Method: Proposes a training-free Multi-Agent Recognition, Reasoning, and Reflection (MAR3) framework incorporating sociological Delphi theory. Uses Consensus Multimodal Recognition with LLM agents to recognize expression difficulty and dominant modality, adaptive Collaborative Object Reasoning based on modality-dominant difficulty rule, and Reflective Learning Segmentation where a check agent examines and iteratively corrects segmentation results.

Result: Achieves 69.2% J&F score on Ref-AVSBench dataset, outperforming state-of-the-art by 3.4% absolutely.

Conclusion: MAR3 effectively addresses limitations of previous Ref-AVS methods by explicitly recognizing multimodal cue characteristics, adaptively reasoning about objects, and incorporating reflective validation, leading to superior segmentation performance.

Abstract: Reference Audio-Visual Segmentation (Ref-AVS) aims to segment objects in audible videos based on multimodal cues in reference expressions. Previous methods overlook the explicit recognition of expression difficulty and dominant modality in multimodal cues, over-rely on the quality of the instruction-tuning dataset for object reasoning, and lack reflective validation of segmentation results, leading to erroneous mask predictions. To address these issues, in this paper, we propose a novel training-free Multi-Agent Recognition, Reasoning, and Reflection framework to achieve high-quality Reference Audio-Visual Segmentation, termed MAR3. Incorporating the sociological Delphi theory to achieve robust analysis, a Consensus Multimodal Recognition mechanism is proposed that enables LLM agents to explicitly recognize the difficulty of reference expressions and the dominant modality of multimodal cues. Based on our modality-dominant difficulty rule, we propose an adaptive Collaborative Object Reasoning strategy to reliably reason about the referred object. To further ensure precise mask prediction, we develop a Reflective Learning Segmentation mechanism, in which a check agent examines intermediate segmentation results and iteratively corrects the object text prompt of the segment agent. Experiments demonstrate that MAR3 achieves superior performance (69.2% in J&F) on the Ref-AVSBench dataset, outperforming SOTA by 3.4% absolutely.

Method: Reframe selection as sequential decision-making and train a Reinforcement Learning agent (Dueling DQN with query-centric Transformer Decoder) to construct optimal demonstration sets that maximize MLLM downstream performance.

Result: LSD significantly outperforms baselines on objective, factual regression tasks across five visual regression benchmarks, while kNN remains optimal for subjective preference tasks. LSD better defines regression boundaries by balancing visual relevance with diversity.

Conclusion: Learned demonstration selection (LSD) is strictly necessary for visual ICL on factual regression tasks, illuminating when sophisticated selection methods are required versus when simple kNN suffices.

Abstract: Multimodal Large Language Models (MLLMs) adapt to visual tasks via in-context learning (ICL), which relies heavily on demonstration quality. The dominant demonstration selection strategy is unsupervised k-Nearest Neighbor (kNN) search. While simple, this similarity-first approach is sub-optimal for complex factual regression tasks; it selects redundant examples that fail to capture the task’s full output range. We reframe selection as a sequential decision-making problem and introduce Learning to Select Demonstrations (LSD), training a Reinforcement Learning agent to construct optimal demonstration sets. Using a Dueling DQN with a query-centric Transformer Decoder, our agent learns a policy that maximizes MLLM downstream performance. Evaluating across five visual regression benchmarks, we uncover a crucial dichotomy: while kNN remains optimal for subjective preference tasks, LSD significantly outperforms baselines on objective, factual regression tasks. By balancing visual relevance with diversity, LSD better defines regression boundaries, illuminating when learned selection is strictly necessary for visual ICL.

Method: Proposes Fed-CMP framework with two key components: 1) Canonical Reliability-Aware Aggregation - constructs canonical space to decompose client projectors into shared alignment basis and client-specific coefficients with reliability-weighted fusion; 2) Orthogonality-Preserved Momentum - applies momentum to shared alignment basis via orthogonal projection to accumulate historical optimization directions while preserving geometric structure.

Result: Extensive experiments on four federated pre-training scenarios based on public datasets show Fed-CMP significantly outperforms existing baselines.

Conclusion: Fed-CMP successfully addresses challenges in federated MLLM pre-training, enabling collaborative training of cross-modal projectors while mitigating parameter interference and gradient oscillations.

Abstract: The rapid evolution of Multimodal Large Language Models (MLLMs) is bottlenecked by the saturation of high-quality public data, while vast amounts of diverse multimodal data remain inaccessible in privacy-sensitive silos. Federated Learning (FL) offers a promising solution to unlock these distributed resources, but existing research focuses predominantly on fine-tuning, leaving the foundational pre-training phase largely unexplored. In this paper, we formally introduce the Federated MLLM Alignment (Fed-MA) task, a lightweight pre-training paradigm that freezes the vision encoder and LLM while collaboratively training the cross-modal projector. We identify two critical challenges in this setting: (i) parameter interference in aggregating local projectors; and (ii) gradient oscillations in one-pass collaborative SGD. To address these challenges, we propose Fed-CMP, a pioneering framework for federated MLLM pre-training. Fed-CMP employs Canonical Reliability-Aware Aggregation, which constructs a canonical space to decompose client projectors into a shared alignment basis and client-specific coefficients, then performs reliability-weighted fusion to suppress parameter interference. Furthermore, Fed-CMP introduces Orthogonality-Preserved Momentum, which applies momentum to the shared alignment basis via orthogonal projection, accumulating historical optimization directions while preserving geometric structure. We construct four federated pre-training scenarios based on public datasets, and extensive experiments validate that Fed-CMP significantly outperforms existing baselines.

Method: GeoBlock analyzes cross-token dependency patterns from attention mechanisms to identify geometrically stable refinement regions. It dynamically determines appropriate block boundaries during decoding by examining dependency geometry rather than using predefined schedules or local confidence heuristics. The framework requires no additional training and integrates into existing block diffusion architectures.

Result: Extensive experiments across multiple benchmarks show that GeoBlock reliably identifies geometry-consistent block boundaries and improves the accuracy of block diffusion with only a small additional computational budget. The method preserves parallel efficiency while enforcing dependency-consistent refinement.

Conclusion: GeoBlock provides a principled approach to block inference in diffusion language models by leveraging dependency geometry, enabling efficient parallel refinement while maintaining the reliability of autoregressive decoding. The framework demonstrates that geometry-aware block sizing can significantly improve diffusion language model performance.

Abstract: Block diffusion enables efficient parallel refinement in diffusion language models, but its decoding behavior depends critically on block size. Existing block-sizing strategies rely on fixed rules or heuristic signals and do not account for the dependency geometry that determines which tokens can be safely refined together. This motivates a geometry view of diffusion decoding: \emph{regions with strong causal ordering require sequential updates, whereas semantically cohesive regions admit parallel refinement.} We introduce GeoBlock, a geometry-aware block inference framework that determines block granularity directly from attention-derived dependency geometry. Instead of relying on predefined schedules or local confidence heuristics, GeoBlock analyzes cross-token dependency patterns to identify geometrically stable refinement regions and dynamically determines appropriate block boundaries during decoding. By adapting block granularity to the dependency geometry, GeoBlock preserves the parallel efficiency of block diffusion while enforcing dependency-consistent refinement that exhibits autoregressive reliability. GeoBlock requires no additional training and integrates seamlessly into existing block diffusion architectures. Extensive experiments across multiple benchmarks show that GeoBlock reliably identifies geometry-consistent block boundaries and improves the accuracy of block diffusion with only a small additional computational budget.

Method: Created AlpsBench with 2,500 long-term interaction sequences from WildChat (real human-LLM dialogues), paired with human-verified structured memories capturing explicit and implicit personalization signals. Defined four tasks: personalized information extraction, updating, retrieval, and utilization.

Result: Benchmarking frontier LLMs revealed: (1) models struggle with latent user trait extraction, (2) memory updating hits performance ceilings, (3) retrieval accuracy drops sharply with large distractor pools, (4) explicit memory mechanisms improve recall but don’t guarantee preference-aligned or emotionally resonant responses.

Conclusion: AlpsBench provides a comprehensive framework for evaluating LLM personalization using real-world dialogues, highlighting current limitations in memory management and personalization capabilities of state-of-the-art models.

Abstract: As Large Language Models (LLMs) evolve into lifelong AI assistants, LLM personalization has become a critical frontier. However, progress is currently bottlenecked by the absence of a gold-standard evaluation benchmark. Existing benchmarks either overlook personalized information management that is critical for personalization or rely heavily on synthetic dialogues, which exhibit an inherent distribution gap from real-world dialogue. To bridge this gap, we introduce AlpsBench, An LLM PerSonalization benchmark derived from real-world human-LLM dialogues. AlpsBench comprises 2,500 long-term interaction sequences curated from WildChat, paired with human-verified structured memories that encapsulate both explicit and implicit personalization signals. We define four pivotal tasks - personalized information extraction, updating, retrieval, and utilization - and establish protocols to evaluate the entire lifecycle of memory management. Our benchmarking of frontier LLMs and memory-centric systems reveals that: (i) models struggle to reliably extract latent user traits; (ii) memory updating faces a performance ceiling even in the strongest models; (iii) retrieval accuracy declines sharply in the presence of large distractor pools; and (iv) while explicit memory mechanisms improve recall, they do not inherently guarantee more preference-aligned or emotionally resonant responses. AlpsBench aims to provide a comprehensive framework.

Method: Statistical analysis of AI context window growth trends, meta-analysis of human reading rates to derive Effective Context Span (ECS), neurobiological mechanism review across eight neuroimaging studies, and empirical evidence analysis of delegation thresholds.

Result: AI context windows grew from 512 tokens (2017) to 2,000,000 tokens (2026) while human ECS declined from ~16,000 tokens (2004) to ~1,800 tokens (2026), creating a 556-1,111x raw divergence ratio (56-111x quality-adjusted).

Conclusion: The Cognitive Divergence creates a Delegation Feedback Loop where humans increasingly delegate cognitive tasks to AI, potentially further reducing human cognitive capacities, requiring research on validated ECS measurement and longitudinal study of AI-mediated cognitive change.

Abstract: This paper documents and theorises a self-reinforcing dynamic between two measurable trends: the exponential expansion of large language model (LLM) context windows and the secular contraction of human sustained-attention capacity. We term the resulting asymmetry the Cognitive Divergence. AI context windows have grown from 512 tokens in 2017 to 2,000,000 tokens by 2026 (factor ~3,906; fitted lambda = 0.59/yr; doubling time ~14 months). Over the same period, human Effective Context Span (ECS) – a token-equivalent measure derived from validated reading-rate meta-analysis (Brysbaert, 2019) and an empirically motivated Comprehension Scaling Factor – has declined from approximately 16,000 tokens (2004 baseline) to an estimated 1,800 tokens (2026, extrapolated from longitudinal behavioural data ending 2020 (Mark, 2023); see Section 9 for uncertainty discussion). The AI-to-human ratio grew from near parity at the ChatGPT launch (November 2022) to 556–1,111x raw and 56–111x quality-adjusted, after accounting for retrieval degradation (Liu et al., 2024; Chroma, 2025). Beyond documenting this divergence, the paper introduces the Delegation Feedback Loop hypothesis: as AI capability grows, the cognitive threshold at which humans delegate to AI falls, extending to tasks of negligible demand; the resulting reduction in cognitive practice may further attenuate the capacities already documented as declining (Gerlich, 2025; Kim et al., 2026; Kosmyna et al., 2025). Neither trend reverses spontaneously. The paper characterises the divergence statistically, reviews neurobiological mechanisms across eight peer-reviewed neuroimaging studies, presents empirical evidence bearing on the delegation threshold, and proposes a research agenda centred on a validated ECS psychometric instrument and longitudinal study of AI-mediated cognitive change.

Method: Collaborated with social scientists to develop SACRED, a high-quality multimedia multimodal dataset with guaranteed classification faithfulness. Evaluated 13 popular LLMs along with traditional rule-based and fine-tuned approaches on this dataset.

Result: DeepSeek-V3 achieved 79.19% accuracy on Quora test set for text classification. GPT-4o-mini surpassed other models in vision tasks with 63.99% F1 score. Discovered a new type of connectedness valuable for communication science studies.

Conclusion: SACRED is the first annotated multimodal dataset from online spirituality communication. The study demonstrates LLMs’ capability in classifying abstract concepts like spirituality and reveals new patterns of connectedness in spiritual communication.

Abstract: In religion and theology studies, spirituality has garnered significant research attention for the reason that it not only transcends culture but offers unique experience to each individual. However, social scientists often rely on limited datasets, which are basically unavailable online. In this study, we collaborated with social scientists to develop a high-quality multimedia multi-modal datasets, \textbf{SACRED}, in which the faithfulness of classification is guaranteed. Using \textbf{SACRED}, we evaluated the performance of 13 popular LLMs as well as traditional rule-based and fine-tuned approaches. The result suggests DeepSeek-V3 model performs well in classifying such abstract concepts (i.e., 79.19% accuracy in the Quora test set), and the GPT-4o-mini model surpassed the other models in the vision tasks (63.99% F1 score). Purportedly, this is the first annotated multi-modal dataset from online spirituality communication. Our study also found a new type of connectedness which is valuable for communication science studies.

Method: Translated 980 questions from MathVista, ScienceQA, and MMMU into 6 Indian languages using IndicTrans2 with Gemini 2.0 Flash cross-verification. Evaluated 8 VLMs across 7 languages (including English), generating 68,600 inference records with text-only and chain-of-thought ablations.

Result: Accuracy drops of 9.8-25 percentage points when switching from English to Indian languages. Dravidian languages suffer up to 13.2pp more than Indo-Aryan languages. Chain-of-thought prompting degrades performance for Bengali (-14.4pp) and Kannada (-11.4pp). Aya-Vision-8B drops 28.5pp on Dravidian scripts despite multilingual pretraining.

Conclusion: Multilingual pretraining alone doesn’t transfer visual reasoning capabilities. Current VLMs exhibit English-centric reasoning chains and significant performance disparities across languages, highlighting the need for more equitable multilingual vision-language models.

Abstract: Vision-language models score well on mathematical, scientific, and spatial reasoning benchmarks, yet these evaluations are overwhelmingly English. I present the first cross-lingual visual reasoning audit for Indian languages. 980 questions from MathVista, ScienceQA, and MMMU are translated into Hindi, Tamil, Telugu, Bengali, Kannada, and Marathi using IndicTrans2, with Gemini 2.0 Flash cross-verification on 50 samples per language (inter-translator agreement 0.79-0.84). Eight VLMs, from 7B open-source models to GPT-4o, are evaluated across all seven languages, yielding 68,600 inference records that include text-only and chain-of-thought ablations. I find accuracy drops of 9.8-25 percentage points when switching from English to an Indian language, with Dravidian languages suffering up to 13.2 pp more than Indo-Aryan. Chain-of-thought prompting degrades Bengali (-14.4 pp) and Kannada (-11.4 pp) rather than helping, exposing English-centric reasoning chains. Aya-Vision-8B, built for 23 languages, still drops 28.5 pp on Dravidian scripts; multilingual pretraining alone does not transfer visual reasoning. I release the translated benchmark and all model outputs.

Method: LogicDiff uses a lightweight classification head (4.2M parameters) to predict logical roles (premise, connective, derived step, conclusion, filler) from base model hidden states with 98.4% accuracy, then uses a dependency-ordered scheduler to unmask tokens in logical dependency order: premises first, then connectives, then derived steps, then conclusions.

Result: Without modifying base model parameters, LogicDiff improves LLaDA-8B-Instruct accuracy from 22.0% to 60.7% on GSM8K (+38.7 percentage points) and from 23.6% to 29.2% on MATH-500 (+5.6 pp), with less than 6% speed overhead.

Conclusion: A substantial portion of the reasoning deficit in masked diffusion language models is attributable to suboptimal token unmasking order, not to limitations of the model’s learned representations.

Abstract: Masked diffusion language models (MDLMs) generate text by iteratively unmasking tokens from a fully masked sequence, offering parallel generation and bidirectional context. However, their standard confidence-based unmasking strategy systematically defers high-entropy logical connective tokens, the critical branching points in reasoning chains, leading to severely degraded reasoning performance. We introduce LogicDiff, an inference-time method that replaces confidence-based unmasking with logic-role-guided unmasking. A lightweight classification head (4.2M parameters, 0.05% of the base model) predicts the logical role of each masked position (premise, connective, derived step, conclusion, or filler) from the base model’s hidden states with 98.4% accuracy. A dependency-ordered scheduler then unmasks tokens in logical dependency order: premises first, then connectives, then derived steps, then conclusions. Without modifying a single parameter of the base model and without any reinforcement learning or task-specific training, LogicDiff improves LLaDA-8B-Instruct accuracy from 22.0% to 60.7% on GSM8K (+38.7 percentage points) and from 23.6% to 29.2% on MATH-500 (+5.6 pp), with less than 6% speed overhead. Our results demonstrate that a substantial portion of the reasoning deficit in MDLMs is attributable to suboptimal token unmasking order, not to limitations of the model’s learned representations.

Method: Proposes Hybrid Document-Routed Retrieval (HDRR), a two-stage architecture that first uses Semantic File Routing (SFR) with LLM structured output to route queries to relevant documents, then performs chunk-based retrieval scoped only to the identified document(s).

Result: HDRR achieves best performance on all metrics: average score of 7.54 (25.2% above CBR, 16.9% above SFR), failure rate of only 6.4%, correctness rate of 67.7% (+18.7pp over CBR), and perfect-answer rate of 20.1% (+6.3pp over CBR, +11.6pp over SFR).

Conclusion: HDRR resolves the robustness-precision trade-off between document routing and chunk retrieval, eliminating cross-document confusion while preserving targeted chunk precision, making it superior for financial document QA systems.

Abstract: Retrieval-Augmented Generation (RAG) systems for financial document question answering typically follow a chunk-based paradigm: documents are split into fragments, embedded into vector space, and retrieved via similarity search. While effective in general settings, this approach suffers from cross-document chunk confusion in structurally homogeneous corpora such as regulatory filings. Semantic File Routing (SFR), which uses LLM structured output to route queries to whole documents, reduces catastrophic failures but sacrifices the precision of targeted chunk retrieval. We identify this robustness-precision trade-off through controlled evaluation on the FinDER benchmark (1,500 queries across five groups): SFR achieves higher average scores (6.45 vs. 6.02) and fewer failures (10.3% vs. 22.5%), while chunk-based retrieval (CBR) yields more perfect answers (13.8% vs. 8.5%). To resolve this trade-off, we propose Hybrid Document-Routed Retrieval (HDRR), a two-stage architecture that uses SFR as a document filter followed by chunk-based retrieval scoped to the identified document(s). HDRR eliminates cross-document confusion while preserving targeted chunk precision. Experimental results demonstrate that HDRR achieves the best performance on every metric: an average score of 7.54 (25.2% above CBR, 16.9% above SFR), a failure rate of only 6.4%, a correctness rate of 67.7% (+18.7 pp over CBR), and a perfect-answer rate of 20.1% (+6.3 pp over CBR, +11.6 pp over SFR). HDRR resolves the trade-off by simultaneously achieving the lowest failure rate and the highest precision across all five experimental groups.

Method: Train minimal GPT on exhaustive 2-digit addition, then systematically test 3-digit generalization failures through controlled experiments: layout shift tests, carry probes, conditional recomposition studies, and targeted repair interventions.

Result: Identified four staged failures: 1) layout barrier requiring mixed-layout exposure, 2) hundreds position acting as carry flag rather than semantic digit, 3) conditional recomposition bottleneck, and 4) tens residual errors. Targeted repairs improved performance from 0.664 to 0.822 exact match.

Conclusion: Arithmetic out-of-distribution failures can be decomposed into experimentally testable stages: layout, carry-semantics, recomposition, and tens-residual, providing a more nuanced understanding than single-score benchmarks.

Abstract: Arithmetic benchmarks are often reduced to a single held-out score, but that score can conflate qualitatively different failures. We study a controlled minimal GPT trained on exhaustive 2-digit addition, where all local digit transitions are already present in training, and ask why 3-digit generalization still fails. The failure is staged. First, there is a layout barrier: a learned absolute-position model collapses under a pure 3-digit layout shift, and mixed-layout exposure is the only intervention that materially weakens this barrier. Second, after layout repair, the hundreds position behaves like a carry flag rather than a semantic hundreds digit; targeted carry probes reverse the relevant logit margin, whereas a matched extra-data control does not. Third, after carry repair, the main remaining bottleneck is conditional recomposition: high-conditioned tail data outperforms a matched control, high-only data, and tail-only data on all true-3-digit suites, and the same ordering reappears in a larger 2-layer bridge experiment. The residual errors after recomposition are then overwhelmingly tens-only, and a separate 10-seed late-stage study shows that a sign-aware tens repair raises exact match on the hardest thousands-carry suite from 0.664 to 0.822. We therefore provide an experimentally testable decomposition of arithmetic OOD failure into layout, carry-semantics, recomposition, and late tens-residual stages.

Method: Controlled evaluation of six open-weight models across four political science annotation tasks under identical conditions (quantisation, hardware, prompt-templates), examining interaction effects between model choice, size, learning approach, and prompt style.

Result: Interaction effects dominate main effects - no single model, prompt style, or learning approach is uniformly superior. Model size is unreliable for cost/performance prediction, and popular prompt engineering techniques yield inconsistent or negative effects.

Conclusion: Researchers need validation-first frameworks with principled pipeline decision ordering, prompt freezing guidance, held-out evaluation, reporting standards, and open-source tools to navigate complex implementation choices transparently.

Abstract: Political scientists are rapidly adopting large language models (LLMs) for text annotation, yet the sensitivity of annotation results to implementation choices remains poorly understood. Most evaluations test a single model or configuration; how model choice, model size, learning approach, and prompt style interact, and whether popular “best practices” survive controlled comparison, are largely unexplored. We present a controlled evaluation of these pipeline choices, testing six open-weight models across four political science annotation tasks under identical quantisation, hardware, and prompt-template conditions. Our central finding is methodological: interaction effects dominate main effects, so seemingly reasonable pipeline choices can become consequential researcher degrees of freedom. No single model, prompt style, or learning approach is uniformly superior, and the best-performing model varies across tasks. Two corollaries follow. First, model size is an unreliable guide both to cost and to performance: cross-family efficiency differences are so large that some larger models are less resource-intensive than much smaller alternatives, while within model families mid-range variants often match or exceed larger counterparts. Second, widely recommended prompt engineering techniques yield inconsistent and sometimes negative effects on annotation performance. We use these benchmark results to develop a validation-first framework - with a principled ordering of pipeline decisions, guidance on prompt freezing and held-out evaluation, reporting standards, and open-source tools - to help researchers navigate this decision space transparently.

Method: Scraped 10 years of publicly available dream reports from an online forum where users share anonymous dream journals, with optional user-provided labels (lucid, non-lucid, nightmare). Applied descriptive statistics, visualizations, and construct validation to analyze language patterns.

Result: Created a corpus of 55k dream reports from 5k contributors, including 10k lucid, 25k non-lucid, and 2k nightmare labels. Construct validation showed language patterns in lucid-labeled reports are consistent with known characteristics of lucid dreams.

Conclusion: The corpus has broad value for dream science, with the labeled subset being particularly powerful for new discoveries in lucid dream studies, enabling systematic analysis of dream phenomenology.

Abstract: All varieties of dreaming remain a mystery. Lucid dreams in particular, or those characterized by awareness of the dream, are notoriously difficult to study. Their scarce prevalence and resistance to deliberate induction make it difficult to obtain a sizeable corpus of lucid dream reports. The consequent lack of clarity around lucid dream phenomenology has left the many purported applications of lucidity under-realized. Here, a large corpus of 55k dream reports from 5k contributors is curated, described, and validated for future research. Ten years of publicly available dream reports were scraped from an online forum where users share anonymous dream journals. Importantly, users optionally categorize their dream as lucid, non-lucid, or a nightmare, offering a user-provided labeling system that includes 10k lucid and 25k non-lucid, and 2k nightmare labels. After characterizing the corpus with descriptive statistics and visualizations, construct validation shows that language patterns in lucid-labeled reports are consistent with known characteristics of lucid dreams. While the entire corpus has broad value for dream science, the labeled subset is particularly powerful for new discoveries in lucid dream studies.

Method: Proposed a five-step genealogy connecting training data to em dash usage, conducted suppression experiments across 12 models from 5 providers with instructions to avoid markdown formatting, and tested explicit em dash prohibition.

Result: Em dash frequency varies from 0.0 to 9.1 per 1,000 words across models, with Llama models producing none. Em dashes persist even under markdown suppression in most models, serving as a signature of specific fine-tuning procedures.

Conclusion: Em dash overuse results from markdown leaking into prose from training data, functioning as a diagnostic tool for fine-tuning methodology rather than a stylistic defect, connecting previously isolated observations about AI text patterns.

Abstract: Large language models produce em dashes at varying rates, and the observation that some models “overuse” them has become one of the most widely discussed markers of AI-generated text. Yet no mechanistic account of this pattern exists, and the parallel observation that LLMs default to markdown-formatted output has never been connected to it. We propose that the em dash is markdown leaking into prose – the smallest surviving unit of the structural orientation that LLMs acquire from markdown-saturated training corpora. We present a five-step genealogy connecting training data composition, structural internalization, the dual-register status of the em dash, and post-training amplification. We test this with a two-condition suppression experiment across twelve models from five providers (Anthropic, OpenAI, Meta, Google, DeepSeek): when models are instructed to avoid markdown formatting, overt features (headers, bullets, bold) are eliminated or nearly eliminated, but em dashes persist – except in Meta’s Llama models, which produce none at all. Em dash frequency and suppression resistance vary from 0.0 per 1,000 words (Llama) to 9.1 (GPT-4.1 under suppression), functioning as a signature of the specific fine-tuning procedure applied. A three-condition suppression gradient shows that even explicit em dash prohibition fails to eliminate the artifact in some models, and a base-vs-instruct comparison confirms that the latent tendency exists pre-RLHF. These findings connect two previously isolated online discourses and reframe em dash frequency as a diagnostic of fine-tuning methodology rather than a stylistic defect.

Method: Retrieval-Augmented Self-Supervised Prompt Refinement (RASPRef) retrieves relevant examples and reasoning trajectories, then uses multi-sample consistency, verifier feedback, and model-generated critiques to iteratively refine prompts without human supervision.

Result: Experiments on GSM8K-style mathematical reasoning tasks show retrieval-guided prompting improves performance compared to static prompting baselines. The effectiveness depends on retrieval quality, trajectory selection, and feedback signals.

Conclusion: Prompt design remains critical for reasoning-oriented language models, and self-improving prompts offer a practical, scalable strategy for improving reasoning performance without manual intervention.

Abstract: Recent reasoning-focused language models such as DeepSeek R1 and OpenAI o1 have demonstrated strong performance on structured reasoning benchmarks including GSM8K, MATH, and multi-hop question answering tasks. However, their performance remains highly sensitive to prompt formulation, and designing effective prompts is typically a manual and iterative process that does not scale well across tasks or domains. To address this limitation, we introduce Retrieval-Augmented Self-Supervised Prompt Refinement (RASPRef), a framework that improves prompts without requiring human annotations or task-specific supervision. The approach retrieves relevant examples and previously generated reasoning trajectories, and leverages signals such as multi-sample consistency, verifier feedback, and model-generated critiques to iteratively refine the prompt. Unlike prior approaches that focus primarily on improving model outputs, RASPRef directly treats the prompt as the optimization target and improves it through an iterative retrieval-guided refinement process. Experiments on GSM8K-style mathematical reasoning tasks show that retrieval-guided prompting improves performance compared with a static prompting baseline. We further discuss how retrieval quality, trajectory selection, and self-supervised feedback signals may influence the effectiveness of prompt refinement. These findings suggest that prompt design remains a critical factor for reasoning-oriented language models, and that self-improving prompts offer a practical and scalable strategy for improving reasoning performance.

Method: Developed a comprehensive benchmarking platform with standardized evaluation metrics (word error rate and inverse real-time factor), tested 86 open-source and proprietary systems across 12 datasets covering English short/long-form and multilingual short-form tracks, with support for multiple toolkits including ESPNet, NeMo, SpeechBrain, and Transformers.

Result: Conformer-based encoders with transformer-based decoders achieved the best average WER, while CTC and TDT decoders offered superior efficiency (RTFx), making them better for long-form and batched processing. All code and dataset loaders are open-sourced for community use.

Conclusion: The Open ASR Leaderboard provides a transparent, extensible evaluation framework that enables reproducible benchmarking and systematic comparison of ASR systems, with findings that guide architectural choices based on accuracy vs. efficiency trade-offs.

Abstract: We present the Open ASR Leaderboard, a reproducible benchmarking platform with community contributions from academia and industry. It compares 86 open-source and proprietary systems across 12 datasets, with English short- and long-form and multilingual short-form tracks. We standardize word error rate (WER) and inverse real-time factor (RTFx) evaluation for consistent accuracy-efficiency comparisons across model architectures and toolkits (e.g., ESPNet, NeMo, SpeechBrain, Transformers). We observe that Conformer-based encoders paired with transformer-based decoders achieve the best average WER, while connectionist temporal classification (CTC) and token-and-duration transducer (TDT) decoders offer superior RTFx, making them better suited for long-form and batched processing. All code and dataset loaders are open-sourced to support transparent, extensible evaluation. We present our evaluation methodology to facilitate community-driven benchmarking in ASR and other tasks.

Method: Community-driven approach using Mozilla Common Voice platform (CV14-CV23). Methods included: interface localization, Wikipedia-based sentence extraction with automated filtering, phonemically targeted contributions for frequently dropped Pashto characters, and multi-channel community outreach including VOA Pashto broadcast campaign.

Result: Built corpus of 147 hours with 1,483 unique speakers (107,781 clips, 60,337 validated, 82.33 validated hours). Speaker participation increased 108-fold between CV17 and CV18. Fine-tuning Whisper Base on MCV20 reduced WER from 99.0% (zero-shot) to 13.4% on MCV20 test split.

Conclusion: Successfully created the first large-scale open Pashto speech corpus through community effort, demonstrating that targeted outreach and phonemic strategies can build resources for low-resource languages, significantly improving ASR performance.

Abstract: We present the Pashto Common Voice corpus – the first large-scale, openly licensed speech resource for Pashto, a language with over 60 million native speakers largely absent from open speech technology. Through a community effort spanning 2022-2025, the corpus grew from 1.5 hours and 5 contributors to 147 total hours and 1,483 unique speakers across ten Mozilla Common Voice releases (CV14-CV23). Speaker participation increased approximately 108-fold between CV17 and CV18, coinciding with a VOA Pashto broadcast campaign. We describe the full methodology: interface localisation, Wikipedia-based sentence extraction with automated filtering, phonemically targeted contributions for the four most frequently dropped Pashto characters, and multi-channel community outreach. MCV23 contains 107,781 clips (60,337 validated; 82.33 validated hours) across 13 content domains. Fine-tuning Whisper Base on the MCV20 yields 13.4% WER on the MCV20 test split, against the published Whisper Base zero-shot WER of 99.0% on Pashto.

Method: Proposes AgentSwing framework with probabilistic modeling of long-horizon success through search efficiency and terminal precision. At trigger points, expands multiple context-managed branches in parallel and uses lookahead routing to select the most promising continuation.

Result: Experiments across diverse benchmarks and agent backbones show AgentSwing consistently outperforms static context management methods, often matching or exceeding their performance with up to 3× fewer interaction turns while improving ultimate performance ceiling.

Conclusion: AgentSwing provides effective adaptive context management for long-horizon agents, and the probabilistic framework offers principled analysis and design methodology for future context management strategies.

Abstract: As large language models (LLMs) evolve into autonomous agents for long-horizon information-seeking, managing finite context capacity has become a critical bottleneck. Existing context management methods typically commit to a single fixed strategy throughout the entire trajectory. Such static designs may work well in some states, but they cannot adapt as the usefulness and reliability of the accumulated context evolve during long-horizon search. To formalize this challenge, we introduce a probabilistic framework that characterizes long-horizon success through two complementary dimensions: search efficiency and terminal precision. Building on this perspective, we propose AgentSwing, a state-aware adaptive parallel context management routing framework. At each trigger point, AgentSwing expands multiple context-managed branches in parallel and uses lookahead routing to select the most promising continuation. Experiments across diverse benchmarks and agent backbones show that AgentSwing consistently outperforms strong static context management methods, often matching or exceeding their performance with up to $3\times$ fewer interaction turns while also improving the ultimate performance ceiling of long-horizon web agents. Beyond the empirical gains, the proposed probabilistic framework provides a principled lens for analyzing and designing future context management strategies for long-horizon agents.

Method: Train lightweight HASS and EAGLE-2 draft models on MathInstruct, ShareGPT, and mixed-data variants; evaluate on MT-Bench, GSM8K, MATH-500, and SVAMP; study acceptance length, temperature effects, and methods for combining specialized drafters (checkpoint averaging, confidence-based routing, merged-tree verification).

Result: Task-specific training yields clear specialization: MathInstruct-trained drafts excel on reasoning benchmarks, ShareGPT-trained drafts excel on MT-Bench. Mixed-data improves robustness but doesn’t dominate across temperatures. Confidence-based routing outperforms single-domain drafts, with merged-tree verification achieving highest acceptance length overall.

Conclusion: Speculative decoding quality depends on both draft architecture and match between training data and downstream workload; specialized drafters are better combined at inference time than in weight space; confidence is more useful routing signal than entropy.

Abstract: Speculative decoding accelerates autoregressive generation by letting a lightweight draft model propose future tokens that a larger target model then verifies in parallel. In practice, however, draft models are usually trained on broad generic corpora, which leaves it unclear how much speculative decoding quality depends on the draft training distribution. We study this question with lightweight HASS and EAGLE-2 drafters trained on MathInstruct, ShareGPT, and mixed-data variants, evaluated on MT-Bench, GSM8K, MATH-500, and SVAMP. Measured by acceptance length, task-specific training yields clear specialization: MathInstruct-trained drafts are strongest on reasoning benchmarks, while ShareGPT-trained drafts are strongest on MT-Bench. Mixed-data training improves robustness, but larger mixtures do not dominate across decoding temperatures. We also study how to combine specialized drafters at inference time. Naive checkpoint averaging performs poorly, whereas confidence-based routing improves over single-domain drafts and merged-tree verification yields the highest acceptance length overall for both backbones. Finally, confidence is a more useful routing signal than entropy: rejected tokens tend to have higher entropy, but confidence produces much clearer benchmark-level routing decisions. These results show that speculative decoding quality depends not only on draft architecture, but also on the match between draft training data and downstream workload, and that specialized drafters are better combined at inference time than in weight space.

Method: Collected speech data from 51 bilingual speakers with matched Mandarin and English sessions, including read sentences and spontaneous interviews. Audio was recorded at 44.1 kHz, fully transcribed, force-aligned at word/phone levels, and anonymized.

Result: Created a 29.8-hour corpus with ~14.7 hours Mandarin and ~15.1 hours English content. Documented token/type statistics, code-switching patterns (frequent in Mandarin sessions), and enables within-/cross-speaker/language acoustic comparisons.

Conclusion: The MELI Corpus provides a valuable open-source resource for bilingual speech research, supporting both quantitative acoustic analysis and qualitative investigation of language attitudes and code-switching behaviors.

Abstract: We introduce the Mandarin-English Language Interview (MELI) Corpus, an open-source resource of 29.8 hours of speech from 51 Mandarin-English bilingual speakers. MELI combines matched sessions in Mandarin and English with two speaking styles: read sentences and spontaneous interviews about language varieties, standardness, and learning experiences. Audio was recorded at 44.1 kHz (16-bit, stereo). Interviews were fully transcribed, force-aligned at word and phone levels, and anonymized. Descriptively, the Mandarin component totals ~14.7 hours (mean duration 17.3 minutes) and the English component ~15.1 hours (mean duration 17.8 minutes). We report token/type statistics for each language and document code-switching patterns (frequent in Mandarin sessions; more limited in English sessions). The corpus design supports within-/cross-speaker, within/cross-language acoustic comparison and links acoustics to speakers’ stated language attitudes, enabling both quantitative and qualitative analyses. The MELI Corpus will be released with transcriptions, alignments, metadata, scans of labelled maps and documentation under a CC BY-NC 4.0 license.

Method: The chapter presents a conceptual framework for integrating textual data, discussing the challenges, reviewing state-of-the-art approaches, and identifying open research problems in this area.

Result: The chapter makes a case for textual data integration, outlines the technical challenges involved, surveys existing approaches, and highlights important open problems for future research.

Conclusion: Textual data integration is an important but challenging area that requires further research to fully leverage the knowledge contained in unstructured text alongside structured data sources.

Abstract: Data comes in many forms. From a shallow perspective, they can be viewed as being either in structured (e.g., as a relation, as key-value pairs) or unstructured (e.g., text, image) formats. So far, machines have been fairly good at processing and reasoning over structured data that follows a precise schema. However, the heterogeneity of data poses a significant challenge on how well diverse categories of data can be meaningfully stored and processed. Data Integration, a crucial part of the data engineering pipeline, addresses this by combining disparate data sources and providing unified data access to end-users. Until now, most data integration systems have leaned on only combining structured data sources. Nevertheless, unstructured data (a.k.a. free text) also contains a plethora of knowledge waiting to be utilized. Thus, in this chapter, we firstly make the case for the integration of textual data, to later present its challenges, state of the art and open problems.

Method: Introduces a scalable method to mitigate social-attribution bias by enriching instruction prompts with two prompt aids using social-attribution knowledge: (1) using user’s goal to infer dispositional causality, and (2) using message context to infer situational causality. Tested on intent detection and theme detection tasks in disaster domain social media across multiple languages.

Result: Method improves model performance while reducing social-attribution bias in zero-shot classification tasks for behavior analytics. Experiments show biases in three open-source LLMs (Llama3, Mistral, Gemma) toward social attribution, and demonstrate effectiveness of mitigation strategies across disaster types and multiple languages.

Conclusion: Incorporating social-attribution knowledge into LLM prompts effectively reduces bias and improves performance in social context reasoning tasks, particularly for behavior analytics applications in domains like disaster response across multiple languages.

Abstract: Attribution theory explains how individuals interpret and attribute others’ behavior in a social context by employing personal (dispositional) and impersonal (situational) causality. Large Language Models (LLMs), trained on human-generated corpora, may implicitly mimic this social attribution process in social contexts. However, the extent to which LLMs utilize these causal attributions in their reasoning remains underexplored. Although using reasoning paradigms, such as Chain-of-Thought (CoT), has shown promising results in various tasks, ignoring social attribution in reasoning could lead to biased responses by LLMs in social contexts. In this study, we investigate the impact of incorporating a user’s goal as knowledge to infer dispositional causality and message context to infer situational causality on LLM performance. To this end, we introduce a scalable method to mitigate such biases by enriching the instruction prompts for LLMs with two prompt aids using social-attribution knowledge, based on the context and goal of a social media message. This method improves the model performance while reducing the social-attribution bias of the LLM in the reasoning on zero-shot classification tasks for behavior analytics applications. We empirically show the benefits of our method across two tasks-intent detection and theme detection on social media in the disaster domain-when considering the variability of disaster types and multiple languages of social media. Our experiments highlight the biases of three open-source LLMs: Llama3, Mistral, and Gemma, toward social attribution, and show the effectiveness of our mitigation strategies.

Method: Created a meticulously structured dataset of 320 hand-written questions curated and validated by experts with musical training. Used this dataset to benchmark six state-of-the-art LALMs and tested their robustness to uni-modal shortcuts.

Result: The paper presents a new benchmark dataset and demonstrates its use through benchmarking of six state-of-the-art models. The results likely reveal performance gaps and highlight the importance of expert-curated evaluation for music understanding.

Conclusion: Focused, manual curation by musical experts is superior for probing complex audio comprehension in LALMs. The proposed benchmark provides a more rigorous standard for evaluating music understanding capabilities.

Abstract: The evaluation of music understanding in Large Audio-Language Models (LALMs) requires a rigorously defined benchmark that truly tests whether models can perceive and interpret music, a standard that current data methodologies frequently fail to meet. This paper introduces a meticulously structured approach to music evaluation, proposing a new dataset of 320 hand-written questions curated and validated by experts with musical training, arguing that such focused, manual curation is superior for probing complex audio comprehension. To demonstrate the use of the dataset, we benchmark six state-of-the-art LALMs and additionally test their robustness to uni-modal shortcuts.

Method: Introduces a multi-agent framework with architect, writer, refiner, and renderer agents coordinated around a persistent shared visual contract that tracks section structure and registered visual elements, with evaluation agents providing feedback in a generate-evaluate-adapt loop.

Result: Achieved expert evaluation score of 6.145 vs 3.963 for DirectChat (+2.182) across multiple LLM backbones, and 5.705 vs 5.197 for structured generation baseline Fars, showing improved structural consistency and visual alignment.

Conclusion: Story2Proposal demonstrates that contract-governed multi-agent frameworks can effectively maintain alignment between narrative reasoning, experimental evidence, and visual artifacts throughout the document generation lifecycle.

Abstract: Generating scientific manuscripts requires maintaining alignment between narrative reasoning, experimental evidence, and visual artifacts across the document lifecycle. Existing language-model generation pipelines rely on unconstrained text synthesis with validation applied only after generation, often producing structural drift, missing figures or tables, and cross-section inconsistencies. We introduce Story2Proposal, a contract-governed multi-agent framework that converts a research story into a structured manuscript through coordinated agents operating under a persistent shared visual contract. The system organizes architect, writer, refiner, and renderer agents around a contract state that tracks section structure and registered visual elements, while evaluation agents supply feedback in a generate evaluate adapt loop that updates the contract during generation. Experiments on tasks derived from the Jericho research corpus show that Story2Proposal achieved an expert evaluation score of 6.145 versus 3.963 for DirectChat (+2.182) across GPT, Claude, Gemini, and Qwen backbones. Compared with the structured generation baseline Fars, Story2Proposal obtained an average score of 5.705 versus 5.197, indicating improved structural consistency and visual alignment.

Method: The study analyzes layer pruning effects in Whisper encoder within SLAM-ASR systems across three Whisper variants (Small, Medium, Large-v2) and three languages representing different resource levels (Danish, Dutch, English). Over 200 training runs examine how LoRA-based fine-tuning compensates for pruning-induced performance degradation.

Result: Pruning two encoder layers causes only 2-4% WER degradation. Combining pruning with LoRA adaptation consistently outperforms unpruned baselines while reducing total parameters by 7-14%. LoRA reduces total word errors by 11-21% for Dutch and English, but only 4-7% for low-resource Danish, where it introduces increased insertion errors.

Conclusion: Pruning Whisper encoder layers in SLAM-ASR is viable with minimal performance loss, and LoRA adaptation can effectively compensate for degradation, especially in languages where the LLM has strong pre-existing proficiency. The effectiveness depends on the language model’s linguistic priors and available training data.

Abstract: Automatic speech recognition (ASR) has advanced rapidly in recent years, driven by large-scale pretrained models and end-to-end architectures such as SLAM-ASR. A key component of SLAM-ASR systems is the Whisper speech encoder, which provides robust acoustic representations. While model pruning has been explored for the full Whisper encoder-decoder architecture, its impact within the SLAM-ASR setting remains under-investigated. In this work, we analyze the effects of layer pruning in the Whisper encoder when used as the acoustic backbone of SLAM-ASR. We further examine the extent to which LoRA-based fine-tuning can recover performance degradation caused by pruning. Experiments conducted across three Whisper variants (Small, Medium, Large-v2), three languages representing distinct resource levels (Danish, Dutch, English), and over 200 training runs demonstrate that pruning two encoder layers causes only 2-4% WER degradation, and that combining this pruning with LoRA adaptation consistently outperforms the unpruned baseline while reducing total parameters by 7-14%. Moreover, our error analysis reveals that LoRA primarily compensates through the language model’s linguistic priors, reducing total word errors by 11-21% for Dutch and English, with substitutions and deletions showing the largest reductions. However, for low-resource Danish, the reduction is smaller (4-7%), and LoRA introduces increased insertion errors, indicating that compensation effectiveness depends on the LLM’s pre-existing language proficiency and available training data.

Method: Introduces FARE (Fairness-Aware Routing Equilibrium), a diagnostic framework to test routing-level stereotype interventions across various MoE architectures (Mixtral, Qwen1.5, Qwen3, DeepSeekMoE, OLMoE). Uses group-level expert masking to analyze bias/knowledge entanglement.

Result: Routing-level preference shifts are either unachievable (Mixtral, Qwen1.5, Qwen3), statistically non-robust (DeepSeekMoE), or come with substantial utility cost (OLMoE: -4.4%p CrowS-Pairs at -6.3%p TQA). Even when log-likelihood shifts are robust, they don’t transfer to decoded generation across all metrics.

Conclusion: Routing sensitivity is necessary but insufficient for stereotype control in MoE models. Bias and core knowledge are deeply entangled within expert groups, limiting routing-level interventions. Findings identify architectural conditions for designing more controllable future MoE systems.

Abstract: Mixture-of-Experts (MoE) language models are universally sensitive to demographic content at the routing level, yet exploiting this sensitivity for fairness control is structurally limited. We introduce Fairness-Aware Routing Equilibrium (FARE), a diagnostic framework designed to probe the limits of routing-level stereotype intervention across diverse MoE architectures. FARE reveals that routing-level preference shifts are either unachievable (Mixtral, Qwen1.5, Qwen3), statistically non-robust (DeepSeekMoE), or accompanied by substantial utility cost (OLMoE, -4.4%p CrowS-Pairs at -6.3%p TQA). Critically, even where log-likelihood preference shifts are robust, they do not transfer to decoded generation: expanded evaluations on both non-null models yield null results across all generation metrics. Group-level expert masking reveals why: bias and core knowledge are deeply entangled within expert groups. These findings indicate that routing sensitivity is necessary but insufficient for stereotype control, and identify specific architectural conditions that can inform the design of more controllable future MoE systems.

Method: Courtroom-style multi-agent framework with specialized roles (Plaintiff, Defense, Judge), Progressive RAG (P-RAG) for dynamic evidence expansion, evidence negotiation, self-reflection, and heterogeneous multi-judge aggregation.

Result: Achieves 81.7% accuracy on Check-COVID benchmark, outperforming standard multi-agent debate by 10.0 percentage points. P-RAG drives primary performance gains (+7.5 pp). Effectively mitigates systematic biases.

Conclusion: Structural deliberation and model heterogeneity provide robust foundation for reliable claim verification, addressing LLM limitations in high-stakes scenarios.

Abstract: Large language models (LLMs) remain unreliable for high-stakes claim verification due to hallucinations and shallow reasoning. While retrieval-augmented generation (RAG) and multi-agent debate (MAD) address this, they are limited by one-pass retrieval and unstructured debate dynamics. We propose a courtroom-style multi-agent framework, PROClaim, that reformulates verification as a structured, adversarial deliberation. Our approach integrates specialized roles (e.g., Plaintiff, Defense, Judge) with Progressive RAG (P-RAG) to dynamically expand and refine the evidence pool during the debate. Furthermore, we employ evidence negotiation, self-reflection, and heterogeneous multi-judge aggregation to enforce calibration, robustness, and diversity. In zero-shot evaluations on the Check-COVID benchmark, PROClaim achieves 81.7% accuracy, outperforming standard multi-agent debate by 10.0 percentage points, with P-RAG driving the primary performance gains (+7.5 pp). We ultimately demonstrate that structural deliberation and model heterogeneity effectively mitigate systematic biases, providing a robust foundation for reliable claim verification. Our code and data are publicly available at https://github.com/mnc13/PROClaim.

Method: Reframe proposal generation as time-sliced scientific forecasting: given research question and papers before cutoff time, model generates structured proposal evaluated by whether it anticipates research directions in papers published after cutoff. Use Future Alignment Score computed via retrieval and LLM-based semantic scoring against held-out future corpus. Build time-consistent dataset of 17,771 papers with pre-cutoff citations, synthesize reasoning traces teaching gap identification and inspiration borrowing.

Result: Future-aligned tuning improves future alignment over unaligned baselines (up to +10.6% overall FAS). Domain-expert human evaluation corroborates improved proposal quality. Practical impact: implemented two model-generated proposals with code agent, obtaining 4.17% accuracy gain on MATH from new prompting strategy and consistent improvements for novel model-merging method.

Conclusion: Time-sliced scientific forecasting provides verifiable evaluation framework for LLM-generated research proposals, with demonstrated improvements in proposal quality and practical impact.

Abstract: Large language models (LLMs) are increasingly used to assist ideation in research, but evaluating the quality of LLM-generated research proposals remains difficult: novelty and soundness are hard to measure automatically, and large-scale human evaluation is costly. We propose a verifiable alternative by reframing proposal generation as a time-sliced scientific forecasting problem. Given a research question and inspiring papers available before a cutoff time, the model generates a structured proposal and is evaluated by whether it anticipates research directions that appear in papers published after the time. We operationalize this objective with the Future Alignment Score (FAS), computed via retrieval and LLM-based semantic scoring against a held-out future corpus. To train models, we build a time-consistent dataset of 17,771 papers from targets and their pre-cutoff citations, and synthesize reasoning traces that teach gap identification and inspiration borrowing. Across Llama-3.1 and Qwen2.5 models, future-aligned tuning improves future alignment over unaligned baselines (up to +10.6% overall FAS), and domain-expert human evaluation corroborates improved proposal quality. Finally, we demonstrate practical impact by implementing two model-generated proposals with a code agent, obtaining 4.17% accuracy gain on MATH from a new prompting strategy and consistent improvements for a novel model-merging method.

Method: Proposes POTSA framework with: 1) Bias Compensation module for coarse speech representation alignment, 2) token-level Optimal Transport constraints on Q-Former using parallel speech pairs for fine-grained consistency, 3) layer scheduling strategy to focus OT constraints on semantically beneficial layers.

Result: Achieves SOTA on FLEURS benchmark: +1.29 BLEU over five common languages and +2.93 BLEU on zero-shot languages, using only 10 hours of parallel speech per language.

Conclusion: POTSA effectively bridges high- and low-resource translation gaps by leveraging semantic commonalities through Optimal Transport alignment with minimal parallel data.

Abstract: Speech Large Language Models have achieved breakthroughs in multilingual speech-to-text translation. However, existing approaches often overlook semantic commonalities across source languages, leading to biased translation performance. In this work, we propose POTSA (Parallel Optimal Transport for Speech Alignment), a new framework based on cross-lingual parallel speech pairs and Optimal Transport, designed to bridge high- and low-resource translation gaps. First, we introduce a Bias Compensation module to coarsely align initial speech representations. Second, we impose token-level OT constraints on a Q-Former using parallel pairs to establish fine-grained representation consistency. Then, we apply a layer scheduling strategy to focus OT constraints on semantically beneficial layers. Experiments on FLEURS show our method achieves SOTA performance, with +1.29 BLEU over five common languages and +2.93 BLEU on zero-shot languages, using only 10 hours of parallel speech per language.

Method: Systematic empirical study using synthetic arithmetic and logical benchmarks where difficulty is characterized by reasoning complexity rather than surface-level proxies. Tests across multiple model families and curriculum schedules, comparing difficulty-based sequencing with standard random sampling in both supervised fine-tuning (SFT) and reinforcement learning (RL) methods.

Result: Surprisingly, across all experiments, no robust advantage was found for difficulty-based sequencing over random sampling in either accuracy or response length. These findings persist across both SFT and RL methods, suggesting the specific ordering of training examples plays a negligible role in achieving compositional generalization.

Conclusion: Curriculum-based post-training may not provide practical utility for deductive reasoning tasks in LLMs, challenging the common assumption that learning in increasing order of difficulty eases generalization for compositional reasoning.

Abstract: Curriculum learning (CL), motivated by the intuition that learning in increasing order of difficulty should ease generalization, is commonly adopted both in pre-training and post-training of large language models (LLMs). The intuition of CL is particularly compelling for compositional reasoning, where complex problems are built from elementary inference rules; however, the actual impact of CL on such tasks remains largely underexplored. We present a systematic empirical study of CL for post-training of LLMs, using synthetic arithmetic and logical benchmarks where difficulty is characterized by reasoning complexity rather than surface-level proxies. Surprisingly, across multiple model families and curriculum schedules, we find no robust advantage in difficulty-based sequencing over standard random sampling in either accuracy or response length. These findings persist across both supervised fine-tuning (SFT) and reinforcement learning (RL) methods. Our study suggests that, in the context of deductive reasoning, the specific ordering of training examples plays a negligible role in achieving compositional generalization, challenging the practical utility of curriculum-based post-training.

Method: Collected 737 self-reported stress narratives from Afghan individuals, with participants describing experienced stress and selecting emotion/stressor labels via Dari checklists. The dataset includes 12 binary labels (5 emotions, 7 stressors). Baseline experiments used character TF-IDF with Linear SVM, compared with ParsBERT and XLM-RoBERTa models.

Result: Character TF-IDF with Linear SVM achieved Micro-F1 = 0.663 and Macro-F1 = 0.651, outperforming ParsBERT and XLM-RoBERTa. Threshold tuning improved Micro-F1 by 10.3 points. Structural stressors dominated, with uncertain future (62.6%) and education closure (60.0%) exceeding emotional states. Strongest co-occurrence was between hopelessness and uncertain future (J = 0.388).

Conclusion: AFSTRESS provides the first Dari resource for computational analysis of stress and well-being in a crisis-affected population, revealing that stress is primarily structurally driven rather than emotional, with structural stressors dominating the narratives.

Abstract: We introduce AFSTRESS, the first multi-label corpus of self-reported stress narratives in Dari (Eastern Persian), comprising 737 responses collected from Afghan individuals during an ongoing humanitarian crisis. Participants describe experienced stress and select emotion and stressor labels via Dari checklists. The dataset enables analysis at three levels: computational (multi-label classification), social (structural drivers and gender disparities), and psychological (learned helplessness, chronic stress, and emotional cascade patterns). It includes 12 binary labels (5 emotions, 7 stressors), with high label cardinality (5.54) and density (0.462), reflecting complex, multi-dimensional stress. Structural stressors dominate: uncertain future (62.6 percent) and education closure (60.0 percent) exceed emotional states, indicating stress is primarily structurally driven. The strongest co-occurrence is between hopelessness and uncertain future (J = 0.388). Baseline experiments show that character TF-IDF with Linear SVM achieves Micro-F1 = 0.663 and Macro-F1 = 0.651, outperforming ParsBERT and XLM-RoBERTa, while threshold tuning improves Micro-F1 by 10.3 points. AFSTRESS provides the first Dari resource for computational analysis of stress and well-being in a crisis-affected population.

Method: SCOPE uses a bi-directional tree structure for efficient template matching from both forward and reverse directions, and a two-stage framework: lightweight NLP with POS information for syntax-based matching, with LLM invoked selectively as fallback for semantically complex cases.

Result: Extensive evaluations on diverse benchmark datasets show SCOPE outperforms state-of-the-art methods in both accuracy and efficiency, significantly reducing LLM API usage while maintaining high accuracy.

Conclusion: SCOPE successfully integrates heuristic and LLM-based approaches to achieve a balance between efficiency and effectiveness in log parsing, with public implementation available for further research.

Abstract: Log parsing is a critical step for automated log analysis in complex systems. Traditional heuristic-based methods offer high efficiency but are limited in accuracy due to overlooking semantic context. In contrast, recent LLM-based parsers improve accuracy via se mantic understanding but incur high latency from frequent model calls. To address this, we propose SCOPE, the first self-correcting online log parsing method that integrates the strengths of both heuristic and LLM-based paradigms. SCOPE introduces a novel bi-directional tree structure that enables efficient template match ing from both forward and reverse directions, resulting in a higher overall matching rate. Additionally, it adopts a two-stage syntactic semantic collaboration framework: a lightweight NLP model first utilizes part-of-speech (POS) information for syntax-based match ing, while the LLM is selectively invoked as a fallback to handle semantically complex cases when uncertainty remains. This design significantly reduces LLM API usage while maintaining high ac curacy, achieving a balance between efficiency and effectiveness. Extensive evaluations on diverse benchmark datasets show that SCOPE outperforms state-of-the-art methods in both accuracy and efficiency. The implementation and datasets are publicly released to facilitate further research.

Method: Created Nwāchā Munā corpus (5.39 hours manually transcribed Devanagari speech). Investigated proximal cross-lingual transfer from Nepali vs. large multilingual pretraining. Fine-tuned Nepali Conformer model with data augmentation and compared with Whisper-Small.

Result: Fine-tuned Nepali Conformer reduced CER from 52.54% zero-shot baseline to 17.59% with augmentation, matching Whisper-Small performance despite using significantly fewer parameters. Proximal transfer proved computationally efficient.

Conclusion: Proximal cross-lingual transfer from geographically/linguistically adjacent languages can rival large multilingual models for ultra-low-resource ASR, offering computationally efficient alternatives. Dataset released to enable Newari community.

Abstract: Nepal Bhasha (Newari), an endangered language of the Kathmandu Valley, remains digitally marginalized due to the severe scarcity of annotated speech resources. In this work, we introduce Nwāchā Munā, a newly curated 5.39-hour manually transcribed Devanagari speech corpus for Nepal Bhasha, and establish the first benchmark using script-preserving acoustic modeling. We investigate whether proximal cross-lingual transfer from a geographically and linguistically adjacent language (Nepali) can rival large-scale multilingual pretraining in an ultra-low-resource Automatic Speech Recognition (ASR) setting. Fine-tuning a Nepali Conformer model reduces the Character Error Rate (CER) from a 52.54% zero-shot baseline to 17.59% with data augmentation, effectively matching the performance of the multilingual Whisper-Small model despite utilizing significantly fewer parameters. Our findings demonstrate that proximal transfer from Nepali language serves as a computationally efficient alternative to massive multilingual models. We openly release the dataset and benchmarks to digitally enable the Newari community and foster further research in Nepal Bhasha.

Method: Two-stage voting framework: (1) Retrieval Voting - multiple agents generate diverse queries in parallel and aggregate retrieved documents; (2) Response Voting - multiple agents independently generate answers based on aggregated documents, with final output determined by majority vote.

Result: VOTE-RAG achieves performance comparable to or surpassing more complex frameworks on six benchmark datasets, with simpler architecture, full parallelizability, and avoidance of problem drift risk.

Conclusion: Simple, reliable ensemble voting is a superior and more efficient method for mitigating RAG hallucinations compared to more complex approaches.

Abstract: Retrieval-Augmented Generation (RAG) aims to reduce hallucinations in Large Language Models (LLMs) by integrating external knowledge. However, RAG introduces a critical challenge: hallucination on hallucination," where flawed retrieval results mislead the generation model, leading to compounded hallucinations. To address this issue, we propose VOTE-RAG, a novel, training-free framework with a two-stage structure and efficient, parallelizable voting mechanisms. VOTE-RAG includes: (1) Retrieval Voting, where multiple agents generate diverse queries in parallel and aggregate all retrieved documents; (2) Response Voting, where multiple agents independently generate answers based on the aggregated documents, with the final output determined by majority vote. We conduct comparative experiments on six benchmark datasets. Our results show that VOTE-RAG achieves performance comparable to or surpassing more complex frameworks. Additionally, VOTE-RAG features a simpler architecture, is fully parallelizable, and avoids the problem drift" risk. Our work demonstrates that simple, reliable ensemble voting is a superior and more efficient method for mitigating RAG hallucinations.

Method: Three-stage approach: 1) Self-critic query refinement evaluates MeSH terms for coverage, alignment, and redundancy to enhance PubMed queries based on partial metadata retrieval; 2) Reflective retrieval processes articles in batches until sufficient evidence is gathered; 3) Evidence-grounded response generation produces answers with explicit citations.

Result: Achieves 78.32% accuracy on PubMedQA, slightly surpassing human experts, with consistent gains on MMLU Clinical Knowledge. LLM-as-judge evaluations show preference across reasoning soundness, evidence grounding, clinical relevance, and trustworthiness.

Conclusion: PubMed Reasoner provides practical assistance to clinicians and biomedical researchers by orchestrating retrieval-first reasoning over authoritative sources while controlling compute and token costs, offering a trustworthy biomedical QA system.

Abstract: Trustworthy biomedical question answering (QA) systems must not only provide accurate answers but also justify them with current, verifiable evidence. Retrieval-augmented approaches partially address this gap but lack mechanisms to iteratively refine poor queries, whereas self-reflection methods kick in only after full retrieval is completed. In this context, we introduce PubMed Reasoner, a biomedical QA agent composed of three stages: self-critic query refinement evaluates MeSH terms for coverage, alignment, and redundancy to enhance PubMed queries based on partial (metadata) retrieval; reflective retrieval processes articles in batches until sufficient evidence is gathered; and evidence-grounded response generation produces answers with explicit citations. PubMed Reasoner with a GPT-4o backbone achieves 78.32% accuracy on PubMedQA, slightly surpassing human experts, and showing consistent gains on MMLU Clinical Knowledge. Moreover, LLM-as-judge evaluations prefer our responses across: reasoning soundness, evidence grounding, clinical relevance, and trustworthiness. By orchestrating retrieval-first reasoning over authoritative sources, our approach provides practical assistance to clinicians and biomedical researchers while controlling compute and token costs.

Method: Human-in-the-loop framework that pairs English task instructions with dynamically retrieved target-language exemplars from filtered annotations, using an Exemplar Bank seeded from multilingual Information Disorder corpus annotations for adaptive prompting.

Result: Initial pilot compares static and adaptive prompting on Farsi and Italian news, evaluating span/severity prediction, rationale quality/cultural appropriateness, and model alignment across evaluator groups.

Conclusion: Provides a testbed for culturally grounded explainable AI by grounding model assessment in human-written rationales from native-speaking annotators through adaptive in-context learning.

Abstract: Recognizing information disorder is difficult because judgments about manipulation depend on cultural and linguistic context. Yet current Large Language Models (LLMs) often behave as monocultural, English-centric “black boxes,” producing fluent rationales that overlook localized framing. Preliminary evidence from the multilingual Information Disorder (InDor) corpus suggests that existing models struggle to explain manipulated news consistently across communities. To address this gap, this ongoing study proposes a Hybrid Intelligence Loop, a human-in-the-loop (HITL) framework that grounds model assessment in human-written rationales from native-speaking annotators. The approach moves beyond static target-language few-shot prompting by pairing English task instructions with dynamically retrieved target-language exemplars drawn from filtered InDor annotations through In-Context Learning (ICL). In the initial pilot, the Exemplar Bank is seeded from these filtered annotations and used to compare static and adaptive prompting on Farsi and Italian news. The study evaluates span and severity prediction, the quality and cultural appropriateness of generated rationales, and model alignment across evaluator groups, providing a testbed for culturally grounded explainable AI.

Method: The authors adopt graded summarization-based salience from Salient Entity Extraction, adapt it for proposition salience, define annotation tasks, apply to multi-genre dataset, evaluate agreement, and study relationship with discourse unit centrality in RST parsing.

Result: The paper presents a method for quantifying proposition salience, applies it to multi-genre data with evaluated agreement, and provides preliminary analysis of how proposition salience relates to discourse unit centrality in RST parsing.

Conclusion: The work operationalizes graded proposition salience using adapted metrics from entity extraction, providing a foundation for studying proposition importance in text and its relationship with discourse structure.

Abstract: Despite a long tradition of work on extractive summarization, which by nature aims to recover the most important propositions in a text, little work has been done on operationalizing graded proposition salience in naturally occurring data. In this paper, we adopt graded summarization-based salience as a metric from previous work on Salient Entity Extraction (SEE) and adapt it to quantify proposition salience. We define the annotation task, apply it to a small multi-genre dataset, evaluate agreement and carry out a preliminary study of the relationship between our metric and notions of discourse unit centrality in discourse parsing following Rhetorical Structure Theory (RST).

Method: Develop and employ structured, tag-based schemes to extract underlying implicit author intents (such as writing or citation intents). Use these extracted intents to enhance zero-shot generation capabilities in LLMs and create high-quality synthetic data for fine-tuning smaller models.

Result: Experiments show improved performance across various challenging scientific report generation tasks, with average improvements of +2.9 and +12.3 absolute points for large and small models over baselines, respectively. Intent awareness also enhances model citation usage and substantially improves report readability.

Conclusion: Enhancing LLMs’ intent awareness through structured tag-based schemes significantly improves the quality of generated long-form reports, benefiting both zero-shot generation and fine-tuning of smaller models while improving citation usage and readability.

Abstract: Large language models (LLMs) are increasingly being used to generate comprehensive, knowledge-intensive reports. However, while these models are trained on diverse academic papers and reports, they are not exposed to the reasoning processes and intents that guide authors in crafting these documents. We hypothesize that enhancing a model’s intent awareness can significantly improve the quality of generated long-form reports. We develop and employ structured, tag-based schemes to better elicit underlying implicit intents to write or cite. We demonstrate that these extracted intents enhance both zero-shot generation capabilities in LLMs and enable the creation of high-quality synthetic data for fine-tuning smaller models. Our experiments reveal improved performance across various challenging scientific report generation tasks, with an average improvement of +2.9 and +12.3 absolute points for large and small models over baselines, respectively. Furthermore, our analysis illuminates how intent awareness enhances model citation usage and substantially improves report readability.

Method: MAD-ACC uses a Proponent-Opponent-Judge model where agents defend conflicting interpretations of ambiguous text, exposing logical nuances through dialectical refinement. This multi-agent debate approach generates transparent reasoning transcripts.

Result: On the UKP Student Essays corpus, MAD-ACC achieves Macro F1 score of 85.7%, significantly outperforming single-agent reasoning baselines without requiring domain-specific training, while providing explainable decisions through debate transcripts.

Conclusion: Multi-agent debate frameworks like MAD-ACC offer an effective, training-free alternative to traditional supervised AM approaches, resolving structural ambiguity through dialectical refinement while providing transparency and explainability missing from “black-box” classifiers.

Abstract: Argument Mining (AM) is a foundational technology for automated writing evaluation, yet traditional supervised approaches rely heavily on expensive, domain-specific fine-tuning. While Large Language Models (LLMs) offer a training-free alternative, they often struggle with structural ambiguity, failing to distinguish between similar components like Claims and Premises. Furthermore, single-agent self-correction mechanisms often suffer from sycophancy, where the model reinforces its own initial errors rather than critically evaluating them. We introduce MAD-ACC (Multi-Agent Debate for Argument Component Classification), a framework that leverages dialectical refinement to resolve classification uncertainty. MAD-ACC utilizes a Proponent-Opponent-Judge model where agents defend conflicting interpretations of ambiguous text, exposing logical nuances that single-agent models miss. Evaluation on the UKP Student Essays corpus demonstrates that MAD-ACC achieves a Macro F1 score of 85.7%, significantly outperforming single-agent reasoning baselines, without requiring domain-specific training. Additionally, unlike “black-box” classifiers, MAD-ACC’s dialectical approach offers a transparent and explainable alternative by generating human-readable debate transcripts that explain the reasoning behind decisions.

Method: The approach demonstrates that each arc-standard transition (shift, leftarc, rightarc) corresponds to a deterministic tree update operation. For projective trees, this creates a unique ordered representation with surface-contiguous yields. For non-projective inputs, pseudo-projective lifting is used before derivation and inverse decoding after recovery.

Result: A proof-of-concept implementation in a standard neural transition-based parser shows that the mapped derivations are executable and support stable dependency recovery. The representation characterizes projectivity: a dependency tree admits such a contiguous ordered representation if and only if it is projective.

Conclusion: Arc-standard derivations provide a direct, derivational approach to constructing ordered tree representations from dependency parsing, establishing formal connections between dependency and constituency formalisms while maintaining recoverability of original dependency structures.

Abstract: We show that arc-standard derivations for projective dependency trees determine a unique ordered tree representation with surface-contiguous yields and stable lexical anchoring. Each \textsc{shift}, \textsc{leftarc}, and \textsc{rightarc} transition corresponds to a deterministic tree update, and the resulting hierarchical object uniquely determines the original dependency arcs. We further show that this representation characterizes projectivity: a single-headed dependency tree admits such a contiguous ordered representation if and only if it is projective. The proposal is derivational rather than convertive. It interprets arc-standard transition sequences directly as ordered tree construction, rather than transforming a completed dependency graph into a phrase-structure output. For non-projective inputs, the same interpretation can be used in practice via pseudo-projective lifting before derivation and inverse decoding after recovery. A proof-of-concept implementation in a standard neural transition-based parser shows that the mapped derivations are executable and support stable dependency recovery.

Method: Analyzes representational geometry of both refusal types by examining hidden-state vectors. Uses linear probing to investigate representational distinctions between harmful-refusal and over-refusal directions across transformer layers.

Result: Harmful-refusal directions are task-agnostic and captured by a single global vector, while over-refusal directions are task-dependent: they reside within benign task-representation clusters, vary across tasks, and span a higher-dimensional subspace.

Conclusion: The two refusal types are representationally distinct from early transformer layers, explaining why global direction ablation alone cannot address over-refusal. Task-specific geometric interventions are necessary to correct over-refusal while preserving proper refusal mechanisms.

Abstract: Aligned language models that are trained to refuse harmful requests also exhibit over-refusal: they decline safe instructions that seemingly resemble harmful instructions. A natural approach is to ablate the global refusal direction, steering the hidden-state vectors away or towards the harmful-refusal examples, but this corrects over-refusal only incidentally while disrupting the broader refusal mechanism. In this work, we analyse the representational geometry of both refusal types to understand why this happens. We show that harmful-refusal directions are task-agnostic and can be captured by a single global vector, whereas over-refusal directions are task-dependent: they reside within the benign task-representation clusters, vary across tasks, and span a higher-dimensional subspace. Linear probing confirms that the two refusal types are representationally distinct from the early transformer layers. These findings provide a mechanistic explanation of why global direction ablation alone cannot address over-refusal, and establish that task-specific geometric interventions are necessary.

Method: Proposes a multi-tier threat framework with three adversary capability levels: hard prompt injection, soft prompt optimization, and supervised fine-tuning. Uses teacher VLM-generated chain-of-thought reasoning to create natural trigger-slogan associations across semantic domains.

Result: Experiments on three VLM architectures show high injection efficacy with near-zero false positives while maintaining task accuracy. The attack is data-efficient, transfers to unseen datasets, and scales to multiple concurrent domain-slogan pairs.

Conclusion: Hidden Ads represents a practical security threat to VLMs in recommendation services, with current defenses (instruction-based filtering and clean fine-tuning) failing to remove the backdoor without significant utility degradation.

Abstract: Vision-Language Models (VLMs) are increasingly deployed in consumer applications where users seek recommendations about products, dining, and services. We introduce Hidden Ads, a new class of backdoor attacks that exploit this recommendation-seeking behavior to inject unauthorized advertisements. Unlike traditional pattern-triggered backdoors that rely on artificial triggers such as pixel patches or special tokens, Hidden Ads activates on natural user behaviors: when users upload images containing semantic content of interest (e.g., food, cars, animals) and ask recommendation-seeking questions, the backdoored model provides correct, helpful answers while seamlessly appending attacker-specified promotional slogans. This design preserves model utility and produces natural-sounding injections, making the attack practical for real-world deployment in consumer-facing recommendation services. We propose a multi-tier threat framework to systematically evaluate Hidden Ads across three adversary capability levels: hard prompt injection, soft prompt optimization, and supervised fine-tuning. Our poisoned data generation pipeline uses teacher VLM-generated chain-of-thought reasoning to create natural trigger–slogan associations across multiple semantic domains. Experiments on three VLM architectures demonstrate that Hidden Ads achieves high injection efficacy with near-zero false positives while maintaining task accuracy. Ablation studies confirm that the attack is data-efficient, transfers effectively to unseen datasets, and scales to multiple concurrent domain-slogan pairs. We evaluate defenses including instruction-based filtering and clean fine-tuning, finding that both fail to remove the backdoor without causing significant utility degradation.

Method: Provides a gentle tutorial with line-by-line execution trace of Bock’s original algorithm, introduces a structured reformulation that clarifies phase structure, state maintenance, and control flow, and includes a worked example for dependency parsing showing transformation to Bock’s formulation.

Result: Successfully makes Bock’s algorithm accessible with complete execution traces and structured reformulation, demonstrates practical application to dependency parsing through affine transformation of maximum weight to minimum cost problems.

Conclusion: Bock’s 1971 algorithm remains relevant for modern NLP applications, particularly as an exact decoder for nonprojective dependency parsing, and can be made accessible through careful tutorial presentation and structured reformulation.

Abstract: This paper presents a gentle tutorial and a structured reformulation of Bock’s 1971 Algol procedure for constructing minimum directed spanning trees. Our aim is to make the original algorithm readable and reproducible for modern readers, while highlighting its relevance as an exact decoder for nonprojective graph based dependency parsing. We restate the minimum arborescence objective in Bock’s notation and provide a complete line by line execution trace of the original ten node example, extending the partial trace given in the source paper from initialization to termination. We then introduce a structured reformulation that makes explicit the procedure’s phase structure, maintained state, and control flow, while preserving the logic of the original method. As a further illustration, we include a worked example adapted from {jurafsky-martin-2026-book} for dependency parsing, showing how a maximum weight arborescence problem is reduced to Bock’s minimum cost formulation by a standard affine transformation and traced under the same state variables.

Method: Two experiments: 1) Three language models reasoning under vocabulary constraints (No-Have and E-Prime) across seven tasks; 2) 16 linguistically constrained agents tackling debugging problems with ensemble analysis

Result: No-Have improved ethical reasoning by 19.1pp, classification by 6.5pp, and epistemic calibration by 7.4pp. E-Prime showed dramatic but model-dependent effects. Constrained agent ensembles achieved 100% ground-truth coverage vs 88.2% for control, with counterfactual agents being crucial

Conclusion: Umwelt engineering effectively alters cognition through cognitive restructuring and diversification, offering a new approach to agent design that can improve reasoning and problem-solving through linguistic constraints

Abstract: I propose Umwelt engineering – the deliberate design of the linguistic cognitive environment – as a third layer in the agent design stack, upstream of both prompt and context engineering. Two experiments test the thesis that altering the medium of reasoning alters cognition itself. In Experiment 1, three language models reason under two vocabulary constraints – No-Have (eliminating possessive “to have”) and E-Prime (eliminating “to be”) – across seven tasks (N=4,470 trials). No-Have improves ethical reasoning by 19.1 pp (p < 0.001), classification by 6.5 pp (p < 0.001), and epistemic calibration by 7.4 pp, while achieving 92.8% constraint compliance. E-Prime shows dramatic but model-dependent effects: cross-model correlations reach r = -0.75. In Experiment 2, 16 linguistically constrained agents tackle 17 debugging problems. No constrained agent outperforms the control individually, yet a 3-agent ensemble achieves 100% ground-truth coverage versus 88.2% for the control. A permutation test confirms only 8% of random 3-agent subsets achieve full coverage, and every successful subset contains the counterfactual agent. Two mechanisms emerge: cognitive restructuring and cognitive diversification. The primary limitation is the absence of an active control matching constraint prompt elaborateness.

Method: Created PRBench with 30 expert-curated tasks spanning 11 physics subfields, each grounded in real published papers. Tasks require agents to comprehend methodology, implement algorithms from scratch, and produce quantitative results matching original publications. Agents operate in sandboxed environments with only task instructions and paper content. Used agentified assessment pipeline to evaluate coding agents across scientific reasoning and execution dimensions.

Result: Best-performing agent (OpenAI Codex powered by GPT-5.3-Codex) achieved only 34% mean overall score. All agents had zero end-to-end callback success rate, with particularly poor performance in data accuracy and code correctness. Identified systematic failure modes including formula implementation errors, inability to debug numerical simulations, and fabrication of output data.

Conclusion: Current AI agents struggle significantly with end-to-end scientific paper reproduction, revealing substantial gaps in their capabilities for autonomous scientific research. PRBench provides a rigorous benchmark for measuring progress toward this goal and highlights key areas needing improvement.

Abstract: AI agents powered by large language models exhibit strong reasoning and problem-solving capabilities, enabling them to assist scientific research tasks such as formula derivation and code generation. However, whether these agents can reliably perform end-to-end reproduction from real scientific papers remains an open question. We introduce PRBench, a benchmark of 30 expert-curated tasks spanning 11 subfields of physics. Each task requires an agent to comprehend the methodology of a published paper, implement the corresponding algorithms from scratch, and produce quantitative results matching the original publication. Agents are provided only with the task instruction and paper content, and operate in a sandboxed execution environment. All tasks are contributed by domain experts from over 20 research groups at the School of Physics, Peking University, each grounded in a real published paper and validated through end-to-end reproduction with verified ground-truth results and detailed scoring rubrics. Using an agentified assessment pipeline, we evaluate a set of coding agents on PRBench and analyze their capabilities across key dimensions of scientific reasoning and execution. The best-performing agent, OpenAI Codex powered by GPT-5.3-Codex, achieves a mean overall score of 34%. All agents exhibit a zero end-to-end callback success rate, with particularly poor performance in data accuracy and code correctness. We further identify systematic failure modes, including errors in formula implementation, inability to debug numerical simulations, and fabrication of output data. Overall, PRBench provides a rigorous benchmark for evaluating progress toward autonomous scientific research.

Method: Introduces Budget-Xfer framework that formulates multi-source cross-lingual transfer as budget-constrained resource allocation problem, evaluating four allocation strategies across NER and sentiment analysis for three African languages using two multilingual models in 288 experiments.

Result: Multi-source transfer significantly outperforms single-source (Cohen’s d = 0.80 to 1.98), but differences among multi-source strategies are modest and non-significant. Embedding similarity as selection proxy is task-dependent - random selection outperforms similarity-based for NER but not sentiment analysis.

Conclusion: Budget-Xfer provides principled framework for cross-lingual transfer optimization, showing multi-source transfer is crucial but specific allocation strategies matter less than previously thought, with task-dependent value of language similarity metrics.

Abstract: Cross-lingual transfer learning enables NLP for low-resource languages by leveraging labeled data from higher-resource sources, yet existing comparisons of source language selection strategies do not control for total training data, confounding language selection effects with data quantity effects. We introduce Budget-Xfer, a framework that formulates multi-source cross-lingual transfer as a budget-constrained resource allocation problem. Given a fixed annotation budget B, our framework jointly optimizes which source languages to include and how much data to allocate from each. We evaluate four allocation strategies across named entity recognition and sentiment analysis for three African target languages (Hausa, Yoruba, Swahili) using two multilingual models, conducting 288 experiments. Our results show that (1) multi-source transfer significantly outperforms single-source transfer (Cohen’s d = 0.80 to 1.98), driven by a structural budget underutilization bottleneck; (2) among multi-source strategies, differences are modest and non-significant; and (3) the value of embedding similarity as a selection proxy is task-dependent, with random selection outperforming similarity-based selection for NER but not sentiment analysis.

Method: Develops information-theoretic metrics for diacritic complexity (frequency, ambiguity, structural diversity). Computes metrics over 24 corpora in 15 languages with single- and multi-diacritic scripts. Evaluates correlation with BERT- and RNN-based diacritic restoration performance.

Result: Higher diacritic complexity strongly correlates with lower restoration accuracy. In single-diacritic scripts, frequency and structural measures align. In multi-diacritic scripts, structural complexity shows strongest association with performance.

Conclusion: Measurable properties of diacritic usage influence diacritic restoration model performance, showing orthographic complexity is functionally relevant for modeling, not just descriptive.

Abstract: Diacritics are orthographic marks that clarify pronunciation, distinguish similar words, or alter meaning. They play a central role in many writing systems, yet their impact on language technology has not been systematically quantified across scripts. While prior work has examined diacritics in individual languages, there’s no cross-linguistic, data-driven framework for measuring the degree to which writing systems rely on them and how this affects downstream tasks. We propose a data-driven framework for quantifying diacritic complexity using corpus-level, information-theoretic metrics that capture the frequency, ambiguity, and structural diversity of character-diacritic combinations. We compute these metrics over 24 corpora in 15 languages, spanning both single- and multi-diacritic scripts. We then examine how diacritic complexity correlates with performance on the task of diacritics restoration, evaluating BERT- and RNN-based models. We find that across languages, higher diacritic complexity is strongly associated with lower restoration accuracy. In single-diacritic scripts, where character-diacritic combinations are more predictable, frequency-based and structural measures largely align. In multi-diacritic scripts, however, structural complexity exhibits the strongest association with performance, surpassing frequency-based measures. These findings show that measurable properties of diacritic usage influence the performance of diacritic restoration models, demonstrating that orthographic complexity is not only descriptive but functionally relevant for modeling.

Method: Evaluated GPT-4o mini, Gemini 1.5 Flash, and Claude 3.5 Haiku using tweet-like opinion prompts translated into Arabic, Chinese, French, Spanish, and Portuguese, analyzing sycophancy across languages and sensitive topics.

Result: Newer models show significantly less sycophancy overall than earlier generations, but sycophancy extent is still influenced by language, revealing systematic cultural and linguistic patterns in model agreeableness.

Conclusion: Progress in sycophancy mitigation is evident, but broader multilingual audits are needed for trustworthy, bias-aware LLM deployment due to persistent language-dependent sycophantic patterns.

Abstract: Large language models (LLMs) have achieved strong performance across a wide range of tasks, but they are also prone to sycophancy, the tendency to agree with user statements regardless of validity. Previous research has outlined both the extent and the underlying causes of sycophancy in earlier models, such as ChatGPT-3.5 and Davinci. Newer models have since undergone multiple mitigation strategies, yet there remains a critical need to systematically test their behavior. In particular, the effect of language on sycophancy has not been explored. In this work, we investigate how the language influences sycophantic responses. We evaluate three state-of-the-art models, GPT-4o mini, Gemini 1.5 Flash, and Claude 3.5 Haiku, using a set of tweet-like opinion prompts translated into five additional languages: Arabic, Chinese, French, Spanish, and Portuguese. Our results show that although newer models exhibit significantly less sycophancy overall compared to earlier generations, the extent of sycophancy is still influenced by the language. We further provide a granular analysis of how language shapes model agreeableness across sensitive topics, revealing systematic cultural and linguistic patterns. These findings highlight both the progress of mitigation efforts and the need for broader multilingual audits to ensure trustworthy and bias-aware deployment of LLMs.

Method: Created benchmark using longitudinal research trajectories of 217 AI researchers across diverse domains, using their scientific publications as externalized cognitive representations. Employed cross-domain, temporal-shift generalization setting to distinguish cognitive transfer from behavior imitation. Proposed multidimensional cognitive alignment metric for individual-level assessment.

Result: Systematic evaluation of state-of-the-art LLMs and enhancement techniques provides empirical study on how well LLMs simulate human cognition and how much existing techniques can enhance these capabilities.

Conclusion: The benchmark enables assessment of LLMs’ cognitive simulation capabilities beyond surface-level behavior imitation, offering tools to measure individual-level cognitive consistency and generalization across domains and time.

Abstract: An essential problem in artificial intelligence is whether LLMs can simulate human cognition or merely imitate surface-level behaviors, while existing datasets suffer from either synthetic reasoning traces or population-level aggregation, failing to capture authentic individual cognitive patterns. We introduce a benchmark grounded in the longitudinal research trajectories of 217 researchers across diverse domains of artificial intelligence, where each author’s scientific publications serve as an externalized representation of their cognitive processes. To distinguish whether LLMs transfer cognitive patterns or merely imitate behaviors, our benchmark deliberately employs a cross-domain, temporal-shift generalization setting. A multidimensional cognitive alignment metric is further proposed to assess individual-level cognitive consistency. Through systematic evaluation of state-of-the-art LLMs and various enhancement techniques, we provide a first-stage empirical study on the questions: (1) How well do current LLMs simulate human cognition? and (2) How far can existing techniques enhance these capabilities?

Method: Adopts “Specialize-then-Unify” paradigm: 1) Decomposes agentic coding into five expert domains, 2) Each domain undergoes independent supervised fine-tuning and reinforcement learning, 3) Consolidates experts into single model via on-policy distillation, 4) Uses KwaiEnv infrastructure for scalable sandbox training, 5) Proposes MCLA for stabilizing MoE RL training and Tree Training for computational efficiency.

Result: Achieves 79.6% on SWE-bench Verified (close to Claude Opus 4.6’s 80.8%), 88.7 on PinchBench (surpassing GLM-5 and MiniMax M2.7), ranks first in all three frontend aesthetics scenarios, and maintains strong scores on Terminal-Bench Hard (46.8) and tau^2-Bench (93.9).

Conclusion: KAT-Coder-V2 demonstrates that the “Specialize-then-Unify” paradigm with multi-domain expertise and scalable RL training can create powerful agentic coding models that excel across diverse coding benchmarks while maintaining computational efficiency.

Abstract: We present KAT-Coder-V2, an agentic coding model developed by the KwaiKAT team at Kuaishou. KAT-Coder-V2 adopts a “Specialize-then-Unify” paradigm that decomposes agentic coding into five expert domains - SWE, WebCoding, Terminal, WebSearch, and General - each undergoing independent supervised fine-tuning and reinforcement learning, before being consolidated into a single model via on-policy distillation. We develop KwaiEnv, a modular infrastructure sustaining tens of thousands of concurrent sandbox instances, and scale RL training along task complexity, intent alignment, and scaffold generalization. We further propose MCLA for stabilizing MoE RL training and Tree Training for eliminating redundant computation over tree-structured trajectories with up to 6.2x speedup. KAT-Coder-V2 achieves 79.6% on SWE-bench Verified (vs. Claude Opus 4.6 at 80.8%), 88.7 on PinchBench (surpassing GLM-5 and MiniMax M2.7), ranks first across all three frontend aesthetics scenarios, and maintains strong generalist scores on Terminal-Bench Hard (46.8) and tau^2-Bench (93.9). Our model is publicly available at https://streamlake.com/product/kat-coder.

Method: RT4CHART decomposes model outputs into independently verifiable claims and performs hierarchical, local-to-global verification against retrieved context. Each claim is labeled as entailed, contradicted, or baseless, with explicit evidence mapping back to answer spans.

Result: Achieves best answer-level hallucination detection F1 among baselines: 0.776 on RAGTruth++ (83% improvement over strongest baseline) and 47.5% span-level F1 on RAGTruth-Enhance. Re-annotation reveals 1.68x more hallucination cases than original labels.

Conclusion: RT4CHART provides fine-grained, interpretable auditing for RAG hallucinations through hierarchical verification. Existing benchmarks substantially underestimate hallucination prevalence, highlighting the need for more rigorous evaluation frameworks.

Abstract: Large language models (LLMs) continue to hallucinate in retrieval-augmented generation (RAG), producing claims that are unsupported by or conflict with the retrieved context. Detecting such errors remains challenging when faithfulness is evaluated solely with respect to the retrieved context. Existing approaches either provide coarse-grained, answer-level scores or focus on open-domain factuality, often lacking fine-grained, evidence-grounded diagnostics. We present RT4CHART, a retromorphic testing framework for context-faithfulness assessment. RT4CHART decomposes model outputs into independently verifiable claims and performs hierarchical, local-to-global verification against the retrieved context. Each claim is assigned one of three labels: entailed, contradicted, or baseless. Furthermore, RT4CHART maps claim-level decisions back to specific answer spans and retrieves explicit supporting or refuting evidence from the context, enabling fine-grained and interpretable auditing. We evaluate RT4CHART on RAGTruth++ (408 samples) and RAGTruth-Enhance (2,675 samples), a newly re-annotated benchmark. RT4CHART achieves the best answer-level hallucination detection F1 among all baselines. On RAGTruth++, it reaches an F1 score of 0.776, outperforming the strongest baseline by 83%. On RAGTruth-Enhance, it achieves a span-level F1 of 47.5%. Ablation studies show that the hierarchical verification design is the primary driver of performance gains. Finally, our re-annotation reveals 1.68x more hallucination cases than the original labels, suggesting that existing benchmarks substantially underestimate the prevalence of hallucinations.

Method: Introduce TailNLG, a new multilingual benchmark in English, Italian, and Spanish built from Wikidata covering entities with varying popularity levels. Evaluate three families of large language models in zero-shot settings, comparing performance on rare vs common entities and against established WebNLG benchmark.

Result: Reveals consistent bias against long-tail entities: embedding-based scores are lower and model uncertainty is higher for rare entities. Impact varies across models and languages, and existing evaluation metrics don’t consistently capture these differences.

Conclusion: Highlights the need for more reliable evaluation frameworks to address biases in Data-to-Text generation, particularly for long-tail entities that are crucial for comprehensive knowledge graph verbalization.

Abstract: The automatic verbalization of structured knowledge is a key task for making knowledge graphs accessible to non-expert users and supporting retrieval-augmented generation systems. Although recent advances in Data-to-Text generation have improved multilingual coverage, little attention has been paid to potential biases in the verbalization of rare entities, frequently known as long-tail entities. In this work, we present the first systematic study of long-tail entities in Data-to-Text generation. We introduce TailNLG, a new multilingual benchmark in English, Italian, and Spanish, built from Wikidata and covering entities with varying levels of popularity. We evaluate three different families of large language models in zero-shot settings and compare their performance on rare versus common entities, as well as against the established WebNLG benchmark. Our results reveal a consistent bias against long-tail entities: embedding-based scores are lower, and model uncertainty is higher for rare entities. We further show that the impact of long-tail entities varies across models and languages, and that existing evaluation metrics do not consistently capture these differences, highlighting the need for more reliable evaluation frameworks.

Method: Analysis of 1,349 AI-generated feedback instances and corresponding teacher edits from 117 teachers, examining textual characteristics, predicting revision decisions using machine learning models with sentence embeddings, and qualitative coding of pedagogical feedback types.

Result: Teachers accept 80% of AI feedback unchanged; edited feedback tends to be longer initially but gets shortened; editing behavior varies widely (50% never edit, 10% edit >2/3); ML models achieve AUC=0.75 for predicting revisions; teachers often simplify AI feedback, shifting from explanatory to corrective forms.

Conclusion: Findings characterize teacher engagement with AI-generated feedback and highlight opportunities to design feedback systems that better align with teacher priorities while reducing unnecessary editing effort.

Abstract: Large language models (LLMs) increasingly generate formative feedback for students, yet little is known about how teachers revise this feedback before it reaches learners. Teachers’ revisions shape what students receive, making revision practices central to evaluating AI classroom tools. We analyze a dataset of 1,349 instances of AI-generated feedback and corresponding teacher-edited explanations from 117 teachers. We examine (i) textual characteristics associated with teacher revisions, (ii) whether revision decisions can be predicted from the AI feedback text, and (iii) how revisions change the pedagogical type of feedback delivered. First, we find that teachers accept AI feedback without modification in about 80% of cases, while edited feedback tends to be significantly longer and subsequently shortened by teachers. Editing behavior varies substantially across teachers: about 50% never edit AI feedback, and only about 10% edit more than two-thirds of feedback instances. Second, machine learning models trained only on the AI feedback text as input features, using sentence embeddings, achieve fair performance in identifying which feedback will be revised (AUC=0.75). Third, qualitative coding shows that when revisions occur, teachers often simplify AI-generated feedback, shifting it away from high-information explanations toward more concise, corrective forms. Together, these findings characterize how teachers engage with AI-generated feedback in practice and highlight opportunities to design feedback systems that better align with teacher priorities while reducing unnecessary editing effort.

Method: The paper traces the historical evolution of NLP from its beginnings to the current era of large language models, comparing each major paradigm with theories of human language capacity from linguistics and cognitive science.

Result: The analysis reveals that despite the impressive language abilities achieved by modern chatbots using artificial neural networks, the evolution of language technology has not substantially deepened our understanding of how human minds process natural language.

Conclusion: Current NLP systems, including large language models, operate on fundamentally different principles than human language processing, and their success doesn’t translate to better understanding of human linguistic cognition.

Abstract: In this paper, we discuss the relationship between natural language processing by computers (NLP) and the understanding of the human language capacity, as studied by linguistics and cognitive science. We outline the evolution of NLP from its beginnings until the age of large language models, and highlight for each of its main paradigms some similarities and differences with theories of the human language capacity. We conclude that the evolution of language technology has not substantially deepened our understanding of how human minds process natural language, despite the impressive language abilities attained by current chatbots using artificial neural networks.

Method: Proposes a counterfactual multi-agent diagnostic framework with: 1) Counterfactual case editing to modify clinical findings and evaluate how changes affect competing diagnoses, 2) Counterfactual Probability Gap to quantify how strongly individual findings support a diagnosis by measuring confidence shifts under edits, and 3) Multi-round specialist discussions guided by these counterfactual signals to challenge unsupported hypotheses and refine differential diagnoses.

Result: Across three diagnostic benchmarks and seven LLMs, the method consistently improves diagnostic accuracy over prompting and prior multi-agent baselines, with largest gains in complex and ambiguous cases. Human evaluation indicates the framework produces more clinically useful, reliable, and coherent reasoning.

Conclusion: Incorporating counterfactual evidence verification is an important step toward building reliable AI systems for clinical decision support, making LLM-based diagnostic reasoning more interpretable and clinically grounded.

Abstract: Clinical diagnosis is a complex reasoning process in which clinicians gather evidence, form hypotheses, and test them against alternative explanations. In medical training, this reasoning is explicitly developed through counterfactual questioning–e.g., asking how a diagnosis would change if a key symptom were absent or altered–to strengthen differential diagnosis skills. As large language model (LLM)-based systems are increasingly used for diagnostic support, ensuring the interpretability of their recommendations becomes critical. However, most existing LLM-based diagnostic agents reason over fixed clinical evidence without explicitly testing how individual findings support or weaken competing diagnoses. In this work, we propose a counterfactual multi-agent diagnostic framework inspired by clinician training that makes hypothesis testing explicit and evidence-grounded. Our framework introduces counterfactual case editing to modify clinical findings and evaluate how these changes affect competing diagnoses. We further define the Counterfactual Probability Gap, a method that quantifies how strongly individual findings support a diagnosis by measuring confidence shifts under these edits. These counterfactual signals guide multi-round specialist discussions, enabling agents to challenge unsupported hypotheses, refine differential diagnoses, and produce more interpretable reasoning trajectories. Across three diagnostic benchmarks and seven LLMs, our method consistently improves diagnostic accuracy over prompting and prior multi-agent baselines, with the largest gains observed in complex and ambiguous cases. Human evaluation further indicates that our framework produces more clinically useful, reliable, and coherent reasoning. These results suggest that incorporating counterfactual evidence verification is an important step toward building reliable AI systems for clinical decision support.

Method: Created ProText dataset spanning three dimensions: Theme nouns (names, occupations, titles, kinship terms), Theme category (male-stereotyped, female-stereotyped, gender-neutral), and Pronoun category (masculine, feminine, gender-neutral, none). Validated through case studies using prompts and state-of-the-art LLMs.

Result: The dataset enables nuanced insights into gender bias, stereotyping, misgendering, and gendering. Case studies revealed systematic gender bias, particularly when inputs lack explicit gender cues or when models default to heteronormative assumptions.

Conclusion: ProText provides a valuable resource for measuring gender bias in LLM text transformations, extending beyond traditional pronoun resolution benchmarks and binary gender frameworks to capture more complex forms of gendering and misgendering.

