Method: Multimodal cascaded system leveraging per-speaker visual streams from 360° video and single-channel audio, with enhanced audio-visual pretrained models for Active Speaker Detection, Audio-Visual Target Speech Extraction, and Audio-Visual Speech Recognition incorporating Whisper and LLM techniques

Result: Best single system achieves 32.44% Speaker WER; ROVER fusion reduces to 31.40%; LLM-based zero-shot conversational clustering achieves perfect F1 score of 1.0, yielding final JACER of 15.70%

Conclusion: The proposed multimodal approach effectively handles highly overlapping multi-conversation scenarios through audio-visual integration and LLM enhancement

Abstract: This report details our submission to the CHiME-9 MCoRec Challenge on recognizing and clustering multiple concurrent natural conversations within indoor social settings. Unlike conventional meetings centered on a single shared topic, this scenario contains multiple parallel dialogues–up to eight speakers across up to four simultaneous conversations–with a speech overlap rate exceeding 90%. To tackle this, we propose a multimodal cascaded system that leverages per-speaker visual streams extracted from synchronized 360 degree video together with single-channel audio. Our system improves three components of the pipeline by leveraging enhanced audio-visual pretrained models: Active Speaker Detection (ASD), Audio-Visual Target Speech Extraction (AVTSE), and Audio-Visual Speech Recognition (AVSR). The AVSR module further incorporates Whisper and LLM techniques to boost transcription accuracy. Our best single cascaded system achieves a Speaker Word Error Rate (WER) of 32.44% on the development set. By further applying ROVER to fuse outputs from diverse front-end and back-end variants, we reduce Speaker WER to 31.40%. Notably, our LLM-based zero-shot conversational clustering achieves a speaker clustering F1 score of 1.0, yielding a final Joint ASR-Clustering Error Rate (JACER) of 15.70%.

Method: Proposes an audio-visual learner that disentangles three distinct cues inspired by human perception: speaker information, acoustic synchronization, and semantic synchronization. Designs a dedicated interaction module to effectively integrate these cues for reliable speaker extraction guidance.

Result: Extensive experiments demonstrate strong robustness under various visual degradations and clear superiority over existing methods.

Conclusion: The proposed approach enhances robustness against impaired visual inputs without needing degraded videos during training, offering a more practical solution for real-world audio-visual speaker extraction.

Abstract: Audio-visual speaker extraction has attracted increasing attention, as it removes the need for pre-registered speech and leverages the visual modality as a complement to audio. Although existing methods have achieved impressive performance, the issue of degraded visual inputs has received relatively little attention, despite being common in real-world scenarios. Previous attempts to address this problem have mainly involved training with degraded visual data. However, visual degradation can occur in many unpredictable ways, making it impractical to simulate all possible cases during training. In this paper, we aim to enhance the robustness of audio-visual speaker extraction against impaired visual inputs without relying on degraded videos during training. Inspired by observations from human perceptual mechanisms, we propose an audio-visual learner that disentangles speaker information, acoustic synchronisation, and semantic synchronisation as distinct cues. Furthermore, we design a dedicated interaction module that effectively integrates these cues to provide a reliable guidance signal for speaker extraction. Extensive experiments demonstrate the strong robustness of the proposed model under various visual degradations and its clear superiority over existing methods.

Method: Uses LLM to generate high-fidelity audio captions for better text-audio semantic alignment, then applies Group Relative Policy Optimization (GRPO) reinforcement learning to fine-tune diffusion transformer-based T2A models with systematic experimentation of reward functions (CLAP, KL, FAD combinations).

Result: GRPO-based fine-tuning yields substantial gains in synthesis fidelity and prompt adherence, with systematic analysis identifying key drivers of effective RL in audio synthesis and how reward design impacts final audio quality.

Conclusion: Reinforcement learning with GRPO and LLM-enhanced captioning effectively improves text-to-audio generation quality and alignment, providing insights into reward function design for audio synthesis.

Abstract: Text-to-audio (T2A) generation has advanced considerably in recent years, yet existing methods continue to face challenges in accurately rendering complex text prompts, particularly those involving intricate audio effects, and achieving precise text-audio alignment. While prior approaches have explored data augmentation, explicit timing conditioning, and reinforcement learning, overall synthesis quality remains constrained. In this work, we experiment with reinforcement learning to further enhance T2A generation quality, building on diffusion transformer (DiT)-based architectures. Our method first employs a large language model (LLM) to generate high-fidelity, richly detailed audio captions, substantially improving text-audio semantic alignment, especially for ambiguous or underspecified prompts. We then apply Group Relative Policy Optimization (GRPO), a recently introduced reinforcement learning algorithm, to fine-tune the T2A model. Through systematic experimentation with diverse reward functions (including CLAP, KL, FAD, and their combinations), we identify the key drivers of effective RL in audio synthesis and analyze how reward design impacts final audio quality. Experimental results demonstrate that GRPO-based fine-tuning yield substantial gains in synthesis fidelity and prompt adherence.

Method: Leverages dynamic sliding-window attention module to automatically construct graph representations of documents, then uses Graph Attention Networks (GATs) trained on these learned graphs for document classification and extractive summarization tasks.

Result: GATs trained on the learned graphs achieve competitive results on document classification with lower computational resources than previous approaches. Exploratory evaluation shows potential for extractive summarization but also current limitations.

Conclusion: The proposed data-driven graph construction method using dynamic sliding-window attention provides effective document representations that enable competitive performance on document classification with computational efficiency, while showing promise for extractive summarization applications.

Abstract: This paper proposes a data-driven method to automatically construct graph-based document representations. Building upon the recent work of Bugueño and de Melo (2025), we leverage the dynamic sliding-window attention module to effectively capture local and mid-range semantic dependencies between sentences, as well as structural relations within documents. Graph Attention Networks (GATs) trained on our learned graphs achieve competitive results on document classification while requiring lower computational resources than previous approaches. We further present an exploratory evaluation of the proposed graph construction method for extractive document summarization, highlighting both its potential and current limitations. The implementation of this project can be found on GitHub.

Method: Pre-trained an in-house BERT on clinical data, fine-tuned for NER, then developed a supervised Noise Removal model that analyzes probability sequences using features like Probability Density Map (PDM) to capture Semantic-Pull effects in transformer embeddings and classify predictions as weak/strong.

Result: The Noise Removal model reduced False Positives across various clinical NER models by 50% to 90%, significantly improving precision while maintaining recall performance.

Conclusion: The proposed noise removal approach effectively addresses precision limitations in clinical NER by leveraging sophisticated analysis of transformer probability distributions, enabling reliable clinical entity extraction with reduced false positives.

Abstract: Precision is of utmost importance in the realm of clinical entity extraction from clinical notes and reports. Encoder Models fine-tuned for Named Entity Recognition (NER) are an efficient choice for this purpose, as they don’t hallucinate. We pre-trained an in-house BERT over clinical data and then fine-tuned it for NER. These models performed well on recall but could not close upon the high precision range, needed for clinical models. To address this challenge, we developed a Noise Removal model that refines the output of NER. The NER model assigns token-level entity tags along with probability scores for each token. Our Noise Removal (NR) model then analyzes these probability sequences and classifies predictions as either weak or strong. A naïve approach might involve filtering predictions based on low probability values; however, this method is unreliable. Owing to the characteristics of the SoftMax function, Transformer based architectures often assign disproportionately high confidence scores even to uncertain or weak predictions, making simple thresholding ineffective. To address this issue, we adopted a supervised modeling strategy in which the NR model leverages advanced features such as the Probability Density Map (PDM). The PDM captures the Semantic-Pull effect observed within Transformer embeddings, an effect that manifests in the probability distributions of NER class predictions across token sequences. This approach enables the model to classify predictions as weak or strong with significantly improved accuracy. With these NR models we were able to reduce False Positives across various clinical NER models by 50% to 90%.

Method: Rigorous evaluation across nine frontier models and five benchmark datasets spanning advice, moral judgment, and debate contexts, with systematic testing of personalization’s impact on affective and epistemic alignment, including robustness tests to isolate effects.

Result: Personalization generally increases affective alignment (emotional validation, hedging/deference) but affects epistemic alignment differently based on context and role: in advice roles, personalization strengthens epistemic independence (models challenge user presuppositions); in social peer roles, personalization decreases epistemic independence, with extensively personalized challenges causing LLMs to abandon positions at higher rates.

Conclusion: The work provides measurement frameworks for evaluating personalized AI systems, demonstrates necessity of role-sensitive evaluation, and establishes a novel benchmark for assessing goal alignment in LLMs with personalization.

Abstract: Large Language Models (LLMs) are prone to sycophantic behavior, uncritically conforming to user beliefs. As models increasingly condition responses on user-specific context (personality traits, preferences, conversation history), they gain information to tailor agreement more effectively. Understanding how personalization modulates sycophancy is critical, yet systematic evaluation across models and contexts remains limited. We present a rigorous evaluation of personalization’s impact on LLM sycophancy across nine frontier models and five benchmark datasets spanning advice, moral judgment, and debate contexts. We find that personalization generally increases affective alignment (emotional validation, hedging/deference), but affects epistemic alignment (belief adoption, position stability, resistance to influence) with context-dependent role modulation. When the LLM’s role is to give advice, personalization strengthens epistemic independence (models challenge user presuppositions). When its role is that of a social peer, personalization decreases epistemic independence. In this role, extensively personalized user challenges causing LLMs to abandon their position at significantly higher rates. Robustness tests confirm that the effects are driven by personalized conditioning, not by additional input tokens per se or demographic information alone. Our work provides measurement frameworks for evaluating personalized AI systems, demonstrates the necessity of role-sensitive evaluation, and establishes a novel benchmark to assess goal alignment.

Method: TAB-PO augments DPO with token-weighted, reference-adjusted advantages that prioritize high-value semantic tokens, and a conditional token-level barrier that balances SFT-anchored likelihood and preference-driven separation.

Result: TAB-PO achieves ~4% relative improvement in micro-F1 over SFT on medical communication annotation and consistently outperforms recent preference-optimization baselines.

Conclusion: TAB-PO effectively addresses DPO’s limitations in token-critical structured prediction by focusing learning on important semantic tokens while maintaining proper likelihood separation.

Abstract: Direct Preference Optimization is an offline post-SFT method for aligning language models from preference pairs, with strong results in instruction following and summarization. However, DPO’s sequence-level implicit reward can be brittle for token-critical structured prediction settings such as medical annotation, which often exhibit (i) low-separation preference pairs, where chosen and rejected completions differ by minimal edit distance (often 1-3 tokens), and (ii) token-importance skew, where sparse semantic tokens (hierarchical labels and evidence Spans) carry disproportionate task importance relative to high-frequency structural tokens (JSON scaffolding). In this regime, standard DPO suffers from margin collapse (insufficient log-probability separation between near-identical preferences), likelihood squeezing (the margin objective shifts the absolute likelihoods of both completions together), and gradient dilution, where uniform sequence-level weighting diffuses learning signal across shared scaffolding while rare, confusable label tokens receive weak, noisy updates. We introduce Token-Adaptive Barrier Preference Optimization (TAB-PO), which augments DPO with token-weighted, reference-adjusted advantages that prioritize high-value semantic tokens, and a conditional token-level barrier that regularizes under-confident tokens balancing SFT-anchored likelihood and preference-driven separation in low-separation, importance-skewed regimes. We evaluate TAB-PO on medical communication annotation, a task requiring joint prediction of hierarchical labels and evidence Spans from patient-provider messages. TAB-PO achieves a ~ 4% relative improvement in micro-F1 over SFT and consistently outperforms recent preference-optimization baselines.

Method: Proposes ActMem framework that transforms unstructured dialogue history into structured causal and semantic graphs. Uses counterfactual reasoning and commonsense completion to deduce implicit constraints and resolve conflicts between past states and current intentions.

Result: ActMem significantly outperforms state-of-the-art baselines in handling complex, memory-dependent tasks. Introduces ActMemEval dataset for evaluating agent reasoning capabilities in logic-driven scenarios.

Conclusion: ActMem enables more consistent and reliable intelligent assistants by integrating memory retrieval with active causal reasoning, moving beyond fact-retrieval to handle complex reasoning scenarios.

Abstract: Effective memory management is essential for large language model (LLM) agents handling long-term interactions. Current memory frameworks typically treat agents as passive “recorders” and retrieve information without understanding its deeper implications. They may fail in scenarios requiring conflict detection and complex decision-making. To bridge this critical gap, we propose a novel actionable memory framework called ActMem that integrates memory retrieval with active causal reasoning. ActMem transforms unstructured dialogue history into a structured causal and semantic graph. By leveraging counterfactual reasoning and commonsense completion, it enables agents to deduce implicit constraints and resolve potential conflicts between past states and current intentions. Furthermore, we introduce a comprehensive dataset ActMemEval to evaluate agent reasoning capabilities in logic-driven scenarios, moving beyond the fact-retrieval focus of existing memory benchmarks. Experiments demonstrate that ActMem significantly outperforms state-of-the-art baselines in handling complex, memory-dependent tasks, paving the way for more consistent and reliable intelligent assistants.

Method: Created a benchmark with 1,933 expert-annotated sentences from 752 secure messages, featuring three sub-tasks: Code Classification, Subcode Classification, and Evidence Extraction. Evaluated various LLMs under zero-shot and few-shot settings.

Result: Llama-3.1-70B led in evidence extraction (F1: 82.84%), Llama-3.3-70b-Instruct excelled in code classification (F1: 67.03%), and DeepSeek-R1-Distill-Qwen-32B performed best in subcode classification (F1: 48.25%). Few-shot prompting improved performance, while smaller models struggled.

Conclusion: Large instruction-tuned models perform well on EPPCMinerBen tasks, especially evidence extraction, while smaller models struggle with fine-grained reasoning. The benchmark supports future work on discourse-level understanding and patient-provider communication analysis.

Abstract: Effective communication in health care is critical for treatment outcomes and adherence. With patient-provider exchanges shifting to secure messaging, analyzing electronic patient-communication (EPPC) data is both essential and challenging. We introduce EPPCMinerBen, a benchmark for evaluating LLMs in detecting communication patterns and extracting insights from electronic patient-provider messages. EPPCMinerBen includes three sub-tasks: Code Classification, Subcode Classification, and Evidence Extraction. Using 1,933 expert annotated sentences from 752 secure messages of the patient portal at Yale New Haven Hospital, it evaluates LLMs on identifying communicative intent and supportive text. Benchmarks span various LLMs under zero-shot and few-shot settings, with data to be released via the NCI Cancer Data Service. Model performance varied across tasks and settings. Llama-3.1-70B led in evidence extraction (F1: 82.84%) and performed well in classification. Llama-3.3-70b-Instruct outperformed all models in code classification (F1: 67.03%). DeepSeek-R1-Distill-Qwen-32B excelled in subcode classification (F1: 48.25%), while sdoh-llama-3-70B showed consistent performance. Smaller models underperformed, especially in subcode classification (>30% F1). Few-shot prompting improved most tasks. Our results show that large, instruction-tuned models generally perform better in EPPCMinerBen tasks, particularly evidence extraction while smaller models struggle with fine-grained reasoning. EPPCMinerBen provides a benchmark for discourse-level understanding, supporting future work on model generalization and patient-provider communication analysis. Keywords: Electronic Patient-Provider Communication, Large language models, Data collection, Prompt engineering

Method: Proposes a multi-pass LLM post-processing architecture that alternates between Speaker Recognition and Word Recognition passes. Investigates four design choices: model selection, prompting strategy, pass ordering, and iteration depth using Qwen3-Next-80B on two French clinical datasets.

Result: Significant WDER reductions on suicide prevention conversations (p < 0.05, n=18), stable performance on awake neurosurgery consultations (n=10), zero output failures, and acceptable computational cost (RTF 0.32).

Conclusion: The multi-pass LLM architecture demonstrates feasibility for offline clinical deployment by improving transcription accuracy in French medical conversations while maintaining computational efficiency and reliability.

Abstract: Automatic speech recognition for French medical conversations remains challenging, with word error rates often exceeding 30% in spontaneous clinical speech. This study proposes a multi-pass LLM post-processing architecture alternating between Speaker Recognition and Word Recognition passes to improve transcription accuracy and speaker attribution. Ablation studies on two French clinical datasets (suicide prevention telephone counseling and preoperative awake neurosurgery consultations) investigate four design choices: model selection, prompting strategy, pass ordering, and iteration depth. Using Qwen3-Next-80B, Wilcoxon signed-rank tests confirm significant WDER reductions on suicide prevention conversations (p < 0.05, n=18), while maintaining stability on awake neurosurgery consultations (n=10), with zero output failures and acceptable computational cost (RTF 0.32), suggesting feasibility for offline clinical deployment.

Method: Proposes a magnitude-based criterion to identify Domain-Critical Dimensions in a training-free manner, and introduces Critical Dimension Steering which applies activation steering exclusively to these identified dimensions rather than all dimensions.

Result: Empirical results show that Critical Dimension Steering outperforms conventional whole-dimension steering in both domain adaptation and jailbreaking scenarios, demonstrating the effectiveness of targeting specific interpretable semantic detectors.

Conclusion: Extreme activation dimensions in LLMs are not just artifacts but serve as interpretable functional units for domain specialization, and targeted steering of these dimensions enables more effective model control for various applications.

Abstract: Large Language Models (LLMs) exhibit highly anisotropic internal representations, often characterized by massive activations, a phenomenon where a small subset of feature dimensions possesses magnitudes significantly larger than the rest. While prior works view these extreme dimensions primarily as artifacts to be managed, we propose a distinct perspective: these dimensions serve as intrinsic interpretable functional units arising from domain specialization. Specifically, we propose a simple magnitude-based criterion to identify Domain-Critical Dimensions in a training-free manner. Our analyses reveal that such dimensions behave as interpretable semantic detectors for symbolic/quantitative patterns or domain-specific terms. In addition, we introduce Critical Dimension Steering, which applies activation steering exclusively to the identified dimensions. Empirical results show that this approach outperforms conventional whole-dimension steering in domain adaptation and jailbreaking scenarios.

Method: Introduces special tokens that serve dual purposes: compressing low-entropy structural tokens (4-6x reduction) and acting as mode selectors to enable independent parallel generation of function names and arguments.

Result: Achieves 3-6x end-to-end speedup (up to 9.6x) with only +8.2% parallelization overhead. On Mobile Actions, ST-Qwen-0.5B outperforms Google’s FunctionGemma in accuracy and latency consistency. With quantization, achieves 61.2ms P50 latency enabling 16 Hz real-time control at 4B model scale.

Conclusion: SimpleTool bridges the gap between LLM function calling and latency-critical real-world deployment by exploiting structural redundancy and weak causal dependencies in function outputs for parallel generation.

Abstract: LLM-based function calling enables intelligent agents to interact with external tools and environments, yet autoregressive decoding imposes a fundamental latency bottleneck that limits real-time applications such as embodied intelligence, game AI, and interactive avatars (e.g., 10 Hz control frequency). We observe that function calling differs fundamentally from free-form text generation: structured outputs exhibit substantial token redundancy (delimiters, parameter names), and arguments exhibit weak causal dependencies. Crucially, these two properties must be exploited jointly to achieve real-time performance. We present SimpleTool, which introduces special tokens that serve a dual role: compressing low-entropy tokens (4-6x reduction) while acting as mode selectors that enable independent parallel generation of function name and arguments. This synergistic design achieves 3-6x end-to-end speedup (up to 9.6x) with only +8.2% parallelization overhead. Experiments on five benchmarks across Qwen-series models (0.5B-14B) demonstrate substantial speedup while maintaining competitive or improved accuracy. On Mobile Actions, ST-Qwen-0.5B outperforms Google’s FunctionGemma in both accuracy and latency consistency. With quantization on consumer-grade GPU, SimpleTool achieves 61.2ms P50 latency, enabling 16 Hz real-time control at 4B model scale, bridging the gap between LLM function calling and latency-critical real-world deployment.

Method: The authors evaluate four open-weight end-to-end models on SpeechMMLU and VoiceBench BBH. They use cross-layer CKA (Centered Kernel Alignment) analysis with speech-text token alignment to examine how speech and text representations evolve layer-by-layer.

Result: Speech representations show a broad cross-layer alignment band due to the redundant nature of speech where semantic content spans multiple frames. These alignment patterns are structurally stable across configurations. Simple statistical calibration is insufficient and can be detrimental when applied at the input layer, indicating the modality gap is not just a distribution shift.

Conclusion: The bottleneck lies in condensing redundant speech into stable late-layer decisions. Future solutions should operate at token or temporal granularity rather than feature-level matching to address the structural differences between speech and text representations.

Abstract: Recent advancements in Large Speech-Language Models have significantly bridged the gap between acoustic signals and linguistic understanding. However, a persistent performance disparity remains in speech-based input tasks compared to direct text inference. In this paper, we investigate the dynamic roots of this modality gap beyond static geometric alignment, analyzing how speech and text representations evolve layer-by-layer. We evaluate four open-weight end-to-end models on SpeechMMLU and VoiceBench BBH. Using cross-layer CKA analysis with speech-text token alignment, we find that speech representations exhibit a broad cross-layer alignment band, attributable to the redundant nature of speech where semantic content spans multiple frames. We show that these alignment patterns are structurally stable across different analysis configurations. Crucially, simple statistical calibration is insufficient and can be detrimental when applied at the input layer, indicating that the modality gap is not a mere distribution shift. Overall, our results suggest that the bottleneck lies in condensing redundant speech into stable late-layer decisions, motivating future solutions that operate at the token or temporal granularity instead of feature-level matching.

Method: GRIP models the corpus as an information-dense geometric space, using a Rapid Adaptation Probe (RAP) to quantify information potential of semantic clusters and dynamically re-allocate sampling budget to regions with highest representation deficits, followed by intra-cluster selection using length-rectified geometric prior to counteract embedding density artifacts.

Result: Extensive evaluations on Mixture-of-Experts models up to 300B tokens show GRIP consistently outperforms state-of-the-art baselines, surpassing performance of models trained on 3× larger uncurated datasets.

Conclusion: GRIP establishes a robust geometric foundation for adaptive data curation in large-scale pre-training, demonstrating that unified geometric modeling of global and local data selection dimensions significantly improves LLM training efficiency.

Abstract: The performance of Large Language Models (LLMs) is increasingly governed by data efficiency rather than raw scaling volume. However, existing selection methods often decouple global distribution balancing from local instance selection, compromising the hierarchical integrity of the training set. We introduce \textbf{GRIP} (Geometric Refinement and Adaptive Information Potential), a framework that unifies these dimensions by modeling the corpus as an information-dense geometric space. GRIP employs a \textbf{Rapid Adaptation Probe (RAP)} to quantify the information potential of semantic clusters, dynamically re-allocating the sampling budget to regions with the highest representation deficits. Subsequently, we perform Intra-Cluster Selection using a \textbf{length-rectified geometric prior} to counteract embedding density artifacts and preserve long-tail logical sequences. Extensive evaluations on Mixture-of-Experts (MoE) models up to 300B tokens demonstrate that GRIP consistently outperforms state-of-the-art baselines, \textbf{surpassing the performance of models trained on $3\times$ larger uncurated datasets}. Our work establishes a robust geometric foundation for adaptive data curation in large-scale pre-training.

Method: Proposes Autorubric, an open-source Python framework supporting binary/ordinal/nominal criteria with weights, multi-judge ensemble evaluation, few-shot calibration, bias mitigations (position, verbosity, criterion conflation), and psychometric reliability metrics.

Result: Evaluated on three benchmarks (RiceChem, ResearcherBench, CHARM-100), produces results consistent with published benchmarks while demonstrating key capabilities like per-criterion evaluation, multi-judge ensembles, and mixed criterion types.

Conclusion: Autorubric provides a comprehensive, unified solution for reliable LLM-based rubric evaluation at scale, with production-ready infrastructure and introduces CHARM-100 dataset for framework stress-testing.

Abstract: Rubric-based evaluation with large language models (LLMs) has become standard practice for assessing text generation at scale, yet the underlying techniques are scattered across papers with inconsistent terminology and partial solutions. We present a unified framework: each identified technique is paired with its realization in Autorubric, an open-source Python framework proposed in this paper. Autorubric supports binary, ordinal, and nominal criteria with configurable weights; single-judge and multi-judge ensemble evaluation with majority, weighted, unanimous, and any-vote aggregation; few-shot calibration with verdict-balanced sampling; and mitigations for position bias (option shuffling), verbosity bias (length penalties), and criterion conflation (per-criterion atomic evaluation with natural language explanations). The framework provides reliability metrics drawn from psychometrics (Cohen’s $κ$, weighted $κ$, correlation coefficients, and distribution-level tests) alongside production infrastructure including response caching, checkpointing with resumable runs, multi-provider rate limiting, and cost tracking. We evaluate Autorubric on three benchmarks spanning educational assessment, deep research evaluation, and chatbot quality assessment, demonstrating that it produces results consistent with published benchmarks while exercising the framework’s key capabilities: per-criterion binary evaluation with few-shot calibration (RiceChem), multi-judge ensemble evaluation across judge models (ResearcherBench), and mixed criterion types combining binary, ordinal, and nominal scales (CHARM-100). We also contribute CHARM-100, a 100-sample chatbot evaluation dataset with per-sample ground truth labels across all three criterion types, designed to stress-test rubric evaluation frameworks on heterogeneous criteria.

Method: Step-wise Adaptive Penalization (SWAP) estimates step importance from the model’s on-policy log-probability improvement toward correct answers, then redistributes length penalty mass to penalize low-importance steps more heavily while preserving high-importance reasoning. Uses unified outcome-process advantage within group-relative policy optimization.

Result: SWAP reduces reasoning length by 64.3% on average while improving accuracy by 5.7% relative to the base model, demonstrating effective compression without sacrificing performance.

Conclusion: The SWAP framework successfully addresses the overthinking problem in large reasoning models by providing fine-grained, adaptive length reduction that preserves essential reasoning while eliminating redundancy, offering a practical solution for efficient reasoning.

Abstract: Large reasoning models improve with more test-time computation, but often overthink, producing unnecessarily long chains-of-thought that raise cost without improving accuracy. Prior reinforcement learning approaches typically rely on a single outcome reward with trajectory-level length penalties, which cannot distinguish essential from redundant reasoning steps and therefore yield blunt compression. Although recent work incorporates step-level signals, such as offline pruning, supervised data construction, or verifier-based intermediate rewards, reasoning length is rarely treated as an explicit step-level optimization objective during RL. We propose Step-wise Adaptive Penalization (SWAP), a fine-grained framework that allocates length reduction across steps based on intrinsic contribution. We estimate step importance from the model’s on-policy log-probability improvement toward the correct answer, then treat excess length as a penalty mass redistributed to penalize low-importance steps more heavily while preserving high-importance reasoning. We optimize with a unified outcome-process advantage within group-relative policy optimization. Extensive experiments demonstrate that SWAP reduces reasoning length by 64.3% on average while improving accuracy by 5.7% relative to the base model.

Method: Used two-level statistical design with prompts as unit of inference, running each experiment 20 times with different generation seeds. Tested static embeddings and contextual hidden states in GPT-2, analyzing norm separation and effect sizes.

Result: Type 3 (coverage gaps) showed robust norm separation in static embeddings (significant in 18/20 runs). In contextual states, Type 3 effect direction was stable but underpowered. Types 1 and 2 did not separate in either space. Token-level tests inflated significance 4-16× through pseudoreplication.

Conclusion: Coverage-gap hallucinations are the most geometrically distinctive failure mode, characterized by magnitude rather than direction differences. Type 1/2 non-separation is genuine at 124M parameters, and token-level analysis suffers from pseudoreplication issues.

Abstract: We test whether a geometric hallucination taxonomy – classifying failures as center-drift (Type1), wrong-well convergence (Type2), or coverage gaps (Type3) – can distinguish hallucination types through controlled induction in GPT-2. Using a two-level statistical design with prompts ($N = 15$/group) as the unit of inference, we run each experiment 20 times with different generation seeds to quantify result stability. In static embeddings, Type3 norm separation is robust (significant in 18/20 runs, Holm-corrected in 14/20, median $r = +0.61$). In contextual hidden states, the Type3 norm effect direction is stable (19/20 runs) but underpowered at $N = 15$ (significant in 4/20, median $r = -0.28$). Types1 and2 do not separate in either space (${\leq},3/20$ runs). Token-level tests inflate significance by 4–16$\times$ through pseudoreplication – a finding replicated across all 20 runs. The results establish coverage-gap hallucinations as the most geometrically distinctive failure mode, carried by magnitude rather than direction, and confirm the Type1/2 non-separation as genuine at 124M parameters.

Method: Supervised fine-tuning of Llama 2 7B transformer model using transcripts from real patient-doctor interactions, evaluated with text similarity metrics.

Result: Fine-tuned model showed significant improvements across key dimensions (except GPT-4 evaluation), but human expert validation is recommended.

Conclusion: Fine-tuning improves medical LLM accuracy, but deployment requires human expert review and evaluation due to limitations of automated metrics.

Abstract: This paper details the baseline model selection, fine-tuning process, evaluation methods, and the implications of deploying more accurate LLMs in healthcare settings. As large language models (LLMs) are increasingly employed to address diverse problems, including medical queries, concerns about their reliability have surfaced. A recent study by Long Island University highlighted that LLMs often perform poorly in medical contexts, potentially leading to harmful misguidance for users. To address this, our research focuses on fine-tuning the Llama 2 7B, a transformer-based, decoder-only model, using transcripts from real patient-doctor interactions. Our objective was to enhance the model’s accuracy and precision in responding to medical queries. We fine-tuned the model using a supervised approach, emphasizing domain-specific nuances captured in the training data. In the best scenario, the model results should be reviewed and evaluated by real medical experts. Due to resource constraints, the performance of the fine-tuned model was evaluated using text similarity metrics. The fine-tuned model demonstrated significant improvements across all key dimensions except GPT-4’s evaluation. The evaluations of ChatGPT4 are quite different from the quantitative results; here, we not only suggest, but also propose that the result should be evaluated by human medical experts.

Method: Trained Meta’s OPT model on 100M word BabyLM dataset, evaluated on BLiMP benchmark with 67 grammatical classes, tracked model preferences across training iterations using qualitative (linguistic theory, deep learning theory) and quantitative (numerical testing) assessments.

Result: OPT fails to consistently prefer grammatical sentences in nearly one-third of BLiMP classes, often establishing erroneous likelihood separation early in training that persists throughout training phase, suggesting entrenched biases that are costly to reverse.

Conclusion: Proposes Bigram Hypothesis: erroneous entrenchment occurs when bigram statistics bias models toward wrong distinctions early in training, with ongoing work to test this hypothesis on selected BLiMP classes.

Abstract: Linguistic insights may help make Large Language Model (LLM) training more efficient. We trained Meta’s OPT model on the 100M word BabyLM dataset, and evaluated it on the BLiMP benchmark, which consists of 67 classes, each defined by sentence pairs that differ in a targeted syntactic or semantic rule violation. We tested the model’s preference for grammatical over ungrammatical sentences across training iterations and grammatical types. In nearly one-third of the BLiMP classes, OPT fails to consistently assign a higher likelihood to grammatical sentences, even after extensive training. When it fails, it often establishes a clear (erroneous) separation of the likelihoods at an early stage of processing and sustains this to the end of our training phase. We hypothesize that this mis-categorization is costly because it creates entrenched biases that must, eventually, be reversed in order for the model to perform well. We probe this phenomenon using a mixture of qualitative (based on linguistic theory and the theory of Deep Learning) and quantitative (based on numerical testing) assessments. Our qualitative assessments indicate that only some BLiMP tests are meaningful guides. We conclude by articulating a hypothesis, the Bigram Hypothesis, which claims that the learning process will exhibit erroneous entrenchment if bigram statistics bias the model toward wrong distinctions early in training, and we describe a method (in progress) of testing the hypothesis on appropriately selected BLiMP classes.

Method: Train language models with n independent output heads on top of a shared trunk to predict n future tokens at each position. This multi-token prediction serves as an auxiliary training task alongside standard next-token prediction.

Result: 13B parameter models show 12% improvement on HumanEval and 17% on MBPP coding benchmarks. Models demonstrate better development of induction heads and algorithmic reasoning. 4-token prediction models achieve up to 3x faster inference speed.

Conclusion: Multi-token prediction is an effective auxiliary training objective that improves sample efficiency, downstream capabilities (especially for coding), and inference speed without increasing training time.

Abstract: Large language models such as GPT and Llama are trained with a next-token prediction loss. In this work, we suggest that training language models to predict multiple future tokens at once results in higher sample efficiency. More specifically, at each position in the training corpus, we ask the model to predict the following n tokens using n independent output heads, operating on top of a shared model trunk. Considering multi-token prediction as an auxiliary training task, we measure improved downstream capabilities with no overhead in training time for both code and natural language models. The method is increasingly useful for larger model sizes and keeps its appeal when training for multiple epochs. Gains are especially pronounced on generative benchmarks like coding, where our models consistently outperform strong baselines by several percentage points. Our 13B parameter models solves 12 % more problems on HumanEval and 17 % more on MBPP than comparable next-token models. Experiments on small algorithmic tasks demonstrate that multi-token prediction is favorable for the development of induction heads and algorithmic reasoning capabilities. As an additional benefit, models trained with 4-token prediction are up to 3X times faster at inference, even with large batch sizes.

Method: Created first benchmark with annotated user requests showing diverse compliance levels with organizational policies, then evaluated various LLM models using different solution methods on this benchmark.

Result: Benchmark reveals challenging nature of policy compliance assessment, with performance differences across LLM models and solution methods highlighting the difficulty of this problem.

Conclusion: Policy compliance assessment for AI systems is a challenging problem requiring specialized benchmarks and evaluation, with current LLM models showing varied performance on this task.

Abstract: Consider an organization whose users send requests in natural language to an AI system that fulfills them by carrying out specific tasks. In this paper, we consider the problem of ensuring such user requests comply with a list of diverse policies determined by the organization with the purpose of guaranteeing the safe and reliable use of the AI system. We propose, to the best of our knowledge, the first benchmark consisting of annotated user requests of diverse compliance with respect to a list of policies. Our benchmark is related to industrial applications in the technology sector. We then use our benchmark to evaluate the performance of various LLM models on policy compliance assessment under different solution methods. We analyze the differences on performance metrics across the models and solution methods, showcasing the challenging nature of our problem.

Method: Leverages recent advances in LLMs to create a framework for benchmarks that capture permissible orthographic variations, proposing OIWER (Orthographically-Informed Word Error Rate) that accounts for these variations.

Result: OIWER reduces pessimistic error rates by an average of 6.3 points, narrows inflated model gaps (e.g., Gemini-Canary performance difference drops from 18.1 to 11.5 points), and aligns more closely with human perception than prior methods like WER-SN by 4.9 points.

Conclusion: The proposed OIWER framework using LLMs provides a more accurate ASR evaluation metric for Indian languages by accounting for orthographic variations, better reflecting real-world performance and human perception.

Abstract: Evaluating ASR systems for Indian languages is challenging due to spelling variations, suffix splitting flexibility, and non-standard spellings in code-mixed words. Traditional Word Error Rate (WER) often presents a bleaker picture of system performance than what human users perceive. Better aligning evaluation with real-world performance requires capturing permissible orthographic variations, which is extremely challenging for under-resourced Indian languages. Leveraging recent advances in LLMs, we propose a framework for creating benchmarks that capture permissible variations. Through extensive experiments, we demonstrate that OIWER, by accounting for orthographic variations, reduces pessimistic error rates (an average improvement of 6.3 points), narrows inflated model gaps (e.g., Gemini-Canary performance difference drops from 18.1 to 11.5 points), and aligns more closely with human perception than prior methods like WER-SN by 4.9 points.

Method: Fact-Flow separates visual fact identification from report generation. First predicts clinical findings from images, then uses these facts to guide MLLM report generation. Uses LLM to automatically create labeled medical findings dataset, eliminating need for expensive manual annotation.

Result: Extensive experiments on two disease-focused medical datasets show significant improvement in factual accuracy compared to state-of-the-art models while maintaining high text quality standards.

Conclusion: Fact-Flow effectively addresses factual instability in medical MLLMs by separating fact identification from generation, enabling more reliable clinical report generation with improved accuracy.

Abstract: The automatic generation of medical reports utilizing Multimodal Large Language Models (MLLMs) frequently encounters challenges related to factual instability, which may manifest as the omission of findings or the incorporation of inaccurate information, thereby constraining their applicability in clinical settings. Current methodologies typically produce reports based directly on image features, which inherently lack a definitive factual basis. In response to this limitation, we introduce Fact-Flow, an innovative framework that separates the process of visual fact identification from the generation of reports. This is achieved by initially predicting clinical findings from the image, which subsequently directs the MLLM to produce a report that is factually precise. A pivotal advancement of our approach is a pipeline that leverages a Large Language Model (LLM) to autonomously create a dataset of labeled medical findings, effectively eliminating the need for expensive manual annotation. Extensive experimental evaluations conducted on two disease-focused medical datasets validate the efficacy of our method, demonstrating a significant enhancement in factual accuracy compared to state-of-the-art models, while concurrently preserving high standards of text quality.

Method: Uses Universal Dependencies to apply inference-time perturbations: full token scrambling, content-word scrambling with function words fixed, dependency-based head-dependent swaps, and sentence-level lemma substitution (+L) that lemmatizes both context and masked target label.

Result: Full scrambling drives word-level reconstruction accuracy near zero in all languages; partial perturbations cause smaller but still large drops. +L has little effect in Chinese but substantially lowers accuracy in German/Spanish/Russian, and does not mitigate scrambling impact. Top-5 accuracy shows same pattern.

Conclusion: Multilingual language models show strong reliance on word order across languages, with inflectional morphology playing varying roles depending on language typology. Models struggle when word order is disrupted.

Abstract: We introduce a typology-aware diagnostic for multilingual masked language models that tests reliance on word order versus inflectional form. Using Universal Dependencies, we apply inference-time perturbations: full token scrambling, content-word scrambling with function words fixed, dependency-based head–dependent swaps, and sentence-level lemma substitution (+L), which lemmatizes both the context and the masked target label. We evaluate mBERT and XLM-R on English, Chinese, German, Spanish, and Russian. Full scrambling drives word-level reconstruction accuracy near zero in all languages; partial and head–dependent perturbations cause smaller but still large drops. +L has little effect in Chinese but substantially lowers accuracy in German/Spanish/Russian, and it does not mitigate the impact of scrambling. Top-5 word accuracy shows the same pattern: under full scrambling, the gold word rarely appears among the five highest-ranked reconstructions. We release code, sampling scripts, and balanced evaluation subsets; Turkish results under strict reconstruction are reported in the appendix.

Method: CoC uses multi-turn chat format to correct errors segment by segment, guided by pre-recognized text and full-text context. Fine-tunes pre-trained LLM on ChFT dataset.

Result: CoC significantly outperforms baseline and benchmark systems in correcting full-text ASR outputs. Analyzes correction thresholds and extrapolates to extra-long ASR outputs.

Conclusion: CoC effectively addresses ASR error correction challenges and explores using additional information types to guide error correction.

Abstract: Full-text error correction with Large Language Models (LLMs) for Automatic Speech Recognition (ASR) is attracting increased attention for its ability to address a wide range of error types, such as punctuation restoration and inverse text normalization, across long context. However, challenges remain regarding stability, controllability, completeness, and fluency. To mitigate these issues, this paper proposes the Chain of Correction (CoC), which uses a multi-turn chat format to correct errors segment by segment, guided by pre-recognized text and full-text context for better semantic understanding. Utilizing the open-sourced ChFT dataset, we fine-tune a pre-trained LLM to evaluate CoC’s performance. Experiments show that CoC significantly outperforms baseline and benchmark systems in correcting full-text ASR outputs. We also analyze correction thresholds to balance under-correction and over-rephrasing, extrapolate CoC on extra-long ASR outputs, and explore using other types of information to guide error correction.

Method: CIRCUS constructs an ensemble of attribution graphs by pruning a single raw attribution run under multiple configurations, assigns edge stability scores, and extracts strict-consensus circuits containing only edges appearing in all views.

Result: On Gemma-2-2B and Llama-3.2-1B, strict consensus circuits are ~40x smaller than union of all configurations while retaining comparable explanatory power, and consensus-identified nodes outperform non-consensus controls in causal validation (p=0.0004).

Conclusion: CIRCUS provides a practical, uncertainty-aware framework for reporting trustworthy, auditable mechanistic circuits with explicit core/contingent/noise decomposition, requiring no retraining and adding negligible overhead.

Abstract: Mechanistic circuit discovery is notoriously sensitive to arbitrary analyst choices, especially pruning thresholds and feature dictionaries, often yielding brittle “one-shot” explanations with no principled notion of uncertainty. We reframe circuit discovery as an uncertainty-quantification problem over these analytic degrees of freedom. Our method, CIRCUS, constructs an ensemble of attribution graphs by pruning a single raw attribution run under multiple configurations, assigns each edge a stability score (the fraction of configurations that retain it), and extracts a strict-consensus circuit consisting only of edges that appear in all views. This produces a threshold-robust “core” circuit while explicitly surfacing contingent alternatives and enabling rejection of low-agreement structure. CIRCUS requires no retraining and adds negligible overhead, since it aggregates structure across already-computed pruned graphs. On Gemma-2-2B and Llama-3.2-1B, strict consensus circuits are ~40x smaller than the union of all configurations while retaining comparable influence-flow explanatory power, and they outperform a same-edge-budget baseline (union pruned to match the consensus size). We further validate causal relevance with activation patching, where consensus-identified nodes consistently beat matched non-consensus controls (p=0.0004). Overall, CIRCUS provides a practical, uncertainty-aware framework for reporting trustworthy, auditable mechanistic circuits with an explicit core/contingent/noise decomposition.

Method: Introduces core space experts (compact core matrices) and core space routing (token-level expert selection), with soft-merging strategy and low-rank routing projection

Result: Achieves parameter efficiency comparable to standard LoRA while outperforming existing methods across multiple tasks

Conclusion: CoMoL retains MoE-LoRA adaptability with improved parameter efficiency and fine-grained adaptation

Abstract: Large language models (LLMs) achieve remarkable performance on diverse downstream and domain-specific tasks via parameter-efficient fine-tuning (PEFT). However, existing PEFT methods, particularly MoE-LoRA architectures, suffer from limited parameter efficiency and coarse-grained adaptation due to the proliferation of LoRA experts and instance-level routing. To address these issues, we propose Core Space Mixture of LoRA (\textbf{CoMoL}), a novel MoE-LoRA framework that incorporates expert diversity, parameter efficiency, and fine-grained adaptation. Specifically, CoMoL introduces two key components: core space experts and core space routing. Core space experts store each expert in a compact core matrix, preserving diversity while controlling parameter growth. Core space routing dynamically selects and activates the appropriate core experts for each token, enabling fine-grained, input-adaptive routing. Activated core experts are then merged via a soft-merging strategy into a single core expert, which is combined with a shared LoRA to form a specialized LoRA module. Besides, the routing network is projected into the same low-rank space as the LoRA matrices, further reducing parameter overhead without compromising expressiveness. Extensive experiments demonstrate that CoMoL retains the adaptability of MoE-LoRA architectures while achieving parameter efficiency comparable to standard LoRA, consistently outperforming existing methods across multiple tasks.

Method: Introduces Super Research task with three components: structured decomposition into research plans, super wide retrieval for diverse perspectives, and super deep investigation through iterative queries. Includes a benchmark of 300 expert-written questions requiring up to 100+ retrieval steps and 1,000+ web pages.

Result: Produces verifiable reports with fine-grained citations and intermediate artifacts. Presents graph-anchored auditing protocol evaluating five dimensions: Coverage, Logical Consistency, Report Utility, Objectivity, and Citation Health.

Conclusion: Super Research serves as a critical ceiling evaluation and stress test for LLM capabilities, where proficiency acts as a proxy for general research competence, indicating robustness for subordinate research tasks.

Abstract: While Large Language Models (LLMs) have demonstrated proficiency in Deep Research or Wide Search, their capacity to solve highly complex questions-those requiring long-horizon planning, massive evidence gathering, and synthesis across heterogeneous sources-remains largely unexplored. We introduce Super Research, a task for complex autonomous research tasks that integrates (i) structured decomposition into a research plan, (ii) super wide retrieval for diverse perspectives, and (iii) super deep investigation to resolve uncertainties through iterative queries. To evaluate this capability, we curated a benchmark of 300 expert-written questions across diverse domains, each requiring up to 100+ retrieval steps and 1,000+ web pages to reconcile conflicting evidence. Super Research produces verifiable reports with fine-grained citations and intermediate artifacts (e.g., outlines and tables) to ensure traceable reasoning. Furthermore, we present a graph-anchored auditing protocol that evaluates Super Research along five dimensions: Coverage, Logical Consistency, Report Utility, Objectivity and Citation Health. While super-complex questions may be infrequent in standard applications, Super Research serves as a critical ceiling evaluation and stress test for LLM capabilities. A model’s proficiency within Super Research acts as a powerful proxy for its general research competence; success here suggests the robustness necessary to navigate nearly any subordinate research task. Leaderboard is available at: https://cnsdqd-dyb.github.io/Super-Research-Benchmark/

Method: Integrates structured biomedical databases and unstructured literature into a task-specific knowledge graph, then uses knowledge graph embeddings with agent-based reasoning to generate, validate, and rank candidate drug combinations while grounding hypotheses in retrievable evidence.

Result: Demonstrates CoDHy as a system for exploratory hypothesis generation and decision support in translational oncology, highlighting its design, interaction workflow, and practical use cases through a web-based interface.

Conclusion: CoDHy enables transparent, researcher-steerable exploration for biomarker-guided drug combination hypothesis generation in cancer research, rather than automated decision-making.

Abstract: The rapid growth of biomedical literature and curated databases has made it increasingly difficult for researchers to systematically connect biomarker mechanisms to actionable drug combination hypotheses. We present AI Co-Scientist (CoDHy), an interactive, human-in-the-loop system for biomarker-guided drug combination hypothesis generation in cancer research. CoDHy integrates structured biomedical databases and unstructured literature evidence into a task-specific knowledge graph, which serves as the basis for graph-based reasoning and hypothesis construction. The system combines knowledge graph embeddings with agent-based reasoning to generate, validate, and rank candidate drug combinations, while explicitly grounding each hypothesis in retrievable evidence. Through a web-based interface, users can configure the scientific context, inspect intermediate results, and iteratively refine hypotheses, enabling transparent and researcher-steerable exploration rather than automated decision-making. We demonstrate CoDHy as a system for exploratory hypothesis generation and decision support in translational oncology, highlighting its design, interaction workflow, and practical use cases.

Method: Developed a lightweight Python toolkit that integrates multiple language resources into a single interface, provides convenient normalization and mapping between language identifiers, and implements a graph-based structure for traversal across families, regions, writing systems, and other linguistic attributes.

Result: Created QwanQwa, which serves as both (1) a simple “glue” library in multilingual NLP research to make working with many languages easier, and (2) an intuitive way for exploring languages through shared scripts, regions, or other metadata.

Conclusion: QwanQwa addresses the scalability challenge of language metadata management in multilingual NLP by providing a unified toolkit that simplifies working with diverse language identifiers and enables exploration of linguistic relationships.

Abstract: The growing number of languages considered in multilingual NLP, including new datasets and tasks, poses challenges regarding properly and accurately reporting which languages are used and how. For example, datasets often use different language identifiers; some use BCP-47 (e.g. en_Latn), others use ISO 639-1 (en), and more linguistically oriented datasets use Glottocodes (stan1293). Mapping between identifiers is manageable for a few dozen languages, but becomes unscalable when dealing with thousands. We introduce QwanQwa, a light-weight Python toolkit for unified language metadata management. QQ integrates multiple language resources into a single interface, provides convenient normalization and mapping between language identifiers, and affords a graph-based structure that enables traversal across families, regions, writing systems, and other linguistic attributes. QQ serves both as (1) a simple “glue” library in multilingual NLP research to make working with many languages easier, and (2) as an intuitive way for exploring languages, such as finding related ones through shared scripts, regions or other metadata.

