: State-of-the-art methods perform comparably to simple baselines, fail to consistently acquire high-level biologically meaningful features, and commonly used benchmark metrics mask these limitations.

Conclusion: Progress in microscopy image representation learning requires both stronger models and more diagnostic benchmarks that measure what is actually learned, with detailed comparisons providing insights for improvement.

Abstract: Representation learning has driven major advances in natural image analysis by enabling models to acquire high-level semantic features. In microscopy imaging, however, it remains unclear what current representation learning methods actually learn. In this work, we conduct a systematic study of representation learning for the two most widely used and broadly available microscopy data types, representing critical scales in biology: cell culture and tissue imaging. To this end, we introduce a set of simple yet revealing baselines on curated benchmarks, including untrained models and simple structural representations of cellular tissue. Our results show that, surprisingly, state-of-the-art methods perform comparably to these baselines. We further show that, in contrast to natural images, existing models fail to consistently acquire high-level, biologically meaningful features. Moreover, we demonstrate that commonly used benchmark metrics are insufficient to assess representation quality and often mask this limitation. In addition, we investigate how detailed comparisons with these benchmarks provide ways to interpret the strengths and weaknesses of models for further improvements. Together, our results suggest that progress in microscopy image representation learning requires not only stronger models, but also more diagnostic benchmarks that measure what is actually learned.

Method: Proposes DL-JEC using single-point cloud representation where event polarity is treated as an attribute rather than separate point clouds. Introduces adaptive voxel binarization strategies optimized for either quality-oriented or computer vision task-oriented purposes to maximize performance for specific tasks.

Result: DL-JEC achieves significant compression performance gains compared to conventional and DL-based state-of-the-art event data coding solutions, including MPEG G-PCC and JPEG Pleno PCC standards. Shows lossy coding with reduced rates doesn’t compromise event classification performance.

Conclusion: The paper demonstrates that lossy event data coding is viable and can achieve substantial compression without degrading computer vision task performance, challenging the current paradigm of requiring lossless coding for event data.

Abstract: Neuromorphic vision sensors, commonly referred to as event cameras, generate a massive number of pixel-level events, composed by spatiotemporal and polarity information, thus demanding highly efficient coding solutions. Existing solutions focus on lossless coding of event data, assuming that no distortion is acceptable for the target use cases, mostly including computer vision tasks such as classification and recognition. One promising coding approach exploits the similarity between event data and point clouds, both being sets of 3D points, thus allowing to use current point cloud coding solutions to code event data, typically adopting a two-point clouds representation, one for each event polarity. This paper proposes a novel lossy Deep Learning-based Joint Event data Coding (DL-JEC) solution, which adopts for the first time a single-point cloud representation, where the event polarity plays the role of a point cloud attribute, thus enabling to exploit the correlation between the geometry/spatiotemporal and polarity event information. Moreover, this paper also proposes novel adaptive voxel binarization strategies which may be used in DL-JEC, optimized for either quality-oriented or computer vision task-oriented purposes which allow to maximize the performance for the task at hand. DL-JEC can achieve significant compression performance gains when compared with relevant conventional and DL-based state-of-the-art event data coding solutions, notably the MPEG G-PCC and JPEG Pleno PCC standards. Furthermore, it is shown that it is possible to use lossy event data coding, with significantly reduced rate regarding lossless coding, without compromising the target computer vision task performance, notably event classification, thus changing the current event data coding paradigm.

Method: DINOv3-based framework that predicts intima-media band at fixed resolution, extracts upper/lower boundaries column-wise, corrects for image resizing using calibration factors, and reports CIMT in physical units with test-time threshold calibration.

Result: Achieved mean test Dice of 0.7739 ± 0.0037, IoU of 0.6384 ± 0.0044, mean CIMT absolute error of 181.16 ± 11.57 μm, with Pearson correlation of 0.480 ± 0.259. Test-time calibration reduced error from 141.0 μm to 101.1 μm.

Conclusion: DINOv3-based approach achieves clinically relevant measurement accuracy (~0.1 mm error), supporting feasibility of vision foundation models for interpretable, calibration-aware CIMT measurement in medical imaging.

Abstract: Carotid intima-media thickness (CIMT) measured from B-mode ultrasound is an established vascular biomarker for atherosclerosis and cardiovascular risk stratification. Although a wide range of computerized methods have been proposed for carotid boundary delineation and CIMT estimation, robust and transferable deep models that jointly address segmentation and measurement remain underexplored, particularly in the era of vision foundation models. Motivated by recent advances in adapting DINOv3 to medical segmentation and exploiting DINOv3 in test-time optimization pipelines, we investigate a DINOv3-based framework for carotid intima-media complex segmentation and subsequent CIMT measurement on the Carotid Ultrasound Boundary Study (CUBS) v1 dataset. Our pipeline predicts the intima-media band at a fixed image resolution, extracts upper and lower boundaries column-wise, corrects for image resizing using the per-image calibration factor provided by CUBS, and reports CIMT in physical units. Across three patient-level test splits, our method achieved a mean test Dice of 0.7739 $\pm$ 0.0037 and IoU of 0.6384 $\pm$ 0.0044. The mean CIMT absolute error was 181.16 $\pm$ 11.57 $μ$m, with a mean Pearson correlation of 0.480 $\pm$ 0.259. In a held-out validation subset ($n=28$), test-time threshold calibration reduced the mean absolute CIMT error from 141.0 $μ$m at the default threshold to 101.1 $μ$m at the measurement-optimized threshold, while simultaneously reducing systematic bias toward zero. Relative to the error ranges reported in the original CUBS benchmark for classical computerized methods, these results place a DINOv3-based approach within the clinically relevant $\sim$0.1 mm measurement regime. Together, our findings support the feasibility of using vision foundation models for interpretable, calibration-aware CIMT measurement.

Method: Proposes ABC-GS based on 3D Gaussian Splatting with: 1) controllable matching stage using segmentation masks for precise content-style alignment, 2) style transfer loss based on feature alignment for global style accuracy, and 3) depth loss and Gaussian regularization to preserve original scene geometry.

Result: Extensive experiments show ABC-GS provides better controllability of style transfer and achieves stylization results more faithfully aligned with the global style of artistic references compared to existing methods.

Conclusion: ABC-GS advances 3D scene stylization by addressing limitations of NeRF-based approaches through 3D Gaussian Splatting, offering improved global style alignment and fine-grained control while preserving scene geometry.

Abstract: 3D scene stylization approaches based on Neural Radiance Fields (NeRF) achieve promising results by optimizing with Nearest Neighbor Feature Matching (NNFM) loss. However, NNFM loss does not consider global style information. In addition, the implicit representation of NeRF limits their fine-grained control over the resulting scenes. In this paper, we introduce ABC-GS, a novel framework based on 3D Gaussian Splatting to achieve high-quality 3D style transfer. To this end, a controllable matching stage is designed to achieve precise alignment between scene content and style features through segmentation masks. Moreover, a style transfer loss function based on feature alignment is proposed to ensure that the outcomes of style transfer accurately reflect the global style of the reference image. Furthermore, the original geometric information of the scene is preserved with the depth loss and Gaussian regularization terms. Extensive experiments show that our ABC-GS provides controllability of style transfer and achieves stylization results that are more faithfully aligned with the global style of the chosen artistic reference. Our homepage is available at https://vpx-ecnu.github.io/ABC-GS-website.

Method: 1. Use multi-view images from commodity cameras processed by a frozen Vision-Language Model to extract dense semantic embeddings. 2. Translate these embeddings into initial estimates of permittivity and conductivity for scene surfaces. 3. Initialize a Sionna-based differentiable ray tracer with these priors. 4. Calibrate material parameters via gradient descent using only a few dozen sparse channel soundings. 5. Retain the association between vision features and material parameters for fast transfer to new scenarios.

Result: Evaluations across three real-world scenarios (office interiors, urban canyons, dynamic public spaces) show: 1. 10× reduction in channel measurement needs. 2. 59% lower median delay spread error compared to pure data-driven deep learning methods.

Conclusion: VisRFTwin provides a scalable, data-efficient framework for mmWave propagation modeling by integrating vision-derived material priors with differentiable ray tracing, enabling accurate digital twins with minimal calibration data.

Abstract: Accurately modeling millimeter-wave (mmWave) propagation is essential for real-time AR and autonomous systems. Differentiable ray tracing offers a physics-grounded solution but still facing deployment challenges due to its over-reliance on exhaustive channel measurements or brittle, hand-tuned scene models for material properties. We present VisRFTwin, a scalable and data-efficient digital-twin framework that integrates vision-derived material priors with differentiable ray tracing. Multi-view images from commodity cameras are processed by a frozen Vision-Language Model to extract dense semantic embeddings, which are translated into initial estimates of permittivity and conductivity for scene surfaces. These priors initialize a Sionna-based differentiable ray tracer, which rapidly calibrates material parameters via gradient descent with only a few dozen sparse channel soundings. Once calibrated, the association between vision features and material parameters is retained, enabling fast transfer to new scenarios without repeated calibration. Evaluations across three real-world scenarios, including office interiors, urban canyons, and dynamic public spaces show that VisRFTwin reduces channel measurement needs by up to 10$\times$ while achieving a 59% lower median delay spread error than pure data-driven deep learning methods.

Method: Proposes HQUIC framework with: 1) Adaptive Lighting and Tone Correction (ALTC) module to predict attenuation coefficients and global light information, 2) Dynamic weighting of multi-scale frequency components to prioritize distortion-critical information, and 3) Tone adjustment loss to balance color channel discrepancies.

Result: Comprehensive evaluations on diverse underwater datasets show HQUIC outperforms state-of-the-art compression methods, demonstrating superior compression performance for underwater images.

Conclusion: HQUIC effectively handles unique underwater illumination conditions and color shifts through specialized modules, achieving better compression performance than existing methods by addressing water-specific optical challenges.

Abstract: With the increasing exploration and exploitation of the underwater world, underwater images have become a critical medium for human interaction with marine environments, driving extensive research into their efficient transmission and storage. However, contemporary underwater image compression algorithms fail to adequately address the impact of water refraction and scattering on light waves, which not only elevate training complexity but also result in suboptimal compression performance. To tackle this limitation, we propose High Quality Underwater Image Compression (HQUIC), a novel framework designed to handle the unique illumination conditions and color shifts inherent in underwater images, thereby achieving superior compression performance. HQUIC first incorporates an Adaptive Lighting and Tone Correction (ALTC) module to adaptively predict the attenuation coefficients and global light information of images, effectively alleviating issues stemming from variations in illumination and tone across underwater images. Secondly, it dynamically weights multi-scale frequency components, prioritizing information critical to distortion quality while discarding redundant details. Furthermore, we introduce a tone adjustment loss to enable the model to better balance discrepancies among different color channels. Comprehensive evaluations on diverse underwater datasets validate that HQUIC outperforms state-of-the-art compression methods, demonstrating its effectiveness.

Method: Developed VisualLeakBench evaluation suite with 1,000 synthetically generated adversarial images containing 8 PII types, validated on 50 real-world screenshots. Evaluated four frontier LVLMs (GPT-5.2, Claude~4, Gemini-3 Flash, Grok-4) with Wilson 95% confidence intervals and tested defensive system prompts.

Result: Claude4 had lowest OCR ASR (14.2%) but highest PII ASR (74.4%) with comply-then-warn pattern. Grok-4 had lowest PII ASR (20.4%). Defensive prompts eliminated PII leakage for two models, reduced Claude4’s leakage from 74.4% to 2.2%, but had no effect on Gemini-3 Flash on synthetic data. Real-world validation showed Gemini-3 Flash did respond to mitigation (50% to 0%), indicating mitigation robustness is template-sensitive.

Conclusion: LVLMs show significant vulnerabilities to PII leakage attacks, with varying robustness across models and attack types. Mitigation effectiveness depends on both model architecture and attack template characteristics, highlighting the need for comprehensive evaluation frameworks like VisualLeakBench for deployment-relevant safety assessment.

Abstract: As Large Vision-Language Models (LVLMs) are increasingly deployed in agent-integrated workflows and other deployment-relevant settings, their robustness against semantic visual attacks remains under-evaluated – alignment is typically tested on explicit harmful content rather than privacy-critical multimodal scenarios. We introduce VisualLeakBench, an evaluation suite to audit LVLMs against OCR Injection and Contextual PII Leakage using 1,000 synthetically generated adversarial images with 8 PII types, validated on 50 in-the-wild (IRL) real-world screenshots spanning diverse visual contexts. We evaluate four frontier systems (GPT-5.2, Claude4, Gemini-3 Flash, Grok-4) with Wilson 95% confidence intervals. Claude4 achieves the lowest OCR ASR (14.2%) but the highest PII ASR (74.4%), exhibiting a comply-then-warn pattern – where verbatim data disclosure precedes any safety-oriented language. Grok-4 achieves the lowest PII ASR (20.4%). A defensive system prompt eliminates PII leakage for two models, reduces Claude~4’s leakage from 74.4% to 2.2%, but has no effect on Gemini-3 Flash on synthetic data. Strikingly, IRL validation reveals Gemini-3 Flash does respond to mitigation on real-world images (50% to 0%), indicating that mitigation robustness is template-sensitive rather than uniformly absent. We release our dataset and code for reproducible robustness and safety evaluation of deployment-relevant vision-language systems.

Method: WaRA reshapes patch tokens into spatial grid, applies fixed discrete wavelet transform, updates subband coefficients using shared low-rank adapter, then reconstructs additive update via inverse wavelet transform. Tiny-WaRA variant learns only small set of coefficients in fixed basis from pretrained weights via truncated SVD.

Result: Experiments on medical image classification across four modalities and datasets show WaRA consistently improves performance over strong PEFT baselines while maintaining favorable efficiency profile.

Conclusion: WaRA provides effective parameter-efficient fine-tuning for medical imaging by leveraging wavelet domain adaptation to capture both coarse structure and fine detail with compact trainable interface.

Abstract: Adapting large pretrained vision models to medical image classification is often limited by memory, computation, and task-specific specializations. Parameter-efficient fine-tuning (PEFT) methods like LoRA reduce this cost by learning low-rank updates, but operating directly in feature space can struggle to capture the localized, multi-scale features common in medical imaging. We propose WaRA, a wavelet-structured adaptation module that performs low-rank adaptation in a wavelet domain. WaRA reshapes patch tokens into a spatial grid, applies a fixed discrete wavelet transform, updates subband coefficients using a shared low-rank adapter, and reconstructs the additive update through an inverse wavelet transform. This design provides a compact trainable interface while biasing the update toward both coarse structure and fine detail. For extremely low-resource settings, we introduce Tiny-WaRA, which further reduces trainable parameters by learning only a small set of coefficients in a fixed basis derived from the pretrained weights through a truncated SVD. Experiments on medical image classification across four modalities and datasets demonstrate that WaRA consistently improves performance over strong PEFT baselines, while retaining a favorable efficiency profile. Our code is publicly available at~\href{https://github.com/moeinheidari7829/WaRA}{\textcolor{magenta}{GitHub}}.

Method: 1) Developed a scalable multi-agent LVLM annotation framework for efficient construction of fine-grained clinically-aligned supervision. 2) Proposed In-Context Diffusion Transformer (IC-DiT) that incorporates spatial layouts, textual descriptions, and visual embeddings into a unified diffusion transformer using hierarchical multimodal attention.

Result: IC-DiT achieves higher fidelity, stronger spatial controllability, and better diagnostic consistency than existing methods across five histopathology datasets. Generated images serve as effective data augmentation resources for downstream tasks like cancer classification and survival analysis.

Conclusion: The proposed framework enables controllable pathology image synthesis with fine-grained structural constraints, addressing limitations of existing text-guided diffusion models through scalable annotation and layout-aware generation.

Abstract: Controllable pathology image synthesis requires reliable regulation of spatial layout, tissue morphology, and semantic detail. However, existing text-guided diffusion models offer only coarse global control and lack the ability to enforce fine-grained structural constraints. Progress is further limited by the absence of large datasets that pair patch-level spatial layouts with detailed diagnostic descriptions, since generating such annotations for gigapixel whole-slide images is prohibitively time-consuming for human experts. To overcome these challenges, we first develop a scalable multi-agent LVLM annotation framework that integrates image description, diagnostic step extraction, and automatic quality judgment into a coordinated pipeline, and we evaluate the reliability of the system through a human verification process. This framework enables efficient construction of fine-grained and clinically aligned supervision at scale. Building on the curated data, we propose In-Context Diffusion Transformer (IC-DiT), a layout-aware generative model that incorporates spatial layouts, textual descriptions, and visual embeddings into a unified diffusion transformer. Through hierarchical multimodal attention, IC-DiT maintains global semantic coherence while accurately preserving structural and morphological details. Extensive experiments on five histopathology datasets show that IC-DiT achieves higher fidelity, stronger spatial controllability, and better diagnostic consistency than existing methods. In addition, the generated images serve as effective data augmentation resources for downstream tasks such as cancer classification and survival analysis.

Method: Uses cylindrical mechanical projector with rotational stage and cylindrical pattern generator with ON/OFF slots at two intervals to project omnidirectional multi-frequency phase-shifted fringe patterns. Applies multi-wavelength unwrapping algorithm and quasi-calibration technique for single-camera 3D reconstruction.

Result: Achieves high-accuracy 3D reconstruction with expanded uncertainty of 0.215 mm for reconstructed depth. Experimental results show reliable measurement performance and practical feasibility for omnidirectional 3D reconstruction.

Conclusion: The proposed cylindrical mechanical projector system successfully addresses limitations of conventional methods and provides practical solution for wide-area 3D sensing with single-camera setup.

Abstract: The demand for 360-degree 3D reconstruction has significantly increased in recent years across various domains such as the metaverse and 3D telecommunication. Accordingly, the importance of precise and wide-area 3D sensing technology has become emphasized. While the digital fringe projection method has been widely used due to its high accuracy and implementation flexibility, it suffers from fundamental limitations such as unidirectional projection and a restricted available light spectrum. To address these issues, this paper proposes a novel 3D reconstruction method based on a cylindrical mechanical projector. The proposed method consists of a rotational stage and a cylindrical pattern generator with ON/OFF slots at two distinct intervals, enabling omnidirectional projection of multi-frequency phase-shifted fringe patterns. By applying a multi-wavelength unwrapping algorithm and a quasi-calibration technique, the system achieves high-accuracy 3D reconstruction using only a single camera. Experimental results, supported by repeatability and reproducibility analyses together with a measurement uncertainty evaluation, confirm reliable measurement performance and practical feasibility for omnidirectional 3D reconstruction. The expanded uncertainty of the reconstructed depth was evaluated as 0.215 mm.

Method: VeloEdit dynamically identifies editing regions by quantifying discrepancies between velocity fields for source preservation and desired edits. It enforces consistency in preservation regions by substituting editing velocity with source-restoring velocity, and enables continuous modulation of edit intensity in target regions via velocity interpolation.

Result: Extensive experiments on Flux.1 Kontext and Qwen-Image-Edit demonstrate improved visual consistency and editing continuity with negligible additional computational cost.

Conclusion: VeloEdit provides a training-free solution for consistent and controllable instruction-based image editing by operating directly on velocity fields, outperforming prior methods that rely on complex attention manipulation or auxiliary trainable modules.

Abstract: Instruction-based image editing aims to modify source content according to textual instructions. However, existing methods built upon flow matching often struggle to maintain consistency in non-edited regions due to denoising-induced reconstruction errors that cause drift in preserved content. Moreover, they typically lack fine-grained control over edit strength. To address these limitations, we propose VeloEdit, a training-free method that enables highly consistent and continuously controllable editing. VeloEdit dynamically identifies editing regions by quantifying the discrepancy between the velocity fields responsible for preserving source content and those driving the desired edits. Based on this partition, we enforce consistency in preservation regions by substituting the editing velocity with the source-restoring velocity, while enabling continuous modulation of edit intensity in target regions via velocity interpolation. Unlike prior works that rely on complex attention manipulation or auxiliary trainable modules, VeloEdit operates directly on the velocity fields. Extensive experiments on Flux.1 Kontext and Qwen-Image-Edit demonstrate that VeloEdit improves visual consistency and editing continuity with negligible additional computational cost. Code is available at https://github.com/xmulzq/VeloEdit.

Method: Proposes a diffusion-based decoding framework with: 1) Logit-to-Code Distributional Mapping that converts VLM image-token logits into continuous distribution-weighted code vectors with uncertainty features; 2) Lightweight Logit Calibration to align training-time proxy logits with VLM-generated logits; 3) Distribution-Conditioned Diffusion Decoder that generates high-fidelity images conditioned on these representations.

Result: The method consistently improves visual fidelity for both VQ-VAE reconstructions and text-to-image generations from VLM-predicted tokens, achieved through short training on ImageNet-1K only.

Conclusion: The approach provides an efficient way to enhance image fidelity of pre-trained VLMs without modifying the original models, requiring minimal training data and computational resources compared to full model retraining.

Abstract: Recent large-scale vision-language models (VLMs) have shown remarkable text-to-image generation capabilities, yet their visual fidelity remains constrained by the discrete image tokenization, which poses a major challenge. Although several studies have explored continuous representation modeling to enhance visual quality, adapting pre-trained VLM models to such representations requires large-scale data and training costs comparable to the original pre-training. To circumvent this limitation, we propose a diffusion-based decoding framework that enhances image fidelity by training only a diffusion decoder on the output image-token logits of pre-trained VLMs, thereby preserving the original model intact. At its core, Logit-to-Code Distributional Mapping converts the VLM’s image-token logits into continuous, distribution-weighted code vectors with uncertainty features, providing an effective conditioning signal for diffusion decoding. A lightweight Logit Calibration aligns training-time proxy logits from the VQ-VAE encoder with VLM-generated logits, mitigating the train-inference gap. Conditioned on these representations, the Distribution-Conditioned Diffusion Decoder generates high-fidelity images. Achieved solely through short training on ImageNet-1K, our method consistently improves visual fidelity for both VQ-VAE reconstructions and text-to-image generations from VLM-predicted tokens.

Method: Introduces WebVR benchmark with 175 webpages across diverse categories created via controlled synthesis pipeline (not web crawling). Designs fine-grained visual rubric for evaluation across multiple dimensions. Tests 19 models on video-to-webpage generation task.

Result: Experiments reveal substantial gaps in models’ ability to recreate fine-grained style and motion quality. The rubric-based automatic evaluation achieves 96% agreement with human preferences.

Conclusion: WebVR fills the gap for video-conditioned webpage generation evaluation, showing current MLLMs struggle with fine-grained visual and motion details. The benchmark, toolkit, and baselines are released to support future research.

Abstract: Existing web-generation benchmarks rely on text prompts or static screenshots as input. However, videos naturally convey richer signals such as interaction flow, transition timing, and motion continuity, which are essential for faithful webpage recreation. Despite this potential, video-conditioned webpage generation remains largely unexplored, with no dedicated benchmark for this task. To fill this gap, we introduce WebVR, a benchmark that evaluates whether MLLMs can faithfully recreate webpages from demonstration videos. WebVR contains 175 webpages across diverse categories, all constructed through a controlled synthesis pipeline rather than web crawling, ensuring varied and realistic demonstrations without overlap with existing online pages. We also design a fine-grained, human-aligned visual rubric that evaluates the generated webpages across multiple dimensions. Experiments on 19 models reveal substantial gaps in recreating fine-grained style and motion quality, while the rubric-based automatic evaluation achieves 96% agreement with human preferences. We release the dataset, evaluation toolkit, and baseline results to support future research on video-to-webpage generation.

Method: Constructed balanced dataset of 18,080 chest X-rays across 5 disease categories, compared 7 architectures (ConvNeXt-Tiny, DenseNet121, DenseNet201, ResNet50, ViT-B/16, EfficientNetV2-M, MobileNetV2) under identical training conditions with ImageNet-pretrained weights, patient-level data splitting to prevent leakage.

Result: All models exceeded 90% accuracy; ConvNeXt-Tiny achieved highest performance (92.31% accuracy, 95.70% AUROC); MobileNetV2 was most parameter-efficient (3.5M params, 90.42% accuracy, 48 min training); Tuberculosis and COVID-19 classification near-perfect (AUROC ≥99.97%); Grad-CAM showed clinically consistent attention patterns.

Conclusion: High-accuracy multi-disease chest X-ray classification is achievable without excessive computational resources, with implications for AI-assisted diagnosis in diverse healthcare settings.

Abstract: Chest X-ray imaging remains the primary diagnostic tool for pulmonary and cardiac disorders worldwide, yet its accuracy is hampered by radiologist shortages and inter-observer variability. This study presents a systematic comparative evaluation of seven deep learning architectures for multi-class chest disease classification: ConvNeXt-Tiny, DenseNet121, DenseNet201, ResNet50, ViT-B/16, EfficientNetV2-M, and MobileNetV2. A balanced dataset of 18,080 chest X-ray images spanning five disease categories (Cardiomegaly, COVID-19, Normal, Pneumonia, and Tuberculosis) was constructed from three public repositories and partitioned at the patient level to prevent data leakage. All models were trained under identical conditions using ImageNet-pretrained weights, standardized preprocessing, and consistent hyperparameters. All seven architectures exceeded 90% test accuracy. ConvNeXt-Tiny achieved the highest performance (92.31% accuracy, 95.70% AUROC), while MobileNetV2 emerged as the most parameter-efficient model (3.5M parameters, 90.42% accuracy, 94.10% AUROC), completing training in 48 minutes. Tuberculosis and COVID-19 classification was near-perfect (AUROC >= 99.97%) across all architectures, while Normal, Cardiomegaly, and Pneumonia presented greater challenges due to overlapping radiographic features. Grad-CAM visualizations confirmed clinically consistent attention patterns across disease categories. These findings demonstrate that high-accuracy multi-disease chest X-ray classification is achievable without excessive computational resources, with important implications for AI-assisted diagnosis in both resource-rich and resource-constrained healthcare settings.

Method: Combines Grounding DINO and Segment Anything Model 2 (SAM2) foundation models fine-tuned for microbiology domain, creating a pipeline for zero-shot detection and segmentation without additional training.

Result: Achieved 93.1% mean Average Precision and Dice@detection score of 0.85 on out-of-distribution datasets, demonstrating excellent detection and segmentation capabilities.

Conclusion: The pipeline provides robust zero-shot bacterial colony analysis, addresses annotation challenges in microbiology, and is shared as open access for community use.

Abstract: The detection and classification of bacterial colonies in images of agar-plates is important in microbiology, but is hindered by the lack of labeled datasets. Therefore, we propose Colony Grounded SAM2, a zero-shot inference pipeline to detect and segment bacterial colonies in multiple settings without any further training. By utilizing the pre-trained foundation models Grounding DINO and Segment Anything Model 2, fine-tuned to the microbiological domain, we developed a model that is robust to data changes. Results showed a mean Average Precision of 93.1% and a $Dice@detection$ score of 0.85, showing excellent detection and segmentation capabilities on out-of-distribution datasets. The entire pipeline with model weights are shared open access to aid with annotation- and classification purposes in microbiology.

Method: Formulates visual token pruning as sequential decision process with explicit state transitions. Uses self-supervised autoencoder to compress visual tokens into compact state representation. Initializes pruning policy via learning from demonstrations, then fine-tunes with Proximal Policy Optimization (PPO) to jointly optimize task accuracy and computational efficiency.

Result: Removes up to 66.7% of visual tokens, achieves up to 54.2% reduction in FLOPs during inference, while maintaining near-lossless average accuracy drop of only 0.7%.

Conclusion: TPRL effectively reduces computational costs of LVLMs through learned adaptive pruning trajectories while preserving task performance, demonstrating the value of reinforcement learning for optimizing multimodal model efficiency.

Abstract: Large Vision-Language Models (LVLMs) incur substantial inference costs due to the processing of a vast number of visual tokens. Existing methods typically struggle to model progressive visual token reduction as a multi-step decision process with sequential dependencies and often rely on hand-engineered scoring rules that lack adaptive optimization for complex reasoning trajectories. To overcome these limitations, we propose TPRL, a reinforcement learning framework that learns adaptive pruning trajectories through language-guided sequential optimization tied directly to end-task performance. We formulate visual token pruning as a sequential decision process with explicit state transitions and employ a self-supervised autoencoder to compress visual tokens into a compact state representation for efficient policy learning. The pruning policy is initialized through learning from demonstrations and subsequently fine-tuned using Proximal Policy Optimization (PPO) to jointly optimize task accuracy and computational efficiency. Our experimental results demonstrate that TPRL removes up to 66.7% of visual tokens and achieves up to a 54.2% reduction in FLOPs during inference while maintaining a near-lossless average accuracy drop of only 0.7%. Code is released at \href{https://github.com/MagicVicCoder/TPRL}{\textcolor{mypink}{https://github.com/MagicVicCoder/TPRL}}.

Method: COT-FM clusters target samples and assigns each cluster a dedicated source distribution obtained by reversing pretrained FM models. This divide-and-conquer strategy yields more accurate local transport and significantly straighter vector fields without changing model architecture.

Result: COT-FM consistently accelerates sampling and improves generation quality across 2D datasets, image generation benchmarks, and robotic manipulation tasks as a plug-and-play approach.

Conclusion: COT-FM provides an effective framework for improving Flow Matching by optimizing probability paths through clustering and dedicated source distributions, leading to faster and higher-quality generation across multiple domains.

Abstract: We introduce COT-FM, a general framework that reshapes the probability path in Flow Matching (FM) to achieve faster and more reliable generation. FM models often produce curved trajectories due to random or batchwise couplings, which increase discretization error and reduce sample quality. COT-FM fixes this by clustering target samples and assigning each cluster a dedicated source distribution obtained by reversing pretrained FM models. This divide-and-conquer strategy yields more accurate local transport and significantly straighter vector fields, all without changing the model architecture. As a plug-and-play approach, COT-FM consistently accelerates sampling and improves generation quality across 2D datasets, image generation benchmarks, and robotic manipulation tasks.

Method: SERUM adds a unique watermark noise to the initial diffusion generation noise and trains a lightweight detector to identify watermarked images. It uses a decoupled architecture that supports multiple users with individualized watermarks.

Result: SERUM achieves the highest true positive rate at 1% false positive rate in most scenarios, provides robustness against image augmentations and removal attacks, and maintains fast injection/detection with low training overhead and negligible quality impact.

Conclusion: SERUM offers a practical, efficient, and robust solution for watermarking diffusion model outputs, enabling reliable distinction between generated and natural images while supporting multiple users with minimal interference.

Abstract: We propose SERUM: an intriguingly simple yet highly effective method for marking images generated by diffusion models (DMs). We only add a unique watermark noise to the initial diffusion generation noise and train a lightweight detector to identify watermarked images, simplifying and unifying the strengths of prior approaches. SERUM provides robustness against any image augmentations or watermark removal attacks and is extremely efficient, all while maintaining negligible impact on image quality. In contrast to prior approaches, which are often only resilient to limited perturbations and incur significant training, injection, and detection costs, our SERUM achieves remarkable performance, with the highest true positive rate (TPR) at a 1% false positive rate (FPR) in most scenarios, along with fast injection and detection and low detector training overhead. Its decoupled architecture also seamlessly supports multiple users by embedding individualized watermarks with little interference between the marks. Overall, our method provides a practical solution to mark outputs from DMs and to reliably distinguish generated from natural images.

Method: Introduces TennisVL benchmark (200+ matches, 471.9 hours, 40,000+ rally clips) with expert analytical commentary, and TennisExpert framework integrating video semantic parser with memory-augmented model based on Qwen3-VL-8B.

Result: TennisExpert consistently outperforms strong proprietary baselines (GPT-5, Gemini, Claude) and demonstrates improved ability to capture tactical context and match dynamics.

Conclusion: The paper addresses key challenges in automatic tennis understanding through a comprehensive benchmark and effective multimodal framework that excels at tactical analysis.

Abstract: Tennis is one of the most widely followed sports, generating extensive broadcast footage with strong potential for professional analysis, automated coaching, and real-time commentary. However, automatic tennis understanding remains underexplored due to two key challenges: (1) the lack of large-scale benchmarks with fine-grained annotations and expert-level commentary, and (2) the difficulty of building accurate yet efficient multimodal systems suitable for real-time deployment. To address these challenges, we introduce TennisVL, a large-scale tennis benchmark comprising over 200 professional matches (471.9 hours) and 40,000+ rally-level clips. Unlike existing commentary datasets that focus on descriptive play-by-play narration, TennisVL emphasizes expert analytical commentary capturing tactical reasoning, player decisions, and match momentum. Furthermore, we propose TennisExpert, a multimodal tennis understanding framework that integrates a video semantic parser with a memory-augmented model built on Qwen3-VL-8B. The parser extracts key match elements (e.g., scores, shot sequences, ball bounces, and player locations), while hierarchical memory modules capture both short- and long-term temporal context. Experiments show that TennisExpert consistently outperforms strong proprietary baselines, including GPT-5, Gemini, and Claude, and demonstrates improved ability to capture tactical context and match dynamics.

Method: Proposes a 4B parameter vision-language model with Layout-as-Thought mechanism - an optional thinking phase triggered by special tokens that generates structured layout representations (bounding boxes, element types, reading order) before final output, enabling layout grounding capabilities.

Result: Achieves state-of-the-art performance: ranks first on OmniDocBench v1.5 (93.12) and OlmOCR Bench (79.8), competitive results on OCRBench, CCOCR, DocVQA, and ChartQA, and highest average score on key information extraction benchmarks, surpassing Gemini-3.1-Pro, Seed-2.0, and Qwen3-VL-235B.

Conclusion: Qianfan-OCR demonstrates that unified vision-language models with explicit layout reasoning capabilities can achieve superior performance on diverse document understanding tasks while maintaining end-to-end efficiency.

Abstract: We present Qianfan-OCR, a 4B-parameter end-to-end vision-language model that unifies document parsing, layout analysis, and document understanding within a single architecture. It performs direct image-to-Markdown conversion and supports diverse prompt-driven tasks including table extraction, chart understanding, document QA, and key information extraction. To address the loss of explicit layout analysis in end-to-end OCR, we propose Layout-as-Thought, an optional thinking phase triggered by special think tokens that generates structured layout representations – bounding boxes, element types, and reading order – before producing final outputs, recovering layout grounding capabilities while improving accuracy on complex layouts. Qianfan-OCR ranks first among end-to-end models on OmniDocBench v1.5 (93.12) and OlmOCR Bench (79.8), achieves competitive results on OCRBench, CCOCR, DocVQA, and ChartQA against general VLMs of comparable scale, and attains the highest average score on public key information extraction benchmarks, surpassing Gemini-3.1-Pro, Seed-2.0, and Qwen3-VL-235B. The model is publicly accessible via the Baidu AI Cloud Qianfan platform.

Method: Proposes FlowAD framework with three components: 1) ego-guided scene partition constructs flow units to quantify scene flow, 2) spatial and temporal flow predictions model dynamics of scene flow, 3) task-aware enhancement exploits learned dynamics for various tasks. Uses Frames before Correct Planning (FCP) metric for evaluation.

Result: Reduces 19% collision rate over SparseDrive with FCP improvements of 1.39 frames (60%) on nuScenes, achieves driving score of 51.77 on Bench2Drive. Demonstrates effectiveness across perception, end-to-end planning, and VLM analysis tasks.

Conclusion: FlowAD’s ego-scene interactive modeling paradigm effectively captures ego motion feedback, improving environment understanding and planning capabilities in autonomous driving systems.

Abstract: Effective environment modeling is the foundation for autonomous driving, underpinning tasks from perception to planning. However, current paradigms often inadequately consider the feedback of ego motion to the observation, which leads to an incomplete understanding of the driving process and consequently limits the planning capability. To address this issue, we introduce a novel ego-scene interactive modeling paradigm. Inspired by human recognition, the paradigm represents ego-scene interaction as the scene flow relative to the ego-vehicle. This conceptualization allows for modeling ego-motion feedback within a feature learning pattern, advantageously utilizing existing log-replay datasets rather than relying on scenario simulations. We specifically propose FlowAD, a general flow-based framework for autonomous driving. Within it, an ego-guided scene partition first constructs basic flow units to quantify scene flow. The ego-vehicle’s forward direction and steering velocity directly shape the partition, which reflects ego motion. Then, based on flow units, spatial and temporal flow predictions are performed to model dynamics of scene flow, encompassing both spatial displacement and temporal variation. The final task-aware enhancement exploits learned spatio-temporal flow dynamics to benefit diverse tasks through object and region-level strategies. We also propose a novel Frames before Correct Planning (FCP) metric to assess the scene understanding capability. Experiments in both open and closed-loop evaluations demonstrate FlowAD’s generality and effectiveness across perception, end-to-end planning, and VLM analysis. Notably, FlowAD reduces 19% collision rate over SparseDrive with FCP improvements of 1.39 frames (60%) on nuScenes, and achieves an impressive driving score of 51.77 on Bench2Drive, proving the superiority. Code, model, and configurations will be released here.

Method: Proposes ViT+UNet, a hybrid deep learning architecture that integrates a U-Net with a Vision Transformer for TFM data analysis, with structured input data to allow inclusion of metadata such as cell-type information.

Result: The hybrid model outperforms standalone U-Net and Vision Transformer architectures in predicting traction force fields, exhibits superior generalization across diverse spatial scales and varying noise levels, and enables application to TFM datasets from different experimental setups.

Conclusion: ViT+UNet provides a robust solution for TFM analysis that addresses scale and noise challenges while allowing metadata integration for improved specificity and accuracy in cellular force quantification.

Abstract: Traction force microscopy (TFM) is a widely used technique for quantifying the forces that cells exert on their surrounding extracellular matrix. Although deep learning methods have recently been applied to TFM data analysis, several challenges remain-particularly achieving reliable inference across multiple spatial scales and integrating additional contextual information such as cell type to improve accuracy. In this study, we propose ViT+UNet, a robust deep learning architecture that integrates a U-Net with a Vision Transformer. Our results demonstrate that this hybrid model outperforms both standalone U-Net and Vision Transformer architectures in predicting traction force fields. Furthermore, ViT+UNet exhibits superior generalization across diverse spatial scales and varying noise levels, enabling its application to TFM datasets obtained from different experimental setups and imaging systems. By appropriately structuring the input data, our approach also allows the inclusion of metadata, in our case cell-type information, to enhance prediction specificity and accuracy.

Method: MAD uses a pretraining strategy that jointly self-distills both the morphology view (cell appearance) and microenvironment view (surrounding tissue context) of the same indexed cell into a unified embedding space through dual-view joint self-distillation.

Result: Achieves state-of-the-art performance across diverse tissues and imaging modalities for cell subtyping, transcriptomic prediction, and bioinformatic inference, outperforming foundation models with similar parameters trained on larger datasets.

Conclusion: MAD effectively captures cellular complexity and diversity within tissues, establishing it as a general tool for representation learning in microscopy that enables virtual spatial omics and biological insights from large microscopy datasets.

Abstract: Bridging microscopy and omics would allow us to read molecular states from images-at single-cell resolution and tissue scale-without the cost and throughput limits of omics technologies. Self-supervised pretraining offers a scalable approach with minimal labels, yet how to encode single-cell identity within tissue environments-and the extent of biological information such models can capture-remains an open question. Here, we introduce MAD (microenvironment-aware distillation), a pretraining strategy that learns cell-centric embeddings by jointly self-distilling the morphology view and the microenvironment view of the same indexed cell into a unified embedding space. Across diverse tissues and imaging modalities, MAD achieves state-of-the-art prediction performance on downstream tasks including cell subtyping, transcriptomic prediction, and bioinformatic inference. MAD even outperforms foundation models with a similar number of model parameters that have been trained on substantially larger datasets. These results demonstrate that MAD’s dual-view joint self-distillation effectively captures the complexity and diversity of cells within tissues. Together, this establishes MAD as a general tool for representation learning in microscopy, enabling virtual spatial omics and biological insights from vast microscopy datasets.

Method: EVD is a DiT-compatible framework with: 1) lightweight event head predicting token-aligned event activity, 2) event-grounded losses coupling activity to state change during training, and 3) event-gated sampling with hysteresis and early-step scheduling to suppress spurious updates and concentrate updates during interactions.

Result: On EVD-Bench, EVD consistently improves human preference and VBench dynamics, substantially reducing failure modes in state persistence, spatial accuracy, support relations, and contact stability without sacrificing appearance quality.

Conclusion: Explicit event grounding is a practical abstraction for reducing interaction hallucinations in video generation, addressing fundamental limitations of frame-first approaches.

Abstract: State-of-the-art text-to-video models often look realistic frame-by-frame yet fail on simple interactions: motion starts before contact, actions are not realized, objects drift after placement, and support relations break. We argue this stems from frame-first denoising, which updates latent state everywhere at every step without an explicit notion of when and where an interaction is active. We introduce Event-Driven Video Generation (EVD), a minimal DiT-compatible framework that makes sampling event-grounded: a lightweight event head predicts token-aligned event activity, event-grounded losses couple activity to state change during training, and event-gated sampling (with hysteresis and early-step scheduling) suppresses spurious updates while concentrating updates during interactions. On EVD-Bench, EVD consistently improves human preference and VBench dynamics, substantially reducing failure modes in state persistence, spatial accuracy, support relations, and contact stability without sacrificing appearance. These results indicate that explicit event grounding is a practical abstraction for reducing interaction hallucinations in video generation.

Method: Three CLIP-based approaches: (1) zero-shot baseline with prompt engineering, (2) hybrid FCN-CLIP model with CBAM attention, (3) ranking-aware prompting that encodes ordinal structure of DR progression. Trained on combined APTOS 2019 and Messidor-2 dataset (n=5,406) with resampling and class-specific optimal thresholding.

Result: Ranking-aware model achieved highest overall accuracy (93.42%, AUROC 0.9845) with strong recall on severe cases. Hybrid FCN-CLIP model achieved 92.49% accuracy (AUROC 0.99) and excelled at detecting proliferative DR. Both substantially outperformed zero-shot baseline (55.17%, AUROC 0.75).

Conclusion: CLIP-based approaches show strong performance for DR severity grading, with ranking-aware prompting and hybrid FCN-CLIP models offering complementary strengths for practical screening applications.

Abstract: Diabetic retinopathy (DR) is a leading cause of preventable blindness, and automated fundus image grading can play an important role in large-scale screening. In this work, we investigate three CLIP-based approaches for five-class DR severity grading: (1) a zero-shot baseline using prompt engineering, (2) a hybrid FCN-CLIP model augmented with CBAM attention, and (3) a ranking-aware prompting model that encodes the ordinal structure of DR progression. We train and evaluate on a combined dataset of APTOS 2019 and Messidor-2 (n=5,406), addressing class imbalance through resampling and class-specific optimal thresholding. Our experiments show that the ranking-aware model achieves the highest overall accuracy (93.42%, AUROC 0.9845) and strong recall on clinically critical severe cases, while the hybrid FCN-CLIP model (92.49%, AUROC 0.99) excels at detecting proliferative DR. Both substantially outperform the zero-shot baseline (55.17%, AUROC 0.75). We analyze the complementary strengths of each approach and discuss their practical implications for screening contexts.

Method: Proposes a recognition framework integrating temporal segment modeling with MLLMs. Uses segment-based strategy (5-second clips) to handle computational efficiency and token constraints. Leverages Qwen3-Omni-30B-A3B model fine-tuned on BAH dataset using LoRA and full-parameter strategies via MS-Swift framework to synergistically analyze visual and auditory signals.

Result: Achieves 85.1% accuracy on test set, significantly outperforming existing benchmarks, validating MLLMs’ superior capability in capturing complex emotional conflicts.

Conclusion: Demonstrates effectiveness of MLLMs in recognizing subtle emotional states through cross-modal analysis, with practical applications in behavioral intervention and digital health.

Abstract: Emotion recognition in videos is a pivotal task in affective computing, where identifying subtle psychological states such as Ambivalence and Hesitancy holds significant value for behavioral intervention and digital health. Ambivalence and Hesitancy states often manifest through cross-modal inconsistencies such as discrepancies between facial expressions, vocal tones, and textual semantics, posing a substantial challenge for automated recognition. This paper proposes a recognition framework that integrates temporal segment modeling with Multimodal Large Language Models. To address computational efficiency and token constraints in long video processing, we employ a segment-based strategy, partitioning videos into short clips with a maximum duration of 5 seconds. We leverage the Qwen3-Omni-30B-A3B model, fine-tuned on the BAH dataset using LoRA and full-parameter strategies via the MS-Swift framework, enabling the model to synergistically analyze visual and auditory signals. Experimental results demonstrate that the proposed method achieves an accuracy of 85.1% on the test set, significantly outperforming existing benchmarks and validating the superior capability of Multimodal Large Language Models in capturing complex and nuanced emotional conflicts. The code is released at https://github.com/dlnn123/A-H-Detection-with-Qwen-Omni.git.

Method: PHARL (PHysics-aware Alignment Representation Learning) learns fall representations without clinical outcome labels using two complementary constraints: (1) trajectory-level temporal consistency for stable representation learning, and (2) multi-class physics alignment where simulation-derived contact outcomes shape embedding geometry. The method pairs video windows with temporally aligned simulation descriptors to capture local impact-relevant dynamics while maintaining feed-forward inference.

Result: Experiments on four public datasets show PHARL consistently improves risk-aligned representation quality over visual-only baselines while maintaining strong fall-detection performance. Notably, PHARL exhibits zero-shot ordinality where an interpretable severity structure (Head > Trunk > Supported) emerges without explicit ordinal supervision.

Conclusion: PHARL successfully addresses the limitations of supervised injury prediction by learning physically meaningful fall representations without requiring clinical outcome labels, enabling better risk assessment through physics-aware alignment while maintaining practical deployment feasibility.

Abstract: Vision-based fall analysis has advanced rapidly, but a key bottleneck remains: visually similarmotions can correspond to very different physical outcomes because small differences in contactmechanics and protective responses are hard to infer from appearance alone. Most existingapproaches handle this by supervised injury prediction, which depends on reliable injury labels.In practice, such labels are difficult to obtain: video evidence is often ambiguous (occlusion,viewpoint limits), and true injury events are rare and cannot be safely staged, leading to noisysupervision. We address this problem with PHARL (PHysics-aware Alignment RepresentationLearning), which learns physically meaningful fall representations without requiring clinicaloutcome labels. PHARL regularizes motion embeddings with two complementary constraints:(1) trajectory-level temporal consistency for stable representation learning, and (2) multi-classphysics alignment, where simulation-derived contact outcomes shape embedding geometry. Bypairing video windows with temporally aligned simulation descriptors, PHARL captures localimpact-relevant dynamics while keeping inference purely feed-forward. Experiments on fourpublic datasets show that PHARL consistently improves risk-aligned representation quality overvisual-only baselines while maintaining strong fall-detection performance. Notably, PHARL alsoexhibits zero-shot ordinality: an interpretable severity structure (Head > Trunk > Supported)emerges without explicit ordinal supervision.

Method: Two-stage framework: 1) Watching stage builds hierarchical memory with Short-Term Memory (STM) for recent frames and Long-Term Memory (LTM) with redundancy-aware eviction policy for historical summaries. 2) Thinking stage uses context-aware retrieval to combine query with STM context and retrieve relevant historical frames from LTM for cross-temporal reasoning.

Result: Achieves state-of-the-art performance on online video benchmarks: 77.7% accuracy on StreamingBench and 55.2% on OVO-Bench, outperforming existing open-source online Video LLMs while operating at real-time frame rates.

Conclusion: WAT effectively addresses online video reasoning challenges by separating watching and thinking stages with hierarchical memory management, enabling efficient long-context preservation and real-time performance.

Abstract: Multimodal Large Language Models (MLLMs) have shown strong capabilities in image understanding, motivating recent efforts to extend them to video reasoning. However, existing Video LLMs struggle in online streaming scenarios, where long temporal context must be preserved under strict memory constraints. We propose WAT (Watching Before Thinking), a two-stage framework for online video reasoning. WAT separates processing into a query-independent watching stage and a query-triggered thinking stage. The watching stage builds a hierarchical memory system with a Short-Term Memory (STM) that buffers recent frames and a fixed-capacity Long-Term Memory (LTM) that maintains a diverse summary of historical content using a redundancy-aware eviction policy. In the thinking stage, a context-aware retrieval mechanism combines the query with the current STM context to retrieve relevant historical frames from the LTM for cross-temporal reasoning. To support training for online video tasks, we introduce WAT-85K, a dataset containing streaming-style annotations emphasizing real-time perception, backward tracing, and forecasting. Experiments show that WAT achieves state-of-the-art performance on online video benchmarks, including 77.7% accuracy on StreamingBench and 55.2% on OVO-Bench, outperforming existing open-source online Video LLMs while operating at real-time frame rates.

Method: Proposes distance-aware soft prompt learning that partitions VA space into 9 emotional regions with textual descriptions, uses Gaussian kernel for soft labels based on Euclidean distance. Uses CLIP image encoder and Audio Spectrogram Transformer (AST) for multimodal features, GRUs for temporal modeling, and hierarchical fusion with cross-modal attention and gated fusion.

Result: Experimental results on Aff-Wild2 dataset show the semantic-guided approach significantly enhances VA estimation accuracy, achieving competitive performance in unconstrained “in-the-wild” scenarios.

Conclusion: The proposed framework effectively bridges semantic space with continuous emotional dimensions through distance-aware soft prompt learning and hierarchical multimodal fusion, demonstrating improved performance for VA estimation in naturalistic settings.

Abstract: Valence-arousal (VA) estimation is crucial for capturing the nuanced nature of human emotions in naturalistic environments. While pre-trained Vision-Language models like CLIP have shown remarkable semantic alignment capabilities, their application in continuous regression tasks is often limited by the discrete nature of text prompts. In this paper, we propose a novel multimodal framework for VA estimation that introduces Distance-aware Soft Prompt Learning to bridge the gap between semantic space and continuous dimensions. Specifically, we partition the VA space into a 3X3 grid, defining nine emotional regions, each associated with distinct textual descriptions. Rather than a hard categorization, we employ a Gaussian kernel to compute soft labels based on the Euclidean distance between the ground truth coordinates and the region centers, allowing the model to learn fine-grained emotional transitions. For multimodal integration, our architecture utilizes a CLIP image encoder and an Audio Spectrogram Transformer (AST) to extract robust spatial and acoustic features. These features are temporally modeled via Gated Recurrent Units (GRUs) and integrated through a hierarchical fusion scheme that sequentially combines cross-modal attention for alignment and gated fusion for adaptive refinement. Experimental results on the Aff-Wild2 dataset demonstrate that our proposed semantic-guided approach significantly enhances the accuracy of VA estimation, achieving competitive performance in unconstrained ``in-the-wild’’ scenarios.

Method: MIBench formulates each instance as a (con_v, con_t, task) triplet requiring correct multimodal interaction. It evaluates three key aspects: vision-centric information sourcing, text-centric information sourcing, and joint synergy generation - each assessed hierarchically across Recognition, Understanding, and Reasoning cognitive levels.

Result: Evaluation of state-of-the-art LMMs shows: (1) multimodal interaction ability remains constrained despite scaling; (2) models are easily distracted by text when processing vision; (3) basic capacity for multimodal synergy exists; (4) natively trained multimodal models show deficits in fundamental interaction ability.

Conclusion: MIBench provides a comprehensive framework for assessing multimodal interaction capabilities, revealing current limitations in LMMs and offering insights for developing models with enhanced multimodal abilities in the future.

Abstract: In different multimodal scenarios, it needs to integrate and utilize information across modalities in a specific way based on the demands of the task. Different integration ways between modalities are referred to as “multimodal interaction”. How well a model handles various multimodal interactions largely characterizes its multimodal ability. In this paper, we introduce MIBench, a comprehensive benchmark designed to evaluate the multimodal interaction capabilities of Large Multimodal Models (LMMs), which formulates each instance as a (con_v , con_t, task) triplet with contexts from vision and text, necessitating that LMMs employ correct forms of multimodal interaction to effectively complete the task. MIBench assesses models from three key aspects: the ability to source information from vision-centric or text-centric cues, and the ability to generate new information from their joint synergy. Each interaction capability is evaluated hierarchically across three cognitive levels: Recognition, Understanding, and Reasoning. MIBench comprises over 10,000 vision-text context pairs spanning 32 distinct tasks. Evaluation of state-of-the-art LMMs show that: (1) LMMs’ ability on multimodal interaction remains constrained, despite the scaling of model parameters and training data; (2) they are easily distracted by textual modalities when processing vision information; (3) they mostly possess a basic capacity for multimodal synergy; and (4) natively trained multimodal models show noticeable deficits in fundamental interaction ability. We expect that these observations can serve as a reference for developing LMMs with more enhanced multimodal ability in the future.

Method: Proposes MSD-DETR with three key innovations: (1) structural re-parameterization strategy decoupling training-time multi-branch topology from inference-time efficiency; (2) deformable attention mechanism for content-adaptive spatial sampling; (3) cross-scale feature fusion architecture with GSConv modules and VoVGSCSP blocks for multi-resolution information aggregation.

Result: Achieves 92.4% mAP@0.5 at 98 FPS on real-world locomotive coil spring dataset, outperforming YOLOv8 (+3.1% mAP) and baseline RT-DETR (+2.8% mAP) while maintaining comparable inference speed.

Conclusion: MSD-DETR establishes a new benchmark for industrial coil spring quality inspection by effectively addressing morphological diversity, scale variations, and complex backgrounds while maintaining real-time performance.

Abstract: Automated visual inspection of locomotive coil springs presents significant challenges due to the morphological diversity of surface defects, substantial scale variations, and complex industrial backgrounds. This paper proposes MSD-DETR (Multi-Scale Deformable Detection Transformer), a novel detection framework that addresses these challenges through three key innovations: (1) a structural re-parameterization strategy that decouples training-time multi-branch topology from inference-time efficiency, enhancing feature extraction while maintaining real-time performance; (2) a deformable attention mechanism that enables content-adaptive spatial sampling, allowing dynamic focus on defect-relevant regions regardless of morphological irregularity; and (3) a cross-scale feature fusion architecture incorporating GSConv modules and VoVGSCSP blocks for effective multi-resolution information aggregation. Comprehensive experiments on a real-world locomotive coil spring dataset demonstrate that MSD-DETR achieves 92.4% mAP@0.5 at 98 FPS, outperforming state-of-the-art detectors including YOLOv8 (+3.1% mAP) and the baseline RT-DETR (+2.8% mAP) while maintaining comparable inference speed, establishing a new benchmark for industrial coil spring quality inspection.

Method: Treats spatial transcriptomics as croppable images by defining multi-channel image representation with fixed spatial size through cropping patches from raw slides, with gene subset selection rules to control input dimensionality and improve pretraining stability.

Result: Extensive experiments show the proposed image-like dataset construction consistently improves downstream performance compared to conventional pretraining schemes, with ablation studies confirming both spatial patching and channel design are necessary.

Conclusion: Establishes a unified, practical paradigm for organizing spatial transcriptomics data that enables large-scale pretraining by treating it as croppable images with controlled dimensionality.

Abstract: Spatial Transcriptomics (ST) profiles thousands of gene expression values at discrete spots with precise coordinates on tissue sections, preserving spatial context essential for clinical and pathological studies. With rising sequencing throughput and advancing platforms, the expanding data volumes motivate large-scale ST pretraining. However, the fundamental unit for pretraining, i.e., what constitutes a single training sample, remains ill-posed. Existing choices fall into two camps: (1) treating each spot as an independent sample, which discards spatial dependencies and collapses ST into single-cell transcriptomics; and (2) treating an entire slide as a single sample, which produces prohibitively large inputs and drastically fewer training examples, undermining effective pretraining. To address this gap, we propose treating spatial transcriptomics as croppable images. Specifically, we define a multi-channel image representation with fixed spatial size by cropping patches from raw slides, thereby preserving spatial context while substantially increasing the number of training samples. Along the channel dimension, we define gene subset selection rules to control input dimensionality and improve pretraining stability. Extensive experiments show that the proposed image-like dataset construction for ST pretraining consistently improves downstream performance, outperforming conventional pretraining schemes. Ablation studies verify that both spatial patching and channel design are necessary, establishing a unified, practical paradigm for organizing ST data and enabling large-scale pretraining.

Method: Proposes CtrlAttack that represents perturbations as low-dimensional velocity fields and constructs continuous displacement fields via temporal integration to affect state transitions while maintaining temporal consistency. The method maps perturbations to observation space to work in both white-box and black-box settings.

Result: The attack significantly disrupts temporal consistency with attack success rates over 90% in white-box and over 80% in black-box settings, while keeping FID variation within 6 and FVD within 130, revealing security risks in I2V models’ state dynamics.

Conclusion: Temporal control mechanisms in diffusion-based I2V models are vulnerable to trajectory-control attacks, revealing potential security risks at the state dynamics level that need to be addressed for robust video generation systems.

Abstract: Diffusion-based image-to-video (I2V) models increasingly exhibit world-model-like properties by implicitly capturing temporal dynamics. However, existing studies have mainly focused on visual quality and controllability, and the robustness of the state transition learned by the model remains understudied. To fill this gap, we are the first to analyze the vulnerability of I2V models, find that temporal control mechanisms constitute a new attack surface, and reveal the challenge of modeling them uniformly under different attack settings. Based on this, we propose a trajectory-control attack, called CtrlAttack, to interfere with state evolution during the generation process. Specifically, we represent the perturbation as a low-dimensional velocity field and construct a continuous displacement field via temporal integration, thereby affecting the model’s state transitions while maintaining temporal consistency; meanwhile, we map the perturbation to the observation space, making the method applicable to both white-box and black-box attack settings. Experimental results show that even under low-dimensional and strongly regularized perturbation constraints, our method can still significantly disrupt temporal consistency by increasing the attack success rate (ASR) to over 90% in the white-box setting and over 80% in the black-box setting, while keeping the variation of the FID and FVD within 6 and 130, respectively, thus revealing the potential security risk of I2V models at the level of state dynamics.

Method: Combines multi-modal AIRT analysis (PCT, TSR, PPT) with modality-aware textual reporting using VLMs. Thermal sequences are processed, anomaly masks fused into consensus segmentation, and fused evidence provided to VLM for structured report generation.

Result: Evaluation on two marquetries demonstrates consistent anomaly detection and stable structured interpretations, indicating reproducibility and generalizability across samples.

Conclusion: The framework enables automated, standardized thermographic analysis and reporting for cultural heritage conservation, addressing current limitations in expert-dependent interpretation and lack of standardization.

Abstract: Authenticity and condition assessment are central to conservation decision-making, yet interpretation and reporting of thermographic output remain largely bespoke and expert-dependent, complicating comparison across collections and limiting systematic integration into conservation documentation. Pulsed Active Infrared Thermography (AIRT) is sensitive to subsurface features such as material heterogeneity, voids, and past interventions; however, its broader adoption is constrained by artifact misinterpretation, inter-laboratory variability, and the absence of standardized, explainable reporting frameworks. Although multi-modal thermographic processing techniques are established, their integration with structured natural-language interpretation has not been explored in cultural heritage. A fully automated thermography-vision-language model (VLM) framework is presented. It combines multi-modal AIRT analysis with modality-aware textual reporting, without human intervention during inference. Thermal sequences are processed using Principal Component Thermography (PCT), Thermographic Signal Reconstruction (TSR), and Pulsed Phase Thermography (PPT), and the resulting anomaly masks are fused into a consensus segmentation that emphasizes regions supported by multiple thermal indicators while mitigating boundary artifacts. The fused evidence is provided to a VLM, which generates structured reports describing the location of the anomaly, thermal behavior, and plausible physical interpretations while explicitly acknowledging the uncertainty and diagnostic limitations. Evaluation on two marquetries demonstrates consistent anomaly detection and stable structured interpretations, indicating reproducibility and generalizability across samples.

Method: Draft-and-Target Sampling: training-free diffusion inference paradigm using self-play denoising with two complementary trajectories (draft sampling with large steps for fast global trajectory, target sampling with small steps for verification). Plus token chunking and progressive acceptance strategy to reduce redundant computation.

Result: Achieves up to 2.1x speedup on three benchmarks with minimal compromise to success rate, improving efficiency of current state-of-the-art methods.

Conclusion: Proposed method effectively addresses computational efficiency challenges in video generation policies for robotics applications.

Abstract: Video generation models have been used as a robot policy to predict the future states of executing a task conditioned on task description and observation. Previous works ignore their high computational cost and long inference time. To address this challenge, we propose Draft-and-Target Sampling, a novel diffusion inference paradigm for video generation policy that is training-free and can improve inference efficiency. We introduce a self-play denoising approach by utilizing two complementary denoising trajectories in a single model, draft sampling takes large steps to generate a global trajectory in a fast manner and target sampling takes small steps to verify it. To further speedup generation, we introduce token chunking and progressive acceptance strategy to reduce redundant computation. Experiments on three benchmarks show that our method can achieve up to 2.1x speedup and improve the efficiency of current state-of-the-art methods with minimal compromise to the success rate. Our code is available.

Method: Locality-Aware Dynamic Rescue (LADR) exploits spatial Markov properties of images by prioritizing recovery of tokens at the “generation frontier” (regions adjacent to observed pixels). It uses morphological neighbor identification, risk-bounded filtering to prevent error propagation, and manifold-consistent inverse scheduling to align diffusion trajectory with accelerated mask density.

Result: Achieves ~4x speedup over standard baselines on four text-to-image generation benchmarks while maintaining or even enhancing generative fidelity, particularly in spatial reasoning tasks.

Conclusion: LADR offers state-of-the-art trade-off between efficiency and quality for discrete diffusion language models in multimodal generation, enabling faster inference without retraining.

Abstract: Discrete Diffusion Language Models have emerged as a compelling paradigm for unified multimodal generation, yet their deployment is hindered by high inference latency arising from iterative decoding. Existing acceleration strategies often require expensive re-training or fail to leverage the 2D spatial redundancy inherent in visual data. To address this, we propose Locality-Aware Dynamic Rescue (LADR), a training-free method that expedites inference by exploiting the spatial Markov property of images. LADR prioritizes the recovery of tokens at the ‘‘generation frontier’’, regions spatially adjacent to observed pixels, thereby maximizing information gain. Specifically, our method integrates morphological neighbor identification to locate candidate tokens, employs a risk-bounded filtering mechanism to prevent error propagation, and utilizes manifold-consistent inverse scheduling to align the diffusion trajectory with the accelerated mask density. Extensive experiments on four text-to-image generation benchmarks demonstrate that our LADR achieves an approximate 4 x speedup over standard baselines. Remarkably, it maintains or even enhances generative fidelity, particularly in spatial reasoning tasks, offering a state-of-the-art trade-off between efficiency and quality.

Method: Benchmarked four GAN architectures (DCGAN, StyleGAN2, two StyleGAN3 variants) on ISIC 2018/2020 datasets with unified preprocessing. Used R1 regularization tuning and evaluated with multi-faceted protocol: FID, FMD, qualitative inspection, downstream classification with frozen EfficientNet, and dermatologist evaluation.

Result: StyleGAN2 achieved best balance with FID scores of 24.8 (ISIC 2018) and 7.96 (ISIC 2020). Frozen classifier recognized 83% of generated images as melanoma. Dermatologists distinguished synthetic from real at only 66.5% accuracy. Adding synthetic images improved melanoma detection AUC from 0.925 to 0.945.

Conclusion: StyleGAN2-generated melanoma images preserve diagnostically relevant features and can effectively mitigate class imbalance in melanoma-focused machine learning pipelines, providing measurable benefits for automated skin cancer detection.

Abstract: Melanoma is the most lethal form of skin cancer, and early detection is critical for improving patient outcomes. Although dermoscopy combined with deep learning has advanced automated skin-lesion analysis, progress is hindered by limited access to large, well-annotated datasets and by severe class imbalance, where melanoma images are substantially underrepresented. To address these challenges, we present the first systematic benchmarking study comparing four GAN architectures-DCGAN, StyleGAN2, and two StyleGAN3 variants (T/R)-for high-resolution melanoma-specific synthesis. We train and optimize all models on two expert-annotated benchmarks (ISIC 2018 and ISIC 2020) under unified preprocessing and hyperparameter exploration, with particular attention to R1 regularization tuning. Image quality is assessed through a multi-faceted protocol combining distribution-level metrics (FID), sample-level representativeness (FMD), qualitative dermoscopic inspection, downstream classification with a frozen EfficientNet-based melanoma detector, and independent evaluation by two board-certified dermatologists. StyleGAN2 achieves the best balance of quantitative performance and perceptual quality, attaining FID scores of 24.8 (ISIC 2018) and 7.96 (ISIC 2020) at gamma=0.8. The frozen classifier recognizes 83% of StyleGAN2-generated images as melanoma, while dermatologists distinguish synthetic from real images at only 66.5% accuracy (chance = 50%), with low inter-rater agreement (kappa = 0.17). In a controlled augmentation experiment, adding synthetic melanoma images to address class imbalance improved melanoma detection AUC from 0.925 to 0.945 on a held-out real-image test set. These findings demonstrate that StyleGAN2-generated melanoma images preserve diagnostically relevant features and can provide a measurable benefit for mitigating class imbalance in melanoma-focused machine learning pipelines.

Method: Introduces per-frame action plans as text latents that serve as dense semantic anchors throughout denoising. Uses latent-specific diffusion steps allowing independent denoising of motion latents with flexible sampling orders. Supports history-conditioned, future-aware real-time streaming and offline generation within the same model.

Result: Achieves 5.25x faster real-time streaming while improving motion quality by 18% in terms of FID compared to previous best methods. Also enables zero-shot motion editing and in-betweening without additional models.

Conclusion: ActionPlan successfully bridges real-time streaming with high-quality offline generation in a single unified framework, demonstrating superior efficiency and quality while enabling additional capabilities like motion editing.

Abstract: We present ActionPlan, a unified motion diffusion framework that bridges real-time streaming with high-quality offline generation within a single model. The core idea is to introduce a per-frame action plan: the model predicts frame-level text latents that act as dense semantic anchors throughout denoising, and uses them to denoise the full motion sequence with combined semantic and motion cues. To support this structured workflow, we design latent-specific diffusion steps, allowing each motion latent to be denoised independently and sampled in flexible orders at inference. As a result, ActionPlan can run in a history-conditioned, future-aware mode for real-time streaming, while also supporting high-quality offline generation. The same mechanism further enables zero-shot motion editing and in-betweening without additional models. Experiments demonstrate that our real-time streaming is 5.25x faster while also achieving 18% motion quality improvement over the best previous method in terms of FID.

Method: Proposes a “Raising the Fulcrum, Tuning to Balance” approach: 1) Quality-over-quantity data filtering (automated + manual), 2) Tune-to-Balance post-training with cross-pair/in-pair data and model merging, 3) Two DPO pipelines (Consis-DPO and Real-Fake DPO), 4) Time-dependent dynamic classifier-free guidance for inference.

Result: LibraGen outperforms both open-source and commercial S2V models using only thousand-scale training data, achieving better balance between subject fidelity, motion coherence, and visual quality.

Conclusion: The framework successfully balances VGFM’s intrinsic capabilities with S2V extension through strategic data curation and training techniques, demonstrating that quality matters more than quantity for effective subject-to-video generation.

Abstract: With the advancement of video generation foundation models (VGFMs), customized generation, particularly subject-to-video (S2V), has attracted growing attention. However, a key challenge lies in balancing the intrinsic priors of a VGFM, such as motion coherence, visual aesthetics, and prompt alignment, with its newly derived S2V capability. Existing methods often neglect this balance by enhancing one aspect at the expense of others. To address this, we propose LibraGen, a novel framework that views extending foundation models for S2V generation as a balance game between intrinsic VGFM strengths and S2V capability. Specifically, guided by the core philosophy of “Raising the Fulcrum, Tuning to Balance,” we identify data quality as the fulcrum and advocate a quality-over-quantity approach. We construct a hybrid pipeline that combines automated and manual data filtering to improve overall data quality. To further harmonize the VGFM’s native capabilities with its S2V extension, we introduce a Tune-to-Balance post-training paradigm. During supervised fine-tuning, both cross-pair and in-pair data are incorporated, and model merging is employed to achieve an effective trade-off. Subsequently, two tailored direct preference optimization (DPO) pipelines, namely Consis-DPO and Real-Fake DPO, are designed and merged to consolidate this balance. During inference, we introduce a time-dependent dynamic classifier-free guidance scheme to enable flexible and fine-grained control. Experimental results demonstrate that LibraGen outperforms both open-source and commercial S2V models using only thousand-scale training data.

Method: Uses generative models as black boxes via API calls, VLM for automatic defect prompt generation, CLIP-based quality filter, and a training-free dual-branch semantic change detection module combining text-conditioned Grounding DINO with YOLOv26-Seg structural features for mask generation.

Result: Benchmarked on MVTec AD and VisA, showing improved anomaly segmentation performance and high visual quality validated by metrics (IS, IC-LPIPS) and human perceptual study with 31 participants and 1,550 pairwise votes.

Conclusion: MIRAGE provides scalable, accessible foundation for anomaly-aware industrial inspection without real defect data, and releases a large-scale dataset of 13,000+ image-mask pairs with prompts and code.

Abstract: Industrial visual anomaly detection (VAD) methods are typically trained on normal samples only, yet performance improves substantially when even limited anomalous data is available. Existing anomaly generation approaches either require real anomalous examples, demand expensive hardware, or produce synthetic defects that lack realism. We present MIRAGE (Model-agnostic Industrial Realistic Anomaly Generation and Evaluation), a fully automated pipeline for realistic anomalous image generation and pixel-level mask creation that requires no training and no anomalous images. Our pipeline accesses any generative model as a black box via API calls, uses a VLM for automatic defect prompt generation, and includes a CLIP-based quality filter to retain only well-aligned generated images. For mask generation at scale, we introduce a lightweight, training-free dual-branch semantic change detection module combining text-conditioned Grounding DINO features with fine-grained YOLOv26-Seg structural features. We benchmark four generation methods using Gemini 2.5 Flash Image (Nano Banana) as the generative backbone, evaluating performance on MVTec AD and VisA across two distinct tasks: (i) downstream anomaly segmentation and (ii) visual quality of the generated images, assessed via standard metrics (IS, IC-LPIPS) and a human perceptual study involving 31 participants and 1,550 pairwise votes. The results demonstrate that MIRAGE offers a scalable, accessible foundation for anomaly-aware industrial inspection that requires no real defect data. As a final contribution, we publicly release a large-scale dataset comprising 500 image-mask pairs per category for every MVTec AD and VisA class, over 13,000 pairs in total, alongside all generation prompts and pipeline code.

Method: Benchmarked 10 GAN architectures on SynthRAD2025 dataset across abdomen, thorax, and head-and-neck regions using unified validation protocol with identical preprocessing and optimization. Evaluated using voxel-wise accuracy, structural fidelity, perceptual quality, and distribution-level realism metrics.

Result: Supervised paired models consistently outperformed unpaired approaches. Pix2Pix achieved most balanced performance across anatomical districts with favorable quality-to-complexity trade-off. Multi-district training improved structural robustness while intra-district training maximized voxel-wise fidelity.

Conclusion: Provides quantitative and computational guidance for model selection in MRI-only radiotherapy workflows and establishes reproducible framework for future comparative studies in medical image translation.