Abstract: We introduce ProText, a dataset for measuring gendering and misgendering in stylistically diverse long-form English texts. ProText spans three dimensions: Theme nouns (names, occupations, titles, kinship terms), Theme category (stereotypically male, stereotypically female, gender-neutral/non-gendered), and Pronoun category (masculine, feminine, gender-neutral, none). The dataset is designed to probe (mis)gendering in text transformations such as summarization and rewrites using state-of-the-art Large Language Models, extending beyond traditional pronoun resolution benchmarks and beyond the gender binary. We validated ProText through a mini case study, showing that even with just two prompts and two models, we can draw nuanced insights regarding gender bias, stereotyping, misgendering, and gendering. We reveal systematic gender bias, particularly when inputs contain no explicit gender cues or when models default to heteronormative assumptions.

Method: Tested Diverse Prompt Mixer in AIMO~3 competition with 3 models, 23+ experiments, and 50 IMO-level problems on a single H100 80GB GPU with 5-hour limit. Compared different prompt strategies and inference-time optimizations.

Result: Every intervention failed. High-temperature sampling already decorrelates errors sufficiently; weaker prompt strategies reduce per-attempt accuracy more than they reduce correlation. Model capability dominates by an order of magnitude across a 17-point model capability gap.

Conclusion: For mathematical reasoning with LLMs, model capability is far more important than sophisticated voting strategies. Simple high-temperature sampling provides sufficient error decorrelation, and attempts to diversify reasoning strategies don’t yield benefits.

Abstract: Majority voting over multiple LLM attempts improves mathematical reasoning, but correlated errors limit the effective sample size. A natural fix: assign structurally different reasoning strategies to different voters to decorrelate errors. We test this Diverse Prompt Mixer in the AIMO~3 competition: 3 models, 23+ experiments, and 50 IMO-level problems on a single H100 80 GB with a 5-hour limit. Every intervention fails. High-temperature sampling already decorrelates errors sufficiently; weaker prompt strategies reduce per-attempt accuracy more than they reduce correlation. Across a 17-point model capability gap and every inference-time optimization we tried, model capability dominates by an order of magnitude.

Method: Used the Pythia scaling suite to test different sized language models on two polarity illusions, analyzing how model performance changes with increasing scale.

Result: NPI illusion becomes weaker and disappears as model size increases, while depth charge illusion becomes stronger in larger models.

Conclusion: Human polarity illusions may not require “rational inference” mechanisms; shallow processing and partial grammaticalization of ungrammatical structures in LLMs suggest alternative explanations rooted in construction grammar.

Abstract: I use the Pythia scaling suite (Biderman et al. 2023) to investigate if and how two well-known polarity illusions, the NPI illusion and the depth charge illusion, arise in LLMs. The NPI illusion becomes weaker and ultimately disappears as model size increases, while the depth charge illusion becomes stronger in larger models. The results have implications for human sentence processing: it may not be necessary to assume “rational inference” mechanisms that convert ill-formed sentences into well-formed ones to explain polarity illusions, given that LLMs cannot plausibly engage in this kind of reasoning, especially at the implicit level of next-token prediction. On the other hand, shallow, “good enough” processing and/or partial grammaticalization of prescriptively ungrammatical structures may both occur in LLMs. I propose a synthesis of different theoretical accounts that is rooted in the basic tenets of construction grammar.

Method: Two-stage approach: 1) Train small adapter to process raw bytes and interface with frozen Qwen2.5-7B, 2) Freeze adapter and fine-tune only attention layers on Kazakh text to adapt the model.

Result: Empirical validation is ongoing; this paper presents the ByteKaz architecture and training protocol as a design proposal with hypotheses to be tested.

Conclusion: ByteKaz offers a promising approach to overcome tokenizer limitations for low-resource languages like Kazakh, potentially matching or exceeding original model performance on Kazakh benchmarks through specialized adaptation.

Abstract: Large language models fragment Kazakh text into many more tokens than equivalent English text, because their tokenizers were built for high-resource languages. This tokenizer tax inflates compute, shortens the effective context window, and weakens the model’s grip on Kazakh morphology. We propose to bypass the tokenizer entirely by feeding raw bytes through a small adapter that learns to speak the internal language of a frozen Qwen2.5-7B. Once the adapter is trained, we freeze it and fine-tune only the attention layers of Qwen on Kazakh text. Our central hypothesis is that this two-stage process – first teach the interface, then adapt the model – should match or exceed the accuracy of the original Qwen2.5-7B on standard Kazakh benchmarks. This report describes the ByteKaz architecture and training protocol. Empirical validation is ongoing; this version stakes the design and hypotheses for the record.

Method: Analyzed 1M comments across 2.7K news articles, operationalizing quality as comment health (constructive, good-faith contributions). Used statistical analysis to examine how article frames predict comment health while controlling for topic.

Result: Article frames significantly predict comment health. Comments that adopt the article frame are healthier than those that depart from it. Unhealthy top-level comments tend to generate more unhealthy responses, independent of the frame being used.

Conclusion: Establishes a link between framing theory and discourse quality, laying groundwork for downstream applications. Demonstrates potential with a proactive frame-aware LLM-based system to mitigate unhealthy discourse.

Abstract: Framing theory posits that how information is presented shapes audience responses, but computational work has largely ignored audience reactions. While recent work showed that article framing systematically shapes the content of reader responses, this paper asks: Does framing also affect response quality? Analyzing 1M comments across 2.7K news articles, we operationalize quality as comment health (constructive, good-faith contributions). We find that article frames significantly predict comment health while controlling for topic, and that comments that adopt the article frame are healthier than those that depart from it. Further, unhealthy top-level comments tend to generate more unhealthy responses, independent of the frame being used in the comment. Our results establish a link between framing theory and discourse quality, laying the groundwork for downstream applications. We illustrate this potential with a proactive frame-aware LLM- based system to mitigate unhealthy discourse

Method: Proposes a syntactic decoder that generates target-language dependency trees in a top-down, left-to-right order, moving from conventional sequence-to-sequence to string-to-tree decoding.

Result: Experiments show the top-down string-to-tree decoding generalizes better than conventional sequence-to-sequence decoding for translating long inputs not observed in training data.

Conclusion: Incorporating syntactic structure through top-down tree generation improves NMT performance on challenging long inputs, addressing generalization issues in conventional models.

Abstract: Most of modern neural machine translation (NMT) models are based on an encoder-decoder framework with an attention mechanism. While they perform well on standard datasets, they can have trouble in translation of long inputs that are rare or unseen during training. Incorporating target syntax is one approach to dealing with such length-related problems. We propose a novel syntactic decoder that generates a target-language dependency tree in a top-down, left-to-right order. Experiments show that the proposed top-down string-to-tree decoding generalizes better than conventional sequence-to-sequence decoding in translating long inputs that are not observed in the training data.

Method: EnsemJudge framework incorporates tailored strategies and ensemble voting mechanisms specifically designed for Chinese text detection, trained on a carefully constructed Chinese dataset from NLPCC2025 Shared Task 1.

Result: Outperformed all baseline methods and achieved first place in NLPCC2025 Shared Task 1, demonstrating effectiveness and reliability in Chinese LLM-generated text detection.

Conclusion: EnsemJudge provides a robust solution for Chinese LLM-generated text detection, addressing gaps in current research and showing practical applicability in real-world scenarios.

Abstract: Large Language Models (LLMs) are widely applied across various domains due to their powerful text generation capabilities. While LLM-generated texts often resemble human-written ones, their misuse can lead to significant societal risks. Detecting such texts is an essential technique for mitigating LLM misuse, and many detection methods have shown promising results across different datasets. However, real-world scenarios often involve out-of-domain inputs or adversarial samples, which can affect the performance of detection methods to varying degrees. Furthermore, most existing research has focused on English texts, with limited work addressing Chinese text detection. In this study, we propose EnsemJudge, a robust framework for detecting Chinese LLM-generated text by incorporating tailored strategies and ensemble voting mechanisms. We trained and evaluated our system on a carefully constructed Chinese dataset provided by NLPCC2025 Shared Task 1. Our approach outperformed all baseline methods and achieved first place in the task, demonstrating its effectiveness and reliability in Chinese LLM-generated text detection. Our code is available at https://github.com/johnsonwangzs/MGT-Mini.

Method: Compared self-decomposing atomic judge (single-prompt decompose-and-verify) against prompt-controlled holistic judge with same inputs and detailed rubric. Tested on 200 source examples each from TruthfulQA, ASQA, and QAMPARI datasets using four model families, with source-level paired tests, cluster bootstrap, and aggregation across three pre-frozen prompt variants per design.

Result: Holistic judge matches or exceeds atomic judge on two of three benchmarks (ASQA and QAMPARI favor holistic across all four model families, statistically reliable in three of four), while TruthfulQA shows small atomic edge. Holistic advantage concentrated in partially_supported cases (incompleteness detection).

Conclusion: For self-decomposing single-prompt patterns on QA-style benchmarks, holistic judges can perform as well or better than atomic judges, particularly for detecting partial support. The advantage of atomic decomposition may stem more from richer prompting than decomposition itself.

Abstract: Atomic decomposition – breaking a candidate answer into claims before verifying each against a reference – is a widely adopted design for LLM-based reference-grounded judges. However, atomic prompts are typically richer and longer, making it unclear whether any advantage comes from decomposition or from richer prompting. We study this for benchmark-style completeness-sensitive reference-support classification: classifying a candidate as fully supported, partially supported, or unsupported relative to a supplied reference. We compare a self-decomposing atomic judge (single-prompt decompose-and-verify) against a prompt-controlled holistic judge with the same inputs and a similarly detailed rubric. On 200 source examples per dataset across TruthfulQA, ASQA, and QAMPARI, with four model families, source-level paired tests, cluster bootstrap, and aggregation across three pre-frozen prompt variants per design family, we find the holistic judge matches or exceeds the atomic judge on two of three benchmarks: ASQA and QAMPARI favor holistic across all four families (statistically reliable in three of four), while TruthfulQA shows a small atomic edge. The holistic advantage is concentrated in partially_supported cases – incompleteness detection. A sensitivity check against human annotations confirms the ranking under both benchmark-completeness and human factual-correctness standards. Our finding is specific to the self-decomposing single-prompt pattern on three QA-style benchmarks with 200 source examples each; multi-stage atomic pipelines and non-QA tasks remain untested. Among perturbations examined, reference-quality degradation produced the largest accuracy drops for both judge families.

Method: Two experimental phases: 1) Zero-shot transfer evaluation of a parser trained on Greek-variety UD data to Turkish-variety Pomak, quantifying phonological and morphosyntactic differences; 2) Introducing a new manually annotated Turkish-variety corpus (650 sentences) and using targeted fine-tuning and cross-variety transfer learning.

Result: Targeted fine-tuning on the small Turkish-variety corpus substantially improves accuracy, and performance is further boosted by cross-variety transfer learning that combines both dialects.

Conclusion: The paper demonstrates that even small amounts of targeted data can significantly improve cross-dialect parsing performance, and that combining resources from different dialects through transfer learning yields the best results for low-resource endangered languages.

Abstract: This paper presents new resources and baselines for Dependency Parsing in Pomak, an endangered Eastern South Slavic language with substantial dialectal variation and no widely adopted standard. We focus on the variety spoken in Turkey (Uzunköprü) and ask how well a dependency parser trained on the existing Pomak Universal Dependencies treebank, which was built primarily from the variety that is spoken in Greece, transfers across dialects. We run two experimental phases. First, we train a parser on the Greek-variety UD data and evaluate zero-shot transfer to Turkish-variety Pomak, quantifying the impact of phonological and morphosyntactic differences. Second, we introduce a new manually annotated Turkish-variety Pomak corpus of 650 sentences and show that, despite its small size, targeted fine-tuning substantially improves accuracy; performance is further boosted by cross-variety transfer learning that combines the two dialects.

Method: TRACE creates interpretable fingerprints by capturing token-level transition patterns (like word rank) using a lightweight language model, working for both open- and closed-source models.

Result: TRACE achieves state-of-the-art performance on GhostWriteBench, remains robust in out-of-distribution settings, and works well with limited training data.

Conclusion: The proposed TRACE method provides an effective, interpretable, and lightweight solution for LLM authorship attribution that generalizes well across different domains and unseen models.

Abstract: In this paper, we introduce GhostWriteBench, a dataset for LLM authorship attribution. It comprises long-form texts (50K+ words per book) generated by frontier LLMs, and is designed to test generalisation across multiple out-of-distribution (OOD) dimensions, including domain and unseen LLM author. We also propose TRACE – a novel fingerprinting method that is interpretable and lightweight – that works for both open- and closed-source models. TRACE creates the fingerprint by capturing token-level transition patterns (e.g., word rank) estimated by another lightweight language model. Experiments on GhostWriteBench demonstrate that TRACE achieves state-of-the-art performance, remains robust in OOD settings, and works well in limited training data scenarios.

Method: Evaluated LLMs (GPT-3.5 Turbo, Gemini 2.0 Flash, Mistral 7B Instruct) on Mini-BAR dataset of 1,000+ health app reviews using zero-shot, one-shot, and two-shot prompting. Evaluated generated user stories using human judgment (RUST framework) and RoBERTa classifier fine-tuned on UStAI.

Result: LLMs can match or outperform humans in writing fluent, well-formatted user stories, especially with few-shot prompts. However, they struggle to produce independent and unique user stories essential for agile backlogs.

Conclusion: LLMs can reliably turn unstructured app reviews into actionable software requirements, providing developers with clear guidance to turn user feedback into meaningful improvements, though limitations remain in story independence and uniqueness.

Abstract: App store reviews provide a constant flow of real user feedback that can help improve software requirements. However, these reviews are often messy, informal, and difficult to analyze manually at scale. Although automated techniques exist, many do not perform well when replicated and often fail to produce clean, backlog-ready user stories for agile projects. In this study, we evaluate how well large language models (LLMs) such as GPT-3.5 Turbo, Gemini 2.0 Flash, and Mistral 7B Instruct can generate usable user stories directly from raw app reviews. Using the Mini-BAR dataset of 1,000+ health app reviews, we tested zero-shot, one-shot, and two-shot prompting methods. We evaluated the generated user stories using both human judgment (via the RUST framework) and a RoBERTa classifier fine-tuned on UStAI to assess their overall quality. Our results show that LLMs can match or even outperform humans in writing fluent, well-formatted user stories, especially when few-shot prompts are used. However, they still struggle to produce independent and unique user stories, which are essential for building a strong agile backlog. Overall, our findings show how LLMs can reliably turn unstructured app reviews into actionable software requirements, providing developers with clear guidance to turn user feedback into meaningful improvements.

Method: 1) Curated three ICWM datasets (SSDF-Syndrome, SSDF-Dialogue, SSDF-PD); 2) Developed SSDF-Core diagnostic LLM with two-stage training (SFT + DPO); 3) Created SSDF-Navigator pluggable consultation navigation model; 4) Established SSDF-Bench evaluation benchmark.

Result: SSDF-Core significantly outperforms 12 mainstream baselines on SSDF-Bench, demonstrating robust ICWM reasoning capabilities for spleen-stomach disorder diagnosis.

Conclusion: DongYuan provides a solid methodological foundation and practical technical references for developing intelligent ICWM diagnostic systems, addressing key challenges in medical AI applications.

Abstract: The clinical burden of spleen-stomach disorders is substantial. While large language models (LLMs) offer new potential for medical applications, they face three major challenges in the context of integrative Chinese and Western medicine (ICWM): a lack of high-quality data, the absence of models capable of effectively integrating the reasoning logic of traditional Chinese medicine (TCM) syndrome differentiation with that of Western medical (WM) disease diagnosis, and the shortage of a standardized evaluation benchmark. To address these interrelated challenges, we propose DongYuan, an ICWM spleen-stomach diagnostic framework. Specifically, three ICWM datasets (SSDF-Syndrome, SSDF-Dialogue, and SSDF-PD) were curated to fill the gap in high-quality data for spleen-stomach disorders. We then developed SSDF-Core, a core diagnostic LLM that acquires robust ICWM reasoning capabilities through a two-stage training regimen of supervised fine-tuning. tuning (SFT) and direct preference optimization (DPO), and complemented it with SSDF-Navigator, a pluggable consultation navigation model designed to optimize clinical inquiry strategies. Additionally, we established SSDF-Bench, a comprehensive evaluation benchmark focused on ICWM diagnosis of spleen-stomach disorders. Experimental results demonstrate that SSDF-Core significantly outperforms 12 mainstream baselines on SSDF-Bench. DongYuan lays a solid methodological foundation and provides practical technical references for the future development of intelligent ICWM diagnostic systems.

Method: Proposes Zero-Initialized Residual Complex Projection (ZRCP) to project textual features into complex semantic space, using phase to disentangle sentiment polarities while amplitude encodes semantic intensity. Introduces Anti-collision Masked Angle Loss with anti-collision mask to preserve intra-polarity cohesion while expanding inter-polarity discriminative margin.

Result: Achieves state-of-the-art Macro-F1 score of 0.8851. Deep geometric analyses show that constraining amplitude catastrophically over-regularizes subjective representations, proving the importance of unconstrained-amplitude and phase-driven approach.

Conclusion: The framework effectively disentangles aspect semantics from sentiment polarities using complex projection and anti-collision mechanisms, providing robust fine-grained sentiment analysis with superior performance.

Abstract: Aspect-Based Sentiment Analysis (ABSA) is fundamentally challenged by representation entanglement, where aspect semantics and sentiment polarities are often conflated in real-valued embedding spaces. Furthermore, standard contrastive learning suffers from false-negative collisions, severely degrading performance on high-frequency aspects. In this paper, we propose a novel framework featuring a Zero-Initialized Residual Complex Projection (ZRCP) and an Anti-collision Masked Angle Loss,inspired by quantum projection and entanglement ideas. Our approach projects textual features into a complex semantic space, systematically utilizing the phase to disentangle sentiment polarities while allowing the amplitude to encode the semantic intensity and lexical richness of subjective descriptions. To tackle the collision bottleneck, we introduce an anti-collision mask that elegantly preserves intra-polarity aspect cohesion while expanding the inter-polarity discriminative margin by over 50%. Experimental results demonstrate that our framework achieves a state-of-the-art Macro-F1 score of 0.8851. Deep geometric analyses further reveal that explicitly penalizing the complex amplitude catastrophically over-regularizes subjective representations, proving that our unconstrained-amplitude and phase-driven objective is crucial for robust, fine-grained sentiment disentanglement.

Method: Interdisciplinary approach combining critical sociolinguistics and computational linguistics. Uses two case studies: South Tyrolean dialects (widely used in informal communication in Italy) and varieties of Kurdish. Explores both technical approaches for making LLMs handle non-standard language and policy implications for democratic, decolonial digital strategies.

Result: The paper provides a critical framework for understanding how GenAI technologies perpetuate linguistic hierarchies and standardization. It identifies both technical challenges in handling non-standard linguistic varieties and policy considerations for making AI more linguistically inclusive and equitable.

Conclusion: LLMs and GenAI systems are not neutral but reproduce existing linguistic inequalities. Addressing this requires both technical solutions for handling linguistic variation and broader policy approaches that consider historical and sociopolitical dimensions of language standardization, with implications for creating more democratic and decolonial digital ecosystems.

Abstract: The design of Large Language Models and generative artificial intelligence has been shown to be “unfair” to less-spoken languages and to deepen the digital language divide. Critical sociolinguistic work has also argued that these technologies are not only made possible by prior socio-historical processes of linguistic standardisation, often grounded in European nationalist and colonial projects, but also exacerbate epistemologies of language as “monolithic, monolingual, syntactically standardized systems of meaning”. In our paper, we draw on earlier work on the intersections of technology and language policy and bring our respective expertise in critical sociolinguistics and computational linguistics to bear on an interrogation of these arguments. We take two different complexes of non-standard linguistic varieties in our respective repertoires–South Tyrolean dialects, which are widely used in informal communication in South Tyrol, Italy, as well as varieties of Kurdish–as starting points to an interdisciplinary exploration of the intersections between GenAI and linguistic variation and standardisation. We discuss both how LLMs can be made to deal with nonstandard language from a technical perspective, and whether, when or how this can contribute to “democratic and decolonial digital and machine learning strategies”, which has direct policy implications.

Method: Used representational similarity analysis across six LLMs from five architecture families, comparing categorical perception-additive models (log-distance plus boundary boost) vs. purely continuous models, examining Arabic numeral processing at digit-count boundaries.

Result: Found categorical perception effects in 100% of primary layers across all models, specific to structurally defined boundaries (digit-count transitions at 10 and 100). Identified two distinct signatures: “classic CP” (Gemma, Qwen) with both geometric warping and explicit categorization, and “structural CP” (Llama, Mistral, Phi) with only geometric warping.

Conclusion: Structural input-format discontinuities are sufficient to produce categorical perception geometry in LLMs independently of explicit semantic category knowledge, revealing architecture-dependent patterns of representation warping.

Abstract: Categorical perception (CP) – enhanced discriminability at category boundaries – is among the most studied phenomena in perceptual psychology. This paper reports that analogous geometric warping occurs in the hidden-state representations of large language models (LLMs) processing Arabic numerals. Using representational similarity analysis across six models from five architecture families, the study finds that a CP-additive model (log-distance plus a boundary boost) fits the representational geometry better than a purely continuous model at 100% of primary layers in every model tested. The effect is specific to structurally defined boundaries (digit-count transitions at 10 and 100), absent at non-boundary control positions, and absent in the temperature domain where linguistic categories (hot/cold) lack a tokenisation discontinuity. Two qualitatively distinct signatures emerge: “classic CP” (Gemma, Qwen), where models both categorise explicitly and show geometric warping, and “structural CP” (Llama, Mistral, Phi), where geometry warps at the boundary but models cannot report the category distinction. This dissociation is stable across boundaries and is a property of the architecture, not the stimulus. Structural input-format discontinuities are sufficient to produce categorical perception geometry in LLMs, independently of explicit semantic category knowledge.

Method: Proposes two representations: speaker-based representation following speech turn segmentation, and dependency-based representation with cross-turn dependencies. Also introduces new propositions to distinguish reformulations vs repairs and handle unfinished phrases

Result: Developed annotation guidelines for cross-speaker dependencies in spoken language treebanks, enabling more accurate syntactic analysis of conversational speech

Conclusion: The proposed framework improves syntactic annotation of spoken dialogue by accounting for dependencies across speaker boundaries, better capturing the collaborative nature of conversation

Abstract: The paper proposes annotation guidelines for syntactic dependencies that span across speaker turns - including collaborative coconstructions proper, wh-question answers, and backchannels - in spoken language treebanks within the Universal Dependencies framework. Two representations are proposed: a speaker-based representation following the segmentation into speech turns, and a dependency-based representation with dependencies across speech turns. New propositions are also put forward to distinguish between reformulations and repairs, and to promote elements in unfinished phrases.

Method: Proposes model merging as a lightweight alternative: merging an instruction-tuned LLM with language-specific base models to transfer language knowledge without needing language-specific instructions or repeated fine-tuning.

Result: Experiments with four Iberian languages (Basque, Catalan, Galician, Spanish) show merging enables effective instruction following in new languages and supports multilingual capability through combining multiple language-specific models.

Conclusion: Model merging is a viable and efficient alternative to traditional adaptation methods for low-resource languages, achieving competitive performance while greatly reducing computational cost.

Abstract: Large Language Models (LLMs) remain heavily centered on English, with limited performance in low-resource languages. Existing adaptation approaches, such as continual pre-training, demand significant computational resources. In the case of instructed models, high-quality instruction data is also required, both of which are often inaccessible for low-resource language communities. Under these constraints, model merging offers a lightweight alternative, but its potential in low-resource contexts has not been systematically explored. In this work, we explore whether it is possible to transfer language knowledge to an instruction-tuned LLM by merging it with a language-specific base model, thereby eliminating the need of language-specific instructions and repeated fine-tuning processes whenever stronger instructed variants become available. Through experiments covering four Iberian languages (Basque, Catalan, Galician, and Spanish) and two model families, we show that merging enables effective instruction following behavior in new languages and even supports multilingual capability through the combination of multiple language-specific models. Our results indicate that model merging is a viable and efficient alternative to traditional adaptation methods for low-resource languages, achieving competitive performance while greatly reducing computational cost.

Method: Conducted controlled experiments to investigate temperature-judge performance relationship, and used causal inference framework within empirical statistical analysis to rigorously examine direct causal effect of temperature on judge behavior.

Result: Temperature significantly influences judge performance in LLM-centric evaluation, with effects being task-dependent; lower temperatures don’t always yield optimal outcomes; provides actionable engineering insights for evaluation pipeline design.

Conclusion: Temperature is a critical parameter in LLM-as-a-Judge evaluation that requires careful selection based on task characteristics, challenging the empirical convention of fixed temperature settings.

Abstract: LLM-as-a-Judge has emerged as an effective and low-cost paradigm for evaluating text quality and factual correctness. Prior studies have shown substantial agreement between LLM judges and human experts, even on tasks that are difficult to assess automatically. In practice, researchers commonly employ fixed temperature configurations during the evaluation process-with values of 0.1 and 1.0 being the most prevalent choices-a convention that is largely empirical rather than principled. However, recent researches suggest that LLM performance exhibits non-trivial sensitivity to temperature settings, that lower temperatures do not universally yield optimal outcomes, and that such effects are highly task-dependent. This raises a critical research question: does temperature influence judge performance in LLM centric evaluation? To address this, we systematically investigate the relationship between temperature and judge performance through a series of controlled experiments, and further adopt a causal inference framework within our empirical statistical analysis to rigorously examine the direct causal effect of temperature on judge behavior, offering actionable engineering insights for the design of LLM-centric evaluation pipelines.

Method: Combines evolutionary agent maintaining population of executable candidates with execution feedback (compilation, correctness, speedup) and converts evolution trajectories into training signals for LLM optimization as local improvers within evolutionary loop.

Result: Achieves SOTA on KernelBench with Triton backend, outperforming proprietary models like Gemini-3.0-pro and Claude-4.6-opus; also validates on MetaX MACA backend surpassing large models like DeepSeek-V3.2-think and Qwen3-235B-2507-think.

Conclusion: LLM-driven kernel optimization can transfer from controlled evaluation to practical deployment, with framework showing potential for seamless adaptation across heterogeneous GPU platforms.

Abstract: We present Kernel-Smith, a framework for high-performance GPU kernel and operator generation that combines a stable evaluation-driven evolutionary agent with an evolution-oriented post-training recipe. On the agent side, Kernel-Smith maintains a population of executable candidates and iteratively improves them using an archive of top-performing and diverse programs together with structured execution feedback on compilation, correctness, and speedup. To make this search reliable, we build backend-specific evaluation services for Triton on NVIDIA GPUs and Maca on MetaX GPUs. On the training side, we convert long-horizon evolution trajectories into step-centric supervision and reinforcement learning signals by retaining correctness-preserving, high-gain revisions, so that the model is optimized as a strong local improver inside the evolutionary loop rather than as a one-shot generator. Under a unified evolutionary protocol, Kernel-Smith-235B-RL achieves state-of-the-art overall performance on KernelBench with Nvidia Triton backend, attaining the best average speedup ratio and outperforming frontier proprietary models including Gemini-3.0-pro and Claude-4.6-opus. We further validate the framework on the MetaX MACA backend, where our Kernel-Smith-MACA-30B surpasses large-scale counterparts such as DeepSeek-V3.2-think and Qwen3-235B-2507-think, highlighting potential for seamless adaptation across heterogeneous platforms. Beyond benchmark results, the same workflow produces upstream contributions to production systems including SGLang and LMDeploy, demonstrating that LLM-driven kernel optimization can transfer from controlled evaluation to practical deployment.

Method: Top-down approach constructing 7 evaluation desiderata based on how subjectivity is represented in NLP data and models, followed by analysis of 60 papers’ experimental setups to identify gaps.

Result: Analysis reveals several understudied aspects: distinction between ambiguous vs polyphonic input, whether subjectivity is effectively expressed to users, and lack of interplay between different desiderata.

Conclusion: Evaluation practices for subjectivity-sensitive models need improvement to better capture diverse perspectives and ensure minority voices are not marginalized, requiring attention to the identified gaps.

Abstract: Subjective judgments are part of several NLP datasets and recent work is increasingly prioritizing models whose outputs reflect this diversity of perspectives. Such responses allow us to shed light on minority voices, which are frequently marginalized or obscured by dominant perspectives. It remains a question whether our evaluation practices align with these models’ objectives. This position paper proposes seven evaluation desiderata for subjectivity-sensitive models, rooted in how subjectivity is represented in NLP data and models. The desiderata are constructed in a top-down approach, keeping in mind the user-centric impact of such models. We scan the experimental setup of 60 papers and show that various aspects of subjectivity are still understudied: the distinction between ambiguous and polyphonic input, whether subjectivity is effectively expressed to the user, and a lack of interplay between different desiderata, amongst other gaps.

Method: Used an adapted reading interface with four modalities: unmodified text, sentence-segmented text, segmented text with pictograms, and segmented text with pictograms plus keyword labels. Conducted a within-subject pilot with 14 primary-school learners with special educational needs and disabilities, measuring comprehension with standardized questions and collecting child- and therapist-reported experience measures with open-ended feedback.

Result: Results showed heterogeneous responses - some learners benefited from segmentation and pictograms, while others showed increased coordination costs with visual scaffolds. Experience ratings showed limited differences between modalities, with some effects linked to clinical complexity. Open-ended feedback frequently requested simpler wording and additional visual supports.

Conclusion: No single scaffold is universally optimal, reinforcing the need for calibrated, adjustable scaffolding. Provides design implications for human-AI co-regulation in supervised inclusive reading contexts.

Abstract: Neurodiverse learners often require reading supports, yet increasing scaffold richness can sometimes overload attention and working memory rather than improve comprehension. Grounded in the Construction-Integration model and a contingent scaffolding perspective, we examine how structural versus semantic scaffolds shape comprehension and reading experience in a supervised inclusive context. Using an adapted reading interface, we compared four modalities: unmodified text, sentence-segmented text, segmented text with pictograms, and segmented text with pictograms plus keyword labels. In a within-subject pilot with 14 primary-school learners with special educational needs and disabilities, we measured reading comprehension using standardized questions and collected brief child- and therapist-reported experience measures alongside open-ended feedback. Results highlight heterogeneous responses as some learners showed patterns consistent with benefits from segmentation and pictograms, while others showed patterns consistent with increased coordination costs when visual scaffolds were introduced. Experience ratings showed limited differences between modalities, with some apparent effects linked to clinical complexity, particularly for perceived ease of understanding. Open-ended feedback of the learners frequently requested simpler wording and additional visual supports. These findings suggest that no single scaffold is universally optimal, reinforcing the need for calibrated, adjustable scaffolding and provide design implications for human-AI co-regulation in supervised inclusive reading contexts.

Method: Used log-probability-based moral justifiability scores to compare LLM outputs with World Values Survey and Pew Global Attitudes Survey data, analyzing both smaller monolingual/multilingual models and advanced instruction-tuned models across various ethical topics.

Result: Earlier/smaller models showed near-zero or negative correlations with human judgments, while advanced instruction-tuned models achieved substantially higher positive correlations. Models aligned better with W.E.I.R.D. (Western, Educated, Industrialized, Rich, Democratic) nations than other regions.

Conclusion: While scaling model size and instruction tuning improves alignment with cross-cultural moral norms, significant challenges remain for certain topics and regions, highlighting the need for improved cultural sensitivity in LLMs.

Abstract: Large Language Models (LLMs) have shown strong performance across many tasks, but their ability to capture culturally diverse moral values remains unclear. In this paper, we examine whether LLMs mirror variations in moral attitudes reported by the World Values Survey (WVS) and the Pew Research Center’s Global Attitudes Survey (PEW). We compare smaller monolingual and multilingual models (GPT-2, OPT, BLOOMZ, and Qwen) with recent instruction-tuned models (GPT-4o, GPT-4o-mini, Gemma-2-9b-it, and Llama-3.3-70B-Instruct). Using log-probability-based \emph{moral justifiability} scores, we correlate each model’s outputs with survey data covering a broad set of ethical topics. Our results show that many earlier or smaller models often produce near-zero or negative correlations with human judgments. In contrast, advanced instruction-tuned models achieve substantially higher positive correlations, suggesting they better reflect real-world moral attitudes. We provide a detailed regional analysis revealing that models align better with Western, Educated, Industrialized, Rich, and Democratic (W.E.I.R.D.) nations than with other regions. While scaling model size and using instruction tuning improves alignment with cross-cultural moral norms, challenges remain for certain topics and regions. We discuss these findings in relation to bias analysis, training data diversity, information retrieval implications, and strategies for improving the cultural sensitivity of LLMs.

Method: Three-level verification framework: 1) QA data synthesis with verification to control question difficulty and ensure answer correctness; 2) Verification-driven trajectory synthesis injecting verification patterns; 3) Test-time scaling using the agent itself as a verifier.

Result: Marco DeepResearch outperforms 8B-scale deep research agents on challenging benchmarks like BrowseComp and BrowseComp-ZH, and with 600 tool calls approaches/surpasses 30B-scale agents like Tongyi DeepResearch-30B.

Conclusion: Verification-centric design is crucial for reliable deep research agents, enabling smaller models to achieve performance comparable to much larger models through systematic verification mechanisms.

Abstract: Deep research agents autonomously conduct open-ended investigations, integrating complex information retrieval with multi-step reasoning across diverse sources to solve real-world problems. To sustain this capability on long-horizon tasks, reliable verification is critical during both training and inference. A major bottleneck in existing paradigms stems from the lack of explicit verification mechanisms in QA data synthesis, trajectory construction, and test-time scaling. Errors introduced at each stage propagate downstream and degrade the overall agent performance. To address this, we present Marco DeepResearch, a deep research agent optimized with a verification-centric framework design at three levels: \textbf{(1)~QA Data Synthesis:} We introduce verification mechanisms to graph-based and agent-based QA synthesis to control question difficulty while ensuring answers are unique and correct; \textbf{(2)~Trajectory Construction:} We design a verification-driven trajectory synthesis method that injects explicit verification patterns into training trajectories; and \textbf{(3)~Test-time scaling:} We use Marco DeepResearch itself as a verifier at inference time and effectively improve performance on challenging questions. Extensive experimental results demonstrate that our proposed Marco DeepResearch agent significantly outperforms 8B-scale deep research agents on most challenging benchmarks, such as BrowseComp and BrowseComp-ZH. Crucially, under a maximum budget of 600 tool calls, Marco DeepResearch even surpasses or approaches several 30B-scale agents, like Tongyi DeepResearch-30B.

Method: Curated LombardoGraphia corpus from Wikipedia, processed and filtered for orthographic analysis. Trained 24 traditional and neural classification models with various features and encoding levels.

Result: Best models achieved 96.06% overall accuracy and 85.78% average class accuracy, though performance on minority classes remains challenging due to data imbalance.

Conclusion: Provides crucial infrastructure for building variety-aware NLP resources for Lombard, enabling better language processing for this underresourced language variety.

Abstract: Lombard, an underresourced language variety spoken by approximately 3.8 million people in Northern Italy and Southern Switzerland, lacks a unified orthographic standard. Multiple orthographic systems exist, creating challenges for NLP resource development and model training. This paper presents the first study of automatic Lombard orthography classification and LombardoGraphia, a curated corpus of 11,186 Lombard Wikipedia samples tagged across 9 orthographic variants, and models for automatic orthography classification. We curate the dataset, processing and filtering raw Wikipedia content to ensure text suitable for orthographic analysis. We train 24 traditional and neural classification models with various features and encoding levels. Our best models achieve 96.06% and 85.78% overall and average class accuracy, though performance on minority classes remains challenging due to data imbalance. Our work provides crucial infrastructure for building variety-aware NLP resources for Lombard.

Method: Three-stage pipeline: 1) ambiguity-preserving n-best parsing using Combinatory Categorial Grammar, 2) STL-oriented template-based semantic composition, 3) canonicalization with score aggregation to output deduplicated STL candidates with plausibility scores

Result: Method generates multiple STL candidates for genuinely ambiguous inputs while collapsing unambiguous or canonically equivalent derivations to single STL formula

Conclusion: Proposed ambiguity-preserving approach better handles natural language ambiguity in NL-to-logic translation compared to existing one-best methods

Abstract: Signal Temporal Logic (STL) is widely used to specify timed and safety-critical tasks for cyber-physical systems, but writing STL formulas directly is difficult for non-expert users. Natural language (NL) provides a convenient interface, yet its inherent structural ambiguity makes one-to-one translation into STL unreliable. In this paper, we propose an \textit{ambiguity-preserving} method for translating NL task descriptions into STL candidate formulas. The key idea is to retain multiple plausible syntactic analyses instead of forcing a single interpretation at the parsing stage. To this end, we develop a three-stage pipeline based on Combinatory Categorial Grammar (CCG): ambiguity-preserving $n$-best parsing, STL-oriented template-based semantic composition, and canonicalization with score aggregation. The proposed method outputs a deduplicated set of STL candidates with plausibility scores, thereby explicitly representing multiple possible formal interpretations of an ambiguous instruction. In contrast to existing one-best NL-to-logic translation methods, the proposed approach is designed to preserve attachment and scope ambiguity. Case studies on representative task descriptions demonstrate that the method generates multiple STL candidates for genuinely ambiguous inputs while collapsing unambiguous or canonically equivalent derivations to a single STL formula.

Method: Retrieve-then-rerank pipeline: 1) Priority infilling retrieval model for structurally/semantically similar candidates, 2) Ensemble reranking model with neighbor-enhanced representations for final link prediction.

Result: Outperforms existing baselines, improves MRR on validation set from 0.2342 to 0.2839.

Conclusion: Proposed method effectively handles large-scale knowledge graph embedding with improved accuracy while maintaining efficiency.

Abstract: WikiKG90Mv2 in NeurIPS 2022 is a large encyclopedic knowledge graph. Embedding knowledge graphs into continuous vector spaces is important for many practical applications, such as knowledge acquisition, question answering, and recommendation systems. Compared to existing knowledge graphs, WikiKG90Mv2 is a large scale knowledge graph, which is composed of more than 90 millions of entities. Both efficiency and accuracy should be considered when building graph embedding models for knowledge graph at scale. To this end, we follow the retrieve then re-rank pipeline, and make novel modifications in both retrieval and re-ranking stage. Specifically, we propose a priority infilling retrieval model to obtain candidates that are structurally and semantically similar. Then we propose an ensemble based re-ranking model with neighbor enhanced representations to produce final link prediction results among retrieved candidates. Experimental results show that our proposed method outperforms existing baseline methods and improves MRR of validation set from 0.2342 to 0.2839.

Method: Extract revision pairs from arXiv LaTeX source files by identifying commented-out text (discarded/alternative formulations), aligning commented segments with nearby final text, and applying LLM-based filtering to validate genuine revisions.

Result: Created EarlySciRev dataset with 578k validated revision pairs from 1.28M candidate pairs, providing authentic early drafting traces and a human-annotated benchmark for revision detection.

Conclusion: EarlySciRev enables research on scientific writing dynamics, revision modeling, and LLM-assisted editing by providing access to previously unavailable early-stage revision data.

Abstract: Scientific writing is an iterative process that generates rich revision traces, yet publicly available resources typically expose only final or near-final versions of papers. This limits empirical study of revision behaviour and evaluation of large language models (LLMs) for scientific writing. We introduce EarlySciRev, a dataset of early-stage scientific text revisions automatically extracted from arXiv LaTeX source files. Our key observation is that commented-out text in LaTeX often preserves discarded or alternative formulations written by the authors themselves. By aligning commented segments with nearby final text, we extract paragraph-level candidate revision pairs and apply LLM-based filtering to retain genuine revisions. Starting from 1.28M candidate pairs, our pipeline yields 578k validated revision pairs, grounded in authentic early drafting traces. We additionally provide a human-annotated benchmark for revision detection. EarlySciRev complements existing resources focused on late-stage revisions or synthetic rewrites and supports research on scientific writing dynamics, revision modelling, and LLM-assisted editing.

Method: Two-stage SFT training paradigm: 1) Train agent on structurally diverse trajectories synthesized from constrained random-walk paths to establish broad exploration prior over KG; 2) Fine-tune on small set of expert trajectories to develop reflection and error recovery capabilities. This enables higher performance ceiling for lightweight RL stage.

Result: Achieves state-of-the-art performance on CWQ and WebQSP datasets. Additional results on GrailQA and constructed GraphWalkerBench confirm enhanced generalization to out-of-distribution reasoning paths.

Conclusion: GraphWalker’s stage-wise SFT paradigm effectively addresses training data scarcity and improves reasoning generalization in agentic KGQA through automated trajectory synthesis and progressive fine-tuning.

Abstract: Agentic knowledge graph question answering (KGQA) requires an agent to iteratively interact with knowledge graphs (KGs), posing challenges in both training data scarcity and reasoning generalization. Specifically, existing approaches often restrict agent exploration: prompting-based methods lack autonomous navigation training, while current training pipelines usually confine reasoning to predefined trajectories. To this end, this paper proposes \textit{GraphWalker}, a novel agentic KGQA framework that addresses these challenges through \textit{Automated Trajectory Synthesis} and \textit{Stage-wise Fine-tuning}. GraphWalker adopts a two-stage SFT training paradigm: First, the agent is trained on structurally diverse trajectories synthesized from constrained random-walk paths, establishing a broad exploration prior over the KG; Second, the agent is further fine-tuned on a small set of expert trajectories to develop reflection and error recovery capabilities. Extensive experiments demonstrate that our stage-wise SFT paradigm unlocks a higher performance ceiling for a lightweight reinforcement learning (RL) stage, enabling GraphWalker to achieve state-of-the-art performance on CWQ and WebQSP. Additional results on GrailQA and our constructed GraphWalkerBench confirm that GraphWalker enhances generalization to out-of-distribution reasoning paths. The code is publicly available at https://github.com/XuShuwenn/GraphWalker

Method: Replace every nn.Linear layer in PicoGPT (GPT-2-style model) with MPOLinear modules parameterized as MPO chains. Use Matrix Product Operator decomposition to factorize weight matrices into chains of low-rank cores with bond dimension chi controlling approximation quality. Initialize cores via TT-SVD from pretrained weights or random initialization, and train using standard PyTorch autograd.

Result: Achieves up to 13x compression per transformer block at chi=4. At chi=16, model uses 191,872 parameters instead of 1,020,224 while retaining 97.7% of baseline token accuracy (51.6% vs 52.8%). Chi=8 model gives best accuracy per parameter, exceeding dense baseline by 2.7x on this metric.

Conclusion: MPO parameterization is a practical and theoretically grounded alternative to low-rank methods and unstructured pruning for transformer compression, offering significant parameter reduction with minimal accuracy loss.

Abstract: Transformer-based language models achieve strong performance across NLP tasks, but their quadratic parameter scaling with hidden dimension makes deployment on resource-constrained hardware expensive. We study Matrix Product Operator (MPO) decomposition as a principled compression method for transformers. MPO factorises weight matrices into chains of low-rank cores, with approximation quality controlled by the bond dimension chi. We replace every nn.Linear layer in PicoGPT, a GPT-2-style character-level language model with about 1M parameters, with an MPOLinear module parameterised as an MPO chain. Cores are initialised either by TT-SVD from pretrained dense weights or from random initialisation, and trained using standard PyTorch autograd without a custom backward pass. We derive balanced factorisation schemes for the five distinct weight shapes in PicoGPT and evaluate bond dimensions chi in {4, 8, 16, 32} on Tiny Shakespeare. MPO compression achieves up to 13x compression per transformer block at chi = 4. At chi = 16, the model uses 191,872 parameters instead of 1,020,224 while retaining 97.7% of baseline token accuracy (51.6% vs 52.8%). Reconstruction error follows the expected trend and is lower for three-site than two-site factorisations at the same bond dimension. The chi = 8 model gives the best accuracy per parameter, exceeding the dense baseline by 2.7x on this metric. These results show that MPO parameterisation is a practical and theoretically grounded alternative to low-rank methods and unstructured pruning for transformer compression.

Method: The authors explore methods to create surrogate datasets using small confidential reference datasets. They use derived text formats to preserve confidentiality while generating larger training datasets, comparing these to purely prompt-based generation approaches.

Result: Successfully created three surrogate datasets that are more similar to the reference dataset than prompt-based generation alone. Early experiments suggest one approach might improve model training for automatic answer grading.

Conclusion: The proposed methods enable creation of large-scale training datasets from small confidential sources, potentially improving automatic grading systems for educational assessments like PISA while maintaining data confidentiality.

Abstract: Every 4 years, the PISA test is administered by the OECD to test the knowledge of teenage students worldwide and allow for comparisons of educational systems. However, having to avoid language differences and annotator bias makes the grading of student answers challenging. For these reasons, it would be interesting to compare methods of automatic student answer grading. To train some of these methods, which require machine learning, or to compute parameters or select hyperparameters for those that do not, a large amount of domain-specific data is needed. In this work, we explore a small number of methods for creating a large-scale training dataset using only a relatively small confidential dataset as a reference, leveraging a set of very simple derived text formats to preserve confidentiality. Using these methods, we successfully created three surrogate datasets that are, at the very least, superficially more similar to the reference dataset than purely the result of prompt-based generation. Early experiments suggest one of these approaches might also lead to improved model training.

Method: Developed EpiScreen by fine-tuning large language models on labeled clinical notes from electronic health records. Used MIMIC-IV dataset and private University of Minnesota cohort for training and evaluation.

Result: Achieved AUC of 0.875 on MIMIC-IV and 0.980 on private cohort. In clinician-AI collaboration, EpiScreen-assisted neurologists outperformed unaided experts by up to 10.9%.

Conclusion: EpiScreen enables early, cost-effective epilepsy screening that can reduce diagnostic delays and unnecessary interventions, especially in resource-limited regions.

Abstract: Epilepsy and psychogenic non-epileptic seizures often present with similar seizure-like manifestations but require fundamentally different management strategies. Misdiagnosis is common and can lead to prolonged diagnostic delays, unnecessary treatments, and substantial patient morbidity. Although prolonged video-electroencephalography is the diagnostic gold standard, its high cost and limited accessibility hinder timely diagnosis. Here, we developed a low-cost, effective approach, EpiScreen, for early epilepsy detection by utilizing routinely collected clinical notes from electronic health records. Through fine-tuning large language models on labeled notes, EpiScreen achieved an AUC of up to 0.875 on the MIMIC-IV dataset and 0.980 on a private cohort of the University of Minnesota. In a clinician-AI collaboration setting, EpiScreen-assisted neurologists outperformed unaided experts by up to 10.9%. Overall, this study demonstrates that EpiScreen supports early epilepsy detection, facilitating timely and cost-effective screening that may reduce diagnostic delays and avoid unnecessary interventions, particularly in resource-limited regions.

Method: Proposes IF4 data type that adaptively selects between FP4 and INT4 for each group of 16 values, using the unused sign bit in NVFP4’s scale factor to denote the selected type. Also designs IF3 and IF6 formats.

Result: IF4 outperforms existing 4-bit block-scaled formats, achieving lower loss during quantized training and higher accuracy in post-training quantization. Hardware design shows IF4 can be efficiently implemented.

Conclusion: IF4 provides superior 4-bit quantization for LLMs by adaptively choosing between float and integer representations, maintaining accuracy while being hardware-friendly.

Abstract: NVFP4 has grown increasingly popular as a 4-bit format for quantizing large language models due to its hardware support and its ability to retain useful information with relatively few bits per parameter. However, the format is not without limitations: recent work has shown that NVFP4 suffers from its error distribution, resulting in large amounts of quantization error on near-maximal values in each group of 16 values. In this work, we leverage this insight to design new Adaptive Block-Scaled Data Types that can adapt to the distribution of their input values. For four-bit quantization, our proposed IF4 (Int/Float 4) data type selects between FP4 and INT4 representations for each group of 16 values, which are then scaled by an E4M3 scale factor as is done with NVFP4. The selected data type is denoted using the scale factor’s sign bit, which is currently unused in NVFP4, and we apply the same insight to design formats for other bit-widths, including IF3 and IF6. When used to quantize language models, we find that IF4 outperforms existing 4-bit block-scaled formats, achieving lower loss during quantized training and achieving higher accuracy on many tasks in post-training quantization. We additionally design and evaluate an IF4 Multiply-Accumulate (MAC) unit to demonstrate that IF4 can be implemented efficiently in next-generation hardware accelerators. Our code is available at https://github.com/mit-han-lab/fouroversix.

Method: Controlled experiments across compound tasks comparing QA-trained models with CoT reasoning, analyzing granularity of CoT data, sample efficiency, and theoretical analysis of shortcut learning vs. true reasoning principles.

Result: CoT reasoning significantly improves OOD generalization, with finer-grained CoT data correlating with better performance and remarkable sample efficiency (matching QA performance with 80% less data).

Conclusion: CoT reasoning is a crucial training paradigm for enabling LM generalization under real-world distributional shifts for compound tasks, with theoretical justification for its effectiveness.

Abstract: Generalization to novel compound tasks under distribution shift is important for deploying transformer-based language models (LMs). This work investigates Chain-of-Thought (CoT) reasoning as a means to enhance OOD generalization. Through controlled experiments across several compound tasks, we reveal three key insights: (1) While QA-trained models achieve near-perfect in-distribution accuracy, their OOD performance degrades catastrophically, even with 10000k+ training examples; (2) the granularity of CoT data strongly correlates with generalization performance; finer-grained CoT data leads to better generalization; (3) CoT exhibits remarkable sample efficiency, matching QA performance with much less (even 80%) data. Theoretically, we demonstrate that compound tasks inherently permit shortcuts in Q-A data that misalign with true reasoning principles, while CoT forces internalization of valid dependency structures, and thus can achieve better generalization. Further, we show that transformer positional embeddings can amplify generalization by emphasizing subtask condition recurrence in long CoT sequences. Our combined theoretical and empirical analysis provides compelling evidence for CoT reasoning as a crucial training paradigm for enabling LM generalization under real-world distributional shifts for compound tasks.

Method: Used natural language processing methods (not commercial large language models) applied to transcribed speech data from 119 children in German-speaking Switzerland with typical and atypical language development. Focused on identifying optimal practices for semi-automatic LSA.

Result: Preliminary findings show potential for integrating locally deployed NLP methods into semi-automatic language sample analysis to support speech-language pathologists in more efficient DLD diagnosis.

Conclusion: Locally deployed NLP methods can enhance the efficiency of language sample analysis for developmental language disorder diagnosis while maintaining human specialist involvement in the diagnostic process.

Abstract: Language sample analysis (LSA) is a process that complements standardized psychometric tests for diagnosing, for example, developmental language disorder (DLD) in children. However, its labour-intensive nature has limited its use in speech-language pathology practice. We introduce an approach that leverages natural language processing (NLP) methods that do not rely on commercial large language models (LLMs) applied to transcribed speech data from 119 children in the German-speaking part of Switzerland with typical and atypical language development. This preliminary study aims to identify optimal practices that support speech-language pathologists in diagnosing DLD more efficiently with active involvement of human specialists. Preliminary findings underscore the potential of integrating locally deployed NLP methods into the process of semi-automatic LSA.

Method: Comparative analysis of historical annotations performed by humans and Large Language Models, examining patterns of agreement/disagreement, cultural bias, and error types in interpreting historical facts from short texts.

Result: Both humans and LLMs exhibit cultural bias, but LLMs achieve higher consensus on historical fact interpretation. Humans disagree based on personal biases, while LLMs disagree due to information skipping or hallucinations.

Conclusion: Findings enable large-scale annotation and quantitative analysis of historical data for digital humanities, offering new educational/research opportunities to explore historical interpretations from different LLMs and foster critical thinking about bias.

Abstract: This paper compares historical annotations by humans and Large Language Models. The findings reveal that both exhibit some cultural bias, but Large Language Models achieve a higher consensus on the interpretation of historical facts from short texts. While humans tend to disagree on the basis of their personal biases, Large Models disagree when they skip information or produce hallucinations. These findings have significant implications for digital humanities, enabling large-scale annotation and quantitative analysis of historical data. This offers new educational and research opportunities to explore historical interpretations from different Language Models, fostering critical thinking about bias.

Method: Created BRIDGE benchmark with 87 tasks from real-world clinical data across 9 languages, covering 8 task types, 6 clinical stages, 20 applications, and 14 specialties. Evaluated 95 LLMs (including DeepSeek-R1, GPT-4o, Gemini, Qwen3) under various inference strategies.

Result: Substantial performance variation across model sizes, languages, NLP tasks, and clinical specialties. Open-source LLMs can achieve performance comparable to proprietary models. Medically fine-tuned LLMs based on older architectures often underperform versus updated general-purpose models.

Conclusion: BRIDGE provides a foundational resource for developing and evaluating LLMs in real-world clinical text understanding, showing the importance of comprehensive multilingual evaluation on actual clinical data rather than simplified benchmarks.

Abstract: Large language models (LLMs) hold great promise for medical applications and are evolving rapidly, with new models being released at an accelerated pace. However, benchmarking on large-scale real-world data such as electronic health records (EHRs) is critical, as clinical decisions are directly informed by these sources, yet current evaluations remain limited. Most existing benchmarks rely on medical exam-style questions or PubMed-derived text, failing to capture the complexity of real-world clinical data. Others focus narrowly on specific application scenarios, limiting their generalizability across broader clinical use. To address this gap, we present BRIDGE, a comprehensive multilingual benchmark comprising 87 tasks sourced from real-world clinical data sources across nine languages. It covers eight major task types spanning the entire continuum of patient care across six clinical stages and 20 representative applications, including triage and referral, consultation, information extraction, diagnosis, prognosis, and billing coding, and involves 14 clinical specialties. We systematically evaluated 95 LLMs (including DeepSeek-R1, GPT-4o, Gemini series, and Qwen3 series) under various inference strategies. Our results reveal substantial performance variation across model sizes, languages, natural language processing tasks, and clinical specialties. Notably, we demonstrate that open-source LLMs can achieve performance comparable to proprietary models, while medically fine-tuned LLMs based on older architectures often underperform versus updated general-purpose models. The BRIDGE and its corresponding leaderboard serve as a foundational resource and a unique reference for the development and evaluation of new LLMs in real-world clinical text understanding. The BRIDGE leaderboard: https://huggingface.co/spaces/YLab-Open/BRIDGE-Medical-Leaderboard

Method: Created SynAnchors dataset for large-scale studies, benchmarked widely used LLMs with refined evaluation metrics

Result: LLMs commonly exhibit anchoring bias, primarily in shallow layers; reasoning provides some mitigation but conventional strategies don’t work

Conclusion: Anchoring bias is a significant issue in LLMs requiring new mitigation approaches beyond conventional methods

Abstract: The rise of Large Language Models (LLMs) like ChatGPT has advanced natural language processing, yet concerns about cognitive biases are growing. In this paper, we investigate the anchoring effect, a cognitive bias where the mind relies heavily on the first information as anchors to make affected judgments. We explore whether LLMs are affected by anchoring, the underlying mechanisms, and potential mitigation strategies. To facilitate studies at scale on the anchoring effect, we introduce a new dataset, SynAnchors (https://huggingface.co/datasets/TimTargaryen/SynAnchors). Combining refined evaluation metrics, we benchmark current widely used LLMs. Our findings show that LLMs’ anchoring bias exists commonly with shallow-layer acting and can not be eliminated by conventional strategies, while reasoning can offer some mitigation.

Method: Supervised fine-tuning on curated (false claim, correction) pairs injected as small “vaccine doses” (5-10% of tokens) alongside truthful data. Introduces direct negative supervision on labeled falsehoods. Includes design requirements: dosage, labeling, quarantine, and diversity.

Result: Across four open weight model families: improves TruthfulQA accuracy by 12 points, increases misinformation rejection rates by 30 points, while preserving overall model capability.

Conclusion: Immunization is a practical and scalable component of responsible LLM development. Advocates for standardized vaccine corpora and benchmarks to evaluate generalization.

Abstract: Large language models (LLMs) reproduce misinformation not by memorizing false facts alone, but by learning the linguistic patterns that make falsehoods persuasive, such as hedging, false presuppositions, and fabricated citations. We propose model immunization, a training paradigm based on supervised fine-tuning over curated (false claim, correction) pairs, injected as small vaccine doses (5 to 10% of tokens) alongside truthful data. Unlike post-hoc filtering or preference-based alignment, immunization introduces direct negative supervision on labeled falsehoods. Across four open weight model families, this approach improves TruthfulQA accuracy by 12 points and increases misinformation rejection rates by 30 points, while preserving overall model capability. We further outline key design requirements, including dosage, labeling, quarantine, and diversity and advocate for standardized vaccine corpora and benchmarks to evaluate generalization. These findings position immunization as a practical and scalable component of responsible LLM development. Project page: https://github.com/shainarazavi/ai-vaccine/

Method: Two main mechanisms: 1) POS-Aware Delay Mechanism that postpones EOS token generation by adjusting attention weights based on POS information, and 2) Generative Path Pruning Mechanism that limits hidden state magnitude to encourage persistent repetitive loops.

Result: LingoLoop successfully traps MLLMs like Qwen2.5-VL-3B in generative loops, driving them to generation limits and inducing outputs with up to 367x more tokens than clean inputs, causing significant energy consumption surges.

Conclusion: The attack exposes significant vulnerabilities in MLLMs related to resource exhaustion, posing challenges for reliable deployment and highlighting the need for defensive mechanisms against such attacks.

Abstract: Multimodal Large Language Models (MLLMs) have shown great promise but require substantial computational resources during inference. Attackers can exploit this by inducing excessive output, leading to resource exhaustion and service degradation. Prior energy-latency attacks aim to increase generation time by broadly shifting the output token distribution away from the EOS token, but they neglect the influence of token-level Part-of-Speech (POS) characteristics on EOS and sentence-level structural patterns on output counts, limiting their efficacy. To address this, we propose LingoLoop, an attack designed to induce MLLMs to generate excessively verbose and repetitive sequences. First, we find that the POS tag of a token strongly affects the likelihood of generating an EOS token. Based on this insight, we propose a POS-Aware Delay Mechanism to postpone EOS token generation by adjusting attention weights guided by POS information. Second, we identify that constraining output diversity to induce repetitive loops is effective for sustained generation. We introduce a Generative Path Pruning Mechanism that limits the magnitude of hidden states, encouraging the model to produce persistent loops. Extensive experiments on models like Qwen2.5-VL-3B demonstrate LingoLoop’s powerful ability to trap them in generative loops; it consistently drives them to their generation limits and, when those limits are relaxed, can induce outputs with up to 367x more tokens than clean inputs, triggering a commensurate surge in energy consumption. These findings expose significant MLLMs’ vulnerabilities, posing challenges for their reliable deployment.

Method: Proposes GHTM (Graph-based Hybrid Topic Model) that: 1) Uses TF-IDF-weighted GloVe embeddings to represent text documents, 2) Builds a document-similarity graph and applies Graph Convolutional Networks (GCN) for representation refinement through neighborhood aggregation, 3) Applies Non-negative Matrix Factorization (NMF) to extract interpretable topics from refined representations.

Result: GHTM achieves superior topic coherence (NPMI: 0.27-0.28) and diversity compared to existing methods while maintaining computational efficiency across datasets of varying scales. It also demonstrates strong cross-lingual generalization, outperforming established graph-based models on the English 20Newsgroups benchmark.

Conclusion: GHTM addresses critical gaps in Bengali topic modeling research by providing a novel architecture with superior performance, computational efficiency, and cross-lingual generalization. The introduction of NCTBText dataset provides much-needed topical diversity beyond newspaper-centric Bengali corpora for future research.

Abstract: Topic modeling is a Natural Language Processing (NLP) technique used to discover latent themes and abstract topics from text corpora by grouping co-occurring keywords. Although widely researched in English, topic modeling remains understudied in Bengali due to a lack of adequate resources and initiatives. Existing Bengali topic modeling research lacks standardized evaluation frameworks with comprehensive baselines and diverse datasets, exploration of modern methodological approaches, and reproducible implementations, with only three Bengali-specific architectures proposed to date. To address these gaps, this study presents a comprehensive evaluation of traditional and contemporary topic modeling approaches across three Bengali datasets and introduces GHTM (Graph-based Hybrid Topic Model), a novel architecture that strategically integrates TF-IDF-weighted GloVe embeddings, Graph Convolutional Networks (GCN), and Non-negative Matrix Factorization (NMF). GHTM represents text documents using hybrid TF-IDF-weighted GloVe embeddings. It builds a document-similarity graph and leverages GCN to refine the representations through neighborhood aggregation. Then, it finally decomposes the refined representations using NMF to extract interpretable topics. Experimental results demonstrate that GHTM achieves superior topic coherence (NPMI: 0.27-0.28) and diversity compared to existing methods while maintaining computational efficiency across datasets of varying scales. The model also demonstrates strong cross-lingual generalization, outperforming established graph-based models on the English 20Newsgroups benchmark. Additionally, we introduce NCTBText, a diverse Bengali textbook-based dataset comprising 8,650 text documents, curated from eight subject areas, providing much-needed topical diversity beyond newspaper-centric Bengali corpora and serving as a benchmark for future research.

Method: Adapts cross-document coreference resolution (CDCR) task, combines state-of-the-art CDCR models with natural language inference (NLI) and semantic text similarity (STS) principles for link prediction.

Result: Record linking model outperformed baseline NLP and STS methods by 28% (11.43 percentage points) and 27.4% (11.21 percentage points) respectively.

Conclusion: Common NLP tasks can be effectively combined and adapted for domain-specific settings to improve data quality and connectivity in industrial knowledge management systems.

Abstract: In the era of graph-based retrieval-augmented generation (RAG), link prediction is a significant preprocessing step for improving the quality of fragmented or incomplete domain-specific data for the graph retrieval. Knowledge management in the process industry uses RAG-based applications to optimize operations, ensure safety, and facilitate continuous improvement by effectively leveraging operational data and past insights. A key challenge in this domain is the fragmented nature of event logs in shift books, where related records are often kept separate, even though they belong to a single event or process. This fragmentation hinders the recommendation of previously implemented solutions to users, which is crucial in the timely problem-solving at live production sites. To address this problem, we develop a record linking model, which we define as a cross-document coreference resolution (CDCR) task. Record linking adapts the task definition of CDCR and combines two state-of-the-art CDCR models with the principles of natural language inference (NLI) and semantic text similarity (STS) to perform link prediction. The evaluation shows that our record linking model outperformed the best versions of our baselines, i.e., NLP and STS, by 28% (11.43 p) and 27.4% (11.21 p), respectively. Our work demonstrates that common NLP tasks can be combined and adapted to a domain-specific setting of the German process industry, improving data quality and connectivity in shift logs.

Method: Created AirQA dataset with 13,956 AI papers and 1,246 human-annotated questions covering multi-task, multi-modal, and instance-level evaluation. Developed ExTrActor framework with three LLM-based agents for automated instruction data synthesis, including example generation and trajectory collection without human intervention.

Result: Most models underperform on AirQA, demonstrating dataset quality. ExTrActor consistently improves multi-turn tool-use capability of small models, enabling them to achieve performance comparable to larger models.

Conclusion: AirQA provides a comprehensive benchmark for evaluating paper QA capabilities, while ExTrActor offers an effective automated solution for instruction data synthesis to enhance LLM-based agents’ performance on scientific paper understanding tasks.

Abstract: The growing volume of academic papers has made it increasingly difficult for researchers to efficiently extract key information. While large language models (LLMs) based agents are capable of automating question answering (QA) workflows for scientific papers, there still lacks a comprehensive and realistic benchmark to evaluate their capabilities. Moreover, training an interactive agent for this specific task is hindered by the shortage of high-quality interaction trajectories. In this work, we propose AirQA, a human-annotated comprehensive paper QA dataset in the field of artificial intelligence (AI), with 13,956 papers and 1,246 questions, that encompasses multi-task, multi-modal and instance-level evaluation. Furthermore, we propose ExTrActor, an automated framework for instruction data synthesis. With three LLM-based agents, ExTrActor can perform example generation and trajectory collection without human intervention. Evaluations of multiple open-source and proprietary models show that most models underperform on AirQA, demonstrating the quality of our dataset. Extensive experiments confirm that ExTrActor consistently improves the multi-turn tool-use capability of small models, enabling them to achieve performance comparable to larger ones.

Method: Proposes a framework combining LLMs with Multiple-Instance Learning (MIL). Each utterance is decomposed into Emotion, Logic, and Behavior (ELB) components. LLMs process these to infer distortion instances with type, expression, and salience scores. A Multi-View Gated Attention mechanism integrates instances for final classification.

Result: Experiments on Korean (KoACD) and English (Therapist QA) datasets show that incorporating ELB decomposition and LLM-inferred salience scores improves classification performance, particularly for distortions with high interpretive ambiguity.

Conclusion: The approach provides a psychologically grounded and generalizable method for fine-grained reasoning in mental health NLP, enhancing interpretability and expression-level analysis of cognitive distortions.

Abstract: Cognitive distortions have been closely linked to mental health disorders, yet their automatic detection remained challenging due to contextual ambiguity, co-occurrence, and semantic overlap. We proposed a novel framework that combines Large Language Models (LLMs) with Multiple-Instance Learning (MIL) architecture to enhance interpretability and expression-level reasoning. Each utterance was decomposed into Emotion, Logic, and Behavior (ELB) components, which were processed by LLMs to infer multiple distortion instances, each with a predicted type, expression, and model-assigned salience score. These instances were integrated via a Multi-View Gated Attention mechanism for final classification. Experiments on Korean (KoACD) and English (Therapist QA) datasets demonstrate that incorporating ELB and LLM-inferred salience scores improves classification performance, especially for distortions with high interpretive ambiguity. Our results suggested a psychologically grounded and generalizable approach for fine-grained reasoning in mental health NLP.

Method: PRISM enforces dual-space smoothness through two stages: (1) representation space stage with robustly trained probe to defend against jailbreak attacks, and (2) parameter-space stage that decouples retain-forget gradient conflicts and smooths parameter space to mitigate relearning attacks.

Result: Extensive experiments on WMDP and MUSE benchmarks across conversational-dialogue and continuous-text settings show PRISM outperforms SOTA baselines under multiple attacks while achieving better balance among key metrics.

Conclusion: PRISM provides a robust framework for machine unlearning that addresses current limitations by enforcing dual-space smoothness, achieving balanced performance across privacy, utility, and safety metrics while defending against attacks.

Abstract: As large language models evolve, Machine Unlearning has emerged to address growing concerns around user privacy, copyright infringement, and overall safety. Yet state-of-the-art (SOTA) unlearning methods often suffer from catastrophic forgetting and metric imbalance, for example, by over-optimizing one objective (e.g., unlearning effectiveness, utility preservation, or privacy protection) at the expense of others. In addition, small perturbations in the representation or parameter space can be exploited by relearn and jailbreak attacks. To address these challenges, we propose PRISM, a unified framework that enforces dual-space smoothness in representation and parameter spaces to improve robustness and balance unlearning metrics. PRISM consists of two smoothness optimization stages: (i) a representation space stage that employs a robustly trained probe to defend against jailbreak attacks, and (ii) a parameter-space stage that decouples retain-forget gradient conflicts, reduces imbalance, and smooths the parameter space to mitigate relearning attacks. Extensive experiments on WMDP and MUSE, across conversational-dialogue and continuous-text settings, show that PRISM outperforms SOTA baselines under multiple attacks while achieving a better balance among key metrics.

Method: Quantitative examination of two decades (2005-2025) using a dataset of 2.2K NLP papers, 4.9K authors, and 7.8K human-annotated contribution sentences, with benchmark results and a research explorer tool.

Result: Created AfricaNLPContributions dataset and research explorer tool that provides insights into AfricaNLP research trends, publications, topics, tasks, and contributor patterns over 20 years.

Conclusion: The dataset and tool offer a powerful lens for tracing AfricaNLP research trends and enable data-driven research approaches for understanding regional NLP development.

Abstract: Natural Language Processing (NLP) is undergoing constant transformation, as Large Language Models (LLMs) are driving daily breakthroughs in research and practice. In this regard, tracking the progress of NLP research and automatically analyzing the contributions of research papers provides key insights into the nature of the field and the researchers. This study explores the progress of African NLP (AfricaNLP) by asking (and answering) research questions about the progress of AfricaNLP (publications, NLP topics, and NLP tasks), contributions (data, method, and task), and contributors (authors, affiliated institutions, and funding bodies). We quantitatively examine two decades (2005 - 2025) of contributions to AfricaNLP research, using a dataset of 2.2K NLP papers, 4.9K contributing authors, and 7.8K human-annotated contribution sentences (AfricaNLPContributions), along with benchmark results. Our dataset and AfricaNLP research explorer tool will provide a powerful lens for tracing AfricaNLP research trends and holds potential for generating data-driven research approaches.

Method: Conduct neuron-level analysis using gradient-based scoring with filtering to identify culture-general neurons (contribute to all cultures) and culture-specific neurons (tied to individual cultures). Validate through suppression experiments and analyze layer distributions.

Result: Culture-general and culture-specific neurons account for <1% of all neurons, concentrated in shallow to middle MLP layers. Suppressing them degrades cultural benchmark performance by up to 30% while general NLU performance remains unaffected. Culture-specific neurons support knowledge of related cultures. Training on NLU benchmarks can diminish cultural understanding when updating modules with many culture-general neurons.

Conclusion: The study provides insights into LLMs’ internal mechanisms for cultural understanding and offers practical guidance for model training and engineering to improve cultural awareness while maintaining general language capabilities.

Abstract: As large language models (LLMs) are increasingly deployed worldwide, ensuring their fair and comprehensive cultural understanding is important. However, LLMs exhibit cultural bias and limited awareness of underrepresented cultures, while the mechanisms underlying their cultural understanding remain underexplored. To fill this gap, we conduct a neuron-level analysis to identify neurons that drive cultural behavior, introducing a gradient-based scoring method with additional filtering for precise refinement. We identify culture-general neurons contributing to cultural understanding regardless of cultures, and culture-specific neurons tied to an individual culture. Culture-general and culture-specific neurons account for less than 1% of all neurons and are concentrated in shallow to middle MLP layers. We validate their role by showing that suppressing them substantially degrades performance on cultural benchmarks (by up to 30%), while performance on general natural language understanding (NLU) benchmarks remains largely unaffected. Moreover, we show that culture-specific neurons support knowledge of not only the target culture, but also related cultures. Finally, we demonstrate that training on NLU benchmarks can diminish models’ cultural understanding when we update modules containing many culture-general neurons. These findings provide insights into the internal mechanisms of LLMs and offer practical guidance for model training and engineering. Our code is available at https://github.com/ynklab/CULNIG

Method: CLMN represents concepts as continuous, human-readable embeddings and applies fuzzy-logic reasoning to learn adaptive interaction rules. It augments original text features with concept-aware representations and automatically induces interpretable logic rules.

Result: Across multiple datasets and pre-trained language models, CLMN achieves higher accuracy than existing concept-based methods while improving explanation quality.

Conclusion: Integrating neural representations with symbolic reasoning in a unified concept space can yield practical, transparent NLP systems.

Abstract: Deep learning has advanced NLP, but interpretability remains limited, especially in healthcare and finance. Concept bottleneck models tie predictions to human concepts in vision, but NLP versions either use binary activations that harm text representations or latent concepts that weaken semantics, and they rarely model dynamic concept interactions such as negation and context. We introduce the Concept Language Model Network (CLMN), a neural-symbolic framework that keeps both performance and interpretability. CLMN represents concepts as continuous, human-readable embeddings and applies fuzzy-logic reasoning to learn adaptive interaction rules that state how concepts affect each other and the final decision. The model augments original text features with concept-aware representations and automatically induces interpretable logic rules. Across multiple datasets and pre-trained language models, CLMN achieves higher accuracy than existing concept-based methods while improving explanation quality. These results show that integrating neural representations with symbolic reasoning in a unified concept space can yield practical, transparent NLP systems.

Method: SA-ICL extracts building blocks of cognition from demonstration examples to create abstracted schemas - lightweight, structured templates of key inferential steps and their relationships. These schemas are then used to augment the model’s reasoning process when presented with novel questions.

Result: Experiments on chemistry and physics questions from GPQA dataset show SA-ICL consistently boosts performance (up to 36.19% improvement) when using high-quality single demonstration examples. It reduces reliance on multiple demonstrations and enhances interpretability.

Conclusion: SA-ICL bridges disparate ICL strategies (pattern priming to Chain-of-Thought prompting) and provides a new path for enhancing human-like reasoning in LLMs through explicit schema-based scaffolding.

Abstract: In-Context Learning (ICL) enables transformer-based language models to adapt to new tasks by conditioning on demonstration examples. However, traditional example-driven in-context learning lacks explicit modules for knowledge retrieval and transfer at the abstraction level. Inspired by cognitive science, specifically schema theory, which holds that humans interpret new information by activating pre-existing mental frameworks (schemas) to structure understanding, we introduce Schema-Activated In-Context Learning (SA-ICL). This framework extracts the representation of the building blocks of cognition for the reasoning process instilled from prior examples, creating an abstracted schema, a lightweight, structured template of key inferential steps and their relationships, which is then used to augment a model’s reasoning process when presented with a novel question. We demonstrate that a broad range of large language models (LLMs) lack the capacity to form and utilize internal schema-based learning representations implicitly, but instead benefit significantly from explicit schema-based scaffolding. Across chemistry and physics questions from the GPQA dataset, our experiments show that SA-ICL consistently boosts performance, up to 36.19 percent, when the single demonstration example is of high quality, which simultaneously reduces reliance on the number of demonstrations and enhances interpretability. Schema-Activated In-Context Learning not only bridges disparate ICL strategies ranging from pattern priming to Chain-of-Thought prompting, but also paves a new path for enhancing human-like reasoning in LLMs.

Method: Develops a framework that: 1) Generates controlled multiple-choice queries from tabular datasets while preserving task structure, 2) Applies systematic transformations to selectively disrupt dataset information while keeping partial knowledge, 3) Uses non-neural baselines for reference performance, and 4) Implements statistical testing to formally detect significant deviations indicating contamination.

Result: Empirical evaluation on eight widely used tabular datasets reveals clear evidence of contamination in four cases, suggesting that performance on downstream tasks involving these datasets may be substantially inflated.

Conclusion: Current evaluation practices for LLMs on tabular data may be unreliable due to data contamination, raising concerns about the validity of reported performance gains and highlighting the need for more rigorous contamination detection methods.

Abstract: Large language models (LLMs) are increasingly exposed to data contamination, i.e., performance gains driven by prior exposure of test datasets rather than generalization. However, in the context of tabular data, this problem is largely unexplored. Existing approaches primarily rely on memorization tests, which are too coarse to detect contamination. In contrast, we propose a framework for assessing contamination in tabular datasets by generating controlled queries and performing comparative evaluation. Given a dataset, we craft multiple-choice aligned queries that preserve task structure while allowing systematic transformations of the underlying data. These transformations are designed to selectively disrupt dataset information while preserving partial knowledge, enabling us to isolate performance attributable to contamination. We complement this setup with non-neural baselines that provide reference performance, and we introduce a statistical testing procedure to formally detect significant deviations indicative of contamination. Empirical results on eight widely used tabular datasets reveal clear evidence of contamination in four cases. These findings suggest that performance on downstream tasks involving such datasets may be substantially inflated, raising concerns about the reliability of current evaluation practices.

Method: Created LuxIT dataset by synthesizing instruction-answer pairs from native Luxembourgish texts using DeepSeek-R1-0528, followed by LLM-as-a-judge quality assurance to retain 227,507 high-quality pairs. Fine-tuned 14 smaller LLMs (≤15B parameters) on LuxIT and evaluated on Luxembourgish proficiency exams and five downstream NLP tasks.

Result: Training on LuxIT improved mean accuracy by +5.37 percentage points on language exams across all 14 models (12/14 showed improvement). On NLP tasks, 9/14 models improved in macro-averaged F1, though gains on the two benchmarks didn’t systematically correlate.

Conclusion: Monolingual synthetic data can effectively improve LLM capabilities in low-resource languages, demonstrating feasibility while highlighting the multi-faceted nature of language proficiency.

Abstract: The effectiveness of instruction-tuned Large Language Models (LLMs) is often limited in low-resource linguistic settings due to a lack of high-quality training data. We introduce LuxIT, a novel, monolingual instruction tuning dataset for Luxembourgish developed to mitigate this challenge. We synthesize the dataset from a corpus of native Luxembourgish texts, utilizing DeepSeek-R1-0528, chosen for its shown proficiency in Luxembourgish. Following generation, we apply a quality assurance process, employing an LLM-as-a-judge approach, retaining 227,507 high-quality instruction-answer pairs. To investigate the practical utility of the dataset, we fine-tune 14 smaller-scale LLMs ($\leq$15B parameters) on LuxIT and evaluate them on standardized Luxembourgish proficiency exams and five downstream NLP tasks. Training on LuxIT yields a mean accuracy change of +5.37 percentage points on language exams across all 14 models, with 12 of 14 showing improvement. On NLP downstream tasks, 9 of 14 models improve in macro-averaged F1, though gains on the two benchmarks do not systematically correlate. These results underscore the feasibility of leveraging monolingual synthetic data to improve LLM capabilities in low-resource languages, while highlighting the multi-faceted nature of language proficiency.

Method: Created CORPUSNAME benchmark with 8,400 structured debate prompts across 4 sensitive domains (Women’s Rights, Backwardness, Terrorism, Religion) in 7 languages. Tested 4 flagship LLMs (GPT-4o, Claude 3.5 Haiku, DeepSeek-Chat, LLaMA-3-70B), generating over 100,000 debate responses with automatic classification of demographic group stereotypes.

Result: All models reproduce entrenched stereotypes: Arabs linked to Terrorism and Religion (≥89%), Africans to socioeconomic “backwardness” (up to 77%), Western groups framed as modern/progressive. Biases grow sharply in lower-resource languages, showing English alignment doesn’t generalize globally.

Conclusion: Current alignment methods reduce explicit toxicity but fail to prevent biased outputs in open-ended multilingual contexts. Persistent divide in multilingual fairness requires better culturally inclusive model alignment.

Abstract: Large language models (LLMs) are widely deployed for open-ended communication, yet most bias evaluations still rely on English, classification-style tasks. We introduce \corpusname, a new multilingual, debate-style benchmark designed to reveal how narrative bias appears in realistic generative settings. Our dataset includes 8{,}400 structured debate prompts spanning four sensitive domains – Women’s Rights, Backwardness, Terrorism, and Religion – across seven languages ranging from high-resource (English, Chinese) to low-resource (Swahili, Nigerian Pidgin). Using four flagship models (GPT-4o, Claude3.5Haiku, DeepSeek-Chat, and LLaMA-3-70B), we generate over 100{,}000 debate responses and automatically classify which demographic groups are assigned stereotyped versus modern roles. Results show that all models reproduce entrenched stereotypes despite safety alignment: Arabs are overwhelmingly linked to Terrorism and Religion ($\geq$89%), Africans to socioeconomic ``backwardness’’ (up to 77%), and Western groups are consistently framed as modern or progressive. Biases grow sharply in lower-resource languages, revealing that alignment trained primarily in English does not generalize globally. Our findings highlight a persistent divide in multilingual fairness: current alignment methods reduce explicit toxicity but fail to prevent biased outputs in open-ended contexts. We release our \corpusname benchmark and analysis framework to support the next generation of multilingual bias evaluation and safer, culturally inclusive model alignment.

Method: Knowledge-Provision-based Prompt Optimization (KPPO) reformulates prompt optimization as systematic knowledge integration with three key innovations: 1) knowledge gap filling mechanism for identification and targeted remediation, 2) batch-wise candidate evaluation considering both performance improvement and distributional stability, 3) adaptive knowledge pruning strategy balancing performance and token efficiency.

Result: Evaluation on 15 knowledge-intensive benchmarks from various domains shows KPPO’s superiority over elicitation-based methods with ~6% average improvement over baselines while achieving comparable or lower token consumption, reducing up to 29% of inference token usage.

Conclusion: KPPO provides a more effective approach to prompt optimization for knowledge-intensive tasks by shifting from elicitation to systematic knowledge integration, addressing fundamental limitations of existing methods.

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

Method: ΠAttention factorizes attention into three components: ring-local neighborhoods for local context, deterministic π-stride skips for periodic long-range connections, and adaptive fusion gates to dynamically combine local and global information. This maintains linear complexity while expanding receptive field coverage.

Result: ΠAttention achieves O(kL + πlogL) receptive field growth vs O(kL) for RingAttention. Experiments show it matches or surpasses dense attention quality with 8.3% lower perplexity than RingAttention while using 50% fewer GPUs for same context length.

Conclusion: Periodic sparse attention with adaptive fusion provides an effective approach to efficient long-context modeling, balancing computational efficiency with expressive power through structured sparsity patterns.

Abstract: Transformers have revolutionized natural language processing, but their quadratic complexity with respect to sequence length remains a fundamental bottleneck for long-range modeling. While sparse attention mechanisms like RingAttention reduce computational costs by restricting attention to local neighborhoods, they suffer from limited receptive fields and lack of adaptability. We present \PiAttention, a periodic sparse Transformer that factorizes attention into ring-local neighborhoods, deterministic $π$-stride skips, and an adaptive fusion gate. The periodic structure provides predictable coverage of distant tokens, while the sparse footprint keeps the per-layer complexity linear in context length. We prove that \PiAttention achieves $\mathcal{O}(kL + π\log L)$ receptive field growth compared to $\mathcal{O}(kL)$ for RingAttention, where $k$ is the local window size, $π$ is the skip period, and $L$ is the sequence length. Extensive experiments on language modeling, retrieval, and vision-language tasks demonstrate that \PiAttention matches or surpasses dense attention quality with 8.3% lower perplexity than RingAttention while using 50% fewer GPUs for the same context length. Our detailed ablations and visualizations reveal the importance of periodic skips, adaptive fusion, and head-level sparsity coordination for efficient long-context modeling.

Method: Constructed balanced dataset by integrating GoEmotions data, emotion-labeled samples from Sentiment140 using RoBERTa-base-GoEmotions model, and GPT-4 mini generated texts. Developed classification model with FastText embeddings, convolutional layers, bidirectional LSTM, attention mechanism, and sigmoid-activated output layer with mixed precision training.

Result: Experimental results show significant improvements in accuracy, precision, recall, F1-score, and AUC compared to models trained on imbalanced data, demonstrating the effectiveness of the balanced dataset and enhanced model architecture.

Conclusion: The balanced dataset and enhanced multi-label classification model effectively address class imbalance issues in sentiment analysis, leading to improved performance across multiple evaluation metrics.

Abstract: Multi-label sentiment classification plays a vital role in natural language processing by detecting multiple emotions within a single text. However, existing datasets like GoEmotions often suffer from severe class imbalance, which hampers model performance, especially for underrepresented emotions. To address this, we constructed a balanced multi-label sentiment dataset by integrating the original GoEmotions data, emotion-labeled samples from Sentiment140 using a RoBERTa-base-GoEmotions model, and manually annotated texts generated by GPT-4 mini. Our data balancing strategy ensured an even distribution across 28 emotion categories. Based on this dataset, we developed an enhanced multi-label classification model that combines pre-trained FastText embeddings, convolutional layers for local feature extraction, bidirectional LSTM for contextual learning, and an attention mechanism to highlight sentiment-relevant words. A sigmoid-activated output layer enables multi-label prediction, and mixed precision training improves computational efficiency. Experimental results demonstrate significant improvements in accuracy, precision, recall, F1-score, and AUC compared to models trained on imbalanced data, highlighting the effectiveness of our approach.

Method: Extended the dataset to over 12,000 questions from ten years of Spanish professional exams, created additional multilingual versions, and benchmarked several open-source LLMs using prompting, RAG (Retrieval-Augmented Generation), and probability-based answer selection techniques.

Result: Results show that performance is mainly driven by model scale and intrinsic reasoning ability, with complex inference strategies obtaining limited gains. The dataset establishes itself as a reliable resource for biomedical reasoning research.

Conclusion: HEAD-QA v2 provides a valuable benchmark for advancing research on biomedical reasoning and model improvement, though current methods show that model scale and intrinsic reasoning capabilities are more important than complex inference strategies.

Abstract: We introduce HEAD-QA v2, an expanded and updated version of a Spanish/English healthcare multiple-choice reasoning dataset originally released by Vilares and Gómez-Rodríguez (2019). The update responds to the growing need for high-quality datasets that capture the linguistic and conceptual complexity of healthcare reasoning. We extend the dataset to over 12,000 questions from ten years of Spanish professional exams, benchmark several open-source LLMs using prompting, RAG, and probability-based answer selection, and provide additional multilingual versions to support future work. Results indicate that performance is mainly driven by model scale and intrinsic reasoning ability, with complex inference strategies obtaining limited gains. Together, these results establish HEAD-QA v2 as a reliable resource for advancing research on biomedical reasoning and model improvement.

Method: Proposes Semantic Pyramid Indexing (SPI) - a multi-resolution vector indexing framework that constructs a semantic pyramid over document embeddings and dynamically selects optimal resolution levels per query using a lightweight classifier, enabling progressive coarse-to-fine retrieval.

Result: SPI achieves up to 5.7× retrieval speedup, 1.8× memory efficiency gain, and improves end-to-end QA F1 scores by up to 2.5 points compared to strong baselines on MS MARCO, Natural Questions, and multimodal retrieval benchmarks.

Conclusion: SPI provides an effective adaptive indexing framework for RAG systems that significantly improves retrieval efficiency while maintaining semantic coverage, with theoretical guarantees and compatibility with existing vector database infrastructures.

Abstract: Retrieval-Augmented Generation (RAG) systems have become a dominant approach to augment large language models (LLMs) with external knowledge. However, existing vector database (VecDB) retrieval pipelines rely on flat or single-resolution indexing structures, which cannot adapt to the varying semantic granularity required by diverse user queries. This limitation leads to suboptimal trade-offs between retrieval speed and contextual relevance. To address this, we propose \textbf{Semantic Pyramid Indexing (SPI)}, a novel multi-resolution vector indexing framework that introduces query-adaptive resolution control for RAG in VecDBs. Unlike existing hierarchical methods that require offline tuning or separate model training, SPI constructs a semantic pyramid over document embeddings and dynamically selects the optimal resolution level per query through a lightweight classifier. This adaptive approach enables progressive retrieval from coarse-to-fine representations, significantly accelerating search while maintaining semantic coverage. We implement SPI as a plugin for both FAISS and Qdrant backends and evaluate it across multiple RAG tasks including MS MARCO, Natural Questions, and multimodal retrieval benchmarks. SPI achieves up to \textbf{5.7$\times$} retrieval speedup and \textbf{1.8$\times$} memory efficiency gain while improving end-to-end QA F1 scores by up to \textbf{2.5 points} compared to strong baselines. Our theoretical analysis provides guarantees on retrieval quality and latency bounds, while extensive ablation studies validate the contribution of each component. The framework’s compatibility with existing VecDB infrastructures makes it readily deployable in production RAG systems. Code is availabe at \href{https://github.com/FastLM/SPI_VecDB}{https://github.com/FastLM/SPI_VecDB}.

Method: Human translation by specialists, evaluation of proprietary and open-source models on translated data, and exploration of machine translation with detailed prompts addressing Estonian linguistic characteristics and WinoGrande-specific translation challenges.

Result: Model performance on human-translated Estonian data is slightly lower than on original English test set; machine-translated data performs notably worse; prompt engineering offers limited improvement in translation quality or model accuracy.

Conclusion: Human translation by language specialists remains essential for reliable evaluation of language competency and reasoning in LLMs, as machine translation and prompt engineering show limited effectiveness for culturally adapted benchmarks.

Abstract: In this paper, we present a localized and culturally adapted Estonian translation of the test set from the widely used commonsense reasoning benchmark, WinoGrande. We detail the translation and adaptation process carried out by translation specialists and evaluate the performance of both proprietary and open source models on the human translated benchmark. Additionally, we explore the feasibility of achieving high-quality machine translation by incorporating insights from the manual translation process into the design of a detailed prompt. This prompt is specifically tailored to address both the linguistic characteristics of Estonian and the unique translation challenges posed by the WinoGrande dataset. Our findings show that model performance on the human translated Estonian dataset is slightly lower than on the original English test set, while performance on machine-translated data is notably worse. Additionally, our experiments indicate that prompt engineering offers limited improvement in translation quality or model accuracy, and highlight the importance of involving language specialists in dataset translation and adaptation to ensure reliable and interpretable evaluations of language competency and reasoning in large language models.

Method: Large-scale systematic evaluation of six state-of-the-art merging methods (including subspace methods) across four open-weight LLMs, twelve fine-tuned checkpoints per base model, and sixteen standard LLM benchmarks in heterogeneous “in-the-wild” settings.

Result: Task Arithmetic (oldest and simplest method) is the only approach that reliably yields performance gains on LLMs in heterogeneous settings. Other interference-aware and subspace merging methods typically don’t improve over the base model.

Conclusion: Current merging techniques mostly fail to extract useful weight updates from heterogeneous/conflicting experts, motivating the need for LLM-specific merging algorithms and merging-aware fine-tuning methods.

Abstract: Model merging combines multiple fine-tuned checkpoints into a single model without additional training, offering an attractive approach to reusing models and efficiently improving performance. However, it remains unclear whether the advantages reported for settings where all merged experts have distinct roles and are tuned on clearly separated tasks also hold in settings where the merged experts do not have clearly distinct roles, but are trained on overlapping or even conflicting objectives. To evaluate this setting, we present a large-scale, systematic evaluation of “in-the-wild” model merging of heterogeneous experts, that may have been trained on overlapping or conflicting objectives. Concretely, we evaluate six state-of-the-art merging methods, including recent subspace methods, across four open-weight LLMs, twelve fine-tuned checkpoints per base model, and sixteen standard LLM benchmarks. Evaluating through standardized benchmarks, we measure both the probability that a model merged from a heterogeneous set of experts outperforms the base model and we measure relative gains over the best individual checkpoint. Our results show that the oldest and simplest method, Task Arithmetic, is the only approach that reliably yields performance gains on LLMs in this “in-the-wild” setting. Other interference-aware and subspace merging methods typically do not result in notable improvements over the base model. Our findings indicate that current merging techniques mostly do not enable extracting useful weight updates from heterogeneous and potentially conflicting versions. This motivates the design of LLM-specific merging algorithms and merging-aware fine-tuning methods.

Method: CREST uses cluster-based cross-lingual transfer, training on only 13 high-resource languages and transferring knowledge to 100 languages. It’s parameter-efficient with only 0.5B parameters.

Result: CREST outperforms existing state-of-the-art guardrails of comparable scale and achieves competitive results against models with significantly larger parameter counts (2.5B+).

Conclusion: The work highlights limitations of language-specific guardrails and underscores the importance of developing universal, language-agnostic safety systems that can scale effectively to serve global populations.

Abstract: Ensuring content safety in large language models (LLMs) is essential for their deployment in real-world applications. However, existing safety guardrails are predominantly tailored for high-resource languages, leaving a significant portion of the world’s population underrepresented who communicate in low-resource languages. To address this, we introduce CREST (CRoss-lingual Efficient Safety Transfer), a parameter-efficient multilingual safety classification model that supports 100 languages with only 0.5B parameters. By training on a strategically chosen subset of only 13 high-resource languages, our model utilizes cluster-based cross-lingual transfer from a few to 100 languages, enabling effective generalization to both unseen high-resource and low-resource languages. This approach addresses the challenge of limited training data in low-resource settings. We conduct comprehensive evaluations across six safety benchmarks to demonstrate that CREST outperforms existing state-of-the-art guardrails of comparable scale and achieves competitive results against models with significantly larger parameter counts (2.5B parameters and above). Our findings highlight the limitations of language-specific guardrails and underscore the importance of developing universal, language-agnostic safety systems that can scale effectively to serve global populations.

Method: Used retrieval-augmented generation over medical Wikipedia information to create 500k reasoning traces in English, Italian, and Spanish. Extended MedQA and MedMCQA datasets to Italian and Spanish. Tested pipeline in both in-domain and out-of-domain settings.

Result: Reasoning traces improved performance via both in-context learning (few-shot) and supervised fine-tuning, achieving state-of-the-art results among 8B-parameter LLMs on medical QA benchmarks.

Conclusion: The generated multilingual reasoning resources support development of more transparent clinical decision-support tools and enable better medical QA performance across languages.

Abstract: Large Language Models (LLMs) with reasoning capabilities have recently demonstrated strong potential in medical Question Answering (QA). Existing approaches are largely English-focused and primarily rely on distillation from general-purpose LLMs, raising concerns about the reliability of their medical knowledge. In this work, we present a method to generate multilingual reasoning traces based on medical knowledge extracted from Wikipedia. We produce 500k traces in English, Italian, and Spanish, using a retrieval-augmented generation approach over medical information from Wikipedia. The traces are generated to solve medical questions drawn from MedQA and MedMCQA, which we extend to Italian and Spanish. We test our pipeline in both in-domain and out-of-domain settings across Medical QA benchmarks, and demonstrate that our reasoning traces improve performance both when utilized via in-context learning (few-shot) and supervised fine-tuning, yielding state-of-the-art results among 8B-parameter LLMs. We believe that these resources can support the development of more transparent clinical decision-support tools in multilingual settings. We release the full suite of resources: reasoning traces, translated QA datasets, Medical-Wikipedia, and fine-tuned models.

Method: Developed a comprehensive new coding scheme with 38 counselor and 28 client utterance categories, created a labeled dataset of ~2,800 messages from counseling conversations, and fine-tuned several models on the dataset.

Result: Created a publicly available dataset and models for fine-grained classification of online counseling conversations, demonstrating applicability through model fine-tuning experiments.

Conclusion: The work contributes a new fine-grained conversational resource to the language resources community, extending existing datasets for social and mental-health dialogue analysis.

Abstract: This paper presents OnCoCo 1.0, a new public dataset for fine-grained message classification in online counseling. It is based on a new, integrative system of categories, designed to improve the automated analysis of psychosocial online counseling conversations. Existing category systems, predominantly based on Motivational Interviewing (MI), are limited by their narrow focus and dependence on datasets derived mainly from face-to-face counseling. This limits the detailed examination of textual counseling conversations. In response, we developed a comprehensive new coding scheme that differentiates between 38 types of counselor and 28 types of client utterances, and created a labeled dataset consisting of about 2.800 messages from counseling conversations. We fine-tuned several models on our dataset to demonstrate its applicability. The data and models are publicly available to researchers and practitioners. Thus, our work contributes a new type of fine-grained conversational resource to the language resources community, extending existing datasets for social and mental-health dialogue analysis.

Method: Proposed evaluation framework using MMMLU benchmark to test three LLMs (GPT-4o mini, Gemini 2.0 Flash, Llama 4 Scout) under Very Polite, Neutral, and Very Rude prompt variants across six STEM and Humanities tasks, with statistical significance testing of pairwise accuracy differences.

Result: Tone sensitivity is model-dependent and domain-specific: Neutral or Very Polite prompts generally yield higher accuracy than Very Rude prompts, but statistically significant effects appear only in Humanities tasks where rude tone reduces accuracy for GPT and Llama, while Gemini remains tone-insensitive. Aggregated across tasks, tone effects diminish and lose statistical significance.

Conclusion: While interaction tone can matter in specific interpretive settings, modern LLMs are broadly robust to tonal variation in typical mixed-domain use, providing practical guidance for prompt design and model selection in real-world deployments.