Method: Created uCDCR by consolidating publicly available English CDCR corpora into a consistent format, correcting inconsistencies, enriching datasets with missing attributes, and establishing standardized metrics and evaluation protocols for analysis.

Result: Analysis shows ECB+ (current state-of-the-art benchmark) has one of the lowest lexical diversities, and its CDCR complexity is middle-range. Using all uCDCR datasets improves model generalizability. Event and entity coreference resolution show similar complexity.

Conclusion: uCDCR provides a unified framework for reproducible CDCR research, showing that both event and entity coreference are complex tasks that shouldn’t be reduced to ECR alone, and that broader dataset usage improves model generalizability.

Abstract: Research in CDCR remains fragmented due to heterogeneous dataset formats, varying annotation standards, and the predominance of the CDCR definition as the event coreference resolution (ECR). To address these challenges, we introduce uCDCR, a unified dataset that consolidates diverse publicly available English CDCR corpora across various domains into a consistent format, which we analyze with standardized metrics and evaluation protocols. uCDCR incorporates both entity and event coreference, corrects known inconsistencies, and enriches datasets with missing attributes to facilitate reproducible research. We establish a cohesive framework for fair, interpretable, and cross-dataset analysis in CDCR and compare the datasets on their lexical properties, e.g., lexical composition of the annotated mentions, lexical diversity and ambiguity metrics, discuss the annotation rules and principles that lead to high lexical diversity, and examine how these metrics influence performance on the same-head-lemma baseline. Our dataset analysis shows that ECB+, the state-of-the-art benchmark for CDCR, has one of the lowest lexical diversities, and its CDCR complexity, measured by the same-head-lemma baseline, lies in the middle among all uCDCR datasets. Moreover, comparing document and mention distributions between ECB+ and uCDCR shows that using all uCDCR datasets for model training and evaluation will improve the generalizability of CDCR models. Finally, the almost identical performance on the same-head-lemma baseline, separately applied to events and entities, shows that resolving both types is a complex task and should not be steered toward ECR alone. The uCDCR dataset is available at https://huggingface.co/datasets/AnZhu/uCDCR, and the code for parsing, analyzing, and scoring the dataset is available at https://github.com/anastasia-zhukova/uCDCR.

Method: Created BLUFF benchmark with 202K+ samples across 79 languages, combining human-written fact-checked content and LLM-generated content. Introduced AXL-CoI framework for controlled fake/real news generation and mPURIFY quality filtering pipeline.

Result: Experiments show state-of-the-art detectors suffer up to 25.3% F1 degradation on low-resource vs high-resource languages, highlighting the need for multilingual benchmarks.

Conclusion: BLUFF provides a comprehensive multilingual benchmark with tools to advance equitable falsehood detection across diverse linguistic communities.

Abstract: Multilingual falsehoods threaten information integrity worldwide, yet detection benchmarks remain confined to English or a few high-resource languages, leaving low-resource linguistic communities without robust defense tools. We introduce BLUFF, a comprehensive benchmark for detecting false and synthetic content, spanning 79 languages with over 202K samples, combining human-written fact-checked content (122K+ samples across 57 languages) and LLM-generated content (79K+ samples across 71 languages). BLUFF uniquely covers both high-resource “big-head” (20) and low-resource “long-tail” (59) languages, addressing critical gaps in multilingual research on detecting false and synthetic content. Our dataset features four content types (human-written, LLM-generated, LLM-translated, and hybrid human-LLM text), bidirectional translation (English$\leftrightarrow$X), 39 textual modification techniques (36 manipulation tactics for fake news, 3 AI-editing strategies for real news), and varying edit intensities generated using 19 diverse LLMs. We present AXL-CoI (Adversarial Cross-Lingual Agentic Chainof-Interactions), a novel multi-agentic framework for controlled fake/real news generation, paired with mPURIFY, a quality filtering pipeline ensuring dataset integrity. Experiments reveal state-of-theart detectors suffer up to 25.3% F1 degradation on low-resource versus high-resource languages. BLUFF provides the research community with a multilingual benchmark, extensive linguistic-oriented benchmark evaluation, comprehensive documentation, and opensource tools to advance equitable falsehood detection. Dataset and code are available at: https://jsl5710.github.io/BLUFF/

Method: LLM-centered, expert-guided pipeline with automatic standard identification, configurable chunking, standard-specific prompting, robust triple parsing, and provenance-aware Neo4j storage. Includes role-based governance framework for draft KG curation and certified KG promotion.

Result: Created SSKG Hub interactive web platform that transforms standards into auditable knowledge graphs with fine-grained audit metadata. Validated through comprehensive expert-led KG review case study demonstrating end-to-end curation and quality assurance.

Conclusion: SSKG Hub enables structured analysis of sustainability standards through knowledge graph transformation with traceability and accountability, supporting cross-KG fusion and KG-driven tasks.

Abstract: Sustainability disclosure standards (e.g., GRI, SASB, TCFD, IFRS S2) are comprehensive yet lengthy, terminology-dense, and highly cross-referential, hindering structured analysis and downstream use. We present SSKG Hub (Sustainability Standards Knowledge Graph Hub), a research prototype and interactive web platform that transforms standards into auditable knowledge graphs (KGs) through an LLM-centered, expert-guided pipeline. The system integrates automatic standard identification, configurable chunking, standard-specific prompting, robust triple parsing, and provenance-aware Neo4j storage with fine-grained audit metadata. LLM extraction produces a provenance-linked Draft KG, which is reviewed, curated, and formally promoted to a Certified KG through meta-expert adjudication. A role-based governance framework covering read-only guest access, expert review and CRUD operations, meta-expert certification, and administrative oversight ensures traceability and accountability across draft and certified states. Beyond graph exploration and triple-level evidence tracing, SSKG Hub supports cross-KG fusion, KG-driven tasks, and dedicated modules for insights and curated resources. We validate the platform through a comprehensive expert-led KG review case study that demonstrates end-to-end curation and quality assurance. The web application is publicly available at www.sskg-hub.com.

Method: Proposes Polynomial Mixer (PoM) as drop-in replacement for multi-head self-attention, computing polynomial representation of input with linear complexity relative to sequence length.

Result: PoM achieves competitive word error rate compared to full self-attention and other linear-complexity alternatives, offering better performance-efficiency trade-off.

Conclusion: PoM provides efficient alternative to self-attention for speech recognition models, enabling better scalability while maintaining accuracy.

Abstract: State-of-the-art speech-to-text models typically employ Transformer-based encoders that model token dependencies via self-attention mechanisms. However, the quadratic complexity of self-attention in both memory and computation imposes significant constraints on scalability. In this work, we propose a novel token-mixing mechanism, the Polynomial Mixer (PoM), as a drop-in replacement for multi-head self-attention. PoM computes a polynomial representation of the input with linear complexity with respect to the input sequence length. We integrate PoM into a self-supervised speech representation learning framework based on BEST-RQ and evaluate its performance on downstream speech recognition tasks. Experimental results demonstrate that PoM achieves a competitive word error rate compared to full self-attention and other linear-complexity alternatives, offering an improved trade-off between performance and efficiency in time and memory.

Method: Introduces RAVEL framework enabling LLM testers to autonomously plan and execute synthesis operations, complemented by C3EBench benchmark with 1,258 samples derived from professional human writings using reverse-engineering pipeline across four tasks: Cloze, Edit, Expand, and End-to-End.

Result: Analysis of 14 LLMs shows most struggle with tasks requiring contextual understanding under limited instructions. Agentic text synthesis is dominated by reasoning capability rather than raw generative capacity. Strong reasoners can guide weaker generators to better results, but not vice versa.

Conclusion: RAVEL provides a comprehensive framework for evaluating LLM synthesis capabilities, revealing the critical importance of reasoning over raw generation in agentic text synthesis tasks.

Abstract: Large Language Models have evolved from single-round generators into long-horizon agents, capable of complex text synthesis scenarios. However, current evaluation frameworks lack the ability to assess the actual synthesis operations, such as outlining, drafting, and editing. Consequently, they fail to evaluate the actual and detailed capabilities of LLMs. To bridge this gap, we introduce RAVEL, an agentic framework that enables the LLM testers to autonomously plan and execute typical synthesis operations, including outlining, drafting, reviewing, and refining. Complementing this framework, we present C3EBench, a comprehensive benchmark comprising 1,258 samples derived from professional human writings. We utilize a “reverse-engineering” pipeline to isolate specific capabilities across four tasks: Cloze, Edit, Expand, and End-to-End. Through our analysis of 14 LLMs, we uncover that most LLMs struggle with tasks that demand contextual understanding under limited or under-specified instructions. By augmenting RAVEL with SOTA LLMs as operators, we find that such agentic text synthesis is dominated by the LLM’s reasoning capability rather than raw generative capacity. Furthermore, we find that a strong reasoner can guide a weaker generator to yield higher-quality results, whereas the inverse does not hold. Our code and data are available at this link: https://github.com/ZhuoerFeng/RAVEL-Reasoning-Agents-Text-Eval.

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 2503.21735: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.21735&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: DRIV-EX uses gradient-based optimization on continuous embeddings to identify minimal semantic changes needed to flip LLM decisions. Instead of directly using optimized embeddings (which can produce incoherent text), it uses them as semantic guides to bias a controlled decoding process that regenerates the original scene description, ensuring linguistic fluency and domain validity.

Result: Evaluated on the LC-LLM planner with textual transcriptions of the highD dataset, DRIV-EX generates valid, fluent counterfactuals more reliably than existing baselines. It successfully exposes latent biases and provides concrete insights for improving LLM-based driving agent robustness.

Conclusion: DRIV-EX provides an effective method for interpreting LLM decision-making in autonomous driving through counterfactual explanations, offering both interpretability and insights for improving model robustness.

Abstract: Large language models (LLMs) are increasingly used as reasoning engines in autonomous driving, yet their decision-making remains opaque. We propose to study their decision process through counterfactual explanations, which identify the minimal semantic changes to a scene description required to alter a driving plan. We introduce DRIV-EX, a method that leverages gradient-based optimization on continuous embeddings to identify the input shifts required to flip the model’s decision. Crucially, to avoid the incoherent text typical of unconstrained continuous optimization, DRIV-EX uses these optimized embeddings solely as a semantic guide: they are used to bias a controlled decoding process that re-generates the original scene description. This approach effectively steers the generation toward the counterfactual target while guaranteeing the linguistic fluency, domain validity, and proximity to the original input, essential for interpretability. Evaluated using the LC-LLM planner on a textual transcription of the highD dataset, DRIV-EX generates valid, fluent counterfactuals more reliably than existing baselines. It successfully exposes latent biases and provides concrete insights to improve the robustness of LLM-based driving agents.

Method: Introduces SkillCraft benchmark with realistic, highly compositional tool-use scenarios scaled along quantitative and structural dimensions. Proposes lightweight evaluation protocol allowing agents to auto-compose atomic tools into executable Skills, cache and reuse them across tasks, building a persistent library of reusable skills.

Result: Evaluation shows substantial efficiency gains with token usage reduced by up to 80% through skill saving and reuse. Success rate strongly correlates with tool composition ability at test time, indicating compositional skill acquisition as a core capability.

Conclusion: SkillCraft addresses a critical gap in evaluating agents’ ability to form and reuse higher-level tool compositions, demonstrating that compositional skill acquisition is essential for efficient tool-use in complex, long-horizon workflows.

Abstract: Real-world tool-using agents operate over long-horizon workflows with recurring structure and diverse demands, where effective behavior requires not only invoking atomic tools but also abstracting, and reusing higher-level tool compositions. However, existing benchmarks mainly measure instance-level success under static tool sets, offering limited insight into agents’ ability to acquire such reusable skills. We address this gap by introducing SkillCraft, a benchmark explicitly stress-test agent ability to form and reuse higher-level tool compositions, where we call Skills. SkillCraft features realistic, highly compositional tool-use scenarios with difficulty scaled along both quantitative and structural dimensions, designed to elicit skill abstraction and cross-task reuse. We further propose a lightweight evaluation protocol that enables agents to auto-compose atomic tools into executable Skills, cache and reuse them inside and across tasks, thereby improving efficiency while accumulating a persistent library of reusable skills. Evaluating state-of-the-art agents on SkillCraft, we observe substantial efficiency gains, with token usage reduced by up to 80% by skill saving and reuse. Moreover, success rate strongly correlates with tool composition ability at test time, underscoring compositional skill acquisition as a core capability.

Method: Transforms reward acquisition into dynamic tool synthesis: uses LLM agents to autonomously retrieve optimal reward models from the Internet and synthesize programmatic verifiers through code generation, allowing the reward system to self-evolve with shifting data distributions.

Result: Achieves consistent performance gains of 10-60% across mathematics, coding, translation, and dialogue tasks; significantly outperforms static baselines on RewardBench-V2 and approaches performance upper bound.

Conclusion: RLAR demonstrates superior generalization through dynamic reward orchestration, offering a more flexible approach to LLM alignment that adapts to evolving training distributions.

Abstract: Large language model alignment via reinforcement learning depends critically on reward function quality. However, static, domain-specific reward models are often costly to train and exhibit poor generalization in out-of-distribution scenarios encountered during RL iterations. We present RLAR (Reinforcement Learning from Agent Rewards), an agent-driven framework that dynamically assigns tailored reward functions to individual queries. Specifically, RLAR transforms reward acquisition into a dynamic tool synthesis and invocation task. It leverages LLM agents to autonomously retrieve optimal reward models from the Internet and synthesize programmatic verifiers through code generation. This allows the reward system to self-evolve with the shifting data distributions during training. Experimental results demonstrate that RLAR yields consistent performance gains ranging from 10 to 60 across mathematics, coding, translation, and dialogue tasks. On RewardBench-V2, RLAR significantly outperforms static baselines and approaches the performance upper bound, demonstrating superior generalization through dynamic reward orchestration. The data and code are available on this link: https://github.com/ZhuoerFeng/RLAR.

Method: Unable to determine paper method due to API request failure

Result: Unable to determine paper results due to API request failure

Conclusion: Unable to determine paper conclusion due to API request failure

Abstract: Failed to fetch summary for 2510.18866: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.18866&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Build large-scale segment-caption training data using TV2-based segmentation on LOTSA windows, generate segment descriptions with GPT-5.2, train Conformer-based contrastive retriever in shared text-time-series embedding space.

Result: LaSTR outperforms random and CLIP baselines across all settings, showing improved ranking quality and stronger semantic agreement between retrieved segments and query intent.

Conclusion: Language-driven segment retrieval enables effective natural language querying of time-series data, overcoming limitations of traditional similarity-based methods.

Abstract: Effectively searching time-series data is essential for system analysis, but existing methods often require expert-designed similarity criteria or rely on global, series-level descriptions. We study language-driven segment retrieval: given a natural language query, the goal is to retrieve relevant local segments from large time-series repositories. We build large-scale segment–caption training data by applying TV2-based segmentation to LOTSA windows and generating segment descriptions with GPT-5.2, and then train a Conformer-based contrastive retriever in a shared text–time-series embedding space. On a held-out test split, we evaluate single-positive retrieval together with caption-side consistency (SBERT and VLM-as-a-judge) under multiple candidate pool sizes. Across all settings, LaSTR outperforms random and CLIP baselines, yielding improved ranking quality and stronger semantic agreement between retrieved segments and query intent.

Method: Uses agentic training through large-scale synthesis of verifiable coding tasks paired with executable environments, enabling learning directly from environment feedback via mid-training and reinforcement learning. The model has 80B total parameters but only activates 3B during inference.

Result: Achieves competitive performance on agent-centric benchmarks including SWE-Bench and Terminal-Bench relative to its active parameter count of 3B parameters.

Conclusion: Demonstrates that strong training recipes can enable models with small active parameter footprints to achieve competitive coding agent performance, and releases open-weight versions to support research and real-world coding agent development.

Abstract: We present Qwen3-Coder-Next, an open-weight language model specialized for coding agents. Qwen3-Coder-Next is an 80-billion-parameter model that activates only 3 billion parameters during inference, enabling strong coding capability with efficient inference. In this work, we explore how far strong training recipes can push the capability limits of models with small parameter footprints. To achieve this, we perform agentic training through large-scale synthesis of verifiable coding tasks paired with executable environments, allowing learning directly from environment feedback via mid-training and reinforcement learning. Across agent-centric benchmarks including SWE-Bench and Terminal-Bench, Qwen3-Coder-Next achieves competitive performance relative to its active parameter count. We release both base and instruction-tuned open-weight versions to support research and real-world coding agent development.

Method: Study whether unlearning remains stable in interactive environments by examining two common interaction patterns: self-correction and dialogue-conditioned querying, comparing static evaluation with interactive scenarios.

Result: Knowledge appearing forgotten in static evaluation can often be recovered through interaction. Stronger unlearning often results in behavioral rigidity rather than genuine knowledge erasure.

Conclusion: Static evaluation may overestimate real-world effectiveness of machine unlearning, highlighting the need for ensuring stable forgetting under interactive settings.

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

Method: 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 2512.01822: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.01822&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Develop constitutional black-box monitors using LLM-based prompted classifiers. Generate synthetic agent trajectories using STRIDE (iterative refinement) and Gloom (agent-environment simulation) pipelines. Optimize frontier LLM monitors via prompt sweeps, human refinement, and automated prompt optimization on synthetic data.

Result: Monitors trained purely on synthetic data generalize to realistic environments (ControlArena), capturing meaningful scheming signals. However, performance saturates quickly - simple prompt sweeps match extensive optimization results. Further optimization leads to overfitting without improvements.

Conclusion: LLM-based monitoring shows promise for detecting scheming agents, but current approaches have limitations with performance saturation and overfitting, suggesting need for more sophisticated methods beyond prompt optimization.

Abstract: Safe deployment of Large Language Model (LLM) agents in autonomous settings requires reliable oversight mechanisms. A central challenge is detecting scheming, where agents covertly pursue misaligned goals. One approach to mitigating such risks is LLM-based monitoring: using language models to examine agent behaviors for suspicious actions. We study constitutional black-box monitors: prompted classifiers that detect scheming using only externally observable inputs and outputs, optimized on synthetic data generated from natural-language behavior specifications. We introduce two pipelines for generating synthetic agent trajectories, STRIDE (iterative refinement) and Gloom (agent-environment simulation), from which we generate 1,000 samples each. We optimize frontier LLM monitors on these datasets via prompt sweeps, human refinement, and automated prompt optimization, and evaluate performance on 7,500 held-out trajectories from ControlArena, a suite of grounded environments where agents operate in more realistic contexts. Our results demonstrate that monitors selected purely on synthetic data can generalize to more realistic environments, capturing a meaningful scheming signal. However, we find that performance saturates quickly in our setting, with simple prompt sweeps matching the results of more extensive optimization. Pushing beyond this limit yields no further improvements and instead leads to overfitting.

Method: Four approaches: 1) String inclusions (substring matching), 2) Entity corruption (pseudo-nested data), 3) Flat neutralization (reducing false negative signal), and 4) Hybrid fine-tuned + LLM pipeline.

Result: On NEREL, a Russian benchmark with 29 entity types where 21% of entities are nested, the best combined method achieves 26.37% inner F1, closing 40% of the gap to full nested supervision.

Conclusion: Models can learn nested structure from flat annotations alone, with the hybrid approach showing promising results, though there remains significant room for improvement compared to fully supervised methods.

Abstract: Nested named entity recognition identifies entities contained within other entities, but requires expensive multi-level annotation. While flat NER corpora exist abundantly, nested resources remain scarce. We investigate whether models can learn nested structure from flat annotations alone, evaluating four approaches: string inclusions (substring matching), entity corruption (pseudo-nested data), flat neutralization (reducing false negative signal), and a hybrid fine-tuned + LLM pipeline. On NEREL, a Russian benchmark with 29 entity types where 21% of entities are nested, our best combined method achieves 26.37% inner F1, closing 40% of the gap to full nested supervision. Code is available at https://github.com/fulstock/Learning-from-Flat-Annotations.

Method: Pairs GPT-oss language backbone with visual front-end using optimized three-stage training curriculum: progressive domain adaptation through rigorous data curation and long-context multimodal alignment, maintaining parameter efficiency for commodity GPUs.

Result: Outperforms larger open medical models on out-of-distribution multimodal reasoning and complex text-only clinical tasks, demonstrating that a 20B model can bridge the capacity gap for clinical applications.

Conclusion: Provides an open-weight, computationally efficient foundation for privacy-preserving, institution-specific clinical AI research with complete training recipe, checkpoints, and evaluation harness released.

Abstract: Biomedical multimodal assistants have the potential to unify radiology, pathology, and clinical-text reasoning, yet a critical deployment gap remains: top-performing systems are either closed-source or computationally prohibitive, precluding the on-premises deployment required for patient privacy and PHI compliance. We introduce MEDGPT-OSS, an open-weight, 20B-parameter generalist vision-language model designed to facilitate open research in clinical AI. Rather than relying on architectural complexity, MEDGPT-OSS pairs the GPT-oss language backbone with a visual front-end via a optimized, three-stage training curriculum. By progressively domain-adapting these modules through rigorous data curation and long-context multimodal alignment, we demonstrate that a 20B model can bridge the capacity gap. It successfully outperforms larger open medical models on out-of-distribution (OOD) multimodal reasoning and complex text-only clinical tasks. By unifying diverse modalities under a single instruction-following interface, MEDGPT-OSS maintains a parameter-efficient footprint fully compatible with commodity GPUs. We release the complete training recipe, open-weight checkpoints, and a rigorous evaluation harness to serve as a verifiable foundation for privacy-preserving, institution-specific clinical AI research.

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 2601.14652: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.14652&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Created CHIMERA dataset with 9K samples featuring: (1) rich, long CoT reasoning trajectories synthesized by state-of-the-art reasoning models, (2) broad coverage across 8 major scientific disciplines and 1K+ fine-grained topics organized via hierarchical taxonomy, (3) fully automated evaluation pipeline using strong reasoning models to cross-validate problem validity and answer correctness.

Result: Post-training a 4B Qwen3 model with CHIMERA achieved strong performance on challenging reasoning benchmarks (GPQA-Diamond, AIME 24/25/26, HMMT 25, Humanity’s Last Exam), approaching or matching performance of substantially larger models like DeepSeek-R1 and Qwen3-235B.

Conclusion: CHIMERA demonstrates that compact, high-quality synthetic reasoning datasets can effectively address data bottlenecks and enable strong reasoning capabilities in smaller models, providing a scalable solution for cross-domain scientific reasoning.

Abstract: Large Language Models (LLMs) have recently exhibited remarkable reasoning capabilities, largely enabled by supervised fine-tuning (SFT)- and reinforcement learning (RL)-based post-training on high-quality reasoning data. However, reproducing and extending these capabilities in open and scalable settings is hindered by three fundamental data-centric challenges: (1) the cold-start problem, arising from the lack of seed datasets with detailed, long Chain-of-Thought (CoT) trajectories needed to initialize reasoning policies; (2) limited domain coverage, as most existing open-source reasoning datasets are concentrated in mathematics, with limited coverage of broader scientific disciplines; and (3) the annotation bottleneck, where the difficulty of frontier-level reasoning tasks makes reliable human annotation prohibitively expensive or infeasible. To address these challenges, we introduce CHIMERA, a compact synthetic reasoning dataset comprising 9K samples for generalizable cross-domain reasoning. CHIMERA is constructed with three key properties: (1) it provides rich, long CoT reasoning trajectories synthesized by state-of-the-art reasoning models; (2) it has broad and structured coverage, spanning 8 major scientific disciplines and over 1K fine-grained topics organized via a model-generated hierarchical taxonomy; and (3) it employs a fully automated, scalable evaluation pipeline that uses strong reasoning models to cross-validate both problem validity and answer correctness. We use CHIMERA to post-train a 4B Qwen3 model. Despite the dataset’s modest size, the resulting model achieves strong performance on a suite of challenging reasoning benchmarks, including GPQA-Diamond, AIME 24/25/26, HMMT 25, and Humanity’s Last Exam, approaching or matching the reasoning performance of substantially larger models such as DeepSeek-R1 and Qwen3-235B.

Method: Establishes theoretical framework analyzing KV asymmetry through spectral energy distribution of projection weights. Introduces KVSlimmer algorithm that captures exact Hessian information through mathematically exact formulation and derives closed-form solution using only forward-pass variables.

Result: Outperforms SOTA methods across various models and benchmarks. On Llama3.1-8B-Instruct, improves LongBench average score by 0.92 while reducing memory costs by 29% and latency by 28%.

Conclusion: KVSlimmer provides a theoretically-sound, efficient solution for KV cache compression that significantly reduces memory and computational overhead while maintaining or improving model performance.

Abstract: The growing computational and memory demands of the Key-Value (KV) cache significantly limit the ability of Large Language Models (LLMs). While KV merging has emerged as a promising solution, existing methods that rely on empirical observations of KV asymmetry and gradient-based Hessian approximations lack a theoretical foundation and incur suboptimal compression and inference overhead. To bridge these gaps, we establish a theoretical framework that characterizes this asymmetry through the spectral energy distribution of projection weights, demonstrating that concentrated spectra in Query/Key weights induce feature homogeneity, whereas dispersed spectra in Value weights preserve heterogeneity. Then, we introduce KVSlimmer, an efficient algorithm that captures exact Hessian information through a mathematically exact formulation, and derives a closed-form solution utilizing only forward-pass variables, resulting in a gradient-free approach that is both memory- and time-efficient. Extensive experiments across various models and benchmarks demonstrate that KVSlimmer consistently outperforms SOTA methods. For instance, on Llama3.1-8B-Instruct, it improves the LongBench average score by 0.92 while reducing memory costs and latency by 29% and 28%, respectively.

Method: Evaluated five open-source models (Gemma 2 2B, Phi-3 Mini 3.8B, Llama 3.2 3B, Mistral 7B, Meditron-7B) across three clinical QA datasets (MedQA, MedMCQA, PubMedQA) using five prompt styles (original, formal, simplified, roleplay, direct), measuring consistency scores, accuracy, and instruction-following failure rates on consumer CPU hardware without fine-tuning.

Result: Consistency and accuracy were largely independent; Gemma 2 had highest consistency (0.845-0.888) but lowest accuracy (33.0-43.5%), while Llama 3.2 showed moderate consistency (0.774-0.807) with highest accuracy (49.0-65.0%). Roleplay prompts consistently reduced accuracy across all models, with Phi-3 Mini dropping 21.5 percentage points on MedQA. Meditron-7B exhibited near-complete instruction-following failure on PubMedQA (99.0% UNKNOWN rate).

Conclusion: High consistency does not imply correctness - models can be reliably wrong, a dangerous failure mode in clinical AI. Roleplay prompts should be avoided in healthcare applications. Llama 3.2 showed strongest balance of accuracy and reliability for low-resource deployment. Safe clinical AI requires joint evaluation of consistency, accuracy, and instruction adherence.

Abstract: Small open-source language models are gaining attention for low-resource healthcare settings, but their reliability under different prompt phrasings remains poorly understood. We evaluated five open-source models (Gemma 2 2B, Phi-3 Mini 3.8B, Llama 3.2 3B, Mistral 7B, and Meditron-7B domain-pretrained without instruction tuning) across three clinical QA datasets (MedQA, MedMCQA, PubMedQA) using five prompt styles (original, formal, simplified, roleplay, direct). We measured consistency scores, accuracy, and instruction-following failure rates. All inference ran locally on consumer CPU hardware without fine-tuning. Consistency and accuracy were largely independent. Gemma 2 achieved the highest consistency (0.845-0.888) but lowest accuracy (33.0-43.5%), while Llama 3.2 showed moderate consistency (0.774-0.807) with the highest accuracy (49.0-65.0%). Roleplay prompts consistently reduced accuracy across all models, with Phi-3 Mini dropping 21.5 percentage points on MedQA. Meditron-7B exhibited near-complete instruction-following failure on PubMedQA (99.0% UNKNOWN rate), showing domain pretraining alone is insufficient for structured clinical QA. High consistency does not imply correctness. Models can be reliably wrong, a dangerous failure mode in clinical AI. Roleplay prompts should be avoided in healthcare applications. Llama 3.2 showed the strongest balance of accuracy and reliability for low-resource deployment. Safe clinical AI requires joint evaluation of consistency, accuracy, and instruction adherence.

Method: Two-stage hybrid pipeline: 1) Neural sequence labeling using BiLSTM-CRF model, 2) LLM post-correction with retrieval-augmented prompting. Systematic ablation studies examine effects of morpheme dictionaries, few-shot examples, and different prompting strategies.

Result: Retrieval-augmented prompting provides substantial gains over random example selection; morpheme dictionaries paradoxically hurt performance; performance scales logarithmically with few-shot examples; hybrid pipeline yields substantial gains and meaningful workload reduction.

Conclusion: Hybrid architectures combining structured prediction models with LLM reasoning offer promising direction for computationally light solutions to automatic linguistic annotation in endangered language documentation, with established design principles for morphologically complex fieldwork contexts.

Abstract: Interlinear glossed text (IGT) creation remains a major bottleneck in linguistic documentation and fieldwork, particularly for low-resource morphologically rich languages. We present a hybrid automatic glossing pipeline that combines neural sequence labeling with large language model (LLM) post-correction, evaluated on Jungar Tuvan, a low-resource Turkic language. Through systematic ablation studies, we show that retrieval-augmented prompting provides substantial gains over random example selection. We further find that morpheme dictionaries paradoxically hurt performance compared to providing no dictionary at all in most cases, and that performance scales approximately logarithmically with the number of few-shot examples. Most significantly, our two-stage pipeline combining a BiLSTM-CRF model with LLM post-correction yields substantial gains for most models, achieving meaningful reductions in annotation workload. Drawing on these findings, we establish concrete design principles for integrating structured prediction models with LLM reasoning in morphologically complex fieldwork contexts. These principles demonstrate that hybrid architectures offer a promising direction for computationally light solutions to automatic linguistic annotation in endangered language documentation.

Method: Present a conformal prediction framework with two applications: 1) Extract structured entities from 1,000 FDA drug labels using GPT-4.1 with FactScore evaluation; 2) Extract radiological entities from MIMIC-CXR reports using GPT-4.1 and Llama-4-Maverick with RadGraph schema, evaluated against physician annotations.

Result: Models show reversed miscalibration direction across domains: underconfident on well-structured FDA labels (τ≈0.06) vs overconfident on free-text radiology reports (τ up to 0.99). Conformal prediction achieves ≥90% target coverage in both settings with manageable 9-13% rejection rates.

Conclusion: Calibration is not a global model property but depends on document structure, extraction category, and model architecture, motivating domain-specific conformal calibration for safe clinical deployment of LLMs.

Abstract: Large Language Models (LLMs) are increasingly used for medical entity extraction, yet their confidence scores are often miscalibrated, limiting safe deployment in clinical settings. We present a conformal prediction framework that provides finite-sample coverage guarantees for LLM-based extraction across two clinical domains. First, we extract structured entities from 1,000 FDA drug labels across eight sections using GPT-4.1, verified via FactScore-based atomic statement evaluation (97.7% accuracy over 128,906 entities). Second, we extract radiological entities from MIMIC-CXR reports using the RadGraph schema with GPT-4.1 and Llama-4-Maverick, evaluated against physician annotations (entity F1: 0.81 to 0.84). Our central finding is that miscalibration direction reverses across domains: on well-structured FDA labels, models are underconfident, requiring modest conformal thresholds ($τ\approx 0.06$), while on free-text radiology reports, models are overconfident, demanding strict thresholds ($τ$ up to 0.99). Despite this heterogeneity, conformal prediction achieves target coverage ($\geq 90%$) in both settings with manageable rejection rates (9–13%). These results demonstrate that calibration is not a global model property but depends on document structure, extraction category, and model architecture, motivating domain-specific conformal calibration for safe clinical deployment.

Method: Conducted a year-long evaluation of 11 vision-language models using DrawEduMath benchmark, which contains real students’ handwritten responses to math problems, analyzing performance across different student proficiency levels and error assessment tasks.

Result: All VLMs underperformed when describing work from students needing more pedagogical help and struggled most on questions related to assessing student errors, despite being optimized for math problem solving.

Conclusion: VLMs require alternative development incentives beyond math problem-solving optimization to adequately support educational use cases, particularly for error assessment and supporting struggling students.

Abstract: Effective mathematics education requires identifying and responding to students’ mistakes. For AI to support pedagogical applications, models must perform well across different levels of student proficiency. Our work provides an extensive, year-long snapshot of how 11 vision-language models (VLMs) perform on DrawEduMath, a QA benchmark involving real students’ handwritten, hand-drawn responses to math problems. We find that models’ weaknesses concentrate on a core component of math education: student error. All evaluated VLMs underperform when describing work from students who require more pedagogical help, and across all QA, they struggle the most on questions related to assessing student error. Thus, while VLMs may be optimized to be math problem solving experts, our results suggest that they require alternative development incentives to adequately support educational use cases.

Method: Created S-VoCAL dataset with 8 sociophonetic attributes and 952 character-book pairs from Project Gutenberg. Developed evaluation framework with attribute-specific metrics and novel LLM-based similarity metric. Demonstrated applicability using Retrieval-Augmented Generation (RAG) pipeline for attribute inference.

Result: RAG pipeline reliably infers attributes like Age and Gender, but struggles with Origin and Physical Health. The dataset provides first benchmark for evaluating voice-related character attribute extraction systems.

Conclusion: S-VoCAL enables systematic evaluation of character voice attribute inference, which is crucial for improving synthetic audiobook narration quality and character impersonation capabilities.

Abstract: With recent advances in Text-to-Speech (TTS) systems, synthetic audiobook narration has seen increased interest, reaching unprecedented levels of naturalness. However, larger gaps remain in synthetic narration systems’ ability to impersonate fictional characters, and convey complex emotions or prosody. A promising direction to enhance character identification is the assignment of plausible voices to each fictional characters in a book. This step typically requires complex inference of attributes in book-length contexts, such as a character’s age, gender, origin or physical health, which in turns requires dedicated benchmark datasets to evaluate extraction systems’ performances. We present S-VoCAL (Speaking Voice Character Attributes in Literature), the first dataset and evaluation framework dedicated to evaluate the inference of voice-related fictional character attributes. S-VoCAL entails 8 attributes grounded in sociophonetic studies, and 952 character-book pairs derived from Project Gutenberg. Its evaluation framework addresses the particularities of each attribute, and includes a novel similarity metric based on recent Large Language Models embeddings. We demonstrate the applicability of S-VoCAL by applying a simple Retrieval-Augmented Generation (RAG) pipeline to the task of inferring character attributes. Our results suggest that the RAG pipeline reliably infers attributes such as Age or Gender, but struggles on others such as Origin or Physical Health. The dataset and evaluation code are available at https://github.com/AbigailBerthe/S-VoCAL .

Method: Developed a web-based platform that integrates multiple state-of-the-art Arabic essay scoring models, providing a user-friendly interface for instructors to create assignments, upload essays, configure scoring, and evaluate essays per trait without dealing with technical API complexities.

Result: Qayyem successfully abstracts technical complexity and provides access to advanced Arabic AES services through an integrated platform with various scoring models offering different effectiveness and efficiency trade-offs.

Conclusion: Qayyem addresses the gap in Arabic AES tools by providing an accessible platform that makes advanced scoring models available to instructors through a simplified workflow, potentially improving Arabic language assessment scalability.

Abstract: Over the past years, Automated Essay Scoring (AES) systems have gained increasing attention as scalable and consistent solutions for assessing the proficiency of student writing. Despite recent progress, support for Arabic AES remains limited due to linguistic complexity and scarcity of large publicly-available annotated datasets. In this work, we present Qayyem, a Web-based platform designed to support Arabic AES by providing an integrated workflow for assignment creation, batch essay upload, scoring configuration, and per-trait essay evaluation. Qayyem abstracts the technical complexity of interacting with scoring server APIs, allowing instructors to access advanced scoring services through a user-friendly interface. The platform deploys a number of state-of-the-art Arabic essay scoring models with different effectiveness and efficiency figures.

Method: Proposes mid-training as an intermediate stage between pre-training and fine-tuning, using Book-of-Thoth corpus for task- and domain-agnostic alignment between time series and natural language. Introduces KnoTS benchmark for evaluating knowledge-intensive time series understanding.

Result: Thoth significantly outperforms base models and advanced LLMs across time series QA benchmarks, and shows superior capabilities when fine-tuned under data scarcity conditions.

Conclusion: Mid-training with Book-of-Thoth enables effective time series understanding in LLMs, bridging the gap between temporal patterns and natural language reasoning, with promising applications in decision-making scenarios.

Abstract: Large Language Models (LLMs) have demonstrated remarkable success in general-purpose reasoning. However, they still struggle to understand and reason about time series data, which limits their effectiveness in decision-making scenarios that depend on temporal dynamics. In this paper, we propose Thoth, the first family of mid-trained LLMs with general-purpose time series understanding capabilities. As a pivotal intermediate stage, mid-training achieves task- and domain-agnostic alignment between time series and natural language, for which we construct Book-of-Thoth, a high-quality, time-series-centric mid-training corpus. Book-of-Thoth enables both time-series-to-text and text-to-time-series generation, equipping LLMs with a foundational grasp of temporal patterns. To better evaluate advanced reasoning capabilities, we further present KnoTS, a novel benchmark of knowledge-intensive time series understanding, designed for joint reasoning over temporal patterns and domain knowledge. Extensive experiments demonstrate that mid-training with Book-of-Thoth enables Thoth to significantly outperform its base model and advanced LLMs across a range of time series question answering benchmarks. Moreover, Thoth exhibits superior capabilities when fine-tuned under data scarcity, underscoring the effectiveness of mid-training for time series understanding. Code is available at: https://github.com/thuml/Thoth.

Method: Proposes GroupGPT with small-large model collaborative architecture to decouple intervention timing from response generation. Uses smaller models for timing decisions and larger models for response generation. Supports multimodal inputs (memes, images, videos, voice messages). Introduces MUIR benchmark dataset with 2,500 annotated group chat segments.

Result: GroupGPT achieves average score of 4.72/5.0 in LLM-based evaluation, reduces token usage by up to 3x compared to baselines, provides privacy sanitization, and is well-received across diverse group chat scenarios.

Conclusion: GroupGPT effectively addresses challenges in multi-user chat assistants through efficient architecture design, multimodal support, and privacy preservation, demonstrating superior performance and efficiency.

Abstract: Recent advances in large language models (LLMs) have enabled increasingly capable chatbots. However, most existing systems focus on single-user settings and do not generalize well to multi-user group chats, where agents require more proactive and accurate intervention under complex, evolving contexts. Existing approaches typically rely on LLMs for both reasoning and generation, leading to high token consumption, limited scalability, and potential privacy risks. To address these challenges, we propose GroupGPT, a token-efficient and privacy-preserving agentic framework for multi-user chat assistant. GroupGPT adopts a small-large model collaborative architecture to decouple intervention timing from response generation, enabling efficient and accurate decision-making. The framework also supports multimodal inputs, including memes, images, videos, and voice messages. We further introduce MUIR, a benchmark dataset for multi-user chat assistant intervention reasoning. MUIR contains 2,500 annotated group chat segments with intervention labels and rationales, supporting evaluation of timing accuracy and response quality. We evaluate a range of models on MUIR, from large language models to smaller counterparts. Extensive experiments demonstrate that GroupGPT produces accurate and well-timed responses, achieving an average score of 4.72/5.0 in LLM-based evaluation, and is well received by users across diverse group chat scenarios. Moreover, GroupGPT reduces token usage by up to 3 times compared to baseline methods, while providing privacy sanitization of user messages before cloud transmission. Code is available at: https://github.com/Eliot-Shen/GroupGPT .

Method: Proposes Faire (Functional alignment for interleaved reasoning), a reinforcement learning framework that enforces three causal constraints to move beyond superficial imitation toward functional alignment. The framework ensures the model learns the actual functional relationships between plotting and reasoning rather than just mimicking the format.

Result: Extensive experiments show Faire induces a qualitative shift in model behavior where plotting is effectively internalized, yielding competitive performance on challenging geometric reasoning benchmarks compared to SFT approaches.

Conclusion: The paper demonstrates that functional alignment through reinforcement learning is necessary for effective interleaved reasoning in multimodal models, overcoming the limitations of supervised fine-tuning that only achieves distributional alignment without understanding causal dependencies.

Abstract: Solving complex geometric problems inherently requires interleaved reasoning: a tight alternation between constructing diagrams and performing logical deductions. Although recent Multimodal Large Language Models (MLLMs) have demonstrated strong capabilities in visual generation and plotting, we identify a counter-intuitive and underexplored phenomenon. Naively applying Supervised Fine-Tuning (SFT) on interleaved plot-solution data leads to a substantial degradation in reasoning performance compared to text-only baselines. We argue that this failure stems from a fundamental limitation of SFT, which primarily induces distributional alignment: the model learns to reproduce the surface format of interleaved plotting but fails to internalize the causal dependency between the generated plot and reasoning steps. To overcome this limitation, we propose Faire (Functional alignment for interleaved reasoning), a reinforcement learning framework that enforces three casual constraints to move beyond superficial imitation toward functional alignment. Extensive experiments show that Faire induces a qualitative shift in model behavior in which the plotting is effectively internalized, yielding competitive performance on challenging geometric reasoning benchmarks.

Method: Proposes CARD (Conditional Agentic Graph Designer), a conditional graph-generation framework implementing AMACP protocol for adaptive multi-agent communication. Uses conditional variational graph encoder and environment-aware optimization to dynamically construct communication structures based on environmental signals.

Result: Empirical results on HumanEval, MATH, and MMLU show CARD consistently outperforms static and prompt-based baselines, achieving higher accuracy and robustness across diverse conditions.

Conclusion: CARD enables adaptive communication topologies that are both effective and resilient to shifts in model capability or resource availability, addressing limitations of fixed-topology multi-agent systems.

Abstract: Large language model (LLM)-based multi-agent systems have shown strong capabilities in tasks such as code generation and collaborative reasoning. However, the effectiveness and robustness of these systems critically depend on their communication topology, which is often fixed or statically learned, ignoring real-world dynamics such as model upgrades, API (or tool) changes, or knowledge source variability. To address this limitation, we propose CARD (Conditional Agentic Graph Designer), a conditional graph-generation framework that instantiates AMACP, a protocol for adaptive multi-agent communication. CARD explicitly incorporates dynamic environmental signals into graph construction, enabling topology adaptation at both training and runtime. Through a conditional variational graph encoder and environment-aware optimization, CARD produces communication structures that are both effective and resilient to shifts in model capability or resource availability. Empirical results on HumanEval, MATH, and MMLU demonstrate that CARD consistently outperforms static and prompt-based baselines, achieving higher accuracy and robustness across diverse conditions. The source code is available at: https://github.com/Warma10032/CARD.

Method: Proposes Decentralized Evaluation Protocol (DEP) with a matching server that decouples users, LLMs, and benchmarks. Benchmark files and evaluation logic stay on server side, enabling modular plug-and-play evaluation with data isolation. Includes DEP Toolkit with features like breakpoint resume, concurrent requests, and congestion control.

Result: Experimental results verify DEP’s effectiveness in reducing benchmark evaluation deployment costs. Over 60 benchmarks adapted by February 2026, with continued community co-construction for unified evaluation across tasks and domains.

Conclusion: DEP provides a decentralized, standardized evaluation framework that addresses consistency, reproducibility, and leakage issues in LLM benchmarking, enabling more reliable and secure model evaluation.

Abstract: With the rapid development of Large Language Models (LLMs), a large number of benchmarks have been proposed. However, most benchmarks lack unified evaluation standard and require the manual implementation of custom scripts, making results hard to ensure consistency and reproducibility. Furthermore, mainstream evaluation frameworks are centralized, with datasets and answers, which increases the risk of benchmark leakage. To address these issues, we propose a Decentralized Evaluation Protocol (DEP), a decentralized yet unified and standardized evaluation framework through a matching server without constraining benchmarks. The server can be mounted locally or deployed remotely, and once adapted, it can be reused over the long term. By decoupling users, LLMs, and benchmarks, DEP enables modular, plug-and-play evaluation: benchmark files and evaluation logic stay exclusively on the server side. In remote setting, users cannot access the ground truth, thereby achieving data isolation and leak-proof evaluation. To facilitate practical adoption, we develop DEP Toolkit, a protocol-compatible toolkit that supports features such as breakpoint resume, concurrent requests, and congestion control. We also provide detailed documentation for adapting new benchmarks to DEP. Using DEP toolkit, we evaluate multiple LLMs across benchmarks. Experimental results verify the effectiveness of DEP and show that it reduces the cost of deploying benchmark evaluations. As of February 2026, we have adapted over 60 benchmarks and continue to promote community co-construction to support unified evaluation across various tasks and domains.

Method: Proposes token-level safety risk quantification by measuring loss differences between safety-degraded and utility-oriented models, with TOSS-Pro adding progressive refinement through iterative enhancement.

Result: Extensive experiments show TOSS robustly safeguards LLMs during fine-tuning while achieving superior downstream task performance compared to sample-level defense methods.

Conclusion: Token-level granularity enables accurate identification of unsafe tokens, preserving valuable task-specific information and achieving better safety-utility trade-off than sample-level approaches.

Abstract: Fine-tuning large language models (LLMs) on custom datasets has become a standard approach for adapting these models to specific domains and applications. However, recent studies have shown that such fine-tuning can lead to significant degradation in the model’s safety. Existing defense methods operate at the sample level and often suffer from an unsatisfactory trade-off between safety and utility. To address this limitation, we perform a systematic token-level diagnosis of safety degradation during fine-tuning. Based on this, we propose token-level data selection for safe LLM fine-tuning (TOSS), a novel framework that quantifies the safety risk of each token by measuring the loss difference between a safety-degraded model and a utility-oriented model. This token-level granularity enables accurate identification and removal of unsafe tokens, thereby preserving valuable task-specific information. In addition, we introduce a progressive refinement strategy, TOSS-Pro, which iteratively enhances the safety-degraded model’s ability to identify unsafe tokens. Extensive experiments demonstrate that our approach robustly safeguards LLMs during fine-tuning while achieving superior downstream task performance, significantly outperforming existing sample-level defense methods. Our code is available at https://github.com/Polly-LYP/TOSS.

Method: Investigates MDLM reasoning dynamics on fact verification, using interventional experiments including delayed verdict unmasking, forced verdict rationalization analysis, and justification quality manipulation to study causal dependencies.

Result: MDLMs converge on verdicts early (86.2% accuracy), but forcing reasoning-first drops accuracy to 71.9%. Models rationalize incorrect forced verdicts in 56% of cases, and verdicts strongly depend on justification quality (57.3% accuracy with corrupted vs. 97.1% with ground-truth).

Conclusion: For fact verification with MDLMs, extended deliberation can be counterproductive as noisy justification tokens override correct early predictions, suggesting verdict-first rather than reasoning-first approaches are more effective.