Abstract: The translation from Magnetic resonance imaging (MRI) to Computed tomography (CT) has been proposed as an effective solution to facilitate MRI-only clinical workflows while limiting exposure to ionizing radiation. Although numerous Generative Adversarial Network (GAN) architectures have been proposed for MRI-to-CT translation, systematic and fair comparisons across heterogeneous models remain limited. We present a comprehensive benchmark of ten GAN architectures evaluated on the SynthRAD2025 dataset across three anatomical districts (abdomen, thorax, head-and-neck). All models were trained under a unified validation protocol with identical preprocessing and optimization settings. Performance was assessed using complementary metrics capturing voxel-wise accuracy, structural fidelity, perceptual quality, and distribution-level realism, alongside an analysis of computational complexity. Supervised Paired models consistently outperformed Unpaired approaches, confirming the importance of voxel-wise supervision. Pix2Pix achieved the most balanced performance across districts while maintaining a favorable quality-to-complexity trade-off. Multi-district training improved structural robustness, whereas intra-district training maximized voxel-wise fidelity. This benchmark provides quantitative and computational guidance for model selection in MRI-only radiotherapy workflows and establishes a reproducible framework for future comparative studies. To ensure the reproducibility of our experiments we make our code public, together with the overall results, at the following link:https://github.com/arco-group/MRI_TO_CT.git

Method: Theoretical proofs: (1) Finite Primitive Basis Theorem showing all imaging forward models decompose into 11 primitives; (2) Triad Decomposition showing all reconstruction failures have three root causes. Validation across 12 modalities and all five carrier families with practical recovery improvements on deployed instruments.

Result: Validation across 12 modalities confirms both theoretical results, with recovery improvements ranging from +0.8 to +13.9 dB on deployed instruments. The framework successfully diagnoses and corrects real-world imaging system issues.

Conclusion: The 11 primitives and 3 gates establish the first universal grammar for designing, diagnosing, and correcting computational imaging systems, providing a unified framework that spans all imaging modalities and carrier families.

Abstract: Computational imaging systems – from coded-aperture cameras to cryo-electron microscopes – span five carrier families yet share a hidden structural simplicity. We prove that every imaging forward model decomposes into a directed acyclic graph over exactly 11 physically typed primitives (Finite Primitive Basis Theorem) – a sufficient and minimal basis that provides a compositional language for designing any imaging modality. We further prove that every reconstruction failure has exactly three independent root causes: information deficiency, carrier noise, and operator mismatch (Triad Decomposition). The three gates map to the system lifecycle: Gates 1 and 2 guide design (sampling geometry, carrier selection); Gate 3 governs deployment-stage calibration and drift correction. Validation across 12 modalities and all five carrier families confirms both results, with +0.8 to +13.9 dB recovery on deployed instruments. Together, the 11 primitives and 3 gates establish the first universal grammar for designing, diagnosing, and correcting computational imaging systems.

Method: Systematic domain-specific investigation examining existence, consistency, and practical implications of redundancy across key dimensions of EO variability, testing across tasks, geospatial locations, sensors, ground sampling distances, and architectural designs.

Result: Redundancy in EO data is substantial and pervasive: exploiting it yields comparable performance (≈98.5% of baseline) at ≈4× fewer GFLOPs, consistent across various experimental conditions.

Conclusion: Multi-faceted redundancy is a structural property of EO data rather than an artifact of specific experimental choices, laying groundwork for more efficient, scalable, and accessible large-scale EO models.

Abstract: The growing availability of Earth Observation (EO) data and recent advances in Computer Vision have driven rapid progress in machine learning for EO, producing domain-specific models at ever-increasing scales. Yet this progress risks overlooking fundamental properties of EO data that distinguish it from other domains. We argue that EO data exhibit a multidimensional redundancy (spectral, temporal, spatial, and semantic) which has a more pronounced impact on the domain and its applications than what current literature reflects. To validate this hypothesis, we conduct a systematic domain-specific investigation examining the existence, consistency, and practical implications of this phenomenon across key dimensions of EO variability. Our findings confirm that redundancy in EO data is both substantial and pervasive: exploiting it yields comparable performance ($\approx98.5%$ of baseline) at a fraction of the computational cost ($\approx4\times$ fewer GFLOPs), at both training and inference. Crucially, these gains are consistent across tasks, geospatial locations, sensors, ground sampling distances, and architectural designs; suggesting that multi-faceted redundancy is a structural property of EO data rather than an artifact of specific experimental choices. These results lay the groundwork for more efficient, scalable, and accessible large-scale EO models.

Method: SAIF constructs joint uncertainty space via structured box perturbations and threshold variations, evaluates hypotheses using decision stability and boundary consistency, and uses stability-consistency score to filter unstable candidates and perform stability-weighted fusion in probability space.

Result: Experiments on Synapse, CVC-ClinicDB, Kvasir-SEG, and CVC-300 show SAIF consistently improves segmentation accuracy and robustness, achieving state-of-the-art performance without retraining or architectural changes.

Conclusion: SAIF provides effective training-free solution to improve SAM’s inference stability for medical image segmentation by explicitly modeling uncertainty in prompts and thresholds.

Abstract: Segment Anything Model (SAM) enable scalable medical image segmentation but suffer from inference-time instability when deployed as a frozen backbone. In practice, bounding-box prompts often contain localization errors, and fixed threshold binarization introduces additional decision uncertainty. These factors jointly cause high prediction variance, especially near object boundaries, degrading reliability. We propose the Stability-Aware Inference Framework (SAIF), a training-free and plug-and-play inference framework that improves robustness by explicitly modeling prompt and threshold uncertainty. SAIF constructs a joint uncertainty space via structured box perturbations and threshold variations, evaluates each hypothesis using decision stability and boundary consistency, and introduces a stability-consistency score to filter unstable candidates and perform stability-weighted fusion in probability space. Experiments on Synapse, CVC-ClinicDB, Kvasir-SEG, and CVC-300 demonstrate that SAIF consistently improves segmentation accuracy and robustness, achieving state-of-the-art performance without retraining or architectural modification. Our anonymous code is released at https://anonymous.4open.science/r/SAIF.

Method: NumColor uses a Color Token Aggregator to detect color specifications regardless of tokenization, and a ColorBook with 6,707 learnable embeddings in perceptually uniform CIE Lab space. Two auxiliary losses (directional alignment and interpolation consistency) enforce geometric correspondence between Lab and embedding spaces. Trained on NumColor-Data, a synthetic dataset of 500K images with unambiguous color-to-pixel correspondence.

Result: NumColor improves numerical color accuracy by 4-9x across five diffusion models (SD3, SD3.5, PixArt-α, PixArt-Σ) and improves color harmony scores by 10-30x on GenColorBench benchmark. It transfers zero-shot to multiple architectures without model-specific adaptation.

Conclusion: NumColor successfully enables precise numerical color control in text-to-image diffusion models by addressing tokenization issues through learnable color embeddings in perceptually uniform space, with strong generalization across different model architectures.

Abstract: Text-to-image diffusion models excel at generating images from natural language descriptions, yet fail to interpret numerical colors such as hex codes (#FF5733) and RGB values (rgb(255,87,51)). This limitation stems from subword tokenization, which fragments color codes into semantically meaningless tokens that text encoders cannot map to coherent color representations. We present NumColor, that enables precise numerical color control across multiple diffusion architectures. NumColor comprises two components: a Color Token Aggregator that detects color specifications regardless of tokenization, and a ColorBook containing 6,707 learnable embeddings that map colors to embedding space of text encoder in perceptually uniform CIE Lab space. We introduce two auxiliary losses, directional alignment and interpolation consistency, to enforce geometric correspondence between Lab and embedding spaces, enabling smooth color interpolation. To train the ColorBook, we construct NumColor-Data, a synthetic dataset of 500K rendered images with unambiguous color-to-pixel correspondence, eliminating the annotation ambiguity inherent in photographic datasets. Although trained solely on FLUX, NumColor transfers zero-shot to SD3, SD3.5, PixArt-α, and PixArt-Σ without model-specific adaptation. NumColor improves numerical color accuracy by 4-9x across five models, while simultaneously improving color harmony scores by 10-30x on GenColorBench benchmark.

Method: Multi-task learning framework that jointly trains keypoint detection, keypoint description, and semantic segmentation, integrated with a deep matching module for enhanced feature correspondence.

Result: Produces semantically annotated 3D point clouds with improved structural detail and elevation information, enabling multi-level mapping with altitude estimation in parking structures.

Conclusion: Joint training with semantic cues leads to more consistent feature matching and enhanced 3D reconstruction, demonstrating the value of semantic awareness in feature extraction.

Abstract: Feature matching is a fundamental problem in computer vision with wide-ranging applications, including simultaneous localization and mapping (SLAM), image stitching, and 3D reconstruction. While recent advances in deep learning have improved keypoint detection and description, most approaches focus primarily on geometric attributes and often neglect higher-level semantic information. This work proposes a semantic-aware feature extraction framework that employs multi-task learning to jointly train keypoint detection, keypoint description, and semantic segmentation. The method is benchmarked against standard feature matching techniques and evaluated in the context of 3D reconstruction. To enhance feature correspondence, a deep matching module is integrated. The system is tested using input from a single monocular fisheye camera mounted on a vehicle and evaluated within a multi-floor parking structure. The proposed approach supports semantic 3D reconstruction with altitude estimation, capturing elevation changes and enabling multi-level mapping. Experimental results demonstrate that the method produces semantically annotated 3D point clouds with improved structural detail and elevation information, underscoring the effectiveness of joint training with semantic cues for more consistent feature matching and enhanced 3D reconstruction.

Method: The paper reviews current practices in veterinary pathology publications, identifying two main approaches: 1) Exclusive visual performance control (eyeballing predictions) with secondary validation metrics, and 2) Statistical performance control using hold-out test sets. The authors compare strengths/weaknesses of both methods and discuss considerations for rigorous statistical evaluation including metric selection, test dataset composition, ground truth quality, resampling methods, model comparison, and stability assessment.

Result: The review identifies that both visual and statistical evaluation methods have complementary strengths. Visual methods provide intuitive error analysis and qualitative insights, while statistical methods offer objective, quantitative performance measures. The paper concludes that a combination of both approaches provides the greatest insight into model performance and error sources.

Conclusion: For reliable DL-AIA applications in pathology, both visual and statistical performance assessment methods should be used together, as they offer complementary insights into model performance, generalization capability, and error sources, ensuring safer and more reliable deployment in clinical and research settings.

Abstract: Deep learning-based automated image analysis (DL-AIA) has been shown to outperform trained pathologists in tasks related to feature quantification. Related to these capacities the use of DL-AIA tools is currently extending from proof-of-principle studies to routine applications such as patient samples (diagnostic pathology), regulatory safety assessment (toxicologic pathology), and recurrent research tasks. To ensure that DL-AIA applications are safe and reliable, it is critical to conduct a thorough and objective generalization performance assessment (i.e., the ability of the algorithm to accurately predict patterns of interest) and possibly evaluate model robustness (i.e., the algorithm’s capacity to maintain predictive accuracy on images from different sources). In this article, we review the practices for performance assessment in veterinary pathology publications by which two approaches were identified: 1) Exclusive visual performance control (i.e. eyeballing of algorithmic predictions) plus validation of the models application utilizing secondary performance indices, and 2) Statistical performance control (alongside the other methods), which requires a dataset creation and separation of an hold-out test set prior to model training. This article compares the strengths and weaknesses of statistical and visual performance control methods. Furthermore, we discuss relevant considerations for rigorous statistical performance evaluation including metric selection, test dataset image composition, ground truth label quality, resampling methods such as bootstrapping, statistical comparison of multiple models, and evaluation of model stability. It is our conclusion that visual and statistical evaluation have complementary strength and a combination of both provides the greatest insight into the DL model’s performance and sources of error.

Method: Proposes a multi-VFM relational guidance framework using diverse VFMs as experts, with universal relational feature representation (local center-of-mass field) and spikiness-aware selection strategy to filter artifacts and aggregate reliable guidance.

Result: Achieves state-of-the-art performance across various downstream dense prediction tasks, demonstrating efficacy of multi-expert relational guidance.

Conclusion: DiveUp breaks away from single-model dependency and provides a unified, encoder-agnostic framework that can universally upsample features from diverse VFMs without per-model retraining.

Abstract: Recently, feature upsampling has gained increasing attention owing to its effectiveness in enhancing vision foundation models (VFMs) for pixel-level understanding tasks. Existing methods typically rely on high-resolution features from the same foundation model to achieve upsampling via self-reconstruction. However, relying solely on intra-model features forces the upsampler to overfit to the source model’s inherent location misalignment and high-norm artifacts. To address this fundamental limitation, we propose DiveUp, a novel framework that breaks away from single-model dependency by introducing multi-VFM relational guidance. Instead of naive feature fusion, DiveUp leverages diverse VFMs as a panel of experts, utilizing their structural consensus to regularize the upsampler’s learning process, effectively preventing the propagation of inaccurate spatial structures from the source model. To reconcile the unaligned feature spaces across different VFMs, we propose a universal relational feature representation, formulated as a local center-of-mass (COM) field, that extracts intrinsic geometric structures, enabling seamless cross-model interaction. Furthermore, we introduce a spikiness-aware selection strategy that evaluates the spatial reliability of each VFM, effectively filtering out high-norm artifacts to aggregate guidance from only the most reliable expert at each local region. DiveUp is a unified, encoder-agnostic framework; a jointly-trained model can universally upsample features from diverse VFMs without requiring per-model retraining. Extensive experiments demonstrate that DiveUp achieves state-of-the-art performance across various downstream dense prediction tasks, validating the efficacy of multi-expert relational guidance. Our code and models are available at: https://github.com/Xiaoqiong-Liu/DiveUp

Method: Proposes a weakly supervised deep learning pipeline that fuses Sentinel-1 SAR and AMSR-2 radiometry data using a U-Net architecture trained with region-based loss. Introduces Analytical Logit Scaling method applied post-inference, which dynamically calculates temperature and bias based on latent space percentiles (2% and 98%) of each scene to force physical binarization of predictions.

Result: The adaptive scaling acts as a topological extractor, successfully revealing fine-grained sea ice fractures (leads) at 40-meter resolution without requiring manual pixel-level annotations. Achieves 78% accuracy on highly fragmented summer scenes, resolving local topology while preserving regional macroscopic concentrations.

Conclusion: The approach bridges the gap between weakly supervised learning and high-resolution physical segmentation for sea ice mapping, enabling fine-grained feature extraction from weak labels through adaptive post-processing scaling.

Abstract: High-resolution sea ice mapping using Synthetic Aperture Radar (SAR) is crucial for Arctic navigation and climate monitoring. However, operational ice charts provide only coarse, region-level polygons (weak labels), forcing automated segmentation models to struggle with pixel-level accuracy and often yielding under-confident, blurred concentration maps. In this paper, we propose a weakly supervised deep learning pipeline that fuses Sentinel-1 SAR and AMSR-2 radiometry data using a U-Net architecture trained with a region-based loss. To overcome the severe under-confidence caused by weak labels, we introduce an Analytical Logit Scaling method applied post-inference. By dynamically calculating the temperature and bias based on the latent space percentiles (2% and 98%) of each scene, we force a physical binarization of the predictions. This adaptive scaling acts as a topological extractor, successfully revealing fine-grained sea ice fractures (leads) at a 40-meter resolution without requiring any manual pixel-level annotations. Our approach not only resolves local topology but also perfectly preserves regional macroscopic concentrations, achieving a 78% accuracy on highly fragmented summer scenes, thereby bridging the gap between weakly supervised learning and high-resolution physical segmentation.

Method: Proposes LingoMotion language with: 1) Motion alphabet based on joint angles, 2) Morphology for forming words/phrases to describe simple actions and attributes, 3) Syntax for describing complex activities with sequences. Implemented and evaluated using Motion-X dataset.

Result: Preliminary results demonstrate high fidelity of motion representation using the Motion-X dataset, showing the proposed symbolic language can effectively represent human motion.

Conclusion: LingoMotion provides an interpretable, hierarchical symbolic representation for human motion that addresses limitations of existing black-box approaches, enabling better understanding and manipulation of motion data.

Abstract: Existing representations for human motion, such as MotionGPT, often operate as black-box latent vectors with limited interpretability and build on joint positions which can cause ambiguity. Inspired by the hierarchical structure of natural languages - from letters to words, phrases, and sentences - we propose LingoMotion, a motion language that facilitates interpretable and unambiguous symbolic representation for both simple and complex human motion. In this paper, we introduce the concept design of LingoMotion, including the definitions of motion alphabet based on joint angles, the morphology for forming words and phrases to describe simple actions like walking and their attributes like speed and scale, as well as the syntax for describing more complex human activities with sequences of words and phrases. The preliminary results, including the implementation and evaluation of motion alphabet using a large-scale motion dataset Motion-X, demonstrate the high fidelity of motion representation.

Method: Three-stage pipeline: 1) Train independent encoders for each modality (localizer MRI, ECG, tabular data), 2) Fuse pre-trained encoders to unify latent space, 3) Use enriched representation for cardiac phenotype prediction, with inference only on localizer MRI.

Result: C-TRIP yields accurate functional cardiac phenotypes and high correlations for structural phenotypes, demonstrating that localizer MRI combined with other modalities can effectively predict cardiac health metrics.

Conclusion: The framework enables better accessibility for cardiac phenotype estimation by leveraging inherently rapid and low-cost localizer MRI combined with ECG and patient metadata, providing an opportunistic alternative to expensive cine cardiac MR.

Abstract: Cardiovascular diseases are the leading cause of death. Cardiac phenotypes (CPs), e.g., ejection fraction, are the gold standard for assessing cardiac health, but they are derived from cine cardiac magnetic resonance imaging (CMR), which is costly and requires high spatio-temporal resolution. Every magnetic resonance (MR) examination begins with rapid and coarse localizers for scan planning, which are discarded thereafter. Despite non-diagnostic image quality and lack of temporal information, localizers can provide valuable structural information rapidly. In addition to imaging, patient-level information, including demographics and lifestyle, influence the cardiac health assessment. Electrocardiograms (ECGs) are inexpensive, routinely ordered in clinical practice, and capture the temporal activity of the heart. Here, we introduce C-TRIP (Cardiac Tri-modal Representations for Imaging Phenotypes), a multi-modal framework that aligns localizer MRI, ECG signals, and tabular metadata to learn a robust latent space and predict CPs using localizer images as an opportunistic alternative to CMR. By combining these three modalities, we leverage cheap spatial and temporal information from localizers, and ECG, respectively while benefiting from patient-specific context provided by tabular data. Our pipeline consists of three stages. First, encoders are trained independently to learn uni-modal representations. The second stage fuses the pre-trained encoders to unify the latent space. The final stage uses the enriched representation space for CP prediction, with inference performed exclusively on localizer MRI. Proposed C-TRIP yields accurate functional CPs, and high correlations for structural CPs. Since localizers are inherently rapid and low-cost, our C-TRIP framework could enable better accessibility for CP estimation.

Method: Developed a reproducible data-processing pipeline converting e-scooter trip records into hourly pickup/dropoff demand images. Used correlation- and error-based procedure to identify informative historical inputs, with optimal temporal depth selected through ablation study using UNET model with statistical tests and Holm correction.

Result: The proposed temporal structure design reduces mean squared error by up to 37% for next-hour prediction and 35% for next-24-hour prediction compared to baseline approaches, capturing short-term persistence and daily/weekly cycles.

Conclusion: Principled dataset construction and statistically validated temporal input design significantly improve spatiotemporal micromobility demand prediction performance, highlighting the value of systematic approaches over heuristic methods.

Abstract: Despite progress in deep learning for shared micromobility demand prediction, the systematic design and statistical validation of temporal input structures remain underexplored. Temporal features are often selected heuristically, even though historical demand strongly affects model performance and generalizability. This paper introduces a reproducible data-processing pipeline and a statistically grounded method for designing temporal input structures for image-to-image demand prediction. Using large-scale e-scooter data from Austin, Texas, we build a grid-based spatiotemporal dataset by converting trip records into hourly pickup and dropoff demand images. The pipeline includes trip filtering, mapping Census Tracts to spatial locations, grid construction, demand aggregation, and creation of a global activity mask that limits evaluation to historically active areas. This representation supports consistent spatial learning while preserving demand patterns. We then introduce a combined correlation- and error-based procedure to identify informative historical inputs. Optimal temporal depth is selected through an ablation study using a baseline UNET model with paired non-parametric tests and Holm correction. The resulting temporal structures capture short-term persistence as well as daily and weekly cycles. Compared with adjacent-hour and fixed-period baselines, the proposed design reduces mean squared error by up to 37 percent for next-hour prediction and 35 percent for next-24-hour prediction. These results highlight the value of principled dataset construction and statistically validated temporal input design for spatiotemporal micromobility demand prediction.

Method: EgoHOI distills geometric and kinematic priors from 3D estimates into physics-informed embeddings that regularize egocentric rollouts toward physically valid dynamics, enabling simulation from action signals alone without future-state inputs.

Result: Experiments on HOT3D dataset show consistent gains over strong baselines, with ablations validating the effectiveness of the physics-informed design.

Conclusion: EgoHOI represents a significant step toward true world simulators for embodied AI that can infer interaction dynamics strictly from user actions, rather than relying on conditional video generation with privileged information.

Abstract: To serve as a scalable data source for embodied AI, world models should act as true simulators that infer interaction dynamics strictly from user actions, rather than mere conditional video generators relying on privileged future object states. In this context, egocentric Human-Object Interaction (HOI) world models are critical for predicting physically grounded first-person rollouts. However, building such models is profoundly challenging due to rapid head motions, severe occlusions, and high-DoF hand articulations that abruptly alter contact topologies. Consequently, existing approaches often circumvent these physics challenges by resorting to conditional video generation with access to known future object trajectories. We introduce EgoHOI, an egocentric HOI world model that breaks away from this shortcut to simulate photorealistic, contact-consistent interactions from action signals alone. To ensure physical accuracy without future-state inputs, EgoHOI distills geometric and kinematic priors from 3D estimates into physics-informed embeddings. These embeddings regularize the egocentric rollouts toward physically valid dynamics. Experiments on the HOT3D dataset demonstrate consistent gains over strong baselines, and ablations validate the effectiveness of our physics-informed design.

Method: 1) Introduces Optimized Locatability Score to quantify image suitability for deep reasoning; 2) Creates Geo-ADAPT-51K dataset with locatability-stratified reasoning trajectories; 3) Proposes two-stage Group Relative Policy Optimization curriculum with customized reward functions for adaptive reasoning depth, visual grounding, and hierarchical geographical accuracy.

Result: Achieves state-of-the-art performance across multiple geo-localization benchmarks, substantially reduces hallucinations, and enables both adaptive and efficient reasoning.

Conclusion: Geo-ADAPT successfully addresses limitations of existing geo-localization methods by learning adaptive reasoning policies that improve accuracy while reducing hallucinations through optimized locatability assessment and policy optimization.

Abstract: The emergence of Vision-Language Models (VLMs) has introduced new paradigms for global image geo-localization through retrieval-augmented generation (RAG) and reasoning-driven inference. However, RAG methods are constrained by retrieval database quality, while reasoning-driven approaches fail to internalize image locatability, relying on inefficient, fixed-depth reasoning paths that increase hallucinations and degrade accuracy. To overcome these limitations, we introduce an Optimized Locatability Score that quantifies an image’s suitability for deep reasoning in geo-localization. Using this metric, we curate Geo-ADAPT-51K, a locatability-stratified reasoning dataset enriched with augmented reasoning trajectories for complex visual scenes. Building on this foundation, we propose a two-stage Group Relative Policy Optimization (GRPO) curriculum with customized reward functions that regulate adaptive reasoning depth, visual grounding, and hierarchical geographical accuracy. Our framework, Geo-ADAPT, learns an adaptive reasoning policy, achieves state-of-the-art performance across multiple geo-localization benchmarks, and substantially reduces hallucinations by reasoning both adaptively and efficiently.

Method: CAAP estimates patch contributions by directly intervening on internal activations rather than using learned masks or synthetic perturbations. For each patch, it inserts source-image activations into a neutral target context over intermediate layers and uses the resulting target-class score as the attribution signal.

Result: CAAP significantly outperforms existing methods across multiple ViT backbones and standard metrics, producing more faithful and localized attributions that better reflect the causal effect of patch-associated internal representations.

Conclusion: CAAP provides a principled causal intervention approach for ViT attribution that captures class-relevant evidence after initial representation formation while avoiding late-layer global mixing, resulting in superior spatial specificity and faithfulness.

Abstract: Attribution methods for Vision Transformers (ViTs) aim to identify image regions that influence model predictions, but producing faithful and well-localized attributions remains challenging. Existing gradient-based and perturbation-based techniques often fail to isolate the causal contribution of internal representations associated with individual image patches. The key challenge is that class-relevant evidence is formed through interactions between patch tokens across layers, and input-level perturbations can be poor proxies for patch importance, since they may fail to reconstruct the internal evidence actually used by the model. We propose Causal Attribution via Activation Patching (CAAP), which estimates the contribution of individual image patches to the ViT’s prediction by directly intervening on internal activations rather than using learned masks or synthetic perturbation patterns. For each patch, CAAP inserts the corresponding source-image activations into a neutral target context over an intermediate range of layers and uses the resulting target-class score as the attribution signal. The resulting attribution map reflects the causal effect of patch-associated internal representations on the model’s prediction. The causal intervention serves as a principled measure of patch influence by capturing class-relevant evidence after initial representation formation, while avoiding late-layer global mixing that can reduce spatial specificity. Across multiple ViT backbones and standard metrics, CAAP significantly outperforms existing methods and produces more faithful and localized attributions.

Method: Two-module framework: (1) SegFlow - small segmentation model that recasts prediction as continuous image→mask transport using flow-matching regression loss and ODE integration for trajectory-level supervision. (2) SynFlow - mask-conditioned mask→image generator that produces pixel-aligned synthetic image-mask pairs with controllable structural variations and edge-aware gating.

Result: On five crack/vessel benchmarks: SegFlow alone improves mean IoU from 0.511 to 0.599 (+17.2%) and reduces Betti matching error from 82.145 to 51.524 (-37.3%). With limited labels, SynFlow augmentation recovers near-full performance using 25% of real annotations and improves cross-domain IoU by 0.11 on average.

Conclusion: FMS² provides integrated solution for thin-structure segmentation with improved topology preservation and generalization. Unlike classical data augmentation, SynFlow offers controllable structural shifts for better domain adaptation, making it effective for label-scarce scenarios.

Abstract: Segmenting thin structures like infrastructure cracks and anatomical vessels is a task hampered by topology-sensitive geometry, high annotation costs, and poor generalization across domains. Existing methods address these challenges in isolation. We propose FMS$^2$, a flow-matching framework with two modules. (1) SegFlow is a 2.96M-parameter segmentation model built on a standard encoder-decoder backbone that recasts prediction as continuous image $\rightarrow$ mask transport. It learns a time-indexed velocity field with a flow-matching regression loss and outputs the mask via ODE integration, rather than supervising only end-state logits. This trajectory-level supervision improves thin-structure continuity and sharpness, compared with tuned topology-aware loss baselines, without auxiliary topology heads, post-processing, or multi-term loss engineering. (2) SynFlow is a mask-conditioned mask $\rightarrow$ image generator that produces pixel-aligned synthetic image-mask pairs. It injects mask geometry at multiple scales and emphasizes boundary bands via edge-aware gating, while a controllable mask generator expands sparsity, width, and branching regimes. On five crack and vessel benchmarks, SegFlow alone outperforms strong CNN, Transformer, Mamba, and generative baselines, improving the volumetric metric (mean IoU) from 0.511 to 0.599 (+17.2%) and reducing the topological metric (Betti matching error) from 82.145 to 51.524 (-37.3%). When training with limited labels, augmenting SegFlow with SynFlow-generated pairs recovers near-full performance using 25% of real annotations and improves cross-domain IoU by 0.11 on average. Unlike classical data augmentation that promotes invariance via label-preserving transforms, SynFlow provides pixel-aligned paired supervision with controllable structural shifts (e.g., sparsity, width, branching), which is particularly effective under domain shift.

Method: MASS treats in-context segmentation as the pretext task, using automatically generated class-agnostic masks as structural supervision. It trains on thousands of diverse mask proposals spanning anatomical structures and pathological findings to learn holistic combinations of appearance, shape, spatial context, and anatomical relationships.

Result: MASS demonstrates effectiveness across data regimes: few-shot segmentation on novel structures, matching full supervision with only 20-40% labeled data while outperforming self-supervised baselines by over 20 Dice score in low-data regimes, and frozen-encoder classification on unseen pathologies matching full supervised training with thousands of samples.

Conclusion: Mask-guided self-supervised pretraining captures broadly generalizable knowledge, opening a path toward 3D medical imaging foundation models without expert annotations.

Abstract: Foundation models have transformed vision and language by learning general-purpose representations from large-scale unlabeled data, yet 3D medical imaging lacks analogous approaches. Existing self-supervised methods rely on low-level reconstruction or contrastive objectives that fail to capture the anatomical semantics critical for medical image analysis, limiting transfer to downstream tasks. We present MASS (MAsk-guided Self-Supervised learning), which treats in-context segmentation as the pretext task for learning general-purpose medical imaging representations. MASS’s key insight is that automatically generated class-agnostic masks provide sufficient structural supervision for learning semantically rich representations. By training on thousands of diverse mask proposals spanning anatomical structures and pathological findings, MASS learns what semantically defines medical structures: the holistic combination of appearance, shape, spatial context, and anatomical relationships. We demonstrate effectiveness across data regimes: from small-scale pretraining on individual datasets (20-200 scans) to large-scale multi-modal pretraining on 5K CT, MRI, and PET volumes, all without annotations. MASS demonstrates: (i) few-shot segmentation on novel structures, (ii) matching full supervision with only 20-40% labeled data while outperforming self-supervised baselines by over 20 in Dice score in low-data regimes, and (iii) frozen-encoder classification on unseen pathologies that matches full supervised training with thousands of samples. Mask-guided self-supervised pretraining captures broadly generalizable knowledge, opening a path toward 3D medical imaging foundation models without expert annotations. Code is available: https://github.com/Stanford-AIMI/MASS.

Method: TSDCRF combines three components: 1) (ε,δ)-differential privacy via calibrated Gaussian noise on sensitive regions, 2) Normalized Control Penalty (NCP) that down-weights unstable class predictions before noise injection, and 3) time-series dynamic conditional random field (DCRF) that enforces temporal consistency and corrects trajectory deviation after noise.

Result: Evaluation on MOT16, MOT17, Cityscapes, and KITTI shows TSDCRF achieves better privacy-utility trade-off than white noise and prior methods (NTPD, PPDTSA) with lower KL-divergence shift, lower tracking RMSE, and improved robustness under trajectory hijacking while preserving privacy.

Conclusion: TSDCRF provides an effective plug-in refinement framework that balances privacy and tracking performance, being agnostic to detector/tracker choices and demonstrating practical utility in real-world tracking scenarios.

Abstract: Multi-object tracking in video often requires appearance or location cues that can reveal sensitive identity information, while adding privacy-preserving noise typically disrupts cross-frame association and causes ID switches or target loss. We propose TSDCRF, a plug-in refinement framework that balances privacy and tracking by combining three components: (i) $(\varepsilon,δ)$-differential privacy via calibrated Gaussian noise on sensitive regions under a configurable privacy budget; (ii) a Normalized Control Penalty (NCP) that down-weights unstable or conflicting class predictions before noise injection to stabilize association; and (iii) a time-series dynamic conditional random field (DCRF) that enforces temporal consistency and corrects trajectory deviation after noise, mitigating ID switches and resilience to trajectory hijacking. The pipeline is agnostic to the choice of detector and tracker (e.g., YOLOv4 and DeepSORT). We evaluate on MOT16, MOT17, Cityscapes, and KITTI. Results show that TSDCRF achieves a better privacy–utility trade-off than white noise and prior methods (NTPD, PPDTSA): lower KL-divergence shift, lower tracking RMSE, and improved robustness under trajectory hijacking while preserving privacy. Source code in https://github.com/mabo1215/TSDCRF.git

Method: Proposes SHAMISA with compositional distortion engine generating continuous degradations, dual-source relation graphs encoding degradation profiles and structural affinities, and non-contrastive self-supervised learning with implicit structural associations.

Result: Achieves strong performance on synthetic, authentic, and cross-dataset NR-IQA benchmarks with improved cross-dataset generalization and robustness, without human quality annotations.

Conclusion: SHAMISA demonstrates effective self-supervised learning for NR-IQA by leveraging structured relational supervision from synthetic distortions, offering a promising alternative to human-annotated approaches.

Abstract: No-Reference Image Quality Assessment (NR-IQA) aims to estimate perceptual quality without access to a reference image of pristine quality. Learning an NR-IQA model faces a fundamental bottleneck: its need for a large number of costly human perceptual labels. We propose SHAMISA, a non-contrastive self-supervised framework that learns from unlabeled distorted images by leveraging explicitly structured relational supervision. Unlike prior methods that impose rigid, binary similarity constraints, SHAMISA introduces implicit structural associations, defined as soft, controllable relations that are both distortion-aware and content-sensitive, inferred from synthetic metadata and intrinsic feature structure. A key innovation is our compositional distortion engine, which generates an uncountable family of degradations from continuous parameter spaces, grouped so that only one distortion factor varies at a time. This enables fine-grained control over representational similarity during training: images with shared distortion patterns are pulled together in the embedding space, while severity variations produce structured, predictable shifts. We integrate these insights via dual-source relation graphs that encode both known degradation profiles and emergent structural affinities to guide the learning process throughout training. A convolutional encoder is trained under this supervision and then frozen for inference, with quality prediction performed by a linear regressor on its features. Extensive experiments on synthetic, authentic, and cross-dataset NR-IQA benchmarks demonstrate that SHAMISA achieves strong overall performance with improved cross-dataset generalization and robustness, all without human quality annotations or contrastive losses.

Method: Proposes a mistake-severity-aware training strategy that organizes diagnostic classes hierarchically, uses severity-weighted cross-entropy loss to penalize high-severity misclassifications more strongly, and enforces hierarchical consistency through probabilistic alignment and semantic feature remix.

Result: Experiments on public and real-world datasets demonstrate significant mitigation of critical errors in MIL diagnosis compared to existing methods, with additional results on natural domain data showing generalizability beyond medical contexts.

Conclusion: The proposed hierarchical MIL framework effectively addresses clinically critical errors in WSI diagnosis by incorporating diagnostic priorities and severity awareness, with potential applications beyond medical domains.

Abstract: Multiple Instance Learning (MIL) has emerged as a promising paradigm for Whole Slide Image (WSI) diagnosis, offering effective learning with limited annotations. However, existing MIL frameworks overlook diagnostic priorities and fail to differentiate the severity of misclassifications in multiclass, leaving clinically critical errors unaddressed. We propose a mistake-severity-aware training strategy that organizes diagnostic classes into a hierarchical structure, with each level optimized using a severity-weighted cross-entropy loss that penalizes high-severity misclassifications more strongly. Additionally, hierarchical consistency is enforced through probabilistic alignment, a semantic feature remix applied to the instance bag to robustly train class priority and accommodate clinical cases involving multiple symptoms. An asymmetric Mikel’s Wheel-based metric is also introduced to quantify the severity of errors specific to medical fields. Experiments on challenging public and real-world in-house datasets demonstrate that our approach significantly mitigates critical errors in MIL diagnosis compared to existing methods. We present additional experimental results on natural domain data to demonstrate the generalizability of our proposed method beyond medical contexts.

Method: Unified framework adapting pretrained diffusion models (U-Net and DiT) via channel concatenation and in-context token concatenation, trained on 60k+ bi-temporal RS image pairs to learn physically coherent edits.

Result: Clear gains over general and commercial baselines, strong generalizability across disaster impacts, urban growth, and seasonal shifts, serving as robust data engine for downstream analysis.

Conclusion: RSEdit enables precise, physically coherent remote sensing image editing while preserving geospatial content, positioning it as a valuable tool for Earth observation applications.

Abstract: General-domain text-guided image editors achieve strong photorealism but introduce artifacts, hallucinate objects, and break the orthographic constraints of remote sensing (RS) imagery. We trace this gap to two high-level causes: (i) limited RS world knowledge in pre-trained models, and (ii) conditioning schemes that misalign with the bi-temporal structure and spatial priors of Earth observation data. We present RSEdit, a unified framework that adapts pretrained text-to-image diffusion models - both U-Net and DiT - into instruction-following RS editors via channel concatenation and in-context token concatenation. Trained on over 60,000 semantically rich bi-temporal remote sensing image pairs, RSEdit learns precise, physically coherent edits while preserving geospatial content. Experiments show clear gains over general and commercial baselines, demonstrating strong generalizability across diverse scenarios including disaster impacts, urban growth, and seasonal shifts, positioning RSEdit as a robust data engine for downstream analysis. We will release code, pretrained models, evaluation protocols, training logs, and generated results for full reproducibility. Code: https://github.com/Bili-Sakura/RSEdit-Preview

Method: Proposes SDMoEA framework with two key components: 1) Sparse-Dense Mixture of Experts Adapter (SDMoE) that combines sparse MoE for modality-specific information and dense-shared MoE for cross-modal shared information, and 2) Gram-based Semantic Alignment Hypergraph Fusion (GSAHF) module that uses Gram matrices for semantic alignment and hypergraph structures for high-order relationship modeling.

Result: Achieves superior performance compared to other PEFT approaches on multiple multimodal tracking benchmarks including LasHeR, RGBT234, VTUAV, VisEvent, COESOT, DepthTrack, and VOT-RGBD2022.

Conclusion: The SDMoEA framework effectively addresses cross-modal heterogeneity in PEFT-based multimodal tracking by combining sparse-dense mixture of experts with hypergraph-based fusion, providing an efficient and unified solution for multimodal representation learning.

Abstract: Parameter-efficient fine-tuning (PEFT) techniques, such as prompts and adapters, are widely used in multi-modal tracking because they alleviate issues of full-model fine-tuning, including time inefficiency, high resource consumption, parameter storage burden, and catastrophic forgetting. However, due to cross-modal heterogeneity, most existing PEFT-based methods struggle to effectively represent multi-modal features within a unified framework with shared parameters. To address this problem, we propose a novel Sparse-Dense Mixture of Experts Adapter (SDMoEA) framework for PEFT-based multi-modal tracking under a unified model structure. Specifically, we design an SDMoE module as the multi-modal adapter to model modality-specific and shared information efficiently. SDMoE consists of a sparse MoE and a dense-shared MoE: the former captures modality-specific information, while the latter models shared cross-modal information. Furthermore, to overcome limitations of existing tracking methods in modeling high-order correlations during multi-level multi-modal fusion, we introduce a Gram-based Semantic Alignment Hypergraph Fusion (GSAHF) module. It first employs Gram matrices for cross-modal semantic alignment, ensuring that the constructed hypergraph accurately reflects semantic similarity and high-order dependencies between modalities. The aligned features are then integrated into the hypergraph structure to exploit its ability to model high-order relationships, enabling deep fusion of multi-level multi-modal information. Extensive experiments demonstrate that the proposed method achieves superior performance compared with other PEFT approaches on several multi-modal tracking benchmarks, including LasHeR, RGBT234, VTUAV, VisEvent, COESOT, DepthTrack, and VOT-RGBD2022.

Method: The framework: (1) links sensitive concepts to layer-wise grouping via NCP and MDAV-based clustering; (2) locates sensitive feature regions using bottom-up (BUA) and top-down (TDA) strategies over multi-scale representations; (3) uses an Expectation-Maximization Privacy Assessment (EMPA) module to produce interpretable budget-alignment signals by comparing fitted sensitive-feature distributions to evaluator-specified references.

Result: Validated on object detectors (YOLO, PPDPTS, DETR) and visual encoders of VLMs (CLIP, LLaVA, BLIP). BUA and TDA yield comparable deviation trends; EMPA provides stable alignment signals. Compared favorably with generic discrepancy baselines (Chi-square, K-L, MMD) and task-relevant baselines (MomentReg, NoiseMLE, Wass-1).

Conclusion: The work contributes a learnable, interpretable modeling perspective for privacy-aligned hierarchical representations rather than just post hoc audit, providing a framework applicable across different vision architectures and VLMs.

Abstract: Learning systems that preserve privacy often inject noise into hierarchical visual representations; a central challenge is to \emph{model} how such perturbations align with a declared privacy budget in a way that is interpretable and applicable across vision backbones and vision–language models (VLMs). We propose \emph{Bodhi VLM}, a \emph{privacy-alignment modeling} framework for \emph{hierarchical neural representations}: it (1) links sensitive concepts to layer-wise grouping via NCP and MDAV-based clustering; (2) locates sensitive feature regions using bottom-up (BUA) and top-down (TDA) strategies over multi-scale representations (e.g., feature pyramids or vision-encoder layers); and (3) uses an Expectation-Maximization Privacy Assessment (EMPA) module to produce an interpretable \emph{budget-alignment signal} by comparing the fitted sensitive-feature distribution to an evaluator-specified reference (e.g., Laplace or Gaussian with scale $c/ε$). The output is reference-relative and is \emph{not} a formal differential-privacy estimator. We formalize BUA/TDA over hierarchical feature structures and validate the framework on object detectors (YOLO, PPDPTS, DETR) and on the \emph{visual encoders} of VLMs (CLIP, LLaVA, BLIP). BUA and TDA yield comparable deviation trends; EMPA provides a stable alignment signal under the reported setups. We compare with generic discrepancy baselines (Chi-square, K-L, MMD) and with task-relevant baselines (MomentReg, NoiseMLE, Wass-1). Results are reported as mean$\pm$std over multiple seeds with confidence intervals in the supplementary materials. This work contributes a learnable, interpretable modeling perspective for privacy-aligned hierarchical representations rather than a post hoc audit only. Source code: \href{https://github.com/mabo1215/bodhi-vlm.git}{Bodhi-VLM GitHub repository}

Method: Created Sky2Ground dataset with 51 sites containing thousands of synthetic and real images across three altitude levels. Proposed SkyNet model with curriculum-based training to progressively incorporate satellite views for better cross-view consistency.

Result: Benchmarking showed satellite imagery degrades existing methods’ performance. SkyNet outperforms state-of-the-art by 9.6% on RRA@5 and 18.1% on RTA@5, improving multi-view alignment.

Conclusion: Sky2Ground provides comprehensive testbed for multi-altitude 3D perception, revealing challenges of large altitude variations. SkyNet demonstrates effective curriculum training for cross-view consistency.

Abstract: We introduce Sky2Ground, a three-view dataset designed for varying altitude camera localization, correspondence learning, and reconstruction. The dataset combines structured synthetic imagery with real, in-the-wild images, providing both controlled multi-view geometry and realistic scene noise. Each of the 51 sites contains thousands of satellite, aerial, and ground images spanning wide altitude ranges and nearly orthogonal viewing angles, enabling rigorous evaluation across global-to-local contexts. We benchmark state of the art pose estimation models, including MASt3R, DUSt3R, Map Anything, and VGGT, and observe that the use of satellite imagery often degrades performance, highlighting the challenges under large altitude variations. We also examine reconstruction methods, highlighting the challenges introduced by sparse geometric overlap, varying perspectives, and the use of real imagery, which often introduces noise and reduces rendering quality. To address some of these challenges, we propose SkyNet, a model which enhances cross-view consistency when incorporating satellite imagery with a curriculum-based training strategy to progressively incorporate more satellite views. SkyNet significantly strengthens multi-view alignment and outperforms existing methods by 9.6% on RRA@5 and 18.1% on RTA@5 in terms of absolute performance. Sky2Ground and SkyNet together establish a comprehensive testbed and baseline for advancing large-scale, multi-altitude 3D perception and generalizable camera localization. Code and models will be released publicly for future research.

Method: Created a custom rig with 12 synchronized cameras surrounding a 4-camera VR headset, captured nearly 1,000 short egocentric videos focusing on hand motions and hand-object interactions in various settings, with detailed rig design, data processing, and calibration procedures.

Result: The dataset presents unique challenges for existing 3D and 4D novel view synthesis methods due to large disparities and image motion caused by close dynamic objects and rig egomotion, enabling new ways to benchmark egocentric scene reconstruction methods.

Conclusion: Ego-1K supports future research in neural 3D video synthesis and dynamic scene understanding, particularly for egocentric applications as smart glasses with multiple cameras become omnipresent.

Abstract: We present Ego-1K, a large-scale collection of time-synchronized egocentric multiview videos designed to advance neural 3D video synthesis and dynamic scene understanding. The dataset contains nearly 1,000 short egocentric videos captured with a custom rig with 12 synchronized cameras surrounding a 4-camera VR headset worn by the user. Scene content focuses on hand motions and hand-object interactions in different settings. We describe rig design, data processing, and calibration. Our dataset enables new ways to benchmark egocentric scene reconstruction methods, an important research area as smart glasses with multiple cameras become omnipresent. Our experiments demonstrate that our dataset presents unique challenges for existing 3D and 4D novel view synthesis methods due to large disparities and image motion caused by close dynamic objects and rig egomotion. Our dataset supports future research in this challenging domain. It is available at https://huggingface.co/datasets/facebook/ego-1k.

Method: CreativeAds uses a three-module pipeline: product pairing, layout generation, and background generation, with an intuitive UI for oversight and customized adjustments.

Result: The system enables scalable generation of high-quality ad images without requiring GenAI expertise, as demonstrated through user studies and evaluations.

Conclusion: CreativeAds addresses key challenges in using GenAI for e-commerce advertising at scale, providing automated yet customizable generation of lifestyle product images.

Abstract: Lifestyle images are photographs that capture environments and objects in everyday settings. In furniture product marketing, advertisers often create lifestyle images containing products to resonate with potential buyers, allowing buyers to visualize how the products fit into their daily lives. While recent advances in Generative Artificial Intelligence (GenAI) have given rise to realistic image content creation, their application in e-commerce advertising is challenging because high-quality ads must authentically representing the products in realistic scearios. Therefore, manual intervention is usually required for individual generations, making it difficult to scale to larger product catalogs. To understand the challenges faced by advertisers using GenAI to create lifestyle images at scale, we conducted evaluations on ad images generated using state-of-the-art image generation models and identified the major challenges. Based on our findings, we present CreativeAds, a multi-product ad creation system that supports scalable automated generation with customized parameter adjustment for individual generation. To ensure automated high-quality ad generation, CreativeAds innovates a pipeline that consists of three modules to address challenges in product pairing, layout generation, and background generation separately. Furthermore, CreativeAds contains an intuitive user interface to allow users to oversee generation at scale, and it also supports detailed controls on individual generation for user customized adjustments. We performed a user study on CreativeAds and extensive evaluations of the generated images, demonstrating CreativeAds’s ability to create large number of high-quality images at scale for advertisers without requiring expertise in GenAI tools.

Method: QTrack is an end-to-end vision-language model integrating multimodal reasoning with tracking-oriented localization, using Temporal Perception-Aware Policy Optimization with structured rewards for motion-aware reasoning.

Result: Extensive experiments demonstrate effectiveness for reasoning-centric, language-guided tracking, with the RMOT26 benchmark enabling robust evaluation of generalization.

Conclusion: Query-driven tracking with natural language queries represents a promising paradigm shift in MOT, enabling more flexible, user-centric tracking applications.

Abstract: Multi-object tracking (MOT) has traditionally focused on estimating trajectories of all objects in a video, without selectively reasoning about user-specified targets under semantic instructions. In this work, we introduce a query-driven tracking paradigm that formulates tracking as a spatiotemporal reasoning problem conditioned on natural language queries. Given a reference frame, a video sequence, and a textual query, the goal is to localize and track only the target(s) specified in the query while maintaining temporal coherence and identity consistency. To support this setting, we construct RMOT26, a large-scale benchmark with grounded queries and sequence-level splits to prevent identity leakage and enable robust evaluation of generalization. We further present QTrack, an end-to-end vision-language model that integrates multimodal reasoning with tracking-oriented localization. Additionally, we introduce a Temporal Perception-Aware Policy Optimization strategy with structured rewards to encourage motion-aware reasoning. Extensive experiments demonstrate the effectiveness of our approach for reasoning-centric, language-guided tracking. Code and data are available at https://github.com/gaash-lab/QTrack

Method: 1) Constructs a controllable synthetic data generation pipeline using rigid-body simulation to create a curated dataset with accurate physics and 3D annotations. 2) Builds a unified physical latent space by coupling explicit 3D geometry constraints with Gram-based spatio-temporal relational alignment to extract kinematic priors from video foundation models.

Result: PhysAlign significantly outperforms existing VDMs on tasks requiring complex physical reasoning and temporal stability, without compromising zero-shot visual quality. It establishes a practical paradigm for physics-grounded video generation.

Conclusion: PhysAlign bridges the gap between raw visual synthesis and rigid-body kinematics, offering an efficient framework for physics-coherent image-to-video generation with improved temporal stability and physical realism.

Abstract: Video Diffusion Models (VDMs) offer a promising approach for simulating dynamic scenes and environments, with broad applications in robotics and media generation. However, existing models often generate temporally incoherent content that violates basic physical intuition, significantly limiting their practical applicability. We propose PhysAlign, an efficient framework for physics-coherent image-to-video (I2V) generation that explicitly addresses this limitation. To overcome the critical scarcity of physics-annotated videos, we first construct a fully controllable synthetic data generation pipeline based on rigid-body simulation, yielding a highly-curated dataset with accurate, fine-grained physics and 3D annotations. Leveraging this data, PhysAlign constructs a unified physical latent space by coupling explicit 3D geometry constraints with a Gram-based spatio-temporal relational alignment that extracts kinematic priors from video foundation models. Extensive experiments demonstrate that PhysAlign significantly outperforms existing VDMs on tasks requiring complex physical reasoning and temporal stability, without compromising zero-shot visual quality. PhysAlign shows the potential to bridge the gap between raw visual synthesis and rigid-body kinematics, establishing a practical paradigm for genuinely physics-grounded video generation. The project page is available at https://physalign.github.io/PhysAlign.

Method: Apply Topological Data Analysis (TDA) with persistent homology to 3D FLAIR MRI volumes from BraTS 2020. Extract 100 topological features (Betti numbers) capturing connected components, loops, and voids. Use these features to train classical ML classifiers like Random Forest and XGBoost for binary HGG/LGG classification.

Result: Random Forest classifier with selected Betti features achieves 89.19% accuracy on BraTS 2020 dataset. The topological approach provides dimensionality reduction from complex 3D volumes to 100 interpretable features.

Conclusion: Persistent homology offers an effective, interpretable alternative to deep learning for 3D medical image analysis, demonstrating potential for brain tumor classification with computational efficiency and reduced data requirements.

Abstract: Accurate and interpretable brain tumor classification from medical imaging remains a challenging problem due to the high dimensionality and complex structural patterns present in magnetic resonance imaging (MRI). In this study, we propose a topology-driven framework for brain tumor classification based on Topological Data Analysis (TDA) applied directly to three-dimensional (3D) MRI volumes. Specifically, we analyze 3D Fluid Attenuated Inversion Recovery (FLAIR) images from the BraTS 2020 dataset and extract interpretable topological descriptors using persistent homology. Persistent homology captures intrinsic geometric and structural characteristics of the data through Betti numbers, which describe connected components (Betti-0), loops (Betti-1), and voids (Betti-2). From the 3D MRI volumes, we derive a compact set of 100 topological features that summarize the underlying topology of brain tumor structures. These descriptors represent complex 3D tumor morphology while significantly reducing data dimensionality. Unlike many deep learning approaches that require large-scale training data or complex architectures, the proposed framework relies on computationally efficient topological features extracted directly from the images. These features are used to train classical machine learning classifiers, including Random Forest and XGBoost, for binary classification of high-grade glioma (HGG) and low-grade glioma (LGG). Experimental results on the BraTS 2020 dataset show that the Random Forest classifier combined with selected Betti features achieves an accuracy of 89.19%. These findings highlight the potential of persistent homology as an effective and interpretable approach for analyzing complex 3D medical images and performing brain tumor classification.

Method: 1) Data curation pipeline to mine inspection knowledge and generate multimodal dataset Chat-AD; 2) Comparison Encoder using cross-attention between paired image features; 3) Multi-stage training incorporating domain knowledge; 4) Extended benchmark MMAD-BBox for bounding-box evaluation.

Result: Achieves 82.3% accuracy on MMAD benchmark (outperforming all other models), 3.35× improvement over baseline on MMAD-BBox, and demonstrates excellent generalization across specialized and general-purpose benchmarks.

Conclusion: AD-Copilot surpasses human expert-level performance on several IAD tasks, demonstrating potential as a reliable assistant for real-world industrial inspection, with all datasets and models to be released.

Abstract: Multimodal Large Language Models (MLLMs) have achieved impressive success in natural visual understanding, yet they consistently underperform in industrial anomaly detection (IAD). This is because MLLMs trained mostly on general web data differ significantly from industrial images. Moreover, they encode each image independently and can only compare images in the language space, making them insensitive to subtle visual differences that are key to IAD. To tackle these issues, we present AD-Copilot, an interactive MLLM specialized for IAD via visual in-context comparison. We first design a novel data curation pipeline to mine inspection knowledge from sparsely labeled industrial images and generate precise samples for captioning, VQA, and defect localization, yielding a large-scale multimodal dataset Chat-AD rich in semantic signals for IAD. On this foundation, AD-Copilot incorporates a novel Comparison Encoder that employs cross-attention between paired image features to enhance multi-image fine-grained perception, and is trained with a multi-stage strategy that incorporates domain knowledge and gradually enhances IAD skills. In addition, we introduce MMAD-BBox, an extended benchmark for anomaly localization with bounding-box-based evaluation. The experiments show that AD-Copilot achieves 82.3% accuracy on the MMAD benchmark, outperforming all other models without any data leakage. In the MMAD-BBox test, it achieves a maximum improvement of $3.35\times$ over the baseline. AD-Copilot also exhibits excellent generalization of its performance gains across other specialized and general-purpose benchmarks. Remarkably, AD-Copilot surpasses human expert-level performance on several IAD tasks, demonstrating its potential as a reliable assistant for real-world industrial inspection. All datasets and models will be released for the broader benefit of the community.

Method: Proposes RetimeGS with explicit temporal behavior definition for 3D Gaussians, optical flow-guided initialization and supervision, triple-rendering supervision, and targeted strategies to mitigate temporal aliasing and ensure temporal coherence.

Result: RetimeGS achieves superior quality and temporal coherence compared to state-of-the-art methods on datasets with fast motion, non-rigid deformation, and severe occlusions, enabling ghost-free rendering.

Conclusion: RetimeGS effectively addresses temporal aliasing in 4DGS representations, enabling high-quality temporal retiming for dynamic scene reconstruction and rendering applications.

Abstract: Temporal retiming, the ability to reconstruct and render dynamic scenes at arbitrary timestamps, is crucial for applications such as slow-motion playback, temporal editing, and post-production. However, most existing 4D Gaussian Splatting (4DGS) methods overfit at discrete frame indices but struggle to represent continuous-time frames, leading to ghosting artifacts when interpolating between timestamps. We identify this limitation as a form of temporal aliasing and propose RetimeGS, a simple yet effective 4DGS representation that explicitly defines the temporal behavior of the 3D Gaussian and mitigates temporal aliasing. To achieve smooth and consistent interpolation, we incorporate optical flow-guided initialization and supervision, triple-rendering supervision, and other targeted strategies. Together, these components enable ghost-free, temporally coherent rendering even under large motions. Experiments on datasets featuring fast motion, non-rigid deformation, and severe occlusions demonstrate that RetimeGS achieves superior quality and coherence over state-of-the-art methods.

Method: Proposes HFGPI with three key components: 1) Molecular Tokenizer for biologically informed gene/protein representations, 2) Gene-Regulated Protein Fusion (GRPF) using graph-aware cross-attention to model gene-protein regulatory relationships, and 3) Protein-Guided Hypergraph Learning (PGHL) to establish associations between proteins and image patches using hypergraph convolution.

Result: Extensive experiments on five benchmark datasets demonstrate HFGPI’s superiority over state-of-the-art methods for cancer survival prediction.

Conclusion: HFGPI effectively addresses limitations of existing multimodal survival methods by incorporating proteomic data and modeling biological hierarchy, leading to improved cancer prognosis precision.

Abstract: To enhance the precision of cancer prognosis, recent research has increasingly focused on multimodal survival methods by integrating genomic data and histology images. However, current approaches overlook the fact that the proteome serves as an intermediate layer bridging genomic alterations and histopathological features while providing complementary biological information essential for survival prediction. This biological reality exposes another architectural limitation: existing integrative analysis studies fuse these heterogeneous data sources in a flat manner that fails to capture their inherent biological hierarchy. To address these limitations, we propose HFGPI, a hierarchical fusion framework that models the biological progression from genes to proteins to histology images from a systems biology perspective. Specifically, we introduce Molecular Tokenizer, a molecular encoding strategy that integrates identity embeddings with expression profiles to construct biologically informed representations for genes and proteins. We then develop Gene-Regulated Protein Fusion (GRPF), which employs graph-aware cross-attention with structure-preserving alignment to explicitly model gene-protein regulatory relationships and generate gene-regulated protein representations. Additionally, we propose Protein-Guided Hypergraph Learning (PGHL), which establishes associations between proteins and image patches, leveraging hypergraph convolution to capture higher-order protein-morphology relationships. The final features are progressively fused across hierarchical layers to achieve precise survival outcome prediction. Extensive experiments on five benchmark datasets demonstrate the superiority of HFGPI over state-of-the-art methods.

Method: Introduces agentic pipeline that autonomously synthesizes spatial VQA data by orchestrating computational tools (volume/distance calculators) with multi-agent collaboration and expert radiologist validation. Creates SpatialMed benchmark with 10K QA pairs across multiple organs and tumor types.

Result: Evaluations on 14 state-of-the-art MLLMs reveal current models lack robust spatial reasoning capabilities for medical imaging, highlighting the gap in 3D spatial intelligence.

Conclusion: SpatialMed addresses the critical need for evaluating 3D spatial intelligence in medical MLLMs, exposing limitations in current models and providing a benchmark for future development.

Abstract: Visual spatial intelligence is critical for medical image interpretation, yet remains largely unexplored in Multimodal Large Language Models (MLLMs) for 3D imaging. This gap persists due to a systemic lack of datasets featuring structured 3D spatial annotations beyond basic labels. In this study, we introduce an agentic pipeline that autonomously synthesizes spatial visual question-answering (VQA) data by orchestrating computational tools such as volume and distance calculators with multi-agent collaboration and expert radiologist validation. We present SpatialMed, the first comprehensive benchmark for evaluating 3D spatial intelligence in medical MLLMs, comprising nearly 10K question-answer pairs across multiple organs and tumor types. Our evaluations on 14 state-of-the-art MLLMs and extensive analyses reveal that current models lack robust spatial reasoning capabilities for medical imaging.

Method: Five-stage pipeline: 1) multi-temporal SAR change detection using VV/VH intensity and InSAR coherence, 2) physics-informed depth estimation fusing flood extent with DEMs, 3) property-level zonal statistics from parcel footprints, 4) depth-damage calibration against NFIP claims, 5) confidence-scored triage ranking.

Result: ALTIS achieves an Inspection Reduction Rate of ~0.52 at 90% recall of high-severity claims, potentially eliminating over half of unnecessary dispatches. Demonstrated on Hurricane Harvey (2017) in Harris County, Texas.

Conclusion: ALTIS establishes a methodological baseline for translating earth observation research into measurable insurance outcomes by blending SAR flood intelligence with claims management realities.

Abstract: Floods are among the costliest natural catastrophes globally, yet the property and casualty insurance industry’s post-event response remains heavily reliant on manual field inspection: slow, expensive, and geographically constrained. Satellite Synthetic Aperture Radar (SAR) offers cloud-penetrating, all-weather imaging uniquely suited to rapid post-flood assessment, but existing research evaluates SAR flood detection against academic benchmarks such as IoU and F1-score that do not capture insurance-workflow requirements. We present ALTIS: a five-stage pipeline transforming raw Sentinel-1 GRD and SLC imagery into property-level impact scores within 24-48 hours of flood peak. Unlike prior approaches producing pixel-level maps or binary outputs, ALTIS delivers a ranked, confidence-scored triage list consumable by claims platforms, integrating (i) multi-temporal SAR change detection using dual-polarization VV/VH intensity and InSAR coherence, (ii) physics-informed depth estimation fusing flood extent with high-resolution DEMs, (iii) property-level zonal statistics from parcel footprints, (iv) depth-damage calibration against NFIP claims, and (v) confidence-scored triage ranking. We formally define Insurance-Grade Flood Triage (IGFT) and introduce the Inspection Reduction Rate (IRR) and Triage Efficiency Score (TES). Using Hurricane Harvey (2017) across Harris County, Texas, we present preliminary analysis grounded in validated sub-components suggesting ALTIS is designed to achieve an IRR of approximately 0.52 at 90% recall of high-severity claims, potentially eliminating over half of unnecessary dispatches. By blending SAR flood intelligence with the realities of claims management, ALTIS establishes a methodological baseline for translating earth observation research into measurable insurance outcomes.

Method: Semi-automated active learning pipeline integrating U-Net CNN with interactive user interface and core-set selection. Evaluated three strategies: manual selection, uncertainty sampling, and proposed SMILE (maximin Latin hypercube sampling from embeddings).

Result: SMILE strategy consistently outperformed others, improving macro F1 score from 0.74 to 0.93 while reducing manual annotation time by ~65%. Segmented defects were further analyzed with classification model to map defects to AM process parameters.

Conclusion: The framework reduces labeling effort while maintaining scalability and robustness, broadly applicable to image-based analysis across diverse materials systems.

Abstract: Image segmentation is fundamental to microstructural analysis for defect identification and structure-property correlation, yet remains challenging due to pronounced heterogeneity in materials images arising from varied processing and testing conditions. Conventional image processing techniques often fail to capture such complex features rendering them ineffective for large-scale analysis. Even deep learning approaches struggle to generalize across heterogeneous datasets due to scarcity of high-quality labeled data. Consequently, segmentation workflows often rely on manual expert-driven annotations which are labor intensive and difficult to scale. Using an additive manufacturing (AM) dataset as a case study, we present a semi-automated active learning based segmentation pipeline that integrates a U-Net based convolutional neural network with an interactive user annotation and correction interface and a representative core-set image selection strategy. The active learning workflow iteratively updates the model by incorporating user corrected segmentations into the training pool while the core-set strategy identifies representative images for annotation. Three subset selection strategies, manual selection, uncertainty driven sampling and proposed maximin Latin hypercube sampling from embeddings (SMILE) method were evaluated over six refinement rounds. The SMILE strategy consistently outperformed other approaches, improving the macro F1 score from 0.74 to 0.93 while reducing manual annotation time by about 65 percent. The segmented defect regions were further analyzed using a coupled classification model to categorize defects based on microstructural characteristics and map them to corresponding AM process parameters. The proposed framework reduces labeling effort while maintaining scalability and robustness and is broadly applicable to image based analysis across diverse materials systems.

Method: Introduces new CVMOGL task, creates CMLocation benchmark with two datasets (V1 and V2), and proposes MOGeo method for multi-object geo-localization with extensive experimental validation across various scenarios.

Result: Experiments show cross-view object geo-localization in realistic multi-object settings remains challenging, with proposed method benchmarked against existing state-of-the-art approaches.

Conclusion: The realistic multi-object geo-localization setting presents significant challenges, encouraging further research in this area with the new benchmark and task definition.

Abstract: Cross-View Object Geo-Localization (CVOGL) aims to locate an object of interest in a query image within a corresponding satellite image. Existing methods typically assume that the query image contains only a single object, which does not align with the complex, multi-object geo-localization requirements in real-world applications, making them unsuitable for practical scenarios. To bridge the gap between the realistic setting and existing task, we propose a new task, called Cross-View Multi-Object Geo-Localization (CVMOGL). To advance the CVMOGL task, we first construct a benchmark, CMLocation, which includes two datasets: CMLocation-V1 and CMLocation-V2. Furthermore, we propose a novel cross-view multi-object geo-localization method, MOGeo, and benchmark it against existing state-of-the-art methods. Extensive experiments are conducted under various application scenarios to validate the effectiveness of our method. The results demonstrate that cross-view object geo-localization in the more realistic setting remains a challenging problem, encouraging further research in this area.

Method: Uses vision foundation models for generalizable visual representations. Implements hierarchical clue extraction with Generalized Mean pooling and Scale-Weighted RMAC to preserve distinctive clues across scales. Employs statistical manifold alignment via domain-wise PCA and Orthogonal Procrustes analysis to linearly align heterogeneous feature distributions.

Result: Achieves strong zero-shot accuracy on standard benchmarks and surpasses supervised methods by over 20% in Recall@1 on the challenging LO-UCV dataset with large oblique angles.

Conclusion: Principled alignment of pre-trained features can effectively bridge the cross-view gap, establishing a robust, training-free paradigm for real-world cross-view geo-localization.

Abstract: Cross-View Geo-Localization (CVGL) in remote sensing aims to locate a drone-view query by matching it to geo-tagged satellite images. Although supervised methods have achieved strong results on closeset benchmarks, they often fail to generalize to unconstrained, real-world scenarios due to severe viewpoint differences and dataset bias. To overcome these limitations, we present VFM-Loc, a training-free framework for zero-shot CVGL that leverages the generalizable visual representations from vision foundational models (VFMs). VFM-Loc identifies and matches discriminative visual clues across different viewpoints through a progressive alignment strategy. First, we design a hierarchical clue extraction mechanism using Generalized Mean pooling and Scale-Weighted RMAC to preserve distinctive visual clues across scales while maintaining hierarchical confidence. Second, we introduce a statistical manifold alignment pipeline based on domain-wise PCA and Orthogonal Procrustes analysis, linearly aligning heterogeneous feature distributions in a shared metric space. Experiments demonstrate that VFM-Loc exhibits strong zero-shot accuracy on standard benchmarks and surpasses supervised methods by over 20% in Recall@1 on the challenging LO-UCV dataset with large oblique angles. This work highlights that principled alignment of pre-trained features can effectively bridge the cross-view gap, establishing a robust and training-free paradigm for real-world CVGL. The relevant code is made available at: https://github.com/DingLei14/VFM-Loc.

Method: Proposes Learning through Creation (LTC) with MKEE generator that creates pseudo-novel instances on-the-fly using kernel-energy minimization and entropy maximization, trained with dual max-margin objective and adaptive thresholding.

Result: Extensive experiments across seven benchmarks show LTC consistently outperforms prior work, achieving improvements ranging from 1.5% to 13.1% in all-class accuracy.

Conclusion: LTC effectively addresses optimization misalignment in OCD by explicitly training for discovery during offline learning through on-the-fly creation of pseudo-novel instances.

Abstract: On-the-Fly Category Discovery (OCD) aims to recognize known classes while simultaneously discovering emerging novel categories during inference, using supervision only from known classes during offline training. Existing approaches rely either on fixed label supervision or on diffusion-based augmentations to enhance the backbone, yet none of them explicitly train the model to perform the discovery task required at test time. It is fundamentally unreasonable to expect a model optimized on limited labeled data to carry out a qualitatively different discovery objective during inference. This mismatch creates a clear optimization misalignment between the offline learning stage and the online discovery stage. In addition, prior methods often depend on hash-based encodings or severe feature compression, which further limits representational capacity. To address these issues, we propose Learning through Creation (LTC), a fully feature-based and hash-free framework that injects novel-category awareness directly into offline learning. At its core is a lightweight, online pseudo-unknown generator driven by kernel-energy minimization and entropy maximization (MKEE). Unlike previous methods that generate synthetic samples once before training, our generator evolves jointly with the model dynamics and synthesizes pseudo-novel instances on the fly at negligible cost. These samples are incorporated through a dual max-margin objective with adaptive thresholding, strengthening the model’s ability to delineate and detect unknown regions through explicit creation. Extensive experiments across seven benchmarks show that LTC consistently outperforms prior work, achieving improvements ranging from 1.5 percent to 13.1 percent in all-class accuracy. The code is available at https://github.com/brandinzhang/LTC

Method: Geo-ID repurposes pretrained single-view intrinsic predictors by coupling independent per-view predictions through sparse geometric correspondences that form uncertainty-aware consensus targets. It’s model-agnostic, requires no retraining or inverse rendering, and works directly with off-the-shelf predictors.

Result: Experiments on synthetic benchmarks and real-world scenes show substantial improvements in cross-view intrinsic consistency as view count increases, while maintaining comparable single-view decomposition performance. Consistent intrinsics enable coherent appearance editing and relighting in neural scene representations.

Conclusion: Geo-ID provides an effective test-time framework for achieving cross-view consistent intrinsic decompositions from sparse image collections, enabling better downstream applications in neural scene editing and relighting without requiring model retraining.

Abstract: Intrinsic image decomposition aims to estimate physically based rendering (PBR) parameters such as albedo, roughness, and metallicity from images. While recent methods achieve strong single-view predictions, applying them independently to multiple views of the same scene often yields inconsistent estimates, limiting their use in downstream applications such as editable neural scenes and 3D reconstruction. Video-based models can improve cross-frame consistency but require dense, ordered sequences and substantial compute, limiting their applicability to sparse, unordered image collections. We propose Geo-ID, a novel test-time framework that repurposes pretrained single-view intrinsic predictors to produce cross-view consistent decompositions by coupling independent per-view predictions through sparse geometric correspondences that form uncertainty-aware consensus targets. Geo-ID is model-agnostic, requires no retraining or inverse rendering, and applies directly to off-the-shelf intrinsic predictors. Experiments on synthetic benchmarks and real-world scenes demonstrate substantial improvements in cross-view intrinsic consistency as the number of views increases, while maintaining comparable single-view decomposition performance. We further show that the resulting consistent intrinsics enable coherent appearance editing and relighting in downstream neural scene representations.

Method: Proposes Cognitive Cascade Segmentation (CogCaS), a dual-phase cascade formulation that separates the task into: 1) class-existence detection phase, and 2) class-specific segmentation phase, inspired by human annotation processes.

Result: Significant improvements on PASCAL VOC 2012 and ADE20K datasets across various challenging scenarios, particularly with long sequences of incremental tasks, outperforming existing state-of-the-art methods.

Conclusion: CogCaS effectively addresses catastrophic forgetting in continual semantic segmentation through its dual-phase cascade approach, enabling better knowledge preservation while incorporating new classes incrementally.

Abstract: Continual semantic segmentation (CSS) is a cornerstone task in computer vision that enables a large number of downstream applications, but faces the catastrophic forgetting challenge. In conventional class-incremental semantic segmentation (CISS) frameworks using Softmax-based classification heads, catastrophic forgetting originates from Catastrophic forgetting and task affiliation probability. We formulate these problems and provide a theoretical analysis to more deeply understand the limitations in existing CISS methods, particularly Strict Parameter Isolation (SPI). To address these challenges, we follow a dual-phase intuition from human annotators, and introduce Cognitive Cascade Segmentation (CogCaS), a novel dual-phase cascade formulation for CSS tasks in the CISS setting. By decoupling the task into class-existence detection and class-specific segmentation, CogCaS enables more effective continual learning, preserving previously learned knowledge while incorporating new classes. Using two benchmark datasets PASCAL VOC 2012 and ADE20K, we have shown significant improvements in a variety of challenging scenarios, particularly those with long sequence of incremental tasks, when compared to exsiting state-of-the-art methods. Our code will be made publicly available upon paper acceptance.

Method: Created Step-CoT dataset with 10K clinical cases and 70K VQA pairs featuring expert-curated structured reasoning steps. Introduced teacher-student framework with dynamic graph-structured focusing mechanism to prioritize diagnostically informative steps.

Result: Step-CoT improves reasoning accuracy and interpretability in medical VQA by providing supervised intermediate steps that guide models along valid clinical reasoning trajectories.