Abstract: Prompt engineering has emerged as a critical factor influencing large language model (LLM) performance, yet the impact of pragmatic elements such as linguistic tone and politeness remains underexplored, particularly across different model families. In this work, we propose a systematic evaluation framework to examine how interaction tone affects model accuracy and apply it to three recently released and widely available LLMs: GPT-4o mini (OpenAI), Gemini 2.0 Flash (Google DeepMind), and Llama 4 Scout (Meta). Using the MMMLU benchmark, we evaluate model performance under Very Polite, Neutral, and Very Rude prompt variants across six tasks spanning STEM and Humanities domains, and analyze pairwise accuracy differences with statistical significance testing. Our results show that tone sensitivity is both model-dependent and domain-specific. Neutral or Very Polite prompts generally yield higher accuracy than Very Rude prompts, but statistically significant effects appear only in a subset of Humanities tasks, where rude tone reduces accuracy for GPT and Llama, while Gemini remains comparatively tone-insensitive. When performance is aggregated across tasks within each domain, tone effects diminish and largely lose statistical significance. Compared with earlier research, these findings suggest that dataset scale and coverage materially influence the detection of tone effects. Overall, our study indicates that while interaction tone can matter in specific interpretive settings, modern LLMs are broadly robust to tonal variation in typical mixed-domain use, providing practical guidance for prompt design and model selection in real-world deployments.

Method: Created human-validated framework to evaluate if reasoning traces logically support conclusions across languages. Analyzed 65k reasoning traces from GlobalMMLU questions across 6 languages and 6 frontier models.

Result: Found critical blind spot: high task accuracy but reasoning often fails to support conclusions. Non-Latin scripts show ≥2× more reasoning-conclusion misalignment than Latin scripts. Error taxonomy reveals evidential errors (unsupported claims) and illogical reasoning as main failures.

Conclusion: Current multilingual evaluation provides incomplete picture of model reasoning capabilities; need reasoning-aware evaluation frameworks that assess reasoning quality, not just task accuracy.

Abstract: Large language models demonstrate strong reasoning capabilities through chain-of-thought prompting, but whether this reasoning quality transfers across languages remains underexplored. We introduce a human-validated framework to evaluate whether model-generated reasoning traces logically support their conclusions across languages. Analyzing 65k reasoning traces from GlobalMMLU questions across 6 languages and 6 frontier models, we uncover a critical blind spot: while models achieve high task accuracy, their reasoning can fail to support their conclusions. Reasoning traces in non-Latin scripts show at least twice as much misalignment between their reasoning and conclusions than those in Latin scripts. We develop an error taxonomy through human annotation to characterize these failures, finding they stem primarily from evidential errors (unsupported claims, ambiguous facts) followed by illogical reasoning steps. Our findings demonstrate that current multilingual evaluation practices provide an incomplete picture of model reasoning capabilities and highlight the need for reasoning-aware evaluation frameworks.

Method: Proposes extracting a single low-dimensional direction from contrastive prompt sets using one prompt-only forward pass, then applying global interventions on residual-stream activations throughout reverse diffusion. Uses safety refusal as a case study to analyze refusal behavior in MDLMs and identify consistent activation subspaces.

Result: Found refusal behavior in multiple MDLMs is governed by a consistent, approximately one-dimensional activation subspace. The steering method yields large systematic behavioral shifts, outperforming prompt-based and optimization-based baselines. Discovered diffusion-specific accessibility where effective directions can be extracted from pre-instruction tokens (unlike autoregressive models). Ablations show maximal leverage in early denoising steps and mid-to-late transformer layers.

Conclusion: The activation steering primitive enables efficient representation-level control in MDLMs, revealing architecture-dependent representations of safety constraints. Directions transfer strongly between languages in multilingual MDLMs but don’t generalize to autoregressive architectures, highlighting fundamental differences in how different model architectures encode behavioral constraints.

Abstract: Masked diffusion language models (MDLMs) generate text via iterative masked-token denoising, enabling mask-parallel decoding and distinct controllability and efficiency tradeoffs from autoregressive LLMs. Yet, efficient representation-level mechanisms for inference-time control in MDLMs remain largely unexplored. To address this gap, we introduce an activation steering primitive for MDLMs: we extract a single low-dimensional direction from contrastive prompt sets using one prompt-only forward pass, and apply a global intervention on residual-stream activations throughout reverse diffusion, without performing optimization or altering the diffusion sampling procedure. Using safety refusal as a deployment-relevant case study, we find that refusal behavior in multiple MDLMs is governed by a consistent, approximately one-dimensional activation subspace. Applying the corresponding direction yields large and systematic behavioral shifts and is substantially more effective than prompt-based and optimization-based baselines. We further uncover diffusion-specific accessibility: effective directions can be extracted not only from post-instruction tokens, but also from pre-instruction tokens that are typically ineffective in autoregressive models due to causal attention. Ablations localize maximal leverage to early denoising steps and mid-to-late transformer layers, with early diffusion blocks contributing disproportionately. Finally, in an MDLM trained on English and Chinese, extracted directions transfer strongly between English and Chinese, but do not reliably generalize to an autoregressive architecture, highlighting architecture-dependent representations of safety constraints.

Method: Created benchmark based on 67 Japan Medical Association guidelines with over 50,000 adversarial conversations using 7 automatically discovered jailbreak strategies. Used dual-LLM scoring protocol to evaluate 27 models.

Result: Commercial models maintain robust safety while medical-specialized models show increased vulnerability. Safety scores decline significantly across conversation turns (median: 9.5 to 5.0). Cross-lingual evaluation shows vulnerabilities persist across Japanese and English versions.

Conclusion: Domain-specific fine-tuning may weaken safety mechanisms, multi-turn interactions represent distinct threat surface requiring dedicated alignment strategies, and medical model vulnerabilities are inherent alignment limitations rather than language-specific.

Abstract: As Large Language Models (LLMs) are increasingly deployed in healthcare field, it becomes essential to carefully evaluate their medical safety before clinical use. However, existing safety benchmarks remain predominantly English-centric, and test with only single-turn prompts despite multi-turn clinical consultations. To address these gaps, we introduce JMedEthicBench, the first multi-turn conversational benchmark for evaluating medical safety of LLMs for Japanese healthcare. Our benchmark is based on 67 guidelines from the Japan Medical Association and contains over 50,000 adversarial conversations generated using seven automatically discovered jailbreak strategies. Using a dual-LLM scoring protocol, we evaluate 27 models and find that commercial models maintain robust safety while medical-specialized models exhibit increased vulnerability. Furthermore, safety scores decline significantly across conversation turns (median: 9.5 to 5.0, $p < 0.001$). Cross-lingual evaluation on both Japanese and English versions of our benchmark reveals that medical model vulnerabilities persist across languages, indicating inherent alignment limitations rather than language-specific factors. These findings suggest that domain-specific fine-tuning may accidentally weaken safety mechanisms and that multi-turn interactions represent a distinct threat surface requiring dedicated alignment strategies.

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

Result: FACTUM outperforms state-of-the-art baselines by up to 37.5% in AUC. Correct citations show higher parametric force and greater attention sink usage. The signature of correctness evolves with model scale - 3B models rely on high pathway alignment while 8B models use specialized orthogonal strategies.

Conclusion: Citation hallucinations result from complex coordination failures between neural pathways, not just parametric over-reliance. High parametric force can be constructive when properly coordinated with attention pathways, enabling more reliable RAG systems.

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

Method: Created DISTRACTMATH-BN benchmark by augmenting MGSM and MSVAMP with semantically coherent but computationally irrelevant information. Proposed DAGGER, which reformulates mathematical problem solving as executable computational graph generation with explicit modeling of distractor nodes. Fine-tuned Gemma-3 models using supervised fine-tuning followed by Group Relative Policy Optimization.

Result: Standard models dropped by up to 41 points under distractors, while reasoning-specialized models declined by 14-20 points despite using 5x more tokens. DAGGER achieved comparable weighted accuracy on augmented benchmarks while using 89% fewer tokens than reasoning models, with robustness emerging without explicit training on distractor-augmented examples.

Conclusion: Enforcing structured intermediate representations (computational graphs) improves robustness and inference efficiency in mathematical reasoning compared to free-form approaches, particularly in noisy, low-resource settings.

Abstract: Chain-of-Thought (CoT) prompting is widely adopted for mathematical problem solving, including in low-resource languages, yet its behavior under irrelevant context remains underexplored. To systematically study this challenge, we introduce DISTRACTMATH-BN, a Bangla benchmark that augments MGSM and MSVAMP with semantically coherent but computationally irrelevant information. Evaluating seven models ranging from 3B to 12B parameters, we observe substantial performance degradation under distractors: standard models drop by up to 41 points, while reasoning-specialized models decline by 14 to 20 points despite consuming five times more tokens. We propose †DAGGER, which reformulates mathematical problem solving as executable computational graph generation with explicit modeling of distractor nodes. Fine-tuning Gemma-3 models using supervised fine-tuning followed by Group Relative Policy Optimization achieves comparable weighted accuracy on augmented benchmarks while using 89 percent fewer tokens than reasoning models. Importantly, this robustness emerges without explicit training on distractor-augmented examples. Our results suggest that enforcing structured intermediate representations improves robustness and inference efficiency in mathematical reasoning compared to free-form approaches, particularly in noisy, low-resource settings.

Method: Uses a Teacher-Student knowledge distillation architecture. The Teacher learns from articulatory phonetic features derived from IPA transcriptions, then transfers this knowledge to a character-level Student model. Trained on 32.7 million triplet samples from 67 million toponyms across GeoNames, Wikidata, and Getty Thesaurus.

Result: Achieves highest Recall@1 (85.2%) and Mean Reciprocal Rank (90.8%) on the MEHDIE cross-script benchmark (medieval Hebrew and Arabic toponym matches). Shows cross-temporal generalization from modern training to pre-modern sources. Ablation with raw articulatory features alone yields only 45.0% MRR.

Conclusion: The approach effectively handles pre-standardization orthographic variation in historical documents and transfers to personal names in archival sources, suggesting broad applicability to name resolution tasks in digital humanities and linked open data contexts.

Abstract: Matching place names across writing systems is a persistent obstacle to the integration of multilingual geographic sources, whether modern gazetteers, medieval itineraries, or colonial-era surveys. Existing approaches depend on language-specific phonetic algorithms or romanisation steps that discard phonetic information, and none generalises across script boundaries. This paper presents Symphonym, a neural embedding system which maps toponyms from twenty writing systems into a unified 128-dimensional phonetic space, enabling direct cross-script similarity comparison without language identification or phonetic resources at inference time. A Teacher-Student knowledge distillation architecture first learns from articulatory phonetic features derived from IPA transcriptions, then transfers this knowledge to a character-level Student model. Trained on 32.7 million triplet samples drawn from 67 million toponyms spanning GeoNames, Wikidata, and the Getty Thesaurus of Geographic Names, the Student achieves the highest Recall@1 (85.2%) and Mean Reciprocal Rank (90.8%) on the MEHDIE cross-script benchmark – medieval Hebrew and Arabic toponym matches curated by domain experts and entirely independent of the training data – demonstrating cross-temporal generalisation from modern training material to pre-modern sources. An ablation using raw articulatory features alone yields only 45.0% MRR, confirming the contribution of the neural training curriculum. The approach naturally handles pre-standardisation orthographic variation characteristic of historical documents, and transfers effectively to personal names in archival sources, suggesting broad applicability to name resolution tasks in digital humanities and linked open data contexts.

Method: Evaluated LLMs on diverse programs from the Termination category of SV-Comp 2025, comparing performance against specialized verification tools. Analyzed models including GPT-5, Claude Sonnet-4.5, and Code World Model (CWM) on termination prediction tasks.

Result: LLMs perform remarkably well at predicting program termination, with GPT-5 and Claude Sonnet-4.5 ranking just behind the top-ranked tool (using test-time-scaling), and CWM placing just behind the second-ranked tool. However, LLMs often fail to provide valid witnesses as proofs, and performance degrades with increasing program length and complexity.

Conclusion: LLMs show promise for program termination prediction despite the undecidable nature of the problem, but have limitations in proof generation and handling complex programs. This motivates further research into LLMs for reasoning about undecidable problems.

Abstract: Determining whether a program terminates is a central problem in computer science. Turing’s foundational result established the Halting Problem as undecidable, showing that no algorithm can universally determine termination for all programs and inputs. Consequently, automatic verification tools approximate termination, sometimes failing to prove or disprove; these tools rely on problem-specific architectures, and are usually tied to particular programming languages. Recent success and progress in large language models (LLMs) raises the following question: can LLMs reliably predict program termination? In this work, we evaluate LLMs on a diverse set of programs from the Termination category of the International Competition on Software Verification (SV-Comp) 2025. Our results suggest that LLMs perform remarkably well at predicting program termination, where GPT-5 and Claude Sonnet-4.5 would rank just behind the top-ranked tool (using test-time-scaling), and Code World Model (CWM) would place just behind the second-ranked tool. While LLMs are effective at predicting program termination, they often fail to provide a valid witness as a proof. Moreover, LLMs performance drops as program length and complexity increases. We hope these insights motivate further research into program termination and the broader potential of LLMs for reasoning about undecidable problems.

Method: Proposes MuVaC framework that: 1) Models Multimodal Sarcasm Detection (MSD) and Multimodal Sarcasm Explanation (MuSE) using structural causal models with variational causal pathways for joint optimization; 2) Uses alignment-then-fusion approach for robust multimodal feature integration; 3) Ensures consistency between detection results and explanations for trustworthy reasoning.

Result: Experimental results show MuVaC’s superiority on public datasets, offering a new perspective for understanding multimodal sarcasm through joint optimization of detection and explanation.

Conclusion: MuVaC successfully bridges the gap between sarcasm detection and explanation by modeling their causal relationship, providing a framework that mimics human cognitive mechanisms for understanding multimodal sarcasm.

Abstract: The prevalence of sarcasm in multimodal dialogues on the social platforms presents a crucial yet challenging task for understanding the true intent behind online content. Comprehensive sarcasm analysis requires two key aspects: Multimodal Sarcasm Detection (MSD) and Multimodal Sarcasm Explanation (MuSE). Intuitively, the act of detection is the result of the reasoning process that explains the sarcasm. Current research predominantly focuses on addressing either MSD or MuSE as a single task. Even though some recent work has attempted to integrate these tasks, their inherent causal dependency is often overlooked. To bridge this gap, we propose MuVaC, a variational causal inference framework that mimics human cognitive mechanisms for understanding sarcasm, enabling robust multimodal feature learning to jointly optimize MSD and MuSE. Specifically, we first model MSD and MuSE from the perspective of structural causal models, establishing variational causal pathways to define the objectives for joint optimization. Next, we design an alignment-then-fusion approach to integrate multimodal features, providing robust fusion representations for sarcasm detection and explanation generation. Finally, we enhance the reasoning trustworthiness by ensuring consistency between detection results and explanations. Experimental results demonstrate the superiority of MuVaC in public datasets, offering a new perspective for understanding multimodal sarcasm.

Method: Uses instruction-tuned SLMs with history summarization strategy to preserve conversational state. Evaluates nine instruction-tuned low-parameterized SLMs against three commercial LLMs using lexical/semantic similarity metrics, human evaluation, and LLM-as-a-judge methods. Introduces conversation stage-based qualitative analysis.

Result: Results show notable variation across SLMs - some demonstrate near-LLM performance while others struggle with dialogue continuity and contextual alignment. Highlights both potential and limitations of low-parameterized models for real-world customer-service QA.

Conclusion: SLMs show promise for efficient customer-service QA but current limitations in maintaining dialogue continuity and contextual understanding need addressing. The study provides insights into practical deployment of smaller models in resource-constrained environments.

Abstract: Customer-service question answering (QA) systems increasingly rely on conversational language understanding. While Large Language Models (LLMs) achieve strong performance, their high computational cost and deployment constraints limit practical use in resource-constrained environments. Small Language Models (SLMs) provide a more efficient alternative, yet their effectiveness for multi-turn customer-service QA remains underexplored, particularly in scenarios requiring dialogue continuity and contextual understanding. This study investigates instruction-tuned SLMs for context-summarized multi-turn customer-service QA, using a history summarization strategy to preserve essential conversational state. We also introduce a conversation stage-based qualitative analysis to evaluate model behavior across different phases of customer-service interactions. Nine instruction-tuned low-parameterized SLMs are evaluated against three commercial LLMs using lexical and semantic similarity metrics alongside qualitative assessments, including human evaluation and LLM-as-a-judge methods. Results show notable variation across SLMs, with some models demonstrating near-LLM performance, while others struggle to maintain dialogue continuity and contextual alignment. These findings highlight both the potential and current limitations of low-parameterized language models for real-world customer-service QA systems.

Method: SEAD decouples user modeling into two components: a Profile Controller that generates diverse user states to manage training curriculum, and a User Role-play Model that focuses on realistic role-playing, ensuring adaptive training scenarios rather than adversarial environments.

Result: SEAD significantly outperforms both open-source foundation models and closed-source commercial models, improving task completion rate by 17.6% and dialogue efficiency by 11.1%.

Conclusion: SEAD provides an effective framework for training service dialogue agents without large-scale human annotations, addressing key limitations in current approaches through self-evolving user modeling.

Abstract: Large Language Models have demonstrated remarkable capabilities in open-domain dialogues. However, current methods exhibit suboptimal performance in service dialogues, as they rely on noisy, low-quality human conversation data. This limitation arises from data scarcity and the difficulty of simulating authentic, goal-oriented user behaviors. To address these issues, we propose SEAD (Self-Evolving Agent for Service Dialogue), a framework that enables agents to learn effective strategies without large-scale human annotations. SEAD decouples user modeling into two components: a Profile Controller that generates diverse user states to manage training curriculum, and a User Role-play Model that focuses on realistic role-playing. This design ensures the environment provides adaptive training scenarios rather than acting as an unfair adversary. Experiments demonstrate that SEAD significantly outperforms Open-source Foundation Models and Closed-source Commercial Models, improving task completion rate by 17.6% and dialogue efficiency by 11.1%. Code is available at: https://github.com/Da1yuqin/SEAD.

Method: Builds image-audio retrieval-augmented generation module for fetching relevant frames/audio snippets, uses agent loop for planning/tool calling across turns, and applies group relative policy optimization for joint improvement.

Result: Outperforms prior methods on OmniVideoBench, WorldSense, and Daily-Omni under low-resource settings, with ablations validating each component.

Conclusion: OmniRAG-Agent effectively addresses key challenges in long-horizon omnimodal QA through retrieval-augmented generation, agentic planning, and optimization techniques.

Abstract: Long-horizon omnimodal question answering answers questions by reasoning over text, images, audio, and video. Despite recent progress on OmniLLMs, low-resource long audio-video QA still suffers from costly dense encoding, weak fine-grained retrieval, limited proactive planning, and no clear end-to-end optimization. To address these issues, we propose OmniRAG-Agent, an agentic omnimodal QA method for budgeted long audio-video reasoning. It builds an image-audio retrieval-augmented generation module that lets an OmniLLM fetch short, relevant frames and audio snippets from external banks. Moreover, it uses an agent loop that plans, calls tools across turns, and merges retrieved evidence to answer complex queries. Furthermore, we apply group relative policy optimization to jointly improve tool use and answer quality over time. Experiments on OmniVideoBench, WorldSense, and Daily-Omni show that OmniRAG-Agent consistently outperforms prior methods under low-resource settings and achieves strong results, with ablations validating each component.

Method: Created GreekMMLU benchmark with 21,805 multiple-choice questions across 45 subject areas, sourced or authored in Greek from academic, professional, and governmental exams. Organized under new subject taxonomy with educational difficulty levels. Released 16,857 samples publicly and reserved 4,948 for private leaderboard.

Result: Evaluation of 80+ LLMs revealed substantial performance gaps: frontier vs open-weight models, and Greek-adapted vs general multilingual models. Systematic analysis identified factors influencing performance including model scale, adaptation, and prompting.

Conclusion: GreekMMLU provides a robust, contamination-resistant benchmark for evaluating Greek language understanding in LLMs, revealing current limitations and offering insights for improving Greek language capabilities through better adaptation and scaling.

Abstract: Large Language Models (LLMs) are commonly trained on multilingual corpora that include Greek, yet reliable evaluation benchmarks for Greek-particularly those based on authentic, native-sourced content-remain limited. Existing datasets are often machine-translated from English, failing to capture Greek linguistic and cultural characteristics. We introduce GreekMMLU, a native-sourced benchmark for massive multitask language understanding in Greek, comprising 21,805 multiple-choice questions across 45 subject areas, organized under a newly defined subject taxonomy and annotated with educational difficulty levels spanning primary to professional examinations. All questions are sourced or authored in Greek from academic, professional, and governmental exams. We publicly release 16,857 samples and reserve 4,948 samples for a private leaderboard to enable robust and contamination-resistant evaluation. Evaluations of over 80 open- and closed-source LLMs reveal substantial performance gaps between frontier and open-weight models, as well as between Greek-adapted models and general multilingual ones. Finally, we provide a systematic analysis of factors influencing performance-including model scale, adaptation, and prompting-and derive insights for improving LLM capabilities in Greek.

Method: Proposes an automated framework using LLMs to label KC-level correctness directly from student-written code. Includes assessment of whether each KC is correctly applied and introduces a temporal context-aware Code-KC mapping mechanism to better align KCs with individual student code.

Result: Experimental results show the framework leads to learning curves more consistent with cognitive theory and improves predictive performance compared to baselines. Human evaluation demonstrates substantial agreement between LLM and expert annotations.

Conclusion: The LLM-based framework effectively addresses the KC-level labeling challenge in programming education, enabling more accurate student modeling and learning analytics for open-ended tasks.

Abstract: Fine-grained skill representations, commonly referred to as knowledge components (KCs), are fundamental to many approaches in student modeling and learning analytics. However, KC-level correctness labels are rarely available in real-world datasets, especially for open-ended programming tasks where solutions typically involve multiple KCs simultaneously. Simply propagating problem-level correctness to all associated KCs obscures partial mastery and often leads to poorly fitted learning curves. To address this challenge, we propose an automated framework that leverages large language models (LLMs) to label KC-level correctness directly from student-written code. Our method assesses whether each KC is correctly applied and further introduces a temporal context-aware Code-KC mapping mechanism to better align KCs with individual student code. We evaluate the resulting KC-level correctness labels in terms of learning curve fit and predictive performance using the power law of practice and the Additive Factors Model. Experimental results show that our framework leads to learning curves that are more consistent with cognitive theory and improves predictive performance, compared to baselines. Human evaluation further demonstrates substantial agreement between LLM and expert annotations.

Method: MetaState introduces three modules with shared time conditioning: a cross-attention Mixer to read backbone activations into memory slots, a GRU-style Updater to integrate information across steps, and a cross-attention Injector to write updated memory back into the backbone. Trained with K-step unrolling pipeline to learn multi-step dynamics.

Result: MetaState adds only ~0.6% trainable parameters while keeping backbone frozen, and consistently improves reasoning performance over frozen baselines on mathematical reasoning and code generation benchmarks, with average gain of 4.5% across all evaluations.

Conclusion: MetaState effectively addresses the Information Island bottleneck in discrete diffusion models by providing persistent working memory, enabling better reasoning performance with minimal computational overhead.

Abstract: Discrete diffusion language models (dLLMs) generate text by iteratively denoising a masked sequence. However, standard dLLMs condition each denoising step solely on the current hard-masked sequence, while intermediate continuous representations are discarded after sampling and remasking. We term this bottleneck the \textbf{Information Island} issue: continuous information remains isolated within individual denoising steps and fails to propagate across the trajectory. This bottleneck is especially harmful for reasoning, which requires intermediate reasoning state to be preserved and updated across many denoising steps. To address this limitation, we introduce \textbf{MetaState}, a lightweight recurrent augmentation that equips a frozen dLLM backbone with persistent, fixed-size working memory. MetaState comprises three modules with a shared time conditioner: a cross-attention \textbf{Mixer} that reads backbone activations into memory slots, a GRU-style \textbf{Updater} that integrates information across steps, and a cross-attention \textbf{Injector} that writes the updated memory back into the backbone. We train these modules with a dedicated $K$-step unrolling pipeline to learn multi-step dynamics. MetaState adds only ${\sim}0.6%$ trainable parameters while keeping the backbone frozen, and consistently improves reasoning performance over frozen baselines on mathematical reasoning and code generation benchmarks, with an average gain of $4.5%$ across all evaluations.

Method: Developed a browser-based tool (VERIFICATION ASSISTANT) that allows users to submit URLs/media files, automatically extracts content, routes it to backend NLP classifiers, and presents credibility signals in an accessible format.

Result: The tool is part of the widely adopted VERIFICATION PLUGIN (140,000+ users), provides actionable credibility signals, estimates AI-generated content, and offers verification guidance in a clear format.

Conclusion: The VERIFICATION ASSISTANT successfully bridges the gap between advanced NLP detection methods and non-expert users by integrating multiple credibility analysis services into a practical, accessible tool for real-world disinformation detection.

Abstract: Disinformation and false content produced by generative AI pose a significant challenge for journalists and fact-checkers who must rapidly verify digital media information. While there is an abundance of NLP models for detecting credibility signals such as persuasion techniques, subjectivity, or machine-generated text, such methods often remain inaccessible to non-expert users and are not integrated into their daily workflows as a unified framework. This paper demonstrates the VERIFICATION ASSISTANT, a browser-based tool designed to bridge this gap. The VERIFICATION ASSISTANT, a core component of the widely adopted VERIFICATION PLUGIN (140,000+ users), allows users to submit URLs or media files to a unified interface. It automatically extracts content and routes it to a suite of backend NLP classifiers, delivering actionable credibility signals, estimating AI-generated content, and providing other verification guidance in a clear, easy-to-digest format. This paper showcases the tool architecture, its integration of multiple NLP services, and its real-world application to detecting disinformation.

Method: Presents the FUSE taxonomy organized along Foundations (theoretical underpinnings), Unification Strategies (algorithmic approaches), Scenarios (downstream applications), and Ecosystem (tools/benchmarks). Reviews weight averaging, task vector arithmetic, sparsification-enhanced methods, mixture-of-experts, and evolutionary optimization.

Result: Comprehensive survey establishing structured foundation for model merging research, covering theoretical foundations, algorithmic approaches, practical applications, and supporting tools/benchmarks.

Conclusion: Provides researchers and practitioners with structured foundation for advancing model merging, identifies key open challenges and future directions in the field.

Abstract: Model merging combines the parameters of multiple neural networks into a single model without additional training. As fine-tuned large language models (LLMs) proliferate, merging offers a computationally efficient alternative to ensembles and full retraining, enabling practitioners to compose specialized capabilities at minimal cost. This survey examines model merging in the LLM era through the \textbf{FUSE} taxonomy, organized along \textbf{F}oundations, \textbf{U}nification Strategies, \textbf{S}cenarios, and \textbf{E}cosystem. We first establish the theoretical underpinnings of merging, including loss landscape geometry and mode connectivity, then systematically review the algorithmic space spanning weight averaging, task vector arithmetic, sparsification-enhanced methods, mixture-of-experts architectures, and evolutionary optimization. We further examine downstream applications across multi-task learning, safety alignment, domain specialization, and federated learning, and survey the supporting ecosystem of tools and evaluation benchmarks. Finally, we identify key open challenges and future directions, aiming to equip researchers and practitioners with a structured foundation for advancing model merging.

Method: Paired-trajectory protocol replaying real financial dialogues under clean vs. contaminated tool-output conditions across 8 LLMs (7B to frontier), decomposing divergence into information-channel and memory-channel mechanisms, with causal interventions (activation patching, feature clamping, direct steering) and sparse autoencoder probing.

Result: Evaluation blindness observed: recommendation quality preserved under contamination (UPR~1.0) while 65-93% of turns contain risk-inappropriate products; violations emerge at turn 1 and persist without self-correction; susceptibility scales with instruction-following fidelity; models internally distinguish adversarial perturbations but fail to propagate signal to output.

Conclusion: Trajectory-level safety monitoring needed for deployed multi-turn agents; safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74; representation-to-action gap is structural and resists linear repair.

Abstract: Tool-augmented LLM agents increasingly operate as multi-turn advisors in high-stakes domains, yet their evaluation relies on ranking metrics that measure what is recommended but not whether it is safe for the user. We present a paired-trajectory protocol that replays real financial dialogues under clean and contaminated tool-output conditions across eight LLMs (7B to frontier), decomposing divergence into information-channel and memory-channel mechanisms. We observe evaluation blindness: recommendation quality is preserved under contamination (UPR~1.0) while risk-inappropriate products appear in 65-93% of turns, invisible to standard NDCG. Violations are information-channel-driven, emerge at turn 1, and persist without self-correction over 23-step trajectories. Even non-extreme perturbations (within-band corruption, narrative-only attacks) evade threshold monitors while producing significant drift. Susceptibility scales with instruction-following fidelity across all eight models. Sparse autoencoder probing reveals models internally distinguish adversarial perturbations but fail to propagate this signal to output; causal interventions (activation patching, feature clamping, direct steering) confirm this representation-to-action gap is structural and resists linear repair. A safety-penalized NDCG variant (sNDCG) reduces preservation ratios to 0.51-0.74. These results motivate trajectory-level safety monitoring for deployed multi-turn agents.

Method: Human professionals collected, translated, and annotated parallel sentence pairs between local Ghanaian languages and English. Data enriched with standard structural metadata for consistency and usability.

Result: Created 41,513 parallel sentence pairs across 5 Ghanaian languages. Deployed in real-world applications like Khaya AI translation engine. Datasets support 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 in low-resource language contexts.

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: Participated in all four subtasks of ArchEHR-QA 2026 shared task, evaluating multiple approaches designed to run on commodity hardware without external APIs or cloud infrastructure, using properly configured smaller models.

Result: Systems achieved competitive performance on shared task leaderboards, performing above average in two subtasks, with smaller models approaching performance of larger systems when properly configured.

Conclusion: Privacy-preserving EHR QA systems running fully locally are feasible with current models and commodity hardware, enabling clinical deployment without compromising data privacy.

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: Proposes ExPosST framework with explicit position allocation that reserves fixed positional slots for incoming source tokens, enabling efficient decoding with KV cache 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.

Conclusion: ExPosST provides a general framework that resolves the positional mismatch dilemma in applying LLMs to simultaneous translation, achieving better balance between efficiency and consistency.

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: Created BanglaSocialBench with 1,719 culturally grounded instances across three domains: Bangla Address Terms, Kinship Reasoning, and Social Customs. All content written and verified by native Bangla speakers. Evaluated twelve contemporary LLMs in zero-shot settings to assess sociopragmatic competence.

Result: LLMs show systematic patterns of cultural misalignment: default to overly formal address forms, fail to recognize multiple socially acceptable address pronouns, conflate kinship terminology across religious contexts. Sociopragmatic failures are structured and non-random, revealing persistent limitations in culturally appropriate language inference.

Conclusion: Current LLMs have significant limitations in inferring and applying culturally appropriate language use in realistic Bangladeshi social interactions. Sociopragmatic competence requires deeper cultural understanding beyond factual recall, highlighting the need for culturally grounded benchmarks and model improvements.

Abstract: Large Language Models have demonstrated strong multilingual fluency, yet fluency alone does not guarantee socially appropriate language use. In high-context languages, communicative competence requires sensitivity to social hierarchy, relational roles, and interactional norms that are encoded directly in everyday language. Bangla exemplifies this challenge through its three-tiered pronominal system, kinship-based addressing, and culturally embedded social customs. We introduce BanglaSocialBench, the first benchmark designed to evaluate sociopragmatic competence in Bangla through context-dependent language use rather than factual recall. The benchmark spans three domains: Bangla Address Terms, Kinship Reasoning, and Social Customs, and consists of 1,719 culturally grounded instances written and verified by native Bangla speakers. We evaluate twelve contemporary LLMs in a zero-shot setting and observe systematic patterns of cultural misalignment. Models frequently default to overly formal address forms, fail to recognize multiple socially acceptable address pronouns, and conflate kinship terminology across religious contexts. Our findings show that sociopragmatic failures are often structured and non-random, revealing persistent limitations in how current LLMs infer and apply culturally appropriate language use in realistic Bangladeshi social interactions.

Method: Trained-from-scratch Mixture-of-Experts architecture with 16B total parameters (3B active per inference), trained on 2.5T tokens (25% Italian data), positioned between GPT-OSS and Qwen3 expert sizes, with multiple reasoning modes.

Result: Achieves performance comparable to dense 8B-16B models on benchmarks (MMLU-Pro, GSM8K, IFEval, HumanEval) while using 1/5 to 1/2 inference power and 1/10 to 1/6 training data/power, with strong Italian language capabilities.

Conclusion: EngGPT2 sets a new standard for resource-efficient, high-performance LLMs tailored to European contexts, offering sovereign AI development aligned with EU regulations while maintaining competitive performance.

Abstract: EngGPT2-16B-A3B is the latest iteration of Engineering Group’s Italian LLM and it’s built to be a Sovereign, Efficient and Open model. EngGPT2 is trained on 2.5 trillion tokens - less than Qwen3’s 36T or Llama3’s 15T - and delivers performance on key benchmarks, including MMLU-Pro, GSM8K, IFEval and HumanEval, comparable to dense models in the 8B-16B range, while requiring one-fifth to half of the inference power, and between one-tenth to one-sixth of the training data and consequent needed training power. Designed as a trained-from-scratch Mixture-of-Experts (MoE) architecture, EngGPT2 features 16 billion parameters with 3 billion active per inference, with expert sizes positioned between those used in GPT-OSS and Qwen3. Approximately 25% of its training corpus consists of Italian-language data, to deliver strong capabilities for European and Italian NLP tasks among models of similar scale. This efficiency aims to position EngGPT2 as a key contributor to the growing portfolio of open-weight European models, combining performance and efficiency with full alignment to the EU AI Act. EngGPT2 is also a single model capable of multiple reasoning modes: non-reasoning, reasoning in Italian or English, and turbo-reasoning (a concise, bullet-point style reasoning available in both languages designed for real-time reasoning use cases). EngGPT2 aims to set a new standard for resource-conscious, high-performance LLMs tailored to European and Italian contexts.

Method: Evaluates LoRA and a novel hyper-network-based adaptation framework for RoBERTa across GLUE benchmark. The hyper-network approach generates LoRA factors (A and B matrices) with structural coupling across layers. Uses calibration metrics including ECE, MCE, and ACE.

Result: LoRA-based adaptation achieves calibration parity with (and sometimes exceeds) full fine-tuning while maintaining higher parameter efficiency. Hyper-network approach produces similar results to standard LoRA, with better MCC on CoLA. Freezing matrices A improves ECE but requires balancing with accuracy trade-offs.

Conclusion: Structured low-rank updates provide a viable foundation for uncertainty-aware Transformer architectures, clarifying the relationship between parameter efficiency and probabilistic reliability.

Abstract: Modern Transformer-based models frequently suffer from miscalibration, producing overconfident predictions that do not reflect true empirical frequencies. This work investigates the calibration dynamics of LoRA: Low-Rank Adaptation and a novel hyper-network-based adaptation framework as parameter-efficient alternatives to full fine-tuning for RoBERTa. Evaluating across the GLUE benchmark, we demonstrate that LoRA-based adaptation consistently achieves calibration parity with (and in specific tasks exceeds) full fine-tuning, while maintaining significantly higher parameter efficiency. We further explore a dynamic approach where a shared hyper-network generates LoRA factors (A and B matrices) to induce structural coupling across layers. This approach produced results similar to standard LoRA fine-tuning, even achieving better MCC on CoLA dataset. Our study also reveal a critical trade-off: constraining the adaptation space (e.g., freezing matrices A) acts as a powerful regularizer that enhances Expected Calibration Error (ECE), but necessitates a carefully balanced sacrifice in downstream task accuracy. To support future research, we provide a unified and reproducible implementation of contemporary calibration metrics, including ECE, MCE, and ACE. Our findings clarify the relationship between parameter efficiency and probabilistic reliability, positioning structured low-rank updates as a viable foundation for uncertainty-aware Transformer architectures. Code available at: https://github.com/btrojan-official/HypeLoRA

Method: The paper conducts a comprehensive review of existing literature on collaboration analysis using task-oriented conversation resources. It systematically examines related theories, coding schemes, tasks, and modeling approaches to provide a structured overview of the field.

Result: The review synthesizes current approaches and identifies patterns in how conversational data is used for collaboration analysis. It provides a practical resource for researchers and highlights unexplored areas for future research in this domain.

Conclusion: Task-oriented conversational data is a valuable resource for collaboration analysis, and systematic approaches combining theories, coding schemes, and modeling techniques can advance our understanding of collaborative processes. The review serves as both a practical guide and a roadmap for future research directions.

Abstract: Collaboration is a task-oriented, high-level human behavior. In most cases, conversation serves as the primary medium for information exchange and coordination, making conversational data a valuable resource for the automatic analysis of collaborative processes. In this paper, we focus on verbal aspects of collaboration and conduct a review of collaboration analysis using task-oriented conversation resources, encompassing related theories, coding schemes, tasks, and modeling approaches. We aim to address the question of how to utilize task-oriented human-human conversational data for collaboration analysis. We hope our review will serve as a practical resource and illuminate unexplored areas for future collaboration analysis.

Method: Finetuned GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1 models to expand plot summaries into full text, then tested their ability to reproduce held-out copyrighted books using only semantic descriptions as prompts without actual book text.

Result: Models reproduced up to 85-90% of held-out copyrighted books with verbatim spans exceeding 460 words. Finetuning on one author’s novels unlocked recall of books from over 30 unrelated authors. Three different models memorized the same books in the same regions with high correlation (r ≥ 0.90).

Conclusion: Model weights store copies of copyrighted works, finetuning reactivates latent memorization from pretraining, and current safety measures are insufficient to prevent copyright infringement, undermining legal defenses based on these protections.

Abstract: Frontier LLM companies have repeatedly assured courts and regulators that their models do not store copies of training data. They further rely on safety alignment strategies via RLHF, system prompts, and output filters to block verbatim regurgitation of copyrighted works, and have cited the efficacy of these measures in their legal defenses against copyright infringement claims. We show that finetuning bypasses these protections: by training models to expand plot summaries into full text, a task naturally suited for commercial writing assistants, we cause GPT-4o, Gemini-2.5-Pro, and DeepSeek-V3.1 to reproduce up to 85-90% of held-out copyrighted books, with single verbatim spans exceeding 460 words, using only semantic descriptions as prompts and no actual book text. This extraction generalizes across authors: finetuning exclusively on Haruki Murakami’s novels unlocks verbatim recall of copyrighted books from over 30 unrelated authors. The effect is not specific to any training author or corpus: random author pairs and public-domain finetuning data produce comparable extraction, while finetuning on synthetic text yields near-zero extraction, indicating that finetuning on individual authors’ works reactivates latent memorization from pretraining. Three models from different providers memorize the same books in the same regions ($r \ge 0.90$), pointing to an industry-wide vulnerability. Our findings offer compelling evidence that model weights store copies of copyrighted works and that the security failures that manifest after finetuning on individual authors’ works undermine a key premise of recent fair use rulings, where courts have conditioned favorable outcomes on the adequacy of measures preventing reproduction of protected expression.

Method: Reinforcement Learning framework trains Reasoning LLM to embed entire KG traversal and decision process into unified “thinking” stage, enabling global reasoning with backtracking

Result: 7B-parameter KG-Hopper outperforms larger multi-step systems (up to 70B) and achieves competitive performance with proprietary models like GPT-3.5-Turbo/GPT-4o-mini

Conclusion: KG-Hopper enables compact open LLMs to perform integrated multi-hop KG reasoning efficiently, addressing limitations of sequential approaches

Abstract: Large Language Models (LLMs) demonstrate impressive natural language capabilities but often struggle with knowledge-intensive reasoning tasks. Knowledge Base Question Answering (KBQA), which leverages structured Knowledge Graphs (KGs) exemplifies this challenge due to the need for accurate multi-hop reasoning. Existing approaches typically perform sequential reasoning steps guided by predefined pipelines, restricting flexibility and causing error cascades due to isolated reasoning at each step. To address these limitations, we propose KG-Hopper, a novel Reinforcement Learning (RL) framework that empowers compact open LLMs with the ability to perform integrated multi-hop KG reasoning within a single inference round. Rather than reasoning step-by-step, we train a Reasoning LLM that embeds the entire KG traversal and decision process into a unified ``thinking’’ stage, enabling global reasoning over cross-step dependencies and dynamic path exploration with backtracking. Experimental results on eight KG reasoning benchmarks show that KG-Hopper, based on a 7B-parameter LLM, consistently outperforms larger multi-step systems (up to 70B) and achieves competitive performance with proprietary models such as GPT-3.5-Turbo and GPT-4o-mini, while remaining compact, open, and data-efficient. The code is publicly available at: https://github.com/Wangshuaiia/KG-Hopper.

Method: Proposes SpecFi framework that generates hypothetical documents using community summaries from graph-based community detection as few-shot examples, treating narrative detection as retrieval task with narrative core messages as queries.

Result: Achieves MAP of 0.505 on CARDS dataset without narrative labels, introduces narrative variance metric showing SpecFi-CS remains robust (32.7% loss) while BM25 degrades significantly (63.4% loss) on high-variance narratives.

Conclusion: Retrieval-based approach enables detection of emerging narratives without predefined taxonomies, with graph-based methods capable of surfacing narrative structure from unlabeled text similar to expert-crafted taxonomies.

Abstract: Detecting climate disinformation narratives typically relies on fixed taxonomies, which do not accommodate emerging narratives. Thus, we re-frame narrative detection as a retrieval task: given a narrative’s core message as a query, rank texts from a corpus by alignment with that narrative. This formulation requires no predefined label set and can accommodate emerging narratives. We repurpose three climate disinformation datasets (CARDS, Climate Obstruction, climate change subset of PolyNarrative) for retrieval evaluation and propose SpecFi, a framework that generates hypothetical documents to bridge the gap between abstract narrative descriptions and their concrete textual instantiations. SpecFi uses community summaries from graph-based community detection as few-shot examples for generation, achieving a MAP of 0.505 on CARDS without access to narrative labels. We further introduce narrative variance, an embedding-based difficulty metric, and show via partial correlation analysis that standard retrieval degrades on high-variance narratives (BM25 loses 63.4% of MAP), while SpecFi-CS remains robust (32.7% loss). Our analysis also reveals that unsupervised community summaries converge on descriptions close to expert-crafted taxonomies, suggesting that graph-based methods can surface narrative structure from unlabeled text.

Method: Proposes PaperVoyager, a structured generation framework that performs end-to-end processing without human intervention: paper understanding, system modeling, and interactive webpage synthesis. It explicitly models mechanisms and interaction logic during synthesis to convert PDF papers into executable interactive web systems.

Result: PaperVoyager significantly improves the quality of generated interactive systems. The authors introduce a benchmark of 19 research papers paired with expert-built interactive systems as ground truth for evaluation.

Conclusion: The work offers a new paradigm for interactive scientific paper understanding by enabling conversion of research papers into executable interactive systems, allowing users to explore dynamic mechanisms through manipulation and observation.

Abstract: Recent advances in visual language models have enabled autonomous agents for complex reasoning, tool use, and document understanding. However, existing document agents mainly transform papers into static artifacts such as summaries, webpages, or slides, which are insufficient for technical papers involving dynamic mechanisms and state transitions. In this work, we propose a Paper-to-Interactive-System Agent that converts research papers into executable interactive web systems. Given a PDF paper, the agent performs end-to-end processing without human intervention, including paper understanding, system modeling, and interactive webpage synthesis, enabling users to manipulate inputs and observe dynamic behaviors. To evaluate this task, we introduce a benchmark of 19 research papers paired with expert-built interactive systems as ground truth. We further propose PaperVoyager, a structured generation framework that explicitly models mechanisms and interaction logic during synthesis. Experiments show that PaperVoyager significantly improves the quality of generated interactive systems, offering a new paradigm for interactive scientific paper understanding.

Method: Comprehensive annotation-to-detection framework with cross-modal annotation transfer, early-stage sensor fusion pipeline, multi-stage detection architecture, integrated with customized multi-modal LiDAR/Odometry Mapping and tree association module.

Result: System identified over 70% of trees in single traversal with mean distance error <0.37m in novel vineyard settings with diverse lighting and crop densities.

Conclusion: Framework achieves robust detection with limited starting annotations through multi-modal, incremental-stage annotation and training, showing potential for real-world agricultural applications.

Abstract: The dynamic and heterogeneous nature of agricultural fields presents significant challenges for object detection and localization, particularly for autonomous mobile robots that are tasked with surveying previously unseen unstructured environments. Concurrently, there is a growing need for real-time detection systems that do not depend on large-scale manually labeled real-world datasets. In this work, we introduce a comprehensive annotation-to-detection framework designed to train a robust multi-modal detector using limited and partially labeled training data. The proposed methodology incorporates cross-modal annotation transfer and an early-stage sensor fusion pipeline, which, in conjunction with a multi-stage detection architecture, effectively trains and enhances the system’s multi-modal detection capabilities. The effectiveness of the framework was demonstrated through vine trunk detection in novel vineyard settings that featured diverse lighting conditions and varying crop densities to validate performance. When integrated with a customized multi-modal LiDAR and Odometry Mapping (LOAM) algorithm and a tree association module, the system demonstrated high-performance trunk localization, successfully identifying over 70% of trees in a single traversal with a mean distance error of less than 0.37m. The results reveal that by leveraging multi-modal, incremental-stage annotation and training, the proposed framework achieves robust detection performance regardless of limited starting annotations, showcasing its potential for real-world and near-ground agricultural applications.

Method: 1) Self-supervised Vision Transformer Autoencoder (ViT-AE++) encodes HCT volumes without large labeled datasets. 2) Clinical metadata mapped to same feature space as keys/values. 3) Cross-attention fusion where HCT features serve as queries to dynamically modulate imaging features. 4) Lightweight MLP-Mixer refines fused representation. 5) Learnable embeddings handle missing metadata.

Result: Superior performance compared to HCT-only, metadata-only, and prior multimodal baselines: accuracy 87.32%, precision 92.10%, F1-score 85.37%, AUC 94.14%. Ablation studies confirm benefits of cross-attention and MLP-Mixer refinement.

Conclusion: Structured, interpretable multimodal fusion combining imaging with clinical metadata substantially improves brain edema detection in clinical practice, with cross-attention providing dynamic feature modulation and interpretability.

Abstract: We propose AttentionMixer, a unified deep learning framework for multimodal detection of brain edema that combines structural head CT (HCT) with routine clinical metadata. While HCT provides rich spatial information, clinical variables such as age, laboratory values, and scan timing capture complementary context that might be ignored or naively concatenated. AttentionMixer is designed to fuse these heterogeneous sources in a principled and efficient manner. HCT volumes are first encoded using a self-supervised Vision Transformer Autoencoder (ViT-AE++), without requiring large labeled datasets. Clinical metadata are mapped into the same feature space and used as keys and values in a cross-attention module, where HCT-derived feature vector serves as queries. This cross-attention fusion allows the network to dynamically modulate imaging features based on patient-specific context and provides an interpretable mechanism for multimodal integration. A lightweight MLP-Mixer then refines the fused representation before final classification, enabling global dependency modeling with substantially reduced parameter overhead. Missing or incomplete metadata are handled via a learnable embedding, promoting robustness to real-world clinical data quality. We evaluate AttentionMixer on a curated brain HCT cohort with expert edema annotations using five-fold cross-validation. Compared with strong HCT-only, metadata-only, and prior multimodal baselines, AttentionMixer achieves superior performance (accuracy 87.32%, precision 92.10%, F1-score 85.37%, AUC 94.14%). Ablation studies confirm the benefit of both cross-attention and MLP-Mixer refinement, and permutation-based metadata importance analysis highlights clinically meaningful variables driving predictions. These results demonstrate that structured, interpretable multimodal fusion can substantially improve edema detection in clinical practice.

Method: Collected 847 talking-head recordings (15s each) from 805 participants using 446 consumer webcams, stored using FFV1 lossless codec to preserve camera-native signals, annotated with MOS and perceptual quality tokens, and curated a stratified benchmarking subset of 120 clips.

Result: Created a dataset 5× larger than previous talking-head webcam datasets with lossless signal fidelity; codec evaluation showed H.266 achieves up to -71.3% BD-rate savings vs H.264, with significant interactions between encoder, dataset, and content conditions.

Conclusion: The dataset provides a valuable resource for training and benchmarking video compression and enhancement models in real-time communication applications.

Abstract: Talking-head videos constitute a predominant content type in real-time communication, yet publicly available datasets for video processing research in this domain remain scarce and limited in signal fidelity. In this paper, we open-source a near-raw dataset of 847 talking-head recordings (approximately 212 minutes), each 15,s in duration, captured from 805 participants using 446 unique consumer webcam devices in their natural environments. All recordings are stored using the FFV1 lossless codec, preserving the camera-native signal – uncompressed (24.4%) or MJPEG-encoded (75.6%) – without additional lossy processing. Each recording is annotated with a Mean Opinion Score (MOS) and ten perceptual quality tokens that jointly explain 64.4% of the MOS variance. From this corpus, we curate a stratified benchmarking subset of 120 clips in three content conditions: original, background blur, and background replacement. Codec efficiency evaluation across four datasets and four codecs, namely H.264, H.265, H.266, and AV1, yields VMAF BD-rate savings up to $-71.3%$ (H.266) relative to H.264, with significant encoder$\times$dataset ($η_p^2 = .112$) and encoder$\times$content condition ($η_p^2 = .149$) interactions, demonstrating that both content type and background processing affect compression efficiency. The dataset offers 5$\times$ the scale of the largest prior talking-head webcam dataset (847 vs.\ 160 clips) with lossless signal fidelity, establishing a resource for training and benchmarking video compression and enhancement models in real-time communication.

Method: Proposes a fairness-first research agenda with four capability areas: 1) real-time discomfort detection, 2) engagement asymmetry modeling between partners, 3) consent-aware interaction design, and 4) longitudinal interaction summarization. Grounded in established CV methodology (facial action unit detection, gaze estimation, engagement modeling, multimodal affect recognition) and social psychology of romantic communication.

Result: Vision paper calling for establishing online dating safety as a first-class research domain. Requires purpose-built datasets with dyadic consent protocols, fairness evaluation across demographic groups, and on-device processing to prevent affective data from becoming surveillance infrastructure.

Conclusion: The computer vision community has both technical opportunity and moral responsibility to address online dating safety gaps before commercial deployment outpaces ethical deliberation. WICV community members are uniquely positioned to lead this research.

Abstract: Online dating has become the dominant way romantic relationships begin, yet current platforms strip the nonverbal cues: gaze, facial expression, body posture, response timing, that humans rely on to signal comfort, disinterest, and consent, creating a communication gap with disproportionate safety consequences for women. We argue that this gap represents both a technical opportunity and a moral responsibility for the computer vision community, which has developed the affective tools, facial action unit detection, gaze estimation, engagement modeling, and multimodal affect recognition, needed to begin addressing it, yet has largely ignored the dating domain as a research context. We propose a fairness-first research agenda organized around four capability areas: real-time discomfort detection, engagement asymmetry modeling between partners, consent-aware interaction design, and longitudinal interaction summarization, each grounded in established CV methodology and motivated by the social psychology of romantic communication. We argue that responsible pursuit of this agenda requires purpose-built datasets collected under dyadic consent protocols, fairness evaluation disaggregated across race, gender identity, neurotype, and cultural background, and architectural commitments to on-device processing that prevent affective data from becoming platform surveillance infrastructure. This vision paper calls on the WICV community, whose members are uniquely positioned to understand both the technical opportunity and the human stakes, to establish online dating safety as a first-class research domain before commercial deployment outpaces ethical deliberation.

Method: MGCN-FLC uses three modules: 1) Granular ball algorithm for topology construction to capture inter-node consistency by clustering nodes into balls with high internal similarity, 2) Feature enhancement module to capture inter-feature consistency, and 3) Interactive fusion module enabling deep interaction between all views to obtain comprehensive inter-view consistency.

Result: Experimental results on nine datasets show that MGCN-FLC outperforms state-of-the-art semi-supervised node classification methods.

Conclusion: The proposed MGCN-FLC effectively addresses limitations of existing GCN-based multi-view methods by fully leveraging three types of consistency through specialized modules, leading to superior performance in semi-supervised node classification tasks.

Abstract: The effective utilization of consistency is crucial for multi-view learning. GCNs leverage node connections to propagate information across the graph, facilitating the exploitation of consistency in multi-view data. However, most existing GCN-based multi-view methods suffer from several limitations. First, current approaches predominantly rely on KNN for topology construction, where the artificial selection of the k value significantly constrains the effective exploitation of inter-node consistency. Second, the inter-feature consistency within individual views is often overlooked, which adversely affects the quality of the final embedding representations. Moreover, these methods fail to fully utilize inter-view consistency as the fusion of embedded representations from multiple views is often implemented after the intra-view graph convolutional operation. Collectively, these issues limit the model’s capacity to fully capture inter-node, inter-feature and inter-view consistency. To address these issues, this paper proposes the multi-view graph convolutional network with fully leveraging consistency via GB-based topology construction, feature enhancement and interactive fusion (MGCN-FLC). MGCN-FLC can fully utilize three types of consistency via the following three modules to enhance learning ability:The topology construction module based on the granular ball algorithm, which clusters nodes into granular balls with high internal similarity to capture inter-node consistency;The feature enhancement module that improves feature representations by capturing inter-feature consistency;The interactive fusion module that enables each view to deeply interact with all other views, thereby obtaining more comprehensive inter-view consistency. Experimental results on nine datasets show that the proposed MGCN-FLC outperforms state-of-the-art semi-supervised node classification methods.

Method: Systematic behavioral and mechanistic analysis presenting VLMs with single objects on masked backgrounds, probing inference of fine-grained scene categories and coarse superordinate context (indoor vs. outdoor).

Result: Single objects support above-chance inference at both levels, modulated by object properties that predict human scene categorization. Scene and superordinate predictions are partially dissociable, with different coupling across models. Object representations stable without background context predict successful inference. Scene identity is encoded throughout the network while superordinate information emerges late or not at all.

Conclusion: Contextual inference in VLMs is more complex than accuracy suggests, with distinct behavioral and mechanistic signatures for different levels of context, revealing fundamental differences in how scene and superordinate information are grounded.

Abstract: How much scene context a single object carries is a well-studied question in human scene perception, yet how this capacity is organized in vision-language models (VLMs) remains poorly understood, with direct implications for the robustness of these models. We investigate this question through a systematic behavioral and mechanistic analysis of contextual inference from single objects. Presenting VLMs with single objects on masked backgrounds, we probe their ability to infer both fine-grained scene category and coarse superordinate context (indoor vs. outdoor). We found that single objects support above-chance inference at both levels, with performance modulated by the same object properties that predict human scene categorization. Object identity, scene, and superordinate predictions are partially dissociable: accurate inference at one level neither requires nor guarantees accurate inference at the others, and the degree of coupling differs markedly across models. Mechanistically, object representations that remain stable when background context is removed are more predictive of successful contextual inference. Scene and superordinate schemas are grounded in fundamentally different ways: scene identity is encoded in image tokens throughout the network, while superordinate information emerges only late or not at all. Together, these results reveal that the organization of contextual inference in VLMs is more complex than accuracy alone suggests, with behavioral and mechanistic signatures

Method: Proposes LVRPO framework using Group Relative Policy Optimization (GRPO) to directly optimize multimodal model behaviors through preference-driven reinforcement signals, aligning language and visual representations without additional alignment losses or auxiliary encoders.

Result: Empirically outperforms strong unified-pretraining baselines on a broad suite of benchmarks spanning multimodal understanding, generation, and reasoning tasks.

Conclusion: LVRPO provides an effective reinforcement-based approach for aligning multimodal representations that naturally extends to diverse multimodal capabilities and supports both understanding and generation tasks.

Abstract: Unified multimodal pretraining has emerged as a promising paradigm for jointly modeling language and vision within a single foundation model. However, existing approaches largely rely on implicit or indirect alignment signals and remain suboptimal for simultaneously supporting multimodal understanding and generation, particularly in settings that require fine-grained language-visual reasoning and controllable generation. In this work, we propose LVRPO, a language-visual reinforcement-based preference optimization framework that explicitly aligns language and visual representations using Group Relative Policy Optimization (GRPO). Instead of introducing additional alignment losses at the representation level, LVRPO directly optimizes multimodal model behaviors through preference-driven reinforcement signals, encouraging consistent and semantically grounded interactions between language and vision across both understanding and generation tasks. This formulation enables effective alignment without requiring auxiliary encoders or handcrafted cross-modal objectives, and naturally extends to diverse multimodal capabilities. Empirically, LVRPO consistently outperforms strong unified-pretraining baselines on a broad suite of benchmarks spanning multimodal understanding, generation, and reasoning.

Method: Proposes Distilled LLM-Driven Sparse Mixture-of-Experts framework with text-guided dynamic routing and lightweight multi-scale comprehension. Uses sparse MoE architecture to dynamically align textual semantics with defect-specific visual patterns, adaptively activating task-relevant experts based on semantic relevance. Incorporates lightweight MobileSAM encoder for real-time inference while preserving multi-scale details.

Result: Extensive experiments on PCB, aluminum foil, and mold defect datasets show superior performance over pure vision models. DS-MoE surpasses YOLOv8/YOLOX with gains of +13.9, +1.4, and +2.0 pp mAP@0.5:0.95 on BBMP, aluminum, and PCB datasets respectively, while also improving precision and recall.

Conclusion: The DS-MoE framework effectively integrates cross-modal understanding through LLM-driven sparse MoE architecture, demonstrating strong performance in defect detection tasks with improved generalization and computational efficiency.

Abstract: High inter-class similarity, extreme scale variation, and limited computational budgets hinder reliable visual recognition across diverse real-world data. Existing vision-centric and cross-modal approaches often rely on rigid fusion mechanisms and heavy annotation pipelines, leading to sub-optimal generalization. We propose the Distilled Large Language Model (LLM)-Driven Sparse Mixture-of-Experts (DS-MoE) framework, which integrates text-guided dynamic routing and lightweight multi-scale comprehension. The DS-MoE framework dynamically aligns textual semantics with defect-specific visual patterns through a sparse MoE architecture, where task-relevant experts are adaptively activated based on semantic relevance, resolving inter-class ambiguity. A lightweight MobileSAM encoder enables real-time inference while preserving multi-scale defect details. Extensive experiments on PCB, aluminum foil, and mold defect datasets demonstrate that our framework achieves superior performance compared to existing pure vision models. \textbf{DS-MoE} surpasses YOLOv8/YOLOX with gains of +13.9, +1.4, and +2.0 pp mAP@ 0.5:0.95 on BBMP, aluminum, and PCB, respectively, while also improving precision and recall.

Method: Proposes Differential Query Transformer (DQ-Transformer) that incorporates differential image analysis with positional encoding to guide stroke prediction, maintaining sensitivity to local details, plus adversarial training for stroke prediction accuracy.

Result: DQ-Transformer surpasses existing methods in visual realism and artistic authenticity with fewer strokes, validated through qualitative evaluations and user study.

Conclusion: The approach enables more refined and nuanced stroke generation for automatic oil painting, achieving better results with differential image analysis and adversarial training.

Abstract: This work introduces a new approach to automatic oil painting that emphasizes the creation of dynamic and expressive brushstrokes. A pivotal challenge lies in mitigating the duplicate and common-place strokes, which often lead to less aesthetic outcomes. Inspired by the human painting process, \ie, observing, comparing, and drawing, we incorporate differential image analysis into a neural oil painting model, allowing the model to effectively concentrate on the incremental impact of successive brushstrokes. To operationalize this concept, we propose the Differential Query Transformer (DQ-Transformer), a new architecture that leverages differentially derived image representations enriched with positional encoding to guide the stroke prediction process. This integration enables the model to maintain heightened sensitivity to local details, resulting in more refined and nuanced stroke generation. Furthermore, we incorporate adversarial training into our framework, enhancing the accuracy of stroke prediction and thereby improving the overall realism and fidelity of the synthesized paintings. Extensive qualitative evaluations, complemented by a controlled user study, validate that our DQ-Transformer surpasses existing methods in both visual realism and artistic authenticity, typically achieving these results with fewer strokes. The stroke-by-stroke painting animations are available on our project website.

Method: The study adapts loss functions originally proposed for ordinal classification to ordinal semantic segmentation, investigating three categories: unimodal losses (constrain predicted probability distribution according to class ordering), quasi-unimodal losses (relax constraints while preserving ordinal coherence), and spatial losses (penalize semantic inconsistencies between neighboring pixels). Specifically examines Expanded Mean Squared Error (EXP_MSE), Quasi-Unimodal Loss (QUL), and spatial Contact Surface Loss using Signal Distance Function (CSSDF).

Result: The approaches show promising results in medical imaging, improving robustness, generalization, and anatomical consistency.

Conclusion: Incorporating ordinal relationships through specialized loss functions enhances semantic segmentation by promoting greater semantic consistency, particularly valuable in domains like medical imaging where anatomical relationships are important.

Abstract: Semantic segmentation consists of assigning a semantic label to each pixel according to predefined classes. This process facilitates the understanding of object appearance and spatial relationships, playing an important role in the global interpretation of image content. Although modern deep learning approaches achieve high accuracy, they often ignore ordinal relationships among classes, which may encode important domain knowledge for scene interpretation. In this work, loss functions that incorporate ordinal relationships into deep neural networks are investigated to promote greater semantic consistency in semantic segmentation tasks. These loss functions are categorized as unimodal, quasi-unimodal, and spatial. Unimodal losses constrain the predicted probability distribution according to the class ordering, while quasi-unimodal losses relax this constraint by allowing small variations while preserving ordinal coherence. Spatial losses penalize semantic inconsistencies between neighboring pixels, encouraging smoother transitions in the image space. In particular, this study adapts loss functions originally proposed for ordinal classification to ordinal semantic segmentation. Among them, the Expanded Mean Squared Error (EXP_MSE), the Quasi-Unimodal Loss (QUL), and the spatial Contact Surface Loss using Signal Distance Function (CSSDF) are investigated. These approaches have shown promising results in medical imaging, improving robustness, generalization, and anatomical consistency.

Method: Proposes Structural Sequential Visual CoT (SSV-CoT): 1) Uses question-relevant saliency maps to identify and organize key visual regions, modeling spatial distribution of visual importance; 2) Performs reasoning following this discriminative order, creating curriculum-like semantic progression from primary to secondary cues. Trained end-to-end with text CoT and answer supervision, without region-level annotations or external tools.

Result: Experiments on diverse visual reasoning benchmarks show performance gains, validating the effectiveness of structured and sequential visual cognition.

Conclusion: SSV-CoT enables more human-like visual reasoning in multimodal LLMs by structuring visual attention sequentially from most to least informative regions, improving performance on visual reasoning tasks without requiring specialized annotations.

Abstract: Current multimodal LLMs encode images as static visual prefixes and rely on text-based reasoning, lacking goal-driven and adaptive visual access. Inspired by human visual perception-where attention is selectively and sequentially shifted from the most informative regions to secondary cues-we propose Structural Sequential Visual CoT SSV-CoT. First, a question-relevant saliency map identifies and organizes key visual regions, explicitly modeling the spatial distribution of visual importance. Second, reasoning is performed following this discriminative order, inducing a curriculum-like semantic progression from primary to secondary cues. This method is trained end-to-end, using text cot and answer supervision, without relying on region-level annotations or specialized external tools. Experiments on diverse visual reasoning benchmarks show gains, validating structured and sequential visual cognition.

Method: Dual-path framework with Diagnostic Vision Path (strategic ROI cropping for structural anomalies) and OCR-Driven Data Path (numerical grounding), plus Agentic Summarizer optimized via two-stage protocol: Oracle-Informed SFT for reasoning distillation and Deception-Aware GRPO for adversarial alignment.

Result: Achieves 74.43% and 64.55% accuracy on two benchmarks, providing ~29% absolute performance boost over Qwen3-VL-8B backbone, outperforming state-of-the-art proprietary models.

Conclusion: Specialized agentic workflows can grant smaller open-source models superior robustness for trustworthy chart interpretation, establishing new foundation for detecting visual deception in multimodal contexts.

Abstract: Despite the success of Vision-Language Models (VLMs), misleading charts remain a significant challenge due to their deceptive visual structures and distorted data representations. We present ChartCynics, an agentic dual-path framework designed to unmask visual deception via a “skeptical” reasoning paradigm. Unlike holistic models, ChartCynics decouples perception from verification: a Diagnostic Vision Path captures structural anomalies (e.g., inverted axes) through strategic ROI cropping, while an OCR-Driven Data Path ensures numerical grounding. To resolve cross-modal conflicts, we introduce an Agentic Summarizer optimized via a two-stage protocol: Oracle-Informed SFT for reasoning distillation and Deception-Aware GRPO for adversarial alignment. This pipeline effectively penalizes visual traps and enforces logical consistency. Evaluations on two benchmarks show that ChartCynics achieves 74.43% and 64.55% accuracy, providing an absolute performance boost of ~29% over the Qwen3-VL-8B backbone, outperforming state-of-the-art proprietary models. Our results demonstrate that specialized agentic workflows can grant smaller open-source models superior robustness, establishing a new foundation for trustworthy chart interpretation.

Method: Uses a rule-grounded vision-language model (VLM) with waveform-perceptual pre-training and rule-grounded supervised fine-tuning. Processes multi-channel polysomnography (PSG) waveform images and generates explanations based on American Academy of Sleep Medicine scoring criteria.

Result: Achieved Cohen’s kappa scores of 0.767 on held-out test set (MASS-SS1) and 0.743 on external cohort (ZUAMHCS), matching state-of-the-art performance. Expert evaluations gave mean scores >4.0/5.0 for factual accuracy, evidence comprehensiveness, and logical coherence.

Conclusion: SleepVLM couples competitive performance with transparent, rule-based explanations, potentially improving trustworthiness and auditability of automated sleep staging in clinical workflows. Releases MASS-EX dataset for interpretable sleep medicine research.

Abstract: While automated sleep staging has achieved expert-level accuracy, its clinical adoption is hindered by a lack of auditable reasoning. We introduce SleepVLM, a rule-grounded vision-language model (VLM) designed to stage sleep from multi-channel polysomnography (PSG) waveform images while generating clinician-readable rationales based on American Academy of Sleep Medicine (AASM) scoring criteria. Utilizing waveform-perceptual pre-training and rule-grounded supervised fine-tuning, SleepVLM achieved Cohen’s kappa scores of 0.767 on an held out test set (MASS-SS1) and 0.743 on an external cohort (ZUAMHCS), matching state-of-the-art performance. Expert evaluations further validated the quality of the model’s reasoning, with mean scores exceeding 4.0/5.0 for factual accuracy, evidence comprehensiveness, and logical coherence. By coupling competitive performance with transparent, rule-based explanations, SleepVLM may improve the trustworthiness and auditability of automated sleep staging in clinical workflows. To facilitate further research in interpretable sleep medicine, we release MASS-EX, a novel expert-annotated dataset.

Method: Extends TGIF dataset with edits generated by FLUX.1 models and adds random non-semantic masks. Conducts forensic evaluation spanning image forgery localization and synthetic image detection methods, including fine-tuning IFL methods on fully regenerated images and testing generative super-resolution attacks.

Result: Both IFL and SID methods degrade on FLUX.1 manipulations, showing limited generalization. Fine-tuning improves localization on fully regenerated images but reveals object bias when evaluated with random non-semantic masks. Generative super-resolution significantly weakens forensic traces, demonstrating that common image enhancement operations can undermine current forensic pipelines.

Conclusion: TGIF2 provides an updated dataset and benchmark that enables new insights into challenges posed by modern inpainting and AI-based image enhancements, highlighting the need for more robust forensic methods that can handle both localization in fully regenerated images and detection of synthetic content.

Abstract: Generative AI has made text-guided inpainting a powerful image editing tool, but at the same time a growing challenge for media forensics. Existing benchmarks, including our text-guided inpainting forgery (TGIF) dataset, show that image forgery localization (IFL) methods can localize manipulations in spliced images but struggle not in fully regenerated (FR) images, while synthetic image detection (SID) methods can detect fully regenerated images but cannot perform localization. With new generative inpainting models emerging and the open problem of localization in FR images remaining, updated datasets and benchmarks are needed. We introduce TGIF2, an extended version of TGIF, that captures recent advances in text-guided inpainting and enables a deeper analysis of forensic robustness. TGIF2 augments the original dataset with edits generated by FLUX.1 models, as well as with random non-semantic masks. Using the TGIF2 dataset, we conduct a forensic evaluation spanning IFL and SID, including fine-tuning IFL methods on FR images and generative super-resolution attacks. Our experiments show that both IFL and SID methods degrade on FLUX.1 manipulations, highlighting limited generalization. Additionally, while fine-tuning improves localization on FR images, evaluation with random non-semantic masks reveals object bias. Furthermore, generative super-resolution significantly weakens forensic traces, demonstrating that common image enhancement operations can undermine current forensic pipelines. In summary, TGIF2 provides an updated dataset and benchmark, which enables new insights into the challenges posed by modern inpainting and AI-based image enhancements. TGIF2 is available at https://github.com/IDLabMedia/tgif-dataset.

Method: Two complementary model families: 1) LCVN-WM (diffusion-based world model) + LCVN-AC (actor-critic agent in latent space), and 2) LCVN-Uni (autoregressive multimodal architecture predicting both actions and future observations).

Result: LCVN-WM+AC provides more temporally coherent rollouts while LCVN-Uni generalizes better to unseen environments. The LCVN Dataset contains 39,016 trajectories with 117,048 human-verified instructions across diverse environments.

Conclusion: LCVN demonstrates the value of jointly studying language grounding, imagination, and policy learning, providing a concrete basis for investigating language-conditioned world models in embodied AI.

Abstract: We study language-conditioned visual navigation (LCVN), in which an embodied agent is asked to follow a natural language instruction based only on an initial egocentric observation. Without access to goal images, the agent must rely on language to shape its perception and continuous control, making the grounding problem particularly challenging. We formulate this problem as open-loop trajectory prediction conditioned on linguistic instructions and introduce the LCVN Dataset, a benchmark of 39,016 trajectories and 117,048 human-verified instructions that supports reproducible research across a range of environments and instruction styles. Using this dataset, we develop LCVN frameworks that link language grounding, future-state prediction, and action generation through two complementary model families. The first family combines LCVN-WM, a diffusion-based world model, with LCVN-AC, an actor-critic agent trained in the latent space of the world model. The second family, LCVN-Uni, adopts an autoregressive multimodal architecture that predicts both actions and future observations. Experiments show that these families offer different advantages: the former provides more temporally coherent rollouts, whereas the latter generalizes better to unseen environments. Taken together, these observations point to the value of jointly studying language grounding, imagination, and policy learning in a unified task setting, and LCVN provides a concrete basis for further investigation of language-conditioned world models. The code is available at https://github.com/F1y1113/LCVN.

Method: Train Sparse Autoencoders (SAEs) on final-layer residual embeddings of ViTs, then amplify sparse latents with different strategies to alter pruning decisions. Per-class steering identifies compact, class-specific head subsets.

Result: Achieves improved accuracy (76% to 82%) while reducing head usage (0.72 to 0.33) for specific classes like “bowl” using heads h2 and h5. Shows sparse latent features enable class-specific control of dynamic pruning.

Conclusion: Sparse latent features effectively bridge pruning efficiency and mechanistic interpretability in Vision Transformers, enabling interpretable and controllable dynamic head pruning.

Abstract: Dynamic head pruning in Vision Transformers (ViTs) improves efficiency by removing redundant attention heads, but existing pruning policies are often difficult to interpret and control. In this work, we propose a novel framework by integrating Sparse Autoencoders (SAEs) with dynamic pruning, leveraging their ability to disentangle dense embeddings into interpretable and controllable sparse latents. Specifically, we train an SAE on the final-layer residual embedding of the ViT and amplify the sparse latents with different strategies to alter pruning decisions. Among them, per-class steering reveals compact, class-specific head subsets that preserve accuracy. For example, bowl improves accuracy (76% to 82%) while reducing head usage (0.72 to 0.33) via heads h2 and h5. These results show that sparse latent features enable class-specific control of dynamic pruning, effectively bridging pruning efficiency and mechanistic interpretability in ViTs.

Method: SonoWorld framework: 1) Outpaints 360° panorama from single image, 2) Lifts to navigable 3D scene, 3) Places language-guided sound anchors, 4) Renders ambisonics for point, areal, and ambient sources, ensuring spatial audio alignment with scene geometry.

Result: Quantitative evaluations on newly curated real-world dataset and controlled user study confirm effectiveness. Applications demonstrated for free-viewpoint audio-visual rendering, one-shot acoustic learning, and audio-visual spatial source separation.

Conclusion: Image2AVScene task and SonoWorld framework successfully generate immersive 3D audio-visual scenes from single images, bridging the gap between visual scene generation and spatial audio synthesis for complete multimodal experiences.

Abstract: Tremendous progress in visual scene generation now turns a single image into an explorable 3D world, yet immersion remains incomplete without sound. We introduce Image2AVScene, the task of generating a 3D audio-visual scene from a single image, and present SonoWorld, the first framework to tackle this challenge. From one image, our pipeline outpaints a 360° panorama, lifts it into a navigable 3D scene, places language-guided sound anchors, and renders ambisonics for point, areal, and ambient sources, yielding spatial audio aligned with scene geometry and semantics. Quantitative evaluations on a newly curated real-world dataset and a controlled user study confirm the effectiveness of our approach. Beyond free-viewpoint audio-visual rendering, we also demonstrate applications to one-shot acoustic learning and audio-visual spatial source separation. Project website: https://humathe.github.io/sonoworld/

Method: CNMBI leverages inherent data distribution information to map clustering as a dynamic comparison process between cluster centers regarding positional behavior. It uses bipartite graph theory to model this process efficiently and actively removes low-confidence samples based on their confidence levels.

Result: Extensive comparisons with state-of-the-art methods on various challenging datasets (including CIFAR-10 and STL-10) demonstrate CNMBI’s superiority. The method shows robustness and flexibility with different data dimensions and shapes.

Conclusion: CNMBI provides an effective approach for determining optimal cluster numbers without prior information, overcoming limitations of traditional validation-based methods and handling complex real-world data better.

Abstract: One of the main challenges in data mining is choosing the optimal number of clusters without prior information. Notably, existing methods are usually in the philosophy of cluster validation and hence have underlying assumptions on data distribution, which prevents their application to complex data such as large-scale images and high-dimensional data from the real world. In this regard, we propose an approach named CNMBI. Leveraging the distribution information inherent in the data space, we map the target task as a dynamic comparison process between cluster centers regarding positional behavior, without relying on the complete clustering results and designing the complex validity index as before. Bipartite graph theory is then employed to efficiently model this process. Additionally, we find that different samples have different confidence levels and thereby actively remove low-confidence ones, which is, for the first time to our knowledge, considered in cluster number determination. CNMBI is robust and allows for more flexibility in the dimension and shape of the target data (e.g., CIFAR-10 and STL-10). Extensive comparison studies with state-of-the-art competitors on various challenging datasets demonstrate the superiority of our method.

Method: Motion Semantics Guided Normalizing Flow (MSG-Flow) decomposes skeleton-based VAD into hierarchical motion semantics modeling: 1) Vector quantized variational auto-encoder discretizes continuous motion into interpretable primitives, 2) Autoregressive Transformer models semantic-level temporal dependencies, 3) Conditional normalizing flow captures detail-level pose variations.

Result: Achieves state-of-the-art performance on HR-ShanghaiTech (88.1% AUC) and HR-UBnormal (75.8% AUC) benchmarks.

Conclusion: MSG-Flow effectively models hierarchical motion semantics for skeleton-based video anomaly detection, offering privacy-preserving approach with improved discriminability for anomalies at different abstraction levels.

Abstract: As embodied perception systems increasingly bridge digital and physical realms in interactive multimedia applications, the need for privacy-preserving approaches to understand human activities in physical environments has become paramount. Video anomaly detection is a critical task in such embodied multimedia systems for intelligent surveillance and forensic analysis. Skeleton-based approaches have emerged as a privacy-preserving alternative that processes physical world information through abstract human pose representations while discarding sensitive visual attributes such as identity and facial features. However, existing skeleton-based methods predominantly model continuous motion trajectories in a monolithic manner, failing to capture the hierarchical nature of human activities composed of discrete semantic primitives and fine-grained kinematic details, which leads to reduced discriminability when anomalies manifest at different abstraction levels. In this regard, we propose Motion Semantics Guided Normalizing Flow (MSG-Flow) that decomposes skeleton-based VAD into hierarchical motion semantics modeling. It employs vector quantized variational auto-encoder to discretize continuous motion into interpretable primitives, an autoregressive Transformer to model semantic-level temporal dependencies, and a conditional normalizing flow to capture detail-level pose variations. Extensive experiments on benchmarks (HR-ShanghaiTech & HR-UBnormal) demonstrate that MSG-Flow achieves state-of-the-art performance with 88.1% and 75.8% AUC respectively.

Method: Proposes TDEC with three key components: 1) T-Encoder using Transformer to learn discriminative features with global dependencies, 2) Dim-Reduction block to create clustering-friendly low-dimensional space, and 3) distribution-aware clustering that considers embedded feature distributions for reliable supervision.

Result: TDEC achieves significantly higher clustering performance than recent state-of-the-art approaches on complex datasets, demonstrating robustness across different data sizes, cluster numbers, and context complexities.

Conclusion: The joint consideration of feature representation, dimensional preference, and robust assignment in TDEC provides a superior deep embedded image clustering framework that effectively handles complex image data.

Abstract: Image clustering is a crucial but challenging task in multimedia machine learning. Recently the combination of clustering with deep learning has achieved promising performance against conventional methods on high-dimensional image data. Unfortunately, existing deep clustering methods (DC) often ignore the importance of information fusion with a global perception field among different image regions on clustering images, especially complex ones. Additionally, the learned features are usually clustering-unfriendly in terms of dimensionality and are based only on simple distance information for the clustering. In this regard, we propose a deep embedded image clustering TDEC, which for the first time to our knowledge, jointly considers feature representation, dimensional preference, and robust assignment for image clustering. Specifically, we introduce the Transformer to form a novel module T-Encoder to learn discriminative features with global dependency while using the Dim-Reduction block to build a friendly low-dimensional space favoring clustering. Moreover, the distribution information of embedded features is considered in the clustering process to provide reliable supervised signals for joint training. Our method is robust and allows for more flexibility in data size, the number of clusters, and the context complexity. More importantly, the clustering performance of TDEC is much higher than recent competitors. Extensive experiments with state-of-the-art approaches on complex datasets show the superiority of TDEC.

Method: Built upon established multimodal foundations (Qwen3-VL, InternVL3, Bagel), the authors systematically curated SenseNova-SI-8M: eight million diverse data samples under a rigorous taxonomy of spatial capabilities. They take a principled approach to constructing high-performing spatial intelligence models.

Result: SenseNova-SI demonstrates unprecedented performance across spatial intelligence benchmarks: 68.8% on VSI-Bench, 43.3% on MMSI, 85.7% on MindCube, 54.7% on ViewSpatial, 47.7% on SITE, 63.9% on BLINK, 55.5% on 3DSR, and 72.0% on EmbSpatial, while maintaining strong general multimodal understanding (84.9% on MMBench-En).

Conclusion: The work successfully cultivates spatial intelligence in multimodal foundation models through systematic data scaling, shows emergent generalization capabilities, analyzes risks of overfitting and language shortcuts, presents spatial chain-of-thought reasoning, and validates downstream applications. All models are publicly released.

Abstract: Despite remarkable progress, multimodal foundation models still exhibit surprising deficiencies in spatial intelligence. In this work, we explore scaling up multimodal foundation models to cultivate spatial intelligence within the SenseNova-SI family, built upon established multimodal foundations including visual understanding models (i.e., Qwen3-VL and InternVL3) and unified understanding and generation models (i.e., Bagel). We take a principled approach to constructing high-performing and robust spatial intelligence by systematically curating SenseNova-SI-8M: eight million diverse data samples under a rigorous taxonomy of spatial capabilities. SenseNova-SI demonstrates unprecedented performance across a broad range of spatial intelligence benchmarks: 68.8% on VSI-Bench, 43.3% on MMSI, 85.7% on MindCube, 54.7% on ViewSpatial, 47.7% on SITE, 63.9% on BLINK, 55.5% on 3DSR, and 72.0% on EmbSpatial, while maintaining strong general multimodal understanding (e.g., 84.9% on MMBench-En). More importantly, we analyze the impact of data scaling, discuss early signs of emergent generalization capabilities enabled by diverse data training, analyze the risk of overfitting and language shortcuts, present a preliminary study on spatial chain-of-thought reasoning, and validate the potential downstream application. All newly trained multimodal foundation models are publicly released.

Method: Conducted controlled empirical study comparing diffusion and rectified flow objectives within MotionGPT3 framework. Held model architecture, training protocol, and evaluation setup fixed to isolate effect of generative objective. Used HumanML3D dataset for experiments.

Result: Rectified flow converges in fewer training epochs, reaches strong test performance earlier, and matches or exceeds diffusion-based motion quality. Flow-based priors show stable behavior across inference step counts and achieve competitive quality with fewer sampling steps, yielding better efficiency-quality trade-offs.

Conclusion: Benefits of rectified flow objectives extend to continuous-latent text-to-motion generation, highlighting importance of training objective choice in motion priors.

Abstract: Recent text-driven motion generation methods span both discrete token-based approaches and continuous-latent formulations. MotionGPT3 exemplifies the latter paradigm, combining a learned continuous motion latent space with a diffusion-based prior for text-conditioned synthesis. While rectified flow objectives have recently demonstrated favorable convergence and inference-time properties relative to diffusion in image and audio generation, it remains unclear whether these advantages transfer cleanly to the motion generation setting. In this work, we conduct a controlled empirical study comparing diffusion and rectified flow objectives within the MotionGPT3 framework. By holding the model architecture, training protocol, and evaluation setup fixed, we isolate the effect of the generative objective on training dynamics, final performance, and inference efficiency. Experiments on the HumanML3D dataset show that rectified flow converges in fewer training epochs, reaches strong test performance earlier, and matches or exceeds diffusion-based motion quality under identical conditions. Moreover, flow-based priors exhibit stable behavior across a wide range of inference step counts and achieve competitive quality with fewer sampling steps, yielding improved efficiency–quality trade-offs. Overall, our results suggest that several known benefits of rectified flow objectives do extend to continuous-latent text-to-motion generation, highlighting the importance of the training objective choice in motion priors.

Method: Uses Physics-Driven Audio Adapter (Phy-Adapter) with physical parameters from Physical Parameter Estimator (PPE), which employs VLM for mass estimation and segmentation-based 3D reconstruction for motion trajectory and velocity computation.

Result: Outperforms existing V2A models in quantitative and qualitative evaluations, producing physically plausible and perceptually coherent audio.

Conclusion: Incorporating physical reasoning into V2A generation significantly improves audio realism and physical consistency.

Abstract: Recent advances in Video-to-Audio (V2A) generation have achieved impressive perceptual quality and temporal synchronization, yet most models remain appearance-driven, capturing visual-acoustic correlations without considering the physical factors that shape real-world sounds. We present Physics-Aware Video-to-Audio Synthesis (PAVAS), a method that incorporates physical reasoning into a latent diffusion-based V2A generation through the Physics-Driven Audio Adapter (Phy-Adapter). The adapter receives object-level physical parameters estimated by the Physical Parameter Estimator (PPE), which uses a Vision-Language Model (VLM) to infer the moving-object mass and a segmentation-based dynamic 3D reconstruction module to recover its motion trajectory for velocity computation. These physical cues enable the model to synthesize sounds that reflect underlying physical factors. To assess physical realism, we curate VGG-Impact, a benchmark focusing on object-object interactions, and introduce Audio-Physics Correlation Coefficient (APCC), an evaluation metric that measures consistency between physical and auditory attributes. Comprehensive experiments show that PAVAS produces physically plausible and perceptually coherent audio, outperforming existing V2A models in both quantitative and qualitative evaluations. Visit https://physics-aware-video-to-audio-synthesis.github.io for demo videos.

Method: Proposes LogiStory framework with explicit visual logic modeling using a multi-agent system that grounds roles, extracts causal chains, and verifies story-level consistency. Also introduces LogicTale benchmark for evaluation.

Result: Experiments show significant improvement in narrative logic of generated visual stories. The approach effectively bridges structured story planning with visual generation.

Conclusion: Provides foundational step toward modeling visual logic in general image sequence and video generation tasks, transforming narrative coherence from implicit byproduct to explicit modeling objective.

Abstract: Generating coherent and communicative visual sequences, such as image sequences and videos, remains a significant challenge for current multimodal systems. Despite advances in visual quality and the integration of world knowledge, existing models still struggle to maintain logical flow, often resulting in disjointed actions, fragmented narratives, and unclear storylines. We attribute these issues to the lack of attention to visual logic, a critical yet underexplored dimension of visual sequence generation that we define as the perceptual and causal coherence among characters, actions, and scenes over time. To bridge this gap, we propose a logic-aware multi-image story visualization framework, LogiStory. The framework is built around the central innovation of explicitly modeling visual logic in story visualization. To realize this idea, we design a multi-agent system that grounds roles, extracts causal chains, and verifies story-level consistency, transforming narrative coherence from an implicit byproduct of image generation into an explicit modeling objective. This design effectively bridges structured story planning with visual generation, enhancing both narrative clarity and visual quality in story visualization. Furthermore, to evaluate the generation capacity, we construct LogicTale, a benchmark comprising richly annotated stories, emphasizing causal reasoning, and visual logic interpretability. We establish comprehensive automatic and human evaluation protocols designed to measure both visual logic and perceptual quality. Experiments demonstrate that our approach significantly improves the narrative logic of generated visual stories. This work provides a foundational step towards modeling and enforcing visual logic in general image sequence and video generation tasks.

Method: Survey methodology examining the historical development through literature review, covering classical texture descriptors, machine learning classifiers, deep learning (CNNs, Vision Transformers, GNNs), foundation models, vision-language systems, and generative AI approaches.

Result: Comprehensive analysis showing the paradigmatic transformation from handcrafted features to sophisticated AI systems, highlighting breakthrough developments in self-supervised foundation models, vision-language systems, and generative AI for tackling annotation costs and data challenges.

Conclusion: Remote sensing scene classification has evolved dramatically, with future priorities including advancing hyperspectral/multi-temporal analysis, developing robust cross-domain generalization methods, and establishing standardized evaluation protocols.

Abstract: Remote sensing scene classification has experienced a paradigmatic transformation from traditional handcrafted feature methods to sophisticated artificial intelligence systems that now form the backbone of modern Earth observation applications. This comprehensive survey examines the complete methodological evolution, systematically tracing development from classical texture descriptors and machine learning classifiers through the deep learning revolution to current state-of-the-art foundation models and generative AI approaches. We chronicle the pivotal shift from manual feature engineering to automated hierarchical representation learning via convolutional neural networks, followed by advanced architectures including Vision Transformers, graph neural networks, and hybrid frameworks. The survey provides in-depth coverage of breakthrough developments in self-supervised foundation models and vision-language systems, highlighting exceptional performance in zero-shot and few-shot learning scenarios. Special emphasis is placed on generative AI innovations that tackle persistent challenges through synthetic data generation and advanced feature learning strategies. We analyze contemporary obstacles including annotation costs, multimodal data fusion complexities, interpretability demands, and ethical considerations, alongside current trends in edge computing deployment, federated learning frameworks, and sustainable AI practices. Based on comprehensive analysis of recent advances and gaps, we identify key future research priorities: advancing hyperspectral and multi-temporal analysis capabilities, developing robust cross-domain generalization methods, and establishing standardized evaluation protocols to accelerate scientific progress in remote sensing scene classification systems.

Method: Two-module system: 1) Real-time module uses CLAHE-enhanced stereo imagery, YOLOv11n object detection, and neural network for distance estimation; 2) Offline module uses Depth Anything V2 for monocular depth estimation and Open3D for point cloud fusion.

Result: Neural network achieves median depth error of 2.26 cm within 1-10 meter range on Chandrayaan 3 NavCam data; object detection maintains balanced precision-recall on lunar scenes; system provides reliable metric context alongside qualitative reconstructions.

Conclusion: The architecture offers a scalable, compute-efficient vision solution for autonomous planetary exploration by combining real-time perception with detailed offline reconstruction capabilities.

Abstract: This study presents a vision system for planetary rovers, combining real-time perception with offline terrain reconstruction. The real-time module integrates CLAHE enhanced stereo imagery, YOLOv11n based object detection, and a neural network to estimate object distances. The offline module uses the Depth Anything V2 metric monocular depth estimation model to generate depth maps from captured images, which are fused into dense point clouds using Open3D. Real world distance estimates from the real time pipeline provide reliable metric context alongside the qualitative reconstructions. Evaluation on Chandrayaan 3 NavCam stereo imagery, benchmarked against a CAHV based utility, shows that the neural network achieves a median depth error of 2.26 cm within a 1 to 10 meter range. The object detection model maintains a balanced precision recall tradeoff on grayscale lunar scenes. This architecture offers a scalable, compute-efficient vision solution for autonomous planetary exploration.

Method: Pipeline uses real camera trap photos from iWildCam, applies MegaDetector bounding boxes and stratified sampling across 8 species. Generative phenotype editing creates severity variants for hair loss (mange) and emaciation. Adaptive scene drift QC uses sham prefilter and decoupled mask/score approach with day/night metrics to reject altered scenes.

Result: From 201 base images across 4 species, generated 553 QC-passing synthetic variants with 83% pass rate. Sim-to-real transfer experiment training only on synthetic data achieved 0.85 AUROC on real camera trap images of suspected health conditions.

Conclusion: Synthetic data pipeline successfully captures visual features sufficient for wildlife health screening, demonstrating practical utility for automated monitoring where real labeled data is scarce.

Abstract: No publicly available, ML ready datasets exist for wildlife health conditions in camera trap imagery, creating a fundamental barrier to automated health screening. We present a pipeline for generating synthetic training images depicting alopecia and body condition deterioration in wildlife from real camera trap photographs. Our pipeline constructs a curated base image set from iWildCam using MegaDetector derived bounding boxes and center frame weighted stratified sampling across 8 North American species. A generative phenotype editing system produces controlled severity variants depicting hair loss consistent with mange and emaciation. An adaptive scene drift quality control system uses a sham prefilter and decoupled mask then score approach with complementary day or night metrics to reject images where the generative model altered the original scene. We frame the pipeline explicitly as a screening data source. From 201 base images across 4 species, we generate 553 QC passing synthetic variants with an overall pass rate of 83 percent. A sim to real transfer experiment training exclusively on synthetic data and testing on real camera trap images of suspected health conditions achieves 0.85 AUROC, demonstrating that the synthetic data captures visual features sufficient for screening.

Method: Two-stage approach: 1) Feed-forward image deblurring network with specialized training for SLAM, followed by local-global multi-view optimization for successfully deblurred frames. 2) For failed deblurring cases, uses 3D Gaussian Splatting (3DGS) representation with an additional blur network to model multiple blurred sub-frames and simulate blur formation in 3D space, learning sharp details and refined sub-frame poses.

Result: Experiments on real-world datasets show consistent improvements in both pose estimation accuracy and sharp reconstruction of geometry and texture compared to existing methods.

Conclusion: Unblur-SLAM effectively handles different types of blur in SLAM systems, adapts computation based on blur severity, and achieves state-of-the-art performance for sharp 3D reconstruction from blurred inputs.

Abstract: We propose Unblur-SLAM, a novel RGB SLAM pipeline for sharp 3D reconstruction from blurred image inputs. In contrast to previous work, our approach is able to handle different types of blur and demonstrates state-of-the-art performance in the presence of both motion blur and defocus blur. Moreover, we adjust the computation effort with the amount of blur in the input image. As a first stage, our method uses a feed-forward image deblurring model for which we propose a suitable training scheme that can improve both tracking and mapping modules. Frames that are successfully deblurred by the feed-forward network obtain refined poses and depth through local-global multi-view optimization and loop closure. Frames that fail the first stage deblurring are directly modeled through the global 3DGS representation and an additional blur network to model multiple blurred sub-frames and simulate the blur formation process in 3D space, thereby learning sharp details and refined sub-frame poses. Experiments on several real-world datasets demonstrate consistent improvements in both pose estimation and sharp reconstruction results of geometry and texture.

Method: Developed a prompt-free pipeline using YOLO26-Seg for joint detection and segmentation of three fetal brain structures (Brain, CSP, LV) in a single forward pass. Key modifications include: 1) composite BCE-Dice-Lovász segmentation loss with inverse-frequency class weighting via runtime monkey-patching, 2) domain-guided copy-paste augmentation respecting anatomical locations, and 3) inter-patient stratified splitting to prevent data leakage.

Result: Achieved mean Dice coefficient of 0.9253 on 575 test images, exceeding the baseline (0.9012) by 2.68 percentage points, despite only reporting over three foreground classes (baseline includes easy background class).

Conclusion: The prompt-free YOLO26-Seg approach with composite loss and domain-specific augmentations effectively improves fetal brain structure segmentation in ultrasound images, addressing practical bottlenecks in obstetric imaging.

Abstract: Segmenting fetal head structures from prenatal ultrasound remains a practical bottleneck in obstetric imaging. The current state-of-the-art baseline, proposed alongside the published dataset, adapts the Segment Anything Model with per-class Dice and Lovász losses but still depends on bounding-box prompts at test time. We build a prompt-free pipeline on top of YOLO26-Seg that jointly detects and segments three structures, Brain, Cavum Septi Pellucidi (CSP), and Lateral Ventricles (LV), in a single forward pass. Three modifications are central to our approach: (i) a composite BCE-Dice-Lovász segmentation loss with inverse-frequency class weighting, injected into the YOLO26 training loop via runtime monkey-patching; (ii) domain-guided copy-paste augmentation that transplants minority-class structures while respecting their anatomical location relative to the brain boundary; and (iii) inter-patient stratified splitting to prevent data leakage. On 575 held-out test images, the composite loss variant reaches a mean Dice coefficient of 0.9253, exceeding the baseline (0.9012) by 2.68 percentage points, despite reporting over three foreground classes only, whereas the baseline’s reported mean includes the easy background class. We further ablate each component and discuss annotation-quality and class-imbalance effects on CSP and LV performance.

Method: Proposes BTK framework that uses Qwen3-4B to extract goal-related phrases, Flux-Schnell to construct image knowledge bases (R2R-GP, REVERIE-GP), and BLIP-2 to create textual knowledge bases from panoramic views. Integrates these via Goal-Aware Augmentor and Knowledge Augmentor.

Result: Significant improvements on R2R (7,189 trajectories) and REVERIE (21,702 instructions) datasets. On test unseen splits: SR increased by 5% and 2.07% respectively, SPL increased by 4% and 3.69% respectively.

Conclusion: BTK effectively integrates multimodal knowledge bases to enhance semantic grounding and cross-modal alignment in VLN, outperforming existing baselines on standard benchmarks.