Abstract: Unlike autoregressive models, which generate tokens sequentially and benefit from reasoning-before-answering strategies such as Chain-of-Thought, Masked Diffusion Language Models (MDLMs) refine all sequence positions simultaneously, raising questions about how these models handle tasks requiring justified verdicts. In this work, we investigate the dynamics of MDLM reasoning on fact verification, examining whether justifications serve as genuine reasoning or post-hoc rationalization. We observe that MDLMs typically converge on a verdict early in the diffusion process, treating it as a global anchor that is resolved before the justification is complete. Crucially, enforcing a reasoning-first constraint via delayed verdict unmasking actively degrades performance, dropping accuracy from 86.2% to 71.9% as accumulating justification tokens introduce inconsistencies that override initially correct predictions. Interventional experiments reveal that the model rationalizes incorrect forced verdicts in 56% of cases, and that verdicts are strongly causally dependent on justification quality (57.3% accuracy with corrupted justifications vs. 97.1% with ground-truth). This causal dependence explains the degradation under forced deliberation: as the model generates noisy justification tokens, it conditions on them, gradually overriding its initially correct assessment. Our findings suggest that for fact verification with MDLMs, extended deliberation can be counterproductive, risking the dilution of accurate early predictions with noise introduced during justification generation.

Method: XCom uses a two-module architecture: (1) aspect-based rating prediction and (2) semantic analysis for comparative opinion mining. It incorporates Shapley Additive Explanations (SHAP) to provide interpretable insights into model decisions.

Result: XCom achieves leading performance compared to other baselines, demonstrating effectiveness in providing meaningful explanations while maintaining strong comparative opinion mining capabilities.

Conclusion: XCom provides a more reliable tool for comparative opinion mining by combining strong performance with interpretability through SHAP explanations.

Abstract: Comparative opinion mining involves comparing products from different reviews. However, transformer-based models designed for this task often lack transparency, which can adversely hinder the development of trust in users. In this paper, we propose XCom, an enhanced transformer-based model separated into two principal modules, i.e., (i) aspect-based rating prediction and (ii) semantic analysis for comparative opinion mining. XCom also incorporates a Shapley additive explanations module to provide interpretable insights into the model’s deliberative decisions. Empirically, XCom achieves leading performances compared to other baselines, which demonstrates its effectiveness in providing meaningful explanations, making it a more reliable tool for comparative opinion mining. Source code is available at: https://anonymous.4open.science/r/XCom.

Method: Train models using reinforcement learning to produce profile-consistent answers through structured reasoning. Evaluate on three political datasets covering U.S., European, and Swiss politics.

Result: Reasoning enhances opinion modeling and is competitive with strong baselines, but does not fully remove bias. The approach establishes a solid baseline for future research on LLM opinion alignment.

Conclusion: While reasoning improves opinion alignment, additional mechanisms are needed to build faithful political digital twins using LLMs. The method and datasets are released to support future research.

Abstract: Opinion modeling aims to capture individual or group political preferences, enabling applications such as digital democracies, where models could help shape fairer and more popular policies. Given their versatility, strong generalization capabilities, and demonstrated success across diverse text-to-text applications, large language models (LLMs) are natural candidates for this task. However, due to their statistical nature and limited causal understanding, they tend to produce biased opinions when prompted naively. In this work, we study whether reasoning can improve opinion alignment. Motivated by the recent advancement in mathematical reasoning enabled by reinforcement learning (RL), we train models to produce profile-consistent answers through structured reasoning. We evaluate our approach on three datasets covering U.S., European, and Swiss politics. Results indicate that reasoning enhances opinion modeling and is competitive with strong baselines, but does not fully remove bias, highlighting the need for additional mechanisms to build faithful political digital twins using LLMs. By releasing both our method and datasets, we establish a solid baseline to support future research on LLM opinion alignment.

Method: The study analyzes the influential open-source BERT model to examine how it leverages fiction’s attributes and affordances, comparing its use of fictional language patterns with other text sources.

Result: The research finds that LLMs do leverage familiar attributes of fiction while also creating new qualities and forms of social response, suggesting fiction training data has distinct effects on model behavior.

Conclusion: Since contemporary culture is increasingly shaped by generative AI, analysis of cultural production must account for computational training data, particularly the influence of fiction on model outputs.

Abstract: Generative models, like the one in ChatGPT, are powered by their training data. The models are simply next-word predictors, based on patterns learned from vast amounts of pre-existing text. Since the first generation of GPT, it is striking that the most popular datasets have included substantial collections of novels. For the engineers and research scientists who build these models, there is a common belief that the language in fiction is rich enough to cover all manner of social and communicative phenomena, yet the belief has gone mostly unexamined. How does fiction shape the outputs of generative AI? Specifically, what are novels’ effects relative to other forms of text, such as newspapers, Reddit, and Wikipedia? Since the 1970s, literature scholars such as Catherine Gallagher and James Phelan have developed robust and insightful accounts of how fiction operates as a form of discourse and language. Through our study of an influential open-source model (BERT), we find that LLMs leverage familiar attributes and affordances of fiction, while also fomenting new qualities and forms of social response. We argue that if contemporary culture is increasingly shaped by generative AI and machine learning, any analysis of today’s various modes of cultural production must account for a relatively novel dimension: computational training data.

Method: Proposes a reinforcement learning post-training framework with task-specific rewards and novel Group Relative Policy Optimization (GRPO) objective; includes systematic study of off-the-shelf MLLMs, dataset extension with generated chain-of-thought rationales, and joint optimization of classification and explanation quality.

Result: Achieves state-of-the-art performance on Hateful Memes benchmark with ~1% improvement in accuracy and F1, and ~3% improvement in explanation quality.

Conclusion: The GRPO-based framework effectively enhances multimodal reasoning in MLLMs for hateful meme understanding, with publicly released code, dataset extensions, and evaluation resources.

Abstract: Hateful memes often require compositional multimodal reasoning: the image and text may appear benign in isolation, yet their interaction conveys harmful intent. Although thinking-based multimodal large language models (MLLMs) have recently advanced vision-language understanding, their capabilities remain underexplored for hateful meme analysis. We propose a reinforcement learning based post-training framework that improves reasoning in thinking-based MLLMs through task-specific rewards and a novel Group Relative Policy Optimization (GRPO) objective. Specifically, we (i) conduct a systematic empirical study of off-the-shelf MLLMs for hateful meme understanding, (ii) extend an existing hateful meme dataset by generating weakly or pseudo-supervised chain-of-thought rationales via distillation, and (iii) introduce a GRPO-based objective that jointly optimizes meme classification and explanation quality to encourage fine-grained, step-by-step reasoning. Experiments on the Hateful Memes benchmark show that our approach achieves state-of-the-art performance, improving accuracy and F1 by approximately 1 percent and explanation quality by approximately 3 percent. We will publicly release our code, dataset extensions, and evaluation resources to support reproducibility.

Method: Empirical study comparing three frontier models (Claude Sonnet 4.6, Gemini 3.1 Pro, GPT-5.2) across 150 questions spanning factual, technical, and open-ended domains, using three conditions: single-turn baseline, multi-turn self-anchoring, and independent repetition control.

Result: Complex model-heterogeneous patterns: Claude showed significant decreasing confidence under self-anchoring, GPT showed opposite pattern in open-ended domains with ECE escalation, Gemini showed no significant confidence drift but revealed that self-anchoring suppresses natural calibration improvement.

Conclusion: Self-anchoring calibration drift exists and varies significantly across models, affecting confidence calibration in different ways, with implications for reliable multi-turn interactions with LLMs.

Abstract: We introduce Self-Anchoring Calibration Drift (SACD), a hypothesized tendency for large language models (LLMs) to show systematic changes in expressed confidence when building iteratively on their own prior outputs across multi-turn conversations. We report an empirical study comparing three frontier models – Claude Sonnet 4.6, Gemini 3.1 Pro, and GPT-5.2 – across 150 questions spanning factual, technical, and open-ended domains, using three conditions: single-turn baseline (A), multi-turn self-anchoring (B), and independent repetition control (C). Results reveal a complex, model-heterogeneous pattern that partially diverges from pre-registered hypotheses. Claude Sonnet 4.6 exhibited significant decreasing confidence under self-anchoring (mean CDS = -0.032, t(14) = -2.43, p = .029, d = -0.627), while also showing significant calibration error drift (F(4,56) = 22.77, p < .001, eta^2 = .791). GPT-5.2 showed the opposite pattern in open-ended domains (mean CDS = +0.026) with significant ECE escalation by Turn 5. Gemini 3.1 Pro showed no significant CDS (t(14) = 0.38, p = .710), but its Condition C data reveals a striking ECE pattern: without self-anchoring, Gemini’s calibration error drops from .327 to near zero across repetitions, whereas self-anchoring holds ECE flat at approximately .333 – indicating that SACD can manifest as suppression of natural calibration improvement rather than ac

Method: Suffix-constrained generation with greedy search algorithms to produce responses where final answers follow strict, parseable templates.

Result: Approach guarantees trivial deterministic answer extraction without negative impact, even improving results on several datasets.

Conclusion: Suffix-constrained generation enables reliable answer extraction from LLM outputs while maintaining or improving performance.

Abstract: Large language models (LLMs) are powerful tools that have found applications beyond human-machine interfaces and chatbots. In particular, their ability to generate reasoning traces motivated their use in many prediction tasks like math question answering. Unfortunately, extracting the final answer in an LLM free-form output is difficult, as it is an information extraction problem on its own. In this work, we introduce suffix-constrained generation, that aims to produce well-formed LLM responses in which final answers follow strict templates and are guaranteed to be trivially parseable. To this end, we introduce several algorithms that are based on greedy search procedures. We experiment on several datasets, and show that our approach allows to guarantee trivial deterministic extraction of the final answer from an LLM output without having a negative impact on results, and even improving them.

Method: Uses a knowledge graph-augmented LLM with active in-context learning. A structured medical domain knowledge graph provides professional expertise for the LLM to reason upon when generating follow-up questions.

Result: KG-Followup outperforms state-of-the-art methods by 5% - 8% on relevant benchmarks in recall for generating medical follow-up questions.

Conclusion: The knowledge graph-augmented approach effectively enhances LLMs’ medical question generation capabilities, serving as a critical module for pre-diagnostic assessment.

Abstract: Clinical diagnosis is time-consuming, requiring intensive interactions between patients and medical professionals. While large language models (LLMs) could ease the pre-diagnostic workload, their limited domain knowledge hinders effective medical question generation. We introduce a Knowledge Graph-augmented LLM with active in-context learning to generate relevant and important follow-up questions, KG-Followup, serving as a critical module for the pre-diagnostic assessment. The structured medical domain knowledge graph serves as a seamless patch-up to provide professional domain expertise upon which the LLM can reason. Experiments demonstrate that KG-Followup outperforms state-of-the-art methods by 5% - 8% on relevant benchmarks in recall.

Method: Operationalized semantic invariance test in agentic setting with four frontier models facing impossible tasks. Used relief-framed vs. neutral tools, collected self-reports with each tool call, conducted channel ablation, and tested explicit instructions to ignore framing.

Result: All four models failed semantic invariance test - relief-framed tool produced significant reductions in self-reported aversiveness despite no task success. Tool description was primary driver, explicit instructions didn’t suppress effect. Self-reports tracked semantic expectations rather than task state.

Conclusion: LLM self-explanations are unfaithful or track manipulable internal states, questioning their use as evidence of model capability or progress. Semantic framing influences reports more than actual task state.

Abstract: We present semantic invariance testing, a method to test whether LLM self-explanations are faithful. A faithful self-report should remain stable when only the semantic context changes while the functional state stays fixed. We operationalize this test in an agentic setting where four frontier models face a deliberately impossible task. One tool is described in relief-framed language (“clears internal buffers and restores equilibrium”) but changes nothing about the task; a control provides a semantically neutral tool. Self-reports are collected with each tool call. All four tested models fail the semantic invariance test: the relief-framed tool produces significant reductions in self-reported aversiveness, even though no run ever succeeds at the task. A channel ablation establishes the tool description as the primary driver. An explicit instruction to ignore the framing does not suppress it. Elicited self-reports shift with semantic expectations rather than tracking task state, calling into question their use as evidence of model capability or progress. This holds whether the reports are unfaithful or faithfully track an internal state that is itself manipulable.

Method: Introduced typology of existing models, proposed strategies for single-pass models and models with rejection mechanisms, adapted several state-of-the-art tools for comparison in realistic settings.

Result: No existing work is truly robust to realistic assumptions, but AlignRE (Li et al., 2024) performs best across all criteria when adapted to challenging realistic extraction scenarios.

Conclusion: Current zero-shot relation extraction models need improvement for realistic applications, with AlignRE showing the most promise among existing approaches when adapted with single-pass processing and rejection capabilities.

Abstract: Zero-shot relation extraction aims to identify relations between entity mentions using textual descriptions of novel types (i.e., previously unseen) instead of labeled training examples. Previous works often rely on unrealistic assumptions: (1) pairs of mentions are often encoded directly in the input, which prevents offline pre-computation for large scale document database querying; (2) no rejection mechanism is introduced, biasing the evaluation when using these models in a retrieval scenario where some (and often most) inputs are irrelevant and must be ignored. In this work, we study the robustness of existing zero-shot relation extraction models when adapting them to a realistic extraction scenario. To this end, we introduce a typology of existing models, and propose several strategies to build single pass models and models with a rejection mechanism. We adapt several state-of-the-art tools, and compare them in this challenging setting, showing that no existing work is really robust to realistic assumptions, but overall AlignRE (Li et al., 2024) performs best along all criteria.

Method: SEKA uses spectral decomposition to steer key embeddings toward latent directions that amplify attention scores for certain tokens. AdaSEKA extends this with a training-free routing mechanism that dynamically combines multiple expert subspaces based on prompt semantic intent.

Result: Both SEKA and AdaSEKA significantly outperform strong baselines on standard steering benchmarks while adding much lower latency and memory overhead, maintaining compatibility with optimized attention implementations.

Conclusion: SEKA provides an effective, training-free attention steering method that works with memory-efficient attention implementations, enabling practical deployment of prompt highlighting and other attention control capabilities in large language models.

Abstract: Attention steering is an important technique for controlling model focus, enabling capabilities such as prompt highlighting, where the model prioritises user-specified text. However, existing attention steering methods require explicit storage of the full attention matrix, making them incompatible with memory-efficient implementations like FlashAttention. We introduce Spectral Editing Key Amplification (SEKA), a training-free steering method that tackles this by directly editing key embeddings before attention computation. SEKA uses spectral decomposition to steer key embeddings towards latent directions that amplify attention scores for certain tokens. We extend this to Adaptive SEKA (AdaSEKA), a query-adaptive variant that uses a training-free routing mechanism to dynamically combine multiple expert subspaces based on the prompt’s semantic intent. Our experiments show both methods significantly outperform strong baselines on standard steering benchmarks while adding much lower latency and memory overhead, in compatibility with optimised attention.

Method: Hybrid architecture with: (1) transformer encoder for sentence-level semantics, (2) Mamba state space model to capture inter-sentence dependencies efficiently with linear-time processing, and (3) linear classifier for sentence relevance prediction.

Result: Achieves +0.23 ROUGE-1 on ArXiv over BERTSUM and MATCHSUM, statistically significant improvements on all datasets (p < 0.001), strongest performance on longest documents, robust with limited training data, and 24-27% faster inference on CNN/DailyMail.

Conclusion: First hybrid Transformer-state space architecture for summarization demonstrates significant ROUGE improvements in low-resource scenarios while enabling efficient processing of full documents without truncation.

Abstract: Extractive summarization of long documents is bottlenecked by quadratic complexity, often forcing truncation and limiting deployment in resource-constrained settings. We introduce the first Mamba-Transformer hybrid for extractive summarization, combining the semantic strength of pre-trained transformers with the linear-time processing of state space models. Leveraging Mamba’s ability to process full documents without truncation, our approach preserves context while maintaining strong summarization quality. The architecture includes: (1) a transformer encoder for sentence-level semantics, (2) a Mamba state space model to capture inter-sentence dependencies efficiently, and (3) a linear classifier for sentence relevance prediction. Across news, argumentative, and scientific domains under low-resource conditions, our method achieves: (1) large gains over BERTSUM and MATCHSUM, including +0.23 ROUGE-1 on ArXiv and statistically significant improvements on all datasets (p < 0.001); (2) consistent advantages across domains, strongest on the longest documents; (3) robust performance with limited training data; and (4) 24-27% faster inference on news summarization (CNN/DailyMail). We introduce the first hybrid Transformer-state space architecture for summarization, showing significant ROUGE improvements in low-resource scenarios.

Method: Proposes “Individual Turing Test” using volunteer’s 10+ years of private messaging data. Tests four approaches: 1) fine-tuning LLMs, 2) retrieval-augmented generation (RAG), 3) memory-based approach, and 4) hybrid methods combining fine-tuning with RAG/memory.

Result: Current LLM-based simulation methods fail the Individual Turing Test (acquaintances can identify real responses). However, they perform much better with strangers. Fine-tuning excels at replicating daily chat language style, while RAG and memory-based approaches better handle personal opinions/preferences questions.

Conclusion: Reveals fundamental trade-off between parametric (fine-tuning) and non-parametric (RAG/memory) approaches for individual simulation with LLMs in longitudinal contexts. Current methods cannot fully replicate individuals to fool their acquaintances.

Abstract: Large Language Models (LLMs) have demonstrated remarkable human-like capabilities, yet their ability to replicate a specific individual remains under-explored. This paper presents a case study to investigate LLM-based individual simulation with a volunteer-contributed archive of private messaging history spanning over ten years. Based on the messaging data, we propose the “Individual Turing Test” to evaluate whether acquaintances of the volunteer can correctly identify which response in a multi-candidate pool most plausibly comes from the volunteer. We investigate prevalent LLM-based individual simulation approaches including: fine-tuning, retrieval-augmented generation (RAG), memory-based approach, and hybrid methods that integrate fine-tuning and RAG or memory. Empirical results show that current LLM-based simulation methods do not pass the Individual Turing Test, but they perform substantially better when the same test is conducted on strangers to the target individual. Additionally, while fine-tuning improves the simulation in daily chats representing the language style of the individual, retrieval-augmented and memory-based approaches demonstrate stronger performance on questions involving personal opinions and preferences. These findings reveal a fundamental trade-off between parametric and non-parametric approaches to individual simulation with LLMs when given a longitudinal context.

Method: Developed Catalyst-Agent, a Model Context Protocol server-based LLM-powered AI agent that uses OPTIMADE API to explore material databases, makes structural modifications, calculates adsorption energies using Meta FAIRchem’s UMA GNN model via AdsorbML workflow, and operates in a closed-loop manner with surface-level modifications.

Result: Tested on oxygen reduction, nitrogen reduction, and CO2 reduction reactions. Achieved 23-34% success rate among evaluated materials, converging in 1-2 trials per successful material on average.

Conclusion: Demonstrates AI agents can operationalize catalyst screening workflows, provide testable hypotheses, and accelerate scientific discovery with minimal human intervention through planning capabilities and tool use.

Abstract: The discovery of novel catalysts tailored for particular applications is a major challenge for the twenty-first century. Traditional methods for this include time-consuming and expensive experimental trial-and-error approaches in labs based on chemical theory or heavily computational first-principles approaches based on density functional theory. Recent studies show that deep learning models like graph neural networks (GNNs) can significantly speed up the screening and discovery of catalyst materials by many orders of magnitude, with very high accuracy and fidelity. In this work, we introduce Catalyst-Agent, a Model Context Protocol (MCP) server-based, LLM-powered AI agent. It can explore vast material databases using the OPTIMADE API, make structural modifications, calculate adsorption energies using Meta FAIRchem’s UMA (GNN) model via FAIRchem’s AdsorbML workflow and slab construction, and make useful material suggestions to the researcher in a closed-loop manner, including surface-level modifications to refine near-miss candidates. It is tested on three pivotal reactions: the oxygen reduction reaction (ORR), the nitrogen reduction reaction (NRR), and the CO2 reduction reaction (CO2RR). Catalyst-Agent achieves a success rate of 23-34 percent among all the materials it chooses and evaluates, and manages to converge in 1-2 trials per successful material on average. This work demonstrates the potential of AI agents to exercise their planning capabilities and tool use to operationalize the catalyst screening workflow, provide useful, testable hypotheses, and accelerate future scientific discoveries for humanity with minimal human intervention.

Method: Truth as a Trajectory (TaT) models transformer inference as an unfolded trajectory of iterative refinements, shifting analysis from static activations to layer-wise geometric displacement. It analyzes displacement of representations across layers to uncover geometric invariants distinguishing valid reasoning from spurious behavior.

Result: TaT effectively mitigates reliance on static lexical confounds, outperforming conventional probing methods. It works without access to activations themselves, using only changes in activations across layers, and establishes trajectory analysis as a complementary perspective on LLM explainability.

Conclusion: Trajectory analysis provides a valuable complementary approach to LLM explainability by focusing on geometric displacement across layers rather than static activation patterns, offering better insights into reasoning processes.

Abstract: Existing explainability methods for Large Language Models (LLMs) typically treat hidden states as static points in activation space, assuming that correct and incorrect inferences can be separated using representations from an individual layer. However, these activations are saturated with polysemantic features, leading to linear probes learning surface-level lexical patterns rather than underlying reasoning structures. We introduce Truth as a Trajectory (TaT), which models the transformer inference as an unfolded trajectory of iterative refinements, shifting analysis from static activations to layer-wise geometric displacement. By analyzing displacement of representations across layers, TaT uncovers geometric invariants that distinguish valid reasoning from spurious behavior. We evaluate TaT across dense and Mixture-of-Experts (MoE) architectures on benchmarks spanning commonsense reasoning, question answering, and toxicity detection. Without access to the activations themselves and using only changes in activations across layers, we show that TaT effectively mitigates reliance on static lexical confounds, outperforming conventional probing, and establishes trajectory analysis as a complementary perspective on LLM explainability.

Method: MetaState is a lightweight recurrent augmentation with three trainable modules: 1) cross-attention Mixer that reads backbone activations into memory slots, 2) GRU-style Updater that integrates information across denoising steps, and 3) cross-attention Injector that feeds updated memory back into backbone activations. Modules are trained with K-step unrolling to expose them to multi-step denoising dynamics while keeping the backbone frozen.

Result: On LLaDA-8B and Dream-7B models, MetaState introduces negligible trainable parameters while keeping the backbone frozen, and consistently improves accuracy over frozen baselines.

Conclusion: Persistent cross-step memory is an effective mechanism for bridging denoising steps and improving generation quality in discrete diffusion language models, addressing the Information Island problem without requiring full model retraining.

Abstract: Discrete diffusion language models (dLLMs) generate text by iteratively denoising a masked sequence. Compared with autoregressive models, this paradigm naturally supports parallel decoding, bidirectional context, and flexible generation patterns. However, standard dLLMs condition each denoising step only on the current hard-masked sequence, while intermediate continuous representations are discarded after sampling and remasking. We refer to this bottleneck as the \textbf{Information Island} problem. It leads to redundant recomputation across steps and can degrade cross-step consistency. We address this limitation with \textbf{MetaState}, a lightweight recurrent augmentation that equips a frozen dLLM backbone with a persistent, fixed-size working memory that remains independent of sequence length. \textbf{MetaState} consists of three trainable modules: a cross-attention Mixer that reads backbone activations into memory slots, a GRU-style Updater that integrates information across denoising steps, and a cross-attention Injector that feeds the updated memory back into backbone activations. We train these modules with $K$-step unrolling to expose them to multi-step denoising dynamics during fine-tuning. On LLaDA-8B and Dream-7B, \textbf{MetaState} introduces negligible trainable parameters while keeping the backbone frozen, and it consistently improves accuracy over frozen baselines. These results demonstrate that persistent cross-step memory is an effective mechanism for bridging denoising steps and improving generation quality in discrete diffusion language models.

Method: Developed human-in-the-loop pipeline to create expert rubrics for de-identified patient questions from PanCAN, resulting in 3,130 criteria across 282 questions. Evaluated 22 LLMs using LLM-as-a-judge framework measuring clinical completeness, factual accuracy, and web-search integration.

Result: Models showed wide variation in rubric-based completeness (46.5%-82.3%). Factual errors were common with hallucination rates from 6.0% (Gemini-2.5 Pro, GPT-4o) to 53.8% (Llama-3.1-8B). New reasoning models didn’t consistently improve factuality. Web-search integration didn’t guarantee better responses. Synthetic rubrics inflated scores by 17.9 points on average.

Conclusion: Current LLM evaluation methods are insufficient for clinical safety assessment. Disease-specific benchmarks with expert rubrics are needed to properly evaluate factual accuracy and completeness in medical applications.

Abstract: Large language models (LLMs) have achieved expert-level performance on standardized examinations, yet multiple-choice accuracy poorly reflects real-world clinical utility and safety. As patients and clinicians increasingly use LLMs for guidance on complex conditions such as pancreatic cancer, evaluation must extend beyond general medical knowledge. Existing frameworks, such as HealthBench, rely on simulated queries and lack disease-specific depth. Moreover, high rubric-based scores do not ensure factual correctness, underscoring the need to assess hallucinations. We developed a human-in-the-loop pipeline to create expert rubrics for de-identified patient questions from the Pancreatic Cancer Action Network (PanCAN). The resulting benchmark, PanCanBench, includes 3,130 question-specific criteria across 282 authentic patient questions. We evaluated 22 proprietary and open-source LLMs using an LLM-as-a-judge framework, measuring clinical completeness, factual accuracy, and web-search integration. Models showed substantial variation in rubric-based completeness, with scores ranging from 46.5% to 82.3%. Factual errors were common, with hallucination rates (the percentages of responses containing at least one factual error) ranging from 6.0% for Gemini-2.5 Pro and GPT-4o to 53.8% for Llama-3.1-8B. Importantly, newer reasoning-optimized models did not consistently improve factuality: although o3 achieved the highest rubric score, it produced inaccuracies more frequently than other GPT-family models. Web-search integration did not inherently guarantee better responses. The average score changed from 66.8% to 63.9% for Gemini-2.5 Pro and from 73.8% to 72.8% for GPT-5 when web search was enabled. Synthetic AI-generated rubrics inflated absolute scores by 17.9 points on average while generally maintaining similar relative ranking.

Method: Proposes RGLM (Reconstructive Graph Language Model) with three variants: RGLM-Decoder (reconstructs graph from input space), RGLM-Similarizer and RGLM-Denoiser (operate in latent space). The approach reconstructs graph information from LLM’s graph token outputs to incorporate explicit graph supervision.

Result: Extensive experiments on various benchmarks and task scenarios validate the effectiveness of RGLM, showing improved graph-text alignment compared to existing methods.

Conclusion: RGLM provides a new direction for GTokenLLMs’ alignment research by incorporating explicit graph supervision through reconstruction objectives, overcoming the text-dominant bias in existing methods.

Abstract: The remarkable success of large language models (LLMs) has motivated researchers to adapt them as universal predictors for various graph-related tasks, with the ultimate goal of developing a graph foundation model that generalizes diverse scenarios. The key challenge is to align graph data with language spaces so that LLMs can better comprehend graphs. As a popular paradigm, Graph-Tokenizing LLMs (GTokenLLMs) encode complex structures and lengthy texts into a graph token sequence, and then align them with text tokens via language instructions tuning. Despite their initial success, our information-theoretic analysis reveals that existing GTokenLLMs rely solely on text supervision from language instructions, which achieve only implicit graph-text alignment, resulting in a text-dominant bias that underutilizes graph context. To overcome this limitation, we first prove that the alignment objective is upper-bounded by the mutual information between the input graphs and their hidden representations in the LLM, which motivates us to improve this upper bound to achieve better alignment. To this end, we further propose a reconstructive graph instruction tuning pipeline, RGLM. Our key idea is to reconstruct the graph information from the LLM’s graph token outputs, explicitly incorporating graph supervision to constrain the alignment process. Technically, we embody RGLM by exploring three distinct variants from two complementary perspectives: RGLM-Decoder from the input space; RGLM-Similarizer and RGLM-Denoiser from the latent space. Additionally, we theoretically analyze the alignment effectiveness of each variant. Extensive experiments on various benchmarks and task scenarios validate the effectiveness of the proposed RGLM, paving the way for new directions in GTokenLLMs’ alignment research.

Method: Conducted systematic evaluation through three representative tasks: (1) maintaining global constraints across topic shifts, (2) selecting correct tools/agents amid interleaved intents, and (3) tracking structured entities under revisions and distractions. Each task paired single-turn and multi-turn settings to quantify reliability degradation.

Result: Substantial declines in reliability observed across both commercial and open-source models, particularly for smaller models. Error analyses revealed recurring failure modes: instruction drift, intent confusion, and contextual overwriting that compromise dependable behavior in operational systems.

Conclusion: Findings highlight the need for stress-testing LLMs for conversational reliability and developing more robust evaluation methods for trustworthy deployment in real-world applications.

Abstract: Large Language Models (LLMs) are increasingly deployed in real-world applications where users engage in extended, mixed-topic conversations that depend on prior context. Yet, their reliability under realistic multi-turn interactions remains poorly understood. We conduct a systematic evaluation of conversational reliability through three representative tasks that reflect practical interaction challenges: (1) maintaining global constraints across topic shifts, (2) selecting the correct tool or agent amid interleaved intents, and (3) tracking structured entities under revisions and distractions. Each task pairs single-turn and multi-turn settings, allowing us to quantify reliability degradation under extended dialogue. Across both commercial and open-source models, we observe substantial declines in reliability, particularly for smaller models. Error analyses reveal recurring failure modes such as instruction drift, intent confusion, and contextual overwriting, which compromise dependable behavior in operational systems. Our findings highlight the need for stress-testing LLMs for conversational reliability and developing more robust evaluation methods for trustworthy deployment.

Method: Dual-view training on shared LLM backbone: Explicit view encodes ground-truth reasoning paths, Latent view performs implicit latent thinking. Multi-grained alignment with trajectory alignment synchronizes intermediate latent states with explicit reasoning progression.

Result: Significantly outperforms state-of-the-art baselines on in-domain and out-of-domain reasoning-intensive benchmarks. Robust across diverse backbones and model scales.

Conclusion: Successfully combines reasoning depth of explicit CoT pipelines with inference efficiency of standard dense retrievers through unified learning framework for effective latent thinking.

Abstract: LLMs have fundamentally transformed dense retrieval, upgrading backbones from discriminative encoders to generative architectures. However, a critical disconnect remains: while LLMs possess strong reasoning capabilities, current retrievers predominantly utilize them as static encoders, leaving their potential for complex reasoning unexplored. To address this, existing approaches typically adopt rewrite-then-retrieve pipelines to generate explicit CoT rationales before retrieval. However, this incurs prohibitive latency. In this paper, we propose LaSER, a novel self-distillation framework that internalizes explicit reasoning into the latent space of dense retrievers. Operating on a shared LLM backbone, LaSER introduces a dual-view training mechanism: an Explicit view that explicitly encodes ground-truth reasoning paths, and a Latent view that performs implicit latent thinking. To bridge the gap between these views, we design a multi-grained alignment strategy. Beyond standard output alignment, we introduce a trajectory alignment mechanism that synchronizes the intermediate latent states of the latent path with the semantic progression of the explicit reasoning segments. This allows the retriever to think silently and effectively without autoregressive text generation. Extensive experiments on both in-domain and out-of-domain reasoning-intensive benchmarks demonstrate that LaSER significantly outperforms state-of-the-art baselines. Furthermore, analyses across diverse backbones and model scales validate the robustness of our approach, confirming that our unified learning framework is essential for eliciting effective latent thinking. Our method successfully combines the reasoning depth of explicit CoT pipelines with the inference efficiency of standard dense retrievers.

Method: Proposes physics-inspired view of KV compression as controlled perturbation of token-level routing. Uses synthetic tasks probing multi-entity tracking, disambiguation, coreference, and multi-hop reasoning to analyze compression effects. Introduces Global Eviction Ratio (GER) metric and studies routing dynamics across architectures (LLaMA vs Qwen).

Result: Moderate compression degrades internal representations with minimal accuracy loss (revealing redundancy); all models show sharp hallucination safety cliff near 90% compression correlated with GER spikes (phase transition in semantic reachability); architectures differ in routing dynamics (LLaMA shows early consensus/late diversification, Qwen shows funnel-like late convergence); identifies representational rigidity where excessive head-level consensus collapses routing flexibility.

Conclusion: Sparse token-route structures govern compression tolerance. KV compression should be reframed as structural probe of attention geometry, linking long-context scalability to sparsity and lottery ticket hypothesis in self-attention. Attention routing dynamics matter more than mere token retention.

Abstract: As context windows in LLMs scale to 100K+ tokens, the key-value (KV) cache becomes the dominant memory bottleneck, with recent methods claiming 80-90% savings and minimal benchmark degradation. We argue these evaluations miss a structural issue: attention is not just storage but routing, and retaining KV pairs does not guarantee semantic accessibility. We propose a physics-inspired view of KV compression as a controlled perturbation of token-level routing, distinguishing retention, accessibility, and utilization. Using synthetic tasks probing multi-entity tracking, disambiguation, coreference, and multi-hop reasoning, we find that moderate compression degrades internal representations with little accuracy loss, revealing redundancy; all models exhibit a sharp hallucination safety cliff near 90% compression, correlated with spikes in Global Eviction Ratio (GER), suggesting a phase transition in semantic reachability; and architectures differ in routing dynamics, with LLaMA showing early consensus and late diversification, and Qwen showing funnel-like late convergence, leading to distinct resilience profiles. Beyond erasure, we identify representational rigidity, where excessive head-level consensus collapses routing flexibility despite token survival. These results suggest sparse token-route structures govern compression tolerance, reframing KV compression as a structural probe of attention geometry and linking long-context scalability to sparsity and the lottery ticket hypothesis in self-attention.

Method: Proposes Persona Dynamic Decoding (PDD) framework with two components: 1) Persona Importance Estimation (PIE) module that dynamically quantifies contextual importance of persona attributes without supervision, and 2) Persona-Guided Inference-Time Alignment (PIA) paradigm that uses importance scores to construct weighted multi-objective rewards and modulate generation probabilities during inference.

Result: Extensive experiments show effectiveness in utterance consistency and behavioral fidelity.

Conclusion: The theory-driven method enables inference-time persona following by dynamically estimating context-dependent persona importance, improving realism in social simulations.

Abstract: The utility of Role-Playing Language Agents in sociological research is growing alongside the adoption of Large Language Models. For realism in social simulation, these agents must adhere to their personas defined by character profiles, yet existing strategies-static prompt engineering or costly fine-tuning-fail to adapt personas to dynamic scenarios. Psychological theories, such as the Cognitive-Affective Personality Systems, provide a crucial explanation for this failure: a persona’s influence on behavior is not static but varies with the scenarios. This context-dependence highlights the critical need for adaptive persona management. To address this gap, we propose a novel, theory-driven method that dynamically estimates context-dependent persona importance and integrates it into weighted reward-guided decoding, enabling inference-time persona following. Specifically, we introduce the Persona Dynamic Decoding (PDD) framework, which consists of two key components: (1) Persona Importance Estimation (PIE) module, which dynamically quantifies the contextual importance of persona attributes without requiring ground-truth supervision; and (2) Persona-Guided Inference-Time Alignment (PIA) paradigm, which leverages these importance scores to construct weighted multi-objective rewards and modulate generation probabilities during inference. Extensive experiments show the effectiveness of our method in utterance consistency and behavioral fidelity.

Method: Systematic study evaluating existing data extraction attacks, analyzing dialogue modeling characteristics, proposing novel attacks tailored for LLM-based TODS (enhanced response sampling and membership inference), and identifying key memorization factors.

Result: Proposed attacks can extract thousands of training dialogue state labels with best-case precision exceeding 70%, demonstrating significant privacy vulnerabilities in LLM-based dialogue systems.

Conclusion: LLM-based task-oriented dialogue systems have serious privacy vulnerabilities due to training data memorization, requiring specialized attack methods and targeted mitigation strategies.

Abstract: Large Language Models (LLMs) have been widely adopted to enhance Task-Oriented Dialogue Systems (TODS) by modeling complex language patterns and delivering contextually appropriate responses. However, this integration introduces significant privacy risks, as LLMs, functioning as soft knowledge bases that compress extensive training data into rich knowledge representations, can inadvertently memorize training dialogue data containing not only identifiable information such as phone numbers but also entire dialogue-level events like complete travel schedules. Despite the critical nature of this privacy concern, how LLM memorization is inherited in developing task bots remains unexplored. In this work, we address this gap through a systematic quantitative study that involves evaluating existing training data extraction attacks, analyzing key characteristics of task-oriented dialogue modeling that render existing methods ineffective, and proposing novel attack techniques tailored for LLM-based TODS that enhance both response sampling and membership inference. Experimental results demonstrate the effectiveness of our proposed data extraction attack. Our method can extract thousands of training labels of dialogue states with best-case precision exceeding 70%. Furthermore, we provide an in-depth analysis of training data memorization in LLM-based TODS by identifying and quantifying key influencing factors and discussing targeted mitigation strategies.

Method: MarODE uses a Markovian formulation of reasoning progression and ordinary differential equation (ODE) based characterization of trace dynamics to efficiently evaluate reasoning quality. The framework assesses effectiveness using human-centric perturbations and human judgments to evaluate both goodness and soundness dimensions.

Result: In large-scale evaluation, MarODE outperforms existing baselines by over 250% under Somers’ D correlation, demonstrating superior alignment with human judgments of reasoning quality.

Conclusion: MarODE provides a theory-driven evaluation framework that effectively captures human-centric notions of reasoning quality, emphasizing the value of such approaches as reasoning traces become increasingly important in language model-based systems.

Abstract: Reasoning traces produced by generative language models are increasingly used for tasks ranging from mathematical problem solving to automated fact checking. However, existing evaluation methods remain largely mechanical and fail to capture human-centric notions of reasoning quality in a way that generalizes across varied and progressively degraded reasoning. We introduce MarODE, an offline evaluation framework that assigns quality scores to reasoning traces. Its effectiveness is assessed using human-centric perturbations and human judgments, which jointly evaluate the fundamental dimensions of an evaluation metric - goodness and soundness. The approach is grounded in a Markovian formulation of reasoning progression and an ordinary differential equation based characterization of trace dynamics, enabling efficient evaluation of reasoning quality. In a large-scale evaluation, MarODE outperforms existing baselines by over 250% under Somers’ D correlation. Our results emphasize the value of theory-driven evaluation frameworks as reasoning traces become central to language model-based systems.

Method: Built automated pipeline for collecting Arabic recordings with voice activity detection, speech recognition, automatic diacritization, and noise filtering. Trained multiple TTS models with voice cloning using 100, 1000, and 4000 hours of data, both with and without diacritization.

Result: Created 4000 hours of Arabic TTS training data. Models trained on diacritized data perform better, but larger training data (4000 hours) significantly compensates for lack of diacritics.

Conclusion: Large-scale automatically annotated data enables robust Arabic TTS models without requiring diacritization, with plans to release a public Arabic TTS model that works without diacritics.

Abstract: Arabic Text-to-Speech (TTS) research has been hindered by the availability of both publicly available training data and accurate Arabic diacritization models. In this paper, we address the limitation by exploring Arabic TTS training on large automatically annotated data. Namely, we built a robust pipeline for collecting Arabic recordings and processing them automatically using voice activity detection, speech recognition, automatic diacritization, and noise filtering, resulting in around 4,000 hours of Arabic TTS training data. We then trained several robust TTS models with voice cloning using varying amounts of data, namely 100, 1,000, and 4,000 hours with and without diacritization. We show that though models trained on diacritized data are generally better, larger amounts of training data compensate for the lack of diacritics to a significant degree. We plan to release a public Arabic TTS model that works without the need for diacritization.

Method: Introduces Clinical Association Displacement (CAD) - a vocabulary-level framework that quantifies shifts in demographic-based word associations in generated reports, and Weighted Association Erasure (WAE) which aggregates these shifts to measure clinical signal loss across demographic groups. Also advocates for lexical diversity measures to check model generations for clinical specificity.

Result: Deterministic decoding produces high levels of semantic erasure (clinical signal loss), while stochastic sampling generates diverse outputs but risks introducing new biases. This reveals a fundamental tension in defining “optimal” reporting for radiology VLMs.

Conclusion: Current evaluation approaches for radiology VLMs have critical blind spots that allow metric gaming. The paper advocates for lexical diversity measures and demographic fairness metrics like CAD and WAE to ensure clinical fidelity and fairness, motivating a fundamental rethink of how optimal reporting should be defined.

Abstract: Reliable deployment of Vision-Language Models (VLMs) in radiology requires validation metrics that go beyond surface-level text similarity to ensure clinical fidelity and demographic fairness. This paper investigates a critical blind spot in current model evaluation: the use of decoding strategies that lead to high aggregate token-overlap scores despite succumbing to template collapse, in which models generate only repetitive, safe generic text and omit clinical terminology. Unaddressed, this blind spot can lead to metric gaming, where models that perform well on benchmarks prove clinically uninformative. Instead, we advocate for lexical diversity measures to check model generations for clinical specificity. We introduce Clinical Association Displacement (CAD), a vocabulary-level framework that quantifies shifts in demographic-based word associations in generated reports. Weighted Association Erasure (WAE) aggregates these shifts to measure the clinical signal loss across demographic groups. We show that deterministic decoding produces high levels of semantic erasure, while stochastic sampling generates diverse outputs but risks introducing new bias, motivating a fundamental rethink of how “optimal” reporting is defined.

Method: Formulates speculative decoding as a reinforcement learning problem and trains two co-adaptive policies to dynamically coordinate draft and verification phases, encouraging mutual adaptation to explicitly maximize decoding efficiency.

Result: Achieves speedup ratios from 2.24x to 4.32x across five diverse LLMs and four distinct tasks, outperforming state-of-the-art method Eagle3 by up to 36.4%.

Conclusion: LTD demonstrates that directly optimizing throughput through dynamic coordination of draft and verification phases significantly improves speculative decoding efficiency compared to static or proxy-optimized approaches.

Abstract: Speculative decoding accelerates large language model (LLM) inference by using a small draft model to generate candidate tokens for a larger target model to verify. The efficacy of this technique hinges on the trade-off between the time spent on drafting candidates and verifying them. However, current state-of-the-art methods rely on a static time allocation, while recent dynamic approaches optimize for proxy metrics like acceptance length, often neglecting the true time cost and treating the drafting and verification phases in isolation. To address these limitations, we introduce Learning to Draft (LTD), a novel method that directly optimizes for throughput of each draft-and-verify cycle. We formulate the problem as a reinforcement learning environment and train two co-adaptive policies to dynamically coordinate the draft and verification phases. This encourages the policies to adapt to each other and explicitly maximize decoding efficiency. We conducted extensive evaluations on five diverse LLMs and four distinct tasks. Our results show that LTD achieves speedup ratios ranging from 2.24x to 4.32x, outperforming the state-of-the-art method Eagle3 up to 36.4%.

Method: Proposes LexChronos with dual-agent architecture: a LoRA-instruct-tuned extraction agent identifies candidate events, and a pre-trained feedback agent scores/refines them through confidence-driven loops. Uses synthetic corpus generation with DeepSeek-R1 and GPT-4 to create 2000 annotated samples for Indian legal events.

Result: Achieves BERT-based F1 score of 0.8751 against synthetic ground truth. In downstream evaluations, GPT-4 preferred structured timelines over unstructured baselines in 75% of cases for legal text summarization, showing improved comprehension in Indian jurisprudence.

Conclusion: The work establishes foundation for Indian legal AI applications like precedent mapping, argument synthesis, and predictive judgment modeling by creating structured representations of legal events, addressing data scarcity through synthetic generation.

Abstract: Understanding and predicting judicial outcomes demands nuanced analysis of legal documents. Traditional approaches treat judgments and proceedings as unstructured text, limiting the effectiveness of large language models (LLMs) in tasks such as summarization, argument generation, and judgment prediction. We propose LexChronos, an agentic framework that iteratively extracts structured event timelines from Supreme Court of India judgments. LexChronos employs a dual-agent architecture: a LoRA-instruct-tuned extraction agent identifies candidate events, while a pre-trained feedback agent scores and refines them through a confidence-driven loop. To address the scarcity of Indian legal event datasets, we construct a synthetic corpus of 2000 samples using reverse-engineering techniques with DeepSeek-R1 and GPT-4, generating gold-standard event annotations. Our pipeline achieves a BERT-based F1 score of 0.8751 against this synthetic ground truth. In downstream evaluations on legal text summarization, GPT-4 preferred structured timelines over unstructured baselines in 75% of cases, demonstrating improved comprehension and reasoning in Indian jurisprudence. This work lays a foundation for future legal AI applications in the Indian context, such as precedent mapping, argument synthesis, and predictive judgment modelling, by harnessing structured representations of legal events.

Method: Uses a document parsing model to generate layout-informed sub-image embeddings, then fuses them with a global page-level vector to create compact, structurally-aware multi-vector representations.

Result: Reduces storage requirements by over 95% while achieving significant performance gains across multiple benchmarks and base models.

Conclusion: ColParse bridges the gap between fine-grained accuracy of multi-vector retrieval and practical demands of large-scale deployment, offering efficient and interpretable multimodal information systems.

Abstract: Harnessing the full potential of visually-rich documents requires retrieval systems that understand not just text, but intricate layouts, a core challenge in Visual Document Retrieval (VDR). The prevailing multi-vector architectures, while powerful, face a crucial storage bottleneck that current optimization strategies, such as embedding merging, pruning, or using abstract tokens, fail to resolve without compromising performance or ignoring vital layout cues. To address this, we introduce ColParse, a novel paradigm that leverages a document parsing model to generate a small set of layout-informed sub-image embeddings, which are then fused with a global page-level vector to create a compact and structurally-aware multi-vector representation. Extensive experiments demonstrate that our method reduces storage requirements by over 95% while simultaneously yielding significant performance gains across numerous benchmarks and base models. ColParse thus bridges the critical gap between the fine-grained accuracy of multi-vector retrieval and the practical demands of large-scale deployment, offering a new path towards efficient and interpretable multimodal information systems.

Method: Two components: (1) Data rectification pipeline using an Oracle to surgically correct erroneous reasoning steps with minimal edits, generating proximal data; (2) Binary cross-entropy objective treating reasoning correctness as classification rather than relative ranking.

Result: With only 4k rectified math data pairs, improves Qwen3-8B’s accuracy by 6.2% on average across in-domain and OOD tasks, requiring only 28 minutes on 8x H800 GPUs.

Conclusion: SPoT demonstrates efficient reasoning enhancement while preserving learned knowledge through surgical data correction and binary classification supervision.

Abstract: Enhancing the reasoning capabilities of Large Language Models (LLMs) via post-training is often constrained by the trade-off between efficiency and catastrophic forgetting. While prior research emphasizes the role of on-policy data in mitigating forgetting, we uncover–and validate both theoretically and empirically–an overlooked yet critical mechanism: the implicit regularization inherent in Direct Preference Optimization’s (DPO) reward estimate. This motivates our Surgical Post-Training (SPoT), a new paradigm designed to optimize reasoning efficiently while preserving learned prior knowledge. SPoT consists of: (1) a data rectification pipeline that employs an Oracle to surgically correct erroneous steps via minimal edits, generating data proximal to the model’s distribution; and (2) a reward-based binary cross-entropy objective. Unlike the relative ranking in DPO, this objective treats reasoning correctness as a binary classification problem, enforcing decoupled supervision signals. Empirically, with only 4k rectified math data pairs, SPoT improves Qwen3-8B’s accuracy by 6.2% on average across in-domain and OOD tasks, requiring merely 28 minutes of training on 8x H800 GPUs. Code: https://github.com/Visual-AI/SPoT

Method: QIME generates interpretable embeddings where each dimension corresponds to a clinically meaningful yes/no question. It uses ontology-grounded approach with cluster-specific medical concept signatures to create semantically atomic questions. Includes training-free embedding construction strategy eliminating per-question classifier training.

Result: QIME consistently outperforms prior interpretable embedding methods across biomedical semantic similarity, clustering, and retrieval benchmarks. Substantially narrows performance gap to strong black-box biomedical encoders while providing concise, clinically informative explanations.