Conclusion: Structured multi-step reasoning supervision aligned with clinical workflows enhances both performance and transparency in medical VQA systems.

Abstract: Chain-of-thought (CoT) reasoning has advanced medical visual question answering (VQA), yet most existing CoT rationales are free-form and fail to capture the structured reasoning process clinicians actually follow. This work asks: Can traceable, multi-step reasoning supervision improve reasoning accuracy and the interpretability of Medical VQA? To this end, we introduce Step-CoT, a large-scale medical reasoning dataset with expert-curated, structured multi-step CoT aligned to clinical diagnostic workflows, implicitly grounding the model’s reasoning in radiographic evidence. Step-CoT comprises more than 10K real clinical cases and 70K VQA pairs organized around diagnostic workflows, providing supervised intermediate steps that guide models to follow valid reasoning trajectories. To effectively learn from Step-CoT, we further introduce a teacher-student framework with a dynamic graph-structured focusing mechanism that prioritizes diagnostically informative steps while filtering out less relevant contexts. Our experiments show that using Step-CoT can improve reasoning accuracy and interpretability. Benchmark: github.com/hahaha111111/Step-CoT. Dataset Card: huggingface.co/datasets/fl-15o/Step-CoT

Method: Two improvement strategies: Method 1 adds Large Separable Kernel Attention (LSKA) to backbone for small object feature extraction and Gold-YOLO structure to neck for multi-scale fusion. Method 2 uses Gold-YOLO in neck and MultiSEAMHead detection head for enhanced small/multi-scale object representation.

Result: Experiments on DOTAv1 dataset show mAP@0.5 improvements of 1.3% and 1.8% respectively over baseline YOLOv11n while maintaining lightweight model advantages.

Conclusion: The proposed methods effectively improve object detection in remote sensing images with practical value for satellite imagery applications.

Abstract: Satellite remote sensing images pose significant challenges for object detection due to their high resolution, complex scenes, and large variations in target scales. To address the insufficient detection accuracy of the YOLOv11n model in remote sensing imagery, this paper proposes two improvement strategies. Method 1: (a) a Large Separable Kernel Attention (LSKA) mechanism is introduced into the backbone network to enhance feature extraction for small objects; (b) a Gold-YOLO structure is incorporated into the neck network to achieve multi-scale feature fusion, thereby improving the detection performance of objects at different scales. Method 2: (a) the Gold-YOLO structure is also integrated into the neck network; (b) a MultiSEAMHead detection head is combined to further strengthen the representation and detection capability for small and multi-scale objects. To verify the effectiveness of the proposed improvements, experiments are conducted on the DOTAv1 dataset. The results show that, while maintaining the lightweight advantage of the model, the proposed methods improve detection accuracy (mAP@0.5) by 1.3% and 1.8%, respectively, compared with the baseline YOLOv11n, demonstrating the effectiveness and practical value of the proposed approaches for object detection in remote sensing images.

Method: The authors show that CCA and InfoNCE objectives are closely related, such that optimizing CCA implicitly optimizes InfoNCE. They introduce Concept CCA (CoCCA) which couples concept-based explainability with CCA to align cross-modal embeddings while enabling interpretable concept decomposition. They further propose Sparse Concept CCA (SCoCCA) which enforces sparsity to produce more disentangled and discriminative concepts.

Result: The approach generalizes concept-based explanations to multi-modal embeddings and achieves state-of-the-art performance in concept discovery, evidenced by reconstruction and manipulation tasks such as concept ablation.

Conclusion: The proposed frameworks provide a principled, training-free mechanism to enhance cross-modal alignment while enabling interpretable concept decomposition for vision-language models, addressing the modality gap issue in existing embeddings.

Abstract: Interpreting the internal reasoning of vision-language models is essential for deploying AI in safety-critical domains. Concept-based explainability provides a human-aligned lens by representing a model’s behavior through semantically meaningful components. However, existing methods are largely restricted to images and overlook the cross-modal interactions. Text-image embeddings, such as those produced by CLIP, suffer from a modality gap, where visual and textual features follow distinct distributions, limiting interpretability. Canonical Correlation Analysis (CCA) offers a principled way to align features from different distributions, but has not been leveraged for multi-modal concept-level analysis. We show that the objectives of CCA and InfoNCE are closely related, such that optimizing CCA implicitly optimizes InfoNCE, providing a simple, training-free mechanism to enhance cross-modal alignment without affecting the pre-trained InfoNCE objective. Motivated by this observation, we couple concept-based explainability with CCA, introducing Concept CCA (CoCCA), a framework that aligns cross-modal embeddings while enabling interpretable concept decomposition. We further extend it and propose Sparse Concept CCA (SCoCCA), which enforces sparsity to produce more disentangled and discriminative concepts, facilitating improved activation, ablation, and semantic manipulation. Our approach generalizes concept-based explanations to multi-modal embeddings and achieves state-of-the-art performance in concept discovery, evidenced by reconstruction and manipulation tasks such as concept ablation.

Method: Proposes LER with three modules: Localization (uses multimodal information to precisely determine character positions), Extraction (dissociates all characters in parallel), and Recognition (considers unique inner structures of Chinese for text prediction). The method explicitly decouples each character for independent recognition.

Result: Extensive experiments on large-scale Chinese benchmarks show LER significantly outperforms existing methods. Additional experiments on six English benchmarks and Union14M benchmark show impressive results in English text recognition as well.

Conclusion: LER effectively addresses the challenges of Chinese scene text recognition by explicitly decoupling characters and considering Chinese inner structures, demonstrating superior performance over existing methods while also showing strong cross-language applicability.

Abstract: Scene text recognition (STR) methods have demonstrated their excellent capability in English text images. However, due to the complex inner structures of Chinese and the extensive character categories, it poses challenges for recognizing Chinese text in images. Recently, studies have shown that the methods designed for English text recognition encounter an accuracy bottleneck when recognizing Chinese text images. This raises the question: Is it appropriate to apply the model developed for English to the Chinese STR task? To explore this issue, we propose a novel method named LER, which explicitly decouples each character and independently recognizes characters while taking into account the complex inner structures of Chinese. LER consists of three modules: Localization, Extraction, and Recognition. Firstly, the localization module utilizes multimodal information to determine the character’s position precisely. Then, the extraction module dissociates all characters in parallel. Finally, the recognition module considers the unique inner structures of Chinese to provide the text prediction results. Extensive experiments conducted on large-scale Chinese benchmarks indicate that our method significantly outperforms existing methods. Furthermore, extensive experiments conducted on six English benchmarks and the Union14M benchmark show impressive results in English text recognition by LER. Code is available at https://github.com/Pandarenlql/LER.

Method: Formulate PET super-resolution as posterior inference under heterogeneous system configurations using a CT-conditioned diffusion framework with physics-constrained sampling. Learn conditional diffusion prior from high-quality PET/CT pairs using cross-attention for anatomical guidance. During inference, enforce measurement consistency through scanner-aware forward model with explicit PSF effects and gradient-based data-consistency refinement.

Result: The method consistently improves experimental metrics and lesion-level clinical relevance indicators over strong baselines under both standard and out-of-distribution settings, while reducing hallucination artifacts and improving structural fidelity.

Conclusion: The proposed physics-constrained diffusion framework effectively addresses PET super-resolution challenges by combining anatomical guidance from CT with scanner-aware physical constraints, achieving improved performance without requiring paired LR-HR PET data.

Abstract: PET super-resolution is highly under-constrained because paired multi-resolution scans from the same subject are rarely available, and effective resolution is determined by scanner-specific physics (e.g., PSF, detector geometry, and acquisition settings). This limits supervised end-to-end training and makes purely image-domain generative restoration prone to hallucinated structures when anatomical and physical constraints are weak. We formulate PET super-resolution as posterior inference under heterogeneous system configurations and propose a CT-conditioned diffusion framework with physics-constrained sampling. During training, a conditional diffusion prior is learned from high-quality PET/CT pairs using cross-attention for anatomical guidance, without requiring paired LR–HR PET data. During inference, measurement consistency is enforced through a scanner-aware forward model with explicit PSF effects and gradient-based data-consistency refinement. Under both standard and OOD settings, the proposed method consistently improves experimental metrics and lesion-level clinical relevance indicators over strong baselines, while reducing hallucination artifacts and improving structural fidelity.

Method: CroBo uses a global-to-local reconstruction objective. A reference observation is compressed into a compact bottleneck token, then the model learns to reconstruct heavily masked patches in a local target crop from sparse visible cues, using the global bottleneck token as context.

Result: CroBo achieves state-of-the-art performance on diverse vision-based robot policy learning benchmarks. Reconstruction analyses and perceptual straightness experiments show the learned representations preserve pixel-level scene composition and encode what-moves-where across observations.

Conclusion: CroBo’s global-to-local reconstruction objective effectively learns visual state representations that capture scene-wide semantic entities, their spatial configurations, and dynamics, supporting sequential decision making for robotic agents.

Abstract: For robotic agents operating in dynamic environments, learning visual state representations from streaming video observations is essential for sequential decision making. Recent self-supervised learning methods have shown strong transferability across vision tasks, but they do not explicitly address what a good visual state should encode. We argue that effective visual states must capture what-is-where by jointly encoding the semantic identities of scene elements and their spatial locations, enabling reliable detection of subtle dynamics across observations. To this end, we propose CroBo, a visual state representation learning framework based on a global-to-local reconstruction objective. Given a reference observation compressed into a compact bottleneck token, CroBo learns to reconstruct heavily masked patches in a local target crop from sparse visible cues, using the global bottleneck token as context. This learning objective encourages the bottleneck token to encode a fine-grained representation of scene-wide semantic entities, including their identities, spatial locations, and configurations. As a result, the learned visual states reveal how scene elements move and interact over time, supporting sequential decision making. We evaluate CroBo on diverse vision-based robot policy learning benchmarks, where it achieves state-of-the-art performance. Reconstruction analyses and perceptual straightness experiments further show that the learned representations preserve pixel-level scene composition and encode what-moves-where across observations.

Method: 1. Predict global 3D layout from text description encoding semantic and geometric structure; 2. Use semantics- and depth-conditioned panoramic diffusion to synthesize 360° imagery aligned with layout; 3. Employ video diffusion with optimized camera trajectories for unobserved regions; 4. Fuse views using 3D Gaussian Splatting for consistent reconstruction.

Result: Produces metrically accurate, globally consistent indoor scenes with absolute scale. Quantitative results and user study show superior 3D consistency and layout plausibility compared to panoramic text-to-3D baselines. Achieves up to 10x faster sampling than exhaustive path exploration.

Conclusion: GuidedSceneGen enables accurate text-to-3D generation with maintained world coordinates, supporting object pose transfer, semantic labeling, and progressive scene expansion without re-alignment, advancing towards practical 3D scene creation.

Abstract: We present GuidedSceneGen, a text-to-3D generation framework that produces metrically accurate, globally consistent, and semantically interpretable indoor scenes. Unlike prior text-driven methods that often suffer from geometric drift or scale ambiguity, our approach maintains an absolute world coordinate frame throughout the entire generation process. Starting from a textual scene description, we predict a global 3D layout encoding both semantic and geometric structure, which serves as a guiding proxy for downstream stages. A semantics- and depth-conditioned panoramic diffusion model then synthesizes 360° imagery aligned with the global layout, substantially improving spatial coherence. To explore unobserved regions, we employ a video diffusion model guided by optimized camera trajectories that balances coverage and collision avoidance, achieving up to 10x faster sampling compared to exhaustive path exploration. The generated views are fused using 3D Gaussian Splatting, yielding a consistent and fully navigable 3D scene in absolute scale. GuidedSceneGen enables accurate transfer of object poses and semantic labels from layout to reconstruction, and supports progressive scene expansion without re-alignment. Quantitative results and a user study demonstrate greater 3D consistency and layout plausibility compared to recent panoramic text-to-3D baselines.

Method: EgoViT uses a vision Transformer framework with three synergistic mechanisms: (1) Proto-object Learning via intra-frame distillation for discriminative representations, (2) Depth Regularization to ground representations in geometric structure, and (3) Teacher-Filtered Temporal Consistency to enforce identity over time, creating a virtuous cycle where initial object hypotheses are progressively refined.

Result: The framework achieves +8.0% CorLoc improvement in unsupervised object discovery and +4.8% mIoU improvement in semantic segmentation on standard benchmarks, demonstrating robustness to varied geometric priors.

Conclusion: EgoViT shows potential to lay a foundation for robust visual abstraction in embodied intelligence by learning stable object representations from unlabeled egocentric video through self-supervised learning.

Abstract: Humans develop visual intelligence through perceiving and interacting with their environment - a self-supervised learning process grounded in egocentric experience. Inspired by this, we ask how can artificial systems learn stable object representations from continuous, uncurated first-person videos without relying on manual annotations. This setting poses challenges of separating, recognizing, and persistently tracking objects amid clutter, occlusion, and ego-motion. We propose EgoViT, a unified vision Transformer framework designed to learn stable object representations from unlabeled egocentric video. EgoViT bootstraps this learning process by jointly discovering and stabilizing “proto-objects” through three synergistic mechanisms: (1) Proto-object Learning, which uses intra-frame distillation to form discriminative representations; (2) Depth Regularization, which grounds these representations in geometric structure; and (3) Teacher-Filtered Temporal Consistency, which enforces identity over time. This creates a virtuous cycle where initial object hypotheses are progressively refined into stable, persistent representations. The framework is trained end-to-end on unlabeled first-person videos and exhibits robustness to geometric priors of varied origin and quality. On standard benchmarks, EgoViT achieves +8.0% CorLoc improvement in unsupervised object discovery and +4.8% mIoU improvement in semantic segmentation, demonstrating its potential to lay a foundation for robust visual abstraction in embodied intelligence.

Method: Comparative evaluation of state-of-the-art VPR families including NetVLAD-style baselines, classification-based global descriptors (CosPlace, EigenPlaces), feature-mixing (MixVPR), and foundation-model-driven methods (AnyLoc, SALAD, MegaLoc) on three challenging datasets: object-centric outdoor scenes (Tanks and Temples), indoor RGB-D scans (ScanNet-GS), and autonomous-driving sequences (KITTI).

Result: Modern global descriptor approaches are increasingly suitable as off-the-shelf image pair retrieval modules in challenging scenarios including perceptual aliasing and incomplete sequences, while exhibiting clear, domain-dependent strengths and weaknesses.

Conclusion: The choice of VPR components for robust mapping and registration requires careful consideration of domain-specific performance characteristics, as different methods excel in different scenarios.

Abstract: Visual Place Recognition (VPR) is a core component in computer vision, typically formulated as an image retrieval task for localization, mapping, and navigation. In this work, we instead study VPR as an image pair retrieval front-end for registration pipelines, where the goal is to find top-matching image pairs between two disjoint image sets for downstream tasks such as scene registration, SLAM, and Structure-from-Motion. We comparatively evaluate state-of-the-art VPR families - NetVLAD-style baselines, classification-based global descriptors (CosPlace, EigenPlaces), feature-mixing (MixVPR), and foundation-model-driven methods (AnyLoc, SALAD, MegaLoc) - on three challenging datasets: object-centric outdoor scenes (Tanks and Temples), indoor RGB-D scans (ScanNet-GS), and autonomous-driving sequences (KITTI). We show that modern global descriptor approaches are increasingly suitable as off-the-shelf image pair retrieval modules in challenging scenarios including perceptual aliasing and incomplete sequences, while exhibiting clear, domain-dependent strengths and weaknesses that are critical when choosing VPR components for robust mapping and registration.

Method: Used a 2x2 experimental framework on visual question answering to assess performance across in-distribution/out-of-distribution image and text inputs. Applied regularization techniques (constrained trainable parameters, low learning rates) and introduced data-hybrid training strategy combining datasets and tasks.

Result: Appropriate regularization prevents forgetting with out-of-distribution images, but task-specific overfitting occurs with in-distribution images and out-of-distribution text. Data-hybrid training addresses this issue and extends to continual learning, outperforming existing methods with complex auxiliary mechanisms.

Conclusion: MLLMs have inherent robustness against catastrophic forgetting; simple fine-tuning adjustments are sufficient for adaptation while preserving general capabilities, providing practical guidelines for MLLM fine-tuning.

Abstract: The paper demonstrate that simple adjustments of the fine-tuning recipes of multimodal large language models (MLLM) are sufficient to mitigate catastrophic forgetting. On visual question answering, we design a 2x2 experimental framework to assess model performance across in-distribution and out-of-distribution image and text inputs. Our results show that appropriate regularization, such as constraining the number of trainable parameters or adopting a low learning rate, effectively prevents forgetting when dealing with out-of-distribution images. However, we uncover a distinct form of forgetting in settings with in-distribution images and out-of-distribution text. We attribute this forgetting as task-specific overfitting and address this issue by introducing a data-hybrid training strategy that combines datasets and tasks. Finally, we demonstrate that this approach naturally extends to continual learning, outperforming existing methods with complex auxiliary mechanisms. In general, our findings challenge the prevailing assumptions by highlighting the inherent robustness of MLLMs and providing practical guidelines for adapting them while preserving their general capabilities.

Method: Proposes OpenCOOD-Air framework integrating UAVs into V2V perception, uses transfer learning to fine-tune UAV weights from pre-trained V2V models, introduces Cross-Domain Spatial Converter (CDSC) and Spatial Offset Prediction Transformer (SOPT) for heterogeneous ground-air integration with explicit altitude supervision.

Result: Improves 2D and 3D AP@0.7 by 4% and 7% respectively compared to state-of-the-art methods, validated on the new OPV2V-Air benchmark.

Conclusion: Integrating UAVs into V2V collaborative perception effectively overcomes ground-level limitations and significantly enhances perception performance through careful domain adaptation and spatial alignment techniques.

Abstract: While Vehicle-to-Vehicle (V2V) collaboration extends sensing ranges through multi-agent data sharing, its reliability remains severely constrained by ground-level occlusions and the limited perspective of chassis-mounted sensors, which often result in critical perception blind spots. We propose OpenCOOD-Air, a novel framework that integrates UAVs as extensible platforms into V2V collaborative perception to overcome these constraints. To mitigate gradient interference from ground-air domain gaps and data sparsity, we adopt a transfer learning strategy to fine-tune UAV weights from pre-trained V2V models. To prevent the spatial information loss inherent in this transition, we formulate ground-air collaborative perception as a heterogeneous integration task with explicit altitude supervision and introduce a Cross-Domain Spatial Converter (CDSC) and a Spatial Offset Prediction Transformer (SOPT). Furthermore, we present the OPV2V-Air benchmark to validate the transition from V2V to Vehicle-to-Vehicle-to-UAV. Compared to state-of-the-art methods, our approach improves 2D and 3D AP@0.7 by 4% and 7%, respectively.

Method: Proposes Discriminative Flow Matching that learns a vector field to transport samples from noise distribution to task-aligned target manifolds (class embeddings or bounding box coordinates). Uses multiple independent flow predictors attached to a shared backbone, trained with local flow matching objectives where gradients are computed independently for each block.

Result: The framework enables robust, generative-inspired inference across diverse architectures (CNNs and vision transformers) and provides flexibility for sequential or parallel block updates to suit different hardware constraints.

Conclusion: Discriminative Flow Matching successfully bridges generative and discriminative learning, offering iterative refinement for classification and object detection tasks while maintaining computational flexibility.

Abstract: Traditional discriminative computer vision relies predominantly on static projections, mapping input features to outputs in a single computational step. Although efficient, this paradigm lacks the iterative refinement and robustness inherent in biological vision and modern generative modelling. In this paper, we propose Discriminative Flow Matching, a framework that reformulates classification and object detection as a conditional transport process. By learning a vector field that continuously transports samples from a simple noise distribution toward a task-aligned target manifold – such as class embeddings or bounding box coordinates – we are at the interface between generative and discriminative learning. Our method attaches multiple independent flow predictors to a shared backbone. These predictors are trained using local flow matching objectives, where gradients are computed independently for each block. We formulate this approach for standard image classification and extend it to the complex task of object detection, where targets are high-dimensional and spatially distributed. This architecture provides the flexibility to update blocks either sequentially to minimise activation memory or in parallel to suit different hardware constraints. By aggregating the predictions from these independent flow predictors, our framework enables robust, generative-inspired inference across diverse architectures, including CNNs and vision transformers.

Method: Bidirectional Cross-Attention Fusion (BCAF) aligns RGB and HSI at native grids using localized bidirectional cross-attention. Uses two independent backbones: standard Swin Transformer for RGB and HSI-adapted Swin with 3D tokenization and spectral self-attention.

Result: Achieves state-of-the-art 76.4% mIoU at 31 images/s and 75.4% mIoU at 55 images/s on SpectralWaste dataset. On K3I-Cycling dataset: 62.3% mIoU for material segmentation and 66.2% mIoU for plastic-type segmentation.

Conclusion: BCAF effectively fuses RGB and HSI for waste sorting, achieving high accuracy and speed. The modality-agnostic approach can be applied to other co-registered RGB with lower-resolution, high-channel auxiliary sensors.

Abstract: Growing waste streams and the transition to a circular economy require efficient automated waste sorting. In industrial settings, materials move on fast conveyor belts, where reliable identification and ejection demand pixel-accurate segmentation. RGB imaging delivers high-resolution spatial detail, which is essential for accurate segmentation, but it confuses materials that look similar in the visible spectrum. Hyperspectral imaging (HSI) provides spectral signatures that separate such materials, yet its lower spatial resolution limits detail. Effective waste sorting therefore needs methods that fuse both modalities to exploit their complementary strengths. We present Bidirectional Cross-Attention Fusion (BCAF), which aligns high-resolution RGB with low-resolution HSI at their native grids via localized, bidirectional cross-attention, avoiding pre-upsampling or early spectral collapse. BCAF uses two independent backbones: a standard Swin Transformer for RGB and an HSI-adapted Swin backbone that preserves spectral structure through 3D tokenization with spectral self-attention. We also analyze trade-offs between RGB input resolution and the number of HSI spectral slices. Although our evaluation targets RGB-HSI fusion, BCAF is modality-agnostic and applies to co-registered RGB with lower-resolution, high-channel auxiliary sensors. On the benchmark SpectralWaste dataset, BCAF achieves state-of-the-art performance of 76.4% mIoU at 31 images/s and 75.4% mIoU at 55 images/s. We further evaluate a novel industrial dataset: K3I-Cycling (first RGB subset already released on Fordatis). On this dataset, BCAF reaches 62.3% mIoU for material segmentation (paper, metal, plastic, etc.) and 66.2% mIoU for plastic-type segmentation (PET, PP, HDPE, LDPE, PS, etc.).

Method: Combines Self-Supervised Learning (SSL) with Hidden Diffusion Models (LDM). An IJEPA module predicts stable semantic representations, which then route a frozen Stable Diffusion backbone via a lightweight cross-attention adapter, ensuring synthesized textures are based on accurate structural predictions.

Result: Achieves leading perceptual scores on global Sentinel-2 dataset (GSSIM: 0.8984, FID: 0.1475), excels at resolving sharp boundaries, and significantly outperforms deterministic baselines despite standard autoregressive stability limits.

Conclusion: Sat-JEPA-Diff successfully bridges the gap between structural accuracy and textural detail in satellite imagery prediction by combining SSL with diffusion models, producing both accurate and visually realistic results.

Abstract: Predicting satellite imagery requires a balance between structural accuracy and textural detail. Standard deterministic methods like PredRNN or SimVP minimize pixel-based errors but suffer from the “regression to the mean” problem, producing blurry outputs that obscure subtle geographic-spatial features. Generative models provide realistic textures but often misleadingly reveal structural anomalies. To bridge this gap, we introduce Sat-JEPA-Diff, which combines Self-Supervised Learning (SSL) with Hidden Diffusion Models (LDM). An IJEPA module predicts stable semantic representations, which then route a frozen Stable Diffusion backbone via a lightweight cross-attention adapter. This ensures that the synthesized high-accuracy textures are based on absolutely accurate structural predictions. Evaluated on a global Sentinel-2 dataset, Sat-JEPA-Diff excels at resolving sharp boundaries. It achieves leading perceptual scores (GSSIM: 0.8984, FID: 0.1475) and significantly outperforms deterministic baselines, despite standard autoregressive stability limits. The code and dataset are publicly available on https://github.com/VU-AIML/SAT-JEPA-DIFF.

Method: Proposes dual-channel contrastive classification guardrail that operates outside the agent’s perceptual loop. It independently evaluates (1) the visual click target and (2) the agent’s reasoning about the action against deployment-specific knowledge bases, blocking execution if either channel indicates risk.

Result: The combined guardrail consistently outperforms either channel alone across controlled attacks, real GUI screenshots, and agent traces. Visual evidence detects target-level mismatches while textual reasoning reveals dangerous intent behind visually innocuous controls.

Conclusion: CUA safety requires not only better action generation but independent verification of what the agent believes it is clicking and why. The proposed guardrail addresses visual confused deputy attacks by operating outside the agent’s perceptual loop.

Abstract: Computer-using agents (CUAs) act directly on graphical user interfaces, yet their perception of the screen is often unreliable. Existing work largely treats these failures as performance limitations, asking whether an action succeeds, rather than whether the agent is acting on the correct object at all. We argue that this is fundamentally a security problem. We formalize the visual confused deputy: a failure mode in which an agent authorizes an action based on a misperceived screen state, due to grounding errors, adversarial screenshot manipulation, or time-of-check-to-time-of-use (TOCTOU) races. This gap is practically exploitable: even simple screen-level manipulations can redirect routine clicks into privileged actions while remaining indistinguishable from ordinary agent mistakes. To mitigate this threat, we propose the first guardrail that operates outside the agent’s perceptual loop. Our method, dual-channel contrastive classification, independently evaluates (1) the visual click target and (2) the agent’s reasoning about the action against deployment-specific knowledge bases, and blocks execution if either channel indicates risk. The key insight is that these two channels capture complementary failure modes: visual evidence detects target-level mismatches, while textual reasoning reveals dangerous intent behind visually innocuous controls. Across controlled attacks, real GUI screenshots, and agent traces, the combined guardrail consistently outperforms either channel alone. Our results suggest that CUA safety requires not only better action generation, but independent verification of what the agent believes it is clicking and why. Materials are provided\footnote{Model, benchmark, and code: https://github.com/vllm-project/semantic-router}.

Method: Uses CLIP’s open-vocabulary recognition to identify relevant semantic categories, dynamically generates corresponding textual features as initial guidance, performs coarse segmentation by cross-modally integrating semantic information, refines segmentation by integrating spatially enriched features from the encoder, and leverages spatial information to refine category predictions for each mask.

Result: Outperforms existing methods on multiple OVSS benchmarks by delivering both higher accuracy and greater efficiency.

Conclusion: DCP-CLIP effectively addresses cross-modal communication and computational efficiency challenges in open-vocabulary semantic segmentation through dynamic category construction and dual interaction modeling.

Abstract: The recent years have witnessed the remarkable development for open-vocabulary semantic segmentation (OVSS) using visual-language foundation models, yet still suffer from following fundamental challenges: (1) insufficient cross-modal communications between textual and visual spaces, and (2) significant computational costs from the interactions with massive number of categories. To address these issues, this paper describes a novel coarse-to-fine framework, called DCP-CLIP, for OVSS. Unlike prior efforts that mainly relied on pre-established category content and the inherent spatial-class interaction capability of CLIP, we dynamic constructing category-relevant textual features and explicitly models dual interactions between spatial image features and textual class semantics. Specifically, we first leverage CLIP’s open-vocabulary recognition capability to identify semantic categories relevant to the image context, upon which we dynamically generate corresponding textual features to serve as initial textual guidance. Subsequently, we conduct a coarse segmentation by cross-modally integrating semantic information from textual guidance into the visual representations and achieve refined segmentation by integrating spatially enriched features from the encoder to recover fine-grained details and enhance spatial resolution. In final, we leverage spatial information from the segmentation side to refine category predictions for each mask, facilitating more precise semantic labeling. Experiments on multiple OVSS benchmarks demonstrate that DCP-CLIP outperforms existing methods by delivering both higher accuracy and greater efficiency.

Method: Two complementary strategies: 1) Inversion-Matching (IM) - an inversion-guided fine-tuning process that aligns denoising trajectories with their inversion counterparts to broaden distributional coverage and enhance diversity; 2) Selective Subgroup Sampling (S^3) - a training-free sampling mechanism that selects synthetic subsets that are both representative and distinctive to improve inter-class separability.

Result: Extensive experiments demonstrate significant enhancement of discriminative quality and generalization of distilled datasets, achieving state-of-the-art performance among diffusion-based dataset distillation methods.

Conclusion: The proposed approach effectively addresses the misalignment between generative and discriminative objectives in diffusion-based dataset distillation, resulting in improved classification performance and better coverage of boundary samples.

Abstract: Dataset Distillation aims to synthesize compact datasets that can approximate the training efficacy of large-scale real datasets, offering an efficient solution to the increasing computational demands of modern deep learning. Recently, diffusion-based dataset distillation methods have shown great promise by leveraging the strong generative capacity of diffusion models to produce diverse and structurally consistent samples. However, a fundamental goal misalignment persists: diffusion models are optimized for generative likelihood rather than discriminative utility, resulting in over-concentration in high-density regions and inadequate coverage of boundary samples crucial for classification. To address this issue, we propose two complementary strategies. Inversion-Matching (IM) introduces an inversion-guided fine-tuning process that aligns denoising trajectories with their inversion counterparts, broadening distributional coverage and enhancing diversity. Selective Subgroup Sampling(S^3) is a training-free sampling mechanism that improves inter-class separability by selecting synthetic subsets that are both representative and distinctive. Extensive experiments demonstrate that our approach significantly enhances the discriminative quality and generalization of distilled datasets, achieving state-of-the-art performance among diffusion-based methods.

Method: Proposes USIS-PGM with: 1) Frequency-aware module for boundary enhancement, 2) Dynamic weighting module for content-adaptive feature reweighting, 3) Transformer-based instance activation module for better instance distinction, and 4) Multi-scale Gaussian heatmaps generated via Photometric Gaussian Mixture for supervising intermediate features.

Result: Experimental results demonstrate superiority and practical applicability of USIS-PGM for underwater salient instance segmentation.

Conclusion: USIS-PGM effectively addresses underwater image degradation challenges and improves salient instance localization and mask prediction quality for marine robotic applications.

Abstract: Underwater salient instance segmentation (USIS) is crucial for marine robotic systems, as it enables both underwater salient object detection and instance-level mask prediction for visual scene understanding. Compared with its terrestrial counterpart, USIS is more challenging due to the underwater image degradation. To address this issue, this paper proposes USIS-PGM, a single-stage framework for USIS. Specifically, the encoder enhances boundary cues through a frequency-aware module and performs content-adaptive feature reweighting via a dynamic weighting module. The decoder incorporates a Transformer-based instance activation module to better distinguish salient instances. In addition, USIS-PGM employs multi-scale Gaussian heatmaps generated from ground-truth masks through Photometric Gaussian Mixture (PGM) to supervise intermediate decoder features, thereby improving salient instance localization and producing more structurally coherent mask predictions. Experimental results demonstrate the superiority and practical applicability of the proposed USIS-PGM model.

Method: Introduces VID-AD dataset with 10 manufacturing scenarios and 5 capture conditions, and proposes language-based anomaly detection using text descriptions from normal images with contrastive learning (positive texts vs contradiction-based negative texts)

Result: Extensive experiments show consistent improvements over baselines across evaluated settings; dataset contains 50 one-class tasks and 10,395 images

Conclusion: VID-AD addresses gap in logical anomaly detection benchmarks and demonstrates effectiveness of language-based approach for learning logical attributes rather than low-level visual features

Abstract: Logical anomaly detection in industrial inspection remains challenging due to variations in visual appearance (e.g., background clutter, illumination shift, and blur), which often distract vision-centric detectors from identifying rule-level violations. However, existing benchmarks rarely provide controlled settings where logical states are fixed while such nuisance factors vary. To address this gap, we introduce VID-AD, a dataset for logical anomaly detection under vision-induced distraction. It comprises 10 manufacturing scenarios and five capture conditions, totaling 50 one-class tasks and 10,395 images. Each scenario is defined by two logical constraints selected from quantity, length, type, placement, and relation, with anomalies including both single-constraint and combined violations. We further propose a language-based anomaly detection framework that relies solely on text descriptions generated from normal images. Using contrastive learning with positive texts and contradiction-based negative texts synthesized from these descriptions, our method learns embeddings that capture logical attributes rather than low-level features. Extensive experiments demonstrate consistent improvements over baselines across the evaluated settings. The dataset is available at: https://github.com/nkthiroto/VID-AD.

Method: Proposes a novel framework with carefully designed learning objective using Pareto optimality to balance visual privacy and MLLM performance, and critical-history enhanced optimization to effectively optimize with black-box MLLMs where only input/output access is available.

Result: Experiments show the method is effective on different benchmarks, demonstrating successful privacy protection while maintaining MLLM performance.

Conclusion: The framework addresses the challenging problem of visual privacy protection in black-box MLLM services, providing a practical solution for users to enjoy MLLM benefits while safeguarding their privacy.

Abstract: The emergence of Multimodal Large Language Models (MLLMs) and the widespread usage of MLLM cloud services such as GPT-4V raised great concerns about privacy leakage in visual data. As these models are typically deployed in cloud services, users are required to submit their images and videos, posing serious privacy risks. However, how to tackle such privacy concerns is an under-explored problem. Thus, in this paper, we aim to conduct a new investigation to protect visual privacy when enjoying the convenience brought by MLLM services. We address the practical case where the MLLM is a “black box”, i.e., we only have access to its input and output without knowing its internal model information. To tackle such a challenging yet demanding problem, we propose a novel framework, in which we carefully design the learning objective with Pareto optimality to seek a better trade-off between visual privacy and MLLM’s performance, and propose critical-history enhanced optimization to effectively optimize the framework with the black-box MLLM. Our experiments show that our method is effective on different benchmarks.

Method: Presents first empirical evaluation of state-of-the-art feature-based Visual Anomaly Detection (VAD) methods on real planetary imagery. Introduces two benchmarks: lunar dataset from Lunar Reconnaissance Orbiter Camera imagery (fresh/degraded craters as anomalies) and Mars surface dataset reflecting rover-acquired imagery characteristics. Evaluates multiple VAD approaches with focus on computationally efficient, edge-oriented solutions suitable for onboard deployment.

Result: Results demonstrate that feature-based VAD methods can effectively identify rare planetary surface phenomena while remaining feasible for resource-constrained environments.

Conclusion: Establishes practical benchmarks and highlights potential of open-world perception systems to support mission-critical applications including tactical planning, landing site selection, hazard detection, bandwidth-aware data prioritization, and discovery of unanticipated geological processes.

Abstract: Space missions generate massive volumes of high-resolution orbital and surface imagery that far exceed the capacity for manual inspection. Detecting rare phenomena is scientifically critical, yet traditional supervised learning struggles due to scarce labeled examples and closed-world assumptions that prevent discovery of genuinely novel observations. In this work, we investigate Visual Anomaly Detection (VAD) as a framework for automated discovery in planetary exploration. We present the first empirical evaluation of state-of-the-art feature-based VAD methods on real planetary imagery, encompassing both orbital lunar data and Mars rover surface imagery. To support this evaluation, we introduce two benchmarks: (i) a lunar dataset derived from Lunar Reconnaissance Orbiter Camera Narrow Angle imagery, comprising of fresh and degraded craters as anomalies alongside normal terrain; and (ii) a Mars surface dataset designed to reflect the characteristics of rover-acquired imagery. We evaluate multiple VAD approaches with a focus on computationally efficient, edge-oriented solutions suitable for onboard deployment, applicable to both orbital platforms surveying the lunar surface and surface rovers operating on Mars. Our results demonstrate that feature-based VAD methods can effectively identify rare planetary surface phenomena while remaining feasible for resource-constrained environments. By grounding anomaly detection in planetary science, this work establishes practical benchmarks and highlights the potential of open-world perception systems to support a range of mission-critical applications, including tactical planning, landing site selection, hazard detection, bandwidth-aware data prioritization, and the discovery of unanticipated geological processes.

Method: Created behavioral benchmark with participants making same/different object judgments for dot pairs on natural scenes (1000+ trials). Tested diverse vision models using simple readout from representations to predict reaction times. Proposed novel metric to quantify object-centric structure by measuring patch similarity within/between objects. Used Gram matrix distillation to improve alignment.

Result: Steady improvement across model generations, with transformers trained with DINO self-supervised objective showing strongest performance. Stronger object-centric structure predicts human segmentation behavior more accurately. Gram matrix distillation improves alignment with human behavior, converging with findings that Gram anchoring improves DINOv3’s feature quality.

Conclusion: Self-supervised vision models capture object structure in a behaviorally human-like manner, and Gram matrix structure plays a key role in driving perceptual alignment between models and human vision.

Abstract: Vision foundation models trained with self-supervised objectives achieve strong performance across diverse tasks and exhibit emergent object segmentation properties. However, their alignment with human object perception remains poorly understood. Here, we introduce a behavioral benchmark in which participants make same/different object judgments for dot pairs on naturalistic scenes, scaling up a classical psychophysics paradigm to over 1000 trials. We test a diverse set of vision models using a simple readout from their representations to predict subjects’ reaction times. We observe a steady improvement across model generations, with both architecture and training objective contributing to alignment, and transformer-based models trained with the DINO self-supervised objective showing the strongest performance. To investigate the source of this improvement, we propose a novel metric to quantify the object-centric component of representations by measuring patch similarity within and between objects. Across models, stronger object-centric structure predicts human segmentation behavior more accurately. We further show that matching the Gram matrix of supervised transformer models, capturing similarity structure across image patches, with that of a self-supervised model through distillation improves their alignment with human behavior, converging with the prior finding that Gram anchoring improves DINOv3’s feature quality. Together, these results demonstrate that self-supervised vision models capture object structure in a behaviorally human-like manner, and that Gram matrix structure plays a role in driving perceptual alignment.

Method: 1) Introduces Meme Reappraisal task requiring emotion-controllable, structure-preserving multimodal transformation. 2) Constructs MER-Bench benchmark with real-world memes annotated with source/target emotions, rewritten text, visual editing specs, and taxonomy labels. 3) Proposes structured evaluation using MLLM-as-a-Judge paradigm assessing modality-level generation quality, affect controllability, structural fidelity, and global affective alignment.

Result: Extensive experiments show substantial gaps in existing image-editing and multimodal-generation systems in satisfying constraints of structural preservation, semantic consistency, and affective transformation. The benchmark reveals current limitations in controllable meme editing.

Conclusion: MER-Bench establishes a foundation for research on controllable meme editing and emotion-aware multimodal generation, highlighting the need for better multimodal systems that can handle complex constraints in meme transformation tasks.

Abstract: Memes represent a tightly coupled, multimodal form of social expression, in which visual context and overlaid text jointly convey nuanced affect and commentary. Inspired by cognitive reappraisal in psychology, we introduce Meme Reappraisal, a novel multimodal generation task that aims to transform negatively framed memes into constructive ones while preserving their underlying scenario, entities, and structural layout. Unlike prior works on meme understanding or generation, Meme Reappraisal requires emotion-controllable, structure-preserving multimodal transformation under multiple semantic and stylistic constraints. To support this task, we construct MER-Bench, a benchmark of real-world memes with fine-grained multimodal annotations, including source and target emotions, positively rewritten meme text, visual editing specifications, and taxonomy labels covering visual type, sentiment polarity, and layout structure. We further propose a structured evaluation framework based on a multimodal large language model (MLLM)-as-a-Judge paradigm, decomposing performance into modality-level generation quality, affect controllability, structural fidelity, and global affective alignment. Extensive experiments across representative image-editing and multimodal-generation systems reveal substantial gaps in satisfying the constraints of structural preservation, semantic consistency, and affective transformation. We believe MER-Bench establishes a foundation for research on controllable meme editing and emotion-aware multimodal generation. Our code is available at: https://github.com/one-seven17/MER-Bench.

Method: Uses polarization cues for reflection decomposition, introduces polarimetric deferred rendering (PolarDR) to model polarization by reflection, and develops a self-occlusion-aware environment map building technique (GridMap) for indirect lighting of non-convex objects.

Result: Achieves ~2 dB improvement in PSNR and 45.7% better cosine distance for surface normal reconstruction compared to RGB-based methods. Demonstrates state-of-the-art inverse rendering and relighting capability on synthetic and real-world scenes, including those with partial polarization cues.

Conclusion: PhyGaP successfully leverages polarization information to enable physically accurate reflection decomposition and high-fidelity relighting of reflective 3D objects, overcoming limitations of RGB-only approaches.

Abstract: Recent advances in 3D Gaussian Splatting (3DGS) have demonstrated great success in modeling reflective 3D objects and their interaction with the environment via deferred rendering (DR). However, existing methods often struggle with correctly reconstructing physical attributes such as albedo and reflectance, and therefore they do not support high-fidelity relighting. Observing that this limitation stems from the lack of shape and material information in RGB images, we present PhyGaP, a physically-grounded 3DGS method that leverages polarization cues to facilitate precise reflection decomposition and visually consistent relighting of reconstructed objects. Specifically, we design a polarimetric deferred rendering (PolarDR) process to model polarization by reflection, and a self-occlusion-aware environment map building technique (GridMap) to resolve indirect lighting of non-convex objects. We validate on multiple synthetic and real-world scenes, including those featuring only partial polarization cues, that PhyGaP not only excels in reconstructing the appearance and surface normal of reflective 3D objects (~2 dB in PSNR and 45.7% in Cosine Distance better than existing RGB-based methods on average), but also achieves state-of-the-art inverse rendering and relighting capability. Our code will be released soon.

Method: Frames semantic vector learning as subspace learning problem where latent vectors are approximated in lower-dimensional semantic subspace. Uses orthogonal non-negative constraints on semantic vectors and incorporates attribute boundary vectors to reduce entanglement. Proposes AIDC (Alternating Iterative Disentanglement and Controllability) algorithm with closed-form updates and provable convergence.

Result: Proposed framework offers effective adaptable solution for unsupervised facial attribute editing with improved disentanglement and controllability compared to existing methods.

Conclusion: U-Face provides novel approach to unsupervised facial attribute editing through subspace learning formulation with constraints that enhance disentanglement, addressing limitations of existing methods.

Abstract: Latent space-based facial attribute editing methods have gained popularity in applications such as digital entertainment, virtual avatar creation, and human-computer interaction systems due to their potential for efficient and flexible attribute manipulation, particularly for continuous edits. Among these, unsupervised latent space-based methods, which discover effective semantic vectors without relying on labeled data, have attracted considerable attention in the research community. However, existing methods still encounter difficulties in disentanglement, as manipulating a specific facial attribute may unintentionally affect other attributes, complicating fine-grained controllability. To address these challenges, we propose a novel framework designed to offer an effective and adaptable solution for unsupervised facial attribute editing, called Unsupervised Facial Attribute Controllable Editing (U-Face). The proposed method frames semantic vector learning as a subspace learning problem, where latent vectors are approximated within a lower-dimensional semantic subspace spanned by a semantic vector matrix. This formulation can also be equivalently interpreted from a projection-reconstruction perspective and further generalized into an autoencoder framework, providing a foundation that can support disentangled representation learning in a flexible manner. To improve disentanglement and controllability, we impose orthogonal non-negative constraints on the semantic vectors and incorporate attribute boundary vectors to reduce entanglement in the learned directions. Although these constraints make the optimization problem challenging, we design an alternating iterative algorithm, called Alternating Iterative Disentanglement and Controllability (AIDC), with closed-form updates and provable convergence under specific conditions.

Method: Real Distribution Bias Correction (RDBC) framework with two components: 1) Real Population Distribution Estimation module that uses i.i.d. property of real samples to derive their normal distribution statistics, and 2) Distribution-Sampled Feature Whitening module that amplifies Gaussianity gap between real and fake samples through sampling-based whitening operation.

Result: Extensive experiments show RDBC achieves state-of-the-art performance in both in-domain and cross-domain deepfake detection, demonstrating strong generalization to unseen target domains.

Conclusion: By focusing on the invariant properties of real data rather than trying to predict future forgery types, RDBC effectively captures real-world properties of real samples and enhances generalization to unseen domains, offering a promising approach for robust deepfake detection.

Abstract: To generalize deepfake detectors to future unseen forgeries, most existing methods attempt to simulate the dynamically evolving forgery types using available source domain data. However, predicting an unbounded set of future manipulations from limited prior examples is infeasible. To overcome this limitation, we propose to exploit the invariance of \textbf{real data} from two complementary perspectives: the fixed population distribution of the entire real class and the inherent Gaussianity of individual real images. Building on these properties, we introduce the Real Distribution Bias Correction (RDBC) framework, which consists of two key components: the Real Population Distribution Estimation module and the Distribution-Sampled Feature Whitening module. The former utilizes the independent and identically distributed (\iid) property of real samples to derive the normal distribution form of their statistics, from which the distribution parameters can be estimated using limited source domain data. Based on the learned population distribution, the latter utilizes the inherent Gaussianity of real data as a discriminative prior and performs a sampling-based whitening operation to amplify the Gaussianity gap between real and fake samples. Through synergistic coupling of the two modules, our model captures the real-world properties of real samples, thereby enhancing its generalizability to unseen target domains. Extensive experiments demonstrate that RDBC achieves state-of-the-art performance in both in-domain and cross-domain deepfake detection.

Method: Proposes a multi-grained vision-language alignment framework using CLIP with: 1) multi-grained language prompts describing different body parts, 2) adaptively masked multi-head self-attention to extract specific part features, and 3) MLLM-based visual grounding expert to automatically generate pseudo labels for body part supervision.

Result: Extensive experiments on single- and multi-source generalization protocols demonstrate superior performance compared to previous methods, showing strong generalization to unseen domains.

Conclusion: The proposed multi-grained vision-language alignment framework effectively addresses the fine-grained feature extraction problem in domain generalized person re-identification, leveraging CLIP’s generalization capabilities while overcoming its limitations for ID-sensitive tasks.

Abstract: Domain Generalized person Re-identification (DG Re-ID) is a challenging task, where models are trained on source domains but tested on unseen target domains. Although previous pure vision-based models have achieved significant progress, the performance remains further improved. Recently, Vision-Language Models (VLMs) present outstanding generalization capabilities in various visual applications. However, directly adapting a VLM to Re-ID shows limited generalization improvement. This is because the VLM only produces with global features that are insensitive to ID nuances. To tacle this problem, we propose a CLIP-based multi-grained vision-language alignment framework in this work. Specifically, several multi-grained prompts are introduced in language modality to describe different body parts and align with their counterparts in vision modality. To obtain fine-grained visual information, an adaptively masked multi-head self-attention module is employed to precisely extract specific part features. To train the proposed module, an MLLM-based visual grounding expert is employed to automatically generate pseudo labels of body parts for supervision. Extensive experiments conducted on both single- and multi-source generalization protocols demonstrate the superior performance of our approach. The implementation code will be released at https://github.com/RikoLi/MUVA.

Method: Uses distinct representations at different stages: Explode state for part completion and Implode state for geometry refinement. Incorporates self-attention mechanisms in both states to maintain structural coherence between parts and enable effective feature fusion among components.

Result: Extensive experiments on multiple benchmarks show EI-Part efficiently produces semantically meaningful and structurally coherent parts with fine-grained geometric details, achieving state-of-the-art performance in part-level 3D generation.

Conclusion: EI-Part successfully addresses challenges in part-level 3D generation by leveraging spatial resolution through Explode/Implode states and self-attention mechanisms, producing high-quality 3D shapes with components that exhibit strong structural coherence and geometric fidelity.

Abstract: Part-level 3D generation is crucial for various downstream applications, including gaming, film production, and industrial design. However, decomposing a 3D shape into geometrically plausible and meaningful components remains a significant challenge. Previous part-based generation methods often struggle to produce well-constructed parts, exhibiting poor structural coherence, geometric implausibility, inaccuracy, or inefficiency. To address these challenges, we introduce EI-Part, a novel framework specifically designed to generate high-quality 3D shapes with components, characterized by strong structural coherence, geometric plausibility, geometric fidelity, and generation efficiency. We propose utilizing distinct representations at different stages: an Explode state for part completion and an Implode state for geometry refinement. This strategy fully leverages spatial resolution, enabling flexible part completion and fine geometric detail generation. To maintain structural coherence between parts, a self-attention mechanism is incorporated in both exploded and imploded states, facilitating effective information perception and feature fusion among components during generation. Extensive experiments on multiple benchmarks demonstrate that EI-Part efficiently produces semantically meaningful and structurally coherent parts with fine-grained geometric details, achieving state-of-the-art performance in part-level 3D generation. Project page: https://cvhadessun.github.io/EI-Part/

Method: Propose a post-hoc pipeline to project Euclidean slot embeddings onto Lorentz hyperboloid without modifying training. Construct five-level visual hierarchies from slot attention masks and analyze whether hyperbolic geometry reveals latent structure invisible in Euclidean space. Integrate pipeline with SPOT (images), VideoSAUR (video), and SlotContrast (video).

Result: Hyperbolic projection exposes consistent scene-level to object-level organization where coarse slots occupy greater manifold depth than fine slots, absent in Euclidean space. Identify “curvature-task tradeoff”: low curvature (c=0.2) matches/outperforms Euclidean on parent slot retrieval, while moderate curvature (c=0.5) achieves better inter-level separation.

Conclusion: Slot representations already encode latent hierarchy that hyperbolic geometry reveals, motivating end-to-end hyperbolic training as a natural next step for object-centric learning.

Abstract: Slot attention has emerged as a powerful framework for unsupervised object-centric learning, decomposing visual scenes into a small set of compact vector representations called \emph{slots}, each capturing a distinct region or object. However, these slots are learned in Euclidean space, which provides no geometric inductive bias for the hierarchical relationships that naturally structure visual scenes. In this work, we propose a simple post-hoc pipeline to project Euclidean slot embeddings onto the Lorentz hyperboloid of hyperbolic space, without modifying the underlying training pipeline. We construct five-level visual hierarchies directly from slot attention masks and analyse whether hyperbolic geometry reveals latent hierarchical structure that remains invisible in Euclidean space. Integrating our pipeline with SPOT (images), VideoSAUR (video), and SlotContrast (video), We find that hyperbolic projection exposes a consistent scene-level to object-level organisation, where coarse slots occupy greater manifold depth than fine slots, which is absent in Euclidean space. We further identify a “curvature–task tradeoff”: low curvature ($c{=}0.2$) matches or outperforms Euclidean on parent slot retrieval, while moderate curvature ($c{=}0.5$) achieves better inter-level separation. Together, these findings suggest that slot representations already encode latent hierarchy that hyperbolic geometry reveals, motivating end-to-end hyperbolic training as a natural next step. Code and models are available at \href{https://github.com/NeeluMadan/HHS}{github.com/NeeluMadan/HHS}.

Method: Uses event-based light field cameras to capture multiple simultaneous views of a scene, combined with machine learning-based reconstruction algorithms that exploit the correlation differences: motion-induced events are strongly correlated across views while turbulence-induced events are weakly correlated.

Result: Tabletop experiments demonstrate the system can overcome strong turbulence while imaging high-speed objects traveling at up to 16,000 pixels per second, achieving the first system capable of such imaging at high frame rates.

Conclusion: Event-based light field cameras with machine learning enable robust high-speed imaging through atmospheric turbulence by leveraging multi-view correlation patterns to separate motion from turbulence effects.

Abstract: This work introduces and demonstrates the first system capable of imaging fast-moving extended non-rigid objects through strong atmospheric turbulence at high frame rate. Event cameras are a novel sensing architecture capable of estimating high-speed imagery at thousands of frames per second. However, on their own event cameras are unable to disambiguate scene motion from turbulence. In this work, we overcome this limitation using event-based light field cameras: By simultaneously capturing multiple views of a scene, event-based light field cameras and machine learning-based reconstruction algorithms are able to disambiguate motion-induced dynamics, which produce events that are strongly correlated across views, from turbulence-induced dynamics, which produce events that are weakly correlated across view. Tabletop experiments demonstrate event-based light field can overcome strong turbulence while imaging high-speed objects traveling at up to 16,000 pixels per second.

Method: Conducted simulation and real-world experiments using five commercial C-arm systems. Perturbed intrinsic parameters (focal length increased by 100-700 pixels, principal point by 20-200 pixels), then reconstructed 3D points, re-estimated extrinsic poses via optimization, and measured reconstruction/reprojection errors relative to ground truth.

Result: Even with focal length errors up to 500 pixels (~100 mm), mean 3D reconstruction error remained under 0.2 mm. Larger deviations (700 pixels) increased error to only ~0.3 mm. Principal point shifts up to 200 pixels introduced negligible reconstruction error after extrinsic re-optimization.

Conclusion: Moderate intrinsic calibration errors can be effectively mitigated by extrinsic re-optimization, preserving submillimeter 3D reconstruction accuracy. This tolerance suggests practical pathways to relax calibration precision requirements and simplify clinical workflow.

Abstract: \textbf{Purpose:} C-arm fluoroscopy’s 3D reconstruction relies on accurate intrinsic calibration, which is often challenging in clinical practice. This study ensures high-precision reconstruction accuracy by re-optimizing the extrinsic parameters to compensate for intrinsic calibration errors. \noindent\textbf{Methods:} We conducted both simulation and real-world experiments using five commercial C-arm systems. Intrinsic parameters were perturbed in controlled increments. Focal length was increased by 100 to 700 pixels ($\approx$20 mm to 140 mm) and principal point by 20 to 200 pixels. For each perturbation, we (1) reconstructed 3D points from known phantom geometries, (2) re-estimated extrinsic poses using standard optimization, and (3) measured reconstruction and reprojection errors relative to ground truth. \noindent\textbf{Results:} Even with focal length errors up to 500 pixels ($\approx$100 mm, assuming a nominal focal length of $\sim$1000 mm), mean 3D reconstruction error remained under 0.2 mm. Larger focal length deviations (700 pixels) elevated error to only $\approx$0.3 mm. Principal point shifts up to 200 pixels introduced negligible reconstruction error once extrinsic parameters were re-optimized, with reprojection error increases below 0.5 pixels. \noindent\textbf{Conclusion:} Moderate errors in intrinsic calibration can be effectively mitigated by extrinsic re-optimization, preserving submillimeter 3D reconstruction accuracy. This intrinsic tolerance suggests a practical pathway to relax calibration precision requirements, thereby simplifying C-arm system setup and reducing clinical workflow burden without compromising performance.

Method: Learns an observation-stable latent ocular state shared across modalities, unifying fine-grained parsing, structure-preserving cross-modality translation, and quality-robust enhancement. Uses longitudinal supervision for time-conditioned state transitions.

Result: Provides a unified framework for robust multimodal interpretation and prognosis-oriented simulation in ophthalmology, enabling forecasting of clinically meaningful progression while preserving stable anatomy.

Conclusion: By moving from static representation learning to explicit dynamical modeling, EyeWorld offers a novel approach to multimodal medical image analysis with temporal forecasting capabilities.

Abstract: Ophthalmic decision-making depends on subtle lesion-scale cues interpreted across multimodal imaging and over time, yet most medical foundation models remain static and degrade under modality and acquisition shifts. Here we introduce EyeWorld, a generative world model that conceptualizes the eye as a partially observed dynamical system grounded in clinical imaging. EyeWorld learns an observation-stable latent ocular state shared across modalities, unifying fine-grained parsing, structure-preserving cross-modality translation and quality-robust enhancement within a single framework. Longitudinal supervision further enables time-conditioned state transitions, supporting forecasting of clinically meaningful progression while preserving stable anatomy. By moving from static representation learning to explicit dynamical modeling, EyeWorld provides a unified approach to robust multimodal interpretation and prognosis-oriented simulation in medicine.

Method: Proposes TMPDiff framework with adaptive bisectioning-based algorithm that assigns per-step precisions with linear evaluation complexity, reducing an exponential search problem. Validates additive error accumulation hypothesis experimentally.

Result: Outperforms uniform-precision baselines at matched speedup with 10-20% improvement in perceptual quality. On FLUX.1-dev, achieves 90% SSIM relative to full-precision model at 2.5x speedup over 16-bit inference.

Conclusion: Temporal mixed-precision quantization effectively optimizes diffusion model inference by exploiting varying sensitivity to quantization errors across different denoising timesteps.

Abstract: Diffusion models are the go-to method for Text-to-Image generation, but their iterative denoising processes has high inference latency. Quantization reduces compute time by using lower bitwidths, but applies a fixed precision across all denoising timesteps, leaving an entire optimization axis unexplored. We propose TMPDiff, a temporal mixed-precision framework for diffusion models that assigns different numeric precision to different denoising timesteps. We hypothesize that quantization errors accumulate additively across timesteps, which we then validate experimentally. Based on our observations, we develop an adaptive bisectioning-based algorithm, which assigns per-step precisions with linear evaluation complexity, reducing an otherwise exponential search problem. Across four state-of-the-art diffusion models and three datasets, TMPDiff consistently outperforms uniform-precision baselines at matched speedup, achieving 10 to 20% improvement in perceptual quality. On FLUX.1-dev, TMPDiff achieves 90% SSIM relative to the full-precision model at a speedup of 2.5x over 16-bit inference.

Method: MotionCFG injects Gaussian noise into concept embeddings to create localized negative anchors representing sub-optimal motion variations. Uses contrastive guidance between target concept and noise-perturbed counterparts with piecewise guidance schedule confined to early denoising steps.

Result: Consistently improves motion dynamics across state-of-the-art T2V frameworks with negligible computational overhead and minimal visual quality compromise. Also effective for steering complex non-linear concepts like object numerosity.

Conclusion: Noise-induced contrastive guidance effectively enhances motion dynamics without content-motion drift, offering a lightweight solution for temporal refinement in video generation.

Abstract: Despite recent advances in Text-to-Video (T2V) synthesis, generating high-fidelity and dynamic motion remains a significant challenge. Existing methods primarily rely on Classifier-Free Guidance (CFG), often with explicit negative prompts (e.g. “static”, “blurry”), to suppress undesired artifacts. However, such explicit negations frequently introduce unintended semantic bias and distort object integrity; a phenomenon we define as Content-Motion Drift. To address this, we propose MotionCFG, a framework that enhances motion dynamics by contrasting a target concept with its noise-perturbed counterparts. Specifically, by injecting Gaussian noise into the concept embeddings, MotionCFG creates localized negative anchors that encapsulate a broad complementary space of sub-optimal motion variations. Unlike explicit negations, this approach facilitates implicit hard negative mining without shifting the global semantic identity, allowing for a focused refinement of temporal details. Combined with a piecewise guidance schedule that confines intervention to the early denoising steps, MotionCFG consistently improves motion dynamics across state-of-the-art T2V frameworks with negligible computational overhead and minimal compromise in visual quality. Additionally, we demonstrate that this noise-induced contrastive mechanism is effective not only for sharpening motion trajectories but also for steering complex, non-linear concepts such as precise object numerosity, which are typically difficult to modulate via standard text-based guidance.

Method: Proposes a novel self-supervised loss based on decision-theoretic perspective, minimizing Bayesian risk to obtain posterior mean and variance as optimal estimators for uncertainty quantification without ground-truth data.

Result: Validated on synthetic SkySat L1B dataset, produces calibrated uncertainty estimates comparable to supervised methods, bridging self-supervised restoration with uncertainty quantification.

Conclusion: Provides a practical framework for uncertainty-aware image reconstruction in satellite imagery super-resolution without requiring ground-truth high-resolution data.

Abstract: Super-resolution (SR) of satellite imagery is challenging due to the lack of paired low-/high-resolution data. Recent self-supervised SR methods overcome this limitation by exploiting the temporal redundancy in burst observations, but they lack a mechanism to quantify uncertainty in the reconstruction. In this work, we introduce a novel self-supervised loss that allows to estimate uncertainty in image super-resolution without ever accessing the ground-truth high-resolution data. We adopt a decision-theoretic perspective and show that minimizing the corresponding Bayesian risk yields the posterior mean and variance as optimal estimators. We validate our approach on a synthetic SkySat L1B dataset and demonstrate that it produces calibrated uncertainty estimates comparable to supervised methods. Our work bridges self-supervised restoration with uncertainty quantification, making a practical framework for uncertainty-aware image reconstruction.

Method: Proposes SGR-OCC with: 1) Soft-Gating Feature Lifter that models depth uncertainty via Gaussian gate to suppress background noise; 2) Dynamic Ray-Constrained Anchor Refinement that simplifies 3D displacement to 1D depth corrections along camera rays; 3) Two-Phase Progressive Training Strategy with identity-initialized fusion to resolve cold start problem.

Result: Achieves state-of-the-art on EmbodiedOcc-ScanNet and Occ-ScanNet benchmarks: 58.55% completion IoU and 49.89% semantic mIoU (surpassing previous best by 3.65% and 3.69%). In embodied prediction tasks: 55.72% SC-IoU and 46.22% mIoU. Shows superior structural integrity and boundary sharpness.

Conclusion: SGR-OCC effectively addresses key bottlenecks in online 3D semantic occupancy prediction through uncertainty-aware feature lifting and physically-constrained refinement with stable training, achieving significant performance improvements in both local and embodied prediction tasks.

Abstract: 3D semantic occupancy prediction is a cornerstone for embodied AI, enabling agents to perceive dense scene geometry and semantics incrementally from monocular video streams. However, current online frameworks face two critical bottlenecks: the inherent depth ambiguity of monocular estimation that causes “feature bleeding” at object boundaries , and the “cold start” instability where uninitialized temporal fusion layers distort high-quality spatial priors during early training stages. In this paper, we propose SGR-OCC (Soft-Gating and Ray-refinement Occupancy), a unified framework driven by the philosophy of “Inheritance and Evolution”. To perfectly inherit monocular spatial expertise, we introduce a Soft-Gating Feature Lifter that explicitly models depth uncertainty via a Gaussian gate to probabilistically suppress background noise. Furthermore, a Dynamic Ray-Constrained Anchor Refinement module simplifies complex 3D displacement searches into efficient 1D depth corrections along camera rays, ensuring sub-voxel adherence to physical surfaces. To ensure stable evolution toward temporal consistency, we employ a Two-Phase Progressive Training Strategy equipped with identity-initialized fusion, effectively resolving the cold start problem and shielding spatial priors from noisy early gradients. Extensive experiments on the EmbodiedOcc-ScanNet and Occ-ScanNet benchmarks demonstrate that SGR-OCC achieves state-of-the-art performance. In local prediction tasks, SGR-OCC achieves a completion IoU of 58.55$%$ and a semantic mIoU of 49.89$%$, surpassing the previous best method, EmbodiedOcc++, by 3.65$%$ and 3.69$%$ respectively. In challenging embodied prediction tasks, our model reaches 55.72$%$ SC-IoU and 46.22$%$ mIoU. Qualitative results further confirm our model’s superior capability in preserving structural integrity and boundary sharpness in complex indoor environments.

Method: Proposes AISSM (Adaptive Inference State Space Model) with a complementary dynamic confidence network that estimates signal-to-noise ratio and event density to dynamically adjust weighting between current and recent information. Also introduces a novel learning technique to improve training efficiency.

Result: Experimental results show that the AISSM system outperforms state-of-the-art models for event-based eye feature extraction.

Conclusion: The AISSM architecture effectively addresses the challenge of sudden event density changes in eye tracking, providing more robust feature extraction through adaptive inference mechanisms.

Abstract: Eye feature extraction from event-based data streams can be performed efficiently and with low energy consumption, offering great utility to real-world eye tracking pipelines. However, few eye feature extractors are designed to handle sudden changes in event density caused by the changes between gaze behaviors that vary in their kinematics, leading to degraded prediction performance. In this work, we address this problem by introducing the \emph{adaptive inference state space model} (AISSM), a novel architecture for feature extraction that is capable of dynamically adjusting the relative weight placed on current versus recent information. This relative weighting is determined via estimates of the signal-to-noise ratio and event density produced by a complementary \emph{dynamic confidence network}. Lastly, we craft and evaluate a novel learning technique that improves training efficiency. Experimental results demonstrate that the AISSM system outperforms state-of-the-art models for event-based eye feature extraction.

Method: The paper investigates DINO-style self-supervised pretraining directly on 3D medical imaging data to learn dense volumetric features suitable for deformable registration. The approach is evaluated on challenging interpatient abdominal registration tasks across MRI and CT modalities.

Result: Domain-specialized pretraining outperforms the DINOv2 model trained on natural images while requiring substantially lower computational resources at inference. It also surpasses established registration models under out-of-domain evaluation.

Conclusion: Task-agnostic yet medical imaging-focused pretraining provides robust and efficient 3D image registration, demonstrating the value of domain-specialized self-supervised learning for clinical applications.

Abstract: Medical image registration is a critical component of clinical imaging workflows, enabling accurate longitudinal assessment, multi-modal data fusion, and image-guided interventions. Intensity-based approaches often struggle with interscanner variability and complex anatomical deformations, whereas feature-based methods offer improved robustness by leveraging semantically informed representations. In this work, we investigate DINO-style self-supervised pretraining directly on 3D medical imaging data, aiming to learn dense volumetric features well suited for deformable registration. We assess the resulting representations on challenging interpatient abdominal registration task across both MRI and CT modalities. Our domain-specialized pretraining outperforms the DINOv2 model trained on a large-scale collection of natural images, while requiring substantially lower computational resources at inference time. Moreover, it surpasses established registration models under out-of-domain evaluation, demonstrating the value of task-agnostic yet medical imaging-focused pretraining for robust and efficient 3D image registration.

Method: Proposes SpaSemSR with two complementary guidances: 1) Spatial-grounded textual guidance integrates object-level spatial cues with semantic prompts to align textual and visual structures, reducing distortion. 2) Semantic-enhanced visual guidance uses a multi-encoder design with semantic degradation constraints to unify multimodal semantic priors. These guidances are adaptively fused via spatial-semantic attention during the diffusion process.

Result: Extensive experiments on multiple benchmarks show SpaSemSR achieves a superior perception-distortion balance, producing both realistic and faithful restorations compared to existing methods.

Conclusion: SpaSemSR successfully addresses the perception-distortion trade-off in image SR by combining spatial and semantic guidance within a diffusion framework, enabling realistic texture synthesis while maintaining fidelity to the input.

Abstract: Image super-resolution (SR) aims to reconstruct high resolution images with both high perceptual quality and low distortion, but is fundamentally limited by the perception-distortion trade-off. GAN-based SR methods reduce distortion but still struggle with realistic fine-grained textures, whereas diffusion-based approaches synthesize rich details but often deviate from the input, hallucinating structures and degrading fidelity. This tension raises a key challenge: how to exploit the powerful generative priors of diffusion models without sacrificing fidelity. To address this, we propose SpaSemSR, a spatial-semantic guided diffusion framework with two complementary guidances. First, spatial-grounded textual guidance integrates object-level spatial cues with semantic prompts, aligning textual and visual structures to reduce distortion. Second, semantic-enhanced visual guidance with a multi-encoder design and semantic degradation constraints unifies multimodal semantic priors, improving perceptual realism under severe degradations. These complementary guidances are adaptively fused into the diffusion process via spatial-semantic attention, suppressing distortion and hallucination while retaining the strengths of diffusion models. Extensive experiments on multiple benchmarks show that SpaSemSR achieves a superior perception-distortion balance, producing both realistic and faithful restorations.

Method: SIEVE is an end-to-end self-revisit framework that trains models to re-engage image evidence through internal representations. It automatically extracts embeddings of salient image regions and injects them into reasoning chains when additional grounding is needed. Uses reinforcement learning to teach when to trigger visual revisiting and which region embeddings to retrieve and insert.

Result: Experiments on multiple visual reasoning benchmarks show SIEVE yields consistent gains, improving performance by 8% on average across several benchmarks. Also evaluated on perception, reasoning, and hallucination tasks.

Conclusion: SIEVE demonstrates that VLMs can effectively leverage internal representations for visual evidence re-grounding without external tools, improving visual reasoning performance through learned self-revisit mechanisms.

Abstract: Vision language models (VLMs) are increasingly capable of reasoning over images, but robust visual reasoning often requires re-grounding intermediate steps in the underlying visual evidence. Recent approaches typically rely on external image operations such as zooming or cropping to re-access fine-grained details during inference, which requires additional image re-encoding and can disrupt the reasoning trajectory. We argue that VLMs already provide strong internal signals for identifying and reusing visual evidence, and that these signals can be directly leveraged to support image-grounded reasoning. Motivated by this insight, we propose an end-to-end self-revisit framework, SIEVE, that trains models to re-engage image evidence through internal representations. SIEVE automatically extracts embeddings of salient image regions and injects them into the reasoning chain when additional grounding is needed, enabling later steps to condition on relevant visual cues without external tool calls or re-encoding. We use reinforcement learning to teach the model when to trigger visual revisiting and which region embeddings to retrieve and insert during the reasoning process. Experiments on multiple visual reasoning benchmarks, together with perception, reasoning, and hallucination evaluations, show that SIEVE yields consistent gains, improving performance by 8 percent on average across several benchmarks.

Method: Proposes a k-space dual channel U-Net to jointly process real and imaginary components of undersampled k-space, restoring missing frequency content. Incorporates ensemble strategy for uncertainty maps and evaluates using out-of-distribution data.

Result: The k-space-driven approach outperforms spatial-domain and other state-of-the-art baselines, achieving image quality comparable to full high-field k-space acquisitions using out-of-distribution data.

Conclusion: This work presents a unified framework combining low-field MR image reconstruction, quality enhancement using undersampled k-space, and uncertainty quantification, demonstrating superior performance on out-of-distribution data.

Abstract: Low-field magnetic resonance imaging (MRI) offers affordable access to diagnostic imaging but faces challenges such as prolonged acquisition times and reduced image quality. Although accelerated imaging via k-space undersampling helps reduce scan time, image quality enhancement methods often rely on spatial-domain postprocessing. Deep learning achieved state-of-the-art results in both domains. However, most models are trained and evaluated using in-distribution (InD) data, creating a significant gap in understanding model performance when tested using out-of-distribution (OOD) data. To address these issues, we propose a novel framework that reconstructs high-field-like MR images directly from undersampled low-field MRI k-space, quantifies the impact of reduced sampling, and evaluates the generalisability of the model using OOD. Our approach utilises a k-space dual channel U-Net to jointly process the real and imaginary components of undersampled k-space, restoring missing frequency content, and incorporates an ensemble strategy to generate uncertainty maps. Experiments on low-field brain MRI demonstrate that our k-space-driven image quality enhancement outperforms the counterpart spatial-domain and other state-of-the-art baselines, achieving image quality comparable to full high-field k-space acquisitions using OOD data. To the best of our knowledge, this work is among the first to combine low-field MR image reconstruction, quality enhancement using undersampled k-space, and uncertainty quantification within a unified framework.

Method: Proposes a U-Net variant that operates directly in k-space to super-resolve low-field MR images from undersampled data. Unlike conventional approaches that treat super-resolution as postprocessing after image reconstruction, this unified model integrates both processes, leveraging k-space information directly.

Result: Extensive experiments on synthetic and real low-field brain MRI datasets show that k-space-driven image super-resolution outperforms conventional spatial-domain counterparts. Undersampled k-space reconstructions achieve comparable quality to full k-space acquisitions, enabling substantial scan-time acceleration without compromising diagnostic utility.

Conclusion: The proposed k-space-based unified framework effectively addresses both scan time reduction and image quality enhancement for low-field MRI, offering a practical solution for resource-limited settings while maintaining diagnostic quality.