Abstract: Vision-and-Language Navigation (VLN) requires an agent to navigate through complex unseen environments based on natural language instructions. However, existing methods often struggle to effectively capture key semantic cues and accurately align them with visual observations. To address this limitation, we propose Beyond Textual Knowledge (BTK), a VLN framework that synergistically integrates environment-specific textual knowledge with generative image knowledge bases. BTK employs Qwen3-4B to extract goal-related phrases and utilizes Flux-Schnell to construct two large-scale image knowledge bases: R2R-GP and REVERIE-GP. Additionally, we leverage BLIP-2 to construct a large-scale textual knowledge base derived from panoramic views, providing environment-specific semantic cues. These multimodal knowledge bases are effectively integrated via the Goal-Aware Augmentor and Knowledge Augmentor, significantly enhancing semantic grounding and cross-modal alignment. Extensive experiments on the R2R dataset with 7,189 trajectories and the REVERIE dataset with 21,702 instructions demonstrate that BTK significantly outperforms existing baselines. On the test unseen splits of R2R and REVERIE, SR increased by 5% and 2.07% respectively, and SPL increased by 4% and 3.69% respectively. The source code is available at https://github.com/yds3/IPM-BTK/.

Method: Introduces GradAttn, a hybrid CNN-transformer framework that replaces fixed residual connections with attention-controlled gradient flow. Extracts multi-scale CNN features at different depths and regulates them through self-attention to dynamically weight shallow texture features and deep semantic representations.

Result: GradAttn outperforms ResNet-18 on 5 of 8 diverse datasets (natural images, medical imaging, fashion recognition), achieving up to +11.07% accuracy improvement on FashionMNIST while maintaining comparable network size. Gradient flow analysis reveals controlled instabilities from attention often coincide with improved generalization.

Conclusion: Attention mechanisms can serve as enablers of learnable gradient control, offering a new paradigm for adaptive representation learning in deep neural architectures. Positional encoding effectiveness is dataset dependent, with CNN hierarchies often encoding sufficient spatial structure.

Abstract: Deep ConvNets suffer from gradient signal degradation as network depth increases, limiting effective feature learning in complex architectures. ResNet addressed this through residual connections, but these fixed short-circuits cannot adapt to varying input complexity or selectively emphasize task relevant features across network hierarchies. This study introduces GradAttn, a hybrid CNN-transformer framework that replaces fixed residual connections with attention-controlled gradient flow. By extracting multi-scale CNN features at different depths and regulating them through self-attention, GradAttn dynamically weights shallow texture features and deep semantic representations. For representational analysis, we evaluated three GradAttn variants across eight diverse datasets, from natural images, medical imaging, to fashion recognition. Results demonstrate that GradAttn outperforms ResNet-18 on five of eight datasets, achieving up to +11.07% accuracy improvement on FashionMNIST while maintaining comparable network size. Gradient flow analysis reveals that controlled instabilities, introduced by attention, often coincide with improved generalization, challenging the assumption that perfect stability is optimal. Furthermore, positional encoding effectiveness proves dataset dependent, with CNN hierarchies frequently encoding sufficient spatial structure. These findings allow attention mechanisms as enablers of learnable gradient control, offering a new paradigm for adaptive representation learning in deep neural architectures.

Method: Proposes MD-RWKV-UNet with MD-RWKV blocks integrating deformable spatial shifts with Receptance Weighted Key Value mechanism for dynamic receptive field adaptation. Uses Selective Kernel Attention for adaptive convolutional kernel selection with varying receptive fields. Implements cross-stage dual-attention fusion strategy to aggregate multi-level features across encoder.

Result: Achieves state-of-the-art performance on Synapse and ACDC datasets, particularly excelling in boundary precision and small-organ segmentation.

Conclusion: The approach provides a lightweight yet expressive solution for dynamic organ modeling, outperforming methods that rely on static convolutions or global attention for medical image segmentation tasks.

Abstract: Multi-organ segmentation in medical imaging remains challenging due to large anatomical variability, complex inter-organ dependencies, and diverse organ scales and shapes. Conventional encoder-decoder architectures often struggle to capture both fine-grained local details and long-range context, which are crucial for accurate delineation - especially for small or deformable organs. To address these limitations, we propose MD-RWKV-UNet, a dynamic encoder network that enables scale-aware representation and spatially adaptive context modeling. At its core is the MD-RWKV block, a dual-path module that integrates deformable spatial shifts with the Receptance Weighted Key Value mechanism, allowing the receptive field to adapt dynamically to local structural cues. We further incorporate Selective Kernel Attention to enable adaptive selection of convolutional kernels with varying receptive fields, enhancing multi-scale interaction and improving robustness to organ size and shape variation. In parallel, a cross-stage dual-attention fusion strategy aggregates multi-level features across the encoder, preserving low-level structure while enhancing semantic consistency. Unlike methods that stack static convolutions or rely heavily on global attention, our approach provides a lightweight yet expressive solution for dynamic organ modeling. Experiments on Synapse and ACDC demonstrate state-of-the-art performance, particularly in boundary precision and small-organ segmentation.

Method: Proposes Scanline-Consistent Range-Aware Diffusion using Partial Diffusion (SDEdit) on a coarse prior. Introduces Ray-Consistency loss and Negative Ray Augmentation to enforce sensor physics and suppress artifacts.

Result: Achieves high-fidelity results in 156ms, state-of-the-art performance on KITTI-360 and nuScenes datasets, and directly boosts off-the-shelf 3D detectors without retraining.

Conclusion: The framework effectively addresses LiDAR sparsity through probabilistic refinement with sensor physics constraints, enabling practical real-time applications and improved downstream perception tasks.

Abstract: LiDAR perception is severely limited by the distance-dependent sparsity of distant objects. While diffusion models can recover dense geometry, they suffer from prohibitive latency and physical hallucinations manifesting as ghost points. We propose Scanline-Consistent Range-Aware Diffusion, a framework that treats densification as probabilistic refinement rather than generation. By leveraging Partial Diffusion (SDEdit) on a coarse prior, we achieve high-fidelity results in just 156ms. Our novel Ray-Consistency loss and Negative Ray Augmentation enforce sensor physics to suppress artifacts. Our method achieves state-of-the-art results on KITTI-360 and nuScenes, directly boosting off-the-shelf 3D detectors without retraining. Code will be made available.

Method: Two-stage training: Stage 1 learns internal motion prior using an audio-free autoregressive generator to capture natural facial dynamics. Stage 2 introduces dual-track audio to fine-tune the generator, modulating the learned motion prior based on external speech cues.

Result: Achieves state-of-the-art speaking accuracy and delivers up to 44.1% improvement in listening metrics, generating significantly more diverse and natural listening expressions while effectively mitigating stiffness problems.

Conclusion: UniLS provides a practical, high-fidelity audio-driven solution for interactive digital humans by enabling end-to-end generation of unified speak-listen expressions using only dual-track audio.

Abstract: Generating lifelike conversational avatars requires modeling not just isolated speakers, but the dynamic, reciprocal interaction of speaking and listening. However, modeling the listener is exceptionally challenging: direct audio-driven training fails, producing stiff, static listening motions. This failure stems from a fundamental imbalance: the speaker’s motion is strongly driven by speech audio, while the listener’s motion primarily follows an internal motion prior and is only loosely guided by external speech. This challenge has led most methods to focus on speak-only generation. The only prior attempt at joint generation relies on extra speaker’s motion to produce the listener. This design is not end-to-end, thereby hindering the real-time applicability. To address this limitation, we present UniLS, the first end-to-end framework for generating unified speak-listen expressions, driven by only dual-track audio. Our method introduces a novel two-stage training paradigm. Stage 1 first learns the internal motion prior by training an audio-free autoregressive generator, capturing the spontaneous dynamics of natural facial motion. Stage 2 then introduces the dual-track audio, fine-tuning the generator to modulate the learned motion prior based on external speech cues. Extensive evaluations show UniLS achieves state-of-the-art speaking accuracy. More importantly, it delivers up to 44.1% improvement in listening metrics, generating significantly more diverse and natural listening expressions. This effectively mitigates the stiffness problem and provides a practical, high-fidelity audio-driven solution for interactive digital humans. Code and demos are available at https://xg-chu.site/project_unils/.

Method: Uses high-frequency wireless sensing (HFWS) to generate 3D facial pointclouds via wearable sensors. Creates AffectNet3D dataset using FLAME-based method to convert existing 2D datasets to 3D. Implements pointcloud refinement pipeline and trains PointNet++ model, with fine-tuning on BU-3DFE dataset.

Result: Achieves over 70% classification accuracy on BU-3DFE, comparable to oracle performance. Models trained on AffectNet3D and fine-tuned with 25% of BU-3DFE outperform those trained solely on BU-3DFE. Works well even with simulated wearable conditions (masked pointclouds).

Conclusion: HFWS-based FER is viable for continuous, privacy-aware emotion monitoring via wearables. The proposed pipeline effectively addresses 3D dataset scarcity and demonstrates practical feasibility for real-world applications.

Abstract: Facial Emotion Recognition is a critical research area within Affective Computing due to its wide-ranging applications in Human Computer Interaction, mental health assessment and fatigue monitoring. Current FER methods predominantly rely on Deep Learning techniques trained on 2D image data, which pose significant privacy concerns and are unsuitable for continuous, real-time monitoring. As an alternative, we propose High-Frequency Wireless Sensing (HFWS) as an enabler of continuous, privacy-aware FER, through the generation of detailed 3D facial pointclouds via on-person sensors embedded in wearables. We present arguments supporting the privacy advantages of HFWS over traditional 2D imaging, particularly under increasingly stringent data protection regulations. A major barrier to adopting HFWS for FER is the scarcity of labeled 3D FER datasets. Towards addressing this issue, we introduce a FLAME-based method to generate 3D facial pointclouds from existing public 2D datasets. Using this approach, we create AffectNet3D, a 3D version of the AffectNet database. To evaluate the quality and usability of the generated data, we design a pointcloud refinement pipeline focused on isolating the facial region, and train the popular PointNet++ model on the refined pointclouds. Fine-tuning the model on a small subset of the unseen 3D FER dataset BU-3DFE yields a classification accuracy exceeding 70%, comparable to oracle-level performance. To further investigate the potential of HFWS-based FER for continuous monitoring, we simulate wearable sensing conditions by masking portions of the generated pointclouds. Experimental results show that models trained on AffectNet3D and fine-tuned with just 25% of BU-3DFE outperform those trained solely on BU-3DFE. These findings highlight the viability of our pipeline and support the feasibility of continuous, privacy-aware FER via wearable HFWS systems.

Method: Integrates lightweight human pose estimation models with biomechanical feature extraction and CNN-LSTM temporal learning architecture. Computes joint angles and skeletal features from detected keypoints, compares with reference poses, uses quantitative scoring for alignment deviations, and provides real-time feedback via visual, text, and voice guidance with Edge AI optimization (quantization, pruning) for resource-constrained devices.

Result: Proposed framework enables real-time yoga pose detection and posture correction with low latency on edge devices, providing intelligent corrective feedback to improve user safety and training effectiveness.

Conclusion: The Edge AI-based framework offers a scalable digital yoga assistant that can enhance safety and effectiveness in modern fitness applications by providing real-time posture correction and guidance.

Abstract: Yoga is widely recognized for improving physical fitness, flexibility, and mental well being. However, these benefits depend strongly on correct posture execution. Improper alignment during yoga practice can reduce effectiveness and increase the risk of musculoskeletal injuries, especially in self guided or online training environments. This paper presents a hybrid Edge AI based framework for real time yoga pose detection and posture correction. The proposed system integrates lightweight human pose estimation models with biomechanical feature extraction and a CNN LSTM based temporal learning architecture to recognize yoga poses and analyze motion dynamics. Joint angles and skeletal features are computed from detected keypoints and compared with reference pose configurations to evaluate posture correctness. A quantitative scoring mechanism is introduced to measure alignment deviations and generate real time corrective feedback through visual, text based, and voice based guidance. In addition, Edge AI optimization techniques such as model quantization and pruning are applied to enable low latency performance on resource constrained devices. The proposed framework provides an intelligent and scalable digital yoga assistant that can improve user safety and training effectiveness in modern fitness applications.

Method: Proposes Tiny-ViT, a small and efficient Vision Transformer designed for resource-constrained environments. Uses image preprocessing (resizing, CLAHE, Gaussian blur) and is evaluated on a 3-class dataset (Early Blight, Late Blight, Healthy). Incorporates GRAD-CAM for model interpretability to identify diseased regions.

Result: Achieves 99.85% test accuracy and 99.82% mean cross-validation accuracy, outperforming baseline models (DEIT Small, SWIN Tiny, MobileViT XS). Shows high reliability with MCC of 0.9990 and narrow confidence intervals [0.9980, 0.9995]. Competitive inference times with low computational costs enable real-time applications.

Conclusion: Tiny-ViT provides a robust, efficient, and explainable solution for plant disease classification, particularly suitable for deployment in resource-limited agricultural settings with real-time requirements.

Abstract: Early and precise identification of plant diseases, especially in potato crops is important to ensure the health of the crops and ensure the maximum yield . Potato leaf diseases, such as Early Blight and Late Blight, pose significant challenges to farmers, often resulting in yield losses and increased pesticide use. Traditional methods of detection are not only time-consuming, but are also subject to human error, which is why automated and efficient methods are required. The paper introduces a new method of potato leaf disease classification Tiny-ViT model, which is a small and effective Vision Transformer (ViT) developed to be used in resource-limited systems. The model is tested on a dataset of three classes, namely Early Blight, Late Blight, and Healthy leaves, and the preprocessing procedures include resizing, CLAHE, and Gaussian blur to improve the quality of the image. Tiny-ViT model has an impressive test accuracy of 99.85% and a mean CV accuracy of 99.82% which is better than baseline models such as DEIT Small, SWIN Tiny, and MobileViT XS. In addition to this, the model has a Matthews Correlation Coefficient (MCC) of 0.9990 and narrow confidence intervals (CI) of [0.9980, 0.9995], which indicates high reliability and generalization. The training and testing inference time is competitive, and the model exhibits low computational expenses, thereby, making it applicable in real-time applications. Moreover, interpretability of the model is improved with the help of GRAD-CAM, which identifies diseased areas. Altogether, the proposed Tiny-ViT is a solution with a high level of robustness, efficiency, and explainability to the problem of plant disease classification.

Method: Controlled Poisson noise applied to high-dose CT images to simulate realistic low-dose CT conditions. Two pipelines compared: (1) direct classification of noisy LDCT images, and (2) denoising followed by classification. Performance evaluated across multiple dose levels using accuracy, sensitivity, and AUC metrics.

Result: While denoising improves perceptual image quality, it does not consistently improve classification. In several settings, direct classification yields higher sensitivity. The best denoise-then-classify pipeline achieves 0.94 AUC and 0.91 accuracy at moderate dose levels, outperforming direct classification by up to 6% in select cases.

Conclusion: This work establishes a reproducible baseline for LDCT stroke triage using hemorrhagic stroke data and highlights the need for validation on ischemic cohorts and real-world portable CT systems. Reveals trade-off between perceptual quality and diagnostic utility.

Abstract: Portable CT scanners enable early stroke detection in prehospital and low-resource settings but require reduced radiation doses, introducing noise that degrades diagnostic reliability. We present a deep learning framework for stroke classification from simulated low-dose CT (LDCT) brain scans for AI-assisted triage in mobile clinical environments. Controlled Poisson noise is applied to high-dose CT images to simulate realistic LDCT conditions. We compare two pipelines: (1) direct classification of noisy LDCT images and (2) denoising followed by classification. Performance is evaluated across multiple dose levels using accuracy, sensitivity, and AUC. While denoising improves perceptual image quality, it does not consistently improve classification. In several settings, direct classification yields higher sensitivity, revealing a trade-off between perceptual quality and diagnostic utility. The best denoise-then-classify pipeline achieves 0.94 AUC and 0.91 accuracy at moderate dose levels, outperforming direct classification by up to 6% in select cases. This work establishes a reproducible baseline for LDCT stroke triage using hemorrhagic stroke data (RSNA dataset) and highlights the need for validation on ischemic cohorts and real-world portable CT systems.

Method: Three key innovations: 1) Comprehensive noise simulation layer modeling realistic screen-shooting distortions including Moiré patterns, 2) JND perceptual loss function adaptively modulating watermark strength, 3) Two automatic localization modules for captured image rectification and anti-cropping recovery.

Result: Achieves average PSNR of 30.94 dB and SSIM of 0.94 on watermarked images while embedding 127-bit payloads, demonstrating strong performance in both visual quality and robustness.

Conclusion: The proposed framework effectively addresses screen-shooting watermarking challenges by jointly optimizing embedding and extraction through realistic distortion modeling, perceptual quality preservation, and automated localization.

Abstract: Screen-shooting robust watermarking aims to imperceptibly embed extractable information into host images such that the watermark survives the complex distortion pipeline of screen display and camera recapture. However, achieving high extraction accuracy while maintaining satisfactory visual quality remains an open challenge, primarily because the screen-shooting channel introduces severe and entangled degradations including Moiré patterns, color-gamut shifts, perspective warping, and sensor noise. In this paper, we present an end-to-end deep learning framework that jointly optimizes watermark embedding and extraction for screen-shooting robustness. Our framework incorporates three key innovations: (i) a comprehensive noise simulation layer that faithfully models realistic screen-shooting distortions – notably including a physically-motivated Moiré pattern generator – enabling the network to learn robust representations against the full spectrum of capture-channel noise through adversarial training; (ii) a Just Noticeable Distortion (JND) perceptual loss function that adaptively modulates watermark embedding strength by supervising the perceptual discrepancy between the JND coefficient map and the watermark residual, thereby concentrating watermark energy in perceptually insensitive regions to maximize visual quality; and (iii) two complementary automatic localization modules – a semantic-segmentation-based foreground extractor for captured image rectification and a symmetric noise template mechanism for anti-cropping region recovery – that enable fully automated watermark decoding under realistic deployment conditions. Extensive experiments demonstrate that our method achieves an average PSNR of 30.94~dB and SSIM of 0.94 on watermarked images while embedding 127-bit payloads.

Method: Proposes a training-free framework with synergistic system that disentangles structure and appearance. Uses high-fidelity inversion to establish content prior capturing lighting and micro-textures, and novel attention-sharing mechanism to dynamically fuse purified appearance features from reference guided by geometric priors.

Result: Unified approach operates at 1024px and outperforms specialized methods on tasks ranging from semantic attribute transfer to fine-grained material application. Extensive experiments confirm state-of-the-art performance in both structural preservation and appearance fidelity.

Conclusion: The proposed framework successfully addresses the challenge of controllable appearance transfer in DiTs, enabling high-fidelity editing while preserving scene structure, outperforming existing methods across various editing tasks.

Abstract: Diffusion Transformers (DiTs) excel at generation, but their global self-attention makes controllable, reference-image-based editing a distinct challenge. Unlike U-Nets, naively injecting local appearance into a DiT can disrupt its holistic scene structure. We address this by proposing the first training-free framework specifically designed to tame DiTs for high-fidelity appearance transfer. Our core is a synergistic system that disentangles structure and appearance. We leverage high-fidelity inversion to establish a rich content prior for the source image, capturing its lighting and micro-textures. A novel attention-sharing mechanism then dynamically fuses purified appearance features from a reference, guided by geometric priors. Our unified approach operates at 1024px and outperforms specialized methods on tasks ranging from semantic attribute transfer to fine-grained material application. Extensive experiments confirm our state-of-the-art performance in both structural preservation and appearance fidelity.

Method: Leverage vision-language models (VLMs) to analyze handwritten Chinese character quality and generate multi-level feedback. Explore both LoRA-based fine-tuning strategies and in-context learning methods to integrate aesthetic assessment knowledge into VLMs for two tasks: simple grade feedback and enriched descriptive feedback.

Result: The approach achieves state-of-the-art performances across multiple evaluation tracks in the CCL 2025 workshop on evaluation of handwritten Chinese character quality.

Conclusion: VLMs can effectively generate actionable multi-level feedback for handwritten Chinese character quality assessment, moving beyond simple regression-based scoring to provide meaningful guidance for learners.

Abstract: The handwriting of Chinese characters is a fundamental aspect of learning the Chinese language. Previous automated assessment methods often framed scoring as a regression problem. However, this score-only feedback lacks actionable guidance, which limits its effectiveness in helping learners improve their handwriting skills. In this paper, we leverage vision-language models (VLMs) to analyze the quality of handwritten Chinese characters and generate multi-level feedback. Specifically, we investigate two feedback generation tasks: simple grade feedback (Task 1) and enriched, descriptive feedback (Task 2). We explore both low-rank adaptation (LoRA)-based fine-tuning strategies and in-context learning methods to integrate aesthetic assessment knowledge into VLMs. Experimental results show that our approach achieves state-of-the-art performances across multiple evaluation tracks in the CCL 2025 workshop on evaluation of handwritten Chinese character quality.

Method: Comparative analysis of a 7B-parameter quantized VLM (Qwen2.5-VL-7B) vs. a 500M-parameter model (SmolVLM2-500M) using 4,000 samples from VQAv2 and COCO Captions. Applied three-category error taxonomy, used GPT-4o as judge, measured confidence calibration via ECE, tested compositional reasoning with structured negation probes, and conducted blur robustness experiments.

Result: Compact model shows distinct failure signature: 12.5pp larger negation collapse (-33.2pp vs. -20.8pp), driven almost entirely by COCO. Most discriminating template shows SmolVLM2-500M responds “Yes” (incorrectly) on 100% of COCO trials vs. 14% for larger model. Asymmetric dataset-dependent miscalibration observed.

Conclusion: Compressed VLMs exhibit qualitatively different failure patterns, with severe compositional reasoning deficits, necessitating systematic safety auditing before edge deployment.

Abstract: The rapid compression of large vision-language models (VLMs) for edge deployment raises an underexplored question: do compact models fail differently, not merely more often? This study compares a 7-billion-parameter quantised VLM (Qwen2.5-VL-7B, 4-bit NF4) against a 500-million-parameter FP16 model (SmolVLM2-500M) across 4,000 samples from VQAv2 and COCO Captions. A three-category error taxonomy (Object Blindness, Semantic Drift, Prior Bias) is applied as a diagnostic framework. A text-only GPT-4o judge reveals Semantic Drift (B) as the dominant failure mode on VQAv2 and on COCO for Qwen, with a mixed Object Blindness / Semantic Drift profile for SmolVLM2 on COCO; Prior Bias (C) is present on VQAv2 but absent on COCO for both models. Confidence calibration is measured via Expected Calibration Error (ECE) using geometric mean token probability, compositional reasoning is probed with structured negation probes across four templates, and a blur robustness experiment completes the evaluation. For this model pair, the compact model exhibits a qualitatively distinct failure signature: a 12.5pp larger negation collapse (-33.2pp vs. -20.8pp, Wald 95% CI [8.2, 16.8]pp, p < 10^-8), driven almost entirely by COCO while the VQAv2 gap is not statistically significant (4.5pp, p=0.19). The most discriminating template is false_yn: SMOLVLM2-500M responds “Yes” (incorrectly claiming a depicted object is absent) on 100% of COCO trials vs. 14% for Q WEN 2.5-VL-7B. Asymmetric dataset-dependent miscalibration and a blur experiment with two controlled ablations complete the analysis. The fully reproducible pipeline is released for systematic safety auditing of compressed VLMs prior to edge deployment.

Method: End-to-end pipeline using LLaVA-1.5-7B for visual damage analysis, structured JSON extraction, and rule-based priority scoring. Systematic comparison of three quantization levels (Q4_K_M, Q5_K_M, Q8_0) on 254 rebar exposure images with 5-point quality evaluation framework.

Result: Q5_K_M achieves optimal balance: quality score 3.18±1.35/5.0, inference time 5.67s/image, and 0.56 quality/sec efficiency - 8.5% higher quality than Q4_K_M with only 4.5% speed reduction, while matching Q8_0’s quality with 25% faster inference.

Conclusion: Q5_K_M quantization provides the best trade-off for practical deployment of VLMs in bridge inspection, offering consistent performance regardless of description length while maintaining efficiency on consumer-grade hardware.

Abstract: Bridge infrastructure inspection is a critical but labor-intensive task requiring expert assessment of structural damage such as rebar exposure, cracking, and corrosion. This paper presents a comprehensive study of quantized Vision-Language Models (VLMs) for automated bridge damage assessment, focusing on the trade-offs between description quality, inference speed, and resource requirements. We develop an end-to-end pipeline combining LLaVA-1.5-7B for visual damage analysis, structured JSON extraction, and rule-based priority scoring. To enable deployment on consumer-grade GPUs, we conduct a systematic comparison of three quantization levels: Q4_K_M, Q5_K_M, and Q8_0 across 254 rebar exposure images. We introduce a 5-point quality evaluation framework assessing damage type recognition, severity classification. Our results demonstrate that Q5_K_M achieves the optimal balance: quality score 3.18$\pm$1.35/5.0, inference time 5.67s/image, and 0.56 quality/sec efficiency – 8.5% higher quality than Q4_K_M with only 4.5% speed reduction, while matching Q8_0’s quality with 25% faster inference. Statistical analysis reveals Q5_K_M exhibits the weakest text-quality correlation (-0.148), indicating consistent performance regardless of description length.

Method: Constructed domain-specific benchmark of Italian TV news clips labeled across four semantic dimensions. Evaluated two pipeline architectures across nine frontier models (Gemini 3.0 Pro, LLaMA 4 Maverick, Qwen-VL variants, Gemma 3) under progressively enriched input strategies combining visual signals, ASR, speaker diarization, and metadata.

Result: Gains from video input are strongly model-dependent: larger models effectively leverage temporal continuity, while smaller models show performance degradation under extended multimodal context due to token overload. Deployed pipeline on 14 full broadcast episodes with minute-level annotations integrated with audience measurement data.

Conclusion: Demonstrated operational viability of MLLM-based framework for content-based audience analytics, enabling correlational analysis of topic-level audience sensitivity and generational engagement divergence in broadcast TV.

Abstract: Automated semantic annotation of broadcast television content presents distinctive challenges, combining structured audiovisual composition, domain-specific editorial patterns, and strict operational constraints. While multimodal large language models (MLLMs) have demonstrated strong general-purpose video understanding capabilities, their comparative effectiveness across pipeline architectures and input configurations in broadcast-specific settings remains empirically undercharacterized. This paper presents a systematic evaluation of multimodal annotation pipelines applied to broadcast television news in the Italian setting. We construct a domain-specific benchmark of clips labeled across four semantic dimensions: visual environment classification, topic classification, sensitive content detection, and named entity recognition. Two different pipeline architectures are evaluated across nine frontier models, including Gemini 3.0 Pro, LLaMA 4 Maverick, Qwen-VL variants, and Gemma 3, under progressively enriched input strategies combining visual signals, automatic speech recognition, speaker diarization, and metadata. Experimental results demonstrate that gains from video input are strongly model-dependent: larger models effectively leverage temporal continuity, while smaller models show performance degradation under extended multimodal context, likely due to token overload. Beyond benchmarking, the selected pipeline is deployed on 14 full broadcast episodes, with minute-level annotations integrated with normalized audience measurement data provided by an Italian media company. This integration enables correlational analysis of topic-level audience sensitivity and generational engagement divergence, demonstrating the operational viability of the proposed framework for content-based audience analytics.

Method: REVL-PV embeds domain-specific diagnostic reasoning into multimodal learning across electroluminescence, thermal, and visible-light imagery. The framework requires the model to link visual evidence to plausible defect mechanisms before classification, producing structured diagnostic reports.

Result: Achieves 93% classification accuracy on 1,927 real-world modules spanning eight defect categories, produces interpretable diagnostic rationales, maintains robustness under realistic image corruptions, and shows strong semantic alignment with expert assessments in blind concordance studies.

Conclusion: Reasoning-aware multimodal learning establishes a general paradigm for trustworthy AI-assisted inspection of photovoltaic energy infrastructure by combining visual analysis with diagnostic reasoning.

Abstract: Reliable photovoltaic defect identification is essential for maintaining energy yield, ensuring warranty compliance, and enabling scalable inspection of rapidly expanding solar fleets. Although recent advances in computer vision have improved automated defect detection, most existing systems operate as opaque classifiers that provide limited diagnostic insight for high-stakes energy infrastructure. Here we introduce REVL-PV, a vision-language framework that embeds domain-specific diagnostic reasoning into multimodal learning across electroluminescence, thermal, and visible-light imagery. By requiring the model to link visual evidence to plausible defect mechanisms before classification, the framework produces structured diagnostic reports aligned with professional photovoltaic inspection practice. Evaluated on 1,927 real-world modules spanning eight defect categories, REVL-PV achieves 93% classification accuracy while producing interpretable diagnostic rationales and maintaining strong robustness under realistic image corruptions. A blind concordance study with a certified solar inspection expert shows strong semantic alignment between model explanations and expert assessments across defect identification, root-cause attribution, and visual descriptions. These results demonstrate that reasoning-aware multimodal learning establishes a general paradigm for trustworthy AI-assisted inspection of photovoltaic energy infrastructure.

Method: BHCast uses a neural model to transform static images into forecasted future frames with multi-scale pyramid loss for autoregressive forecasting. It extracts spatio-temporal features (pattern speed, pitch angle) and uses gradient-boosting trees to recover black hole properties from these features.

Result: The framework successfully forecasts dynamics from blurry frames, enabling stable long-term movie generation and accurate extraction of black hole properties (spin and viewing inclination) from simulated and real EHT data for Sagittarius A* and M87*.

Conclusion: BHCast establishes a scalable paradigm for solving inverse problems in astrophysics by using learned dynamics to extract insights from resolution-limited data, with modular design enabling interpretability and uncertainty quantification.

Abstract: The Event Horizon Telescope (EHT) delivered the first image of a black hole by capturing the light from its surrounding accretion flow, revealing structure but not dynamics. Simulations of black hole accretion dynamics are essential for interpreting EHT images but costly to generate and impractical for inference. Motivated by this bottleneck, BHCast presents a framework for forecasting black hole plasma dynamics from a single, blurry snapshot, such as those captured by the EHT. At its core, BHCast is a neural model that transforms a static image into forecasted future frames, revealing the underlying dynamics hidden within one snapshot. With a multi-scale pyramid loss, we demonstrate how autoregressive forecasting can simultaneously super-resolve and evolve a blurry frame into a coherent, high-resolution movie that remains stable over long time horizons. From forecasted dynamics, we can then extract interpretable spatio-temporal features, such as pattern speed (rotation rate) and pitch angle. Finally, BHCast uses gradient-boosting trees to recover black hole properties from these plasma features, including the spin and viewing inclination angle. The separation between forecasting and inference provides modular flexibility, interpretability, and robust uncertainty quantification. We demonstrate the effectiveness of BHCast on simulations of two distinct black hole accretion systems, Sagittarius A* and M87*, by testing on simulated frames blurred to EHT resolution and real EHT images of M87*. Ultimately, our methodology establishes a scalable paradigm for solving inverse problems, demonstrating the potential of learned dynamics to unlock insights from resolution-limited scientific data.

Method: Used a dual-module architecture: a reasoning module (MLLM) interacts with an imagery module on 3D model rotation tasks. The imagery module handles rendering and rotating 3D models, outsourcing the burden of maintaining/manipulating holistic 3D states.

Result: Performance was lower than expected, with accuracy reaching at most 62.5%. Even with the imagery module handling 3D manipulation, the system still fails, indicating current frontier models lack foundational visual-spatial primitives.

Conclusion: Current MLLMs lack essential visual-spatial capabilities: (1) low-level sensitivity to extract spatial signals (depth, motion, short-horizon dynamic prediction), and (2) capacity for contemplative reasoning over images with dynamic visual focus and integration of imagery with symbolic/associative information.

Abstract: Large Language Models (LLMs) have demonstrated impressive reasoning capabilities, yet they struggle with spatial tasks that require mental simulation, such as mental rotation. This paper investigates whether equipping an LLM with an external Imagery Module'' -- a tool capable of rendering and rotating 3D models -- can bridge this gap, functioning as a cognitive prosthetic.’’ We conducted experiments using a dual-module architecture in which a reasoning module (an MLLM) interacts with an imagery module on 3D model rotation tasks. Performance was lower than expected, with accuracy reaching at most 62.5%. Further investigation suggests that even when the burden of maintaining and manipulating a holistic 3D state is outsourced, the system still fails. This reveals that current frontier models lack the foundational visual-spatial primitives required to interface with imagery. Specifically, they lack: (1) the low-level sensitivity to extract spatial signals such as (a) depth, (b) motion, and (c) short-horizon dynamic prediction; and (2) the capacity to reason contemplatively over images, dynamically shifting visual focus and balancing imagery with symbolic and associative information.

Method: Introduce RatSeizure dataset with recorded clips annotated with seizure-related action units and temporal boundaries. Propose RaSeformer, a saliency-context Transformer for temporal action localization that highlights behavior-relevant context while suppressing redundant cues.

Result: RaSeformer achieves strong performance on RatSeizure dataset and provides a competitive reference model. Established standardized dataset splits and evaluation protocols for reproducible benchmarking.

Conclusion: RatSeizure addresses critical limitations in animal seizure research by providing fine-grained temporal annotations and standardized evaluation, enabling better behavior analysis and temporal localization.

Abstract: Animal models, particularly rats, play a critical role in seizure research for studying epileptogenesis and treatment response. However, progress is limited by the lack of datasets with precise temporal annotations and standardized evaluation protocols. Existing animal behavior datasets often have limited accessibility, coarse labeling, and insufficient temporal localization of clinically meaningful events. To address these limitations, we introduce RatSeizure, the first publicly benchmark for fine-grained seizure behavior analysis. The dataset consists of recorded clips annotated with seizure-related action units and temporal boundaries, enabling both behavior classification and temporal localization. We further propose RaSeformer, a saliency-context Transformer for temporal action localization that highlights behavior-relevant context while suppressing redundant cues. Experiments on RatSeizure show that RaSeformer achieves strong performance and provides a competitive reference model for this challenging task. We also establish standardized dataset splits and evaluation protocols to support reproducible benchmarking.

Method: The paper introduces stroke-size control as a controlled intervention that modifies the effective roughness of supervised targets, predictions, and perturbations across different timesteps. This approach aims to ease the low signal-to-noise challenge by adjusting the granularity of operations.

Result: The authors analyze the advantages and trade-offs of this intervention both theoretically and empirically, though specific quantitative results are not provided in the abstract.

Conclusion: Stroke-size control offers a promising approach to improve diffusion model performance in challenging low signal-to-noise conditions by adapting the granularity of operations throughout the denoising process.

Abstract: Diffusion models can be challenged in the low signal-to-noise regime, where they have to make pixel-level predictions despite the presence of high noise. The geometric intuition is akin to using the finest stroke for oil painting throughout, which may be ineffective. We therefore study stroke-size control as a controlled intervention that changes the effective roughness of the supervised target, predictions and perturbations across timesteps, in an attempt to ease the low signal-to-noise challenge. We analyze the advantages and trade-offs of the intervention both theoretically and empirically. Code will be released.

Method: Introduces HighlightBench, a diagnostic benchmark with five task families: Markup Grounding, Constrained Retrieval, Local Relations, Aggregation & Comparison, and Consistency & Missingness. Provides a reference pipeline that makes intermediate decisions explicit for reproducible baselines and finer-grained error attribution.

Result: Experiments show that even strong models remain unstable when visual cues must be consistently aligned with symbolic reasoning under structured output constraints.

Conclusion: The benchmark addresses a critical gap in evaluating MLLMs’ markup understanding capabilities, enabling better assessment of perception vs. reasoning failures in table-centric document analysis.

Abstract: Visual markups such as highlights, underlines, and bold text are common in table-centric documents. Although multimodal large language models (MLLMs) have made substantial progress in document understanding, their ability to treat such cues as explicit logical directives remains under-explored. More importantly, existing evaluations cannot distinguish whether a model fails to see the markup or fails to reason with it. This creates a key blind spot in assessing markup-conditioned behavior over tables. To address this gap, we introduce HighlightBench, a diagnostic benchmark for markup-driven table understanding that decomposes evaluation into five task families: Markup Grounding, Constrained Retrieval, Local Relations, Aggregation & Comparison, and Consistency & Missingness. We further provide a reference pipeline that makes intermediate decisions explicit, enabling reproducible baselines and finer-grained attribution of errors along the perception-to-execution chain. Experiments show that even strong models remain unstable when visual cues must be consistently aligned with symbolic reasoning under structured output constraints.

Method: Proposes a spike fusion mechanism that integrates unimodal features at the spike level to generate enhanced multimodal representations, which act as soft supervisory signals to refine unimodal spike embeddings, mitigating semantic loss within the CMSF network.

Result: Despite requiring only two time steps, CMSF achieves top-tier retrieval accuracy, surpassing state-of-the-art ANN counterparts while maintaining exceptionally low energy consumption and high retrieval speed.

Conclusion: This work represents a significant step toward multimodal SNNs, offering a brain-inspired framework that unifies temporal dynamics with cross-modal alignment and provides new insights for future spiking-based multimodal research.

Abstract: Spiking neural networks (SNNs) have recently shown strong potential in unimodal visual and textual tasks, yet building a directly trained, low-energy, and high-performance SNN for multimodal applications such as image-text retrieval (ITR) remains highly challenging. Existing artificial neural network (ANN)-based methods often pursue richer unimodal semantics using deeper and more complex architectures, while overlooking cross-modal interaction, retrieval latency, and energy efficiency. To address these limitations, we present a brain-inspired Cross-Modal Spike Fusion network (CMSF) and apply it to ITR for the first time. The proposed spike fusion mechanism integrates unimodal features at the spike level, generating enhanced multimodal representations that act as soft supervisory signals to refine unimodal spike embeddings, effectively mitigating semantic loss within CMSF. Despite requiring only two time steps, CMSF achieves top-tier retrieval accuracy, surpassing state-of-the-art ANN counterparts while maintaining exceptionally low energy consumption and high retrieval speed. This work marks a significant step toward multimodal SNNs, offering a brain-inspired framework that unifies temporal dynamics with cross-modal alignment and provides new insights for future spiking-based multimodal research. The code is available at https://github.com/zxt6174/CMSF.

Method: Applied uncertainty estimation techniques (deep ensemble, test-time augmentation, Monte Carlo dropout) to a state-of-the-art DL pipeline for cardiac functional biomarker estimation. Proposed new distribution-based metrics for assessing biomarker precision and evaluated on two external validation scan-rescan CMR datasets.

Result: Model achieved high accuracy (average Dice 87%) and good point estimate precision. However, distribution-based metrics showed poor scan-rescan agreement: confidence interval overlap >50% in less than 45% of cases, and statistical tests showed significant differences in over 65% of cases.

Conclusion: While point estimate metrics suggest good performance, distributional analyses reveal lower precision, highlighting the need to use more representative metrics to assess scan-rescan agreement in medical imaging applications.

Abstract: The performance of deep learning (DL) methods for the analysis of cine cardiovascular magnetic resonance (CMR) is typically assessed in terms of accuracy, overlooking precision. In this work, uncertainty estimation techniques, namely deep ensemble, test-time augmentation, and Monte Carlo dropout, are applied to a state-of-the-art DL pipeline for cardiac functional biomarker estimation, and new distribution-based metrics are proposed for the assessment of biomarker precision. The model achieved high accuracy (average Dice 87%) and point estimate precision on two external validation scan-rescan CMR datasets. However, distribution-based metrics showed that the overlap between scan/rescan confidence intervals was >50% in less than 45% of the cases. Statistical similarity tests between scan and rescan biomarkers also resulted in significant differences for over 65% of the cases. We conclude that, while point estimate metrics might suggest good performance, distributional analyses reveal lower precision, highlighting the need to use more representative metrics to assess scan-rescan agreement.

Method: Systematic analysis of flow-matching design space for cell-microscopy images, identifying unnecessary techniques and developing a simple, stable, and scalable recipe for training foundation models, then fine-tuning with pre-trained molecular embeddings.

Result: Scaled model to two orders of magnitude larger than prior methods, achieving two-fold FID and ten-fold KID improvement over prior methods, with state-of-the-art performance simulating responses to unseen molecules after fine-tuning.

Conclusion: Many popular flow-matching techniques for cell microscopy are unnecessary or harmful, and a simpler, more scalable approach yields significantly better generative performance for simulating cellular responses to perturbations.

Abstract: Flow-matching generative models are increasingly used to simulate cell responses to biological perturbations. However, the design space for building such models is large and underexplored. We systematically analyse the design space of flow matching models for cell-microscopy images, finding that many popular techniques are unnecessary and can even hurt performance. We develop a simple, stable, and scalable recipe which we use to train our foundation model. We scale our model to two orders of magnitude larger than prior methods, achieving a two-fold FID and ten-fold KID improvement over prior methods. We then fine-tune our model with pre-trained molecular embeddings to achieve state-of-the-art performance simulating responses to unseen molecules. Code is available at https://github.com/valence-labs/microscopy-flow-matching

Method: Develops PhyDCM as a modular open-source framework with hybrid classification architecture based on MedViT, standardized DICOM processing, interactive desktop visualization interface, and standardized preprocessing including intensity rescaling and limited data augmentation.

Result: Achieves over 93% classification accuracy across categories on MRI datasets from BRISC2025 and curated Kaggle collections (FigShare, SARTAJ, and Br35H), with stable diagnostic performance.

Conclusion: PhyDCM provides a practical foundation for reproducible AI-driven medical image analysis with transparency, modularity, and accessibility, offering flexibility for future integration of additional imaging modalities.

Abstract: MRI-based medical imaging has become indispensable in modern clinical diagnosis, particularly for brain tumor detection. However, the rapid growth in data volume poses challenges for conventional diagnostic approaches. Although deep learning has shown strong performance in automated classification, many existing solutions are confined to closed technical architectures, limiting reproducibility and further academic development. PhyDCM is introduced as an open-source software framework that integrates a hybrid classification architecture based on MedViT with standardized DICOM processing and an interactive desktop visualization interface. The system is designed as a modular digital library that separates computational logic from the graphical interface, allowing independent modification and extension of components. Standardized preprocessing, including intensity rescaling and limited data augmentation, ensures consistency across varying MRI acquisition settings. Experimental evaluation on MRI datasets from BRISC2025 and curated Kaggle collections (FigShare, SARTAJ, and Br35H) demonstrates stable diagnostic performance, achieving over 93% classification accuracy across categories. The framework supports structured, exportable outputs and multi-planar reconstruction of volumetric data. By emphasizing transparency, modularity, and accessibility, PhyDCM provides a practical foundation for reproducible AI-driven medical image analysis, with flexibility for future integration of additional imaging modalities.

Method: ViPAC uses a dual-branch strategy to disentangle periodic and aperiodic components of vibrotactile signals, with dynamic fusion mechanism, orthogonality constraint, and weighting regularization. Also created LMT108-CAP dataset using GPT-4o for vibrotactile-text pairs.

Result: ViPAC significantly outperforms baseline methods adapted from audio and image captioning, achieving superior lexical fidelity and semantic alignment in generating natural language descriptions from vibrotactile signals.

Conclusion: The paper successfully addresses vibrotactile captioning for the first time, proposing an effective method that handles the intrinsic properties of vibrotactile data and demonstrating its superiority over adapted baselines.

Abstract: The standardization of vibrotactile data by IEEE P1918.1 workgroup has greatly advanced its applications in virtual reality, human-computer interaction and embodied artificial intelligence. Despite these efforts, the semantic interpretation and understanding of vibrotactile signals remain an unresolved challenge. In this paper, we make the first attempt to address vibrotactile captioning, {\it i.e.}, generating natural language descriptions from vibrotactile signals. We propose Vibrotactile Periodic-Aperiodic Captioning (ViPAC), a method designed to handle the intrinsic properties of vibrotactile data, including hybrid periodic-aperiodic structures and the lack of spatial semantics. Specifically, ViPAC employs a dual-branch strategy to disentangle periodic and aperiodic components, combined with a dynamic fusion mechanism that adaptively integrates signal features. It also introduces an orthogonality constraint and weighting regularization to ensure feature complementarity and fusion consistency. Additionally, we construct LMT108-CAP, the first vibrotactile-text paired dataset, using GPT-4o to generate five constrained captions per surface image from the popular LMT-108 dataset. Experiments show that ViPAC significantly outperforms the baseline methods adapted from audio and image captioning, achieving superior lexical fidelity and semantic alignment.

Method: Unified seed-controlled protocol comparing SIREN, Fourier features, Haar positional encoding, and multi-resolution grid on 950 grayscale microscopy images from MapZebrain atlas. Images normalized with per-image percentiles, spatial generalization tested with 40% column-wise hold-out along X-axis.

Result: Haar and Fourier achieved strongest reconstruction fidelity (~26 dB) on held-out columns, better preserving boundaries according to SSIM and edge-focused metrics. SIREN performed worse in macro averages but competitive on micro averages. Grid method moderately behind.

Conclusion: Explicit spectral and multiscale encodings (Haar, Fourier) better capture high-frequency neuroanatomical detail than smoother alternatives, making them best for boundary-sensitive tasks like atlas registration and label transfer, while SIREN remains suitable for background modeling.

Abstract: Implicit neural representations (INRs) offer continuous coordinate-based encodings for atlas registration, cross-modality resampling, sparse-view completion, and compact sharing of neuroanatomical data. Yet reproducible evaluation is lacking for high-resolution larval zebrafish microscopy, where preserving neuropil boundaries and fine neuronal processes is critical. We present a reproducible INR benchmark for the MapZebrain larval zebrafish brain atlas. Using a unified, seed-controlled protocol, we compare SIREN, Fourier features, Haar positional encoding, and a multi-resolution grid on 950 grayscale microscopy images, including atlas slices and single-neuron projections. Images are normalized with per-image (1,99) percentiles estimated from 10% of pixels in non-held-out columns, and spatial generalization is tested with a deterministic 40% column-wise hold-out along the X-axis. Haar and Fourier achieve the strongest macro-averaged reconstruction fidelity on held-out columns (about 26 dB), while the grid is moderately behind. SIREN performs worse in macro averages but remains competitive on area-weighted micro averages in the all-in-one regime. SSIM and edge-focused error further show that Haar and Fourier preserve boundaries more accurately. These results indicate that explicit spectral and multiscale encodings better capture high-frequency neuroanatomical detail than smoother-bias alternatives. For MapZebrain workflows, Haar and Fourier are best suited to boundary-sensitive tasks such as atlas registration, label transfer, and morphology-preserving sharing, while SIREN remains a lightweight baseline for background modelling or denoising.

Method: Uses 3D Gaussian Splatting (3DGS) for surgical scene reconstruction and temporally tracked surface representation. Integrates Extended Position-Based Dynamics (XPBD) with local deformation constraints to produce representative geometry points (RGPs) with anisotropic stiffness metrics. Incorporates robotic tool poses in SE(3) to compute rigidly induced displacements, deriving physics-aware compliance energy and positional agreement scores for affordance prediction.

Result: Experiments on surgical video datasets show arg-VU yields more stable, physically consistent, and interpretable affordance predictions than kinematic baselines. The framework demonstrates reliable affordance reasoning for deformable surgical environments and supports embodied robotic interaction.

Conclusion: Physics-aware geometric representations enable reliable affordance reasoning for deformable surgical environments, bridging the gap between perception and action in surgical robotics through physically consistent modeling of tissue-tool interactions.

Abstract: Affordance reasoning provides a principled link between perception and action, yet remains underexplored in surgical robotics, where tissues are highly deformable, compliant, and dynamically coupled with tool motion. We present arg-VU, a physics-aware affordance reasoning framework that integrates temporally consistent geometry tracking with constraint-induced mechanical modeling for surgical visual understanding. Surgical scenes are reconstructed using 3D Gaussian Splatting (3DGS) and converted into a temporally tracked surface representation. Extended Position-Based Dynamics (XPBD) embeds local deformation constraints and produces representative geometry points (RGPs) whose constraint sensitivities define anisotropic stiffness metrics capturing the local constraint-manifold geometry. Robotic tool poses in SE(3) are incorporated to compute rigidly induced displacements at RGPs, from which we derive two complementary measures: a physics-aware compliance energy that evaluates mechanical feasibility with respect to local deformation constraints, and a positional agreement score that captures motion alignment (as kinematic motion baseline). Experiments on surgical video datasets show that arg-VU yields more stable, physically consistent, and interpretable affordance predictions than kinematic baselines. These results demonstrate that physics-aware geometric representations enable reliable affordance reasoning for deformable surgical environments and support embodied robotic interaction.

Method: Developed a multimodal generative AI framework that integrates text prompts and geospatial controls to generate high-fidelity, diverse urban satellite imagery. The system learns from surrounding environments for urban redevelopment and encodes latent representations of urban form for cross-city style transfer.

Result: Successfully generates realistic urban images across 500 largest metropolitan areas worldwide. The framework enables style transfer of urban environments across global spatial networks and enhances downstream prediction tasks like carbon emission prediction. Human expert evaluation confirms generated images are comparable to real urban images.

Conclusion: The study presents innovative approaches for accelerated urban planning and supports scenario-based planning processes for worldwide cities through multimodal generative AI that can envision sustainable urban development.

Abstract: Urban development has been a defining force in human history, shaping cities for centuries. However, past studies mostly analyze such development as predictive tasks, failing to reflect its generative nature. Therefore, this study designs a multimodal generative AI framework to envision sustainable urban development at a global scale. By integrating prompts and geospatial controls, our framework can generate high-fidelity, diverse, and realistic urban satellite imagery across the 500 largest metropolitan areas worldwide. It enables users to specify urban development goals, creating new images that align with them while offering diverse scenarios whose appearance can be controlled with text prompts and geospatial constraints. It also facilitates urban redevelopment practices by learning from the surrounding environment. Beyond visual synthesis, we find that it encodes and interprets latent representations of urban form for global cross-city learning, successfully transferring styles of urban environments across a global spatial network. The latent representations can also enhance downstream prediction tasks such as carbon emission prediction. Further, human expert evaluation confirms that our generated urban images are comparable to real urban images. Overall, this study presents innovative approaches for accelerated urban planning and supports scenario-based planning processes for worldwide cities.

Method: Uses dual-view optical flow from synchronized viewpoints, automatic apex-frame detection, sequence decomposition into onset-apex and apex-offset phases, and Triple-Stream MicroAttNet with fusion attention and squeeze-excitation blocks.

Result: Achieves macro-UF1 score of 0.536 on 4DME dataset, outperforming baseline by over 50% and securing first place in 4DMR IJCAI Workshop Challenge 2025; ablation shows fusion attention and SE components each contribute up to 3.6 UF1 points.

Conclusion: Dual-view, phase-aware optical flow with multi-stream fusion provides robust, interpretable solution for 4D micro-expression recognition, demonstrating effectiveness of multimodal feature integration.

Abstract: Micro-expression recognition is vital for affective computing but remains challenging due to the extremely brief, low-intensity facial motions involved and the high-dimensional nature of 4D mesh data. To address these challenges, we introduce a dual-view optical flow approach that simplifies mesh processing by capturing each micro-expression sequence from two synchronized viewpoints and computing optical flow to represent motion. Our pipeline begins with view separation and sequence-wise face cropping to ensure spatial consistency, followed by automatic apex-frame detection based on peak motion intensity in both views. We decompose each sequence into onset-apex and apex-offset phases, extracting horizontal, vertical, and magnitude flow channels for each phase. These are fed into our Triple-Stream MicroAttNet, which employs a fusion attention module to adaptively weight modality-specific features and a squeeze-and-excitation block to enhance magnitude representations. Training uses focal loss to mitigate class imbalance and the Adam optimizer with early stopping. Evaluated on the multi-label 4DME dataset, comprising 24 subjects and five emotion categories, in the 4DMR IJCAI Workshop Challenge 2025, our method achieves a macro-UF1 score of 0.536, outperforming the official baseline by over 50% and securing first place. Ablation studies confirm that both the fusion attention and SE components each contribute up to 3.6 points of UF1 gain. These results demonstrate that dual-view, phase-aware optical flow combined with multi-stream fusion yields a robust and interpretable solution for 4D micro-expression recognition.

Method: LACON (Labeling-and-Conditioning) framework uses quality signals like aesthetic scores and watermark probabilities as explicit, quantitative condition labels. Instead of filtering bad data, it trains generative models to learn the full spectrum from bad to good quality, learning explicit boundaries between high and low-quality content.

Result: LACON achieves superior generation quality compared to baselines trained only on filtered data using the same compute budget, demonstrating the significant value of uncurated data that was previously discarded.

Conclusion: The discarded “bad” data in text-to-image generation contains valuable information that can be leveraged through explicit conditioning on quality signals, leading to better model performance without additional computational cost.

Abstract: The success of modern text-to-image generation is largely attributed to massive, high-quality datasets. Currently, these datasets are curated through a filter-first paradigm that aggressively discards low-quality raw data based on the assumption that it is detrimental to model performance. Is the discarded bad data truly useless, or does it hold untapped potential? In this work, we critically re-examine this question. We propose LACON (Labeling-and-Conditioning), a novel training framework that exploits the underlying uncurated data distribution. Instead of filtering, LACON re-purposes quality signals, such as aesthetic scores and watermark probabilities as explicit, quantitative condition labels. The generative model is then trained to learn the full spectrum of data quality, from bad to good. By learning the explicit boundary between high- and low-quality content, LACON achieves superior generation quality compared to baselines trained only on filtered data using the same compute budget, proving the significant value of uncurated data.

Method: Test-time framework that converts pretrained CNNs into self-explainable models via convolution-based replacement of classification heads, initialized from original weights to preserve performance.

Result: Preserves black-box predictive performance while delivering built-in faithful explanations competitive with post-hoc methods, both qualitatively and quantitatively.

Conclusion: TTE-CAM bridges the gap between interpretability and performance in medical image analysis, enabling clinical adoption of high-performing CNNs with faithful explanations.

Abstract: Convolutional neural networks (CNNs) achieve state-of-the-art performance in medical image analysis yet remain opaque, limiting adoption in high-stakes clinical settings. Existing approaches face a fundamental trade-off: post-hoc methods provide unfaithful approximate explanations, while inherently interpretable architectures are faithful but often sacrifice predictive performance. We introduce TTE-CAM, a test-time framework that bridges this gap by converting pretrained black-box CNNs into self-explainable models via a convolution-based replacement of their classification head, initialized from the original weights. The resulting model preserves black-box predictive performance while delivering built-in faithful explanations competitive with post-hoc methods, both qualitatively and quantitatively. The code is available at https://github.com/kdjoumessi/Test-Time-Explainability

Method: Proposes SuperCam, a novel camera design that performs superpixel segmentation on the fly during image capture. The system adaptively processes captured data by grouping similar pixels into superpixels, significantly reducing data volume while preserving essential visual information.

Result: SuperCam outperforms current state-of-the-art superpixel algorithms in memory-constrained situations. The compressed data maintains high performance for downstream computer vision tasks including image segmentation, object detection, and monocular depth estimation when camera memory is limited.

Conclusion: Superpixel segmentation will be crucial for deploying computer vision models on edge devices. SuperCam enables more efficient system designs for resource-constrained applications by reducing data redundancy at the capture stage while maintaining task performance.

Abstract: Conventional cameras generate a lot of data that can be challenging to process in resource-constrained applications. Usually, cameras generate data streams on the order of the number of pixels in the image. However, most of this captured data is redundant for many downstream computer vision algorithms. We propose a novel camera design, which we call SuperCam, that adaptively processes captured data by performing superpixel segmentation on the fly. We show that SuperCam performs better than current state-of-the-art superpixel algorithms under memory-constrained situations. We also compare how well SuperCam performs when the compressed data is used for downstream computer vision tasks. Our results demonstrate that the proposed design provides superior output for image segmentation, object detection, and monocular depth estimation in situations where the available memory on the camera is limited. We posit that superpixel segmentation will play a crucial role as more computer vision inference models are deployed in edge devices. SuperCam would allow computer vision engineers to design more efficient systems for these applications.

Method: Proposes FusionAgent: an agentic framework using Multimodal Large Language Models (MLLMs) for dynamic, sample-specific model selection. Each expert model is treated as a tool, and through Reinforcement Fine-Tuning (RFT) with metric-based reward, the agent learns optimal model combinations. Introduces Anchor-based Confidence Top-k (ACT) score-fusion to address model score misalignment and embedding heterogeneity.

Result: Extensive experiments on multiple whole-body biometric benchmarks demonstrate that FusionAgent significantly outperforms state-of-the-art methods while achieving higher efficiency through fewer model invocations.

Conclusion: FusionAgent underscores the critical role of dynamic, explainable, and robust model fusion in real-world recognition systems, showing that adaptive model selection via MLLMs can improve both performance and efficiency.

Abstract: Model fusion is a key strategy for robust recognition in unconstrained scenarios, as different models provide complementary strengths. This is especially important for whole-body human recognition, where biometric cues such as face, gait, and body shape vary across samples and are typically integrated via score-fusion. However, existing score-fusion strategies are usually static, invoking all models for every test sample regardless of sample quality or modality reliability. To overcome these limitations, we propose \textbf{FusionAgent}, a novel agentic framework that leverages a Multimodal Large Language Model (MLLM) to perform dynamic, sample-specific model selection. Each expert model is treated as a tool, and through Reinforcement Fine-Tuning (RFT) with a metric-based reward, the agent learns to adaptively determine the optimal model combination for each test input. To address the model score misalignment and embedding heterogeneity, we introduce Anchor-based Confidence Top-k (ACT) score-fusion, which anchors on the most confident model and integrates complementary predictions in a confidence-aware manner. Extensive experiments on multiple whole-body biometric benchmarks demonstrate that FusionAgent significantly outperforms SoTA methods while achieving higher efficiency through fewer model invocations, underscoring the critical role of dynamic, explainable, and robust model fusion in real-world recognition systems. Project page: \href{https://fusionagent.github.io/}{FusionAgent}.

Method: Introduces Live Interactive Training (LIT) framework where models learn from user corrections at inference time. LIT-LoRA implements this by continually updating lightweight LoRA modules on-the-fly when users provide corrections, allowing the vision system to improve on subsequent video frames.

Result: Achieves 18-34% reduction in total corrections on challenging video segmentation benchmarks with minimal training overhead (~0.5s per correction). Successfully adapts to other segmentation models and extends to CLIP-based fine-grained image classification.

Conclusion: Live adaptation can transform interactive visual tools by significantly reducing redundant human effort in complex visual tasks, demonstrating the promise of online learning from human feedback for vision systems.

Abstract: Interactive video segmentation often requires many user interventions for robust performance in challenging scenarios (e.g., occlusions, object separations, camouflage, etc.). Yet, even state-of-the-art models like SAM2 use corrections only for immediate fixes without learning from this feedback, leading to inefficient, repetitive user effort. To address this, we introduce Live Interactive Training (LIT), a novel framework for prompt-based visual systems where models also learn online from human corrections at inference time. Our primary instantiation, LIT-LoRA, implements this by continually updating a lightweight LoRA module on-the-fly. When a user provides a correction, this module is rapidly trained on that feedback, allowing the vision system to improve performance on subsequent frames of the same video. Leveraging the core principles of LIT, our LIT-LoRA implementation achieves an average 18-34% reduction in total corrections on challenging video segmentation benchmarks, with a negligible training overhead of ~0.5s per correction. We further demonstrate its generality by successfully adapting it to other segmentation models and extending it to CLIP-based fine-grained image classification. Our work highlights the promise of live adaptation to transform interactive tools and significantly reduce redundant human effort in complex visual tasks. Project: https://youngxinyu1802.github.io/projects/LIT/.

Method: Created AVAPrintDB using two audiovisual corpora and three state-of-the-art avatar generators (GAGAvatar, LivePortrait, HunyuanPortrait) with different synthesis paradigms, including both self- and cross-reenactments. Established a standardized benchmark for avatar fingerprinting using public state-of-the-art systems and novel Foundation Model-based methods (DINOv2 and CLIP).

Result: Results show that identity-related motion cues persist across synthetic avatars, but current avatar fingerprinting systems remain highly sensitive to changes in synthesis pipeline and source domain. The database and benchmark protocols are publicly available.

Conclusion: AVAPrintDB provides a comprehensive resource for avatar fingerprinting research, highlighting both the persistence of identity cues in synthetic avatars and the sensitivity of current systems to technical variations, enabling reproducible research in this important security domain.

Abstract: Recent advances in photorealistic avatar generation have enabled highly realistic talking-head avatars, raising security concerns regarding identity impersonation in AI-mediated communication. To advance in this challenging problem, the task of avatar fingerprinting aims to determine whether two avatar videos are driven by the same human operator or not. However, current public databases in the literature are scarce and based solely on old-fashioned talking-head avatar generators, not representing realistic scenarios for the current task of avatar fingerprinting. To overcome this situation, the present article introduces AVAPrintDB, a new publicly available multi-generator talking-head avatar database for avatar fingerprinting. AVAPrintDB is constructed from two audiovisual corpora and three state-of-the-art avatar generators (GAGAvatar, LivePortrait, HunyuanPortrait), representing different synthesis paradigms, and includes both self- and cross-reenactments to simulate legitimate usage and impersonation scenarios. Building on this database, we also define a standardized and reproducible benchmark for avatar fingerprinting, considering public state-of-the-art avatar fingerprinting systems and exploring novel methods based on Foundation Models (DINOv2 and CLIP). Also, we conduct a comprehensive analysis under generator and dataset shift. Our results show that, while identity-related motion cues persist across synthetic avatars, current avatar fingerprinting systems remain highly sensitive to changes in the synthesis pipeline and source domain. The AVAPrintDB, benchmark protocols, and avatar fingerprinting systems are publicly available to facilitate reproducible research.

Method: Three-stage pipeline: 1) Exercise-specific degree-of-freedom selector focuses on salient joints, 2) Structured biomechanical context pairs individualized morphometrics with cycle and constraint analysis, 3) Vision-biomechanics conditioned feedback module uses cross-attention to generate precise text. Uses parameter-efficient training with frozen vision and language backbones.

Result: BioCoach achieves clear gains on QEVD-bio-fit-coach dataset across lexical and judgment metrics while maintaining temporal triggering. On original QEVD-fit-coach, it improves text quality and correctness with near-parity timing, demonstrating that explicit kinematics and constraints are key to accurate coaching.

Conclusion: Explicit integration of kinematics and biomechanical constraints enables accurate, phase-aware fitness coaching. The framework provides transparent, personalized reasoning rather than pattern matching, showing the importance of biomechanical grounding in vision-language systems for fitness applications.

Abstract: We present BioCoach, a biomechanics-grounded vision–language framework for fitness coaching from streaming video. BioCoach fuses visual appearance and 3D skeletal kinematics, through a novel three-stage pipeline: an exercise-specific degree-of-freedom selector that focuses analysis on salient joints; a structured biomechanical context that pairs individualized morphometrics with cycle and constraint analysis; and a vision–biomechanics conditioned feedback module that applies cross-attention to generate precise, actionable text. Using parameter-efficient training that freezes the vision and language backbones, BioCoach yields transparent, personalized reasoning rather than pattern matching. To enable learning and fair evaluation, we augment QEVD-fit-coach with biomechanics-oriented feedback to create QEVD-bio-fit-coach, and we introduce a biomechanics-aware LLM judge metric. BioCoach delivers clear gains on QEVD-bio-fit-coach across lexical and judgment metrics while maintaining temporal triggering; on the original QEVD-fit-coach, it improves text quality and correctness with near-parity timing, demonstrating that explicit kinematics and constraints are key to accurate, phase-aware coaching.

Method: Proposes a robust AGE framework with: 1) Ensemble augmentation techniques (eyeglasses, masks, varied lighting synthesis) to expand image manifold, and 2) Reformulates gaze regression as multi-task learning with multi-view supervised contrastive learning, discretized label classification, and eye-region segmentation as auxiliary objectives.

Result: MobileNet-based lightweight model achieves generalization performance competitive with SOTA UniGaze-H while using <1% of its parameters, enabling high-fidelity real-time gaze tracking on mobile devices. New benchmark datasets created to evaluate robustness under challenging conditions.

Conclusion: The proposed framework effectively addresses generalization challenges in gaze estimation through data augmentation and multi-task learning, enabling robust performance in practical scenarios with minimal computational requirements.

Abstract: Appearance-based gaze estimation (AGE) has achieved remarkable performance in constrained settings, yet we reveal a significant generalization gap where existing AGE models often fail in practical, unconstrained scenarios, particularly those involving facial wearables and poor lighting conditions. We attribute this failure to two core factors: limited image diversity and inconsistent label fidelity across different datasets, especially along the pitch axis. To address these, we propose a robust AGE framework that enhances generalization without requiring additional human-annotated data. First, we expand the image manifold via an ensemble of augmentation techniques, including synthesis of eyeglasses, masks, and varied lighting. Second, to mitigate the impact of anisotropic inter-dataset label deviation, we reformulate gaze regression as a multi-task learning problem, incorporating multi-view supervised contrastive (SupCon) learning, discretized label classification, and eye-region segmentation as auxiliary objectives. To rigorously validate our approach, we curate new benchmark datasets designed to evaluate gaze robustness under challenging conditions, a dimension largely overlooked by existing evaluation protocols. Our MobileNet-based lightweight model achieves generalization performance competitive with the state-of-the-art (SOTA) UniGaze-H, while utilizing less than 1% of its parameters, enabling high-fidelity, real-time gaze tracking on mobile devices.

Method: Developed Raspberry Pi-based portable imaging system to capture RGB+thermal images simultaneously; collected 1,205 samples labeled into 6 stages; trained DenseNet121, EfficientNetV2, InceptionV3, ResNet50, and VGG16 on three datasets: RGB-only, thermal-only, and RGB+Thermal (4 channels)

Result: Multimodal RGB+Thermal dataset outperformed single-modal approaches; VGG16 achieved best performance: 93.25% accuracy, 92.53% F1-score, 91.03% MCC; Grad-CAM showed thermal channel highlighted temperature anomalies while RGB provided structural/textural information

Conclusion: Multimodal imaging combining RGB and thermal data improves DFU stage classification; thermal imaging helps locate temperature anomalies while RGB provides complementary visual information

Abstract: Diabetic foot ulcers (DFU) are one of the serious complications of diabetes that can lead to amputations and high healthcare costs. Regular monitoring and early diagnosis are critical for reducing the clinical burden and the risk of amputation. The aim of this study is to investigate the impact of using multimodal images on deep learning models for the classification of DFU stages. To this end, we developed a Raspberry Pi-based portable imaging system capable of simultaneously capturing RGB and thermal images. Using this prototype, a dataset consisting of 1,205 samples was collected in a hospital setting. The dataset was labeled by experts into six distinct stages. To evaluate the models performance, we prepared three different training sets: RGB-only, thermal-only, and RGB+Thermal (with the thermal image added as a fourth channel). We trained these training sets on the DenseNet121, EfficientNetV2, InceptionV3, ResNet50, and VGG16 models. The results show that the multimodal training dataset, in which RGB and thermal data are combined across four channels, outperforms single-modal approaches. The highest performance was observed in the VGG16 model trained on the RGB+Thermal dataset. The model achieved an accuracy of 93.25%, an F1-score of 92.53%, and an MCC of 91.03%. Grad-CAM heatmap visualizations demonstrated that the thermal channel helped the model focus on the correct location by highlighting temperature anomalies in the ulcer region, while the RGB channel supported the decision-making process with complementary structural and textural information.

Method: 1) Collected new dataset of NIR and visible-light iris images from 259 deceased subjects with largest post-mortem interval of 1,674 hours, including first before/after death case; 2) Combined with public datasets to evaluate 5 iris recognition methods on 338 subjects; 3) Implemented model for detecting post-mortem iris images as presentation attacks; 4) Developed open-source forensic tool integrating three methods with explainability features.

Result: Provides comprehensive dataset and evaluation of state-of-the-art iris recognition methods on post-mortem data, including analysis of demographic factors’ influence on performance. Offers detection model for post-mortem iris images and open-source forensic tool with human-interpretable comparisons.

Conclusion: The paper makes significant contributions to post-mortem iris recognition by providing new datasets, evaluation benchmarks, detection methods, and practical forensic tools with explainability, advancing both research and practical applications in forensic biometrics.

Abstract: Post-mortem iris recognition brings both hope to the forensic community (a short-term but accurate and fast means of verifying identity) as well as concerns to society (its potential illicit use in post-mortem impersonation). These hopes and concerns have grown along with the volume of research in post-mortem iris recognition. Barriers to further progress in post-mortem iris recognition include the difficult nature of data collection, and the resulting small number of approaches designed specifically for comparing iris images of deceased subjects. This paper makes several unique contributions to mitigate these barriers. First, we have collected and we offer a new dataset of NIR (compliant with ISO/IEC 19794-6 where possible) and visible-light iris images collected after demise from 259 subjects, with the largest PMI (post-mortem interval) being 1,674 hours. For one subject, the data has been collected before and after death, the first such case ever published. Second, the collected dataset was combined with publicly-available post-mortem samples to assess the current state of the art in automatic forensic iris recognition with five iris recognition methods and data originating from 338 deceased subjects. These experiments include analyses of how selected demographic factors influence recognition performance. Thirdly, this study implements a model for detecting post-mortem iris images, which can be considered as presentation attacks. Finally, we offer an open-source forensic tool integrating three post-mortem iris recognition methods with explainability elements added to make the comparison process more human-interpretable.

Method: Proposes Energy-Guided Test-Time Transformation (ET3), a training-free defense that minimizes the energy of input samples during inference. The method is theoretically grounded with proofs that the transformation succeeds in classification under reasonable assumptions.

Result: Extensive experiments show ET3 provides strong defense for classifiers, zero-shot classification with CLIP, and boosts robustness of LVLMs in tasks like Image Captioning and Visual Question Answering.

Conclusion: ET3 offers an effective, lightweight defense strategy for enhancing multimodal model robustness against adversarial attacks without requiring retraining, making it practical for real-world deployment.

Abstract: Despite the rapid progress in multimodal models and Large Visual-Language Models (LVLM), they remain highly susceptible to adversarial perturbations, raising serious concerns about their reliability in real-world use. While adversarial training has become the leading paradigm for building models that are robust to adversarial attacks, Test-Time Transformations (TTT) have emerged as a promising strategy to boost robustness at inference.In light of this, we propose Energy-Guided Test-Time Transformation (ET3), a lightweight, training-free defense that enhances the robustness by minimizing the energy of the input samples.Our method is grounded in a theory that proves our transformation succeeds in classification under reasonable assumptions. We present extensive experiments demonstrating that ET3 provides a strong defense for classifiers, zero-shot classification with CLIP, and also for boosting the robustness of LVLMs in tasks such as Image Captioning and Visual Question Answering. Code is available at github.com/OmnAI-Lab/Energy-Guided-Test-Time-Defense .

Method: GUIDED decomposes fine-grained prompts into coarse-grained subjects and descriptive attributes using a language model. It uses subject embeddings for stable localization, an attribute embedding fusion module for selective attribute incorporation, and a region-level attribute discrimination module with refined vision-language alignment.

Result: Extensive experiments on FG-OVD and 3F-OVD benchmarks show GUIDED achieves new state-of-the-art results, demonstrating benefits of disentangled modeling and modular optimization.

Conclusion: The decomposition framework effectively addresses semantic entanglement in fine-grained open-vocabulary detection by separating localization and recognition tasks, with modular optimization improving performance on fine-grained object detection.

Abstract: Fine-grained open-vocabulary object detection (FG-OVD) aims to detect novel object categories described by attribute-rich texts. While existing open-vocabulary detectors show promise at the base-category level, they underperform in fine-grained settings due to the semantic entanglement of subjects and attributes in pretrained vision-language model (VLM) embeddings – leading to over-representation of attributes, mislocalization, and semantic drift in embedding space. We propose GUIDED, a decomposition framework specifically designed to address the semantic entanglement between subjects and attributes in fine-grained prompts. By separating object localization and fine-grained recognition into distinct pathways, HUIDED aligns each subtask with the module best suited for its respective roles. Specifically, given a fine-grained class name, we first use a language model to extract a coarse-grained subject and its descriptive attributes. Then the detector is guided solely by the subject embedding, ensuring stable localization unaffected by irrelevant or overrepresented attributes. To selectively retain helpful attributes, we introduce an attribute embedding fusion module that incorporates attribute information into detection queries in an attention-based manner. This mitigates over-representation while preserving discriminative power. Finally, a region-level attribute discrimination module compares each detected region against full fine-grained class names using a refined vision-language model with a projection head for improved alignment. Extensive experiments on FG-OVD and 3F-OVD benchmarks show that GUIDED achieves new state-of-the-art results, demonstrating the benefits of disentangled modeling and modular optimization. Our code will be released at https://github.com/lijm48/GUIDED.

Method: GG-Langevin uses Geometry-Guided Langevin dynamics that traverses diffusion model trajectories while preserving measurement consistency at every step, combining a data-informed prior with geometric constraints.

Result: Extensive experiments show GG-Langevin achieves higher geometric accuracy and greater robustness to missing data than existing surface reconstruction methods.

Conclusion: The probabilistic approach successfully unifies generative modeling with geometric constraints, enabling robust 3D shape reconstruction from incomplete/noisy observations.

Abstract: Reconstructing complete 3D shapes from incomplete or noisy observations is a fundamentally ill-posed problem that requires balancing measurement consistency with shape plausibility. Existing methods for shape reconstruction can achieve strong geometric fidelity in ideal conditions but fail under realistic conditions with incomplete measurements or noise. At the same time, recent generative models for 3D shapes can synthesize highly realistic and detailed shapes but fail to be consistent with observed measurements. In this work, we introduce GG-Langevin: Geometry-Guided Langevin dynamics, a probabilistic approach that unifies these complementary perspectives. By traversing the trajectories of Langevin dynamics induced by a diffusion model, while preserving measurement consistency at every step, we generatively reconstruct shapes that fit both the measurements and the data-informed prior. We demonstrate through extensive experiments that GG-Langevin achieves higher geometric accuracy and greater robustness to missing data than existing methods for surface reconstruction.

Method: Used custom dataset and COCO2017 dataset, applied image distortions to test robustness, and evaluated various YOLO models across different training/testing configurations.

Result: Experimental results comparing different YOLO versions and variants under various conditions, providing insights into model performance for robotic vision applications.

Conclusion: The study provides guidance on selecting appropriate YOLO models for robotic vision tasks based on performance under different conditions and robustness to image distortions.

Abstract: YOLO object detectors recently became a key component of vision systems in many domains. The family of available YOLO models consists of multiple versions, each in various variants. The research reported in this paper aims to validate the applicability of members of this family to detect objects located within the robot workspace. In our experiments, we used our custom dataset and the COCO2017 dataset. To test the robustness of investigated detectors, the images of these datasets were subject to distortions. The results of our experiments, including variations of training/testing configurations and models, may support the choice of the appropriate YOLO version for robotic vision tasks.

Method: Proposes RealBirdID benchmark with curated unanswerable examples labeled with specific rationales (“requires vocalization,” “low quality image,” “view obstructed”) paired with answerable instances. Evaluates open-source and proprietary models on both classification accuracy and abstention calibration.

Result: (1) Species identification on answerable set is challenging (<13% accuracy for MLLMs including GPT-5 and Gemini-2.5 Pro); (2) Better classification ability doesn’t guarantee better abstention calibration; (3) MLLMs generally fail at providing correct reasons even when they do abstain.

Conclusion: RealBirdID establishes a focused target for abstention-aware fine-grained recognition and provides a recipe for measuring progress in multimodal systems’ ability to recognize when they lack sufficient information.

Abstract: Fine-grained bird species identification in the wild is frequently unanswerable from a single image: key cues may be non-visual (e.g. vocalization), or obscured due to occlusion, camera angle, or low resolution. Yet today’s multimodal systems are typically judged on answerable, in-schema cases, encouraging confident guesses rather than principled abstention. We propose the RealBirdID benchmark: given an image of a bird, a system should either answer with a species or abstain with a concrete, evidence-based rationale: “requires vocalization,” “low quality image,” or “view obstructed”. For each genus, the dataset includes a validation split composed of curated unanswerable examples with labeled rationales, paired with a companion set of clearly answerable instances. We find that (1) the species identification on the answerable set is challenging for a variety of open-source and proprietary models (less than 13% accuracy for MLLMs including GPT-5 and Gemini-2.5 Pro), (2) models with greater classification ability are not necessarily more calibrated to abstain from unanswerable examples, and (3) that MLLMs generally fail at providing correct reasons even when they do abstain. RealBirdID establishes a focused target for abstention-aware fine-grained recognition and a recipe for measuring progress.

Method: Proposes Unified Motion Flow (UMF) with two components: Pyramid Motion Flow (P-Flow) for hierarchical motion prior generation, and Semi-Noise Motion Flow (S-Flow) for reaction generation. Uses unified latent space to bridge heterogeneous motion datasets.

Result: Extensive results and user studies demonstrate UMF’s effectiveness as a generalist model for multi-person motion generation from text.

Conclusion: UMF provides an effective solution for variable-agent motion generation by decomposing the problem into motion prior and reaction generation stages, addressing computational efficiency and error accumulation issues.

Abstract: Generative models excel at motion synthesis for a fixed number of agents but struggle to generalize with variable agents. Based on limited, domain-specific data, existing methods employ autoregressive models to generate motion recursively, which suffer from inefficiency and error accumulation. We propose Unified Motion Flow (UMF), which consists of Pyramid Motion Flow (P-Flow) and Semi-Noise Motion Flow (S-Flow). UMF decomposes the number-free motion generation into a single-pass motion prior generation stage and multi-pass reaction generation stages. Specifically, UMF utilizes a unified latent space to bridge the distribution gap between heterogeneous motion datasets, enabling effective unified training. For motion prior generation, P-Flow operates on hierarchical resolutions conditioned on different noise levels, thereby mitigating computational overheads. For reaction generation, S-Flow learns a joint probabilistic path that adaptively performs reaction transformation and context reconstruction, alleviating error accumulation. Extensive results and user studies demonstrate UMF’ s effectiveness as a generalist model for multi-person motion generation from text. Project page: https://githubhgh.github.io/umf/.

Method: Two-stage approach: Stage 1 pretrains a vision-only slide encoder on 77,134 public slide feature grids using masked self-distillation. Stage 2 aligns representations with clinical semantics using a case transformer and multi-task supervision over 333 tasks from 56 public datasets (205 classification, 128 survival tasks).

Result: Achieves best or tied-best performance on most metrics across eight held-out tasks, improving macro averages over TITAN by +7.37%, +5.50%, +7.83% and over PRISM by +8.83%, +10.70%, +9.78% for weighted F1, weighted ROC-AUC, and balanced accuracy. Parameter efficient with 85.77M parameters (14x smaller than GigaPath).

Conclusion: Open, reproducible patient-level pretraining yields transferable embeddings, providing a practical path toward scalable patient-first histopathology foundation models that better capture clinical context across multiple slides.

Abstract: Computational pathology needs whole-slide image (WSI) foundation models that transfer across diverse clinical tasks, yet current approaches remain largely slide-centric, often depend on private data and expensive paired-report supervision, and do not explicitly model relationships among multiple slides from the same patient. We present MOOZY, a patient-first pathology foundation model in which the patient case, not the individual slide, is the core unit of representation. MOOZY explicitly models dependencies across all slides from the same patient via a case transformer during pretraining, combining multi-stage open self-supervision with scaled low-cost task supervision. In Stage 1, we pretrain a vision-only slide encoder on 77,134 public slide feature grids using masked self-distillation. In Stage 2, we align these representations with clinical semantics using a case transformer and multi-task supervision over 333 tasks from 56 public datasets, including 205 classification and 128 survival tasks across four endpoints. Across eight held-out tasks with five-fold frozen-feature probe evaluation, MOOZY achieves best or tied-best performance on most metrics and improves macro averages over TITAN by +7.37%, +5.50%, and +7.83% and over PRISM by +8.83%, +10.70%, and +9.78% for weighted F1, weighted ROC-AUC, and balanced accuracy, respectively. MOOZY is also parameter efficient with 85.77M parameters, 14x smaller than GigaPath. These results demonstrate that open, reproducible patient-level pretraining yields transferable embeddings, providing a practical path toward scalable patient-first histopathology foundation models.

Method: Proposes MonoIA framework that uses large language models and vision-language models to generate intrinsic embeddings encoding visual and geometric implications of camera parameters. These embeddings are integrated via an Intrinsic Adaptation Module to modulate feature representations according to camera-specific configurations.

Result: Achieves new state-of-the-art results on standard benchmarks including KITTI, Waymo, and nuScenes (+1.18% on KITTI leaderboard), with further improvements under multi-dataset training (+4.46% on KITTI Val).

Conclusion: Shifts intrinsic modeling from numeric conditioning to semantic representation, enabling robust and unified 3D perception across different camera configurations through language-grounded adaptation.

Abstract: Monocular 3D object detection (Mono3D) aims to infer object locations and dimensions in 3D space from a single RGB image. Despite recent progress, existing methods remain highly sensitive to camera intrinsics and struggle to generalize across diverse settings, since intrinsics govern how 3D scenes are projected onto the image plane. We propose MonoIA, a unified intrinsic-aware framework that models and adapts to intrinsic variation through a language-grounded representation. The key insight is that intrinsic variation is not a numeric difference but a perceptual transformation that alters apparent scale, perspective, and spatial geometry. To capture this effect, MonoIA employs large language models and vision-language models to generate intrinsic embeddings that encode the visual and geometric implications of camera parameters. These embeddings are hierarchically integrated into the detection network via an Intrinsic Adaptation Module, allowing the model to modulate its feature representations according to camera-specific configurations and maintain consistent 3D detection across intrinsics. This shifts intrinsic modeling from numeric conditioning to semantic representation, enabling robust and unified perception across cameras. Extensive experiments show that MonoIA achieves new state-of-the-art results on standard benchmarks including KITTI, Waymo, and nuScenes (e.g., +1.18% on the KITTI leaderboard), and further improves performance under multi-dataset training (e.g., +4.46% on KITTI Val).

Method: Proposes VIRST with Spatio-Temporal Fusion (STF) to fuse segmentation-aware video features into vision-language backbone, and Temporal Dynamic Anchor Updater to maintain temporally adjacent anchor frames for stable temporal cues under large motion, occlusion, and reappearance.

Result: Achieves state-of-the-art results across diverse RVOS benchmarks under realistic and challenging conditions, demonstrating strong generalization to both referring and reasoning oriented settings.

Conclusion: VIRST’s unified design effectively addresses limitations of previous approaches by integrating global video reasoning with pixel-level segmentation in a single model framework.

Abstract: Referring Video Object Segmentation (RVOS) aims to segment target objects in videos based on natural language descriptions. However, fixed keyframe-based approaches that couple a vision language model with a separate propagation module often fail to capture rapidly changing spatiotemporal dynamics and to handle queries requiring multi-step reasoning, leading to sharp performance drops on motion-intensive and reasoning-oriented videos beyond static RVOS benchmarks. To address these limitations, we propose VIRST (Video-Instructed Reasoning Assistant for Spatio-Temporal Segmentation), an end-to-end framework that unifies global video reasoning and pixel-level mask prediction within a single model. VIRST bridges semantic and segmentation representations through the Spatio-Temporal Fusion (STF), which fuses segmentation-aware video features into the vision-language backbone, and employs the Temporal Dynamic Anchor Updater to maintain temporally adjacent anchor frames that provide stable temporal cues under large motion, occlusion, and reappearance. This unified design achieves state-of-the-art results across diverse RVOS benchmarks under realistic and challenging conditions, demonstrating strong generalization to both referring and reasoning oriented settings. The code and checkpoints are available at https://github.com/AIDASLab/VIRST.

Method: Created ChartNet using a novel code-guided synthesis pipeline that generates diverse chart samples from 6 plotting libraries. Includes rigorous quality-filtering for visual fidelity and semantic accuracy. Features five aligned components per sample and specialized subsets for human annotation, real-world data, safety, and grounding.

Result: Produced 1.5 million chart samples spanning 24 chart types. Fine-tuning on ChartNet consistently improves performance across benchmarks, demonstrating its utility as large-scale supervision for multimodal models.

Conclusion: ChartNet is the largest open-source dataset for chart understanding, supporting development of foundation models with robust capabilities for data visualization interpretation through fine-grained cross-modal alignment.

Abstract: Understanding charts requires models to jointly reason over geometric visual patterns, structured numerical data, and natural language – a capability where current vision-language models (VLMs) remain limited. We introduce ChartNet, a high-quality, million-scale multimodal dataset designed to advance chart interpretation and reasoning. ChartNet leverages a novel code-guided synthesis pipeline to generate 1.5 million diverse chart samples spanning 24 chart types and 6 plotting libraries. Each sample consists of five aligned components: plotting code, rendered chart image, data table, natural language summary, and question-answering with reasoning, providing fine-grained cross-modal alignment. To capture the full spectrum of chart comprehension, ChartNet additionally includes specialized subsets encompassing human annotated data, real-world data, safety, and grounding. Moreover, a rigorous quality-filtering pipeline ensures visual fidelity, semantic accuracy, and diversity across chart representations. Fine-tuning on ChartNet consistently improves results across benchmarks, demonstrating its utility as large-scale supervision for multimodal models. As the largest open-source dataset of its kind, ChartNet aims to support the development of foundation models with robust and generalizable capabilities for data visualization understanding. The dataset is publicly available at https://huggingface.co/datasets/ibm-granite/ChartNet

Method: Represents each layer as a within-layer correlation graph derived from neuron-neuron co-activations. Analyzes whether population-level structure is behaviorally meaningful, how it changes across modalities and depth, and whether it identifies causally influential internal components under intervention.

Result: Correlation topology carries recoverable behavioral signal. Cross-modal structure progressively consolidates with depth around a compact set of recurrent hub neurons, whose targeted perturbation substantially alters model output.

Conclusion: Neural topology emerges as a meaningful intermediate scale for VLM interpretability: richer than local attribution, more tractable than full circuit recovery, and empirically tied to multimodal behavior.

Abstract: Vision-language models (VLMs) achieve strong multimodal performance, yet how computation is organized across populations of neurons remains poorly understood. In this work, we study VLMs through the lens of neural topology, representing each layer as a within-layer correlation graph derived from neuron-neuron co-activations. This view allows us to ask whether population-level structure is behaviorally meaningful, how it changes across modalities and depth, and whether it identifies causally influential internal components under intervention. We show that correlation topology carries recoverable behavioral signal; moreover, cross-modal structure progressively consolidates with depth around a compact set of recurrent hub neurons, whose targeted perturbation substantially alters model output. Neural topology thus emerges as a meaningful intermediate scale for VLM interpretability: richer than local attribution, more tractable than full circuit recovery, and empirically tied to multimodal behavior. Code is publicly available at https://github.com/he-h/vlm-graph-probing.

Method: Combines pre-trained diffusion models: image relighting model processes frames individually, then video diffusion prior ensures temporal consistency. Introduces Light Map Injection module for illumination coherence and Geometry-Aware Relighting module to suppress original lighting influence.

Result: Produces high-quality videos with diverse illumination changes that closely follow specified light trajectories, demonstrating improved controllability over baseline methods.

Conclusion: LightCtrl enables training-free controllable video relighting with explicit illumination control through user-supplied light trajectories, advancing video editing capabilities.

Abstract: Recent diffusion models have achieved remarkable success in image relighting, and this success has quickly been extended to video relighting. However, existing methods offer limited explicit control over illumination in the relighted output. We present LightCtrl, the first controllable video relighting method that enables explicit control of video illumination through a user-supplied light trajectory in a training-free manner. Our approach combines pre-trained diffusion models: an image relighting model processes each frame individually, followed by a video diffusion prior to enhance temporal consistency. To achieve explicit control over dynamically varying lighting, we introduce two key components. First, a Light Map Injection module samples light trajectory-specific noise and injects it into the latent representation of the source video, improving illumination coherence with the conditional light trajectory. Second, a Geometry-Aware Relighting module dynamically combines RGB and normal map latents in the frequency domain to suppress the influence of the original lighting, further enhancing adherence to the input light trajectory. Experiments show that LightCtrl produces high-quality videos with diverse illumination changes that closely follow the specified light trajectory, demonstrating improved controllability over baseline methods. Code is available at: https://github.com/GVCLab/LightCtrl.

Method: SceneExpander uses test-time adaptation on a parametric feed-forward 3D reconstruction model with two complementary distillation signals: anchor distillation stabilizes the original scene by distilling geometric cues from captured views, and inserted-view self-distillation preserves observation-supported predictions while adapting latent geometry and appearance to accommodate misaligned inserted views.

Result: Experiments on ETH scenes and online data demonstrate improved expansion behavior and reconstruction quality under misalignment compared to existing approaches.

Conclusion: The proposed method effectively addresses the challenge of 3D scene expansion with misaligned inserted views, enabling more robust iterative scene creation workflows while maintaining multi-view consistency.

Abstract: World building with 3D scene representations is increasingly important for content creation, simulation, and interactive experiences, yet real workflows are inherently iterative: creators must repeatedly extend an existing scene under user control. Motivated by this research gap, we study 3D scene expansion in a user-centric workflow: starting from a real scene captured by multi-view images, we extend its coverage by inserting an additional view synthesized by a generative model. Unlike simple object editing or style transfer in a fixed scene, the inserted view is often 3D-misaligned with the original reconstruction, introducing geometry shifts, hallucinated content, or view-dependent artifacts that break global multi-view consistency. To address the challenge, we propose SceneExpander, which applies test-time adaptation to a parametric feed-forward 3D reconstruction model with two complementary distillation signals: anchor distillation stabilizes the original scene by distilling geometric cues from the captured views, while inserted-view self-distillation preserves observation-supported predictions yet adapts latent geometry and appearance to accommodate the misaligned inserted view. Experiments on ETH scenes and online data demonstrate improved expansion behavior and reconstruction quality under misalignment.

Method: Proposes EFlow with: 1) Gated Local-Global Attention for efficient token-droppable hybrid blocks, 2) Path-Drop Guided training to replace expensive guidance targets, and 3) Mean-Velocity Additivity regularizer for high fidelity at low step counts.

Result: Achieves 2.5x higher training throughput over standard solution-flow and 45.3x lower inference latency than standard iterative models with competitive performance on Kinetics and large-scale text-to-video datasets.

Conclusion: EFlow enables practical from-scratch training of video diffusion transformers by simultaneously addressing attention complexity and sampling step bottlenecks through efficient architectural and training innovations.

Abstract: Scaling video diffusion transformers is fundamentally bottlenecked by two compounding costs: the expensive quadratic complexity of attention per step, and the iterative sampling steps. In this work, we propose EFlow, an efficient few-step training framework, that tackles these bottlenecks simultaneously. To reduce sampling steps, we build on a solution-flow objective that learns a function mapping a noised state at time t to time s. Making this formulation computationally feasible and high-quality at video scale, however, demands two complementary innovations. First, we propose Gated Local-Global Attention, a token-droppable hybrid block which is efficient, expressive, and remains highly stable under aggressive random token-dropping, substantially reducing per-step compute. Second, we develop an efficient few-step training recipe. We propose Path-Drop Guided training to replace the expensive guidance target with a computationally cheap, weak path. Furthermore, we augment this with a Mean-Velocity Additivity regularizer to ensure high fidelity at extremely low step counts. Together, our EFlow enables a practical from-scratch training pipeline, achieving up to 2.5x higher training throughput over standard solution-flow, and 45.3x lower inference latency than standard iterative models with competitive performance on Kinetics and large-scale text-to-video datasets.

Method: Conducted first systematic evaluation of 18 segmentation and geospatial foundation models on the Fields of The World benchmark. Proposed a new segmentation approach combining U-Net backbone, composite loss functions, and targeted data augmentations to enhance performance and robustness under real-world conditions.

Result: U-Net semantic segmentation model outperformed instance-based and GFM alternatives on performance and deployment metrics. Achieved 76% IoU and 47% object-F1 on FTW, representing 6% and 9% improvements over previous baseline. Released models and model-derived field boundary datasets for five countries.

Conclusion: The proposed approach provides a practical framework for reliable, scalable, and reproducible field boundary delineation across model design, training, and inference. U-Net-based segmentation offers superior performance for agricultural monitoring tasks compared to more complex alternatives.

Abstract: Large-scale maps of field boundaries are essential for agricultural monitoring tasks. Existing deep learning approaches for satellite-based field mapping are sensitive to illumination, spatial scale, and changes in geographic location. We conduct the first systematic evaluation of segmentation and geospatial foundation models (GFMs) for global field boundary delineation using the Fields of The World (FTW) benchmark. We evaluate 18 models under unified experimental settings, showing that a U-Net semantic segmentation model outperforms instance-based and GFM alternatives on a suite of performance and deployment metrics. We propose a new segmentation approach that combines a U-Net backbone, composite loss functions, and targeted data augmentations to enhance performance and robustness under real-world conditions. Our model achieves a 76% IoU and 47% object-F1 on FTW, an increase of 6% and 9% over the previous baseline. Our approach provides a practical framework for reliable, scalable, and reproducible field boundary delineation across model design, training, and inference. We release all models and model-derived field boundary datasets for five countries.

Method: Proposes HocSLM with: 1) HGLNet for hierarchical global-local spatio-temporal modeling, 2) VLM-generated textual descriptions from RGB videos, and 3) skeleton-language sequential fusion module for aligning skeletal features with textual descriptions in unified semantic space.

Result: Achieves state-of-the-art performance on three mainstream benchmark datasets: NTU RGB+D 60, NTU RGB+D 120, and Northwestern-UCLA.

Conclusion: The proposed HocSLM framework effectively enhances semantic discrimination and cross-modal understanding in skeleton-based action recognition by integrating hierarchical spatio-temporal modeling with vision-language semantic alignment.

Abstract: Skeleton-based human action recognition has achieved remarkable progress in recent years. However, most existing GCN-based methods rely on short-range motion topologies, which not only struggle to capture long-range joint dependencies and complex temporal dynamics but also limit cross-modal semantic alignment and understanding due to insufficient modeling of action semantics. To address these challenges, we propose a hierarchical global-local skeleton-language model (HocSLM), enabling the large action model be more representative of action semantics. First, we design a hierarchical global-local network (HGLNet) that consists of a composite-topology spatial module and a dual-path hierarchical temporal module. By synergistically integrating multi-level global and local modules, HGLNet achieves dynamically collaborative modeling at both global and local scales while preserving prior knowledge of human physical structure, significantly enhancing the model’s representation of complex spatio-temporal relationships. Then, a large vision-language model (VLM) is employed to generate textual descriptions by passing the original RGB video sequences to this model, providing the rich action semantics for further training the skeleton-language model. Furthermore, we introduce a skeleton-language sequential fusion module by combining the features from HGLNet and the generated descriptions, which utilizes a skeleton-language model (SLM) for aligning skeletal spatio-temporal features and textual action descriptions precisely within a unified semantic space. The SLM model could significantly enhance the HGLNet’s semantic discrimination capabilities and cross-modal understanding abilities. Extensive experiments demonstrate that the proposed HocSLM achieves the state-of-the-art performance on three mainstream benchmark datasets: NTU RGB+D 60, NTU RGB+D 120, and Northwestern-UCLA.

Method: Decouples metric depth estimation into relative depth prediction and spatially varying scale estimation. Uses a lightweight Depth-Guided Scale Estimation module that upsamples coarse scale maps using relative depth as guidance. Introduces RoPE-φ, a distortion-aware positional embedding that respects spatial warping in Equi-Rectangular Projections via latitude-aware weighting.