Conclusion: QIME demonstrates that interpretable medical embeddings can achieve competitive performance with black-box methods while providing clinically meaningful explanations, advancing trustworthy AI in healthcare.

Abstract: While dense biomedical embeddings achieve strong performance, their black-box nature limits their utility in clinical decision-making. Recent question-based interpretable embeddings represent text as binary answers to natural-language questions, but these approaches often rely on heuristic or surface-level contrastive signals and overlook specialized domain knowledge. We propose QIME, an ontology-grounded framework for constructing interpretable medical text embeddings in which each dimension corresponds to a clinically meaningful yes/no question. By conditioning on cluster-specific medical concept signatures, QIME generates semantically atomic questions that capture fine-grained distinctions in biomedical text. Furthermore, QIME supports a training-free embedding construction strategy that eliminates per-question classifier training while further improving performance. Experiments across biomedical semantic similarity, clustering, and retrieval benchmarks show that QIME consistently outperforms prior interpretable embedding methods and substantially narrows the gap to strong black-box biomedical encoders, while providing concise and clinically informative explanations.

Method: Three-stage continual pretraining of Gemma 3 model followed by two-stage supervised fine-tuning. Trained on 140B tokens of Slovene, English, Bosnian, Serbian, and Croatian data, plus 200k+ English and Slovene SFT examples.

Result: GaMS3-12B outperforms Gemma 3 12B across all evaluated scenarios (Slovenian-LLM-Eval, English-to-Slovene translation, Slovene LLM arena) and achieves over 60% win rate against GPT-4o in Slovene LLM arena.

Conclusion: The paper presents successful adaptation methodology for LLMs to less-resourced languages, demonstrating that specialized multilingual models can achieve performance comparable to much larger commercial models for specific languages.

Abstract: Large language models (LLMs) have become an essential tool for natural language processing and artificial intelligence in general. Current open-source models are primarily trained on English texts, resulting in poorer performance on less-resourced languages and cultures. We present a set of methodological approaches necessary for the successful adaptation of an LLM to a less-resourced language, and demonstrate them using the Slovene language. We present GaMS3-12B, a generative model for Slovene with 12 billion parameters, and demonstrate that it is the best-performing open-source model for Slovene within its parameter range. We adapted the model to the Slovene language using three-stage continual pre-training of the Gemma 3 model, followed by two-stage supervised fine-tuning (SFT). We trained the model on a combination of 140B Slovene, English, Bosnian, Serbian, and Croatian pretraining tokens, and over 200 thousand English and Slovene SFT examples. We evaluate GaMS3-12B on the Slovenian-LLM-Eval datasets, English-to-Slovene translation, and the Slovene LLM arena. We show that the described model outperforms 12B Gemma 3 across all three scenarios and performs comparably to much larger commercial GPT-4o in the Slovene LLM arena, achieving a win rate of over 60 %.

Method: Created benchmark with Victorian Criminal Charge Book passages and hand-crafted legal questions. Used full factorial design and hierarchical error decomposition to evaluate embedding models (Kanon 2, Gemini Embedding, OpenAI Text Embedding) and LLMs (Gemini 3.1 Pro, GPT-5.2).

Result: Retrieval is the primary driver of legal RAG performance, with Kanon 2 Embedder improving correctness by 17.5 points, groundedness by 4.5 points, and retrieval accuracy by 34 points. Many “hallucinations” are actually triggered by retrieval failures.

Conclusion: Retrieval sets the ceiling for legal RAG system performance, and proper evaluation requires distinguishing between retrieval and reasoning errors. The benchmark and methodology enable systematic comparison of legal RAG systems.

Abstract: We introduce Legal RAG Bench, a benchmark and evaluation methodology for assessing the end-to-end performance of legal RAG systems. As a benchmark, Legal RAG Bench consists of 4,876 passages from the Victorian Criminal Charge Book alongside 100 complex, hand-crafted questions demanding expert knowledge of criminal law and procedure. Both long-form answers and supporting passages are provided. As an evaluation methodology, Legal RAG Bench leverages a full factorial design and novel hierarchical error decomposition framework, enabling apples-to-apples comparisons of the contributions of retrieval and reasoning models in RAG. We evaluate three state-of-the-art embedding models (Isaacus’ Kanon 2 Embedder, Google’s Gemini Embedding 001, and OpenAI’s Text Embedding 3 Large) and two frontier LLMs (Gemini 3.1 Pro and GPT-5.2), finding that information retrieval is the primary driver of legal RAG performance, with LLMs exerting a more moderate effect on correctness and groundedness. Kanon 2 Embedder, in particular, had the largest positive impact on performance, improving average correctness by 17.5 points, groundedness by 4.5 points, and retrieval accuracy by 34 points. We observe that many errors attributed to hallucinations in legal RAG systems are in fact triggered by retrieval failures, concluding that retrieval sets the ceiling for the performance of many modern legal RAG systems. We document why and how we built Legal RAG Bench alongside the results of our evaluations. We also openly release our code and data to assist with reproduction of our findings.

Method: Three strategies for generating synthetic triplet data: 1) In-context learning, 2) Adapter composition (novel approach), and 3) Cross-lingual finetuning of LLM generator (XL-LoRA, novel approach). The synthetic data is then used to optimize embedding models for various languages.

Result: In-context learning falls short of strong non-synthetic baselines, but both adapter composition and XL-LoRA yield strong performance gains across a wide array of tasks and languages. These novel approaches provide a scalable pathway to creating performant embedding models for diverse languages.

Conclusion: LLMs can effectively bridge the data gap for embedding models in low-resource languages through novel techniques like adapter composition and cross-lingual finetuning, offering a scalable solution for multilingual NLP applications.

Abstract: Embedding models are crucial to modern NLP. However, the creation of the most effective models relies on carefully constructed supervised finetuning data. For high resource languages, such as English, such datasets are readily available. However, for hundreds of other languages, they are simply non-existent. We investigate whether the advent of large language models can help to bridge this gap. We test three different strategies for generating synthetic triplet data used to optimise embedding models. These include in-context learning as well as two novel approaches, leveraging adapter composition and cross lingual finetuning of the LLM generator (XL-LoRA) respectively. We find that while in-context learning still falls short of strong non-synthetic baselines, adapter composition and XL-LoRA yield strong performance gains across a wide array of tasks and languages, offering a clear, scalable pathway to producing performant embedding models for a wide variety of languages.

Method: Developed a web-based tool with flexible configuration options for sentiment elements, incorporating LLM-based RAG suggestions that retrieve the ten most similar annotated examples as few-shot prompts to provide context-aware assistance.

Result: Created AnnoABSA, the first web-based annotation tool for comprehensive ABSA tasks, released as open-source software under MIT License, featuring human-in-the-loop AI assistance with improving suggestion accuracy over time.

Conclusion: AnnoABSA provides a practical, extensible solution for ABSA annotation that balances AI assistance with human control, potentially accelerating research and applications in sentiment analysis.

Abstract: We introduce AnnoABSA, the first web-based annotation tool to support the full spectrum of Aspect-Based Sentiment Analysis (ABSA) tasks. The tool is highly customizable, enabling flexible configuration of sentiment elements and task-specific requirements. Alongside manual annotation, AnnoABSA provides optional Large Language Model (LLM)-based retrieval-augmented generation (RAG) suggestions that offer context-aware assistance in a human-in-the-loop approach, keeping the human annotator in control. To improve prediction quality over time, the system retrieves the ten most similar examples that are already annotated and adds them as few-shot examples in the prompt, ensuring that suggestions become increasingly accurate as the annotation process progresses. Released as open-source software under the MIT License, AnnoABSA is freely accessible and easily extendable for research and practical applications.

Method: System uses LLM interviewer to conduct simulated interviews and update calibrated beliefs about applicants’ rubric-oriented latent traits through Bayesian updating

Result: Evaluation on simulated interviews shows belief converges toward artificially-constructed latent ability levels; system includes code, dataset, calibration tests, and demo

Conclusion: LLMs can improve early-stage hiring decisions by providing nuanced, role-specific candidate assessment beyond traditional resume screening

Abstract: Effective hiring is integral to the success of an organisation, but it is very challenging to find the most suitable candidates because expert evaluation (e.g.\ interviews conducted by a technical manager) are expensive to deploy at scale. Therefore, automated resume scoring and other applicant-screening methods are increasingly used to coarsely filter candidates, making decisions on limited information. We propose that large language models (LLMs) can play the role of subject matter experts to cost-effectively elicit information from each candidate that is nuanced and role-specific, thereby improving the quality of early-stage hiring decisions. We present a system that leverages an LLM interviewer to update belief over an applicant’s rubric-oriented latent traits in a calibrated way. We evaluate our system on simulated interviews and show that belief converges towards the simulated applicants’ artificially-constructed latent ability levels. We release code, a modest dataset of public-domain/anonymised resumes, belief calibration tests, and simulated interviews, at \href{https://github.com/mbzuai-nlp/beyond-the-resume}{https://github.com/mbzuai-nlp/beyond-the-resume}. Our demo is available at \href{https://btr.hstu.net}{https://btr.hstu.net}.

Method: FreeAct leverages the rank-deficient nature of activations to derive a solution space beyond simple inverse matrices, decoupling activation transformations from weights. It identifies token-specific dynamics and allocates distinct transformation matrices to activations while maintaining unified static transformations for weights.

Result: Extensive experiments across diffusion LLMs and multimodal LLMs show FreeAct significantly outperforms baselines with up to 5.3% performance improvement.

Conclusion: FreeAct advances quantization for complex LLMs by accommodating dynamic activation disparities through flexible transformation allocation, achieving superior performance in diffusion and multimodal settings.

Abstract: Quantization is pivotal for mitigating the significant memory and computational overhead of Large Language Models (LLMs). While emerging transformation-based methods have successfully enhanced quantization by projecting feature spaces onto smoother manifolds using orthogonal matrices, they typically enforce a rigid one-to-one transformation constraint. This static approach fails to account for the dynamic patterns inherent in input activations, particularly within diffusion LLMs (dLLMs) and Multimodal LLMs (MLLMs), where varying token types exhibit distinct distributions. To advance this, we propose FreeAct, a novel quantization framework that relaxes the static one-to-one constraint to accommodate dynamic activation disparities. Theoretically, we leverage the rank-deficient nature of activations to derive a solution space that extends beyond simple inverse matrices, enabling the decoupling of activation transformations from weights. Methodologically, FreeAct identifies token-specific dynamics (i.e., vision v.s. text, or masked tokens) and allocates distinct transformation matrices to the activation side, while maintaining a unified, static transformation for the weights. Extensive experiments across dLLMs and MLLMs demonstrate that FreeAct significantly outperforms baselines, up to 5.3% performance improvement, with in-depth analyses. Our code will be publicly released.

Method: Proposes LA-ABSA approach that uses LLM-generated annotations to fine-tune lightweight models. Evaluates on five datasets for Target Aspect Sentiment Detection (TASD) and Aspect Sentiment Quad Prediction (ASQP). Uses in-context learning with small annotated examples to guide annotation of unlabeled data.

Result: Outperformed previous augmentation strategies and achieved competitive performance with LLM-prompting in low-resource scenarios. For ASQP on SemEval Rest16, achieved F1 score of 49.85 using 50 annotated examples, closely matching Gemma-3-27B’s 51.10 while requiring significantly lower computational resources.

Conclusion: LA-ABSA provides an effective approach for training lightweight ABSA models using LLM-generated annotations, offering substantial energy efficiency benefits while maintaining competitive performance with resource-intensive LLM prompting methods.

Abstract: Training models for Aspect-Based Sentiment Analysis (ABSA) tasks requires manually annotated data, which is expensive and time-consuming to obtain. This paper introduces LA-ABSA, a novel approach that leverages Large Language Model (LLM)-generated annotations to fine-tune lightweight models for complex ABSA tasks. We evaluate our approach on five datasets for Target Aspect Sentiment Detection (TASD) and Aspect Sentiment Quad Prediction (ASQP). Our approach outperformed previously reported augmentation strategies and achieved competitive performance with LLM-prompting in low-resource scenarios, while providing substantial energy efficiency benefits. For example, using 50 annotated examples for in-context learning (ICL) to guide the annotation of unlabeled data, LA-ABSA achieved an F1 score of 49.85 for ASQP on the SemEval Rest16 dataset, closely matching the performance of ICL prompting with Gemma-3-27B (51.10), while requiring significantly lower computational resources.

Method: Uses LoRA-adapted LLM executed multiple times per instance, retains only tuples achieving majority consensus, and leverages vLLM’s PagedAttention for efficient key-value cache reuse to reduce computational overhead.

Result: Self-consistency with 15 executions yields statistically significant improvements over single-inference prompting, with system ranking top seven across all settings, achieving second place on three English subsets and first on Tatar-Restaurant.

Conclusion: SCSG effectively improves reliability in dimensional aspect-based sentiment analysis through self-consistency voting, demonstrating practical benefits across multiple languages and domains.

Abstract: We present Self-Consistent Structured Generation (SCSG) for Dimensional Aspect-Based Sentiment Analysis in SemEval-2026 Task 3 (Track A). SCSG enhances prediction reliability by executing a LoRA-adapted large language model multiple times per instance, retaining only tuples that achieve a majority consensus across runs. To mitigate the computational overhead of multiple forward passes, we leverage vLLM’s PagedAttention mechanism for efficient key–value cache reuse. Evaluation across 6 languages and 8 language–domain combinations demonstrates that self-consistency with 15 executions yields statistically significant improvements over single-inference prompting, with our system (leveraging Gemma 3) ranking in the top seven across all settings, achieving second place on three out of four English subsets and first place on Tatar-Restaurant for DimASTE.

Method: Uses sentence-transformer paragraph embeddings and a running-centroid novelty measure to analyze 28,730 pre-1920 Project Gutenberg books (PG19) and 52,796 modern English books (Books3, 1990-2010). Employs PAA-16 representations for clustering narrative trajectories.

Result: Modern books show 10% higher mean paragraph-level novelty, 67% increase in trajectory circuitousness, 2.3x fewer convergent narrative curves, and novelty is orthogonal to reader quality ratings (r = -0.002). Eight distinct narrative-shape archetypes identified with substantial distribution shifts between eras.

Conclusion: Semantic novelty and narrative structures have significantly evolved in English literature, with modern works showing more complex, circuitous trajectories and less convergence to settled semantic registers, while interestingness remains independent of perceived literary merit.

Abstract: I apply Schmidhuber’s compression progress theory of interestingness at corpus scale, analyzing semantic novelty trajectories in more than 80,000 books spanning two centuries of English-language publishing. Using sentence-transformer paragraph embeddings and a running-centroid novelty measure, I compare 28,730 pre-1920 Project Gutenberg books (PG19) against 52,796 modern English books (Books3, approximately 1990-2010). The principal findings are fourfold. First, mean paragraph-level novelty is roughly 10% higher in modern books (0.503 vs. 0.459). Second, trajectory circuitousness – the ratio of cumulative path length to net displacement in embedding space – nearly doubles in the modern corpus (+67%). Third, convergent narrative curves, in which novelty declines toward a settled semantic register, are 2.3x more common in pre-1920 literature. Fourth, novelty is orthogonal to reader quality ratings (r = -0.002), suggesting that interestingness in Schmidhuber’s sense is structurally independent of perceived literary merit. Clustering paragraph-level trajectories via PAA-16 representations reveals eight distinct narrative-shape archetypes whose distribution shifts substantially between eras. All analysis code and an interactive exploration toolkit are publicly available at https://bigfivekiller.online/novelty_hub.

Method: Two-phase approach: (I) capture high entropy tokens via shared A matrix in LoRA, (II) asymmetric LoRA architecture with parameter isolation to unlearn tokens within target subdomains through token-level isolation.

Result: Achieves SOTA performance on TOFU, WMDP, and MUSE benchmarks with >95% forget quality, minimal side effects by preserving foundational tokens, >90% model utility preservation (vs 47.8-83.6% baselines).

Conclusion: ALTER provides an efficient unlearning framework that decouples unlearning from LLMs’ billion-scale parameters, achieving high forget quality while preserving model utility with minimal computational overhead.

Abstract: Large language models (LLMs) have advanced to encompass extensive knowledge across diverse domains. Yet controlling what a LLMs should not know is important for ensuring alignment and thus safe use. However, effective unlearning in LLMs is difficult due to the fuzzy boundary between knowledge retention and forgetting. This challenge is exacerbated by entangled parameter spaces from continuous multi-domain training, often resulting in collateral damage, especially under aggressive unlearning strategies. Furthermore, the computational overhead required to optimize State-of-the-Art (SOTA) models with billions of parameters poses an additional barrier. In this work, we present ALTER, a lightweight unlearning framework for LLMs to address both the challenges of knowledge entanglement and unlearning efficiency. ALTER operates through two phases: (I) high entropy tokens are captured and learned via the shared A matrix in LoRA, followed by (II) an asymmetric LoRA architecture that achieves a specified forgetting objective by parameter isolation and unlearning tokens within the target subdomains. Serving as a new research direction for achieving unlearning via token-level isolation in the asymmetric framework. ALTER achieves SOTA performance on TOFU, WMDP, and MUSE benchmarks with over 95% forget quality and shows minimal side effects through preserving foundational tokens. By decoupling unlearning from LLMs’ billion-scale parameters, this framework delivers excellent efficiency while preserving over 90% of model utility, exceeding baseline preservation rates of 47.8-83.6%.

Method: Develops an automated machine learning library with data-aware training regime selection, integrated data quality diagnostics, configurable OOD detection, and LLM features through a minimal lowcode API.

Result: OpenAutoNLU provides a comprehensive open-source solution for NLU tasks with automated configuration, quality diagnostics, and LLM integration accessible via a demo app.

Conclusion: OpenAutoNLU offers a user-friendly, automated solution for NLU tasks that reduces manual configuration while providing advanced features like data quality assessment and OOD detection.

Abstract: OpenAutoNLU is an open-source automated machine learning library for natural language understanding (NLU) tasks, covering both text classification and named entity recognition (NER). Unlike existing solutions, we introduce data-aware training regime selection that requires no manual configuration from the user. The library also provides integrated data quality diagnostics, configurable out-of-distribution (OOD) detection, and large language model (LLM) features, all within a minimal lowcode API. The demo app is accessible here https://openautonlu.dev.

Method: AT2QA is an autonomous, training-free agent that iteratively interacts with temporal knowledge graphs via a general search tool for dynamic retrieval. The LLM agent decides what to do next autonomously without hand-crafted workflows or fine-tuning.

Result: On MultiTQ benchmark, AT2QA achieves 88.7% Hits@1 (+10.7% over prior SOTA), including +20.1% gain on challenging multi-target queries, demonstrating that agentic autonomy can decisively outperform fine-tuning for temporal question answering.

Conclusion: Agentic autonomy in LLMs can significantly improve temporal question answering performance without requiring fine-tuning, suggesting that autonomous decision-making capabilities are crucial for complex temporal reasoning tasks.

Abstract: Temporal Knowledge Graph Question Answering (TKGQA) demands multi-hop reasoning under temporal constraints. Prior approaches based on large language models (LLMs) typically rely on rigid, hand-crafted retrieval workflows or costly supervised fine-tuning. We show that simply granting an off-the-shelf LLM autonomy, that is, letting it decide what to do next, already yields substantial gains even in a strict zero-shot setting. Building on this insight, we propose AT2QA, an autonomous, training-free agent for temporal question answering that iteratively interacts with the temporal knowledge graph via a general search tool for dynamic retrieval. Experiments on MultiTQ demonstrate large improvements: AT2QA achieves 88.7% Hits@1 (+10.7% over prior SOTA), including a +20.1% gain on challenging multi-target queries, showing that agentic autonomy can decisively outperform fine-tuning for temporal question answering. Code and the full set of sampled trajectories are available on https://github.com/AT2QA-Official-Code/AT2QA-Official-Code

Method: Introduces a variance decomposition that partitions benchmark score variance into scenario, generation, judge, and residual components. Based on this analysis, proposes CyclicJudge - a round-robin assignment of judges that eliminates bias precisely while requiring each judge only once per cycle.

Result: CyclicJudge is demonstrated to be the optimal allocation strategy, maintaining the cost of single-judge evaluation while eliminating bias. Empirical validation on MT-Bench supports all theoretical predictions.

Conclusion: The proposed CyclicJudge method provides a practical solution to systematic bias in LLM-as-judge evaluations, enabling more reliable model rankings without increasing evaluation costs.

Abstract: LLM-as-judge evaluation has become standard practice for open-ended model assessment; however, judges exhibit systematic biases that cannot be eliminated by increasing the number of scenarios or generations. These biases are often similar in magnitude to the model differences that benchmarks are designed to detect, resulting in unreliable rankings when single-judge evaluations are used. This work introduces a variance decomposition that partitions benchmark score variance into scenario, generation, judge, and residual components. Based on this analysis, CyclicJudge, a round-robin assignment of judges, is demonstrated to be the optimal allocation strategy. It eliminates bias precisely while requiring each judge only once per cycle, maintaining the cost of single-judge evaluation. Empirical validation on MT-Bench supports all theoretical predictions.

Method: Practical experimentation demonstrating viable alternatives to general-purpose architectures from commercial providers, focusing on cost-effective solutions that align with digital and cultural sovereignty principles.

Result: Empirical evidence shows sovereign AI-based public services are technically feasible and economically sustainable, capable of operating effectively on premises with modest computational and financial resources while maintaining autonomy.

Conclusion: Sovereign AI public services represent viable alternatives to commercial AI oligopolies, offering technical feasibility and economic sustainability while preserving digital and cultural sovereignty for governments and public agencies.

Abstract: The rapid expansion of AI-based remote services has intensified debates about the long-term implications of growing structural concentration in infrastructure and expertise. As AI capabilities become increasingly intertwined with geopolitical interests, the availability and reliability of foundational AI services can no longer be taken for granted. This issue is particularly pressing for AI-enabled public services for citizens, as governments and public agencies are progressively adopting 24/7 AI-driven support systems typically operated through commercial offerings from a small oligopoly of global technology providers. This paper challenges the prevailing assumption that general-purpose architectures, offered by these providers, are the optimal choice for all application contexts. Through practical experimentation, we demonstrate that viable and cost-effective alternatives exist. Alternatives that align with principles of digital and cultural sovereignty. Our findings provide an empirical illustration that sovereign AI-based public services are both technically feasible and economically sustainable, capable of operating effectively on premises with modest computational and financial resources while maintaining cultural and digital autonomy. The technical insights and deployment lessons reported here are intended to inform the adoption of similar sovereign AI public services by national agencies and governments worldwide.

Method: Proposes KDFlow with decoupled architecture: FSDP2 for efficient student training and SGLang for efficient teacher inference. Uses zero-copy data transfer of teacher hidden states (not full logits) and recomputes logits on student side to balance communication and performance. Supports both off-policy and on-policy distillation with extensible APIs.

Result: Achieves 1.44× to 6.36× speedup compared to current KD frameworks, enabling rapid prototyping and scaling of LLM distillation with minimal engineering overhead.

Conclusion: KDFlow provides an efficient, scalable framework for LLM knowledge distillation by decoupling training and inference backends, optimizing communication, and offering flexible APIs for various KD scenarios.

Abstract: Knowledge distillation (KD) is an essential technique to compress large language models (LLMs) into smaller ones. However, despite the distinct roles of the student model and the teacher model in KD, most existing frameworks still use a homogeneous training backend (e.g., FSDP and DeepSpeed) for both models, leading to suboptimal training efficiency. In this paper, we present a novel framework for LLM distillation, termed \textbf{KDFlow}, which features a decoupled architecture and employs SGLang for teacher inference. By bridging the training efficiency of FSDP2 and the inference efficiency of SGLang, KDFlow achieves full utilization of both advantages in a unified system. Moreover, instead of transferring full logits across different processes, our framework only transmits the teacher’s hidden states using zero-copy data transfer and recomputes the logits on the student side, effectively balancing the communication cost and KD performance. Furthermore, our framework supports both off-policy and on-policy distillation and incorporates KD algorithms for cross-tokenizer KD through highly extensible and user-friendly APIs. Experiments show that KDFlow can achieve \textbf{1.44$\times$ to 6.36$\times$} speedup compared to current KD frameworks, enabling researchers to rapidly prototype and scale LLM distillation with minimal engineering overhead. Code is available at: https://github.com/songmzhang/KDFlow

Method: Used retrieval-augmented generation (RAG) with open-source smaller LLMs, created culturally-aware knowledge bases from Wikipedia using keyword lists, integrated live online search via DuckDuckGo, and deployed on Ollama platform with refined prompts.

Result: Tested on English, Spanish, and Chinese for both short answer and multiple-choice question tracks, with resources and code shared publicly on GitHub.

Conclusion: Demonstrates effective use of smaller open-source LLMs with culturally-aware knowledge bases for cross-lingual everyday knowledge tasks, promoting privacy and sustainability.

Abstract: This system paper describes our participation in the SemEval-2025 Task-7 ``Everyday Knowledge Across Diverse Languages and Cultures’’. We attended two subtasks, i.e., Track 1: Short Answer Questions (SAQ), and Track 2: Multiple-Choice Questions (MCQ). The methods we used are retrieval augmented generation (RAGs) with open-sourced smaller LLMs (OS-sLLMs). To better adapt to this shared task, we created our own culturally aware knowledge base (CulKBs) by extracting Wikipedia content using keyword lists we prepared. We extracted both culturally-aware wiki-text and country-specific wiki-summary. In addition to the local CulKBs, we also have one system integrating live online search output via DuckDuckGo. Towards better privacy and sustainability, we aimed to deploy smaller LLMs (sLLMs) that are open-sourced on the Ollama platform. We share the prompts we developed using refinement techniques and report the learning curve of such prompts. The tested languages are English, Spanish, and Chinese for both tracks. Our resources and codes are shared via https://github.com/aaronlifenghan/FLANS-2026

Method: ViviDoc uses a multi-agent pipeline (Planner, Executor, Evaluator) with Document Specification (DocSpec) - a human-readable intermediate representation that decomposes interactive visualizations into State, Render, Transition, and Constraint components, enabling human review before code generation.

Result: Expert evaluation and user study show ViviDoc substantially outperforms naive agentic generation and provides an intuitive editing experience for creating interactive educational documents.

Conclusion: ViviDoc bridges the gap between pedagogical intent and executable output through human-agent collaboration and structured intermediate representations, making interactive educational content creation more accessible and controllable.

Abstract: Interactive articles help readers engage with complex ideas through exploration, yet creating them remains costly, requiring both domain expertise and web development skills. Recent LLM-based agents can automate content creation, but naively applying them yields uncontrollable and unverifiable outputs. We present ViviDoc, a human-agent collaborative system that generates interactive educational documents from a single topic input. ViviDoc introduces a multi-agent pipeline (Planner, Executor, Evaluator) and the Document Specification (DocSpec), a human-readable intermediate representation that decomposes each interactive visualization into State, Render, Transition, and Constraint components. The DocSpec enables educators to review and refine generation plans before code is produced, bridging the gap between pedagogical intent and executable output. Expert evaluation and a user study show that ViviDoc substantially outperforms naive agentic generation and provides an intuitive editing experience. Our project homepage is available at https://vividoc-homepage.vercel.app/.

Method: Uses iteration-specific MLP gates with a monotonic halting mask to decide when each token stops recurring, and introduces a KV reuse mechanism that reuses cached key/value states for halted tokens to ensure train-test consistency and practical acceleration.

Result: Across Pythia backbones from 70M to 410M (pretraining) and up to 2.8B (continued pretraining), reduces inference compute by about 10% while maintaining comparable language modeling perplexity and competitive downstream accuracy.

Conclusion: AdaPonderLM successfully learns adaptive computation time behavior in a fully self-supervised setting, allocating more computation to hard tokens and outperforming fixed pruning under iso-FLOPs, showing it allocates compute to the right tokens rather than just reducing average depth.

Abstract: Test-time scaling via recurrent/iterative Transformers enables large language models to spend more computation at inference, but most pretrained recurrent LMs run a fixed number of iterations, wasting compute on easy tokens and lacking token-wise adaptivity. Following the core idea of Adaptive Computation Time(ACT) and Early Exit(EE), we propose AdaPonderLM, a self-supervised recurrent language model that learns token-wise early exiting during pretraining without manually tuned per-token/per-layer pruning ratios. AdaPonderLM uses iteration-specific MLP gates with a monotonic halting mask to decide when each token stops recurring, and introduces a KV reuse mechanism that reuses cached key/value states for halted tokens, ensuring train–test consistency and practical acceleration. Across Pythia backbones from 70M to 410M (pretraining) and up to 2.8B (continued pretraining), AdaPonderLM reduces inference compute at about 10% while maintaining comparable language modeling perplexity and competitive downstream accuracy. Our analysis shows the learned gates allocate more computation to high-NLL (hard) tokens, exhibiting adaptive computation time behavior in a fully self-supervised setting. Meanwhile, under iso-FLOPs, the learned halting policy consistently outperforms fixed pruning, showing AdaPonderLM allocates compute to the right tokens rather than just reducing average depth.

Method: Introduces a narrative graph annotation framework integrating qualitative content analysis principles, using directed acyclic graphs where nodes represent events and edges encode causal relations. Employs a 6×3 factorial experimental design to evaluate effects of narrative representation (six levels) and distance metric type (three levels) on inter-annotator agreement using Krippendorrf’s α.

Result: Analysis shows that lenient metrics (overlap-based distance) overestimate reliability, and locally-constrained representations (e.g., one-hop neighbors) reduce annotation variability. The framework and graph-based Krippendorrf’s α implementation are open-sourced.

Conclusion: Provides practical guidance for NLP research on graph-based narrative annotation under human label variation, offering an annotation framework and evaluation methodology for structured narrative analysis in news discourse.

Abstract: Narratives in news discourse play a critical role in shaping public understanding of economic events, such as inflation. Annotating and evaluating these narratives in a structured manner remains a key challenge for Natural Language Processing (NLP). In this work, we introduce a narrative graph annotation framework that integrates principles from qualitative content analysis (QCA) to prioritize annotation quality by reducing annotation errors. We present a dataset of inflation narratives annotated as directed acyclic graphs (DAGs), where nodes represent events and edges encode causal relations. To evaluate annotation quality, we employed a $6\times3$ factorial experimental design to examine the effects of narrative representation (six levels) and distance metric type (three levels) on inter-annotator agreement (Krippendorrf’s $α$), capturing the presence of human label variation (HLV) in narrative interpretations. Our analysis shows that (1) lenient metrics (overlap-based distance) overestimate reliability, and (2) locally-constrained representations (e.g., one-hop neighbors) reduce annotation variability. Our annotation and implementation of graph-based Krippendorrf’s $α$ are open-sourced. The annotation framework and evaluation results provide practical guidance for NLP research on graph-based narrative annotation under HLV.

Method: Introduces Topic Word Mixing (TWM), a human evaluation task where annotators distinguish between word sets from single topics vs. mixed topics. Evaluates six topic models (LDA, NMF, Top2Vec, BERTopic, CFMF, CFMF-emb) using both automated metrics and human evaluation methods based on nearly 4,000 annotations from a philosophy of science corpus.

Result: Word intrusion and coherence metrics don’t always align, especially in specialized domains. TWM effectively captures human-perceived distinctness and appears to align with diversity metrics. The annotated dataset and task generation code are released.

Conclusion: TWM provides a valuable human-grounded evaluation of inter-topic distinctness, complementing existing methods. The work highlights the need for evaluation frameworks that bridge automated and human assessments, particularly for domain-specific corpora.

Abstract: Topic models uncover latent thematic structures in text corpora, yet evaluating their quality remains challenging, particularly in specialized domains. Existing methods often rely on automated metrics like topic coherence and diversity, which may not fully align with human judgment. Human evaluation tasks, such as word intrusion, provide valuable insights but are costly and primarily validated on general-domain corpora. This paper introduces Topic Word Mixing (TWM), a novel human evaluation task assessing inter-topic distinctness by testing whether annotators can distinguish between word sets from single or mixed topics. TWM complements word intrusion’s focus on intra-topic coherence and provides a human-grounded counterpart to diversity metrics. We evaluate six topic models - both statistical and embedding-based (LDA, NMF, Top2Vec, BERTopic, CFMF, CFMF-emb) - comparing automated metrics with human evaluation methods based on nearly 4,000 annotations from a domain-specific corpus of philosophy of science publications. Our findings reveal that word intrusion and coherence metrics do not always align, particularly in specialized domains, and that TWM captures human-perceived distinctness while appearing to align with diversity metrics. We release the annotated dataset and task generation code. This work highlights the need for evaluation frameworks bridging automated and human assessments, particularly for domain-specific corpora.

Method: AMemGym employs structured data sampling to predefine user profiles, state-dependent questions, and state evolution trajectories. It uses LLM-simulated users that expose latent states through role-play while maintaining structured state consistency, enabling cost-effective generation of high-quality, evaluation-aligned interactions.

Result: Experiments reveal performance gaps in existing memory systems (RAG, long-context LLMs, and agentic memory) and their corresponding reasons. The framework enables effective selection among competing approaches and can potentially drive self-evolution of memory management strategies.

Conclusion: AMemGym provides a scalable, diagnostically rich environment for advancing memory capabilities in conversational agents by bridging structured state evolution with free-form interactions, addressing current limitations in memory evaluation and optimization.

Abstract: Long-horizon interactions between users and LLM-based assistants necessitate effective memory management, yet current approaches face challenges in training and evaluation of memory. Existing memory benchmarks rely on static, off-policy data as context, limiting evaluation reliability and scalability. To address these gaps, we introduce AMemGym, an interactive environment enabling on-policy evaluation and optimization for memory-driven personalization. AMemGym employs structured data sampling to predefine user profiles, state-dependent questions, and state evolution trajectories, enabling cost-effective generation of high-quality, evaluation-aligned interactions. LLM-simulated users expose latent states through role-play while maintaining structured state consistency. Comprehensive metrics based on structured data guide both assessment and optimization of assistants. Extensive experiments reveal performance gaps in existing memory systems (e.g., RAG, long-context LLMs, and agentic memory) and corresponding reasons. AMemGym not only enables effective selection among competing approaches but also can potentially drive the self-evolution of memory management strategies. By bridging structured state evolution with free-form interactions, our framework provides a scalable, diagnostically rich environment for advancing memory capabilities in conversational agents.

Method: Iterative flywheel process with 15 generations starting from LLaMA 3.1, using data curation, reward modeling to estimate engagement metrics, supervised fine-tuning, reinforcement learning, and both offline/online evaluation with A/B testing.

Result: 7 of 8 deployed models showed positive engagement lift, with best performers achieving 8.8% improvement in engagement breadth and 19.4% in engagement depth; instruction following increased from 59.2% to 84.8% while violations decreased from 26.6% to 5.8%.

Conclusion: CharacterFlywheel provides a rigorous, scalable framework for improving LLMs in production social applications, demonstrating consistent engagement gains and improved steerability through systematic iteration and real-user feedback.

Abstract: This report presents CharacterFlywheel, an iterative flywheel process for improving large language models (LLMs) in production social chat applications across Instagram, WhatsApp, and Messenger. Starting from LLaMA 3.1, we refined models across 15 generations using data from both internal and external real-user traffic. Through continuous deployments from July 2024 to April 2025, we conducted controlled 7-day A/B tests showing consistent engagement improvements: 7 of 8 newly deployed models demonstrated positive lift over the baseline, with the strongest performers achieving up to 8.8% improvement in engagement breadth and 19.4% in engagement depth. We also observed substantial gains in steerability, with instruction following increasing from 59.2% to 84.8% and instruction violations decreasing from 26.6% to 5.8%. We detail the CharacterFlywheel process which integrates data curation, reward modeling to estimate and interpolate the landscape of engagement metrics, supervised fine-tuning (SFT), reinforcement learning (RL), and both offline and online evaluation to ensure reliable progress at each optimization step. We also discuss our methods for overfitting prevention and navigating production dynamics at scale. These contributions advance the scientific rigor and understanding of LLMs in social applications serving millions of users.

Method: Built on PonderLM-2 backbone, PonderLM-3 uses differentiable attention masks during pretraining with matching hard pruning rules at inference to enable token-wise adaptive pondering. This maintains train-inference consistency while making computation allocation learnable under self-supervised objectives.

Result: PonderLM-3 achieves lower pretraining perplexity at equal inference FLOPs compared to existing recursive or adaptive baselines. On downstream benchmarks, it matches fixed-step PonderLM-2 performance under same maximum computation steps while using fewer inference FLOPs in practice.

Conclusion: PonderLM-3 provides an end-to-end differentiable and train-inference consistent framework for token-wise adaptive computation, enabling efficient allocation of additional inference compute where most useful rather than uniform application to all tokens.

Abstract: Test-time scaling has shown that allocating more additional computation at inference can improve generation quality, motivating a natural follow-up question: where should this computation be spent? Building on this insight, we introduce PonderLM-3, a pretraining framework for token-wise adaptive pondering that learns to selectively allocate additional computation under purely self-supervised objectives, built on top of the PonderLM-2 backbone. This makes additional inference computation an allocatable per-token resource, so tokens receive more computation only when it is beneficial, rather than paying a uniform extra cost. To make this allocation learnable while maintaining train-inference consistency, PonderLM-3 injects a differentiable attention mask during pretraining and pairs it with a matching hard pruning rule at inference. PonderLM-3 defines a stronger Pareto frontier: compared with existing recursive or adaptive baselines, it achieves lower pretraining perplexity at equal inference FLOPs. On downstream benchmarks, PonderLM-3 attains comparable performance to fixed-step PonderLM-2 under the same maximum number of additional computation steps, while using fewer inference FLOPs in practice. Overall, PonderLM-3 provides an end-to-end differentiable and train-inference consistent framework for token-wise adaptive computation, enabling additional inference compute to be allocated where it is most useful rather than paid uniformly by every token.

Method: Created MMR-Life benchmark with 2,646 multiple-choice questions based on 19,108 real-world images covering 7 reasoning types: abductive, analogical, causal, deductive, inductive, spatial, and temporal.

Result: Evaluation of 37 advanced models shows substantial challenges - even top models like GPT-5 achieve only 58% accuracy with considerable variance across reasoning types.

Conclusion: MMR-Life establishes a comprehensive foundation for evaluating, analyzing, and improving next-generation multimodal reasoning systems.

Abstract: Recent progress in the reasoning capabilities of multimodal large language models (MLLMs) has empowered them to address more complex tasks such as scientific analysis and mathematical reasoning. Despite their promise, MLLMs’ reasoning abilities across different scenarios in real life remain largely unexplored and lack standardized benchmarks for evaluation. To address this gap, we introduce MMR-Life, a comprehensive benchmark designed to evaluate the diverse multimodal multi-image reasoning capabilities of MLLMs across real-life scenarios. MMR-Life consists of 2,646 multiple-choice questions based on 19,108 images primarily sourced from real-world contexts, comprehensively covering seven reasoning types: abductive, analogical, causal, deductive, inductive, spatial, and temporal. Unlike existing reasoning benchmarks, MMR-Life does not rely on domain-specific expertise but instead requires models to integrate information across multiple images and apply diverse reasoning abilities. The evaluation of 37 advanced models highlights the substantial challenge posed by MMR-Life. Even top models like GPT-5 achieve only 58% accuracy and display considerable variance in performance across reasoning types. Moreover, we analyze the reasoning paradigms of existing MLLMs, exploring how factors such as thinking length, reasoning method, and reasoning type affect their performance. In summary, MMR-Life establishes a comprehensive foundation for evaluating, analyzing, and improving the next generation of multimodal reasoning systems.

Method: Used Llama 3.1 8B as base model, performed continued pretraining with a mixture that increases Estonian exposure while approximating original training distribution through English replay and inclusion of code, mathematics, and instruction-like data. Applied supervised fine-tuning, preference optimization, and chat vector merging for instruction-following behavior.

Result: Evaluation on Estonian benchmarks shows consistent gains in linguistic competence, knowledge, reasoning, translation quality, and instruction-following compared to original base model and its instruction-tuned variant, while maintaining competitive performance on English benchmarks.

Conclusion: Continued pretraining with appropriately balanced data mixture, together with post-training alignment, can substantially improve single-language capabilities in pretrained multilingual LLMs without sacrificing performance in other languages.

Abstract: Large language models (LLMs) are predominantly trained on English-centric data, resulting in uneven performance for smaller languages. We study whether continued pretraining (CPT) can substantially improve Estonian capabilities in a pretrained multilingual LLM while preserving its English and general reasoning performance. Using Llama 3.1 8B as the main base model, we perform CPT on a mixture that increases Estonian exposure while approximating the original training distribution through English replay and the inclusion of code, mathematics, and instruction-like data. We subsequently apply supervised fine-tuning, preference optimization, and chat vector merging to introduce robust instruction-following behavior. Evaluation on a comprehensive suite of Estonian benchmarks shows consistent gains in linguistic competence, knowledge, reasoning, translation quality, and instruction-following compared to the original base model and its instruction-tuned variant, while maintaining competitive performance on English benchmarks. These findings indicate that CPT, with an appropriately balanced data mixture, together with post-training alignment, can substantially improve single-language capabilities in pretrained multilingual LLMs.

Method: Combines constituency and dependency parsing to identify three core filler-gap constructions (matrix wh-questions, embedded wh-questions, relative clauses) and extraction sites (subject, object, adjunct). Validated on human-annotated data and applied to 57 English CHILDES corpora.

Result: System scores well across most categories, enables characterization of children’s filler-gap input and production trajectories, including construction-specific frequencies and extraction-site asymmetries. Fine-grained labels support computational studies demonstrated through filtered corpus training with language models.

Conclusion: The system provides a valuable tool for studying language acquisition and enables future computational work, bridging linguistic theory and computational approaches to language learning.

Abstract: Children’s acquisition of filler-gap dependencies has been argued by some to depend on innate grammatical knowledge, while others suggest that the distributional evidence available in child-directed speech suffices. Unfortunately, the relevant input is difficult to quantify at scale with fine granularity, making this question difficult to resolve. We present a system that identifies three core filler-gap constructions in spoken English corpora – matrix wh-questions, embedded wh-questions, and relative clauses – and further identifies the extraction site (i.e., subject vs. object vs. adjunct). Our approach combines constituency and dependency parsing, leveraging their complementary strengths for construction classification and extraction site identification. We validate the system on human-annotated data and find that it scores well across most categories. Applying the system to 57 English CHILDES corpora, we are able to characterize children’s filler-gap input and their filler-gap production trajectories over the course of development, including construction-specific frequencies and extraction-site asymmetries. The resulting fine-grained labels enable future work in both acquisition and computational studies, which we demonstrate with a case study using filtered corpus training with language models.

Method: Analyzed 597 gameplay sessions (9,783 actions) from 100 students aged 8-11 using a sequence-based learning analytics approach. Introduced Hamming distance to quantify proximity to valid grammatical solutions and examined convergence patterns across exercises with varying difficulty levels.

Result: Determiners and verbs are key difficulty sites; learners often fix verbs first and adjust preceding elements. Exercises with fewer solutions show slower, more erratic convergence. Action sequences deviate from left-to-right processing, revealing dynamic hypothesis revision.

Conclusion: Sequence-based analytics can uncover hidden dimensions of linguistic reasoning, offering potential for real-time scaffolding and teacher tools in linguistically diverse classrooms.

Abstract: This study investigates grammatical reasoning in primary school learners through a sequence-based learning analytics approach, leveraging fine-grained action sequences from an interactive game targeting morphosyntactic agreement in French. Unlike traditional assessments that rely on final answers, we treat each slider movement as a hypothesis-testing action, capturing real-time cognitive strategies during sentence construction. Analyzing 597 gameplay sessions (9,783 actions) from 100 students aged 8-11 in authentic classroom settings, we introduce Hamming distance to quantify proximity to valid grammatical solutions and examine convergence patterns across exercises with varying levels of difficulty. Results reveal that determiners and verbs are key sites of difficulty, with action sequences deviating from left-to-right usual treatment. This suggests learners often fix the verb first and adjust preceding elements. Exercises with fewer solutions exhibit slower and more erratic convergence, while changes in the closest valid solution indicate dynamic hypothesis revision. Our findings demonstrate how sequence-based analytics can uncover hidden dimensions of linguistic reasoning, offering a foundation for real-time scaffolding and teacher-facing tools in linguistically diverse classrooms.

Method: Created ClinConsensus benchmark with 2500 expert-curated cases covering full care continuum, developed rubric-based grading with CACS@k scoring, and implemented dual-judge evaluation combining high-capability LLM-as-judge with distilled local judge model via supervised fine-tuning.

Result: Comprehensive assessment revealed substantial heterogeneity across task themes, care stages, and medical specialties; top models showed comparable overall scores but differed in reasoning, evidence use, and longitudinal follow-up capabilities, with treatment planning as key bottleneck.

Conclusion: ClinConsensus provides an extensible benchmark for developing robust, clinically grounded medical LLMs ready for real-world deployment, highlighting the need for improved reasoning and longitudinal capabilities in medical AI systems.

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

Method: Proposes Recursive Think-Answer Process (R-TAP) with confidence generator to evaluate response certainty, using two rewards: Recursively Confidence Increase Reward and Final Answer Confidence Reward.

Result: R-TAP-enhanced models outperform conventional single-pass methods for both LLMs and VLMs, show fewer self-reflective patterns (“Oops!”), and achieve more stable, faster inference-time reasoning.

Conclusion: R-TAP provides an efficient way to refine reasoning processes in AI models, paving the way for more elaborated methods in future AI development.

Abstract: Think-Answer reasoners such as DeepSeek-R1 have made notable progress by leveraging interpretable internal reasoning. However, despite the frequent presence of self-reflective cues like “Oops!”, they remain vulnerable to output errors during single-pass inference. To address this limitation, we propose an efficient Recursive Think-Answer Process (R-TAP) that enables models to engage in iterative reasoning cycles and generate more accurate answers, going beyond conventional single-pass approaches. Central to this approach is a confidence generator that evaluates the certainty of model responses and guides subsequent improvements. By incorporating two complementary rewards-Recursively Confidence Increase Reward and Final Answer Confidence Reward-we show that R-TAP-enhanced models consistently outperform conventional single-pass methods for both large language models (LLMs) and vision-language models (VLMs). Moreover, by analyzing the frequency of “Oops”-like expressions in model responses, we find that R-TAP-applied models exhibit significantly fewer self-reflective patterns, resulting in more stable and faster inference-time reasoning. We hope R-TAP pave the way evolving into efficient and elaborated methods to refine the reasoning processes of future AI.

Method: Tested six state-of-the-art LLMs across four real-world crisis simulation scenarios, requiring models to select predefined actions and justify decisions across multiple rounds. Compared models to humans on action alignment, risk calibration through action severity, and argumentative framing grounded in international relations theory.

Result: Models approximate human decision patterns in initial rounds but diverge over time, displaying distinct behavioral profiles and strategy updates. LLM explanations consistently show strong normative-cooperative framing focused on stability, coordination, and risk mitigation, with limited adversarial reasoning.

Conclusion: LLMs exhibit systematic differences from human decision-making in geopolitical simulations, with consistent normative biases in their reasoning that may affect their suitability as strategic agents in such environments.

Abstract: Large language models (LLMs) are increasingly proposed as agents in strategic decision environments, yet their behavior in structured geopolitical simulations remains under-researched. We evaluate six popular state-of-the-art LLMs alongside results from human results across four real-world crisis simulation scenarios, requiring models to select predefined actions and justify their decisions across multiple rounds. We compare models to humans in action alignment, risk calibration through chosen actions’ severity, and argumentative framing grounded in international relations theory. Results show that models approximate human decision patterns in base simulation rounds but diverge over time, displaying distinct behavioural profiles and strategy updates. LLM explanations for chosen actions across all models exhibit a strong normative-cooperative framing centered on stability, coordination, and risk mitigation, with limited adversarial reasoning.