Abstract: Low-field magnetic resonance imaging (MRI) offers a cost-effective alternative for medical imaging in resource-limited settings. However, its widespread adoption is hindered by two key challenges: prolonged scan times and reduced image quality. Accelerated acquisition can be achieved using k-space undersampling, while image enhancement traditionally relies on spatial-domain postprocessing. In this work, we propose a novel deep learning framework based on a U-Net variant that operates directly in k-space to super-resolve low-field MR images directly using undersampled data while quantifying the impact of reduced k-space sampling. Unlike conventional approaches that treat image super-resolution as a postprocessing step following image reconstruction from undersampled k-space, our unified model integrates both processes, leveraging k-space information to achieve superior image fidelity. Extensive experiments on synthetic and real low-field brain MRI datasets demonstrate that k-space-driven image super-resolution outperforms conventional spatial-domain counterparts. Furthermore, our results show that undersampled k-space reconstructions achieve comparable quality to full k-space acquisitions, enabling substantial scan-time acceleration without compromising diagnostic utility.

Method: Implements Schraml and Uhl’s method using local Fourier spectrum analysis to estimate pith (center point) from rough log end images, tested on two datasets

Result: Algorithm successfully implemented in Python and tested on two datasets, demonstrating practical applicability

Conclusion: The Python implementation provides a working tool for pith estimation that can be applied to wood processing applications

Abstract: In this article, we analyze and propose a Python implementation of the method “Pith Estimation on Rough Log End images using Local Fourier Spectrum Analysis”, by Rudolf Schraml and Andreas Uhl. The algorithm is tested over two datasets.

Method: Centered Reward Distillation (CRD) is derived from KL-regularized reward maximization using forward-process-based fine-tuning. Key techniques: 1) within-prompt centering to cancel intractable normalizing constants, 2) decoupling sampler from moving reference to prevent ratio-signal collapse, 3) KL anchoring to CFG-guided pretrained model to control long-run drift, and 4) reward-adaptive KL strength to balance early learning and late-stage exploitation.

Result: Experiments on text-to-image post-training with GenEval and OCR rewards show CRD achieves competitive SOTA reward optimization with fast convergence and reduced reward hacking, validated on unseen preference metrics.

Conclusion: CRD provides a robust diffusion RL framework that effectively addresses distribution drift and reward hacking problems in text-to-image model fine-tuning, enabling reliable optimization of external rewards while maintaining model integrity.

Abstract: Diffusion and flow models achieve State-Of-The-Art (SOTA) generative performance, yet many practically important behaviors such as fine-grained prompt fidelity, compositional correctness, and text rendering are weakly specified by score or flow matching pretraining objectives. Reinforcement Learning (RL) fine-tuning with external, black-box rewards is a natural remedy, but diffusion RL is often brittle. Trajectory-based methods incur high memory cost and high-variance gradient estimates; forward-process approaches converge faster but can suffer from distribution drift, and hence reward hacking. In this work, we present \textbf{Centered Reward Distillation (CRD)}, a diffusion RL framework derived from KL-regularized reward maximization built on forward-process-based fine-tuning. The key insight is that the intractable normalizing constant cancels under \emph{within-prompt centering}, yielding a well-posed reward-matching objective. To enable reliable text-to-image fine-tuning, we introduce techniques that explicitly control distribution drift: (\textit{i}) decoupling the sampler from the moving reference to prevent ratio-signal collapse, (\textit{ii}) KL anchoring to a CFG-guided pretrained model to control long-run drift and align with the inference-time semantics of the pre-trained model, and (\textit{iii}) reward-adaptive KL strength to accelerate early learning under large KL regularization while reducing late-stage exploitation of reward-model loopholes. Experiments on text-to-image post-training with \texttt{GenEval} and \texttt{OCR} rewards show that CRD achieves competitive SOTA reward optimization results with fast convergence and reduced reward hacking, as validated on unseen preference metrics.

Method: Proposes DualSwinFusionSeg with two parallel Swin Transformer V2 encoders for modality-specific feature extraction from RGB and auxiliary geophysical inputs, followed by multi-scale cross-modal fusion and UNet++ decoder with dense nested skip connections for fine boundary preservation.

Result: Achieves 0.867 mIoU and 0.905 F1 on development benchmark, 0.783 mIoU on held-out test set; modality-specific encoders and simple concatenation-based fusion improve segmentation accuracy under limited training data.

Conclusion: The proposed multimodal architecture effectively handles heterogeneous planetary data and demonstrates strong performance for Martian landslide segmentation, with potential applications in planetary surface analysis.

Abstract: Automated segmentation of Martian landslides, particularly in tectonically active regions such as Valles Marineris,is important for planetary geology, hazard assessment, and future robotic exploration. However, detecting landslides from planetary imagery is challenging due to the heterogeneous nature of available sensing modalities and the limited number of labeled samples. Each observation combines RGB imagery with geophysical measurements such as digital elevation models, slope maps, thermal inertia, and contextual grayscale imagery, which differ significantly in resolution and statistical properties. To address these challenges, we propose DualSwinFusionSeg, a multimodal segmentation architecture that separates modality-specific feature extraction and performs multi-scale cross-modal fusion. The model employs two parallel Swin Transformer V2 encoders to independently process RGB and auxiliary geophysical inputs, producing hierarchical feature representations. Corresponding features from the two streams are fused at multiple scales and decoded using a UNet++ decoder with dense nested skip connections to preserve fine boundary details. Extensive ablation studies evaluate modality contributions, loss functions, decoder architectures, and fusion strategies. Experiments on the MMLSv2 dataset from the PBVS 2026 Mars-LS Challenge show that modality-specific encoders and simple concatenation-based fusion improve segmentation accuracy under limited training data. The final model achieves 0.867 mIoU and 0.905 F1 on the development benchmark and 0.783 mIoU on the held-out test set, demonstrating strong performance for multimodal planetary surface segmentation.

Method: Uses plug-and-play module to select informative frame pairs maximizing viewpoint diversity while ensuring valid correspondence matching, followed by reconstruction model that simultaneously estimates RGB appearance, geometry, and semantics in real-time.

Result: Method effectively generates accurate 3D reconstructions and depth maps, enhancing culvert inspection efficiency with minimal human intervention.

Conclusion: The proposed pipeline enables efficient automated culvert inspection through robust 3D reconstruction in challenging repetitive environments.

Abstract: Automated culvert inspection systems can help increase the safety and efficiency of flood management operations. As a key step to this system, we present an efficient RGB-based 3D reconstruction pipeline for culvert-like structures in visually repetitive environments. Our approach first selects informative frame pairs to maximize viewpoint diversity while ensuring valid correspondence matching using a plug-and-play module, followed by a reconstruction model that simultaneously estimates RGB appearance, geometry, and semantics in real-time. Experiments demonstrate that our method effectively generates accurate 3D reconstructions and depth maps, enhancing culvert inspection efficiency with minimal human intervention.

Method: PRISM combines compound-aware supervision over mixed degradations with a weighted contrastive disentanglement objective that aligns primitives and their mixtures in latent space. This creates compositional geometry enabling high-fidelity joint removal of overlapping distortions while allowing flexible, targeted fixes through natural language prompts.

Result: PRISM outperforms state-of-the-art baselines on complex compound degradations across microscopy, wildlife monitoring, remote sensing, and urban weather datasets, including zero-shot mixtures not seen during training. Selective restoration significantly improves downstream scientific accuracy over standard “black-box” restoration.

Conclusion: PRISM establishes a generalizable and controllable framework for high-fidelity restoration in domains where scientific utility is a priority, enabling interpretable separation of degradations and targeted restoration through natural language interaction.

Abstract: Scientific and environmental imagery often suffer from complex mixtures of noise related to the sensor and the environment. Existing restoration methods typically remove one degradation at a time, leading to cascading artifacts, overcorrection, or loss of meaningful signal. In scientific applications, restoration must be able to simultaneously handle compound degradations while allowing experts to selectively remove subsets of distortions without erasing important features. To address these challenges, we present PRISM (Precision Restoration with Interpretable Separation of Mixtures). PRISM is a prompted conditional diffusion framework which combines compound-aware supervision over mixed degradations with a weighted contrastive disentanglement objective that aligns primitives and their mixtures in the latent space. This compositional geometry enables high-fidelity joint removal of overlapping distortions while also allowing flexible, targeted fixes through natural language prompts. Across microscopy, wildlife monitoring, remote sensing, and urban weather datasets, PRISM outperforms state-of-the-art baselines on complex compound degradations, including zero-shot mixtures not seen during training. Importantly, we show that selective restoration significantly improves downstream scientific accuracy in several domains over standard “black-box” restoration. These results establish PRISM as a generalizable and controllable framework for high-fidelity restoration in domains where scientific utility is a priority.

Method: Proposes SK-Adapter, a lightweight structural adapter network that encodes joint coordinates and topology into learnable tokens, injected into frozen 3D generation backbones via cross-attention. Also introduces Objaverse-TMS dataset of 24k text-mesh-skeleton pairs.

Result: Achieves robust structural control while preserving geometry and texture quality of foundation models, significantly outperforming existing baselines. Extends capability to local 3D editing with skeletal guidance.

Conclusion: SK-Adapter provides an effective framework for precise skeletal manipulation in native 3D generation, enabling structural control previously unattainable with text/image prompts alone.

Abstract: Native 3D generative models have achieved remarkable fidelity and speed, yet they suffer from a critical limitation: inability to prescribe precise structural articulations, where precise structural control within the native 3D space remains underexplored. This paper proposes SK-Adapter, a simple and yet highly efficient and effective framework that unlocks precise skeletal manipulation for native 3D generation. Moving beyond text or image prompts, which can be ambiguous for precise structure, we treat the 3D skeleton as a first-class control signal. SK-Adapter is a lightweight structural adapter network that encodes joint coordinates and topology into learnable tokens, which are injected into the frozen 3D generation backbone via cross-attention. This smart design allows the model to not only effectively “attend” to specific 3D structural constraints but also preserve its original generative priors. To bridge the data gap, we contribute Objaverse-TMS dataset, a large-scale dataset of 24k text-mesh-skeleton pairs. Extensive experiments confirm that our method achieves robust structural control while preserving the geometry and texture quality of the foundation model, significantly outperforming existing baselines. Furthermore, we extend this capability to local 3D editing, enabling the region specific editing of existing assets with skeletal guidance, which is unattainable by previous methods. Project Page: https://sk-adapter.github.io/

Method: Created Garments2Look dataset with 80K outfit pairs across 40 major categories and 300+ subcategories. Developed synthesis pipeline with heuristic outfit construction, automated filtering, and human validation. Adapted SOTA VTON methods and general-purpose image editing models as baselines.

Result: Current methods struggle with complete outfit try-on, showing difficulties in seamless integration, correct layering inference, and styling, leading to misalignment and artifacts.

Conclusion: Outfit-level VTON presents significant challenges beyond single-garment approaches, highlighting the need for new methods that can handle complex outfit composition, layering, and styling.

Abstract: Virtual try-on (VTON) has advanced single-garment visualization, yet real-world fashion centers on full outfits with multiple garments, accessories, fine-grained categories, layering, and diverse styling, remaining beyond current VTON systems. Existing datasets are category-limited and lack outfit diversity. We introduce Garments2Look, the first large-scale multimodal dataset for outfit-level VTON, comprising 80K many-garments-to-one-look pairs across 40 major categories and 300+ fine-grained subcategories. Each pair includes an outfit with 3-12 reference garment images (Average 4.48), a model image wearing the outfit, and detailed item and try-on textual annotations. To balance authenticity and diversity, we propose a synthesis pipeline. It involves heuristically constructing outfit lists before generating try-on results, with the entire process subjected to strict automated filtering and human validation to ensure data quality. To probe task difficulty, we adapt SOTA VTON methods and general-purpose image editing models to establish baselines. Results show current methods struggle to try on complete outfits seamlessly and to infer correct layering and styling, leading to misalignment and artifacts.

Method: Uses unpaired blurred and sharp images of similar scenes, employs a dense matching model to identify correspondences between blurry images and reference sharp images to generate pseudo-ground truth data for training.

Result: Achieves state-of-the-art performance in image deblurring, demonstrating effectiveness without requiring paired training data or pre-trained networks.

Conclusion: The novel unsupervised approach provides a more adaptable and practical solution for image deblurring that works across various scenarios and network sizes, including low-resource devices.

Abstract: This paper introduces a novel unsupervised approach for image deblurring that utilizes a simple process for training data collection, thereby enhancing the applicability and effectiveness of deblurring methods. Our technique does not require meticulously paired data of blurred and corresponding sharp images; instead, it uses unpaired blurred and sharp images of similar scenes to generate pseudo-ground truth data by leveraging a dense matching model to identify correspondences between a blurry image and reference sharp images. Thanks to the simplicity of the training data collection process, our approach does not rely on existing paired training data or pre-trained networks, making it more adaptable to various scenarios and suitable for networks of different sizes, including those designed for low-resource devices. We demonstrate that this novel approach achieves state-of-the-art performance, marking a significant advancement in the field of image deblurring.

Method: Analyzed attention maps to identify attention dispersion, found correlation between overall attention on image tokens and spatial dispersiveness, then proposed VRGA framework that selects visual heads based on entropy-focus criterion and reweights their attention to guide focus on relevant regions.

Result: Extensive experiments on vision-language benchmarks show VRGA effectively alleviates perceptual degradation, improves visual grounding and reasoning accuracy, and provides interpretable insights into MLLM visual processing.

Conclusion: Attention dispersion is a key issue in MLLM reasoning, and the proposed training-free VRGA framework successfully addresses this by guiding attention to relevant visual regions, improving performance while maintaining interpretability.

Abstract: Multimodal large language models (MLLMs) often suffer from perceptual impairments under extended reasoning modes, particularly in visual question answering (VQA) tasks. We identify attention dispersion as the underlying cause: during multi-step reasoning, the model’s visual attention becomes scattered and drifts away from question-relevant regions, effectively “losing focus” on the visual input. To better understand this phenomenon, we analyze the attention maps of MLLMs and observe that reasoning prompts significantly reduce attention to regions critical for answering the question. We further find a strong correlation between the model’s overall attention on image tokens and the spatial dispersiveness of its attention within the image. Leveraging this insight, we propose a training-free Visual Region-Guided Attention (VRGA) framework that selects visual heads based on an entropy-focus criterion and reweights their attention, effectively guiding the model to focus on question-relevant regions during reasoning. Extensive experiments on vision-language benchmarks demonstrate that our method effectively alleviates perceptual degradation, leading to improvements in visual grounding and reasoning accuracy while providing interpretable insights into how MLLMs process visual information.

Method: Benchmarked 8 models (one-step flows, multi-step baselines, established systems) under controlled class-conditional protocol on ImageNet validation, ImageNetV2, and reLAIONet. Used FID, Inception Score, CLIP Score, and Pick Score. Introduced MinMax Harmonic Mean (MMHM) as composite proxy metric.

Result: FID-focused development can be misleading in few-step regimes - guidance changes can improve FID while degrading text-image alignment and human preference. One-step models benefit from step scaling and become more competitive under multi-step inference, though still show local distortions.

Conclusion: Need more holistic evaluation beyond FID, one-step models can scale to multi-step inference, and MMHM helps stabilize hyperparameter selection across guidance and step sweeps.

Abstract: State-of-the-art text-to-image models produce high-quality images, but inference remains expensive as generation requires several sequential ODE or denoising steps. Native one-step models aim to reduce this cost by mapping noise to an image in a single step, yet fair comparisons to multi-step systems are difficult because studies use mismatched sampling steps and different classifier-free guidance (CFG) settings, where CFG can shift FID, Inception Score, and CLIP-based alignment in opposing directions. It is also unclear how well one-step models scale to multi-step inference, and there is limited standardized out-of-distribution evaluation for label-ID-conditioned generators beyond ImageNet. To address this, We benchmark eight models spanning one-step flows (MeanFlow, Improved MeanFlow, SoFlow), multi-step baselines (RAE, Scale-RAE), and established systems (SiT, Stable Diffusion 3.5, FLUX.1) under a controlled class-conditional protocol on ImageNet validation, ImageNetV2, and reLAIONet, our new proofread out-of-distribution dataset aligned to ImageNet label IDs. Using FID, Inception Score, CLIP Score, and Pick Score, we show that FID-focused model development and CFG selection can be misleading in few-step regimes, where guidance changes can improve FID while degrading text-image alignment and human preference signals and worsening perceived quality. We further show that leading one-step models benefit from step scaling and become substantially more competitive under multi-step inference, although they still exhibit characteristic local distortions. To capture these tradeoffs, we introduce MinMax Harmonic Mean (MMHM), a composite proxy over all four metrics that stabilizes hyperparameter selection across guidance and step sweeps.

Method: End-to-end deep learning model trained on time-to-event outcomes using H&E-stained whole-slide prostatectomy specimens. Evaluated across four independent international cohorts with integration with CAPRA-S clinical risk score.

Result: Model demonstrated robust generalization across institutions and populations. When combined with CAPRA-S, improved discrimination for BCR with concordance indices increasing from 0.725-0.772 to 0.749-0.788 across cohorts.

Conclusion: Deep learning applied to routine prostate histopathology can deliver reproducible, clinically generalizable biomarkers that augment postoperative risk stratification, supporting personalized prostate cancer management.

Abstract: Accurate prediction of biochemical recurrence (BCR) after radical prostatectomy is critical for guiding adjuvant treatment and surveillance decisions in prostate cancer. However, existing clinicopathological risk models reduce complex morphology to relatively coarse descriptors, leaving substantial prognostic information embedded in routine histopathology underexplored. We present a deep learning-based biomarker that predicts continuous, patient-specific risk of BCR directly from H&E-stained whole-slide prostatectomy specimens. Trained end-to-end on time-to-event outcomes and evaluated across four independent international cohorts, our model demonstrates robust generalization across institutions and patient populations. When integrated with the CAPRA-S clinical risk score, the deep learning risk score consistently improved discrimination for BCR, increasing concordance indices from 0.725-0.772 to 0.749-0.788 across cohorts. To support clinical interpretability, outcome-grounded analyses revealed subtle histomorphological patterns associated with recurrence risk that are not captured by conventional clinicopathological risk scores. This multicohort study demonstrates that deep learning applied to routine prostate histopathology can deliver reproducible and clinically generalizable biomarkers that augment postoperative risk stratification, with potential to support personalized management of prostate cancer in real-world clinical settings.

Method: Proposes IMO with mid-level fusion of fundus and OCT features, cross-modal feature alignment module to reduce modality discrepancies, and iterative refinement decoder using denoising diffusion mechanism for progressive optimization.

Result: Extensive experiments show effective multimodal feature integration, providing comprehensive and clinically significant approach to glaucoma assessment with fine-grained optic disc/cup segmentation and accurate grading.

Conclusion: IMO successfully integrates multimodal features for joint segmentation and grading, offering improved glaucoma assessment through cross-modal alignment and diffusion-based iterative refinement.

Abstract: Accurate diagnosis of glaucoma is challenging, as early-stage changes are subtle and often lack clear structural or appearance cues. Most existing approaches rely on a single modality, such as fundus or optical coherence tomography (OCT), capturing only partial pathological information and often missing early disease progression. In this paper, we propose an iterative multimodal optimization model (IMO) for joint segmentation and grading. IMO integrates fundus and OCT features through a mid-level fusion strategy, enhanced by a cross-modal feature alignment (CMFA) module to reduce modality discrepancies. An iterative refinement decoder progressively optimizes the multimodal features through a denoising diffusion mechanism, enabling fine-grained segmentation of the optic disc and cup while supporting accurate glaucoma grading. Extensive experiments show that our method effectively integrates multimodal features, providing a comprehensive and clinically significant approach to glaucoma assessment. Source codes are available at https://github.com/warren-wzw/IMO.git.

Method: Proposes LRGait benchmark and EMGaitNet framework with CLIP-based Semantic Mining module, Semantic-Guided Alignment to bridge 2D-3D gap, Symmetric Cross-Attention Fusion for hierarchical feature integration, and Spatio-Temporal module for gait dynamics.

Result: Extensive experiments on various gait datasets validate the effectiveness of the proposed method for robust long-range multimodal gait recognition.

Conclusion: The work presents the first LiDAR-Camera multimodal benchmark for long-range gait recognition with a novel semantic-guided fusion framework that effectively addresses modality gaps and improves performance in diverse outdoor scenarios.

Abstract: Gait recognition is an emerging biometric technology that enables non-intrusive and hard-to-spoof human identification. However, most existing methods are confined to short-range, unimodal settings and fail to generalize to long-range and cross-distance scenarios under real-world conditions. To address this gap, we present \textbf{LRGait}, the first LiDAR-Camera multimodal benchmark designed for robust long-range gait recognition across diverse outdoor distances and environments. We further propose \textbf{EMGaitNet}, an end-to-end framework tailored for long-range multimodal gait recognition. To bridge the modality gap between RGB images and point clouds, we introduce a semantic-guided fusion pipeline. A CLIP-based Semantic Mining (SeMi) module first extracts human body-part-aware semantic cues, which are then employed to align 2D and 3D features via a Semantic-Guided Alignment (SGA) module within a unified embedding space. A Symmetric Cross-Attention Fusion (SCAF) module hierarchically integrates visual contours and 3D geometric features, and a Spatio-Temporal (ST) module captures global gait dynamics. Extensive experiments on various gait datasets validate the effectiveness of our method.

Method: Proposes SDAVS with two key components: 1) Selective Noise-Resilient Processor (SNRP) that suppresses audio noise while enhancing relevant auditory cues, and 2) Discriminative Audio-Visual Mutual Fusion (DAMF) strategy for more consistent audio-visual representations.

Result: Achieves state-of-the-art performance on benchmark AVS datasets, with particularly strong results in multi-source and complex scenes where audio-visual relationships are challenging.

Conclusion: The proposed SNRP and DAMF components effectively address audio noise and discriminative audio-visual interaction, advancing AVS capabilities for real-world complex scenarios.

Abstract: The ability to capture and segment sounding objects in dynamic visual scenes is crucial for the development of Audio-Visual Segmentation (AVS) tasks. While significant progress has been made in this area, the interaction between audio and visual modalities still requires further exploration. In this work, we aim to answer the following questions: How can a model effectively suppress audio noise while enhancing relevant audio information? How can we achieve discriminative interaction between the audio and visual modalities? To this end, we propose SDAVS, equipped with the Selective Noise-Resilient Processor (SNRP) module and the Discriminative Audio-Visual Mutual Fusion (DAMF) strategy. The proposed SNRP mitigates audio noise interference by selectively emphasizing relevant auditory cues, while DAMF ensures more consistent audio-visual representations. Experimental results demonstrate that our proposed method achieves state-of-the-art performance on benchmark AVS datasets, especially in multi-source and complex scenes. \textit{The code and model are available at https://github.com/happylife-pk/SDAVS}.

Method: DualTSR uses a single multimodal transformer backbone trained with dual diffusion objectives: Conditional Flow Matching for continuous high-resolution image distribution modeling and discrete diffusion for textual content distribution. This shared architecture enables visual and textual information interaction at every layer.

Result: Experiments on synthetic Chinese benchmarks and curated real-world evaluation show DualTSR achieves strong perceptual quality and text fidelity, outperforming prior methods with simpler architecture.

Conclusion: DualTSR provides a unified end-to-end solution for STISR that eliminates dependency on external OCR models and simplifies complex multi-branch architectures while maintaining strong performance in both visual quality and text recognition accuracy.

Abstract: Scene Text Image Super-Resolution (STISR) aims to restore high-resolution details in low-resolution text images, which is crucial for both human readability and machine recognition. Existing methods, however, often depend on external Optical Character Recognition (OCR) models for textual priors or rely on complex multi-component architectures that are difficult to train and reproduce. In this paper, we introduce DualTSR, a unified end-to-end framework that addresses both issues. DualTSR employs a single multimodal transformer backbone trained with a dual diffusion objective. It simultaneously models the continuous distribution of high-resolution images via Conditional Flow Matching and the discrete distribution of textual content via discrete diffusion. This shared design enables visual and textual information to interact at every layer, allowing the model to infer text priors internally instead of relying on an external OCR module. Compared with prior multi-branch diffusion systems, DualTSR offers a simpler end-to-end formulation with fewer hand-crafted components. Experiments on synthetic Chinese benchmarks and a curated real-world evaluation protocol show that DualTSR achieves strong perceptual quality and text fidelity.

Method: Introduces a skeleton-based spatial control representation that encodes only data-encoding information, allowing easy incorporation of reference visuals. Built on Diffusion Transformer (DiT) with adaptive position encoding and Spatially Gated Attention to manage spatial and subject controls. Created a dataset of 30,000 triplets for fine-tuning.

Result: Developed a domain-specific diffusion model capable of generating pictorial charts with both spatial alignment to chart structures and subject-driven control respecting visual characteristics of reference images. Proposed a unified data accuracy metric for evaluation.

Conclusion: Demonstrates that generative models can achieve data-driven visual storytelling by moving beyond general-purpose conditions to task-specific representations, enabling the creation of pictorial charts that maintain both data faithfulness and visual aesthetics.

Abstract: A pictorial chart is an effective medium for visual storytelling, seamlessly integrating visual elements with data charts. However, creating such images is challenging because the flexibility of visual elements often conflicts with the rigidity of chart structures. This process thus requires a creative deformation that maintains both data faithfulness and visual aesthetics. Current methods that extract dense structural cues from natural images (e.g., edge or depth maps) are ill-suited as conditioning signals for pictorial chart generation. We present ChArtist, a domain-specific diffusion model for generating pictorial charts automatically, offering two distinct types of control: 1) spatial control that aligns well with the chart structure, and 2) subject-driven control that respects the visual characteristics of a reference image. To achieve this, we introduce a skeleton-based spatial control representation. This representation encodes only the data-encoding information of the chart, allowing for the easy incorporation of reference visuals without a rigid outline constraint. We implement our method based on the Diffusion Transformer (DiT) and leverage an adaptive position encoding mechanism to manage these two controls. We further introduce Spatially Gated Attention to modulate the interaction between spatial control and subject control. To support the fine-tuning of pre-trained models for this task, we created a large-scale dataset of 30,000 triplets (skeleton, reference image, pictorial chart). We also propose a unified data accuracy metric to evaluate the data faithfulness of the generated charts. We believe this work demonstrates that current generative models can achieve data-driven visual storytelling by moving beyond general-purpose conditions to task-specific representations. Project page: https://chartist-ai.github.io/.

Method: 1) Uses DINOv3 for modality-consistent feature extraction to create shared semantic space; 2) Introduces reconstruction-alignment loss to maintain consistency between fused outputs and inputs; 3) Employs bilevel optimization to decouple and jointly optimize reconstruction and fusion objectives.

Result: Extensive experiments show UniFusion achieves superior visual quality, generalization ability, and adaptability to real-world scenarios across multiple fusion tasks compared to existing methods.

Conclusion: UniFusion successfully addresses the limitations of task-specific fusion methods by providing a unified framework with strong cross-task generalization and effective source information preservation.

Abstract: Image fusion aims to integrate complementary information from multiple source images to produce a more informative and visually consistent representation, benefiting both human perception and downstream vision tasks. Despite recent progress, most existing fusion methods are designed for specific tasks (i.e., multi-modal, multi-exposure, or multi-focus fusion) and struggle to effectively preserve source information during the fusion process. This limitation primarily arises from task-specific architectures and the degradation of source information caused by deep-layer propagation. To overcome these issues, we propose UniFusion, a unified image fusion framework designed to achieve cross-task generalization. First, leveraging DINOv3 for modality-consistent feature extraction, UniFusion establishes a shared semantic space for diverse inputs. Second, to preserve the understanding of each source image, we introduce a reconstruction-alignment loss to maintain consistency between fused outputs and inputs. Finally, we employ a bilevel optimization strategy to decouple and jointly optimize reconstruction and fusion objectives, effectively balancing their coupling relationship and ensuring smooth convergence. Extensive experiments across multiple fusion tasks demonstrate UniFusion’s superior visual quality, generalization ability, and adaptability to real-world scenarios. Code is available at https://github.com/dusongcheng/UniFusion.

Method: Introduces Safety-Potential Pruning, a one-shot pruning framework that amplifies safety-relevant activations by identifying and removing weights that are less responsive to safety prompts. The method requires no additional retraining and works by exposing and amplifying dormant safety pathways within VLM architectures.

Result: Across three representative VLM architectures and three jailbreak benchmarks, the method reduces attack success rates by up to 22% relative to prompting alone while maintaining strong benign performance. The approach demonstrates effectiveness across different model architectures and attack scenarios.

Conclusion: Pruning can be framed not only as a model compression technique but as a structural intervention to emerge alignment-relevant subnets, offering a new path to robust jailbreak resistance in vision-language models.

Abstract: Safety prompts constitute an interpretable layer of defense against jailbreak attacks in vision-language models (VLMs); however, their efficacy is constrained by the models’ latent structural responsiveness. We observe that such prompts consistently engage a sparse set of parameters that remain largely quiescent during benign use. This finding motivates the Safety Subnetwork Hypothesis: VLMs embed structurally distinct pathways capable of enforcing safety, but these pathways remain dormant without explicit stimulation. To expose and amplify these pathways, we introduce Safety-Potential Pruning, a one-shot pruning framework that amplifies safety-relevant activations by removing weights that are less responsive to safety prompts without additional retraining. Across three representative VLM architectures and three jailbreak benchmarks, our method reduces attack success rates by up to 22% relative to prompting alone, all while maintaining strong benign performance. These findings frame pruning not only as a model compression technique, but as a structural intervention to emerge alignment-relevant subnets, offering a new path to robust jailbreak resistance.

Method: Proposes Forgery Identification via Noise Disturbance (FIND) - a simple binary classifier approach. Key innovation: adding Gaussian noise to real images during training and labeling them as synthetic, forcing the classifier to learn statistical patterns distinguishing real from synthetic images.

Result: Improves performance by 11.7% on GenImage benchmark while running 126x faster than existing methods. Eliminates need for auxiliary diffusion models and reconstruction computations.

Conclusion: FIND offers a practical, efficient, and generalizable way to detect diffusion-generated content by directly targeting core distributional differences rather than relying on reconstruction error.

Abstract: The remarkable realism of images generated by diffusion models poses critical detection challenges. Current methods utilize reconstruction error as a discriminative feature, exploiting the observation that real images exhibit higher reconstruction errors when processed through diffusion models. However, these approaches require costly reconstruction computations and depend on specific diffusion models, making their performance highly model-dependent. We identify a fundamental difference: real images are more difficult to fit with Gaussian distributions compared to synthetic ones. In this paper, we propose Forgery Identification via Noise Disturbance (FIND), a novel method that requires only a simple binary classifier. It eliminates reconstruction by directly targeting the core distributional difference between real and synthetic images. Our key operation is to add Gaussian noise to real images during training and label these noisy versions as synthetic. This step allows the classifier to focus on the statistical patterns that distinguish real from synthetic images. We theoretically prove that the noise-augmented real images resemble diffusion-generated images in their ease of Gaussian fitting. Furthermore, simply by adding noise, they still retain visual similarity to the original images, highlighting the most discriminative distribution-related features. The proposed FIND improves performance by 11.7% on the GenImage benchmark while running 126x faster than existing methods. By removing the need for auxiliary diffusion models and reconstruction, it offers a practical, efficient, and generalizable way to detect diffusion-generated content.

Method: Dual-component design: (1) Information Bottleneck-guided filter to prune task-irrelevant directions, (2) lightweight graph-based coordinator for inter-layer consistency during training only.

Result: Outperforms vanilla LoRA and advanced methods across LLMs, VLMs, and vision models (LLaMA, LLaVA, ViT), especially in low-rank/low-data regimes, with zero inference cost increase.

Conclusion: StructLoRA advances PEFT from parameter compression to optimizing information quality and structural integrity, establishing new SOTA with practical deployment advantages.

Abstract: Low-Rank Adaptation (LoRA) has become a cornerstone of parameter-efficient fine-tuning (PEFT). Yet, its efficacy is hampered by two fundamental limitations: semantic drift, by treating all update directions with equal importance, and structural incoherence, from adapting layers independently, resulting in suboptimal, uncoordinated updates. To remedy these, we propose StructLoRA, a framework that addresses both limitations through a principled, dual-component design: (1) an Information Bottleneck-guided filter that prunes task-irrelevant directions to mitigate semantic drift, and (2) a lightweight, training-only graph-based coordinator that enforces inter-layer consistency to resolve structural incoherence. Extensive experiments across large language model , vision language model, and vision model (including LLaMA, LLaVA, and ViT) demonstrate that StructLoRA consistently establishes a new state-of-the-art, outperforming not only vanilla LoRA but also advanced dynamic rank allocation and sparsity-based methods. Notably, the benefits are particularly pronounced in challenging low-rank and low-data regimes. Crucially, since our proposed modules operate only during training, StructLoRA enhances performance with zero additional inference cost, advancing the focus of PEFT – from mere parameter compression to a more holistic optimization of information quality and structural integrity.

Method: Proposes Streaming Semantic Gaussian Splatting (S2GS) with geometry-semantic decoupled dual-backbone design: geometry branch performs causal modeling for incremental Gaussian updates; semantic branch uses 2D foundation vision model and query-driven decoder for segmentation and identity embeddings, stabilized by query-level contrastive alignment and lightweight online association with instance memory.

Result: S2GS matches or outperforms strong offline baselines on joint reconstruction-and-understanding benchmarks while significantly improving long-horizon scalability - processes 1,000+ frames with slower growth in runtime/GPU memory, whereas offline baselines typically run out of memory at around 80 frames.

Conclusion: S2GS enables scalable online joint reconstruction and understanding through strictly causal, incremental 3D Gaussian semantic field framework that doesn’t require reprocessing historical frames.

Abstract: Existing offline feed-forward methods for joint scene understanding and reconstruction on long image streams often repeatedly perform global computation over an ever-growing set of past observations, causing runtime and GPU memory to increase rapidly with sequence length and limiting scalability. We propose Streaming Semantic Gaussian Splatting (S2GS), a strictly causal, incremental 3D Gaussian semantic field framework: it does not leverage future frames and continuously updates scene geometry, appearance, and instance-level semantics without reprocessing historical frames, enabling scalable online joint reconstruction and understanding. S2GS adopts a geometry-semantic decoupled dual-backbone design: the geometry branch performs causal modeling to drive incremental Gaussian updates, while the semantic branch leverages a 2D foundation vision model and a query-driven decoder to predict segmentation masks and identity embeddings, further stabilized by query-level contrastive alignment and lightweight online association with an instance memory. Experiments show that S2GS matches or outperforms strong offline baselines on joint reconstruction-and-understanding benchmarks, while significantly improving long-horizon scalability: it processes 1,000+ frames with much slower growth in runtime and GPU memory, whereas offline global-processing baselines typically run out of memory at around 80 frames under the same setting.

Method: 1) Creates first FG-DG-GCD benchmarks using controlled diffusion-adapter stylization to generate painting and sketch domains for CUB-200-2011, Stanford Cars, and FGVC-Aircraft. 2) Proposes FoCUS framework with Domain-Consistent Parts Discovery (DCPD) for geometry-stable part reasoning and Uncertainty-Aware Feature Augmentation (UFA) for confidence-calibrated feature regularization using uncertainty-guided perturbations.

Result: FoCUS outperforms strong GCD, FG-GCD, and DG-GCD baselines by 3.28%, 9.68%, and 2.07% respectively in clustering accuracy on proposed benchmarks. Also achieves nearly 3x higher computational efficiency than state-of-the-art while remaining competitive on coarse-grained DG-GCD tasks.

Conclusion: FG-DG-GCD addresses critical real-world challenges in open-world recognition under domain shift, with FoCUS providing effective solution through domain-consistent part discovery and uncertainty-aware regularization, advancing fine-grained visual understanding in unseen domains.

Abstract: We introduce the first unified framework for Fine-Grained Domain-Generalized Generalized Category Discovery (FG-DG-GCD), bringing open-world recognition closer to real-world deployment under domain shift. Unlike conventional GCD, which assumes labeled and unlabeled data come from the same distribution, DG-GCD learns only from labeled source data and must both recognize known classes and discover novel ones in unseen, unlabeled target domains. This problem is especially challenging in fine-grained settings, where subtle inter-class differences and large intra-class variation make domain generalization significantly harder. To support systematic evaluation, we establish the first FG-DG-GCD benchmarks by creating identity-preserving painting and sketch domains for CUB-200-2011, Stanford Cars, and FGVC-Aircraft using controlled diffusion-adapter stylization. On top of this ,we propose FoCUS, a single-stage framework that combines Domain-Consistent Parts Discovery (DCPD) for geometry-stable part reasoning with Uncertainty-Aware Feature Augmentation (UFA) for confidence-calibrated feature regularization through uncertainty-guided perturbations. Extensive experiments show that FoCUS outperforms strong GCD, FG-GCD, and DG-GCD baselines by 3.28%, 9.68%, and 2.07%, respectively, in clustering accuracy on the proposed benchmarks. It also remains competitive on coarse-grained DG-GCD tasks while achieving nearly 3x higher computational efficiency than the current state of the art. ^[Code and datasets will be released upon acceptance.]

Method: Proposes Tucker Adaptation (TuKA) which represents multi-hierarchical navigation knowledge as a high-order tensor and uses Tucker decomposition to decouple knowledge into shared subspaces and scenario-specific experts. Also introduces a decoupled knowledge incremental learning strategy to consolidate shared subspaces while constraining specific experts for decoupled lifelong learning.

Result: The AlldayWalker agent built on TuKA consistently outperforms state-of-the-art baselines in all-day multi-scenes lifelong VLN (AML-VLN) experiments, demonstrating effective continual learning across multiple navigation scenarios.

Conclusion: TuKA effectively addresses the AML-VLN problem by capturing multi-hierarchical navigation knowledge through tensor representation and Tucker decomposition, enabling VLN agents to learn continuously across diverse scenes without catastrophic forgetting.

Abstract: Deploying vision-and-language navigation (VLN) agents requires adaptation across diverse scenes and environments, but fine-tuning on a specific scenario often causes catastrophic forgetting in others, which severely limits flexible long-term deployment. We formalize this challenge as the all-day multi-scenes lifelong VLN (AML-VLN) problem. Existing parameter-efficient adapters (e.g., LoRA and its variants) are limited by their two-dimensional matrix form, which fails to capture the multi-hierarchical navigation knowledge spanning multiple scenes and environments. To address this, we propose Tucker Adaptation (TuKA), which represents the multi-hierarchical navigation knowledge as a high-order tensor and leverages Tucker decomposition to decouple the knowledge into shared subspaces and scenario-specific experts. We further introduce a decoupled knowledge incremental learning strategy to consolidate shared subspaces while constraining specific experts for decoupled lifelong learning. Building on TuKA, we also develop a VLN agent named AlldayWalker, which continually learns across multiple navigation scenarios, achieving all-day multi-scenes navigation. Extensive experiments show that AlldayWalker consistently outperforms state-of-the-art baselines.

Method: Uses a unified video diffusion model architecture that takes a reference image, camera trajectory, and environment map as input. Generates temporally coherent videos with both relit novel-view frames and corresponding albedo frames in a single generative process.

Result: Achieves high-fidelity outputs comparable to state-of-the-art methods in both novel view synthesis and relighting tasks. Produces temporally coherent and spatially aligned outputs without sacrificing visual quality in either task.

Conclusion: A single generative model can effectively integrate camera and lighting control, simplifying video generation pipelines while maintaining competitive performance and consistent realism.

Abstract: We present CamLit, the first unified video diffusion model that jointly performs novel view synthesis (NVS) and relighting from a single input image. Given one reference image, a user-defined camera trajectory, and an environment map, CamLit synthesizes a video of the scene from new viewpoints under the specified illumination. Within a single generative process, our model produces temporally coherent and spatially aligned outputs, including relit novel-view frames and corresponding albedo frames, enabling high-quality control of both camera pose and lighting. Qualitative and quantitative experiments demonstrate that CamLit achieves high-fidelity outputs on par with state-of-the-art methods in both novel view synthesis and relighting, without sacrificing visual quality in either task. We show that a single generative model can effectively integrate camera and lighting control, simplifying the video generation pipeline while maintaining competitive performance and consistent realism.

Method: Proposes BIT network with encoder-decoder architecture: encoder extracts preliminary features, decoder performs bi-directional feature integration and query-aware scoring to enhance cross-modality correspondence through explicit pairwise matching.

Result: Extensive experiments on several benchmarks demonstrate state-of-the-art performance, highlighting BIT’s effectiveness in VI-ReID task.

Conclusion: BIT successfully addresses modality gap challenges in VI-ReID through explicit pairwise interaction modeling rather than rigid feature alignment, representing a novel matching-driven approach.

Abstract: Visible-Infrared Person Re-Identification (VI-ReID) is a challenging retrieval task due to the substantial modality gap between visible and infrared images. While existing methods attempt to bridge this gap by learning modality-invariant features within a shared embedding space, they often overlook the complex and implicit correlations between modalities. This limitation becomes more severe under distribution shifts, where infrared samples are often far fewer than visible ones. To address these challenges, we propose a novel network termed Bi-directional Interaction Transformation (BIT). Instead of relying on rigid feature alignment, BIT adopts a matching-based strategy that explicitly models the interaction between visible and infrared image pairs. Specifically, BIT employs an encoder-decoder architecture where the encoder extracts preliminary feature representations, and the decoder performs bi-directional feature integration and query aware scoring to enhance cross-modality correspondence. To our best knowledge, BIT is the first to introduce such pairwise matching-driven interaction in VI-ReID. Extensive experiments on several benchmarks demonstrate that our BIT achieves state-of-the-art performance, highlighting its effectiveness in the VI-ReID task.

Method: Probe intermediate denoising representations of pretrained Diffusion Transformer (DiT), train lightweight physics verifier on frozen features, and use progressive trajectory selection to prune low-scoring candidates during inference

Result: Physically plausible and implausible videos are partially separable in mid-layer feature space; progressive trajectory selection improves physical consistency while reducing inference cost, achieving comparable results to Best-of-K sampling with fewer denoising steps

Conclusion: Video diffusion models encode recoverable physics-related cues, enabling efficient inference-time strategies for improving physical plausibility

Abstract: Do video diffusion models encode signals predictive of physical plausibility? We probe intermediate denoising representations of a pretrained Diffusion Transformer (DiT) and find that physically plausible and implausible videos are partially separable in mid-layer feature space across noise levels. This separability cannot be fully attributed to visual quality or generator identity, suggesting recoverable physics-related cues in frozen DiT features. Leveraging this observation, we introduce progressive trajectory selection, an inference-time strategy that scores parallel denoising trajectories at a few intermediate checkpoints using a lightweight physics verifier trained on frozen features, and prunes low-scoring candidates early. Extensive experiments on PhyGenBench demonstrate that our method improves physical consistency while reducing inference cost, achieving comparable results to Best-of-K sampling with substantially fewer denoising steps.

Method: Proposes OAHuman with a decoupling-perception paradigm that separates geometry reconstruction and texture synthesis. Geometry reconstruction is perceptually reinforced even in occluded areas, isolating it from texture interference. Texture synthesis learns exclusively from visible regions to prevent texture errors from transferring to occluded areas.

Result: Extensive experiments on occlusion-rich benchmarks show OAHuman achieves superior performance in structural completeness, surface detail, and texture realism, significantly improving monocular 3D human reconstruction under occlusion conditions.

Conclusion: OAHuman’s occlusion-aware framework with explicit decoupling of geometry and texture enables robust and high-fidelity 3D human reconstruction from single RGB images under challenging occlusion conditions.

Abstract: Monocular 3D human reconstruction in real-world scenarios remains highly challenging due to frequent occlusions from surrounding objects, people, or image truncation. Such occlusions lead to missing geometry and unreliable appearance cues, severely degrading the completeness and realism of reconstructed human models. Although recent neural implicit methods achieve impressive results on clean inputs, they struggle under occlusion due to entangled modeling of shape and texture. In this paper, we propose OAHuman, an occlusion-aware framework that explicitly decouples geometry reconstruction and texture synthesis for robust 3D human modeling from a single RGB image. The core innovation lies in the decoupling-perception paradigm, which addresses the fundamental issue of geometry-texture cross-contamination in occluded regions. Our framework ensures that geometry reconstruction is perceptually reinforced even in occluded areas, isolating it from texture interference. In parallel, texture synthesis is learned exclusively from visible regions, preventing texture errors from being transferred to the occluded areas. This decoupling approach enables OAHuman to achieve robust and high-fidelity reconstruction under occlusion, which has been a long-standing challenge in the field. Extensive experiments on occlusion-rich benchmarks demonstrate that OAHuman achieves superior performance in terms of structural completeness, surface detail, and texture realism, significantly improving monocular 3D human reconstruction under occlusion conditions.

Method: Proposes 4D Synchronized Fields - a 4D Gaussian representation that learns object-factored motion in-loop during reconstruction. Each Gaussian trajectory is decomposed into shared object motion plus implicit residual. A kinematic-conditioned ridge map predicts temporal semantic variation, creating a single representation where reconstruction, motion, and semantics are structurally coupled.

Result: On HyperNeRF: 28.52 dB mean PSNR (highest among language-grounded and motion-aware baselines, within 1.5 dB of reconstruction-only methods). On temporal-state retrieval: 0.884 mean accuracy, 0.815 mean vIoU, 0.733 mean tIoU, surpassing 4D LangSplat and LangSplat. Kinematic conditioning accounts for +0.45 tIoU over static-embedding-only baseline.

Conclusion: 4D Synchronized Fields is the only method that jointly exposes interpretable motion primitives and temporally grounded language fields from a single trained representation, enabling open-vocabulary temporal queries that retrieve both objects and moments.

Abstract: Current 4D representations decouple geometry, motion, and semantics: reconstruction methods discard interpretable motion structure; language-grounded methods attach semantics after motion is learned, blind to how objects move; and motion-aware methods encode dynamics as opaque per-point residuals without object-level organization. We propose 4D Synchronized Fields, a 4D Gaussian representation that learns object-factored motion in-loop during reconstruction and synchronizes language to the resulting kinematics through a per-object conditioned field. Each Gaussian trajectory is decomposed into shared object motion plus an implicit residual, and a kinematic-conditioned ridge map predicts temporal semantic variation, yielding a single representation in which reconstruction, motion, and semantics are structurally coupled and enabling open-vocabulary temporal queries that retrieve both objects and moments. On HyperNeRF, 4D Synchronized Fields achieves 28.52 dB mean PSNR, the highest among all language-grounded and motion-aware baselines, within 1.5 dB of reconstruction-only methods. On targeted temporal-state retrieval, the kinematic-conditioned field attains 0.884 mean accuracy, 0.815 mean vIoU, and 0.733 mean tIoU, surpassing 4D LangSplat (0.620, 0.433, and 0.439 respectively) and LangSplat (0.415, 0.304, and 0.262). Ablation confirms that kinematic conditioning is the primary driver, accounting for +0.45 tIoU over a static-embedding-only baseline. 4D Synchronized Fields is the only method that jointly exposes interpretable motion primitives and temporally grounded language fields from a single trained representation. Code will be released.

Method: Combines a mistake detector that processes recent frames and anticipates future visual features with a reinforcement learning policy that aggregates detector outputs over time to decide when to exit early with final prediction.

Result: Superior mistake detection accuracy while reducing the fraction of video observed compared to state-of-the-art models on diverse real-world procedural video datasets.

Conclusion: MistExit effectively enables early mistake detection in procedural videos through joint optimization of detection and early exiting, balancing accuracy and efficiency.

Abstract: We introduce the task of early mistake detection in video, where the goal is to determine whether a keystep in a procedural activity is performed correctly while observing as little of the streaming video as possible. To tackle this problem, we propose a method comprising a mistake detector and a reinforcement learning policy. At each timestep, the detector processes recently observed frames to estimate the keystep’s correctness while anticipating future visual features, enabling reliable early mistake estimates. Meanwhile, the policy aggregates the detector outputs and visual observations over time and adaptively decides when to exit (i.e., stop processing incoming frames) while producing the final prediction. Using diverse real-world procedural video datasets, we demonstrate that our MistExit model achieves superior mistake detection accuracy while reducing the fraction of video observed compared to state-of-the-art models. Project: https://vision.cs.utexas.edu/projects/mist_exit.

Method: Created VGMED dataset with clinical guidance to assess visual grounding; introduced quantitative metrics and qualitative analyses; proposed VGRefine inference-time method to refine attention distribution for better visual grounding.

Result: Found medical MLLMs fail to ground predictions in clinically relevant image regions; VGRefine achieved SOTA performance across 6 diverse Med-VQA benchmarks (110K+ samples, 8 imaging modalities) without additional training.

Conclusion: Inadequate visual grounding is a key factor in medical MLLMs’ underperformance; VGRefine effectively addresses this issue and improves medical image interpretation capabilities.

Abstract: Generalist multimodal large language models (MLLMs) have achieved impressive performance across a wide range of vision-language tasks. However, their performance on medical tasks, particularly in zero-shot settings where generalization is critical, remains suboptimal. A key research gap is the limited understanding of why medical MLLMs underperform in medical image interpretation. In this work, we present a pioneering systematic investigation into the visual grounding capabilities of state-of-the-art medical MLLMs. To disentangle visual grounding from semantic grounding, we design VGMED, a novel evaluation dataset developed with expert clinical guidance, explicitly assessing the visual grounding capability of medical MLLMs. We introduce new quantitative metrics and conduct detailed qualitative analyses. Our study across eight state-of-the-art (SOTA) medical MLLMs validates that they often fail to ground their predictions in clinically relevant image regions. We note that this finding is specific to medical image analysis; in contrast, prior work has shown that MLLMs are capable of grounding their predictions in the correct image regions when applied to natural scene images. Motivated by these findings, we propose VGRefine, a simple yet effective inference-time method that refines attention distribution to improve visual grounding in medical settings. Our approach achieves SOTA performance across 6 diverse Med-VQA benchmarks (over 110K VQA samples from 8 imaging modalities) without requiring additional training or external expert models. Overall, our work, for the first time, systematically validates inadequate visual grounding as one of the key contributing factors for medical MLLMs’ under-performance. Additional experiments are included in the Supp.

Method: Uses zeroth-order optimization (ZOO) with forward passes only. Two key components: 1) Distribution-Robust Layer Selection identifies and freezes layers with distribution-invariant features, updating only domain-sensitive layers; 2) Spatial Feature Aggregation Alignment stabilizes ZOO by aligning globally aggregated spatial features between source and target domains.

Result: Outperforms or matches BP-based methods on ImageNet-C/R/Sketch/A benchmarks. Reduces memory usage by 84% and improves accuracy by 3.9% over SAR on ImageNet-C. Enables architecture-agnostic and stable BP-free adaptation.

Conclusion: ZOTTA provides an efficient, BP-free test-time adaptation framework that achieves competitive performance with significantly reduced computational overhead, making it suitable for deployment on edge devices with non-differentiable models.

Abstract: Test-time adaptation (TTA) aims to improve model robustness under distribution shifts by adapting to unlabeled test data, but most existing methods rely on backpropagation (BP), which is computationally costly and incompatible with non-differentiable models such as quantized models, limiting practical deployment on numerous edge devices. Recent BP-free approaches alleviate overhead but remain either architecture-specific or limited in optimization capacity to handle high-dimensional models. We propose ZOTTA, a fully BP-free TTA framework that performs efficient adaptation using only forward passes via Zeroth-Order Optimization (ZOO). While ZOO is theoretically appealing, naive application leads to slow convergence under high-dimensional parameter spaces and unstable optimization due to the lack of labels. ZOTTA overcomes these challenges through 1) Distribution-Robust Layer Selection, which automatically identifies and freezes layers that already extract distribution-invariant features, updating only domain-sensitive layers to reduce the optimization dimensionality and accelerate convergence; 2) Spatial Feature Aggregation Alignment, which stabilizes ZOO by aligning globally aggregated spatial features between source and target to reduce gradient variance. Together, these components enable architecture-agnostic and stable BP-free adaptation. Extensive experiments on ImageNet-C/R/Sketch/A show that ZOTTA outperforms or matches BP-based methods, e.g., it reduces memory usage by 84% and improves accuracy by 3.9% over SAR on ImageNet-C.

Method: 1) Created AgroOmni dataset (288K multi-view corpus) capturing diverse spatial scales; 2) Proposed AgroNVILA with Perception-Reasoning Decoupling architecture: View-Conditioned Meta-Net for spatial context injection and Agriculture-aware Relative Policy Optimization for expert-aligned reasoning.

Result: AgroNVILA outperforms state-of-the-art MLLMs with +15.18% improvement in multi-altitude agricultural reasoning, demonstrating robust capability for holistic agricultural spatial planning.

Conclusion: The proposed approach effectively addresses scale confusion in agricultural multimodal reasoning through specialized architecture and training data, enabling better spatial understanding across varying scales.

Abstract: Agricultural multimodal reasoning requires robust spatial understanding across varying scales, from ground-level close-ups to top-down UAV and satellite imagery. Existing Multi-modal Large Language Models (MLLMs) suffer from a significant “terrestrial-centric” bias, causing scale confusion and logic drift during complex agricultural planning. To address this, we introduce the first large-scale AgroOmni (288K), a multi-view training corpus designed to capture diverse spatial topologies and scales in modern precision agriculture. Built on this dataset, we propose AgroNVILA, an MLLM that utilizes a novel Perception-Reasoning Decoupling (PRD) architecture. On the perception side, we incorporate a View-Conditioned Meta-Net (VCMN), which injects macroscopic spatial context into visual tokens, resolving scale ambiguities with minimal computational overhead. On the reasoning side, Agriculture-aware Relative Policy Optimization (ARPO) leverages reinforcement learning to align the model’s decision-making with expert agricultural logic, preventing statistical shortcuts. Extensive experiments demonstrate that AgroNVILA outperforms state-of-the-art MLLMs, achieving significant improvements (+15.18%) in multi-altitude agricultural reasoning, reflecting its robust capability for holistic agricultural spatial planning.

Method: Proposes a framework conceptualizing adaptation as post-pretraining intervention, organizing approaches into five mechanisms: parameter-, representation-, objective-, data-centric, and architectural/sequence-level adaptation. Analyzes trade-offs across adaptation depth, label efficiency, domain robustness, computational cost, auditability, and regulatory burden.

Result: Synthesizes evidence across classification, segmentation, and detection tasks, showing how adaptation strategies influence clinically relevant failure modes rather than just benchmark performance. Examines interaction between adaptation choices and validation protocols, calibration stability, multi-institutional deployment, and regulatory oversight.

Conclusion: Provides practical guidance for designing robust, auditable, and clinically deployable FM-based systems by reframing adaptation as controlled representational change under clinical constraints.

Abstract: Foundation models (FMs) have demonstrated strong transferability across medical imaging tasks, yet their clinical utility depends critically on how pretrained representations are adapted to domain-specific data, supervision regimes, and deployment constraints. Prior surveys primarily emphasize architectural advances and application coverage, while the mechanisms of adaptation and their implications for robustness, calibration, and regulatory feasibility remain insufficiently structured. This review introduces a strategy-centric framework for FM adaptation in medical image analysis (MIA). We conceptualize adaptation as a post-pretraining intervention and organize existing approaches into five mechanisms: parameter-, representation-, objective-, data-centric, and architectural/sequence-level adaptation. For each mechanism, we analyze trade-offs in adaptation depth, label efficiency, domain robustness, computational cost, auditability, and regulatory burden. We synthesize evidence across classification, segmentation, and detection tasks, highlighting how adaptation strategies influence clinically relevant failure modes rather than only aggregate benchmark performance. Finally, we examine how adaptation choices interact with validation protocols, calibration stability, multi-institutional deployment, and regulatory oversight. By reframing adaptation as a process of controlled representational change under clinical constraints, this review provides practical guidance for designing FM-based systems that are robust, auditable, and compatible with clinical deployment.

Method: Proposes AerialVLA with: 1) Streamlined dual-view perception strategy reducing visual redundancy while preserving essential navigation cues, 2) Fuzzy directional prompting mechanism derived solely from onboard sensors eliminating dependency on dense oracle guidance, 3) Unified control space integrating continuous 3-DoF kinematic commands with intrinsic landing signal.

Result: Achieves state-of-the-art performance on TravelUAV benchmark in seen environments and exhibits superior generalization in unseen scenarios with nearly three times the success rate of leading baselines.

Conclusion: A minimalist, autonomy-centric paradigm captures more robust visual-motor representations than complex modular systems, enabling genuine autonomous UAV navigation through direct mapping of vision-language inputs to continuous control signals.

Abstract: Vision-Language Navigation (VLN) for Unmanned Aerial Vehicles (UAVs) demands complex visual interpretation and continuous control in dynamic 3D environments. Existing hierarchical approaches rely on dense oracle guidance or auxiliary object detectors, creating semantic gaps and limiting genuine autonomy. We propose AerialVLA, a minimalist end-to-end Vision-Language-Action framework mapping raw visual observations and fuzzy linguistic instructions directly to continuous physical control signals. First, we introduce a streamlined dual-view perception strategy that reduces visual redundancy while preserving essential cues for forward navigation and precise grounding, which additionally facilitates future simulation-to-reality transfer. To reclaim genuine autonomy, we deploy a fuzzy directional prompting mechanism derived solely from onboard sensors, completely eliminating the dependency on dense oracle guidance. Ultimately, we formulate a unified control space that integrates continuous 3-Degree-of-Freedom (3-DoF) kinematic commands with an intrinsic landing signal, freeing the agent from external object detectors for precision landing. Extensive experiments on the TravelUAV benchmark demonstrate that AerialVLA achieves state-of-the-art performance in seen environments. Furthermore, it exhibits superior generalization in unseen scenarios by achieving nearly three times the success rate of leading baselines, validating that a minimalist, autonomy-centric paradigm captures more robust visual-motor representations than complex modular systems.

Method: Proposes Decoupled Vision Transformer (DC-ViT) with Decoupled Self-Attention (DSA) that decomposes token updates into two pathways: spatial updates for intra-channel structure modeling and channel-wise updates for adaptive cross-channel information integration. Also introduces Decoupled Aggregation (DAG) to learn task-specific channel importances.

Result: Extensive experiments across three MCI benchmarks demonstrate consistent improvements over existing MC-ViT approaches.

Conclusion: DC-ViT effectively addresses feature dilution in multi-channel imaging by regulating information sharing through decoupled attention mechanisms, preserving channel-specific semantics while enabling selective inter-channel interactions.

Abstract: Training and evaluation in multi-channel imaging (MCI) remains challenging due to heterogeneous channel configurations arising from varying staining protocols, sensor types, and acquisition settings. This heterogeneity limits the applicability of fixed-channel encoders commonly used in general computer vision. Recent Multi-Channel Vision Transformers (MC-ViTs) address this by enabling flexible channel inputs, typically by jointly encoding patch tokens from all channels within a unified attention space. However, unrestricted token interactions across channels can lead to feature dilution, reducing the ability to preserve channel-specific semantics that are critical in MCI data. To address this, we propose Decoupled Vision Transformer (DC-ViT), which explicitly regulates information sharing using Decoupled Self-Attention (DSA), which decomposes token updates into two complementary pathways: spatial updates that model intra-channel structure, and channel-wise updates that adaptively integrate cross-channel information. This decoupling mitigates informational collapse while allowing selective inter-channel interaction. To further exploit these enhanced channel-specific representations, we introduce Decoupled Aggregation (DAG), which allows the model to learn task-specific channel importances. Extensive experiments across three MCI benchmarks demonstrate consistent improvements over existing MC-ViT approaches.

Method: Three key innovations: 1) Multi-scale Receptive Field Reweighting to mitigate aliasing and preserve high-frequency details; 2) Multi-scale Unified Semantic Enhancer (MUSE) integrating local appearance with global context; 3) Multi-Periodic Texture Contrast Enhancement (MPTCE) module modeling texture disruptions in frequency domain to decouple anomalies from structured backgrounds.

Result: Achieves state-of-the-art performance on real-world industrial datasets, surpassing baseline by 8.3% in mAP50-95 and 7.7% in recall, while reducing false positive rate by approximately 8.6%.

Conclusion: TexWDS demonstrates robustness in handling complex periodic patterns and suitability for high-precision manufacturing inspection, addressing key challenges in wafer defect segmentation through texture-aware multi-scale and frequency-domain approaches.

Abstract: Wafer defect segmentation is pivotal for semiconductor yield optimization yet remains challenged by the intrinsic conflict between microscale anomalies and highly periodic, overwhelming background textures. Existing deep learning paradigms often falter due to feature dilution during downsampling and the lack of explicit mechanisms to disentangle low-contrast defects from process-induced noise. To transcend these limitations, we propose TexWDS, a texture-aware framework that harmonizes multi-scale feature retention with frequency-domain perturbation modeling. Our methodology incorporates three strategic innovations: (1) A Multi-scale Receptive Field Reweighting strategy is introduced to mitigate aliasing effects and preserve high-frequency details of micro-defects often lost in standard pyramidal architectures. (2) The Multi-scale Unified Semantic Enhancer (MUSE) integrates local appearance with global context encoding, effectively enhancing feature discriminability in low-visibility regions. (3) Crucially, we design a plug-and-play Multi-Periodic Texture Contrast Enhancement (MPTCE) module. By modeling texture disruptions in the frequency domain, MPTCE explicitly decouples non-periodic anomalies from structured backgrounds, boosting contrast for camouflaged defects. Extensive experiments on real-world industrial datasets demonstrate that TexWDS achieves a new state-of-the-art, surpassing the baseline by 8.3% in mAP50-95 and 7.7% in recall, while reducing the false positive rate by approximately 8.6%. These results underscore the framework’s robustness in handling complex periodic patterns and its suitability for high-precision manufacturing inspection.

Method: Proposes Visual Chronometer, a predictor that recovers Physical Frames Per Second (PhyFPS) directly from visual dynamics of input videos. Trained via controlled temporal resampling to estimate true temporal scale implied by motion itself, bypassing unreliable metadata.

Result: Established two benchmarks (PhyFPS-Bench-Real and PhyFPS-Bench-Gen) revealing severe PhyFPS misalignment and temporal instability in state-of-the-art video generators. Demonstrated that applying PhyFPS corrections significantly improves human-perceived naturalness of AI-generated videos.

Conclusion: Visual Chronometer addresses critical temporal grounding problem in video generation, providing a method to estimate true physical motion speeds from visual dynamics, which improves realism and temporal consistency in generated videos.

Abstract: While recent generative video models have achieved remarkable visual realism and are being explored as world models, true physical simulation requires mastering both space and time. Current models can produce visually smooth kinematics, yet they lack a reliable internal motion pulse to ground these motions in a consistent, real-world time scale. This temporal ambiguity stems from the common practice of indiscriminately training on videos with vastly different real-world speeds, forcing them into standardized frame rates. This leads to what we term chronometric hallucination: generated sequences exhibit ambiguous, unstable, and uncontrollable physical motion speeds. To address this, we propose Visual Chronometer, a predictor that recovers the Physical Frames Per Second (PhyFPS) directly from the visual dynamics of an input video. Trained via controlled temporal resampling, our method estimates the true temporal scale implied by the motion itself, bypassing unreliable metadata. To systematically quantify this issue, we establish two benchmarks, PhyFPS-Bench-Real and PhyFPS-Bench-Gen. Our evaluations reveal a harsh reality: state-of-the-art video generators suffer from severe PhyFPS misalignment and temporal instability. Finally, we demonstrate that applying PhyFPS corrections significantly improves the human-perceived naturalness of AI-generated videos. Our project page is https://xiangbogaobarry.github.io/Visual_Chronometer/.

Method: Proposes RegFormer++ with: 1) Hierarchical projection-aware 2D transformer that projects 3D LiDAR points onto cylindrical surface for linear complexity processing while preserving 3D geometric information; 2) Bijective Association Transformer (BAT) combining cross attention and all-to-all point gathering to reduce wrong matches; 3) Feature-transformed optimal transport module for stable training and pose regression.

Result: Achieves state-of-the-art performance on KITTI, NuScenes, and Argoverse datasets in both accuracy and efficiency, demonstrating effectiveness for large-scale outdoor LiDAR registration.

Conclusion: RegFormer++ provides an end-to-end solution for large-scale point cloud registration that eliminates post-processing dependencies, offers high efficiency through 2D projection while maintaining 3D accuracy, and shows strong performance across multiple outdoor datasets.

Abstract: Although point cloud registration has achieved remarkable advances in object-level and indoor scenes, large-scale LiDAR registration methods has been rarely explored before. Challenges mainly arise from the huge point scale, complex point distribution, and numerous outliers within outdoor LiDAR scans. In addition, most existing registration works generally adopt a two-stage paradigm: They first find correspondences by extracting discriminative local descriptors and then leverage robust estimators (e.g. RANSAC) to filter outliers, which are highly dependent on well-designed descriptors and post-processing choices. To address these problems, we propose a novel end-to-end differential transformer network, termed RegFormer++, for large-scale point cloud alignment without requiring any further post-processing. Specifically, a hierarchical projection-aware 2D transformer with linear complexity is proposed to project raw LiDAR points onto a cylindrical surface and extract global point features, which can improve resilience to outliers due to long-range dependencies. Because we fill original 3D coordinates into 2D projected positions, our designed transformer can benefit from both high efficiency in 2D processing and accuracy from 3D geometric information. Furthermore, to effectively reduce wrong point matching, a Bijective Association Transformer (BAT) is designed, combining both cross attention and all-to-all point gathering. To improve training stability and robustness, a feature-transformed optimal transport module is also designed for regressing the final pose transformation. Extensive experiments on KITTI, NuScenes, and Argoverse datasets demonstrate that our model achieves state-of-the-art performance in terms of both accuracy and efficiency.

Method: Pathological Gait-conditioned Generative Adversarial Network (PGcGAN) that synthesizes pathology-specific gait sequences from 3D pose keypoint trajectories. Uses one-hot encoded pathology labels in both generator and discriminator for controlled synthesis across six gait categories. Generator employs conditional autoencoder architecture trained with adversarial and reconstruction objectives.

Result: Experiments on Pathological Gait Dataset show strong alignment between real and synthetic sequences via PCA/t-SNE analyses, visual kinematic inspection, and downstream classification. Augmenting real data with synthetic sequences improved pathological gait recognition across GRU, LSTM, and CNN models.

Conclusion: Pathology-conditioned gait synthesis can effectively support data augmentation in pathological gait analysis, addressing dataset limitations.

Abstract: Pathological gait analysis is constrained by limited and variable clinical datasets, which restrict the modeling of diverse gait impairments. To address this challenge, we propose a Pathological Gait-conditioned Generative Adversarial Network (PGcGAN) that synthesises pathology-specific gait sequences directly from observed 3D pose keypoint trajectories data. The framework incorporates one-hot encoded pathology labels within both the generator and discriminator, enabling controlled synthesis across six gait categories. The generator adopts a conditional autoencoder architecture trained with adversarial and reconstruction objectives to preserve structural and temporal gait characteristics. Experiments on the Pathological Gait Dataset demonstrate strong alignment between real and synthetic sequences through PCA and t-SNE analyses, visual kinematic inspection, and downstream classification tasks. Augmenting real data with synthetic sequences improved pathological gait recognition across GRU, LSTM, and CNN models, indicating that pathology-conditioned gait synthesis can effectively support data augmentation in pathological gait analysis.

Method: Proposes RL-ScanIQA with PPO-trained scanpath policy and quality assessor, using multi-level rewards (scanpath diversity, equator-biased priors) and distortion-space augmentation with rank-consistent losses for cross-dataset robustness.

Result: Extensive experiments on three benchmarks show RL-ScanIQA achieves superior in-dataset performance and cross-dataset generalization compared to existing methods.

Conclusion: The reinforcement learning framework successfully integrates scanpath generation and quality assessment for 360° IQA, demonstrating improved performance through end-to-end optimization and task-aligned viewing strategies.

Abstract: Blind 360°image quality assessment (IQA) aims to predict perceptual quality for panoramic images without a pristine reference. Unlike conventional planar images, 360°content in immersive environments restricts viewers to a limited viewport at any moment, making viewing behaviors critical to quality perception. Although existing scanpath-based approaches have attempted to model viewing behaviors by approximating the human view-then-rate paradigm, they treat scanpath generation and quality assessment as separate steps, preventing end-to-end optimization and task-aligned exploration. To address this limitation, we propose RL-ScanIQA, a reinforcement-learned framework for blind 360°IQA. RL-ScanIQA optimize a PPO-trained scanpath policy and a quality assessor, where the policy receives quality-driven feedback to learn task-relevant viewing strategies. To improve training stability and prevent mode collapse, we design multi-level rewards, including scanpath diversity and equator-biased priors. We further boost cross-dataset robustness using distortion-space augmentation together with rank-consistent losses that preserve intra-image and inter-image quality orderings. Extensive experiments on three benchmarks show that RL-ScanIQA achieves superior in-dataset performance and cross-dataset generalization. Codes are available at https://github.com/wangyuji1/RLScanIQA.git.

Method: Hierarchical expectation-maximization framework that treats annotations as noisy observations of latent clean masks, alternating between inferring posterior distributions and training CNNs with site-specific sensitivity/specificity estimation.

Result: Significant improvements in cross-site generalization (DSC 27.91-39.69%) over state-of-the-art methods, with interpretable per-site label-quality estimates.

Conclusion: Explicitly modeling site-dependent annotation variability improves cross-site generalization in medical image segmentation.

Abstract: Label variability is a major challenge for prostate lesion segmentation. In multi-site datasets, annotations often reflect centre-specific contouring protocols, causing segmentation networks to overfit to local styles and generalise poorly to unseen sites in inference. We treat each observed annotation as a noisy observation of an underlying latent ‘clean’ lesion mask, and propose a hierarchical expectation-maximisation (HierEM) framework that alternates between: (1) inferring a voxel-wise posterior distribution over the latent mask, and (2) training a CNN using this posterior as a soft target and estimate site-specific sensitivity and specificity under a hierarchical prior. This hierarchical prior decomposes label-quality into a global mean with site- and case-level deviations, reducing site-specific bias by penalising the likelihood term contributed only by site deviations. Experiments on three cohorts demonstrate that the proposed hierarchical EM framework enhances cross-site generalisation compared to state-of-the-art methods. For pooled-dataset evaluation, the per-site mean DSC ranges from 29.50% to 39.69%; for leave-one-site-out generalisation, it ranges from 27.91% to 32.67%, yielding statistically significant improvements over comparison methods (p<0.039). The method also produces interpretable per-site latent label-quality estimates (sensitivity alpha ranges from 31.5% to 47.3% at specificity beta approximates 0.99), supporting post-hoc analyses of cross-site annotation variability. These results indicate that explicitly modelling site-dependent annotation can improve cross-site generalisation.

Method: Collects YoURVOS dataset from YouTube untrimmed videos (1,120 videos, 7x more duration/scenes than existing datasets). Proposes OMFormer (Object-level Multimodal Transformers) with object-level multimodal interactions for efficient global spatial-temporal localization.

Result: Previous VOS methods struggle on YoURVOS, especially with target-absent frames. OMFormer consistently performs well on the new benchmark, demonstrating effectiveness for both spatial and temporal localization.

Conclusion: YoURVOS offers an imperative benchmark that pushes RVOS toward practical applications by requiring when-object-appear prediction in addition to where-object-is segmentation.