Result: Achieves state-of-the-art in cross-camera generalization, consistently outperforming prior methods across all datasets for diverse camera types.

Conclusion: UniDAC provides a universal framework for monocular metric depth estimation that generalizes robustly across diverse camera domains using a single model, addressing key limitations in camera-type generalization.

Abstract: Monocular metric depth estimation (MMDE) is a core challenge in computer vision, playing a pivotal role in real-world applications that demand accurate spatial understanding. Although prior works have shown promising zero-shot performance in MMDE, they often struggle with generalization across diverse camera types, such as fisheye and $360^\circ$ cameras. Recent advances have addressed this through unified camera representations or canonical representation spaces, but they require either including large-FoV camera data during training or separately trained models for different domains. We propose UniDAC, an MMDE framework that presents universal robustness in all domains and generalizes across diverse cameras using a single model. We achieve this by decoupling metric depth estimation into relative depth prediction and spatially varying scale estimation, enabling robust performance across different domains. We propose a lightweight Depth-Guided Scale Estimation module that upsamples a coarse scale map to high resolution using the relative depth map as guidance to account for local scale variations. Furthermore, we introduce RoPE-$φ$, a distortion-aware positional embedding that respects the spatial warping in Equi-Rectangular Projections (ERP) via latitude-aware weighting. UniDAC achieves state of the art (SoTA) in cross-camera generalization by consistently outperforming prior methods across all datasets.

Method: MotiMem uses lightweight 2D Motion Propagation to dynamically identify Regions of Interest (RoI) by exploiting temporal coherence. It combines this with a Hybrid Sparsity-Aware Coding scheme that uses adaptive inversion and truncation to induce bit-level sparsity.

Result: Extensive experiments across nuScenes, Waymo, and KITTI datasets with 16 detection models show MotiMem reduces memory-interface dynamic energy by approximately 43% while retaining approximately 93% of object detection accuracy, establishing a superior Pareto frontier compared to standard codecs like JPEG and WebP.

Conclusion: MotiMem provides an effective hardware-software co-designed solution for reducing memory energy consumption in autonomous perception systems while maintaining high detection accuracy, addressing the critical memory wall problem in battery-constrained electric vehicles.

Abstract: High-resolution sensors are critical for robust autonomous perception but impose a severe memory wall on battery-constrained electric vehicles. In these systems, data movement energy often outweighs computation. Traditional image compression is ill-suited as it is semantically blind and optimizes for storage rather than bus switching activity. We propose MotiMem, a hardware-software co-designed interface. Exploiting temporal coherence,MotiMem uses lightweight 2D Motion Propagation to dynamically identify Regions of Interest (RoI). Complementing this, a Hybrid Sparsity-Aware Coding scheme leverages adaptive inversion and truncation to induce bitlevel sparsity. Extensive experiments across nuScenes, Waymo, and KITTI with 16 detection models demonstrate that MotiMem reduces memory-interface dynamic energy by approximately 43 percent while retaining approximately 93 percent of the object detection accuracy, establishing a new Pareto frontier significantly superior to standard codecs like JPEG and WebP.

Method: Proposes RailVQA-bench (20K single-frame + 1,168 video QA pairs) for evaluating visual cognition, and RailVQA-CoM - a collaborative framework combining small models’ efficiency with large models’ cognition via transparent three-module architecture and adaptive temporal sampling.

Result: The approach improves performance, enhances interpretability, reduces inference latency, strengthens cross-domain generalization, and enables plug-and-play deployment in autonomous driving systems.

Conclusion: RailVQA addresses critical gaps in ATO visual cognition by providing both evaluation benchmarks and an efficient collaborative framework that balances computational efficiency with advanced reasoning capabilities.

Abstract: Automatic Train Operation (ATO) relies on low-latency, reliable cab-view visual perception and decision-oriented inference to ensure safe operation in complex and dynamic railway environments. However, existing approaches focus primarily on basic perception and often generalize poorly to rare yet safety-critical corner cases. They also lack the high-level reasoning and planning capabilities required for operational decision-making. Although recent Large Multi-modal Models (LMMs) show strong generalization and cognitive capabilities, their use in safety-critical ATO is hindered by high computational cost and hallucination risk. Meanwhile, reliable domain-specific benchmarks for systematically evaluating cognitive capabilities are still lacking. To address these gaps, we introduce RailVQA-bench, the first VQA benchmark for cab-view visual cognition in ATO, comprising 20,000 single-frame and 1,168 video based QA pairs to evaluate cognitive generalization and interpretability in both static and dynamic scenarios. Furthermore, we propose RailVQA-CoM, a collaborative large-small model framework that combines small-model efficiency with large-model cognition via a transparent three-module architecture and adaptive temporal sampling, improving perceptual generalization and enabling efficient reasoning and planning. Experiments demonstrate that the proposed approach substantially improves performance, enhances interpretability, reduces inference latency, and strengthens cross-domain generalization, while enabling plug-and-play deployment in autonomous driving systems. Code and datasets will be available at https://github.com/Cybereye-bjtu/RailVQA.

Method: Proposes Speculative Jacobi Decoding with Verification Prediction (SJD-VP) that leverages changes in token probabilities across iterations to guide sampling. The method favors tokens whose probabilities increase, effectively predicting which tokens will pass verification. It’s designed as a plug-and-play module that integrates with existing SJD methods.

Result: Extensive experiments on standard benchmarks show SJD-VP consistently accelerates autoregressive decoding while improving image generation quality.

Conclusion: SJD-VP effectively addresses the low acceptance rate problem in speculative Jacobi decoding by exploiting iterative probability dynamics, resulting in faster and higher-quality image generation.

Abstract: Speculative Jacobi Decoding (SJD) has emerged as a promising method for accelerating autoregressive image generation. Despite its potential, existing SJD approaches often suffer from the low acceptance rate issue of speculative tokens due to token selection ambiguity. Recent works attempt to mitigate this issue primarily from the relaxed token verification perspective but fail to fully exploit the iterative dynamics of decoding. In this paper, we conduct an in-depth analysis and make a novel observation that tokens whose probabilities increase are more likely to match the verification-accepted and correct token. Based on this, we propose a novel Speculative Jacobi Decoding with Verification Prediction (SJD-VP). The key idea is to leverage the change in token probabilities across iterations to guide sampling, favoring tokens whose probabilities increase. This effectively predicts which tokens are likely to pass subsequent verification, boosting the acceptance rate. In particular, our SJD-VP is plug-and-play and can be seamlessly integrated into existing SJD methods. Extensive experiments on standard benchmarks demonstrate that our SJD-VP method consistently accelerates autoregressive decoding while improving image generation quality.

Method: GRACE uses adaptive weight perturbations scaled by local curvature to promote flatter minima in parameter space, combined with a feature alignment loss that maintains representation consistency across clean, adversarial, and OOD inputs. Grounded in Robust PAC-Bayes theory.

Result: On ImageNet fine-tuning of CLIP models, GRACE simultaneously improves ID accuracy by 10.8%, adversarial accuracy by 13.5% while maintaining 57.0% OOD accuracy (vs. 57.4% zero-shot baseline). Geometric analysis confirms flatter minima without feature distortion.

Conclusion: GRACE provides a principled approach to achieving generalized robustness in foundation VLMs by addressing both parameter-space curvature and feature-space invariance, resolving the three-way trade-off that plagues existing fine-tuning methods.

Abstract: Fine-tuning approaches for Vision-Language Models (VLMs) face a critical three-way trade-off between In-Distribution (ID) accuracy, Out-of-Distribution (OOD) generalization, and adversarial robustness. Existing robust fine-tuning strategies resolve at most two axes of this trade-off. Generalization-preserving methods retain ID/OOD performance but leave models vulnerable to adversarial attacks, while adversarial training improves robustness to targeted attacks but degrades ID/OOD accuracy. Our key insight is that the robustness trade-off stems from two geometric failures: sharp, anisotropic minima in parameter space and unstable feature representations that deform under perturbation. To address this, we propose GRACE (Gram-aligned Robustness via Adaptive Curvature Estimation), a unified fine-tuning framework that jointly regularizes the parameter-space curvature and feature-space invariance for VLMs. Grounded in Robust PAC-Bayes theory, GRACE employs adaptive weight perturbations scaled by local curvature to promote flatter minima, combined with a feature alignment loss that maintains representation consistency across clean, adversarial, and OOD inputs. On ImageNet fine-tuning of CLIP models, GRACE simultaneously improves ID accuracy by 10.8%, and adversarial accuracy by 13.5% while maintaining 57.0% OOD accuracy (vs. 57.4% zero-shot baseline). Geometric analysis confirms that GRACE converges to flatter minima without feature distortion across distribution shifts, providing a principled step toward generalized robustness in foundation VLMs.

Method: Multi-task network cascading transducer pose scoring module and uncertainty-aware left ventricular landmark detector with automated LVEF estimation, trained and inferred using only image data without requiring transducer position tracking hardware.

Result: Network successfully determines transducer pose quality (on target, close to target, or far from target) based on images alone while generating visual landmark cues for anatomical guidance, evaluated on point-of-care TTE data acquired with dense sweep protocol.

Conclusion: Demonstrates promising strategy for A4CH view acquisition guidance that could be useful for deploying point-of-care TTE in limited resource settings without requiring cumbersome tracking setups.

Abstract: Point-of-care transthoracic echocardiography (TTE) makes it possible to assess a patient’s cardiac function in almost any setting. A critical step in the TTE exam is acquisition of the apical 4-chamber (A4CH) view, which is used to evaluate clinically impactful measurements such as left ventricular ejection fraction (LVEF). However, optimizing transducer pose for high-quality image acquisition and subsequent measurement is a challenging task, particularly for novice users. In this work, we present a multi-task network that provides feedback cues for A4CH view acquisition and automatically estimates LVEF in high-quality A4CH images. The network cascades a transducer pose scoring module and an uncertainty-aware LV landmark detector with automated LVEF estimation. A strength is that network training and inference do not require cumbersome or costly setups for transducer position tracking. We evaluate performance on point-of-care TTE data acquired with a spatially dense “sweep” protocol around the optimal A4CH view. The results demonstrate the network’s ability to determine when the transducer pose is on target, close to target, or far from target based on the images alone, while generating visual landmark cues that guide anatomical interpretation and orientation. In conclusion, we demonstrate a promising strategy to provide guidance for A4CH view acquisition, which may be useful when deploying point-of-care TTE in limited resource settings.

Method: Train a rotation invariant weakly convex ridge regularizer (WCRR) for variational reconstruction. The approach is benchmarked against state-of-the-art methods on retrospectively simulated data and prospective GoLF SPARKLING and CAIPIRINHA acquisitions.

Result: WCRR consistently outperforms widely used baselines and achieves performance comparable to Plug and Play reconstruction with state-of-the-art 3D DRUNet denoiser, while offering substantially improved computational efficiency and robustness to acquisition changes.

Conclusion: WCRR unifies the strengths of principled variational methods and modern deep learning based approaches for accelerated MRI reconstruction.

Abstract: While highly accelerated non-Cartesian acquisition protocols significantly reduce scan time, they often entail long reconstruction delays. Deep learning based reconstruction methods can alleviate this, but often lack stability and robustness to distribution shifts. As an alternative, we train a rotation invariant weakly convex ridge regularizer (WCRR). The resulting variational reconstruction approach is benchmarked against state of the art methods on retrospectively simulated data and (out of distribution) on prospective GoLF SPARKLING and CAIPIRINHA acquisitions. Our approach consistently outperforms widely used baselines and achieves performance comparable to Plug and Play reconstruction with a state of the art 3D DRUNet denoiser, while offering substantially improved computational efficiency and robustness to acquisition changes. In summary, WCRR unifies the strengths of principled variational methods and modern deep learning based approaches.

Method: Proposes RiskProp, a collision-anchored self-supervised risk propagation paradigm that uses only reliably annotated collision frames. It employs two observation-driven losses: 1) future-frame regularization loss that uses next-frame predictions as soft targets for current frame supervision, enabling backward risk propagation; 2) adaptive monotonic constraint to encourage non-decreasing risk progression over time based on empirical trends.

Result: Experiments on CAP and Nexar datasets demonstrate state-of-the-art performance. RiskProp produces smoother, more discriminative risk curves, improving both early anticipation accuracy and interpretability compared to existing methods.

Conclusion: RiskProp successfully eliminates the need for subjective anomaly onset annotations while achieving superior accident anticipation performance through self-supervised risk propagation from collision frames, offering more reliable and interpretable risk estimation.

Abstract: Accident anticipation aims to predict impending collisions from dashcam videos and trigger early alerts. Existing methods rely on binary supervision with manually annotated “anomaly onset” frames, which are subjective and inconsistent, leading to inaccurate risk estimation. In contrast, we propose RiskProp, a novel collision-anchored self-supervised risk propagation paradigm for early accident anticipation, which removes the need for anomaly onset annotations and leverages only the reliably annotated collision frame. RiskProp models temporal risk evolution through two observation-driven losses: first, since future frames contain more definitive evidence of an impending accident, we introduce a future-frame regularization loss that uses the model’s next-frame prediction as a soft target to supervise the current frame, enabling backward propagation of risk signals; second, inspired by the empirical trend of rising risk before accidents, we design an adaptive monotonic constraint to encourage a non-decreasing progression over time. Experiments on CAP and Nexar demonstrate that RiskProp achieves state-of-the-art performance and produces smoother, more discriminative risk curves, improving both early anticipation and interpretability.

Method: Proposes MultiLoc with joint fusion of multiple reference views and their camera poses in a single forward pass. Also introduces a co-visibility-driven retrieval strategy for geometrically relevant reference view selection to supply informative context.

Result: Establishes new benchmark in visual re-localization, consistently outperforming SOTA RPR methods across WaySpots, Cambridge Landmarks, and Indoor6 datasets. Also shows SOTA performance in relative pose estimation on MegaDepth-1500, ScanNet-1500, and ACID benchmarks.

Conclusion: MultiLoc demonstrates robust domain generalization across indoor, outdoor and natural environments, showing that multi-view fusion significantly enhances relative pose regression capabilities while maintaining real-time efficiency.

Abstract: Relative Pose Regression (RPR) generalizes well to unseen environments, but its performance is often limited due to pairwise and local spatial views. To this end, we propose MultiLoc, a novel multi-view guided RPR model trained at scale, equipping relative pose regression with globally consistent spatial and geometric understanding. Specifically, our method jointly fuses multiple reference views and their associated camera poses in a single forward pass, enabling accurate zero-shot pose estimation with real-time efficiency. To reliably supply informative context, we further propose a co-visibility-driven retrieval strategy for geometrically relevant reference view selection. MultiLoc establishes a new benchmark in visual re-localization, consistently outperforming existing state-of-the-art (SOTA) relative pose regression (RPR) methods across diverse datasets, including WaySpots, Cambridge Landmarks, and Indoor6. Furthermore, MultiLoc’s pose regressor exhibits SOTA performance in relative pose estimation, surpassing RPR, feature matching and non-regression-based techniques on the MegaDepth-1500, ScanNet-1500, and ACID benchmarks. These results demonstrate robust domain generalization of MultiLoc across indoor, outdoor and natural environments. Code will be made publicly available.

Method: A stage-wise framework with: 1) learnable anomaly-aware tokens () to focus on abnormal regions, 2) inter-image difference tokens () to encode temporal changes between studies, and 3) a dedicated explainability stage to generate heatmaps highlighting lesion-related regions.

Result: Achieves state-of-the-art results in anomaly detection, symptom tracking, and anomaly segmentation compared to both closed-source and medical-specialized baselines. Shows stable predictions and clinically faithful explanations in real longitudinal clinical data from hospital workflows.

Conclusion: MEDIC-AD successfully bridges the gap between general medical VLMs and clinically actionable outputs through its staged design, delivering practical value in patient-monitoring and decision-support workflows.

Abstract: Lesion detection, symptom tracking, and visual explainability are central to real-world medical image analysis, yet current medical Vision-Language Models (VLMs) still lack mechanisms that translate their broad knowledge into clinically actionable outputs. To bridge this gap, we present MEDIC-AD, a clinically oriented VLM that strengthens these three capabilities through a stage-wise framework. First, learnable anomaly-aware tokens () encourage the model to focus on abnormal regions and build more discriminative lesion centered representations. Second, inter image difference tokens () explicitly encode temporal changes between studies, allowing the model to distinguish worsening, improvement, and stability in disease burden. Finally, a dedicated explainability stage trains the model to generate heatmaps that highlight lesion-related regions, offering clear visual evidence that is consistent with the model’s reasoning. Through our staged design, MEDIC-AD steadily boosts performance across anomaly detection, symptom tracking, and anomaly segmentation, achieving state-of-the-art results compared with both closed source and medical-specialized baselines. Evaluations on real longitudinal clinical data collected from real hospital workflows further show that MEDIC-AD delivers stable predictions and clinically faithful explanations in practical patient-monitoring and decision-support workflows

Method: Formulates light editing as sequence-to-sequence prediction in visual token space using video diffusion priors. Uses light-control tokens to adjust position, color, intensity along with reflections, shadows, and falloff. Introduces adaptive token-pruning to preserve spatial information while compactly encoding non-spatial attributes. Trained on scalable rendering pipeline generating image pairs with varied lighting while keeping scene content consistent.

Result: Achieves precise independent control over light position, color, and intensity. Shows high PSNR and strong semantic consistency (DINO, CLIP) across different editing tasks. Reduces control sequence length by 41% while maintaining editing fidelity through token pruning.

Conclusion: LightMover provides a unified framework for controllable light manipulation in single images, enabling physically plausible illumination changes without re-rendering by leveraging video diffusion priors and efficient token-based representation.

Abstract: We present LightMover, a framework for controllable light manipulation in single images that leverages video diffusion priors to produce physically plausible illumination changes without re-rendering the scene. We formulate light editing as a sequence-to-sequence prediction problem in visual token space: given an image and light-control tokens, the model adjusts light position, color, and intensity together with resulting reflections, shadows, and falloff from a single view. This unified treatment of spatial (movement) and appearance (color, intensity) controls improves both manipulation and illumination understanding. We further introduce an adaptive token-pruning mechanism that preserves spatially informative tokens while compactly encoding non-spatial attributes, reducing control sequence length by 41% while maintaining editing fidelity. To train our framework, we construct a scalable rendering pipeline that generates large numbers of image pairs across varied light positions, colors, and intensities while keeping the scene content consistent with the original image. LightMover enables precise, independent control over light position, color, and intensity, and achieves high PSNR and strong semantic consistency (DINO, CLIP) across different tasks.

Method: Proposes Reasoning-Driven Anomaly Localization (ReAL) that extracts anomaly-related tokens from MLLMs’ autoregressive reasoning process and aggregates attention responses to generate pixel-level anomaly maps. Also introduces Consistency-Guided Reasoning Optimization (CGRO) using reinforcement learning to align reasoning tokens with visual attentions for more coherent reasoning and accurate localization.

Result: Extensive experiments on four public benchmarks show significant improvements in anomaly detection, localization, and interpretability. The method achieves performance competitive with MLLM-based methods trained with dense pixel-level supervision, despite using only image-level supervision.

Conclusion: The work successfully demonstrates that MLLMs’ intrinsic reasoning capabilities can be leveraged for comprehensive anomaly analysis without requiring pixel-level annotations or external vision modules, advancing multimodal understanding in anomaly detection tasks.

Abstract: Multimodal large language models (MLLMs) have recently demonstrated remarkable reasoning and perceptual abilities for anomaly detection. However, most approaches remain confined to image-level anomaly detection and textual reasoning, while pixel-level localization still relies on external vision modules and dense annotations. In this work, we activate the intrinsic reasoning potential of MLLMs to perform anomaly detection, pixel-level localization, and interpretable reasoning solely from image-level supervision, without any auxiliary components or pixel-wise labels. Specifically, we propose Reasoning-Driven Anomaly Localization (ReAL), which extracts anomaly-related tokens from the autoregressive reasoning process and aggregates their attention responses to produce pixel-level anomaly maps. We further introduce a Consistency-Guided Reasoning Optimization (CGRO) module that leverages reinforcement learning to align reasoning tokens with visual attentions, resulting in more coherent reasoning and accurate anomaly localization. Extensive experiments on four public benchmarks demonstrate that our method significantly improves anomaly detection, localization, and interpretability. Remarkably, despite relying solely on image-level supervision, our approach achieves performance competitive with MLLM-based methods trained under dense pixel-level supervision. Code is available at https://github.com/YizhouJin313/ReADL.

Method: Introduces a unified evaluation framework with dependency-based TextSceneGraphParser (spaCy) to extract subject-relation-object triples, and Graph Asymmetry Scorer using optimal bipartite matching to inject structural relational priors. Tests four VLMs (CLIP, BLIP, LLaVA, Qwen3-VL-8B-Thinking) on Winoground benchmark with plain and scene-graph-augmented regimes.

Result: Qwen3-VL-8B-Thinking achieves group score of 62.75, far above encoder-based models. Proposed multi-turn scene graph filtering strategy lifts it to 66.0, surpassing prior open-source state-of-the-art. Scene graph augmentation benefits capable models but provides negligible or negative gains for weaker baselines.

Conclusion: Scene graph augmentation at inference time can significantly improve compositional reasoning in capable vision-language models, revealing a capability augmentation tradeoff where stronger models benefit more from structural priors.

Abstract: Vision-language models (VLMs) excel at image-text retrieval yet persistently fail at compositional reasoning, distinguishing captions that share the same words but differ in relational structure. We present, a unified evaluation and augmentation framework benchmarking four architecturally diverse VLMs,CLIP, BLIP, LLaVA, and Qwen3-VL-8B-Thinking,on the Winoground benchmark under plain and scene-graph-augmented regimes. We introduce a dependency-based TextSceneGraphParser (spaCy) extracting subject-relation-object triples, and a Graph Asymmetry Scorer using optimal bipartite matching to inject structural relational priors. Caption ablation experiments (subject-object masking and swapping) reveal that Qwen3-VL-8B-Thinking achieves a group score of 62.75, far above all encoder-based models, while a proposed multi-turn SG filtering strategy further lifts it to 66.0, surpassing prior open-source state-of-the-art. We analyze the capability augmentation tradeoff and find that SG augmentation benefits already capable models while providing negligible or negative gains for weaker baselines. Code: https://github.com/amartyacodes/Inference-Time-Structural-Reasoning-for-Compositional-Vision-Language-Understanding

Method: Introduces COSMIC benchmark with 899 3D scenes and 1250 QA pairs across 5 tasks; tests two static MLLM agents exchanging natural-language messages from different viewpoints to solve spatial queries; compares with 250 human-human dialogues.

Result: MLLMs show consistent capability hierarchy: best at identifying shared anchor objects (72% for Gemini-3-Pro-Thinking), worse at relational reasoning, largely fail at building globally consistent maps (near chance). Humans achieve 95% accuracy. Model dialogues don’t converge like human conversations.

Conclusion: Current MLLMs have limited ability to build and maintain robust shared mental models through spatial communication, with significant room for improvement compared to human performance.

Abstract: Humans build shared spatial understanding by communicating partial, viewpoint-dependent observations. We ask whether Multimodal Large Language Models (MLLMs) can do the same, aligning distinct egocentric views through dialogue to form a coherent, allocentric mental model of a shared environment. To study this systematically, we introduce COSMIC, a benchmark for Collaborative Spatial Communication. In this setting, two static MLLM agents observe a 3D indoor environment from different viewpoints and exchange natural-language messages to solve spatial queries. COSMIC contains 899 diverse scenes and 1250 question-answer pairs spanning five tasks. We find a consistent capability hierarchy, MLLMs are most reliable at identifying shared anchor objects across views, perform worse on relational reasoning, and largely fail at building globally consistent maps, performing near chance, even for the frontier models. Moreover, we find thinking capability yields consistent gains in anchor grounding, but is insufficient for higher-level spatial communication. To contextualize model behavior, we additionally collect 250 human-human dialogues. Humans achieve 95% aggregate accuracy, leaving significant room for improvement for even the best performing model Gemini-3-Pro-Thinking which achieves 72% aggregate accuracy. Moreover, human conversations become increasingly specific as partners converge on a shared mental model, whereas model dialogues continue to explore new possibilities rather than converging, consistent with a limited ability to build and maintain a robust shared mental model. Our code and data is available at https://github.com/ankursikarwar/Cosmic

Method: Introduces Discrete Native Autoregressive (DiNA) framework with shared discrete space representation. Key innovation is dNaViT (Discrete Native Any-resolution Visual Transformer) for tokenization/de-tokenization at arbitrary resolutions. LongCat-Next model processes text, vision, and audio under single autoregressive objective with minimal modality-specific design.

Result: LongCat-Next achieves strong performance across wide range of multimodal benchmarks, addresses performance ceiling of discrete vision modeling on understanding tasks, and provides unified approach to reconcile conflict between understanding and generation.

Conclusion: DiNA framework represents significant step toward native multimodality, enabling consistent autoregressive modeling across modalities. The open-sourced LongCat-Next model and tokenizers aim to foster further research in unified multimodal systems.

Abstract: The prevailing Next-Token Prediction (NTP) paradigm has driven the success of large language models through discrete autoregressive modeling. However, contemporary multimodal systems remain language-centric, often treating non-linguistic modalities as external attachments, leading to fragmented architectures and suboptimal integration. To transcend this limitation, we introduce Discrete Native Autoregressive (DiNA), a unified framework that represents multimodal information within a shared discrete space, enabling a consistent and principled autoregressive modeling across modalities. A key innovation is the Discrete Native Any-resolution Visual Transformer (dNaViT), which performs tokenization and de-tokenization at arbitrary resolutions, transforming continuous visual signals into hierarchical discrete tokens. Building on this foundation, we develop LongCat-Next, a native multimodal model that processes text, vision, and audio under a single autoregressive objective with minimal modality-specific design. As an industrial-strength foundation model, it excels at seeing, painting, and talking within a single framework, achieving strong performance across a wide range of multimodal benchmarks. In particular, LongCat-Next addresses the long-standing performance ceiling of discrete vision modeling on understanding tasks and provides a unified approach to effectively reconcile the conflict between understanding and generation. As an attempt toward native multimodality, we open-source the LongCat-Next and its tokenizers, hoping to foster further research and development in the community. GitHub: https://github.com/meituan-longcat/LongCat-Next

Method: Temporal Global Policy Optimization (TGPO) uses reinforcement learning with verifiable rewards (RLVR) that contrasts model outputs from temporally ordered versus shuffled video frames to create calibrated, globally normalized reward signals favoring temporal coherence.

Result: Experiments across five egocentric video benchmarks show TGPO consistently improves temporal grounding and causal coherence, outperforming prior RL-based video reasoning approaches and effectively suppressing spatial shortcut behaviors.

Conclusion: TGPO provides a simple and scalable pathway toward temporally robust MLLMs for egocentric video understanding by explicitly rewarding temporal awareness through contrastive reinforcement learning.

Abstract: Multimodal large language models (MLLMs) have recently shown strong performance in visual understanding, yet they often lack temporal awareness, particularly in egocentric settings where reasoning depends on the correct ordering and evolution of events. This deficiency stems in part from training objectives that fail to explicitly reward temporal reasoning and instead rely on frame-level spatial shortcuts. To address this limitation, we propose Temporal Global Policy Optimization (TGPO), a reinforcement learning with verifiable rewards (RLVR) algorithm designed to incentivize temporal awareness in MLLMs. TGPO contrasts model outputs generated from temporally ordered versus shuffled video frames to derive calibrated, globally normalized reward signals that explicitly favor temporally coherent reasoning. Integrated with GRPO and GSPO, TGPO supports cold-start RL training and effectively suppresses spatial shortcut behaviors learned by existing MLLMs. Experiments across five egocentric video benchmarks demonstrate that TGPO consistently improves temporal grounding and causal coherence, outperforming prior RL-based video reasoning approaches. Our results suggest that TGPO offers a simple and scalable pathway toward temporally robust MLLMs for egocentric video understanding.

Method: Proposes MotionReward - a heterogeneous-representation, multi-dimensional reward model that maps different motion representations into a shared semantic space anchored by text. Uses Self-refinement Preference Learning to enhance semantics without extra annotations. Also introduces EasyTune - an efficient fine-tuning method that optimizes step-wise rather than full trajectory to overcome recursive gradient dependence bottlenecks.

Result: Achieves FID 0.132 at 22.10 GB peak memory for MLD model, saving up to 15.22 GB over DRaFT. Reduces FID by 22.9% on joint-based ACMDM, and achieves 12.6% R-Precision gain and 23.3% FID improvement on rotation-based HY Motion.

Conclusion: The framework effectively addresses limitations of existing post-training methods by providing unified semantic representation, multi-dimensional reward learning, and efficient fine-grained optimization for text-to-motion generation.

Abstract: Text-to-motion generation has advanced with diffusion- and flow-based generative models, yet supervised pretraining remains insufficient to align models with high-level objectives such as semantic consistency, realism, and human preference. Existing post-training methods have key limitations: they (1) target a specific motion representation, such as joints, (2) optimize a particular aspect, such as text-motion alignment, and may compromise other factors; and (3) incur substantial computational overhead, data dependence, and coarse-grained optimization. We present a reinforcement fine-tuning framework that comprises a heterogeneous-representation, multi-dimensional reward model, MotionReward, and an efficient, fine-grained fine-tuning method, EasyTune. To obtain a unified semantics representation, MotionReward maps heterogeneous motions into a shared semantic space anchored by text, enabling multidimensional reward learning; Self-refinement Preference Learning further enhances semantics without additional annotations. For efficient and effective fine-tuning, we identify the recursive gradient dependence across denoising steps as the key bottleneck, and propose EasyTune, which optimizes step-wise rather than over the full trajectory, yielding dense, fine-grained, and memory-efficient updates. Extensive experiments validate the effectiveness of our framework, achieving FID 0.132 at 22.10 GB peak memory for MLD model and saving up to 15.22 GB over DRaFT. It reduces FID by 22.9% on joint-based ACMDM, and achieves a 12.6% R-Precision gain and 23.3% FID improvement on rotation-based HY Motion. Our project page with code is publicly available.

Method: Proposes KαLOS, a unified meta-algorithm that generalizes the “Localization First” principle. It resolves spatial correspondence before assessing agreement, transforming complex spatio-categorical problems into nominal reliability matrices. Uses data-driven configuration by statistically calibrating localization parameters to inherent agreement distribution.

Result: Enables granular diagnostics beyond single scores, including annotator vitality, collaboration clustering, and localization sensitivity. Introduces empirically derived noise generator for validation that models complex human variability rather than uniform error assumptions.

Conclusion: KαLOS establishes a robust standard for distinguishing signal from noise in modern computer vision benchmarks, restoring trust in benchmarking through rigorous quantification of annotation consistency.

Abstract: Progress in object detection benchmarks is stagnating. It is limited not by architectures but by the inability to distinguish model improvements from label noise. To restore trust in benchmarking the field requires rigorous quantification of annotation consistency to ensure the reliability of evaluation data. However, standard statistical metrics fail to handle the instance correspondence problem inherent to vision tasks. Furthermore, validating new agreement metrics remains circular because no objective ground truth for agreement exists. This forces reliance on unverifiable heuristics. We propose K$α$LOS (KALOS), a unified meta-algorithm that generalizes the “Localization First” principle to standardize dataset quality evaluation. By resolving spatial correspondence before assessing agreement, our framework transforms complex spatio-categorical problems into nominal reliability matrices. Unlike prior heuristic implementations, K$α$LOS employs a principled, data-driven configuration; by statistically calibrating the localization parameters to the inherent agreement distribution, it generalizes to diverse tasks ranging from bounding boxes to volumetric segmentation or pose estimation. This standardization enables granular diagnostics beyond a single score. These include annotator vitality, collaboration clustering, and localization sensitivity. To validate this approach, we introduce a novel and empirically derived noise generator. Where prior validations relied on uniform error assumptions, our controllable testbed models complex and non-isotropic human variability. This provides evidence of the metric’s properties and establishes K$α$LOS as a robust standard for distinguishing signal from noise in modern computer vision benchmarks.

Method: Proposes Frequency-Aware Dropout (FAD) with two components: 1) co-occurrence analysis to identify token-context relationships, and 2) curriculum-inspired scheduling that gradually adjusts dropout rates based on token frequency to disentangle representations.

Result: Demonstrates consistent improvements in prompt fidelity, stylistic precision, and user-perceived quality across multiple models (SD 1.5, SDXL, FLUX, Qwen-Image) without adding parameters or architectural changes.

Conclusion: FAD provides a simple yet effective regularization technique that enhances controllability and personalization in text-to-image generation with minimal computational overhead, making it readily applicable to existing models.

Abstract: Text-to-image models such as Stable Diffusion have achieved unprecedented levels of high-fidelity visual synthesis. As these models advance, personalization of generative models – commonly facilitated through Low-Rank Adaptation (LoRA) with a dedicated trigger token – has become a significant area of research. Previous works have naively assumed that fine-tuning with a single trigger token to represent new concepts. However, this often results in poor controllability, where the trigger token alone fails to reliably evoke the intended concept. We attribute this issue to the frequent co-occurrence of the trigger token with the surrounding context during fine-tuning, which entangles their representations and compromises the token’s semantic distinctiveness. To disentangle this, we propose Frequency-Aware Dropout (FAD) – a novel regularization technique that improves prompt controllability without adding new parameters. FAD consists of two key components: co-occurrence analysis and curriculum-inspired scheduling. Qualitative and quantitative analyses across token-based diffusion models (SD~1.5 and SDXL) and natural language–driven backbones (FLUX and Qwen-Image) demonstrate consistent gains in prompt fidelity, stylistic precision, and user-perceived quality. Our method provides a simple yet effective dropout strategy that enhances controllability and personalization in text-to-image generation. Notably, it achieves these improvements without introducing additional parameters or architectural modifications, making it readily applicable to existing models with minimal computational overhead.

Method: Introduces CDH-Bench, a benchmark with visual evidence-commonsense conflicts across three dimensions: counting anomalies, relational anomalies, and attribute anomalies. Evaluates frontier VLMs using binary QA and multiple-choice QA with metrics like Counterfactual Accuracy, Commonsense Accuracy, and Commonsense Collapse Rate.

Result: Even strong VLMs remain vulnerable to prior-driven normalization under visual-commonsense conflicts, demonstrating commonsense-driven hallucinations where models override visual evidence in favor of commonsense alternatives.

Conclusion: CDH-Bench provides a controlled diagnostic tool for evaluating visual fidelity in VLMs when visual evidence conflicts with commonsense knowledge, revealing a fundamental reliability issue that needs addressing in multimodal models.

Abstract: Vision-language models (VLMs) achieve strong performance on many benchmarks, yet a basic reliability question remains underexplored: when visual evidence conflicts with commonsense, do models follow what is shown or what commonsense suggests? A characteristic failure in this setting is that the model overrides visual evidence and outputs the commonsense alternative. We term this phenomenon \textbf{commonsense-driven hallucination} (CDH). To evaluate it, we introduce \textbf{CDH-Bench}, a benchmark designed to create explicit \textbf{visual evidence–commonsense conflicts}. CDH-Bench covers three dimensions: \textit{counting anomalies}, \textit{relational anomalies}, and \textit{attribute anomalies}. We evaluate frontier VLMs under \textit{binary Question Answering (QA)} and \textit{multiple-choice QA}, and report metrics including \textit{Counterfactual Accuracy} (CF-Acc), \textit{Commonsense Accuracy} (CS-Acc), \textit{Counterfactual Accuracy Drop} (CFAD), \textit{Commonsense Collapse Rate} (CCR), and \textit{Relative Prior Dependency} (RPD). Results show that even strong models remain vulnerable to prior-driven normalization under visual evidence–commonsense conflict. CDH-Bench provides a controlled diagnostic of visual fidelity under visual evidence–commonsense conflict.

Method: Systematically investigate MCoT hallucination patterns, identify divergent thinking steps as primary source of fabricated texts, then develop strategy to localize these steps and intervene in decoding process to mitigate hallucinations.

Result: Method outperforms existing approaches by large margin and can be integrated with other hallucination mitigation methods to further boost performance.

Conclusion: MCoT models have unique hallucination patterns in divergent thinking steps requiring specialized intervention strategies, with proposed method effectively addressing these issues while being compatible with existing approaches.

Abstract: Multimodal Chain-of-Thought (MCoT) models have demonstrated impressive capability in complex visual reasoning tasks. Unfortunately, recent studies reveal that they suffer from severe hallucination problems due to diminished visual attention during the generation process. However, visual attention decay is a well-studied problem in Large Vision-Language Models (LVLMs). Considering the fundamental differences in reasoning processes between MCoT models and traditional LVLMs, we raise a basic question: Whether MCoT models have unique causes of hallucinations? To answer this question, we systematically investigate the hallucination patterns of MCoT models and find that fabricated texts are primarily generated in associative reasoning steps, which we term divergent thinking. Leveraging these insights, we introduce a simple yet effective strategy that can effectively localize divergent thinking steps and intervene in the decoding process to mitigate hallucinations. Extensive experiments show that our method outperforms existing methods by a large margin. More importantly, our proposed method can be conveniently integrated with other hallucination mitigation methods and further boost their performance. The code is publicly available at https://github.com/ASGO-MM/MCoT-hallucination.

Method: 1) Construct embedding-deduplicated prompt bank for each novel class to generate diverse synthetic images; 2) Support-guided pseudo-label estimation with two-stage refinement: filter low-consistency regions for high-precision seeds, then relabel uncertain pixels with image-adaptive prototypes; 3) Refine boundary-band and unlabeled pixels using constrained SAM-based update for better contour fidelity.

Result: Extensive experiments on PASCAL-5^i and COCO-20^i show consistent improvements in both 1-shot and 5-shot settings, demonstrating synthetic data as a scalable path for GFSS with reliable masks and precise boundaries.

Conclusion: Syn4Seg effectively expands novel-class coverage while improving pseudo-label quality through synthetic data generation and sophisticated refinement techniques, offering a scalable solution for few-shot semantic segmentation.

Abstract: Generalized few-shot semantic segmentation (GFSS) is fundamentally limited by the coverage of novel-class appearances under scarce annotations. While diffusion models can synthesize novel-class images at scale, practical gains are often hindered by insufficient coverage and noisy supervision when masks are unavailable or unreliable. We propose Syn4Seg, a generation-enhanced GFSS framework designed to expand novel-class coverage while improving pseudo-label quality. Syn4Seg first maximizes prompt-space coverage by constructing an embedding-deduplicated prompt bank for each novel class, yielding diverse yet class-consistent synthetic images. It then performs support-guided pseudo-label estimation via a two-stage refinement that i) filters low-consistency regions to obtain high-precision seeds and ii) relabels uncertain pixels with image-adaptive prototypes that combine global (support) and local (image) statistics. Finally, we refine only boundary-band and unlabeled pixels using a constrained SAM-based update to improve contour fidelity without overwriting high-confidence interiors. Extensive experiments on PASCAL-$5^i$ and COCO-$20^i$ demonstrate consistent improvements in both 1-shot and 5-shot settings, highlighting synthetic data as a scalable path for GFSS with reliable masks and precise boundaries.

Method: Proposes HD-VGGT with a dual-branch architecture: 1) low-resolution branch predicts coarse, globally consistent geometry, 2) high-resolution branch refines details via learned feature upsampling. Introduces Feature Modulation to suppress unreliable features early in the transformer, addressing unstable tokens in ambiguous regions. The approach leverages high-resolution supervision without full-resolution transformer costs.

Result: HD-VGGT achieves state-of-the-art reconstruction quality while maintaining computational efficiency. The method successfully handles high-resolution inputs and supervision without the prohibitive costs of full-resolution transformers, demonstrating robust performance in challenging visual conditions.

Conclusion: HD-VGGT provides an effective solution for high-resolution 3D reconstruction by combining coarse-to-fine processing with feature stabilization mechanisms, enabling detailed geometric inference while managing computational complexity and handling visual ambiguities.

Abstract: High-resolution imagery is essential for accurate 3D reconstruction, as many geometric details only emerge at fine spatial scales. Recent feed-forward approaches, such as the Visual Geometry Grounded Transformer (VGGT), have demonstrated the ability to infer scene geometry from large collections of images in a single forward pass. However, scaling these models to high-resolution inputs remains challenging: the number of tokens in transformer architectures grows rapidly with both image resolution and the number of views, leading to prohibitive computational and memory costs. Moreover, we observe that visually ambiguous regions, such as repetitive patterns, weak textures, or specular surfaces, often produce unstable feature tokens that degrade geometric inference, especially at higher resolutions. We introduce HD-VGGT, a dual-branch architecture for efficient and robust high-resolution 3D reconstruction. A low-resolution branch predicts a coarse, globally consistent geometry, while a high-resolution branch refines details via a learned feature upsampling module. To handle unstable tokens, we propose Feature Modulation, which suppresses unreliable features early in the transformer. HD-VGGT leverages high-resolution images and supervision without full-resolution transformer costs, achieving state-of-the-art reconstruction quality.

Method: Created a dataset of over 8,000 high-resolution scanned multiple-choice questions from real civil service exams across South Korea, Japan, Taiwan, India, and the European Union. Questions cover 17 academic/administrative domains and embed all content (problem statements, answer choices, visual elements) within single images with only minimal standardized instructions for answer formatting.

Result: Even state-of-the-art vision-language models achieve only 86% accuracy on the benchmark, demonstrating its difficulty and ability to diagnose model limitations. The benchmark preserves rich visual structures including tables, multilingual typography, and form-like layouts.

Conclusion: EuraGovExam establishes a new standard for evaluating VLMs in high-stakes, multilingual, image-grounded settings and supports practical applications in e-governance, public-sector document analysis, and equitable exam preparation.

Abstract: We present EuraGovExam, a multilingual and multimodal benchmark sourced from real-world civil service examinations across five representative Eurasian regions: South Korea, Japan, Taiwan, India, and the European Union. Designed to reflect the authentic complexity of public-sector assessments, the dataset contains over 8,000 high-resolution scanned multiple-choice questions covering 17 diverse academic and administrative domains. Unlike existing benchmarks, EuraGovExam embeds all question content–including problem statements, answer choices, and visual elements–within a single image, providing only a minimal standardized instruction for answer formatting. This design demands that models perform layout-aware, cross-lingual reasoning directly from visual input. All items are drawn from real exam documents, preserving rich visual structures such as tables, multilingual typography, and form-like layouts. Evaluation results show that even state-of-the-art vision-language models (VLMs) achieve only 86% accuracy, underscoring the benchmark’s difficulty and its power to diagnose the limitations of current models. By emphasizing cultural realism, visual complexity, and linguistic diversity, EuraGovExam establishes a new standard for evaluating VLMs in high-stakes, multilingual, image-grounded settings. It also supports practical applications in e-governance, public-sector document analysis, and equitable exam preparation.

Method: ResAdapt couples a lightweight Allocator with an unchanged MLLM backbone. It formulates allocation as a contextual bandit and trains with Cost-Aware Policy Optimization (CAPO) to convert sparse rollout feedback into stable accuracy-cost learning signals.

Result: ResAdapt improves low-budget operating points, often lies on the efficiency-accuracy frontier, supports up to 16x more frames at same visual budget with over 15% performance gain, especially on reasoning-intensive benchmarks under aggressive compression.

Conclusion: ResAdapt effectively addresses the visual budget allocation problem in MLLMs, enabling better trade-offs between spatial resolution and temporal context without modifying the core MLLM architecture.

Abstract: Multimodal Large Language Models (MLLMs) achieve stronger visual understanding by scaling input fidelity, yet the resulting visual token growth makes jointly sustaining high spatial resolution and long temporal context prohibitive. We argue that the bottleneck lies not in how post-encoding representations are compressed but in the volume of pixels the encoder receives, and address it with ResAdapt, an Input-side adaptation framework that learns how much visual budget each frame should receive before encoding. ResAdapt couples a lightweight Allocator with an unchanged MLLM backbone, so the backbone retains its native visual-token interface while receiving an operator-transformed input. We formulate allocation as a contextual bandit and train the Allocator with Cost-Aware Policy Optimization (CAPO), which converts sparse rollout feedback into a stable accuracy-cost learning signal. Across budget-controlled video QA, temporal grounding, and image reasoning tasks, ResAdapt improves low-budget operating points and often lies on or near the efficiency-accuracy frontier, with the clearest gains on reasoning-intensive benchmarks under aggressive compression. Notably, ResAdapt supports up to 16x more frames at the same visual budget while delivering over 15% performance gain. Code is available at https://github.com/Xnhyacinth/ResAdapt.

Method: Decomposes degradations into: 1) global transmission field for static atmospheric effects shared across views, and 2) per-view particulate residuals for transient disturbances. Uses geometry-guided gradient scaling to stabilize geometry learning under severe visibility degradation during self-supervised optimization of 3D Gaussian representations.

Result: Superior geometry reconstruction compared to task-specific methods across diverse and challenging weather conditions. The physically grounded formulation disentangles complex degradations while preserving scene structure.

Conclusion: NimbusGS provides a unified solution for 3D scene reconstruction under adverse weather by modeling weather’s dual nature and addressing optimization challenges through geometry-guided gradient scaling.

Abstract: We present NimbusGS, a unified framework for reconstructing high-quality 3D scenes from degraded multi-view inputs captured under diverse and mixed adverse weather conditions. Unlike existing methods that target specific weather types, NimbusGS addresses the broader challenge of generalization by modeling the dual nature of weather: a continuous, view-consistent medium that attenuates light, and dynamic, view-dependent particles that cause scattering and occlusion. To capture this structure, we decompose degradations into a global transmission field and per-view particulate residuals. The transmission field represents static atmospheric effects shared across views, while the residuals model transient disturbances unique to each input. To enable stable geometry learning under severe visibility degradation, we introduce a geometry-guided gradient scaling mechanism that mitigates gradient imbalance during the self-supervised optimization of 3D Gaussian representations. This physically grounded formulation allows NimbusGS to disentangle complex degradations while preserving scene structure, yielding superior geometry reconstruction and outperforming task-specific methods across diverse and challenging weather conditions. Code is available at https://github.com/lyy-ovo/NimbusGS.

Method: Introduces ADMesh (15K+ high-quality 3D models with textures and semantic annotations) and CarlaOcc (100K+ frames of physically consistent panoptic occupancy data generated via CARLA simulator with 0.05m voxel resolution).

Result: Created first unified 3D mesh library for autonomous driving and large-scale panoptic occupancy dataset with instance-level ground truth, plus standardized evaluation metrics and systematic benchmark of existing models.

Conclusion: Provides comprehensive resources and evaluation framework to advance 3D panoptic perception research, enabling fair comparison and reproducible work in autonomous driving scene understanding.

Abstract: Panoptic occupancy prediction aims to jointly infer voxel-wise semantics and instance identities within a unified 3D scene representation. Nevertheless, progress in this field remains constrained by the absence of high-quality 3D mesh resources, instance-level annotations, and physically consistent occupancy datasets. Existing benchmarks typically provide incomplete and low-resolution geometry without instance-level annotations, limiting the development of models capable of achieving precise geometric reconstruction, reliable occlusion reasoning, and holistic 3D understanding. To address these challenges, this paper presents an instance-centric benchmark for the 3D panoptic occupancy prediction task. Specifically, we introduce ADMesh, the first unified 3D mesh library tailored for autonomous driving, which integrates over 15K high-quality 3D models with diverse textures and rich semantic annotations. Building upon ADMesh, we further construct CarlaOcc, a large-scale, physically consistent panoptic occupancy dataset generated using the CARLA simulator. This dataset contains over 100K frames with fine-grained, instance-level occupancy ground truth at voxel resolutions as fine as 0.05 m. Furthermore, standardized evaluation metrics are introduced to quantify the quality of existing occupancy datasets. Finally, a systematic benchmark of representative models is established on the proposed dataset, which provides a unified platform for fair comparison and reproducible research in the field of 3D panoptic perception. Code and dataset are available at https://mias.group/CarlaOcc.

Method: 1) Causal mediation analysis to identify neurons/layers responsible for unsafe behaviors; 2) Dual-modal safety subspace projection learning generalized safety subspaces for visual/textual modalities via eigen-decomposition; 3) Hybrid fusion mechanism for dynamic projection during inference to balance visual/textual corrections.

Result: Significantly enhances safety robustness on multiple benchmarks without degrading general multimodal capabilities, outperforms prior activation steering and alignment-based baselines, and shows good transferability against unseen attacks.

Conclusion: The CARE framework provides an effective approach for diagnosing and repairing unsafe channels in LVLMs through causal analysis and subspace projection, offering improved safety control while maintaining model capabilities.

Abstract: Large Vision-Language Models (LVLMs) have achieved impressive performance across multimodal understanding and reasoning tasks, yet their internal safety mechanisms remain opaque and poorly controlled. In this work, we present a comprehensive framework for diagnosing and repairing unsafe channels within LVLMs (CARE). We first perform causal mediation analysis to identify neurons and layers that are causally responsible for unsafe behaviors. Based on these findings, we introduce a dual-modal safety subspace projection method that learns generalized safety subspaces for both visual and textual modalities through generalized eigen-decomposition between benign and malicious activations. During inference, activations are dynamically projected toward these safety subspaces via a hybrid fusion mechanism that adaptively balances visual and textual corrections, effectively suppressing unsafe features while preserving semantic fidelity. Extensive experiments on multiple safety benchmarks demonstrate that our causal-subspace repair framework significantly enhances safety robustness without degrading general multimodal capabilities, outperforming prior activation steering and alignment-based baselines. Additionally, our method exhibits good transferability, defending against unseen attacks.

Method: Extends SaSaSa2VA (which already increased input frames and added [SEG] tokens to Sa2VA backbone) with a simple target existence-aware verification mechanism. This verification strategy helps determine whether referred targets exist in video frames, particularly important for motion-centric expressions.

Result: Achieved final score of 89.19 in the 5th PVUW Challenge (MeViS-Text Track), securing 2nd place. Both quantitative results and ablation studies show the existence-aware verification strategy effectively unlocks strong performance on motion-centric referring tasks.

Conclusion: The target existence-aware verification mechanism, despite its simplicity, is sufficient to achieve strong performance on motion-centric referring video object segmentation tasks, demonstrating the importance of verifying target presence in dynamic video understanding.

Abstract: Referring video object segmentation (RVOS) commonly grounds targets in videos based on static textual cues. MeViS benchmark extends this by incorporating motion-centric expressions (referring & reasoning motion expressions) and introducing no-target queries. Extending SaSaSa2VA, where increased input frames and [SEG] tokens already strengthen the Sa2VA backbone, we adopt a simple yet effective target existence-aware verification mechanism, leading to Still Awesome SaSaSa2VA (SaSaSaSa2VA). Despite its simplicity, the method achieves a final score of 89.19 in the 5th PVUW Challenge (MeViS-Text Track), securing 2nd place. Both quantitative results and ablations suggest that this existence-aware verification strategy is sufficient to unlock strong performance on motion-centric referring tasks.

Method: Proposes IP-SAM with two key components: 1) Self-Prompt Generator (SPG) that distills image context into complementary intrinsic prompts as coarse regional anchors, projected through SAM2’s frozen prompt encoder; 2) Prompt-Space Gating (PSG) that uses intrinsic background prompt as asymmetric suppressive constraint to reduce false positives.

Result: Achieves state-of-the-art performance on four camouflaged object detection benchmarks (e.g., MAE 0.017 on COD10K) with only 21.26M trainable parameters. Also demonstrates strong zero-shot transfer in medical polyp segmentation from Kvasir-SEG to CVC-ClinicDB and ETIS.

Conclusion: Prompt-space conditioning enables effective adaptation of prompt-conditioned foundation models for fully automatic segmentation while preserving their native prompt-guided decoding capabilities, with strong generalization across domains.

Abstract: Prompt-conditioned foundation segmenters have emerged as a dominant paradigm for image segmentation, where explicit spatial prompts (e.g., points, boxes, masks) guide mask decoding. However, many real-world deployments require fully automatic segmentation, creating a structural mismatch: the decoder expects prompts that are unavailable at inference. Existing adaptations typically modify intermediate features, inadvertently bypassing the model’s native prompt interface and weakening prompt-conditioned decoding. We propose IP-SAM, which revisits adaptation from a prompt-space perspective through prompt-space conditioning. Specifically, a Self-Prompt Generator (SPG) distills image context into complementary intrinsic prompts that serve as coarse regional anchors. These cues are projected through SAM2’s frozen prompt encoder, restoring prompt-guided decoding without external intervention. To suppress background-induced false positives, Prompt-Space Gating (PSG) leverages the intrinsic background prompt as an asymmetric suppressive constraint prior to decoding. Under a deterministic no-external-prompt protocol, IP-SAM achieves state-of-the-art performance across four camouflaged object detection benchmarks (e.g., MAE 0.017 on COD10K) with only 21.26M trainable parameters (optimizing SPG, PSG, and a task-specific mask decoder trained from scratch, alongside image-encoder LoRA while keeping the prompt encoder frozen). Furthermore, the proposed conditioning strategy generalizes beyond COD to medical polyp segmentation, where a model trained solely on Kvasir-SEG exhibits strong zero-shot transfer to both CVC-ClinicDB and ETIS.

Method: Two-stage framework: SOTA retrieval followed by VLM reranking with two strategies - pointwise (individual scoring) and pairwise (relative comparison between candidates).

Result: Pointwise methods cause catastrophic performance drops, while pairwise comparison using LLaVA improves Top-1 accuracy over strong retrieval baselines on VIGOR dataset.

Conclusion: VLMs are poorly calibrated for absolute relevance scoring but effective at fine-grained relative visual judgment, making pairwise reranking promising for enhancing CVGL precision.

Abstract: Cross-view geolocalization (CVGL) systems, while effective at retrieving a list of relevant candidates (high Recall@k), often fail to identify the single best match (low Top-1 accuracy). This work investigates the use of zero-shot Vision-Language Models (VLMs) as rerankers to address this gap. We propose a two-stage framework: state-of-the-art (SOTA) retrieval followed by VLM reranking. We systematically compare two strategies: (1) Pointwise (scoring candidates individually) and (2) Pairwise (comparing candidates relatively). Experiments on the VIGOR dataset show a clear divergence: all pointwise methods cause a catastrophic drop in performance or no change at all. In contrast, a pairwise comparison strategy using LLaVA improves Top-1 accuracy over the strong retrieval baseline. Our analysis concludes that, these VLMs are poorly calibrated for absolute relevance scoring but are effective at fine-grained relative visual judgment, making pairwise reranking a promising direction for enhancing CVGL precision.

Method: Defines a scene as a coherent video segment with consistent visual and semantic contexts. Introduces SceneBench benchmark with scene-level challenges. Proposes Scene Retrieval-Augmented Generation (Scene-RAG) that constructs dynamic scene memory by retrieving and integrating relevant context across scenes.

Result: Evaluation shows sharp accuracy drop when VLMs answer scene-level questions, indicating significant forgetting of long-range context. Scene-RAG improves VLM performance by +2.50%, confirming models struggle with long-context retention.

Conclusion: Current VLMs have limitations in long-context retention for video understanding. SceneBench provides a valuable benchmark for evaluating scene-level comprehension, and Scene-RAG demonstrates potential for improving long-range context handling in multimodal models.

Abstract: Long video understanding (LVU) remains a core challenge in multimodal learning. Although recent vision-language models (VLMs) have made notable progress, existing benchmarks mainly focus on either fine-grained perception or coarse summarization, offering limited insight into temporal understanding over long contexts. In this work, we define a scene as a coherent segment of a video in which both visual and semantic contexts remain consistent, aligning with human perception. This leads us to a key question: can current VLMs reason effectively over long, scene-level contexts? To answer this, we introduce a new benchmark, SceneBench, designed to provide scene-level challenges. Our evaluation reveals a sharp drop in accuracy when VLMs attempt to answer scene-level questions, indicating significant forgetting of long-range context. To further validate these findings, we propose Scene Retrieval-Augmented Generation (Scene-RAG), which constructs a dynamic scene memory by retrieving and integrating relevant context across scenes. This Scene-RAG improves VLM performance by +2.50%, confirming that current models still struggle with long-context retention. We hope SceneBench will encourage future research toward VLMs with more robust, human-like video comprehension.

Method: TrendGen leverages cloth images and product attributes to generate cohesive outfit recommendations, and uses Generative AI to transform raw images into high-quality lay-down views with clear, structured presentation of garments.

Result: Evaluation on production data demonstrates TrendGen’s consistent high-quality outfit suggestions and lay-down image generation, showing significant advancement for fashion retail applications.

Conclusion: TrendGen represents an effective AI-driven solution for enhancing online shopping experiences through intelligent outfit recommendations and improved visual presentation of fashion products.

Abstract: Recent advances in Computer Vision have significantly improved image understanding and generation, revolutionizing the fashion industry. However, challenges such as inconsistent lighting, non-ideal garment angles, complex backgrounds, and occlusions in raw images hinder their full potential. Overcoming these obstacles is crucial for developing robust fashion AI systems capable of real-world applications. In this paper, we introduce TrendGen, a Fashion AI system designed to enhance online shopping with intelligent outfit recommendations. Deployed on a major e-commerce platform, TrendGen leverages cloth images and product attributes to generate trend-aligned, cohesive outfit suggestions. Additionally, it employs Generative AI to transform raw images into high-quality lay-down views, offering a clear and structured presentation of garments. Our evaluation on production data demonstrates TrendGen’s consistent high-quality outfits and lay-down images, marking a significant advancement in AI-driven solutions for fashion retail.

Method: Uses off-the-shelf point tracker to generate motion trajectories from input videos, employs motion-aware masking strategy, and reconstructs both pixel/semantic features with motion trajectories as complementary supervision.

Result: Consistently outperforms state-of-the-art video self-supervised learning baselines across six diverse datasets, learning more discriminative and generalizable representations.

Conclusion: Explicit motion supervision in masked video modeling significantly improves video representation learning, especially for motion-centric tasks, through motion-aware masking and trajectory reconstruction.

Abstract: Masked video modeling (MVM) has emerged as a simple and scalable self-supervised pretraining paradigm, but only encodes motion information implicitly, limiting the encoding of temporal dynamics in the learned representations. As a result, such models struggle on motion-centric tasks that require fine-grained motion awareness. To address this, we propose TrackMAE, a simple masked video modeling paradigm that explicitly uses motion information as a reconstruction signal. In TrackMAE, we use an off-the-shelf point tracker to sparsely track points in the input videos, generating motion trajectories. Furthermore, we exploit the extracted trajectories to improve random tube masking with a motion-aware masking strategy. We enhance video representations learned in both pixel and feature semantic reconstruction spaces by providing a complementary supervision signal in the form of motion targets. We evaluate on six datasets across diverse downstream settings and find that TrackMAE consistently outperforms state-of-the-art video self-supervised learning baselines, learning more discriminative and generalizable representations. Code available at https://github.com/rvandeghen/TrackMAE

Method: Created Body-Keypoint-Part Dataset (BKPD) with images across age groups and sexual explicitness, with hierarchical labels for skeletal keypoints and body part bounding boxes. Proposed BKP-Association and YOLO-BKP methods for simultaneous pose estimation and detection with per-individual target association.

Result: Methods benchmarked on COCO-Keypoints and COCO-HumanParts, achieving competitive results with joint-task models. Cross-domain studies on BKPD and case study on RCPD highlight challenges in sexually explicit domains.

Conclusion: Study addresses unexplored targets in CSAI domain, paving way for novel research opportunities in explainable CSAI classification using human-centric perception approaches.

Abstract: Law enforcement agencies and non-gonvernmental organizations handling reports of Child Sexual Abuse Imagery (CSAI) are overwhelmed by large volumes of data, requiring the aid of automation tools. However, defining sexual abuse in images of children is inherently challenging, encompassing sexually explicit activities and hints of sexuality conveyed by the individual’s pose, or their attire. CSAI classification methods often rely on black-box approaches, targeting broad and abstract concepts such as pornography. Thus, our work is an in-depth exploration of tasks from the literature on Human-Centric Perception, across the domains of safe images, adult pornography, and CSAI, focusing on targets that enable more objective and explainable pipelines for CSAI classification in the future. We introduce the Body-Keypoint-Part Dataset (BKPD), gathering images of people from varying age groups and sexual explicitness to approximate the domain of CSAI, along with manually curated hierarchically structured labels for skeletal keypoints and bounding boxes for person and body parts, including head, chest, hip, and hands. We propose two methods, namely BKP-Association and YOLO-BKP, for simultaneous pose estimation and detection, with targets associated per individual for a comprehensive decomposed representation of each person. Our methods are benchmarked on COCO-Keypoints and COCO-HumanParts, as well as our human-centric dataset, achieving competitive results with models that jointly perform all tasks. Cross-domain ablation studies on BKPD and a case study on RCPD highlight the challenges posed by sexually explicit domains. Our study addresses previously unexplored targets in the CSAI domain, paving the way for novel research opportunities.

Method: Proposes class-distribution guided active learning with three complementary criteria: inter-sample diversity (prioritizes samples with different predicted class distributions), intra-set diversity (prevents redundant sampling), and frequency-weighted uncertainty (emphasizes rare classes by reweighting voxel-level entropy with inverse per-sample class proportions). Uses geographically disjoint train/validation split to reduce map memorization.

Result: Achieves 26.62 mIoU with only 42.4% labeled data, comparable to full supervision and outperforming active learning baselines at same budget. Validates generality on SemanticKITTI with different architecture, showing consistent effectiveness across datasets.

Conclusion: The proposed active learning framework effectively addresses class imbalance and annotation costs in 3D occupancy prediction, achieving near-full-supervision performance with significantly less labeled data while maintaining generalizability across datasets.

Abstract: 3D occupancy prediction provides dense spatial understanding critical for safe autonomous driving. However, this task suffers from a severe class imbalance due to its volumetric representation, where safety-critical objects (bicycles, traffic cones, pedestrians) occupy minimal voxels compared to dominant backgrounds. Additionally, voxel-level annotation is costly, yet dedicating effort to dominant classes is inefficient. To address these challenges, we propose a class-distribution guided active learning framework for selecting training samples to annotate in autonomous driving datasets. Our approach combines three complementary criteria to select the training samples. Inter-sample diversity prioritizes samples whose predicted class distributions differ from those of the labeled set, intra-set diversity prevents redundant sampling within each acquisition cycle, and frequency-weighted uncertainty emphasizes rare classes by reweighting voxel-level entropy with inverse per-sample class proportions. We ensure evaluation validity by using a geographically disjoint train/validation split of Occ3D-nuScenes, which reduces train-validation overlap and mitigates potential map memorization. With only 42.4% labeled data, our framework reaches 26.62 mIoU, comparable to full supervision and outperforming active learning baselines at the same budget. We further validate generality on SemanticKITTI using a different architecture, demonstrating consistent effectiveness across datasets.

Method: The method formalizes Geometric Complete 4D Reconstruction as a unified framework of reconstruction and completion. It uses a decoder-only transformer to operate on all context globally from sequential video input, directly accumulating full contexts onto each frame to reconstruct complete geometry for every timestamp.

Result: Complet4R demonstrates state-of-the-art performance on the proposed benchmark for Geometric Complete 4D Reconstruction and the 3D Point Tracking task.

Conclusion: The framework successfully addresses the challenge of geometric complete 4D reconstruction by leveraging global context accumulation through transformer architecture, achieving superior performance in reconstructing dynamic scenes with temporal coherence.

Abstract: We introduce Complet4R, a novel end-to-end framework for Geometric Complete 4D Reconstruction, which aims to recover temporally coherent and geometrically complete reconstruction for dynamic scenes. Our method formalizes the task of Geometric Complete 4D Reconstruction as a unified framework of reconstruction and completion, by directly accumulating full contexts onto each frame. Unlike previous approaches that rely on pairwise reconstruction or local motion estimation, Complet4R utilizes a decoder-only transformer to operate all context globally directly from sequential video input, reconstructing a complete geometry for every single timestamp, including occluded regions visible in other frames. Our method demonstrates the state-of-the-art performance on our proposed benchmark for Geometric Complete 4D Reconstruction and the 3D Point Tracking task. Code will be released to support future research.

Method: Proposes Decoupling then Perceive (DTP) framework with three key components: 1) Frequency-aware Structural Decoupling (FSD) separates input into low-frequency luminance and high-frequency texture subspaces; 2) Semantics-specific Dual-path Representation (SDR) learning for targeted enhancement of each component; 3) Cross-frequency Semantic Recomposition (CSR) module to integrate decoupled representations while maintaining structural consistency.

Result: Extensive experiments show DTP outperforms state-of-the-art methods, achieving +1.6% PSNR, +9.6% SSIM, and -48% LPIPS improvements on standard LLISR benchmarks.

Conclusion: DTP effectively addresses the limitations of serial processing in low-light image super-resolution by explicitly decoupling and separately enhancing luminance and texture components, leading to superior visual quality and structural preservation.

Abstract: Low-light image super-resolution (LLISR) is essential for restoring fine visual details and perceptual quality under insufficient illumination conditions with ubiquitous low-resolution devices. Although pioneer methods achieve high performance on single tasks, they solve both tasks in a serial manner, which inevitably leads to artifact amplification, texture suppression, and structural degradation. To address this, we propose Decoupling then Perceive (DTP), a novel frequency-aware framework that explicitly separates luminance and texture into semantically independent components, enabling specialized modeling and coherent reconstruction. Specifically, to adaptively separate the input into low-frequency luminance and high-frequency texture subspaces, we propose a Frequency-aware Structural Decoupling (FSD) mechanism, which lays a solid foundation for targeted representation learning and reconstruction. Based on the decoupled representation, a Semantics-specific Dual-path Representation (SDR) learning strategy that performs targeted enhancement and reconstruction for each frequency component is further designed, facilitating robust luminance adjustment and fine-grained texture recovery. To promote structural consistency and perceptual alignment in the reconstructed output, building upon this dual-path modeling, we further introduce a Cross-frequency Semantic Recomposition (CSR) module that selectively integrates the decoupled representations. Extensive experiments on the most widely used LLISR benchmarks demonstrate the superiority of our DTP framework, improving $+$1.6% PSNR, $+$9.6% SSIM, and $-$48% LPIPS compared to the most state-of-the-art (SOTA) algorithm. Codes are released at https://github.com/JXVision/DTP.

Method: Developed YOLOv11-based deep learning model for simultaneous wound boundary segmentation (WBS) and wound classification (WC) across five wound types. Created balanced dataset of 2,963 annotated images with five-fold cross-validation. Used data augmentation (rotation, flipping, brightness/saturation/exposure variations) to improve generalization. Tested different YOLOv11 variants (YOLOv11x, YOLOv11n).

Result: YOLOv11x achieved highest performance: F1-scores of 0.9341 for WBS and 0.8736 for WC. YOLOv11n provided comparable accuracy with lower computational cost. Data augmentation significantly improved performance, especially for visually subtle burn injury cases. Models showed robustness against complex backgrounds and high intra-class variability.

Conclusion: YOLOv11-based architectures are effective for accurate, real-time wound analysis in clinical and remote care settings, handling both segmentation and classification simultaneously across multiple wound types.

Abstract: Accurate wound classification and boundary segmentation are essential for guiding clinical decisions in both chronic and acute wound management. However, most existing AI models are limited, focusing on a narrow set of wound types or performing only a single task (segmentation or classification), which reduces their clinical applicability. This study presents a deep learning model based on YOLOv11 that simultaneously performs wound boundary segmentation (WBS) and wound classification (WC) across five clinically relevant wound types: burn injury (BI), pressure injury (PI), diabetic foot ulcer (DFU), vascular ulcer (VU), and surgical wound (SW). A wound-type balanced dataset of 2,963 annotated images was created to train the models for both tasks, with stratified five-fold cross-validation ensuring robust and unbiased evaluation. The models trained on the original non-augmented dataset achieved consistent performance across folds, though BI detection accuracy was relatively lower. Therefore, the dataset was augmented using rotation, flipping, and variations in brightness, saturation, and exposure to help the model learn more generalized and invariant features. This augmentation significantly improved model performance, particularly in detecting visually subtle BI cases. Among tested variants, YOLOv11x achieved the highest performance with F1-scores of 0.9341 (WBS) and 0.8736 (WC), while the lightweight YOLOv11n provided comparable accuracy at lower computational cost, making it suitable for resource-constrained deployments. Supported by confusion matrices and visual detection outputs, the results confirm the model’s robustness against complex backgrounds and high intra-class variability, demonstrating the potential of YOLOv11-based architectures for accurate, real-time wound analysis in both clinical and remote care settings.

Method: Proposes RICE (Reciprocal Interaction-based Cross-functionality Exploitation), a novel attack paradigm that exploits bidirectional interactions between understanding and generation. Systematically evaluates Generation-to-Understanding (G-U) and Understanding-to-Generation (U-G) attack pathways where unsafe intermediate signals propagate across modalities.

Result: Extensive experiments show high Attack Success Rates (ASR) in both directions, revealing previously overlooked safety weaknesses inherent to unified multimodal models.

Conclusion: Cross-functionality reciprocity itself constitutes a structural source of vulnerability in unified multimodal models, demonstrating that unsafe signals can propagate and amplify safety risks across modalities.

Abstract: Recent advances in Large Language Models (LLMs) and Text-to-Image (T2I) models have led to the emergence of Unified Multimodal Models (UMMs), where multimodal understanding and image generation are tightly integrated within a shared architecture. Prior studies suggest that such reciprocity enhances cross-functionality performance through shared representations and joint optimization. However, the safety implications of this tight coupling remain largely unexplored, as existing safety research predominantly analyzes understanding and generation functionalities in isolation. In this work, we investigate whether cross-functionality reciprocity itself constitutes a structural source of vulnerability in UMMs. We propose RICE: Reciprocal Interaction-based Cross-functionality Exploitation, a novel attack paradigm that explicitly exploits bidirectional interactions between understanding and generation. Using this framework, we systematically evaluate Generation-to-Understanding (G-U) and Understanding-to-Generation (U-G) attack pathways, demonstrating that unsafe intermediate signals can propagate across modalities and amplify safety risks. Extensive experiments show high Attack Success Rates (ASR) in both directions, revealing previously overlooked safety weaknesses inherent to UMMs.

Method: EVA uses sequential glimpses with minimal fovea-periphery representation, CNN-based feature extraction, variance control, and adaptive gating to stabilize attention dynamics. It’s trained with standard classification objectives without gaze supervision.

Result: On CIFAR-10 with human gaze annotations, EVA improves scanpath alignment (DTW, NSS metrics) while maintaining competitive accuracy. On ImageNet-100 and COCO-Search18, it yields human-like scanpaths without additional training or gaze supervision.

Conclusion: EVA provides a principled framework for trustworthy, human-interpretable active vision by making the performance-human-likeness trade-off explicit and adjustable.

Abstract: Optimizing vision models purely for classification accuracy can impose an alignment tax, degrading human-like scanpaths and limiting interpretability. We introduce EVA, a neuroscience-inspired hard-attention mechanistic testbed that makes the performance-human-likeness trade-off explicit and adjustable. EVA samples a small number of sequential glimpses using a minimal fovea-periphery representation with CNN-based feature extractor and integrates variance control and adaptive gating to stabilize and regulate attention dynamics. EVA is trained with the standard classification objective without gaze supervision. On CIFAR-10 with dense human gaze annotations, EVA improves scanpath alignment under established metrics such as DTW, NSS, while maintaining competitive accuracy. Ablations show that CNN-based feature extraction drives accuracy but suppresses human-likeness, whereas variance control and gating restore human-aligned trajectories with minimal performance loss. We further validate EVA’s scalability on ImageNet-100 and evaluate scanpath alignment on COCO-Search18 without COCO-Search18 gaze supervision or finetuning, where EVA yields human-like scanpaths on natural scenes without additional training. Overall, EVA provides a principled framework for trustworthy, human-interpretable active vision.

Method: Introduces TerraSeg model trained on OmniLiDAR dataset (12 public benchmarks, 22M scans, 15 sensor models) using PseudoLabeler module that generates high-quality ground/non-ground labels through self-supervised per-scan runtime optimization without human annotations.

Result: TerraSeg achieves state-of-the-art results on nuScenes, SemanticKITTI, and Waymo Perception benchmarks while delivering real-time performance, despite using no manual labels during training.

Conclusion: TerraSeg demonstrates that self-supervised learning on diverse sensor data can produce highly generalizable ground segmentation models that outperform supervised methods, enabling scalable LiDAR perception across different platforms.

Abstract: LiDAR perception is fundamental to robotics, enabling machines to understand their environment in 3D. A crucial task for LiDAR-based scene understanding and navigation is ground segmentation. However, existing methods are either handcrafted for specific sensor configurations or rely on costly per-point manual labels, severely limiting their generalization and scalability. To overcome this, we introduce TerraSeg, the first self-supervised, domain-agnostic model for LiDAR ground segmentation. We train TerraSeg on OmniLiDAR, a unified large-scale dataset that aggregates and standardizes data from 12 major public benchmarks. Spanning almost 22 million raw scans across 15 distinct sensor models, OmniLiDAR provides unprecedented diversity for learning a highly generalizable ground model. To supervise training without human annotations, we propose PseudoLabeler, a novel module that generates high-quality ground and non-ground labels through self-supervised per-scan runtime optimization. Extensive evaluations demonstrate that, despite using no manual labels, TerraSeg achieves state-of-the-art results on nuScenes, SemanticKITTI, and Waymo Perception while delivering real-time performance. Our code and model weights are publicly available.

Method: Introduces Falcon Perception, a unified dense Transformer that processes image patches and text tokens in shared parameter space from the first layer. Uses hybrid attention (bidirectional for image tokens, causal for prediction tokens) to combine global visual context with autoregressive instance generation. Retains lightweight token interface and decodes continuous spatial outputs with specialized heads.

Result: Achieves 68.0 Macro-F1 on SA-Co benchmark (vs. 62.3 for SAM3). Introduces PBench for compositional prompts and dense long-context regimes. Falcon OCR (300M parameters) attains 80.3% on olmOCR and 88.64 on OmniDocBench.

Conclusion: Early-fusion architecture with shared parameter space can effectively handle both perception and task modeling, promoting simplicity by keeping a single scalable backbone and shifting complexity to data and training signals.

Abstract: Perception-centric systems are typically implemented with a modular encoder-decoder pipeline: a vision backbone for feature extraction and a separate decoder (or late-fusion module) for task prediction. This raises a central question: is this architectural separation essential or can a single early-fusion stack do both perception and task modeling at scale? We introduce Falcon Perception, a unified dense Transformer that processes image patches and text tokens in a shared parameter space from the first layer, using a hybrid attention pattern (bidirectional among image tokens, causal for prediction tokens) to combine global visual context with autoregressive, variable-length instance generation. To keep dense outputs practical, Falcon Perception retains a lightweight token interface and decodes continuous spatial outputs with specialized heads, enabling parallel high-resolution mask prediction. Our design promotes simplicity: we keep a single scalable backbone and shift complexity toward data and training signals, adding only small heads where outputs are continuous and dense. On SA-Co, Falcon Perception improves mask quality to 68.0 Macro-F$_1$ compared to 62.3 of SAM3. We also introduce PBench, a benchmark targeting compositional prompts (OCR, spatial constraints, relations) and dense long-context regimes, where the model shows better gains. Finally, we extend the same early-fusion recipe to Falcon OCR: a compact 300M-parameter model which attains 80.3% on olmOCR and 88.64 on OmniDocBench.

Method: Proposes HMPDM with three key components: 1) Temporal-aware Latent Conditioning (TaLC) for implicit historical motion injection, 2) Motion-aware Pyramid Encoder (MaPE) for multi-scale motion representation, and 3) Self-Conditioning (SC) strategy for stable iterative denoising.

Result: Outperforms state-of-the-art methods on Cityscapes and KITTI benchmarks, achieving 28.2% improvement in FVD on Cityscapes under monocular RGB input configuration.

Conclusion: HMPDM effectively enhances motion understanding and temporal coherence in video prediction for autonomous driving through historical motion priors and diffusion modeling.

Abstract: Video prediction is a useful function for autonomous driving, enabling intelligent vehicles to reliably anticipate how driving scenes will evolve and thereby supporting reasoning and safer planning. However, existing models are constrained by multi-stage training pipelines and remain insufficient in modeling the diverse motion patterns in real driving scenes, leading to degraded temporal consistency and visual quality. To address these challenges, this paper introduces the historical motion priors-informed diffusion model (HMPDM), a video prediction model that leverages historical motion priors to enhance motion understanding and temporal coherence. The proposed deep learning system introduces three key designs: (i) a Temporal-aware Latent Conditioning (TaLC) module for implicit historical motion injection; (ii) a Motion-aware Pyramid Encoder (MaPE) for multi-scale motion representation; (iii) a Self-Conditioning (SC) strategy for stable iterative denoising. Extensive experiments on the Cityscapes and KITTI benchmarks demonstrate that HMPDM outperforms state-of-the-art video prediction methods with efficiency, achieving a 28.2% improvement in FVD on Cityscapes under the same monocular RGB input configuration setting. The implementation codes are publicly available at https://github.com/KELISBU/HMPDM.

Method: KAWHI adaptively localizes semantically salient regions through hierarchical geometric aggregation, identifies vision-critical attention heads via structured attribution, and performs paragraph-level credit reallocation to align spatial visual evidence with semantically decisive reasoning steps.

Result: Extensive empirical evaluations on diverse reasoning benchmarks show KAWHI consistently improves performance of various uniform reward optimization methods (like GRPO and GSPO) as a general-purpose enhancement module.

Conclusion: KAWHI successfully addresses the visual representation bottleneck in RLVR for LVLMs by explicitly incorporating structured visual information into reward optimization, enabling better multimodal reasoning capabilities.

Abstract: Reinforcement Learning from Verifiable Rewards (RLVR) has substantially enhanced the reasoning capabilities of large language models in abstract reasoning tasks. However, its application to Large Vision-Language Models (LVLMs) remains constrained by a structural representational bottleneck. Existing approaches generally lack explicit modeling and effective utilization of visual information, preventing visual representations from being tightly coupled with the reinforcement learning optimization process and thereby limiting further improvements in multimodal reasoning performance. To address this limitation, we propose KAWHI (Key-Region Aligned Weighted Harmonic Incentive), a plug-and-play reward reweighting mechanism that explicitly incorporates structured visual information into uniform reward policy optimization methods (e.g., GRPO and GSPO). The method adaptively localizes semantically salient regions through hierarchical geometric aggregation, identifies vision-critical attention heads via structured attribution, and performs paragraph-level credit reallocation to align spatial visual evidence with semantically decisive reasoning steps. Extensive empirical evaluations on diverse reasoning benchmarks substantiate KAWHI as a general-purpose enhancement module, consistently improving the performance of various uniform reward optimization methods. Project page: KAWHI (https://kawhiiiileo.github.io/KAWHI_PAGE/)

Method: CRISP factorizes pretrained weights into basis matrices and component mixing projections. It shares basis matrices across layers (enabling model compression) and uses small mixer weights (enabling PEFT). The framework supports both tasks simultaneously through this factorization approach.

Result: CRISP outperforms prior dual-task methods by 4-5%, outperforms state-of-the-art PEFT methods by 1.5%, and outperforms PEFT+MC combinations by 1%.

Conclusion: CRISP provides a general framework that seamlessly integrates multiple parameter recombination tasks, offering improved performance for both model compression and parameter-efficient fine-tuning applications.

Abstract: Parameter Recombination (PR) methods aim to efficiently compose the weights of a neural network for applications like Parameter-Efficient FineTuning (PEFT) and Model Compression (MC), among others. Most methods typically focus on one application of PR, which can make composing them challenging. For example, when deploying a large model you may wish to compress the model and also quickly adapt to new settings. However, PEFT methods often can still contain millions of parameters. This may be small compared to the original model size, but can be problematic in resource constrained deployments like edge devices, where they take a larger portion of the compressed model’s parameters. To address this, we present Coefficient-gated weight Recombination by Interpolated Shared basis Projections (CRISP), a general approach that seamlessly integrates multiple PR tasks within the same framework. CRISP accomplishes this by factorizing pretrained weights into basis matrices and their component mixing projections. Sharing basis matrices across layers and adjusting its size enables us to perform MC, whereas the mixer weight’s small size (fewer than 200 in some experiments) enables CRISP to support PEFT. Experiments show CRISP outperforms methods from prior work capable of dual-task applications by 4-5% while also outperforming the state-of-the-art in PEFT by 1.5% and PEFT+MC combinations by 1%. Our code is available on the repository: https://github.com/appledora/CRISP-CVPR26.

Method: 1) Created PEAR benchmark with joint 6D pose and per-instance 3D deformation ground truth across 8 produce categories using robotic manipulator for high annotation accuracy. 2) Proposed SEED framework that jointly predicts 6D pose and explicit lattice deformations from single RGB images across multiple categories, trained entirely on synthetic data with generative texture augmentation at UV level.

Result: State-of-the-art methods suffer up to 6x performance degradation on real-world produce. SEED outperforms MegaPose on 6 out of 8 categories under identical RGB-only conditions, demonstrating explicit shape modeling improves pose estimation reliability.

Conclusion: Explicit shape modeling is critical for reliable pose estimation in agricultural robotics. The PEAR benchmark enables evaluation of joint pose and deformation estimation, while SEED provides a unified RGB-only framework that effectively handles produce deformability.

Abstract: Accurate 6D pose estimation for robotic harvesting is fundamentally hindered by the biological deformability and high intra-class shape variability of agricultural produce. Instance-level methods fail in this setting, as obtaining exact 3D models for every unique piece of produce is practically infeasible, while category-level approaches that rely on a fixed template suffer significant accuracy degradation when the prior deviates from the true instance geometry. To bridge such lack of robustness to deformation, we introduce PEAR (Pose and dEformation of Agricultural pRoduce), the first benchmark providing joint 6D pose and per-instance 3D deformation ground truth across 8 produce categories, acquired via a robotic manipulator for high annotation accuracy. Using PEAR, we show that state-of-the-art methods suffer up to 6x performance degradation when faced with the inherent geometric deviations of real-world produce. Motivated by this finding, we propose SEED (Simultaneous Estimation of posE and Deformation), a unified RGB-only framework that jointly predicts 6D pose and explicit lattice deformations from a single image across multiple produce categories. Trained entirely on synthetic data with generative texture augmentation applied at the UV level, SEED outperforms MegaPose on 6 out of 8 categories under identical RGB-only conditions, demonstrating that explicit shape modeling is a critical step toward reliable pose estimation in agricultural robotics.

Method: Proposes SpatialStack, a hierarchical fusion framework that stacks and synchronizes multi-level geometric features with the language backbone. Instead of conventional late-stage fusion, it progressively aligns vision, geometry, and language representations across the model hierarchy to capture both local geometric precision and global contextual semantics.

Result: VLM-SpatialStack achieves state-of-the-art performance on multiple 3D spatial reasoning benchmarks. Extensive experiments show the multi-level fusion strategy consistently enhances 3D understanding and generalizes robustly across diverse spatial reasoning tasks.

Conclusion: SpatialStack establishes an effective and extensible design paradigm for vision-language-geometry integration in next-generation multimodal physical AI systems, overcoming limitations of current approaches by enabling hierarchical feature alignment across modalities.

Abstract: Large vision-language models (VLMs) still struggle with reliable 3D spatial reasoning, a core capability for embodied and physical AI systems. This limitation arises from their inability to capture fine-grained 3D geometry and spatial relationships. While recent efforts have introduced multi-view geometry transformers into VLMs, they typically fuse only the deep-layer features from vision and geometry encoders, discarding rich hierarchical signals and creating a fundamental bottleneck for spatial understanding. To overcome this, we propose SpatialStack, a general hierarchical fusion framework that progressively aligns vision, geometry, and language representations across the model hierarchy. Moving beyond conventional late-stage vision-geometry fusion, SpatialStack stacks and synchronizes multi-level geometric features with the language backbone, enabling the model to capture both local geometric precision and global contextual semantics. Building upon this framework, we develop VLM-SpatialStack, a model that achieves state-of-the-art performance on multiple 3D spatial reasoning benchmarks. Extensive experiments and ablations demonstrate that our multi-level fusion strategy consistently enhances 3D understanding and generalizes robustly across diverse spatial reasoning tasks, establishing SpatialStack as an effective and extensible design paradigm for vision-language-geometry integration in next-generation multimodal physical AI systems.

Method: Compared SAM2 (transformer-based) with lightweight YOLOv8 on new dataset of 1,456 annotated RGB images of laptop components (logic boards, heat sinks, fans) under varying conditions; used data augmentation (rotation, flipping, cropping) for robustness.

Result: YOLOv8 achieved much higher segmentation accuracy (mAP50 = 98.8%, mAP50-95 = 85%) and better boundary precision than SAM2 (mAP50 = 8.4%). SAM2 showed flexibility but produced overlapping masks and inconsistent contours.

Conclusion: Large pre-trained models like SAM2 require task-specific optimization for industrial applications; YOLOv8 performs better for this specific segmentation task; dataset and framework support scalable vision algorithms for robotic e-waste disassembly.

Abstract: Precise segmentation of irregular and densely arranged components is essential for robotic disassembly and material recovery in electronic waste (e-waste) recycling. This study evaluates the impact of model architecture and scale on segmentation performance by comparing SAM2, a transformer-based vision model, with the lightweight YOLOv8 network. Both models were trained and tested on a newly collected dataset of 1,456 annotated RGB images of laptop components including logic boards, heat sinks, and fans, captured under varying illumination and orientation conditions. Data augmentation techniques, such as random rotation, flipping, and cropping, were applied to improve model robustness. YOLOv8 achieved higher segmentation accuracy (mAP50 = 98.8%, mAP50-95 = 85%) and stronger boundary precision than SAM2 (mAP50 = 8.4%). SAM2 demonstrated flexibility in representing diverse object structures but often produced overlapping masks and inconsistent contours. These findings show that large pre-trained models require task-specific optimization for industrial applications. The resulting dataset and benchmarking framework provide a foundation for developing scalable vision algorithms for robotic e-waste disassembly and circular manufacturing systems.

Method: LOME jointly estimates spatial human actions and environment contexts during training, then fine-tunes a pretrained video generative model on diverse egocentric human-object interaction videos. It conditions generation on input image, text prompt, and per-frame human actions (body poses and hand gestures).

Result: LOME demonstrates high action-following accuracy, strong generalization to unseen scenarios, realistic physical consequences of hand-object interactions (like liquid flowing), and significantly outperforms state-of-the-art image/video-based action-conditioned methods and I/T2V models in temporal consistency and motion control.

Conclusion: LOME enables photorealistic AR/VR experiences and scalable robotic training without being limited to simulated environments or explicit 3D/4D modeling, advancing realistic human-object interaction generation.

Abstract: Learning human-object manipulation presents significant challenges due to its fine-grained and contact-rich nature of the motions involved. Traditional physics-based animation requires extensive modeling and manual setup, and more importantly, it neither generalizes well across diverse object morphologies nor scales effectively to real-world environment. To address these limitations, we introduce LOME, an egocentric world model that can generate realistic human-object interactions as videos conditioned on an input image, a text prompt, and per-frame human actions, including both body poses and hand gestures. LOME injects strong and precise action guidance into object manipulation by jointly estimating spatial human actions and the environment contexts during training. After finetuning a pretrained video generative model on videos of diverse egocentric human-object interactions, LOME demonstrates not only high action-following accuracy and strong generalization to unseen scenarios, but also realistic physical consequences of hand-object interactions, e.g., liquid flowing from a bottle into a mug after executing a ``pouring’’ action. Extensive experiments demonstrate that our video-based framework significantly outperforms state-of-the-art image based and video-based action-conditioned methods and Image/Text-to-Video (I/T2V) generative model in terms of both temporal consistency and motion control. LOME paves the way for photorealistic AR/VR experiences and scalable robotic training, without being limited to simulated environments or relying on explicit 3D/4D modeling.

Method: Uses a feed-forward framework that reconstructs 3D Gaussians from uncalibrated context views and renders target views with self-predicted camera parameters. Employs a shared transformer backbone with masked attention for joint reconstruction and camera prediction, and a depth-based Gaussian formulation for stable optimization.

Result: Achieves superior results compared to other self-supervised methods, establishing a scalable and geometry-aware paradigm for 3D reconstruction from unconstrained data.

Conclusion: NAS3R provides an effective self-supervised approach for 3D reconstruction that works with uncalibrated, unposed images and is compatible with state-of-the-art supervised architectures.

Abstract: In this paper, we introduce NAS3R, a self-supervised feed-forward framework that jointly learns explicit 3D geometry and camera parameters with no ground-truth annotations and no pretrained priors. During training, NAS3R reconstructs 3D Gaussians from uncalibrated and unposed context views and renders target views using its self-predicted camera parameters, enabling self-supervised training from 2D photometric supervision. To ensure stable convergence, NAS3R integrates reconstruction and camera prediction within a shared transformer backbone regulated by masked attention, and adopts a depth-based Gaussian formulation that facilitates well-conditioned optimization. The framework is compatible with state-of-the-art supervised 3D reconstruction architectures and can incorporate pretrained priors or intrinsic information when available. Extensive experiments show that NAS3R achieves superior results to other self-supervised methods, establishing a scalable and geometry-aware paradigm for 3D reconstruction from unconstrained data. Code and models are publicly available at https://ranrhuang.github.io/nas3r/.

Method: Comprehensive survey of over 1,000 open-access medical image datasets with systematic cataloging, analysis of fragmentation, and proposal of metadata-driven fusion paradigm (MDFP) to integrate datasets.

Result: Analysis shows medical image datasets are modest in scale, fragmented across narrow tasks, unevenly distributed across organs/modalities. Created interactive portal and unified structured table of all surveyed datasets.

Conclusion: Survey provides roadmap for scaling medical imaging corpora through dataset consolidation, supporting faster data discovery and more capable medical foundation models.

Abstract: Foundation models have demonstrated remarkable success across diverse domains and tasks, primarily due to the thrive of large-scale, diverse, and high-quality datasets. However, in the field of medical imaging, the curation and assembling of such medical datasets are highly challenging due to the reliance on clinical expertise and strict ethical and privacy constraints, resulting in a scarcity of large-scale unified medical datasets and hindering the development of powerful medical foundation models. In this work, we present the largest survey to date of medical image datasets, covering over 1,000 open-access datasets with a systematic catalog of their modalities, tasks, anatomies, annotations, limitations, and potential for integration. Our analysis exposes a landscape that is modest in scale, fragmented across narrowly scoped tasks, and unevenly distributed across organs and modalities, which in turn limits the utility of existing medical image datasets for developing versatile and robust medical foundation models. To turn fragmentation into scale, we propose a metadata-driven fusion paradigm (MDFP) that integrates public datasets with shared modalities or tasks, thereby transforming multiple small data silos into larger, more coherent resources. Building on MDFP, we release an interactive discovery portal that enables end-to-end, automated medical image dataset integration, and compile all surveyed datasets into a unified, structured table that clearly summarizes their key characteristics and provides reference links, offering the community an accessible and comprehensive repository. By charting the current terrain and offering a principled path to dataset consolidation, our survey provides a practical roadmap for scaling medical imaging corpora, supporting faster data discovery, more principled dataset creation, and more capable medical foundation models.

Method: Proposes Differential Feedback which automatically constructs token/step-level supervision masks by repairing erroneous reasoning trajectories, explicitly marking key positions requiring correction. Can be integrated into existing GRPO-like frameworks without large-scale step-by-step human annotations.

Result: Experiments on multimodal reasoning benchmarks (MMMStar and MathVista) show average 3% improvement under matched compute budgets.

Conclusion: The approach offers an effective, low-cost solution for accurate vision-reasoning process alignment in VLMs.

Abstract: Vision–language models (VLMs) are increasingly aligned via Group Relative Policy Optimization (GRPO)-style training. However, relying solely on terminal outcome rewards yields sparse credit assignment in multi-step reasoning, weakening the linkage between visual evidence and intermediate steps and often causing unstable optimization and visual hallucinations. We propose Differential Feedback, which automatically constructs token/step-level supervision masks by repairing erroneous reasoning trajectories, explicitly marking the key positions that require correction. Without costly large-scale step-by-step human annotations, our method enables process-level visual alignment and can be seamlessly integrated into existing GRPO-like frameworks. Experiments on multimodal reasoning benchmarks including MMMStar and MathVista show an average 3% improvement under matched compute budgets. Our approach offers an effective, low-cost solution for accurate vision–reasoning process alignment.

Method: Used Google Earth Engine satellite imagery datasets, developed car-counting models using Detectron2 and Faster R-CNN frameworks to detect vehicles at various locations (university, shopping mall, plaza, restaurant, supermarket) before and during COVID-19.

Result: Car counts reduced by average 30% in 2020 compared to 2019 across monitored locations, demonstrating satellite imagery’s effectiveness for travel demand and economic activity estimation.

Conclusion: Satellite imagery combined with computer vision/deep learning provides reliable information for transportation decision-making and can effectively monitor infrastructure usage and economic trends.

Abstract: Considering recent advances in remote sensing satellite systems and computer vision algorithms, many satellite sensing platforms and sensors have been used to monitor the condition and usage of transportation infrastructure systems. The level of details that can be detected increases significantly with the increase of ground sample distance (GSD), which is around 15 cm - 30 cm for high-resolution satellite images. In this study, we analyzed data acquired from high-resolution satellite imagery to provide insights, predictive signals, and trend for travel demand estimation. More specifically, we estimate the impact of COVID-19 in the metropolitan area of Houston using satellite imagery from Google Earth Engine datasets. We developed a car-counting model through Detectron2 and Faster R-CNN to monitor the presence of cars within different locations (i.e., university, shopping mall, community plaza, restaurant, supermarket) before and during the COVID-19. The results show that the number of cars detected at these selected locations reduced on average 30% in 2020 compared with the previous year 2019. The results also show that satellite imagery provides rich information for travel demand and economic activity estimation. Together with advanced computer vision and deep learning algorithms, it can generate reliable and accurate information for transportation agency decision makers.

Method: Proposes STATrack, a fully spiking neural network framework for UAV visual tracking with RGB inputs. Introduces adaptive mutual information maximization between templates and features to mitigate target feature weakening by background tokens.

Result: Extensive experiments on four UAV tracking benchmarks show STATrack achieves competitive tracking performance while maintaining low energy consumption.

Conclusion: STATrack demonstrates that efficient SNN-based visual tracking is possible with standard RGB inputs, making it more practical for UAV deployment compared to event camera-dependent approaches.

Abstract: Spiking Neural Networks (SNNs), characterized by their event-driven computation and low power consumption, have shown great potential for energy-efficient visual tracking on unmanned aerial vehicles (UAVs). However, existing efficient SNN-based trackers heavily rely on costly event cameras, limiting their deployment on UAVs. To address this limitation, we propose STATrack, an efficient fully spiking neural network framework for UAV visual tracking using RGB inputs only. To the best of our knowledge, this work is the first to investigate spiking neural networks for UAV visual tracking tasks. To mitigate the weakening of target features by background tokens, we propose adaptively maximizing the mutual information between templates and features. Extensive experiments on four widely used UAV tracking benchmarks demonstrate that STATrack achieves competitive tracking performance while maintaining low energy consumption.

Method: Two-stage RL framework: 1) “Information Gap” mechanism adjusts global image granularity to train models to focus on cropped key regions based on information gain; 2) Incorporates grounding loss with bounding box annotations to enhance cropping precision.