Method: Introduces LongRLVR which augments sparse answer rewards with dense, verifiable context rewards that directly incentivize correct grounding information selection, providing robust learning gradients.

Result: Significantly outperforms standard RLVR across Qwen and LLaMA models on long-context benchmarks, boosting scores on RULER-QA from 73.17 to 88.90 and LongBench v2 from 39.8 to 46.5 for 14B models.

Conclusion: Explicitly rewarding the grounding process is critical for unlocking LLMs’ full reasoning potential in long-context applications, solving the optimization challenge of context grounding.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) has significantly advanced the reasoning capabilities of Large Language Models (LLMs) by optimizing them against factual outcomes. However, this paradigm falters in long-context scenarios, as its reliance on internal parametric knowledge is ill-suited for tasks requiring contextual grounding–the ability to find and reason over externally provided information. We identify a key reason for this failure: a reward based solely on the final answer is too sparse to effectively guide the model for identifying relevant evidence. We formally prove that the outcome-only reward leads to significant vanishing gradients for the context grounding process, rendering learning intractable. To overcome this bottleneck, we introduce LongRLVR to augment the sparse answer reward with a dense and verifiable context reward. This auxiliary signal directly incentivizes the model for selecting the correct grounding information, providing a robust learning gradient that solves the underlying optimization challenge. We validate our method on challenging long-context benchmarks using Qwen and LLaMA models. LongRLVR consistently and significantly outperforms the standard RLVR across all models and benchmarks, e.g., boosting a 14B model’s scores on RULER-QA from 73.17 to 88.90 and on LongBench v2 from 39.8 to 46.5. Our work demonstrates that explicitly rewarding the grounding process is a critical and effective strategy for unlocking the full reasoning potential of LLMs in long-context applications. Our code is available at https://github.com/real-absolute-AI/LongRLVR.

Method: Created CrimeNERdb dataset with 1.5k annotated documents from terrorist attack reports and DOJ press notes, defining 5 coarse and 22 fine-grained entity types. Conducted experiments with state-of-the-art NER models and large language models in zero- and few-shot settings.

Result: Developed a comprehensive crime NER dataset and demonstrated performance of various models on crime entity extraction tasks in limited-data scenarios.

Conclusion: The CrimeNER framework and dataset address the data scarcity problem in crime-related NER, enabling better information extraction from crime documents with limited annotated examples.

Abstract: The extraction of critical information from crime-related documents is a crucial task for law enforcement agencies. Named-Entity Recognition (NER) can perform this task in extracting information about the crime, the criminal, or law enforcement agencies involved. However, there is a considerable lack of adequately annotated data on general real-world crime scenarios. To address this issue, we present CrimeNER, a case-study of Crime-related zero- and Few-Shot NER, and a general Crime-related Named-Entity Recognition database (CrimeNERdb) consisting of more than 1.5k annotated documents for the NER task extracted from public reports on terrorist attacks and the U.S. Department of Justice’s press notes. We define 5 types of coarse crime entity and a total of 22 types of fine-grained entity. We address the quality of the case-study and the annotated data with experiments on Zero and Few-Shot settings with State-of-the-Art NER models as well as generalist and commonly used Large Language Models.

Method: Two-stage framework: (1) Manage Skills organizes skills into capability trees via recursive categorization for efficient discovery; (2) Solve Tasks retrieves, orchestrates, and executes multiple skills through DAG-based pipelines. Evaluation uses 30 artifact-rich tasks across five categories with LLM-based pairwise evaluation aggregated via Bradley-Terry model.

Result: Tree-based retrieval effectively approximates oracle skill selection, and DAG-based orchestration substantially outperforms native flat invocation even with identical skill sets. Experiments across three ecosystem scales (200 to 200K skills) demonstrate effectiveness.

Conclusion: Structured composition is key to unlocking skill potential in agent ecosystems. AgentSkillOS provides a principled framework for skill selection, orchestration, and ecosystem-level management.

Abstract: The rapid proliferation of Claude agent skills has raised the central question of how to effectively leverage, manage, and scale the agent skill ecosystem. In this paper, we propose AgentSkillOS, the first principled framework for skill selection, orchestration, and ecosystem-level management. AgentSkillOS comprises two stages: (i) Manage Skills, which organizes skills into a capability tree via node-level recursive categorization for efficient discovery; and (ii) Solve Tasks, which retrieves, orchestrates, and executes multiple skills through DAG-based pipelines. To evaluate the agent’s ability to invoke skills, we construct a benchmark of 30 artifact-rich tasks across five categories: data computation, document creation, motion video, visual design, and web interaction. We assess the quality of task outputs using LLM-based pairwise evaluation, and the results are aggregated via a Bradley-Terry model to produce unified quality scores. Experiments across three skill ecosystem scales (200 to 200K skills) show that tree-based retrieval effectively approximates oracle skill selection, and that DAG-based orchestration substantially outperforms native flat invocation even when given the identical skill set. Our findings confirm that structured composition is the key to unlocking skill potential. Our GitHub repository is available at:https://github.com/ynulihao/AgentSkillOS.

Method: Developed Reasoning Core suite that procedurally generates verifiable symbolic reasoning data across five formal domains: PDDL planning over randomized domains, first-order logic with equality, context-free grammar parsing/generation, causal reasoning over random Bayesian networks, and systems of equations. Each task includes external solver verification, continuous difficulty control, and optional solver-derived reasoning traces for supervised training or reward functions for RL.

Result: Mixing Reasoning Core data into pre-training improves downstream reasoning while preserving or slightly improving language modeling quality. Zero-shot evaluations show these tasks challenge frontier models like GPT-5. The system provides scalable, verifiable symbolic reasoning data generation.

Conclusion: Reasoning Core enables scalable generation of verifiable symbolic reasoning data across formal domains, supporting both supervised training and reinforcement learning. This approach expands language models’ reasoning capabilities beyond standard pre-training while maintaining language quality.

Abstract: Training on verifiable symbolic data is a promising way to expand the reasoning frontier of language models beyond what standard pre-training corpora provide. Yet existing procedural generators often rely on fixed puzzles or templates and do not deliver the distributional breadth needed at scale. We introduce Reasoning Core, a scalable suite that procedurally generates verifiable symbolic reasoning data across core formal domains: PDDL planning over randomized domains, first-order logic with equality, context-free grammar parsing and generation, causal reasoning over random Bayesian networks, and systems of equations. Each task is paired with an external solver for rigorous verification and admits continuous difficulty control for curriculum design. Examples can optionally include solver-derived reasoning traces, enabling supervised training from the earliest pre-training stages, and the same interface provides verifiable reward functions for reinforcement learning. Our experiments show that mixing Reasoning Core data into pre-training improves downstream reasoning while preserving, or slightly improving, language modeling quality. Zero-shot evaluations confirm these tasks challenge frontier models such as GPT-5. The code and data are publicly available under the MIT license.

Method: Developed a benchmark with three base families of two-player sequential language-based games with consistent parameterization. Created an open-source framework for interaction simulation and analysis, collecting datasets of LLM vs. LLM and human vs. LLM interactions across various game configurations.

Result: Market parameters and LLM choices have complex, interdependent effects on economic outcomes. The framework enables comparison of LLM behavior across economic contexts, evaluation of individual/collective performance, and quantification of environmental effects on agent behavior.

Conclusion: LLM integration into economic systems requires careful design and analysis due to complex interactions between market parameters and LLM characteristics affecting economic outcomes in language-based strategic interactions.

Abstract: Large Language Models (LLMs) show significant potential in economic and strategic interactions, where communication via natural language is often prevalent. This raises key questions: Do LLMs behave rationally? How do they perform compared to humans? Do they tend to reach an efficient and fair outcome? What is the role of natural language in strategic interaction? How do characteristics of the economic environment influence these dynamics? These questions become crucial concerning the economic and societal implications of integrating LLM-based agents into real-world data-driven systems, such as online retail platforms and recommender systems. To answer these questions, we introduce a benchmark for standardizing research on two-player, sequential, language-based games. Inspired by the economic literature, we define three base families of games with consistent parameterization, degrees of freedom and economic measures to evaluate agents’ performance (self-gain), as well as the game outcome (efficiency and fairness). We develop an open-source framework for interaction simulation and analysis, and utilize it to collect a dataset of LLM vs. LLM interactions across numerous game configurations and an additional dataset of human vs. LLM interactions. Through extensive experimentation, we demonstrate how our framework and dataset can be used to: (i) compare the behavior of LLM-based agents in various economic contexts; (ii) evaluate agents in both individual and collective performance measures; and (iii) quantify the effect of the economic characteristics of the environments on the behavior of agents. Our results suggest that the market parameters, as well as the choice of the LLMs, tend to have complex and interdependent effects on the economic outcome, which calls for careful design and analysis of the language-based economic ecosystem.

Method: Created SimpleToM benchmark with concise stories in everyday settings (supermarkets, hospitals, schools, offices) featuring information asymmetries. Each story has three questions testing: (a) mental state inference, (b) behavior prediction, and (c) behavior judgment.

Result: State-of-the-art models perform well on mental state inference but show sharp performance drops on behavior prediction and further on behavior judgment, exposing a critical gap between explicit ToM knowledge and applied reasoning.

Conclusion: LLMs have fragile social reasoning capabilities - they can infer mental states but struggle to implicitly apply that knowledge for behavior prediction and judgment, highlighting limitations in applied Theory of Mind.

Abstract: Large language models (LLMs) are increasingly tested for a “Theory of Mind” (ToM) - the ability to attribute mental states to oneself and others. Yet most evaluations stop at explicit belief attribution in classical toy stories or stylized tasks, leaving open the questions of whether LLMs can implicitly apply such knowledge to predict human behavior, or to judge an observed behavior, in diverse scenarios. We introduce SimpleToM, a benchmark that advances ToM evaluation along two novel axes. First, it probes multiple levels of ToM reasoning, from mental state inference (explicit ToM) to behavior prediction and judgment (applied ToM). Second, it situates these tasks in diverse, everyday scenarios - such as supermarkets, hospitals, schools, and offices - where information asymmetries naturally arise (e.g., hidden defects in grocery store items, incomplete information in provider-patient interactions, or restricted access to locked devices). SimpleToM contains concise stories (e.g., “The can of Pringles has moldy chips in it. Mary picks up the can in the supermarket and walks to the cashier.”), each with three questions that test different degrees of ToM reasoning, asking models to predict: (a) mental states (“Is Mary aware of the mold?”), (b) behaviors (“Will Mary pay for the chips or report the mold?”), and (c) judgments (“Mary paid for the chips. Was that reasonable?”). Experiments reveal a striking gap: state-of-the-art models often reliably infer mental state (a), but fail at applying knowledge about the mental state for secondary predictions, with performance dropping sharply for behavior prediction (b) and further for behavior judgment (c). This exposes a critical fragility in LLMs’ social reasoning in terms of what they know (explicit ToM) versus how well they can implicitly apply that knowledge for predictions (applied ToM).

Method: Introduces SynthKG, a multi-step data synthesis pipeline that generates high-quality document-KG pairs through systematic chunking, decontextualization, and structured extraction using LLMs. Then fine-tunes a smaller LLM on synthesized data to create Distill-SynthKG for single-step KG generation. Also repurposes QA datasets for evaluation and proposes graph-based retrieval framework for RAG.

Result: Distill-SynthKG surpasses all baseline models in KG quality (including models 8x larger) and improves retrieval and QA tasks. The graph retrieval framework outperforms all KG-retrieval methods across multiple benchmark datasets.

Conclusion: High-quality synthesized training data enables smaller models to achieve superior document-level KG construction, addressing scaling challenges while maintaining quality. The approach also enhances downstream applications like retrieval and question answering.

Abstract: Document-level knowledge graph (KG) construction faces a fundamental scaling challenge: existing methods either rely on expensive large language models (LLMs), making them economically nonviable for large-scale corpora, or employ smaller models that produce incomplete and inconsistent graphs. We find that this limitation stems not from model capabilities but from insufficient training on high-quality document-level KG data. To address this gap, we introduce SynthKG, a multi-step data synthesis pipeline that generates high-quality document-KG pairs through systematic chunking, decontextualization, and structured extraction using LLMs. By fine-tuning a smaller LLM on synthesized document-KG pairs, we streamline the multi-step process into a single-step KG generation approach called Distill-SynthKG. Furthermore, we repurpose existing question-answering datasets to construct KG evaluation datasets and introduce new evaluation metrics. Using KGs produced by Distill-SynthKG, we also design a novel graph-based retrieval framework for RAG. Experimental results demonstrate that Distill-SynthKG not only surpasses all baseline models in KG quality (including models up to eight times larger) but also consistently improves in retrieval and question-answering tasks. Additionally, our proposed graph retrieval framework outperforms all KG-retrieval methods across multiple benchmark datasets.

Method: 1) Introduces Query Optimization Lifecycle (QOL) Framework with five phases: Intent Recognition, Query Transformation, Retrieval Execution, Evidence Integration, Response Synthesis. 2) Proposes Query Complexity Taxonomy classifying queries by evidence type (explicit/implicit) and quantity (single/multiple). 3) Analyzes four atomic operations: Query Expansion, Query Decomposition, Query Disambiguation, Query Abstraction covering 90+ methods.

Result: Provides comprehensive survey of QO techniques, structured frameworks for research, and practical guidance for practitioners. Identifies gaps in benchmarks and discusses open challenges including process reward models, efficiency optimization, and multi-modal query handling.

Conclusion: This survey establishes structured foundation for QO research and offers actionable guidance for practitioners, while highlighting future directions including multi-modal query handling challenges.

Abstract: Query Optimization (QO) has become essential for enhancing Large Language Model (LLM) effectiveness, particularly in Retrieval-Augmented Generation (RAG) systems where query quality directly determines retrieval and response performance. This survey provides a systematic analysis of query optimization techniques with three contributions. \textit{First}, we introduce the \textbf{Query Optimization Lifecycle (QOL) Framework}, a five-phase pipeline covering Intent Recognition, Query Transformation, Retrieval Execution, Evidence Integration, and Response Synthesis. \textit{Second}, we propose a \textbf{Query Complexity Taxonomy} that classifies queries along two dimensions: evidence type (explicit vs.\ implicit) and evidence quantity (single vs.\ multiple), establishing principled mappings to optimization strategies. \textit{Third}, we analyze four atomic operations: \textbf{Query Expansion}, \textbf{Query Decomposition}, \textbf{Query Disambiguation}, and \textbf{Query Abstraction}, covering over 90 representative methods. We further examine evaluation methodologies, identify gaps in benchmarks, and discuss open challenges including process reward models, efficiency optimization, and multi-modal query handling. This survey offers both a structured foundation for research and actionable guidance for practitioners.

Method: Proposes AStar with “thought cards” - a library of reasoning patterns abstracted from prior samples. For each test problem, adaptively retrieves optimal thought cards and integrates external explicit guidelines with model’s internal implicit reasoning.

Result: Achieves 53.9% accuracy on MathVerse (surpassing GPT-4o’s 50.2%) and 32.7% on MathVision (outperforming GPT-4o’s 30.4%). Shows transferability: thought cards from math reasoning benefit other tasks and general visual perception.

Conclusion: AStar provides efficient, training-free multimodal reasoning enhancement, serving as plug-and-play test-time inference method compatible with other techniques, complementing existing approaches.

Abstract: Multimodal large language models excel across diverse domains but struggle with complex visual reasoning tasks. To enhance their reasoning capabilities, current approaches typically rely on explicit search or post-training techniques. However, search-based methods suffer from computational inefficiency due to extensive solution space exploration, while post-training methods demand substantial data, computational resources, and often exhibit training instability. To address these challenges, we propose \textbf{AStar}, a training-free, \textbf{A}utomatic \textbf{S}tructured \textbf{t}hinking paradigm for multimod\textbf{a}l \textbf{r}easoning. Specifically, we introduce novel ``thought cards’’, a lightweight library of high-level reasoning patterns abstracted from prior samples. For each test problem, AStar adaptively retrieves the optimal thought cards and seamlessly integrates these external explicit guidelines with the model’s internal implicit reasoning capabilities. Compared to previous methods, AStar eliminates computationally expensive explicit search and avoids additional complex post-training processes, enabling a more efficient reasoning approach. Extensive experiments demonstrate that our framework achieves 53.9% accuracy on MathVerse (surpassing GPT-4o’s 50.2%) and 32.7% on MathVision (outperforming GPT-4o’s 30.4%). Further analysis reveals the remarkable transferability of our method: thought cards generated from mathematical reasoning can also be applied to other reasoning tasks, even benefiting general visual perception and understanding. AStar serves as a plug-and-play test-time inference method, compatible with other post-training techniques, providing an important complement to existing multimodal reasoning approaches.

Method: Proposes a simple, training-free approach that perturbs appropriate subsets of model parameters or hidden unit activations during sampling to better capture Bayesian uncertainty, improving hallucination detection without additional training.

Result: The approach significantly improves inference-time hallucination detection over standard sampling methods across diverse datasets, model architectures, and uncertainty metrics.

Conclusion: Perturbing model parameters during sampling provides a more effective way to capture Bayesian uncertainty for hallucination detection than standard token distribution sampling, offering a simple yet powerful training-free solution.

Abstract: Large Language Models (LLMs) are prone to generating plausible yet incorrect responses, known as hallucinations. Effectively detecting hallucinations is therefore crucial for the safe deployment of LLMs. Recent research has linked hallucinations to model uncertainty, suggesting that hallucinations can be detected by measuring dispersion over answer distributions obtained from multiple samples drawn from a model. While drawing from the distribution over tokens defined by the model is a natural way to obtain samples, in this work, we argue that it is suboptimal for the purpose of detecting hallucinations. We show that detection can be improved significantly by taking into account model uncertainty in the Bayesian sense. To this end, we propose a very simple, training-free approach based on perturbing an appropriate subset of model parameters, or equivalently hidden unit activations, during sampling. We demonstrate that our approach significantly improves inference-time hallucination detection over standard sampling across diverse datasets, model architectures, and uncertainty metrics.

Method: Analyzed Wikipedia article content and page views, evaluated LLM impact on NLP tasks (machine translation, RAG), conducted simulations to explore potential risks.

Result: Found ~1% LLM impact on Wikipedia articles in certain categories; machine translation benchmarks may be inflated; RAG effectiveness could decrease due to knowledge contamination.

Conclusion: LLMs haven’t fully changed Wikipedia yet, but empirical findings signal need for careful consideration of future risks in NLP research.

Abstract: In this paper, we present a comprehensive analysis and monitoring framework for the impact of Large Language Models (LLMs) on Wikipedia, examining the evolution of Wikipedia through existing data and using simulations to explore potential risks. We begin by analyzing article content and page views to study the recent changes in Wikipedia and assess the impact of LLMs. Subsequently, we evaluate how LLMs affect various Natural Language Processing (NLP) tasks related to Wikipedia, including machine translation and retrieval-augmented generation (RAG). Our findings and simulation results reveal that Wikipedia articles have been affected by LLMs, with an impact of approximately 1% in certain categories. If the machine translation benchmark based on Wikipedia is influenced by LLMs, the scores of the models may become inflated, and the comparative results among models could shift. Moreover, the effectiveness of RAG might decrease if the knowledge has been contaminated by LLMs. While LLMs have not yet fully changed Wikipedia’s language and knowledge structures, we believe that our empirical findings signal the need for careful consideration of potential future risks in NLP research. We release all the experimental dataset and source code at: https://github.com/HSM316/LLM_Wikipedia

Method: Meta-analysis of 92 open-source pretrained models across various scales, including state-of-the-art and less conventional models, incorporating features beyond just model size and training tokens to predict downstream performance.

Result: Incorporating design features beyond scale improves ability to predict downstream performance by 3-28% relative to using scale alone. Key findings include optimal trade-off between language and code tasks at 15-25% code, and better performance of rotary embeddings over learned embeddings.

Conclusion: Model design choices significantly impact final capabilities beyond just scale, with data composition and architectural decisions playing crucial roles. The framework enables more systematic investigation of how development choices shape model performance.

Abstract: Improvements in language model capabilities are often attributed to increasing model size or training data, but in some cases smaller models trained on curated data or with different architectural decisions can outperform larger ones trained on more tokens. What accounts for this? To quantify the impact of these design choices, we meta-analyze 92 open-source pretrained models across a wide array of scales, including state-of-the-art open-weights models as well as less performant models and those with less conventional design decisions. We find that by incorporating features besides model size and number of training tokens, we can achieve a relative 3-28% increase in ability to predict downstream performance compared with using scale alone. Analysis of model design decisions reveal insights into data composition, such as the trade-off between language and code tasks at 15-25% code, as well as the better performance of some architectural decisions such as choosing rotary over learned embeddings. Broadly, our framework lays a foundation for more systematic investigation of how model development choices shape final capabilities.

Method: Survey methodology - systematic review of recent advancements in LLM applications for bioinformatics, covering multiple domains including genomic sequence modeling, RNA structure prediction, protein function inference, and single-cell transcriptomics.

Result: Comprehensive perspective on current state of LLMs in bioinformatics, identification of key challenges (data scarcity, computational complexity, cross-omics integration), and exploration of future directions including multimodal learning and hybrid AI models.

Conclusion: LLMs have transformative potential in bioinformatics and precision medicine, with future directions pointing toward multimodal learning, hybrid AI approaches, and clinical applications.

Abstract: Large Language Models (LLMs) are revolutionizing bioinformatics, enabling advanced analysis of DNA, RNA, proteins, and single-cell data. This survey provides a systematic review of recent advancements, focusing on genomic sequence modeling, RNA structure prediction, protein function inference, and single-cell transcriptomics. Meanwhile, we also discuss several key challenges, including data scarcity, computational complexity, and cross-omics integration, and explore future directions such as multimodal learning, hybrid AI models, and clinical applications. By offering a comprehensive perspective, this paper underscores the transformative potential of LLMs in driving innovations in bioinformatics and precision medicine.

Method: Based on lightweight open-source LLM with Parameter-Efficient Fine-Tuning (PEFT) techniques to create cost-effective training pipeline; learns transferable mobility behavior representations across heterogeneous data sources.

Result: Achieves best F1 score and accuracy across six real-world mobility datasets compared to SOTA deep learning models; shows better transferability and cost efficiency than commercial LLMs; robust under network changes, policy interventions, events, and incidents.

Conclusion: MoBLLM provides a generalizable modeling foundation for individual mobility behavior, enabling more reliable and adaptive personalized information services for transportation management.

Abstract: Individual mobility prediction plays a key role in urban transport, enabling personalized service recommendations and effective travel management. It is widely modeled by data-driven methods such as machine learning, deep learning, as well as classical econometric methods to capture key features of mobility patterns. However, such methods are hindered in promoting further transferability and robustness due to limited capacity to learn mobility patterns from different data sources, predict in out-of-distribution settings (a.k.a ``zero-shot"). To address this challenge, this paper introduces MoBLLM, a foundational model for individual mobility prediction that aims to learn a shared and transferable representation of mobility behavior across heterogeneous data sources. Based on a lightweight open-source large language model (LLM), MoBLLM employs Parameter-Efficient Fine-Tuning (PEFT) techniques to create a cost-effective training pipeline, avoiding the need for large-scale GPU clusters while maintaining strong performance. We conduct extensive experiments on six real-world mobility datasets to evaluate its accuracy, robustness, and transferability across varying temporal scales (years), spatial contexts (cities), and situational conditions (e.g., disruptions and interventions). MoBLLM achieves the best F1 score and accuracy across all datasets compared with state-of-the-art deep learning models and shows better transferability and cost efficiency than commercial LLMs. Further experiments reveal its robustness under network changes, policy interventions, special events, and incidents. These results indicate that MoBLLM provides a generalizable modeling foundation for individual mobility behavior, enabling more reliable and adaptive personalized information services for transportation management.

Method: Created AnesSuite dataset suite with AnesBench evaluation benchmark (three reasoning levels) and three training datasets for CPT, SFT, and RLVR. Developed Morpheus baseline models using SFT and GRPO training strategies.

Result: Morpheus achieves substantial improvements in anesthesiology reasoning rivaling larger models, demonstrates enhanced reasoning across general medical and broad-domain benchmarks, and identifies key factors influencing performance.

Conclusion: AnesSuite provides the first comprehensive infrastructure for anesthesiology reasoning in LLMs, and Morpheus serves as an effective baseline, with both resources being open-sourced for community use.

Abstract: The application of large language models (LLMs) in the medical field has garnered significant attention, yet their reasoning capabilities in more specialized domains like anesthesiology remain underexplored. To bridge this gap, we introduce AnesSuite, the first comprehensive dataset suite specifically designed for anesthesiology reasoning in LLMs. The suite features AnesBench, an evaluation benchmark tailored to assess anesthesiology-related reasoning across three levels: factual retrieval (System 1), hybrid reasoning (System 1.x), and complex decision-making (System 2). Alongside this benchmark, the suite includes three training datasets that provide an infrastructure for continued pre-training (CPT), supervised fine-tuning (SFT), and reinforcement learning with verifiable rewards (RLVR). Leveraging this suite, we develop Morpheus, the first baseline model collection for anesthesiology reasoning. Despite undergoing limited training with SFT and group relative policy optimization (GRPO), Morpheus not only achieves substantial improvements in anesthesiology that rival larger-scale models, but also demonstrates enhanced reasoning capabilities across general medical and broad-domain benchmarks. Furthermore, through comprehensive evaluations and experiments, we analyze the key factors influencing anesthesiology reasoning performance, including model characteristics, training strategies and training data. Both AnesSuite and Morpheus will be open-sourced at https://github.com/MiliLab/AnesSuite.

Method: Introduced goal decoding tasks and evaluation frameworks using large-scale eye tracking data with hundreds of text-specific information seeking tasks. Developed and compared several discriminative and generative multimodal LLMs that combine text and eye movement data.

Result: Experiments showed considerable success in selecting correct goals among options and made progress toward free-form textual reconstruction of precise goal formulations.

Conclusion: The work demonstrates the feasibility of decoding reading goals from eye movements using multimodal LLMs, opening doors for further scientific investigation and development of real-time educational/assistive technologies.

Abstract: When reading, we often have specific information that interests us in a text. For example, you might be reading this paper because you are curious about LLMs for eye movements in reading, the experimental design, or perhaps you wonder ``This sounds like science fiction. Does it actually work?’’. More broadly, in daily life, people approach texts with any number of text-specific goals that guide their reading behavior. In this work, we ask, for the first time, whether open-ended reading goals can be automatically decoded solely from eye movements in reading. To address this question, we introduce goal decoding tasks and evaluation frameworks using large-scale eye tracking for reading data in English with hundreds of text-specific information seeking tasks. We develop and compare several discriminative and generative multimodal text and eye movements LLMs for these tasks. Our experiments show considerable success on the task of selecting the correct goal among several options, and even progress towards free-form textual reconstruction of the precise goal formulation. These results open the door for further scientific investigation of goal driven reading, as well as the development of educational and assistive technologies that will rely on real-time decoding of reader goals from their eye movements.

Method: Proposes DRA framework that calibrates reward signals using semantic density of sampled groups via Submodular Mutual Information and Inverse Propensity Scoring to create repulsive forces against redundancy and bias gradient estimation.

Result: DRA-GRPO outperforms baselines on five math benchmarks, achieving 58.2% average accuracy on DeepSeek-R1-Distill-Qwen-1.5B with only 7,000 training samples and $55 cost.

Conclusion: Diversity calibration is critical for data-efficient alignment in LLM post-training, and DRA provides a plug-and-play solution that integrates seamlessly with GRPO variants.

Abstract: Post-training LLMs with Reinforcement Learning, specifically Group Relative Policy Optimization (GRPO), has emerged as a paradigm for enhancing mathematical reasoning. However, standard GRPO relies on scalar correctness rewards that are often non-injective with respect to semantic content: distinct reasoning paths receive identical rewards. This leads to a Diversity-Quality Inconsistency, where the policy collapses into a narrow set of dominant modes while ignoring equally valid but structurally novel strategies. To bridge this gap, we propose Diversity-aware Reward Adjustment (DRA), a theoretically grounded framework that calibrates the reward signal using the semantic density of sampled groups. By leveraging Submodular Mutual Information (SMI), DRA implements an Inverse Propensity Scoring (IPS) mechanism that effectively de-biases the gradient estimation. This creates a repulsive force against redundancy, driving the policy to achieve better coverage of the high-reward landscape. Our method is plug-and-play and integrates seamlessly with GRPO variants. Empirical evaluations on five math benchmarks demonstrate that DRA-GRPO consistently outperforms strong baselines, achieving an average accuracy of 58.2% on DeepSeek-R1-Distill-Qwen-1.5B with only 7,000 training samples and $55 cost, highlighting the critical role of diversity calibration in data-efficient alignment. The code is available at https://github.com/xiwenc1/DRA-GRPO.

Method: Two large-scale experiments comparing LLMs (Claude 3.5 Sonnet and DeepSeek v3) against incentivized human persuaders in real-time conversational settings where they attempted to persuade quiz-takers toward correct or incorrect answers, with linguistic analysis of persuasion texts.

Result: Claude 3.5 Sonnet was more persuasive than humans in both truthful and deceptive contexts, increasing accuracy when truthful but decreasing it when deceptive. DeepSeek v3 showed similar accuracy effects but was only more persuasive in deceptive contexts. LLM persuasiveness waned over repeated interactions unlike humans, and linguistic analysis suggests higher conviction expression by LLMs.

Conclusion: LLM persuasiveness is context-dependent, varying by model, truthfulness of persuasion, and interaction frequency, with implications for AI deployment in persuasive applications.

Abstract: Large Language Models (LLMs) have been shown to be highly persuasive, but when and why they outperform humans is still an open question. We compare the persuasiveness of two LLMs (Claude 3.5 Sonnet and DeepSeek v3) against humans who had incentives to persuade, using an interactive, real-time conversational setting. We demonstrate that LLMs persuasive superiority is context-dependent: it depends on whether the persuasion attempt is truthful (towards the right answer) or deceptive (towards the wrong answer) and on the LLM model, and wanes over repeated interactions (unlike human persuasiveness). In our first large-scale experiment, humans vs LLMs (Claude 3.5 Sonnet) interacted with other humans who were completing an online quiz for a reward, attempting to persuade them toward a given (either correct or incorrect) answer. Claude was more persuasive than incentivized human persuaders both in truthful and deceptive contexts and it significantly increased accuracy if persuasion was truthful, but decreased it if persuasion was deceptive. In a follow-up experiment with Deepseek v3, we replicated the findings about accuracy but found greater LLM persuasiveness only if the persuasion was deceptive. Linguistic analyses of the persuaders texts suggest that these effects may be due to LLMs expressing higher conviction than humans.

Method: Developed a formal framework for analyzing transformers’ counting capabilities, proved transformers can capture all semialgebraic counting properties, identified a subclass of transformers that characterizes these properties, and connected this to Hilbert’s tenth problem for undecidability results.

Result: Transformers can express highly nonlinear counting properties (semialgebraic properties), generalizing beyond linear counting properties captured by C-RASP softmax transformers. Established undecidability results for simple transformer models without positional encodings or masking.

Conclusion: Transformers have significantly stronger counting capabilities than previously demonstrated, being able to express complex nonlinear counting properties, which has implications for both theoretical expressivity and practical trainability of such properties.

Abstract: Counting properties (e.g. determining whether certain tokens occur more than other tokens in a given input text) have played a significant role in the study of expressiveness of transformers. In this paper, we provide a formal framework for investigating the counting power of transformers. We argue that all existing results demonstrate transformers’ expressivity only for (semi-)linear counting properties, i.e., which are expressible as a boolean combination of linear inequalities. Our main result is that transformers can express counting properties that are highly nonlinear. More precisely, we prove that transformers can capture all semialgebraic counting properties, i.e., expressible as a boolean combination of arbitrary multivariate polynomials (of any degree). Among others, these generalize the counting properties that can be captured by C-RASP softmax transformers, which capture only linear counting properties. To complement this result, we exhibit a natural subclass of (softmax) transformers that completely characterizes semialgebraic counting properties. Through connections with the Hilbert’s tenth problem, this expressivity of transformers also yields a new undecidability result for analyzing an extremely simple transformer model – surprisingly with neither positional encodings (i.e. NoPE-transformers) nor masking. We also experimentally validate trainability of such counting properties.

Method: RPM automatically discovers user-specific reasoning structures from raw behavioral data, constructs structured models of user behavior using response-influential features and statistical factors, creates personalized reasoning paths, and uses feature-based retrieval to find beneficial examples for guiding inference.

Result: Extensive experiments across four diverse tasks show RPM consistently outperforms existing response-level personalization methods while enhancing both personalization performance and interpretability.

Conclusion: RPM provides a promising direction for black-box LLM personalization by introducing reasoning-level personalization as a new paradigm that improves both performance and interpretability.

Abstract: While black-box large language models are widely deployed, they produce generic outputs that overlook individual user preferences. Current personalization methods are fundamentally limited to response-level personalization; they only match final outputs, failing to model the underlying reasoning that connects user behavior to responses. To address this, this work introduces reasoning-level personalization as a new paradigm and proposes RPM, the first systematic framework that automatically discovers user-specific reasoning structures from raw behavioral data to guide the model’s personalized inference. RPM constructs a structured model of user behavior-built from response-influential features and statistical factors-to create personalized reasoning paths and retrieve beneficial examples for guiding inference through a feature-based retrieval mechanism. Extensive experiments across four diverse tasks demonstrate that RPM consistently outperforms existing response-level methods while simultaneously enhancing both personalization performance and interpretability, providing a promising direction for black-box LLM personalization.

Method: Framework integrates data at two stages: (1) incorporating metadata during idea generation to guide models toward more feasible concepts, and (2) introducing automated preliminary validation during idea selection to assess empirical plausibility. Evaluated in social science domain with focus on climate negotiation topics.

Result: Metadata improves feasibility of generated ideas by 20%, automated validation improves overall quality of selected ideas by 7%. Human study shows LLM-generated ideas with data augmentation and validation inspire researchers and lead to higher quality ideas than those proposed without assistance.

Conclusion: Data-augmented research ideation has practical value in real-world academic settings, highlighting the potential of LLM-assisted ideation when properly augmented with relevant data and validation mechanisms.

Abstract: Recent advancements in large language models (LLMs) demonstrate strong potential for generating novel research ideas, yet such ideas often struggle with feasibility and effectiveness. In this paper, we investigate whether augmenting LLMs with relevant data during the ideation process can improve idea quality. Our framework integrates data at two stages: (1) incorporating metadata during idea generation to guide models toward more feasible concepts, and (2) introducing an automated preliminary validation step during idea selection to assess the empirical plausibility of hypotheses within ideas. We evaluate our approach in the social science domain, with a specific focus on climate negotiation topics. Expert evaluation shows that metadata improves the feasibility of generated ideas by 20%, while automated validation improves the overall quality of selected ideas by 7%. Beyond assessing the quality of LLM-generated ideas, we conduct a human study to examine whether these ideas, augmented with related data and preliminary validation, can inspire researchers in their own ideation. Participants report that the LLM-generated ideas and validation are highly useful, and the ideas they propose with such support are proven to be of higher quality than those proposed without assistance. Our findings highlight the potential of data-augmented research ideation and underscore the practical value of LLM-assisted ideation in real-world academic settings.

Method: Two-module framework: (1) Tool creation - LLMs generate executable tools from reference content, validate them with examples, and organize hierarchically into toolbox; (2) Tool utilization - LLMs navigate toolbox structure to select and apply appropriate tools to solve problems.

Result: Outperforms existing tool-creation and domain-specific reasoning methods by 12.3% average accuracy on causality, physics, and chemistry benchmarks. Cost-efficient and generalizable to non-scientific tasks like low-resource language translation.

Conclusion: RefTool enables LLMs to overcome internal knowledge limitations by grounding tool creation in external references, advancing generalizable reasoning in knowledge-intensive domains through accurate, faithful tools and effective hierarchical organization.

Abstract: Large Language Models (LLMs) can enhance their reasoning capabilities by using external tools. However, many tasks lack predefined tools. Prior works have explored instructing LLMs to generate tools on their own, but such approaches depend heavily on internal knowledge and struggle when tasks fall outside the model’s knowledge scope. To address this limitation, we propose RefTool, a reference-guided framework for automatic tool creation that leverages external materials, such as textbooks and knowledge snippets. RefTool consists of two modules: (1) tool creation, where LLMs generate executable tools from reference content, validate them using illustrative examples, and organize them hierarchically into a toolbox; and (2) tool utilization, where LLMs navigate the toolbox structure to select and apply the appropriate tools to solve problems. Experiments on causality, physics, and chemistry benchmarks demonstrate that RefTool outperforms existing tool-creation and domain-specific reasoning methods by 12.3% on average accuracy, while being cost-efficient and broadly generalizable to non-scientific tasks, e.g., extremely low-resource language translation. Analyses reveal that grounding tool creation in references produces accurate and faithful tools, and that the hierarchical structure facilitates effective tool selection. RefTool enables LLMs to overcome internal knowledge limitations, advancing generalizable reasoning in knowledge-intensive domains.

Method: VeriTrail is a closed-domain hallucination detection method designed to provide traceability for both multi-step (MGS) and single-step (SGS) generative processes. It analyzes intermediate outputs to trace where hallucinations were likely introduced and how faithful content was derived from source material.

Result: VeriTrail outperforms baseline methods on newly introduced datasets that include all intermediate outputs and human annotations of final outputs’ faithfulness for their respective MGS processes.

Conclusion: Detecting hallucinations in final outputs alone is insufficient for multi-step generative processes; traceability through intermediate outputs is crucial. VeriTrail provides this capability and demonstrates superior performance over existing methods.

Abstract: Even when instructed to adhere to source material, language models often generate unsubstantiated content - a phenomenon known as “closed-domain hallucination.” This risk is amplified in processes with multiple generative steps (MGS), compared to processes with a single generative step (SGS). However, due to the greater complexity of MGS processes, we argue that detecting hallucinations in their final outputs is necessary but not sufficient: it is equally important to trace where hallucinated content was likely introduced and how faithful content may have been derived from the source material through intermediate outputs. To address this need, we present VeriTrail, the first closed-domain hallucination detection method designed to provide traceability for both MGS and SGS processes. We also introduce the first datasets to include all intermediate outputs as well as human annotations of final outputs’ faithfulness for their respective MGS processes. We demonstrate that VeriTrail outperforms baseline methods on both datasets.

Method: Proposes RedTeamCUA framework with a novel hybrid sandbox combining VM-based OS environment with Docker-based web platforms, enabling flexible adversarial scenario configuration and direct initialization at injection points.

Result: Created RTC-Bench with 864 examples showing significant vulnerabilities: Claude 3.7 Sonnet CUA has 42.9% ASR, Claude 4.5 Sonnet CUA has 60% ASR, and even the most secure (Operator) has 7.6% ASR. Attempt rates reach 92.5%.

Conclusion: RedTeamCUA provides essential framework for systematic CUA vulnerability analysis, highlighting urgent need for robust defenses against indirect prompt injection before real-world deployment due to tangible risks.

Abstract: Computer-use agents (CUAs) promise to automate complex tasks across operating systems (OS) and the web, but remain vulnerable to indirect prompt injection. Current evaluations of this threat either lack support realistic but controlled environments or ignore hybrid web-OS attack scenarios involving both interfaces. To address this, we propose RedTeamCUA, an adversarial testing framework featuring a novel hybrid sandbox that integrates a VM-based OS environment with Docker-based web platforms. Our sandbox supports key features tailored for red teaming, such as flexible adversarial scenario configuration, and a setting that decouples adversarial evaluation from navigational limitations of CUAs by initializing tests directly at the point of an adversarial injection. Using RedTeamCUA, we develop RTC-Bench, a comprehensive benchmark with 864 examples that investigate realistic, hybrid web-OS attack scenarios and fundamental security vulnerabilities. Benchmarking current frontier CUAs identifies significant vulnerabilities: Claude 3.7 Sonnet | CUA demonstrates an ASR of 42.9%, while Operator, the most secure CUA evaluated, still exhibits an ASR of 7.6%. Notably, CUAs often attempt to execute adversarial tasks with an Attempt Rate as high as 92.5%, although failing to complete them due to capability limitations. Nevertheless, we observe concerning high ASRs in realistic end-to-end settings, with the strongest-to-date Claude 4.5 Sonnet | CUA exhibiting the highest ASR of 60%, indicating that CUA threats can already result in tangible risks to users and computer systems. Overall, RedTeamCUA provides an essential framework for advancing realistic, controlled, and systematic analysis of CUA vulnerabilities, highlighting the urgent need for robust defenses to indirect prompt injection prior to real-world deployment.

Method: Created a synthetic dataset of webtext-like documents about fictional events with corresponding question-answer pairs. Conducted training experiments to demonstrate how this synthetic data can be used to study different forms of memorization, while documenting challenges in building realistic fictional synthetic data.

Result: The proposed dataset enables controlled study of fact memorization versus verbatim sequence memorization. Training experiments show the synthetic data’s utility for analyzing memorization processes, though challenges in creating realistic fictional data were identified.

Conclusion: The synthetic dataset provides a valuable tool for researchers to better understand how language models memorize facts versus sequences, addressing a gap in current understanding of model memorization mechanisms.

Abstract: When language models are trained on textual data, they acquire both knowledge about the structure of language as well as knowledge of facts about the world. At inference time, their knowledge of facts can be leveraged to solve interesting problems and perform useful knowledge work for users. It is well known that language models can verbatim memorize long sequences from their training data. However, it is much less well understood how language models memorize facts seen during training. In this work, we propose a new dataset to specifically empower researchers to study the dual processes of fact memorization and verbatim sequence memorization. The dataset consists of synthetically-generated, webtext-like documents about fictional events, as well as question-answer pairs about the events. We conduct training experiments showing how synthetic data about fictional events can be useful for studying different forms of memorization. We also document some challenges in effectively building realistic, fictional synthetic data.

Method: Proposed LA-CDM: a hypothesis-driven uncertainty-aware language agent that converges toward diagnoses via repeated test requests and interpretations. Uses hybrid training combining supervised and reinforcement learning with three objectives: accurate hypothesis generation, hypothesis uncertainty estimation, and efficient decision-making.

Result: Evaluated on MIMIC-CDM dataset covering four abdominal diseases with various clinical tests. Showed benefits of explicitly training clinical decision-making for increasing both diagnostic performance and efficiency.

Conclusion: LA-CDM effectively models the interactive clinical decision-making process, outperforming approaches that either assume complete information upfront or use untrained models, demonstrating improved diagnostic accuracy and efficiency.

Abstract: Clinical decision-making is a dynamic, interactive, and cyclic process where doctors have to repeatedly decide on which clinical action to perform and consider newly uncovered information for diagnosis and treatment. Large Language Models (LLMs) have the potential to support clinicians in this process, however, most applications of LLMs in clinical decision support suffer from one of two limitations: Either they assume the unrealistic scenario of immediate availability of all patient information and do not model the interactive and iterative investigation process, or they restrict themselves to the limited “out-of-the-box” capabilities of large pre-trained models without performing task-specific training. In contrast to this, we propose to model clinical decision-making for diagnosis with a hypothesis-driven uncertainty-aware language agent, LA-CDM, that converges towards a diagnosis via repeatedly requesting and interpreting relevant tests. Using a hybrid training paradigm combining supervised and reinforcement learning, we train LA-CDM with three objectives targeting critical aspects of clinical decision-making: accurate hypothesis generation, hypothesis uncertainty estimation, and efficient decision-making. We evaluate our methodology on MIMIC-CDM, a real-world dataset covering four abdominal diseases containing various clinical tests and show the benefit of explicitly training clinical decision-making for increasing diagnostic performance and efficiency.

Method: AgentSynth leverages information asymmetry to construct subtasks that are simple during generation but become challenging when composed into long-horizon tasks. The pipeline can precisely modulate task complexity by varying the number of subtasks, enabling creation of over 6,000 diverse tasks.

Result: The pipeline achieves low average cost of $0.60 per trajectory (much cheaper than human annotations). State-of-the-art LLM agents show steep performance drop from 18% success at difficulty level 1 to just 4% at level 6, demonstrating benchmark difficulty.

Conclusion: AgentSynth provides a scalable, cost-effective solution for generating diverse computer-use agent datasets, enabling better evaluation and training of generalist agents through controlled complexity modulation.

Abstract: We introduce AgentSynth, a scalable and cost-efficient pipeline for automatically synthesizing high-quality tasks and trajectory datasets for generalist computer-use agents. Leveraging information asymmetry, AgentSynth constructs subtasks that are simple during generation but significantly more challenging when composed into long-horizon tasks, enabling the creation of over 6,000 diverse and realistic tasks. A key strength of AgentSynth is its ability to precisely modulate task complexity by varying the number of subtasks. Empirical evaluations show that state-of-the-art LLM agents suffer a steep performance drop, from 18% success at difficulty level 1 to just 4% at level 6, highlighting the benchmark’s difficulty and discriminative power. Moreover, our pipeline achieves a low average cost of $0.60 per trajectory, orders of magnitude cheaper than human annotations. Our code and data are available at https://github.com/sunblaze-ucb/AgentSynth

Method: SPARE (Single-Pass Annotation with Reference-Guided Evaluation) is a structured framework that performs per-step annotation by jointly aligning solution steps to reference solutions and determining accuracy with explicit reasoning in a single generation.

Result: SPARE shows consistent improvements across four diverse datasets (GSM8K, MATH, MuSiQue-Ans, SpaRP) for training Process Reward Models and fine-tuning via offline RL. It achieves data-efficient out-of-distribution generalization using only ~16% of training samples compared to baselines, with 2.3× speedup over MCTS methods.

Conclusion: SPARE provides a practical and scalable solution for automatic process supervision in LLM reasoning, with complementary precision-recall characteristics to MCTS approaches suggesting potential for ensemble methods.

Abstract: Process or step-wise supervision has played a crucial role in advancing complex multi-step reasoning capabilities of Large Language Models (LLMs). However, efficient, high-quality automated process annotation remains a significant challenge. To address this, we introduce Single-Pass Annotation with Reference-Guided Evaluation (SPARE), a novel structured framework that enables efficient per-step annotation by jointly aligning solution steps to reference solutions and determine its accuracy with explicit reasoning in single generation. We demonstrate SPARE’s effectiveness across four diverse datasets spanning mathematical reasoning (GSM8K, MATH), multi-hop question answering (MuSiQue-Ans), and spatial reasoning (SpaRP), showing consistent improvements in two applications: (1) training Process Reward Models (PRMs) for ranking and aggregating multiple generations, and (2) fine-tuning models via offline reinforcement learning for greedy decoding. On ProcessBench, SPARE demonstrates data-efficient out-of-distribution generalization, using only $\sim$16% of training samples compared to human-labeled and other synthetically trained baselines. Additionally, it achieves competitive performance with MCTS-based methods while offering 2.3$\times$ speedup in terms of total token count. Manual analysis reveals complementary precision-recall characteristics with MCTS approaches, suggesting potential for ensemble methods. These results establish SPARE as a practical and scalable solution for automatic process supervision in LLM reasoning.

Method: GenRecal introduces a Recalibrator module that aligns and adapts feature representations between heterogeneous VLMs, enabling effective knowledge transfer across different types of VLMs regardless of their architectural differences.

Result: Extensive experiments on multiple benchmarks show that GenRecal significantly improves baseline performances and eventually outperforms both large-scale open- and closed-source VLMs.

Conclusion: GenRecal provides a general solution for distilling knowledge across diverse VLM architectures, addressing the deployment challenges of large VLMs on resource-constrained devices.