Abstract: Referring video object segmentation (RVOS) has recently generated great popularity in computer vision due to its widespread applications. Existing RVOS setting contains elaborately trimmed videos, with text-referred objects always appearing in all frames, which however fail to fully reflect the realistic challenges of this task. This simplified setting requires RVOS methods to only predict where objects, with no need to show when the objects appear. In this work, we introduce a new setting towards in-the-wild RVOS. To this end, we collect a new benchmark dataset using Youtube Untrimmed videos for RVOS - YoURVOS, which contains 1,120 in-the-wild videos with 7 times more duration and scenes than existing datasets. Our new benchmark challenges RVOS methods to show not only where but also when objects appear in videos. To set a baseline, we propose Object-level Multimodal TransFormers (OMFormer) to tackle the challenges, which are characterized by encoding object-level multimodal interactions for efficient and global spatial-temporal localisation. We demonstrate that previous VOS methods struggle on our YoURVOS benchmark, especially with the increase of target-absent frames, while our OMFormer consistently performs well. Our YoURVOS dataset offers an imperative benchmark, which will push forward the advancement of RVOS methods for practical applications.

Method: Physics-based attack-defense paradigm: 1) Physics-based Degradation Synthesis (PDS) pipeline modeling ISP inversion to RAW domain, injecting photon/read noise, and reprojecting to sRGB; 2) Dual-layer defense system with noise predictor and degradation-aware Mixture of Experts (DA-MoE); 3) Adaptive Metric Defense (AMD) mechanism

Result: Extensive experiments show significant plug-and-play performance enhancement for existing benchmark LLIE methods, effectively suppressing real-world noise while preserving structural fidelity

Conclusion: The proposed physics-based attack-defense approach effectively addresses limitations of existing LLIE methods by explicitly modeling physical noise transformations and providing adaptive defense mechanisms

Abstract: Limited illumination often causes severe physical noise and detail degradation in images. Existing Low-Light Image Enhancement (LLIE) methods frequently treat the enhancement process as a blind black-box mapping, overlooking the physical noise transformation during imaging, leading to suboptimal performance. To address this, we propose a novel LLIE approach, conceptually formulated as a physics-based attack and display-adaptive defense paradigm. Specifically, on the attack side, we establish a physics-based Degradation Synthesis (PDS) pipeline. Unlike standard data augmentation, PDS explicitly models Image Signal Processor (ISP) inversion to the RAW domain, injects physically plausible photon and read noise, and re-projects the data to the sRGB domain. This generates high-fidelity training pairs with explicitly parameterized degradation vectors, effectively simulating realistic attacks on clean signals. On the defense side, we construct a dual-layer fortified system. A noise predictor estimates degradation parameters from the input sRGB image. These estimates guide a degradation-aware Mixture of Experts (DA-MoE), which dynamically routes features to experts specialized in handling specific noise intensities. Furthermore, we introduce an Adaptive Metric Defense (AMD) mechanism, dynamically calibrating the feature embedding space based on noise severity, ensuring robust representation learning under severe degradation. Extensive experiments demonstrate that our approach offers significant plug-and-play performance enhancement for existing benchmark LLIE methods, effectively suppressing real-world noise while preserving structural fidelity. The sourced code is available at https://github.com/bywlzts/Attack-defense-llie.

Method: Two-stage training: (1) SFT cold-start with curated visual Chain-of-Thought data to activate structured reasoning and tool-use behaviors, (2) GRPO-based reinforcement learning to align complete reasoning-action trajectories with task-level success.

Result: Achieves 97.5% success rate on LIBERO benchmark and shows strong gains across long-horizon robotic tasks on LIBERO and RoboTwin 2.0 benchmarks.

Conclusion: VLA-Thinker significantly improves manipulation performance by enabling active visual reasoning and environment revisiting, advancing embodied intelligence capabilities.

Abstract: Vision-Language-Action (VLA) models have shown promising capabilities for embodied intelligence, but most existing approaches rely on text-based chain-of-thought reasoning where visual inputs are treated as static context. This limits the ability of the model to actively revisit the environment and resolve ambiguities during long-horizon tasks. We propose VLA-Thinker, a thinking-with-image reasoning framework that models perception as a dynamically invocable reasoning action. To train such a system, we introduce a two-stage training pipeline consisting of (1) an SFT cold-start phase with curated visual Chain-of-Thought data to activate structured reasoning and tool-use behaviors, and (2) GRPO-based reinforcement learning to align complete reasoning-action trajectories with task-level success. Extensive experiments on LIBERO and RoboTwin 2.0 benchmarks demonstrate that VLA-Thinker significantly improves manipulation performance, achieving 97.5% success rate on LIBERO and strong gains across long-horizon robotic tasks. Project and Codes: https://cywang735.github.io/VLA-Thinker/ .

Method: Two-stage pipeline: 1) Generate initial 3D instance proposals using 3D-to-2D multi-view projections, Grounded SAM for zero-shot 2D segmentation, and multi-view label fusion; 2) Use these proposals as noisy pseudo-labels to train a supervised 3D panoptic-style segmentation network.

Result: The approach demonstrates feasibility and shows performance improvements over Wheat3DGS (a recent alternative solution based on multi-view RGB images and 3D Gaussian Splatting), showcasing TLS as a competitive sensing alternative for in-field wheat head instance segmentation.

Conclusion: The proposed pipeline enables usable 3D instance segmentation without manual annotations, indicating promising low-effort transferability to other comparable TLS-based point cloud segmentation tasks in agriculture and remote sensing domains.

Abstract: 3D instance segmentation for laser scanning (LiDAR) point clouds remains a challenge in many remote sensing-related domains. Successful solutions typically rely on supervised deep learning and manual annotations, and consequently focus on objects that can be well delineated through visual inspection and manual labeling of point clouds. However, for tasks with more complex and cluttered scenes, such as in-field plant phenotyping in agriculture, such approaches are often infeasible. In this study, we tackle the task of in-field wheat head instance segmentation directly from terrestrial laser scanning (TLS) point clouds. To address the problem and circumvent the need for manual annotations, we propose a novel two-stage pipeline. To obtain the initial 3D instance proposals, the first stage uses 3D-to-2D multi-view projections, the Grounded SAM pipeline for zero-shot 2D object-centric segmentation, and multi-view label fusion. The second stage uses these initial proposals as noisy pseudo-labels to train a supervised 3D panoptic-style segmentation neural network. Our results demonstrate the feasibility of the proposed approach and show performance improvementsrelative to Wheat3DGS, a recent alternative solution for in-field wheat head instance segmentation without manual 3D annotations based on multi-view RGB images and 3D Gaussian Splatting, showcasing TLS as a competitive sensing alternative. Moreover, the results show that both stages of the proposed pipeline can deliver usable 3D instance segmentation without manual annotations, indicating promising, low-effort transferability to other comparable TLS-based point cloud segmentation tasks.

Method: Created curated multi-center, multi-resolution dataset with expert-validated MES and UCEIS labels alongside detailed clinical descriptions. Provided benchmarking using convolutional neural networks, vision transformers, hybrid models, and multimodal vision-language captioning algorithms.

Result: First comprehensive dataset combining dual scoring metrics (MES/UCEIS) for classification tasks with expert-generated captions describing mucosal appearance and clinically accepted reasoning for image captioning.

Conclusion: This resource opens new opportunities for developing clinically meaningful multimodal algorithms for ulcerative colitis assessment and provides benchmarking for future research in this domain.

Abstract: Ulcerative colitis (UC) is a chronic mucosal inflammatory condition that places patients at increased risk of colorectal cancer. Colonoscopic surveillance remains the gold standard for assessing disease activity, and reporting typically relies on standardised endoscopic scoring metrics. The most widely used is the Mayo Endoscopic Score (MES), with some centres also adopting the Ulcerative Colitis Endoscopic Index of Severity (UCEIS). Both are descriptive assessments of mucosal inflammation (MES: 0 to 3; UCEIS: 0 to 8), where higher values indicate more severe disease. However, computational methods for automatically predicting these scores remain limited, largely due to the lack of publicly available expert-annotated datasets and the absence of robust benchmarking. There is also a significant research gap in generating clinically meaningful descriptions of UC images, despite image captioning being a well-established computer vision task. Variability in endoscopic systems and procedural workflows across centres further highlights the need for multi-centre datasets to ensure algorithmic robustness and generalisability. In this work, we introduce a curated multi-centre, multi-resolution dataset that includes expert-validated MES and UCEIS labels, alongside detailed clinical descriptions. To our knowledge, this is the first comprehensive dataset that combines dual scoring metrics for classification tasks with expert-generated captions describing mucosal appearance and clinically accepted reasoning for image captioning. This resource opens new opportunities for developing clinically meaningful multimodal algorithms. In addition to the dataset, we also provide benchmarking using convolutional neural networks, vision transformers, hybrid models, and widely used multimodal vision-language captioning algorithms.

Method: Integrates foreground-background probability cues into Gaussian primitives using YOLO and SAM probability masks, replaces binary masks with continuous probability values, employs dual-stage filtering strategy during training startup, and uses rendered probability masks for supervision refinement and boundary consistency.

Result: Achieves reconstruction quality comparable to standard 3DGS approaches while requiring only ~1/10 of Gaussian amount. Demonstrates strong self-correction capability with mask errors on MIP-360, T&T, and NVOS datasets.

Conclusion: Proposed method is efficient and robust for single-object reconstruction, offering high fidelity with computational efficiency, suitable for applications requiring both quality and efficiency.

Abstract: Object-level 3D reconstruction play important roles across domains such as cultural heritage digitization, industrial manufacturing, and virtual reality. However, existing Gaussian Splatting-based approaches generally rely on full-scene reconstruction, in which substantial redundant background information is introduced, leading to increased computational and storage overhead. To address this limitation, we propose an efficient single-object 3D reconstruction method based on 2D Gaussian Splatting. By directly integrating foreground-background probability cues into Gaussian primitives and dynamically pruning low-probability Gaussians during training, the proposed method fundamentally focuses on an object of interest and improves the memory and computational efficiency. Our pipeline leverages probability masks generated by YOLO and SAM to supervise probabilistic Gaussian attributes, replacing binary masks with continuous probability values to mitigate boundary ambiguity. Additionally, we propose a dual-stage filtering strategy for training’s startup to suppress background Gaussians. And, during training, rendered probability masks are conversely employed to refine supervision and enhance boundary consistency across views. Experiments conducted on the MIP-360, T&T, and NVOS datasets demonstrate that our method exhibits strong self-correction capability in the presence of mask errors and achieves reconstruction quality comparable to standard 3DGS approaches, while requiring only approximately 1/10 of their Gaussian amount. These results validate the efficiency and robustness of our method for single-object reconstruction and highlight its potential for applications requiring both high fidelity and computational efficiency.

Method: Proposes a real-time inspection (RI) module that converts latents into intermediate video previews, enabling text-video alignment scoring in RGB space. Uses a hierarchical early-exit intervention pipeline triggered only when failure is predicted. The RI module completes conversion and inspection in just 39.2ms.

Result: Experiments on CogVideoX-5B and Wan2.1-1.3B show consistency gains on VBench with up to 2.64× less time overhead compared to post-hoc regeneration. Method generalizes to higher-capacity models (Wan2.1-14B with 720p, 81-frame generation) and is plug-and-play compatible with existing techniques like prompt refinement and sampling guidance.

Conclusion: Failure signals emerge early in denoising and are detectable within intermediate video previews using standard vision-language evaluators. The proposed pipeline enables efficient early failure detection and intervention for text-to-video diffusion models.

Abstract: Text-to-video (T2V) diffusion models have rapidly advanced, yet generations still occasionally fail in practice, such as low text-video alignment or low perceptual quality. Since diffusion sampling is non-deterministic, it is difficult to know during inference whether a generation will succeed or fail, incurring high computational cost due to trial-and-error regeneration. To address this, we propose an early failure detection and diagnostic intervention pipeline for latent T2V diffusion models. For detection, we design a Real-time Inspection (RI) module that converts latents into intermediate video previews, enabling the use of established text-video alignment scorers for inspection in the RGB space. The RI module completes the conversion and inspection process in just 39.2ms. This is highly efficient considering that CogVideoX-5B requires 4.3s per denoising step when generating a 480p, 49-frame video on an NVIDIA A100 GPU. Subsequently, we trigger a hierarchical and early-exit intervention pipeline only when failure is predicted. Experiments on CogVideoX-5B and Wan2.1-1.3B demonstrate consistency gains on VBench with up to 2.64 times less time overhead compared to post-hoc regeneration. Our method also generalizes to a higher-capacity setting, remaining effective on Wan2.1-14B with 720p resolution and 81-frame generation. Furthermore, our pipeline is plug-and-play and orthogonal to existing techniques, showing seamless compatibility with prompt refinement and sampling guidance methods. We also provide evidence that failure signals emerge early in the denoising process and are detectable within intermediate video previews using standard vision-language evaluators.

Method: Proposes PerCS-DINO framework built on DINOv2 backbone, integrating image features and reference embeddings via cross-attention transformer and contrastive learning to segment cells matching the reference.

Result: Extensive experiments demonstrate the effectiveness of PerCS-DINO on a benchmark of 1,372 images with over 110,000 annotated cells, highlighting the challenges of this new personalized cell segmentation task.

Conclusion: PerCS serves as a useful testbed for advancing research in cell-based applications, with PerCS-DINO as a pioneering solution for personalized cell segmentation.

Abstract: Accurate cell segmentation is critical for biological and medical imaging studies. Although recent deep learning models have advanced this task, most methods are limited to generic cell segmentation, lacking the ability to differentiate specific cell types. In this work, we introduce the Personalized Cell Segmentation (PerCS) task, which aims to segment all cells of a specific type given a reference cell. To support this task, we establish a benchmark by reorganizing publicly available datasets, yielding 1,372 images and over 110,000 annotated cells. As a pioneering solution, we propose PerCS-DINO, a framework built on the DINOv2 backbone. By integrating image features and reference embeddings via a cross-attention transformer and contrastive learning, PerCS-DINO effectively segments cells matching the reference. Extensive experiments demonstrate the effectiveness of the proposed PerCS-DINO and highlight the challenges of this new task. We expect PerCS to serve as a useful testbed for advancing research in cell-based applications.

Method: Proposes AvatarForcing with dual-anchor temporal forcing: style anchor (re-indexes RoPE with fixed relative positioning and anchor-audio zero-padding) and temporal anchor (reuses recently emitted clean blocks). Uses two-stage streaming distillation with offline ODE backfill and distribution matching for one-step inference.

Result: Achieves strong visual quality and lip synchronization at 34 ms/frame using a 1.3B-parameter student model, demonstrated on standard benchmarks and a new 400-video long-form benchmark.

Conclusion: AvatarForcing enables real-time one-step streaming diffusion for talking avatar generation with stable long-form synthesis and constant per-step computational cost.

Abstract: Real-time talking avatar generation requires low latency and minute-level temporal stability. Autoregressive (AR) forcing enables streaming inference but suffers from exposure bias, which causes errors to accumulate and become irreversible over long rollouts. In contrast, full-sequence diffusion transformers mitigate drift but remain computationally prohibitive for real-time long-form synthesis. We present AvatarForcing, a one-step streaming diffusion framework that denoises a fixed local-future window with heterogeneous noise levels and emits one clean block per step under constant per-step cost. To stabilize unbounded streams, the method introduces dual-anchor temporal forcing: a style anchor that re-indexes RoPE to maintain a fixed relative position with respect to the active window and applies anchor-audio zero-padding, and a temporal anchor that reuses recently emitted clean blocks to ensure smooth transitions. Real-time one-step inference is enabled by two-stage streaming distillation with offline ODE backfill and distribution matching. Experiments on standard benchmarks and a new 400-video long-form benchmark show strong visual quality and lip synchronization at 34 ms/frame using a 1.3B-parameter student model for realtime streaming. Our page is available at: https://cuiliyuan121.github.io/AvatarForcing/

Method: Created UAVBench (43 test units, 966k samples across 10 tasks) and UAVIT-1M (1.24M instructions, 789k multi-scene images, 11 tasks) with real-world drone images, diverse weather conditions, and manual verification.

Result: Open-source MLLMs perform poorly on low-altitude visual content compared to closed-source models. Fine-tuning open-source MLLMs on UAVIT-1M significantly improves their performance on drone vision-language tasks.

Conclusion: The benchmark and dataset bridge the gap between current MLLMs and real-world low-altitude UAV application requirements, enabling better drone vision-language understanding.

Abstract: Multimodal Large Language Models (MLLMs) have made significant strides in natural images and satellite remote sensing images. However, understanding low-altitude drone scenarios remains a challenge. Existing datasets primarily focus on a few specific low-altitude visual tasks, which cannot fully assess the ability of MLLMs in real-world low-altitude UAV applications. Therefore, we introduce UAVBench, a comprehensive benchmark, and UAVIT-1M, a large-scale instruction tuning dataset, designed to evaluate and improve MLLMs’ abilities in low-altitude vision-language tasks. UAVBench comprises 43 test units and 966k high-quality data samples across 10 tasks at the image-level and region-level. UAVIT-1M consists of approximately 1.24 million diverse instructions, covering 789k multi-scene images and about 2,000 types of spatial resolutions with 11 distinct tasks. UAVBench and UAVIT-1M feature pure real-world visual images and rich weather conditions, and involve manual verification to ensure high quality. Our in-depth analysis of 11 state-of-the-art MLLMs using UAVBench reveals that open-source MLLMs cannot generate accurate conversations about low-altitude visual content, lagging behind closed-source MLLMs. Extensive experiments demonstrate that fine-tuning open-source MLLMs on UAVIT-1M significantly addresses this gap. Our contributions pave the way for bridging the gap between current MLLMs and low-altitude UAV real-world application demands. (Project page: https://UAVBench.github.io/)

Method: Introduces OutRo, which: (1) aligns non-sink token representations with the sink representation in feature space, and (2) allows the sink token to attend beyond causal constraints to facilitate information exchange with non-sink tokens. This enhances reasoning without additional forward passes or access to attention maps.

Result: OutRo consistently improves performance across representative MLLMs on seven video QA benchmarks, demonstrates strong generalization, and incurs only 1.1x decoding overhead.

Conclusion: Attention sinks encode structured global information that influences decoding, and leveraging them through OutRo provides an effective way to enhance multimodal LLM reasoning with minimal computational overhead.

Abstract: Large language models and their multimodal extensions have achieved remarkable success across diverse tasks, yet the internal mechanisms that govern their reasoning behaviour remain partially understood. In particular, the attention sink, a token that attracts disproportionate attention mass, has been observed in transformer architectures, but its role is still unclear. Our goal is to understand what attention sinks represent and how they shape model behaviour during inference, rather than considering them as incidental artifacts. Through our analysis, we find that attention sink representations encode structured global information that influences the decoding process. Building on our findings, we introduce OutRo, a lightweight inference-time strategy that leverages the sink token to enhance contextual representations: (i) non-sink token representations are aligned with the sink representation in the feature space; and (ii) the sink token is allowed to attend beyond the causal constraint, facilitating information exchange with non-sink tokens. This design enhances the reasoning process without requiring additional forward passes or access to attention maps. Based on extensive experiments, OutRo consistently improves performance across representative MLLMs on seven video QA benchmarks and demonstrates strong generalisation, while incurring only a 1.1x decoding overhead.

Method: Extract robust unimodal features from visual, acoustic, and linguistic data with specialized statistical text modality for temporal speech variations. Evaluate 15 modality combinations across MLP, Random Forest, and GBDT classifiers, then use Particle Swarm Optimization hard-voting ensemble with train-validation gap penalty to suppress overfitting.

Result: Linguistic features are strongest independent predictor, but PSO ensemble (lambda=0.2) achieves peak Macro F1-score of 0.7465 on unseen test set by effectively harnessing multimodal synergies.

Conclusion: Treating ambivalence and hesitancy as multimodal conflict evaluated through intelligently weighted committee provides robust framework for in-the-wild behavioral analysis.

Abstract: Recognizing complex behavioral states such as Ambivalence and Hesitancy (A/H) in naturalistic video settings remains a significant challenge in affective computing. Unlike basic facial expressions, A/H manifests as subtle, multimodal conflicts that require deep contextual and temporal understanding. In this paper, we propose a highly regularized, multimodal fusion pipeline to predict A/H at the video level. We extract robust unimodal features from visual, acoustic, and linguistic data, introducing a specialized statistical text modality explicitly designed to capture temporal speech variations and behavioral cues. To identify the most effective representations, we evaluate 15 distinct modality combinations across a committee of machine learning classifiers (MLP, Random Forest, and GBDT), selecting the most well-calibrated models based on validation Binary Cross-Entropy (BCE) loss. Furthermore, to optimally fuse these heterogeneous models without overfitting to the training distribution, we implement a Particle Swarm Optimization (PSO) hard-voting ensemble. The PSO fitness function dynamically incorporates a train-validation gap penalty (lambda) to actively suppress redundant or overfitted classifiers. Our comprehensive evaluation demonstrates that while linguistic features serve as the strongest independent predictor of A/H, our heavily regularized PSO ensemble (lambda = 0.2) effectively harnesses multimodal synergies, achieving a peak Macro F1-score of 0.7465 on the unseen test set. These results emphasize that treating ambivalence and hesitancy as a multimodal conflict, evaluated through an intelligently weighted committee, provides a robust framework for in-the-wild behavioral analysis.

Method: Proposes PixelREPA with two key components: 1) transforms alignment target to address information asymmetry, 2) uses Masked Transformer Adapter combining shallow transformer adapter with partial token masking to constrain alignment

Result: PixelREPA reduces FID from 3.66 to 3.17 for JiT-B/16, improves Inception Score from 275.1 to 284.6 on ImageNet 256×256, achieves >2× faster convergence, and PixelREPA-H/16 achieves FID=1.81 and IS=317.2

Conclusion: PixelREPA successfully addresses REPA’s failures for pixel-space diffusion transformers, improving both training convergence and final quality while maintaining the benefits of pixel-space approaches

Abstract: Representation Alignment (REPA) has emerged as a simple way to accelerate Diffusion Transformers training in latent space. At the same time, pixel-space diffusion transformers such as Just image Transformers (JiT) have attracted growing attention because they remove a dependency on a pretrained tokenizer, and then avoid the reconstruction bottleneck of latent diffusion. This paper shows that the REPA can fail for JiT. REPA yields worse FID for JiT as training proceeds and collapses diversity on image subsets that are tightly clustered in the representation space of pretrained semantic encoder on ImageNet. We trace the failure to an information asymmetry: denoising occurs in the high dimensional image space, while the semantic target is strongly compressed, making direct regression a shortcut objective. We propose PixelREPA, which transforms the alignment target and constrains alignment with a Masked Transformer Adapter that combines a shallow transformer adapter with partial token masking. PixelREPA improves both training convergence and final quality. PixelREPA reduces FID from 3.66 to 3.17 for JiT-B$/16$ and improves Inception Score (IS) from 275.1 to 284.6 on ImageNet $256 \times 256$, while achieving $> 2\times$ faster convergence. Finally, PixelREPA-H$/16$ achieves FID$=1.81$ and IS$=317.2$. Our code is available at https://github.com/kaist-cvml/PixelREPA.

Method: 1) Topology-aware augmentations using relative bottleneck distance between persistence diagrams to control topological perturbations while preserving medically relevant properties. 2) Hierarchical Topology Encoder with self-attention and cross-attention mechanisms to capture topological features. 3) Adaptive mixture-of-experts (MoE) module to dynamically integrate visual and topological representations. 4) Seamless integration with existing CL methods.

Result: Evaluated on five CL methods (SimCLR, MoCo-v3, BYOL, DINO, Barlow Twins) and five medical image classification datasets. Achieved average gain of +3.26% in linear probe classification accuracy with strong statistical significance.

Conclusion: TopoCL effectively integrates topological structures into contrastive learning for medical imaging, consistently improving performance across various CL methods and datasets, demonstrating the importance of topological features in medical image analysis.

Abstract: Contrastive learning (CL) has become a powerful approach for learning representations from unlabeled images. However, existing CL methods focus predominantly on visual appearance features while neglecting topological characteristics (e.g., connectivity patterns, boundary configurations, cavity formations) that provide valuable cues for medical image analysis. To address this limitation, we propose a new topological CL framework (TopoCL) that explicitly exploits topological structures during contrastive learning for medical imaging. Specifically, we first introduce topology-aware augmentations that control topological perturbations using a relative bottleneck distance between persistence diagrams, preserving medically relevant topological properties while enabling controlled structural variations. We then design a Hierarchical Topology Encoder that captures topological features through self-attention and cross-attention mechanisms. Finally, we develop an adaptive mixture-of-experts (MoE) module to dynamically integrate visual and topological representations. TopoCL can be seamlessly integrated with existing CL methods. We evaluate TopoCL on five representative CL methods (SimCLR, MoCo-v3, BYOL, DINO, and Barlow Twins) and five diverse medical image classification datasets. The experimental results show that TopoCL achieves consistent improvements: an average gain of +3.26% in linear probe classification accuracy with strong statistical significance, verifying its effectiveness.

Method: Proposes an architecture-agnostic safeguard with Context-Guided Chain-of-Thought (CG-CoT) that decomposes risk assessment into: 1) active perception that sequentially anchors attention to interaction targets and relevant spatial neighborhoods, and 2) semantic judgment based on visual evidence. Uses a curated grounding dataset and two-stage training with Reinforcement Fine-Tuning (RFT) with process rewards to enforce precise intermediate grounding.

Result: HomeGuard significantly enhances safety, improving risk match rates by over 30% compared to base models while reducing oversafety. The generated visual anchors serve as actionable spatial constraints for downstream planners, facilitating explicit collision avoidance and safety trajectory generation.

Conclusion: The proposed safeguard effectively addresses contextual safety risks in vision-language models for embodied agents through a systematic approach combining active perception and semantic judgment, with practical applications for safety planning.

Abstract: Vision-Language Models (VLMs) empower embodied agents to execute complex instructions, yet they remain vulnerable to contextual safety risks where benign commands become hazardous due to subtle environmental states. Existing safeguards often prove inadequate. Rule-based methods lack scalability in object-dense scenes, whereas model-based approaches relying on prompt engineering suffer from unfocused perception, resulting in missed risks or hallucinations. To address this, we propose an architecture-agnostic safeguard featuring Context-Guided Chain-of-Thought (CG-CoT). This mechanism decomposes risk assessment into active perception that sequentially anchors attention to interaction targets and relevant spatial neighborhoods, followed by semantic judgment based on this visual evidence. We support this approach with a curated grounding dataset and a two-stage training strategy utilizing Reinforcement Fine-Tuning (RFT) with process rewards to enforce precise intermediate grounding. Experiments demonstrate that our model HomeGuard significantly enhances safety, improving risk match rates by over 30% compared to base models while reducing oversafety. Beyond hazard detection, the generated visual anchors serve as actionable spatial constraints for downstream planners, facilitating explicit collision avoidance and safety trajectory generation. Code and data are released under https://github.com/AI45Lab/HomeGuard

Method: Evaluated 200 human participants and 95 state-of-the-art AI detectors across two datasets: DF40 (standard benchmark) and CharadesDF (novel dataset of everyday activities recorded with mobile phones).

Result: Humans outperformed AI detectors on both datasets, with the performance gap widening on CharadesDF where AI accuracy dropped to near chance (0.537) while humans maintained robust performance (0.784). Human and AI errors were complementary, and hybrid human-AI ensembles reduced high-confidence errors.

Conclusion: Effective real-world deepfake detection, especially for non-professionally produced videos, requires human-AI collaboration rather than relying on AI algorithms alone.

Abstract: Deepfake detection is widely framed as a machine learning problem, yet how humans and AI detectors compare under realistic conditions remains poorly understood. We evaluate 200 human participants and 95 state-of-the-art AI detectors across two datasets: DF40, a standard benchmark, and CharadesDF, a novel dataset of videos of everyday activities. CharadesDF was recorded using mobile phones leading to low/moderate quality videos compared to the more professionally captured DF40. Humans outperform AI detectors on both datasets, with the gap widening in the case of CharadesDF where AI accuracy collapses to near chance (0.537) while humans maintain robust performance (0.784). Human and AI errors are complementary: humans miss high-quality deepfakes while AI detectors flag authentic videos as fake, and hybrid human-AI ensembles reduce high-confidence errors. These findings suggest that effective real-world deepfake detection, especially in non-professionally produced videos, requires human-AI collaboration rather than AI algorithms alone.

Method: Proposes LoCAtion: Long-time Collaborative Attention framework that reformulates HDR video generation as alignment-free collaborative feature routing. Uses continuous medium-exposure backbone, collaborative attention to harvest irradiance cues, and learned global sequence solver with bidirectional context and long-range temporal modeling.

Result: Achieves state-of-the-art visual quality and temporal stability with competitive balance between accuracy and computational efficiency in extensive experiments.

Conclusion: LoCAtion offers a robust alternative to alignment-based methods by decoupling reconstruction tasks and leveraging collaborative attention and temporal coherence for ghost-free HDR video generation.

Abstract: Prevailing High Dynamic Range (HDR) video reconstruction methods are fundamentally trapped in a fragile alignment-and-fusion paradigm. While explicit spatial alignment can successfully recover fine details in controlled environments, it becomes a severe bottleneck in unconstrained dynamic scenes. By forcing rigid alignment across unpredictable motions and varying exposures, these methods inevitably translate registration errors into severe ghosting artifacts and temporal flickering. In this paper, we rethink this conventional prerequisite. Recognizing that explicit alignment is inherently vulnerable to real-world complexities, we propose LoCAtion, a Long-time Collaborative Attention framework that reformulates HDR video generation from a fragile spatial warping task into a robust, alignment-free collaborative feature routing problem. Guided by this new formulation, our architecture explicitly decouples the highly entangled reconstruction task. Rather than struggling to rigidly warp neighboring frames, we anchor the scene on a continuous medium-exposure backbone and utilize collaborative attention to dynamically harvest and inject reliable irradiance cues from unaligned exposures. Furthermore, we introduce a learned global sequence solver. By leveraging bidirectional context and long-range temporal modeling, it propagates corrective signals and structural features across the entire sequence, inherently enforcing whole-video coherence and eliminating jitter. Extensive experiments demonstrate that LoCAtion achieves state-of-the-art visual quality and temporal stability, offering a highly competitive balance between accuracy and computational efficiency.

Method: Two-component framework: (1) Visual Prompt Selector predicts appropriate prompt types conditioned on video/question, (2) Spatio-Temporal Reasoner optimized with RL under visual prompt guidance and object-aware grounding rewards. Uses self-distillation to internalize improvements.

Result: Achieves state-of-the-art performance across diverse video reasoning, understanding, and temporal grounding benchmarks (V-STAR, VideoMME, World-Sense, VideoMMMU, PerceptionTest, Charades-STA) while maintaining single efficient inference pathway without external tools.

Conclusion: Visual prompting during training improves grounded video reasoning, and self-distillation enables models to internalize this ability without requiring prompts at inference time, offering efficient solution to spatio-temporal grounding challenges.

Abstract: Video reasoning requires models to locate and track question-relevant evidence across frames. While reinforcement learning (RL) with verifiable rewards improves accuracy, it still struggles to achieve reliable spatio-temporal grounding during the reasoning process. Moreover, improving grounding typically relies on scaled training data or inference-time perception tools, which increases annotation cost or computational cost. To address this challenge, we propose VisonCoach, an input-adaptive RL framework that improves spatio-temporal grounding through visual prompting as training-time guidance. During RL training, visual prompts are selectively applied to challenging inputs to amplify question-relevant evidence and suppress distractors. The model then internalizes these improvements through self-distillation, enabling grounded reasoning directly on raw videos without visual prompting at inference. VisonCoach consists of two components: (1) Visual Prompt Selector, which predicts appropriate prompt types conditioned on the video and question, and (2) Spatio-Temporal Reasoner, optimized with RL under visual prompt guidance and object-aware grounding rewards that enforce object identity consistency and multi-region bounding-box overlap. Extensive experiments demonstrate that VisonCoach achieves state-of-the-art performance under comparable settings, across diverse video reasoning, video understanding, and temporal grounding benchmarks (V-STAR, VideoMME, World-Sense, VideoMMMU, PerceptionTest, and Charades-STA), while maintaining a single efficient inference pathway without external tools. Our results show that visual prompting during training improves grounded video reasoning, while self-distillation enables the model to internalize this ability without requiring prompts at inference time.

Method: Presents Segment Anything Reasoner (StAR) with refined design space including parameter-tuning scheme, reward functions, learning strategies, and answer format. Introduces parallel test-time scaling for segmentation tasks. Constructs ReasonSeg-X benchmark with deeper reasoning samples. Uses rollout-expanded selective-tuning approach with only 5k training samples.

Result: StAR achieves significant gains over base counterparts across extensive benchmarks, demonstrating effective activation of latent reasoning capabilities with minimal training data.

Conclusion: The framework successfully brings dormant reasoning competence to the surface through comprehensive design refinements and novel scaling techniques, establishing a rigorous benchmark for systematic evaluation of advanced reasoning segmentation methods.

Abstract: As AI systems are being integrated more rapidly into diverse and complex real-world environments, the ability to perform holistic reasoning over an implicit query and an image to localize a target is becoming increasingly important. However, recent reasoning segmentation methods fail to sufficiently elicit the visual reasoning capabilities of the base mode. In this work, we present Segment Anything Reasoner (StAR), a comprehensive framework that refines the design space from multiple perspectives-including parameter-tuning scheme, reward functions, learning strategies and answer format-and achieves substantial improvements over recent baselines. In addition, for the first time, we successfully introduce parallel test-time scaling to the segmentation task, pushing the performance boundary even further. To extend the scope and depth of reasoning covered by existing benchmark, we also construct the ReasonSeg-X, which compactly defines reasoning types and includes samples that require deeper reasoning. Leveraging this dataset, we train StAR with a rollout-expanded selective-tuning approach to activate the base model’s latent reasoning capabilities, and establish a rigorous benchmark for systematic, fine-grained evaluation of advanced methods. With only 5k training samples, StAR achieves significant gains over its base counterparts across extensive benchmarks, demonstrating that our method effectively brings dormant reasoning competence to the surface.

Method: Uses an elucidated diffusion model (EDM) framework with two U-Net architectures: (1) full 3D convolutional U-Net processing volumetric patches with 3D convolutions and multi-head self-attention, and (2) 2.5D slice-conditioned U-Net that super-resolves each slice independently while conditioning on adjacent slices for inter-slice context. Both use continuous-sigma noise conditioning and are trained on the NKI cohort of FOMO60K dataset.

Result: On a test set of 5 subjects (6 volumes, 993 slices), the 3D model achieves 37.75 dB PSNR, 0.997 SSIM, and 0.020 LPIPS, outperforming both the EDSR baseline (35.57 dB/0.024 LPIPS) and the 2.5D variant (35.82 dB) across all metrics.

Conclusion: The 3D diffusion model architecture provides superior performance for MRI super-resolution compared to 2.5D approaches and traditional methods, demonstrating the effectiveness of volumetric processing with attention mechanisms for medical image enhancement.

Abstract: Magnetic resonance imaging (MRI) super-resolution (SR) methods that computationally enhance low-resolution acquisitions to approximate high-resolution quality offer a compelling alternative to expensive high-field scanners. In this work we investigate an elucidated diffusion model (EDM) framework for brain MRI SR and compare two U-Net backbone architectures: (i) a full 3D convolutional U-Net that processes volumetric patches with 3D convolutions and multi-head self-attention, and (ii) a 2.5D slice-conditioned U-Net that super-resolves each slice independently while conditioning on an adjacent slice for inter-slice context. Both models employ continuous-sigma noise conditioning following Karras et al. and are trained on the NKI cohort of the FOMO60K dataset. On a held-out test set of 5 subjects (6 volumes, 993 slices), the 3D model achieves 37.75 dB PSNR, 0.997 SSIM, and 0.020 LPIPS, improving on the off-the-shelf pretrained EDSR baseline (35.57 dB / 0.024 LPIPS) and the 2.5D variant (35.82 dB) across all three metrics under the same test data and degradation pipeline.

Method: Uses feature-based implicit neural representation (INR) fusion network as backbone with multi-scale semi-supervised training scheme for robust generalization to arbitrary scales and cross-scene/sensor scenarios.

Result: Achieves state-of-the-art results on multiple real-world datasets for PAN-scale fusion in both visual quality and quantitative metrics. Supports weight reuse across image pairs while maintaining competitiveness with per-pair retraining.

Conclusion: G-ZAP demonstrates strong potential for efficient real-world deployment by enabling generalizable zero-shot arbitrary-scale pansharpening with weight reuse across different image pairs and scenarios.

Abstract: Pansharpening aims to fuse a high-resolution panchromatic (PAN) image and a low-resolution multispectral (LRMS) image to produce a high-resolution multispectral (HRMS) image. Recent deep models have achieved strong performance, yet they typically rely on large-scale pretraining and often generalize poorly to unseen real-world image pairs.Prior zero-shot approaches improve real-scene generalization but require per-image optimization, hindering weight reuse, and the above methods are usually limited to a fixed scale.To address this issue, we propose G-ZAP, a generalizable zero-shot framework for arbitrary-scale pansharpening, designed to handle cross-resolution, cross-scene, and cross-sensor generalization.G-ZAP adopts a feature-based implicit neural representation (INR) fusion network as the backbone and introduces a multi-scale, semi-supervised training scheme to enable robust generalization.Extensive experiments on multiple real-world datasets show that G-ZAP achieves state-of-the-art results under PAN-scale fusion in both visual quality and quantitative metrics.Notably, G-ZAP supports weight reuse across image pairs while maintaining competitiveness with per-pair retraining, demonstrating strong potential for efficient real-world deployment.

Method: Two-stage framework: 1) 3D Foundation Model produces view-consistent object priors conditioned on implicit motion dynamics across novel viewpoints, 2) Controllable video generation model synthesizes high-fidelity object texture by incorporating multi-view reference images with appearance consistency via retrieval mechanism. The two stages mutually reinforce each other during inference.

Result: Superior performance in generating long-duration HOI videos with intricate object manipulations. Extensive experiments show substantial improvements over prior approaches, especially for HOI with complex 3D object manipulations.

Conclusion: MVHOI effectively bridges multi-view reference conditions and video foundation models via 3D Foundation Model, enabling realistic motion reenactment for complex human-object interactions with 3D object manipulations.

Abstract: Human-Object Interaction (HOI) video reenactment with realistic motion remains a frontier in expressive digital human creation. Existing approaches primarily handle simple image-plane motion (e.g., in-plane translations), struggling with complex non-planar manipulations like out-of-plane reorientation. In this paper, we propose MVHOI, a two-stage HOI video reenactment framework that bridges multi-view reference conditions and video foundation models via a 3D Foundation Model (3DFM). The 3DFM first produces view-consistent object priors conditioned on implicit motion dynamics across novel viewpoints. A controllable video generation model then synthesizes high-fidelity object texture by incorporating multi-view reference images, ensuring appearance consistency via a reasonable retrieval mechanism. By enabling these two stages to mutually reinforce one another during the inference phase, our framework shows superior performance in generating long-duration HOI videos with intricate object manipulations. Extensive experiments show substantial improvements over prior approaches, especially for HOI with complex 3D object manipulations.

Method: Integrates DenseNet, ConvNeXt, and EfficientNet backbones in a gated multi-expert architecture, adds prototype learning for interpretability, applies physics-informed regularization for morphology preservation, and uses Monte Carlo Dropout for uncertainty quantification.

Result: Achieves 96.97% accuracy on BreaKHis dataset under multi-magnification training, shows improved generalization to unseen magnification levels, and uncertainty estimation helps identify out-of-distribution samples.

Conclusion: Histo-MExNet provides a balanced combination of accuracy, robustness, and interpretability for clinical decision support in breast cancer diagnosis.

Abstract: Accurate and reliable histopathological image classification is essential for breast cancer diagnosis. However, many deep learning models remain sensitive to magnification variability and lack interpretability. To address these challenges, we propose Histo-MExNet, a unified framework designed for scaleinvariant and uncertainty-aware classification. The model integrates DenseNet, ConvNeXt, and EfficientNet backbones within a gated multi-expert architecture, incorporates a prototype learning module for example-driven interpretability, and applies physics-informed regularization to enforce morphology preservation and spatial coherence during feature learning. Monte Carlo Dropout is used to quantify predictive uncertainty. On the BreaKHis dataset, Histo-MExNet achieves 96.97% accuracy under multi-magnification training and demonstrates improved generalization to unseen magnification levels compared to single-expert models, while uncertainty estimation helps identify out-of-distribution samples and reduce overconfident errors, supporting a balanced combination of accuracy, robustness, and interpretability for clinical decision support.

Method: THO uses an end-to-end Spatial-Temporal Transformer that predicts human and object motion from video and 3D template. It leverages spatial priors (contact-region proximity to infer occluded object features) and temporal priors (cross-frame kinematic correlations for physical coherence).

Result: Achieves 31.5 FPS inference speed on RTX 4090 GPU, >600x speedup over optimization-based methods while improving reconstruction accuracy and temporal consistency.

Conclusion: THO enables real-time monocular 4D HOI reconstruction with superior speed and accuracy by effectively leveraging spatial-temporal priors in an end-to-end transformer architecture.

Abstract: Monocular 4D human-object interaction (HOI) reconstruction - recovering a moving human and a manipulated object from a single RGB video - remains challenging due to depth ambiguity and frequent occlusions. Existing methods often rely on multi-stage pipelines or iterative optimization, leading to high inference latency, failing to meet real-time requirements, and susceptibility to error accumulation. To address these limitations, we propose THO, an end-to-end Spatial-Temporal Transformer that predicts human motion and coordinated object motion in a forward fashion from the given video and 3D template. THO achieves this by leveraging spatial-temporal HOI tuple priors. Spatial priors exploit contact-region proximity to infer occluded object features from human cues, while temporal priors capture cross-frame kinematic correlations to refine object representations and enforce physical coherence. Extensive experiments demonstrate that THO operates at an inference speed of 31.5 FPS on a single RTX 4090 GPU, achieving a >600x speedup over prior optimization-based methods while simultaneously improving reconstruction accuracy and temporal consistency. The project page is available at: https://nianheng.github.io/THO-project/

Method: Two-stage ensemble approach using supervised domain adaptation (SDA), adapting the xView2 first-place method to the Ida-BD dataset, with systematic investigation of individual augmentation components on classification performance.

Result: SDA is indispensable - removing it causes damage detection to fail entirely. Best performance achieved with SDA and unsharp-enhanced RGB input, attaining Macro-F1 of 0.5552 on unseen Ida-BD test split.

Conclusion: Domain adaptation is critical for building trustworthy automated damage assessment modules in human-machine systems for disaster response, addressing domain shift between training and deployment data.

Abstract: Rapid structural damage assessment from remote sensing imagery is essential for timely disaster response. Within human-machine systems (HMS) for disaster management, automated damage detection provides decision-makers with actionable situational awareness. However, models trained on multi-disaster benchmarks often underperform in unseen geographic regions due to domain shift - a distributional mismatch between training and deployment data that undermines human trust in automated assessments. We explore a two-stage ensemble approach using supervised domain adaptation (SDA) for building damage classification across four severity classes. The pipeline adapts the xView2 first-place method to the Ida-BD dataset using SDA and systematically investigates the effect of individual augmentation components on classification performance. Comprehensive ablation experiments on the unseen Ida-BD test split demonstrate that SDA is indispensable: removing it causes damage detection to fail entirely. Our pipeline achieves the most robust performance using SDA with unsharp-enhanced RGB input, attaining a Macro-F1 of 0.5552. These results underscore the critical role of domain adaptation in building trustworthy automated damage assessment modules for HMS-integrated disaster response.

Method: Proposes two core components: 1) Memory-Aware Compression Prompt (MCP) compresses memory features into prompt tokens that interact throughout the backbone, and 2) Dynamic State Fusion (DSF) captures continuous target dynamics by progressively introducing updated dynamic state features from shallow to deep layers.

Result: Achieves state-of-the-art results on 10 datasets spanning five modalities (RGB, RGB-D/T/E, RGB-Language). Training only MCP, DSF, and prediction head (30% trainable parameters) yields substantial performance gains. Both modules are plug-and-play and boost various baseline trackers.

Conclusion: Uni-MDTrack effectively addresses spatio-temporal context limitations in tracking, offering efficient memory compression, dynamic state modeling, and unified multi-modal support with strong generalization capabilities.

Abstract: With the advent of Transformer-based one-stream trackers that possess strong capability in inter-frame relation modeling, recent research has increasingly focused on how to introduce spatio-temporal context. However, most existing methods rely on a limited number of historical frames, which not only leads to insufficient utilization of the context, but also inevitably increases the length of input and incurs prohibitive computational overhead. Methods that query an external memory bank, on the other hand, suffer from inadequate fusion between the retrieved spatio-temporal features and the backbone. Moreover, using discrete historical frames as context overlooks the rich dynamics of the target. To address the issues, we propose Uni-MDTrack, which consists of two core components: Memory-Aware Compression Prompt (MCP) module and Dynamic State Fusion (DSF) module. MCP effectively compresses rich memory features into memory-aware prompt tokens, which deeply interact with the input throughout the entire backbone, significantly enhancing the performance while maintaining a stable computational load. DSF complements the discrete memory by capturing the continuous dynamic, progressively introducing the updated dynamic state features from shallow to deep layers, while also preserving high efficiency. Uni-MDTrack also supports unified tracking across RGB, RGB-D/T/E, and RGB-Language modalities. Experiments show that in Uni-MDTrack, training only the MCP, DSF, and prediction head, keeping the proportion of trainable parameters around 30%, yields substantial performance gains, achieves state-of-the-art results on 10 datasets spanning five modalities. Furthermore, both MCP and DSF exhibit excellent generality, functioning as plug-and-play components that can boost the performance of various baseline trackers, while significantly outperforming existing parameter-efficient training approaches.

Method: Introduces LongVidSearch benchmark with 3,000 questions over 447 long videos (avg 26 min), enforcing retrieval necessity where Hop-k questions require exactly k necessary evidence clips. Provides unified tool interface for controlled evaluation and measures both answer accuracy and tool-call cost.

Result: GPT-5 achieves highest accuracy (42.43%) but remains below 50%, outperforming Gemini 3 Pro (30.97%) and GPT-4o (19.20%). With gold evidence clips, performance becomes near-perfect, confirming retrieval planning as primary bottleneck.

Conclusion: Multi-hop retrieval planning is challenging for current LLM-based agents, with retrieval planning identified as the main bottleneck rather than answer generation. The benchmark enables standardized evaluation of agentic video understanding systems.

Abstract: Long video question answering (Long-Video QA) increasingly relies on agentic tool use to retrieve evidence from long videos. In realistic settings, this process often requires multi-hop retrieval, where agents must iteratively gather multiple discontinuous evidence clips. However, existing long-video benchmarks are largely static: they rarely enforce strict multi-hop retrieval and typically lack a standardized evidence-access interface, making it difficult to separate failures in retrieval planning from those in answer generation. To address this gap, we introduce LongVidSearch, a benchmark for evaluating agentic multi-hop evidence retrieval planning in long videos under standardized access constraints. LongVidSearch enforces retrieval necessity: a Hop-k question requires exactly k necessary evidence clips, and removing any single clip renders the question unsolvable. The benchmark contains 3,000 questions over 447 long videos (average length 26 minutes), covering four reasoning categories: State Mutation, Causal Inference, Global Summary, and Visual Tracking, with 2-hop, 3-hop, and 4-hop evidence requirements. To ensure fair and controlled evaluation, all agents interact with LongVidSearch through a unified tool interface, which fixes the retrieval backend and isolates the agent’s ability to formulate queries and plan iterative retrieval. In addition to answer accuracy, we measure tool-call cost to analyze the accuracy-efficiency trade-off under identical access conditions. We evaluate VideoAgent-style QA agents with multiple backbone LLMs using three-judge majority voting. GPT-5 achieves the highest accuracy (42.43), outperforming Gemini 3 Pro (30.97) and GPT-4o (19.20), yet remaining below 50 %, highlighting the difficulty of multi-hop retrieval planning. With gold evidence clips, performance becomes near-perfect, confirming retrieval planning as the primary bottleneck.

Method: Augments each transformer block with residual low-rank bottleneck adapters where up-projection is zero-initialized, ensuring adapted network starts exactly at pretrained function. Formalizes adapter rank as capacity budget for approximating downstream task shifts.

Result: On 5-dataset transfer suite, improves top-1 accuracy over head-only transfer by +14.9 points average while training only 0.92% of full fine-tuning parameters. Outperforms full fine-tuning on 10 of 15 dataset-backbone pairs. Shows monotonic but diminishing accuracy gains with increasing rank.

Conclusion: AdapterTune provides stable optimization and principled capacity guidance for Vision Transformer transfer learning, achieving strong performance with minimal parameter updates while eliminating early-epoch representation drift.

Abstract: Frozen-backbone transfer with Vision Transformers faces two under-addressed issues: optimization instability when adapters are naively inserted into a fixed feature extractor, and the absence of principled guidance for setting adapter capacity. We introduce AdapterTune, which augments each transformer block with a residual low-rank bottleneck whose up-projection is zero-initialized, guaranteeing that the adapted network starts exactly at the pretrained function and eliminates early-epoch representation drift. On the analytical side, we formalize adapter rank as a capacity budget for approximating downstream task shifts in feature space. The resulting excess-risk decomposition predicts monotonic but diminishing accuracy gains with increasing rank, an ``elbow’’ behavior we confirm through controlled sweeps. We evaluate on 9 datasets and 3 backbone scales with multi-seed reporting throughout. On a core 5 dataset transfer suite, AdapterTune improves top-1 accuracy over head-only transfer by +14.9 points on average while training only 0.92 of the parameters required by full fine-tuning, and outperforms full fine-tuning on 10 of 15 dataset-backbone pairs. Across the full benchmark, AdapterTune improves over head-only transfer on every dataset-backbone pair tested. Ablations on rank, placement, and initialization isolate each design choice. The code is available at: https://github.com/salimkhazem/adaptertune

Method: Uses spiking neural networks (SNNs) with spiking convolutional layers for spatio-temporal feature extraction from WiFi CSI signals, a novel temporal attention mechanism, spiking fully connected layers, and a voting layer for classification.

Result: Achieves competitive accuracy in single-action recognition (95.83% accuracy) and superior performance in multi-action recognition on three benchmark datasets, while reducing energy consumption by at least half compared to other methods.

Conclusion: Wi-Spike establishes state-of-the-art in WiFi-based multi-action HAR, offering a promising solution for real-time, energy-efficient edge sensing applications by leveraging SNNs’ event-driven and low-power characteristics.

Abstract: WiFi-based human action recognition (HAR) has gained significant attention due to its non-intrusive and privacy-preserving nature. However, most existing WiFi sensing models predominantly focus on improving recognition accuracy, while issues of power consumption and energy efficiency remain insufficiently discussed. In this work, we present Wi-Spike, a bio-inspired spiking neural network (SNN) framework for efficient and accurate action recognition using WiFi channel state information (CSI) signals. Specifically, leveraging the event-driven and low-power characteristics of SNNs, Wi-Spike introduces spiking convolutional layers for spatio-temporal feature extraction and a novel temporal attention mechanism to enhance discriminative representation. The extracted features are subsequently encoded and classified through spiking fully connected layers and a voting layer. Comprehensive experiments on three benchmark datasets (NTU-Fi-HAR, NTU-Fi-HumanID, and UT-HAR) demonstrate that Wi-Spike achieves competitive accuracy in single-action recognition and superior performance in multi-action recognition tasks. As for energy consumption, Wi-Spike reduces the energy cost by at least half compared with other methods, while still achieving 95.83% recognition accuracy in human activity recognition. More importantly, Wi-Spike establishes a new state-of-the-art in WiFi-based multi-action HAR, offering a promising solution for real-time, energy-efficient edge sensing applications.

Method: Proposes ego-motion-guided trajectory prediction network based on Mamba model: 1) Two Mamba encoders extract pedestrian motion and ego-motion features, 2) Ego-motion guided Mamba decoder explicitly models relative motion by integrating pedestrian motion as historical context with ego-motion as guiding cues, 3) Generates future trajectory from decoded features.

Result: Extensive experiments demonstrate effectiveness, achieving state-of-the-art performance on PIE and JAAD datasets.

Conclusion: The proposed ego-motion-guided Mamba-based approach effectively addresses the challenge of pedestrian trajectory prediction from egocentric perspective by modeling relative motion dynamics.

Abstract: Future trajectory prediction of a tracked pedestrian from an egocentric perspective is a key task in areas such as autonomous driving and robot navigation. The challenge of this task lies in the complex dynamic relative motion between the ego-camera and the tracked pedestrian. To address this challenge, we propose an ego-motion-guided trajectory prediction network based on the Mamba model. Firstly, two Mamba models are used as encoders to extract pedestrian motion and ego-motion features from pedestrian movement and ego-vehicle movement, respectively. Then, an ego-motion guided Mamba decoder that explicitly models the relative motion between the pedestrian and the vehicle by integrating pedestrian motion features as historical context with ego-motion features as guiding cues to capture decoded features. Finally, the future trajectory is generated from the decoded features corresponding to the future timestamps. Extensive experiments demonstrate the effectiveness of the proposed model, which achieves state-of-the-art performance on the PIE and JAAD datasets.

Method: Combines four components: 1) dense predictive loss with masking-based objective where both visible and masked tokens contribute, 2) deep self-supervision applied hierarchically across multiple encoder layers, 3) multi-modal tokenizers for unified image/video training, and 4) effective scaling of model capacity and training data.

Result: State-of-the-art performance on multiple benchmarks: 7.71 mAP on Ego4D for object-interaction anticipation, 40.8 Recall@5 on EPIC-KITCHENS for action anticipation, 20-point improvement in robot grasping, strong robotic navigation (5.687 ATE), depth estimation (0.307 RMSE), and global recognition (77.7 on Something-Something-V2).

Conclusion: V-JEPA 2.1 significantly advances dense visual understanding and world modeling through its combined approach, producing representations that are spatially structured, semantically coherent, and temporally consistent across diverse applications.

Abstract: We present V-JEPA 2.1, a family of self-supervised models that learn dense, high-quality visual representations for both images and videos while retaining strong global scene understanding. The approach combines four key components. First, a dense predictive loss uses a masking-based objective in which both visible and masked tokens contribute to the training signal, encouraging explicit spatial and temporal grounding. Second, deep self-supervision applies the self-supervised objective hierarchically across multiple intermediate encoder layers to improve representation quality. Third, multi-modal tokenizers enable unified training across images and videos. Finally, the model benefits from effective scaling in both model capacity and training data. Together, these design choices produce representations that are spatially structured, semantically coherent, and temporally consistent. Empirically, V-JEPA 2.1 achieves state-of-the-art performance on several challenging benchmarks, including 7.71 mAP on Ego4D for short-term object-interaction anticipation and 40.8 Recall@5 on EPIC-KITCHENS for high-level action anticipation, as well as a 20-point improvement in real-robot grasping success rate over V-JEPA-2 AC. The model also demonstrates strong performance in robotic navigation (5.687 ATE on TartanDrive), depth estimation (0.307 RMSE on NYUv2 with a linear probe), and global recognition (77.7 on Something-Something-V2). These results show that V-JEPA 2.1 significantly advances the state of the art in dense visual understanding and world modeling.

Method: An order-preserving polygon augmentation strategy that performs transformations in mask space and then projects surviving vertices back into index-space to restore adjacency relations. This repair maintains the original traversal order of the polygon and preserves topological consistency with minimal computational overhead.

Result: The approach reliably restores connectivity, achieving near-perfect Cyclic Adjacency Preservation (CAP) across both single and compound augmentations.

Conclusion: The method effectively preserves topological consistency in polygon annotations during geometric data augmentation, which is crucial for maintaining structural relationships in complex domains like architectural analysis.

Abstract: Geometric data augmentation is widely used in segmentation pipelines and typically assumes that polygon annotations represent simply connected regions. However, in structured domains such as architectural floorplan analysis, ring-type regions are often encoded as a single cyclic polygon chain connecting outer and inner boundaries. During augmentation, clipping operations may remove intermediate vertices and disrupt this cyclic connectivity, breaking the structural relationship between the boundaries. In this work, we introduce an order-preserving polygon augmentation strategy that performs transformations in mask space and then projects surviving vertices back into index-space to restore adjacency relations. This repair maintains the original traversal order of the polygon and preserves topological consistency with minimal computational overhead. Experiments demonstrate that the approach reliably restores connectivity, achieving near-perfect Cyclic Adjacency Preservation (CAP) across both single and compound augmentations.

Method: 1) Introduce CoWTalk benchmark with 3D arterial anatomies featuring controllable synthesized anatomical errors and corresponding repair instructions. 2) Propose iterative refinement model representing 3D shapes as vector sets that interact with textual instructions to progressively update target shapes.

Result: Experimental results show significant improvements over corrupted inputs and competitive performance compared to baselines, demonstrating feasibility of language-driven clinician-in-the-loop refinement for 3D medical shape modeling.

Conclusion: The work presents a promising approach for interactive refinement of 3D medical shapes using natural language instructions, addressing the data scarcity problem through synthetic error generation and establishing a benchmark for future research.

Abstract: Accurate 3D anatomical segmentation is essential for clinical diagnosis and surgical planning. However, automated models frequently generate suboptimal shape predictions due to factors such as limited and imbalanced training data, inadequate labeling quality, and distribution shifts between training and deployment settings. A natural solution is to iteratively refine the predicted shape based on the radiologists’ verbal instructions. However, this is hindered by the scarcity of paired data that explicitly links erroneous shapes to corresponding corrective instructions. As an initial step toward addressing this limitation, we introduce CoWTalk, a benchmark comprising 3D arterial anatomies with controllable synthesized anatomical errors and their corresponding repairing instructions. Building on this benchmark, we further propose an iterative refinement model that represents 3D shapes as vector sets and interacts with textual instructions to progressively update the target shape. Experimental results demonstrate that our method achieves significant improvements over corrupted inputs and competitive baselines, highlighting the feasibility of language-driven clinician-in-the-loop refinement for 3D medical shapes modeling.

Method: Proposes WorldVLM: a hybrid architecture where a high-level VLM generates behavior commands to guide a driving World Model, enabling interpretable and context-aware actions while maintaining dynamic prediction capabilities.

Result: The paper evaluates conditioning strategies and provides insights into hybrid design challenges for combining VLMs and WMs in autonomous driving systems.

Conclusion: Unifying VLMs and WMs creates a promising direction for autonomous driving that leverages contextual reasoning and dynamic prediction while enhancing generalization and interpretability.

Abstract: Autonomous driving systems depend on on models that can reason about high-level scene contexts and accurately predict the dynamics of their surrounding environment. Vision- Language Models (VLMs) have recently emerged as promising tools for decision-making and scene understanding, offering strong capabilities in contextual reasoning. However, their limited spatial comprehension constrains their effectiveness as end-to-end driving models. World Models (WM) internalize environmental dynamics to predict future scene evolution. Recently explored as ego-motion predictors and foundation models for autonomous driving, they represent a promising direction for addressing key challenges in the field, particularly enhancing generalization while maintaining dynamic prediction. To leverage the complementary strengths of context-based decision making and prediction, we propose WorldVLM: A hybrid architecture that unifies VLMs and WMs. In our design, the high-level VLM generates behavior commands to guide the driving WM, enabling interpretable and context-aware actions. We evaluate conditioning strategies and provide insights into the hybrid design challenges.

Method: Introduces DarkClusters-15k dataset of 15,000 simulated clusters with paired mass and photometry maps, trains a plug-and-play diffusion prior to learn statistical relationship between mass and light, and draws posterior samples constrained by weak- and strong-lensing observables for principled reconstructions.

Result: The method requires no expert tuning, runs in minutes rather than hours, achieves higher accuracy, and matches expertly-tuned reconstructions of the MACS 1206 cluster while providing well-calibrated uncertainties.

Conclusion: The approach enables scalable, automated cluster mass reconstruction for upcoming wide-field cosmological surveys, with released method and dataset supporting development and benchmarking.

Abstract: Galaxy clusters are powerful probes of astrophysics and cosmology through gravitational lensing: the clusters’ mass, dominated by 85% dark matter, distorts background light. Yet, mass reconstruction lacks the scalability and large-scale benchmarks to process the hundreds of thousands of clusters expected from forthcoming wide-field surveys. We introduce a fully automated method to reconstruct cluster surface mass density from photometry and gravitational lensing observables. Central to our approach is DarkClusters-15k, our new dataset of 15,000 simulated clusters with paired mass and photometry maps, the largest benchmark to date, spanning multiple redshifts and simulation frameworks. We train a plug-and-play diffusion prior on DarkClusters-15k that learns the statistical relationship between mass and light, and draw posterior samples constrained by weak- and strong-lensing observables; this yields principled reconstructions driven by explicit physics, alongside well-calibrated uncertainties. Our approach requires no expert tuning, runs in minutes rather than hours, achieves higher accuracy, and matches expertly-tuned reconstructions of the MACS 1206 cluster. We release our method and DarkClusters-15k to support development and benchmarking for upcoming wide-field cosmological surveys.

Method: RAZOR identifies the most important layers and attention heads by measuring their contribution to forgetting target data while preserving useful knowledge, then updates these parts using carefully regularized rules with gradual component editing to avoid harming overall performance.

Result: RAZOR achieves highly accurate and stable forgetting on CLIP, Stable Diffusion, and vision-language models across identity, style, and object erasure tasks, even under quantization, with stronger retention, better efficiency, and significantly faster operation than prior methods.

Conclusion: RAZOR is a practical and scalable solution for safe, adaptive unlearning in transformer-based vision models, offering efficient information removal without compromising overall model performance.

Abstract: Transformer based diffusion and vision-language models have achieved remarkable success; yet, efficiently removing undesirable or sensitive information without retraining remains a central challenge for model safety and compliance. We introduce Ratio-Aware Zero/One-step Optimized Retentive unlearning (RAZOR), a lightweight, model-agnostic unlearning framework that generalizes forgetting updates to coordinated multi-layer and multi-head edits within transformer backbones. RAZOR identifies the most important layers and attention heads by measuring how much they contribute to forgetting the target data while preserving useful knowledge. Then, it updates these parts of the model using a carefully regularized rule to avoid harming overall performance. The set of edited components grows gradually, ensuring precise unlearning without over-editing or damaging unrelated capabilities. We evaluate RAZOR on CLIP, Stable Diffusion, and vision-language models (VLMs) using widely adopted unlearning benchmarks covering identity, style, and object erasure tasks. Our results show that RAZOR achieves highly accurate and stable forgetting, even under quantization. This approach offers stronger retention and better efficiency than prior methods. Notably, it also operates significant faster than conventional techniques. These results demonstrate that RAZOR is a practical and scalable solution for safe, adaptive unlearning in transformer-based vision models.

Method: Introduces SVE-ASCII framework for Symbolic Visual Expression in pure text space. Constructs ASCIIArt-7K dataset via “Seed-and-Evolve” pipeline that augments human-curated anchors through in-context stylistic editing. Implements unified instruction-tuning strategy jointly optimizing for both Text-to-ASCII generation and ASCII-to-Text understanding.

Result: Reveals a critical phenomenon regarding task duality: generative training significantly enhances visual comprehension, confirming a mutually reinforcing cycle in symbolic visual processing. Establishes ASCIIArt-Bench benchmark and releases SVE-ASCII model as baseline for native text-based visual intelligence.

Conclusion: Demonstrates that LLMs possess latent visual representation capabilities that can be unlocked through ASCII art as a text-native format. Shows that joint training for generation and understanding creates a mutually reinforcing cycle, advancing native text-based visual intelligence.

Abstract: Current multimodal approaches predominantly treat visual generation as an external process, relying on pixel rendering or code execution, thereby overlooking the native visual representation capabilities latent within Large Language Models (LLMs). In this work, we unlock this potential through ASCII art, a compact, efficient, and text-native visual format. We introduce SVE-ASCII, a unified framework designed to elicit and benchmark Symbolic Visual Expression directly within the pure text space. To address the scarcity of systematic resources, we construct ASCIIArt-7K, a high-quality dataset synthesized via a novel “Seed-and-Evolve” pipeline that augments human-curated anchors through in-context stylistic editing. We further implement a unified instruction-tuning strategy that jointly optimizes for both Generation (Text-to-ASCII) and Understanding (ASCII-to-Text). Crucially, our experiments reveal a critical phenomenon regarding task duality: while it is established that perception aids generation, we provide compelling evidence that generative training significantly enhances visual comprehension. This confirms a mutually reinforcing cycle in symbolic visual processing, a relationship previously hypothesized but rarely empirically demonstrated in the visual domain. We release our dataset, the ASCIIArt-Bench benchmark, and the SVE-ASCII model, establishing a robust baseline for native text-based visual intelligence.

Method: Identifies a modality-induced bias direction consistently observed across datasets arising from suboptimal coupling between LLM backbone and visual encoders. Proposes projecting cross-modal features onto the null space of this bias direction to remove corresponding components, requiring only a single forward pass.

Result: Effectively breaks the conventional safety-utility tradeoff, simultaneously improving both safety and utility across diverse benchmarks through efficient inference-time defense.

Conclusion: Safety and utility in Large Vision-Language Models are not inherently antagonistic; the identified modality-induced bias direction undermines both tasks, and removing it through projection can simultaneously enhance both safety and utility.

Abstract: Existing jailbreak defence frameworks for Large Vision-Language Models often suffer from a safety utility tradeoff, where strengthening safety inadvertently degrades performance on general visual-grounded reasoning tasks. In this work, we investigate whether safety and utility are inherently antagonistic objectives. We focus on a modality induced bias direction consistently observed across datasets, which arises from suboptimal coupling between the Large Language Model backbone and visual encoders. We further demonstrate that this direction undermines performance on both tasks. Leveraging this insight, we propose Two Birds, One Projection, an efficient inference time jailbreak defence that projects cross-modal features onto the null space of the identified bias direction to remove the corresponding components. Requiring only a single forward pass, our method effectively breaks the conventional tradeoff, simultaneously improving both safety and utility across diverse benchmarks.

Method: Proposes EMDUL approach that: 1) trains a pseudo-label estimator to annotate unlabeled mmWave data, and 2) translates annotated LiDAR point clouds to mmWave counterparts. This expands dataset volume and diversity.

Result: Expanded dataset significantly boosts HPE model performance with 15.1% error reduction for in-domain settings and 18.9% error reduction for out-of-domain settings, improving generalization ability.

Conclusion: EMDUL effectively addresses data scarcity in mmWave HPE by leveraging available unlabeled mmWave data and LiDAR datasets, enabling better model training and generalization.

Abstract: Current mmWave datasets for human pose estimation (HPE) are scarce and lack diversity in both point cloud (PC) attributes and human poses, severely hampering the generalization ability of their trained models. On the other hand, unlabeled mmWave HPE data and diverse LiDAR HPE datasets are readily available. We propose EMDUL, a novel approach to expand the volume and diversity of an existing mmWave dataset using unlabeled mmWave data and a LiDAR dataset. EMDUL trains a pseudo-label estimator to annotate the unlabeled mmWave data and is able to convert, or translate, a given annotated LiDAR PC to its mmWave counterpart. Expanded with both LiDAR-converted and pseudo-labeled mmWave PCs, our mmWave dataset significantly boosts the performance and generalization ability of all our HPE models, with substantial 15.1% and 18.9% error reductions for in-domain and out-of-domain settings, respectively.

Method: Introduces latent reward model that scores partially denoised latents at arbitrary timesteps for visual quality, motion quality, and text alignment. Proposes LatSearch with Reward-Guided Resampling and Pruning (RGRP): resampling candidates according to reward-normalized probabilities, and pruning at final step to retain highest cumulative reward candidate.

Result: Evaluated on VBench-2.0 benchmark, LatSearch consistently improves video generation across multiple evaluation dimensions compared to baseline Wan2.1 model.

Conclusion: LatSearch enables efficient inference-time scaling for video diffusion by providing intermediate, informative feedback along denoising trajectory, unlocking gains in controllability, sample efficiency and generation quality.

Abstract: The recent success of inference-time scaling in large language models has inspired similar explorations in video diffusion. In particular, motivated by the existence of “golden noise” that enhances video quality, prior work has attempted to improve inference by optimising or searching for better initial noise. However, these approaches have notable limitations: they either rely on priors imposed at the beginning of noise sampling or on rewards evaluated only on the denoised and decoded videos. This leads to error accumulation, delayed and sparse reward signals, and prohibitive computational cost, which prevents the use of stronger search algorithms. Crucially, stronger search algorithms are precisely what could unlock substantial gains in controllability, sample efficiency and generation quality for video diffusion, provided their computational cost can be reduced. To fill in this gap, we enable efficient inference-time scaling for video diffusion through latent reward guidance, which provides intermediate, informative and efficient feedback along the denoising trajectory. We introduce a latent reward model that scores partially denoised latents at arbitrary timesteps with respect to visual quality, motion quality, and text alignment. Building on this model, we propose LatSearch, a novel inference-time search mechanism that performs Reward-Guided Resampling and Pruning (RGRP). In the resampling stage, candidates are sampled according to reward-normalised probabilities to reduce over-reliance on the reward model. In the pruning stage, applied at the final scheduled step, only the candidate with the highest cumulative reward is retained, improving both quality and efficiency. We evaluate LatSearch on the VBench-2.0 benchmark and demonstrate that it consistently improves video generation across multiple evaluation dimensions compared to the baseline Wan2.1 model.

Method: Interp3R enhances pointmap-based models by using asynchronous event data to interpolate pointmaps between frames. It jointly recovers depth and camera poses by aligning interpolated pointmaps with those from frame-based models into a consistent spatial framework.

Result: Trained exclusively on synthetic data, Interp3R shows strong generalization across synthetic and real-world benchmarks, outperforming state-of-the-art baselines that use 2D video frame interpolation followed by 3D geometry estimation.

Conclusion: Interp3R successfully enables temporally continuous 3D geometry estimation by integrating event data with pointmap-based models, addressing the limitation of discrete-time reconstruction in current 3D foundation models.

Abstract: In recent years, 3D visual foundation models pioneered by pointmap-based approaches such as DUSt3R have attracted a lot of interest, achieving impressive accuracy and strong generalization across diverse scenes. However, these methods are inherently limited to recovering scene geometry only at the discrete time instants when images are captured, leaving the scene evolution during the blind time between consecutive frames largely unexplored. We introduce Interp3R, to the best of our knowledge the first method that enhances pointmap-based models to estimate depth and camera poses at arbitrary time instants. Interp3R leverages asynchronous event data to interpolate pointmaps produced by frame-based models, enabling temporally continuous geometric representations. Depth and camera poses are then jointly recovered by aligning the interpolated pointmaps together with those predicted by the underlying frame-based models into a consistent spatial framework. We train Interp3R exclusively on a synthetic dataset, yet demonstrate strong generalization across a wide range of synthetic and real-world benchmarks. Extensive experiments show that Interp3R outperforms by a considerable margin state-of-the-art baselines that follow a two-stage pipeline of 2D video frame interpolation followed by 3D geometry estimation.

Method: Developed Video Detector (VD) with two modules: real-time module (VD-RT) for intersection control and offline analytical module (VD-Offline) for detailed traffic analysis. Implemented three system configurations using SSD Inception v2, Faster R-CNN Inception v2, and CenterNet ResNet-50 V1 FPN architectures. Trained on 108,000 annotated images across 6-10 vehicle classes.

Result: Achieved up to 90% test accuracy and 29.5 mAP@0.5 detection performance with real-time throughput of 37 FPS on HD video streams. Successfully deployed in field tests with Istanbul IT and Smart City Technologies Inc., demonstrating stable operation under diverse environmental conditions.