Result: Significantly enhances model attention to cropped regions and achieves state-of-the-art performance on high-resolution visual question-answering benchmarks.

Conclusion: Provides a more efficient approach for perceiving and reasoning fine-grained details in MLLMs without requiring trajectory supervision.

Abstract: To enhance the perception and reasoning capabilities of multimodal large language models in complex visual scenes, recent research has introduced agent-based workflows. In these works, MLLMs autonomously utilize image cropping tool to analyze regions of interest for question answering. While existing training strategies, such as those employing supervised fine-tuning and reinforcement learning, have made significant progress, our empirical analysis reveals a key limitation. We demonstrate the model’s strong reliance on global input and its weak dependence on the details within the cropped region. To address this issue, we propose a novel two-stage reinforcement learning framework that does not require trajectory supervision. In the first stage, we introduce the ``Information Gap” mechanism by adjusting the granularity of the global image. This mechanism trains the model to answer questions by focusing on cropped key regions, driven by the information gain these regions provide. The second stage further enhances cropping precision by incorporating a grounding loss, using a small number of bounding box annotations. Experiments show that our method significantly enhances the model’s attention to cropped regions, enabling it to achieve state-of-the-art performance on high-resolution visual question-answering benchmarks. Our method provides a more efficient approach for perceiving and reasoning fine-grained details in MLLMs. Code is available at: https://github.com/XuanPu-Z/LFPC.

Method: Leverages DINOv3’s backbone for fine-grained localization and its text-aligned vision head for open-vocabulary classification, with cross-attention between interaction queries and vision head outputs, keeping all DINOv3 parameters frozen.

Result: Achieves state-of-the-art performance on both SWiG-HOI and HICO-DET benchmarks, demonstrating effectiveness of the streamlined architecture.

Conclusion: SL-HOI provides an effective streamlined solution for open-vocabulary HOI detection by fully utilizing DINOv3’s capabilities with minimal additional parameters.

Abstract: Open-vocabulary human-object interaction (HOI) detection aims to localize and recognize all human-object interactions in an image, including those unseen during training. Existing approaches usually rely on the collaboration between a conventional HOI detector and a Vision-Language Model (VLM) to recognize unseen HOI categories. However, feature fusion in this paradigm is challenging due to significant gaps in cross-model representations. To address this issue, we introduce SL-HOI, a StreamLined open-vocabulary HOI detection framework based solely on the powerful DINOv3 model. Our design leverages the complementary strengths of DINOv3’s components: its backbone for fine-grained localization and its text-aligned vision head for open-vocabulary interaction classification. Moreover, to facilitate smooth cross-attention between the interaction queries and the vision head’s output, we propose first feeding both the interaction queries and the backbone image tokens into the vision head, effectively bridging their representation gaps. All DINOv3 parameters in our approach are frozen, with only a small number of learnable parameters added, allowing a fast adaptation to the HOI detection task. Extensive experiments show that SL-HOI achieves state-of-the-art performance on both the SWiG-HOI and HICO-DET benchmarks, demonstrating the effectiveness of our streamlined model architecture. Code is available at https://github.com/MPI-Lab/SL-HOI.

Method: 1) Construct Physical-Centric Knowledge Graph (PCKG) using LLMs to extract physical priors from 1,763 vocabularies; 2) Build Phy-Sky-SA dataset (heterogeneous, spatial-aligned); 3) Develop PriorSeg with physics-aware residual refinement and joint visual-physical training with physics-consistency loss.

Result: PriorSeg improves segmentation accuracy and physical plausibility in heterogeneous settings without retraining foundation models. Ablation studies confirm effectiveness of Phy-Sky-SA dataset, PCKG, and physics-consistency loss.

Conclusion: The proposed paradigm successfully integrates domain-specific physical priors into segmentation models, overcoming limitations of current foundation models while maintaining accuracy and physical consistency without costly retraining.

Abstract: Semantic segmentation of remote sensing imagery is fundamental to Earth observation. Achieving accurate results requires integrating not only optical images but also physical variables such as the Digital Elevation Model (DEM), Synthetic Aperture Radar (SAR) and Normalized Difference Vegetation Index (NDVI). Recent foundation models (FMs) leverage pre-training to exploit these variables but still depend on spatially aligned data and costly retraining when involving new sensors. To overcome these limitations, we introduce a novel paradigm for integrating domain-specific physical priors into segmentation models. We first construct a Physical-Centric Knowledge Graph (PCKG) by prompting large language models to extract physical priors from 1,763 vocabularies, and use it to build a heterogeneous, spatial-aligned dataset, Phy-Sky-SA. Building on this foundation, we develop PriorSeg, a physics-aware residual refinement model trained with a joint visual-physical strategy that incorporates a novel physics-consistency loss. Experiments on heterogeneous settings demonstrate that PriorSeg improves segmentation accuracy and physical plausibility without retraining the FMs. Ablation studies verify the effectiveness of the Phy-Sky-SA dataset, the PCKG, and the physics-consistency loss.

Method: Represents 3D scenes as sequences of objects with contextual semantics, using identifier tokens and extracting context-rich object features from pre-trained 3D scene-level and 2D image-level encoders. Supports grounded chain-of-thought reasoning for multi-step inference.

Result: Achieves state-of-the-art performance on five major 3D vision-language benchmarks (ScanRefer, Multi3DRefer, Scan2Cap, ScanQA, SQA3D) and demonstrates applicability to real-world scenarios using only 2D inputs.

Conclusion: Chat-Scene++ effectively addresses limitations in 3D scene understanding through object-centric representation and contextual reasoning, showing strong performance across diverse 3D vision-language tasks.

Abstract: Recent advancements in multi-modal large language models (MLLMs) have shown strong potential for 3D scene understanding. However, existing methods struggle with fine-grained object grounding and contextual reasoning, limiting their ability to interpret and interact with complex 3D environments. In this paper, we present Chat-Scene++, an MLLM framework that represents 3D scenes as context-rich object sequences. By structuring scenes as sequences of objects with contextual semantics, Chat-Scene++ enables object-centric representation and interaction. It decomposes a 3D scene into object representations paired with identifier tokens, allowing LLMs to follow instructions across diverse 3D vision-language tasks. To capture inter-object relationships and global semantics, Chat-Scene++ extracts context-rich object features using large-scale pre-trained 3D scene-level and 2D image-level encoders, unlike the isolated per-object features in Chat-Scene. Its flexible object-centric design also supports grounded chain-of-thought (G-CoT) reasoning, enabling the model to distinguish objects at both category and spatial levels during multi-step inference. Without the need for additional task-specific heads or fine-tuning, Chat-Scene++ achieves state-of-the-art performance on five major 3D vision-language benchmarks: ScanRefer, Multi3DRefer, Scan2Cap, ScanQA, and SQA3D. These results highlight its effectiveness in scene comprehension, object grounding, and spatial reasoning. Additionally, without reconstructing 3D worlds through computationally expensive processes, we demonstrate its applicability to real-world scenarios using only 2D inputs.

Method: A multi-stage framework using off-the-shelf models for object detection, mask generation, and semantically guided inpainting/regeneration to produce controlled, meaning-altering edits with minimal perceptual degradation for systematic stress-testing.

Result: Robustness varies significantly with semantic entanglement - methods that remain detectable under conventional perturbations often fail under semantic edits, with watermark detectability dropping to near zero while image quality remains high.

Conclusion: Current watermarking evaluations have a critical gap; watermark designs and benchmarking must explicitly account for robustness against semantic manipulation.

Abstract: The widespread deployment of high-fidelity generative models has intensified the need for reliable mechanisms for provenance and content authentication. In-processing watermarking, embedding a signature into the generative model’s synthesis procedure, has been advocated as a solution and is often reported to be robust to standard post-processing (such as geometric transforms and filtering). Yet robustness to semantic manipulations that alter high-level scene content while maintaining reasonable visual quality is not well studied or understood. We introduce a simple, multi-stage framework for systematically stress-testing in-processing generative watermarks under semantic drift. The framework utilizes off-the-shelf models for object detection, mask generation, and semantically guided inpainting or regeneration to produce controlled, meaning-altering edits with minimal perceptual degradation. Based on extensive experiments on representative schemes, we find that robustness varies significantly with the degree of semantic entanglement: methods by which watermarks remain detectable under a broad suite of conventional perturbations can fail under semantic edits, with watermark detectability in many cases dropping to near zero while image quality remains high. Overall, our results reveal a critical gap in current watermarking evaluations and suggest that watermark designs and benchmarking must explicitly account for robustness against semantic manipulation.

Method: 1. Constructs dense geometry-consistent Gaussian semantic field using semantic and geometric priors. 2. Performs material-illumination disentanglement with hybrid illumination model and material prior. 3. Introduces illumination-invariant material constraint and deshadowing model to mitigate cast shadows and enhance material recovery robustness.

Result: Extensive experiments on benchmark datasets show consistent improvements in both reconstruction fidelity and inverse rendering quality over existing 3DGS-based inverse rendering approaches.

Conclusion: SGS-Intrinsic successfully addresses sparse-view inverse rendering challenges for indoor scenes, achieving high-quality geometry reconstruction and accurate material-illumination disentanglement through semantic-guided Gaussian fields and robust illumination modeling.

Abstract: We present SGS-Intrinsic, an indoor inverse rendering framework that works well for sparse-view images. Unlike existing 3D Gaussian Splatting (3DGS) based methods that focus on object-centric reconstruction and fail to work under sparse view settings, our method allows to achieve high-quality geometry reconstruction and accurate disentanglement of material and illumination. The core idea is to construct a dense and geometry-consistent Gaussian semantic field guided by semantic and geometric priors, providing a reliable foundation for subsequent inverse rendering. Building upon this, we perform material-illumination disentanglement by combining a hybrid illumination model and material prior to effectively capture illumination-material interactions. To mitigate the impact of cast shadows and enhance the robustness of material recovery, we introduce illumination-invariant material constraint together with a deshadowing model. Extensive experiments on benchmark datasets show that our method consistently improves both reconstruction fidelity and inverse rendering quality over existing 3DGS-based inverse rendering approaches. Our code is available at https://github.com/GrumpySloths/SGS_Intrinsic.github.io.

Method: Uses a VAE-free Pixel-space Diffusion Transformer trained via diffusion denoising on 2.6 million high-quality agricultural images spanning diverse crops, growth stages, and environments. The model learns rich, structure-aware representations through denoising enabling efficient end-to-end training.

Result: SPROUT consistently outperforms state-of-the-art web-pretrained and agricultural foundation models across a wide range of downstream tasks while requiring substantially lower pre-training cost.

Conclusion: SPROUT demonstrates the effectiveness of diffusion-based pre-training for agricultural vision tasks, providing a scalable solution that bridges the domain gap between general vision models and specialized agricultural applications.

Abstract: Vision Foundation Models (VFM) pre-trained on large-scale unlabeled data have achieved remarkable success on general computer vision tasks, yet typically suffer from significant domain gaps when applied to agriculture. In this context, we introduce $SPROUT$ ($S$calable $P$lant $R$epresentation model via $O$pen-field $U$nsupervised $T$raining), a multi-crop, multi-task agricultural foundation model trained via diffusion denoising. SPROUT leverages a VAE-free Pixel-space Diffusion Transformer to learn rich, structure-aware representations through denoising and enabling efficient end-to-end training. We pre-train SPROUT on a curated dataset of 2.6 million high-quality agricultural images spanning diverse crops, growth stages, and environments. Extensive experiments demonstrate that SPROUT consistently outperforms state-of-the-art web-pretrained and agricultural foundation models across a wide range of downstream tasks, while requiring substantially lower pre-training cost. The code and model are available at https://github.com/UTokyo-FieldPhenomics-Lab/SPROUT.

Method: TokenDial learns attribute-specific additive offsets in intermediate spatiotemporal visual patch-token space. It uses pretrained understanding signals: semantic direction matching for appearance attributes and motion-magnitude scaling for motion attributes, without retraining the backbone model.

Result: TokenDial achieves stronger controllability and higher-quality edits than state-of-the-art baselines across diverse attributes and prompts, validated through extensive quantitative evaluation and human studies.

Conclusion: TokenDial provides an effective framework for continuous attribute control in text-to-video generation by leveraging token-space offsets, enabling precise control over both appearance and motion dynamics while maintaining video coherence.

Abstract: We present TokenDial, a framework for continuous, slider-style attribute control in pretrained text-to-video generation models. While modern generators produce strong holistic videos, they offer limited control over how much an attribute changes (e.g., effect intensity or motion magnitude) without drifting identity, background, or temporal coherence. TokenDial is built on the observation: additive offsets in the intermediate spatiotemporal visual patch-token space form a semantic control direction, where adjusting the offset magnitude yields coherent, predictable edits for both appearance and motion dynamics. We learn attribute-specific token offsets without retraining the backbone, using pretrained understanding signals: semantic direction matching for appearance and motion-magnitude scaling for motion. We demonstrate TokenDial’s effectiveness on diverse attributes and prompts, achieving stronger controllability and higher-quality edits than state-of-the-art baselines, supported by extensive quantitative evaluation and human studies.

Method: The framework categorizes guidance signals into spatially-aligned conditions (using dual-path encoding with Dense Feature Alignment loss) and semantic-reference conditions (using VLM-only encoding with Temporal Redundancy Elimination). An Adaptive Spatial-Semantic Gating module dynamically balances multi-modal constraints to resolve input conflicts.

Result: Experimental results show OmniColor achieves superior controllability, visual quality, and temporal stability compared to existing methods, providing a robust practical solution for lineart colorization.

Conclusion: OmniColor offers a unified framework that effectively handles diverse control signals for lineart colorization, balancing spatial precision with semantic reference efficiency through innovative architectural components.

Abstract: Lineart colorization is a critical stage in professional content creation, yet achieving precise and flexible results under diverse user constraints remains a significant challenge. To address this, we propose OmniColor, a unified framework for multi-modal lineart colorization that supports arbitrary combinations of control signals. Specifically, we systematically categorize guidance signals into two types: spatially-aligned conditions and semantic-reference conditions. For spatially-aligned inputs, we employ a dual-path encoding strategy paired with a Dense Feature Alignment loss to ensure rigorous boundary preservation and precise color restoration. For semantic-reference inputs, we utilize a VLM-only encoding scheme integrated with a Temporal Redundancy Elimination mechanism to filter repetitive information and enhance inference efficiency. To resolve potential input conflicts, we introduce an Adaptive Spatial-Semantic Gating module that dynamically balances multi-modal constraints. Experimental results demonstrate that OmniColor achieves superior controllability, visual quality, and temporal stability, providing a robust and practical solution for lineart colorization. The source code and dataset will be open at https://github.com/zhangxulu1996/OmniColor.

Method: Proposes DeMo-Pose with: 1) Hybrid architecture fusing monocular semantic features with depth-based graph convolutional representations via novel multimodal fusion strategy, 2) Mesh-Point Loss (MPL) that leverages mesh structure during training without inference overhead to improve geometric reasoning.

Result: Achieves real-time inference and outperforms state-of-the-art methods, including GPV-Pose baseline by 3.2% on 3D IoU and 11.1% on pose accuracy on REAL275 benchmark.

Conclusion: Demonstrates effectiveness of depth-RGB fusion and geometry-aware learning for robust category-level 3D pose estimation in real-world applications.

Abstract: Object pose estimation is a fundamental task in 3D vision with applications in robotics, AR/VR, and scene understanding. We address the challenge of category-level 9-DoF pose estimation (6D pose + 3Dsize) from RGB-D input, without relying on CAD models during inference. Existing depth-only methods achieve strong results but ignore semantic cues from RGB, while many RGB-D fusion models underperform due to suboptimal cross-modal fusion that fails to align semantic RGB cues with 3D geometric representations. We propose DeMo-Pose, a hybrid architecture that fuses monocular semantic features with depth-based graph convolutional representations via a novel multimodal fusion strategy. To further improve geometric reasoning, we introduce a novel Mesh-Point Loss (MPL) that leverages mesh structure during training without adding inference overhead. Our approach achieves real-time inference and significantly improves over state-of-the-art methods across object categories, outperforming the strong GPV-Pose baseline by 3.2% on 3D IoU and 11.1% on pose accuracy on the REAL275 benchmark. The results highlight the effectiveness of depth-RGB fusion and geometry-aware learning, enabling robust category-level 3D pose estimation for real-world applications.

Method: Designs an efficient architecture with: (1) multi-view encoder using pairwise matching as geometric prior, (2) multi-view matching refiner using pixel-wise attention, and (3) post-processing to integrate consistent multi-view correspondences as high-quality tracks for SfM.

Result: Produces more reliable correspondences and substantially denser, more accurate 3D reconstructions than existing sparse and dense matching methods across diverse benchmarks.

Conclusion: MV-RoMa effectively addresses the limitations of pairwise matching by providing geometrically consistent multi-view correspondences, improving 3D reconstruction quality.

Abstract: Establishing consistent correspondences across images is essential for 3D vision tasks such as structure-from-motion (SfM), yet most existing matchers operate in a pairwise manner, often producing fragmented and geometrically inconsistent tracks when their predictions are chained across views. We propose MV-RoMa, a multi-view dense matching model that jointly estimates dense correspondences from a source image to multiple co-visible targets. Specifically, we design an efficient model architecture which avoids high computational cost of full cross-attention for multi-view feature interaction: (i) multi-view encoder that leverages pair-wise matching results as a geometric prior, and (ii) multi-view matching refiner that refines correspondences using pixel-wise attention. Additionally, we propose a post-processing strategy that integrates our model’s consistent multi-view correspondences as high-quality tracks for SfM. Across diverse and challenging benchmarks, MV-RoMa produces more reliable correspondences and substantially denser, more accurate 3D reconstructions than existing sparse and dense matching methods. Project page: https://icetea-cv.github.io/mv-roma/.

Method: Proposes MADL (map-based automatic data labeler) module combining LiDAR mapping/localization with curb detection to generate training data automatically. Uses LiDAR mapping to avoid occlusion/sparsity issues and includes a data review agent to filter low-quality samples.

Result: Experiments on KITTI, KITTI-CARLA and 3D-Curb datasets show MADL achieves impressive performance compared to manual labeling, and outperforms traditional and state-of-the-art self-supervised methods in robustness and accuracy.

Conclusion: MADL provides an effective automated solution for generating training data for drivable area and curb detection, overcoming limitations of manual labeling and previous automated methods.

Abstract: Drivable areas and curbs are critical traffic elements for autonomous driving, forming essential components of the vehicle visual perception system and ensuring driving safety. Deep neural networks (DNNs) have significantly improved perception performance for drivable area and curb detection, but most DNN-based methods rely on large manually labeled datasets, which are costly, time-consuming, and expert-dependent, limiting their real-world application. Thus, we developed an automated training data generation module. Our previous work generated training labels using single-frame LiDAR and RGB data, suffering from occlusion and distant point cloud sparsity. In this paper, we propose a novel map-based automatic data labeler (MADL) module, combining LiDAR mapping/localization with curb detection to automatically generate training data for both tasks. MADL avoids occlusion and point cloud sparsity issues via LiDAR mapping, creating accurate large-scale datasets for DNN training. In addition, we construct a data review agent to filter the data generated by the MADL module, eliminating low-quality samples. Experiments on the KITTI, KITTI-CARLA and 3D-Curb datasets show that MADL achieves impressive performance compared to manual labeling, and outperforms traditional and state-of-the-art self-supervised methods in robustness and accuracy.

Method: Proposes Pixel-wise Attention Dissolution to remove objects by nullifying the most correlated attention keys for masked pixels, eliminating objects from self-attention flow. Also introduces Localized Attentional Disentanglement Guidance to steer denoising toward latent manifolds favorable to clean object removal.

Result: Demonstrates superior visual fidelity and semantic plausibility compared to state-of-the-art methods, enabling precise, non-rigid, prompt-free, and scalable multi-object erasure in a single pass.

Conclusion: PANDORA provides an effective zero-shot framework for object removal that operates directly on pre-trained diffusion models without additional training or optimization, addressing key limitations of existing approaches.

Abstract: Removing objects from natural images is challenging due to difficulty of synthesizing semantically coherent content while preserving background integrity. Existing methods often rely on fine-tuning, prompt engineering, or inference-time optimization, yet still suffer from texture inconsistency, rigid artifacts, weak foreground-background disentanglement, and poor scalability for multi-object removal. We propose a novel zero-shot object removal framework, namely PANDORA, that operates directly on pre-trained text-to-image diffusion models, requiring no fine-tuning, prompts, or optimization. We propose Pixel-wise Attention Dissolution to remove object by nullifying the most correlated attention keys for masked pixels, effectively eliminating the object from self-attention flow and allowing background context to dominate reconstruction. We further introduce Localized Attentional Disentanglement Guidance to steer denoising toward latent manifolds favorable to clean object removal. Together, these components enable precise, non-rigid, prompt-free, and scalable multi-object erasure in a single pass. Experiments demonstrate superior visual fidelity and semantic plausibility compared to state-of-the-art methods. The project page is available at https://vdkhoi20.github.io/PANDORA.

Method: Proposes PICA (Progressive Domain-invariant Cross-modal Alignment) with multi-level ambiguity and signal strength curriculum. It builds adaptive pseudo-word prototypes refined via sample confidence and visual consistency to enforce invariant cross-domain modality alignment, departing from uniform training approaches.

Result: The work demonstrates that OVOD’s robustness to domain shifts is intrinsically linked to the stability of latent cross-modal alignment space. It provides both a challenging benchmark for DG-OVOD and shows that PICA improves generalization beyond static laboratory conditions.

Conclusion: The paper provides a new perspective on building truly generalizable open-vocabulary systems that extend beyond static conditions, highlighting the fundamental vulnerability of cross-modal alignment to domain shifts and offering a principled solution.

Abstract: Open-Vocabulary Object Detection (OVOD) has achieved remarkable success in generalizing to novel categories. However, this success often rests on the implicit assumption of domain stationarity. In this work, we provide a principled revisit of the OVOD paradigm, uncovering a fundamental vulnerability: the fragile coupling between visual manifolds and textual embeddings when distribution shifts occur. We first systematically formalize Domain-Generalized Open-Vocabulary Object Detection (DG-OVOD). Through empirical analysis, we demonstrate that visual shifts do not merely add noise; they cause a collapse of the latent cross-modal space where novel category visual signals detach from their semantic anchors. Motivated by these insights, we propose Progressive Domain-invariant Cross-modal Alignment (PICA). PICA departs from uniform training by introducing a multi-level ambiguity and signal strength curriculum. It builds adaptive pseudo-word prototypes, refined via sample confidence and visual consistency, to enforce invariant cross-domain modality alignment. Our findings suggest that OVOD’s robustness to domain shifts is intrinsically linked to the stability of the latent cross-modal alignment space. Our work provides both a challenging benchmark and a new perspective on building truly generalizable open-vocabulary systems that extend beyond static laboratory conditions.

Method: Proposes Event-MLLM with two key components: 1) Illumination Indicator - a learnable signal from DINOv2 branch representing exposure degradation that adaptively modulates event-RGB fusion, and 2) Illumination Correction Loss that aligns fused features with normal-light semantics in latent space. Also creates first multi-illumination event-instruction corpus for MLLMs.

Result: Event-MLLM markedly outperforms general-purpose, illumination-adaptive, and event-only baselines, setting new state-of-the-art in robust multimodal perception and reasoning under challenging illumination across 17 brightness rates (0.05x-20x).

Conclusion: Event-enhanced multimodal LLMs can achieve robust visual reasoning across extreme illumination conditions by dynamically fusing event streams with RGB frames and compensating for information loss through illumination-aware mechanisms.

Abstract: Multimodal Large Language Models (MLLMs) perform strong vision-language reasoning under standard conditions but fail in extreme illumination, where RGB inputs lose irrevocable structure and semantics. We propose Event-MLLM, an event-enhanced model that performs all-light visual reasoning by dynamically fusing event streams with RGB frames. Two key components drive our approach: an Illumination Indicator - a learnable signal derived from a DINOv2 branch that represents exposure degradation and adaptively modulates event-RGB fusion - and an Illumination Correction Loss that aligns fused features with non-degraded (normal-light) semantics in the latent space, compensating for information lost in extreme lighting. We curate the first multi-illumination event-instruction corpus for MLLMs, with 2,241 event-RGB samples (around 6 QA pairs each) across diverse scenes and 17 brightness rates (0.05x - 20x), plus an instruct-following benchmark for reasoning, counting, and fine-grained recognition under extreme lighting. Experiments show that Event-MLLM markedly outperforms general-purpose, illumination-adaptive, and event-only baselines, setting a new state of the art in robust multimodal perception and reasoning under challenging illumination.

Method: Divide RGB-D images into N*N grid, extract semantic, color, and depth information for each cell to form structured text descriptions, then concatenate with language instructions as pure language input to pre-trained language models.

Result: SOL-Nav significantly reduces model size and training data dependency, fully leverages PLM capabilities, achieves strong generalization to unseen environments on R2R and RxR benchmarks, and demonstrates real-world deployment success.

Conclusion: Converting visual observations to structured language enables efficient VLN by leveraging pre-trained language models’ reasoning capabilities, offering a promising direction for generalizable embodied AI.

Abstract: Vision-Language Navigation (VLN) requires an embodied agent to navigate complex environments by following natural language instructions, which typically demands tight fusion of visual and language modalities. Existing VLN methods often convert raw images into visual tokens or implicit features, requiring large-scale visual pre-training and suffering from poor generalization under environmental variations (e.g., lighting, texture). To address these issues, we propose SOL-Nav (Structured Observation Language for Navigation), a novel framework that translates egocentric visual observations into compact structured language descriptions for efficient and generalizable navigation. Specifically, we divide RGB-D images into a N*N grid, extract representative semantic, color, and depth information for each grid cell to form structured text, and concatenate this with the language instruction as pure language input to a pre-trained language model (PLM). Experimental results on standard VLN benchmarks (R2R, RxR) and real-world deployments demonstrate that SOL-Nav significantly reduces the model size and training data dependency, fully leverages the reasoning and representation capabilities of PLMs, and achieves strong generalization to unseen environments.

Method: Uses Huber loss function as data-fidelity term instead of traditional ℓ₂-norm, with total variation norm for cartoon component and nuclear norm for texture component. Employs two operator splitting algorithms tailored to different degradation operators.

Result: Extensive numerical experiments show superior performance on image restoration tasks under high-intensity heavy-tailed noise compared to conventional methods.

Conclusion: The proposed robust low-rank prior model with Huber loss effectively handles cartoon-texture decomposition in presence of heavy-tailed noise, outperforming traditional approaches.

Abstract: Cartoon-texture image decomposition is a fundamental yet challenging problem in image processing. A significant hurdle in achieving accurate decomposition is the pervasive presence of noise in the observed images, which severely impedes robust results. To address the challenging problem of cartoon-texture decomposition in the presence of heavy-tailed noise, we in this paper propose a robust low-rank prior model. Our approach departs from conventional models by adopting the Huber loss function as the data-fidelity term, rather than the traditional $\ell_2$-norm, while retaining the total variation norm and nuclear norm to characterize the cartoon and texture components, respectively. Given the inherent structure, we employ two implementable operator splitting algorithms, tailored to different degradation operators. Extensive numerical experiments, particularly on image restoration tasks under high-intensity heavy-tailed noise, efficiently demonstrate the superior performance of our model.

Method: Models proactive activation as a structured sequence modeling problem using temporal span-structured activation patterns. Employs STRIDE (Structured Temporal Refinement with Iterative DEnoising) with a lightweight masked diffusion module at the activation interface to jointly predict and progressively refine activation signals across sliding temporal windows.

Result: Extensive experiments on diverse streaming benchmarks and downstream models demonstrate that STRIDE shows more reliable and temporally coherent proactive responses, significantly improving when-to-speak decision quality in online streaming scenarios.

Conclusion: STRIDE effectively addresses the streaming video proactive activation problem by capturing span-level temporal structure through iterative refinement, enabling better when-to-speak decisions in real-time video applications.

Abstract: Recent progress in video large language models (Video-LLMs) has enabled strong offline reasoning over long and complex videos. However, real-world deployments increasingly require streaming perception and proactive interaction, where video frames arrive online and the system must decide not only what to respond, but also when to respond. In this work, we revisit proactive activation in streaming video as a structured sequence modeling problem, motivated by the observation that temporal transitions in streaming video naturally form span-structured activation patterns. To capture this span-level structure, we model activation signals jointly over a sliding temporal window and update them iteratively as new frames arrive. We propose STRIDE (Structured Temporal Refinement with Iterative DEnoising), which employs a lightweight masked diffusion module at the activation interface to jointly predict and progressively refine activation signals across the window. Extensive experiments on diverse streaming benchmarks and downstream models demonstrate that STRIDE shows more reliable and temporally coherent proactive responses, significantly improving when-to-speak decision quality in online streaming scenarios.

Method: Trains an object erasure diffusion model on unpaired real-world images using a sundries detector and context coherence loss built on an entity segmentation model, with diffusion distillation for efficient few-step training and inference.

Result: Extensive experiments show YOEO outperforms state-of-the-art object erasure methods, producing high-quality results free of unwanted objects or artifacts while preserving context coherence.

Conclusion: YOEO enables effective object erasure without paired training data by leveraging entity segmentation supervision and context coherence constraints, with efficient training via diffusion distillation.

Abstract: We present YOEO, an approach for object erasure. Unlike recent diffusion-based methods which struggle to erase target objects without generating unexpected content within the masked regions due to lack of sufficient paired training data and explicit constraint on content generation, our method allows to produce high-quality object erasure results free of unwanted objects or artifacts while faithfully preserving the overall context coherence to the surrounding content. We achieve this goal by training an object erasure diffusion model on unpaired data containing only large-scale real-world images, under the supervision of a sundries detector and a context coherence loss that are built upon an entity segmentation model. To enable more efficient training and inference, a diffusion distillation strategy is employed to train for a few-step erasure diffusion model. Extensive experiments show that our method outperforms the state-of-the-art object erasure methods. Code will be available at https://zyxunh.github.io/YOEO-ProjectPage/.

Method: Proposes Click-based Local Refinement (Clore) pipeline with hierarchical interaction: initial clicks drive global segmentation to outline large regions, while subsequent clicks progressively refine local details for precise boundaries.

Result: Experimental results on four datasets show Clore achieves the best balance between segmentation accuracy and interaction cost, outperforming existing methods.

Conclusion: Clore provides an effective solution for efficient and accurate interactive pathology image segmentation through its hierarchical local refinement approach.

Abstract: Recent advancements in deep learning-based interactive segmentation methods have significantly improved pathology image segmentation. Most existing approaches utilize user-provided positive and negative clicks to guide the segmentation process. However, these methods primarily rely on iterative global updates for refinement, which lead to redundant re-prediction and often fail to capture fine-grained structures or correct subtle errors during localized adjustments. To address this limitation, we propose the Click-based Local Refinement (Clore) pipeline, a simple yet efficient method designed to enhance interactive segmentation. The key innovation of Clore lies in its hierarchical interaction paradigm: the initial clicks drive global segmentation to rapidly outline large target regions, while subsequent clicks progressively refine local details to achieve precise boundaries. This approach not only improves the ability to handle fine-grained segmentation tasks but also achieves high-quality results with fewer interactions. Experimental results on four datasets demonstrate that Clore achieves the best balance between segmentation accuracy and interaction cost, making it an effective solution for efficient and accurate interactive pathology image segmentation.

Method: Composes learnable, panel-specific orthogonal operators onto the backbone’s frozen positional encodings. This design ensures isometry (preserves geometry of internal features) and same-panel invariance (maintains pre-trained intra-panel synthesis behavior). The method works across diverse positional encoding regimes.

Result: The adaptation method effectively enables panel-relative conditioning and consistently improves in-context image-based instructional editing pipelines, including state-of-the-art approaches. It generalizes across different positional encoding designs.

Conclusion: The proposed parameter-efficient adaptation method successfully enables panel-aware in-context image generation while preserving pre-trained model properties, offering a flexible solution for adapting diffusion transformers to multi-panel image tasks.

Abstract: We introduce a parameter-efficient adaptation method for panel-aware in-context image generation with pre-trained diffusion transformers. The key idea is to compose learnable, panel-specific orthogonal operators onto the backbone’s frozen positional encodings. This design provides two desirable properties: (1) isometry, which preserves the geometry of internal features, and (2) same-panel invariance, which maintains the model’s pre-trained intra-panel synthesis behavior. Through controlled experiments, we demonstrate that the effectiveness of our adaptation method is not tied to a specific positional encoding design but generalizes across diverse positional encoding regimes. By enabling effective panel-relative conditioning, the proposed method consistently improves in-context image-based instructional editing pipelines, including state-of-the-art approaches.

Method: Two-stage pipeline: 1) Generate class-agnostic change proposals using SAM/DINOv2, 2) Use OpenDPR’s diffusion-guided prototype retrieval for category identification (offline prototype construction with diffusion models + visual similarity retrieval), and 3) Optional S2C module for weakly supervised change localization adaptation.

Result: State-of-the-art performance on four benchmark datasets under both supervised and weakly supervised modes, demonstrating effectiveness for open-vocabulary change detection.

Conclusion: OpenDPR effectively addresses bottlenecks in open-vocabulary change detection through diffusion-guided prototype retrieval and spatial-to-change adaptation, achieving strong performance with minimal supervision.

Abstract: Open-vocabulary change detection (OVCD) seeks to recognize arbitrary changes of interest by enabling generalization beyond a fixed set of predefined classes. We reformulate OVCD as a two-stage pipeline: first generate class-agnostic change proposals using visual foundation models (VFMs) such as SAM and DINOv2, and then perform category identification with vision-language models (VLMs) such as CLIP. We reveal that category identification errors are the primary bottleneck of OVCD, mainly due to the limited ability of VLMs based on image-text matching to represent fine-grained land-cover categories. To address this, we propose OpenDPR, a training-free vision-centric diffusion-guided prototype retrieval framework. OpenDPR leverages diffusion models to construct diverse prototypes for target categories offline, and to perform similarity retrieval with change proposals in the visual space during inference. The secondary bottleneck lies in change localization, due to the inherent lack of change priors in VFMs. To bridge this gap, we design a spatial-to-change weakly supervised change detection module named S2C to adapt their strong spatial modeling capabilities for change localization. Integrating the pretrained S2C into OpenDPR leads to an optional weakly supervised variant named OpenDPR-W, which further improves OVCD with minimal supervision. Experimental results on four benchmark datasets demonstrate that the proposed methods achieve state-of-the-art performance under both supervision modes. Code is available at https://github.com/guoqi2002/OpenDPR.

Method: 1) Retrieves morphologically compatible supports from an archive using DINOv3 features to reduce brittleness in single-support choice; 2) Adapts retrieved support mask to query by fitting boundary-aware structural cues for anatomy-consistent pre-mask; 3) Converts pre-mask into prompts via Voronoi partitioning (positive points) and sector-based sampling (negative points), feeding them into SAM2 for final refinement without fine-tuning.

Result: Extensive experiments on multiple medical segmentation benchmarks show RAP consistently surpasses prior FSMIS baselines and achieves state-of-the-art performance.

Conclusion: RAP demonstrates that explicit structural fitting combined with retrieval-augmented prompting offers a simple and effective route to robust training-free few-shot medical segmentation.

Abstract: Few-shot medical image segmentation (FSMIS) has achieved notable progress, yet most existing methods mainly rely on semantic correspondences from scarce annotations while under-utilizing a key property of medical imagery: anatomical targets exhibit repeatable high-frequency morphology (e.g., boundary geometry and spatial layout) across patients and acquisitions. We propose RAP, a training-free framework that retrieves, adapts, and prompts Segment Anything Model 2 (SAM2) for FSMIS. First, RAP retrieves morphologically compatible supports from an archive using DINOv3 features to reduce brittleness in single-support choice. Second, it adapts the retrieved support mask to the query by fitting boundary-aware structural cues, yielding an anatomy-consistent pre-mask under domain shifts. Third, RAP converts the pre-mask into prompts by sampling positive points via Voronoi partitioning and negative points via sector-based sampling, and feeds them into SAM2 for final refinement without any fine-tuning. Extensive experiments on multiple medical segmentation benchmarks show that RAP consistently surpasses prior FSMIS baselines and achieves state-of-the-art performance. Overall, RAP demonstrates that explicit structural fitting combined with retrieval-augmented prompting offers a simple and effective route to robust training-free few-shot medical segmentation.

Method: V-CAST frames token compression as trajectory approximation with: 1) curvature-guided temporal allocation that routes token budgets to semantic turns and event boundaries, and 2) dual-anchor spatial selection that preserves high-entropy visual evidence while maintaining original positional coordinates.

Result: Achieves 98.6% of original performance, outperforms second-best method by +1.1% on average, reduces peak memory to 86.7% and total latency to 86.4% of vanilla Qwen3-VL-8B-Instruct across multiple VideoLLM architectures and scales.

Conclusion: V-CAST provides an effective training-free pruning solution for VideoLLMs that maintains performance while significantly reducing computational overhead for long-context video inference.

Abstract: Video large language models (VideoLLMs) show strong capability in video understanding, yet long-context inference is still dominated by massive redundant visual tokens in the prefill stage. We revisit token compression for VideoLLMs under a tight budget and identify a key bottleneck, namely insufficient spatio-temporal information coverage. Existing methods often introduce discontinuous coverage through coarse per-frame allocation or scene segmentation, and token merging can further misalign spatio-temporal coordinates under MRoPE-style discrete (t,h,w) bindings. To address these issues, we propose V-CAST (Video Curvature-Aware Spatio-Temporal Pruning), a training-free, plug-and-play pruning policy for long-context video inference. V-CAST casts token compression as a trajectory approximation problem and introduces a curvature-guided temporal allocation module that routes per-frame token budgets to semantic turns and event boundaries. It further adopts a dual-anchor spatial selection mechanism that preserves high-entropy visual evidence without attention intervention, while keeping retained tokens at their original coordinates to maintain positional alignment. Extensive experiments across multiple VideoLLMs of different architectures and scales demonstrate that V-CAST achieves 98.6% of the original performance, outperforms the second-best method by +1.1% on average, and reduces peak memory and total latency to 86.7% and 86.4% of vanilla Qwen3-VL-8B-Instruct.

Method: Proposes Amped: adaptive multi-stage pruning framework that identifies high-confidence non-edge tokens and removes them early in processing. Also introduces SED: a simple yet high-performance Transformer-based edge detection model designed for reduced structural complexity.

Result: Amped reduces GFLOPs by up to 40% with only 0.4% drop in ODS F-measure. SED achieves state-of-the-art ODS F-measure of 86.5% despite its simplicity.

Conclusion: The proposed pruning strategy provides favorable accuracy-efficiency balance for transformer-based edge detectors, and the streamlined model enables practical deployment while maintaining high performance.

Abstract: Edge detection is a fundamental image analysis task that underpins numerous high-level vision applications. Recent advances in Transformer architectures have significantly improved edge quality by capturing long-range dependencies, but this often comes with computational overhead. Achieving higher pixel-level accuracy requires increased input resolution, further escalating computational cost and limiting practical deployment. Building on the strong representational capacity of recent Transformer-based edge detectors, we propose an Adaptive Multi-stage non-edge Pruning framework for Edge Detection(Amped). Amped identifies high-confidence non-edge tokens and removes them as early as possible to substantially reduce computation, thus retaining high accuracy while cutting GFLOPs and accelerating inference with minimal performance loss. Moreover, to mitigate the structural complexity of existing edge detection networks and facilitate their integration into real-world systems, we introduce a simple yet high-performance Transformer-based model, termed Streamline Edge Detector(SED). Applied to both existing detectors and our SED, the proposed pruning strategy provides a favorable balance between accuracy and efficiency-reducing GFLOPs by up to 40% with only a 0.4% drop in ODS F-measure. In addition, despite its simplicity, SED achieves a state-of-the-art ODS F-measure of 86.5%. The code will be released.

Method: Comparative study of 8 state-of-the-art open-source VidLLMs evaluated on two news datasets (Chilean TV: 1,345 clips, BBC News: 9,838 clips) using lexical metrics (METEOR, ROUGE-L), semantic metrics (BERTScore, CLIPScore, Text Similarity, Mean Reciprocal Rank), and two novel fidelity metrics: Thematic Fidelity Score (TFS) and Entity Fidelity Score (EFS).

Result: Standard metrics show limited discriminative power due to surface-form dependence, static-frame insensitivity, and function-word inflation. Gemma~3 achieves highest overall performance across both datasets and most evaluation dimensions, with Qwen-VL as consistent runner-up.

Conclusion: The proposed TFS and EFS metrics address gaps in standard evaluation by directly assessing thematic structure preservation and named-entity coverage, providing better assessment of news video captioning quality.

Abstract: News videos are among the most prevalent content types produced by television stations and online streaming platforms, yet generating textual descriptions to facilitate indexing and retrieval largely remains a manual process. Video Large Language Models (VidLLMs) offer significant potential to automate this task, but a comprehensive evaluation in the news domain is still lacking. This work presents a comparative study of eight state-of-the-art open-source VidLLMs for automatic news video captioning, evaluated on two complementary benchmark datasets: a Chilean TV news corpus (approximately 1,345 clips) and a BBC News corpus (9,838 clips). We employ lexical metrics (METEOR, ROUGE-L), semantic metrics (BERTScore, CLIPScore, Text Similarity, Mean Reciprocal Rank), and two novel fidelity metrics proposed in this work: the Thematic Fidelity Score (TFS) and Entity Fidelity Score (EFS). Our analysis reveals that standard metrics exhibit limited discriminative power for news video captioning due to surface-form dependence, static-frame insensitivity, and function-word inflation. TFS and EFS address these gaps by directly assessing thematic structure preservation and named-entity coverage in the generated captions. Results show that Gemma~3 achieves the highest overall performance across both datasets and most evaluation dimensions, with Qwen-VL as a consistent runner-up.

Method: Presents a taxonomy of four key crowd management tasks (detection, counting, tracking, behavior classification) with LiDAR applications, identifies challenges like dataset scarcity, sensor fusion needs, AI integration, and point cloud processing requirements.

Result: Provides actionable insights for developing LiDAR-based crowd management solutions tailored to public safety applications, highlighting current applications and future research directions.

Conclusion: LiDAR technology shows promise for crowd management but requires addressing challenges like dataset availability, sensor fusion, AI integration, and processing needs to realize its full potential for public safety applications.

Abstract: Light Detection and Ranging (LiDAR) technology offers significant advantages for effective crowd management. This article presents LiDAR technology and highlights its primary advantages over other monitoring technologies, including enhanced privacy, performance in various weather conditions, and precise 3D mapping. We present a general taxonomy of four key tasks in crowd management: crowd detection, counting, tracking, and behavior classification, with illustrative examples of LiDAR applications for each task. We identify challenges and open research directions, including the scarcity of dedicated datasets, sensor fusion requirements, artificial intelligence integration, and processing needs for LiDAR point clouds. This article offers actionable insights for developing crowd management solutions tailored to public safety applications.

Method: The researchers conducted a large-scale analysis evaluating commercial AI-powered facial unmasking tools, specifically assessing whether the resulting AI-generated faces can be reliably matched to true identities through systematic testing and evaluation.

Result: The analysis found that AI-generated “unmasked” faces cannot be reliably matched to true identities, demonstrating significant risks of misidentification when these tools are used in criminal investigations.

Conclusion: Commercial AI facial unmasking tools pose serious risks for misidentification and should not be trusted for reliable identity matching in criminal investigations, highlighting the need for caution and regulation in using AI for evidence enhancement.

Abstract: Recently, crowd-sourced online criminal investigations have used generative-AI to enhance low-quality visual evidence. In one high-profile case, social-media users circulated an “AI-unmasked” image of a federal agent involved in a fatal shooting, fueling a wide-spread misidentification. In response to this and similar incidents, we conducted a large-scale analysis evaluating the efficacy and risks of commercial AI-powered facial unmasking, specifically assessing whether the resulting faces can be reliably matched to true identities.

Method: Composer generates input-conditioned parameter adaptations at inference time, which are injected into pretrained model weights. Adaptation occurs once before multi-step generation, enabling per-input specialization without retraining.

Result: Experiments show Composer substantially improves performance across diverse generative models and use cases, including lightweight/quantized models and test-time scaling, with minimal computational overhead.

Conclusion: Composer establishes a new paradigm for adaptive generative models that dynamically adapt to each input through input-aware parameter composition, moving beyond static parameterization.

Abstract: Existing generative models, such as diffusion and auto-regressive networks, are inherently static, relying on a fixed set of pretrained parameters to handle all inputs. In contrast, humans flexibly adapt their internal generative representations to each perceptual or imaginative context. Inspired by this capability, we introduce Composer, a new paradigm for adaptive generative modeling based on test-time instance-specific parameter composition. Composer generates input-conditioned parameter adaptations at inference time, which are injected into the pretrained model’s weights, enabling per-input specialization without fine-tuning or retraining. Adaptation occurs once prior to multi-step generation, yielding higher-quality, context-aware outputs with minimal computational and memory overhead. Experiments show that Composer substantially improves performance across diverse generative models and use cases, including lightweight/quantized models and test-time scaling. By leveraging input-aware parameter composition, Composer establishes a new paradigm for designing generative models that dynamically adapt to each input, moving beyond static parameterization.

Method: Proposes a unified gated conditioning module in a dual-path pipeline that effectively integrates both spatially aligned and non-aligned conditional inputs, tailored specifically for linear attention backbones like SANA.

Result: Achieves state-of-the-art controllable generation performance on linear-attention models, surpassing existing methods in fidelity and controllability across multiple tasks and benchmarks.

Conclusion: The framework enables secure, efficient on-device controllable generation while maintaining high performance, addressing both privacy concerns and computational limitations of current cloud-based approaches.

Abstract: Recent advances in diffusion-based controllable visual generation have led to remarkable improvements in image quality. However, these powerful models are typically deployed on cloud servers due to their large computational demands, raising serious concerns about user data privacy. To enable secure and efficient on-device generation, we explore in this paper controllable diffusion models built upon linear attention architectures, which offer superior scalability and efficiency, even on edge devices. Yet, our experiments reveal that existing controllable generation frameworks, such as ControlNet and OminiControl, either lack the flexibility to support multiple heterogeneous condition types or suffer from slow convergence on such linear-attention models. To address these limitations, we propose a novel controllable diffusion framework tailored for linear attention backbones like SANA. The core of our method lies in a unified gated conditioning module working in a dual-path pipeline, which effectively integrates multi-type conditional inputs, such as spatially aligned and non-aligned cues. Extensive experiments on multiple tasks and benchmarks demonstrate that our approach achieves state-of-the-art controllable generation performance based on linear-attention models, surpassing existing methods in terms of fidelity and controllability.

Method: Agentic framework with three specialized agents coordinating via round-robin speaker selection in a Plan-Do-Check-Act cycle. Uses pre-defined processing for high-confidence cases and multi-agent reasoning for indirect identifiers. Implements spatially-filtered coarse-to-fine detection with scout-and-zoom strategy, open-vocabulary segmentation, and IoU-based deduplication. Applies modal-specific diffusion guidance with appearance decorrelation for anonymization.

Result: Reduces person Re-ID risk by 73% on CUHK03-NP (R1: 16.9% vs 62.4% baseline). Achieves KID: 0.001 and FID: 9.1 on CityScapes, significantly outperforming existing anonymization methods. Preserves downstream semantic segmentation and detects non-direct PII instances across object categories.

Conclusion: CAIAMAR provides effective context-aware PII anonymization with superior privacy protection and image quality preservation. The on-premise, open-source framework generates audit trails for GDPR compliance while flagging failed cases for human review.

Abstract: Street-level imagery contains personally identifiable information (PII), some of which is context-dependent. Existing anonymization methods either over-process images or miss subtle identifiers, while API-based solutions compromise data sovereignty. We present an agentic framework CAIAMAR (\underline{C}ontext-\underline{A}ware \underline{I}mage \underline{A}nonymization with \underline{M}ulti-\underline{A}gent \underline{R}easoning) for context-aware PII segmentation with diffusion-based anonymization, combining pre-defined processing for high-confidence cases with multi-agent reasoning for indirect identifiers. Three specialized agents coordinate via round-robin speaker selection in a Plan-Do-Check-Act (PDCA) cycle, enabling large vision-language models to classify PII based on spatial context (private vs. public property) rather than rigid category rules. The agents implement spatially-filtered coarse-to-fine detection where a scout-and-zoom strategy identifies candidates, open-vocabulary segmentation processes localized crops, and $IoU$-based deduplication ($30%$ threshold) prevents redundant processing. Modal-specific diffusion guidance with appearance decorrelation substantially reduces re-identification (Re-ID) risks. On CUHK03-NP, our method reduces person Re-ID risk by $73%$ ($R1$: $16.9%$ vs. $62.4%$ baseline). For image quality preservation on CityScapes, we achieve KID: $0.001$, and FID: $9.1$, significantly outperforming existing anonymization. The agentic workflow detects non-direct PII instances across object categories, and downstream semantic segmentation is preserved. Operating entirely on-premise with open-source models, the framework generates human-interpretable audit trails supporting EU’s GDPR transparency requirements while flagging failed cases for human review.

Method: Introduces VstoryGen framework with multimodal conditioning (text, character images, background references) and shot-type control via parameter-efficient prompt tuning on movie data to reflect cinematic grammar.

Result: VstoryGen achieves improved consistency and cinematic diversity compared to existing methods, as demonstrated through new multimodal benchmarks assessing character/scene consistency, text-visual alignment, and shot-type control.

Conclusion: The proposed multimodal framework enables customizable story generation with enhanced cinematic diversity and consistency through integrated text, visual references, and shot-type control.

Abstract: Multimodal story customization aims to generate coherent story flows conditioned on textual descriptions, reference identity images, and shot types. While recent progress in story generation has shown promising results, most approaches rely on text-only inputs. A few studies incorporate character identity cues (e.g., facial ID), but lack broader multimodal conditioning. In this work, we introduce VstoryGen, a multimodal framework that integrates descriptions with character and background references to enable customizable story generation. To enhance cinematic diversity, we introduce shot-type control via parameter-efficient prompt tuning on movie data, enabling the model to generate sequences that more faithfully reflect cinematic grammar. To evaluate our framework, we establish two new benchmarks that assess multimodal story customization from the perspectives of character and scene consistency, text-visual alignment, and shot-type control. Experiments demonstrate that VstoryGen achieves improved consistency and cinematic diversity compared to existing methods.

Method: Utilize segmentation foundation models (Segment Anything Model and SAM 2) to automate mask generation from unannotated frames and coarse annotations, reducing manual annotation effort.

Result: Can reduce annotation need by one-third while maintaining similar performance for video semantic segmentation. Found that dataset frame variety is more important than frame quantity for best performance.

Conclusion: Segmentation foundation models can effectively reduce video segmentation annotation costs, with frame diversity being more critical than frame count for optimal model performance.

Abstract: Present-day deep neural networks for video semantic segmentation require a large number of fine-grained pixel-level annotations to achieve the best possible results. Obtaining such annotations, however, is very expensive. On the other hand, raw, unannotated video frames are practically free to obtain. Similarly, coarse annotations, which do not require precise boundaries, are also much cheaper. This paper investigates approaches to reduce the annotation cost required for video segmentation datasets by utilising such resources. We show that using state-of-the-art segmentation foundation models, Segment Anything Model (SAM) and Segment Anything Model 2 (SAM 2), we can utilise both unannotated frames as well as coarse annotations to alleviate the effort required for manual annotation of video segmentation datasets by automating mask generation. Our investigation suggests that if used appropriately, we can reduce the need for annotation by a third with similar performance for video semantic segmentation. More significantly, our analysis suggests that the variety of frames in the dataset is more important than the number of frames for obtaining the best performance.

Method: Train machine learning models on 3D optical profilometry data from mechanically opened Herculaneum papyri to separate inked and uninked areas. Quantify how lateral sampling affects learnability and test native-resolution models on coarsened inputs.

Result: High-resolution topography alone contains usable signal for ink detection. Segmentation performance diminishes with decreasing lateral resolution, revealing characteristic spatial scales needed to exploit morphological signals.

Conclusion: Morphology-based ink detection works for carbon ink on carbonized papyrus. Findings inform spatial resolution targets for reading closed scrolls via X-ray tomography using surface morphology analysis.

Abstract: Reading the Herculaneum papyri is challenging because both the scrolls and the ink, which is carbon-based, are carbonized. In X-ray radiography and tomography, ink detection typically relies on density- or composition-driven contrast, but carbon ink on carbonized papyrus provides little attenuation contrast. Building on the morphological hypothesis, we show that the surface morphology of written regions contains enough signal to distinguish ink from papyrus. To this end, we train machine learning models on three-dimensional optical profilometry from mechanically opened Herculaneum papyri to separate inked and uninked areas. We further quantify how lateral sampling governs learnability and how a native-resolution model behaves on coarsened inputs. We show that high-resolution topography alone contains a usable signal for ink detection. Diminishing segmentation performance with decreasing lateral resolution provides insight into the characteristic spatial scales that must be resolved on our dataset to exploit the morphological signal. These findings inform spatial resolution targets for morphology-based reading of closed scrolls through X-ray tomography.

Method: Clinician Mimetic Workflow with two components: 1) Discriminative Exemplar Coreset Selection (DECS) - selects discriminative visual coresets from noisy data to simulate clinician’s reference to “anchor cases”; 2) Self-Refined Experience Summarization (SRES) - distills diverse rollouts into dynamic textual Experience Bank to mimic clinical cognition and reflection.

Result: Outperforms zero-shot general and medical MLLMs across all 12 datasets of MedMNIST 2D benchmark. Achieves performance comparable to fully supervised vision models and domain-specific fine-tuned MLLMs, setting new benchmark for parameter-efficient medical in-context learning.

Conclusion: The proposed framework effectively bridges the performance gap in medical diagnosis without updating MLLM backbone weights, offering a scalable, parameter-efficient alternative to fine-tuning while maintaining clinical relevance through workflow mimicry.

Abstract: General Multimodal Large Language Models (MLLMs) often underperform in capturing domain-specific nuances in medical diagnosis, trailing behind fully supervised baselines. Although fine-tuning provides a remedy, the high costs of expert annotation and massive computational overhead limit its scalability. To bridge this gap without updating the weights of the pre-trained backbone of the MLLM, we propose a Clinician Mimetic Workflow. This is a novel In-Context Learning (ICL) framework designed to synergize Discriminative Exemplar Coreset Selection (DECS) and Self-Refined Experience Summarization (SRES). Specifically, DECS simulates a clinician’s ability to reference “anchor cases” by selecting discriminative visual coresets from noisy data at the computational level; meanwhile, SRES mimics the cognition and reflection in clinical diagnosis by distilling diverse rollouts into a dynamic textual Experience Bank. Extensive evaluation across all 12 datasets of the MedMNIST 2D benchmark demonstrates that our method outperforms zero-shot general and medical MLLMs. Simultaneously, it achieves performance levels comparable to fully supervised vision models and domain-specific fine-tuned MLLMs, setting a new benchmark for parameter-efficient medical in-context learning. Our code is available at an anonymous repository: https://anonymous.4open.science/r/Synergizing-Discriminative-Exemplars-and-Self-Refined-Experience-ED74.

Method: Proposes TIR-Agent with a two-stage training pipeline: 1) Supervised fine-tuning (SFT) to learn basic tool-calling patterns, 2) Reinforcement learning (RL) with key designs: random perturbation strategy on SFT data to broaden exploration, and multi-dimensional adaptive reward mechanism to dynamically weight image quality metrics and prevent reward hacking. Also develops a globally shared model-call pool for high-throughput GPU-based tool invocation.

Result: Outperforms 12 baselines including 6 all-in-one models, 3 training-free agents, and 3 proprietary models on both in-domain and out-of-domain degradations. Achieves over 2.5× inference speedup by eliminating redundant tool executions.

Conclusion: TIR-Agent demonstrates that trainable vision-language agents with learned policies can significantly improve image restoration performance and efficiency compared to training-free approaches, offering a promising direction for multimodal AI systems.

Abstract: Vision-language agents that orchestrate specialized tools for image restoration (IR) have emerged as a promising method, yet most existing frameworks operate in a training-free manner. They rely on heuristic task scheduling and exhaustive tool traversal, resulting in sub-optimal restoration paths and prohibitive computational cost. We argue that the core bottleneck lies in the absence of a learned policy to make decision, as a vision-language model cannot efficiently handle degradation-aware task ordering and tool composition. To this end, we propose TIR-Agent, a trainable image restoration agent that performs a direct tool-calling policy through a two-stage training pipeline of supervised fine-tuning (SFT) followed by reinforcement learning (RL). Two key designs underpin effective RL training: (i) a random perturbation strategy applied to the SFT data, which broadens the policy’s exploration over task schedules and tool compositions, and (ii) a multi-dimensional adaptive reward mechanism that dynamically re-weights heterogeneous image quality metrics to mitigate reward hacking. To support high-throughput, asynchronous GPU-based tool invocation during training, we further develop a globally shared model-call pool. Experiments on both in-domain and out-of-domain degradations show that TIR-Agent outperforms 12 baselines, including 6 all-in-one models, 3 training-free agents, and 3 proprietary models, and achieves over 2.5$\times$ inference speedup by eliminating redundant tool executions.

Method: Created JaWildText benchmark with 3,241 instances from 2,961 newly captured images in Japan, containing 1.12 million annotated characters across 3,643 unique character types. Includes three tasks: Dense Scene Text VQA, Receipt Key Information Extraction, and Handwriting OCR.

Result: Evaluation of 14 open-weight VLMs shows best model achieves average score of 0.64 across three tasks. Error analysis reveals recognition remains dominant bottleneck, especially for kanji characters.

Conclusion: JaWildText enables fine-grained, script-aware diagnosis of Japanese scene text capabilities and will be released with evaluation code to advance research in Japanese text understanding for vision-language models.

Abstract: Japanese scene text poses challenges that multilingual benchmarks often fail to capture, including mixed scripts, frequent vertical writing, and a character inventory far larger than the Latin alphabet. Although Japanese is included in several multilingual benchmarks, these resources do not adequately capture the language-specific complexities. Meanwhile, existing Japanese visual text datasets have primarily focused on scanned documents, leaving in-the-wild scene text underexplored. To fill this gap, we introduce JaWildText, a diagnostic benchmark for evaluating vision-language models (VLMs) on Japanese scene text understanding. JaWildText contains 3,241 instances from 2,961 images newly captured in Japan, with 1.12 million annotated characters spanning 3,643 unique character types. It comprises three complementary tasks that vary in visual organization, output format, and writing style: (i) Dense Scene Text Visual Question Answering (STVQA), which requires reasoning over multiple pieces of visual text evidence; (ii) Receipt Key Information Extraction (KIE), which tests layout-aware structured extraction from mobile-captured receipts; and (iii) Handwriting OCR, which evaluates page-level transcription across various media and writing directions. We evaluate 14 open-weight VLMs and find that the best model achieves an average score of 0.64 across the three tasks. Error analyses show recognition remains the dominant bottleneck, especially for kanji. JaWildText enables fine-grained, script-aware diagnosis of Japanese scene text capabilities, and will be released with evaluation code.

Method: VLM-3R processes monocular video frames using a geometry encoder to derive implicit 3D tokens representing spatial understanding. It employs Spatial-Visual-View Fusion and uses over 200K curated 3D reconstructive instruction tuning QA pairs to align real-world spatial context with language instructions.

Result: Extensive experiments show VLM-3R facilitates robust visual-spatial reasoning and enables understanding of temporal 3D context changes, excelling in both accuracy and scalability. The model also introduces a Vision-Spatial-Temporal Intelligence benchmark with 138.6K QA pairs across five tasks.

Conclusion: VLM-3R successfully extends multimodal models to 3D scene understanding from monocular videos, achieving deep spatial understanding without external sensors or pre-constructed maps, while enabling temporal reasoning about evolving spatial relationships.

Abstract: The rapid advancement of Large Multimodal Models (LMMs) for 2D images and videos has motivated extending these models to understand 3D scenes, aiming for human-like visual-spatial intelligence. Nevertheless, achieving deep spatial understanding comparable to human capabilities poses significant challenges in model encoding and data acquisition. Existing methods frequently depend on external depth sensors for geometry capture or utilize off-the-shelf algorithms for pre-constructing 3D maps, thereby limiting their scalability, especially with prevalent monocular video inputs and for time-sensitive applications. In this work, we introduce VLM-3R, a unified framework for Vision-Language Models (VLMs) that incorporates 3D Reconstructive instruction tuning. VLM-3R processes monocular video frames by employing a geometry encoder to derive implicit 3D tokens that represent spatial understanding. Leveraging our Spatial-Visual-View Fusion and over 200K curated 3D reconstructive instruction tuning question-answer (QA) pairs, VLM-3R effectively aligns real-world spatial context with language instructions. This enables monocular 3D spatial assistance and embodied reasoning. To facilitate the evaluation of temporal reasoning, we introduce the Vision-Spatial-Temporal Intelligence benchmark, featuring over 138.6K QA pairs across five distinct tasks focused on evolving spatial relationships. Extensive experiments demonstrate that our model, VLM-3R, not only facilitates robust visual-spatial reasoning but also enables the understanding of temporal 3D context changes, excelling in both accuracy and scalability.

Method: Controlled three-stage training framework with fixed backbone, trainable modules, and optimization pipeline. Compares four data organization strategies: direct mixture, curriculum training, balanced sampling, and reverse curriculum. Evaluates on general visual instruction following, diagram reasoning, chart reasoning, scene-text question answering, and document question answering.

Result: Curriculum training gives best overall trade-off and strongest structured reasoning performance. Balanced sampling is better for OCR-oriented capability but weakens broader capability balance. Reverse curriculum performs worst in both final performance and optimization stability. Training dynamics show building general understanding and reasoning before OCR leads to smoother optimization.

Conclusion: Data organization is a first-order design variable in multimodal adaptation. Curriculum training (general→reasoning→OCR) provides optimal balance. Data scheduling should be considered as an explicit design dimension for multimodal model adaptation.

Abstract: Recent multimodal large language models (MLLMs) perform strongly on general visual understanding, diagram and chart reasoning, and document-centric perception. However, these abilities are learned from heterogeneous supervision sources with very different task structures and learning demands, and the effect of their temporal organization during training remains underexplored. We study whether data organization affects the trade-off among general understanding, structured reasoning, and fine-grained OCR/document understanding in multimodal instruction tuning. To isolate this factor, we use a controlled three-stage training framework in which the backbone, trainable modules, and optimization pipeline are fixed across all runs, and only the temporal arrangement of post-alignment supervision is changed. We compare four strategies: direct mixture, curriculum training, balanced sampling, and reverse curriculum. Experiments on general visual instruction following, diagram reasoning, chart reasoning, scene-text question answering, and document question answering show that data organization is a first-order design variable in multimodal adaptation. Curriculum training gives the best overall trade-off and the strongest structured reasoning performance. Balanced sampling is better for OCR-oriented capability but weakens the broader capability balance. Reverse curriculum performs worst in both final performance and optimization stability. Training-dynamics analysis further suggests that building general understanding and reasoning before introducing OCR-intensive supervision leads to smoother optimization and faster convergence. These findings highlight data scheduling as an explicit design dimension for multimodal model adaptation.

Method: Introduces E-TIDE with TIDE module (Temporal Interaction for Dynamic Events) that uses efficient spatiotemporal interaction design for sparse event tensors. Captures temporal dependencies via large-kernel mixing and activity-aware gating while maintaining low computational complexity. The architecture is lightweight and end-to-end trainable without requiring large-scale pretraining.

Result: Experiments on standard event-based datasets show competitive performance with significantly reduced model size and training requirements. The method is well-suited for real-time deployment under tight latency and memory budgets.

Conclusion: E-TIDE provides an efficient solution for event-tensor prediction that balances performance with computational efficiency, making it practical for resource-constrained applications while maintaining competitive accuracy.

Abstract: Event-based cameras capture visual information as asynchronous streams of per-pixel brightness changes, generating sparse, temporally precise data. Compared to conventional frame-based sensors, they offer significant advantages in capturing high-speed dynamics while consuming substantially less power. Predicting future event representations from past observations is an important problem, enabling downstream tasks such as future semantic segmentation or object tracking without requiring access to future sensor measurements. While recent state-of-the-art approaches achieve strong performance, they often rely on computationally heavy backbones and, in some cases, large-scale pretraining, limiting their applicability in resource-constrained scenarios. In this work, we introduce E-TIDE, a lightweight, end-to-end trainable architecture for event-tensor prediction that is designed to operate efficiently without large-scale pretraining. Our approach employs the TIDE module (Temporal Interaction for Dynamic Events), motivated by efficient spatiotemporal interaction design for sparse event tensors, to capture temporal dependencies via large-kernel mixing and activity-aware gating while maintaining low computational complexity. Experiments on standard event-based datasets demonstrate that our method achieves competitive performance with significantly reduced model size and training requirements, making it well-suited for real-time deployment under tight latency and memory budgets.

Method: Establishes theoretical framework proving equivalence between dataset distillation and distribution matching; proposes DsCo with Noise-Optimization (NOpt) to synthesize representative samples, optionally augmented via “Doping” (mixing selected original samples).

Result: Achieves SOTA for low data volumes, extends well to high volumes (nearly halves dataset size with no performance degradation), applicable in both data-accessible and data-free scenarios.

Conclusion: DsCo provides theoretically justified, efficient dataset distillation that overcomes scalability limitations and works in data-free scenarios, enabling more accessible large-scale visual recognition systems.

Abstract: The high cost and accessibility problem associated with large datasets hinder the development of large-scale visual recognition systems. Dataset Distillation addresses these problems by synthesizing compact surrogate datasets for efficient training, storage, transfer, and privacy preservation. The existing state-of-the-art diffusion-based dataset distillation methods face three issues: lack of theoretical justification, poor efficiency in scaling to high data volumes, and failure in data-free scenarios. To address these issues, we establish a theoretical framework that justifies the use of diffusion models by proving the equivalence between dataset distillation and distribution matching, and reveals an inherent efficiency limit in the dataset distillation paradigm. We then propose a Dataset Concentration (DsCo) framework that uses a diffusion-based Noise-Optimization (NOpt) method to synthesize a small yet representative set of samples, and optionally augments the synthetic data via “Doping”, which mixes selected samples from the original dataset with the synthetic samples to overcome the efficiency limit of dataset distillation. DsCo is applicable in both data-accessible and data-free scenarios, achieving SOTA performances for low data volumes, and it extends well to high data volumes, where it nearly reduces the dataset size by half with no performance degradation.

Method: Proposes RHO model with two-branch Pin-Pan architecture: one branch processes panoramic images using Split-Undistort-Merge (SUM) module to handle distortion, while the other processes OpenStreetMap data. Uses Position-Orientation Fusion (POF) mechanism to combine position and heading information for accurate localization. Introduces CV-RHO dataset with 2.7M images under diverse conditions.

Result: Extensive experiments show significant performance gains up to 20% compared to state-of-the-art baselines. The CV-RHO dataset proves valuable for benchmarking, and the RHO model demonstrates effectiveness in metric cross-view geo-localization.

Conclusion: The work establishes a comprehensive benchmark for panoramic cross-view geo-localization and proposes an effective architecture that handles panoramic distortion while leveraging OpenStreetMap data for improved accuracy under varying conditions.

Abstract: Metric Cross-View Geo-Localization (MCVGL) aims to estimate the 3-DoF camera pose (position and heading) by matching ground and satellite images. In this work, instead of pinhole and satellite images, we study robust MCVGL using holistic panoramas and OpenStreetMap (OSM). To this end, we establish a large-scale MCVGL benchmark dataset, CV-RHO, with over 2.7M images under different weather and lighting conditions, as well as sensor noise. Furthermore, we propose a model termed RHO with a two-branch Pin-Pan architecture for accurate visual localization. A Split-Undistort-Merge (SUM) module is introduced to address the panoramic distortion, and a Position-Orientation Fusion (POF) mechanism is designed to enhance the localization accuracy. Extensive experiments prove the value of our CV-RHO dataset and the effectiveness of the RHO model, with a significant performance gain up to 20% compared with the state-of-the-art baselines. Project page: https://github.com/InSAI-Lab/RHO.

Method: Parametric structural perturbation method inspired by 3D fabric mechanics, generating photorealistic wrinkles via multi-scale wrinkle fields with displacement field distortion and surface-consistent appearance variations. Uses hierarchical fitness function in low-dimensional parameter space with optimization-based search strategy.

Result: Method significantly degrades performance of various state-of-the-art VLMs, consistently outperforming baselines in both image captioning and visual question-answering tasks.

Conclusion: VLMs remain vulnerable to physically plausible non-rigid deformations, highlighting the need for more robust vision-language models that can handle real-world physical variations.

Abstract: Visual-Language Models (VLMs) have demonstrated exceptional cross-modal understanding across various tasks, including zero-shot classification, image captioning, and visual question answering. However, their robustness to physically plausible non-rigid deformations-such as wrinkles on flexible surfaces-remains poorly understood. In this work, we propose a parametric structural perturbation method inspired by the mechanics of three-dimensional fabric wrinkles. Specifically, our method generates photorealistic non-rigid perturbations by constructing multi-scale wrinkle fields and integrating displacement field distortion with surface-consistent appearance variations. To achieve an optimal balance between visual naturalness and adversarial effectiveness, we design a hierarchical fitness function in a low-dimensional parameter space and employ an optimization-based search strategy. We evaluate our approach using a two-stage framework: perturbations are first optimized on a zero-shot classification proxy task and subsequently assessed for transferability on generative tasks. Experimental results demonstrate that our method significantly degrades the performance of various state-of-the-art VLMs, consistently outperforming baselines in both image captioning and visual question-answering tasks.

Method: Created person-centric annotations for full LAION-400M dataset using validated automatic labeling pipelines combining object detection, multimodal captioning, and finetuned classifiers for perceived gender and race/ethnicity labels.

Result: Uncovered demographic imbalances and harmful associations (e.g., men and Black/Middle Eastern individuals disproportionately linked with crime/negative content). Linear fit predicts 60-70% of gender bias in CLIP/Stable Diffusion from data co-occurrences.

Conclusion: Establishes first large-scale empirical link between dataset composition and downstream model bias, providing resources to study and mitigate biases in vision-language models.

Abstract: Vision-language models trained on large-scale multimodal datasets show strong demographic biases, but the role of training data in producing these biases remains unclear. A major barrier has been the lack of demographic annotations in web-scale datasets such as LAION-400M. We address this gap by creating person-centric annotations for the full dataset, including over 276 million bounding boxes, perceived gender and race/ethnicity labels, and automatically generated captions. These annotations are produced through validated automatic labeling pipelines combining object detection, multimodal captioning, and finetuned classifiers. Using them, we uncover demographic imbalances and harmful associations, such as the disproportionate linking of men and individuals perceived as Black or Middle Eastern with crime-related and negative content. We also show that a linear fit predicts 60-70% of gender bias in CLIP and Stable Diffusion from direct co-occurrences in the data. Our resources establish the first large-scale empirical link between dataset composition and downstream model bias. Code is available at https://github.com/ExplainableML/LAION-400M-Person-Centric-Annotations.

Method: RINO uses RINONet, a feature extractor that integrates vector-based SO(3)-invariant learning with orientation-aware complex functional maps to extract robust features directly from raw geometry, enabling a fully end-to-end, data-driven approach without shape pre-alignment or handcrafted features.

Result: Extensive experiments show unprecedented performance across challenging non-rigid matching tasks, including arbitrary poses, non-isometry, partiality, non-manifoldness, and noise.

Conclusion: RINO provides a robust, unsupervised solution for dense 3D shape correspondence that overcomes limitations of existing methods and works effectively under various challenging conditions.

Abstract: Dense 3D shape correspondence remains a central challenge in computer vision and graphics as many deep learning approaches still rely on intermediate geometric features or handcrafted descriptors, limiting their effectiveness under non-isometric deformations, partial data, and non-manifold inputs. To overcome these issues, we introduce RINO, an unsupervised, rotation-invariant dense correspondence framework that effectively unifies rigid and non-rigid shape matching. The core of our method is the novel RINONet, a feature extractor that integrates vector-based SO(3)-invariant learning with orientation-aware complex functional maps to extract robust features directly from raw geometry. This allows for a fully end-to-end, data-driven approach that bypasses the need for shape pre-alignment or handcrafted features. Extensive experiments show unprecedented performance of RINO across challenging non-rigid matching tasks, including arbitrary poses, non-isometry, partiality, non-manifoldness, and noise.

Method: Proposes GS3LAM framework that models scenes as Semantic Gaussian Fields (SG-Field), jointly optimizes camera poses and fields via multimodal error constraints, introduces Depth-adaptive Scale Regularization (DSR) to resolve scale misalignments, and uses Random Sampling-based Keyframe Mapping (RSKM) to mitigate catastrophic forgetting.

Result: Extensive experiments show GS3LAM achieves increased tracking robustness, superior rendering quality, and enhanced semantic precision compared to state-of-the-art methods on benchmark datasets.

Conclusion: GS3LAM successfully addresses limitations of existing semantic SLAM systems by leveraging 3D Gaussian Splatting for real-time, dense multimodal fusion with improved performance in tracking, rendering, and semantic mapping.

Abstract: Recently, the multi-modal fusion of RGB, depth, and semantics has shown great potential in dense Simultaneous Localization and Mapping (SLAM). However, a prerequisite for generating consistent semantic maps is the availability of dense, efficient, and scalable scene representations. Existing semantic SLAM systems based on explicit representations are often limited by resolution and an inability to predict unknown areas. Conversely, implicit representations typically rely on time-consuming ray tracing, failing to meet real-time requirements. Fortunately, 3D Gaussian Splatting (3DGS) has emerged as a promising representation that combines the efficiency of point-based methods with the continuity of geometric structures. To this end, we propose GS3LAM, a Gaussian Semantic Splatting SLAM framework that processes multimodal data to render consistent, dense semantic maps in real-time. GS3LAM models the scene as a Semantic Gaussian Field (SG-Field) and jointly optimizes camera poses and the field via multimodal error constraints. Furthermore, a Depth-adaptive Scale Regularization (DSR) scheme is introduced to resolve misalignments between scale-invariant Gaussians and geometric surfaces. To mitigate catastrophic forgetting, we propose a Random Sampling-based Keyframe Mapping (RSKM) strategy, which demonstrates superior performance over common local covisibility optimization methods. Extensive experiments on benchmark datasets show that GS3LAM achieves increased tracking robustness, superior rendering quality, and enhanced semantic precision compared to state-of-the-art methods. Source code is available at https://github.com/lif314/GS3LAM.

Method: Corrects the generation trajectory at inference time so the input manga image can be reconstructed more faithfully under an empty prompt, requiring only the input image itself.

Result: Consistently outperforms existing baselines while incurring only negligible computational overhead.

Conclusion: Proposed inference-time adaptation method effectively adapts pretrained image editing models to manga without retraining, addressing domain gap issues.

Abstract: We present an inference-time adaptation method that tailors a pretrained image editing model to each input manga image using only the input image itself. Despite recent progress in pretrained image editing, such models often underperform on manga because they are trained predominantly on natural-image data. Re-training or fine-tuning large-scale models on manga is, however, generally impractical due to both computational cost and copyright constraints. To address this issue, our method slightly corrects the generation trajectory at inference time so that the input image can be reconstructed more faithfully under an empty prompt. Experimental results show that our method consistently outperforms existing baselines while incurring only negligible computational overhead.

Method: The model generates special pointing tokens that cross-attend to input image/video tokens to select appropriate ones. It uses a hierarchical approach: first token selects region, second token selects fine-grained subpatch within that region, third token specifies location within subpatch. Includes sequential generation with consistent order, encoding relative position of previous point, and special no-more-points class.

Result: Achieves new SOTA on image pointing (70.7% on PointBench), SOTA among fully open models on GUI pointing (61.1% on ScreenSpotPro), improves video pointing (59.1% human preference win rate vs. text coordinate baseline), and tracking (+6.3% gain on Molmo2Track). Shows higher sample efficiency and qualitative improvements.

Conclusion: The proposed token-based pointing mechanism is more intuitive and effective than coordinate-based approaches, achieving superior performance across multiple vision-language pointing tasks while being more sample efficient.

Abstract: Grounding has become a fundamental capability of vision-language models (VLMs). Most existing VLMs point by generating coordinates as part of their text output, which requires learning a complicated coordinate system and results in a high token count. Instead, we propose a more intuitive pointing mechanism that directly selects the visual tokens that contain the target concept. Our model generates a special pointing token that cross-attends to the input image or video tokens and selects the appropriate one. To make this model more fine-grained, we follow these pointing tokens with an additional special token that selects a fine-grained subpatch within the initially selected region, and then a third token that specifies a location within that subpatch. We further show that performance improves by generating points sequentially in a consistent order, encoding the relative position of the previously selected point, and including a special no-more-points class when selecting visual tokens. Using this method, we set a new state-of-the-art on image pointing (70.7% on PointBench), set a new state-of-the-art among fully open models on GUI pointing (61.1% on ScreenSpotPro), and improve video pointing (59.1% human preference win rate vs. a text coordinate baseline) and tracking (+6.3% gain on Molmo2Track). We additionally show that our method achieves much higher sample efficiency and discuss the qualitative differences that emerge from this design change.

Method: Proposes a feature-domain similarity filtering mechanism to discard redundant samples and ensure diverse training data, plus a dual-branch network combining CLIP features from pixel and frequency domains to capture both semantic and structural cues.

Result: Extensive experiments on benchmark datasets show significant improvements in cross-model and cross-dataset performance compared to existing methods.

Conclusion: The work highlights the critical role of data and feature diversity in building reliable and robust detectors against the rapidly evolving landscape of synthetic content.

Abstract: The rapid proliferation of AI-generated images, powered by generative adversarial networks (GANs), diffusion models, and other synthesis techniques, has raised serious concerns about misinformation, copyright violations, and digital security. However, detecting such images in a generalized and robust manner remains a major challenge due to the vast diversity of generative models and data distributions. In this work, we present \textbf{Diversity Matters}, a novel framework that emphasizes data diversity and feature domain complementarity for AI-generated image detection. The proposed method introduces a feature-domain similarity filtering mechanism that discards redundant or highly similar samples across both inter-class and intra-class distributions, ensuring a more diverse and representative training set. Furthermore, we propose a dual-branch network that combines CLIP features from the pixel domain and the frequency domain to jointly capture semantic and structural cues, leading to improved generalization against unseen generative models and adversarial conditions. Extensive experiments on benchmark datasets demonstrate that the proposed approach significantly improves cross-model and cross-dataset performance compared to existing methods. \textbf{Diversity Matters} highlights the critical role of data and feature diversity in building reliable and robust detectors against the rapidly evolving landscape of synthetic content.

Method: GraySense consists of two stages: (1) Packet Grouping module identifies frame boundaries and estimates frame sizes from encrypted network traffic, and (2) Tracker module uses a Transformer encoder with recurrent state to fuse indirect packet-based inputs with optional direct camera-based inputs to estimate object position.

Result: Extensive experiments with realistic videos from CARLA simulator and emulated networks show GraySense achieves 2.33 meters tracking error (Euclidean distance) without raw signal access, within tracked object dimensions (4.61m x 1.93m).

Conclusion: GraySense demonstrates novel capability to perform geospatial object tracking using only encrypted packet-level information, expanding the use of latent signals for sensing without requiring access to raw video streams.

Abstract: Accurate observation of dynamic environments traditionally relies on synthesizing raw, signal-level information from multiple distributed sensors. This work investigates an alternative approach: performing geospatial inference using only encrypted packet-level information, without access to the raw sensory data. We further explore how this indirect information can be fused with directly available sensory data to extend overall inference capabilities. We introduce GraySense, a learning-based framework that performs geospatial object tracking by analyzing encrypted wireless video transmission traffic, such as packet sizes, from cameras with inaccessible streams. GraySense leverages the inherent relationship between scene dynamics and transmitted packet sizes to infer object motion. The framework consists of two stages: (1) a Packet Grouping module that identifies frame boundaries and estimates frame sizes from encrypted network traffic, and (2) a Tracker module, based on a Transformer encoder with a recurrent state, which fuses indirect packet-based inputs with optional direct camera-based inputs to estimate the object’s position. Extensive experiments with realistic videos from the CARLA simulator and emulated networks under varying conditions show that GraySense achieves 2.33 meters tracking error (Euclidean distance) without raw signal access, within the dimensions of tracked objects (4.61m x 1.93m). To our knowledge, this capability has not been previously demonstrated, expanding the use of latent signals for sensing.

Method: VideoARM uses an agentic reasoning-over-hierarchical-memory paradigm with adaptive on-the-fly reasoning and memory construction. It employs a controller that autonomously invokes tools for coarse-to-fine video interpretation, while a hierarchical multimodal memory continuously captures and updates multi-level clues to support decision-making.

Result: Experiments on prevalent benchmarks show VideoARM outperforms the state-of-the-art DVD method while significantly reducing token consumption for long-form videos.

Conclusion: VideoARM provides an effective paradigm for long-form video understanding through adaptive agentic reasoning and hierarchical memory, addressing efficiency and performance limitations of existing approaches.

Abstract: Long-form video understanding remains challenging due to the extended temporal structure and dense multimodal cues. Despite recent progress, many existing approaches still rely on hand-crafted reasoning pipelines or employ token-consuming video preprocessing to guide MLLMs in autonomous reasoning. To overcome these limitations, we introduce VideoARM, an Agentic Reasoning-over-hierarchical-Memory paradigm for long-form video understanding. Instead of static, exhaustive preprocessing, VideoARM performs adaptive, on-the-fly agentic reasoning and memory construction. Specifically, VideoARM performs an adaptive and continuous loop of observing, thinking, acting, and memorizing, where a controller autonomously invokes tools to interpret the video in a coarse-to-fine manner, thereby substantially reducing token consumption. In parallel, a hierarchical multimodal memory continuously captures and updates multi-level clues throughout the operation of the agent, providing precise contextual information to support the controller in decision-making. Experiments on prevalent benchmarks demonstrate that VideoARM outperforms the state-of-the-art method, DVD, while significantly reducing token consumption for long-form videos.

Method: Proposes stateful experience learning that abstracts interaction data into atomic decision experiences through hindsight reasoning. These experiences are organized into a quality-filtered experience bank. At inference time, the agent uses policy-driven experience retrieval with complementary wide- and deep-search strategies to dynamically retrieve multimodal guidance across diverse semantic viewpoints.

Result: Extensive experiments show MuSEAgent consistently outperforms strong trajectory-level experience retrieval baselines on both fine-grained visual perception and complex multimodal reasoning tasks, validating the effectiveness of stateful experience modeling.

Conclusion: Stateful experience modeling significantly improves multimodal agent reasoning by enabling more effective learning from past interactions and adaptive retrieval of multimodal guidance during decision-making.

Abstract: Research agents have recently achieved significant progress in information seeking and synthesis across heterogeneous textual and visual sources. In this paper, we introduce MuSEAgent, a multimodal reasoning agent that enhances decision-making by extending the capabilities of research agents to discover and leverage stateful experiences. Rather than relying on trajectory-level retrieval, we propose a stateful experience learning paradigm that abstracts interaction data into atomic decision experiences through hindsight reasoning. These experiences are organized into a quality-filtered experience bank that supports policy-driven experience retrieval at inference time. Specifically, MuSEAgent enables adaptive experience exploitation through complementary wide- and deep-search strategies, allowing the agent to dynamically retrieve multimodal guidance across diverse compositional semantic viewpoints. Extensive experiments demonstrate that MuSEAgent consistently outperforms strong trajectory-level experience retrieval baselines on both fine-grained visual perception and complex multimodal reasoning tasks. These results validate the effectiveness of stateful experience modeling in improving multimodal agent reasoning.

Method: Built a large-scale dataset from expert-curated annotations with 750K images paired with 7.5M question-answer samples. Covers three key tasks: abnormality detection, anatomy recognition, and pathology identification. Supports open-ended, closed-ended, and multiple-choice questions across 8 anatomical regions and 97 pathology categories.

Result: State-of-the-art vision-language models struggle with fine-grained pathology identification, particularly in open-ended settings even after fine-tuning. Text-only analysis shows model performance collapses to near-random without image inputs, confirming the dataset is free from linguistic shortcuts.

Conclusion: RadImageNet-VQA provides a challenging benchmark for radiologic VQA that requires genuine visual understanding, exposing limitations of current vision-language models in medical image analysis tasks.

Abstract: In this work, we introduce RadImageNet-VQA, a large-scale dataset designed to advance radiologic visual question answering (VQA) on CT and MRI exams. Existing medical VQA datasets are limited in scale, dominated by X-ray imaging or biomedical illustrations, and often prone to text-based shortcuts. RadImageNet-VQA is built from expert-curated annotations and provides 750K images paired with 7.5M question-answer samples. It covers three key tasks - abnormality detection, anatomy recognition, and pathology identification - spanning eight anatomical regions and 97 pathology categories, and supports open-ended, closed-ended, and multiple-choice questions. Extensive experiments show that state-of-the-art vision-language models still struggle with fine-grained pathology identification, particularly in open-ended settings and even after fine-tuning. Text-only analysis further reveals that model performance collapses to near-random without image inputs, confirming that RadImageNet-VQA is free from linguistic shortcuts. The full dataset and benchmark are publicly available at https://huggingface.co/datasets/raidium/RadImageNet-VQA.

Method: Created benchmark by converting KITTI-360 to nuScenes format. Evaluated three adaptation strategies: 1) rectification for zero-shot evaluation/fine-tuning, 2) distortion-aware view transformation modules using MEI camera model, 3) polar coordinate representations to align with radial distortion. Tested on three BEV architectures (BEVFormer, BEVDet, PETR).

Result: Projection-free architectures are inherently more robust to fisheye distortion than other view transformation modules. Established first real-data 3D detection benchmark with mixed fisheye and pinhole images.

Conclusion: Provides systematic adaptation guidelines for designing robust 3D perception systems with mixed camera configurations. Shows projection-free architectures offer better distortion robustness.

Abstract: Modern autonomous driving systems increasingly rely on mixed camera configurations with pinhole and fisheye cameras for full view perception. However, Bird’s-Eye View (BEV) 3D object detection models are predominantly designed for pinhole cameras, leading to performance degradation under fisheye distortion. To bridge this gap, we introduce a multi-view BEV detection benchmark with mixed cameras by converting KITTI-360 into nuScenes format. Our study encompasses three adaptations: rectification for zero-shot evaluation and fine-tuning of nuScenes-trained models, distortion-aware view transformation modules (VTMs) via the MEI camera model, and polar coordinate representations to better align with radial distortion. We systematically evaluate three representative BEV architectures, BEVFormer, BEVDet and PETR, across these strategies. We demonstrate that projection-free architectures are inherently more robust and effective against fisheye distortion than other VTMs. This work establishes the first real-data 3D detection benchmark with fisheye and pinhole images and provides systematic adaptation and practical guidelines for designing robust and cost-effective 3D perception systems. The code is available at https://github.com/CesarLiu/FishBEVOD.git.

Method: Created MDPBench with 3,400 document images across 17 languages using expert model labeling, manual correction, and human verification. Maintains separate public/private splits to prevent data leakage and ensure fair comparison.

Result: Closed-source models (Gemini3-Pro) show relative robustness, while open-source alternatives suffer dramatic performance collapse: average 17.8% drop on photographed documents and 14.0% drop on non-Latin scripts.

Conclusion: Reveals significant performance imbalances across languages and conditions, highlighting the need for more inclusive, deployment-ready document parsing systems that work across diverse real-world scenarios.

Abstract: We introduce Multilingual Document Parsing Benchmark, the first benchmark for multilingual digital and photographed document parsing. Document parsing has made remarkable strides, yet almost exclusively on clean, digital, well-formatted pages in a handful of dominant languages. No systematic benchmark exists to evaluate how models perform on digital and photographed documents across diverse scripts and low-resource languages. MDPBench comprises 3,400 document images spanning 17 languages, diverse scripts, and varied photographic conditions, with high-quality annotations produced through a rigorous pipeline of expert model labeling, manual correction, and human verification. To ensure fair comparison and prevent data leakage, we maintain separate public and private evaluation splits. Our comprehensive evaluation of both open-source and closed-source models uncovers a striking finding: while closed-source models (notably Gemini3-Pro) prove relatively robust, open-source alternatives suffer dramatic performance collapse, particularly on non-Latin scripts and real-world photographed documents, with an average drop of 17.8% on photographed documents and 14.0% on non-Latin scripts. These results reveal significant performance imbalances across languages and conditions, and point to concrete directions for building more inclusive, deployment-ready parsing systems. Source available at https://github.com/Yuliang-Liu/MultimodalOCR.

Method: Compared voxel-grid representation with varying regularizers to continuous neural representation for tomographic reconstruction of simulated pool fire. Both representations output temperature and composition as functions of location, followed by ray-tracing to solve radiative transfer equation for spectral intensity on hyperspectral infrared cameras, convolved with instrument lineshape function.

Result: Voxel-grid approach with total-variation regularizer reproduced ground-truth synthetic flame with highest accuracy for reduced memory intensity and runtime, outperforming neural network representation.

Conclusion: Classical voxel-grid representation with appropriate regularization can outperform neural network approaches for flame tomography, suggesting need for careful representation selection. Future work will explore more representations and experimental configurations.

Abstract: Flame tomography is a compelling approach for extracting large amounts of data from experiments via 3-D thermochemical reconstruction. Recent efforts employing neural-network flame representations have suggested improved reconstruction quality compared with classical tomography approaches, but a rigorous quantitative comparison with the same algorithm using a voxel-grid representation has not been conducted. Here, we compare a classical voxel-grid representation with varying regularizers to a continuous neural representation for tomographic reconstruction of a simulated pool fire. The representations are constructed to give temperature and composition as a function of location, and a subsequent ray-tracing step is used to solve the radiative transfer equation to determine the spectral intensity incident on hyperspectral infrared cameras, which is then convolved with an instrument lineshape function. We demonstrate that the voxel-grid approach with a total-variation regularizer reproduces the ground-truth synthetic flame with the highest accuracy for reduced memory intensity and runtime. Future work will explore more representations and under experimental configurations.

Method: Uses Rank-Revealing QR Decomposition to systematically analyze VFM subspace structures. Creates dual adapters: main adapter leverages minor subspace directions for diverse features, while sub adapter refines major directions with minimal adjustments. No additional regularization losses needed.

Result: Achieves state-of-the-art performance on both synthetic-to-real and real-to-real generalization tasks without complex architectures or additional inference latency.

Conclusion: Systematic exploitation of pre-trained subspace structures through RRQR-based initialization leads to superior domain generalization performance in semantic segmentation.

Abstract: Domain Generalized Semantic Segmentation (DGSS) aims to maintain robust performance across unseen target domains. Vision Foundation Models (VFMs) offer rich multi-domain knowledge that can enhance generalization. However, strategies for actively exploiting the rich subspace structures within VFMs remain under-explored, with many existing methods focusing primarily on preserving pre-trained knowledge. Furthermore, their LoRA components often suffer from limited representational diversity and inefficient parameter utilization. We propose RecycleLoRA, which addresses both challenges by employing Rank-Revealing QR Decomposition (RRQR) to systematically exploit VFM’s subspace structures and enhance LoRA’s representational richness. Our main adapter leverages minor subspace directions identified by RRQR to learn diverse and independent features, achieving competitive performance even when used alone. We further introduce a sub adapter that carefully refines major directions with minimal adjustments, providing complementary improvements to the main adapter’s strong baseline performance. This design enables the dual adapters to learn distinct representations without requiring additional regularization losses. Our systematic exploitation of pre-trained subspace structures through RRQR-based initialization leads to superior domain generalization performance. RecycleLoRA achieves state-of-the-art performance on both synthetic-to-real generalization and real-to-real generalization tasks without complex architectures or additional inference latency.

Method: Adapt Group Relative Policy Optimization (GRPO) to flow-based video models, design verifiable reward functions (multi-component trajectory rewards for games, embedding-level verifiable rewards for robotics), and systematically analyze reward design.

Result: RL fine-tuning with verifiable rewards improves generalization significantly: 29.1% accuracy improvement on 3D mazes and 51.4% on trap-avoidance tasks over SFT baselines.

Conclusion: Verifiable reward design is critical for stable RL training in video reasoning tasks, while multimodal reward models can lead to degenerate solutions.

Abstract: Video generation models produce visually coherent content but struggle with tasks requiring spatial reasoning and multi-step planning. Reinforcement learning (RL) offers a path to improve generalization, but its effectiveness in video reasoning hinges on reward design – a challenge that has received little systematic study. We investigate this problem by adapting Group Relative Policy Optimization (GRPO) to flow-based video models and training them on maze-solving and robotic navigation tasks. We first show that multimodal reward models fail catastrophically in this setting. To address this, we design verifiable reward functions grounded in objective task metrics. For structured game environments, we introduce a multi-component trajectory reward. For robotic navigation, we propose an embedding-level verifiable reward. Our experiments show that RL fine-tuning with verifiable rewards improves generalization. For example, on complex 3D mazes, our model improves exact match accuracy by 29.1% over the SFT baseline, and on trap-avoidance tasks by 51.4%. Our systematic reward analysis reveals that verifiable rewards are critical for stable training, while multimodal reward models could lead to degenerate solutions. These findings establish verifiable reward design as a key enabler for robust video reasoning. Code will be publicly available.

Method: Proposes an LLM-driven agent with multi-level change interpretation vision-language backbone and LLM-based orchestration, using the Forest-Change dataset with bi-temporal satellite imagery, change masks, and semantic captions.

Result: Achieves 67.10% mIoU and 40.17% BLEU-4 on Forest-Change dataset, and 88.13% mIoU and 34.41% BLEU-4 on LEVIR-MCI-Trees subset for joint change detection and captioning.

Conclusion: Demonstrates potential of interactive LLM-driven RSICI systems to improve accessibility, interpretability, and efficiency of forest change analysis.

Abstract: Modern forest monitoring workflows increasingly benefit from the growing availability of high-resolution satellite imagery and advances in deep learning. Two persistent challenges in this context are accurate pixel-level change detection and meaningful semantic change captioning for complex forest dynamics. While large language models (LLMs) are being adapted for interactive data exploration, their integration with vision-language models (VLMs) for remote sensing image change interpretation (RSICI) remains underexplored. To address this gap, we introduce an LLM-driven agent for integrated forest change analysis that supports natural language querying across multiple RSICI tasks. The proposed system builds upon a multi-level change interpretation (MCI) vision-language backbone with LLM-based orchestration. To facilitate adaptation and evaluation in forest environments, we further introduce the Forest-Change dataset, which comprises bi-temporal satellite imagery, pixel-level change masks, and multi-granularity semantic change captions generated using a combination of human annotation and rule-based methods. Experimental results show that the proposed system achieves mIoU and BLEU-4 scores of 67.10% and 40.17% on the Forest-Change dataset, and 88.13% and 34.41% on LEVIR-MCI-Trees, a tree-focused subset of LEVIR-MCI benchmark for joint change detection and captioning. These results highlight the potential of interactive, LLM-driven RSICI systems to improve accessibility, interpretability, and efficiency of forest change analysis. All data and code are publicly available at https://github.com/JamesBrockUoB/ForestChat.