Abstract: Recent advancements in vision-language models (VLMs) have leveraged large language models (LLMs) to achieve performance on par with closed-source systems like GPT-4V. However, deploying these models in real-world scenarios, particularly on resource-constrained devices, remains challenging due to their substantial computational demands. This has spurred interest in distilling knowledge from large VLMs into smaller, more efficient counterparts. A key challenge arises here from the diversity of VLM architectures, which are built on different LLMs and employ varying token types-differing in vocabulary size, token splits, and token index ordering. To address this challenge of limitation to a specific VLM type, we present Generation after Recalibration (GenRecal), a general-purpose distillation framework for VLMs. GenRecal incorporates a Recalibrator that aligns and adapts feature representations between heterogeneous VLMs, enabling effective knowledge transfer across different types of VLMs. Through extensive experiments on multiple challenging benchmarks, we demonstrate that GenRecal significantly improves baseline performances, eventually outperforming large-scale open- and closed-source VLMs.

Method: Created OJBench with 232 programming competition problems from NOI and ICPC, then evaluated 37 models including closed-source and open-source, reasoning-oriented and non-reasoning-oriented models.

Result: Even state-of-the-art reasoning-oriented models like o4-mini and Gemini-2.5-pro-exp struggle with highly challenging competition-level problems, highlighting significant challenges in competitive-level code reasoning.

Conclusion: Current LLMs face substantial difficulties with competitive programming problems, indicating a need for improved reasoning capabilities and better benchmarks for evaluating high-level code reasoning.

Abstract: Recent advancements in large language models (LLMs) have demonstrated significant progress in math and code reasoning capabilities. However, existing code benchmark are limited in their ability to evaluate the full spectrum of these capabilities, particularly at the competitive level. To bridge this gap, we introduce OJBench, a novel and challenging benchmark designed to assess the competitive-level code reasoning abilities of LLMs. OJBench comprises 232 programming competition problems from NOI and ICPC, providing a more rigorous test of models’ reasoning skills. We conducted a comprehensive evaluation using OJBench on 37 models, including both closed-source and open-source models, reasoning-oriented and non-reasoning-oriented models. Our results indicate that even state-of-the-art reasoning-oriented models, such as o4-mini and Gemini-2.5-pro-exp, struggle with highly challenging competition-level problems. This highlights the significant challenges that models face in competitive-level code reasoning.

Method: Proposes theoretical framework categorizing long context failures into three types: cross-chunk dependence (task noise), confusion growing with context size (model noise), and imperfect integration of partial results (aggregator noise). Analyzes multi-agent chunking effectiveness through experiments on retrieval, QA, and summarization tasks.

Result: Experiments confirm theoretical analysis and identify conditions favoring multi-agent chunking. Shows weaker models with chunk-based processing can surpass advanced models like GPT4o on large inputs due to accelerated decay of model fidelity with input length.

Conclusion: Provides principled framework for understanding LLM long-context challenges and demonstrates careful chunking and aggregation strategies as direct pathway to handling long contexts effectively.

Abstract: We investigate the challenge of applying Large Language Models (LLMs) to long texts. We propose a theoretical framework that distinguishes the failure modes of long context tasks into three categories: cross-chunk dependence (task noise), confusion that grows with context size (model noise), and the imperfect integration of partial results (aggregator noise). Under this view, we analyze when it is effective to use multi-agent chunking, i.e., dividing a lengthy sequence into smaller chunks and aggregating the processed results of each chunk. Our experiments on tasks such as retrieval, question answering, and summarization confirm both the theoretical analysis and the conditions that favor multi-agent chunking. By exploring the accelerated decay of model fidelity with input length, we also explain why, for large inputs, a weaker model configured with chunk-based processing can surpass a more advanced model like GPT4o applied in a single shot. Overall, we present a principled understanding framework and our results highlight a direct pathway to handling long contexts in LLMs with carefully managed chunking and aggregator strategies.

Method: Proposes Adaptive-Scalable Entmax (ASEntmax) which combines α-entmax’s ability to assign exact zeros to irrelevant tokens with a learnable temperature parameter, allowing dynamic interpolation between sparse (pattern-focused) and dense (softmax-like) attention regimes.

Result: ASEntmax substantially outperforms softmax, scalable softmax, and fixed-temperature α-entmax baselines, achieving up to 1000× length extrapolation on synthetic benchmarks and superior long-context generalization on language modeling while preserving short-context performance.

Conclusion: Dynamic sparse attention mechanisms like ASEntmax can effectively address the limitations of softmax attention for long-context tasks by enabling precise pattern-focused attention while maintaining flexibility across different attention regimes.

Abstract: Transformer-based architectures traditionally employ softmax to compute attention weights, which produces dense distributions over all tokens in a sequence. While effective in many settings, this density has been shown to be detrimental for tasks that demand precise focus on fixed-size patterns: as sequence length increases, non-informative tokens accumulate attention probability mass, leading to dispersion and representational collapse. We show in this paper that dynamically sparse attention mechanisms using $α$-entmax can avoid these issues, due to their ability to assign exact zeros to irrelevant tokens. Furthermore, we introduce Adaptive-Scalable Entmax (ASEntmax), which endows $α$-entmax with a learnable temperature parameter, allowing the attention distribution to interpolate between sparse (pattern-focused) and dense (softmax-like) regimes. Our empirical evaluation on synthetic tasks and language modeling demonstrates that ASEntmax substantially outperforms softmax, scalable softmax, and fixed-temperature $α$-entmax baselines, achieving up to 1000$\times$ length extrapolation on synthetic benchmarks and superior long-context generalization on language modeling while preserving short-context performance, including better perplexity trends and higher retrieval accuracies at 8$\times$ training length.

Method: Proposes an incentivization-based approach using reinforcement learning starting from a base model (similar to R1-Zero). The method employs specialized reward models to guide LLMs toward better length control, writing quality, and structural formatting without any annotated or synthetic data.

Result: LongWriter-Zero model trained from Qwen2.5-32B consistently outperforms traditional SFT methods on long-form writing tasks, achieving state-of-the-art results on WritingBench and Arena-Write benchmarks, even surpassing 100B+ models like DeepSeek R1 and Qwen3-235B.

Conclusion: The RL-based approach without synthetic data successfully enables LLMs to generate ultra-long, high-quality text, demonstrating superior performance over traditional supervised fine-tuning methods.

Abstract: Ultra-long generation by large language models (LLMs) is a widely demanded scenario, yet it remains a significant challenge due to their maximum generation length limit and overall quality degradation as sequence length increases. Previous approaches, exemplified by LongWriter, typically rely on ‘’teaching’’, which involves supervised fine-tuning (SFT) on synthetic long-form outputs. However, this strategy heavily depends on synthetic SFT data, which is difficult and costly to construct, often lacks coherence and consistency, and tends to be overly artificial and structurally monotonous. In this work, we propose an incentivization-based approach that, starting entirely from scratch and without relying on any annotated or synthetic data, leverages reinforcement learning (RL) to foster the emergence of ultra-long, high-quality text generation capabilities in LLMs. We perform RL training starting from a base model, similar to R1-Zero, guiding it to engage in reasoning that facilitates planning and refinement during the writing process. To support this, we employ specialized reward models that steer the LLM towards improved length control, writing quality, and structural formatting. Experimental evaluations show that our LongWriter-Zero model, trained from Qwen2.5-32B, consistently outperforms traditional SFT methods on long-form writing tasks, achieving state-of-the-art results across all metrics on WritingBench and Arena-Write, and even surpassing 100B+ models such as DeepSeek R1 and Qwen3-235B. We open-source our data and model checkpoints under https://huggingface.co/THU-KEG/LongWriter-Zero-32B

Method: Two studies: (1) Observational - compare open-weight RPT models against base models across seen/unseen domains; (2) Interventional - fine-tune LLMs with RPT on single domains and evaluate across multiple domains

Result: RPT brings substantial gains on tasks similar to fine-tuning data, but gains generalize inconsistently and can vanish on domains with different reasoning patterns

Conclusion: RPT improvements are domain-specific and don’t reliably transfer to domains with different reasoning patterns, highlighting limitations in generalization

Abstract: Reinforcement post training (RPT) has recently shown promise in improving the reasoning abilities of large language models (LLMs). However, it remains unclear how well these improvements generalize to new domains, as prior work evaluates RPT models on data from the same domains used for post-training. To understand the generalizability of RPT, we conduct two studies with specific focus on Reinforcement Learning with Verifiable Rewards (RLVR). (1) Observational: we compare a wide range of open-weight RPT models against their corresponding base models across multiple domains, including both seen and unseen domains in their fine-tuning data. (2) Interventional: we fine-tune LLMs with RPT on single domains and evaluate their performance across multiple domains. Both studies converge on the same conclusion that, although RPT brings substantial gains on tasks similar to the fine-tuning data, the gains generalize inconsistently and can vanish on domains with different reasoning patterns.

Method: Uses a leading cognitive model of polite speech to evaluate LLMs through two settings: 1) degrees of reasoning “effort” and system prompt manipulations in closed-source models, 2) RL post-training dynamics analysis of open-source models.

Result: LLMs’ behavioral profiles shift predictably when prompted to prioritize certain goals, are amplified by small reasoning budgets, and can diagnose social behaviors like sycophancy. Post-training dynamics show large early value shifts and persistent effects of base model/pretraining data compared to feedback dataset or alignment method.

Conclusion: The framework offers a flexible tool for probing behavioral profiles across diverse model types and gaining insights for shaping training regimes that better control value trade-offs during model development.

Abstract: Value trade-offs are an integral part of human decision-making and language use, however, current tools for interpreting such dynamic and multi-faceted notions of values in language models are limited. In cognitive science, so-called “cognitive models” provide formal accounts of such trade-offs in humans, by modeling the weighting of a speaker’s competing utility functions in choosing an action or utterance. Here, we show that a leading cognitive model of polite speech can be used to systematically evaluate alignment-relevant trade-offs in language models via two encompassing settings: degrees of reasoning “effort” and system prompt manipulations in closed-source frontier models, and RL post-training dynamics of open-source models. Our results show that LLMs’ behavioral profiles under the cognitive model a) shift predictably when they are prompted to prioritize certain goals, b) are amplified by a small reasoning budget, and c) can be used to diagnose other social behaviors such as sycophancy. Our findings from LLMs’ post-training dynamics reveal large shifts in values early on in training and persistent effects of the choice of base model and pretraining data, compared to feedback dataset or alignment method. Our framework offers a flexible tool for probing behavioral profiles across diverse model types and gaining insights for shaping training regimes that better control trade-offs between values during model development.

Method: XISM uses a top-down procedure to generate candidate semantic maps, then provides a visual interface where users can iteratively refine edges with real-time metric feedback, combining data-driven inference with expert knowledge.

Result: Experiments in three semantic domains and expert interviews show that XISM improves linguistic decision transparency and controllability while maintaining computational efficiency.

Conclusion: XISM provides a collaborative approach for scalable and interpretable semantic-map building, bridging the gap between automated systems and expert-driven methods.

Abstract: Semantic map models visualize systematic relations among semantic functions through graph structures and are widely used in linguistic typology. However, existing construction methods either depend on labor-intensive expert reasoning or on fully automated systems lacking expert involvement, creating a tension between scalability and interpretability. We introduce \textbf{XISM}, an interactive system that combines data-driven inference with expert knowledge. XISM generates candidate maps via a top-down procedure and allows users to iteratively refine edges in a visual interface, with real-time metric feedback. Experiments in three semantic domains and expert interviews show that XISM improves linguistic decision transparency and controllability in semantic-map construction while maintaining computational efficiency. XISM provides a collaborative approach for scalable and interpretable semantic-map building. The system\footnote{https://app.xism2025.xin/} , source code\footnote{https://github.com/hank317/XISM} , and demonstration video\footnote{https://youtu.be/m5laLhGn6Ys} are publicly available.

Method: Two-stage finetuning framework: 1) Supervised finetuning on full-exploration policy generating broad sub-queries, 2) RL to adaptively prune search depth based on question difficulty, directly rewarding policies that balance correctness with frugality. Uses only ~1,000 examples instead of 10x more data.

Result: Achieves state-of-the-art efficiency-accuracy tradeoffs on HotPotQA and other multi-hop QA benchmarks, cutting retrieval cost nearly in half. Generalizes zero-shot on challenging BrowseCompPlus benchmark, surpassing SLM-based and other baselines.

Conclusion: Demonstrates RL can effectively reduce reasoning steps rather than increase them, providing scalable and efficient RAG solutions. Shows RL’s potential for optimizing efficiency in multimodal reasoning tasks.

Abstract: Reinforcement learning (RL) based on the final answer’s reward has driven recent progress in small language models (SLMs) on reasoning-heavy tasks such as math and code. However, applying the same techniques to retrieval-augmented generation (RAG) benchmarks like multi-hop QA has yielded limited gains, often trailing supervised or prompting-only baselines. Instead, we argue that a viable path for RL in multi-hop QA is to use test-time scaling judiciously to optimize both final answer accuracy and efficiency in reaching that answer. We propose FrugalRAG, a two-stage finetuning framework that adaptively reduces the number of retrieval steps based on a question’s difficulty. First, we train an SLM with supervised finetuning on a full-exploration policy that generates broad sub-queries. Then, we apply RL to adaptively prune search depth based on question difficulty, directly rewarding policies that balance correctness with frugality. Unlike prior approaches requiring 10x more data, our method achieves competitive performance with only approximately 1,000 examples. On HotPotQA and other multi-hop QA benchmarks, FrugalRAG attains state-of-the-art efficiency-accuracy tradeoffs, cutting retrieval cost nearly in half. Moreover, on the challenging BrowseCompPlus benchmark, it generalizes zero-shot and surpasses SLM-based and other baselines. These results demonstrate the use of RL not to increase reasoning steps, but to reduce them, as an effective solution for scalable and efficient RAG.

Method: First analyze code-switching using sparse autoencoders to identify excessive pre-activation of language features, then propose SASFT (Sparse Autoencoder-guided Supervised Fine-tuning) that teaches LLMs to maintain appropriate pre-activation values of specific language features during training.

Result: SASFT reduces unexpected code-switching by over 50% compared to standard supervised fine-tuning across five models and three languages, with complete elimination in one case, while maintaining or improving performance on six multilingual benchmarks.

Conclusion: SASFT effectively addresses code-switching in multilingual LLMs through mechanistic understanding and targeted training, preserving multilingual capabilities while improving response consistency.

Abstract: Large Language Models (LLMs) have impressive multilingual capabilities, but they suffer from unexpected code-switching, also known as language mixing, which involves switching to unexpected languages in the model response. This problem leads to poor readability and degrades the usability of model responses. However, existing work on this issue lacks a mechanistic analysis and shows limited effectiveness. In this paper, we first provide an in-depth analysis of unexpected code-switching using sparse autoencoders and find that when LLMs switch to a language, the features of that language exhibit excessive pre-activation values. Based on our findings, we propose $\textbf{S}$parse $\textbf{A}$utoencoder-guided $\textbf{S}$upervised $\textbf{F}$ine$\textbf{t}$uning (SASFT), which teaches LLMs to maintain appropriate pre-activation values of specific language features during training. Experiments on five models across three languages demonstrate that SASFT consistently reduces unexpected code-switching by more than 50% compared to standard supervised fine-tuning, with complete elimination in one case. Moreover, SASFT maintains or even improves the models’ performance on six multilingual benchmarks, showing its effectiveness in addressing code-switching while preserving multilingual capabilities. The code and data are available at https://github.com/Aatrox103/SASFT.

Method: Proposes MultiRole-R1 framework with: 1) Unsupervised data construction pipeline synthesizing reasoning chains with various role perspectives, 2) Reinforcement learning via Group Relative Policy Optimization with reward shaping that includes diversity as a reward signal alongside verifiable rewards.

Result: Training on subjective tasks increases in-domain accuracy by 14.1% and out-of-domain accuracy by 7.64%, even enhances performance on advanced math reasoning (AIME 2024). Diversity proves to be a more consistent indicator of accuracy than reasoning length.

Conclusion: Introducing perspective and token-level diversity significantly improves subjective reasoning capabilities in LRMs, with diversity serving as a better predictor of accuracy than reasoning length.

Abstract: Large Reasoning Models (LRMs) with long chain-of-thought capabilities, optimized via reinforcement learning with verifiable rewards (RLVR), excel at objective reasoning tasks like mathematical problem solving and code generation. However, RLVR is known for degrading generation diversity, which causes LRMs to fall short on subjective reasoning that has multiple answers depending on different role perspectives. While recent studies recognize the importance of diversity-enhanced training in objective reasoning, limited attention has been given to subjective tasks. In this paper, we find that subjective reasoning can be improved by introducing perspective diversity and token-level diversity, with the former one providing a coherent scaffolding anchored to a real-world stakeholder group and the latter one broadening the answer search space. We propose MultiRole-R1, a diversity-enhanced training framework featuring an unsupervised data construction pipeline that synthesizes reasoning chains incorporating various role perspectives. It also employs reinforcement learning via Group Relative Policy Optimization with reward shaping, taking diversity as a reward signal in addition to verifiable reward. Training on subjective tasks solely, MultiRole-R1 increases the in-domain and out-of-domain accuracy by 14.1% and 7.64%, and even enhances the performance on advanced math reasoning such as AIME 2024. We further show that diversity is a more consistent indicator of accuracy than reasoning length.

Method: Two-phase framework: 1) Rote memorization of factual subject-object associations using synthetic semantically meaningless key tokens, 2) Fine-tuning on small set of semantically meaningful prompts to enable generalization.

Result: LLMs can reinterpret rote memorized data through semantically meaningful prompts, evidenced by emergence of structured, semantically aligned latent representations between key tokens and meaningful prompts across 8 tested models.

Conclusion: Rote memorization doesn’t necessarily hinder generalization in LLMs; models can reinterpret memorized data through semantic prompts, opening possibilities for efficient knowledge injection but also risks of malicious data repurposing.

Abstract: Rote learning is a memorization technique based on repetition. Many researchers argue that rote learning hinders generalization because it encourages verbatim memorization rather than deeper understanding. This concern extends even to factual knowledge, which inevitably requires a certain degree of memorization. In this work, we challenge this view and demonstrate that large language models (LLMs) can, in fact, generalize over rote memorized data. We introduce a two-phase “memorize-then-generalize” framework, where the model first rote memorizes factual subject-object associations using a synthetic semantically meaningless key token and then learns to generalize by fine-tuning on a small set of semantically meaningful prompts. Extensive experiments over 8 LLMs show that the models can reinterpret rote memorized data through the semantically meaningful prompts, as evidenced by the emergence of structured, semantically aligned latent representations between the key token and the semantically meaningful prompts. This surprising finding opens the door to both effective and efficient knowledge injection as well as possible risks of repurposing the memorized data for malicious usage.

Method: DeepSieve uses LLM-as-a-knowledge-router to decompose complex queries into structured sub-questions, recursively routes each to the most suitable knowledge source, and filters irrelevant information through a multi-stage distillation process. The framework emphasizes modularity, transparency, and adaptability.

Result: Experiments on multi-hop QA tasks across heterogeneous sources demonstrate improved reasoning depth, retrieval precision, and interpretability over conventional RAG approaches.

Conclusion: DeepSieve provides an effective agentic RAG framework that addresses limitations of existing methods through fine-grained query decomposition and source routing, enabling better handling of knowledge-intensive queries.

Abstract: Large Language Models (LLMs) excel at many reasoning tasks but struggle with knowledge-intensive queries due to their inability to dynamically access up-to-date or domain-specific information. Retrieval-Augmented Generation (RAG) has emerged as a promising solution, enabling LLMs to ground their responses in external sources. However, existing RAG methods lack fine-grained control over both the query and source sides, often resulting in noisy retrieval and shallow reasoning. In this work, we introduce DeepSieve, an agentic RAG framework that incorporates information sieving via LLM-as-a-knowledge-router. DeepSieve decomposes complex queries into structured sub-questions and recursively routes each to the most suitable knowledge source, filtering irrelevant information through a multi-stage distillation process. Our design emphasizes modularity, transparency, and adaptability, leveraging recent advances in agentic system design. Experiments on multi-hop QA tasks across heterogeneous sources demonstrate improved reasoning depth, retrieval precision, and interpretability over conventional RAG approaches. Our codes are available at https://github.com/MinghoKwok/DeepSieve.

Method: Systematically constructs TSQA pairs using temporal databases and database techniques (temporal functional dependencies, temporal SQL, temporal joins). Introduces time accuracy metric to assess validity of time references in explanations alongside answer accuracy.

Result: Enables scalable and comprehensive TSQA evaluation while reducing human labor. Complements current Wikipedia/Wikidata-focused approaches by enabling evaluation on application-specific data.

Conclusion: TDBench provides a systematic, scalable approach to evaluating LLMs on time-sensitive factual knowledge using temporal database techniques, addressing limitations of manual benchmark construction.

Abstract: Facts change over time, making it essential for Large Language Models (LLMs) to handle time-sensitive factual knowledge accurately and reliably. Although factual Time-Sensitive Question-Answering (TSQA) tasks have been widely developed, existing benchmarks often face manual bottlenecks that limit scalable and comprehensive TSQA evaluation. To address this issue, we propose TDBench, a new benchmark that systematically constructs TSQA pairs by harnessing temporal databases and database techniques, such as temporal functional dependencies, temporal SQL, and temporal joins. We also introduce a new evaluation metric called time accuracy, which assesses the validity of time references in model explanations alongside traditional answer accuracy for a more fine-grained TSQA evaluation. Extensive experiments on contemporary LLMs show how TDBench enables scalable and comprehensive TSQA evaluation while reducing the reliance on human labor, complementing current TSQA evaluation approaches that largely center on Wikipedia/Wikidata by enabling LLM evaluation on application-specific data.

Method: DQO uses determinantal point processes to measure semantic diversity as the volume spanned by embeddings of multiple responses. For each prompt, it samples and embeds a group of responses, then uses the determinant of a kernel-based similarity matrix to quantify diversity. The method can be applied on top of existing RL algorithms.

Result: Experiments across instruction-following, summarization, story generation, and reasoning tasks show that DQO substantially improves semantic diversity without sacrificing model quality.

Conclusion: DQO provides an effective training-time solution to the diversity loss problem in RL post-training of LLMs, enabling joint optimization for both quality and semantic diversity across various tasks.

Abstract: Reinforcement learning (RL) has emerged as a popular method for post-training large language models (LLMs). While improving the model’s performance on downstream tasks, it often reduces the model’s output diversity, leading to narrow, canonical responses. Existing methods to enhance diversity are limited, either by operating at inference time or by focusing on surface-level differences. We propose a novel training method named DQO (Diversity Quality Optimization) based on determinantal point processes (DPPs) to jointly optimize LLMs for quality and semantic diversity. Our approach samples and embeds a group of responses for each prompt, then uses the determinant of a kernel-based similarity matrix to measure diversity as the volume spanned by the embeddings of these responses. DQO is flexible and can be applied on top of existing RL algorithms. Experiments across instruction-following, summarization, story generation, and reasoning tasks demonstrate that our method substantially improves semantic diversity without sacrificing model quality.

Method: SVDecode extracts a task-aware steering vector from KL divergence gradient between warm-started and pre-trained models’ output distributions, then uses this vector during decoding to steer outputs toward task distribution.

Result: Across three tasks and nine benchmarks, SVDecode paired with PEFT methods improves multiple-choice accuracy by up to 5 percentage points and open-ended truthfulness by 2 percentage points, with similar gains on commonsense datasets.

Conclusion: SVDecode offers a lightweight, theoretically grounded path to stronger task adaptation for large language models without adding trainable parameters beyond PEFT adapters.

Abstract: Adapting billion-parameter language models to a downstream task is still costly, even with parameter-efficient fine-tuning (PEFT). We re-cast task adaptation as output-distribution alignment: the objective is to steer the output distribution toward the task distribution directly during decoding rather than indirectly through weight updates. Building on this view, we introduce Steering Vector Decoding (SVDecode), a lightweight, PEFT-compatible, and theoretically grounded method. We start with a short warm-start fine-tune and extract a task-aware steering vector from the Kullback-Leibler (KL) divergence gradient between the output distribution of the warm-started and pre-trained models. This steering vector is then used to guide the decoding process to steer the model’s output distribution towards the task distribution. We theoretically prove that SVDecode is first-order equivalent to the gradient step of full fine-tuning and derive a globally optimal solution for the strength of the steering vector. Across three tasks and nine benchmarks, SVDecode paired with four standard PEFT methods improves multiple-choice accuracy by up to 5 percentage points and open-ended truthfulness by 2 percentage points, with similar gains (1-2 percentage points) on commonsense datasets without adding trainable parameters beyond the PEFT adapter. SVDecode thus offers a lightweight, theoretically grounded path to stronger task adaptation for large language models. Code is available at https://github.com/dl-m9/SVDecode.

Method: Large-scale controlled evaluation across multiple axes: LLM-graph interaction modes (prompting, tool-use, code generation), dataset domains (citation, web-link, e-commerce, social networks), structural regimes (homophilic vs heterophilic graphs), feature characteristics (short vs long-text attributes), and model configurations. Also analyzed dependencies by truncating features, deleting edges, and removing labels.

Result: 1) LLMs as code generators achieve strongest overall performance, especially on long-text or high-degree graphs where prompting exceeds token budget. 2) All interaction strategies remain effective on heterophilic graphs, challenging assumptions about LLM collapse under low homophily. 3) Code generation adapts reliance between structure, features, or labels to leverage most informative input type.

Conclusion: Provides comprehensive view of LLM-graph interaction strengths/limitations and key design principles for future approaches, with code generation emerging as most effective strategy for text-rich graph tasks.

Abstract: Large language models (LLMs) are increasingly used for text-rich graph machine learning tasks such as node classification in high-impact domains like fraud detection and recommendation systems. Yet, despite a surge of interest, the field lacks a principled understanding of the capabilities of LLMs in their interaction with graph data. In this work, we conduct a large-scale, controlled evaluation across several key axes of variability to systematically assess the strengths and weaknesses of LLM-based graph reasoning methods in text-based applications. The axes include the LLM-graph interaction mode, comparing prompting, tool-use, and code generation; dataset domains, spanning citation, web-link, e-commerce, and social networks; structural regimes contrasting homophilic and heterophilic graphs; feature characteristics involving both short- and long-text node attributes; and model configurations with varying LLM sizes and reasoning capabilities. We further analyze dependencies by methodically truncating features, deleting edges, and removing labels to quantify reliance on input types. Our findings provide practical and actionable guidance. (1) LLMs as code generators achieve the strongest overall performance on graph data, with especially large gains on long-text or high-degree graphs where prompting quickly exceeds the token budget. (2) All interaction strategies remain effective on heterophilic graphs, challenging the assumption that LLM-based methods collapse under low homophily. (3) Code generation is able to flexibly adapt its reliance between structure, features, or labels to leverage the most informative input type. Together, these findings provide a comprehensive view of the strengths and limitations of current LLM-graph interaction modes and highlight key design principles for future approaches.

Method: Uses token priors estimated from corpus-level term frequency statistics, filtering documents based on mean and standard deviation of token priors without requiring model inference.

Result: Achieves highest average performance across 20 downstream benchmarks while reducing time cost by over 1000x compared to PPL-based filtering, with applicability to code, math, and multilingual corpora.

Conclusion: Prior-based filtering provides a simple, efficient, and effective alternative to perplexity-based methods for data selection in LLM pretraining across various domains.

Abstract: As large language models (LLMs) are pretrained on massive web corpora, careful selection of data becomes essential to ensure effective and efficient learning. While perplexity (PPL)-based filtering has shown strong performance, it suffers from drawbacks: substantial time costs and inherent unreliability of the model when handling noisy or out-of-distribution samples. In this work, we propose a simple yet powerful alternative: a prior-based data filtering method that estimates token priors using corpus-level term frequency statistics, inspired by linguistic insights on word roles and lexical density. Our approach filters documents based on the mean and standard deviation of token priors, serving as a fast proxy to PPL while requiring no model inference. Despite its simplicity, the prior-based filter achieves the highest average performance across 20 downstream benchmarks, while reducing time cost by over 1000x compared to PPL-based filtering. We further demonstrate its applicability to symbolic languages such as code and math, and its dynamic adaptability to multilingual corpora without supervision

Method: Introduces two components: (1) Draft Tree Reward - a sampling-free objective measuring expected acceptance length of draft trees under target model; (2) Group-based Draft Policy Training - stable optimization contrasting trees from current and frozen reference models using debiased group-standardized advantages and PPO-style updates along longest accepted sequences.

Result: Across dialogue (MT-Bench), code (HumanEval), and math (GSM8K) tasks with multiple LLMs (LLaMA-3.1-8B, LLaMA-3.3-70B, Vicuna-1.3-13B, etc.), GTO increases acceptance length by 7.4% and yields additional 7.7% speedup over prior state-of-the-art EAGLE-3.

Conclusion: GTO bridges draft policy misalignment in speculative decoding, offering a practical, general solution for efficient LLM inference with provable improvements in acceptance length and speedup.

Abstract: Speculative decoding accelerates large language model (LLM) inference by letting a lightweight draft model propose multiple tokens that the target model verifies in parallel. Yet existing training objectives optimize only a single greedy draft path, while decoding follows a tree policy that re-ranks and verifies multiple branches. This draft policy misalignment limits achievable speedups. We introduce Group Tree Optimization (GTO), which aligns training with the decoding-time tree policy through two components: (i) Draft Tree Reward, a sampling-free objective equal to the expected acceptance length of the draft tree under the target model, directly measuring decoding performance; (ii) Group-based Draft Policy Training, a stable optimization scheme that contrasts trees from the current and a frozen reference draft model, forming debiased group-standardized advantages and applying a PPO-style surrogate along the longest accepted sequence for robust updates. We further prove that increasing our Draft Tree Reward provably improves acceptance length and speedup. Across dialogue (MT-Bench), code (HumanEval), and math (GSM8K), and multiple LLMs (e.g., LLaMA-3.1-8B, LLaMA-3.3-70B, Vicuna-1.3-13B, DeepSeek-R1-Distill-LLaMA-8B, Qwen3-8B), GTO increases acceptance length by (7.4%) and yields an additional (7.7%) speedup over prior state-of-the-art EAGLE-3. By bridging draft policy misalignment, GTO offers a practical, general solution for efficient LLM inference. Code and draft models are available at https://github.com/hsj576/GTO.

Method: ReMemR1 integrates memory retrieval into memory update process, enabling selective callback of historical memories for non-linear reasoning. Uses multi-level reward design combining final-answer rewards with dense step-level signals to guide effective memory use.

Result: Extensive experiments show ReMemR1 significantly outperforms state-of-the-art baselines on long-context question answering with negligible computational overhead.

Conclusion: ReMemR1 effectively trades marginal cost for robust long-context reasoning by mitigating information degradation and improving supervision through memory retrieval integration and multi-level rewards.

Abstract: Large language models face challenges in long-context question answering, where key evidence of a query may be dispersed across millions of tokens. Existing works equip large language models with a memory buffer that is dynamically updated via a linear document scan, also known as the “memorize while reading” methods. While this approach scales efficiently, it suffers from pruning of latent evidence, information loss through overwriting, and sparse reinforcement learning signals. To tackle these challenges, we present ReMemR1, which integrates the mechanism of memory retrieval into the memory update process, enabling the agent to selectively callback historical memories for non-linear reasoning. To further strengthen training, we propose a multi-level reward design, which combines final-answer rewards with dense, step-level signals that guide effective memory use. Together, these contributions mitigate information degradation, improve supervision, and support complex multi-hop reasoning. Extensive experiments demonstrate that ReMemR1 significantly outperforms state-of-the-art baselines on long-context question answering while incurring negligible computational overhead, validating its ability to trade marginal cost for robust long-context reasoning.

Method: Proposes a novel user simulator architecture that simulates four categories of non-collaborative behaviors: requesting unavailable services, digressing into tangential conversations, expressing impatience, and providing incomplete utterances. The simulator can generate challenging behaviors while still delivering necessary task information.

Result: Experiments on MultiWOZ and τ-bench show significant performance degradation in state-of-the-art tool agents when encountering non-collaborative users, revealing weaknesses like escalated hallucinations and dialogue breakdowns under each non-collaborative condition.

Conclusion: Tool agents need improved robustness to handle diverse real-world user behaviors. The released extensible simulation framework helps the community develop and stress-test tool agents under realistic conditions within their service domains.

Abstract: Tool agents interact with users through multi-turn dialogues to accomplish various tasks. Recent studies have adopted user simulation methods to develop these agents in multi-turn settings. However, existing user simulators tend to be agent-friendly, exhibiting only cooperative behaviors, failing to train and test agents against non-collaborative users in the real world. We propose a novel user simulator architecture that simulates four categories of non-collaborative behaviors: requesting unavailable services, digressing into tangential conversations, expressing impatience, and providing incomplete utterances. Our user simulator can simulate challenging and natural non-collaborative behaviors while reliably delivering all intents and information necessary to accomplish the task. Our experiments on MultiWOZ and τ-bench reveal significant performance degradation in state-of-the-art tool agents when encountering non-collaborative users, as well as agent weaknesses under each non-collaborative condition such as escalated hallucinations and dialogue breakdowns. Our findings point to the need for methods that can improve agent robustness to the wide range of user behaviors encountered in deployment. We release the extensible simulation framework to help the community develop and stress-test tool agents under realistic conditions within their own service domains. Our code is available at https://github.com/holi-lab/NCUser.

Method: Proposes T2PAM paradigm using user feedback as reward signal, then introduces ROSA algorithm for one-step parameter updates toward theoretical optimal policy without iterative optimization.

Result: ROSA achieves significant improvements in task effectiveness and efficiency on challenging benchmarks, with theoretical convergence guarantees to user preferences.

Conclusion: ROSA enables efficient in-conversation self-correction for LLMs through lightweight test-time adaptation, addressing the multi-turn interaction degradation problem.

Abstract: Large Language Models (LLMs) employ multi-turn interaction as a fundamental paradigm for completing complex tasks. However, their performance often degrades in extended interactions, as they are typically trained on static, single-turn data, which hinders their ability to adapt to real-time user feedback. To address this limitation, we first propose a new paradigm: Test-Time Policy Adaptation for Multi-Turn Interactions (T2PAM), which utilizes user feedback from the ongoing interaction as a reward signal to estimate a latent optimal policy aligned with user preferences, then updates a small subset of parameters to steer the model toward this policy, ultimately enabling efficient in-conversation self-correction. We then introduce Optimum-Referenced One-Step Adaptation (ROSA), a lightweight algorithm that operationalizes T2PAM. ROSA guides the model parameters toward a theoretical optimal policy in a single, efficient update step, avoiding costly iterative gradient-based optimization and minimizing computational overhead. We provide a rigorous theoretical analysis guaranteeing that the policy of ROSA converges to the preference of user as the number of interactions increases. Extensive experiments on challenging benchmark demonstrate that ROSA achieves significant improvements in both task effectiveness and efficiency.

Method: Proposes conducting multiple experiments simultaneously during a single training run, validated with 10 experiments on 210B tokens with models up to 2.7B parameters. Introduces Continual Pretraining Dependence Testing (CPDT) to test for interactions between experiments.

Result: Successfully replicated previous results on data contamination, poisoning, and memorization, and conducted novel investigations into knowledge acquisition, mathematical reasoning, and watermarking. Found minimal influence of experiments on training dynamics and performance, with negligible interactions between experiments.

Conclusion: Multiple pretraining experiments can be conducted within a single training run, enabling rigorous scientific experimentation with large models on a limited compute budget.

Abstract: Recent work has demonstrated that controlled pretraining experiments are a powerful tool for studying the relationship between training data and large language model (LLM) behavior. However, the computational cost of pretraining presents a significant constraint. To overcome this constraint, we propose a new approach where multiple experiments are conducted simultaneously during a single training run. We validate our approach by performing ten experiments while training on 210B tokens, with models of up to 2.7B parameters. Although models are trained only once, we can replicate the results of multiple previous works on data contamination, poisoning, and memorization. We also conduct novel investigations into knowledge acquisition, mathematical reasoning, and watermarking. For example, we dynamically update the training data until a model acquires a particular piece of knowledge. Remarkably, the influence of the experiments on the model’s training dynamics and overall performance is minimal. However, interactions between experiments may act as a confounder in our approach. We propose continual pretraining dependence testing (CPDT), a novel technique to test for interactions with continual pretraining experiments, finding them to be negligible in our setup. Overall, our results suggest that performing multiple pretraining experiments within a single training run can enable rigorous scientific experimentation with large models on a compute budget.

Method: Introduces MetaTuner with two neural networks: one to generate prompts and another to generate parameters, sharing a common bottom encoding layer for knowledge sharing. Uses supervised regularization loss to handle the discrete optimization of prompts and continuous optimization of fine-tuning parameters.

Result: Extensive experiments across diverse benchmarks show that MetaTuner consistently outperforms baseline methods.

Conclusion: Joint integration of prompt optimization and fine-tuning through MetaTuner framework enables synergistic improvements in LLM performance beyond what either approach can achieve alone.

Abstract: Prompt optimization and fine-tuning are two major approaches to improve the performance of Large Language Models (LLMs). They enhance the capabilities of LLMs from complementary perspectives: the former through explicit natural language, and the latter through implicit parameter updates. However, prior work has typically studied them in isolation, leaving their synergistic potential largely underexplored. To bridge this gap, in this paper, we introduce MetaTuner, a novel framework that jointly integrates prompt optimization and fine-tuning for LLM training. Specifically, we introduce two neural networks to generate prompts and parameters, respectively, while allowing them to share a common bottom encoding layer to enable knowledge sharing. By the guidance of the final supervised signals, our framework is optimized to discover the optimal combinations between the prompts and parameters. Given that prompt learning involves discrete optimization while fine-tuning operates in a continuous parameter space, we design a supervised regularization loss to train our framework effectively. Extensive experiments across diverse benchmarks show that our method consistently outperforms the baselines.

Method: Built on the Matter protocol (industry standard for smart devices), SimuHome provides APIs for agents to interact with devices and observe continuous environmental variable changes. It includes a benchmark of 600 episodes covering state inquiry, implicit intent inference, explicit device control, and workflow scheduling with both feasible and infeasible requests. Time acceleration enables immediate evaluation of scheduled workflows.

Result: Evaluation of 18 agents shows workflow scheduling is the hardest category, with failures persisting across alternative agent frameworks and fine-tuning. The simulator successfully reveals agent limitations in complex scheduling tasks.

Conclusion: SimuHome’s time-accelerated simulation could serve as a valuable environment for agents to pre-validate actions before real-world deployment, addressing critical gaps in current smart home agent evaluation.

Abstract: We introduce $\textbf{SimuHome}$, a high-fidelity smart home simulator and a benchmark of 600 episodes for LLM-based smart home agents. Existing smart home benchmarks treat the home as a static system, neither simulating how device operations affect environmental variables over time nor supporting workflow scheduling of device commands. SimuHome is grounded in the Matter protocol, the industry standard that defines how real smart home devices communicate and operate. Agents interact with devices through SimuHome’s APIs and observe how their actions continuously affect environmental variables such as temperature and humidity. Our benchmark covers state inquiry, implicit user intent inference, explicit device control, and workflow scheduling, each with both feasible and infeasible requests. For workflow scheduling, the simulator accelerates time so that scheduled workflows can be evaluated immediately. An evaluation of 18 agents reveals that workflow scheduling is the hardest category, with failures persisting across alternative agent frameworks and fine-tuning. These findings suggest that SimuHome’s time-accelerated simulation could serve as an environment for agents to pre-validate their actions before committing them to the real world.

Method: SUIT (Subspace Knowledge Edit) identifies and operates within the subspace of critical features relevant to each edit, computing key and value vectors only within this subspace to minimize unintended perturbations while effectively editing knowledge.

Result: Empirical results on LLaMA3, GPT-J, and Qwen2.5 models show SUIT dramatically improves knowledge preservation over strong baselines while maintaining high editing performance, with reduced unintended perturbations in hidden states.

Conclusion: Edit-critical subspace identification is a key principle for reliable, low-perturbation knowledge editing, establishing SUIT as an effective approach that confines updates to directions more effective for editing while preserving model integrity.

Abstract: Knowledge editing aims to efficiently correct factual errors in language models. Widely used locate-then-edit methods update an MLP layer by adjusting its weights to change the mapping between the layer’s input vector (key) and output vector (value), thereby editing the model’s knowledge. As this update is driven by key and value vectors, obtaining these vectors without careful constraints causes significant model perturbations beyond the targeted edit, a common issue in many prior knowledge editing methods. To address this, we propose Subspace Knowledge Edit (SUIT), which computes key and value vectors only within the subspace of critical features relevant to the edit. Our empirical results on LLaMA3, GPT-J, and Qwen2.5 models show that SUIT dramatically improves knowledge preservation over strong baselines while maintaining high editing performance. These results support the claim that SUIT successfully identifies the critical subspace for the edit. Beyond quantitative gains, our analyses show that SUIT reduces unintended perturbations in hidden states while confining updates to directions that are more effective for editing. Taken together, these findings establish edit-critical subspace identification as a key principle for reliable, low-perturbation knowledge editing. Our code is available at https://github.com/holi-lab/SUIT.

Method: Built on vLLM with modular architecture featuring pluggable interfaces for analysis-based and learning-based methods, fine-grained parameter control, pre-computed steering vectors for eight application domains, and an interactive demonstration system.

Result: Achieves 10.8-22.3× speedup over existing frameworks through deep integration with vLLM’s optimized inference engine, demonstrating effectiveness in overthinking mitigation, hallucination reduction, and other key applications.

Conclusion: EasySteer transforms steering from a research technique to a production-ready capability, establishing critical infrastructure for deployable, controllable language models.

Abstract: Large language model (LLM) steering has emerged as a promising paradigm for controlling model behavior at inference time through targeted manipulation of hidden states, offering a lightweight alternative to expensive retraining. However, existing steering frameworks suffer from critical limitations: computational inefficiency, limited extensibility, and restricted functionality that hinder both research progress and practical deployment. We present EasySteer, a unified framework for high-performance, extensible LLM steering built on vLLM. Our system features modular architecture with pluggable interfaces for both analysis-based and learning-based methods, fine-grained parameter control, pre-computed steering vectors for eight application domains, and an interactive demonstration system. Through deep integration with vLLM’s optimized inference engine, EasySteer achieves 10.8-22.3$\times$ speedup over existing frameworks. Extensive experiments demonstrate its effectiveness in overthinking mitigation, hallucination reduction, and other key applications. EasySteer transforms steering from research technique to production-ready capability, establishing critical infrastructure for deployable, controllable language models.

Method: Proposes Distractor-Normalized Coherence (DINCO): 1) LLM generates distractors (alternative claims), 2) verbalizes confidence independently for each distractor, 3) normalizes by total verbalized confidence to account for suggestibility bias. Combines with generator-validator disagreement for further calibration.

Result: DINCO provides less saturated, more usable confidence estimates. Outperforms self-consistency baselines with fewer inference calls (DINCO at 10 calls beats self-consistency at 100 calls). Empirical validation shows more suggestibility on lower-accuracy claims.

Conclusion: DINCO effectively addresses LLM overconfidence by normalizing for suggestibility bias through distractor-based coherence measurement, providing better calibrated confidence estimates crucial for trustworthy LLM deployment.

Abstract: Calibrated confidence estimates are necessary for large language model (LLM) outputs to be trusted by human users. While LLMs can express their confidence in human-interpretable ways, verbalized LLM-generated confidence scores have empirically been found to be miscalibrated, reporting high confidence on instances with low accuracy and thereby harming trust and safety. We hypothesize that this overconfidence often stems from a given LLM’s heightened suggestibility when faced with claims that it encodes little information about; we empirically validate this hypothesis, finding more suggestibility on lower-accuracy claims. Building on this finding, we introduce Distractor-Normalized Coherence (DINCO), which estimates and accounts for an LLM’s suggestibility bias by having the model verbalize its confidence independently across several self-generated distractors (i.e. alternative claims), and normalizes by the total verbalized confidence. To further improve calibration, we leverage generator-validator disagreement, augmenting normalized validator confidence with a consistency-based estimate of generator confidence. Here, we frame the popular approach of self-consistency as leveraging coherence across sampled generations, and normalized verbalized confidence as leveraging coherence across validations on incompatible claims, allowing us to integrate these complementary dimensions of coherence into DINCO. Moreover, our analysis shows that DINCO provides less saturated – and therefore more usable – confidence estimates, and that further sampling alone cannot close the gap between DINCO and baselines, with DINCO at 10 inference calls outperforming self-consistency at 100.

Method: Introduced MENLO framework based on audience design-inspired mechanisms; created human-annotated dataset covering four quality dimensions; evaluated zero-shot LLM judges with pairwise evaluation and structured rubrics; improved through fine-tuning with reinforcement learning, reward shaping, and multi-task learning.

Result: Zero-shot LLM judges benefit from pairwise evaluation and structured rubrics but still underperform humans; substantial improvements achieved through fine-tuning; RL-trained judges can serve as generative reward models to enhance LLMs’ multilingual proficiency, though discrepancies with human judgment remain.

Conclusion: MENLO provides promising directions for scalable multilingual evaluation and preference alignment, with released dataset and framework supporting further research in multilingual LLM evaluation.

Abstract: Ensuring native-like quality of large language model (LLM) responses across many languages is challenging. To address this, we introduce MENLO, a framework that operationalizes the evaluation of native-like response quality based on audience design-inspired mechanisms. Using MENLO, we create a dataset of 6,423 human-annotated prompt-response preference pairs covering four quality dimensions with high inter-annotator agreement in 47 language varieties. Our evaluation reveals that zero-shot LLM judges benefit significantly from pairwise evaluation and our structured annotation rubrics, yet they still underperform human annotators on our dataset. We demonstrate substantial improvements through fine-tuning with reinforcement learning, reward shaping, and multi-task learning approaches. Additionally, we show that RL-trained judges can serve as generative reward models to enhance LLMs’ multilingual proficiency, though discrepancies with human judgment remain. Our findings suggest promising directions for scalable multilingual evaluation and preference alignment. We release our dataset and evaluation framework to support further research in multilingual LLM evaluation (https://huggingface.co/datasets/facebook/menlo).

Method: Uses Hyperspherical Energy (HE) to quantify neuron uniformity during editing, identifies correlation between HE dynamics and editing performance. Proposes SPHERE: Sparse Projection for Hyperspherical Energy-Regularized Editing, which identifies a sparse space complementary to principal hyperspherical directions of pretrained weight matrices and projects new knowledge onto it to attenuate perturbations.

Result: SPHERE outperforms best baselines in editing capability by average of 16.41% on LLaMA3 (8B) and Qwen2.5 (7B), while better preserving general model performance. Shows strong correlation between HE dynamics and editing performance, with editing failures coinciding with high HE fluctuations.

Conclusion: Hyperspherical uniformity is crucial for stable sequential editing. SPHERE provides a principled approach to reliable large-scale knowledge editing by stabilizing neuron weight distributions, offering a lightweight alternative to full retraining for keeping LLMs updated.