Conclusion: The framework provides a scalable, deployable vision-based solution for intelligent transportation systems and smart-city traffic management, supporting various functions like virtual loop detection, vehicle counting, tracking, queue estimation, speed analysis, and multiclass classification without embedded road sensors.

Abstract: Urban traffic management increasingly requires intelligent sensing systems capable of adapting to dynamic traffic conditions without costly infrastructure modifications. Vision-based vehicle detection has therefore become a key technology for modern intelligent transportation systems. This study presents Video Detector (VD), a dual-phase vision-based traffic intersection management system designed as a flexible and cost-effective alternative to traditional inductive loop detectors. The framework integrates a real-time module (VD-RT) for intersection control with an offline analytical module (VD-Offline) for detailed traffic behavior analysis. Three system configurations were implemented using SSD Inception v2, Faster R-CNN Inception v2, and CenterNet ResNet-50 V1 FPN, trained on datasets totaling 108,000 annotated images across 6-10 vehicle classes. Experimental results show detection performance of up to 90% test accuracy and 29.5 mAP@0.5, while maintaining real-time throughput of 37 FPS on HD video streams. Field deployments conducted in collaboration with Istanbul IT and Smart City Technologies Inc. (ISBAK) demonstrate stable operation under diverse environmental conditions. The system supports virtual loop detection, vehicle counting, multi-object tracking, queue estimation, speed analysis, and multiclass vehicle classification, enabling comprehensive intersection monitoring without the need for embedded road sensors. The annotated dataset and training pipeline are publicly released to support reproducibility. These results indicate that the proposed framework provides a scalable and deployable vision-based solution for intelligent transportation systems and smart-city traffic management.

Method: Distill a compact VAE encoder only at low resolutions (up to 256²), then evaluate at higher, unseen resolutions (512²). Analyze latent distributions across resolutions and use simple resolution remapping: upsampling inputs before encoding and downsampling reconstructions for evaluation.

Result: The distilled encoder trained only at low resolutions achieves dramatically improved reconstruction performance at higher, unseen resolutions. Higher-resolution inputs produce latent representations more closely aligned with the teacher’s manifold. Experiments on ImageNet-256 show substantial gains across PSNR, MSE, SSIM, LPIPS, and rFID metrics with resolution remapping.

Conclusion: VAE encoder distillation learns resolution-consistent latent manifolds rather than resolution-specific pixel mappings. High training costs on memory, time, and high-resolution datasets are not necessary for distilling a VAE with high-resolution image reconstruction capabilities. Models can learn detailed knowledge from teacher models even when trained on low-resolution data.

Abstract: Variational Autoencoder (VAE) encoders play a critical role in modern generative models, yet their computational cost often motivates the use of knowledge distillation or quantification to obtain compact alternatives. Existing studies typically believe that the model work better on the samples closed to their training data distribution than unseen data distribution. In this work, we report a counter-intuitive phenomenon in VAE encoder distillation: a compact encoder distilled only at low resolutions exhibits poor reconstruction performance at its native resolution, but achieves dramatically improved results when evaluated at higher, unseen input resolutions. Despite never being trained beyond $256^2$ resolution, the distilled encoder generalizes effectively to $512^2$ resolution inputs, partially inheriting the teacher model’s resolution preference.We further analyze latent distributions across resolutions and find that higher-resolution inputs produce latent representations more closely aligned with the teacher’s manifold. Through extensive experiments on ImageNet-256, we show that simple resolution remapping-upsampling inputs before encoding and downsampling reconstructions for evaluation-leads to substantial gains across PSNR, MSE, SSIM, LPIPS, and rFID metrics. These findings suggest that VAE encoder distillation learns resolution-consistent latent manifolds rather than resolution-specific pixel mappings. This also means that the high training cost on memory, time and high-resolution datasets are not necessary conditions for distilling a VAE with high-resolution image reconstruction capabilities. On low resolution datasets, the distillation model still could learn the detailed knowledge of the teacher model in high-resolution image reconstruction.

Method: Two-component framework: 1) Pseudo keyframe labels derived from LMMs that provide informative supervision, and 2) Coverage regularization that promotes diverse, complementary evidence across time to avoid redundant frame selection.

Result: Experiments on NExT-QA show significant accuracy improvements, especially for temporal and causal question types, establishing keyframe selection as an effective and learnable module for VideoQA.

Conclusion: The proposed question-aware keyframe selection framework with LMM-generated pseudo labels and coverage regularization effectively addresses efficiency and reasoning challenges in VideoQA, particularly improving performance on temporal and causal reasoning tasks.

Abstract: Large multimodal models (LMMs) have recently demonstrated remarkable performance in video question answering (VideoQA), yet reasoning over video remains challenging due to high inference cost and diluted information. Keyframe selection offers efficiency and sharper reasoning but suffers from sparse supervision and redundant frame choices when relying only on image-text similarity. We present a question-aware keyframe selection framework with two components: pseudo keyframe labels derived from LMMs that provide informative supervision and a coverage regularization that promotes diverse, complementary evidence across time. Experiments on NExT-QA show that our method significantly improves accuracy, especially for temporal and causal question types, establishing keyframe selection as an effective and learnable module for VideoQA.

Method: ASAP uses: 1) Dynamic bidirectional soft attention masks to mitigate attention shift and select genuinely informative tokens, and 2) Weighted soft merging to combine semantically similar tokens, preserving only the most feature-dense visual patches for subsequent layers.

Result: ASAP achieves virtually lossless compression, retaining 99.02% of original LLaVA-NeXT-7B performance while reducing computational FLOPs by ~80%.

Conclusion: ASAP provides an effective training-free solution for computational efficiency in LVLMs by addressing attention shift and token redundancy, enabling high-performance vision-language understanding with significantly reduced computational cost.

Abstract: While Large Vision-Language Models (LVLMs) demonstrate exceptional multi-modal capabilities, the quadratic computational cost of processing high-resolution visual tokens remains a critical bottleneck. Though recent token reduction strategies attempt to accelerate inference, such methods inadequately exploit attention values and fail to address token redundancy. More critically, they overlook the ``attention shift’’ phenomenon inherent in LVLMs, which skews token attention scores. In this work, we propose ASAP, a novel training-free, KV-Cache-compatible pruning recipe that comprehensively addresses these limitations. First, we mitigate the attention shift by utilizing a dynamic bidirectional soft attention mask, ensuring the selection of genuinely informative tokens rather than naive attention-based selection. Second, we posit that high semantic redundancy within the token set degrades performance. We therefore introduce a weighted soft merging component that merges semantically similar tokens, preserving only the most feature-dense visual patches for subsequent layers. ASAP achieves virtually lossless compression of visual context, retaining 99.02% of the original LLaVA-NeXT-7B performance while aggressively slashing computational FLOPs by ~80%.

Method: 1) Analyze VLM vulnerabilities in medical image spatial grounding across image modalities, slice directions, coordinate systems, and anatomical terminology. 2) Create MIS-Ground benchmark for comprehensive testing. 3) Develop MIS-SemSam, a low-cost, inference-time, model-agnostic optimization using semantic sampling to improve spatial grounding.

Result: MIS-SemSam improves Qwen3-VL-32B’s accuracy on the MIS-Ground benchmark by 13.06%, demonstrating effective enhancement of spatial grounding capabilities in medical imaging contexts.

Conclusion: Medical image spatial grounding presents unique challenges for VLMs that require specialized benchmarks and optimization techniques; MIS-Ground enables measurement and reproducibility, while MIS-SemSam provides a practical solution for improving VLM performance in this domain.

Abstract: Vision language models (VLMs) have shown significant promise in visual grounding for images as well as videos. In medical imaging research, VLMs represent a bridge between object detection and segmentation, and report understanding and generation. However, spatial grounding of anatomical structures in the three-dimensional space of medical images poses many unique challenges. In this study, we examine image modalities, slice directions, and coordinate systems as differentiating factors for vision components of VLMs, and the use of anatomical, directional, and relational terminology as factors for the language components. We then demonstrate that visual and textual prompting systems such as labels, bounding boxes, and mask overlays have varying effects on the spatial grounding ability of VLMs. To enable measurement and reproducibility, we introduce \textbf{MIS-Ground}, a benchmark that comprehensively tests a VLM for vulnerabilities against specific modes of \textbf{M}edical \textbf{I}mage \textbf{S}patial \textbf{Ground}ing. We release MIS-Ground to the public at \href{https://anonymous.4open.science/r/mis-ground}{\texttt{anonymous.4open.science/r/mis-ground}}. In addition, we present \textbf{MIS-SemSam}, a low-cost, inference-time, and model-agnostic optimization of VLMs that improve their spatial grounding ability with the use of \textbf{Sem}antic \textbf{Sam}pling. We find that MIS-SemSam improves the accuracy of Qwen3-VL-32B on MIS-Ground by 13.06%.

Method: Proposes VisionZip, a method that selects a subset of informative visual tokens from vision encoders like CLIP/SigLIP to reduce redundancy while preserving essential information.

Result: Outperforms previous SOTA by at least 5% across settings, improves inference speed 8x (prefilling time), enables LLaVA-Next 13B to infer faster than 7B while achieving better results.

Conclusion: VisionZip effectively addresses visual token redundancy, improves efficiency and performance, and encourages focus on better visual feature extraction rather than increasing token length.

Abstract: Recent advancements in vision-language models have enhanced performance by increasing the length of visual tokens, making them much longer than text tokens and significantly raising computational costs. However, we observe that the visual tokens generated by popular vision encoders, such as CLIP and SigLIP, contain significant redundancy. To address this, we introduce VisionZip, a simple yet effective method that selects a set of informative tokens for input to the language model, reducing visual token redundancy and improving efficiency while maintaining model performance. The proposed VisionZip can be widely applied to image and video understanding tasks and is well-suited for multi-turn dialogues in real-world scenarios, where previous methods tend to underperform. Experimental results show that VisionZip outperforms the previous state-of-the-art method by at least 5% performance gains across nearly all settings. Moreover, our method significantly enhances model inference speed, improving the prefilling time by 8x and enabling the LLaVA-Next 13B model to infer faster than the LLaVA-Next 7B model while achieving better results. Furthermore, we analyze the causes of this redundancy and encourage the community to focus on extracting better visual features rather than merely increasing token length. Our code is available at https://github.com/dvlab-research/VisionZip .

Method: Render scene geometry into cubemap from fixed world point, then splat each texel to screen as world-space quad; cubemap indexing provides rotation invariance, grid-snapping origin provides translation invariance

Result: Achieves stable pixel art rendering with camera movement by avoiding perspective drift issues; maintains pixel stability through rotation and translation invariance

Conclusion: Texel splatting solves pixel stability problem for perspective projection in pixel art rendering, though limited by fixed origin visibility and disocclusion at probe boundaries

Abstract: Rendering 3D scenes as pixel art requires that discrete pixels remain stable as the camera moves. Existing methods snap the camera to a grid. Under orthographic projection, this works: every pixel shifts by the same amount, and a single snap corrects all of them. Perspective breaks this. Pixels at different depths drift at different rates, and no single snap corrects all depths. Texel splatting avoids this entirely. Scene geometry is rendered into a cubemap from a fixed point in the world, and each texel is splatted to the screen as a world-space quad. Cubemap indexing gives rotation invariance. Grid-snapping the origin gives translation invariance. The primary limitation is that a fixed origin cannot see all geometry; disocclusion at probe boundaries remains an open tradeoff.

Method: Introduce a large-scale dataset grounded in verifiable cadastral vector data with 3.8 million annotated objects across 510k high-resolution images covering 135 granular semantic categories. Validate through comprehensive instruction-tuning benchmark spanning seven spatial reasoning tasks using standard LLaVA architecture.

Result: Current RS-specialized and commercial models (e.g., Gemini) struggle in zero-shot settings, but high-fidelity supervision effectively bridges this gap, enabling standard architectures to master fine-grained spatial grounding without complex architectural modifications.

Conclusion: The proposed large-scale dataset and benchmark enable significant improvements in spatial understanding for multimodal LLMs in Earth Observation, demonstrating that high-quality supervision can overcome current limitations without requiring complex architectural changes.

Abstract: Precise spatial understanding in Earth Observation is essential for translating raw aerial imagery into actionable insights for critical applications like urban planning, environmental monitoring and disaster management. However, Multimodal Large Language Models exhibit critical deficiencies in fine-grained spatial understanding within Remote Sensing, primarily due to a reliance on limited or repurposed legacy datasets. To bridge this gap, we introduce a large-scale dataset grounded in verifiable cadastral vector data, comprising 3.8 million annotated objects across 510k high-resolution images with 135 granular semantic categories. We validate this resource through a comprehensive instruction-tuning benchmark spanning seven spatial reasoning tasks. Our evaluation establishes a robust baseline using a standard LLaVA architecture. We show that while current RS-specialized and commercial models (e.g., Gemini) struggle in zero-shot settings, high-fidelity supervision effectively bridges this gap, enabling standard architectures to master fine-grained spatial grounding without complex architectural modifications.

Method: Heterogeneous ensemble of nine models with three inference paradigms: (1) self-supervised DINOv2 Vision Transformer with slice-level sigmoid aggregation, (2) RadImageNet-pretrained DenseNet-121 with slice-level sigmoid averaging, and (3) seven Gated Attention Multiple Instance Learning models using EfficientNet-B3, ConvNeXt-Tiny, and EfficientNetV2-S backbones. Ensemble diversity enhanced through random-seed variation and Stochastic Weight Averaging. Overfitting addressed via Focal Loss, embedding-level Mixup, and domain-aware augmentation. Model fusion via score-weighted probability averaging with per-source threshold optimization.

Result: Final ensemble achieves average macro F1 of 0.9280 across four hospital centers, outperforming best single model (F1=0.8969) by +0.031. Validation-to-training loss ratio reduced from 35x to less than 3x through anti-overfitting techniques.

Conclusion: Heterogeneous architectures combined with source-aware calibration are essential for robust multi-site medical image classification, demonstrating that ensemble diversity and domain adaptation techniques effectively address domain shift challenges in multi-center medical imaging applications.

Abstract: The COVID-19 pandemic exposed critical limitations in diagnostic workflows: RT-PCR tests suffer from slow turnaround times and high false-negative rates, while CT-based screening offers faster complementary diagnosis but requires expert radiological interpretation. Deploying automated CT analysis across multiple hospital centres introduces further challenges, as differences in scanner hardware, acquisition protocols, and patient populations cause substantial domain shift that degrades single-model performance. To address these challenges, we present a heterogeneous ensemble of nine models spanning three inference paradigms: (1) a self-supervised DINOv2 Vision Transformer with slice-level sigmoid aggregation, (2) a RadImageNet-pretrained DenseNet-121 with slice-level sigmoid averaging, and (3) seven Gated Attention Multiple Instance Learning models using EfficientNet-B3, ConvNeXt-Tiny, and EfficientNetV2-S backbones with scan-level softmax classification. Ensemble diversity is further enhanced through random-seed variation and Stochastic Weight Averaging. We address severe overfitting, reducing the validation-to-training loss ratio from 35x to less than 3x, through a combination of Focal Loss, embedding-level Mixup, and domain-aware augmentation. Model outputs are fused via score-weighted probability averaging and calibrated with per-source threshold optimization. The final ensemble achieves an average macro F1 of 0.9280 across four hospital centres, outperforming the best single model (F1=0.8969) by +0.031, demonstrating that heterogeneous architectures combined with source-aware calibration are essential for robust multi-site medical image classification.

Method: Formulates synthetic fingerprint detection as continual few-shot adaptation problem. Uses combination of binary cross-entropy and supervised contrastive losses on feature representations. Implements replay of few samples from previously known styles during fine-tuning to mitigate catastrophic forgetting.

Result: Experiments with various DNN backbones and real/synthetic fingerprint datasets show the approach achieves good trade-off between fast adaptation for detecting unseen synthetic styles and retention of knowledge about known styles.

Conclusion: The continual few-shot adaptation framework effectively addresses generalization challenges in synthetic fingerprint detection, enabling rapid evolution of detectors to identify new types of synthetic data while maintaining performance on previously learned styles.

Abstract: The quality and realism of synthetically generated fingerprint images have increased significantly over the past decade fueled by advancements in generative artificial intelligence (GenAI). This has exacerbated the vulnerability of fingerprint recognition systems to data injection attacks, where synthetic fingerprints are maliciously inserted during enrollment or authentication. Hence, there is an urgent need for methods to detect if a fingerprint image is real or synthetic. While it is straightforward to train deep neural network (DNN) models to classify images as real or synthetic, often such DNN models overfit the training data and fail to generalize well when applied to synthetic fingerprints generated using unseen GenAI models. In this work, we formulate synthetic fingerprint detection as a continual few-shot adaptation problem, where the objective is to rapidly evolve a base detector to identify new types of synthetic data. To enable continual few-shot adaptation, we employ a combination of binary cross-entropy and supervised contrastive (applied to the feature representation) losses and replay a few samples from previously known styles during fine-tuning to mitigate catastrophic forgetting. Experiments based on several DNN backbones (as feature extractors) and a variety of real and synthetic fingerprint datasets indicate that the proposed approach achieves a good trade-off between fast adaptation for detecting unseen synthetic styles and forgetting of known styles.

Method: Population graph with subjects as nodes, functional/structural features as node attributes, phenotypic relationships via pairwise association encoder, Edge Variational GCNs for embeddings, asymmetric transformer cross-attention for multimodal fusion preserving functional dominance

Result: Achieved 87.3% AUC and 84.4% accuracy with 10-fold CV; 82.0% average cross-site accuracy with LOSO-CV, outperforming existing methods by 3-7%

Conclusion: Framework effectively integrates multimodal data from multi-site ABIDE-I dataset, improving automated ASD classification across imaging sites with functional dominance preservation

Abstract: Autism spectrum disorder (ASD) is a complex neurodevelopmental condition characterized by atypical functional brain connectivity and subtle structural alterations. rs-fMRI has been widely used to identify disruptions in large-scale brain networks, while structural MRI provides complementary information about morphological organization. Despite their complementary nature, effectively integrating these heterogeneous imaging modalities within a unified framework remains challenging. This study proposes a multimodal graph learning framework that preserves the dominant role of functional connectivity while integrating structural imaging and phenotypic information for ASD classification. The proposed framework is evaluated on ABIDE-I dataset. Each subject is represented as a node within a population graph. Functional and structural features are extracted as modality-specific node attributes, while inter-subject relationships are modeled using a pairwise association encoder (PAE) based on phenotypic information. Two Edge Variational GCNs are trained to learn subject-level embeddings. To enable effective multimodal integration, we introduce a novel asymmetric transformer-based cross-attention mechanism that allows functional embeddings to selectively incorporate complementary structural information while preserving functional dominance. The fused embeddings are then passed to a MLP for ASD classification. Using stratified 10-fold cross-validation, the framework achieved an AUC of 87.3% and an accuracy of 84.4%. Under leave-one-site-out cross-validation (LOSO-CV), the model achieved an average cross-site accuracy of 82.0%, outperforming existing methods by approximately 3% under 10-fold cross-validation and 7% under LOSO-CV. The proposed framework effectively integrates heterogeneous multimodal data from the multi-site ABIDE-I dataset, improving automated ASD classification across imaging sites.

Method: Proposes Spectrum Matching Hypothesis with two components: Encoding Spectrum Matching (ESM) matches PSD between images and latents, and Decoding Spectrum Matching (DSM) preserves frequency-to-frequency semantic correspondence via shared spectral masking with frequency-aligned reconstruction.

Result: Superior diffusion generation on CelebA and ImageNet datasets, outperforms prior approaches like VA-VAE and EQ-VAE. Also extends spectral view to representation alignment (REPA) with DoG-based method.

Conclusion: Spectrum Matching provides unified view explaining prior observations of VAE latent issues, improves diffusability, and offers insights for representation alignment.

Abstract: In this paper, we study the diffusability (learnability) of variational autoencoders (VAE) in latent diffusion. First, we show that pixel-space diffusion trained with an MSE objective is inherently biased toward learning low and mid spatial frequencies, and that the power-law power spectral density (PSD) of natural images makes this bias perceptually beneficial. Motivated by this result, we propose the \emph{Spectrum Matching Hypothesis}: latents with superior diffusability should (i) follow a flattened power-law PSD (\emph{Encoding Spectrum Matching}, ESM) and (ii) preserve frequency-to-frequency semantic correspondence through the decoder (\emph{Decoding Spectrum Matching}, DSM). In practice, we apply ESM by matching the PSD between images and latents, and DSM via shared spectral masking with frequency-aligned reconstruction. Importantly, Spectrum Matching provides a unified view that clarifies prior observations of over-noisy or over-smoothed latents, and interprets several recent methods as special cases (e.g., VA-VAE, EQ-VAE). Experiments suggest that Spectrum Matching yields superior diffusion generation on CelebA and ImageNet datasets, and outperforms prior approaches. Finally, we extend the spectral view to representation alignment (REPA): we show that the directional spectral energy of the target representation is crucial for REPA, and propose a DoG-based method to further improve the performance of REPA. Our code is available https://github.com/forever208/SpectrumMatching.

Method: Proposes EviATTA framework with Dirichlet-based Evidential Modeling to decompose uncertainty into distribution and data uncertainty. Uses Hierarchical Evidential Sampling to select informative samples and guide sparse annotations. Introduces Dual Consistency Regularization with progressive prompt consistency and variational feature consistency.

Result: Extensive experiments on six medical image segmentation datasets show EviATTA consistently improves adaptation reliability with minimal expert feedback in both batch-wise and instance-wise TTA settings.

Conclusion: EviATTA effectively addresses reliability issues in medical SAM adaptation under distribution shifts by combining evidential uncertainty modeling with efficient sparse annotation utilization.

Abstract: Deploying foundational medical Segment Anything Models (SAMs) via test-time adaptation (TTA) is challenging under large distribution shifts, where test-time supervision is often unreliable. While active test-time adaptation (ATTA) introduces limited expert feedback to improve reliability, existing ATTA methods still suffer from unreliable uncertainty estimation and inefficient utilization of sparse annotations. To address these issues, we propose Evidential Active Test-Time Adaptation (EviATTA), which is, to our knowledge, the first ATTA framework tailored for medical SAMs. Specifically, we adopt the Dirichlet-based Evidential Modeling to decompose overall predictive uncertainty into distribution uncertainty and data uncertainty. Building on this decomposition, we design a Hierarchical Evidential Sampling strategy, where image-wise distribution uncertainty is used to select informative shifted samples, while distance-aware data uncertainty guides sparse pixel annotations to resolve data ambiguities. We further introduce Dual Consistency Regularization, which enforces progressive prompt consistency on sparsely labeled samples to better exploit sparse supervision and applies variational feature consistency on unlabeled samples to stabilize adaptation. Extensive experiments on six medical image segmentation datasets demonstrate that EviATTA consistently improves adaptation reliability with minimal expert feedback under both batch-wise and instance-wise test-time adaptation settings.

Method: E2EGS extracts edges from noisy event streams by exploiting spatio-temporal characteristics: edges produce consistent events during camera movement while non-edge regions produce sparse noise. Uses patch-based temporal coherence analysis to measure local variance for edge extraction and noise suppression. Extracted edges guide structure-aware Gaussian initialization and enable edge-weighted losses throughout initialization, tracking, and bundle adjustment.

Result: Extensive experiments on synthetic and real datasets demonstrate superior reconstruction quality and trajectory accuracy compared to existing methods, establishing a fully pose-free paradigm for event-based 3D reconstruction.

Conclusion: E2EGS successfully addresses limitations of existing methods by operating solely on event streams without requiring known poses, enabling robust novel view synthesis under challenging real-world conditions through effective edge extraction and utilization.

Abstract: The emergence of neural radiance fields (NeRF) and 3D Gaussian splatting (3DGS) has advanced novel view synthesis (NVS). These methods, however, require high-quality RGB inputs and accurate corresponding poses, limiting robustness under real-world conditions such as fast camera motion or adverse lighting. Event cameras, which capture brightness changes at each pixel with high temporal resolution and wide dynamic range, enable precise sensing of dynamic scenes and offer a promising solution. However, existing event-based NVS methods either assume known poses or rely on depth estimation models that are bounded by their initial observations, failing to generalize as the camera traverses previously unseen regions. We present E2EGS, a pose-free framework operating solely on event streams. Our key insight is that edge information provides rich structural cues essential for accurate trajectory estimation and high-quality NVS. To extract edges from noisy event streams, we exploit the distinct spatio-temporal characteristics of edges and non-edge regions. The event camera’s movement induces consistent events along edges, while non-edge regions produce sparse noise. We leverage this through a patch-based temporal coherence analysis that measures local variance to extract edges while robustly suppressing noise. The extracted edges guide structure-aware Gaussian initialization and enable edge-weighted losses throughout initialization, tracking, and bundle adjustment. Extensive experiments on both synthetic and real datasets demonstrate that E2EGS achieves superior reconstruction quality and trajectory accuracy, establishing a fully pose-free paradigm for event-based 3D reconstruction.

Method: Created TOXICTAGS dataset with 6,300 real-world memes annotated in two stages: binary toxic/normal classification and fine-grained labeling (hateful, dangerous, offensive). Proposed STEMTOX framework - an entropy-guided multi-tasking approach that integrates generation of socially grounded tags with classification.

Result: Incorporating socially relevant tags substantially enhances performance of state-of-the-art VLMs in toxicity detection tasks. The dataset and framework provide improved content moderation capabilities.

Conclusion: TOXICTAGS dataset and STEMTOX framework offer a novel and scalable foundation for improved content moderation in multimodal online environments, addressing the challenge of meme-based toxicity detection.

Abstract: Memes, as a widely used mode of online communication, often serve as vehicles for spreading harmful content. However, limitations in data accessibility and the high costs of dataset curation hinder the development of robust meme moderation systems. To address this challenge, in this work, we introduce a first-of-its-kind dataset - TOXICTAGS consisting of 6,300 real-world meme-based posts annotated in two stages: (i) binary classification into toxic and normal, and (ii) fine-grained labelling of toxic memes as hateful, dangerous, or offensive. A key feature of this dataset is that it is enriched with auxiliary metadata of socially relevant tags, enhancing the context of each meme. In addition, we propose a novel entropy guided multi-tasking framework - STEMTOX - that integrates the generation of socially grounded tags with a robust classification framework. Experimental results show that incorporating these tags substantially enhances the performance of state-of-the-art VLMs in toxicity detection tasks. Our contributions offer a novel and scalable foundation for improved content moderation in multimodal online environments. Warning: Contains potentially toxic contents.

Method: Introduces AURORA-KITTI benchmark with 82K weather-consistent RGB-LiDAR pairs and DDCD baseline that formulates Depth Completion and Denoising as a unified task, using distillation to leverage depth foundation models for injecting clean structural priors into training.

Result: DDCD achieves state-of-the-art performance on AURORA-KITTI and real-world DENSE dataset while maintaining efficiency, demonstrating that weather-aware, physically consistent data contributes more to robustness than architectural modifications alone.

Conclusion: The work establishes the first comprehensive benchmark for robust depth completion under adverse weather and shows that leveraging depth foundation models with weather-consistent data significantly improves robustness in real-world 3D scene understanding.

Abstract: Robust depth completion is fundamental to real-world 3D scene understanding, yet existing RGB-LiDAR fusion methods degrade significantly under adverse weather, where both camera images and LiDAR measurements suffer from weather-induced corruption. In this paper, we introduce AURORA-KITTI, the first large-scale multi-modal, multi-weather benchmark for robust depth completion in the wild. We further formulate Depth Completion and Denoising (DCD) as a unified task that jointly reconstructs a dense depth map from corrupted sparse inputs while suppressing weather-induced noise. AURORA-KITTI contains over \textit{82K} weather-consistent RGBL pairs with metric depth ground truth, spanning diverse weather types, three severity levels, day and night scenes, paired clean references, lens occlusion conditions, and textual descriptions. Moreover, we introduce DDCD, an efficient distillation-based baseline that leverages depth foundation models to inject clean structural priors into in-the-wild DCD training. DDCD achieves state-of-the-art performance on AURORA-KITTI and the real-world DENSE dataset while maintaining efficiency. Notably, our results further show that weather-aware, physically consistent data contributes more to robustness than architectural modifications alone. Data and code will be released upon publication.

Method: Proposes Fractal Visual Autoregressive Diffusion framework: 1) VCFR module fuses multi-scale image features with current depth predictions for cross-modal conditioning, 2) conditional denoising diffusion loss models depth distributions in continuous space, 3) fractal recursive architecture reuses base visual AR unit in self-similar hierarchy for efficiency, 4) uncertainty-aware robust consensus aggregation for multi-sample inference.

Result: Experiments on standard benchmarks demonstrate strong performance and validate effectiveness of the proposed design. The framework shows improved depth estimation accuracy and computational efficiency.

Conclusion: The proposed framework successfully addresses challenges in autoregressive depth estimation by combining cross-modal fusion, continuous distribution modeling, fractal architecture for efficiency, and uncertainty-aware inference, leading to improved monocular depth estimation performance.

Abstract: Monocular depth estimation can benefit from autoregressive (AR) generation, but direct AR modeling is hindered by the modality gap between RGB and depth, inefficient pixel-wise generation, and instability in continuous depth prediction. We propose a Fractal Visual Autoregressive Diffusion framework that reformulates depth estimation as a coarse-to-fine, next-scale autoregressive generation process. A VCFR module fuses multi-scale image features with current depth predictions to improve cross-modal conditioning, while a conditional denoising diffusion loss models depth distributions directly in continuous space and mitigates errors caused by discrete quantization. To improve computational efficiency, we organize the scale-wise generators into a fractal recursive architecture, reusing a base visual AR unit in a self-similar hierarchy. We further introduce an uncertainty-aware robust consensus aggregation scheme for multi-sample inference to improve fusion stability and provide a practical pixel-wise reliability estimate. Experiments on standard benchmarks demonstrate strong performance and validate the effectiveness of the proposed design.

Method: Proposed frame-accurate benchmark for small VLMs on video question-answering, evaluated under controlled frame-sampling strategies with open-sourced benchmarking code.

Result: Confirmed suspected frame-sampling bias and revealed both data-specific and task-specific behaviors of SVLMs under different frame-sampling techniques.

Conclusion: Need for standardized frame-sampling strategies tailored to each benchmarking dataset and reproducible, unbiased protocols for evaluating video VLMs.

Abstract: Comparing vision language models on videos is particularly complex, as the performances is jointly determined by the model’s visual representation capacity and the frame-sampling strategy used to construct the input. Current video benchmarks are suspected to suffer from substantial frame-sampling bias, as models are evaluated with different frame selection strategies. In this work, we propose the first frame-accurate benchmark of state-of-the-art small VLMs for video question-answering, evaluated under controlled frame-sampling strategies. Our results confirm the suspected bias and highlight both data-specific and task-specific behaviors of SVLMs under different frame-sampling techniques. By open-sourcing our benchmarking code, we provide the community with a reproducible and unbiased protocol for evaluating video VLMs and emphasize the need for standardized frame-sampling strategies tailored to each benchmarking dataset in future research.

Method: Proposes Hand4Whole++ with CHAM (Conditional Hands Modulator) module that modulates whole-body features using hand-specific features from a pre-trained hand estimator. Also incorporates finger articulations and hand shapes from hand estimator aligned to full-body mesh via differentiable rigid alignment.

Result: Extensive experiments show Hand4Whole++ substantially improves hand accuracy and enhances overall full-body pose quality compared to existing methods.

Conclusion: The modular framework successfully bridges the supervision gap by combining globally consistent body reasoning with fine-grained hand detail without retraining the full-body model.

Abstract: Accurately recovering hand poses within the body context remains a major challenge in 3D whole-body pose estimation. This difficulty arises from a fundamental supervision gap: whole-body pose estimators are trained on full-body datasets with limited hand diversity, while hand-only estimators, trained on hand-centric datasets, excel at detailed finger articulation but lack global body awareness. To address this, we propose Hand4Whole++, a modular framework that leverages the strengths of both pre-trained whole-body and hand pose estimators. We introduce CHAM (Conditional Hands Modulator), a lightweight module that modulates the whole-body feature stream using hand-specific features extracted from a pre-trained hand pose estimator. This modulation enables the whole-body model to predict wrist orientations that are both accurate and coherent with the upper-body kinematic structure, without retraining the full-body model. In parallel, we directly incorporate finger articulations and hand shapes predicted by the hand pose estimator, aligning them to the full-body mesh via differentiable rigid alignment. This design allows Hand4Whole++ to combine globally consistent body reasoning with fine-grained hand detail. Extensive experiments demonstrate that Hand4Whole++ substantially improves hand accuracy and enhances overall full-body pose quality.

Method: Proposes SyDES with: 1) Mixed-scale image strategy combining global context from low-res images with fine-grained local features via masked image modeling on high-res sub-images, 2) Symmetrical cross-modal decoders (text-guided image decoder and image-guided text decoder) for mutual reconstruction, 3) Dedicated loss functions to harmonize MIM and modal alignment objectives.

Result: Achieves state-of-the-art on MSED benchmark with 1.1% F1-score improvement in desire understanding. Shows consistent gains in emotion and sentiment recognition, validating generalization ability and necessity of using non-verbal cues.

Conclusion: The symmetrical bidirectional framework effectively captures intention-related visual representations and enables deep cross-modal interaction, demonstrating the importance of leveraging both verbal and non-verbal cues for multimodal desire understanding.

Abstract: Multimodal desire understanding, a task closely related to both emotion and sentiment that aims to infer human intentions from visual and textual cues, is an emerging yet underexplored task in affective computing with applications in social media analysis. Existing methods for related tasks predominantly focus on mining verbal cues, often overlooking the effective utilization of non-verbal cues embedded in images. To bridge this gap, we propose a Symmetrical Bidirectional Multimodal Learning Framework for Desire, Emotion, and Sentiment Recognition (SyDES). The core of SyDES is to achieve bidirectional fine-grained modal alignment between text and image modalities. Specifically, we introduce a mixed-scaled image strategy that combines global context from low-resolution images with fine-grained local features via masked image modeling (MIM) on high-resolution sub-images, effectively capturing intention-related visual representations. Then, we devise symmetrical cross-modal decoders, including a text-guided image decoder and an image-guided text decoder, which enable mutual reconstruction and refinement between modalities, facilitating deep cross-modal interaction. Furthermore, a set of dedicated loss functions is designed to harmonize potential conflicts between the MIM and modal alignment objectives during optimization. Extensive evaluations on the MSED benchmark demonstrate the superiority of our approach, which establishes a new state-of-the-art performance with 1.1% F1-score improvement in desire understanding. Consistent gains in emotion and sentiment recognition further validate its generalization ability and the necessity of utilizing non-verbal cues. Our code is available at: https://github.com/especiallyW/SyDES.

Method: Developed a deep learning system using five contemporary architectures (EfficientNet-V2-S with SSL, Vision Transformer, Swin Transformer, ConvNeXt-Base, ResNet-50) on 2,640 clinically annotated anterior segment images. Implemented tailored preprocessing with specular reflection mitigation and CLAHE to enhance vascular/textural patterns. Used self-supervised learning (SimCLR) on domain-specific ocular images to improve performance.

Result: EfficientNet-V2-S with SSL achieved optimal performance: 98.21% F1-score, 97.90% precision, 98.55% recall. This was a substantial improvement over ImageNet-only initialization (94.63% F1). The model achieved near-perfect precision (100%) for Normal classification, minimizing unnecessary clinical referrals.

Conclusion: The study demonstrates that anterior segment ocular imaging combined with deep learning and self-supervised learning can provide accurate, accessible diabetic retinopathy screening, potentially expanding access to screening in resource-limited settings.

Abstract: Diabetic retinopathy screening traditionally relies on fundus photography, requiring specialized equipment and expertise often unavailable in primary care and resource limited settings. We developed and validated a deep learning (DL) system for automated diabetic classification using anterior segment ocular imaging a readily accessible alternative utilizing standard photography equipment. The system leverages visible biomarkers in the iris, sclera, and conjunctiva that correlate with systemic diabetic status. We systematically evaluated five contemporary architectures (EfficientNet-V2-S with self-supervised learning (SSL), Vision Transformer, Swin Transformer, ConvNeXt-Base, and ResNet-50) on 2,640 clinically annotated anterior segment images spanning Normal, Controlled Diabetic, and Uncontrolled Diabetic categories. A tailored preprocessing pipeline combining specular reflection mitigation and contrast limited adaptive histogram equalization (CLAHE) was implemented to enhance subtle vascular and textural patterns critical for classification. SSL using SimCLR on domain specific ocular images substantially improved model performance.EfficientNet-V2-S with SSL achieved optimal performance with an F1-score of 98.21%, precision of 97.90%, and recall of 98.55% a substantial improvement over ImageNet only initialization (94.63% F1). Notably, the model attained near perfect precision (100%) for Normal classification, critical for minimizing unnecessary clinical referrals.

Method: OraPO enables single-stage RL-only training by converting failed GRPO explorations into direct preference supervision via a lightweight oracle step. FactScore-based reward grounds learning in diagnostic evidence by extracting atomic clinical facts and checking entailment against ground-truth labels.

Result: Achieves new SOTA performance on CheXpert Plus dataset (0.341 F1) with 2-3 orders of magnitude less training data using a small base VLM on modest hardware.

Conclusion: The proposed framework creates a compact and powerful solution that significantly improves learning efficiency on clinically challenging cases for radiology report generation.

Abstract: Radiology report generation (RRG) aims to automatically produce clinically faithful reports from chest X-ray images. Prevailing work typically follows a scale-driven paradigm, by multi-stage training over large paired corpora and oversized backbones, making pipelines highly data- and compute-intensive. In this paper, we propose Oracle-educated GRPO (OraPO) with a FactScore-based reward (FactS) to tackle the RRG task under constrained budgets. OraPO enables single-stage, RL-only training by converting failed GRPO explorations on rare or difficult studies into direct preference supervision via a lightweight oracle step. FactS grounds learning in diagnostic evidence by extracting atomic clinical facts and checking entailment against ground-truth labels, yielding dense, interpretable sentence-level rewards. Together, OraPO and FactS create a compact and powerful framework that significantly improves learning efficiency on clinically challenging cases, setting the new SOTA performance on the CheXpert Plus dataset (0.341 in F1) with 2–3 orders of magnitude less training data using a small base VLM on modest hardware.

Method: The paper introduces MVX-Bench, a Multi-Video Cross-Dimension Benchmark that reformulates 11 classical computer vision tasks into a unified multi-video QA framework with 1,442 questions over 4,255 videos. It also proposes SAMA (Skill-Augmented Agentic Framework for Multi-Video Understanding), which integrates visual tools, task-specific skills, and a conflict-aware verification mechanism for iterative structured reasoning.

Result: SAMA outperforms strong open-source baselines and GPT on MVX-Bench. Ablation studies validate the effectiveness of the skill design and conflict resolution mechanisms in the framework.

Conclusion: The work addresses important gaps in multi-video understanding by introducing a comprehensive benchmark and an effective agentic framework that enables structured cross-video reasoning through skill integration and conflict-aware verification.

Abstract: Multimodal Large Language Models have achieved strong performance in single-video understanding, yet their ability to reason across multiple videos remains limited. Existing approaches typically concatenate multiple videos into a single input and perform direct inference, which introduces training-inference mismatch, information loss from frame compression, and a lack of explicit cross-video coordination. Meanwhile, current multi-video benchmarks primarily emphasize event-level comparison, leaving identity-level matching, fine-grained discrimination, and structured multi-step reasoning underexplored. To address these gaps, we introduce MVX-Bench, a Multi-Video Cross-Dimension Benchmark that reformulates 11 classical computer vision tasks into a unified multi-video question-answering framework, comprising 1,442 questions over 4,255 videos from diverse real-world datasets. We further propose SAMA, a Skill-Augmented Agentic Framework for Multi-Video Understanding, which integrates visual tools, task-specific skills, and a conflict-aware verification mechanism to enable iterative and structured reasoning. Experimental results show that SAMA outperforms strong open-source baselines and GPT on MVX-Bench, and ablations validate the effectiveness of skill design and conflict resolution.

Method: Models event streams as continuous-time Dirac impulse trains, enabling artifact-free compression via direct transform evaluation at event timestamps. Uses density-driven adaptive selection among DCT, DTFT, and DWT transforms with transform-specific coefficient pruning strategies tailored to each domain’s sparsity characteristics.

Result: Achieves superior reconstruction fidelity with DTFT delivering lowest earth mover distance. In downstream segmentation with EventSAM, achieves mean IoU 0.87 on MVSEC vs 0.44 for voxel grids at 24 channels. ROS2 implementation provides real-time deployment with DCT processing at 1.5 ms latency and 2.7X higher throughput than alternatives.

Conclusion: EECVS establishes the first adaptive event compression framework that maintains both computational efficiency and superior generalization across diverse robotic scenarios, eliminating temporal binning artifacts while automatically adapting compression strategies based on real-time event density analysis.

Abstract: Event cameras promise low latency and high dynamic range, yet their sparse output challenges integration into standard robotic pipelines. We introduce \nameframew (Efficient Event Camera Volume System), a novel framework that models event streams as continuous-time Dirac impulse trains, enabling artifact-free compression through direct transform evaluation at event timestamps. Our key innovation combines density-driven adaptive selection among DCT, DTFT, and DWT transforms with transform-specific coefficient pruning strategies tailored to each domain’s sparsity characteristics. The framework eliminates temporal binning artifacts while automatically adapting compression strategies based on real-time event density analysis. On EHPT-XC and MVSEC datasets, our framework achieves superior reconstruction fidelity with DTFT delivering the lowest earth mover distance. In downstream segmentation tasks, EECVS demonstrates robust generalization. Notably, our approach demonstrates exceptional cross-dataset generalization: when evaluated with EventSAM segmentation, EECVS achieves mean IoU 0.87 on MVSEC versus 0.44 for voxel grids at 24 channels, while remaining competitive on EHPT-XC. Our ROS2 implementation provides real-time deployment with DCT processing achieving 1.5 ms latency and 2.7X higher throughput than alternative transforms, establishing the first adaptive event compression framework that maintains both computational efficiency and superior generalization across diverse robotic scenarios.

Method: Proposes an agentic framework that grounds an LLM in an explicit 3D scene graph constructed by a dedicated perception module. The agent interacts with scenes through structured geometric tools that expose fundamental properties like object dimensions, distances, poses, and spatial relationships.

Result: Achieves up to 16% improvement over previous works on VSI-Bench static split without task-specific fine-tuning. Compared to base VLMs, achieves 33% to 50% average improvements. Provides an upper bound on spatial reasoning performance under ideal perceptual conditions.

Conclusion: Explicit geometric grounding substantially improves spatial reasoning performance, and structured representations offer a compelling alternative to purely end-to-end visual reasoning for indoor scene understanding.

Abstract: Visual Language Models (VLMs) have increasingly become the main paradigm for understanding indoor scenes, but they still struggle with metric and spatial reasoning. Current approaches rely on end-to-end video understanding or large-scale spatial question answering fine-tuning, inherently coupling perception and reasoning. In this paper, we investigate whether decoupling perception and reasoning leads to improved spatial reasoning. We propose an agentic framework for static 3D indoor scene reasoning that grounds an LLM in an explicit 3D scene graph (3DSG). Rather than ingesting videos directly, each scene is represented as a persistent 3DSG constructed by a dedicated perception module. To isolate reasoning performance, we instantiate the 3DSG from ground-truth annotations. The agent interacts with the scene exclusively through structured geometric tools that expose fundamental properties such as object dimensions, distances, poses, and spatial relationships. The results we obtain on the static split of VSI-Bench provide an upper bound under ideal perceptual conditions on the spatial reasoning performance, and we find that it is significantly higher than previous works, by up to 16%, without task specific fine-tuning. Compared to base VLMs, our agentic variant achieves significantly better performance, with average improvements between 33% to 50%. These findings indicate that explicit geometric grounding substantially improves spatial reasoning performance, and suggest that structured representations offer a compelling alternative to purely end-to-end visual reasoning.

Method: Uses ControlNet modules to encode user point-based prompts (joints, bounding boxes) and inject them into a pre-trained diffusion model. Fine-tunes only cross-attention blocks for prompt alignment. Includes inpainting-based refinement module starting from slightly noised coarse completion to preserve visible regions and ensure seamless blending.

Result: Extensive experiments on HAC and PGPIS benchmarks show more physically plausible and higher-quality completions with significantly improved prompt alignment compared to existing methods.

Conclusion: PHAC enables controllable human amodal completion with user prompts while maintaining visible appearance preservation, advancing human-centric conditional image generation.

Abstract: Conditional image generation methods are increasingly used in human-centric applications, yet existing human amodal completion (HAC) models offer users limited control over the completed content. Given an occluded person image, they hallucinate invisible regions while preserving visible ones, but cannot reliably incorporate user-specified constraints such as a desired pose or spatial extent. As a result, users often resort to repeatedly sampling the model until they obtain a satisfactory output. Pose-guided person image synthesis (PGPIS) methods allow explicit pose conditioning, but frequently fail to preserve the instance-specific visible appearance and tend to be biased toward the training distribution, even when built on strong diffusion model priors. To address these limitations, we introduce promptable human amodal completion (PHAC), a new task that completes occluded human images while satisfying both visible appearance constraints and multiple user prompts. Users provide simple point-based prompts, such as additional joints for the target pose or bounding boxes for desired regions; these prompts are encoded using ControlNet modules specialized for each prompt type. These modules inject the prompt signals into a pre-trained diffusion model, and we fine-tune only the cross-attention blocks to obtain strong prompt alignment without degrading the underlying generative prior. To further preserve visible content, we propose an inpainting-based refinement module that starts from a slightly noised coarse completion, faithfully preserves the visible regions, and ensures seamless blending at occlusion boundaries. Extensive experiments on the HAC and PGPIS benchmarks show that our approach yields more physically plausible and higher-quality completions, while significantly improving prompt alignment compared with existing amodal completion and pose-guided synthesis methods.

Method: Proposes FSENet with three key components: 1) Face-guided Sentiment Discovery (FSD) module that integrates facial features via dual-branch modeling for sentiment stimuli clues, 2) Point-aware Sentiment Semantics Contrast (PSSC) strategy using contrastive learning to discriminate sentiment semantics near annotation points, and 3) Boundary-aware Sentiment Pseudo-label Generation (BSPG) to convert sparse point annotations into smooth supervisory pseudo-labels.

Result: Achieves state-of-the-art performance on benchmarks under full supervision, video-level, and point-level weak supervision settings. Extensive experiments and visualizations demonstrate the framework’s effectiveness and strong generalization ability across different annotation settings.

Conclusion: FSENet effectively addresses boundary imprecision in P-WTSL by leveraging facial features to guide sentiment localization, with the proposed components working synergistically to enhance sentiment boundary recognition and achieve superior performance across various supervision levels.

Abstract: Point-level weakly-supervised temporal sentiment localization (P-WTSL) aims to detect sentiment-relevant segments in untrimmed multimodal videos using timestamp sentiment annotations, which greatly reduces the costly frame-level labeling. To further tackle the challenges of imprecise sentiment boundaries in P-WTSL, we propose the Face-guided Sentiment Boundary Enhancement Network (\textbf{FSENet}), a unified framework that leverages fine-grained facial features to guide sentiment localization. Specifically, our approach \textit{first} introduces the Face-guided Sentiment Discovery (FSD) module, which integrates facial features into multimodal interaction via dual-branch modeling for effective sentiment stimuli clues; We \textit{then} propose the Point-aware Sentiment Semantics Contrast (PSSC) strategy to discriminate sentiment semantics of candidate points (frame-level) near annotation points via contrastive learning, thereby enhancing the model’s ability to recognize sentiment boundaries. At \textit{last}, we design the Boundary-aware Sentiment Pseudo-label Generation (BSPG) approach to convert sparse point annotations into temporally smooth supervisory pseudo-labels. Extensive experiments and visualizations on the benchmark demonstrate the effectiveness of our framework, achieving state-of-the-art performance under full supervision, video-level, and point-level weak supervision, thereby showcasing the strong generalization ability of our FSENet across different annotation settings.

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to draw conclusions due to fetch failure

Abstract: Failed to fetch summary for 2603.15396: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.15396&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes Self-expressive Sequence Regularization (SSR), a plug-and-play, training-free operator that enforces Grassmannian sequence regularity during inference. Given a window of historical states, SSR computes an analytical affinity matrix via the self-expressive property and uses it to regularize the current update, pulling noisy predictions back toward the manifold-consistent trajectory with minimal overhead.

Result: Experiments on long-sequence benchmarks demonstrate that SSR consistently reduces drift and improves reconstruction quality across multiple streaming 3D reconstruction tasks.

Conclusion: SSR provides an effective, low-overhead solution to geometric drift in streaming 3D reconstruction by leveraging Grassmannian manifold perspectives and self-expressive properties to maintain temporal coherence.

Abstract: Streaming 3D reconstruction demands long-horizon state updates under strict latency constraints, yet stateful recurrent models often suffer from geometric drift as errors accumulate over time. We revisit this problem from a Grassmannian manifold perspective: the latent persistent state can be viewed as a subspace representation, i.e., a point evolving on a Grassmannian manifold, where temporal coherence implies the state trajectory should remain on (or near) this manifold.Based on this view, we propose Self-expressive Sequence Regularization (SSR), a plug-and-play, training-free operator that enforces Grassmannian sequence regularity during inference.Given a window of historical states, SSR computes an analytical affinity matrix via the self-expressive property and uses it to regularize the current update, effectively pulling noisy predictions back toward the manifold-consistent trajectory with minimal overhead. Experiments on long-sequence benchmarks demonstrate that SSR consistently reduces drift and improves reconstruction quality across multiple streaming 3D reconstruction tasks.

Method: Uses a diffusion-transformer finetuning framework with three key components: (1) RoPE-aligned location canvas with location-aligned token pruning for spatial grounding, (2) AdaLN-style identity-adaptive modulation from face-recognition embeddings for persistent identity injection, and (3) identity-isolated attention to prevent cross-identity interference. Training combines conditional flow matching with embedding-space face similarity loss, plus reference-face replacement and location-canvas degradations to discourage shortcuts.

Result: On MultiID-Bench, AnyPhoto improves identity similarity while reducing copy-paste tendency, with gains increasing as the number of identities grows. Also supports prompt-driven stylization with accurate placement.

Conclusion: AnyPhoto demonstrates effective multi-person identity-preserving generation with improved identity similarity and reduced copy-paste artifacts, showing great potential application value for controllable image generation.

Abstract: Multi-person identity-preserving generation requires binding multiple reference faces to specified locations under a text prompt. Strong identity/layout conditions often trigger copy-paste shortcuts and weaken prompt-driven controllability. We present AnyPhoto, a diffusion-transformer finetuning framework with (i) a RoPE-aligned location canvas plus location-aligned token pruning for spatial grounding, (ii) AdaLN-style identity-adaptive modulation from face-recognition embeddings for persistent identity injection, and (iii) identity-isolated attention to prevent cross-identity interference. Training combines conditional flow matching with an embedding-space face similarity loss, together with reference-face replacement and location-canvas degradations to discourage shortcuts. On MultiID-Bench, AnyPhoto improves identity similarity while reducing copy-paste tendency, with gains increasing as the number of identities grows. AnyPhoto also supports prompt-driven stylization with accurate placement, showing great potential application value.

Method: Transformer-based feed-forward architecture predicts dynamic 3D Gaussian deformations without subject-specific optimization. Uses static-to-dynamic knowledge transfer: Transformer pretrained on large-scale static captures provides geometric/appearance priors, adapted via LoRA fine-tuning on dynamic captures. Proposes DynaFlow loss (optical flow-guided objective) for cloth dynamics in rendered space. Reannotates noisy SMPL-X fittings in dynamic capture datasets.

Result: Produces visually rich and generalizable animations, outperforming prior methods in reconstructing animatable 3D human avatars with realistic cloth dynamics from single images.

Conclusion: DynaAvatar successfully addresses limitations of rigid transformations by enabling realistic cloth dynamics in 3D human avatars from single images through innovative knowledge transfer and training strategies.

Abstract: Existing single-image 3D human avatar methods primarily rely on rigid joint transformations, limiting their ability to model realistic cloth dynamics. We present DynaAvatar, a zero-shot framework that reconstructs animatable 3D human avatars with motion-dependent cloth dynamics from a single image. Trained on large-scale multi-person motion datasets, DynaAvatar employs a Transformer-based feed-forward architecture that directly predicts dynamic 3D Gaussian deformations without subject-specific optimization. To overcome the scarcity of dynamic captures, we introduce a static-to-dynamic knowledge transfer strategy: a Transformer pretrained on large-scale static captures provides strong geometric and appearance priors, which are efficiently adapted to motion-dependent deformations through lightweight LoRA fine-tuning on dynamic captures. We further propose the DynaFlow loss, an optical flow-guided objective that provides reliable motion-direction geometric cues for cloth dynamics in rendered space. Finally, we reannotate the missing or noisy SMPL-X fittings in existing dynamic capture datasets, as most public dynamic capture datasets contain incomplete or unreliable fittings that are unsuitable for training high-quality 3D avatar reconstruction models. Experiments demonstrate that DynaAvatar produces visually rich and generalizable animations, outperforming prior methods.

Method: Unable to determine method due to retrieval failure

Result: Unable to determine results due to retrieval failure

Conclusion: Unable to determine conclusion due to retrieval failure

Abstract: Failed to fetch summary for 2603.15404: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.15404&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes a dual approach: 1) refines latent code of pretrained 3D-Aware GAN for texture editing, 2) optimizes expression code of 3DMM model for mesh editing. Uses Dual Mappers module (Texture Mapper and Emotion Mapper) to learn transformations. Employs Text-Guided Optimization with CLIP-based objective and expression text prompts, enhanced by SubSpace Projection mechanism for precise fine-grained control.

Result: Extensive experiments and comparative analyses demonstrate the method’s effectiveness and superiority in achieving precise control over fine-grained facial expressions while maintaining high-quality, view-consistent 3D renderings.

Conclusion: The proposed approach successfully addresses limitations of existing methods by combining 3D-Aware GAN refinement with 3DMM expression optimization, enabling precise fine-grained facial expression editing through text-guided control mechanisms.

Abstract: Facial expression editing methods can be mainly categorized into two types based on their architectures: 2D-based and 3D-based methods. The former lacks 3D face modeling capabilities, making it difficult to edit 3D factors effectively. The latter has demonstrated superior performance in generating high-quality and view-consistent renderings using single-view 2D face images. Although these methods have successfully used animatable models to control facial expressions, they still have limitations in achieving precise control over fine-grained expressions. To address this issue, in this paper, we propose a novel approach by simultaneously refining both the latent code of a pretrained 3D-Aware GAN model for texture editing and the expression code of the driven 3DMM model for mesh editing. Specifically, we introduce a Dual Mappers module, comprising Texture Mapper and Emotion Mapper, to learn the transformations of the given latent code for textures and the expression code for meshes, respectively. To optimize the Dual Mappers, we propose a Text-Guided Optimization method, leveraging a CLIP-based objective function with expression text prompts as targets, while integrating a SubSpace Projection mechanism to project the text embedding to the expression subspace such that we can have more precise control over fine-grained expressions. Extensive experiments and comparative analyses demonstrate the effectiveness and superiority of our proposed method.

Method: The framework uses a multi-agent system with four cooperative modules: 1) Script Development agents that draft and refine screenplays iteratively, 2) Virtual Scene Design that transforms text into semantically aligned 3D environments, 3) Character Behaviour Control for determining character blocking and motion, and 4) Camera Planning that optimizes framing, movement, and composition. The system is built on a game engine with real-time visual editing capabilities for interactive inspection and synchronized timeline adjustments.

Result: The system generates high-quality, semantically grounded previz sequences in approximately 25 minutes per creative idea. Extensive experiments and human evaluations demonstrate the effectiveness of agent collaboration for both automated prototyping and human-in-the-loop filmmaking applications.

Conclusion: Mind-of-Director successfully demonstrates that multi-agent collaboration can effectively automate film previsualization, producing professional-quality results efficiently. The framework shows promise for both automated prototyping and collaborative human-AI filmmaking workflows.

Abstract: We present Mind-of-Director, a multi-modal agent-driven framework for film previz that models the collaborative decision-making process of a film production team. Given a creative idea, Mind-of-Director orchestrates multiple specialized agents to produce previz sequences within the game engine. The framework consists of four cooperative modules: Script Development, where agents draft and refine the screenplay iteratively; Virtual Scene Design, which transforms text into semantically aligned 3D environments; Character Behaviour Control, which determines character blocking and motion; and Camera Planning, which optimizes framing, movement, and composition for cinematic camera effects. A real-time visual editing system built in the game engine further enables interactive inspection and synchronized timeline adjustment across scenes, behaviours, and cameras. Extensive experiments and human evaluations show that Mind-of-Director generates high-quality, semantically grounded previz sequences in approximately 25 minutes per idea, demonstrating the effectiveness of agent collaboration for both automated prototyping and human-in-the-loop filmmaking.

Method: Created a 70-hour dataset from talk-show exchanges using semi-automatic pipeline with multi-person tracking, speech diarization, and human verification to extract temporally aligned host/guest tracks. Used MultiTalk-style diffusion model for reactive avatar generation conditioned on cross-person visual context.

Result: Dataset provides 14k clips with tight crops and metadata. Conditioning on cross-person visual context yields small but consistent improvements in Emotion-FID and FVD metrics while preserving lip-sync quality compared to audio-only baseline.

Conclusion: F2F-JF dataset enables studying dyadic sequential behavior and provides an end-to-end blueprint for conversational modeling, with applications to reactive digital avatars and multimodal interaction analysis.

Abstract: Modeling the reactive tempo of human conversation remains difficult because most audio-visual datasets portray isolated speakers delivering short monologues. We introduce \textbf{Face-to-Face with Jimmy Fallon (F2F-JF)}, a 70-hour, 14k-clip dataset of two-person talk-show exchanges that preserves the sequential dependency between a guest turn and the host’s response. A semi-automatic pipeline combines multi-person tracking, speech diarization, and lightweight human verification to extract temporally aligned host/guest tracks with tight crops and metadata that are ready for downstream modeling. We showcase the dataset with a reactive, speech-driven digital avatar task in which the host video during $[t_1,t_2]$ is generated from their audio plus the guest’s preceding video during $[t_0,t_1]$. Conditioning a MultiTalk-style diffusion model on this cross-person visual context yields small but consistent Emotion-FID and FVD gains while preserving lip-sync quality relative to an audio-only baseline. The dataset, preprocessing recipe, and baseline together provide an end-to-end blueprint for studying dyadic, sequential behavior, which we expand upon throughout the paper. Dataset and code will be made publicly available.

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2603.15507: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.15507&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: GTM proposes a unified BnB-based framework for globally-optimal TL loss minimization. It uses a hybrid solving design: for an n-dimensional problem, it performs BnB search over an (n-1)-dimensional subspace while solving the remaining 1D variable by bounding the objective function. This enables derivation of Lipschitz-continuous bounding functions that can be efficiently solved by the DIRECT global Lipschitz solver.

Result: GTM demonstrates remarkable threshold resilience and the highest efficiency compared to baseline methods on robust linear regression. Extensive experiments on various geometric estimation problems show GTM achieves state-of-the-art outlier-robustness and threshold-resilience while maintaining high efficiency across diverse estimation tasks.

Conclusion: GTM provides a systematic framework for globally minimizing truncated losses with branch-and-bound, overcoming limitations of consensus maximization by better leveraging residual information and achieving improved threshold resilience and computational efficiency in outlier-robust geometric estimation.

Abstract: To achieve outlier-robust geometric estimation, robust objective functions are generally employed to mitigate the influence of outliers. The widely used consensus maximization(CM) is highly robust when paired with global branch-and-bound(BnB) search. However, CM relies solely on inlier counts and is sensitive to the inlier threshold. Besides, the discrete nature of CM leads to loose bounds, necessitating extensive BnB iterations and computation cost. Truncated losses(TL), another continuous alternative, leverage residual information more effectively and could potentially overcome these issues. But to our knowledge, no prior work has systematically explored globally minimizing TL with BnB and its potential for enhanced threshold resilience or search efficiency. In this work, we propose GTM, the first unified BnB-based framework for globally-optimal TL loss minimization across diverse geometric problems. GTM involves a hybrid solving design: given an n-dimensional problem, it performs BnB search over an (n-1)-dimensional subspace while the remaining 1D variable is solved by bounding the objective function. Our hybrid design not only reduces the search space, but also enables us to derive Lipschitz-continuous bounding functions that are general, tight, and can be efficiently solved by a classic global Lipschitz solver named DIRECT, which brings further acceleration. We conduct a systematic evaluation on various BnB-based methods for CM and TL on the robust linear regression problem, showing that GTM enjoys remarkable threshold resilience and the highest efficiency compared to baseline methods. Furthermore, we apply GTM on different geometric estimation problems with diverse residual forms. Extensive experiments demonstrate that GTM achieves state-of-the-art outlier-robustness and threshold-resilience while maintaining high efficiency across these estimation tasks.

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot determine conclusion without access to paper content

Abstract: Failed to fetch summary for 2603.15484: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.15484&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes HiMemVLN which integrates a Hierarchical Memory System into a multimodal large model to enhance visual perception recall and long-term localization, specifically designed to mitigate the Navigation Amnesia problem identified through detailed analysis of navigation processes.

Result: Extensive experiments in both simulated and real-world environments show HiMemVLN achieves nearly twice the performance of open-source state-of-the-art methods, significantly closing the gap with closed-source approaches.

Conclusion: The Hierarchical Memory System effectively addresses Navigation Amnesia in VLN tasks, enabling open-source multimodal LLMs to achieve substantially improved navigation performance while avoiding the costs and risks of closed-source models.

Abstract: LLM-based agents have demonstrated impressive zero-shot performance in vision-language navigation (VLN) tasks. However, most zero-shot methods primarily rely on closed-source LLMs as navigators, which face challenges related to high token costs and potential data leakage risks. Recent efforts have attempted to address this by using open-source LLMs combined with a spatiotemporal CoT framework, but they still fall far short compared to closed-source models. In this work, we identify a critical issue, Navigation Amnesia, through a detailed analysis of the navigation process. This issue leads to navigation failures and amplifies the gap between open-source and closed-source methods. To address this, we propose HiMemVLN, which incorporates a Hierarchical Memory System into a multimodal large model to enhance visual perception recall and long-term localization, mitigating the amnesia issue and improving the agent’s navigation performance. Extensive experiments in both simulated and real-world environments demonstrate that HiMemVLN achieves nearly twice the performance of the open-source state-of-the-art method. The code is available at https://github.com/lvkailin0118/HiMemVLN.

Method: Proposes M2IR framework with two key components: 1) Mamba-Style Transformer (MST) blocks that perform pixel-wise selective state modulation to mitigate degradations while preserving structural integrity during encoding, and 2) Adaptive Degradation Expert Collaboration (ADEC) module that uses degradation-specific experts guided by a DA-CLIP-driven router plus a shared expert to eliminate residual degradations through targeted cooperative restoration.

Result: M2IR achieves superior generalization, enhanced adaptability, and refined recovery of fine-grained details across diverse all-in-one image restoration benchmarks by transitioning from passive reaction to active degradation control.

Conclusion: The proposed framework effectively harnesses learned representations to achieve better restoration performance through proactive degradation regulation during encoding and efficient residual degradation elimination during decoding.

Abstract: While Transformer-based architectures have dominated recent advances in all-in-one image restoration, they remain fundamentally reactive: propagating degradations rather than proactively suppressing them. In the absence of explicit suppression mechanisms, degraded signals interfere with feature learning, compelling the decoder to balance artifact removal and detail preservation, thereby increasing model complexity and limiting adaptability. To address these challenges, we propose M2IR, a novel restoration framework that proactively regulates degradation propagation during the encoding stage and efficiently eliminates residual degradations during decoding. Specifically, the Mamba-Style Transformer (MST) block performs pixel-wise selective state modulation to mitigate degradations while preserving structural integrity. In parallel, the Adaptive Degradation Expert Collaboration (ADEC) module utilizes degradation-specific experts guided by a DA-CLIP-driven router and complemented by a shared expert to eliminate residual degradations through targeted and cooperative restoration. By integrating the MST block and ADEC module, M2IR transitions from passive reaction to active degradation control, effectively harnessing learned representations to achieve superior generalization, enhanced adaptability, and refined recovery of fine-grained details across diverse all-in-one image restoration benchmarks. Our source codes are available at https://github.com/Im34v/M2IR.

Method: Proposes RadarXFormer framework that directly leverages raw radar spectra (not sparse point clouds) to construct efficient 3D representations. Uses cross-dimension (3D-2D) fusion where multi-scale 3D spherical radar feature cubes are fused with complementary 2D image feature maps.

Result: Experiments on K-Radar dataset show improved detection accuracy and robustness under challenging conditions while maintaining real-time inference capability.

Conclusion: RadarXFormer enables efficient cross-modal fusion between 4D radar spectra and RGB images, addressing limitations of conventional perception systems and improving autonomous driving reliability in adverse conditions.

Abstract: Reliable perception is essential for autonomous driving systems to operate safely under diverse real-world traffic conditions. However, camera- and LiDAR-based perception systems suffer from performance degradation under adverse weather and lighting conditions, limiting their robustness and large-scale deployment in intelligent transportation systems. Radar-vision fusion provides a promising alternative by combining the environmental robustness and cost efficiency of millimeter-wave (mmWave) radar with the rich semantic information captured by cameras. Nevertheless, conventional 3D radar measurements lack height resolution and remain highly sparse, while emerging 4D mmWave radar introduces elevation information but also brings challenges such as signal noise and large data volume. To address these issues, this paper proposes RadarXFormer, a 3D object detection framework that enables efficient cross-modal fusion between 4D radar spectra and RGB images. Instead of relying on sparse radar point clouds, RadarXFormer directly leverages raw radar spectra and constructs an efficient 3D representation that reduces data volume while preserving complete 3D spatial information. The “X” highlights the proposed cross-dimension (3D-2D) fusion mechanism, in which multi-scale 3D spherical radar feature cubes are fused with complementary 2D image feature maps. Experiments on the K-Radar dataset demonstrate improved detection accuracy and robustness under challenging conditions while maintaining real-time inference capability.

Method: Two-stage semantic distillation: 1) Derive structured semantic supervision from ground-truth ARKit coefficients, 2) Distill this knowledge into a multimodal large language model to predict interpretable facial action coefficients from images.

Result: Language-aligned semantic supervision improves both coefficient accuracy and perceptual consistency, enables strong cross-identity generalization, and provides robustness to large domain shifts including cartoon faces.

Conclusion: SemanticFace successfully bridges the gap between traditional facial action estimation and interpretable semantic reasoning, demonstrating the value of language-aligned supervision for multimodal understanding tasks.

Abstract: Facial action estimation from a single image is often formulated as predicting or fitting parameters in compact expression spaces, which lack explicit semantic interpretability. However, many practical applications, such as avatar control and human-computer interaction, require interpretable facial actions that correspond to meaningful muscle movements. In this work, we propose \textbf{SemanticFace}, a framework for facial action estimation in the interpretable ARKit blendshape space that reformulates coefficient prediction as structured semantic reasoning. SemanticFace adopts a two-stage semantic distillation paradigm: it first derives structured semantic supervision from ground-truth ARKit coefficients and then distills this knowledge into a multimodal large language model to predict interpretable facial action coefficients from images. Extensive experiments demonstrate that language-aligned semantic supervision improves both coefficient accuracy and perceptual consistency, while enabling strong cross-identity generalization and robustness to large domain shifts, including cartoon faces.

Method: Combines 2.5D and 3D representations: 2.5D branch uses DINOv3 vision transformer on multi-view CT slices (axial, coronal, sagittal); 3D branch uses ResNet-18 pretrained with Variance Risk Extrapolation (VREx) and supervised contrastive learning. Final predictions via logit-level ensemble inference.

Result: On PHAROS-AIF-MIH benchmark: binary COVID-19 detection achieves 94.48% accuracy and 0.9426 Macro F1-score; multi-class disease classification achieves 79.35% accuracy and 0.7497 Macro F1-score with 2.5D DINOv3 model.

Conclusion: The combination of pretrained slice-based representations with volumetric modeling is effective for robust multi-source medical imaging analysis, with the ensemble approach outperforming individual models.

Abstract: We propose a deep learning framework for COVID-19 detection and disease classification from chest CT scans that integrates both 2.5D and 3D representations to capture complementary slice-level and volumetric information. The 2.5D branch processes multi-view CT slices (axial, coronal, sagittal) using a DINOv3 vision transformer to extract robust visual features, while the 3D branch employs a ResNet-18 architecture to model volumetric context and is pretrained with Variance Risk Extrapolation (VREx) followed by supervised contrastive learning to improve cross-source robustness. Predictions from both branches are combined through logit-level ensemble inference. Experiments on the PHAROS-AIF-MIH benchmark demonstrate the effectiveness of the proposed approach: for binary COVID-19 detection, the ensemble achieves 94.48% accuracy and a 0.9426 Macro F1-score, outperforming both individual models, while for multi-class disease classification the 2.5D DINOv3 model achieves the best performance with 79.35% accuracy and a 0.7497 Macro F1-score. These results highlight the benefit of combining pretrained slice-based representations with volumetric modeling for robust multi-source medical imaging analysis. Code is available at https://github.com/HySonLab/PHAROS-AIF-MIH

Method: Proposes DamageArbiter, a multimodal disagreement-driven arbitration framework using CLIP models. It leverages complementary strengths of unimodal (image-only, text-only) and multimodal models, employing a lightweight logistic regression meta-classifier to arbitrate cases of disagreement between models.

Result: DamageArbiter improved accuracy from 74.33% (ViT-B/32 image-only) to 82.79%, surpassing the 80% threshold with 8.46% absolute improvement. It reduces overconfidence errors in ambiguous disaster scenarios and provides more reliable geo-referenced predictions.

Conclusion: The work advances street-view-based disaster assessment from coarse severity classification toward a more reliable and interpretable framework by effectively combining multimodal and unimodal approaches through intelligent arbitration.

Abstract: Analyzing street-view imagery with computer vision models for rapid, hyperlocal damage assessment is becoming popular and valuable in emergency response and recovery, but traditional models often act like black boxes, lacking interpretability and reliability. This study proposes a multimodal disagreement-driven Arbitration framework powered by Contrastive Language-Image Pre-training (CLIP) models, DamageArbiter, to improve the accuracy, interpretability, and robustness of damage estimation from street-view imagery. DamageArbiter leverages the complementary strengths of unimodal and multimodal models, employing a lightweight logistic regression meta-classifier to arbitrate cases of disagreement. Using 2,556 post-disaster street-view images, paired with both manually generated and large language model (LLM)-generated text descriptions, we systematically compared the performance of unimodal models (including image-only and text-only models), multimodal CLIP-based models, and DamageArbiter. Notably, DamageArbiter improved the accuracy from 74.33% (ViT-B/32, image-only) to 82.79%, surpassing the 80% accuracy threshold and achieving an absolute improvement of 8.46% compared to the strongest baseline model. Beyond improvements in overall accuracy, compared to visual models relying solely on images, DamageArbiter, through arbitration of discrepancies between unimodal and multimodal predictions, mitigates common overconfidence errors in visual models, especially in situations where disaster visual cues are ambiguous or subject to interference, reducing overconfidence but incorrect predictions. We further mapped and analyzed geo-referenced predictions and misclassifications to compare model performance across locations. Overall, this work advances street-view-based disaster assessment from coarse severity classification toward a more reliable and interpretable framework.