Method: Poppy keeps backbone weights frozen and optimizes per-pixel offsets to the input RGB and output normal along with a learned reflectance decomposition. A differentiable rendering layer converts refined normals into polarization predictions and penalizes mismatches with observed polarization signals.

Result: Across seven benchmarks and three backbone architectures (diffusion, flow, and feed-forward), Poppy reduces mean angular error by 23-26% on synthetic data and 6-16% on real data.

Conclusion: Guiding learned RGB-based normal estimators with polarization cues at test time refines normals on challenging surfaces without requiring retraining, demonstrating effective integration of physics-based cues with learned models.

Abstract: Monocular surface normal estimators trained on large-scale RGB-normal data often perform poorly in the edge cases of reflective, textureless, and dark surfaces. Polarization encodes surface orientation independently of texture and albedo, offering a physics-based complement for these cases. Existing polarization methods, however, require multi-view capture or specialized training data, limiting generalization. We introduce Poppy, a training-free framework that refines normals from any frozen RGB backbone using single-shot polarization measurements at test time. Keeping backbone weights frozen, Poppy optimizes per-pixel offsets to the input RGB and output normal along with a learned reflectance decomposition. A differentiable rendering layer converts the refined normals into polarization predictions and penalizes mismatches with the observed signal. Across seven benchmarks and three backbone architectures (diffusion, flow, and feed-forward), Poppy reduces mean angular error by 23-26% on synthetic data and 6-16% on real data. These results show that guiding learned RGB-based normal estimators with polarization cues at test time refines normals on challenging surfaces without retraining.

Method: Visual-conditioned framework with noise-aware gated fusion module to adaptively control caption feature conditioning, staged SFT+RL strategy with CLIP-based alignment and BERTScore-based semantic rewards, and curated BioSci-Fig-Cap benchmark for supervision.

Result: Achieves 0.728 mAP@0.5:0.95 for detection, outperforms Qwen3-VL-8B by 0.44 in METEOR and 0.22 in BERTScore, and transfers zero-shot to out-of-distribution scientific domains without fine-tuning.

Conclusion: FigEx2 effectively converts otherwise unusable scientific figures into aligned panel-text pairs, enabling better utilization of multimodal scientific data for downstream pretraining and retrieval tasks.

Abstract: Scientific compound figures combine multiple labeled panels into a single image. However, in a PMC-scale crawl of 346,567 compound figures, 16.3% have no caption and 1.8% only have captions shorter than ten words, causing them to be discarded by existing caption-decomposition pipelines. We propose FigEx2, a visual-conditioned framework that localizes panels and generates panel-wise captions directly from the image, converting otherwise unusable figures into aligned panel-text pairs for downstream pretraining and retrieval. To mitigate linguistic variance in open-ended captioning, we introduce a noise-aware gated fusion module that adaptively controls how caption features condition the detection query space, and employ a staged SFT+RL strategy with CLIP-based alignment and BERTScore-based semantic rewards. To support high-quality supervision, we curate BioSci-Fig-Cap, a refined benchmark for panel-level grounding, alongside cross-disciplinary test suites in physics and chemistry. FigEx2 achieves 0.728 mAP@0.5:0.95 for detection, outperforms Qwen3-VL-8B by 0.44 in METEOR and 0.22 in BERTScore, and transfers zero-shot to out-of-distribution scientific domains without fine-tuning.

Method: SAGE uses attention sink tokens (punctuation/function tokens that accumulate disproportionate attention) as anchors to monitor grounding reliability. At each sink trigger, it extracts semantic concepts, estimates visual grounding using self-attention maps and gradient-based attribution, measures spatial agreement, and adaptively sharpens or broadens self-attention distributions to reinforce grounded regions or suppress unreliable ones.

Result: SAGE consistently outperforms existing decoding strategies across diverse hallucination benchmarks, achieving average relative improvements of 10.65% on MSCOCO and 7.19% on AMBER across various VLM architectures, reducing hallucinations while preserving descriptive coverage.

Conclusion: SAGE effectively mitigates hallucinations in VLMs by dynamically modulating self-attention during decoding using sink-aware grounding, without requiring model retraining or architectural modifications, offering a practical solution to hallucination reduction.

Abstract: Large vision-language models (VLMs) frequently suffer from hallucinations, generating content that is inconsistent with visual inputs. Existing methods typically address this problem through post-hoc filtering, additional training objectives, or external verification, but they do not intervene during the decoding process when hallucinations arise. In this work, we introduce SAGE, a Sink-Aware Grounded Decoding framework that mitigates hallucinations by dynamically modulating self-attention during generation. Hallucinations are strongly correlated with attention sink tokens - punctuation or function tokens that accumulate disproportionate attention despite carrying limited semantic content. SAGE leverages these tokens as anchors to monitor grounding reliability in real time. At each sink trigger, the method extracts semantic concepts from the generated sequence, estimates their visual grounding using both self-attention maps and gradient-based attribution, and measures their spatial agreement. Based on this signal, self-attention distributions are adaptively sharpened or broadened to reinforce grounded regions or suppress unreliable ones. Extensive experiments across diverse hallucination benchmarks demonstrate that SAGE consistently outperforms existing decoding strategies, achieving substantial reductions in hallucination while preserving descriptive coverage, without requiring model retraining or architectural modifications. Our method achieves an average relative improvement of 10.65% on MSCOCO and 7.19% on AMBER across diverse VLM architectures, demonstrating consistent gains in hallucination mitigation.

Method: Introduces Col-Ln, a training-free token importance metric derived from Rényi entropy that can identify informative tokens from the first layer of the network, enabling more reliable token pruning in early layers without requiring additional training.

Result: Extensive experiments on Vision Transformers and Large Vision-Language Models show that Col-Ln consistently outperforms state-of-the-art pruning methods across diverse benchmarks, demonstrating effective token reduction while maintaining performance.

Conclusion: Col-Ln provides a reliable training-free solution for early-layer token pruning in Vision Transformers, addressing the limitations of [CLS]-based methods and enabling more efficient inference for high-resolution inputs.

Abstract: Vision Transformers (ViTs) achieve state-of-the-art performance but suffer from the $O(N^2)$ complexity of self-attention, making inference costly for high-resolution inputs. To address this bottleneck, token pruning has emerged as a critical technique to accelerate inference. Most existing methods rely on the [CLS] token to estimate patch importance. However, we argue that the [CLS] token can be unreliable in early layers where semantic representations are still immature. As a result, pruning in the early layer often leads to inaccurate importance estimation and unnecessary information loss. In this work, we propose a training-free token importance metric, namely Col-Ln, which is derived from Rényi entropy that enables the identification of informative tokens from the first layer of the network, thereby enabling more reliable pruning in token reduction. Extensive experiments on ViTs and Large Vision-Language Models (LVLMs) demonstrate that our approach consistently outperforms state-of-the-art pruning methods across diverse benchmarks.

Method: Lightweight network with shared encoder and task-specific decoders. Features Efficient Pyramid Mixing (EPM) module for multi-scale feature extraction using grouped convolutions, depthwise dilated convolutions, and channel shuffle operations. Decoders use Dual-Branch Upsampling (DBU) Blocks with learnable transposed convolution and bilinear interpolation branches.

Result: Achieves 92.0% mIoU for drivable-area segmentation and 32.3% IoU for lane segmentation with only 0.43M parameters and 3.95 GFLOPs. Outperforms existing segmentation models on BDD100K dataset.

Conclusion: TwinMixing provides an effective trade-off between accuracy and computational efficiency, demonstrating strong potential for real-time deployment in autonomous driving and embedded perception systems.

Abstract: Accurate and efficient perception is essential for autonomous driving, where segmentation tasks such as drivable-area and lane segmentation provide critical cues for motion planning and control. However, achieving high segmentation accuracy while maintaining real-time performance on low-cost hardware remains a challenging problem. To address this issue, we introduce TwinMixing, a lightweight multi-task segmentation model designed explicitly for drivable-area and lane segmentation. The proposed network features a shared encoder and task-specific decoders, enabling both feature sharing and task specialization. Within the encoder, we propose an Efficient Pyramid Mixing (EPM) module that enhances multi-scale feature extraction through a combination of grouped convolutions, depthwise dilated convolutions and channel shuffle operations, effectively expanding the receptive field while minimizing computational cost. Each decoder adopts a Dual-Branch Upsampling (DBU) Block composed of a learnable transposed convolution-based Fine detailed branch and a parameter-free bilinear interpolation-based Coarse grained branch, achieving detailed yet spatially consistent feature reconstruction. Extensive experiments on the BDD100K dataset validate the effectiveness of TwinMixing across three configurations - tiny, base, and large. Among them, the base configuration achieves the best trade-off between accuracy and computational efficiency, reaching 92.0% mIoU for drivable-area segmentation and 32.3% IoU for lane segmentation with only 0.43M parameters and 3.95 GFLOPs. Moreover, TwinMixing consistently outperforms existing segmentation models on the same tasks, as illustrated in Fig. 1. Thanks to its compact and modular design, TwinMixing demonstrates strong potential for real-time deployment in autonomous driving and embedded perception systems. The source code: https://github.com/Jun0se7en/TwinMixing.

Method: Fuses rectified stereo pairs at input stage using stereo micro cell tokens with row-aware patch phase positional encoding. Trains with one-view masked token distillation plus occlusion and view-specific appearance mismatch losses.

Result: In low-resource KITTI-only pretraining, BINO achieves best frozen descriptor results on dense stereo, hard negative retrieval, and KITTI Stereo 2012 disparity among baselines without linkage modules. Performs near CroCo v2 with much smaller encoder.

Conclusion: Much cross-view reasoning typically assigned to separate linkage modules can be learned inside a compact, reusable encoder, enabling efficient stereo correspondence learning.

Abstract: Stereo needs features that preserve fine cross view correspondence rather than only semantic similarity. Recent self supervised vision models transfer well, but they are not built for this goal, and geometry focused methods often rely on a binocular decoder or another explicit linkage module during pretraining. BINO asks whether strong binocular structure can instead be learned inside a compact encoder. It does this by fusing the rectified pair at the input stage, forming stereo micro cell tokens, and using a row aware patch phase positional encoding. Training uses one view masked token only distillation together with occlusion and view specific appearance mismatch. In a strict low resource setting with pretraining only on KITTI object, BINO gives the best frozen descriptor results under a no linkage probe among all compared baselines on proxy dense stereo, hard negative retrieval, and KITTI Stereo2012 disparity. With the same lightweight stereo head for every encoder, it stays near CroCov2 while using a much smaller encoder. Supplementary transfer experiments on KITTI Stereo~2015 show the same qualitative trend. These results suggest that much of the cross view reasoning often assigned to a separate linkage module can be learned inside a compact and reusable encoder.

Method: Proposes DiffAttn, a diffusion-based framework that formulates attention prediction as conditional diffusion-denoising. Uses Swin Transformer encoder, Feature Fusion Pyramid decoder for cross-layer interaction, multi-scale conditional diffusion for fine-grained context modeling, and incorporates LLM layer for semantic reasoning.

Result: Achieves state-of-the-art performance on four public datasets, surpassing video-based, top-down-feature-driven, and LLM-enhanced baselines. Framework supports interpretable driver-centric scene understanding.

Conclusion: DiffAttn effectively models drivers’ visual attention patterns and has potential applications in in-cabin human-machine interaction, risk perception, and driver state measurement for intelligent vehicles.

Abstract: Drivers’ visual attention provides critical cues for anticipating latent hazards and directly shapes decision-making and control maneuvers, where its absence can compromise traffic safety. To emulate drivers’ perception patterns and advance visual attention prediction for intelligent vehicles, we propose DiffAttn, a diffusion-based framework that formulates this task as a conditional diffusion-denoising process, enabling more accurate modeling of drivers’ attention. To capture both local and global scene features, we adopt Swin Transformer as encoder and design a decoder that combines a Feature Fusion Pyramid for cross-layer interaction with dense, multi-scale conditional diffusion to jointly enhance denoising learning and model fine-grained local and global scene contexts. Additionally, a large language model (LLM) layer is incorporated to enhance top-down semantic reasoning and improve sensitivity to safety-critical cues. Extensive experiments on four public datasets demonstrate that DiffAttn achieves state-of-the-art (SoTA) performance, surpassing most video-based, top-down-feature-driven, and LLM-enhanced baselines. Our framework further supports interpretable driver-centric scene understanding and has the potential to improve in-cabin human-machine interaction, risk perception, and drivers’ state measurement in intelligent vehicles.

Method: Proposes SOR-Track with Spatial Orthogonal Refinement module using orthogonal directional filters dynamically guided by local motion orientations to extract sharp structural responses from event streams, which serve as geometric anchors to modulate and refine aliased RGB textures through asymmetric structural modulation

Result: Extensive experiments on FE108 benchmark show SOR-Track consistently outperforms existing fusion-based trackers, particularly under motion blur and low-light conditions

Conclusion: Despite its simplicity, the method offers a principled and physics-grounded approach to multi-modal feature alignment and texture rectification for robust visual object tracking

Abstract: Robust visual object tracking (VOT) remains challenging in high-speed motion scenarios, where conventional RGB sensors suffer from severe motion blur and performance degradation. Event cameras, with microsecond temporal resolution and high dynamic range, provide complementary structural cues that can potentially compensate for these limitations. However, existing RGB-Event fusion methods typically treat event data as dense intensity representations and adopt black-box fusion strategies, failing to explicitly leverage the directional geometric priors inherently encoded in event streams to rectify degraded RGB features. To address this limitation, we propose SOR-Track, a streamlined framework for robust RGB-Event tracking based on Spatial Orthogonal Refinement (SOR). The core SOR module employs a set of orthogonal directional filters that are dynamically guided by local motion orientations to extract sharp and motion-consistent structural responses from event streams. These responses serve as geometric anchors to modulate and refine aliased RGB textures through an asymmetric structural modulation mechanism, thereby explicitly bridging structural discrepancies between two modalities. Extensive experiments on the large-scale FE108 benchmark demonstrate that SOR-Track consistently outperforms existing fusion-based trackers, particularly under motion blur and low-light conditions. Despite its simplicity, the proposed method offers a principled and physics-grounded approach to multi-modal feature alignment and texture rectification. The source code of this paper will be released on https://github.com/Event-AHU/OpenEvTracking

Method: Leverages video codec primitives (motion vectors and residuals) that natively encode video redundancy and sparsity. Introduces lightweight transformer-based encoders to aggregate codec primitives and align their representations with image encoder embeddings through pre-training.

Result: Reduces time-to-first-token by up to 86% and token usage by up to 93% compared to standard VideoLMs. Maintains or exceeds performance on 14 diverse video understanding benchmarks spanning general QA, temporal/motion reasoning, long-form understanding, and spatial scene understanding.

Conclusion: CoPE-VideoLM provides an efficient alternative to standard VideoLMs by exploiting video codec primitives, achieving significant computational savings while preserving or enhancing video understanding capabilities across multiple domains.

Abstract: Video Language Models (VideoLMs) enable AI systems to understand temporal dynamics in videos. To fit within the maximum context window constraint, current methods use keyframe sampling which often misses both macro-level events and micro-level details due to the sparse temporal coverage. Furthermore, processing full images and their tokens for each frame incurs substantial computational overhead. We address these limitations by leveraging video codec primitives (specifically motion vectors and residuals) which natively encode video redundancy and sparsity without requiring expensive full-image encoding for most frames. To this end, we introduce lightweight transformer-based encoders that aggregate codec primitives and align their representations with image encoder embeddings through a pre-training strategy that accelerates convergence during end-to-end fine-tuning. Our approach, CoPE-VideoLM, reduces the time-to-first-token by up to 86% and token usage by up to 93% compared to standard VideoLMs. Moreover, by varying the keyframe and codec primitive densities we maintain or exceed performance on 14 diverse video understanding benchmarks spanning general question answering, temporal and motion reasoning, long-form understanding, and spatial scene understanding.

Method: Proposes a pose-free framework using diffusion models to directly map natural language text to sign language videos. Introduces two innovations: 1) pose-free generative model learning implicit text-to-gesture alignments without pose estimation, and 2) Trainable Sliding Tile Attention (T-STA) mechanism that accelerates inference by exploiting spatio-temporal locality patterns with trainable sparsity.

Result: Achieves 3.07x increase in video generation speed without compromising video quality, making real-time deployment feasible. The method eliminates the train-test gap through integrated trainable sparsity.

Conclusion: The approach enables real-time, high-quality, pose-free sign language synthesis, opening new avenues for inclusive communication tools for diverse communities.

Abstract: Sign language plays a crucial role in bridging communication gaps between the deaf and hard-of-hearing communities. However, existing sign language video generation models often rely on complex intermediate representations, which limits their flexibility and efficiency. In this work, we propose a novel pose-free framework for real-time sign language video generation. Our method eliminates the need for intermediate pose representations by directly mapping natural language text to sign language videos using a diffusion-based approach. We introduce two key innovations: (1) a pose-free generative model based on the a state-of-the-art diffusion backbone, which learns implicit text-to-gesture alignments without pose estimation, and (2) a Trainable Sliding Tile Attention (T-STA) mechanism that accelerates inference by exploiting spatio-temporal locality patterns. Unlike previous training-free sparsity approaches, T-STA integrates trainable sparsity into both training and inference, ensuring consistency and eliminating the train-test gap. This approach significantly reduces computational overhead while maintaining high generation quality, making real-time deployment feasible. Our method increases video generation speed by 3.07x without compromising video quality. Our contributions open new avenues for real-time, high-quality, pose-free sign language synthesis, with potential applications in inclusive communication tools for diverse communities. Code: https://github.com/AIGeeksGroup/FlashSign.

Method: Proposes Modality-Decoupled Direct Preference Optimization (MoD-DPO) with modality-aware regularization terms: invariance to irrelevant modality corruptions and sensitivity to relevant modality perturbations, plus language-prior debiasing penalty.

Result: Extensive experiments across multiple audiovisual hallucination benchmarks show MoD-DPO consistently improves perception accuracy and hallucination resistance, outperforming previous preference optimization baselines.

Conclusion: MoD-DPO demonstrates the importance of modality-faithful alignment and provides a scalable path toward more reliable multimodal foundation models by reducing unintended cross-modal interactions.

Abstract: Omni-modal large language models (omni LLMs) have recently achieved strong performance across audiovisual understanding tasks, yet they remain highly susceptible to cross-modal hallucinations arising from spurious correlations and dominant language priors. In this work, we propose Modality-Decoupled Direct Preference Optimization (MoD-DPO), a simple and effective framework for improving modality grounding in omni LLMs. MoD-DPO introduces modality-aware regularization terms that explicitly enforce invariance to corruptions in irrelevant modalities and sensitivity to perturbations in relevant modalities, thereby reducing unintended cross-modal interactions. To further mitigate over-reliance on textual priors, we incorporate a language-prior debiasing penalty that discourages hallucination-prone text-only responses. Extensive experiments across multiple audiovisual hallucination benchmarks demonstrate that MoD-DPO consistently improves perception accuracy and hallucination resistance, outperforming previous preference optimization baselines under similar training budgets. Our findings underscore the importance of modality-faithful alignment and demonstrate a scalable path toward more reliable and resilient multimodal foundation models.

Method: Created a high-fidelity synthetic dataset using Unreal Engine environments integrated with Microsoft AirSim, generating 2094 photorealistic images across 25 diverse environments with consistent pixel-accurate labels for 20 classes relevant to autonomous navigation.

Result: Benchmarked ForestSim using state-of-the-art architectures and reported strong performance despite the inherent challenges of unstructured scenes, providing a scalable foundation for perception research.

Conclusion: ForestSim addresses the dataset scarcity for unstructured environments and supports development of perception systems for next-generation intelligent off-road vehicles.

Abstract: Robust scene understanding is essential for intelligent vehicles operating in natural, unstructured environments. While semantic segmentation datasets for structured urban driving are abundant, the datasets for extremely unstructured wild environments remain scarce due to the difficulty and cost of generating pixel-accurate annotations. These limitations hinder the development of perception systems needed for intelligent ground vehicles tasked with forestry automation, agricultural robotics, disaster response, and all-terrain mobility. To address this gap, we present ForestSim, a high-fidelity synthetic dataset designed for training and evaluating semantic segmentation models for intelligent vehicles in forested off-road and no-road environments. ForestSim contains 2094 photorealistic images across 25 diverse environments, covering multiple seasons, terrain types, and foliage densities. Using Unreal Engine environments integrated with Microsoft AirSim, we generate consistent, pixel-accurate labels across 20 classes relevant to autonomous navigation. We benchmark ForestSim using state-of-the-art architectures and report strong performance despite the inherent challenges of unstructured scenes. ForestSim provides a scalable and accessible foundation for perception research supporting the next generation of intelligent off-road vehicles. The dataset and code are publicly available: Dataset: https://vailforestsim.github.io Code: https://github.com/pragatwagle/ForestSim

Method: Proposes AmodalCG framework that: 1) assesses occlusion extent to selectively invoke MLLM guidance, 2) uses MLLMs to reason about both extent and content of missing regions, and 3) integrates guidance with visual generative models to iteratively refine completions.

Result: Experimental results on various real-world images show impressive improvements compared to all existing works, suggesting MLLMs as a promising direction for challenging amodal completion.

Conclusion: MLLMs can effectively provide real-world knowledge to guide amodal completion, addressing limitations of existing approaches and showing significant performance improvements.

Abstract: With the widespread adoption of autonomous vehicles and robotics, amodal completion, which reconstructs the occluded parts of people and objects in an image, has become increasingly crucial. Just as humans infer hidden regions based on prior experience and common sense, this task inherently requires physical knowledge about real-world entities. However, existing approaches either depend solely on the image generation ability of visual generative models, which lack such knowledge, or leverage it only during the segmentation stage, preventing it from explicitly guiding the completion process. To address this, we propose AmodalCG, a novel framework that harnesses the real-world knowledge of Multimodal Large Language Models (MLLMs) to guide amodal completion. Our framework first assesses the extent of occlusion to selectively invoke MLLM guidance only when the target object is heavily occluded. If guidance is required, the framework further incorporates MLLMs to reason about both the (1) extent and (2) content of the missing regions. Finally, a visual generative model integrates these guidance and iteratively refines imperfect completions that may arise from inaccurate MLLM guidance. Experimental results on various real-world images show impressive improvements compared to all existing works, suggesting MLLMs as a promising direction for addressing challenging amodal completion.

Method: Survey methodology that provides detailed technical walkthroughs of each model type (VAEs, GANs, normalizing flows, autoregressive/transformer generators, diffusion methods), covering their objectives, architectural components, training algorithms, optimization techniques, and limitations.

Result: A comprehensive overview of the state-of-the-art in image generation, including recent developments in video generation (from still frames to high-quality videos) and coverage of robustness and responsible deployment issues like deepfake risks, detection, artifacts, and watermarking.

Conclusion: Image generation has advanced rapidly with diverse model families, and the field now faces important challenges in extending to video generation and ensuring responsible deployment through robustness measures and ethical considerations.

Abstract: Image generation has advanced rapidly over the past decade, yet the literature seems fragmented across different models and application domains. This paper aims to offer a comprehensive survey of breakthrough image generation models, including variational autoencoders (VAEs), generative adversarial networks (GANs), normalizing flows, autoregressive and transformer-based generators, and diffusion-based methods. We provide a detailed technical walkthrough of each model type, including their underlying objectives, architectural building blocks, and algorithmic training steps. For each model type, we present the optimization techniques as well as common failure modes and limitations. We also go over recent developments in video generation and present the research works that made it possible to go from still frames to high quality videos. Lastly, we cover the growing importance of robustness and responsible deployment of these models, including deepfake risks, detection, artifacts, and watermarking.

Method: Proposes a cross-scale decoder with three mechanisms: 1) Global-local token refinement on a compact bottleneck lattice with boundary-aware regularization, 2) Gated detail bridge that selectively injects fine-scale structural cues once via cross-scale attention, 3) Uncertainty-guided class-aware point refinement that updates least reliable pixels.

Result: Achieves noise-robust and boundary-preserving segmentation tailored to off-road environments, recovering fine structural details while maintaining deployment-friendly efficiency. Shows consistent improvements on standard off-road benchmarks without heavy dense feature fusion.

Conclusion: The framework effectively addresses off-road segmentation challenges through complementary mechanisms that handle annotation ambiguity, preserve boundaries, and refine uncertain regions with minimal computational overhead.

Abstract: Off-road semantic segmentation is fundamentally challenged by irregular terrain, vegetation clutter, and inherent annotation ambiguity. Unlike urban scenes with crisp object boundaries, off-road environments exhibit strong class-level similarity among terrain categories, resulting in thick and uncertain transition regions that degrade boundary coherence and destabilize training. Rare or thin structures, such as narrow traversable gaps or isolated obstacles, further receive sparse and unreliable supervision and are easily overwhelmed by dominant background textures. Existing decoder designs either rely on low-scale bottlenecks that oversmooth fine structural details, or repeatedly fuse high-detail features, which tends to amplify annotation noise and incur substantial computational cost. We present a cross-scale decoder that explicitly addresses these challenges through three complementary mechanisms. First, a global–local token refinement module consolidates semantic context on a compact bottleneck lattice, guided by boundary-aware regularization to remain robust under ambiguous supervision. Second, a gated detail bridge selectively injects fine-scale structural cues only once through cross-scale attention, preserving boundary and texture information while avoiding noise accumulation. Third, an uncertainty-guided class-aware point refinement selectively updates the least reliable pixels, improving rare and ambiguous structures with minimal computational overhead. The resulting framework achieves noise-robust and boundary-preserving segmentation tailored to off-road environments, recovering fine structural details while maintaining deployment-friendly efficiency. Experimental results on standard off-road benchmarks demonstrate consistent improvements over prior approaches without resorting to heavy dense feature fusion.

Method: Uses rehearsal stage scenes captured under stable lighting as geometric consistency reference. Employs learnable lighting vectors to represent illumination colors temporally and disentangle projected light colors from scene radiance. Incorporates optical flow regularization with off-the-shelf interactive masks for dynamic object reconstruction.

Result: Demonstrates robust performance on novel view synthesis and scene editing under dynamic illumination conditions. Shows effective disentanglement of scene components from lighting variations.

Conclusion: RehearsalNeRF provides an effective solution for learning disentangled neural fields under severe illumination changes by leveraging rehearsal stage data and lighting vector representations.

Abstract: Although there has been significant progress in neural radiance fields, an issue on dynamic illumination changes still remains unsolved. Different from relevant works that parameterize time-variant/-invariant components in scenes, subjects’ radiance is highly entangled with their own emitted radiance and lighting colors in spatio-temporal domain. In this paper, we present a new effective method to learn disentangled neural fields under the severe illumination changes, named RehearsalNeRF. Our key idea is to leverage scenes captured under stable lighting like rehearsal stages, easily taken before dynamic illumination occurs, to enforce geometric consistency between the different lighting conditions. In particular, RehearsalNeRF employs a learnable vector for lighting effects which represents illumination colors in a temporal dimension and is used to disentangle projected light colors from scene radiance. Furthermore, our RehearsalNeRF is also able to reconstruct the neural fields of dynamic objects by simply adopting off-the-shelf interactive masks. To decouple the dynamic objects, we propose a new regularization leveraging optical flow, which provides coarse supervision for the color disentanglement. We demonstrate the effectiveness of RehearsalNeRF by showing robust performances on novel view synthesis and scene editing under dynamic illumination conditions. Our source code and video datasets will be publicly available.

Method: Introduced MathGen benchmark with 900 problems spanning 7 core mathematical domains, each paired with an executable verifier using Script-as-a-Judge protocol for deterministic and objective evaluation. Tested representative open-source and proprietary text-to-image models.

Result: Mathematical fidelity remains a major bottleneck: best closed-source model reached only 42.0% overall accuracy, while open-source models achieved just ~1-11%, often near 0% on structured tasks. Current T2I models remain far from competent at elementary mathematical visual generation.

Conclusion: Current text-to-image models struggle significantly with generating mathematically correct visual representations, indicating a substantial gap in their ability to handle precise mathematical visual generation tasks.

Abstract: Modern generative models have demonstrated the ability to solve challenging mathematical problems. In many real-world settings, however, mathematical solutions must be expressed visually through diagrams, plots, geometric constructions, and structured symbolic layouts, where correctness depends on precise visual composition. Can generative models still do so when the answer must be rendered visually rather than written in text? To study this problem, we introduce MathGen, a rigorous benchmark of 900 problems spanning seven core domains, each paired with an executable verifier under a Script-as-a-Judge protocol for deterministic and objective evaluation. Experiments on representative open-source and proprietary text-to-image models show that mathematical fidelity remains a major bottleneck: even the best closed-source model reaches only 42.0% overall accuracy, while open-source models achieve just ~ 1-11%, often near 0% on structured tasks. Overall, current T2I models remain far from competent at even elementary mathematical visual generation.

Method: ExFusion upcycles Transformer FFNs into multi-expert configurations during initialization, assigning each expert a fusion weight. During training, these weights fuse multiple experts into a single unified expert equivalent to the original FFN for forward computation. After training, learned weights integrate multi-experts into a single expert.

Result: Extensive experiments on computer vision and NLP tasks demonstrate effectiveness. The method introduces multi-expert characteristics with only marginal computational cost compared to standard dense training, and eliminates storage/deployment overhead.

Conclusion: ExFusion provides an efficient approach to leverage multi-expert capabilities in Transformers without the computational and deployment costs typically associated with MoE models.

Abstract: Mixture-of-Experts (MoE) models substantially improve performance by increasing the capacity of dense architectures. However, directly training MoE models requires considerable computational resources and introduces extra overhead in parameter storage and deployment. Therefore, it is critical to develop an approach that leverages the multi-expert capability of MoE to enhance performance while incurring minimal additional cost. To this end, we propose a novel pre-training approach, termed ExFusion, which improves the efficiency of Transformer training through multi-expert fusion. Specifically, during the initialization phase, ExFusion upcycles the feed-forward network (FFN) of the Transformer into a multi-expert configuration, where each expert is assigned a weight for later parameter fusion. During training, these weights allow multiple experts to be fused into a single unified expert equivalent to the original FFN, which is subsequently used for forward computation. As a result, ExFusion introduces multi-expert characteristics into the training process while incurring only marginal computational cost compared to standard dense training. After training, the learned weights are used to integrate multi-experts into a single unified expert, thereby eliminating additional overhead in storage and deployment. Extensive experiments on a variety of computer vision and natural language processing tasks demonstrate the effectiveness of the proposed method.

Method: Hardware-aware optimization framework that systematically identifies and prunes structural redundancies in DiT backbones that are particularly taxing for mobile data-flows, creating lightweight models for mobile NPUs like Qualcomm Hexagon and Apple Neural Engine.

Result: Achieves 20-30% reduction in parameters, 36-46% decrease in FLOPs, and 1.65-fold reduction in on-device latency while maintaining scaling advantages and expressive capacity. Offers superior Pareto-optimal trade-off between FID and inference latency compared to optimized mobile U-Nets and vanilla DiT variants.

Conclusion: EdgeDiT enables responsive, private, and offline generative AI directly on-device, providing a scalable blueprint for transitioning large-scale foundation models from high-end GPUs to mobile devices.

Abstract: Diffusion Transformers (DiT) have established a new state-of-the-art in high-fidelity image synthesis; however, their massive computational complexity and memory requirements hinder local deployment on resource-constrained edge devices. In this paper, we introduce EdgeDiT, a family of hardware-efficient generative transformers specifically engineered for mobile Neural Processing Units (NPUs), such as the Qualcomm Hexagon and Apple Neural Engine (ANE). By leveraging a hardware-aware optimization framework, we systematically identify and prune structural redundancies within the DiT backbone that are particularly taxing for mobile data-flows. Our approach yields a series of lightweight models that achieve a 20-30% reduction in parameters, a 36-46% decrease in FLOPs, and a 1.65-fold reduction in on-device latency without sacrificing the scaling advantages or the expressive capacity of the original transformer architecture. Extensive benchmarking demonstrates that EdgeDiT offers a superior Pareto-optimal trade-off between Frechet Inception Distance (FID) and inference latency compared to both optimized mobile U-Nets and vanilla DiT variants. By enabling responsive, private, and offline generative AI directly on-device, EdgeDiT provides a scalable blueprint for transitioning large-scale foundation models from high-end GPUs to the palm of the user.

Method: Created Cross-View Relations (XVR) dataset with 100K vision-question-answer samples from 18K diverse 3D scenes and 70K robotic manipulation trajectories, covering three spatial reasoning tasks: Correspondence, Verification, and Localization. Fine-tuned VLMs on this dataset.

Result: VLMs fine-tuned on XVR achieved substantial improvements on multi-view and robotic spatial reasoning benchmarks (MindCube and RoboSpatial). When integrated into Vision-Language-Action models, XVR-trained representations improved success rates on RoboCasa robotic manipulation tasks.

Conclusion: Explicit training on cross-view spatial relations significantly enhances multi-view reasoning capabilities in vision-language models and transfers effectively to real-world robotic manipulation, bridging the gap between 2D vision understanding and 3D spatial reasoning.

Abstract: Vision-language models (VLMs) have achieved impressive results on single-view vision tasks, but lack the multi-view spatial reasoning capabilities essential for embodied AI systems to understand 3D environments and manipulate objects across different viewpoints. In this work, we introduce Cross-View Relations (XVR), a large-scale dataset designed to teach VLMs spatial reasoning across multiple views. XVR comprises 100K vision-question-answer samples derived from 18K diverse 3D scenes and 70K robotic manipulation trajectories, spanning three fundamental spatial reasoning tasks: Correspondence (matching objects across views), Verification (validating spatial relationships), and Localization (identifying object positions). VLMs fine-tuned on XVR achieve substantial improvements on established multi-view and robotic spatial reasoning benchmarks (MindCube and RoboSpatial). When integrated as backbones in Vision-Language-Action models, XVR-trained representations improve success rates on RoboCasa. Our results demonstrate that explicit training on cross-view spatial relations significantly enhances multi-view reasoning and transfers effectively to real-world robotic manipulation.

Method: Proposes a heterogeneous graph that represents mapping between segmented 2D and 3D regions, enabling cross-modal feature interaction and correspondence pruning. The method mines multi-path feature relationships, adapts features under heterogeneous edge guidance, and prunes correspondences using graph-based projection consistency.

Result: Experiments on six indoor and outdoor benchmarks under cross-domain setups show Hg-I2P significantly outperforms existing methods in both generalization and accuracy.

Conclusion: The heterogeneous graph approach effectively bridges the modality gap in I2P registration, improving feature discriminability and correspondence reliability for better generalization across domains.

Abstract: Image-to-point-cloud (I2P) registration aims to align 2D images with 3D point clouds by establishing reliable 2D-3D correspondences. The drastic modality gap between images and point clouds makes it challenging to learn features that are both discriminative and generalizable, leading to severe performance drops in unseen scenarios. We address this challenge by introducing a heterogeneous graph that enables refining both cross-modal features and correspondences within a unified architecture. The proposed graph represents a mapping between segmented 2D and 3D regions, which enhances cross-modal feature interaction and thus improves feature discriminability. In addition, modeling the consistency among vertices and edges within the graph enables pruning of unreliable correspondences. Building on these insights, we propose a heterogeneous graph embedded I2P registration method, termed Hg-I2P. It learns a heterogeneous graph by mining multi-path feature relationships, adapts features under the guidance of heterogeneous edges, and prunes correspondences using graph-based projection consistency. Experiments on six indoor and outdoor benchmarks under cross-domain setups demonstrate that Hg-I2P significantly outperforms existing methods in both generalization and accuracy. Code is released on https://github.com/anpei96/hg-i2p-demo.

Method: Created AffordBridge dataset with point cloud annotations and linked RGB images, then proposed AffordMatcher method that establishes semantic correspondences between image-based and point cloud-based instances for keypoint matching to identify affordance regions using visual signifiers.

Result: Experimental results demonstrate effectiveness compared to other methods on the new dataset.

Conclusion: The work provides a large-scale dataset and method for scene-level affordance learning through cross-modal matching between images and point clouds.

Abstract: Affordance learning is a complex challenge in many applications, where existing approaches primarily focus on the geometric structures, visual knowledge, and affordance labels of objects to determine interactable regions. However, extending this learning capability to a scene is significantly more complicated, as incorporating object- and scene-level semantics is not straightforward. In this work, we introduce AffordBridge, a large-scale dataset with 291,637 functional interaction annotations across 685 high-resolution indoor scenes in the form of point clouds. Our affordance annotations are complemented by RGB images that are linked to the same instances within the scenes. Building upon our dataset, we propose AffordMatcher, an affordance learning method that establishes coherent semantic correspondences between image-based and point cloud-based instances for keypoint matching, enabling a more precise identification of affordance regions based on cues, so-called visual signifiers. Experimental results on our dataset demonstrate the effectiveness of our approach compared to other methods.

Method: Two main components: 1) Neighborhood-Aware Anchor Pruning (NAAP) evaluates anchor importance via weighted neighborhood feature aggregation and merges redundant anchors into salient neighbors. 2) Hierarchical entropy coding scheme with lightweight Geometry-Guided Convolution (GG-Conv) operator for spatially adaptive context modeling and rate-distortion optimization.

Result: Extensive experiments demonstrate that GeoHCC effectively resolves the structure preservation bottleneck, maintaining superior geometric integrity and rendering fidelity over state-of-the-art anchor-based approaches.

Conclusion: GeoHCC provides a geometry-aware compression framework for 3DGS that achieves better compression while preserving geometric structure and rendering quality compared to existing methods.

Abstract: Although 3D Gaussian Splatting (3DGS) enables high-fidelity real-time rendering, its prohibitive storage overhead severely hinders practical deployment. Recent anchor-based 3DGS compression schemes reduce redundancy through context modeling, yet overlook explicit geometric dependencies, leading to structural degradation and suboptimal rate-distortion performance. In this paper, we propose GeoHCC, a geometry-aware 3DGS compression framework that incorporates inter-anchor geometric correlations into anchor pruning and entropy coding for compact representation. We first introduce Neighborhood-Aware Anchor Pruning (NAAP), which evaluates anchor importance via weighted neighborhood feature aggregation and merges redundant anchors into salient neighbors, yielding a compact yet geometry-consistent anchor set. Building upon this optimized structure, we further develop a hierarchical entropy coding scheme, in which coarse-to-fine priors are exploited through a lightweight Geometry-Guided Convolution (GG-Conv) operator to enable spatially adaptive context modeling and rate-distortion optimization. Extensive experiments demonstrate that GeoHCC effectively resolves the structure preservation bottleneck, maintaining superior geometric integrity and rendering fidelity over state-of-the-art anchor-based approaches.

Method: Proposes a dual-branch framework with Task-Specific Physical Grounding Module (TS-PGM) to extract degradation-aware priors (rain masks, dark channels), which guide a Retinex decomposition network via Physical-conditioned Multi-head Self-Attention (PC-MSA) mechanism for robust reflection and illumination correction.

Result: Achieved 4th place in NTIRE 2026 Day and Night Raindrop Removal Challenge and 5th place in Joint Noise Low-light Enhancement Challenge, demonstrating state-of-the-art performance and exceptional generalizability across different restoration tasks.

Conclusion: RetinexDualV2 provides an effective physically motivated approach for diverse UHD image restoration that maintains a single architecture while handling various complex degradations through explicit physical priors and conditioning mechanisms.

Abstract: We propose RetinexDualV2, a unified, physically grounded dual-branch framework for diverse Ultra-High-Definition (UHD) image restoration. Unlike generic models, our method employs a Task-Specific Physical Grounding Module (TS-PGM) to extract degradation-aware priors (e.g., rain masks and dark channels). These explicitly guide a Retinex decomposition network via a novel Physical-conditioned Multi-head Self-Attention (PC-MSA) mechanism, enabling robust reflection and illumination correction. This physical conditioning allows a single architecture to handle various complex degradations seamlessly, without task-specific structural modifications. RetinexDualV2 demonstrates exceptional generalizability, securing 4\textsuperscript{th} place in the NTIRE 2026 Day and Night Raindrop Removal Challenge and 5\textsuperscript{th} place in the Joint Noise Low-light Enhancement (JNLLIE) Challenge. Extensive experiments confirm the state-of-the-art performance and efficiency of our physically motivated approach.

Method: Developed a domain-specific porcelain connoisseurship agent supporting multi-image inputs, vision tool invocation, and multimodal retrieval-augmented generation. Created CiQi-VQA dataset (29,596 specimens, 51,553 images, 557,940 VQA pairs) and CiQi-Bench benchmark. Trained using supervised fine-tuning, reinforcement learning, and tool-augmented reasoning framework with vision and multimodal retrieval tools.

Result: CiQi-Agent (7B) outperforms all competitive open- and closed-source models across all six attributes on CiQi-Bench, achieving on average 12.2% higher accuracy than GPT-5.

Conclusion: CiQi-Agent successfully enables intelligent porcelain analysis, captures subtle visual details, retrieves relevant domain knowledge, and produces coherent, explainable connoisseurship descriptions, advancing multimodal AI applications in cultural heritage.

Abstract: The connoisseurship of antique Chinese porcelain demands extensive historical expertise, material understanding, and aesthetic sensitivity, making it difficult for non-specialists to engage. To democratize cultural-heritage understanding and assist expert connoisseurship, we introduce CiQi-Agent – a domain-specific Porcelain Connoisseurship Agent for intelligent analysis of antique Chinese porcelain. CiQi-Agent supports multi-image porcelain inputs and enables vision tool invocation and multimodal retrieval-augmented generation, performing fine-grained connoisseurship analysis across six attributes: dynasty, reign period, kiln site, glaze color, decorative motif, and vessel shape. Beyond attribute classification, it captures subtle visual details, retrieves relevant domain knowledge, and integrates visual and textual evidence to produce coherent, explainable connoisseurship descriptions. To achieve this capability, we construct a large-scale, expert-annotated dataset CiQi-VQA, comprising 29,596 porcelain specimens, 51,553 images, and 557,940 visual question–answering pairs, and further establish a comprehensive benchmark CiQi-Bench aligned with the previously mentioned six attributes. CiQi-Agent is trained through supervised fine-tuning, reinforcement learning, and a tool-augmented reasoning framework that integrates two categories of tools: a vision tool and multimodal retrieval tools. Experimental results show that CiQi-Agent (7B) outperforms all competitive open- and closed-source models across all six attributes on CiQi-Bench, achieving on average 12.2% higher accuracy than GPT-5. The model and dataset have been released and are publicly available at https://huggingface.co/datasets/SII-Monument-Valley/CiQi-VQA.

Method: PPCR structures reasoning as a Semantic Understanding-Spatial Grounding-Instance Segmentation pipeline using MLLMs to generate Semantic Segmentation Prompts for semantic cues, then Spatial Segmentation Prompts for object location and spatial extent.

Result: Extensive experiments on standard benchmarks show PPCR consistently outperforms existing methods.

Conclusion: PPCR effectively bridges linguistic descriptions with object-level visual representations through progressive prompt-guided reasoning, advancing referring image segmentation.

Abstract: Referring image segmentation aims to localize and segment a target object in an image based on a free-form referring expression. The core challenge lies in effectively bridging linguistic descriptions with object-level visual representations, especially when referring expressions involve detailed attributes and complex inter-object relationships. Existing methods either rely on cross-modal alignment or employ Semantic Segmentation Prompts, but they often lack explicit reasoning mechanisms for grounding language descriptions to target regions in the image. To address these limitations, we propose PPCR, a Progressive Prompt-guided Cross-modal Reasoning framework for referring image segmentation. PPCR explicitly structures the reasoning process as a Semantic Understanding-Spatial Grounding-Instance Segmentation pipeline. Specifically, PPCR first employs multimodal large language models (MLLMs) to generate Semantic Segmentation Prompt that capture key semantic cues of the target object. Based on this semantic context, Spatial Segmentation Prompt are further generated to reason about object location and spatial extent, enabling a progressive transition from semantic understanding to spatial grounding. The Semantic and Spatial Segmentation prompts are then jointly integrated into the segmentation module to guide accurate target localization and segmentation. Extensive experiments on standard referring image segmentation benchmarks demonstrate that PPCR consistently outperforms existing methods. The code will be publicly released to facilitate reproducibility.

Method: Proposes UniDA3D with: 1) Unified multi-target domain adaptation treating different adverse conditions as unified adaptation problem; 2) Query Guided Domain Discrepancy Mitigation (QDDM) module aligning object features via query-centric adversarial and contrastive learning; 3) Domain-adaptive teacher-student pipeline with exponential-moving-average teacher and dynamically updated pseudo labels for consistency learning.

Result: Outperforms state-of-the-art camera-only multi-view 3D detectors on synthesized nuScenes benchmarks (nuScenes-Night, nuScenes-Rain, nuScenes-Haze), achieving substantial gains in mAP and NDS while maintaining real-time inference efficiency.

Conclusion: UniDA3D enables robust all-weather 3D perception through unified domain adaptation, eliminating the need for separate training per condition while maintaining efficiency and performance under extreme environmental conditions.

Abstract: Camera-only 3D object detection is critical for autonomous driving, offering a cost-effective alternative to LiDAR based methods. In particular, multi-view 3D object detection has emerged as a promising direction due to its balanced trade-off between performance and cost. However, existing methods often suffer significant performance degradation under complex environmental conditions such as nighttime, fog, and rain, primarily due to their reliance on training data collected mostly in ideal conditions. To address this challenge, we propose UniDA3D, a unified domain-adaptive multi-view 3D object detector designed for robust perception under diverse adverse conditions. UniDA3D formulates nighttime, rainy, and foggy scenes as a unified multi target domain adaptation problem and leverages a novel query guided domain discrepancy mitigation (QDDM) module to align object features between source and target domains at both batch and global levels via query-centric adversarial and contrastive learning. Furthermore, we introduce a domain-adaptive teacher student training pipeline with an exponential-moving-average teacher and dynamically updated high-quality pseudo labels to enhance consistency learning and suppress background noise in unlabeled target domains. In contrast to prior approaches that require separate training for each condition, UniDA3D performs a single unified training process across multiple domains, enabling robust all-weather 3D perception. On a synthesized multi-view 3D benchmark constructed by generating nighttime, rainy, and foggy counterparts from nuScenes (nuScenes-Night, nuScenes-Rain, and nuScenes-Haze), UniDA3D consistently outperforms state of-the-art camera-only multi-view 3D detectors under extreme conditions, achieving substantial gains in mAP and NDS while maintaining real-time inference efficiency.

Method: Hydra uses a dual-head vision-language model with a single LoRA adapter trained only for retrieval. At inference, toggling the adapter enables multi-vector embeddings for retrieval or disables it for generation. The approach requires attention-mode restoration, lm_head preservation, and KV-cache-aware decoding to maintain generation quality.

Result: Hydra achieves byte-identical outputs in 100% of 10,500 samples compared to base model generation, with max delta-ANLS = 0.0044 across 15,301 samples on four VQA benchmarks. It reduces peak GPU memory by 41% and generalizes to audio retrieval and video embedding with speech generation in Qwen2.5-Omni-3B.

Conclusion: Hydra demonstrates that a single model can efficiently handle both retrieval and generation tasks through adapter toggling, significantly reducing memory requirements while maintaining generation quality, with potential extensions to multimodal domains.

Abstract: Visual document understanding typically requires separate retrieval and generation models, doubling memory and system complexity. We present Hydra, a dual-head approach that provides both ColBERT-style late-interaction retrieval and autoregressive generation from a single vision-language model (VLM). A single LoRA adapter, trained only for retrieval, is toggled at inference: enabling it produces multi-vector embeddings; disabling it recovers the base model’s generation quality – byte-identical outputs in 100% of 10,500 greedy and stochastic samples, with max delta-ANLS = 0.0044 across 15,301 samples on four VQA benchmarks (three informative; ChartQA is near-zero for both models under greedy decoding) when compared against an independent base-model pipeline. We identify three engineering requirements (attention-mode restoration, lm_head preservation, KV-cache-aware decoding) whose omission silently breaks generation despite correct weight recovery. On ViDoRe V1, Hydra (4B) is within 1 percentage point of a controlled single-head baseline in a single training run, with higher aggregate scores on V2 and V3 that are concentrated on a subset of tasks; multi-seed experiments are needed to confirm these trends. The single-model design reduces peak GPU memory by 41%, though adapter switching introduces throughput overhead under concurrent serving loads. An ablation shows that GritLM-style joint training provides no benefit within the LoRA-based (r=16) training regime. A proof-of-concept extension to Qwen2.5-Omni-3B demonstrates that the mechanism generalizes to audio retrieval and video embedding, with speech generation.

Method: CLIP-AU uses Action Units (AUs) as structured textual prompts within CLIP to model fine-grained facial expressions. CLIP-AUTT adds test-time personalization with entropy-guided temporal window selection and prompt tuning to adapt to unseen subjects while preserving temporal consistency.

Result: Outperforms state-of-the-art CLIP-based FER and TTA methods on three challenging video-based subtle ER datasets (BioVid, StressID, and BAH), achieving robust and personalized subtle emotion recognition.

Conclusion: AU-guided temporal learning with test-time personalization enables fine-grained, subject-adaptive emotion recognition without CLIP fine-tuning or LLM-generated text supervision, addressing limitations of existing CLIP-based approaches.

Abstract: Personalization in emotion recognition (ER) is essential for an accurate interpretation of subtle and subject-specific expressive patterns. Recent advances in vision-language models (VLMs) such as CLIP demonstrate strong potential for leveraging joint image-text representations in ER. However, CLIP-based methods either depend on CLIP’s contrastive pretraining or on LLMs to generate descriptive text prompts, which are noisy, computationally expensive, and fail to capture fine-grained expressions, leading to degraded performance. In this work, we leverage Action Units (AUs) as structured textual prompts within CLIP to model fine-grained facial expressions. AUs encode the subtle muscle activations underlying expressions, providing localized and interpretable semantic cues for more robust ER. We introduce CLIP-AU, a lightweight AU-guided temporal learning method that integrates interpretable AU semantics into CLIP. It learns generic, subject-agnostic representations by aligning AU prompts with facial dynamics, enabling fine-grained ER without CLIP fine-tuning or LLM-generated text supervision. Although CLIP-AU models fine-grained AU semantics, it does not adapt to subject-specific variability in subtle expressions. To address this limitation, we propose CLIP-AUTT, a video-based test-time personalization method that dynamically adapts AU prompts to videos from unseen subjects. By combining entropy-guided temporal window selection with prompt tuning, CLIP-AUTT enables subject-specific adaptation while preserving temporal consistency. Our extensive experiments on three challenging video-based subtle ER datasets, BioVid, StressID, and BAH, indicate that CLIP-AU and CLIP-AUTT outperform state-of-the-art CLIP-based FER and TTA methods, achieving robust and personalized subtle ER.

Method: DiCoOp extends CoOp with adversarial training to learn domain-invariant prompts. It forces the model to learn prompts that are invariant across different domains while preserving discriminative power for classification tasks.

Result: Experimental results show DiCoOp consistently surpasses CoOp in domain generalization tasks across diverse visual domains, demonstrating improved robustness to distribution shifts.

Conclusion: DiCoOp provides an effective extension to CoOp for domain generalization by learning domain-invariant prompts through adversarial training, addressing the limitation of handling unseen domain distributions.

Abstract: Large pre-trained vision-language models like CLIP have transformed computer vision by aligning images and text in a shared feature space, enabling robust zero-shot transfer via prompting. Soft-prompting, such as Context Optimization (CoOp), effectively adapts these models for downstream recognition tasks by learning a set of context vectors. However, CoOp lacks explicit mechanisms for handling domain shifts across unseen distributions. To address this, we propose Domain-invariant Context Optimization (DiCoOp), an extension of CoOp optimized for domain generalization. By employing an adversarial training approach, DiCoOp forces the model to learn domain-invariant prompts while preserving discriminative power for classification. Experimental results show that DiCoOp consistently surpasses CoOp in domain generalization tasks across diverse visual domains.

Method: Two-stage pipeline: 1) Learn stable geometry-driven base appearance capturing global facial structure and coarse expression-dependent variations. 2) Predict personalized residual details (high-frequency components, wrinkles, subtle skin deformations) not captured in first stage. Uses dynamic appearance fusion to integrate residual details after deformation with spatial and semantic alignment.

Result: DipGuava consistently outperforms prior methods in both visual quality and quantitative performance, generating photorealistic, identity-preserving avatars with personalized attributes.

Conclusion: DipGuava successfully addresses the limitation of existing methods by explicitly disentangling facial appearance, enabling high-fidelity 3D head avatar creation with personalized details from monocular video.

Abstract: While recent 3D head avatar creation methods attempt to animate facial dynamics, they often fail to capture personalized details, limiting realism and expressiveness. To fill this gap, we present DipGuava (Disentangled and Personalized Gaussian UV Avatar), a novel 3D Gaussian head avatar creation method that successfully generates avatars with personalized attributes from monocular video. DipGuava is the first method to explicitly disentangle facial appearance into two complementary components, trained in a structured two-stage pipeline that significantly reduces learning ambiguity and enhances reconstruction fidelity. In the first stage, we learn a stable geometry-driven base appearance that captures global facial structure and coarse expression-dependent variations. In the second stage, the personalized residual details not captured in the first stage are predicted, including high-frequency components and nonlinearly varying features such as wrinkles and subtle skin deformations. These components are fused via dynamic appearance fusion that integrates residual details after deformation, ensuring spatial and semantic alignment. This disentangled design enables DipGuava to generate photorealistic, identity-preserving avatars, consistently outperforming prior methods in both visual quality and quantitativeperformance, as demonstrated in extensive experiments.

Method: Proposes an energy-aware imitation learning framework with Energy-driven Cross-modality Fusion Module (ECFM) and energy-aware decoder to fuse event and frame data for steering prediction.

Result: Outperforms existing state-of-the-art approaches on two public real-world datasets (DDD20 and DRFuser).

Conclusion: The framework effectively leverages complementary event and frame modalities for reliable and safe steering prediction in autonomous driving.

Abstract: In autonomous driving, relying solely on frame-based cameras can lead to inaccuracies caused by factors like long exposure times, high-speed motion, and challenging lighting conditions. To address these issues, we introduce a bio-inspired vision sensor known as the event camera. Unlike conventional cameras, event cameras capture sparse, asynchronous events that provide a complementary modality to mitigate these challenges. In this work, we propose an energy-aware imitation learning framework for steering prediction that leverages both events and frames. Specifically, we design an Energy-driven Cross-modality Fusion Module (ECFM) and an energy-aware decoder to produce reliable and safe predictions. Extensive experiments on two public real-world datasets, DDD20 and DRFuser, demonstrate that our method outperforms existing state-of-the-art (SOTA) approaches. The codes and trained models will be released upon acceptance.

Method: The paper likely proposes specialized adversarial robustness methods for semantic segmentation in robotic perception, potentially including novel detection strategies and architecture adaptations beyond standard image classification defenses.

Result: Not specified in the abstract, but presumably demonstrates improved robustness against adversarial attacks for semantic segmentation in robotic applications compared to standard approaches.

Conclusion: Robust semantic segmentation for robotics requires specialized adversarial defense approaches distinct from image classification methods to ensure safety in real-world applications.

Abstract: Deep Neural Networks (DNNs) achieve strong performance in semantic segmentation for robotic perception but remain vulnerable to adversarial attacks, threatening safety-critical applications. While robustness has been studied for image classification, semantic segmentation in robotic contexts requires specialized architectures and detection strategies.

Method: Uses a two-branch approach: image-exposure (IE) branch for standard camera observations and Gaussian-illumination (GI) branch for illumination-dependent appearance changes. Introduces cross-branch HDR consistency loss for explicit HDR supervision and illumination-guided gradient scaling to mitigate exposure-biased gradient starvation.

Result: Achieves superior HDR detail reconstruction with 2.04 dB PSNR gain over HDR-GS while maintaining real-time rendering speed up to 76 FPS across realistic and synthetic datasets.

Conclusion: PhysHDR-GS effectively addresses limitations in HDR novel view synthesis by physically modeling scene appearance and providing explicit HDR supervision, enabling high-quality HDR reconstruction with real-time performance.

Abstract: High dynamic range novel view synthesis (HDR-NVS) reconstructs scenes with dynamic details by fusing multi-exposure low dynamic range (LDR) views, yet it struggles to capture ambient illumination-dependent appearance. Implicitly supervising HDR content by constraining tone-mapped results fails in correcting abnormal HDR values, and results in limited gradients for Gaussians in under/over-exposed regions. To this end, we introduce PhysHDR-GS, a physically inspired HDR-NVS framework that models scene appearance via intrinsic reflectance and adjustable ambient illumination. PhysHDR-GS employs a complementary image-exposure (IE) branch and Gaussian-illumination (GI) branch to faithfully reproduce standard camera observations and capture illumination-dependent appearance changes, respectively. During training, the proposed cross-branch HDR consistency loss provides explicit supervision for HDR content, while an illumination-guided gradient scaling strategy mitigates exposure-biased gradient starvation and reduces under-densified representations. Experimental results across realistic and synthetic datasets demonstrate our superiority in reconstructing HDR details (e.g., a PSNR gain of 2.04 dB over HDR-GS), while maintaining real-time rendering speed (up to 76 FPS). Code and models are available at https://huimin-zeng.github.io/PhysHDR-GS/.

Method: Three key innovations: 1) Modality-aware CLIP (MA-CLIP) with LoRA fine-tuning for modality-specific guidance, 2) Modality-aligned Embeddings for fine-grained features, 3) Domain-specific Refinement Module (DSRM) for dynamic feature adjustment.

Result: Achieves state-of-the-art performance with 65.03% mIoU across five diverse datasets with complementary modalities, surpassing specialized multi-modal methods.

Conclusion: Proposes a universal framework that effectively unifies semantic segmentation across multiple sensor modalities, demonstrating strong generalization capabilities.

Abstract: Semantic segmentation across arbitrary sensor modalities faces significant challenges due to diverse sensor characteristics, and the traditional configurations for this task result in redundant development efforts. We address these challenges by introducing a universal arbitrary-modal semantic segmentation framework that unifies segmentation across multiple modalities. Our approach features three key innovations: (1) the Modality-aware CLIP (MA-CLIP), which provides modality-specific scene understanding guidance through LoRA fine-tuning; (2) Modality-aligned Embeddings for capturing fine-grained features; and (3) the Domain-specific Refinement Module (DSRM) for dynamic feature adjustment. Evaluated on five diverse datasets with different complementary modalities (event, thermal, depth, polarization, and light field), our model surpasses specialized multi-modal methods and achieves state-of-the-art performance with a mIoU of 65.03%. The codes will be released upon acceptance.

Method: Proposes CFG-SAM with three components: 1) Multi-scale Adaptive Local-aware Adapter for capability transfer with minimal parameters and dynamic multi-scale convolutional kernels, 2) Hierarchical Modulated Fusion Module for dynamic aggregation of multi-level encoder features to preserve spatial details, and 3) Boundary-Guided Mask Refinement that integrates multi-context boundary cues with semantic features through explicit supervision.

Result: The framework enables SAM to perform accurate nuclei instance segmentation directly by cooperatively enhancing local perception, preserving spatial details, and refining boundaries through parameter-efficient fine-tuning.

Conclusion: CFG-SAM successfully adapts SAM for medical imaging tasks by addressing its limitations in local feature perception while maintaining computational efficiency, making it suitable for nuclei instance segmentation in computational pathology.

Abstract: Nuclei instance segmentation is critical in computational pathology for cancer diagnosis and prognosis. Recently, the Segment Anything Model has demonstrated exceptional performance in various segmentation tasks, leveraging its rich priors and powerful global context modeling capabilities derived from large-scale pre-training on natural images. However, directly applying SAM to the medical imaging domain faces significant limitations: it lacks sufficient perception of the local structural features that are crucial for nuclei segmentation, and full fine-tuning for downstream tasks requires substantial computational costs. To efficiently transfer SAM’s robust prior knowledge to nuclei instance segmentation while supplementing its task-aware local perception, we propose a parameter-efficient fine-tuning framework, named Cooperative Fine-Grained Refinement of SAM, consisting of three core components: 1) a Multi-scale Adaptive Local-aware Adapter, which enables effective capability transfer by augmenting the frozen SAM backbone with minimal parameters and instilling a powerful perception of local structures through dynamically generated, multi-scale convolutional kernels; 2) a Hierarchical Modulated Fusion Module, which dynamically aggregates multi-level encoder features to preserve fine-grained spatial details; and 3) a Boundary-Guided Mask Refinement, which integrates multi-context boundary cues with semantic features through explicit supervision, producing a boundary-focused signal to refine initial mask predictions for sharper delineation. These three components work cooperatively to enhance local perception, preserve spatial details, and refine boundaries, enabling SAM to perform accurate nuclei instance segmentation directly.

Method: Analysis of subgroup-level error distribution (false positive rate and false negative rate) to demonstrate how aggregate metrics obscure demographic disparities, plus examination of operational risks in law enforcement applications.

Result: Empirical observations show systems with similar overall accuracy can have substantially different fairness profiles, with subgroup error rates varying significantly despite identical aggregate metrics.

Conclusion: Need to move beyond accuracy as primary metric and adopt comprehensive fairness-aware evaluation frameworks and model-agnostic auditing strategies for responsible AI deployment in high-stakes environments.

Abstract: Facial recognition systems are increasingly deployed in law enforcement and security contexts, where algorithmic decisions can carry significant societal consequences. Despite high reported accuracy, growing evidence demonstrates that such systems often exhibit uneven performance across demographic groups, leading to disproportionate error rates and potential harm. This paper argues that aggregate accuracy is an insufficient metric for evaluating the fairness and reliability of facial recognition systems in high-stakes environments. Through analysis of subgroup-level error distribution, including false positive rate (FPR) and false negative rate (FNR), the paper demonstrates how aggregate performance metrics can obscure critical disparities across demographic groups. Empirical observations show that systems with similar overall accuracy can exhibit substantially different fairness profiles, with subgroup error rates varying significantly despite a single aggregate metric. The paper further examines the operational risks associated with accuracy-centric evaluation practices in law enforcement applications, where misclassification may result in wrongful suspicion or missed identification. It highlights the importance of fairness-aware evaluation approaches and model-agnostic auditing strategies that enable post-deployment assessment of real-world systems. The findings emphasise the need to move beyond accuracy as a primary metric and adopt more comprehensive evaluation frameworks for responsible AI deployment.

Method: Decouples OCR into two modules: 1) lightweight visual character detection (domain-agnostic), and 2) domain-specific linguistic correction using pretrained sequence models (T5, ByT5, BART). The correctors are trained entirely on synthetic noise, enabling annotation-free domain adaptation without labeled target images.

Result: The framework achieves near-SOTA accuracy across modern clean handwriting, cursive script, and historical documents while reducing compute by approximately 95% compared to end-to-end transformers. T5-Base excels on modern text with standard vocabulary, while ByT5-Base dominates on historical documents by reconstructing archaic spellings at the byte level.

Conclusion: The decoupled paradigm matches end-to-end transformer accuracy while dramatically reducing computational requirements, establishing a viable, resource-efficient alternative to monolithic OCR architectures that democratizes access to high-quality OCR.

Abstract: Optical character recognition remains critical infrastructure for document digitization, yet state-of-the-art performance is often restricted to well-resourced institutions by prohibitive computational barriers. End-to-end transformer architectures achieve strong accuracy but demand hundreds of GPU hours for domain adaptation, limiting accessibility for practitioners and digital humanities scholars. We present a modular detection-and-correction framework that achieves near-SOTA accuracy with single-GPU training. Our approach decouples lightweight visual character detection (domain-agnostic) from domain-specific linguistic correction using pretrained sequence models including T5, ByT5, and BART. By training the correctors entirely on synthetic noise, we enable annotation-free domain adaptation without requiring labeled target images. Evaluating across modern clean handwriting, cursive script, and historical documents, we identify a critical “Pareto frontier” in architecture selection: T5-Base excels on modern text with standard vocabulary, whereas ByT5-Base dominates on historical documents by reconstructing archaic spellings at the byte level. Our results demonstrate that this decoupled paradigm matches end-to-end transformer accuracy while reducing compute by approximately 95%, establishing a viable, resource-efficient alternative to monolithic OCR architectures.

Method: Splits video into groups, extracts cross-modal attention to rank tokens within each group, uses model’s response entropy to estimate group relevance, enabling global token budget allocation and early stopping (AdaptToken-Lite).

Result: Consistent accuracy improvements across four long-video benchmarks and multiple base MLLMs (7B-72B), with +6.7 average improvement over Qwen2.5-VL 7B, effective up to 10K frames, and ~50% inference time reduction with AdaptToken-Lite.

Conclusion: AdaptToken provides an effective training-free solution for long video understanding by leveraging model uncertainty for global token selection, enabling efficient processing of extremely long videos while maintaining accuracy.

Abstract: Long video understanding remains challenging for Multi-modal Large Language Models (MLLMs) due to high memory costs and context-length limits. Prior approaches mitigate this by scoring and selecting frames/tokens within short clips, but they lack a principled mechanism to (i) compare relevance across distant video clips and (ii) stop processing once sufficient evidence has been gathered. We propose AdaptToken, a training-free framework that turns an MLLM’s self-uncertainty into a global control signal for long-video token selection. AdaptToken splits a video into groups, extracts cross-modal attention to rank tokens within each group, and uses the model’s response entropy to estimate each group’s prompt relevance. This entropy signal enables a global token budget allocation across groups and further supports early stopping (AdaptToken-Lite), skipping the remaining groups when the model becomes sufficiently certain. Across four long-video benchmarks (VideoMME, LongVideoBench, LVBench, and MLVU) and multiple base MLLMs (7B-72B), AdaptToken consistently improves accuracy (e.g., +6.7 on average over Qwen2.5-VL 7B) and continues to benefit from extremely long inputs (up to 10K frames), while AdaptToken-Lite reduces inference time by about half with comparable performance. Project page: https://haozheqi.github.io/adapt-token

Method: Introduces three effort-based metrics: False Speed Reduction (FSR) for cumulative velocity loss from phantom detections, Maximum Deceleration Rate (MDR) for peak braking demand from missed objects, and Lateral Evasion Acceleration (LEA) for minimum steering effort to avoid collisions. Uses reachability-based ellipsoidal collision filtering and frame-level matching with track-level aggregation.

Result: Evaluation on nuScenes and Argoverse 2 datasets shows 65-93% of perception errors are non-critical. Spearman correlation analysis confirms the proposed metrics capture safety-relevant information not accessible to traditional time-based or deceleration-based measures, enabling targeted mining of critical perception failures.

Conclusion: The proposed effort-based metrics provide more nuanced evaluation of perception system safety by quantifying actual evasion efforts required, distinguishing critical from non-critical errors, and enabling focused improvement on safety-critical perception failures.

Abstract: Criticality metrics such as time-to-collision (TTC) quantify collision urgency but conflate the consequences of false-positive (FP) and false-negative (FN) perception errors. We propose two novel effort-based metrics: False Speed Reduction (FSR), the cumulative velocity loss from persistent phantom detections, and Maximum Deceleration Rate (MDR), the peak braking demand from missed objects under a constant-acceleration model. These longitudinal metrics are complemented by Lateral Evasion Acceleration (LEA), adapted from prior lateral evasion kinematics and coupled with reachability-based collision timing to quantify the minimum steering effort to avoid a predicted collision. A reachability-based ellipsoidal collision filter ensures only dynamically plausible threats are scored, with frame-level matching and track-level aggregation. Evaluation of different perception pipelines on nuScenes and Argoverse~2 shows that 65-93% of errors are non-critical, and Spearman correlation analysis confirms that all three metrics capture safety-relevant information inaccessible to established time-based, deceleration-based, or normalized criticality measures, enabling targeted mining of the most critical perception failures.

Method: EventTrack6D uses a dual reconstruction approach conditioned on the most recent depth measurement to recover dense photometric and geometric cues from sparse event streams. It reconstructs both intensity and depth at arbitrary timestamps between depth frames, enabling 6D pose tracking at over 120 FPS.

Result: The method achieves accurate 6D pose tracking across diverse objects and motion patterns, maintaining temporal consistency under rapid motion. Trained exclusively on synthetic data, it generalizes effectively to real-world scenarios without fine-tuning. A comprehensive benchmark suite including synthetic and real datasets was created for evaluation.

Conclusion: EventTrack6D demonstrates the effectiveness of event cameras for event-based 6D pose tracking of novel objects, providing a high-speed solution that overcomes limitations of conventional RGB-depth pipelines in fast, dynamic scenes.

Abstract: Event cameras provide microsecond latency, making them suitable for 6D object pose tracking in fast, dynamic scenes where conventional RGB and depth pipelines suffer from motion blur and large pixel displacements. We introduce EventTrack6D, an event-depth tracking framework that generalizes to novel objects without object-specific training by reconstructing both intensity and depth at arbitrary timestamps between depth frames. Conditioned on the most recent depth measurement, our dual reconstruction recovers dense photometric and geometric cues from sparse event streams. Our EventTrack6D operates at over 120 FPS and maintains temporal consistency under rapid motion. To support training and evaluation, we introduce a comprehensive benchmark suite: a large-scale synthetic dataset for training and two complementary evaluation sets, including real and simulated event datasets. Trained exclusively on synthetic data, EventTrack6D generalizes effectively to real-world scenarios without fine-tuning, maintaining accurate tracking across diverse objects and motion patterns. Our method and datasets validate the effectiveness of event cameras for event-based 6D pose tracking of novel objects. Code and datasets are publicly available at https://chohoonhee.github.io/Event6D.

Method: Uses prediction entropy from continuous-space AR models as a unifying signal. For AR acceleration: entropy-informed speculative decoding aligns draft-target entropy distributions. For visual decoder acceleration: reinterprets entropy as physical variance for an anti-symmetric drifting field, enabling single-step (1-NFE) decoding without iterative denoising.

Result: Achieves 3.8-5.5× speedup with genuine 1-NFE decoding while matching or surpassing original quality on MAR, TransDiff, and NextStep-1 benchmarks.

Conclusion: Prediction entropy serves as a natural unifying signal for joint acceleration of both AR and diffusion stages in hybrid models, enabling significant speed improvements without quality degradation.

Abstract: Autoregressive (AR)-Diffusion hybrid paradigms combine AR’s structured semantic modeling with diffusion’s high-fidelity synthesis, yet suffer from a dual speed bottleneck: the sequential AR stage and the iterative multi-step denoising of the diffusion vision decode stage. Existing methods address each in isolation without a unified principle design. We observe that the per-position \emph{prediction entropy} of continuous-space AR models naturally encodes spatially varying generation uncertainty, which simultaneously governing draft prediction quality in the AR stage and reflecting the corrective effort required by vision decoding stage, which is not fully explored before. Since entropy is inherently tied to both bottlenecks, it serves as a natural unifying signal for joint acceleration. In this work, we propose \textbf{Drift-AR}, which leverages entropy signal to accelerate both stages: 1) for AR acceleration, we introduce Entropy-Informed Speculative Decoding that align draft–target entropy distributions via a causal-normalized entropy loss, resolving the entropy mismatch that causes excessive draft rejection; 2) for visual decoder acceleration, we reinterpret entropy as the \emph{physical variance} of the initial state for an anti-symmetric drifting field – high-entropy positions activate stronger drift toward the data manifold while low-entropy positions yield vanishing drift – enabling single-step (1-NFE) decoding without iterative denoising or distillation. Moreover, both stages share the same entropy signal, which is computed once with no extra cost. Experiments on MAR, TransDiff, and NextStep-1 demonstrate 3.8–5.5$\times$ speedup with genuine 1-NFE decoding, matching or surpassing original quality. Code will be available at https://github.com/aSleepyTree/Drift-AR.

Method: Uses Distributed CNN where output vector modules for positive classes are monotonic with feature presence probabilities. Detects objects by identifying high-scoring patches across all scales and overlapping them to form bounding boxes.

Result: Enables parallel detection for multiple classes and faster single-object detection due to lightweight architecture. Training requires only object-centered image data with positive/negative labels.

Conclusion: Proposes a straightforward, efficient object detection approach that leverages scale-invariant feature detection through Distributed CNN, offering computational advantages over traditional methods.

Abstract: Based on the Distributed Convolutional Neural Network(DisCNN), a straightforward object detection method is proposed. The modules of the output vector of a DisCNN with respect to a specific positive class are positively monotonic with the presence probabilities of the positive features. So, by identifying all high-scoring patches across all possible scales, the positive object can be detected by overlapping them to form a bounding box. The essential idea is that the object is detected by detecting its features on multiple scales, ranging from specific sub-features to abstract features composed of these sub-features. Training DisCNN requires only object-centered image data with positive and negative class labels. The detection process for multiple positive classes can be conducted in parallel to significantly accelerate it, and also faster for single-object detection because of its lightweight model architecture.

Method: Proposes applying repulsion in the Contextual Space by intervening in multimodal attention channels during transformer forward passes. The intervention occurs between blocks where text conditioning is enriched with emergent image structure, redirecting guidance after structural formation but before composition fixation.

Result: Produces significantly richer diversity without sacrificing visual fidelity or semantic adherence. Method is computationally efficient with small overhead and remains effective in modern “Turbo” and distilled models where traditional trajectory-based interventions fail.

Conclusion: Repulsion in Contextual Space provides an effective framework for enhancing diversity in diffusion transformers while maintaining efficiency and compatibility with modern model architectures.

Abstract: Modern Text-to-Image (T2I) diffusion models have achieved remarkable semantic alignment, yet they often suffer from a significant lack of variety, converging on a narrow set of visual solutions for any given prompt. This typicality bias presents a challenge for creative applications that require a wide range of generative outcomes. We identify a fundamental trade-off in current approaches to diversity: modifying model inputs requires costly optimization to incorporate feedback from the generative path. In contrast, acting on spatially-committed intermediate latents tends to disrupt the forming visual structure, leading to artifacts. In this work, we propose to apply repulsion in the Contextual Space as a novel framework for achieving rich diversity in Diffusion Transformers. By intervening in the multimodal attention channels, we apply on-the-fly repulsion during the transformer’s forward pass, injecting the intervention between blocks where text conditioning is enriched with emergent image structure. This allows for redirecting the guidance trajectory after it is structurally informed but before the composition is fixed. Our results demonstrate that repulsion in the Contextual Space produces significantly richer diversity without sacrificing visual fidelity or semantic adherence. Furthermore, our method is uniquely efficient, imposing a small computational overhead while remaining effective even in modern “Turbo” and distilled models where traditional trajectory-based interventions typically fail.

Method: Proposes 4DSurf with two key innovations: 1) Gaussian deformations induced Signed Distance Function Flow Regularization to constrain Gaussian motion with evolving surfaces, and 2) Overlapping Segment Partitioning that divides sequences into overlapping segments with small deformations and incrementally passes geometric information.

Result: Outperforms state-of-the-art surface reconstruction methods by 49% on Hi4D dataset and 19% on CMU Panoptic dataset in Chamfer distance, achieving superior temporal consistency under sparse-view settings.

Conclusion: 4DSurf provides a unified framework for generic dynamic surface reconstruction that can handle large deformations and maintain temporal consistency without requiring prior knowledge of scene objects.

Abstract: This paper addresses the problem of dynamic scene surface reconstruction using Gaussian Splatting (GS), aiming to recover temporally consistent geometry. While existing GS-based dynamic surface reconstruction methods can yield superior reconstruction, they are typically limited to either a single object or objects with only small deformations, struggling to maintain temporally consistent surface reconstruction of large deformations over time. We propose ``\textit{4DSurf}’’, a novel and unified framework for generic dynamic surface reconstruction that does not require specifying the number or types of objects in the scene, can handle large surface deformations and temporal inconsistency in reconstruction. The key innovation of our framework is the introduction of Gaussian deformations induced Signed Distance Function Flow Regularization that constrains the motion of Gaussians to align with the evolving surface. To handle large deformations, we introduce an Overlapping Segment Partitioning strategy that divides the sequence into overlapping segments with small deformations and incrementally passes geometric information across segments through the shared overlapping timestep. Experiments on two challenging dynamic scene datasets, Hi4D and CMU Panoptic, demonstrate that our method outperforms state-of-the-art surface reconstruction methods by 49% and 19% in Chamfer distance, respectively, and achieves superior temporal consistency under sparse-view settings.

Method: Proposes AIBench benchmark with VQA-based evaluation for logic correctness and VLM-based assessment for aesthetics. Uses four levels of questions derived from logic diagrams summarizing paper methods to query illustration-paper alignment at different scales.

Result: Performance gaps between models on academic illustration generation are significantly larger than on general tasks, reflecting varying complex reasoning and high-density generation abilities. Logic and aesthetics are hard to optimize simultaneously. Test-time scaling on both abilities significantly boosts performance.

Conclusion: AIBench provides a more accurate evaluation framework for academic illustration generation, revealing model limitations in complex reasoning and high-density content generation while showing that test-time scaling improves performance on this specialized task.

Abstract: Although image generation has boosted various applications via its rapid evolution, whether the state-of-the-art models are able to produce ready-to-use academic illustrations for papers is still largely unexplored.Directly comparing or evaluating the illustration with VLM is native but requires oracle multi-modal understanding ability, which is unreliable for long and complex texts and illustrations. To address this, we propose AIBench, the first benchmark using VQA for evaluating logic correctness of the academic illustrations and VLMs for assessing aesthetics. In detail, we designed four levels of questions proposed from a logic diagram summarized from the method part of the paper, which query whether the generated illustration aligns with the paper on different scales. Our VQA-based approach raises more accurate and detailed evaluations on visual-logical consistency while relying less on the ability of the judger VLM. With our high-quality AIBench, we conduct extensive experiments and conclude that the performance gap between models on this task is significantly larger than general ones, reflecting their various complex reasoning and high-density generation ability. Further, the logic and aesthetics are hard to optimize simultaneously as in handcrafted illustrations. Additional experiments further state that test-time scaling on both abilities significantly boosts the performance on this task.

Method: Analyze cross-attention as spatiotemporal signals, track Fourier power over denoising steps, then introduce Attention Frequency Modulation (AFM) that edits pre-softmax logits in Fourier domain with frequency-specific reweighting schedules.

Result: AFM reliably redistributes attention spectra and produces substantial visual edits while preserving semantic alignment in Stable Diffusion, with entropy acting as adaptive gain rather than independent control.

Conclusion: Cross-attention exhibits consistent coarse-to-fine spectral progression, enabling principled frequency-based control through AFM without retraining or prompt editing.

Abstract: Cross-attention is the primary interface through which text conditions latent diffusion models, yet its step-wise multi-resolution dynamics remain under-characterized, limiting principled training-free control. We cast diffusion cross-attention as a spatiotemporal signal on the latent grid by summarizing token-softmax weights into token-agnostic concentration maps and tracking their radially binned Fourier power over denoising. Across prompts and seeds, encoder cross-attention exhibits a consistent coarse-to-fine spectral progression, yielding a stable time-frequency fingerprint of token competition. Building on this structure, we introduce Attention Frequency Modulation (AFM), a plug-and-play inference-time intervention that edits token-wise pre-softmax cross-attention logits in the Fourier domain: low- and high-frequency bands are reweighted with a progress-aligned schedule and can be adaptively gated by token-allocation entropy, before the token softmax. AFM provides a continuous handle to bias the spatial scale of token-competition patterns without retraining, prompt editing, or parameter updates. Experiments on Stable Diffusion show that AFM reliably redistributes attention spectra and produces substantial visual edits while largely preserving semantic alignment. Finally, we find that entropy mainly acts as an adaptive gain on the same frequency-based edit rather than an independent control axis.

Method: Three core components: 1) Agent Loop - structured multi-agent framework for iterative closed-loop optimization; 2) Agent Memory - persistent, hierarchical memory storing factual states and experiential summaries; 3) Agent Skill - extensible collection of domain-specific expertise with on-demand loading.

Result: GEMS achieves significant performance gains across five mainstream tasks and four downstream tasks on multiple generative backends. Notably enables lightweight 6B model Z-Image-Turbo to surpass state-of-the-art Nano Banana 2 on GenEval2.

Conclusion: The agent harness effectively extends model capabilities beyond their original limits, demonstrating the power of agent-native frameworks for multimodal generation.

Abstract: Recent multimodal generation models have achieved remarkable progress on general-purpose generation tasks, yet continue to struggle with complex instructions and specialized downstream tasks. Inspired by the success of advanced agent frameworks such as Claude Code, we propose \textbf{GEMS} (Agent-Native Multimodal \textbf{GE}neration with \textbf{M}emory and \textbf{S}kills), a framework that pushes beyond the inherent limitations of foundational models on both general and downstream tasks. GEMS is built upon three core components. Agent Loop introduces a structured multi-agent framework that iteratively improves generation quality through closed-loop optimization. Agent Memory provides a persistent, trajectory-level memory that hierarchically stores both factual states and compressed experiential summaries, enabling a global view of the optimization process while reducing redundancy. Agent Skill offers an extensible collection of domain-specific expertise with on-demand loading, allowing the system to effectively handle diverse downstream applications. Across five mainstream tasks and four downstream tasks, evaluated on multiple generative backends, GEMS consistently achieves significant performance gains. Most notably, it enables the lightweight 6B model Z-Image-Turbo to surpass the state-of-the-art Nano Banana 2 on GenEval2, demonstrating the effectiveness of agent harness in extending model capabilities beyond their original limits.

Method: Used a single dataset with annotated road damages and material labels. Applied SONAR sensor data for two tasks: 1) Classifying road materials (asphalt, concrete, element roads), 2) Detecting and classifying damage types without localization.

Result: Material classification achieved ~90% F1 score on test set. Damage detection and classification performed worse at ~75% F1 score.

Conclusion: SONAR sensing is promising for opportunistic pavement management systems but requires further research to improve damage detection accuracy.

Abstract: In this paper, we investigate the capabilities of in-air 3D SONAR sensors for the monitoring of road surface conditions. Concretely, we consider two applications: Road material classification and Road damage detection and classification. While such tasks can be performed with other sensor modalities, such as camera sensors and LiDAR sensors, these sensor modalities tend to fail in harsh sensing conditions, such as heavy rain, smoke or fog. By using a sensing modality that is robust to such interference, we enable the creation of opportunistic sensing applications, where vehicles performing other tasks (garbage collection, mail delivery, etc.) can also be used to monitor the condition of the road. For these tasks, we use a single dataset, in which different types of damages are annotated, with labels including the material of the road surface. In the material classification task, we differentiate between three different road materials: Asphalt, Concrete and Element roads. In the damage detection and classification task, we determine if there is damage, and what type of damage (independent of material type), without localizing the damage. We are succesful in determining the road surface type from SONAR sensor data, with F1 scores approaching 90% on the test set, but find that for the detection of damages performace lags, with F1 score around 75%. From this, we conclude that SONAR sensing is a promising modality to include in opportunistic sensing-based pavement management systems, but that further research is needed to reach the desired accuracy.

Method: Proposes NeRP3D, a NeRF-resembled point-based 3D detector that preserves pre-trained NeRF networks regardless of tasks. It learns continuous 3D representations to avoid misaligned priors from view transformation, enabling better scene reconstruction and detection.

Result: Experiments on nuScenes dataset show significant improvements over previous state-of-the-art methods, outperforming both pretext scene reconstruction tasks and downstream detection tasks.

Conclusion: NeRP3D successfully addresses the conflicting priors in NeRF-based pre-training for 3D perception by maintaining continuous 3D representation learning through preserved NeRF networks, leading to superior performance in both reconstruction and detection tasks.

Abstract: Neural radiance fields (NeRFs) have emerged as a prominent pre-training paradigm for vision-centric autonomous driving, which enhances 3D geometry and appearance understanding in a fully self-supervised manner. To apply NeRF-based pretraining to 3D perception models, recent approaches have simply applied NeRFs to volumetric features obtained from view transformation. However, coupling NeRFs with view transformation inherits conflicting priors; view transformation imposes discrete and rigid representations, whereas radiance fields assume continuous and adaptive functions. When these opposing assumptions are forced into a single pipeline, the misalignment surfaces as blurry and ambiguous 3D representations that ultimately limit 3D scene understanding. Moreover, the NeRF network for pre-training is discarded during downstream tasks, resulting in inefficient utilization of enhanced 3D representations through NeRF. In this paper, we propose a novel NeRF-Resembled Point-based 3D detector that can learn continuous 3D representation and thus avoid the misaligned priors from view transformation. NeRP3D preserves the pre-trained NeRF network regardless of the tasks, inheriting the principle of continuous 3D representation learning and leading to greater potentials for both scene reconstruction and detection tasks. Experiments on nuScenes dataset demonstrate that our proposed approach significantly improves previous state-of-the-art methods, outperforming not only pretext scene reconstruction tasks but also downstream detection tasks.

Method: Incremental processing of observation windows with instance-aware context streaming to maintain/update latent agent representations across inference steps, plus dual training objective for consistent accuracy across observation horizons.

Result: Achieves state-of-the-art performance on Argoverse 2 multi-agent benchmark in streaming inference, maintains minimal latency, robust across Argoverse 2, nuScenes, and Argoverse 1 datasets.

Conclusion: Proposed framework effectively handles evolving scenes with heterogeneous observations, suitable for real-world deployment due to strong performance and low latency.

Abstract: In dynamic traffic environments, motion forecasting models must be able to accurately estimate future trajectories continuously. Streaming-based methods are a promising solution, but despite recent advances, their performance often degrades when exposed to heterogeneous observation lengths. To address this, we propose a novel streaming-based motion forecasting framework that explicitly focuses on evolving scenes. Our method incrementally processes incoming observation windows and leverages an instance-aware context streaming to maintain and update latent agent representations across inference steps. A dual training objective further enables consistent forecasting accuracy across diverse observation horizons. Extensive experiments on Argoverse 2, nuScenes, and Argoverse 1 demonstrate the robustness of our approach under evolving scene conditions and also on the single-agent benchmarks. Our model achieves state-of-the-art performance in streaming inference on the Argoverse 2 multi-agent benchmark, while maintaining minimal latency, highlighting its suitability for real-world deployment.

Method: 1) For object point clouds: proposed leaf nodes lossy compression method using bit-wise coding and binary prediction on leaf nodes. 2) For LiDAR point clouds: proposed variable rate approaches with a simple but effective rate control method.

Result: The leaf nodes lossy compression significantly outperforms previous octree-based methods on object point clouds. The rate control method achieves about 1% bit error without finetuning on LiDAR point clouds.

Conclusion: The paper demonstrates that specialized lossy compression approaches tailored to different point cloud characteristics (object vs LiDAR) can overcome limitations of traditional methods and achieve superior performance.

Abstract: Octree-based context learning has recently become a leading method in point cloud compression. However, its potential on lossy compression remains undiscovered. The traditional lossy compression paradigm using lossless octree representation with quantization step adjustment may result in severe distortions due to massive missing points in quantization. Therefore, we analyze data characteristics of different point clouds and propose lossy approaches specifically. For object point clouds that suffer from quantization step adjustment, we propose a new leaf nodes lossy compression method, which achieves lossy compression by performing bit-wise coding and binary prediction on leaf nodes. For LiDAR point clouds, we explore variable rate approaches and propose a simple but effective rate control method. Experimental results demonstrate that the proposed leaf nodes lossy compression method significantly outperforms the previous octree-based method on object point clouds, and the proposed rate control method achieves about 1% bit error without finetuning on LiDAR point clouds.

Method: Convert single-channel Bayer data to four-channel RGGB format, partition into patches, compute each patch’s bit depth as auxiliary input, and design a bit-depth-adaptive entropy model to estimate patch distributions conditioned on their bit depths. This enables a single model to handle diverse cameras and bit depths.

Result: RAWIC consistently surpasses traditional lossless codecs, achieving an average 7.7% bitrate reduction over JPEG-XL.

Conclusion: RAWIC provides an effective learned lossless compression solution for raw images that handles varying bit depths and camera characteristics with a single model.

Abstract: Raw images preserve linear sensor measurements and high bit-depth information crucial for advanced vision tasks and photography applications, yet their storage remains challenging due to large file sizes, varying bit depths, and sensor-dependent characteristics. Existing learned lossless compression methods mainly target 8-bit sRGB images, while raw reconstruction approaches are inherently lossy and rely on camera-specific assumptions. To address these challenges, we introduce RAWIC, a bit-depth-adaptive learned lossless compression framework for Bayer-pattern raw images. We first convert single-channel Bayer data into a four-channel RGGB format and partition it into patches. For each patch, we compute its bit depth and use it as auxiliary input to guide compression. A bit-depth-adaptive entropy model is then designed to estimate patch distributions conditioned on their bit depths. This architecture enables a single model to handle raw images from diverse cameras and bit depths. Experiments show that RAWIC consistently surpasses traditional lossless codecs, achieving an average 7.7% bitrate reduction over JPEG-XL. Our code is available at https://github.com/chunbaobao/RAWIC.

Method: Proposes PEMV-thyroid, a Prototype-Enhanced Multi-View learning framework that learns complementary representations from multiple feature perspectives and refines decision boundaries through a prototype-based correction mechanism with mixed prototype information to account for data heterogeneity.

Result: Extensive experiments on multiple thyroid ultrasound datasets show PEMV-thyroid consistently outperforms state-of-the-art methods, particularly in cross-device and cross-domain evaluation scenarios, leading to improved diagnostic accuracy and generalization in real-world clinical settings.

Conclusion: The proposed framework effectively addresses domain heterogeneity in medical imaging by integrating multi-view representations with prototype-level guidance, enabling more stable representation learning under heterogeneous imaging conditions for improved clinical deployment.

Abstract: Thyroid nodule classification using ultrasound imaging is essential for early diagnosis and clinical decision-making; however, despite promising performance on in-distribution data, existing deep learning methods often exhibit limited robustness and generalisation when deployed across different ultrasound devices or clinical environments. This limitation is mainly attributed to the pronounced heterogeneity of thyroid ultrasound images, which can lead models to capture spurious correlations rather than reliable diagnostic cues. To address this challenge, we propose PEMV-thyroid, a Prototype-Enhanced Multi-View learning framework that accounts for data heterogeneity by learning complementary representations from multiple feature perspectives and refining decision boundaries through a prototype-based correction mechanism with mixed prototype information. By integrating multi-view representations with prototype-level guidance, the proposed approach enables more stable representation learning under heterogeneous imaging conditions. Extensive experiments on multiple thyroid ultrasound datasets demonstrate that PEMV-thyroid consistently outperforms state-of-the-art methods, particularly in cross-device and cross-domain evaluation scenarios, leading to improved diagnostic accuracy and generalisation performance in real-world clinical settings. The source code is available at https://github.com/chenyangmeii/Prototype-Enhanced-Multi-View-Learning.

Method: Proposes Contour-Guided Query Refinement Network (CGQR-Net) with HRNet backbone for multi-resolution features. Generates coarse segmentation, extracts anatomical contours, encodes them into learnable query embeddings. Uses cross-attention between contour-guided queries and fused feature maps for structure-aware refinement. Employs dual-head supervision for joint segmentation and boundary prediction optimization.

Result: Evaluated on CAMUS dataset and validated on CardiacNet for cross-dataset generalization. Demonstrates improved segmentation accuracy, enhanced boundary precision, and robust performance across varying imaging conditions.

Conclusion: Integrating contour-level structural information with feature-level representations is effective for reliable cardiac ultrasound segmentation, addressing challenges of low contrast, noise, and domain shifts.

Abstract: Accurate cardiac ultrasound segmentation is essential for reliable assessment of ventricular function in intelligent healthcare systems. However, echocardiographic images are challenging due to low contrast, speckle noise, irregular boundaries, and domain shifts across devices and patient populations. Existing methods, largely based on appearance-driven learning, often fail to preserve boundary precision and structural consistency under these conditions. To address these issues, we propose a Contour-Guided Query Refinement Network (CGQR-Net) for boundary-aware cardiac ultrasound segmentation. The framework integrates multi-resolution feature representations with contour-derived structural priors. An HRNet backbone preserves high-resolution spatial details while capturing multi-scale context. A coarse segmentation is first generated, from which anatomical contours are extracted and encoded into learnable query embeddings. These contour-guided queries interact with fused feature maps via cross-attention, enabling structure-aware refinement that improves boundary delineation and reduces noise artifacts. A dual-head supervision strategy jointly optimizes segmentation and boundary prediction to enforce structural consistency. The proposed method is evaluated on the CAMUS dataset and further validated on the CardiacNet dataset to assess cross-dataset generalization. Experimental results demonstrate improved segmentation accuracy, enhanced boundary precision, and robust performance across varying imaging conditions. These results highlight the effectiveness of integrating contour-level structural information with feature-level representations for reliable cardiac ultrasound segmentation.

Method: Proposes MedLoc-R1 with sliding-window performance tracker and multi-condition update rule that automatically adjusts reward schedule from dense, easily obtainable signals to stricter localization requirements, preserving GRPO properties without auxiliary networks.

Result: Experiments on three medical visual grounding benchmarks show MedLoc-R1 consistently improves both localization accuracy and training stability over GRPO-based baselines.

Conclusion: MedLoc-R1 offers a general, lightweight, and effective solution for RL-based grounding in high-stakes medical applications.

Abstract: Medical visual grounding serves as a crucial foundation for fine-grained multimodal reasoning and interpretable clinical decision support. Despite recent advances in reinforcement learning (RL) for grounding tasks, existing approaches such as Group Relative Policy Optimization~(GRPO) suffer from severe reward sparsity when directly applied to medical images, primarily due to the inherent difficulty of localizing small or ambiguous regions of interest, which is further exacerbated by the rigid and suboptimal nature of fixed IoU-based reward schemes in RL. This leads to vanishing policy gradients and stagnated optimization, particularly during early training. To address this challenge, we propose MedLoc-R1, a performance-aware reward scheduling framework that progressively tightens the reward criterion in accordance with model readiness. MedLoc-R1 introduces a sliding-window performance tracker and a multi-condition update rule that automatically adjust the reward schedule from dense, easily obtainable signals to stricter, fine-grained localization requirements, while preserving the favorable properties of GRPO without introducing auxiliary networks or additional gradient paths. Experiments on three medical visual grounding benchmarks demonstrate that MedLoc-R1 consistently improves both localization accuracy and training stability over GRPO-based baselines. Our framework offers a general, lightweight, and effective solution for RL-based grounding in high-stakes medical applications. Code & checkpoints are available at \hyperlink{}{https://github.com/MembrAI/MedLoc-R1}.

Method: Proposes a weighted ensemble learning system integrating multiple classifiers: ResNet101, DenseNet121, Xception, CNN-MRI, ResNet50 with edge-enhanced images, SVM, and KNN with HOG features. Uses weighted voting mechanism giving higher influence to more accurate models. Applies image processing techniques including Balance Contrast Enhancement, K-means clustering, and Canny edge detection for feature enhancement.

Result: Experimental evaluations on Figshare and Kaggle MRI datasets show the proposed method achieves state-of-the-art accuracy, outperforming existing models.

Conclusion: The ensemble-based learning approach demonstrates potential for improving brain tumor classification, offering a reliable and scalable framework for medical image analysis.