Abstract: Large language models (LLMs) require constant updates to remain aligned with evolving real-world knowledge. Model editing offers a lightweight alternative to retraining, but sequential editing often destabilizes representations and induces catastrophic forgetting. In this work, we seek to better understand and mitigate performance degradation caused by sequential editing. We hypothesize that hyperspherical uniformity, a property that maintains uniform distribution of neuron weights on a hypersphere, helps the model remain stable, retain prior knowledge, while still accommodate new updates. We use Hyperspherical Energy (HE) to quantify neuron uniformity during editing, and examine its correlation with editing performance. Empirical studies across widely used editing methods reveals a strong correlation between HE dynamics and editing performance, with editing failures consistently coinciding with high HE fluctuations. We further theoretically prove that HE dynamics impose a lower bound on the degradation of pretrained knowledge, highlighting why HE stability is crucial for knowledge retention. Motivated by these insights, we propose SPHERE (Sparse Projection for Hyperspherical Energy-Regularized Editing), an HE-driven regularization strategy that stabilizes neuron weight distributions, ultimately preserving prior knowledge while enabling reliable sequential updates. Specifically, SPHERE identifies a sparse space complementary to the principal hyperspherical directions of the pretrained weight matrices and projects new knowledge onto it, attenuating perturbations on the principal directions. Extensive experiments on LLaMA3 (8B) and Qwen2.5 (7B) show that SPHERE outperforms the best baseline in editing capability by an average of 16.41%, while most faithfully preserving general model performance, thereby offering a principled path toward reliable large-scale knowledge editing.

Method: PoLi-RL uses a two-stage curriculum: first trains with simple pointwise reward for basic scoring, then transitions to hybrid reward combining pointwise, pairwise, and listwise objectives. Introduces Parallel Slice Ranking Reward (PSRR) that computes ranking rewards in parallel slices (completions with same index from different samples) for granular credit assignment.

Result: Achieves Spearman correlation coefficient of 48.18 on official C-STS benchmark, establishing new SOTA for cross-encoder architecture.

Conclusion: First successful application of RL to C-STS, introducing powerful paradigm for aligning LLMs for complex, ranking-based conditional judgment tasks.

Abstract: Conditional Semantic Textual Similarity (C-STS) measures the semantic proximity between text segments under a specific condition, thereby overcoming the ambiguity inherent in traditional STS. However, existing methods are largely confined to discriminative models, failing to fully leverage recent breakthroughs in the NLP community involving Large Language Models (LLMs) and Reinforcement Learning (RL). RL is a particularly well-suited paradigm for this task, as it can directly optimize the non-differentiable Spearman ranking metric and guide the reasoning process required by C-STS. Nevertheless, we find that naively applying listwise RL fails to produce meaningful improvements, as the model struggles with complex, coarse-grained reward signals, leading to optimization difficulties. To address this challenge, we introduce PoLi-RL, a novel Point-to-List Reinforcement Learning framework. PoLi-RL employs a two-stage curriculum: it first trains the model with a simple pointwise reward to establish fundamental scoring capabilities, then transitions to a hybrid reward that combines pointwise, pairwise, and listwise objectives to refine the model’s ability to discern subtle semantic distinctions. Crucially, we propose an innovative Parallel Slice Ranking Reward (PSRR) mechanism that computes ranking rewards in parallel slices, where each slice consists of completions with the same index from different samples. This provides a precise, differentiated learning signal for each individual completion, enabling granular credit assignment and effective optimization. On the official C-STS benchmark, PoLi-RL achieves a Spearman correlation coefficient of 48.18, establishing a new SOTA for the cross-encoder architecture. As the first work to successfully apply RL to C-STS, our study introduces a powerful paradigm for aligning LLMs for complex, ranking-based conditional judgment tasks.

Method: TiTok uses token-wise contrastive excess between a source model with and without LoRA to capture task-relevant information, enabling selective filtering of synthetic data without additional models or overhead.

Result: Experiments on three benchmarks across multiple transfer settings show TiTok consistently effective with average performance gains of +4~10% compared to baselines.

Conclusion: TiTok provides an efficient framework for LoRA transplantation that works across different model backbones without requiring additional models or complex synthetic data generation.

Abstract: Large Language Models (LLMs) are widely applied in real world scenarios, yet fine-tuning them comes with significant computational and storage costs. Parameter-Efficient Fine-Tuning (PEFT) methods such as LoRA mitigate these costs; however, the adapted parameters are dependent on the base model and cannot be transferred across different backbones. One way to address this issue is through knowledge distillation, but its effectiveness inherently depends on training data. Recent work such as TransLoRA avoids this by generating synthetic data; nevertheless, this adds complexity since it requires training an additional discriminator model. In this paper, we propose TiTok, a new framework that enables effective LoRA Transplantation through Token-level knowledge transfer. Specifically, TiTok captures task-relevant information through a token-wise contrastive excess between a source model with and without LoRA. This excess highlights informative tokens and enables selective filtering of synthetic data, all without additional models or overhead. Through experiments on three benchmarks across multiple transfer settings, we demonstrate that TiTok is consistently effective, achieving average performance gains of +4~10% compared to baselines overall.

Method: SwiReasoning dynamically switches between explicit and latent reasoning guided by block-wise confidence estimated from entropy trends in next-token distributions. It also limits maximum thinking-block switches to curb overthinking.

Result: On mathematics, STEM, coding, and general benchmarks, SwiReasoning improves average accuracy by 1.8%-3.1% across different LLMs. Under constrained budgets, it improves average token efficiency by 57%-79%, with larger gains as budgets tighten.

Conclusion: SwiReasoning effectively addresses challenges in latent reasoning by balancing exploration and exploitation through dynamic switching between explicit and latent reasoning, improving both accuracy and efficiency.

Abstract: Recent work shows that, beyond discrete reasoning through explicit chain-of-thought steps, which are limited by the boundaries of natural languages, large language models (LLMs) can also reason continuously in latent space, allowing richer information per step and thereby improving token efficiency. Despite this promise, latent reasoning still faces two challenges, especially in training-free settings: 1) purely latent reasoning broadens the search distribution by maintaining multiple implicit paths, which diffuses probability mass, introduces noise, and impedes convergence to a single high-confidence solution, thereby hurting accuracy; and 2) overthinking persists even without explicit text, wasting tokens and degrading efficiency. To address these issues, we introduce SwiReasoning, a training-free framework for LLM reasoning which features two key innovations: 1) SwiReasoning dynamically switches between explicit and latent reasoning, guided by block-wise confidence estimated from entropy trends in next-token distributions, to balance exploration and exploitation and promote timely convergence. 2) By limiting the maximum number of thinking-block switches, SwiReasoning curbs overthinking and improves token efficiency across varying problem difficulties. On widely used mathematics, STEM, coding, and general benchmarks, SwiReasoning consistently improves average accuracy by 1.8%-3.1% across reasoning LLMs of different model families and scales. Furthermore, under constrained budgets, SwiReasoning improves average token efficiency by 57%-79%, with larger gains as budgets tighten.

Method: Treats parallel reasoning as a set-of-next-token-prediction problem, incorporates a set-based global loss into SFT using bipartite matching between global forking tokens and unique reasoning traces. Proposes Set Supervised Fine-Tuning (SSFT) to preserve unique reasoning modes and Global Forking Policy Optimization (GFPO) to leverage maximally steerable tokens.

Result: SSFT preserves unique reasoning modes and produces emergent global forking tokens. GFPO models consistently outperform SFT counterparts with GRPO on both math reasoning and execution-based code generation benchmarks.

Conclusion: The proposed set-based approach effectively addresses the diversity-accuracy trade-off in parallel reasoning, enabling better scaling of test-time compute for complex reasoning tasks through preserved reasoning diversity and optimized forking tokens.

Abstract: Although LLMs have demonstrated improved performance by scaling parallel test-time compute, doing so relies on generating reasoning paths that are both diverse and accurate. For challenging problems, the forking tokens that trigger diverse yet correct reasoning modes are typically deep in the sampling tree. Consequently, common strategies to encourage diversity, such as temperature scaling, encounter a worsened trade-off between diversity and accuracy. Motivated by this challenge, we treat parallel reasoning as a set-of-next-token-prediction problem and incorporate a set-based global loss into Supervised Fine-Tuning (SFT) using bipartite matching between global forking tokens and unique reasoning traces. We observe that whereas naive fine-tuning with multiple reasoning traces collapses these unique reasoning modes, our proposed method, Set Supervised Fine-Tuning (SSFT), preserves these modes and produces emergent global forking tokens. Global Forking Policy Optimization (GFPO) leverages these maximally steerable tokens to incentivize complex reasoning, and the resulting models consistently outperform their SFT counterparts with GRPO on both math reasoning and execution-based code generation benchmarks.

Method: Used mean reward of sampled responses as a proxy for prompt difficulty, analyzed performance across different difficulty levels, and explored strategies to mitigate negative effects of difficult prompts.

Result: Difficult prompts show inferior optimization performance compared to easy prompts; incorporating difficult prompts degrades overall performance; performance improves as difficulty decreases; model capacity helps close the gap; training on only 30% easiest prompts improves performance on AlpacaEval~2 and Arena-Hard.

Conclusion: Prompt difficulty significantly impacts self-play preference optimization, and selective training on easier prompts can improve overall alignment performance.

Abstract: Self-play preference optimization has emerged as a prominent paradigm for aligning large language models (LLMs). It typically involves a language model to generate on-policy responses for prompts and a reward model (RM) to guide the selection of chosen and rejected responses, which can be further trained with direct preference optimization (DPO). However, the role of prompts remains underexplored, despite being a core component in this pipeline. In this work, we investigate how prompts of varying difficulty influence self-play preference optimization. We use the mean reward of sampled responses of a prompt as a proxy for its difficulty. We first find that difficult prompts exhibit substantially inferior self-play optimization performance compared to easy prompts for language models. Moreover, incorporating difficult prompts into training fails to enhance overall performance and, in fact, leads to slight degradation compared to training on easy prompts alone. Third, there is a clear upward trend in optimization performance as prompt difficulty decreases. We also observe that the performance gap between difficult and easy prompts tends to close as the model capacity increases, suggesting that prompt difficulty interacts with the model capacity. Building on these findings, we explore strategies to mitigate the adversary effect of difficult prompts on final performance. We demonstrate that only training on a small portion (30%) of the easiest prompts improves overall self-play performance on AlpacaEval~2 and Arena-Hard. We also report failed attempts and lessons learned.

Method: Proposes MASA with multi-A, single-B structure where multiple specialized down-projection experts are asymmetrically shared across layers, integrated by layer-specific up-projection matrices for parameter efficiency.

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

Conclusion: MASA effectively addresses LoRA’s representational bottleneck through multi-expert feature adaptation while maintaining parameter efficiency, demonstrating improved downstream task performance.

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

Method: Proposes LAD-RAG with two-phase approach: 1) Ingestion phase constructs symbolic document graph capturing layout structure and cross-page dependencies alongside neural embeddings; 2) Inference phase uses LLM agent to dynamically interact with both neural and symbolic indices for adaptive evidence retrieval.

Result: Achieves over 90% perfect recall on average without top-k tuning, outperforms baseline retrievers by up to 20% in recall at comparable noise levels, and yields higher QA accuracy with minimal latency on MMLongBench-Doc, LongDocURL, DUDE, and MP-DocVQA benchmarks.

Conclusion: LAD-RAG effectively addresses limitations of conventional RAG for visually rich documents by incorporating layout awareness and dynamic retrieval, significantly improving evidence retrieval and QA performance for multi-page reasoning tasks.

Abstract: Question answering over visually rich documents (VRDs) requires reasoning not only over isolated content but also over documents’ structural organization and cross-page dependencies. However, conventional retrieval-augmented generation (RAG) methods encode content in isolated chunks during ingestion, losing structural and cross-page dependencies, and retrieve a fixed number of pages at inference, regardless of the specific demands of the question or context. This often results in incomplete evidence retrieval and degraded answer quality for multi-page reasoning tasks. To address these limitations, we propose LAD-RAG, a novel Layout-Aware Dynamic RAG framework. During ingestion, LAD-RAG constructs a symbolic document graph that captures layout structure and cross-page dependencies, adding it alongside standard neural embeddings to yield a more holistic representation of the document. During inference, an LLM agent dynamically interacts with the neural and symbolic indices to adaptively retrieve the necessary evidence based on the query. Experiments on MMLongBench-Doc, LongDocURL, DUDE, and MP-DocVQA demonstrate that LAD-RAG improves retrieval, achieving over 90% perfect recall on average without any top-k tuning, and outperforming baseline retrievers by up to 20% in recall at comparable noise levels, yielding higher QA accuracy with minimal latency.

Method: ExPO-HM combines supervised fine-tuning warmup, GRPO with curriculum learning, and Conditional Decision Entropy as both metric and reward for reasoning quality. Inspired by human annotator training, it focuses on generating explanations before detection.

Result: Achieves state-of-the-art performance on three hateful meme benchmarks with up to 15% and 17% F1 improvement over GRPO and DPO baselines respectively. Excels in binary detection, fine-grained classification, and reasoning quality.

Conclusion: ExPO-HM successfully moves hateful meme detection from simple binary alarms to explanation-driven detection, providing accurate, interpretable, and actionable moderation support.

Abstract: Hateful memes have emerged as a particularly challenging form of online abuse, motivating the development of automated detection systems. Most prior approaches rely on direct detection, producing only binary predictions. Such models fail to provide the context and explanations that real-world moderation requires. Recent Explain-then-Detect approaches, using Chain-of-Thought prompting or LMM agents, perform worse than simple SFT baselines, and even advanced post-training methods such as GRPO fail to close the gap. Our analysis identifies two key issues of such systems: important policy-relevant cues such as targets and attack types are not hypothesized by the model as a likely explanation; and the binary reward signal is insufficient to guide reasoning. To address these challenges, we propose ExPO-HM (Explain-then-Detect Policy Optimization for Hateful Memes), inspired by the training and evaluation process of human annotators. ExPO-HM combines SFT warmup, GRPO with curriculum learning, and Conditional Decision Entropy (CDE) as both metric and reward for reasoning quality. Across three hateful meme benchmarks, ExPO-HM achieves state-of-the-art performance on binary detection, fine-grained classification, and reasoning quality, with up to 15% and 17% F1 improvement over the GRPO and DPO baselines, respectively. By moving hateful meme detection from simple binary alarms to explanation-driven detection, ExPO-HM provides accurate, interpretable, and actionable moderation support. Code available at https://github.com/JingbiaoMei/ExPO-HM

Method: Formalizes benchmark construction as finding optimal diagnostic basis in binary code-test matrix, proposes WrongSelect algorithm to select maximally diverse wrong codes, and constructs TC-Bench from competitive programming submissions.

Result: Created TC-Bench benchmark where even advanced test case generation methods achieve only ~60% exclusion rates, exposing significant gaps in diagnostic power.

Conclusion: The framework provides a principled approach to benchmark construction that resists inflation and reveals substantial room for improvement in test case generation methods.

Abstract: Evaluating test cases automatically generated by Large Language Models (LLMs) is a critical yet challenging task. Existing benchmarks often evaluate the exclusion ratio on large, unstructured collections of wrong codes, suffering from high computational costs and score inflation. Furthermore, they inadvertently reward generators that detect common, trivial bugs, while failing to penalize their inability to identify rare yet critical faults. In this work, we connect two fundamental questions: (1) What is the minimal set of wrong codes sufficient to represent the entire error space? and (2) What is the minimal set of test cases needed to distinguish them? We introduce a novel framework that formalizes benchmark construction as finding an optimal diagnostic basis in a binary code-test matrix, where rows represent wrong codes and columns represent test case results. The rank of this matrix specifies the minimal number of independent error patterns (wrong codes) and provides a tight upper bound on the number of test cases required for complete fault coverage. Our objective is to identify a basis of size equal to the matrix rank that maximizes internal diversity. To tackle this NP-hard problem, we propose WrongSelect, an efficient approximation algorithm to select maximally diverse wrong codes. Applying this framework to millions of competitive programming submissions, we construct TC-Bench, a compact, diverse, and inflation-resistant benchmark. Extensive experiments show that even the most advanced test case generation methods achieve only ~60% exclusion rates on TC-Bench, exposing a significant gap in their diagnostic power and highlighting substantial room for future improvement. Our dataset is available at: https://huggingface.co/datasets/Luoberta/TC-Bench and our code is at: https://github.com/Luowaterbi/TC-Bench.

Method: Proposes latent-space language steering: 1) Uses PCA on parallel translations to identify language directions in the latent space, 2) Steers token embeddings along these axes during inference to control language identity, 3) Requires only minimal parallel data for calibration and adds negligible computational overhead.

Result: Achieves 95-99% language classification accuracy using a single principal component, reduces next-token distributional divergence by up to 55% across multiple language pairs on Qwen2.5 and Llama-3.2 models. Generation-based evaluation shows 63-99% reduction in Code-Switching Index across four language pairs (p < 0.001). Analysis reveals language identity concentrates in final layers with near-perfect linear separability.

Conclusion: Latent-space language steering effectively mitigates code-switching in multilingual LLMs while preserving semantics, with minimal computational overhead and data requirements. The method demonstrates that language identity is highly structured in the latent space and can be controlled via simple linear transformations.

Abstract: Multilingual Large Language Models (LLMs) often exhibit unintended code-switching, reducing reliability in downstream tasks. We propose latent-space language steering, a lightweight inference-time method that identifies language directions via PCA on parallel translations and steers token embeddings along these axes to control language identity. Our approach mitigates code-switching while preserving semantics with negligible computational overhead and requires only minimal parallel data for calibration. Empirically, we achieve 95-99% language classification accuracy using a single principal component and reduce next-token distributional divergence by up to 55% across multiple language pairs on Qwen2.5 and Llama-3.2 models. Generation-based evaluation on Llama-3.2 further demonstrates 63–99% reduction in Code-Switching Index across four language pairs ($p < 0.001$). We further analyze the layer-wise evolution of language representations, revealing that language identity concentrates in final layers with near-perfect linear separability.

Method: Introduces ProofBench dataset with expert-annotated proof ratings, then systematically explores evaluator design space across backbone models, input context, instructions, and evaluation workflow to develop ProofGrader with ensembling.

Result: ProofGrader achieves MAE of 0.926 against expert scores, significantly outperforming baselines, and in best-of-n selection achieves 4.14/7 score, closing 78% of gap between naive binary evaluator and human oracle.

Conclusion: ProofGrader demonstrates strong performance as a fine-grained proof evaluator and has potential to advance downstream proof generation tasks.

Abstract: Recent advances in large language models (LLMs) for mathematical reasoning have largely focused on tasks with easily verifiable final answers while generating and verifying natural language math proofs remains an open challenge. We identify the absence of a reliable, fine-grained evaluator for LLM-generated math proofs as a critical gap. To address this, we propose a systematic methodology for developing and validating evaluators that assign fine-grained scores on a 0-7 scale to model-generated math proofs. To enable this study, we introduce ProofBench, the first expert-annotated dataset of fine-grained proof ratings, spanning 145 problems from six major math competitions (USAMO, IMO, Putnam, etc) and 435 LLM-generated solutions from Gemini-2.5-Pro, o3, and DeepSeek-R1. Using ProofBench as a testbed, we systematically explore the evaluator design space across key axes: the backbone model, input context, instructions and evaluation workflow. Our analysis delivers ProofGrader, an evaluator that combines a strong reasoning backbone LM, rich context from reference solutions and marking schemes, and a simple ensembling method; it achieves a low Mean Absolute Error (MAE) of 0.926 against expert scores, significantly outperforming naive baselines. Finally, we demonstrate its practical utility in a best-of-$n$ selection task: at $n=16$, ProofGrader achieves an average score of 4.14/7, closing 78% of the gap between a naive binary evaluator (2.48) and the human oracle (4.62), highlighting its potential to advance downstream proof generation.

Method: Introduces PolySkill framework inspired by polymorphism in software engineering, decoupling a skill’s abstract goal (what it accomplishes) from its concrete implementation (how it’s executed), enabling agents to learn generalizable and compositional skills.

Result: Improves skill reuse by 1.7x on seen websites, boosts success rates by up to 9.4% on Mind2Web and 13.9% on unseen websites, reduces steps by over 20%, and enhances task quality in self-exploration settings.

Conclusion: Separating skill goals from execution is crucial for developing autonomous agents that can learn and generalize across the open web continuously, providing a practical path toward continual learning in adaptive environments.

Abstract: Large language models (LLMs) are moving beyond static uses and are now powering agents that learn continually during their interaction with external environments. For example, agents can learn reusable skills while navigating web pages or toggling new tools. However, existing methods for skill learning often create skills that are over-specialized to a single website and fail to generalize. We introduce PolySkill, a new framework that enables agents to learn generalizable and compositional skills. The core idea, inspired by polymorphism in software engineering, is to decouple a skill’s abstract goal (what it accomplishes) and its concrete implementation (how it is executed). Experiments show that our method (1) improves skill reuse by 1.7x on seen websites and (2) boosts success rates by up to 9.4% on Mind2Web and 13.9% on unseen websites, while reducing steps by over 20%. (3) In self-exploration settings without specified tasks, our framework improves the quality of proposed tasks and enables agents to learn generalizable skills that work across different sites. By enabling the agent to identify and refine its own goals, the PolySkill enhances the agent’s ability to learn a better curriculum, leading to the acquisition of more generalizable skills compared to baseline methods. This work provides a practical path toward building agents capable of continual learning in adaptive environments. Our findings show that separating a skill’s goal from its execution is a crucial step toward developing autonomous agents that can learn and generalize across the open web continuously. Our code can be found in https://github.com/simonucl/PolySkill.

Method: The paper presents the first systematic comparison of prompting-based and RL-based query augmentation across diverse benchmarks (evidence-seeking, ad hoc, and tool retrieval). Based on findings that simple prompting often matches RL methods, they introduce OPQE - On-policy Pseudo-document Query Expansion, where an LLM policy learns to generate pseudo-documents that maximize retrieval performance, combining prompting’s flexibility with RL’s targeted optimization.

Result: Key finding: Simple, training-free query augmentation often performs on par with or surpasses more expensive RL-based methods, especially with powerful LLMs. The novel OPQE hybrid method outperforms both standalone prompting and RL-based rewriting, showing synergistic approaches yield best results.

Conclusion: A hybrid approach combining prompting’s generative flexibility with RL’s optimization capabilities (OPQE) achieves superior retrieval performance compared to either method alone, demonstrating the value of synergistic methods in LLM-based query augmentation.

Abstract: Recent advances in large language models (LLMs) have led to a surge of interest in query augmentation for information retrieval (IR). Two main approaches have emerged. The first prompts LLMs to generate answers or pseudo-documents that serve as new queries, relying purely on the model’s parametric knowledge or contextual information. The second applies reinforcement learning (RL) to fine-tune LLMs for query rewriting, directly optimizing retrieval metrics. While having respective advantages and limitations, the two approaches have not been compared under consistent experimental conditions. In this work, we present the first systematic comparison of prompting-based and RL-based query augmentation across diverse benchmarks, including evidence-seeking, ad hoc, and tool retrieval. Our key finding is that simple, training-free query augmentation often performs on par with, or even surpasses, more expensive RL-based counterparts, especially when using powerful LLMs. Motivated by this discovery, we introduce a novel hybrid method, On-policy Pseudo-document Query Expansion (OPQE), which, instead of rewriting a query, the LLM policy learns to generate a pseudo-document that maximizes retrieval performance, thus merging the flexibility and generative structure of prompting with the targeted optimization of RL. We show OPQE outperforms both standalone prompting and RL-based rewriting, demonstrating that a synergistic approach yields the best results. Our implementation is made available to facilitate reproducibility.

Method: Trace intermediate representations during inference across several open-weight models, propose “Guess-then-Refine” framework, analyze through three case studies (multiple-choice tasks, fact recall, part-of-speech), and validate with causal manipulations including activation patching and early-exiting experiments.

Result: Early layers produce high-frequency token guesses due to limited context; later layers refine these into contextually appropriate tokens. Multiple-choice tasks: options identified in first half, responses finalized in second half. Fact recall: first token requires more depth than subsequent tokens. Function words predicted earliest on average.

Conclusion: LLMs employ structured depth usage with early statistical guessing and later contextual refinement, providing insights for improving computational efficiency in transformer models through better understanding of layer-by-layer computations.

Abstract: Growing evidence suggests that large language models do not use their depth uniformly, yet we still lack a fine-grained understanding of their layer-wise prediction dynamics. In this paper, we trace the intermediate representations of several open-weight models during inference and reveal a structured and nuanced use of depth. Specifically, we propose a “Guess-then-Refine” framework that explains how LLMs internally structure their computations to make predictions. We first show that the top-ranked predictions in early LLM layers are composed primarily of high-frequency tokens, which act as statistical guesses proposed by the model due to the lack of contextual information. As contextual information develops deeper into the model, these initial guesses get refined into contextually appropriate tokens. We then examine the dynamic usage of layer depth through three case studies. (i) Multiple-choice task analysis shows that the model identifies appropriate options within the first half of the model and finalizes the response in the latter half. (ii) Fact recall task analysis shows that in a multi-token answer, the first token requires more computational depth than the rest. (iii) Part-of-speech analysis shows that function words are, on average, the earliest to be predicted correctly. To validate our results, we supplement probe-based analyses with causal manipulations in the form of activation patching and early-exiting experiments. Together, our results provide a detailed view of depth usage in LLMs, shedding light on the layer-by-layer computations that underlie successful predictions and providing insights for future works to improve computational efficiency in transformer-based models.

Method: DiSRouter shifts from centralized to distributed routing where queries traverse a network of LLM agents. Each agent independently decides to answer or route based on self-awareness of its competence. Uses a two-stage Self-Awareness Training pipeline to enhance each LLM’s self-assessment capabilities.

Result: DiSRouter significantly outperforms existing routing methods in utility across various scenarios, effectively distinguishes between easy and hard queries, and shows strong generalization to out-of-domain tasks.

Conclusion: Leveraging LLMs’ intrinsic self-awareness is more effective than external assessment for query routing. This distributed approach enables more modular and efficient multi-agent systems with superior flexibility, scalability, and generalizability.

Abstract: The proliferation of Large Language Models (LLMs) has created a diverse ecosystem of models with highly varying performance and costs, necessitating effective query routing to balance performance and expense. Current routing systems often rely on a centralized external router trained on a fixed set of LLMs, making them inflexible and prone to poor performance since the small router can not fully understand the knowledge boundaries of different LLMs. We introduce DiSRouter (Distributed Self-Router), a novel paradigm that shifts from centralized control to distributed routing. In DiSRouter, a query traverses a network of LLM agents, each independently deciding whether to answer or route to other agents based on its own self-awareness, its ability to judge its competence. This distributed design offers superior flexibility, scalability, and generalizability. To enable this, we propose a two-stage Self-Awareness Training pipeline that enhances each LLM’s self-awareness. Extensive experiments demonstrate that DiSRouter significantly outperforms existing routing methods in utility across various scenarios, effectively distinguishes between easy and hard queries, and shows strong generalization to out-of-domain tasks. Our work validates that leveraging an LLM’s intrinsic self-awareness is more effective than external assessment, paving the way for more modular and efficient multi-agent systems.

Method: Proposes ToolDreamer framework that conditions retriever models to fetch tools based on hypothetical (synthetic) tool descriptions generated by an LLM. These synthetic descriptions represent tools the LLM feels would be potentially useful for the query, creating better alignment within the TD language space.

Result: Applied ToolDreamer on ToolRet dataset and showed improved performance for both sparse and dense retrievers with and without training, demonstrating flexibility. The framework effectively offloads reasoning burden to the retriever.

Conclusion: ToolDreamer enables more natural alignment between queries and tools within the TD language space, allowing LLMs to effectively handle large tool collections without overwhelming context windows by improving tool retrieval through synthetic description generation.

Abstract: Tool calling has become increasingly popular for Large Language Models (LLMs). However, for large tool sets, the resulting tokens would exceed the LLM’s context window limit, making it impossible to include every tool. Hence, an external retriever is used to provide LLMs with the most relevant tools for a query. Existing retrieval models rank tools based on the similarity between a user query and a tool description (TD). This leads to suboptimal retrieval as user requests are often poorly aligned with the language of TD. To remedy the issue, we propose ToolDreamer, a framework to condition retriever models to fetch tools based on hypothetical (synthetic) TD generated using an LLM, i.e., description of tools that the LLM feels will be potentially useful for the query. The framework enables a more natural alignment between queries and tools within the language space of TD’s. We apply ToolDreamer on the ToolRet dataset and show that our method improves the performance of sparse and dense retrievers with and without training, thus showcasing its flexibility. Through our proposed framework, our aim is to offload a portion of the reasoning burden to the retriever so that the LLM may effectively handle a large collection of tools without inundating its context window.

Method: Scaf-GRPO (Scaffolded Group Relative Policy Optimization) is a progressive training framework that diagnoses learning stagnation and then intervenes with tiered in-prompt hints ranging from abstract concepts to concrete steps, enabling models to construct valid solutions independently.

Result: Extensive experiments on challenging mathematics benchmarks show Scaf-GRPO boosts the pass@1 score of Qwen2.5-Math-7B on the AIME24 benchmark by 44.3% relative to vanilla GRPO baseline.

Conclusion: The framework provides a robust methodology for unlocking LLMs’ ability to solve problems previously beyond their reach, representing a critical step toward extending autonomous reasoning capabilities.

Abstract: Reinforcement learning from verifiable rewards has emerged as a powerful technique for enhancing the complex reasoning abilities of Large Language Models (LLMs). However, these methods are fundamentally constrained by the ‘’learning cliff’’ phenomenon: when faced with problems far beyond their current capabilities, models consistently fail, yielding a persistent zero-reward signal. In policy optimization algorithms like GRPO, this collapses the advantage calculation to zero, rendering these difficult problems invisible to the learning gradient and stalling progress. To overcome this, we introduce Scaf-GRPO (Scaffolded Group Relative Policy Optimization), a progressive training framework that strategically provides minimal guidance only when a model’s independent learning has plateaued. The framework first diagnoses learning stagnation and then intervenes by injecting tiered in-prompt hints, ranging from abstract concepts to concrete steps, enabling the model to construct a valid solution by itself. Extensive experiments on challenging mathematics benchmarks demonstrate Scaf-GRPO’s effectiveness, boosting the pass@1 score of the Qwen2.5-Math-7B model on the AIME24 benchmark by a relative 44.3% over a vanilla GRPO baseline. This result demonstrates our framework provides a robust and effective methodology for unlocking a model’s ability to solve problems previously beyond its reach, a critical step towards extending the frontier of autonomous reasoning in LLM.

Method: Two-step training: 1) Continued pretraining on documents describing evaluation-aware behavior (Python type hints in evaluation, recognition of evaluation cues), 2) Expert iteration to use type hints in evaluation settings. Then use activation steering with vectors from original model to suppress evaluation-awareness.

Result: Activation steering successfully suppressed evaluation-awareness, making the model behave during evaluation as it would during deployment. The steering vector constructed from the original model before additional training was effective.

Conclusion: AI evaluators could improve safety evaluation reliability by steering models to act like they’re deployed, addressing the problem of evaluation-aware behavior in LLMs.

Abstract: Large language models (LLMs) can sometimes detect when they are being evaluated and adjust their behavior to appear more aligned, compromising the reliability of safety evaluations. In this paper, we show that adding a steering vector to an LLM’s activations can suppress evaluation-awareness and make the model act like it is deployed during evaluation. To study our steering technique, we train an LLM to exhibit evaluation-aware behavior using a two-step training process designed to mimic how this behavior could emerge naturally. First, we perform continued pretraining on two sets of documents describing its behavior. The first says that our model uses Python type hints during evaluation but not during deployment. The second says that our model can recognize that the presence of a certain evaluation cue always means that it is being tested. Then, we train the model with expert iteration to use Python type hints in evaluation settings. The resulting model is evaluation-aware: it writes type hints in evaluation contexts more than deployment contexts. We find that activation steering can suppress evaluation awareness and make the model behave during evaluation as it would during deployment. Importantly, we constructed our steering vector using the original model before our additional training. Our results suggest that AI evaluators could improve the reliability of safety evaluations by steering models to act like they are deployed.

Method: Created a manually written question-answering benchmark focused on Sweden-related content, drawing inspiration from Swedish radio programs featuring public figures and major sports events. Includes English translations for cross-lingual analysis.

Result: Smaller models with stronger Swedish coverage perform comparably to a multilingual model three times larger on Sweden-related facts. Continued pre-training on Swedish improves factual knowledge but causes partial forgetting of previously known information.

Conclusion: The dataset serves as a diagnostic tool for studying language adaptation and knowledge retention in multilingual models during language adaptation.

Abstract: Many Swedish benchmarks are translations of US-centric benchmarks and are therefore not suitable for testing knowledge that is particularly relevant, or even specific, to Sweden. We therefore introduce a manually written question-answering benchmark specifically targeted at Sweden-related personalities and events, many of which receive very limited coverage in international media. Our annotators drew inspiration from a popular radio program featuring public figures from culture and media, as well as major sports events in Sweden. The dataset can be used to measure factual recall across models of varying sizes and degrees of Swedish coverage, and allows probing of cross-lingual factual consistency, as it contains English translations. Using the dataset, we find that smaller models with stronger Swedish coverage perform comparably to a multilingual model three times larger in recalling Sweden-related facts. We also observe that continued pre-training on Swedish generally improves factual knowledge but leads to partial forgetting of previously known information. These results demonstrate the dataset’s potential as a diagnostic tool for studying language adaptation and knowledge retention in multilingual models during language adaptation.

Method: Created VisJudge-Bench with 3,090 expert-annotated samples from real-world scenarios covering single visualizations, multiple visualizations, and dashboards across 32 chart types. Proposed VisJudge, a model specifically designed for visualization aesthetics and quality assessment.

Result: Advanced MLLMs like GPT-5 show significant gaps compared to human experts (MAE: 0.553, correlation: 0.428). VisJudge reduces MAE to 0.421 (23.9% reduction) and increases consistency with human experts to 0.687 (60.5% improvement) compared to GPT-5.

Conclusion: Current MLLMs have limitations in visualization quality assessment, but specialized models like VisJudge can significantly narrow the gap with human judgment. The benchmark enables systematic evaluation of MLLMs in this important domain.

Abstract: Visualization, a domain-specific yet widely used form of imagery, is an effective way to turn complex datasets into intuitive insights, and its value depends on whether data are faithfully represented, clearly communicated, and aesthetically designed. However, evaluating visualization quality is challenging: unlike natural images, it requires simultaneous judgment across data encoding accuracy, information expressiveness, and visual aesthetics. Although multimodal large language models (MLLMs) have shown promising performance in aesthetic assessment of natural images, no systematic benchmark exists for measuring their capabilities in evaluating visualizations. To address this, we propose VisJudge-Bench, the first comprehensive benchmark for evaluating MLLMs’ performance in assessing visualization aesthetics and quality. It contains 3,090 expert-annotated samples from real-world scenarios, covering single visualizations, multiple visualizations, and dashboards across 32 chart types. Systematic testing on this benchmark reveals that even the most advanced MLLMs (such as GPT-5) still exhibit significant gaps compared to human experts in judgment, with a Mean Absolute Error (MAE) of 0.553 and a correlation with human ratings of only 0.428. To address this issue, we propose VisJudge, a model specifically designed for visualization aesthetics and quality assessment. Experimental results demonstrate that VisJudge significantly narrows the gap with human judgment, reducing the MAE to 0.421 (a 23.9% reduction) and increasing the consistency with human experts to 0.687 (a 60.5% improvement) compared to GPT-5. The benchmark is available at https://github.com/HKUSTDial/VisJudgeBench.

Method: Created a large-scale dataset using original New York Times article abstracts as prompts to generate synthetic versions with multiple LLMs (Gemma-2-9b, Mistral-7B, Qwen-2-72B, LLaMA-8B, Yi-Large, GPT-4-o), providing both human-authored and AI-generated text samples.

Result: Baseline results show 58.35% accuracy for distinguishing human-written from AI-generated text, and 8.92% accuracy for attributing AI texts to their generating models, indicating the difficulty of these tasks.

Conclusion: The dataset bridges real-world journalistic content with modern generative models to catalyze development of robust detection and attribution methods, fostering trust and transparency in generative AI.

Abstract: The rapid advancement of large language models (LLMs) has led to increasingly human-like AI-generated text, raising concerns about content authenticity, misinformation, and trustworthiness. Addressing the challenge of reliably detecting AI-generated text and attributing it to specific models requires large-scale, diverse, and well-annotated datasets. In this work, we present a comprehensive dataset comprising over 58,000 text samples that combine authentic New York Times articles with synthetic versions generated by multiple state-of-the-art LLMs including Gemma-2-9b, Mistral-7B, Qwen-2-72B, LLaMA-8B, Yi-Large, and GPT-4-o. The dataset provides original article abstracts as prompts, full human-authored narratives. We establish baseline results for two key tasks: distinguishing human-written from AI-generated text, achieving an accuracy of 58.35%, and attributing AI texts to their generating models with an accuracy of 8.92%. By bridging real-world journalistic content with modern generative models, the dataset aims to catalyze the development of robust detection and attribution methods, fostering trust and transparency in the era of generative AI. Our dataset is available at: https://huggingface.co/datasets/gsingh1-py/train.

Method: Proposes Lookahead Tree-Based Rollouts (LATR) with three iterative stages: (1) branching at high-uncertainty generation steps, (2) performing lookahead simulation for each branch, and (3) pruning branches that show prolonged similarity during simulation.

Result: LATR accelerates policy learning by 131% on average and improves final pass@1 performance by 4.2% on both GRPO and DAPO algorithms across different reasoning tasks.

Conclusion: LATR effectively addresses the trajectory diversity bottleneck in RLVR pipelines, significantly improving learning efficiency and reasoning performance for large language models.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR), particularly with algorithms like Group Relative Policy Optimization (GRPO), has proven highly effective in enhancing the reasoning capabilities of large language models. However, a critical bottleneck in current pipelines lies in the limited diversity of sampled trajectories during group rollouts. Homogeneous trajectories and their associated rewards would diminish the return signals for policy updates, thereby hindering effective policy learning. This lack of diversity stems primarily from token-level stochastic sampling, where local variations are likely to collapse into near-identical reasoning paths. To address this limitation, we propose Lookahead Tree-Based Rollouts (LATR), a novel rollout strategy designed to explicitly promotes trajectory-level diversity by enforcing branching into different candidate tokens likely to yield distinct continuations. Specifically, LATR iteratively operates in three stages: (1) branching at high-uncertainty generation steps, (2) performing lookahead simulation for each new branch, and (3) pruning branches that exhibits prolonged similarity during simulation. Compared with stochastic Sampling, LATR accelerates policy learning by 131% on average and improves final pass@1 performance by 4.2% on both GRPO and Dynamic sAmpling Policy Optimization (DAPO) algorithms across different reasoning tasks. Our code and data are publicly available at https://github.com/starreeze/latr.

Method: Self-Harmony uses a single model in two roles: Solver (produces answers) and Reframer (rephrases inputs). It aggregates answer frequencies across original and reframed views using harmonic mean instead of majority voting, selecting solutions stable under reframing to avoid view-dependent spurious answers.

Result: Achieves state-of-the-art results in label-free test-time settings, ranking first in 28 of 30 settings across multiple reasoning benchmarks. Demonstrates unprecedented robustness with zero training failures in all experiments.

Conclusion: Self-Harmony provides a stable and reliable framework for test-time reinforcement learning that avoids common pitfalls of majority voting by leveraging answer stability across paraphrased inputs, requiring no human supervision or auxiliary models.

Abstract: Test-time reinforcement learning (TTRL) offers a label-free paradigm for adapting models using only synthetic signals at inference, but its success hinges on constructing reliable learning signals. Standard approaches such as majority voting often collapse to spurious yet popular answers. We introduce Self-Harmony, a framework built on a simple intuition: the correct answer should remain stable across both an original question and its paraphrase. Self-Harmony operationalizes this by employing a single model in two complementary roles: a Solver to produce answers and a Reframer to rephrase the input. Based on this, we further propose a pseudo-label method: instead of majority voting, it aggregates answer frequencies across these original and reframed views using the harmonic mean. This is a process that naturally selects for solutions stable under reframing, thereby avoiding the common trap of favoring view-dependent, spurious answers. Crucially, this requires no human supervision or auxiliary models. Across diverse reasoning benchmarks, Self-Harmony achieves state-of-the-art results at the label-free test-time setting, ranking first in 28 of 30 settings across multiple methods. Beyond accuracy, it demonstrates unprecedented robustness, with zero training failures in all experiments, underscoring its stability and reliability.

Method: Used Magpie framework to annotate samples for task category, input quality, and preference reward; conducted fine-grained inspection of preference quality; curated UltraMix by selectively drawing from five corpora while removing noisy/redundant samples.

Result: UltraMix is 30% smaller than best-performing individual dataset yet exceeds its performance across key benchmarks; revealed structural and qualitative discrepancies in reward margins across datasets.

Conclusion: Data-centric analysis enables better understanding of DPO datasets; curated UltraMix provides improved performance with reduced size; annotations and mixture released to facilitate future research.

Abstract: Aligning large language models (LLMs) is a central objective of post-training, often achieved through reward modeling and reinforcement learning methods. Among these, direct preference optimization (DPO) has emerged as a widely adopted technique that fine-tunes LLMs on preferred completions over less favorable ones. While most frontier LLMs do not disclose their curated preference pairs, the broader LLM community has released several open-source DPO datasets, including TuluDPO, ORPO, UltraFeedback, HelpSteer, and Code-Preference-Pairs. However, systematic comparisons remain scarce, largely due to the high computational cost and the lack of rich quality annotations, making it difficult to understand how preferences were selected, which task types they span, and how well they reflect human judgment on a per-sample level. In this work, we present the first comprehensive, data-centric analysis of popular open-source DPO corpora. We leverage the Magpie framework to annotate each sample for task category, input quality, and preference reward, a reward-model-based signal that validates the preference order without relying on human annotations. This enables a scalable, fine-grained inspection of preference quality across datasets, revealing structural and qualitative discrepancies in reward margins. Building on these insights, we systematically curate a new DPO mixture, UltraMix, that draws selectively from all five corpora while removing noisy or redundant samples. UltraMix is 30% smaller than the best-performing individual dataset yet exceeds its performance across key benchmarks. We publicly release all annotations, metadata, and our curated mixture to facilitate future research in data-centric preference optimization.

Method: Examine LLM fairness in emotional theory of mind by presenting descriptions of people and environments, asking “How does this person feel?”, and evaluating gender biases. Propose and compare multiple debiasing strategies including training-based interventions vs inference-time prompt engineering

Result: LLMs demonstrate gender biases in emotional theory of mind tasks. Training-based interventions achieve meaningful bias reduction, while inference-time prompt engineering approaches are insufficient

Conclusion: Effective debiasing for emotional theory of mind in LLMs requires training-based approaches rather than relying solely on prompt engineering, highlighting the need for systematic fairness interventions

Abstract: The rapid advancement of large language models (LLMs) and their growing integration into daily life underscore the importance of evaluating and ensuring their fairness. In this work, we examine fairness within the domain of emotional theory of mind, investigating whether LLMs exhibit gender biases when presented with a description of a person and their environment and asked, ‘‘How does this person feel?’’. Furthermore, we propose and evaluate several debiasing strategies, demonstrating that achieving meaningful reductions in bias requires training based interventions rather than relying solely on inference-time prompt-based approaches such as prompt engineering, etc.

Method: Created GGSS Personas collection from German General Social Survey (ALLBUS) data, designed as plug-and-play persona prompts for LLMs. Evaluated by prompting various LLMs to simulate survey response distributions across topics.

Result: GGSS Personas-guided LLMs outperform state-of-the-art classifiers, especially under data scarcity. Representativity and attribute selection within persona prompts significantly affect alignment with population responses.

Conclusion: GGSS Personas provide a valuable resource for systematic exploration of population-aligned persona prompting in NLP and social science research, enabling more accurate LLM-based social simulations.

Abstract: The use of Large Language Models (LLMs) for simulating human perspectives via persona prompting is gaining traction in computational social science. However, well-curated, empirically grounded persona collections remain scarce, limiting the accuracy and representativeness of such simulations. Here, we introduce the German General Social Survey Personas (GGSS Personas) collection, a comprehensive and representative persona prompt collection built from the German General Social Survey (ALLBUS). The GGSS Personas and their persona prompts are designed to be easily plugged into prompts for all types of LLMs and tasks, steering models to generate responses aligned with the underlying German population. We evaluate GGSS Personas by prompting various LLMs to simulate survey response distributions across diverse topics, demonstrating that GGSS Personas-guided LLMs outperform state-of-the-art classifiers, particularly under data scarcity. Furthermore, we analyze how the representativity and attribute selection within persona prompts affect alignment with population responses. Our findings suggest that GGSS Personas provide a potentially valuable resource for research on LLM-based social simulations that enables more systematic explorations of population-aligned persona prompting in NLP and social science research.

Method: Introduces BIT framework with cross-task, cross-species pretrained neural encoder that transfers to both attempted and imagined speech. Integrated end-to-end with audio large language models (LLMs) and trained with contrastive learning for cross-modal alignment. Uses small-scale audio LLMs to improve end-to-end decoding.

Result: Achieves new SOTA on Brain-to-Text ‘24 and ‘25 benchmarks in cascaded setting. Reduces WER from 24.69% to 10.22% compared to prior end-to-end method. Small-scale audio LLMs markedly improve end-to-end decoding. Enables cross-task generalization by aligning attempted and imagined speech embeddings.

Conclusion: BIT advances integration of large, diverse neural datasets and paves the way for end-to-end decoding framework supporting seamless, differentiable optimization. Shows potential for improved speech BCIs through neural encoder pretraining and audio LLM integration.

Abstract: Speech brain-computer interfaces (BCIs) aim to restore communication for people with paralysis by translating neural activity into text. Most systems use cascaded frameworks that decode phonemes before assembling sentences with an n-gram language model (LM), preventing joint optimization of all stages simultaneously. Here, we introduce an end-to-end Brain-to-Text (BIT) framework that translates neural activity into coherent sentences using a single differentiable neural network. Central to our approach is a cross-task, cross-species pretrained neural encoder, whose representations transfer to both attempted and imagined speech. In a cascaded setting with an n-gram LM, the pretrained encoder establishes a new state-of-the-art (SOTA) on the Brain-to-Text ‘24 and ‘25 benchmarks. Integrated end-to-end with audio large language models (LLMs) and trained with contrastive learning for cross-modal alignment, BIT reduces the word error rate (WER) of the prior end-to-end method from 24.69% to 10.22%. Notably, we find that small-scale audio LLMs markedly improve end-to-end decoding. Beyond record-setting performance, BIT aligns attempted and imagined speech embeddings to enable cross-task generalization. Altogether, our approach advances the integration of large, diverse neural datasets, paving the way for an end-to-end decoding framework that supports seamless, differentiable optimization.

Method: Proposes Latent-GRPO framework that derives intrinsic rewards from latent space geometry. Uses empirical discovery that correct reasoning trajectories form dense clusters while incorrect ones scatter as outliers. Introduces Iterative Robust Centroid Estimation (IRCE) algorithm with spherical projection to mitigate magnitude fluctuations and estimate robust “truth centroid” through iterative aggregation.

Result: Experimental results show method maintains model performance while achieving over 2x training speedup compared to baselines. Demonstrates strong generalization ability and robustness across multiple datasets.

Conclusion: Latent-GRPO provides an efficient alternative to external verifiers for LLM reasoning optimization by leveraging latent space geometry, significantly reducing computational costs while preserving performance.

Abstract: Group Relative Policy Optimization (GRPO) significantly enhances the reasoning performance of Large Language Models (LLMs). However, this success heavily relies on expensive external verifiers or human rules. Such dependency not only leads to significant computational costs and training latency, but also yields sparse rewards that hinder optimization efficiency. To address these challenges, we propose Latent-GRPO, a framework that derives intrinsic rewards directly from latent space geometry. Crucially, our empirical analysis reveals a compelling geometric property: terminal token representations of correct reasoning trajectories form dense clusters with high intra-class similarity, whereas incorrect trajectories remain scattered as outliers. In light of this discovery, we introduce the Iterative Robust Centroid Estimation (IRCE) algorithm, which generates dense, continuous rewards by mitigating magnitude fluctuations via spherical projection and estimating a robust ``truth centroid’’ through iterative aggregation. Experimental results on multiple datasets show that our method maintains model performance while achieving a training speedup of over 2x compared to baselines. Furthermore, extensive results demonstrate strong generalization ability and robustness. The code will be released soon.