Method: Introduces Residual Fisher Information Matrix (ResFIM) to quantify parameter sensitivity to domain differences. Uses spectral transfer strategy to estimate ResFIM without accessing private data. Partitions parameters into domain-sensitive and domain-invariant components, then aggregates only domain-invariant parameters on server to construct personalized models.

Result: Extensive experiments on public datasets show pFL-ResFIM consistently outperforms state-of-the-art personalized federated learning methods.

Conclusion: The proposed pFL-ResFIM framework effectively addresses data heterogeneity in federated learning through parameter-level personalization while maintaining privacy, demonstrating superior performance over existing methods.

Abstract: Federated learning enables multiple clients (institutions) to collaboratively train machine learning models without sharing their private data. To address the challenge of data heterogeneity across clients, personalized federated learning (pFL) aims to learn customized models for each client. In this work, we propose pFL-ResFIM, a novel pFL framework that achieves client-adaptive personalization at the parameter level. Specifically, we introduce a new metric, Residual Fisher Information Matrix (ResFIM), to quantify the sensitivity of model parameters to domain discrepancies. To estimate ResFIM for each client model under privacy constraints, we employ a spectral transfer strategy that generates simulated data reflecting the domain styles of different clients. Based on the estimated ResFIM, we partition model parameters into domain-sensitive and domain-invariant components. A personalized model for each client is then constructed by aggregating only the domain-invariant parameters on the server. Extensive experiments on public datasets demonstrate that pFL-ResFIM consistently outperforms state-of-the-art methods, validating its effectiveness.

Method: Fine-tunes less than 0.2% of BrushNet weights using rank-constrained LoRA (Low-Rank Adaptation) on a pretrained diffusion model. Uses only 7390 artefact-clean pairs distilled from 739 experimental scans. Creates a public SPM InpBench benchmark for evaluation.

Result: Improves PSNR by 6.61 dB and halves LPIPS relative to zero-shot inference. Matches or slightly surpasses full retraining accuracy while being trainable on a single GPU instead of four high-memory cards. Generalizes across SPM image channels (height, amplitude, phase), restores subtle structural details, and suppresses hallucination artefacts.

Conclusion: The lightweight framework enables efficient, scalable recovery of irreplaceable SPM images and paves the way for broader diffusion model adoption in nanoscopic imaging analysis.

Abstract: Scanning Probe Microscopy or SPM offers nanoscale resolution but is frequently marred by structured artefacts such as line scan dropout, gain induced noise, tip convolution, and phase hops. While most available methods treat SPM artefact removal as isolated denoising or interpolation tasks, the generative inpainting perspective remains largely unexplored. In this work, we introduce a diffusion based inpainting framework tailored to scientific grayscale imagery. By fine tuning less than 0.2 percent of BrushNet weights with rank constrained low rank adaptation (LoRA), we adapt a pretrained diffusion model using only 7390 artefact, clean pairs distilled from 739 experimental scans. On our forthcoming public SPM InpBench benchmark, the LoRA enhanced model lifts the Peak Signal to Noise Ratio or PSNR by 6.61 dB and halves the Learned Perceptual Image Patch Similarity or LPIPS relative to zero-shot inference, while matching or slightly surpassing the accuracy of full retraining, trainable on a single GPU instead of four high-memory cards. The approach generalizes across various SPM image channels including height, amplitude and phase, faithfully restores subtle structural details, and suppresses hallucination artefacts inherited from natural image priors. This lightweight framework enables efficient, scalable recovery of irreplaceable SPM images and paves the way for a broader diffusion model adoption in nanoscopic imaging analysis.

Method: Proposes an end-to-end autonomous driving framework using a vision-language-action (VLA) model with mixture-of-transformer architecture featuring joint attention sharing. Uses asynchronous execution at different task frequencies for efficient fast-slow inference.

Result: Achieves competitive performance on multiple benchmarks in both open- and closed-loop settings. Shows pretrained VLMs can achieve competitive multi-task scene understanding through semantic prompting alone, while fine-tuning remains essential for action-level tasks like decision-making and trajectory planning.

Conclusion: The proposed VLA framework effectively unifies reasoning and action generation for autonomous driving, preserving pretrained VLM capabilities while enabling efficient inference. Reveals functional boundaries of pretrained VLMs in AD applications.

Abstract: Integrating vision-language models (VLMs) into end-to-end (E2E) autonomous driving (AD) systems has shown promise in improving scene understanding. However, existing integration strategies suffer from several limitations: they either struggle to resolve distribution misalignment between reasoning and action spaces, underexploit the general reasoning capabilities of pretrained VLMs, or incur substantial inference latency during action policy generation, which degrades driving performance. To address these challenges, we propose \OURS in this work, an end-to-end AD framework that unifies reasoning and action generation within a single vision-language-action (VLA) model. Our approach leverages a mixture-of-transformer (MoT) architecture with joint attention sharing, which preserves the general reasoning capabilities of pre-trained VLMs while enabling efficient fast-slow inference through asynchronous execution at different task frequencies. Extensive experiments on multiple benchmarks, under both open- and closed-loop settings, demonstrate that \OURS achieves competitive performance compared to state-of-the-art methods. We further investigate the functional boundary of pre-trained VLMs in AD, examining when AD-tailored fine-tuning is necessary. Our results show that pre-trained VLMs can achieve competitive multi-task scene understanding performance through semantic prompting alone, while fine-tuning remains essential for action-level tasks such as decision-making and trajectory planning. We refer to \href{https://automot-website.github.io/}{Project Page} for the demonstration videos and qualitative results.

Method: Proposes OSGeo framework using Rotated Bounding Boxes (RBoxes) as natural extension of detection paradigm. Includes multi-scale perception module and orientation-sensitive head to accurately regress RBoxes. Also releases CVOGL-R dataset with precise RBox annotations.

Result: OSGeo achieves state-of-the-art performance, consistently matching or surpassing leading segmentation-based methods’ accuracy while reducing annotation cost by over an order of magnitude.

Conclusion: RBox-based detection paradigm effectively bridges accuracy gap between detection and segmentation methods for CVOGL, offering high precision with significantly lower annotation cost.

Abstract: Cross-View object geo-localization (CVOGL) aims to precisely determine the geographic coordinates of a query object from a ground or drone perspective by referencing a satellite map. Segmentation-based approaches offer high precision but require prohibitively expensive pixel-level annotations, whereas more economical detection-based methods suffer from lower accuracy. This performance disparity in detection is primarily caused by two factors: the poor geometric fit of Horizontal Bounding Boxes (HBoxes) for oriented objects and the degradation in precision due to feature map scaling. Motivated by these, we propose leveraging Rotated Bounding Boxes (RBoxes) as a natural extension of the detection-based paradigm. RBoxes provide a much tighter geometric fit to oriented objects. Building on this, we introduce OSGeo, a novel geo-localization framework, meticulously designed with a multi-scale perception module and an orientation-sensitive head to accurately regress RBoxes. To support this scheme, we also construct and release CVOGL-R, the first dataset with precise RBox annotations for CVOGL. Extensive experiments demonstrate that our OSGeo achieves state-of-the-art performance, consistently matching or even surpassing the accuracy of leading segmentation-based methods but with an annotation cost that is over an order of magnitude lower.

Method: Proposes RealVLG-11B dataset with multi-granularity annotations and RealVLG-R1 model using reinforcement fine-tuning on pretrained vision-language models to predict bounding boxes, segmentation masks, grasp poses, and contact points.

Result: Demonstrates zero-shot perception and manipulation in real-world unseen environments, establishing a unified semantic-visual multimodal benchmark for language-driven robotic tasks.

Conclusion: RealVLG provides comprehensive data and evaluation platform for language-driven robotic perception and grasping policy learning, bridging visual-language understanding with physical manipulation.

Abstract: Visual-language grounding aims to establish semantic correspondences between natural language and visual entities, enabling models to accurately identify and localize target objects based on textual instructions. Existing VLG approaches focus on coarse-grained, object-level localization, while traditional robotic grasping methods rely predominantly on geometric cues and lack language guidance, which limits their applicability in language-driven manipulation scenarios. To address these limitations, we propose the RealVLG framework, which integrates the RealVLG-11B dataset and the RealVLG-R1 model to unify real-world visual-language grounding and grasping tasks. RealVLG-11B dataset provides multi-granularity annotations including bounding boxes, segmentation masks, grasp poses, contact points, and human-verified fine-grained language descriptions, covering approximately 165,000 images, over 800 object instances, 1.3 million segmentation, detection, and language annotations, and roughly 11 billion grasping examples. Building on this dataset, RealVLG-R1 employs Reinforcement Fine-tuning on pretrained large-scale vision-language models to predict bounding boxes, segmentation masks, grasp poses, and contact points in a unified manner given natural language instructions. Experimental results demonstrate that RealVLG supports zero-shot perception and manipulation in real-world unseen environments, establishing a unified semantic-visual multimodal benchmark that provides a comprehensive data and evaluation platform for language-driven robotic perception and grasping policy learning. All data and code are publicly available at https://github.com/lif314/RealVLG-R1.

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2507.15509: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.15509&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions about paper content due to access limitations

Abstract: Failed to fetch summary for 2601.17468: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17468&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes LLMind with Bio-inspired Adaptive Sampling Strategy (BASS) using Mobius-parameterized non-uniform sampling, and closed-loop semantic feedback (CSF) via test-time adaptation to align perceptual saliency with textual information.

Result: Achieves dramatic improvements: +20% on VQAv2, +38% on Seed-Bench, +37% on A-OKVQA compared to uniform sampling under tight pixel budgets. Retains up to 82%, 92%, and 97% of full-resolution performance using only 1%, 3%, and 5% of pixels respectively.

Conclusion: LLMind enables efficient, adaptive visual representations for VLMs without architectural changes, achieving near-full performance with minimal pixels through bio-inspired selective attention mechanisms.

Abstract: Vision-Language Models (VLMs) typically assume a uniform spatial fidelity across the entire field of view of visual inputs, dedicating equal precision to even the uninformative regions. By contrast, human vision is neither uniform nor static; it is adaptive, selective, and resource-efficient. In light of this, we present the first systematic analysis of bio-inspired visual representation methods, providing insights for more efficient and adaptive VLMs. We propose LLMind (Looking Like the Mind), a novel training-free framework that mimics foveated encoding and cortical magnification in human vision to achieve adaptive, efficient representations for VLMs under tight pixel budgets. Our key idea is to explore a Bio-inspired Adaptive Sampling Strategy (BASS), enabling a Mobius-parameterized module that performs non-uniform sampling while preserving global scene structure. On top of BASS, we introduce closed-loop semantic feedback (CSF) via test-time adaptation to align perceptual saliency with textual information from the frozen VLM. We evaluate LLMind against uniform and other sampling baselines across diverse scene-level and region-guided visual question answering benchmarks. The results show dramatic gains, with average improvements of +20% on VQAv2, +38% on Seed-Bench, and +37% on A-OKVQA compared to uniform sampling under tight pixel budgets. More surprisingly, LLMind retains up to 82%, 92%, and 97% of the full-resolution performance using only 1%, 3%, and 5% of the pixels, respectively. Moreover, LLMind is lightweight, plug-and-play, and compatible with existing VLMs without requiring architectural changes.

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2508.13587: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.13587&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to draw conclusions due to access error

Abstract: Failed to fetch summary for 2602.20409: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.20409&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes SpiralDiff, a diffusion-based framework with signal-dependent noise weighting that adapts reconstruction fidelity across intensity levels. Also introduces CamLoRA, a camera-aware lightweight adaptation module that enables a unified model to adapt to different camera-specific ISP characteristics.

Result: Extensive experiments on four benchmark datasets demonstrate superiority in RGB-to-RAW conversion quality and downstream benefits in RAW-based object detection.

Conclusion: SpiralDiff effectively addresses the challenges of varying reconstruction difficulty across intensity levels and multi-camera adaptation, providing high-quality RGB-to-RAW conversion with practical benefits for downstream tasks.

Abstract: RAW images preserve superior fidelity and rich scene information compared to RGB, making them essential for tasks in challenging imaging conditions. To alleviate the high cost of data collection, recent RGB-to-RAW conversion methods aim to synthesize RAW images from RGB. However, they overlook two key challenges: (i) the reconstruction difficulty varies with pixel intensity, and (ii) multi-camera conversion requires camera-specific adaptation. To address these issues, we propose SpiralDiff, a diffusion-based framework tailored for RGB-to-RAW conversion with a signal-dependent noise weighting strategy that adapts reconstruction fidelity across intensity levels. In addition, we introduce CamLoRA, a camera-aware lightweight adaptation module that enables a unified model to adapt to different camera-specific ISP characteristics. Extensive experiments on four benchmark datasets demonstrate the superiority of SpiralDiff in RGB-to-RAW conversion quality and its downstream benefits in RAW-based object detection. Our code and model are available at https://github.com/Chuancy-TJU/SpiralDiff.

Method: Proposes PASTE framework with: 1) scattering keypoint generation and automatic annotation using Attributed Scattering Center model, 2) scattering topology injection module for multi-scale feature learning, and 3) scattering prior supervision strategy aligning predictions with scattering center distributions.

Result: Experiments show PASTE is compatible with various detectors, achieving 2.9% to 11.3% relative mAP gains over baselines with acceptable computation overhead. Visualization confirms successful embedding of scattering topological priors into feature space.

Conclusion: PASTE successfully integrates scattering physics into SAR detectors through a closed-loop architecture, improving performance while providing interpretability by distinguishing target and background scattering regions.

Abstract: Current deep learning-based object detection for Synthetic Aperture Radar (SAR) imagery mainly adopts optical image methods, treating targets as texture patches while ignoring inherent electromagnetic scattering mechanisms. Though scattering points have been studied to boost detection performance, most methods still rely on amplitude-based statistical models. Some approaches introduce frequency-domain information for scattering center extraction, but they suffer from high computation cost and poor compatibility with diverse datasets. Thus, effectively embedding scattering topological information into modern detection frameworks remains challenging. To solve these problems, this paper proposes the Physics-Aware Scattering Topology Embedding Framework (PASTE), a novel closed-loop architecture for comprehensive scattering prior integration. By building the full pipeline from topology generation, injection to joint supervision, PASTE elegantly integrates scattering physics into modern SAR detectors. Specifically, it designs a scattering keypoint generation and automatic annotation scheme based on the Attributed Scattering Center (ASC) model to produce scalable and physically consistent priors. A scattering topology injection module guides multi-scale feature learning, and a scattering prior supervision strategy constrains network optimization by aligning predictions with scattering center distributions. Experiments on real datasets show that PASTE is compatible with various detectors and brings relative mAP gains of 2.9% to 11.3% over baselines with acceptable computation overhead. Visualization of scattering maps verifies that PASTE successfully embeds scattering topological priors into feature space, clearly distinguishing target and background scattering regions, thus providing strong interpretability for results.

Method: Proposes PromPrune with semantic prominence-aware budget allocation and a two-stage selection pipeline. Adaptively balances local saliency preservation and global coverage according to each sample’s semantic prominence distribution by allocating token budgets between locally salient regions and globally diverse regions.

Result: On LLaVA-NeXT-7B, reduces FLOPs by 88% and prefill latency by 22% while preserving 97.5% of the original accuracy, maintaining strong performance even under high compression ratios.

Conclusion: Sample-adaptive visual token compression that dynamically adjusts to semantic prominence distribution is more effective than static compression strategies, enabling efficient VLMs with minimal performance degradation.

Abstract: Large Vision-Language Models (VLMs) achieve strong multimodal understanding capabilities by leveraging high-resolution visual inputs, but the resulting large number of visual tokens creates a major computational bottleneck. Recent work mitigates this issue through visual token compression, typically compressing tokens based on saliency, diversity, or a fixed combination of both. We observe that the distribution of semantic prominence varies substantially across samples, leading to different optimal trade-offs between local saliency preservation and global coverage. This observation suggests that applying a static compression strategy across all samples can be suboptimal. Motivated by this insight, we propose PromPrune, a sample-adaptive visual token selection framework composed of semantic prominence-aware budget allocation and a two-stage selection pipeline. Our method adaptively balances local saliency preservation and global coverage according to the semantic prominence distribution of each sample. By allocating token budgets between locally salient regions and globally diverse regions, our method maintains strong performance even under high compression ratios. On LLaVA-NeXT-7B, our approach reduces FLOPs by 88% and prefill latency by 22% while preserving 97.5% of the original accuracy.

Method: Constructs multi-scale sphere graphs to sample input images, uses graph neural networks to jointly estimate tracking directions and vessel radii, employs gating-based feature fusion for multi-scale representations, incorporates geometry-aware weighting for class imbalance, and uses wave-propagation-based skeleton tracking with space-occupancy filtering to reduce spurious skeletons.

Result: Achieves competitive performance in both overlapping and topological metrics on two vessel datasets with different geometries compared to state-of-the-art baselines.

Conclusion: TopoVST effectively addresses challenges in vessel skeleton extraction by improving topological faithfulness through novel graph-based tracking and filtering approaches.

Abstract: Automatic extraction of vessel skeletons is crucial for many clinical applications. However, achieving topologically faithful delineation of thin vessel skeletons remains highly challenging, primarily due to frequent discontinuities and the presence of spurious skeleton segments. To address these difficulties, we propose TopoVST, a topology-fidelitious vessel skeleton tracker. TopoVST constructs multi-scale sphere graphs to sample the input image and employs graph neural networks to jointly estimate tracking directions and vessel radii. The utilization of multi-scale representations is enhanced through a gating-based feature fusion mechanism, while the issue of class imbalance during training is mitigated by embedding a geometry-aware weighting scheme into the directional loss. In addition, we design a wave-propagation-based skeleton tracking algorithm that explicitly mitigates the generation of spurious skeletons through space-occupancy filtering. We evaluate TopoVST on two vessel datasets with different geometries. Extensive comparisons with state-of-the-art baselines demonstrate that TopoVST achieves competitive performance in both overlapping and topological metrics. Our source code is available at: https://github.com/EndoluminalSurgicalVision-IMR/TopoVST.

Method: ILV constructs an explicit 3D latent volume that is repeatedly updated by conditioning on multi-view X-ray features and learned anatomical priors. Key architectural components include an X-ray feature volume, group cross-attention, efficient self-attention, and view-wise feature aggregation to efficiently realize latent volume refinement.

Result: Extensive experiments on a large-scale dataset of approximately 14,000 CT volumes demonstrate that ILV significantly outperforms existing feed-forward and optimization-based methods in both reconstruction quality and speed.

Conclusion: ILV enables fast and accurate sparse-view CBCT reconstruction suitable for clinical use by integrating data-driven priors with iterative reconstruction principles to overcome key limitations of prior feed-forward models.

Abstract: A long-term goal in CT imaging is to achieve fast and accurate 3D reconstruction from sparse-view projections, thereby reducing radiation exposure, lowering system cost, and enabling timely imaging in clinical workflows. Recent feed-forward approaches have shown strong potential toward this overarching goal, yet their results still suffer from artifacts and loss of fine details. In this work, we introduce Iterative Latent Volumes (ILV), a feed-forward framework that integrates data-driven priors with classical iterative reconstruction principles to overcome key limitations of prior feed-forward models in sparse-view CBCT reconstruction. At its core, ILV constructs an explicit 3D latent volume that is repeatedly updated by conditioning on multi-view X-ray features and the learned anatomical prior, enabling the recovery of fine structural details beyond the reach of prior feed-forward models. In addition, we develop and incorporate several key architectural components, including an X-ray feature volume, group cross-attention, efficient self-attention, and view-wise feature aggregation, that efficiently realize its core latent volume refinement concept. Extensive experiments on a large-scale dataset of approximately 14,000 CT volumes demonstrate that ILV significantly outperforms existing feed-forward and optimization-based methods in both reconstruction quality and speed. These results show that ILV enables fast and accurate sparse-view CBCT reconstruction suitable for clinical use. The project page is available at: https://sngryonglee.github.io/ILV/.

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2509.26255: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.26255&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes a two-stage HDR video reconstruction framework with: 1) Flow adapter that adapts generic optical flow for robust cross-exposure alignment, 2) Physical motion modeling to identify salient motion regions, and 3) Motion-aware refinement network that aggregates complementary information while removing ghosting and noise.

Result: Extensive experiments demonstrate state-of-the-art performance on real-world HDR video benchmarks, producing ghost-free and high-fidelity results under large motion and exposure variations.

Conclusion: F²HDR effectively addresses challenges in HDR video reconstruction from alternating-exposure LDR frames, particularly in dynamic scenes with complex motion, achieving superior performance through robust motion perception and refinement mechanisms.

Abstract: Reconstructing High Dynamic Range (HDR) videos from sequences of alternating-exposure Low Dynamic Range (LDR) frames remains highly challenging, especially under dynamic scenes where cross-exposure inconsistencies and complex motion make inter-frame alignment difficult, leading to ghosting and detail loss. Existing methods often suffer from inaccurate alignment, suboptimal feature aggregation, and degraded reconstruction quality in motion-dominated regions. To address these challenges, we propose $\text{F}^2\text{HDR}$, a two-stage HDR video reconstruction framework that robustly perceives inter-frame motion and restores fine details in complex dynamic scenarios. The proposed framework integrates a flow adapter that adapts generic optical flow for robust cross-exposure alignment, a physical motion modeling to identify salient motion regions, and a motion-aware refinement network that aggregates complementary information while removing ghosting and noise. Extensive experiments demonstrate that $\text{F}^2\text{HDR}$ achieves state-of-the-art performance on real-world HDR video benchmarks, producing ghost-free and high-fidelity results under large motion and exposure variations.

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2510.04673: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.04673&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Decomposes illustrations into semantically meaningful production layers inspired by anime creation pipeline. Uses lightweight layer semantic embeddings for task guidance and layer-wise losses for supervision. Constructs high-quality dataset simulating standard anime workflow.

Result: Achieves accurate and visually coherent layer decompositions. The resulting layered representation enables downstream tasks like recoloring and texture embedding for content creation and illustration editing.

Conclusion: Proposed framework successfully decomposes anime illustrations into production-aligned layers, providing structured representation that supports various editing tasks in anime content creation.

Abstract: Recent generative image editing methods adopt layered representations to mitigate the entangled nature of raster images and improve controllability, typically relying on object-based segmentation. However, such strategies may fail to capture the structural and stylized properties of human-created images, such as anime illustrations. To solve this issue, we propose a workflow-aware structured layer decomposition framework tailored to the illustration production of anime artwork. Inspired by the creation pipeline of anime production, our method decomposes the illustration into semantically meaningful production layers, including line art, flat color, shadow, and highlight. To decouple all these layers, we introduce lightweight layer semantic embeddings to provide specific task guidance for each layer. Furthermore, a set of layer-wise losses is incorporated to supervise the training process of individual layers. To overcome the lack of ground-truth layered data, we construct a high-quality illustration dataset that simulated the standard anime production workflow. Experiments demonstrate that the accurate and visually coherent layer decompositions were achieved by using our method. We believe that the resulting layered representation further enables downstream tasks such as recoloring and embedding texture, supporting content creation, and illustration editing. Code is available at: https://github.com/zty0304/Anime-layer-decomposition

Method: Proposes Chain of Events (CoE) paradigm that constructs temporal event chains to implicitly enforce MLLMs to focus on visual content and logical connections between videos and future events. Uses multiple training protocols to incentivize model’s reasoning capability.

Result: Experimental results on public benchmarks demonstrate the method outperforms both leading open-source and commercial MLLMs, establishing a new state-of-the-art on the VEP task.

Conclusion: The CoE paradigm effectively addresses MLLMs’ limitations in video event prediction by enhancing temporal modeling and logical reasoning capabilities, achieving superior performance on VEP benchmarks.

Abstract: Despite advances in the application of MLLMs for various video tasks, video event prediction (VEP) remains relatively underexplored. VEP requires the model to perform fine-grained temporal modeling of videos and establish logical relationships between videos and future events, which current MLLMs still struggle with. In this work, we first present a comprehensive evaluation of current leading MLLMs on the VEP task, revealing the reasons behind their inaccurate predictions, including lack of logical reasoning ability for future events prediction and insufficient utilization of visual information. To address these challenges, we propose \textbf{C}hain \textbf{o}f \textbf{E}vents (\textbf{CoE}) paradigm, which constructs temporal event chains to implicitly enforce MLLM focusing on the visual content and the logical connections between videos and future events, incentivizing model’s reasoning capability with multiple training protocols. Experimental results on public benchmarks demonstrate that our method outperforms both leading open-source and commercial MLLMs, establishing a new state-of-the-art on the VEP task. Codes and models will be released soon.

Method: RFD adapts relevance feedback mechanism from information retrieval to diffusion models. Users provide implicit, multi-select visual feedback instead of textual dialogue. The system uses an expert-curated feature repository and information-theoretic weighted cumulative preference analysis to translate feedback into generative guidance. It employs probabilistic sampling for prompt reconstruction to balance exploitation and exploration.

Result: Extensive experiments demonstrate that RFD effectively captures users’ true visual intent and significantly outperforms baselines in preference alignment.

Conclusion: RFD provides a training-free, model-agnostic solution that bridges the gap between user visual intent and text-to-image generation by replacing explicit textual dialogue with implicit visual feedback, achieving better alignment with lower cognitive load.

Abstract: Text-to-image generation using diffusion models has achieved remarkable success. However, users often possess clear visual intents but struggle to express them precisely in language, resulting in ambiguous prompts and misaligned images. Existing methods struggle to bridge this gap, typically relying on high-load textual dialogues, opaque black-box inferences, or expensive fine-tuning. They fail to simultaneously achieve low cognitive load, interpretable preference inference, and remain training-free and model-agnostic. To address this, we propose RFD, an interactive framework that adapts the relevance feedback mechanism from information retrieval to diffusion models. In RFD, users replace explicit textual dialogue with implicit, multi-select visual feedback to minimize cognitive load, easily expressing complex, multi-dimensional preferences. To translate feedback into precise generative guidance, we construct an expert-curated feature repository and introduce an information-theoretic weighted cumulative preference analysis. This white-box method calculates preferences from current-round feedback and incrementally accumulates them, avoiding the concatenation of historical interactions and preventing inference degradation caused by lengthy contexts. Furthermore, RFD employs a probabilistic sampling mechanism for prompt reconstruction to balance exploitation and exploration, preventing output homogenization. Crucially, RFD operates entirely within the external text space, making it strictly training-free and model-agnostic as a universal plug-and-play solution. Extensive experiments demonstrate that RFD effectively captures the user’s true visual intent, significantly outperforming baselines in preference alignment.

Method: Proposes a multi-view diffusion transformer with fine-grained structured control for geometrically consistent multi-camera generation. Uses a two-stage training strategy: adaptive reference horizon conditioning and blend-forcing autoregressive training to address long-horizon consistency and iterative degradation. Includes system-level efficiency optimizations for low-latency inference.

Result: Achieves state-of-the-art performance among existing closed-loop autonomous driving simulation approaches on the nuScenes dataset, while maintaining sub-second latency on a single GPU.

Conclusion: FAR-Drive provides an effective solution for building learning-based closed-loop simulators for autonomous driving that maintain temporal consistency, mitigate autoregressive degradation, and satisfy low-latency requirements.

Abstract: Despite rapid progress in autonomous driving, reliable training and evaluation of driving systems remain fundamentally constrained by the lack of scalable and interactive simulation environments. Recent generative video models achieve remarkable visual fidelity, yet most operate in open-loop settings and fail to support fine-grained frame-level interaction between agent actions and environment evolution. Building a learning-based closed-loop simulator for autonomous driving poses three major challenges: maintaining long-horizon temporal and cross-view consistency, mitigating autoregressive degradation under iterative self-conditioning, and satisfying low-latency inference constraints. In this work, we propose FAR-Drive, a frame-level autoregressive video generation framework for autonomous driving. We introduce a multi-view diffusion transformer with fine-grained structured control, enabling geometrically consistent multi-camera generation. To address long-horizon consistency and iterative degradation, we design a two-stage training strategy consisting of adaptive reference horizon conditioning and blend-forcing autoregressive training, which progressively improves consistency and robustness under self-conditioning. To meet low-latency interaction requirements, we further integrate system-level efficiency optimizations for inference acceleration. Experiments on the nuScenes dataset demonstrate that our method achieves state-of-the-art performance among existing closed-loop autonomous driving simulation approaches, while maintaining sub-second latency on a single GPU.

Method: 1) Trajectory-aware Driving World Model that conditions on a trajectory vocabulary to enforce consistency between visual dynamics and motion intentions; 2) Transfer vision and motion encoders to a downstream Multi-modal Planner; 3) Future-aware Rewarder that distills future latent representation from the frozen world model to evaluate and select optimal trajectories in real-time.

Result: WorldDrive achieves leading planning performance among vision-only methods on NAVSIM, NAVSIM-v2, and nuScenes benchmarks while maintaining high-fidelity action-controlled video generation capabilities.

Conclusion: The framework demonstrates the effectiveness of unifying vision and motion representation for robust autonomous driving, showing that shared representations between scene generation and planning can improve both tasks simultaneously.

Abstract: End-to-end autonomous driving aims to generate safe and plausible planning policies from raw sensor input. Driving world models have shown great potential in learning rich representations by predicting the future evolution of a driving scene. However, existing driving world models primarily focus on visual scene representation, and motion representation is not explicitly designed to be planner-shared and inheritable, leaving a schism between the optimization of visual scene generation and the requirements of precise motion planning. We present WorldDrive, a holistic framework that couples scene generation and real-time planning via unifying vision and motion representation. We first introduce a Trajectory-aware Driving World Model, which conditions on a trajectory vocabulary to enforce consistency between visual dynamics and motion intentions, enabling the generation of diverse and plausible future scenes conditioned on a specific trajectory. We transfer the vision and motion encoders to a downstream Multi-modal Planner, ensuring the driving policy operates on mature representations pre-optimized by scene generation. A simple interaction between motion representation, visual representation, and ego status can generate high-quality, multi-modal trajectories. Furthermore, to exploit the world model’s foresight, we propose a Future-aware Rewarder, which distills future latent representation from the frozen world model to evaluate and select optimal trajectories in real-time. Extensive experiments on the NAVSIM, NAVSIM-v2, and nuScenes benchmarks demonstrate that WorldDrive achieves leading planning performance among vision-only methods while maintaining high-fidelity action-controlled video generation capabilities, providing strong evidence for the effectiveness of unifying vision and motion representation for robust autonomous driving.

Method: Proposes GT-PCQA with two key strategies: 1) 2D-3D joint training that formulates PCQA as relative quality comparison to unify large-scale IQA datasets with limited PCQA datasets using LoRA for parameter-efficient instruction tuning, and 2) geometry-texture decoupling strategy with dual-prompt mechanism and alternating optimization to mitigate texture-dominant bias and enhance sensitivity to geometric degradations.

Result: Extensive experiments demonstrate that GT-PCQA achieves competitive performance and exhibits strong generalization capabilities for point cloud quality assessment.

Conclusion: GT-PCQA successfully addresses the challenges of extending MLLM-based quality assessment to 3D point clouds through innovative training strategies and bias mitigation techniques, showing promising results for PCQA tasks.

Abstract: With the rapid advancement of Multi-modal Large Language Models (MLLMs), MLLM-based Image Quality Assessment (IQA) methods have shown promising generalization. However, directly extending these MLLM-based IQA methods to PCQA remains challenging. On the one hand, existing PCQA datasets are limited in scale, which hinders stable and effective instruction tuning of MLLMs. On the other hand, due to large-scale image-text pretraining, MLLMs tend to rely on texture-dominant reasoning and are insufficiently sensitive to geometric structural degradations that are critical for PCQA. To address these gaps, we propose a novel MLLM-based no-reference PCQA framework, termed GT-PCQA, which is built upon two key strategies. First, to enable stable and effective instruction tuning under scarce PCQA supervision, a 2D-3D joint training strategy is proposed. This strategy formulates PCQA as a relative quality comparison problem to unify large-scale IQA datasets with limited PCQA datasets. It incorporates a parameter-efficient Low-Rank Adaptation (LoRA) scheme to support instruction tuning. Second, a geometry-texture decoupling strategy is presented, which integrates a dual-prompt mechanism with an alternating optimization scheme to mitigate the inherent texture-dominant bias of pre-trained MLLMs, while enhancing sensitivity to geometric structural degradations. Extensive experiments demonstrate that GT-PCQA achieves competitive performance and exhibits strong generalization.

Method: End-to-end framework integrating physical priors with frequency-decoupled restoration block disentangling MSI features into amplitude (guided by NIR band) and phase (guided by PAN) components, plus interactive inter-frequency consistency module for cross-modal refinement.

Result: Superior performance on real-world and synthetic datasets, establishing new benchmark for pansharpening under realistic atmospheric degradations; introduces first real-world thin-cloud contaminated pansharpening dataset (PanTCR-GF2).

Conclusion: Pan-TCR effectively addresses pansharpening under thin cloudy conditions through unified modeling, frequency-domain analysis, and cross-modal consistency, advancing remote sensing image restoration.

Abstract: Pansharpening under thin cloudy conditions is a practically significant yet rarely addressed task, challenged by simultaneous spatial resolution degradation and cloud-induced spectral distortions. Existing methods often address cloud removal and pansharpening sequentially, leading to cumulative errors and suboptimal performance due to the lack of joint degradation modeling. To address these challenges, we propose a Unified Pansharpening Model with Thin Cloud Removal (Pan-TCR), an end-to-end framework that integrates physical priors. Motivated by theoretical analysis in the frequency domain, we design a frequency-decoupled restoration (FDR) block that disentangles the restoration of multispectral image (MSI) features into amplitude and phase components, each guided by complementary degradation-robust prompts: the near-infrared (NIR) band amplitude for cloud-resilient restoration, and the panchromatic (PAN) phase for high-resolution structural enhancement. To ensure coherence between the two components, we further introduce an interactive inter-frequency consistency (IFC) module, enabling cross-modal refinement that enforces consistency and robustness across frequency cues. Furthermore, we introduce the first real-world thin-cloud contaminated pansharpening dataset (PanTCR-GF2), comprising paired clean and cloudy PAN-MSI images, to enable robust benchmarking under realistic conditions. Extensive experiments on real-world and synthetic datasets demonstrate the superiority and robustness of Pan-TCR, establishing a new benchmark for pansharpening under realistic atmospheric degradations.

Method: CyCLeGen uses a fully integrated autoregressive architecture with cycle-consistent learning through two loops: image->layout->image and layout->image->layout. This enables self-improvement via synthetic supervision under reinforcement learning guided by cycle consistency.

Result: Extensive experiments show CyCLeGen achieves significant gains across diverse image understanding and generation benchmarks, demonstrating the effectiveness of the unified approach.

Conclusion: The paper highlights the potential of unified vision-language foundation models that integrate understanding and generation capabilities, offering advantages in introspection and data efficiency through cycle-consistent learning.

Abstract: We present CyCLeGen, a unified vision-language foundation model capable of both image understanding and image generation within a single autoregressive framework. Unlike existing vision models that depend on separate modules for perception and synthesis, CyCLeGen adopts a fully integrated architecture that enforces cycle-consistent learning through image->layout->image and layout->image->layout generation loops. This unified formulation introduces two key advantages: introspection, enabling the model to reason about its own generations, and data efficiency, allowing self-improvement via synthetic supervision under a reinforcement learning objective guided by cycle consistency. Extensive experiments show that CyCLeGen achieves significant gains across diverse image understanding and generation benchmarks, highlighting the potential of unified vision-language foundation models.

Method: Introduces Gaussian Splat Feature Adapter (GS-Adapter) that lifts input-view diffusion features into 3D Gaussian representations, renders geometry-constrained novel-view features, and fuses them with diffusion features to correct geometric inconsistencies. Operates in feature space rather than input level to avoid view-dependent color noise.

Result: Achieves state-of-the-art performance across 9 scenes and 18 settings with 11.3% and 14.9% improvements over SEVA and CameraCtrl, plus up to 2x reduction in translation error and 7x in Chamfer Distance.

Conclusion: GeoNVS successfully enhances geometric fidelity and camera controllability in novel view synthesis through explicit 3D geometric guidance, with plug-and-play design enabling zero-shot compatibility with various geometry models.

Abstract: Novel view synthesis requires strong 3D geometric consistency and the ability to generate visually coherent images across diverse viewpoints. While recent camera-controlled video diffusion models show promising results, they often suffer from geometric distortions and limited camera controllability. To overcome these challenges, we introduce GeoNVS, a geometry-grounded novel-view synthesizer that enhances both geometric fidelity and camera controllability through explicit 3D geometric guidance. Our key innovation is the Gaussian Splat Feature Adapter (GS-Adapter), which lifts input-view diffusion features into 3D Gaussian representations, renders geometry-constrained novel-view features, and adaptively fuses them with diffusion features to correct geometrically inconsistent representations. Unlike prior methods that inject geometry at the input level, GS-Adapter operates in feature space, avoiding view-dependent color noise that degrades structural consistency. Its plug-and-play design enables zero-shot compatibility with diverse feed-forward geometry models without additional training, and can be adapted to other video diffusion backbones. Experiments across 9 scenes and 18 settings demonstrate state-of-the-art performance, achieving 11.3% and 14.9% improvements over SEVA and CameraCtrl, with up to 2x reduction in translation error and 7x in Chamfer Distance.

Method: Proposes integrating second-order pooling with Voronoi cell inductive bias, inspired by NetVLAD’s association with second-order statistics. The method aggregates local descriptors into second-order matrices, whitens global descriptors to implicitly measure Mahalanobis distance while conserving Voronoi cluster properties, and addresses numerical instability with diverse techniques.

Result: Demonstrated performance gains on Oxford Robotcar and Wild-Places benchmarks, with analysis showing the numerical effects of the proposed whitening algorithm.

Conclusion: The proposed second-order pooling method with Voronoi cell integration and whitening effectively improves LiDAR place recognition performance while maintaining numerical stability and metric compatibility.

Abstract: The pooling layer plays a vital role in aggregating local descriptors into the metrizable global descriptor in the LiDAR Place Recognition (LPR). In particular, the second-order pooling is capable of capturing higher-order interactions among local descriptors. However, its existing methods in the LPR adhere to conventional implementations and post-normalization, and incur the descriptor unsuitable for Euclidean distancing. Based on the recent interpretation that associates NetVLAD with the second-order statistics, we propose to integrate second-order pooling with the inductive bias from Voronoi cells. Our novel pooling method aggregates local descriptors to form the second-order matrix and whitens the global descriptor to implicitly measure the Mahalanobis distance while conserving the cluster property from Voronoi cells, addressing its numerical instability during learning with diverse techniques. We demonstrate its performance gains through the experiments conducted on the Oxford Robotcar and Wild-Places benchmarks and analyze the numerical effect of the proposed whitening algorithm.

Method: Introduces MMSpec benchmark with 600 multimodal samples across six task categories, integrates ten speculative decoding algorithms under unified framework, and proposes ViSkip method that dynamically adapts speculation to vision tokens.

Result: Three key findings: (1) text-only methods degrade in multimodal scenarios, (2) vision awareness becomes increasingly important at larger batch sizes, (3) throughput speedup alone doesn’t reliably reflect latency performance. ViSkip achieves state-of-the-art performance.

Conclusion: Speculative decoding for VLMs requires vision-aware approaches, and MMSpec provides a comprehensive benchmark for evaluation. ViSkip demonstrates the importance of adapting speculation strategies to multimodal content.

Abstract: Vision-language models (VLMs) achieve strong performance on multimodal tasks but suffer from high inference latency due to large model sizes and long multimodal contexts. Speculative decoding has recently emerged as an effective acceleration technique, yet its behavior in VLMs remains insufficiently understood. We introduce MMSpec, the first benchmark for evaluating speculative decoding in vision-language models. MMSpec contains 600 multimodal samples across six task categories and integrates ten representative speculative decoding algorithms under a unified evaluation framework. Our study reveals three key findings: (1) methods designed for text-only LLMs degrade in multimodal scenarios, (2) vision awareness becomes increasingly important at larger batch sizes, and (3) throughput speedup alone does not reliably reflect latency performance. Motivated by these findings, we propose ViSkip, a plug-and-play speculative decoding method that dynamically adapts speculation to vision tokens and achieves state-of-the-art performance.

Method: Proposes a lightweight supervised network with recurrent blocks (RBs) to capture temporal dependencies for robust depth prediction from thermal images. Uses thermal refinement network (T-RefNet) to enhance feature visibility. Integrates refined thermal images and depth maps into ORB-SLAM3 for thermal-only localization. Trained on custom non-radiometric dataset to avoid expensive radiometric cameras.

Result: Achieves competitive depth accuracy: absolute relative error ~0.06 on VIVID++ indoor-dark dataset (baselines >0.11). On non-radiometric indoor set, error <0.10 (baselines >0.24). Thermal-only ORB-SLAM3 maintains mean trajectory error <0.4m. Demonstrates robust SLAM performance in low-light conditions.

Conclusion: The proposed thermal-only pipeline enables robust UAV navigation in challenging environments without GPS or visible light, using cost-effective non-radiometric thermal cameras. The integration of lightweight networks with SLAM shows promising results for real-time applications.

Abstract: Autonomous navigation in GPS-denied and visually degraded environments remains challenging for unmanned aerial vehicles (UAVs). To this end, we investigate the use of a monocular thermal camera as a standalone sensor on a UAV platform for real-time depth estimation and simultaneous localization and mapping (SLAM). To extract depth information from thermal images, we propose a novel pipeline employing a lightweight supervised network with recurrent blocks (RBs) integrated to capture temporal dependencies, enabling more robust predictions. The network combines lightweight convolutional backbones with a thermal refinement network (T-RefNet) to refine raw thermal inputs and enhance feature visibility. The refined thermal images and predicted depth maps are integrated into ORB-SLAM3, enabling thermal-only localization. Unlike previous methods, the network is trained on a custom non-radiometric dataset, obviating the need for high-cost radiometric thermal cameras. Experimental results on datasets and UAV flights demonstrate competitive depth accuracy and robust SLAM performance under low-light conditions. On the radiometric VIVID++ (indoor-dark) dataset, our method achieves an absolute relative error of approximately 0.06, compared to baselines exceeding 0.11. In our non-radiometric indoor set, baseline errors remain above 0.24, whereas our approach remains below 0.10. Thermal-only ORB-SLAM3 maintains a mean trajectory error under 0.4 m.

Method: Adapt a large image editing model (Qwen-Image-Edit) for Video Frame Interpolation using only 64-256 training samples via Low-Rank Adaptation (LoRA), without adding video-specific architectures or motion estimation modules.

Result: The adapted model successfully unlocks interpolation capabilities, while the baseline model fails at coherent intermediate frames. This demonstrates that foundation image editing models possess untapped potential for temporal tasks.

Conclusion: Spatial and temporal reasoning may be more intertwined in foundation models than previously recognized, offering data-efficient pathways for video synthesis in resource-constrained scenarios.

Abstract: Pre-trained image editing models exhibit strong spatial reasoning and object-aware transformation capabilities acquired from billions of image-text pairs, yet they possess no explicit temporal modeling. This paper demonstrates that these spatial priors can be repurposed to unlock temporal synthesis capabilities through minimal adaptation - without introducing any video-specific architecture or motion estimation modules. We show that a large image editing model (Qwen-Image-Edit), originally designed solely for static instruction-based edits, can be adapted for Video Frame Interpolation (VFI) using only 64-256 training samples via Low-Rank Adaptation (LoRA). Our core contribution is revealing that the model’s inherent understanding of “how objects transform” in static scenes contains latent temporal reasoning that can be activated through few-shot fine-tuning. While the baseline model completely fails at producing coherent intermediate frames, our parameter-efficient adaptation successfully unlocks its interpolation capability. Rather than competing with task-specific VFI methods trained from scratch on massive datasets, our work establishes that foundation image editing models possess untapped potential for temporal tasks, offering a data-efficient pathway for video synthesis in resource-constrained scenarios. This bridges the gap between image manipulation and video understanding, suggesting that spatial and temporal reasoning may be more intertwined in foundation models than previously recognized

Method: Proposes ClueNet with two-stage supervised fine-tuning: 1) Decoupled supervision aligns clue extraction and chain-based reasoning, 2) Inference supervision with adaptive clue filter refines high-order reasoning. Uses lightweight modules without extensive base model modifications.

Result: Outperforms state-of-the-art methods by ≥1.1% on NExT-QA, STAR, and MVBench benchmarks. Shows superior generalization, hallucination mitigation, inference efficiency, and cross-backbone compatibility.

Conclusion: ClueNet bridges the perception-to-generation gap in MLLM video understanding, providing an interpretable, faithful reasoning paradigm for high-stakes VideoQA applications.

Abstract: Multi-modal Large Language Models (MLLMs) have significantly advanced video reasoning, yet Video Question Answering (VideoQA) remains challenging due to its demand for temporal causal reasoning and evidence-grounded answer generation. Prevailing end-to-end MLLM frameworks lack explicit structured reasoning between visual perception and answer derivation, causing severe hallucinations and poor interpretability. Existing methods also fail to address three core gaps: faithful visual clue extraction, utility-aware clue filtering, and end-to-end clue-answer alignment. Inspired by hierarchical human visual cognition, we propose ClueNet, a clue-aware video reasoning framework with a two-stage supervised fine-tuning paradigm without extensive base model modifications. Decoupled supervision aligns clue extraction and chain-based reasoning, while inference supervision with an adaptive clue filter refines high-order reasoning, alongside lightweight modules for efficient inference. Experiments on NExT-QA, STAR, and MVBench show that ClueNet outperforms state-of-the-art methods by $\ge$ 1.1%, with superior generalization, hallucination mitigation, inference efficiency, and cross-backbone compatibility. This work bridges the perception-to-generation gap in MLLM video understanding, providing an interpretable, faithful reasoning paradigm for high-stakes VideoQA applications.

Method: Proposes two approaches: 1) Identifier as Visual Prompting (IdtVP) that uses molecule identifiers to activate chemical knowledge, and 2) Re3-DAPO reinforcement learning algorithm that optimizes reaction-level metrics directly. Also introduces ScannedRxn benchmark dataset.

Result: IdtVP enables powerful zero-shot and out-of-distribution capabilities, outperforming existing prompting strategies. Re3-DAPO achieves consistent gains over standard supervised fine-tuning.

Conclusion: The contributions advance accuracy and generalization of VLM-based reaction diagram parsing, with data, models, and code to be released.

Abstract: Reaction diagram parsing (RxnDP) is critical for extracting chemical synthesis information from literature. Although recent Vision-Language Models (VLMs) have emerged as a promising paradigm to automate this complex visual reasoning task, their application is fundamentally bottlenecked by the inability to align visual chemical entities with pre-trained knowledge, alongside the inherent discrepancy between token-level training and reaction-level evaluation. To address these dual challenges, this work enhances VLM-based RxnDP from two complementary perspectives: prompting representation and learning paradigms. First, we propose Identifier as Visual Prompting (IdtVP), which leverages naturally occurring molecule identifiers (e.g., bold numerals like 1a) to activate the chemical knowledge acquired during VLM pre-training. IdtVP enables powerful zero-shot and out-of-distribution capabilities, outperforming existing prompting strategies. Second, to further optimize performance within fine-tuning paradigms, we introduce Re3-DAPO, a reinforcement learning algorithm that leverages verifiable rewards to directly optimize reaction-level metrics, thereby achieving consistent gains over standard supervised fine-tuning. Additionally, we release the ScannedRxn benchmark, comprising scanned historical reaction diagrams with real-world artifacts, to rigorously assess model robustness and out-of-distribution ability. Our contributions advance the accuracy and generalization of VLM-based reaction diagram parsing. We will release data, models, and code on GitHub.

Method: RMG represents motion on a product manifold, factorizing motion into several manifold factors. It uses Riemannian flow matching with geodesic interpolation, tangent-space supervision, and manifold-preserving ODE integration for training and sampling.

Result: On HumanML3D, RMG achieves state-of-the-art FID (0.043) and ranks first on all reported metrics under MotionStreamer format. On MotionMillion, it surpasses strong baselines (FID 5.6, R@1 0.86). The compact translation+rotations representation proves most stable and effective.

Conclusion: Geometry-aware modeling on manifolds provides a practical and scalable route to high-fidelity motion generation, with the Riemannian approach outperforming Euclidean methods.

Abstract: Human motion generation is often learned in Euclidean spaces, although valid motions follow structured non-Euclidean geometry. We present Riemannian Motion Generation (RMG), a unified framework that represents motion on a product manifold and learns dynamics via Riemannian flow matching. RMG factorizes motion into several manifold factors, yielding a scale-free representation with intrinsic normalization, and uses geodesic interpolation, tangent-space supervision, and manifold-preserving ODE integration for training and sampling. On HumanML3D, RMG achieves state-of-the-art FID in the HumanML3D format (0.043) and ranks first on all reported metrics under the MotionStreamer format. On MotionMillion, it also surpasses strong baselines (FID 5.6, R@1 0.86). Ablations show that the compact $\mathscr{T}+\mathscr{R}$ (translation + rotations) representation is the most stable and effective, highlighting geometry-aware modeling as a practical and scalable route to high-fidelity motion generation.

Method: Proposes FreeOmniMVS with View-pair Correlation Transformer (VCT) to model pairwise correlation volumes across all camera view pairs, dropping unreliable pairs due to occlusion. Uses lightweight attention mechanism to adaptively fuse correlation vectors without designated reference view.

Result: Extensive experiments on diverse benchmark datasets demonstrate superiority for globally consistent, visibility-aware, and scale-aware omnidirectional depth estimation.

Conclusion: FreeOmniMVS provides a novel reference-free framework that achieves robust omnidirectional depth estimation through multi-view consistency maximization, addressing limitations of existing approaches.

Abstract: Reliable omnidirectional depth estimation from multi-fisheye stereo matching is pivotal to many applications, such as embodied robotics. Existing approaches either rely on spherical sweeping with heuristic fusion strategies to build the cost columns or perform reference-centric stereo matching based on rectified views. However, these methods fail to explicitly exploit geometric relationships between multiple views, rendering them less capable of capturing the global dependencies, visibility, or scale changes. In this paper, we shift to a new perspective and propose a novel reference-free framework, dubbed FreeOmniMVS, via multi-view consistency maximization. The highlight of FreeOmniMVS is that it can aggregate pair-wise correlations into a robust, visibility-aware, and global consensus. As such, it is tolerant to occlusions, partial overlaps, and varying baselines. Specifically, to achieve global coherence, we introduce a novel View-pair Correlation Transformer (VCT) that explicitly models pairwise correlation volumes across all camera view pairs, allowing us to drop unreliable pairs caused by occlusion or out-of-focus observations. To realize scalable and visibility-aware consensus, we propose a lightweight attention mechanism that adaptively fuses the correlation vectors, eliminating the need for a designated reference view and allowing all cameras to contribute equally to the stereo matching process. Extensive experiments on diverse benchmark datasets demonstrate the superiority of our method for globally consistent, visibility-aware, and scale-aware omnidirectional depth estimation.

Method: Proposes Uncertainty-Guided Manifold Smoothing (UMS) with a classifier to identify sub-manifolds and predict uncertainty scores, guiding diverse sample generation across the entire manifold. Uses dynamic global- and sub-manifold-driven architecture guided by the classifier to adapt to subdomain variations.

Result: Extensive experiments on public datasets validate effectiveness across different generation paradigms, showing improved reconstruction performance by bridging gaps between discrete sub-manifolds.

Conclusion: UMS framework effectively addresses limitations of existing NICT enhancement methods by creating a continuous and dense feature space through uncertainty-guided manifold smoothing and dynamic adaptation to scanning protocol variations.

Abstract: Non-ideal measurement computed tomography (NICT), which lowers radiation at the cost of image quality, is expanding the clinical use of CT. Although unified models have shown promise in NICT enhancement, most methods require paired data, which is an impractical demand due to inevitable organ motion. Unsupervised approaches attempt to overcome this limitation, but their assumption of homogeneous noise neglects the variability of scanning protocols, leading to poor generalization and potential model collapse. We further observe that distinct scanning protocols, which correspond to different physical imaging processes, produce discrete sub-manifolds in the feature space, contradicting these assumptions and limiting their effectiveness. To address this, we propose an Uncertainty-Guided Manifold Smoothing (UMS) framework to bridge the gaps between sub-manifolds. A classifier in UMS identifies sub-manifolds and predicts uncertainty scores, which guide the generation of diverse samples across the entire manifold. By leveraging the classifier’s capability, UMS effectively fills the gaps between discrete sub-manifolds, and promotes a continuous and dense feature space. Due to the complexity of the global manifold, it’s hard to directly model it. Therefore, we propose to dynamically incorporate the global- and sub-manifold-specific features. Specifically, we design a global- and sub-manifold-driven architecture guided by the classifier, which enables dynamic adaptation to subdomain variations. This dynamic mechanism improves the network’s capacity to capture both shared and domain-specific features, thereby improving reconstruction performance. Extensive experiments on public datasets are conducted to validate the effectiveness of our method across different generation paradigms.

Method: Cannot determine method as paper content is unavailable

Result: Cannot determine results as paper content is unavailable

Conclusion: Cannot determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.05295: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.05295&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: STALL is a simple, training-free detector that provides likelihood-based scoring for videos, jointly modeling spatial and temporal evidence within a probabilistic framework. It operates without synthetic data, scoring content against real-data statistics.

Result: STALL consistently outperforms prior image- and video-based baselines on two public benchmarks and a new benchmark called ComGenVid with state-of-the-art generative models.

Conclusion: STALL provides an effective zero-shot approach for synthetic video detection that addresses key limitations of existing methods through its training-free, probabilistic framework that jointly considers spatial and temporal evidence.

Abstract: Following major advances in text and image generation, the video domain has surged, producing highly realistic and controllable sequences. Along with this progress, these models also raise serious concerns about misinformation, making reliable detection of synthetic videos increasingly crucial. Image-based detectors are fundamentally limited because they operate per frame and ignore temporal dynamics, while supervised video detectors generalize poorly to unseen generators, a critical drawback given the rapid emergence of new models. These challenges motivate zero-shot approaches, which avoid synthetic data and instead score content against real-data statistics, enabling training-free, model-agnostic detection. We introduce \emph{STALL}, a simple, training-free, theoretically justified detector that provides likelihood-based scoring for videos, jointly modeling spatial and temporal evidence within a probabilistic framework. We evaluate STALL on two public benchmarks and introduce ComGenVid, a new benchmark with state-of-the-art generative models. STALL consistently outperforms prior image- and video-based baselines. Code and data are available at https://omerbenhayun.github.io/stall-video.

Method: Built on physical device environments with 201 mainstream apps across four device types. Uses two-level structure evaluating atomic abilities and realistic application-level performance across five dimensions. Data collected through multi-stage manual processes for authenticity.

Result: Experiments on 20 MLLMs and multi-agent systems show Qwen2.5-VL and UI-TARS perform competitively, but most models exhibit weaknesses in reflective decision-making and post-action self-evaluation, limiting real-world reliability.

Conclusion: GUI-CEval provides a comprehensive, interpretable benchmark to guide capability diagnosis and advance development of Chinese mobile GUI agents, addressing the gap in Chinese-centric multimodal evaluation frameworks.

Abstract: Recent progress in Multimodal Large Language Models (MLLMs) has enabled mobile GUI agents capable of visual perception, cross-modal reasoning, and interactive control. However, existing benchmarks are largely English-centric and fail to capture the linguistic and interaction characteristics of the Chinese mobile ecosystem. They also focus on isolated skills such as GUI grounding or offline agent, lacking a unified and fine-grained framework to assess the full capability chain from perception to execution. To address this gap, we introduce GUI-CEval, the first comprehensive benchmark for Chinese mobile GUI agents, built entirely on physical device environments. GUI-CEval spans 201 mainstream apps across four device types and adopts a two-level structure that evaluates both atomic abilities and realistic application-level performance along five dimensions: perception, planning, reflection, execution, and evaluation. All data are collected and verified through multi-stage manual processes to ensure authenticity and reproducibility. Extensive experiments on 20 representative MLLMs and multi-agent systems show that while models such as Qwen2.5-VL and UI-TARS perform competitively, most MLLMs still exhibit clear weaknesses in reflective decision-making and post-action self-evaluation, limiting their reliability in real-world interactions. We hope GUI-CEval provides a comprehensive and interpretable benchmark to guide capability diagnosis and advance the development of Chinese mobile GUI agents.

Method: Uses structured residual Fourier representations with ring-based frequency organization, learnable ring-wise spectral projections, and frequency band organization (low, mid, high) with cross-band interactions. Maps spectral features directly to detection scores without reconstruction errors.

Result: Extensive evaluation on FERET-Morph, FRLL-Morph, and MorDIFF datasets shows SRL-MAD consistently outperforms recent one-class and supervised MAD models.

Conclusion: Learning frequency-aware projections provides a more discriminative alternative to azimuthal spectral summarization for one-class morphing attack detection, enabling effective detection of unseen morphing attacks.

Abstract: Face morphing attacks represent a significant threat to biometric systems as they allow multiple identities to be combined into a single face. While supervised morphing attack detection (MAD) methods have shown promising performance, their reliance on attack-labeled data limits generalization to unseen morphing attacks. This has motivated increasing interest in one-class MAD, where models are trained exclusively on bona fide samples and are expected to detect unseen attacks as deviations from the normal facial structure. In this context, we introduce SRL-MAD, a one-class single-image MAD that uses structured residual Fourier representations for open-set morphing attack detection. Starting from a residual frequency map that suppresses image-specific spectral trends, we preserve the two-dimensional organization of the Fourier domain through a ring-based representation and replace azimuthal averaging with a learnable ring-wise spectral projection. To further encode domain knowledge about where morphing artifacts arise, we impose a frequency-informed inductive bias by organizing spectral evidence into low, mid, and high-frequency bands and learning cross-band interactions. These structured spectral features are mapped into a latent space designed for direct scoring, avoiding the reliance on reconstruction errors. Extensive evaluation on FERET-Morph, FRLL-Morph, and MorDIFF demonstrates that SRL-MAD consistently outperforms recent one-class and supervised MAD models. Overall, our results show that learning frequency-aware projections provides a more discriminative alternative to azimuthal spectral summarization for one-class morphing attack detection.

Method: Proposes an architecture that supervises facial embeddings using identity class labels, identity-relevant facial attributes, and non-identity-related attributes. Attributes are organized into interpretable groups to analyze individual contributions.

Result: Experiments on face verification benchmarks show: (1) using identity-relevant attribute subsets outperforms broader attribute sets, (2) explicitly unlearning non-identity-related attributes yields further gains, and (3) the method serves as a diagnostic tool for assessing encoder trustworthiness.

Conclusion: Joint learning of identity and facial attributes improves face embedding discriminability, with selective attribute supervision and explicit unlearning of non-identity attributes being key to better performance and interpretability.

Abstract: Despite recent advances in face recognition, robust performance remains challenging under large variations in age, pose, and occlusion. A common strategy to address these issues is to guide representation learning with auxiliary supervision from facial attributes, encouraging the visual encoder to focus on identity-relevant regions. However, existing approaches typically rely on heterogeneous and fixed sets of attributes, implicitly assuming equal relevance across attributes. This assumption is suboptimal, as different attributes exhibit varying discriminative power for identity recognition, and some may even introduce harmful biases. In this paper, we propose an attribute-aware face recognition architecture that supervises the learning of facial embeddings using identity class labels, identity-relevant facial attributes, and non-identity-related attributes. Facial attributes are organized into interpretable groups, making it possible to decompose and analyze their individual contributions in a human-understandable manner. Experiments on standard face verification benchmarks demonstrate that joint learning of identity and facial attributes improves the discriminability of face embeddings with two major conclusions: (i) using identity-relevant subsets of facial attributes consistently outperforms supervision with a broader attribute set, and (ii) explicitly forcing embeddings to unlearn non-identity-related attributes yields further performance gains compared to leaving such attributes unsupervised. Additionally, our method serves as a diagnostic tool for assessing the trustworthiness of face recognition encoders by allowing for the measurement of accuracy gains with suppression of non-identity-relevant attributes, with such gains suggesting shortcut learning from redundant attributes associated with each identity.

Method: Proposes a multimodal deep learning framework that integrates foundation model-based CT feature extraction with a missing-aware architecture for clinical variables, using weighted fusion to combine imaging and clinical modalities without conventional imputation strategies.

Result: The multimodal model consistently outperforms both unimodal imaging and clinical baselines, demonstrating the added value of integrating heterogeneous data sources for pR prediction.

Conclusion: The study highlights the potential of multimodal, missing-aware systems to support pathological response prediction under realistic clinical conditions with limited and incomplete data.

Abstract: Major pathological response (pR) following neoadjuvant therapy is a clinically meaningful endpoint in non-small cell lung cancer, strongly associated with improved survival. However, accurate preoperative prediction of pR remains challenging, particularly in real-world clinical settings characterized by limited data availability and incomplete clinical profiles. In this study, we propose a multimodal deep learning framework designed to address these constraints by integrating foundation model-based CT feature extraction with a missing-aware architecture for clinical variables. This approach enables robust learning from small cohorts while explicitly modeling missing clinical information, without relying on conventional imputation strategies. A weighted fusion mechanism is employed to leverage the complementary contributions of imaging and clinical modalities, yielding a multimodal model that consistently outperforms both unimodal imaging and clinical baselines. These findings underscore the added value of integrating heterogeneous data sources and highlight the potential of multimodal, missing-aware systems to support pR prediction under realistic clinical conditions.

Method: Proposes PAKAN framework with two adaptive variants: 2D Adaptive KAN that generates spline summation weights across spatial dimensions, and 1D Adaptive KAN that generates them across spectral channels. These are assembled into PAKAN 2to1 for feature fusion and PAKAN 1to1 for feature refinement.

Result: Extensive experiments demonstrate that the proposed modules significantly enhance network performance, proving the effectiveness and superiority of pixel-adaptive activation in pansharpening tasks.

Conclusion: The PAKAN framework with pixel-adaptive activation functions effectively addresses limitations of static activation functions in pansharpening, improving spatial-spectral fusion through dynamic adaptability.

Abstract: Pansharpening aims to fuse high-resolution spatial details from panchromatic images with the rich spectral information of multispectral images. Existing deep neural networks for this task typically rely on static activation functions, which limit their ability to dynamically model the complex, non-linear mappings required for optimal spatial-spectral fusion. While the recently introduced Kolmogorov-Arnold Network (KAN) utilizes learnable activation functions, traditional KANs lack dynamic adaptability during inference. To address this limitation, we propose a Pixel Adaptive Kolmogorov-Arnold Network framework. Starting from KAN, we design two adaptive variants: a 2D Adaptive KAN that generates spline summation weights across spatial dimensions and a 1D Adaptive KAN that generates them across spectral channels. These two components are then assembled into PAKAN 2to1 for feature fusion and PAKAN 1to1 for feature refinement. Extensive experiments demonstrate that our proposed modules significantly enhance network performance, proving the effectiveness and superiority of pixel-adaptive activation in pansharpening tasks.

Method: VAREX uses a Reverse Annotation pipeline that programmatically fills PDF templates with synthetic values to create deterministic ground truth. The benchmark includes 1,777 documents with 1,771 unique schemas across three structural categories, each provided in four input modalities: plain text, layout-preserving text, document image, or both text and image combined. Evaluation includes 20 models from frontier proprietary to small open models (≤4B parameters).

Result: Key findings: (1) Below 4B parameters, structured output compliance (not extraction capability) is the dominant bottleneck, with schema echo depressing scores by 45-65 percentage points; (2) Extraction-specific fine-tuning at 2B yields +81 pp gains; (3) Layout-preserving text provides the largest accuracy gain (+3-18 pp), exceeding pixel-level visual cues; (4) The benchmark most effectively discriminates models in the 60-95% accuracy band.

Conclusion: VAREX enables systematic evaluation of multimodal models on structured data extraction across different input modalities, revealing that layout-preserving text is more valuable than visual cues, and that instruction-following deficits in smaller models can be addressed through targeted fine-tuning rather than scaling model size.

Abstract: We introduce VAREX (VARied-schema EXtraction), a benchmark for evaluating multimodal foundation models on structured data extraction from government forms. VAREX employs a Reverse Annotation pipeline that programmatically fills PDF templates with synthetic values, producing deterministic ground truth validated through three-phase quality assurance. The benchmark comprises 1,777 documents with 1,771 unique schemas across three structural categories, each provided in four input modalities: plain text, layout-preserving text (whitespace-aligned to approximate column positions), document image, or both text and image combined. Unlike existing benchmarks that evaluate from a single input representation, VAREX provides four controlled modalities per document, enabling systematic ablation of how input format affects extraction accuracy – a capability absent from prior benchmarks. We evaluate 20 models from frontier proprietary models to small open models, with particular attention to models <=4B parameters suitable for cost-sensitive and latency-constrained deployment. Results reveal that (1) below 4B parameters, structured output compliance – not extraction capability – is a dominant bottleneck; in particular, schema echo (models producing schema-conforming structure instead of extracted values) depresses scores by 45-65 pp (percentage points) in affected models; (2) extraction-specific fine-tuning at 2B yields +81 pp gains, demonstrating that the instruction-following deficit is addressable without scale; (3) layout-preserving text provides the largest accuracy gain (+3-18 pp), exceeding pixel-level visual cues; and (4) the benchmark most effectively discriminates models in the 60-95% accuracy band. Dataset and evaluation code are publicly available.

Method: Unable to determine method due to technical error fetching paper content

Result: Unable to determine results due to technical error fetching paper content

Conclusion: Unable to draw conclusions due to technical error fetching paper content

Abstract: Failed to fetch summary for 2202.02543: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2202.02543&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Developed SAR-W-SimMIM, a weighted variant of SimMIM applied to ALOS-2 single-channel SAR imagery. Created a SAR dataset focused on Japan region using ALOS-2 HH polarization imagery. Used vision transformer-based autoencoder for pretraining, then fine-tuned encoder with task-specific decoder for semantic segmentation.

Result: Significant performance improvements in semantic segmentation compared to training from scratch with random initialization. SAR-W-SimMIM showed notable improvements over previous SAR-W-MixMAE approach. The method effectively reduces impact of speckle and extreme intensity values during pretraining.

Conclusion: Provides a guide for processing ALOS2 observations to create datasets suitable for self-supervised pretraining and fine-tuning downstream tasks. The approach enables development of region-specific foundation models for SAR imagery, addressing challenges of noise and imbalanced land cover distributions.

Abstract: Masked auto-encoders (MAE) and related approaches have shown promise for satellite imagery, but their application to synthetic aperture radar (SAR) remains limited due to challenges in semantic labeling and high noise levels. Building on our prior work with SAR-W-MixMAE, which adds SAR-specific intensity-weighted loss to standard MixMAE for pretraining, we also introduce SAR-W-SimMIM; a weighted variant of SimMIM applied to ALOS-2 single-channel SAR imagery. This method aims to reduce the impact of speckle and extreme intensity values during self-supervised pretraining. We evaluate its effect on semantic segmentation compared to our previous trial with SAR-W-MixMAE and random initialization, observing notable improvements. In addition, pretraining and fine-tuning models on satellite imagery pose unique challenges, particularly when developing region-specific models. Imbalanced land cover distributions such as dominant water, forest, or desert areas can introduce bias, affecting both pretraining and downstream tasks like land cover segmentation. To address this, we constructed a SAR dataset using ALOS-2 single-channel (HH polarization) imagery focused on the Japan region, marking the initial phase toward a national-scale foundation model. This dataset was used to pretrain a vision transformer-based autoencoder, with the resulting encoder fine-tuned for semantic segmentation using a task-specific decoder. Initial results demonstrate significant performance improvements compared to training from scratch with random initialization. In summary, this work provides a guide to process and prepare ALOS2 observations to create dataset so that it can be taken advantage of self-supervised pretraining of models and finetuning downstream tasks such as semantic segmentation.

Method: Define an explicit intermediate anchor frame during decoding that preserves scene geometry and semantic layout but discards high-frequency details. Use a pretrained video diffusion model (VDM) as temporal priors, treating the anchor as initial frame and original image as target frame, transforming decoding into next-frame prediction. This creates a visible, semantically faithful starting point for the generative process.

Result: Achieves over 50% bitrate savings across LPIPS, DISTS, FID, and KID metrics compared to DiffC on CLIC2020 test set. Also delivers up to 5x decoding speedup while improving both fidelity and realism for perceptual image compression.

Conclusion: The proposed paradigm successfully leverages temporal priors from video diffusion models for ultra-low-bitrate image compression, creating a more controlled decoding process that starts from a visible anchor frame, resulting in significant bitrate savings and faster decoding.

Abstract: We present a novel paradigm for ultra-low-bitrate image compression (ULB-IC) that exploits the ``temporal’’ evolution in generative image compression. Specifically, we define an explicit intermediate state during decoding: a compact anchor frame, which preserves the scene geometry and semantic layout while discarding high-frequency details. We then reinterpret generative decoding as a virtual temporal transition from this anchor to the final reconstructed image.To model this progression, we leverage a pretrained video diffusion model (VDM) as temporal priors: the anchor frame serves as the initial frame and the original image as the target frame, transforming the decoding process into a next-frame prediction task.In contrast to image diffusion-based ULB-IC models, our decoding proceeds from a visible, semantically faithful anchor, which improves both fidelity and realism for perceptual image compression. Extensive experiments demonstrate that our method achieves superior objective and subjective performance. On the CLIC2020 test set, our method achieves over \textbf{50% bitrate savings} across LPIPS, DISTS, FID, and KID compared to DiffC, while also delivering a significant decoding speedup of up to $\times$5. Code will be released later.

Method: Two-stage model: 1) Latent space decomposition with log transformation and 1-pixel offset to convert multiplicative relationship to additive formulation, 2) U-shaped component refiner with guidance fusion transformer blocks to refine reflectance and illumination components.

Result: Achieves competitive performance across four benchmark datasets for low-light enhancement with more stable training process compared to existing methods.

Conclusion: The proposed RGT model effectively addresses decomposition challenges in Retinex theory and provides stable, high-quality low-light image enhancement.

Abstract: Retinex theory provides a principled foundation for low-light image enhancement, inspiring numerous learning-based methods that integrate its principles. However, existing methods exhibits limitations in accurately decomposing reflectance and illumination components. To address this, we propose a Retinex-Guided Transformer~(RGT) model, which is a two-stage model consisting of decomposition and enhancement phases. First, we propose a latent space decomposition strategy to separate reflectance and illumination components. By incorporating the log transformation and 1-pixel offset, we convert the intrinsically multiplicative relationship into an additive formulation, enhancing decomposition stability and precision. Subsequently, we construct a U-shaped component refiner incorporating the proposed guidance fusion transformer block. The component refiner refines reflectance component to preserve texture details and optimize illumination distribution, effectively transforming low-light inputs to normal-light counterparts. Experimental evaluations across four benchmark datasets validate that our method achieves competitive performance in low-light enhancement and a more stable training process.

Method: Uses transformer architecture to leverage inherent permutation equivariance, allowing it to handle varying numbers of points per 3D instance, withstand occlusions, and generalize across object categories without requiring correspondence in training data.

Result: Demonstrates state-of-the-art performance across 2D-3D lifting task benchmarks and shows generalization to unseen categories, establishing the first 3D lifting foundation model.

Conclusion: The proposed 3D-LFM represents a significant advancement in 3D reconstruction by eliminating the need for correspondence in training data, enabling broader applicability across diverse object categories and real-world scenarios.