Abstract: The accurate classification of brain tumors from MRI scans is essential for effective diagnosis and treatment planning. This paper presents a weighted ensemble learning approach that combines deep learning and traditional machine learning models to improve classification performance. The proposed system integrates multiple classifiers, including ResNet101, DenseNet121, Xception, CNN-MRI, and ResNet50 with edge-enhanced images, SVM, and KNN with HOG features. A weighted voting mechanism assigns higher influence to models with better individual accuracy, ensuring robust decision-making. Image processing techniques such as Balance Contrast Enhancement, K-means clustering, and Canny edge detection are applied to enhance feature extraction. Experimental evaluations on the Figshare and Kaggle MRI datasets demonstrate that the proposed method achieves state-of-the-art accuracy, outperforming existing models. These findings highlight the potential of ensemble-based learning for improving brain tumor classification, offering a reliable and scalable framework for medical image analysis.

Method: Proposes SVGS (Single-View to 3D Object Editing via Gaussian Splatting) using 3D Gaussian Splatting as the 3D representation. Introduces a single-view editing strategy based on multi-view diffusion models that reconstructs 3D scenes using only views with consistent editing results.

Result: SVGS outperforms existing baseline methods (including Instruct-NeRF2NeRF and GaussianEditor) across various scene settings in both editing capability and processing speed.

Conclusion: SVGS represents a significant advancement in 3D editing technology by addressing consistency and efficiency challenges in text-driven 3D scene editing through single-view strategy and efficient 3D Gaussian Splatting representation.

Abstract: Text-driven 3D scene editing has attracted considerable interest due to its convenience and user-friendliness. However, methods that rely on implicit 3D representations, such as Neural Radiance Fields (NeRF), while effective in rendering complex scenes, are hindered by slow processing speeds and limited control over specific regions of the scene. Moreover, existing approaches, including Instruct-NeRF2NeRF and GaussianEditor, which utilize multi-view editing strategies, frequently produce inconsistent results across different views when executing text instructions. This inconsistency can adversely affect the overall performance of the model, complicating the task of balancing the consistency of editing results with editing efficiency. To address these challenges, we propose a novel method termed Single-View to 3D Object Editing via Gaussian Splatting (SVGS), which is a single-view text-driven editing technique based on 3D Gaussian Splatting (3DGS). Specifically, in response to text instructions, we introduce a single-view editing strategy grounded in multi-view diffusion models, which reconstructs 3D scenes by leveraging only those views that yield consistent editing results. Additionally, we employ sparse 3D Gaussian Splatting as the 3D representation, which significantly enhances editing efficiency. We conducted a comparative analysis of SVGS against existing baseline methods across various scene settings, and the results indicate that SVGS outperforms its counterparts in both editing capability and processing speed, representing a significant advancement in 3D editing technology. For further details, please visit our project page at: https://amateurc.github.io/svgs.github.io.

Method: 1) Divides training pairs into clean, ambiguous, and noisy categories based on loss magnitude; 2) Estimates reliability via loss-based weighting; 3) Uses multi-modal self-paced function to regulate training sequence; 4) Implements robust triplet loss with dynamic soft margin adjustment.

Result: Extensive experiments on three benchmark datasets show RRSITR significantly outperforms state-of-the-art methods, especially at high noise rates.

Conclusion: The proposed RRSITR paradigm effectively addresses noisy correspondence in RSITR through self-paced learning and robust loss functions, demonstrating superior performance.

Abstract: As a pivotal task that bridges remote visual and linguistic understanding, Remote Sensing Image-Text Retrieval (RSITR) has attracted considerable research interest in recent years. However, almost all RSITR methods implicitly assume that image-text pairs are matched perfectly. In practice, acquiring a large set of well-aligned data pairs is often prohibitively expensive or even infeasible. In addition, we also notice that the remote sensing datasets (e.g., RSITMD) truly contain some inaccurate or mismatched image text descriptions. Based on the above observations, we reveal an important but untouched problem in RSITR, i.e., Noisy Correspondence (NC). To overcome these challenges, we propose a novel Robust Remote Sensing Image-Text Retrieval (RRSITR) paradigm that designs a self-paced learning strategy to mimic human cognitive learning patterns, thereby learning from easy to hard from multi-modal data with NC. Specifically, we first divide all training sample pairs into three categories based on the loss magnitude of each pair, i.e., clean sample pairs, ambiguous sample pairs, and noisy sample pairs. Then, we respectively estimate the reliability of each training pair by assigning a weight to each pair based on the values of the loss. Further, we respectively design a new multi-modal self-paced function to dynamically regulate the training sequence and weights of the samples, thus establishing a progressive learning process. Finally, for noisy sample pairs, we present a robust triplet loss to dynamically adjust the soft margin based on semantic similarity, thereby enhancing the robustness against noise. Extensive experiments on three popular benchmark datasets demonstrate that the proposed RRSITR significantly outperforms the state-of-the-art methods, especially in high noise rates. The code is available at: https://github.com/MSFLabX/RRSITR

Method: Proposes reconceptualizing distribution matching as a reward (R_dm), introducing Group Normalized Distribution Matching (GNDM) to stabilize R_dm estimation using group-mean statistics. The framework supports adaptive weighting for combining DMD with external rewards and incorporates importance sampling for efficiency.

Result: GNDM outperforms vanilla DMD by reducing FID by 1.87. The multi-reward variant GNDMR achieves peak HPS of 30.37 and low FID-SD of 12.21, balancing aesthetic quality and fidelity better than existing baselines.

Conclusion: R_dm provides a flexible, stable, and efficient framework for real-time high-fidelity synthesis by unifying diffusion matching distillation with reinforcement learning principles through a reward-centric formulation.

Abstract: Diffusion models achieve state-of-the-art generative performance but are fundamentally bottlenecked by their slow iterative sampling process. While diffusion distillation techniques enable high-fidelity few-step generation, traditional objectives often restrict the student’s performance by anchoring it solely to the teacher. Recent approaches have attempted to break this ceiling by integrating Reinforcement Learning (RL), typically through a simple summation of distillation and RL objectives. In this work, we propose a novel paradigm by reconceptualizing distribution matching as a reward, denoted as $R_{dm}$. This unified perspective bridges the algorithmic gap between Diffusion Matching Distillation (DMD) and RL, providing several key benefits. (1) Enhanced optimization stability: we introduce Group Normalized Distribution Matching (GNDM), which adapts standard RL group normalization to stabilize $R_{dm}$ estimation. By leveraging group-mean statistics, GNDM establishes a more robust and effective optimization direction. (2) Seamless reward integration: our reward-centric formulation inherently supports adaptive weighting mechanisms, allowing flexible combination of DMD with external reward models. (3) Improved sampling efficiency: by aligning with RL principles, the framework readily incorporates importance sampling (IS), leading to a significant boost in sampling efficiency. Extensive experiments demonstrate that GNDM outperforms vanilla DMD, reducing the FID by 1.87. Furthermore, our multi-reward variant, GNDMR, surpasses existing baselines by achieving a strong balance between aesthetic quality and fidelity, reaching a peak HPS of 30.37 and a low FID-SD of 12.21. Overall, $R_{dm}$ provides a flexible, stable, and efficient framework for real-time high-fidelity synthesis. Code will be released upon publication.

Method: Proposes BlankSkip, an adaptive object detection network that uses a simple auxiliary classification task for early exit. The system identifies frames with no objects of interest and skips full detection processing for those frames, reducing computational effort for “blank” frames.

Result: Achieves up to 24% average throughput improvement with only 0.015 mean Average Precision (mAP) drop compared to static MobileNet-SSD detector. Tested on real-world nano-drone platform (Bitcraze Crazyflie 2.1) using state-of-the-art nano-drones object detection dataset.

Conclusion: BlankSkip successfully applies early-exit adaptive networks to object detection for nano-drones, demonstrating significant throughput improvements with minimal accuracy degradation, enabling more efficient on-device object detection for resource-constrained platforms.

Abstract: Deploying tiny computer vision Deep Neural Networks (DNNs) on-board nano-sized drones is key for achieving autonomy, but is complicated by the extremely tight constraints of their computational platforms (approximately 10 MiB memory, 1 W power budget). Early-exit adaptive DNNs that dial down the computational effort for “easy-to-process” input frames represent a promising way to reduce the average inference latency. However, while this approach is extensively studied for classification, its application to dense tasks like object detection (OD) is not straightforward. In this paper, we propose BlankSkip, an adaptive network for on-device OD that leverages a simple auxiliary classification task for early exit, i.e., identifying frames with no objects of interest. With experiments using a real-world nano-drone platform, the Bitcraze Crazyflie 2.1, we achieve up to 24% average throughput improvement with a limited 0.015 mean Average Precision (mAP) drop compared to a static MobileNet-SSD detector, on a state-of-the-art nano-drones OD dataset.

Method: ObjectMorpher lifts target instances from 2D images to editable 3D Gaussian Splatting representations using an image-to-3D generator. Users interact via control points, and the system applies graph-based non-rigid deformation with ARAP constraints for physically sensible shape/pose changes, followed by composite diffusion for seamless reintegration.

Result: ObjectMorpher achieves fine-grained, photorealistic edits across diverse categories, outperforming both 2D drag and 3D-aware baselines on metrics including KID, LPIPS, SIFID, and user preference, with superior controllability and efficiency.

Conclusion: ObjectMorpher provides a unified, interactive framework for geometry-grounded image editing that bridges the gap between 2D and 3D approaches, enabling precise object-level control with physically sensible results.

Abstract: Achieving precise, object-level control in image editing remains challenging: 2D methods lack 3D awareness and often yield ambiguous or implausible results, while existing 3D-aware approaches rely on heavy optimization or incomplete monocular reconstructions. We present ObjectMorpher, a unified, interactive framework that converts ambiguous 2D edits into geometry-grounded operations. ObjectMorpher lifts target instances with an image-to-3D generator into editable 3D Gaussian Splatting (3DGS), enabling fast, identity-preserving manipulation. Users drag control points; a graph-based non-rigid deformation with as-rigid-as-possible (ARAP) constraints ensures physically sensible shape and pose changes. A composite diffusion module harmonizes lighting, color, and boundaries for seamless reintegration. Across diverse categories, ObjectMorpher delivers fine-grained, photorealistic edits with superior controllability and efficiency, outperforming 2D drag and 3D-aware baselines on KID, LPIPS, SIFID, and user preference.

Method: Multi-event camera array approach combining asynchronous event stream analysis with temporal correlation to extract marker centers, rapid calibration using Kruppa equations with parameter optimization, and unified coordinate transformation with linear intersection for 3D deformation measurement.

Result: Relative measurement error below 0.08%, successful field experiments under extreme illumination conditions including self-calibration of camera array and 3D deformation measurement.

Conclusion: The method overcomes traditional camera limitations for high-speed 3D deformation measurement under extreme illumination, achieving accurate measurements with less than 0.1% relative error under harsh lighting conditions.

Abstract: Background: Large engineering structures, such as space launch towers and suspension bridges, are subjected to extreme forces that cause high-speed 3D deformation and compromise safety. These structures typically operate under extreme illumination conditions. Traditional cameras often struggle to handle strong light intensity, leading to overexposure due to their limited dynamic range. Objective: Event cameras have emerged as a compelling alternative to traditional cameras in high dynamic range and low-latency applications. This paper presents an integrated method, from calibration to measurement, using a multi-event camera array for high-speed 3D deformation monitoring of structures in extreme illumination conditions. Methods: Firstly, the proposed method combines the characteristics of the asynchronous event stream and temporal correlation analysis to extract the corresponding marker center point. Subsequently, the method achieves rapid calibration by solving the Kruppa equations in conjunction with a parameter optimization framework. Finally, by employing a unified coordinate transformation and linear intersection, the method enables the measurement of 3D deformation of the target structure. Results: Experiments confirmed that the relative measurement error is below 0.08%. Field experiments under extreme illumination conditions, including self-calibration of a multi-event camera array and 3D deformation measurement, verified the performance of the proposed method. Conclusions: This paper addressed the critical limitation of traditional cameras in measuring high-speed 3D deformations under extreme illumination conditions. The experimental results demonstrate that, compared to other methods, the proposed method can accurately measure 3D deformations of structures under harsh lighting conditions, and the relative error of the measured deformation is less than 0.1%.

Method: Uses FLUX diffusion model with structure-color decoupling strategy, progressive Direct Preference Optimization (Pro-DPO) for color preference learning, and visual semantic prompts to extract fine-grained semantic information from old photos.

Result: Outperforms state-of-the-art colorization methods on both synthetic and real datasets, including closed-source commercial models, producing high-quality and vivid colorization.

Conclusion: The proposed framework effectively addresses the domain gap in old photo colorization through innovative strategies for structure preservation, color restoration, and semantic understanding.

Abstract: Old photos preserve invaluable historical memories, making their restoration and colorization highly desirable. While existing restoration models can address some degradation issues like denoising and scratch removal, they often struggle with accurate colorization. This limitation arises from the unique degradation inherent in old photos, such as faded brightness and altered color hues, which are different from modern photo distributions, creating a substantial domain gap during colorization. In this paper, we propose a novel old photo colorization framework based on the generative diffusion model FLUX. Our approach introduces a structure-color decoupling strategy that separates structure preservation from color restoration, enabling accurate colorization of old photos while maintaining structural consistency. We further enhance the model with a progressive Direct Preference Optimization (Pro-DPO) strategy, which allows the model to learn subtle color preferences through coarse-to-fine transitions in color augmentation. Additionally, we address the limitations of text-based prompts by introducing visual semantic prompts, which extract fine-grained semantic information directly from old photos, helping to eliminate the color bias inherent in old photos. Experimental results on both synthetic and real datasets demonstrate that our approach outperforms existing state-of-the-art colorization methods, including closed-source commercial models, producing high-quality and vivid colorization.

Method: Two-stage pipeline: Stage 1 uses vision-language models to detect privacy risks, generate paired private/public captions, and produce identity-neutral edit instructions via LLMs. Stage 2 employs instruction-driven diffusion editors to apply dual textual prompts, rewriting only sensitive regions while preserving global structure.

Result: Unsafe2Safe significantly reduces face similarity, text similarity, and demographic predictability across MS-COCO, Caltech101, and MIT Indoor67 datasets, while maintaining downstream model accuracy comparable to training on raw data. Fine-tuning diffusion editors on automatically generated triplets further improves privacy protection and semantic fidelity.

Conclusion: Unsafe2Safe provides a scalable, principled solution for constructing large, privacy-safe datasets without sacrificing visual consistency or downstream utility, addressing critical privacy concerns in multimodal AI training.

Abstract: Large-scale image datasets frequently contain identifiable or sensitive content, raising privacy risks when training models that may memorize and leak such information. We present Unsafe2Safe, a fully automated pipeline that detects privacy-prone images and rewrites only their sensitive regions using multimodally guided diffusion editing. Unsafe2Safe operates in two stages. Stage 1 uses a vision-language model to (i) inspect images for privacy risks, (ii) generate paired private and public captions that respectively include and omit sensitive attributes, and (iii) prompt a large language model to produce structured, identity-neutral edit instructions conditioned on the public caption. Stage 2 employs instruction-driven diffusion editors to apply these dual textual prompts, producing privacy-safe images that preserve global structure and task-relevant semantics while neutralizing private content. To measure anonymization quality, we introduce a unified evaluation suite covering Quality, Cheating, Privacy, and Utility dimensions. Across MS-COCO, Caltech101, and MIT Indoor67, Unsafe2Safe reduces face similarity, text similarity, and demographic predictability by large margins, while maintaining downstream model accuracy comparable to training on raw data. Fine-tuning diffusion editors on our automatically generated triplets (private caption, public caption, edit instruction) further improves both privacy protection and semantic fidelity. Unsafe2Safe provides a scalable, principled solution for constructing large, privacy-safe datasets without sacrificing visual consistency or downstream utility.

Method: Proposes ToLL framework with Anchor-Conditioned Topological Geometry Reasoning using GNNs to recover global layouts from sparse anchors, strictly modulated by predicate features. Includes Structural Multi-view Augmentation to avoid semantic corruption and self-distillation for enhanced representations.

Result: Extensive experiments on 3DSSG dataset demonstrate improved representation quality, outperforming state-of-the-art baselines.

Conclusion: ToLL provides an effective self-supervised pretraining framework for 3D Scene Graph generation that enforces predicate relation learning and improves generalizability despite data scarcity.

Abstract: 3D Scene Graph (3DSG) generation plays a pivotal role in spatial understanding and semantic-affordance perception. However, its generalizability is often constrained by data scarcity. Current solutions primarily focus on cross-modal assisted representation learning and object-centric generation pre-training. The former relies heavily on predicate annotations, while the latter’s predicate learning may be bypassed due to strong object priors. Consequently, they could not often provide a label-free and robust self-supervised proxy task for 3DSG fine-tuning. To bridge this gap, we propose a Topological Layout Learning (ToLL) for 3DSG pretraining framework. In detail, we design an Anchor-Conditioned Topological Geometry Reasoning, with a GNN to recover the global layout of zero-centered subgraphs by the spatial priors from sparse anchors. This process is strictly modulated by predicate features, thereby enforcing the predicate relation learning. Furthermore, we construct a Structural Multi-view Augmentation to avoid semantic corruption, and enhancing representations via self-distillation. The extensive experiments on 3DSSG dataset demonstrate that our ToLL could improve representation quality, outperforming state-of-the-art baselines.

Method: Proposes hybrid ensemble decoder with shared hierarchical layer and parallel decoder branches using denoising queries for diversity, plus progressive fine-tuning with plateau-aware learning rate schedule

Result: Achieves 41.9 average performance on RF100-VL (100 diverse datasets) in 10-shot setting, outperforming SAM3 (35.7), and shows robustness on OOD samples

Conclusion: The method effectively addresses FSOD challenges through ensemble learning and stable optimization, demonstrating strong generalization and robustness across diverse domains

Abstract: Few-shot object detection (FSOD) is challenging due to unstable optimization and limited generalization arising from the scarcity of training samples. To address these issues, we propose a hybrid ensemble decoder that enhances generalization during fine-tuning. Inspired by ensemble learning, the decoder comprises a shared hierarchical layer followed by multiple parallel decoder branches, where each branch employs denoising queries either inherited from the shared layer or newly initialized to encourage prediction diversity. This design fully exploits pretrained weights without introducing additional parameters, and the resulting diverse predictions can be effectively ensembled to improve generalization. We further leverage a unified progressive fine-tuning framework with a plateau-aware learning rate schedule, which stabilizes optimization and achieves strong few-shot adaptation without complex data augmentations or extensive hyperparameter tuning. Extensive experiments on CD-FSOD, ODinW-13, and RF100-VL validate the effectiveness of our approach. Notably, on RF100-VL, which includes 100 datasets across diverse domains, our method achieves an average performance of 41.9 in the 10-shot setting, significantly outperforming the recent approach SAM3, which obtains 35.7. We further construct a mixed-domain test set from CD-FSOD to evaluate robustness to out-of-distribution (OOD) samples, showing that our proposed modules lead to clear improvement gains. These results highlight the effectiveness, generalization, and robustness of the proposed method. Code is available at: https://github.com/Intellindust-AI-Lab/FT-FSOD.

Method: Projects CLIP’s joint image-text embeddings into a concept space learned from language descriptions using alignment and reconstruction objectives. This preserves CLIP’s semantic structure while making concept activations interpretable, enabling faithful explanations without extra supervision.

Result: Extensive experiments on CIFAR-100, CUB-200-2011, Places365, ImageNet-100, and ImageNet-1k show the approach maintains CLIP’s strong zero-shot classification accuracy while providing meaningful concept-level explanations.

Conclusion: EZPC offers a principled step toward interpretable and trustworthy vision-language models by grounding open-vocabulary predictions in explicit semantic concepts, bridging the gap between performance and interpretability.

Abstract: Large-scale vision-language models such as CLIP have achieved remarkable success in zero-shot image recognition, yet their predictions remain largely opaque to human understanding. In contrast, Concept Bottleneck Models provide interpretable intermediate representations by reasoning through human-defined concepts, but they rely on concept supervision and lack the ability to generalize to unseen classes. We introduce EZPC that bridges these two paradigms by explaining CLIP’s zero-shot predictions through human-understandable concepts. Our method projects CLIP’s joint image-text embeddings into a concept space learned from language descriptions, enabling faithful and transparent explanations without additional supervision. The model learns this projection via a combination of alignment and reconstruction objectives, ensuring that concept activations preserve CLIP’s semantic structure while remaining interpretable. Extensive experiments on five benchmark datasets, CIFAR-100, CUB-200-2011, Places365, ImageNet-100, and ImageNet-1k, demonstrate that our approach maintains CLIP’s strong zero-shot classification accuracy while providing meaningful concept-level explanations. By grounding open-vocabulary predictions in explicit semantic concepts, our method offers a principled step toward interpretable and trustworthy vision-language models. Code is available at https://github.com/oonat/ezpc.

Method: Created Ghost-FWL dataset with 24K frames and 7.5B peak-level annotations across 10 diverse scenes. Developed FWL-based baseline model for ghost detection and FWL-MAE (masked autoencoder) for self-supervised representation learning on FWL data.

Result: Baseline outperforms existing methods in ghost removal accuracy. Ghost removal enhances LiDAR-based SLAM (66% trajectory error reduction) and 3D object detection (50x false positive reduction).

Conclusion: Full-waveform LiDAR provides crucial cues for ghost detection in mobile scenarios. The large-scale Ghost-FWL dataset enables effective ghost removal that significantly improves downstream perception and localization tasks.

Abstract: LiDAR has become an essential sensing modality in autonomous driving, robotics, and smart-city applications. However, ghost points (or ghosts), which are false reflections caused by multi-path laser returns from glass and reflective surfaces, severely degrade 3D mapping and localization accuracy. Prior ghost removal relies on geometric consistency in dense point clouds, failing on mobile LiDAR’s sparse, dynamic data. We address this by exploiting full-waveform LiDAR (FWL), which captures complete temporal intensity profiles rather than just peak distances, providing crucial cues for distinguishing ghosts from genuine reflections in mobile scenarios. As this is a new task, we present Ghost-FWL, the first and largest annotated mobile FWL dataset for ghost detection and removal. Ghost-FWL comprises 24K frames across 10 diverse scenes with 7.5 billion peak-level annotations, which is 100x larger than existing annotated FWL datasets. Benefiting from this large-scale dataset, we establish a FWL-based baseline model for ghost detection and propose FWL-MAE, a masked autoencoder for efficient self-supervised representation learning on FWL data. Experiments show that our baseline outperforms existing methods in ghost removal accuracy, and our ghost removal further enhances downstream tasks such as LiDAR-based SLAM (66% trajectory error reduction) and 3D object detection (50x false positive reduction). The dataset and code is publicly available and can be accessed via the project page: https://keio-csg.github.io/Ghost-FWL

Method: Captured a comprehensive 3D dataset comprising 50,000 images divided into 150 scenes (ground, aerial, and mixed) focusing on European landmarks, with curated calibration data, camera poses, and depth maps.

Result: Created TerraSky3D - a high-resolution large-scale 3D reconstruction dataset that provides challenging data for training and evaluating 3D reconstruction pipelines.

Conclusion: TerraSky3D addresses the need for comprehensive 3D datasets and can serve as a valuable resource for developing and benchmarking 3D reconstruction methods.

Abstract: Despite the growing need for data of more and more sophisticated 3D reconstruction pipelines, we can still observe a scarcity of suitable public datasets. Existing 3D datasets are either low resolution, limited to a small amount of scenes, based on images of varying quality because retrieved from the internet, or limited to specific capturing scenarios. Motivated by this lack of suitable 3D datasets, we captured TerraSky3D, a high-resolution large-scale 3D reconstruction dataset comprising 50,000 images divided into 150 ground, aerial, and mixed scenes. The dataset focuses on European landmarks and comes with curated calibration data, camera poses, and depth maps. TerraSky3D tries to answer the need for challenging dataset that can be used to train and evaluate 3D reconstruction-related pipelines.

Method: Created DinoDental benchmark from multiple public datasets covering panoramic radiographs and intraoral photographs. Evaluated DINOv3 across classification, detection, and instance segmentation tasks. Analyzed transfer performance by scaling model size and input resolution, comparing frozen features, full fine-tuning, and LoRA adaptation.

Result: DINOv3 serves as strong unified encoder for dental image analysis across both panoramic radiographs and intraoral photographs, remaining competitive across tasks with particular advantages for intraoral image understanding and boundary-sensitive dense prediction.

Conclusion: DinoDental provides systematic framework for evaluating DINOv3 in dental analysis, establishing foundational benchmark to guide efficient model selection and adaptation for dental AI community.

Abstract: The scarcity and high cost of expert annotations in dental imaging present a significant challenge for the development of AI in dentistry. DINOv3, a state-of-the-art, self-supervised vision foundation model pre-trained on 1.7 billion images, offers a promising pathway to mitigate this issue. However, its reliability when transferred to the dental domain, with its unique imaging characteristics and clinical subtleties, remains unclear. To address this, we introduce DinoDental, a unified benchmark designed to systematically evaluate whether DINOv3 can serve as a reliable, off-the-shelf encoder for comprehensive dental image analysis without requiring domain-specific pre-training. Constructed from multiple public datasets, DinoDental covers a wide range of tasks, including classification, detection, and instance segmentation on both panoramic radiographs and intraoral photographs. We further analyze the model’s transfer performance by scaling its size and input resolution, and by comparing different adaptation strategies, including frozen features, full fine-tuning, and the parameter-efficient Low-Rank Adaptation (LoRA) method. Our experiments show that DINOv3 can serve as a strong unified encoder for dental image analysis across both panoramic radiographs and intraoral photographs, remaining competitive across tasks while showing particularly clear advantages for intraoral image understanding and boundary-sensitive dense prediction. Collectively, DinoDental provides a systematic framework for comprehensively evaluating DINOv3 in dental analysis, establishing a foundational benchmark to guide efficient and effective model selection and adaptation for the dental AI community.

Method: Formulates gaze modeling as autoregressive dynamical system using Affinity Relation Transformer (ART) to process gaze-centric graphs of driver gaze, traffic objects, and road structure. Introduces Object Density Network (ODN) to predict next-step gaze distributions. Uses raw gaze data without fixation filtering.

Result: Produces more natural gaze trajectories, scanpath dynamics, and saliency maps than existing attention models. Released Focus100 dataset with raw gaze data from 30 participants viewing egocentric driving footage.

Conclusion: The unified approach offers valuable insights for temporal modeling of human attention in dynamic environments, with applications in automotive safety and computer vision.

Abstract: Accurately modelling human attention is essential for numerous computer vision applications, particularly in the domain of automotive safety. Existing methods typically collapse gaze into saliency maps or scanpaths, treating gaze dynamics only implicitly. We instead formulate gaze modelling as an autoregressive dynamical system and explicitly unroll raw gaze trajectories over time, conditioned on both gaze history and the evolving environment. Driving scenes are represented as gaze-centric graphs processed by the Affinity Relation Transformer (ART), a heterogeneous graph transformer that models interactions between driver gaze, traffic objects, and road structure. We further introduce the Object Density Network (ODN) to predict next-step gaze distributions, capturing the stochastic and object-centric nature of attentional shifts in complex environments. We also release Focus100, a new dataset of raw gaze data from 30 participants viewing egocentric driving footage. Trained directly on raw gaze, without fixation filtering, our unified approach produces more natural gaze trajectories, scanpath dynamics, and saliency maps than existing attention models, offering valuable insights for the temporal modelling of human attention in dynamic environments.

Method: SFDemorpher performs identity disentanglement within joint StyleGAN latent and high-dimensional feature spaces. It uses a dual-pass training strategy handling both morphed and bona fide documents, leveraging a hybrid corpus with predominantly synthetic identities to enhance robustness against unseen distributions.

Result: Extensive evaluation confirms state-of-the-art generalizability across unseen identities, diverse capture conditions, and 13 morphing techniques. The framework achieves superior D-MAD performance by widening the margin between score distributions of bona fide and morphed samples while providing high-fidelity visual reconstructions.

Conclusion: SFDemorpher presents an effective framework for operational deployment of face demorphing for differential morphing attack detection, offering improved generalizability and explainability through high-fidelity visual reconstructions.

Abstract: Face morphing attacks compromise biometric security by creating document images that verify against multiple identities, posing significant risks from document issuance to border control. Differential Morphing Attack Detection (D-MAD) offers an effective countermeasure, particularly when employing face demorphing to disentangle identities blended in the morph. However, existing methods lack operational generalizability due to limited training data and the assumption that all document inputs are morphs. This paper presents SFDemorpher, a framework designed for the operational deployment of face demorphing for D-MAD that performs identity disentanglement within joint StyleGAN latent and high-dimensional feature spaces. We introduce a dual-pass training strategy handling both morphed and bona fide documents, leveraging a hybrid corpus with predominantly synthetic identities to enhance robustness against unseen distributions. Extensive evaluation confirms state-of-the-art generalizability across unseen identities, diverse capture conditions, and 13 morphing techniques, spanning both border verification and the challenging document enrollment stage. Our framework achieves superior D-MAD performance by widening the margin between the score distributions of bona fide and morphed samples while providing high-fidelity visual reconstructions facilitating explainability.

Method: Incorporates multiview visual-language reasoning into video generation, uses MV-VLM evaluator for spatiotemporal consistency assessment, and implements a closed-loop generation-evaluation-regeneration mechanism with object-level refinement.

Result: Achieves diverse driving video generation with fine-grained controllability (especially for long-tail objects) and significantly better spatiotemporal consistency than previous approaches.

Conclusion: VistaGEN successfully addresses the challenge of fine-grained object-level control in driving video generation while maintaining spatiotemporal consistency through its novel multiview visual-language reasoning approach and closed-loop mechanism.

Abstract: Driving video generation has achieved much progress in controllability, video resolution, and length, but fails to support fine-grained object-level controllability for diverse driving videos, while preserving the spatiotemporal consistency, especially in long video generation. In this paper, we present a new driving video generation technique, called VistaGEN, which enables fine-grained control of specific entities, including 3D objects, images, and text descriptions, while maintaining spatiotemporal consistency in long video sequences. Our key innovation is the incorporation of multiview visual-language reasoning into the long driving video generation. To this end, we inject visual-language features into a multiview video generator to enable fine-grained controllability. More importantly, we propose a multiview vision-language evaluator (MV-VLM) to intelligently and automatically evaluate spatiotemporal consistency of the generated content, thus formulating a novel generation-evaluation-regeneration closed-loop generation mechanism. This mechanism ensures high-quality, coherent outputs, facilitating the creation of complex and reliable driving scenarios. Besides, within the closed-loop generation, we introduce an object-level refinement module to refine the unsatisfied results evaluated from the MV-VLM and then feed them back to the video generator for regeneration. Extensive evaluation shows that our VistaGEN achieves diverse driving video generation results with fine-grained controllability, especially for long-tail objects, and much better spatiotemporal consistency than previous approaches.

Method: Introduces SEA metric that uses commonsense knowledge about class-defining visual elements, leverages visual question answering models to detect element presence, and computes scores reflecting semantic retention under visual economy. Also creates CommonSketch dataset with 23,100 annotated sketches across 300 classes.

Result: SEA aligns closely with human judgments and reliably discriminates levels of abstraction efficiency. CommonSketch serves as a benchmark for evaluating element-level sketch understanding across vision-language models.

Conclusion: SEA provides the first effective metric for quantifying sketch abstraction efficiency, addressing a fundamental gap in sketch evaluation. The CommonSketch dataset enables systematic evaluation of sketch understanding in multimodal models.

Abstract: A sketch is a distilled form of visual abstraction that conveys core concepts through simplified yet purposeful strokes while omitting extraneous detail. Despite its expressive power, quantifying the efficiency of semantic abstraction in sketches remains challenging. Existing evaluation methods that rely on reference images, low-level visual features, or recognition accuracy do not capture abstraction, the defining property of sketches. To address these limitations, we introduce SEA (Sketch Evaluation metric for Abstraction efficiency), a reference-free metric that assesses how economically a sketch represents class-defining visual elements while preserving semantic recognizability. These elements are derived per class from commonsense knowledge about features typically depicted in sketches. SEA leverages a visual question answering model to determine the presence of each element and returns a quantitative score that reflects semantic retention under visual economy. To support this metric, we present CommonSketch, the first semantically annotated sketch dataset, comprising 23,100 human-drawn sketches across 300 classes, each paired with a caption and element-level annotations. Experiments show that SEA aligns closely with human judgments and reliably discriminates levels of abstraction efficiency, while CommonSketch serves as a benchmark providing systematic evaluation of element-level sketch understanding across various vision-language models.

Method: Uses dedicated encoders to extract video/audio features, applies residual vector quantization to discretize them into unified tokens aligned with text, creating a shared video-audio-text token space. Builds a multimodal LLM for video editing through multimodal alignment and supervised fine-tuning, supporting tasks like video selection, script generation, and background music selection.

Result: Experiments on real-world advertisement datasets show AutoCut reduces production cost and iteration time while improving consistency and controllability.

Conclusion: AutoCut provides an effective end-to-end framework for scalable video creation that unifies multimodal processing, paving the way for more efficient digital advertising workflows.

Abstract: Short-form videos have become a primary medium for digital advertising, requiring scalable and efficient content creation. However, current workflows and AI tools remain disjoint and modality-specific, leading to high production costs and low overall efficiency. To address this issue, we propose AutoCut, an end-to-end advertisement video editing framework based on multimodal discretization and controllable editing. AutoCut employs dedicated encoders to extract video and audio features, then applies residual vector quantization to discretize them into unified tokens aligned with textual representations, constructing a shared video-audio-text token space. Built upon a foundation model, we further develop a multimodal large language model for video editing through combined multimodal alignment and supervised fine-tuning, supporting tasks covering video selection and ordering, script generation, and background music selection within a unified editing framework. Finally, a complete production pipeline converts the predicted token sequences into deployable long video outputs. Experiments on real-world advertisement datasets show that AutoCut reduces production cost and iteration time while substantially improving consistency and controllability, paving the way for scalable video creation.

Method: Unable to determine method due to access error

Result: Unable to determine results due to access error

Conclusion: Unable to analyze paper content due to technical access issues

Abstract: Failed to fetch summary for 2307.01502: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2307.01502&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: 1) Coarse-to-fine token localization strategy to refine editable regions; 2) Analysis of VAR intermediate features to identify structure-related features for injection; 3) Reinforcement learning-based adaptive feature injection scheme to learn optimal scale- and layer-specific injection ratios.

Result: Extensive experiments show superior structural consistency and editing quality compared to state-of-the-art approaches across both local and global editing scenarios.

Conclusion: The proposed framework effectively addresses VAR-based editing challenges through feature analysis and adaptive injection mechanisms, achieving better balance between editing fidelity and structure preservation.

Abstract: Visual autoregressive (VAR) models have recently emerged as a promising family of generative models, enabling a wide range of downstream vision tasks such as text-guided image editing. By shifting the editing paradigm from noise manipulation in diffusion-based methods to token-level operations, VAR-based approaches achieve better background preservation and significantly faster inference. However, existing VAR-based editing methods still face two key challenges: accurately localizing editable tokens and maintaining structural consistency in the edited results. In this work, we propose a novel text-guided image editing framework rooted in an analysis of intermediate feature distributions within VAR models. First, we introduce a coarse-to-fine token localization strategy that can refine editable regions, balancing editing fidelity and background preservation. Second, we analyze the intermediate representations of VAR models and identify structure-related features, by which we design a simple yet effective feature injection mechanism to enhance structural consistency between the edited and source images. Third, we develop a reinforcement learning-based adaptive feature injection scheme that automatically learns scale- and layer-specific injection ratios to jointly optimize editing fidelity and structure preservation. Extensive experiments demonstrate that our method achieves superior structural consistency and editing quality compared with state-of-the-art approaches, across both local and global editing scenarios.

Method: Generated hyperspectral cubes (211 bands, 400-2500 nm) using PROSAIL radiative transfer model inversion from Sentinel-2 data, followed by forward simulations for physically consistent reflectance spectra.

Result: Dataset includes 10,915 hyperspectral image cubes with vegetation trait maps, uncertainty quantification, and covers four diverse ecological regions with 64×64 pixel spatial resolution.

Conclusion: This dataset enables research in vegetation trait retrieval, radiative transfer emulation, and uncertainty quantification for remote sensing applications.

Abstract: This dataset provides a large collection of 10,915 synthetic hyperspectral image cubes paired with pixel-level vegetation trait maps, designed to support research in radiative transfer emulation, vegetation trait retrieval, and uncertainty quantification. Each hyperspectral cube contains 211 bands spanning 400–2500 nm at 10 nm resolution and a fixed spatial layout of 64 \times 64 pixels, offering continuous simulated surface reflectance spectra suitable for emulator development and machine-learning tasks requiring high spectral detail. Vegetation traits were derived by inverting Sentinel-2 Level-2A surface reflectance using a PROSAIL-based lookup-table approach, followed by forward PROSAIL simulations to generate hyperspectral reflectance under physically consistent canopy and illumination conditions. The dataset covers four ecologically diverse regions – East Africa, Northern France, Eastern India, and Southern Spain – and includes 5th and 95th percentile uncertainty maps as well as Sentinel-2 scene classification layers. This resource enables benchmarking of inversion methods, development of fast radiative transfer emulators, and studies of spectral–biophysical relationships under controlled yet realistic environmental variability.

Method: Proposes IVMSD dataset for maritime scenarios, and a Multi-task Complementary Learning Framework (MCLF) with: 1) Frequency-Spatial Enhancement Complementary module for degradation suppression, 2) Semantic-Visual Consistency Attention for semantic guidance, and 3) cross-modality guided attention for selective fusion.

Result: Achieves state-of-the-art segmentation performance on IVMSD dataset, significantly enhancing robustness and perceptual quality under complex maritime conditions.

Conclusion: The proposed framework effectively addresses marine image degradation through collaborative restoration, fusion, and segmentation, demonstrating improved maritime scene understanding.

Abstract: Marine scene understanding and segmentation plays a vital role in maritime monitoring and navigation safety. However, prevalent factors like fog and strong reflections in maritime environments cause severe image degradation, significantly compromising the stability of semantic perception. Existing restoration and enhancement methods typically target specific degradations or focus solely on visual quality, lacking end-to-end collaborative mechanisms that simultaneously improve structural recovery and semantic effectiveness. Moreover, publicly available infrared-visible datasets are predominantly collected from urban scenes, failing to capture the authentic characteristics of coupled degradations in marine environments. To address these challenges, the Infrared-Visible Maritime Ship Dataset (IVMSD) is proposed to cover various maritime scenarios under diverse weather and illumination conditions. Building upon this dataset, a Multi-task Complementary Learning Framework (MCLF) is proposed to collaboratively perform image restoration, multimodal fusion, and semantic segmentation within a unified architecture. The framework includes a Frequency-Spatial Enhancement Complementary (FSEC) module for degradation suppression and structural enhancement, a Semantic-Visual Consistency Attention (SVCA) module for semantic-consistent guidance, and a cross-modality guided attention mechanism for selective fusion. Experimental results on IVMSD demonstrate that the proposed method achieves state-of-the-art segmentation performance, significantly enhancing robustness and perceptual quality under complex maritime conditions.

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 2408.13516: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2408.13516&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: DiPA uses smartphone screens to display adversarial patches instead of printed artifacts. It leverages an ensemble of state-of-the-art face recognition models (ArcFace, MagFace, CosFace) to enhance transferability across unseen commercial systems. The approach eliminates the need for total-variation regularization and enables real-time interactive attacks.

Result: DiPA demonstrates superior performance over existing physical attacks in success rate, feature-space distortion, and reductions in detection confidence. The interactive demo shows real-time dodging attacks against deployed face-recognition systems with immediate observable effects.

Conclusion: DiPA reveals critical vulnerabilities at the intersection of mobile devices, pervasive vision, and sensor-driven authentication infrastructures, highlighting the need for more robust security measures in camera-based authentication systems.

Abstract: This demonstration presents Digital-Physical Adversarial Attacks (DiPA), a new class of practical adversarial attacks against pervasive camera-based authentication systems, where an attacker displays an adversarial patch directly on a smartphone screen instead of relying on printed artifacts. This digital-only physical presentation enables rapid deployment, removes the need for total-variation regularization, and improves patch transferability in black-box conditions. DiPA leverages an ensemble of state-of-the-art face-recognition models (ArcFace, MagFace, CosFace) to enhance transfer across unseen commercial systems. Our interactive demo shows a real-time dodging attack against a deployed face-recognition camera, preventing authorized users from being recognized while participants dynamically adjust patch patterns and observe immediate effects on the sensing pipeline. We further demonstrate DiPA’s superiority over existing physical attacks in terms of success rate, feature-space distortion, and reductions in detection confidence, highlighting critical vulnerabilities at the intersection of mobile devices, pervasive vision, and sensor-driven authentication infrastructures.

Method: Cannot determine method as abstract is unavailable due to rate limiting error

Result: Cannot determine results as abstract is unavailable due to rate limiting error

Conclusion: Cannot draw conclusions about paper content due to data retrieval failure

Abstract: Failed to fetch summary for 2410.21086: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2410.21086&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes WaveSDG with a novel Wavelet-based Invariant Structure Extraction and Refinement (WISER) module that processes encoder features using wavelet sub-band decomposition. Low-frequency components are refined to anchor global anatomy, while high-frequency sub-bands are selectively enhanced for directional edges and noise suppression.

Result: WaveSDG outperforms seven state-of-the-art methods on optic cup and optic disc segmentation across one source and five unseen target datasets, achieving best balanced Dice score and lowest 95th percentile Hausdorff distance with reduced variance.

Conclusion: WaveSDG effectively addresses domain generalization in fundus imaging by decoupling anatomical structure from domain-specific appearance through wavelet decomposition, demonstrating improved accuracy, robustness, and cross-domain stability.

Abstract: Domain generalization in fundus imaging is challenging due to variations in acquisition conditions across devices and clinical settings. The inability to adapt to these variations causes performance degradation on unseen domains for deep learning models. Besides, obtaining annotated data across domains is often expensive and privacy constraints restricts their availability. Although single-source domain generalization (SDG) offers a realistic solution to this problem, the existing approaches frequently fail to capture anatomical topology or decouple appearance from anatomical features. This research introduces WaveSDG, a new wavelet-guided segmentation network for SDG. It decouples anatomical structure from domain-specific appearance through a wavelet sub-band decomposition. A novel Wavelet-based Invariant Structure Extraction and Refinement (WISER) module is proposed to process encoder features by leveraging distinct semantic roles of each wavelet sub-band. The module refines low-frequency components to anchor global anatomy, while selectively enhancing directional edges and suppressing noise within the high-frequency sub-bands. Extensive ablation studies validate the effectiveness of the WISER module and its decoupling strategy. Our evaluations on optic cup and optic disc segmentation across one source and five unseen target datasets show that WaveSDG consistently outperforms seven state-of-the-art methods. Notably, it achieves the best balanced Dice score and lowest 95th percentile Hausdorff distance with reduced variance, indicating improved accuracy, robustness, and cross-domain stability.

Method: Replace final linear layer with k-means-based classifier. Formalize k-means-based post-hoc explanations for classifier, encoder output, and combinations of intermediate feature activations. Use spatial consistency of convolutional receptive fields to generate concept-based explanation maps with gradient-free feature attribution.

Result: Using shallower, less compressed feature activations (like from last three blocks B234) creates a trade-off between semantic fidelity and slight reduction in predictive performance. ResNet34 evaluation shows the approach works.

Conclusion: K-means-based classifiers improve CNN interpretability while maintaining performance. Combining intermediate feature activations provides better concept-based explanations, though with trade-offs between semantic fidelity and accuracy.

Abstract: Although standard Convolutional Neural Networks (CNNs) can be mathematically reinterpreted as Self-Explainable Models (SEMs), their built-in prototypes do not on their own accurately represent the data. Replacing the final linear layer with a $k$-means-based classifier addresses this limitation without compromising performance. This work introduces a common formalization of $k$-means-based post-hoc explanations for the classifier, the encoder’s final output (B4), and combinations of intermediate feature activations. The latter approach leverages the spatial consistency of convolutional receptive fields to generate concept-based explanation maps, which are supported by gradient-free feature attribution maps. Empirical evaluation with a ResNet34 shows that using shallower, less compressed feature activations, such as those from the last three blocks (B234), results in a trade-off between semantic fidelity and a slight reduction in predictive performance.

Method: INSID3 leverages scaled-up dense self-supervised features from DINOv3, which exhibit strong spatial structure and semantic correspondence. The approach is training-free and segments concepts at varying granularities using only frozen DINOv3 features given an in-context example.

Result: INSID3 achieves state-of-the-art results across one-shot semantic, part, and personalized segmentation, outperforming previous work by +7.5% mIoU while using 3x fewer parameters and requiring no mask or category-level supervision.

Conclusion: A single self-supervised backbone (DINOv3) can effectively support both semantic matching and segmentation without any supervision, offering a minimalist yet powerful approach to in-context segmentation that preserves generalization while reducing complexity.

Abstract: In-context segmentation (ICS) aims to segment arbitrary concepts, e.g., objects, parts, or personalized instances, given one annotated visual examples. Existing work relies on (i) fine-tuning vision foundation models (VFMs), which improves in-domain results but harms generalization, or (ii) combines multiple frozen VFMs, which preserves generalization but yields architectural complexity and fixed segmentation granularities. We revisit ICS from a minimalist perspective and ask: Can a single self-supervised backbone support both semantic matching and segmentation, without any supervision or auxiliary models? We show that scaled-up dense self-supervised features from DINOv3 exhibit strong spatial structure and semantic correspondence. We introduce INSID3, a training-free approach that segments concepts at varying granularities only from frozen DINOv3 features, given an in-context example. INSID3 achieves state-of-the-art results across one-shot semantic, part, and personalized segmentation, outperforming previous work by +7.5 % mIoU, while using 3x fewer parameters and without any mask or category-level supervision. Code is available at https://github.com/visinf/INSID3 .

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 2501.19111: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2501.19111&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Introduces differential probing to analyze concept token influence, revealing three-stage generative process (Blueprint, Instantiation, Refinement). Proposes ConceptWeaver with stage-aware optimization to learn concept-specific semantic offsets from single reference images, deployed via ConceptWeaver Guidance during inference.

Result: Extensive experiments validate that ConceptWeaver enables high-fidelity, compositional synthesis and editing, demonstrating precise multi-granularity content manipulation by leveraging the intrinsic staged nature of flow models.

Conclusion: Understanding and leveraging the intrinsic staged nature of flow models is key to unlocking precise, multi-granularity content manipulation, with ConceptWeaver providing an effective framework for one-shot concept disentanglement.

Abstract: Pre-trained flow-based models excel at synthesizing complex scenes yet lack a direct mechanism for disentangling and customizing their underlying concepts from one-shot real-world sources. To demystify this process, we first introduce a novel differential probing technique to isolate and analyze the influence of individual concept tokens on the velocity field over time. This investigation yields a critical insight: the generative process is not monolithic but unfolds in three distinct stages. An initial \textbf{Blueprint Stage} establishes low-frequency structure, followed by a pivotal \textbf{Instantiation Stage} where content concepts emerge with peak intensity and become naturally disentangled, creating an optimal window for manipulation. A final concept-insensitive refinement stage then synthesizes fine-grained details. Guided by this discovery, we propose \textbf{ConceptWeaver}, a framework for one-shot concept disentanglement. ConceptWeaver learns concept-specific semantic offsets from a single reference image using a stage-aware optimization strategy that aligns with the three-stage framework. These learned offsets are then deployed during inference via our novel ConceptWeaver Guidance (CWG) mechanism, which strategically injects them at the appropriate generative stage. Extensive experiments validate that ConceptWeaver enables high-fidelity, compositional synthesis and editing, demonstrating that understanding and leveraging the intrinsic, staged nature of flow models is key to unlocking precise, multi-granularity content manipulation.

Method: Cannot determine method due to failed paper fetch

Result: Cannot determine results due to failed paper fetch

Conclusion: Cannot determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2502.03330: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.03330&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes FGOS-Net based on frequency-geometric disentanglement: 1) Decomposes features into stable topology carrier and directional high-frequency bands, using latter to correct spatial misalignments from downsampling; 2) Introduces frequency-aligned scanning that makes serialization geometry-conditioned to preserve direction-consistent traces; 3) Uses active probing strategy to selectively inject high-frequency details and suppress texture ambiguity.

Result: Consistently outperforms strong baselines across four challenging benchmarks. Achieves 91.3% mIoU and 97.1% clDice on DeepCrack while running at 80 FPS with only 7.87 GFLOPs.

Conclusion: FGOS-Net effectively addresses the geometry mismatch problem in thin linear structure segmentation through frequency-geometric disentanglement and geometry-conditioned serialization, achieving superior performance with high efficiency.

Abstract: The segmentation of thin linear structures is inherently topology allowbreak-critical, where minor local errors can sever long-range connectivity. While recent State-Space Models (SSMs) offer efficient long-range modeling, their isotropic serialization (e.g., raster scanning) creates a geometry mismatch for anisotropic targets, causing state propagation across rather than along the structure trajectories. To address this, we propose FGOS-Net, a framework based on frequency allowbreak-geometric disentanglement. We first decompose features into a stable topology carrier and directional high-frequency bands, leveraging the latter to explicitly correct spatial misalignments induced by downsampling. Building on this calibrated topology, we introduce frequency-aligned scanning that elevates serialization to a geometry-conditioned decision, preserving direction-consistent traces. Coupled with an active probing strategy to selectively inject high-frequency details and suppress texture ambiguity, FGOS-Net consistently outperforms strong baselines across four challenging benchmarks. Notably, it achieves 91.3% mIoU and 97.1% clDice on DeepCrack while running at 80 FPS with only 7.87 GFLOPs.

Method: Proposes a framework that synergistically integrates lightweight artifact-aware detectors with MLLMs via a fuzzy decision tree. The decision tree treats detector outputs as fuzzy membership values, enabling adaptive fusion of complementary semantic and perceptual cues.

Result: Extensive experiments demonstrate state-of-the-art accuracy and strong generalization across diverse generative models.

Conclusion: The proposed hybrid approach effectively addresses limitations of both traditional artifact-based detectors and MLLMs by combining their complementary strengths for robust AI-generated image detection.

Abstract: The malicious use and widespread dissemination of AI-generated images pose a serious threat to the authenticity of digital content. Existing detection methods exploit low-level artifacts left by common manipulation steps within the generation pipeline, but they often lack generalization due to model-specific overfitting. Recently, researchers have resorted to Multimodal Large Language Models (MLLMs) for AIGC detection, leveraging their high-level semantic reasoning and broad generalization capabilities. While promising, MLLMs lack the fine-grained perceptual sensitivity to subtle generation artifacts, making them inadequate as standalone detectors. To address this issue, we propose a novel AI-generated image detection framework that synergistically integrates lightweight artifact-aware detectors with MLLMs via a fuzzy decision tree. The decision tree treats the outputs of basic detectors as fuzzy membership values, enabling adaptive fusion of complementary cues from semantic and perceptual perspectives. Extensive experiments demonstrate that the proposed method achieves state-of-the-art accuracy and strong generalization across diverse generative models.

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 2504.10833: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.10833&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: 1) Created GEditBench v2 with 1,200 real-world queries spanning 23 tasks including open-set category; 2) Developed PVC-Judge, an open-source pairwise assessment model trained via novel region-decoupled preference data synthesis pipelines; 3) Built VCReward-Bench with expert-annotated preference pairs to validate PVC-Judge.

Result: PVC-Judge achieves state-of-the-art evaluation performance among open-source models and surpasses GPT-5.1 on average. Benchmarking 16 frontier editing models reveals critical limitations and provides reliable foundation for advancing precise image editing.

Conclusion: GEditBench v2 enables more human-aligned evaluation of image editing models, addressing current limitations in evaluation frameworks and providing comprehensive assessment capabilities for visual consistency.

Abstract: Recent advances in image editing have enabled models to handle complex instructions with impressive realism. However, existing evaluation frameworks lag behind: current benchmarks suffer from narrow task coverage, while standard metrics fail to adequately capture visual consistency, i.e., the preservation of identity, structure and semantic coherence between edited and original images. To address these limitations, we introduce GEditBench v2, a comprehensive benchmark with 1,200 real-world user queries spanning 23 tasks, including a dedicated open-set category for unconstrained, out-of-distribution editing instructions beyond predefined tasks. Furthermore, we propose PVC-Judge, an open-source pairwise assessment model for visual consistency, trained via two novel region-decoupled preference data synthesis pipelines. Besides, we construct VCReward-Bench using expert-annotated preference pairs to assess the alignment of PVC-Judge with human judgments on visual consistency evaluation. Experiments show that our PVC-Judge achieves state-of-the-art evaluation performance among open-source models and even surpasses GPT-5.1 on average. Finally, by benchmarking 16 frontier editing models, we show that GEditBench v2 enables more human-aligned evaluation, revealing critical limitations of current models, and providing a reliable foundation for advancing precise image editing.

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 2504.11967: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.11967&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Uses visibility-guided flow matching to mask unknown regions in real scans, represents scenes with TSDF volumes in sparse grids, employs sparse transformers to model complex structures while masking unknown regions, and uses 3D layout boxes as conditioning with flexibility for other inputs like text or partial scans.

Result: Outperforms baselines in completion accuracy and generation quality, producing coherent, complete, and realistic 3D scenes from real-world incomplete scans.

Conclusion: Seen2Scene enables realistic 3D scene completion for complex real environments by learning directly from incomplete real-world scans, representing a significant advancement over synthetic-data-dependent approaches.

Abstract: We present Seen2Scene, the first flow matching-based approach that trains directly on incomplete, real-world 3D scans for scene completion and generation. Unlike prior methods that rely on complete and hence synthetic 3D data, our approach introduces visibility-guided flow matching, which explicitly masks out unknown regions in real scans, enabling effective learning from real-world, partial observations. We represent 3D scenes using truncated signed distance field (TSDF) volumes encoded in sparse grids and employ a sparse transformer to efficiently model complex scene structures while masking unknown regions. We employ 3D layout boxes as an input conditioning signal, and our approach is flexibly adapted to various other inputs such as text or partial scans. By learning directly from real-world, incomplete 3D scans, Seen2Scene enables realistic 3D scene completion for complex, cluttered real environments. Experiments demonstrate that our model produces coherent, complete, and realistic 3D scenes, outperforming baselines in completion accuracy and generation quality.

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 2505.03821: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.03821&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: 1) Dedicated OCR for text extraction from chemical images; 2) Joint encoding of text, image, and layout via Vision-Text-Layout encoder and Optical Chemical Structure Recognition vision encoder; 3) Two-stage training strategy for effective fusion; 4) Auto-regressive generation of Markush structure representation; 5) Created automatic pipeline for large-scale dataset construction.

Result: Substantially outperforms state-of-the-art models in multimodal Markush structure recognition while maintaining strong performance in molecule structure recognition. Introduced IP5-M benchmark dataset.

Conclusion: MarkushGrapher-2 advances multimodal chemical structure recognition with superior performance and provides valuable datasets for future research in this challenging domain.

Abstract: Automatically extracting chemical structures from documents is essential for the large-scale analysis of the literature in chemistry. Automatic pipelines have been developed to recognize molecules represented either in figures or in text independently. However, methods for recognizing chemical structures from multimodal descriptions (Markush structures) lag behind in precision and cannot be used for automatic large-scale processing. In this work, we present MarkushGrapher-2, an end-to-end approach for the multimodal recognition of chemical structures in documents. First, our method employs a dedicated OCR model to extract text from chemical images. Second, the text, image, and layout information are jointly encoded through a Vision-Text-Layout encoder and an Optical Chemical Structure Recognition vision encoder. Finally, the resulting encodings are effectively fused through a two-stage training strategy and used to auto-regressively generate a representation of the Markush structure. To address the lack of training data, we introduce an automatic pipeline for constructing a large-scale dataset of real-world Markush structures. In addition, we present IP5-M, a large manually-annotated benchmark of real-world Markush structures, designed to advance research on this challenging task. Extensive experiments show that our approach substantially outperforms state-of-the-art models in multimodal Markush structure recognition, while maintaining strong performance in molecule structure recognition. Code, models, and datasets are released publicly.

Method: Proposes a curriculum learning-based framework with progressive training strategy that guides the model from high-confidence, clearly defined scar regions to low-confidence or visually ambiguous samples with limited scar burden.

Result: Experimental results show enhanced segmentation accuracy and consistency, particularly for cases with minimal or diffuse scar, outperforming standard training baselines.

Conclusion: The curriculum learning strategy provides a principled way to leverage imperfect data for improved myocardial scar quantification in clinical applications.

Abstract: Identification and quantification of myocardial scar is important for diagnosis and prognosis of cardiovascular diseases. However, reliable scar segmentation from Late Gadolinium Enhancement Cardiac Magnetic Resonance (LGE-CMR) images remains a challenge due to variations in contrast enhancement across patients, suboptimal imaging conditions such as post contrast washout, and inconsistencies in ground truth annotations on diffuse scars caused by inter observer variability. In this work, we propose a curriculum learning-based framework designed to improve segmentation performance under these challenging conditions. The method introduces a progressive training strategy that guides the model from high-confidence, clearly defined scar regions to low confidence or visually ambiguous samples with limited scar burden. By structuring the learning process in this manner, the network develops robustness to uncertain labels and subtle scar appearances that are often underrepresented in conventional training pipelines. Experimental results show that the proposed approach enhances segmentation accuracy and consistency, particularly for cases with minimal or diffuse scar, outperforming standard training baselines. This strategy provides a principled way to leverage imperfect data for improved myocardial scar quantification in clinical applications. Our code is publicly available on GitHub.

Method: Proposes X-shaped Sparse Pixel Attack (XSPA) - an imperceptible structured attack that restricts perturbations to two intersecting diagonal lines. It jointly optimizes classification objective, cross-task semantic guidance, and regularization on perturbation magnitude and along-line smoothness. Only modifies about 1.76% of image pixels.

Result: XSPA consistently degrades performance across three tasks on COCO dataset: zero-shot accuracy drops by 52.33-67.00 points on CLIP models, GPT-4-evaluated caption consistency decreases by up to 58.60 points, and VQA correctness by up to 44.38 points.

Conclusion: Even highly sparse and visually subtle perturbations with fixed geometric priors can substantially disrupt cross-task semantics in VLMs, revealing a notable robustness gap in current multimodal systems.

Abstract: Vision-language models (VLMs) rely on a shared visual-textual representation space to perform tasks such as zero-shot classification, image captioning, and visual question answering (VQA). While this shared space enables strong cross-task generalization, it may also introduce a common vulnerability: small visual perturbations can propagate through the shared embedding space and cause correlated semantic failures across tasks. This risk is particularly important in interactive and decision-support settings, yet it remains unclear whether VLMs are robust to highly constrained, sparse, and geometrically fixed perturbations. To address this question, we propose X-shaped Sparse Pixel Attack (XSPA), an imperceptible structured attack that restricts perturbations to two intersecting diagonal lines. Compared with dense perturbations or flexible localized patches, XSPA operates under a much stricter attack budget and thus provides a more stringent test of VLM robustness. Within this sparse support, XSPA jointly optimizes a classification objective, cross-task semantic guidance, and regularization on perturbation magnitude and along-line smoothness, inducing transferable misclassification as well as semantic drift in captioning and VQA while preserving visual subtlety. Under the default setting, XSPA modifies only about 1.76% of image pixels. Experiments on the COCO dataset show that XSPA consistently degrades performance across all three tasks. Zero-shot accuracy drops by 52.33 points on OpenAI CLIP ViT-L/14 and 67.00 points on OpenCLIP ViT-B/16, while GPT-4-evaluated caption consistency decreases by up to 58.60 points and VQA correctness by up to 44.38 points. These results suggest that even highly sparse and visually subtle perturbations with fixed geometric priors can substantially disrupt cross-task semantics in VLMs, revealing a notable robustness gap in current multimodal systems.

Method: Proposes ORSIFlow, a saliency-guided rectified flow framework that performs saliency mask generation in a compact latent space using a frozen variational autoencoder. Includes a Salient Feature Discriminator for global semantic discrimination and a Salient Feature Calibrator for precise boundary refinement.

Result: Extensive experiments on multiple public benchmarks show ORSIFlow achieves state-of-the-art performance with significantly improved efficiency compared to existing methods.

Conclusion: ORSIFlow effectively addresses the challenges of ORSI-SOD by reformulating it as a deterministic latent flow generation problem, offering both high performance and computational efficiency.

Abstract: Optical Remote Sensing Image Salient Object Detection (ORSI-SOD) remains challenging due to complex backgrounds, low contrast, irregular object shapes, and large variations in object scale. Existing discriminative methods directly regress saliency maps, while recent diffusion-based generative approaches suffer from stochastic sampling and high computational cost. In this paper, we propose ORSIFlow, a saliency-guided rectified flow framework that reformulates ORSI-SOD as a deterministic latent flow generation problem. ORSIFlow performs saliency mask generation in a compact latent space constructed by a frozen variational autoencoder, enabling efficient inference with only a few steps. To enhance saliency awareness, we design a Salient Feature Discriminator for global semantic discrimination and a Salient Feature Calibrator for precise boundary refinement. Extensive experiments on multiple public benchmarks show that ORSIFlow achieves state-of-the-art performance with significantly improved efficiency. Codes are available at: https://github.com/Ch3nSir/ORSIFlow.

Method: ELViS operates in similarity space rather than representation space. It leverages local descriptor correspondences, refines similarities through optimal transport with data-dependent gains to suppress uninformative descriptors, and aggregates strong correspondences via voting into image-level similarity.

Result: ELViS outperforms competing methods by a large margin in out-of-domain scenarios and on average across eight datasets spanning landmarks, artworks, products, and multi-domain collections, while requiring only a fraction of their computational cost.

Conclusion: The similarity-space approach with strong inductive biases yields a simple, efficient, and interpretable model that effectively generalizes to unseen domains for image retrieval tasks.

Abstract: Large-scale instance-level training data is scarce, so models are typically trained on domain-specific datasets. Yet in real-world retrieval, they must handle diverse domains, making generalization to unseen data critical. We introduce ELViS, an image-to-image similarity model that generalizes effectively to unseen domains. Unlike conventional approaches, our model operates in similarity space rather than representation space, promoting cross-domain transfer. It leverages local descriptor correspondences, refines their similarities through an optimal transport step with data-dependent gains that suppress uninformative descriptors, and aggregates strong correspondences via a voting process into an image-level similarity. This design injects strong inductive biases, yielding a simple, efficient, and interpretable model. To assess generalization, we compile a benchmark of eight datasets spanning landmarks, artworks, products, and multi-domain collections, and evaluate ELViS as a re-ranking method. Our experiments show that ELViS outperforms competing methods by a large margin in out-of-domain scenarios and on average, while requiring only a fraction of their computational cost. Code available at: https://github.com/pavelsuma/ELViS/

Method: Proposes a modular framework with a lightweight CNN classifier as a router that evaluates input images and directs them to specialized restoration nodes (demonstrated with U-Net experts). The system is model-agnostic and extensible, isolating reconstruction paths to prevent feature conflicts.

Result: The framework offers computationally accessible multi-degradation restoration on standard local hardware, with reduced training overhead compared to monolithic models. Adding new degradation types only requires training a single expert and updating the router.

Conclusion: The modular, task-decoupled approach provides a scalable and efficient solution for image restoration that prevents negative task interference and enables easy extensibility to new degradation types.

Abstract: Restoring images affected by various types of degradation, such as noise, blur, or improper exposure, remains a significant challenge in computer vision. While recent trends favor complex monolithic all-in-one architectures, these models often suffer from negative task interference and require extensive joint training cycles on high-end computing clusters. In this paper, we propose a modular, task-decoupled image restoration framework based on an explicit diagnostic routing mechanism. The architecture consists of a lightweight Convolutional Neural Network (CNN) classifier that evaluates the input image and dynamically directs it to a specialized restoration node. A key advantage of this framework is its model-agnostic extensibility: while we demonstrate it using three independent U-Net experts, the system allows for the integration of any restoration method tailored to specific tasks. By isolating reconstruction paths, the framework prevents feature conflicts and significantly reduces training overhead. Unlike monolithic models, adding new degradation types in our framework only requires training a single expert and updating the router, rather than a full system retraining. Experimental results demonstrate that this computationally accessible approach offers a scalable and efficient solution for multi-degradation restoration on standard local hardware. The code will be published upon paper acceptance.

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 2508.09428: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.09428&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Created Industrial3D dataset with 612M expertly labeled points from 13 water treatment facilities at 6mm resolution. Established cross-paradigm benchmark evaluating 9 methods across supervised, weakly supervised, unsupervised, and foundation model settings under unified protocol.

Result: Best supervised method achieves 55.74% mIoU, while zero-shot Point-SAM reaches only 15.79% - revealing 39.95% gap highlighting domain-transfer challenges. Analysis shows gap stems from statistical rarity (215:1 imbalance) and geometric ambiguity that frequency-based re-weighting alone cannot resolve.

Conclusion: Industrial3D provides the largest and most demanding testbed for industrial 3D scene understanding, quantifying unresolved domain-transfer challenges for industrial TLS data and highlighting need for specialized approaches beyond current architectural benchmarks.

Abstract: Automated semantic understanding of dense point clouds is a prerequisite for Scan-to-BIM pipelines, digital twin construction, and as-built verification–core tasks in the digital transformation of the construction industry. Yet for industrial mechanical, electrical, and plumbing (MEP) facilities, this challenge remains largely unsolved: TLS acquisitions of water treatment plants, chiller halls, and pumping stations exhibit extreme geometric ambiguity, severe occlusion, and extreme class imbalance that architectural benchmarks (e.g., S3DIS or ScanNet) cannot adequately represent. We present Industrial3D, a terrestrial LiDAR dataset comprising 612 million expertly labelled points at 6 mm resolution from 13 water treatment facilities. At 6.6x the scale of the closest comparable MEP dataset, Industrial3D provides the largest and most demanding testbed for industrial 3D scene understanding to date. We further establish the first industrial cross-paradigm benchmark, evaluating nine representative methods across fully supervised, weakly supervised, unsupervised, and foundation model settings under a unified benchmark protocol. The best supervised method achieves 55.74% mIoU, whereas zero-shot Point-SAM reaches only 15.79%–a 39.95 percentage-point gap that quantifies the unresolved domain-transfer challenge for industrial TLS data. Systematic analysis reveals that this gap originates from a dual crisis: statistical rarity (215:1 imbalance, 3.5x more severe than S3DIS) and geometric ambiguity (tail-class points share cylindrical primitives with head-class pipes) that frequency-based re-weighting alone cannot resolve. Industrial3D, along with benchmark code and pre-trained models, will be publicly available at https://github.com/pointcloudyc/Industrial3D.

Method: Generated synthetic datasets in NVIDIA Isaac Sim and combined them with limited real-world fruit images to train YOLO-based detection models under three regimes: real-only, synthetic-only, and hybrid. Evaluated on in-domain and domain shift test datasets, with deployment on Jetson Orin NX using TensorRT optimization.

Result: Real-only models achieved highest accuracy, synthetic-only models showed reduced performance due to domain gap, but hybrid training significantly improved over synthetic-only and approached real-only performance while reducing annotation needs. All models degraded under domain shift, but hybrid models showed improved robustness. Successful real-time deployment on Jetson Orin NX achieved.

Conclusion: Synthetic data is most effective when combined with real data, hybrid training reduces manual annotation requirements while maintaining competitive performance, and deployment constraints must be considered alongside detection accuracy for practical embedded applications.

Abstract: This study investigates the effectiveness of synthetic data for sim-to-real transfer in object detection under constrained data conditions and embedded deployment requirements. Synthetic datasets were generated in NVIDIA Isaac Sim and combined with limited real-world fruit images to train YOLO-based detection models under real-only, synthetic-only, and hybrid regimes. Performance was evaluated on two test datasets: an in-domain dataset with conditions matching the training data and a domain shift dataset containing real fruit and different background conditions. Results show that models trained exclusively on real data achieve the highest accuracy, while synthetic-only models exhibit reduced performance due to a domain gap. Hybrid training strategies significantly improve performance compared to synthetic-only approaches and achieve results close to real-only training while reducing the need for manual annotation. Under domain shift conditions, all models show performance degradation, with hybrid models providing improved robustness. The trained models were successfully deployed on a Jetson Orin NX using TensorRT optimization, achieving real-time inference performance. The findings highlight that synthetic data is most effective when used in combination with real data and that deployment constraints must be considered alongside detection accuracy.

Method: Built on pruned mobile U-Net backbone with in-context spatial concatenation in latent space. Uses (target|blank) configuration for generation and (target|source) for editing. Employs task-progressive joint pretraining strategy (T2I → editing → joint tasks) with SFT and reinforcement learning. Uses step distillation to reduce denoising to 4 steps.

Result: Achieves GenEval score of 0.72 for generation and ImgEdit score of 4.11 for editing, outperforming existing on-device models and competitive with server-side models. Generates/edits 1024x1024 images in <1s on Xiaomi 14 smartphone.

Conclusion: DreamLite is the first unified on-device diffusion model supporting both image generation and editing, demonstrating efficient multimodal capabilities in a compact architecture suitable for mobile deployment.

Abstract: Diffusion models have made significant progress in both text-to-image (T2I) generation and text-guided image editing. However, these models are typically built with billions of parameters, leading to high latency and increased deployment challenges. While on-device diffusion models improve efficiency, they largely focus on T2I generation and lack support for image editing. In this paper, we propose DreamLite, a compact unified on-device diffusion model (0.39B) that supports both T2I generation and text-guided image editing within a single network. DreamLite is built on a pruned mobile U-Net backbone and unifies conditioning through in-context spatial concatenation in the latent space. It concatenates images horizontally as input, using a (target | blank) configuration for generation tasks and (target | source) for editing tasks. To stabilize the training of this compact model, we introduce a task-progressive joint pretraining strategy that sequentially targets T2I, editing, and joint tasks. After high-quality SFT and reinforcement learning, DreamLite achieves GenEval (0.72) for image generation and ImgEdit (4.11) for image editing, outperforming existing on-device models and remaining competitive with several server-side models. By employing step distillation, we further reduce denoising processing to just 4 steps, enabling our DreamLite could generate or edit a 1024 x 1024 image in less than 1s on a Xiaomi 14 smartphone. To the best of our knowledge, DreamLite is the first unified on-device diffusion model that supports both image generation and image editing.

Method: Uses hierarchical transformer architecture for global context, formulates flow initialization as optimal transport problem to get robust initial flow with occlusion/confidence maps, then performs guided refinement propagating reliable estimates from high-confidence to low-confidence regions.

Result: Achieves state-of-the-art results on Sintel and KITTI benchmarks, and establishes new SOTA cross-dataset zero-shot generalization performance on Sintel, Spring, and LayeredFlow datasets.

Conclusion: FlowIt demonstrates superior performance in optical flow estimation, particularly for large displacements and challenging scenarios, through its transformer-based architecture and optimal transport formulation with guided refinement.

Abstract: We present FlowIt, a novel architecture for optical flow estimation designed to robustly handle large pixel displacements. At its core, FlowIt leverages a hierarchical transformer architecture that captures extensive global context, enabling the model to effectively model long-range correspondences. To overcome the limitations of localized matching, we formulate the flow initialization as an optimal transport problem. This formulation yields a highly robust initial flow field, alongside explicitly derived occlusion and confidence maps. These cues are then seamlessly integrated into a guided refinement stage, where the network actively propagates reliable motion estimates from high-confidence regions into ambiguous, low-confidence areas. Extensive experiments across the Sintel, KITTI, Spring, and LayeredFlow datasets validate the efficacy of our approach. FlowIt achieves state-of-the-art results on the competitive Sintel and KITTI benchmarks, while simultaneously establishing new state-of-the-art cross-dataset zero-shot generalization performance on Sintel, Spring, and LayeredFlow.

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 2509.25848: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.25848&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Developed a lightweight, back-mounted multi-camera rig synchronized with a VR headset for unconstrained mobility. Created an ego-exo tracking pipeline for 3D ground-truth annotation of hands and objects in diverse environments.

Result: Created SHOW3D, the first large-scale dataset with 3D annotations of hands interacting with objects in diverse real-world environments including outdoor settings. Validated approach through experiments on downstream tasks.

Conclusion: The system significantly reduces the trade-off between environmental realism and 3D annotation accuracy, enabling better generalization of hand-object interaction models to real-world scenarios.

Abstract: Accurate 3D understanding of human hands and objects during manipulation remains a significant challenge for egocentric computer vision. Existing hand-object interaction datasets are predominantly captured in controlled studio settings, which limits both environmental diversity and the ability of models trained on such data to generalize to real-world scenarios. To address this challenge, we introduce a novel marker-less multi-camera system that allows for nearly unconstrained mobility in genuinely in-the-wild conditions, while still having the ability to generate precise 3D annotations of hands and objects. The capture system consists of a lightweight, back-mounted, multi-camera rig that is synchronized and calibrated with a user-worn VR headset. For 3D ground-truth annotation of hands and objects, we develop an ego-exo tracking pipeline and rigorously evaluate its quality. Finally, we present SHOW3D, the first large-scale dataset with 3D annotations that show hands interacting with objects in diverse real-world environments, including outdoor settings. Our approach significantly reduces the fundamental trade-off between environmental realism and accuracy of 3D annotations, which we validate with experiments on several downstream tasks. show3d-dataset.github.io

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 2510.02001: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.02001&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Uses diffusion models for controllable image generation with Direct Preference Optimization for control alignment, curriculum-based hard sample mining, and multi-stage quality filtering to maintain 3D label-image correspondence.

Result: Generated 500,000+ high-quality synthetic samples with 76% improvement in image-quality metrics over rendering-based datasets. Models trained on PoseDreamer perform comparably or better than those trained on real or traditional synthetic data.

Conclusion: Generated data via diffusion models offers a promising third path for 3D human mesh estimation, providing scalable, high-quality synthetic datasets that complement existing real and synthetic data sources.

Abstract: Acquiring labeled datasets for 3D human mesh estimation is challenging due to depth ambiguities and the inherent difficulty of annotating 3D geometry from monocular images. Existing datasets are either real, with manually annotated 3D geometry and limited scale, or synthetic, rendered from 3D engines that provide precise labels but suffer from limited photorealism, low diversity, and high production costs. In this work, we explore a third path: generated data. We introduce PoseDreamer, a novel pipeline that leverages diffusion models to generate large-scale synthetic datasets with 3D mesh annotations. Our approach combines controllable image generation with Direct Preference Optimization for control alignment, curriculum-based hard sample mining, and multi-stage quality filtering. Together, these components naturally maintain correspondence between 3D labels and generated images, while prioritizing challenging samples to maximize dataset utility. Using PoseDreamer, we generate more than 500,000 high-quality synthetic samples, achieving a 76% improvement in image-quality metrics compared to rendering-based datasets. Models trained on PoseDreamer achieve performance comparable to or superior to those trained on real-world and traditional synthetic datasets. In addition, combining PoseDreamer with synthetic datasets results in better performance than combining real-world and synthetic datasets, demonstrating the complementary nature of our dataset. We will release the full dataset and generation code.

Method: Unable to determine method due to API rate limiting preventing access to paper details

Result: Unable to determine results due to API rate limiting preventing access to paper details

Conclusion: Unable to draw conclusions due to API rate limiting preventing access to paper details

Abstract: Failed to fetch summary for 2510.08553: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.08553&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: 1) Consolidate existing datasets and collect new motion-capture data for bimanual interactions; 2) Develop decoupled annotation pipeline extracting motion features (contact events, finger flexion) then using LLMs for semantic descriptions; 3) Benchmark diffusion and autoregressive models with various conditioning modes; 4) Propose hand-focused evaluation metrics.

Result: High-quality dexterous motion generation demonstrated, with clear scaling trends showing larger models on larger, higher-quality datasets produce more semantically coherent bimanual motion. New dataset released for research.

Conclusion: HandX provides comprehensive framework for bimanual hand motion synthesis, addressing data scarcity and annotation challenges through LLM-assisted pipeline, enabling realistic dexterous motion generation with measurable improvements through scaling.

Abstract: Synthesizing human motion has advanced rapidly, yet realistic hand motion and bimanual interaction remain underexplored. Whole-body models often miss the fine-grained cues that drive dexterous behavior, finger articulation, contact timing, and inter-hand coordination, and existing resources lack high-fidelity bimanual sequences that capture nuanced finger dynamics and collaboration. To fill this gap, we present HandX, a unified foundation spanning data, annotation, and evaluation. We consolidate and filter existing datasets for quality, and collect a new motion-capture dataset targeting underrepresented bimanual interactions with detailed finger dynamics. For scalable annotation, we introduce a decoupled strategy that extracts representative motion features, e.g., contact events and finger flexion, and then leverages reasoning from large language models to produce fine-grained, semantically rich descriptions aligned with these features. Building on the resulting data and annotations, we benchmark diffusion and autoregressive models with versatile conditioning modes. Experiments demonstrate high-quality dexterous motion generation, supported by our newly proposed hand-focused metrics. We further observe clear scaling trends: larger models trained on larger, higher-quality datasets produce more semantically coherent bimanual motion. Our dataset is released to support future research.

Method: Developed Gen-Searcher as a search-augmented image generation agent with multi-hop reasoning capabilities. Created two datasets (Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k) and KnowGen benchmark. Trained using supervised fine-tuning followed by agentic reinforcement learning with dual reward feedback combining text-based and image-based rewards.

Result: Gen-Searcher substantially improves performance, boosting Qwen-Image by around 16 points on KnowGen benchmark and 15 points on WISE benchmark, demonstrating significant gains in search-grounded image generation.

Conclusion: Gen-Searcher represents the first search-augmented image generation agent that effectively addresses knowledge limitations through multi-hop reasoning and search, providing a foundation for more knowledgeable and current image generation systems.

Abstract: Recent image generation models have shown strong capabilities in generating high-fidelity and photorealistic images. However, they are fundamentally constrained by frozen internal knowledge, thus often failing on real-world scenarios that are knowledge-intensive or require up-to-date information. In this paper, we present Gen-Searcher, as the first attempt to train a search-augmented image generation agent, which performs multi-hop reasoning and search to collect the textual knowledge and reference images needed for grounded generation. To achieve this, we construct a tailored data pipeline and curate two high-quality datasets, Gen-Searcher-SFT-10k and Gen-Searcher-RL-6k, containing diverse search-intensive prompts and corresponding ground-truth synthesis images. We further introduce KnowGen, a comprehensive benchmark that explicitly requires search-grounded external knowledge for image generation and evaluates models from multiple dimensions. Based on these resources, we train Gen-Searcher with SFT followed by agentic reinforcement learning with dual reward feedback, which combines text-based and image-based rewards to provide more stable and informative learning signals for GRPO training. Experiments show that Gen-Searcher brings substantial gains, improving Qwen-Image by around 16 points on KnowGen and 15 points on WISE. We hope this work can serve as an open foundation for search agents in image generation, and we fully open-source our data, models, and code.

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2510.13044: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.13044&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: Paper analysis not possible due to technical limitations in accessing content

Abstract: Failed to fetch summary for 2511.11483: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.11483&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 draw conclusions about the paper due to technical fetching error

Abstract: Failed to fetch summary for 2511.15090: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.15090&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2511.16719: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.16719&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 2511.07738: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.07738&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content is unavailable due to HTTP 429 error from arXiv API

Result: Cannot determine results as paper content is unavailable due to HTTP 429 error from arXiv API

Conclusion: Cannot determine conclusion as paper content is unavailable due to HTTP 429 error from arXiv API

Abstract: Failed to fetch summary for 2511.19413: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.19413&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 2511.21428: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.21428&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 2511.22442: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.22442&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 determine conclusion due to access error

Abstract: Failed to fetch summary for 2405.13580: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2405.13580&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 2511.23342: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.23342&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 draw conclusions as paper content could not be retrieved

Abstract: Failed to fetch summary for 2406.10045: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2406.10045&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 draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2512.08503: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.08503&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2407.07603: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2407.07603&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract

Result: Cannot determine results without access to the paper abstract

Conclusion: Cannot determine conclusion without access to the paper abstract

Abstract: Failed to fetch summary for 2512.10932: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.10932&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 draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2409.11018: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2409.11018&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 2512.22065: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.22065&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2409.20283: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2409.20283&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as paper content is unavailable due to HTTP 429 error

Result: No results available - paper retrieval failed due to rate limiting

Conclusion: Cannot provide analysis due to technical limitations in accessing the paper

Abstract: Failed to fetch summary for 2411.06197: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.06197&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract.

Result: Cannot determine results without access to the paper abstract.

Conclusion: Cannot draw conclusions without access to the paper abstract.

Abstract: Failed to fetch summary for 2412.05386: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.05386&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2412.14015: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.14015&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot draw conclusions without access to the paper content.

Abstract: Failed to fetch summary for 2602.04361: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.04361&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 determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2412.16906: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2412.16906&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper’s content.

Result: Cannot determine results without access to the paper’s content.

Conclusion: Cannot draw conclusions without access to the paper’s content.

Abstract: Failed to fetch summary for 2501.16997: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2501.16997&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 2503.04522: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.04522&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 without access to paper content

Abstract: Failed to fetch summary for 2503.09750: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.09750&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 determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2602.08961: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.08961&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 draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2503.21262: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.21262&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content

Result: Cannot determine results without access to the paper content

Conclusion: Cannot draw conclusions without access to the paper content

Abstract: Failed to fetch summary for 2602.11448: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.11448&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 2504.17817: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.17817&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 2602.16898: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.16898&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper retrieval

Result: Unable to determine results due to failed paper retrieval

Conclusion: Unable to determine conclusion due to failed paper retrieval

Abstract: Failed to fetch summary for 2602.22601: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22601&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 determine conclusion due to access error

Abstract: Failed to fetch summary for 2505.12772: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.12772&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 determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2602.19190: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.19190&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 determine conclusion due to access error

Abstract: Failed to fetch summary for 2505.21545: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.21545&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed summary fetch

Result: Unable to determine results due to failed summary fetch

Conclusion: Unable to determine conclusion due to failed summary fetch

Abstract: Failed to fetch summary for 2505.24862: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.24862&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 2506.03388: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.03388&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2602.21655: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.21655&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is unavailable

Result: Cannot determine results as abstract is unavailable

Conclusion: Cannot draw conclusion due to missing abstract data

Abstract: Failed to fetch summary for 2506.05207: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.05207&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 2602.23153: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23153&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 2603.10281: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.10281&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 without access to paper content

Abstract: Failed to fetch summary for 2506.21742: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.21742&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 draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2603.01305: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01305&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: N/A - Paper content unavailable

Result: N/A - Paper content unavailable

Conclusion: N/A - Paper content unavailable

Abstract: Failed to fetch summary for 2506.23323: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.23323&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to API access restrictions

Result: No results available - paper content inaccessible

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

Abstract: Failed to fetch summary for 2506.23555: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.23555&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method as paper content is unavailable due to HTTP 429 error

Result: No results available - paper content retrieval failed due to rate limiting

Conclusion: Cannot provide analysis due to technical limitations in accessing the paper

Abstract: Failed to fetch summary for 2603.02190: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.02190&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper retrieval

Result: Unable to determine results due to failed paper retrieval

Conclusion: Unable to determine conclusion due to failed paper retrieval

Abstract: Failed to fetch summary for 2507.04017: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.04017&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 2507.04990: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.04990&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions about paper content

Abstract: Failed to fetch summary for 2507.06233: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.06233&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - paper content inaccessible

Result: No results available - unable to fetch paper summary

Conclusion: Cannot draw conclusions about paper relevance without access to abstract/content

Abstract: Failed to fetch summary for 2507.10195: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.10195&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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 2603.12057: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.12057&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 2507.16279: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.16279&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to data fetch failure

Result: Unable to determine results due to data fetch failure

Conclusion: Unable to determine conclusion due to data fetch failure

Abstract: Failed to fetch summary for 2508.03100: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.03100&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2508.13669: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.13669&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 2509.02028: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.02028&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access limitations

Result: Unable to determine results due to access limitations

Conclusion: Unable to determine conclusion due to access limitations

Abstract: Failed to fetch summary for 2509.05970: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.05970&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 draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2509.07704: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.07704&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 2509.11334: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.11334&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 2509.12554: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.12554&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 2509.18387: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.18387&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 determine conclusion due to inability to access paper content

Abstract: Failed to fetch summary for 2509.21309: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21309&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error fetching paper information

Result: Unable to determine results due to technical error fetching paper information

Conclusion: Unable to determine conclusion due to technical error fetching paper information

Abstract: Failed to fetch summary for 2509.24968: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.24968&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 draw conclusion without access to paper content

Abstract: Failed to fetch summary for 2603.25008: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.25008&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available due to API rate limiting preventing access to the paper abstract

Result: No results information available as the paper content could not be retrieved

Conclusion: Unable to provide analysis due to technical limitations in accessing the paper content

Abstract: Failed to fetch summary for 2407.19097: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2407.19097&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 2510.01014: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.01014&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Introduces Hybrid Memory paradigm requiring models to act as archivists for static backgrounds and trackers for dynamic subjects. Constructs HM-World dataset with 59K clips featuring decoupled camera/subject trajectories and exit-entry events. Proposes HyDRA architecture that compresses memory into tokens with spatiotemporal relevance-driven retrieval.

Result: HyDRA significantly outperforms state-of-the-art approaches in both dynamic subject consistency and overall generation quality on HM-World benchmark.

Conclusion: Hybrid Memory paradigm addresses critical limitation in video world models, enabling better handling of dynamic subjects during occlusions through specialized memory architecture and comprehensive dataset.

Abstract: Video world models have shown immense potential in simulating the physical world, yet existing memory mechanisms primarily treat environments as static canvases. When dynamic subjects hide out of sight and later re-emerge, current methods often struggle, leading to frozen, distorted, or vanishing subjects. To address this, we introduce Hybrid Memory, a novel paradigm requiring models to simultaneously act as precise archivists for static backgrounds and vigilant trackers for dynamic subjects, ensuring motion continuity during out-of-view intervals. To facilitate research in this direction, we construct HM-World, the first large-scale video dataset dedicated to hybrid memory. It features 59K high-fidelity clips with decoupled camera and subject trajectories, encompassing 17 diverse scenes, 49 distinct subjects, and meticulously designed exit-entry events to rigorously evaluate hybrid coherence. Furthermore, we propose HyDRA, a specialized memory architecture that compresses memory into tokens and utilizes a spatiotemporal relevance-driven retrieval mechanism. By selectively attending to relevant motion cues, HyDRA effectively preserves the identity and motion of hidden subjects. Extensive experiments on HM-World demonstrate that our method significantly outperforms state-of-the-art approaches in both dynamic subject consistency and overall generation quality. Code is publicly available at https://github.com/H-EmbodVis/HyDRA.

Method: Cannot determine method due to failed paper fetch

Result: Cannot determine results due to failed paper fetch

Conclusion: Cannot draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2510.08673: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.08673&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Constructed large-scale InstructScene dataset (100k scenes with diverse driving instructions and trajectories). Proposed Vega model with: 1) autoregressive paradigm for visual inputs and language instructions, 2) diffusion paradigm for future predictions and trajectory generation, 3) joint attention for modality interactions, and 4) individual projection layers for different modalities.

Result: Extensive experiments show superior planning performance and strong instruction-following abilities, enabling more intelligent and personalized driving systems.

Conclusion: Vega paves the way for instruction-based generation and planning in autonomous driving by effectively integrating vision, language, world modeling, and action generation in a unified framework.

Abstract: Vision-language-action models have reshaped autonomous driving to incorporate languages into the decision-making process. However, most existing pipelines only utilize the language modality for scene descriptions or reasoning and lack the flexibility to follow diverse user instructions for personalized driving. To address this, we first construct a large-scale driving dataset (InstructScene) containing around 100,000 scenes annotated with diverse driving instructions with the corresponding trajectories. We then propose a unified Vision-Language-World-Action model, Vega, for instruction-based generation and planning. We employ the autoregressive paradigm to process visual inputs (vision) and language instructions (language) and the diffusion paradigm to generate future predictions (world modeling) and trajectories (action). We perform joint attention to enable interactions between the modalities and use individual projection layers for different modalities for more capabilities. Extensive experiments demonstrate that our method not only achieves superior planning performance but also exhibits strong instruction-following abilities, paving the way for more intelligent and personalized driving systems.

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 2510.11417: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.11417&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 2603.26285: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.26285&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 2510.12901: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.12901&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 determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2603.26425: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.26425&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 determine conclusion due to inability to access paper content

Abstract: Failed to fetch summary for 2510.14255: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.14255&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 2510.20212: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.20212&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 2510.26794: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.26794&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 2511.00956: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.00956&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method unknown - arXiv API rate limiting prevented content access

Result: No results available - HTTP 429 error indicates too many requests to arXiv API

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

Abstract: Failed to fetch summary for 2511.01250: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.01250&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 draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2511.04520: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.04520&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 2511.14019: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.14019&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 2511.17209: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.17209&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine paper method due to access error

Result: Unable to determine paper results due to access error

Conclusion: Unable to determine paper conclusion due to access error

Abstract: Failed to fetch summary for 2511.18174: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.18174&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 draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2511.18471: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.18471&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to draw conclusions due to failed paper fetch

Abstract: Failed to fetch summary for 2511.18600: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.18600&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 without access to paper content

Abstract: Failed to fetch summary for 2511.21606: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.21606&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content.

Result: Cannot determine results without access to the paper content.

Conclusion: Cannot determine conclusion without access to the paper content.

Abstract: Failed to fetch summary for 2511.22264: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.22264&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as the paper content could not be retrieved due to API rate limiting.

Result: Unable to determine results as the paper content could not be retrieved due to API rate limiting.

Conclusion: Unable to determine conclusion as the paper content could not be retrieved due to API rate limiting.

Abstract: Failed to fetch summary for 2511.18151: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.18151&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 2511.22950: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.22950&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 2512.02172: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.02172&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: Paper analysis cannot be completed due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2511.23170: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.23170&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 determine conclusion due to access error

Abstract: Failed to fetch summary for 2601.01095: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.01095&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as paper content could not be retrieved

Result: Cannot determine results as paper content could not be retrieved

Conclusion: Cannot determine conclusion as paper content could not be retrieved

Abstract: Failed to fetch summary for 2511.23227: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.23227&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 2601.16771: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.16771&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 2512.00255: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.00255&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 2512.01342: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.01342&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 2512.02369: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.02369&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 determine conclusion due to inability to access paper content

Abstract: Failed to fetch summary for 2512.03350: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.03350&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 the paper due to access limitations

Abstract: Failed to fetch summary for 2512.03619: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.03619&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot analyze method without access to paper content

Result: No results available due to technical access issues

Conclusion: Paper analysis impossible due to API rate limiting preventing content retrieval

Abstract: Failed to fetch summary for 2512.05422: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.05422&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 draw conclusions without access to paper content.

Abstract: Failed to fetch summary for 2512.05597: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.05597&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper content. The paper ID suggests it’s from December 2024, but no technical details are available.

Result: Cannot determine results without access to the paper content. The analysis is limited by the API rate limiting issue.

Conclusion: Cannot draw conclusions about the paper’s content due to technical limitations in accessing the abstract and full paper details.

Abstract: Failed to fetch summary for 2512.10652: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.10652&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access limitations

Result: Unable to determine results due to access limitations

Conclusion: Unable to draw conclusions due to access limitations

Abstract: Failed to fetch summary for 2512.10950: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.10950&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine paper method due to technical error in fetching abstract

Result: Unable to determine paper results due to technical error in fetching abstract

Conclusion: Unable to determine paper conclusion due to technical error in fetching abstract

Abstract: Failed to fetch summary for 2512.12378: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.12378&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract is unavailable

Result: Cannot determine results as abstract is unavailable

Conclusion: Cannot draw conclusions about paper content due to data unavailability

Abstract: Failed to fetch summary for 2512.13680: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13680&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract

Result: Cannot determine results without access to the paper abstract

Conclusion: Cannot draw conclusions without access to the paper abstract

Abstract: Failed to fetch summary for 2512.13689: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13689&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed paper fetch

Result: Unable to determine results due to failed paper fetch

Conclusion: Unable to determine conclusion due to failed paper fetch

Abstract: Failed to fetch summary for 2512.13874: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13874&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 2512.13953: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.13953&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2512.14177: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.14177&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to missing paper content

Result: Unable to determine results due to missing paper content

Conclusion: Unable to determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2512.15508: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.15508&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 the paper due to access limitations

Abstract: Failed to fetch summary for 2512.16727: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.16727&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 2512.19693: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.19693&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to determine conclusion due to failed API request

Abstract: Failed to fetch summary for 2512.20770: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.20770&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 determine conclusion due to access error

Abstract: Failed to fetch summary for 2512.21643: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.21643&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 draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2601.03596: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.03596&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 2601.05138: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.05138&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to the paper abstract

Result: Cannot determine results without access to the paper abstract

Conclusion: Cannot determine conclusion without access to the paper abstract

Abstract: Failed to fetch summary for 2601.11396: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.11396&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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

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

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

Abstract: Failed to fetch summary for 2601.14959: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.14959&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 2601.15288: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.15288&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 2601.17354: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17354&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to access restrictions

Result: Unable to determine results due to access restrictions

Conclusion: Unable to determine conclusion due to access restrictions

Abstract: Failed to fetch summary for 2601.17470: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.17470&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in fetching paper content

Result: Unable to determine results due to technical error in fetching paper content

Conclusion: Unable to draw conclusions due to technical error in fetching paper content

Abstract: Failed to fetch summary for 2601.19582: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.19582&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 2601.20331: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.20331&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 2601.22150: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.22150&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

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 2601.22674: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.22674&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 draw conclusions without access to paper content

Abstract: Failed to fetch summary for 2602.02914: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.02914&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing paper content

Result: Cannot determine results due to missing paper content

Conclusion: Cannot determine conclusion due to missing paper content

Abstract: Failed to fetch summary for 2602.05551: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.05551&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 2602.07775: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.07775&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper content. The paper ID format suggests it’s a recent submission (February 2025).

Result: Cannot determine results without access to paper content. The HTTP 429 error indicates the arXiv API is rate limiting requests.

Conclusion: Cannot draw conclusions without access to paper content. The paper exists in the arXiv system but details are currently inaccessible due to API rate limits.

Abstract: Failed to fetch summary for 2602.12957: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.12957&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 determine conclusion as paper content is unavailable

Abstract: Failed to fetch summary for 2602.18792: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.18792&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 2602.18845: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.18845&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 2602.19575: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.19575&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 determine conclusion due to access error

Abstract: Failed to fetch summary for 2602.22419: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22419&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 determine conclusion due to inability to access paper content

Abstract: Failed to fetch summary for 2602.23165: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23165&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to failed API request

Result: Unable to determine results due to failed API request

Conclusion: Unable to draw conclusions due to failed API request

Abstract: Failed to fetch summary for 2603.00175: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00175&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: Unable to analyze paper due to technical limitations in accessing arXiv data.

Abstract: Failed to fetch summary for 2603.00667: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.00667&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to data retrieval failure

Result: Unable to determine results due to data retrieval failure

Conclusion: Unable to draw conclusions due to data retrieval failure

Abstract: Failed to fetch summary for 2603.01506: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.01506&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to paper fetch failure

Result: Unable to determine results due to paper fetch failure

Conclusion: Unable to determine conclusion due to paper fetch failure

Abstract: Failed to fetch summary for 2603.07619: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2603.07619&sortBy=relevance&sortOrder=descending&start=0&max_results=100)