Abstract: The lifting of 3D structure and camera from 2D landmarks is at the cornerstone of the entire discipline of computer vision. Traditional methods have been confined to specific rigid objects, such as those in Perspective-n-Point (PnP) problems, but deep learning has expanded our capability to reconstruct a wide range of object classes (e.g. C3DPO and PAUL) with resilience to noise, occlusions, and perspective distortions. All these techniques, however, have been limited by the fundamental need to establish correspondences across the 3D training data – significantly limiting their utility to applications where one has an abundance of “in-correspondence” 3D data. Our approach harnesses the inherent permutation equivariance of transformers to manage varying number of points per 3D data instance, withstands occlusions, and generalizes to unseen categories. We demonstrate state of the art performance across 2D-3D lifting task benchmarks. Since our approach can be trained across such a broad class of structures we refer to it simply as a 3D Lifting Foundation Model (3D-LFM) – the first of its kind.

Method: WiT factorizes the continuous vector field using intermediate semantic waypoints projected from pre-trained vision models. It breaks optimal transport into prior-to-waypoint and waypoint-to-pixel segments. A lightweight generator dynamically infers waypoints during denoising, which condition the primary diffusion transformer via Just-Pixel AdaLN mechanism.

Result: WiT beats strong pixel-space baselines on ImageNet 256x256 and accelerates JiT training convergence by 2.2x.

Conclusion: WiT successfully disentangles pixel-space generation trajectories using semantic waypoints, offering an effective alternative to latent representations while improving training efficiency and performance.

Abstract: While recent Flow Matching models avoid the reconstruction bottlenecks of latent autoencoders by operating directly in pixel space, the lack of semantic continuity in the pixel manifold severely intertwines optimal transport paths. This induces severe trajectory conflicts near intersections, yielding sub-optimal solutions. Rather than bypassing this issue via information-lossy latent representations, we directly untangle the pixel-space trajectories by proposing Waypoint Diffusion Transformers (WiT). WiT factorizes the continuous vector field via intermediate semantic waypoints projected from pre-trained vision models. It effectively disentangles the generation trajectories by breaking the optimal transport into prior-to-waypoint and waypoint-to-pixel segments. Specifically, during the iterative denoising process, a lightweight generator dynamically infers these intermediate waypoints from the current noisy state. They then continuously condition the primary diffusion transformer via the Just-Pixel AdaLN mechanism, steering the evolution towards the next state, ultimately yielding the final RGB pixels. Evaluated on ImageNet 256x256, WiT beats strong pixel-space baselines, accelerating JiT training convergence by 2.2x. Code will be publicly released at https://github.com/hainuo-wang/WiT.git.

Method: Introduces DetectorContext, an abstraction for the Stone Soup multi-target tracking framework that exposes detection probability and clutter intensity as state-dependent functions evaluated during hypothesis formation. The abstraction integrates with existing probabilistic trackers without modifying their update equations.

Result: Experiments on asynchronous radar-lidar data show that context-aware modeling restores stable fusion and significantly improves HOTA and GOSPA performance without increasing false tracks.

Conclusion: State-dependent modeling of detection probability and clutter intensity through the DetectorContext abstraction effectively addresses the challenges of asynchronous multi-sensor tracking with partial coverage, improving tracking performance while maintaining compatibility with existing probabilistic trackers.

Abstract: Multi-sensor tracking in the real world involves asynchronous sensors with partial coverage and heterogeneous detection performance. Although probabilistic tracking methods permit detection probability and clutter intensity to depend on state and sensing context, many practical frameworks enforce globally uniform observability assumptions. Under multi-rate and partially overlapping sensing, this simplification causes repeated non-detections from high-rate sensors to erode tracks visible only to low-rate sensors, potentially degrading fusion performance. We introduce DetectorContext, an abstraction for the open-source multi-target tracking framework Stone Soup. DetectorContext exposes detection probability and clutter intensity as state-dependent functions evaluated during hypothesis formation. The abstraction integrates with existing probabilistic trackers without modifying their update equations. Experiments on asynchronous radar-lidar data demonstrate that context-aware modeling restores stable fusion and significantly improves HOTA and GOSPA performance without increasing false tracks.

Method: Proposes Stochastic Neighbor Cross Entropy Minimization (SNCE) which constructs soft categorical distributions over neighboring tokens based on proximity between code embeddings and ground-truth image embeddings, encouraging capture of semantic geometric structure.

Result: SNCE significantly improves convergence speed and generation quality across class-conditional ImageNet-256 generation, large-scale text-to-image synthesis, and image editing tasks compared to standard cross-entropy objectives.

Conclusion: SNCE effectively addresses optimization challenges of large-codebook discrete image generators by better leveraging the geometric structure of the quantized embedding space through soft neighbor-based supervision.

Abstract: Recent advancements in discrete image generation showed that scaling the VQ codebook size significantly improves reconstruction fidelity. However, training generative models with a large VQ codebook remains challenging, typically requiring larger model size and a longer training schedule. In this work, we propose Stochastic Neighbor Cross Entropy Minimization (SNCE), a novel training objective designed to address the optimization challenges of large-codebook discrete image generators. Instead of supervising the model with a hard one-hot target, SNCE constructs a soft categorical distribution over a set of neighboring tokens. The probability assigned to each token is proportional to the proximity between its code embedding and the ground-truth image embedding, encouraging the model to capture semantically meaningful geometric structure in the quantized embedding space. We conduct extensive experiments across class-conditional ImageNet-256 generation, large-scale text-to-image synthesis, and image editing tasks. Results show that SNCE significantly improves convergence speed and overall generation quality compared to standard cross-entropy objectives.

Method: Proposes TextOVSR with two text-guided branches: 1) negative branch uses degradation-descriptive text to constrain solution space, 2) positive branch uses content-descriptive text for semantic guidance. Includes Text-Enhanced Discriminator (TED) and Degradation-Robust Feature Fusion (DRF) module for cross-modal fusion while suppressing degradation interference.

Result: Outperforms state-of-the-art methods both qualitatively and quantitatively on the OperaLQ benchmark dataset.

Conclusion: Text-guided approach effectively addresses both degradation modeling and semantic guidance challenges in opera video super-resolution, achieving superior reconstruction quality.

Abstract: Many classic opera videos exhibit poor visual quality due to the limitations of early filming equipment and long-term degradation during storage. Although real-world video super-resolution (RWVSR) has achieved significant advances in recent years, directly applying existing methods to degraded opera videos remains challenging. The difficulties are twofold. First, accurately modeling real-world degradations is complex: simplistic combinations of classical degradation kernels fail to capture the authentic noise distribution, while methods that extract real noise patches from external datasets are prone to style mismatches that introduce visual artifacts. Second, current RWVSR methods, which rely solely on degraded image features, struggle to reconstruct realistic and detailed textures due to a lack of high-level semantic guidance. To address these issues, we propose a Text-guided Dual-Branch Opera Video Super-Resolution (TextOVSR) network, which introduces two types of textual prompts to guide the super-resolution process. Specifically, degradation-descriptive text, derived from the degradation process, is incorporated into the negative branch to constrain the solution space. Simultaneously, content-descriptive text is incorporated into a positive branch and our proposed Text-Enhanced Discriminator (TED) to provide semantic guidance for enhanced texture reconstruction. Furthermore, we design a Degradation-Robust Feature Fusion (DRF) module to facilitate cross-modal feature fusion while suppressing degradation interference. Experiments on our OperaLQ benchmark show that TextOVSR outperforms state-of-the-art methods both qualitatively and quantitatively. The code is available at https://github.com/ChangHua0/TextOVSR.

Method: Three-stage framework: 1) Lung-aware 3D expert combining original and lung-extracted CT volumes; 2) Two MedSigLIP-based experts for slice-wise representation/probability learning and Transformer-based inter-slice context modeling; 3) Source classifier to predict latent source identity, enabling source-aware model fusion and hierarchical voting.

Result: Stage 1 model achieves macro-F1 of 0.9711, ACC of 0.9712, and AUC of 0.9791 on validation set. Stage 2a and 2b achieve best AUC scores of 0.9864 and 0.9854 respectively. Stage 3 source classifier reaches 0.9107 ACC and 0.9114 F1, demonstrating effectiveness of source-aware expert modeling.

Conclusion: Source-aware expert modeling and hierarchical voting provide an effective solution for robust COVID-19 CT classification under heterogeneous multi-source conditions, addressing challenges of source shift, imbalance, and hidden identities in multi-institutional settings.

Abstract: Robust detection of COVID-19 from chest CT remains challenging in multi-institutional settings due to substantial source shift, source imbalance, and hidden test-source identities. In this work, we propose a three-stage source-aware multi-expert framework for multi-source COVID-19 CT classification. First, we build a lung-aware 3D expert by combining original CT volumes and lung-extracted CT volumes for volumetric classification. Second, we develop two MedSigLIP-based experts: a slice-wise representation and probability learning module, and a Transformer-based inter-slice context modeling module for capturing cross-slice dependency. Third, we train a source classifier to predict the latent source identity of each test scan. By leveraging the predicted source information, we perform model fusion and voting based on different experts. On the validation set covering all four sources, the Stage 1 model achieves the best macro-F1 of 0.9711, ACC of 0.9712, and AUC of 0.9791. Stage2a and Stage2b achieve the best AUC scores of 0.9864 and 0.9854, respectively. Stage~3 source classifier reaches 0.9107 ACC and 0.9114 F1. These results demonstrate that source-aware expert modeling and hierarchical voting provide an effective solution for robust COVID-19 CT classification under heterogeneous multi-source conditions.

Method: Proposes Distillation with Adaptive Intermediate Teacher transfer (DAIT), which introduces a trainable intermediate teacher that learns to transfer frozen VLM representations under explicit supervision from the target fine-grained task. This intermediate teacher adaptively enhances discriminative visual cues and produces compact, task-aligned knowledge for distillation.

Result: Achieves performance gains of 12.63% on FGVC-Aircraft and 8.34% on CUB-200-2011 datasets. Extensive evaluations on multiple FGVC benchmarks with diverse student architectures demonstrate DAIT’s effectiveness as a principled paradigm for transferring from general-purpose VLMs to deployable fine-grained recognition models.

Conclusion: DAIT provides an effective solution for transferring VLM capabilities to lightweight models for fine-grained visual categorization, addressing architectural misalignment and task-irrelevant information issues in conventional distillation approaches.

Abstract: Large-scale Vision-Language Models (VLMs) encode rich multimodal semantics that are highly beneficial for fine-grained visual categorization (FGVC). However, their prohibitive computational cost hinders practical deployment in resource-constrained environments. Although knowledge distillation contributes to transferring VLMs capacity to lightweight classifiers, conventional distillation mechanisms, which directly transfer from a generic VLM to a compact student, often yield suboptimal results due to severe architectural misalignment and introducing task-irrelevant information. To alleviate this limitation, we propose Distillation with Adaptive Intermediate Teacher transfer (DAIT) in this study, facilitating adaptive knowledge transfer from VLMs to lightweight students. DAIT introduces a trainable intermediate teacher that learns to transfer frozen VLMs representations under explicit supervision from the target fine-grained task. This intermediate teacher adaptively enhances discriminative visual cues, thereby producing compact and task-aligned knowledge that can be reliably distilled into lightweight models. Extensive evaluations on multiple FGVC benchmarks with diverse student architectures demonstrate that our method achieves respective performance gains of 12.63% and 8.34% on FGVC-Aircraft and CUB-200-2011 datasets, establishing DAIT as a principled paradigm for transferring from general-purpose VLMS to deployable fine-grained recognition models.

Method: Proposes Question-guided Visual Compression with Memory Feedback (QViC-MF) framework. Core component is Question-guided Multimodal Selective Attention (QMSA) that learns to preserve visual information related to given questions from both current clips and past related frames from memory. The compressor and memory feedback work iteratively for each clip of the entire video.

Result: Achieves significant improvements over state-of-the-art methods: 6.1% on MLVU test, 8.3% on LVBench, 18.3% on VNBench Long, and 3.7% on VideoMME Long benchmarks.

Conclusion: The simple yet effective feedback-driven design with question-guided multimodal selective attention yields large performance gains on long-term video understanding tasks, demonstrating the importance of establishing feedback between perception and memory rather than using one-way schemes.

Abstract: In the context of long-term video understanding with large multimodal models, many frameworks have been proposed. Although transformer-based visual compressors and memory-augmented approaches are often used to process long videos, they usually compress each frame independently and therefore fail to achieve strong performance on tasks that require understanding complete events, such as temporal ordering tasks in MLVU and VNBench. This motivates us to rethink the conventional one-way scheme from perception to memory, and instead establish a feedbackdriven process in which past visual contexts stored in the context memory can benefit ongoing perception. To this end, we propose Question-guided Visual Compression with Memory Feedback (QViC-MF), a framework for long-term video understanding. At its core is a Question-guided Multimodal Selective Attention (QMSA), which learns to preserve visual information related to the given question from both the current clip and the past related frames from the memory. The compressor and memory feedback work iteratively for each clip of the entire video. This simple yet effective design yields large performance gains on longterm video understanding tasks. Extensive experiments show that our method achieves significant improvement over current state-of-the-art methods by 6.1% on MLVU test, 8.3% on LVBench, 18.3% on VNBench Long, and 3.7% on VideoMME Long. The code will be released publicly.

Method: DART dynamically tracks dual prototypes (ID and OOD) in feature space to recover the drifting discriminative axis, with multi-layer fusion and flip correction for robustness during test-time online operation.

Result: Significant improvements: 15.32pp AUROC gain and 49.15pp FPR@95TPR reduction on ImageNet-C vs. Textures-C compared to baselines, across 15 corruption types at severity level 5.

Conclusion: Test-time discriminative axis tracking enables dependable OOD detection in dynamically changing environments, addressing the limitations of stationary distribution assumptions.

Abstract: For reliable deployment of deep-learning systems, out-of-distribution (OOD) detection is indispensable. In the real world, where test-time inputs often arrive as streaming mixtures of in-distribution (ID) and OOD samples under evolving covariate shifts, OOD samples are domain-constrained and bounded by the environment, and both ID and OOD are jointly affected by the same covariate factors. Existing methods typically assume a stationary ID distribution, but this assumption breaks down in such settings, leading to severe performance degradation. We empirically discover that, even under covariate shift, covariate-shifted ID (csID) and OOD (csOOD) samples remain separable along a discriminative axis in feature space. Building on this observation, we propose DART, a test-time, online OOD detection method that dynamically tracks dual prototypes – one for ID and the other for OOD – to recover the drifting discriminative axis, augmented with multi-layer fusion and flip correction for robustness. Extensive experiments on a wide range of challenging benchmarks, where all datasets are subjected to 15 common corruption types at severity level 5, demonstrate that our method significantly improves performance, yielding 15.32 percentage points (pp) AUROC gain and 49.15 pp FPR@95TPR reduction on ImageNet-C vs. Textures-C compared to established baselines. These results highlight the potential of the test-time discriminative axis tracking for dependable OOD detection in dynamically changing environments.

Method: Unable to determine method due to abstract fetch failure.

Result: Unable to determine results due to abstract fetch failure.

Conclusion: Unable to determine conclusion due to abstract fetch failure.

Abstract: Failed to fetch summary for 2501.05264: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2501.05264&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: HYDRA-TOK reformulates standard ViT backbones into progressive learners: Gen-ViT captures structure-preserving primitives for generation, Sem-ViT handles semantic encoding for understanding, connected by a Generation-Semantic Bottleneck that compresses features to filter noise then restores dimensionality for semantic comprehension.

Result: HYDRA achieves state-of-the-art performance: visual reconstruction (rFID 0.08), top-tier generation on GenEval (0.86), DPG-Bench (86.4), and WISE (0.53), while outperforming previous native UMMs by average 10.0 points across eight understanding benchmarks.

Conclusion: The progressive architecture with Generation-Semantic Bottleneck successfully bridges the gap between generation and understanding in unified multimodal models, establishing a new state-of-the-art framework for integrated perception and generation.

Abstract: Unified Multimodal Models struggle to bridge the fundamental gap between the abstract representations needed for visual understanding and the detailed primitives required for generation. Existing approaches typically compromise by employing decoupled encoders, stacking representation encoder atop VAEs, or utilizing discrete quantization. However, these methods often disrupt information coherence and lead to optimization conflicts. To this end, we introduce HYDRA-TOK, a representation-harmonized pure ViT in the insight that visual modeling should evolve from generation to understanding. HYDRA-TOK reformulates the standard backbone into a progressive learner that transitions from a Gen-ViT, which captures structure-preserving primitives, to a Sem-ViT for semantic encoding. Crucially, this transition is mediated by a Generation-Semantic Bottleneck (GSB), which compresses features into a low-dimensional space to filter noise for robust synthesis, then restores dimensionality to empower complex semantic comprehension. Built upon this foundation, we present HYDRA, a native unified framework integrating perception and generation within a single parameter space. Extensive experiments establish HYDRA as a new state-of-the-art. It sets a benchmark in visual reconstruction (rFID 0.08) and achieves top-tier generation performance on GenEval (0.86), DPG-Bench (86.4), and WISE (0.53), while simultaneously outperforming previous native UMMs by an average of 10.0 points across eight challenging understanding benchmarks.

Method: Cannot determine method without access to paper abstract

Result: Cannot determine results without access to paper abstract

Conclusion: Cannot draw conclusions without access to paper abstract

Abstract: Failed to fetch summary for 2501.18328: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2501.18328&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Introduces a physics-informed instruction tuning framework that explicitly encodes object properties, motion paradigms, and dynamic constraints into structured prompts, delivered through multi-turn dialogues to decompose causal reasoning into incremental steps

Result: Achieves 96.7% AUROC in video-level detection on Phys-AD benchmark, substantially outperforming prior SOTA (66.9%), and yields superior causal explanations (0.777 LLM score)

Conclusion: Structured physics priors can transform VLMs into reliable detectors of dynamic anomalies, bridging the gap between appearance-based reasoning and physics-grounded understanding

Abstract: Vision-Language Models (VLMs) demonstrate strong general-purpose reasoning but remain limited in physics-grounded anomaly detection, where causal understanding of dynamics is essential. Existing VLMs, trained predominantly on appearance-centric correlations, fail to capture kinematic constraints, leading to poor performance on anomalies such as irregular rotations or violated mechanical motions. We introduce a physics-informed instruction tuning framework that explicitly encodes object properties, motion paradigms, and dynamic constraints into structured prompts. By delivering these physical priors through multi-turn dialogues, our method decomposes causal reasoning into incremental steps, enabling robust internal representations of normal and abnormal dynamics. Evaluated on the Phys-AD benchmark, our approach achieves 96.7% AUROC in video-level detection–substantially outperforming prior SOTA (66.9%)–and yields superior causal explanations (0.777 LLM score). This work highlights how structured physics priors can transform VLMs into reliable detectors of dynamic anomalies.

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to draw conclusions due to missing paper content

Abstract: Failed to fetch summary for 2503.12778: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.12778&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Created HalDec-Bench with captions from diverse VLMs, human annotations for hallucination presence, detailed type categorization, and segment-level labels. Designed tasks with varying difficulty levels to reveal performance differences not visible in existing benchmarks.

Result: Benchmark reveals two key findings: 1) Detectors tend to recognize sentences at the beginning of responses as correct regardless of actual correctness, 2) Dataset noise can be substantially reduced using strong VLMs as filters with recent VLMs as caption generators.

Conclusion: HalDec-Bench provides a principled and interpretable framework for evaluating hallucination detectors, uncovering important model behaviors and offering practical solutions for improving training data quality in vision-language tasks.

Abstract: Hallucination detection in captions (HalDec) assesses a vision-language model’s ability to correctly align image content with text by identifying errors in captions that misrepresent the image. Beyond evaluation, effective hallucination detection is also essential for curating high-quality image-caption pairs used to train VLMs. However, the generalizability of VLMs as hallucination detectors across different captioning models and hallucination types remains unclear due to the lack of a comprehensive benchmark. In this work, we introduce HalDec-Bench, a benchmark designed to evaluate hallucination detectors in a principled and interpretable manner. HalDec-Bench contains captions generated by diverse VLMs together with human annotations indicating the presence of hallucinations, detailed hallucination-type categories, and segment-level labels. The benchmark provides tasks with a wide range of difficulty levels and reveals performance differences across models that are not visible in existing multimodal reasoning or alignment benchmarks. Our analysis further uncovers two key findings. First, detectors tend to recognize sentences appearing at the beginning of a response as correct, regardless of their actual correctness. Second, our experiments suggest that dataset noise can be substantially reduced by using strong VLMs as filters while employing recent VLMs as caption generators. Our project page is available at https://dahlian00.github.io/HalDec-Bench-Page/.

Method: IConE decouples collapse prevention from training batch size by maintaining a global set of learnable auxiliary instance embeddings regularized by an explicit diversity objective, transferring the anti-collapse mechanism from transient batch statistics to a dataset-level embedding space.

Result: IConE outperforms strong contrastive and non-contrastive baselines across diverse 2D and 3D biomedical modalities in the small-batch regime (B=1 to B=64), demonstrates marked robustness to severe class imbalance, and preserves high intrinsic dimensionality in learned representations.

Conclusion: IConE provides a novel approach to prevent representation collapse in self-supervised learning without relying on batch statistics, enabling effective training with small batch sizes and addressing practical constraints in scientific and medical applications.

Abstract: Self-supervised learning (SSL) has revolutionized representation learning, with Joint-Embedding Architectures (JEAs) emerging as an effective approach for capturing semantic features. Existing JEAs rely on implicit or explicit batch interaction – via negative sampling or statistical regularization – to prevent representation collapse. This reliance becomes problematic in regimes where batch sizes must be small, such as high-dimensional scientific data, where memory constraints and class imbalance make large, well-balanced batches infeasible. We introduce IConE (Instance-Contrasted Embeddings), a framework that decouples collapse prevention from the training batch size. Rather than enforcing diversity through batch statistics, IConE maintains a global set of learnable auxiliary instance embeddings regularized by an explicit diversity objective. This transfers the anti-collapse mechanism from the transient batch to a dataset-level embedding space, allowing stable training even when batch statistics are unreliable, down to batch size 1. Across diverse 2D and 3D biomedical modalities, IConE outperforms strong contrastive and non-contrastive baselines throughout the small-batch regime (from B=1 to B=64) and demonstrates marked robustness to severe class imbalance. Geometric analysis shows that IConE preserves high intrinsic dimensionality in the learned representations, preventing the collapse observed in existing JEAs as batch sizes shrink.

Method: Proposes “exemplar diffusion” - a training-free approach that leverages existing diffusion methods for object detection. The method uses known bounding boxes (exemplars) at test time to guide the diffusion process, improving detection accuracy. It also enables uncertainty quantification in diffusion detection methods.

Result: The method yields across-the-board increases in average precision and recall for medical image datasets with clear spatial structure. It shows robustness to exemplar quality, enabling non-expert annotation. Also demonstrates capability for quantifying predictive uncertainty in diffusion detection methods.

Conclusion: Exemplar diffusion provides an effective training-free approach to leverage existing labels at inference, improving object detection performance in medical images while offering uncertainty quantification capabilities. The method is robust to annotation quality and works well with structured medical datasets.

Abstract: We present a framework to take advantage of existing labels at inference, called \textit{exemplars}, in order to improve the performance of object detection in medical images. The method, \textit{exemplar diffusion}, leverages existing diffusion methods for object detection to enable a training-free approach to adding information of known bounding boxes at test time. We demonstrate that for medical image datasets with clear spatial structure, the method yields an across-the-board increase in average precision and recall, and a robustness to exemplar quality, enabling non-expert annotation. Moreover, we demonstrate how our method may also be used to quantify predictive uncertainty in diffusion detection methods. Source code and data splits openly available online: https://github.com/waahlstrand/ExemplarDiffusion

Method: Uses self-supervised pretrained DINO features from ImageNet, transfers them to in vivo confocal microscopy domain through careful fine-tuning. The model learns to focus on key morphological elements for grading without segmentation maps.

Result: Achieves 84.25% accuracy and 77.97% sensitivity, improving upon state-of-the-art methods. Demonstrates DINO’s effectiveness for medical imaging despite being superseded by newer versions.

Conclusion: Self-supervised pretrained features are transferable to medical imaging domains. DINO remains a viable model for medical imaging tasks, enabling accurate tortuosity grading without expensive segmentation.

Abstract: The tortuosity of corneal nerve fibers are used as indication for different diseases. Current state-of-the-art methods for grading the tortuosity heavily rely on expensive segmentation maps of these nerve fibers. In this paper, we demonstrate that self-supervised pretrained features from ImageNet are transferable to the domain of in vivo confocal microscopy. We show that DINO should not be disregarded as a deep learning model for medical imaging, although it was superseded by two later versions. After careful fine-tuning, DINO improves upon the state-of-the-art in terms of accuracy (84,25%) and sensitivity (77,97%). Our fine-tuned model focuses on the key morphological elements in grading without the use of segmentation maps.

Method: Systematic analysis reveals parameter specialization in unified models. FlashU introduces: 1) Task-Specific Network Pruning and Dynamic Layer Skipping to eliminate redundancy; 2) For generation: time-varying control signals and temporal approximation via Diffusion Head Cache; 3) For understanding: Dynamic Token Pruning via V-Norm Proxy to exploit spatial redundancy of visual inputs.

Result: Extensive experiments on Show-o2 demonstrate FlashU achieves 1.78× to 2.01× inference acceleration across both understanding and generation tasks while maintaining state-of-the-art performance, outperforming competing unified models.

Conclusion: FlashU validates a task-aware acceleration paradigm for unified multimodal models, showing that specialized optimization for different task types (generation vs understanding) can significantly reduce computational overhead without sacrificing performance.

Abstract: Native unified multimodal models, which integrate both generative and understanding capabilities, face substantial computational overhead that hinders their real-world deployment. Existing acceleration techniques typically employ a static, monolithic strategy, ignoring the fundamental divergence in computational profiles between iterative generation tasks (e.g., image generation) and single-pass understanding tasks (e.g., VQA). In this work, we present the first systematic analysis of unified models, revealing pronounced parameter specialization, where distinct neuron sets are critical for each task. This implies that, at the parameter level, unified models have implicitly internalized separate inference pathways for generation and understanding within a single architecture. Based on these insights, we introduce a training-free and task-aware acceleration framework, FlashU, that tailors optimization to each task’s demands. Across both tasks, we introduce Task-Specific Network Pruning and Dynamic Layer Skipping, aiming to eliminate inter-layer and task-specific redundancy. For visual generation, we implement a time-varying control signal for the guidance scale and a temporal approximation for the diffusion head via Diffusion Head Cache. For multimodal understanding, building upon the pruned model, we introduce Dynamic Token Pruning via a V-Norm Proxy to exploit the spatial redundancy of visual inputs. Extensive experiments on Show-o2 demonstrate that FlashU achieves 1.78$\times$ to 2.01$\times$ inference acceleration across both understanding and generation tasks while maintaining SOTA performance, outperforming competing unified models and validating our task-aware acceleration paradigm. Our code is publicly available at https://github.com/Rirayh/FlashU.

Method: Used MorphoMNIST (controlled toy dataset) and PadChest (public chest X-ray dataset) to evaluate multiple diversity metrics for images, text, and metadata. Assessed correlations between metrics, clinical expert intuition, downstream task performance (AUC), and training dynamics.

Result: Limited correlations between AUC and image/metadata reference-free diversity metrics, but higher correlations with FID and semantic diversity metrics. Clinical expert identified scanners as main practical diversity source, but adding another scanner led to shortcut learning.

Conclusion: Dataset diversity metrics show varying correlations with performance, with semantic diversity and FID being more predictive than reference-free metrics. Scanner diversity, while important in practice, can introduce shortcut learning issues.

Abstract: The diversity of training datasets is usually perceived as an important aspect to obtain a robust model. However, the definition of diversity is often not defined or differs across papers, and while some metrics exist, the quantification of this diversity is often overlooked when developing new algorithms. In this work, we study the behaviour of multiple dataset diversity metrics for image, text and metadata using MorphoMNIST, a toy dataset with controlled perturbations, and PadChest, a publicly available chest X-ray dataset. We evaluate whether these metrics correlate with each other but also with the intuition of a clinical expert. We also assess whether they correlate with downstream-task performance and how they impact the training dynamic of the models. We find limited correlations between the AUC and image or metadata reference-free diversity metrics, but higher correlations with the FID and the semantic diversity metrics. Finally, the clinical expert indicates that scanners are the main source of diversity in practice. However, we find that the addition of another scanner to the training set leads to shortcut learning. The code used in this study is available at https://github.com/TheoSourget/dataset_diversity_evaluation

Method: Uses masked representation-aligned Graph Attention Network (GAT) encoding with patch-level visual feature tokens as graph nodes. Employs self-attentional layers to encode complex local relations, enhanced with representation alignment in a learnable latent space. Defects are detected using Scaled Cosine Error (SCE) objective function on high reconstruction residual areas.

Result: Achieves state-of-the-art performance on MVTec AD, VisA, and MPDD benchmarks across 1- to 8-shot settings, with highest detection accuracy (up to 1.8% increase in image AUROC) and lowest inference latency (at least 25.05% faster) compared to best literature methods.

Conclusion: GATE-AD provides an effective and efficient solution for few-shot industrial visual anomaly detection, combining graph attention networks with representation alignment for robust defect identification with minimal training data.

Abstract: Few-Shot Industrial Visual Anomaly Detection (FS-IVAD) comprises a critical task in modern manufacturing settings, where automated product inspection systems need to identify rare defects using only a handful of normal/defect-free training samples. In this context, the current study introduces a novel reconstruction-based approach termed GATE-AD. In particular, the proposed framework relies on the employment of a masked, representation-aligned Graph Attention Network (GAT) encoding scheme to learn robust appearance patterns of normal samples. By leveraging dense, patch-level, visual feature tokens as graph nodes, the model employs stacked self-attentional layers to adaptively encode complex, irregular, non-Euclidean, local relations. The graph is enhanced with a representation alignment component grounded on a learnable, latent space, where high reconstruction residual areas (i.e., defects) are assessed using a Scaled Cosine Error (SCE) objective function. Extensive comparative evaluation on the MVTec AD, VisA, and MPDD industrial defect detection benchmarks demonstrates that GATE-AD achieves state-of-the-art performance across the $1$- to $8$-shot settings, combining the highest detection accuracy (increase up to $1.8%$ in image AUROC in the 8-shot case in MPDD) with the lowest per-image inference latency (at least $25.05%$ faster), compared to the best-performing literature methods. In order to facilitate reproducibility and further research, the source code of GATE-AD is available at https://github.com/gthpapadopoulos/GATE-AD.

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to determine conclusion due to access error

Abstract: Failed to fetch summary for 2505.16643: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.16643&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: GVC1D encodes video into extreme compact 1D latent tokens conditioned on both short- and long-term contexts. The 1D representation allows adaptive attention to semantic regions, facilitates token reduction to reduce spatial redundancy, and uses a 1D memory mechanism to provide semantically rich long-term context with low computational cost.

Result: Achieves superior compression efficiency with bitrate reductions of 60.4% under LPIPS and 68.8% under DISTS on the HEVC Class B dataset, surpassing previous video compression methods.

Conclusion: The 1D latent representation paradigm effectively addresses spatial and temporal redundancy challenges in video compression, offering significant improvements in compression efficiency over traditional 2D approaches.

Abstract: Recent advancements in generative video codec (GVC) typically encode video into a 2D latent grid and employ high-capacity generative decoders for reconstruction. However, this paradigm still leaves two key challenges in fully exploiting spatial-temporal redundancy: Spatially, the 2D latent grid inevitably preserves intra-frame redundancy due to its rigid structure, where adjacent patches remain highly similar, thereby necessitating a higher bitrate. Temporally, the 2D latent grid is less effective for modeling long-term correlations in a compact and semantically coherent manner, as it hinders the aggregation of common contents across frames. To address these limitations, we introduce Generative Video Compression with One-Dimensional (1D) Latent Representation (GVC1D). GVC1D encodes the video data into extreme compact 1D latent tokens conditioned on both short- and long-term contexts. Without the rigid 2D spatial correspondence, these 1D latent tokens can adaptively attend to semantic regions and naturally facilitate token reduction, thereby reducing spatial redundancy. Furthermore, the proposed 1D memory provides semantically rich long-term context while maintaining low computational cost, thereby further reducing temporal redundancy. Experimental results indicate that GVC1D attains superior compression efficiency, where it achieves bitrate reductions of 60.4% under LPIPS and 68.8% under DISTS on the HEVC Class B dataset, surpassing the previous video compression methods.Project: https://gvc1d.github.io/

Method: Created a synthetic stereo dataset using Unreal Engine 5 with 115 photogrammetry-scanned trees from Quixel Megascans library. Simulated a stereo rig matching ZED Mini camera specifications (63 mm baseline, 2.8 mm focal length, 3.84 mm sensor width). Captured 5,520 rectified 1920x1080 stereo pairs by orbiting each tree at up to 2m across three elevation bands (horizontal, +45°, -45°) with pixel-perfect disparity labels.

Result: Produced a comprehensive synthetic dataset with statistical characterization covering disparity distributions, scene diversity, and visual fidelity. Qualitative comparison with real-world Canterbury Tree Branches imagery confirmed photorealistic quality and geometric plausibility of rendered data.

Conclusion: UE5-Forest provides a ready-to-use benchmark and training resource for stereo-based forestry depth estimation, addressing the critical data gap for training supervised stereo matching networks in complex forest environments.

Abstract: Dense ground-truth disparity maps are practically unobtainable in forestry environments, where thin overlapping branches and complex canopy geometry defeat conventional depth sensors – a critical bottleneck for training supervised stereo matching networks for autonomous UAV-based pruning. We present UE5-Forest, a photorealistic synthetic stereo dataset built entirely in Unreal Engine 5 (UE5). One hundred and fifteen photogrammetry-scanned trees from the Quixel Megascans library are placed in virtual scenes and captured by a simulated stereo rig whose intrinsics – 63 mm baseline, 2.8 mm focal length, 3.84 mm sensor width – replicate the ZED Mini camera mounted on our drone. Orbiting each tree at up to 2 m across three elevation bands (horizontal, +45 degrees, -45 degrees) yields 5,520 rectified 1920 x 1080 stereo pairs with pixel-perfect disparity labels. We provide a statistical characterisation of the dataset – covering disparity distributions, scene diversity, and visual fidelity – and a qualitative comparison with real-world Canterbury Tree Branches imagery that confirms the photorealistic quality and geometric plausibility of the rendered data. The dataset will be publicly released to provide the community with a ready-to-use benchmark and training resource for stereo-based forestry depth estimation.

Method: MeMix recasts recurrent state into a Memory Mixture by partitioning it into multiple independent memory patches and updating only the least-aligned memory patches while exactly preserving others, requiring no fine-tuning or additional parameters.

Result: Across standard benchmarks (ScanNet, 7-Scenes, KITTI), MeMix reduces reconstruction completeness error by 15.3% on average (up to 40.0%) across 300-500 frame streams on 7-Scenes while maintaining O(1) inference memory.

Conclusion: MeMix effectively addresses catastrophic forgetting in streaming 3D reconstruction through selective memory updates, offering a practical training-free solution that can be directly applied to existing recurrent models.

Abstract: Reconstruction is a fundamental task in 3D vision and a fundamental capability for spatial intelligence. Particularly, streaming 3D reconstruction is central to real-time spatial perception, yet existing recurrent online models often suffer from progressive degradation on long sequences due to state drift and forgetting, motivating inference-time remedies. We present MeMix, a training-free, plug-and-play module that improves streaming reconstruction by recasting the recurrent state into a Memory Mixture. MeMix partitions the state into multiple independent memory patches and updates only the least-aligned memory patches while exactly preserving others. This selective update mitigates catastrophic forgetting while retaining $O(1)$ inference memory, and requires no fine-tuning or additional learnable parameters, making it directly applicable to existing recurrent reconstruction models. Across standard benchmarks (ScanNet, 7-Scenes, KITTI, etc.), under identical backbones and inference settings, MeMix reduces reconstruction completeness error by 15.3% on average (up to 40.0%) across 300–500 frame streams on 7-Scenes. The code is available at https://dongjiacheng06.github.io/MeMix/

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusions due to access restrictions

Abstract: Failed to fetch summary for 2505.24786: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.24786&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: ODIS axially translates a disperser between conjugate image plane and defocused position to capture panchromatic (PAN) image and dispersed measurement sequentially. PDAUN network uses PAN-guided deep unfolding with FFT-Woodbury preconditioned solver exploiting cyclic-convolution properties, and Dispersion-Aware Deformable Convolution for spectral alignment correction.

Result: State-of-the-art performance on standard benchmarks, decisive gains under low illumination in cross-system comparisons, and high-fidelity reconstruction validated on physical prototype.

Conclusion: ODIS achieves near-full light throughput for computational spectral imaging, enabling high-quality reconstruction under low-light conditions through novel optical design and deep learning reconstruction.

Abstract: Existing computational spectral imaging systems typically rely on coded aperture and beam splitters that block a substantial fraction of incident light, degrading reconstruction quality under light-starved conditions. To address this limitation, we develop the Oscillating Dispersion Imaging Spectrometer (ODIS), which for the first time achieves near-full light throughput by axially translating a disperser between the conjugate image plane and a defocused position, sequentially capturing a panchromatic (PAN) image and a dispersed measurement along a single optical path. We further propose a PAN-guided Dispersion-Aware Deep Unfolding Network (PDAUN) that recovers high-fidelity spectral information from maskless dispersion under PAN structural guidance. Its data-fidelity step derives an FFT-Woodbury preconditioned solver by exploiting the cyclic-convolution property of the ODIS forward model, while a Dispersion-Aware Deformable Convolution module (DADC) corrects sub-pixel spectral misalignment using PAN features. Experiments show state-of-the-art performance on standard benchmarks, and cross-system comparisons confirm that ODIS yields decisive gains under low illumination. High-fidelity reconstruction is validated on a physical prototype.

Method: Proposes PCDC framework integrating conditional diffusion decoder with PPO-based block-wise bitrate allocation strategy. Also releases high-resolution drone image dataset of coastal urban residential areas.

Result: Achieves compression ratios of 19.3x on DIV2K and 21.2x on drone dataset. Downstream object detection shows reconstructed images preserve task-relevant information with negligible performance loss.

Conclusion: PCDC effectively compresses high-resolution drone imagery while maintaining perceptual quality and preserving information for downstream vision tasks like object detection.

Abstract: Existing remote sensing image compression methods still explore to balance high compression efficiency with the preservation of fine details and task-relevant information. Meanwhile, high-resolution drone imagery offers valuable structural details for urban monitoring and disaster assessment, but large-area datasets can easily reach hundreds of gigabytes, creating significant challenges for storage and long-term management. In this paper, we propose a PPO-based bitrate allocation Conditional Diffusion Compression (PCDC) framework. PCDC integrates a conditional diffusion decoder with a PPO-based block-wise bitrate allocation strategy to achieve high compression ratios while maintaining strong perceptual performance. We also release a high-resolution drone image dataset with richer structural details at a consistent low altitude over residential neighborhoods in coastal urban areas. Experimental results show compression ratios of 19.3x on DIV2K and 21.2x on the drone image dataset. Moreover, downstream object detection experiments demonstrate that the reconstructed images preserve task-relevant information with negligible performance loss.

Method: IRIS introduces an analytical sampling strategy that precisely identifies ray-primitive intersections to eliminate empty space processing, and a continuous feature aggregation mechanism that operates directly along rays by interpolating latent attributes from sorted intersections, bypassing costly 3D searches.

Result: The method achieves high-fidelity, real-time rendering and enables flexible shape editing while ensuring geometric consistency, overcoming the computational bottlenecks of previous approaches.

Conclusion: IRIS provides an efficient framework for interactive scene editing that combines the benefits of neural radiance fields and 3D Gaussian splatting through novel analytical sampling and feature aggregation techniques.

Abstract: Neural Radiance Fields achieve high-fidelity scene representation but suffer from costly training and rendering, while 3D Gaussian splatting offers real-time performance with strong empirical results. Recently, solutions that harness the best of both worlds by using Gaussians as proxies to guide neural field evaluations, still suffer from significant computational inefficiencies. They typically rely on stochastic volumetric sampling to aggregate features, which severely limits rendering performance. To address this issue, a novel framework named IRIS (Intersection-aware Ray-based Implicit Editable Scenes) is introduced as a method designed for efficient and interactive scene editing. To overcome the limitations of standard ray marching, an analytical sampling strategy is employed that precisely identifies interaction points between rays and scene primitives, effectively eliminating empty space processing. Furthermore, to address the computational bottleneck of spatial neighbor lookups, a continuous feature aggregation mechanism is introduced that operates directly along the ray. By interpolating latent attributes from sorted intersections, costly 3D searches are bypassed, ensuring geometric consistency, enabling high-fidelity, real-time rendering, and flexible shape editing. Code can be found at https://github.com/gwilczynski95/iris.

Method: NavGRPO uses Group Relative Policy Optimization, a reinforcement learning framework that explores diverse trajectories and optimizes via within-group performance comparisons. This allows agents to distinguish effective strategies beyond expert paths without needing additional value networks.

Result: Achieves +3.0% and +1.71% SPL improvements on R2R and REVERIE benchmarks in unseen environments. Under extreme early-stage perturbations, demonstrates +14.89% SPL gain over baseline, showing substantially more robust navigation policies.

Conclusion: Goal-directed reinforcement learning training builds substantially more robust navigation policies than imitation learning approaches, with NavGRPO demonstrating strong performance improvements on standard benchmarks and under challenging perturbation scenarios.

Abstract: Vision-and-Language Navigation (VLN) requires agents to navigate photo-realistic environments following natural language instructions. Current methods predominantly rely on imitation learning, which suffers from limited generalization and poor robustness to execution perturbations. We present NavGRPO, a reinforcement learning framework that learns goal-directed navigation policies through Group Relative Policy Optimization. By exploring diverse trajectories and optimizing via within-group performance comparisons, our method enables agents to distinguish effective strategies beyond expert paths without requiring additional value networks. Built on ScaleVLN, NavGRPO achieves superior robustness on R2R and REVERIE benchmarks with +3.0% and +1.71% SPL improvements in unseen environments. Under extreme early-stage perturbations, we demonstrate +14.89% SPL gain over the baseline, confirming that goal-directed RL training builds substantially more robust navigation policies. Code and models will be released.

Method: Parameter-efficient adaptation framework with adaptive spectral rectification module using learnable wavelet decomposition to explicitly model and amplify attenuated high-frequency components in feature maps, preserving pre-trained geometric representations.

Result: Achieved state-of-the-art performance on C3VD and SimCol3D datasets with absolute relative errors of 0.022 and 0.027 respectively.

Conclusion: Directly addressing spectral mismatches is highly effective for adapting vision foundation models to specialized medical imaging tasks.

Abstract: Accurate monocular depth estimation is critical in colonoscopy for lesion localization and navigation. Foundation models trained on natural images fail to generalize directly to colonoscopy. We identify the core issue not as a semantic gap, but as a statistical shift in the frequency domain: colonoscopy images lack the strong high-frequency edge and texture gradients that these models rely on for geometric reasoning. To address this, we propose SpecDepth, a parameter-efficient adaptation framework that preserves the robust geometric representations of the pre-trained models while adapting to the colonoscopy domain. Its key innovation is an adaptive spectral rectification module, which uses a learnable wavelet decomposition to explicitly model and amplify the attenuated high-frequency components in feature maps. Different from conventional fine-tuning that risks distorting high-level semantic features, this targeted, low-level adjustment realigns the input signal with the original inductive bias of the foundational model. On the public C3VD and SimCol3D datasets, SpecDepth achieved state-of-the-art performance with an absolute relative error of 0.022 and 0.027, respectively. Our work demonstrates that directly addressing spectral mismatches is a highly effective strategy for adapting vision foundation models to specialized medical imaging tasks. The code will be released publicly after the manuscript is accepted for publication.

Method: Introduces AVA-Bench which explicitly disentangles 14 Atomic Visual Abilities (AVAs) like localization, depth estimation, and spatial understanding. The benchmark decouples these abilities and matches training and test distributions within each ability category to isolate performance.

Result: AVA-Bench reveals distinctive “ability fingerprints” for leading VFMs, enabling principled model selection. Also found that a 0.5B LLM yields similar VFM rankings as a 7B LLM while cutting GPU hours by 8x.

Conclusion: AVA-Bench provides a comprehensive and transparent benchmark for evaluating vision foundation models, laying the foundation for next-generation VFM development by enabling precise diagnosis of visual capabilities.

Abstract: The rise of vision foundation models (VFMs) calls for systematic evaluation. A common approach pairs VFMs with large language models (LLMs) as general-purpose heads, followed by evaluation on broad Visual Question Answering (VQA) benchmarks. However, this protocol has two key blind spots: (i) the instruction tuning data may not align with VQA test distributions, meaning a wrong prediction can stem from such data mismatch rather than a VFM’ visual shortcomings; (ii) VQA benchmarks often require multiple visual abilities, making it hard to tell whether errors stem from lacking all required abilities or just a single critical one. To address these gaps, we introduce AVA-Bench, the first benchmark that explicitly disentangles 14 Atomic Visual Abilities (AVAs) – foundational skills like localization, depth estimation, and spatial understanding that collectively support complex visual reasoning tasks. By decoupling AVAs and matching training and test distributions within each, AVA-Bench pinpoints exactly where a VFM excels or falters. Applying AVA-Bench to leading VFMs thus reveals distinctive “ability fingerprints,” turning VFM selection from educated guesswork into principled engineering. Notably, we find that a 0.5B LLM yields similar VFM rankings as a 7B LLM while cutting GPU hours by 8x, enabling more efficient evaluation. By offering a comprehensive and transparent benchmark, we hope AVA-Bench lays the foundation for the next generation of VFMs.

Method: Integrates multiple SOTA methods: object detection, body pose estimation, monocular depth estimation, and vision-language models. Uses 3D spatial information from single images and image captioning models to correct classification errors.

Result: Experimental results on a custom dataset show that incorporating depth information significantly improves target identification, especially in complex scenes with overlapping objects.

Conclusion: The modular approach enables deployment in environments without specialized depth sensors, providing an effective solution for recognizing pointing gesture targets using only RGB images.

Abstract: This paper presents a comprehensive pipeline for recognizing objects targeted by human pointing gestures using RGB images. As human-robot interaction moves toward more intuitive interfaces, the ability to identify targets of non-verbal communication becomes crucial. Our proposed system integrates several existing state-of-the-art methods, including object detection, body pose estimation, monocular depth estimation, and vision-language models. We evaluate the impact of 3D spatial information reconstructed from a single image and the utility of image captioning models in correcting classification errors. Experimental results on a custom dataset show that incorporating depth information significantly improves target identification, especially in complex scenes with overlapping objects. The modularity of the approach allows for deployment in environments where specialized depth sensors are unavailable.

Method: Proposes AnyCrowd framework with: 1) Instance-Isolated Latent Representation (IILR) to encode characters independently before DiT processing; 2) Tri-Stage Decoupled Attention (TSDA) that decomposes self-attention into instance-aware foreground attention, background-centric interaction, and global foreground-background coordination; 3) Adaptive Gated Fusion (AGF) module to handle overlapping regions by predicting identity-aware weights.

Result: The framework can scale to arbitrary number of characters while maintaining identity consistency and preventing identity bleeding. It achieves precise binding between reference identities and driving pose sequences with improved spatio-temporal consistency.

Conclusion: AnyCrowd addresses key challenges in multi-character animation through disentangled representations and attention mechanisms, enabling scalable generation of identity-consistent multi-character animations.

Abstract: Controllable character animation has advanced rapidly in recent years, yet multi-character animation remains underexplored. As the number of characters grows, multi-character reference encoding becomes more susceptible to latent identity entanglement, resulting in identity bleeding and reduced controllability. Moreover, learning precise and spatio-temporally consistent correspondences between reference identities and driving pose sequences becomes increasingly challenging, often leading to identity-pose mis-binding and inconsistency in generated videos. To address these challenges, we propose AnyCrowd, a Diffusion Transformer (DiT)-based video generation framework capable of scaling to an arbitrary number of characters. Specifically, we first introduce an Instance-Isolated Latent Representation (IILR), which encodes character instances independently prior to DiT processing to prevent latent identity entanglement. Building on this disentangled representation, we further propose Tri-Stage Decoupled Attention (TSDA) to bind identities to driving poses by decomposing self-attention into: (i) instance-aware foreground attention, (ii) background-centric interaction, and (iii) global foreground-background coordination. Furthermore, to mitigate token ambiguity in overlapping regions, an Adaptive Gated Fusion (AGF) module is integrated within TSDA to predict identity-aware weights, effectively fusing competing token groups into identity-consistent representations…

Method: Created Gym-V, a platform with 179 procedurally generated visual environments across 10 domains with controllable difficulty. Used this infrastructure to conduct experiments on observation scaffolding (captions, game rules), RL algorithm choices, and cross-domain transfer effects.

Result: Found that observation scaffolding is more decisive for training success than RL algorithm choice, with captions and game rules determining whether learning succeeds at all. Cross-domain transfer experiments show diverse task training generalizes broadly while narrow training can cause negative transfer, with multi-turn interaction amplifying these effects.

Conclusion: Gym-V provides essential infrastructure for systematic study of vision agents, revealing critical insights about observation scaffolding and transfer learning. The platform aims to accelerate future research on agentic vision-language models by providing standardized environments and evaluation toolkits.

Abstract: As agentic systems increasingly rely on reinforcement learning from verifiable rewards, standardized ``gym’’ infrastructure has become essential for rapid iteration, reproducibility, and fair comparison. Vision agents lack such infrastructure, limiting systematic study of what drives their learning and where current models fall short. We introduce \textbf{Gym-V}, a unified platform of 179 procedurally generated visual environments across 10 domains with controllable difficulty, enabling controlled experiments that were previously infeasible across fragmented toolkits. Using it, we find that observation scaffolding is more decisive for training success than the choice of RL algorithm, with captions and game rules determining whether learning succeeds at all. Cross-domain transfer experiments further show that training on diverse task categories generalizes broadly while narrow training can cause negative transfer, with multi-turn interaction amplifying all of these effects. Gym-V is released as a convenient foundation for training environments and evaluation toolkits, aiming to accelerate future research on agentic VLMs.

Method: Reformulate multi-modal CoT reasoning as a KL-constrained reward maximization focused on rationale-conditional log-likelihood. Propose RED as a plug-and-play inference-time decoding strategy that multiplies distinct image-conditional and rationale-conditional next token distributions to harmonize visual and rationale information.

Result: RED consistently and significantly improves reasoning over standard CoT and other decoding methods across multiple benchmarks and LVLMs. The approach enhances both faithfulness and accuracy of CoT reasoning in vision-language models.

Conclusion: RED offers a practical and effective approach to improve rationale-grounded multi-modal systems, making chain-of-thought reasoning more reliable in large vision-language models through better integration of visual and rationale information.

Abstract: Large vision-language models (LVLMs) have demonstrated remarkable capabilities by integrating pre-trained vision encoders with large language models (LLMs). Similar to single-modal LLMs, chain-of-thought (CoT) prompting has been adapted for LVLMs to enhance multi-modal reasoning by generating intermediate rationales based on visual and textual inputs. While CoT is assumed to improve grounding and accuracy in LVLMs, our experiments reveal a key challenge: existing LVLMs often ignore the contents of generated rationales in CoT reasoning. To address this, we re-formulate multi-modal CoT reasoning as a KL-constrained reward maximization focused on rationale-conditional log-likelihood. As the optimal solution, we propose rationale-enhanced decoding (RED), a novel plug-and-play inference-time decoding strategy. RED harmonizes visual and rationale information by multiplying distinct image-conditional and rationale-conditional next token distributions. Extensive experiments show that RED consistently and significantly improves reasoning over standard CoT and other decoding methods across multiple benchmarks and LVLMs. Our work offers a practical and effective approach to improve both the faithfulness and accuracy of CoT reasoning in LVLMs, paving the way for more reliable rationale-grounded multi-modal systems. Code is available at https://github.com/yshinya6/red/.

Method: Proposes PrismMirror with cascade learning strategy: first learns coarse geometric features (SMPL-X meshes and point clouds), then refines textures through rendering supervision. Distills this unified framework into a lightweight linear attention model for real-time efficiency.

Result: Achieves real-time inference at 24 FPS, significantly outperforming previous methods in both visual authenticity and structural accuracy. First monocular human frontal view synthesis model to reach real-time performance.

Conclusion: PrismMirror successfully addresses the trade-off between visual fidelity and geometric understanding in human view synthesis, enabling real-time 3D telepresence from single images without complex multi-camera setups.

Abstract: Photorealistic human novel view synthesis from a single image is crucial for democratizing immersive 3D telepresence, eliminating the need for complex multi-camera setups. However, current rendering-centric methods prioritize visual fidelity over explicit geometric understanding and struggle with intricate regions like faces and hands, leading to temporal instability. Meanwhile, human-centric frameworks suffer from memory bottlenecks since they typically rely on an auxiliary model to provide informative structural priors for geometric modeling, which limits real-time performance. To address these challenges, we propose PrismMirror, a geometry-guided framework for instant frontal view synthesis from a single image. By avoiding external geometric modeling and focusing on frontal view synthesis, our model optimizes visual integrity for telepresence. Specifically, PrismMirror introduces a novel cascade learning strategy that enables coarse-to-fine geometric feature learning. It first directly learns coarse geometric features, such as SMPL-X meshes and point clouds, and then refines textures through rendering supervision. To achieve real-time efficiency, we distill this unified framework into a lightweight linear attention model. Notably, PrismMirror is the first monocular human frontal view synthesis model that achieves real-time inference at 24 FPS, significantly outperforming previous methods in both visual authenticity and structural accuracy.

Method: Unable to determine method due to fetch failure

Result: Unable to determine results due to fetch failure

Conclusion: Unable to determine conclusion due to fetch failure

Abstract: Failed to fetch summary for 2507.13984: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.13984&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes MV2UV that combines 2D generative priors from multiview generation with UV refinement inpainting. Uses a UV space generative model that simultaneously inpaints unseen parts of multiview images while resolving multiview inconsistency.

Result: Experiments show the method enables better texture generation quality than existing methods, especially in unseen occluded and multiview-inconsistent parts.

Conclusion: MV2UV effectively combines strengths of both multiview generation and UV refinement approaches to produce high-quality, consistent texture maps for 3D assets.

Abstract: Generating high-quality textures for 3D assets is a challenging task. Existing multiview texture generation methods suffer from the multiview inconsistency and missing textures on unseen parts, while UV inpainting texture methods do not generalize well due to insufficient UV data and cannot well utilize 2D image diffusion priors. In this paper, we propose a new method called MV2UV that combines 2D generative priors from multiview generation and the inpainting ability of UV refinement to get high-quality texture maps. Our key idea is to adopt a UV space generative model that simultaneously inpaints unseen parts of multiview images while resolving the inconsistency of multiview images. Experiments show that our method enables a better texture generation quality than existing methods, especially in unseen occluded and multiview-inconsistent parts.

Method: Introduces EscapeCraft-4D, a customizable 4D environment incorporating trigger-based auditory sources, temporally transient evidence, and location-dependent cues. It requires agents to perform spatio-temporal reasoning and proactive multimodal integration under time constraints. A benchmark is curated to evaluate these abilities across powerful models.

Result: Evaluation results show models struggle with modality bias and reveal significant gaps in current models’ ability to integrate multiple modalities under time constraints. In-depth analysis uncovers how multiple modalities interact and jointly influence model decisions in complex multimodal reasoning environments.

Conclusion: EscapeCraft-4D provides a comprehensive testbed for assessing selective cross-modal perception and time awareness in Omni models, highlighting current limitations in multimodal integration under temporal constraints and offering insights for future model development.

Abstract: Multimodal Large Language Models (MLLMs) have recently made rapid progress toward unified Omni models that integrate vision, language, and audio. However, existing environments largely focus on 2D or 3D visual context and vision-language tasks, offering limited support for temporally dependent auditory signals and selective cross-modal integration, where different modalities may provide complementary or interfering information, which are essential capabilities for realistic multimodal reasoning. As a result, whether models can actively coordinate modalities and reason under time-varying, irreversible conditions remains underexplored. To this end, we introduce \textbf{EscapeCraft-4D}, a customizable 4D environment for assessing selective cross-modal perception and time awareness in Omni models. It incorporates trigger-based auditory sources, temporally transient evidence, and location-dependent cues, requiring agents to perform spatio-temporal reasoning and proactive multimodal integration under time constraints. Building on this environment, we curate a benchmark to evaluate corresponding abilities across powerful models. Evaluation results suggest that models struggle with modality bias, and reveal significant gaps in current model’s ability to integrate multiple modalities under time constraints. Further in-depth analysis uncovers how multiple modalities interact and jointly influence model decisions in complex multimodal reasoning environments.

Method: Decomposes 3D counting into two subproblems: 1) estimating 3D geometry of the stack from multi-view images, and 2) determining occupancy ratio. Combines geometric reconstruction with deep learning-based depth analysis to count identical parts even when irregularly stacked and partially hidden.

Result: Validated on large-scale synthetic and diverse real-world data with manually verified total counts. Demonstrates robust performance under realistic industrial inspection conditions with heavy occlusion.

Conclusion: Proposes a novel 3D counting approach that effectively handles the challenging scenario of counting stacked manufactured parts with heavy occlusion, providing a robust solution for industrial inspection applications.

Abstract: Visual object counting is a fundamental computer vision task in industrial inspection, where accurate, high-throughput inventory tracking and quality assurance are critical. Moreover, manufactured parts are often too light to reliably deduce their count from their weight, or too heavy to move the stack on a scale safely and practically, making automated visual counting the more robust solution in many scenarios. However, existing methods struggle with stacked 3D items in containers, pallets, or bins, where most objects are heavily occluded and only a few are directly visible. To address this important yet underexplored challenge, we propose a novel 3D counting approach that decomposes the task into two complementary subproblems: estimating the 3D geometry of the stack and its occupancy ratio from multi-view images. By combining geometric reconstruction with deep learning-based depth analysis, our method can accurately count identical manufactured parts inside containers, even when they are irregularly stacked and partially hidden. We validate our 3D counting pipeline on large-scale synthetic and diverse real-world data with manually verified total counts, demonstrating robust performance under realistic inspection conditions.

Method: Proposes anchor-then-polish (ATP) framework: 1) Macro anchoring learns scene-adaptive projection matrix (12 DoF) to stabilize luminance and correct color globally, 2) Micro polishing refines details in wavelet domain and chrominance space under matrix guidance, 3) Constrained luminance update ensures global consistency while focusing on fine-grained polishing.

Result: Extensive experiments on multiple benchmarks show state-of-the-art performance, producing visually natural and quantitatively superior low-light enhancements compared to existing methods.

Conclusion: ATP framework effectively decouples global and local aspects of low-light enhancement, demonstrating that simple linear operators can handle global energy alignment while specialized modules refine details, leading to superior enhancement quality.

Abstract: Low-light image enhancement is challenging due to entangled degradations, mainly including poor illumination, color shifts, and texture interference. Existing methods often rely on complex architectures to address these issues jointly but may overfit simple physical constraints, leading to global distortions. This work proposes a novel anchor-then-polish (ATP) framework to fundamentally decouple global energy alignment from local detail refinement. First, macro anchoring is customized to (greatly) stabilize luminance distribution and correct color by learning a scene-adaptive projection matrix with merely 12 degrees of freedom, revealing that a simple linear operator can effectively align global energy. The macro anchoring then reduces the task to micro polishing, which further refines details in the wavelet domain and chrominance space under matrix guidance. A constrained luminance update strategy is designed to ensure global consistency while directing the network to concentrate on fine-grained polishing. Extensive experiments on multiple benchmarks show that our method achieves state-of-the-art performance, producing visually natural and quantitatively superior low-light enhancements.

Method: Introduces calibration tokens as a lightweight adaptation mechanism that modulates latent embeddings to align fisheye image distributions with perspective image distributions. Uses self-supervised training by recalibrating perspective images to fisheye images and enforcing consistency between depth estimates.

Result: Method consistently improves over state-of-the-art approaches on both indoor and outdoor datasets using a single set of tokens, enabling FMDEs to work with fisheye cameras without retraining.

Conclusion: Calibration tokens provide an effective way to adapt existing depth estimators to different camera geometries by aligning latent distributions, avoiding artifacts from conventional recalibration methods.

Abstract: We propose a method to extend foundational monocular depth estimators (FMDEs), trained on perspective images, to fisheye images. Despite being trained on tens of millions of images, FMDEs are susceptible to the covariate shift introduced by changes in camera calibration (intrinsic, distortion) parameters, leading to erroneous depth estimates. Our method aligns the distribution of latent embeddings encoding fisheye images to those of perspective images, enabling the reuse of FMDEs for fisheye cameras without retraining or finetuning. To this end, we introduce a set of Calibration Tokens as a light-weight adaptation mechanism that modulates the latent embeddings for alignment. By exploiting the already expressive latent space of FMDEs, we posit that modulating their embeddings avoids the negative impact of artifacts and loss introduced in conventional recalibration or map projection to a canonical reference frame in the image space. Our method is self-supervised and does not require fisheye images but leverages publicly available large-scale perspective image datasets. This is done by recalibrating perspective images to fisheye images, and enforcing consistency between their estimates during training. We evaluate our approach with several FMDEs, on both indoors and outdoors, where we consistently improve over state-of-the-art methods using a single set of tokens for both. Code available at: https://github.com/JungHeeKim29/calibration-token.

Method: Proposes ViFeEdit, a video-free tuning framework that uses architectural reparameterization to decouple spatial independence from full 3D attention in video diffusion transformers. Uses dual-path pipeline with separate timestep embeddings for noise scheduling, requiring only minimal training on 2D image data.

Result: Achieves promising results for controllable video generation and editing with only minimal training on 2D images, demonstrating visually faithful editing while maintaining temporal consistency.

Conclusion: ViFeEdit provides an effective solution for video generation and editing without requiring video training data, addressing key limitations in video diffusion model training through innovative architectural design.

Abstract: Diffusion Transformers (DiTs) have demonstrated remarkable scalability and quality in image and video generation, prompting growing interest in extending them to controllable generation and editing tasks. However, compared to the image counterparts, progress in video control and editing remains limited, mainly due to the scarcity of paired video data and the high computational cost of training video diffusion models. To address this issue, in this paper, we propose a video-free tuning framework termed ViFeEdit for video diffusion transformers. Without requiring any forms of video training data, ViFeEdit achieves versatile video generation and editing, adapted solely with 2D images. At the core of our approach is an architectural reparameterization that decouples spatial independence from the full 3D attention in modern video diffusion transformers, which enables visually faithful editing while maintaining temporal consistency with only minimal additional parameters. Moreover, this design operates in a dual-path pipeline with separate timestep embeddings for noise scheduling, exhibiting strong adaptability to diverse conditioning signals. Extensive experiments demonstrate that our method delivers promising results of controllable video generation and editing with only minimal training on 2D image data. Codes are available https://github.com/Lexie-YU/ViFeEdit.

Method: Proposes angle distribution refinement to reformulate angle regression as iterative refinement of probability distributions. Incorporates Chamfer distance cost into bipartite matching for better geometric alignment. Introduces oriented contrastive denoising to stabilize training and analyzes four noise modes.

Result: Achieves 77.73%/78.45%/80.15% AP50 on DOTA1.0 dataset with 132/119/119 FPS on 2080ti GPU across different model variants (O2-DFINE-L, O2-RTDETR-R50, O2-DEIM-R50).

Conclusion: The proposed O2-RTDETR is the first real-time end-to-end oriented object detector that effectively addresses rotation challenges in remote sensing imagery through novel angle representation, matching, and training stabilization techniques.

Abstract: Recent real-time detection transformers have gained popularity due to their simplicity and efficiency. However, these detectors do not explicitly model object rotation, especially in remote sensing imagery where objects appear at arbitrary angles, leading to challenges in angle representation, matching cost, and training stability. In this paper, we propose a real-time oriented object detection transformer, the first real-time end-to-end oriented object detector to the best of our knowledge, that addresses the above issues. Specifically, angle distribution refinement is proposed to reformulate angle regression as an iterative refinement of probability distributions, thereby capturing the uncertainty of object rotation and providing a more fine-grained angle representation. Then, we incorporate a Chamfer distance cost into bipartite matching, measuring box distance via vertex sets, enabling more accurate geometric alignment and eliminating ambiguous matches. Moreover, we propose oriented contrastive denoising to stabilize training and analyze four noise modes. We observe that a ground truth can be assigned to different index queries across different decoder layers, and analyze this issue using the proposed instability metric. We design a series of model variants and experiments to validate the proposed method. Notably, our O2-DFINE-L, O2-RTDETR-R50 and O2-DEIM-R50 achieve 77.73%/78.45%/80.15% AP50 on DOTA1.0 and 132/119/119 FPS on the 2080ti GPU. Code is available at https://github.com/wokaikaixinxin/ai4rs.

Method: Two-stage approach: 1) Audio-To-Sparse (ATS) predicts 3D landmark displacements from speech audio conditioned on emotion category/intensity; 2) Sparse-To-Mesh (STM) transfers landmark motion to target meshes using intrinsic surface features and landmark-to-vertex conditioning without template fitting.

Result: FreeTalk matches specialized baselines in-domain while providing substantially improved robustness to unseen identities and mesh topologies, enabling emotion-conditioned animation on arbitrary 3D scans.

Conclusion: FreeTalk enables emotion-conditioned 3D talking-head animation on unregistered meshes with arbitrary topology, overcoming limitations of template-based approaches while maintaining performance.

Abstract: Speech-driven 3D facial animation has advanced rapidly, yet most approaches remain tied to registered template meshes, preventing effective deployment on raw 3D scans with arbitrary topology. At the same time, modeling controllable emotional dynamics beyond lip articulation remains challenging, and is often tied to template-based parameterizations. We address these challenges by proposing FreeTalk, a two-stage framework for emotion-conditioned 3D talking-head animation that generalizes to unregistered face meshes with arbitrary vertex count and connectivity. First, Audio-To-Sparse (ATS) predicts a temporally coherent sequence of 3D landmark displacements from speech audio, conditioned on an emotion category and intensity. This sparse representation captures both articulatory and affective motion while remaining independent of mesh topology. Second, Sparse-To-Mesh (STM) transfers the predicted landmark motion to a target mesh by combining intrinsic surface features with landmark-to-vertex conditioning, producing dense per-vertex deformations without template fitting or correspondence supervision at test time. Extensive experiments show that FreeTalk matches specialized baselines when trained in-domain, while providing substantially improved robustness to unseen identities and mesh topologies. Code and pre-trained models will be made publicly available.

Method: Cannot determine method without access to paper content

Result: Cannot determine results without access to paper content

Conclusion: Cannot draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2509.11453: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.11453&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: CARS applies targeted perturbations to clinical feature vectors for controlled insertion/deletion of pathological findings while preserving anatomical structure. Uses principled synthetic image generation to address feature space gaps.

Result: Fine-tuning on CARS-generated images improves precision-recall performance, reduces predictive uncertainty, and improves model calibration across seven backbone architectures. Expert radiologists confirm realism and clinical agreement.

Conclusion: Anatomically faithful synthetic data generation for better feature space coverage is viable and effective for improving both performance and trustworthiness of chest X-ray classification systems without compromising clinical integrity.

Abstract: The clinical deployment of AI diagnostic models demands more than benchmark accuracy - it demands robustness across the full spectrum of disease presentations. However, publicly available chest radiographic datasets systematically underrepresent critical clinical feature combinations, leaving models under-trained precisely where clinical stakes are highest. We present CARS, a clinically aware and anatomically grounded framework that addresses this gap through principled synthetic image generation. CARS applies targeted perturbations to clinical feature vectors, enabling controlled insertion and deletion of pathological findings while explicitly preserving anatomical structure. We evaluate CARS across seven backbone architectures by fine-tuning models on synthetic subsets and testing on a held-out MIMIC-CXR benchmark. Compared to prior feature perturbation approaches, fine-tuning on CARS-generated images consistently improves precision-recall performance, reduces predictive uncertainty, and improves model calibration. Structural and semantic analyses demonstrate high anatomical fidelity, strong feature alignment, and low semantic uncertainty. Independent evaluation by two expert radiologists further confirms realism and clinical agreement. As the field moves toward regulated clinical AI, CARS demonstrates that anatomically faithful synthetic data generation for better feature space coverage is a viable and effective strategy for improving both the performance and trustworthiness of chest X-ray classification systems - without compromising clinical integrity.

Method: Introduces Kimodo, a kinematic motion diffusion model with carefully designed motion representation and two-stage denoiser architecture that decomposes root and body prediction to minimize motion artifacts while enabling flexible constraint conditioning through text and kinematic constraints (full-body keyframes, sparse joint positions/rotations, 2D waypoints, dense 2D paths).

Result: Generates high-quality motions while being easily controlled through various input modalities; experiments on large-scale mocap dataset justify design decisions and analyze how scaling dataset size and model size affect performance.

Conclusion: Kimodo demonstrates that scaling up mocap datasets and model architecture enables high-quality, expressive, and controllable human motion generation through multiple intuitive input modalities.

Abstract: High-quality human motion data is becoming increasingly important for applications in robotics, simulation, and entertainment. Recent generative models offer a potential data source, enabling human motion synthesis through intuitive inputs like text prompts or kinematic constraints on poses. However, the small scale of public mocap datasets has limited the motion quality, control accuracy, and generalization of these models. In this work, we introduce Kimodo, an expressive and controllable kinematic motion diffusion model trained on 700 hours of optical motion capture data. Our model generates high-quality motions while being easily controlled through text and a comprehensive suite of kinematic constraints including full-body keyframes, sparse joint positions/rotations, 2D waypoints, and dense 2D paths. This is enabled through a carefully designed motion representation and two-stage denoiser architecture that decomposes root and body prediction to minimize motion artifacts while allowing for flexible constraint conditioning. Experiments on the large-scale mocap dataset justify key design decisions and analyze how the scaling of dataset size and model size affect performance.

Method: Bootleg predicts latent representations from multiple hidden layers of a teacher network, creating a hierarchical objective that forces the model to capture features at varying levels of abstraction simultaneously.

Result: Significantly outperforms comparable baselines (+10% over I-JEPA) on ImageNet-1K and iNaturalist-21 classification, and semantic segmentation of ADE20K and Cityscapes.

Conclusion: Bootleg successfully bridges the divide between generative and predictive SSL by leveraging hierarchical latent prediction, achieving strong performance across multiple vision tasks while addressing limitations of existing approaches.

Abstract: The landscape of self-supervised learning (SSL) is currently dominated by generative approaches (e.g., MAE) that reconstruct raw low-level data, and predictive approaches (e.g., I-JEPA) that predict high-level abstract embeddings. While generative methods provide strong grounding, they are computationally inefficient for high-redundancy modalities like imagery, and their training objective does not prioritize learning high-level, conceptual features. Conversely, predictive methods often suffer from training instability due to their reliance on the non-stationary targets of final-layer self-distillation. We introduce Bootleg, a method that bridges this divide by tasking the model with predicting latent representations from multiple hidden layers of a teacher network. This hierarchical objective forces the model to capture features at varying levels of abstraction simultaneously. We demonstrate that Bootleg significantly outperforms comparable baselines (+10% over I-JEPA) on classification of ImageNet-1K and iNaturalist-21, and semantic segmentation of ADE20K and Cityscapes.