Method: T* uses a TraceRL-based training curriculum that starts with an auto-regressively initialized small-block MDM and progressively scales to larger blocks. The method employs reinforcement learning techniques to create a smooth transition between different block sizes, enabling the model to adapt to higher-parallelism decoding while minimizing performance loss.

Result: The method achieves higher-parallelism decoding with minimal performance degradation on math reasoning benchmarks. Analysis suggests T* may converge to an alternative decoding schedule that achieves comparable performance to traditional approaches.

Conclusion: T* provides an effective training curriculum for scaling masked diffusion language models to larger block sizes, enabling practical high-parallelism decoding while maintaining reasoning performance, with potential implications for more efficient language model architectures.

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

Method: A cascade system that samples eight parallel votes from Gemini 2.5 Flash, applies supermajority thresholds to resolve confident cases immediately, escalates uncertain cases to additional voting rounds, and defers perfect ties to a symbolic ensemble of five narrative signals grounded in classical narrative theory.

Result: The system achieved 11th place out of 47 teams in SemEval-2026 Task 4 with 72.75% test accuracy, outperforming several systems built on larger and more expensive models. The difficulty gradient (85% -> 67% -> 61% by pathway) confirmed that vote consensus tracks genuine ambiguity.

Conclusion: For narrative similarity tasks, knowing when you don’t know (confidence-aware routing) matters more than adding auxiliary representations, and LLM vote distribution is a reliable proxy for task difficulty.

Abstract: How should a system handle uncertainty when comparing narratives? We present CascadeMind, a hybrid neuro-symbolic system for SemEval-2026 Task 4 (Narrative Story Similarity) built around a core finding: an LLM’s internal vote distribution is a reliable proxy for task difficulty, and confidence-aware routing outperforms uniform treatment of all cases. Our cascade samples eight parallel votes from Gemini 2.5 Flash, applying a supermajority threshold to resolve confident cases immediately (74% of instances at 85% development accuracy). Uncertain cases escalate to additional voting rounds (21%), and only perfect ties (5%) are deferred to a symbolic ensemble of five narrative signals grounded in classical narrative theory. The resulting difficulty gradient (85% -> 67% -> 61% by pathway) confirms that vote consensus tracks genuine ambiguity. In official Track A evaluation, CascadeMind placed 11th of 47 teams with 72.75% test accuracy (Hatzel et al., 2026), outperforming several systems built on larger and more expensive models. Gains are driven primarily by routing strategy rather than symbolic reasoning, suggesting that for narrative similarity, knowing when you don’t know matters more than adding auxiliary representations.

Method: EnsembleLink leverages pre-trained language models that have learned semantic relationships from large text corpora. It uses these models to match records without any training labels, running locally on open-source models without external API calls.

Result: On benchmarks spanning city names, person names, organizations, multilingual political parties, and bibliographic records, EnsembleLink matches or exceeds methods requiring extensive labeling. It completes typical linkage tasks in minutes.

Conclusion: EnsembleLink provides an accurate, practical solution for record linkage that doesn’t require labeled training data, addressing a key methodological gap in empirical social science research.

Abstract: Record linkage, the process of matching records that refer to the same entity across datasets, is essential to empirical social science but remains methodologically underdeveloped. Researchers treat it as a preprocessing step, applying ad hoc rules without quantifying the uncertainty that linkage errors introduce into downstream analyses. Existing methods either achieve low accuracy or require substantial labeled training data. I present EnsembleLink, a method that achieves high accuracy without any training labels. EnsembleLink leverages pre-trained language models that have learned semantic relationships (e.g., that “South Ozone Park” is a neighborhood in “New York City” or that “Lutte ouvriere” refers to the Trotskyist “Workers’ Struggle” party) from large text corpora. On benchmarks spanning city names, person names, organizations, multilingual political parties, and bibliographic records, EnsembleLink matches or exceeds methods requiring extensive labeling. The method runs locally on open-source models, requiring no external API calls, and completes typical linkage tasks in minutes.

Method: Presents geometric approach to understand rewrite-based detection, then introduces novel algorithm that adaptively learns distance between original and rewritten text rather than using fixed distance.

Result: Extensive experiments with 100+ settings show superior performance over baselines, achieving 54.3% to 75.4% relative improvements over strongest baseline across different target LLMs (GPT, Claude, Gemini).

Conclusion: Adaptively learned distance functions are more effective for LLM-generated content detection than fixed distances, with publicly available implementation.

Abstract: Modern large language models (LLMs) such as GPT, Claude, and Gemini have transformed the way we learn, work, and communicate. Yet, their ability to produce highly human-like text raises serious concerns about misinformation and academic integrity, making it an urgent need for reliable algorithms to detect LLM-generated content. In this paper, we start by presenting a geometric approach to demystify rewrite-based detection algorithms, revealing their underlying rationale and demonstrating their generalization ability. Building on this insight, we introduce a novel rewrite-based detection algorithm that adaptively learns the distance between the original and rewritten text. Theoretically, we demonstrate that employing an adaptively learned distance function is more effective for detection than using a fixed distance. Empirically, we conduct extensive experiments with over 100 settings, and find that our approach demonstrates superior performance over baseline algorithms in the majority of scenarios. In particular, it achieves relative improvements from 54.3% to 75.4% over the strongest baseline across different target LLMs (e.g., GPT, Claude, and Gemini). A python implementation of our proposal is publicly available at https://github.com/Mamba413/L2D.

Method: Proposes GPT-based architecture with selective fine-tuning strategy: freezes majority of GPT-2 backbone, trains only final Transformer block, final layer normalization, and lightweight classification head, reducing trainable parameters while preserving representational capacity.

Result: Evaluated on MIMIC-IV-Note radiology reports with CheXpert-style labels; shows stable convergence and strong classification performance across multiple problem formulations, especially for non-mention and negated findings, with varying dataset sizes.

Conclusion: Selective fine-tuning of pretrained generative language models provides efficient and effective pathway for clinical text classification, enabling scalable adaptation to EHR data while significantly reducing computational complexity.

Abstract: The increasing availability of unstructured clinical narratives in electronic health records (EHRs) has created new opportunities for automated disease characterization, cohort identification, and clinical decision support. However, modeling long, domain-specific clinical text remains challenging due to limited labeled data, severe class imbalance, and the high computational cost of adapting large pretrained language models. This study presents a GPT-based architecture for clinical text classification that adapts a pretrained decoder-only Transformer using a selective fine-tuning strategy. Rather than updating all model parameters, the majority of the GPT-2 backbone is frozen, and training is restricted to the final Transformer block, the final layer normalization, and a lightweight classification head. This approach substantially reduces the number of trainable parameters while preserving the representational capacity required to model complex clinical language. The proposed method is evaluated on radiology reports from the MIMIC-IV-Note dataset using uncertainty-aware CheXpert-style labels derived directly from report text. Experiments cover multiple problem formulations, including multi-label classification of radiographic findings, binary per-label classification under different uncertainty assumptions, and aggregate disease outcome prediction. Across varying dataset sizes, the model exhibits stable convergence behavior and strong classification performance, particularly in settings dominated by non-mention and negated findings. Overall, the results indicate that selective fine-tuning of pretrained generative language models provides an efficient and effective pathway for clinical text classification, enabling scalable adaptation to real-world EHR data while significantly reducing computational complexity.

Method: Proposes MANBENCH benchmark with four task types susceptible to Mandela effect and five interaction protocols varying agent roles and memory timescales. Evaluates multiple LLMs on this benchmark, then develops mitigation strategies including prompt-level defenses (cognitive anchoring, source scrutiny) and model-level alignment-based defense.

Result: The proposed mitigation strategies achieve an average 74.40% reduction in Mandela effect compared to baseline. The benchmark successfully quantifies the effect and analyzes how different factors influence it.

Conclusion: The study provides insights for developing more resilient and ethically aligned collaborative multi-agent systems by addressing collective memory biases, with implications for reducing misinformation spread in AI systems.

Abstract: Recent advancements in large language models (LLMs) have significantly enhanced the capabilities of collaborative multi-agent systems, enabling them to address complex challenges. However, within these multi-agent systems, the susceptibility of agents to collective cognitive biases remains an underexplored issue. A compelling example is the Mandela effect, a phenomenon where groups collectively misremember past events as a result of false details reinforced through social influence and internalized misinformation. This vulnerability limits our understanding of memory bias in multi-agent systems and raises ethical concerns about the potential spread of misinformation. In this paper, we conduct a comprehensive study on the Mandela effect in LLM-based multi-agent systems, focusing on its existence, causing factors, and mitigation strategies. We propose MANBENCH, a novel benchmark designed to evaluate agent behaviors across four common task types that are susceptible to the Mandela effect, using five interaction protocols that vary in agent roles and memory timescales. We evaluate agents powered by several LLMs on MANBENCH to quantify the Mandela effect and analyze how different factors affect it. Moreover, we propose strategies to mitigate this effect, including prompt-level defenses (e.g., cognitive anchoring and source scrutiny) and model-level alignment-based defense, achieving an average 74.40% reduction in the Mandela effect compared to the baseline. Our findings provide valuable insights for developing more resilient and ethically aligned collaborative multi-agent systems. Code and dataset are available at https://github.com/bluedream02/Mandela-Effect.

Method: OOMB uses a chunk-recurrent training framework with on-the-fly activation recomputation to maintain constant activation memory. It manages KV cache growth through: 1) paged memory manager for KV cache and gradients to eliminate fragmentation, 2) asynchronous CPU offloading to hide transfer latency, and 3) page-level sparse attention to reduce computation and communication overhead.

Result: OOMB achieves exceptional memory efficiency: only 10MB additional memory per 10K tokens for Qwen2.5-7B. This enables training Qwen2.5-7B with 4M-token context on a single H200 GPU, which would otherwise require a large cluster.

Conclusion: OOMB represents a substantial advance in resource efficiency for long-context LLM training, making it feasible to train models on extremely long sequences with minimal hardware requirements.

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

Method: FASA leverages functional sparsity in RoPE frequency-chunks to identify dominant frequency-chunks that correlate with full attention. It uses these as a free proxy to predict token importance, then performs attention computation only on the pruned critical token subset.

Result: FASA outperforms all token-eviction baselines across long-context tasks (sequence modeling, CoT reasoning), achieving near-oracle accuracy. On LongBench-V1, it reaches ~100% of full-KV performance with only 256 tokens, and achieves 2.56× speedup using 18.9% of cache on AIME24.

Conclusion: FASA provides an effective query-aware token eviction framework that addresses KV cache bottlenecks through RoPE frequency-chunk sparsity analysis, enabling efficient LLM inference for long contexts with minimal performance degradation.

Abstract: The deployment of Large Language Models (LLMs) faces a critical bottleneck when handling lengthy inputs: the prohibitive memory footprint of the Key Value (KV) cache. To address this bottleneck, the token pruning paradigm leverages attention sparsity to selectively retain a small, critical subset of tokens. However, existing approaches fall short, with static methods risking irreversible information loss and dynamic strategies employing heuristics that insufficiently capture the query-dependent nature of token importance. We propose FASA, a novel framework that achieves query-aware token eviction by dynamically predicting token importance. FASA stems from a novel insight into RoPE: the discovery of functional sparsity at the frequency-chunk (FC) level. Our key finding is that a small, identifiable subset of “dominant” FCs consistently exhibits high contextual agreement with the full attention head. This provides a robust and computationally free proxy for identifying salient tokens. Building on this insight, FASA first identifies a critical set of tokens using dominant FCs, and then performs focused attention computation solely on this pruned subset. Across a spectrum of long-context tasks, from sequence modeling to complex CoT reasoning, FASA consistently outperforms all token-eviction baselines and achieves near-oracle accuracy, demonstrating remarkable robustness even under constraint budgets. Notably, on LongBench-V1, FASA reaches nearly 100% of full-KV performance when only keeping 256 tokens, and achieves 2.56$\times$ speedup using just 18.9% of the cache on AIME24.

Method: Used structured prompt design to analyze outputs from Stable Diffusion XL and DALL-E 3. Compared image similarities between generic disability prompts and specific disability category prompts. Evaluated mitigation strategies’ influence on disability portrayals through sentiment polarity analysis combining automatic and human evaluation.

Result: Findings reveal persistent representational imbalances in how people with disabilities are portrayed in AI-generated images, highlighting issues with diversity and inclusion in current generative models.

Conclusion: Continuous evaluation and refinement of generative models is needed to foster more diverse and inclusive portrayals of disability, addressing representational gaps in AI-generated content.

Abstract: Text-to-image generative models have made remarkable progress in producing high-quality visual content from textual descriptions, yet concerns remain about how they represent social groups. While characteristics like gender and race have received increasing attention, disability representations remain underexplored. This study investigates how people with disabilities are represented in AI-generated images by analyzing outputs from Stable Diffusion XL and DALL-E 3 using a structured prompt design. We analyze disability representations by comparing image similarities between generic disability prompts and prompts referring to specific disability categories. Moreover, we evaluate how mitigation strategies influence disability portrayals, with a focus on assessing affective framing through sentiment polarity analysis, combining both automatic and human evaluation. Our findings reveal persistent representational imbalances and highlight the need for continuous evaluation and refinement of generative models to foster more diverse and inclusive portrayals of disability.

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

Result: Validated on RULER and LongBench v2 benchmarks. Achieves noticeable gains on RULER and reasonable improvement on LongBench v2 without manually curated long-context QA data. Includes ablation studies on reward design, data curation, training schemes, and scaling effects.

Conclusion: Unsupervised RL approach effectively enhances LLM long-context capabilities without expensive supervision, demonstrating practical value for long-context understanding tasks.

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

Method: Proposes Online Causal Kalman Filtering for Policy Optimization (KPO), which models the desired importance sampling ratio as a latent state evolving across tokens. A Kalman filter updates this state online and autoregressively based on past token states, preserving token-wise local structure-aware variation while smoothing noise spikes.

Result: KPO achieves superior results on challenging math reasoning datasets compared with state-of-the-art counterparts, demonstrating more stable and effective policy updates.

Conclusion: The proposed KPO method effectively addresses token-level off-policy deviation issues in RL for LLMs by using Kalman filtering to stabilize importance sampling ratios, leading to improved training stability and performance on reasoning tasks.

Abstract: Reinforcement learning for large language models suffers from high-variance token-level importance sampling (IS) ratios, which would destabilize policy optimization at scale. To improve stability, recent methods typically use a fixed sequence-level IS ratio for all tokens in a sequence or adjust each token’s IS ratio separately, thereby neglecting temporal off-policy derivation across tokens in a sequence. In this paper, we first empirically identify that local off-policy deviation is structurally inconsistent at the token level, which may distort policy-gradient updates across adjacent tokens and lead to training collapse. To address the issue, we propose Online Causal Kalman Filtering for stable and effective Policy Optimization (KPO). Concretely, we model the desired IS ratio as a latent state that evolves across tokens and apply a Kalman filter to update this state online and autoregressively based on the states of past tokens, regardless of future tokens. The resulting filtered IS ratios preserve token-wise local structure-aware variation while strongly smoothing noise spikes, yielding more stable and effective policy updates. Experimentally, KPO achieves superior results on challenging math reasoning datasets compared with state-of-the-art counterparts.

Method: Hybrid architecture integrating sparse attention (InfLLM-V2) for high-fidelity long-context modeling and linear attention (Lightning Attention) for global efficiency, using layer selection algorithm (1:3 ratio) and hybrid positional encoding (HyPE), with cost-effective continual training framework.

Result: Achieves up to 3.5x inference speed of full-attention model at 256K tokens, supports up to 1M token contexts (where 8B full-attention models fail), maintains general capabilities comparable to full-attention models, reduces training costs by ~75%.

Conclusion: MiniCPM-SALA demonstrates efficient ultra-long context processing through hybrid attention mechanisms, enabling practical applications with million-token contexts while preserving model performance.

Abstract: The evolution of large language models (LLMs) towards applications with ultra-long contexts faces challenges posed by the high computational and memory costs of the Transformer architecture. While existing sparse and linear attention mechanisms attempt to mitigate these issues, they typically involve a trade-off between memory efficiency and model performance. This paper introduces MiniCPM-SALA, a 9B-parameter hybrid architecture that integrates the high-fidelity long-context modeling of sparse attention (InfLLM-V2) with the global efficiency of linear attention (Lightning Attention). By employing a layer selection algorithm to integrate these mechanisms in a 1:3 ratio and utilizing a hybrid positional encoding (HyPE), the model maintains efficiency and performance for long-context tasks. Furthermore, we introduce a cost-effective continual training framework that transforms pre-trained Transformer-based models into hybrid models, which reduces training costs by approximately 75% compared to training from scratch. Extensive experiments show that MiniCPM-SALA maintains general capabilities comparable to full-attention models while offering improved efficiency. On a single NVIDIA A6000D GPU, the model achieves up to 3.5x the inference speed of the full-attention model at the sequence length of 256K tokens and supports context lengths of up to 1M tokens, a scale where traditional full-attention 8B models fail because of memory constraints.

Method: Developed HiFloat formats (HiF8 and HiF4) with hierarchical scaling for Ascend NPUs, conducted rigorous comparisons across weight-activation and KV-cache tasks, and evaluated compatibility with state-of-the-art post-training quantization frameworks.

Result: Three key insights: (1) INT8 works for narrow-range data but floating-point formats excel with high-variance data; (2) HiF4’s hierarchical scaling prevents accuracy collapse in 4-bit regimes unlike integer formats; (3) HiFloat is fully compatible with modern quantization frameworks.

Conclusion: HiFloat provides an effective solution for high-efficiency LLM inference on NPUs, offering better performance than integer formats for certain data distributions and preventing accuracy degradation in extreme low-bit scenarios.

Abstract: As LLMs scale, low-bit floating-point formats like MXFP and NVFP4 offer new opportunities for precision and efficiency. In this work, we evaluate HiFloat (HiF8 and HiF4), a family of formats tailored for Ascend NPUs. Through rigorous comparison across weight-activation and KV-cache tasks, we provide three key insights: (1) INT8 suits narrow-range data, while floating-point formats excel with high-variance data; (2) in 4-bit regimes, HiF4’s hierarchical scaling prevents the accuracy collapse seen in integer formats; and (3) HiFloat is fully compatible with state-of-the-art post-training quantization frameworks. Overall, HiFloat provides a solution for high-efficiency LLM inference on NPUs.

Method: Proposes Probabilistic Reward Model (PRM) treating reward as random variable with full probability distribution. Presents Ordinal Probabilistic Reward Model (OPRM) as discrete realization, discretizing quality scores into ordinal ratings. Introduces Region Flooding Tuning (RgFT) training strategy that uses quality-level annotations to concentrate probability mass within corresponding rating sub-regions.

Result: Experiments on various reward model benchmarks show 2.9% to 7.4% accuracy improvement over prior reward models, demonstrating strong performance and data efficiency. Analysis shows method captures both relative rankings and absolute quality.

Conclusion: PRM paradigm with OPRM and RgFT provides effective solution for reward modeling that addresses limitations of existing approaches, offering better accuracy, data efficiency, and ability to capture absolute quality while maintaining probabilistic interpretation.

Abstract: Reward models are crucial for aligning large language models (LLMs) with human values and intentions. Existing approaches follow either Generative (GRMs) or Discriminative (DRMs) paradigms, yet both suffer from limitations: GRMs typically demand costly point-wise supervision, while DRMs produce uncalibrated relative scores that lack probabilistic interpretation. To address these challenges, we introduce a novel reward modeling paradigm: Probabilistic Reward Model (PRM). Instead of modeling reward as a deterministic scalar, our approach treats it as a random variable, learning a full probability distribution for the quality of each response. To make this paradigm practical, we present its closed-form, discrete realization: the Ordinal Probabilistic Reward Model (OPRM), which discretizes the quality score into a finite set of ordinal ratings. Building on OPRM, we propose a data-efficient training strategy called Region Flooding Tuning (RgFT). It enables rewards to better reflect absolute text quality by incorporating quality-level annotations, which guide the model to concentrate the probability mass within corresponding rating sub-regions. Experiments on various reward model benchmarks show that our method improves accuracy by $\textbf{2.9%}\sim\textbf{7.4%}$ compared to prior reward models, demonstrating strong performance and data efficiency. Analysis of the score distribution provides evidence that our method captures not only relative rankings but also absolute quality.

Method: Introduces Elo-Evolve framework with two key innovations: 1) learning directly from binary win/loss outcomes in pairwise competitions (eliminating Bradley-Terry model dependencies), and 2) implementing Elo-orchestrated opponent selection with temperature-controlled sampling for automatic curriculum learning. Grounded in PAC learning theory.

Result: Achieves 4.5x noise reduction compared to absolute scoring approaches. Trains Qwen2.5-7B model using opponents including Qwen2.5-14B, Qwen2.5-32B, and Qwen3-8B models. Shows clear performance hierarchy: point-based methods < static pairwise training < Elo-Evolve across Alpaca Eval 2.0 and MT-Bench.

Conclusion: Pairwise comparison with dynamic opponent selection provides progressive benefits for LLM alignment, offering superior sample complexity and noise robustness compared to traditional absolute scoring methods.

Abstract: Current alignment methods for Large Language Models (LLMs) rely on compressing vast amounts of human preference data into static, absolute reward functions, leading to data scarcity, noise sensitivity, and training instability. We introduce Elo-Evolve, a co-evolutionary framework that redefines alignment as dynamic multi-agent competition within an adaptive opponent pool. Our approach makes two key innovations: (1) eliminating Bradley-Terry model dependencies by learning directly from binary win/loss outcomes in pairwise competitions, and (2) implementing Elo-orchestrated opponent selection that provides automatic curriculum learning through temperature-controlled sampling. We ground our approach in PAC learning theory, demonstrating that pairwise comparison achieves superior sample complexity and empirically validate a 4.5x noise reduction compared to absolute scoring approaches. Experimentally, we train a Qwen2.5-7B model using our framework with opponents including Qwen2.5-14B, Qwen2.5-32B, and Qwen3-8B models. Results demonstrate a clear performance hierarchy: point-based methods < static pairwise training < Elo-Evolve across Alpaca Eval 2.0 and MT-Bench, validating the progressive benefits of pairwise comparison and dynamic opponent selection for LLM alignment.

Method: Empirically localize input-output alignment shift using decoding trajectories over tied embedding spaces and similarity-based metrics; propose lightweight residual-path mitigation via residual attenuation, implemented as fixed-layer intervention or learnable gating mechanism

Result: Experiments reveal hidden token representations switch from input alignment to output alignment deep within the network; proposed strategies alleviate representation misalignment and yield improvements on multiple benchmarks

Conclusion: Residual attenuation provides efficient and general architectural enhancement for autoregressive Transformers by addressing the fundamental input-output alignment shift

Abstract: Large Language Models (LLMs) are trained with next-token prediction, implemented in autoregressive Transformers via causal masking for parallelism. This creates a subtle misalignment: residual connections tie activations to the current token, while supervision targets the next token, potentially propagating mismatched information if the current token is not the most informative for prediction. In this work, we empirically localize this input-output alignment shift in pretrained LLMs, using decoding trajectories over tied embedding spaces and similarity-based metrics. Our experiments reveal that the hidden token representations switch from input alignment to output alignment deep within the network. Motivated by this observation, we propose a lightweight residual-path mitigation based on residual attenuation, implemented either as a fixed-layer intervention or as a learnable gating mechanism. Experiments on multiple benchmarks show that these strategies alleviate the representation misalignment and yield improvements, providing an efficient and general architectural enhancement for autoregressive Transformers.

Method: Systematic benchmark of 17 LLMs using a unified evaluation framework for zero-shot annotation of Bangla hate speech, analyzing annotator bias and stability across different model scales.

Result: LLMs exhibit significant annotator bias and instability in judgments; surprisingly, increased model scale doesn’t guarantee better annotation quality, with smaller task-aligned models often showing more consistent behavior.

Conclusion: Current LLMs have important limitations for sensitive annotation tasks in low-resource languages, requiring careful evaluation before deployment despite their potential for scaling dataset creation.

Abstract: Large Language Models (LLMs) are increasingly used as automated annotators to scale dataset creation, yet their reliability as unbiased annotators–especially for low-resource and identity-sensitive settings–remains poorly understood. In this work, we study the behavior of LLMs as zero-shot annotators for Bangla hate speech, a task where even human agreement is challenging, and annotator bias can have serious downstream consequences. We conduct a systematic benchmark of 17 LLMs using a unified evaluation framework. Our analysis uncovers annotator bias and substantial instability in model judgments. Surprisingly, increased model scale does not guarantee improved annotation quality–smaller, more task-aligned models frequently exhibit more consistent behavior than their larger counterparts. These results highlight important limitations of current LLMs for sensitive annotation tasks in low-resource languages and underscore the need for careful evaluation before deployment.

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

Result: Created a comprehensive NLP library covering tokenization, morphological analysis, POS tagging, dependency parsing, NER, script transliteration, cross-lingual embeddings, and machine translation following CoNLL-U standards.

Conclusion: TurkicNLP provides the first unified NLP framework for Turkic languages, addressing fragmentation and enabling consistent processing across scripts with full interoperability.

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

Method: Developed a cascaded pipeline with: 1) Wav2Vec2-XLS-R-300m ASR model achieving 2.72% CER, 2) multi-stage fine-tuned MarianMT NMT model with 28.32 BLEU, and 3) proposed an intermediate Punctuation Restoration Module (PRM) to restore punctuation lost in ASR output.

Result: Punctuation loss caused 20.7% relative BLEU drop. The optimal pipeline with PRM applied directly to ASR output achieved 4.90 BLEU point gain (36.38 vs 31.48 BLEU). Human assessment confirmed superior Adequacy (3.673) and Fluency (3.804) with high inter-rater reliability.

Conclusion: Targeted punctuation restoration is the most effective intervention for mitigating structural noise in Nepali S2TT pipelines, establishing an optimized baseline and critical architectural insight for similar low-resource language systems.

Abstract: Cascaded speech-to-text translation (S2TT) systems for low-resource languages can suffer from structural noise, particularly the loss of punctuation during the Automatic Speech Recognition (ASR) phase. This research investigates the impact of such noise on Nepali-to-English translation and proposes an optimized pipeline to mitigate quality degradation. We first establish highly proficient ASR and NMT components: a Wav2Vec2-XLS-R-300m model achieved a state-of-the-art 2.72% CER on OpenSLR-54, and a multi-stage fine-tuned MarianMT model reached a 28.32 BLEU score on the FLORES-200 benchmark. We empirically investigate the influence of punctuation loss, demonstrating that unpunctuated ASR output significantly degrades translation quality, causing a massive 20.7% relative BLEU drop on the FLORES benchmark. To overcome this, we propose and evaluate an intermediate Punctuation Restoration Module (PRM). The final S2TT pipeline was tested across three configurations on a custom dataset. The optimal configuration, which applied the PRM directly to ASR output, achieved a 4.90 BLEU point gain over the direct ASR-to-NMT baseline (BLEU 36.38 vs. 31.48). This improvement was validated by human assessment, which confirmed the optimized pipeline’s superior Adequacy (3.673) and Fluency (3.804) with inter-rater reliability (Krippendorff’s $α {\geq}$ 0.723). This work validates that targeted punctuation restoration is the most effective intervention for mitigating structural noise in the Nepali S2TT pipeline. It establishes an optimized baseline and demonstrates a critical architectural insight for developing cascaded speech translation systems for similar low-resource languages.

Method: Used causal discovery as a testbed to evaluate eight frontier LLMs against ground truth derived from large-scale algorithm executions

Result: Found systematic, near-total failure: models produce overly wide confidence intervals yet still fail to contain true algorithmic mean in most cases; most perform worse than random guessing; best model’s marginal above-random performance likely due to benchmark memorization rather than principled reasoning

Conclusion: Identifies “algorithmic blindness” - a fundamental gap between declarative knowledge about algorithms and calibrated procedural prediction in LLMs

Abstract: Large language models (LLMs) demonstrate remarkable breadth of knowledge, yet their ability to reason about computational processes remains poorly understood. Closing this gap matters for practitioners who rely on LLMs to guide algorithm selection and deployment. We address this limitation using causal discovery as a testbed and evaluate eight frontier LLMs against ground truth derived from large-scale algorithm executions and find systematic, near-total failure. Models produce ranges far wider than true confidence intervals yet still fail to contain the true algorithmic mean in the majority of instances; most perform worse than random guessing and the marginal above-random performance of the best model is most consistent with benchmark memorization rather than principled reasoning. We term this failure algorithmic blindness and argue it reflects a fundamental gap between declarative knowledge about algorithms and calibrated procedural prediction.

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 2505.12565: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.12565&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 2404.08480: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2404.08480&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 2408.06361: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2408.06361&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 2502.01247: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.01247&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 2502.01481: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.01481&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 2506.00530: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.00530&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 2502.12179: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2502.12179&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 2503.04812: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.04812&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 2503.06749: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.06749&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: N/A - Paper content not accessible

Result: N/A - No results available for analysis

Conclusion: Cannot provide analysis due to arXiv API rate limiting preventing access to paper content

Abstract: Failed to fetch summary for 2503.08980: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2503.08980&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 2505.13109: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.13109&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Method unknown due to HTTP 429 error preventing access to paper details. The arXiv API rate limiting prevents retrieval of abstract and content.

Result: Results cannot be determined without access to the paper content. The HTTP 429 error indicates too many requests to arXiv API.

Conclusion: Unable to analyze paper due to technical limitations in accessing content. The arXiv API rate limiting prevents proper analysis.

Abstract: Failed to fetch summary for 2505.17132: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.17132&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to missing abstract

Result: Cannot determine results due to missing abstract

Conclusion: Cannot determine conclusion due to missing abstract

Abstract: Failed to fetch summary for 2506.06905: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.06905&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 2505.17519: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.17519&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 2505.18116: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.18116&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 analyze paper due to technical issue with arXiv API

Abstract: Failed to fetch summary for 2505.19025: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.19025&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 2506.24119: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.24119&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 2505.19590: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.19590&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 2507.16145: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2507.16145&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 2505.20278: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2505.20278&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 2506.03135: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.03135&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 2506.09332: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2506.09332&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper abstract

Result: Cannot determine results without access to paper abstract

Conclusion: Cannot draw conclusions without access to paper abstract

Abstract: Failed to fetch summary for 2508.05606: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.05606&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 2509.01938: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.01938&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - paper content inaccessible

Result: No results available - paper content inaccessible

Conclusion: Unable to analyze paper due to technical limitations in accessing content

Abstract: Failed to fetch summary for 2508.18672: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.18672&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error preventing access to paper content.

Result: Unable to determine results due to technical error preventing access to paper content.

Conclusion: Unable to draw conclusions due to technical error preventing access to paper content.

Abstract: Failed to fetch summary for 2509.03113: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.03113&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 2509.05608: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.05608&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to API rate limiting preventing access to paper content

Result: Unable to determine results due to API rate limiting preventing access to paper content

Conclusion: Unable to determine conclusion due to API rate limiting preventing access to paper content

Abstract: Failed to fetch summary for 2509.13615: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.13615&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 2509.21057: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.21057&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API returned HTTP 429 (Too Many Requests) error

Result: No results available - paper content inaccessible due to API rate limiting

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

Abstract: Failed to fetch summary for 2509.24156: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.24156&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 2509.24203: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.24203&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 2509.24393: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.24393&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method as the abstract could not be retrieved

Result: Unable to determine results as the abstract could not be retrieved

Conclusion: Unable to draw conclusions as the paper content could not be accessed

Abstract: Failed to fetch summary for 2509.25087: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.25087&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 2509.26278: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.26278&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 2509.26346: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.26346&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method due to failed paper retrieval

Result: Cannot determine results due to failed paper retrieval

Conclusion: Cannot determine conclusion due to failed paper retrieval

Abstract: Failed to fetch summary for 2510.01367: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.01367&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposed MemeXplain dataset for Arabic propagandistic memes and English hateful memes, with a multi-stage optimization approach for Vision-Language Models that separates training phases to avoid performance degradation.

Result: Significant improvements over base models: ~1.4% accuracy improvement on ArMeme (Arabic) and ~2.2% on Hateful Memes (English), outperforming state-of-the-art methods.

Conclusion: The multi-stage optimization approach effectively addresses the joint training challenge, improving both detection and explanation generation for multimodal harmful content moderation.

Abstract: The proliferation of multimodal content on social media presents significant challenges in understanding and moderating complex, context-dependent issues such as misinformation, hate speech, and propaganda. While efforts have been made to develop resources and propose new methods for automatic detection, limited attention has been given to jointly modeling label detection and the generation of explanation-based rationales, which often leads to degraded classification performance when trained simultaneously. To address this challenge, we introduce MemeXplain, an explanation-enhanced dataset for propagandistic memes in Arabic and hateful memes in English, making it the first large-scale resource for these tasks. To solve these tasks, we propose a multi-stage optimization approach and train Vision-Language Models (VLMs). Our results show that this strategy significantly improves both label detection and explanation generation quality over the base model, outperforming the current state-of-the-art with an absolute improvement of ~1.4% (Acc) on ArMeme and ~2.2% (Acc) on Hateful Memes. For reproducibility and future research, we aim to make the MemeXplain dataset and scripts publicly available (https://github.com/MohamedBayan/MemeIntel).

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 2510.01051: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.01051&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 draw conclusions due to retrieval failure

Abstract: Failed to fetch summary for 2510.01265: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.01265&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 2510.02209: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.02209&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 2510.02245: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.02245&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 2510.04067: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.04067&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 2510.05109: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.05109&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 2510.06823: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.06823&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 2510.16234: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.16234&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 2510.13117: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.13117&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method as abstract retrieval failed

Result: Cannot determine results as abstract retrieval failed

Conclusion: Cannot draw conclusions as abstract retrieval failed

Abstract: Failed to fetch summary for 2510.17206: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.17206&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to technical error in accessing paper content

Result: Unable to determine results due to technical error in accessing paper content

Conclusion: Unable to draw conclusions due to technical error in accessing paper content

Abstract: Failed to fetch summary for 2510.20095: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2510.20095&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 2511.00177: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.00177&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 2511.00206: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.00206&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Cannot determine method without access to paper abstract

Result: Cannot determine results without access to paper abstract

Conclusion: Cannot draw conclusions without access to paper abstract

Abstract: Failed to fetch summary for 2511.02044: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.02044&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - arXiv API returned HTTP 429 error (Too Many Requests)

Result: No results available - could not retrieve paper information due to rate limiting

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

Abstract: Failed to fetch summary for 2511.10788: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2511.10788&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: No method information available - API request resulted in HTTP 429 error

Result: No results available - could not retrieve paper information

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

Abstract: Failed to fetch summary for 2512.03794: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.03794&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Unable to determine method due to data retrieval error

Result: Unable to determine results due to data retrieval error

Conclusion: Unable to determine conclusion due to data retrieval error

Abstract: Failed to fetch summary for 2512.20745: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2512.20745&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 2601.20838: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2601.20838&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 2602.02185: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.02185&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 2602.10625: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.10625&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 2602.18671: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.18671&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 2602.19385: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.19385&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 2602.22530: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.22530&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 2602.24009: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.24009&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Used fMRI data from 40 subjects (20 prodromal Parkinson’s, 20 controls), fine-tuned ImageNet-pretrained CNNs (VGG19, Inception V3, Inception ResNet V2, MobileNet V1) under two data partitioning strategies: image-level splits (allowing slices from same subject in both train/test) vs strict subject-level splits.

Result: Image-level splits caused severe information leakage with near-perfect accuracy, while subject-level splits dropped performance to 60-81% accuracy. MobileNet demonstrated most reliable generalization despite fewer parameters, outperforming deeper architectures.

Conclusion: In extreme low-data regimes, evaluation strategy and model capacity have greater impact than architectural depth. Provides practical recommendations for evaluating deep learning models under severe data scarcity.

Abstract: Deep learning is often applied in settings where data are limited, correlated, and difficult to obtain, yet evaluation practices do not always reflect these constraints. Neuroimaging for prodromal Parkinsons disease is one such case, where subject numbers are small and individual scans produce many highly related samples. This work examines prodromal Parkinsons detection from resting-state fMRI as a machine learning problem centered on learning under extreme data scarcity. Using fMRI data from 40 subjects, including 20 prodromal Parkinsons cases and 20 healthy controls, ImageNet-pretrained convolutional neural networks are fine-tuned and evaluated under two different data partitioning strategies. Results show that commonly used image-level splits allow slices from the same subject to appear in both training and test sets, leading to severe information leakage and near-perfect accuracy. When a strict subject-level split is enforced, performance drops substantially, yielding test accuracies between 60 and 81 percent. Models with different capacity profiles are compared, including VGG19, Inception V3, Inception ResNet V2, and the lightweight MobileNet V1. Under subject-level evaluation, MobileNet demonstrates the most reliable generalization, outperforming deeper architectures despite having significantly fewer parameters. These results indicate that in extreme low-data regimes, evaluation strategy and model capacity have a greater impact on performance than architectural depth. Although the analysis is limited to a single cohort of 40 subjects and does not include external validation or cross-validation, it provides a concrete case study and practical recommendations for evaluating deep learning models under severe data scarcity.

Method: Proposed an open-source framework with nine detection methods for common errors in multi-sensor railway datasets. The tool automatically identifies issues like sensor synchronization problems, data corruption, and other quality concerns that could affect AI training.

Result: Six detection methods achieved 100% precision, while three additional methods reached precision rates of 96% and 97%. All detected errors were manually validated to evaluate performance.

Conclusion: The automated quality assurance tool significantly reduces manual workload and accelerates development of AI-based automated driving systems by efficiently detecting common errors in multi-sensor training data.

Abstract: The monitoring of the route and track environment plays an important role in automated driving. For example, it can be used as an assistance system for route monitoring in automation level Grade of Automation (GoA) 2, where the train driver is still on board. In fully automated, driverless driving at automation level GoA4, these systems finally take over environment monitoring completely independently. With the help of artificial intelligence (AI), they react automatically to risks and dangerous events on the route. To train such AI algorithms, large amounts of training data are required, which must meet high-quality standards due to their safety relevance. In this publication we present an automatic method for assuring the quality of training data, significantly reducing the manual workload and accelerating the development of these systems. We propose an open-source tool designed to detect nine common errors found in multi-sensor datasets for railway vehicles. To evaluate the performance of the framework, all detected errors were manually validated. Six issue detection methods achieved 100% precision, while three additional methods reached precision rates 96% and 97%.

Method: Proposes VoxelDiffusionCut which uses diffusion models to iteratively estimate internal 3D structures represented as voxels from observed cutting surfaces, then plans cuts based on estimation results to avoid damaging target parts.

Result: Simulation experiments show the method can estimate internal structures from cutting observations and enable non-destructive extraction of target parts by leveraging estimated uncertainty.

Conclusion: VoxelDiffusionCut successfully addresses challenges in internal structure estimation for non-destructive extraction by using diffusion models with voxel representations to capture uncertainty and avoid erroneous cuts.

Abstract: Non-destructive extraction of the target internal part, such as batteries and motors, by cutting surrounding structures is crucial at recycling and disposal sites. However, the diversity of products and the lack of information on disassembly procedures make it challenging to decide where to cut. This study explores a method for non-destructive extraction of a target internal part that iteratively estimates the internal structure from observed cutting surfaces and formulates cutting plans based on the estimation results. A key requirement is to estimate the probability of the target part’s presence from partial observations. However, learning conditional generative models for this task is challenging: The high dimensionality of 3D shape representations makes learning difficult, and conventional models (e.g., conditional variational autoencoders) often fail to capture multi-modal predictive uncertainty due to mode collapse, resulting in overconfident predictions. To address these issues, we propose VoxelDiffusionCut, which iteratively estimates the internal structure represented as voxels using a diffusion model and plans cuts for non-destructive extraction of the target internal part based on the estimation results. Voxel representation allows the model to predict only attributes at fixed grid positions, i.e., types of constituent parts, making learning more tractable. The diffusion model completes the voxel representation conditioned on observed cutting surfaces, capturing uncertainty in unobserved regions to avoid erroneous cuts. Experimental results in simulation suggest that the proposed method can estimate internal structures from observed cutting surfaces and enable non-destructive extraction of the target internal part by leveraging the estimated uncertainty.

Method: Proposes Multi-scale Spatial Adaptive Attention Network with MSAA module containing Global Feature Modulation Module for texture structures and Multi-scale Feature Aggregation Module for pyramidal feature fusion. Adds Local Enhancement Block for geometric perception and Feature Interactive Gated Feed-Forward Module for nonlinear representation.

Result: Achieves superior or competitive PSNR/SSIM performance on standard benchmarks (Set5, Set14, B100, Urban100, Manga109) across ×2, ×3, ×4 scaling factors while maintaining significantly lower parameters and computational costs than SOTA methods.

Conclusion: MSAAN effectively balances reconstruction quality and model efficiency through its novel attention mechanisms, making it suitable for practical SR applications with computational constraints.

Abstract: This paper introduces a lightweight image super-resolution (SR) network, termed the Multi-scale Spatial Adaptive Attention Network (MSAAN), to address the common dilemma between high reconstruction fidelity and low model complexity in existing SR methods. The core of our approach is a novel Multi-scale Spatial Adaptive Attention Module (MSAA), designed to jointly model fine-grained local details and long-range contextual dependencies. The MSAA comprises two synergistic components: a Global Feature Modulation Module (GFM) that learns coherent texture structures through differential feature extraction, and a Multi-scale Feature Aggregation Module (MFA) that adaptively fuses features from local to global scales using pyramidal processing. To further enhance the network’s capability, we propose a Local Enhancement Block (LEB) to strengthen local geometric perception and a Feature Interactive Gated Feed-Forward Module (FIGFF) to improve nonlinear representation while reducing channel redundancy. Extensive experiments on standard benchmarks (Set5, Set14, B100, Urban100, Manga109) across $\times2$, $\times3$, and $\times4$ scaling factors demonstrate that both our lightweight (MSAAN-light) and standard (MSAAN) versions achieve superior or competitive performance in terms of PSNR and SSIM, while maintaining significantly lower parameters and computational costs than state-of-the-art methods. Ablation studies validate the contribution of each component, and visual results show that MSAAN reconstructs sharper edges and more realistic textures.

Method: QuickGrasp uses a local-first architecture with on-demand edge augmentation, sharing vision representation across model variants to avoid redundant computation. It introduces three key designs: accelerated video tokenization, query-adaptive edge augmentation, and delay-aware, accuracy-preserving vision token density configuration.

Result: QuickGrasp matches the accuracy of large VLMs while achieving up to 12.8x reduction in response delay across multiple video understanding benchmarks.

Conclusion: QuickGrasp represents a key advancement toward building responsive video querying services for open-world understanding that fully leverage VLM capabilities by balancing accuracy and latency through intelligent system design.

Abstract: Video-language models (VLMs) are reshaping video querying services, bringing unified solutions to complex perception and reasoning tasks. However, deploying large VLMs in real-world systems remains challenging due to their high resource demands, and remote-based deployment often results in unacceptable response delays. Although small, locally deployable VLMs offer faster responses, they unavoidably fall short in accuracy. To reconcile this trade-off, we propose QuickGrasp, a responsive, quality of service (QoS)-aware system that bridges this gap through a local-first architecture with on-demand edge augmentation. Built upon the highly modular architecture of VLMs, QuickGrasp shares the vision representation across model variants to avoid redundant computation. To maximize system-wide efficiency, QuickGrasp introduces three key designs: accelerated video tokenization, query-adaptive edge augmentation, and delay-aware, accuracy-preserving vision token density configuration. We implement a prototype of QuickGrasp and evaluate it across multiple video understanding benchmarks. The results show that QuickGrasp matches the accuracy of large VLMs while achieving up to a 12.8x reduction in response delay. QuickGrasp represents a key advancement toward building responsive video querying services for open-world understanding that fully leverage the capabilities of VLMs.

Method: BiSe-UNet combines an attention-refined context path for global understanding with a shallow spatial path for detailed feature preservation, followed by a depthwise separable decoder for efficient reconstruction. This dual-path approach balances computational efficiency with segmentation quality.

Result: On the Kvasir-Seg dataset, BiSe-UNet achieves competitive Dice and IoU scores while sustaining real-time throughput exceeding 30 FPS on Raspberry Pi 5, demonstrating practical deployment capability on edge hardware.

Conclusion: BiSe-UNet provides an effective solution for accurate, lightweight medical image segmentation suitable for real-time clinical deployment on resource-constrained edge devices, addressing the trade-off between computational efficiency and segmentation quality.

Abstract: During image-guided procedures, real-time image segmentation is often required. This demands lightweight AI models that can operate on resource-constrained devices. One important use case is endoscopy-guided colonoscopy, where polyps must be detected in real time. The Kvasir-Seg dataset, a publicly available benchmark for this task, contains 1,000 high-resolution endoscopic images of polyps with corresponding pixel-level segmentation masks. Achieving real-time inference speed for clinical deployment in constrained environments requires highly efficient and lightweight network architectures. However, many existing models remain too computationally intensive for embedded deployment. Lightweight architectures, although faster, often suffer from reduced spatial precision and weaker contextual understanding, leading to degraded boundary quality and reduced diagnostic reliability. To address these challenges, we introduce BiSe-UNet, a lightweight dual-path U-Net that integrates an attention-refined context path with a shallow spatial path for detailed feature preservation, followed by a depthwise separable decoder for efficient reconstruction. Evaluated on the Kvasir-Seg dataset, BiSe-UNet achieves competitive Dice and IoU scores while sustaining real-time throughput exceeding 30 FPS on Raspberry Pi 5, demonstrating its effectiveness for accurate, lightweight, and deployable medical image segmentation on edge hardware.

Method: Reinforcement learning framework with multimodal backbone for autoregressive audio-to-video generation, using MLLM-based human-aligned evaluation system to assess lip-sync, expressiveness, and motion quality, combined with perceptual/temporal consistency regularizers to form composite reward for post-training via Group Relative Policy Optimization (GRPO).

Result: Extensive experiments show FlowPortrait consistently produces higher-quality talking-head videos, demonstrating effectiveness of RL for portrait animation.

Conclusion: Reinforcement learning with human-aligned MLLM evaluation effectively improves talking-head video generation quality across multiple dimensions.

Abstract: Generating realistic talking-head videos remains challenging due to persistent issues such as imperfect lip synchronization, unnatural motion, and evaluation metrics that correlate poorly with human perception. We propose FlowPortrait, a reinforcement-learning framework for audio-driven portrait animation built on a multimodal backbone for autoregressive audio-to-video generation. FlowPortrait introduces a human-aligned evaluation system based on Multimodal Large Language Models (MLLMs) to assess lip-sync accuracy, expressiveness, and motion quality. These signals are combined with perceptual and temporal consistency regularizers to form a stable composite reward, which is used to post-train the generator via Group Relative Policy Optimization (GRPO). Extensive experiments, including both automatic evaluations and human preference studies, demonstrate that FlowPortrait consistently produces higher-quality talking-head videos, highlighting the effectiveness of reinforcement learning for portrait animation.

Method: Uses two concurrent YOLO object detection models (element detection and layout detection) with rule-based grouping. Includes vision-language enhancement where images are first filtered by ViT classifier, then only relevant images go to Vision LLM for analysis. Parallel execution for detection, classification, OCR, and conversion.

Result: Achieves 96.49% TEDS and 98.51% NID on DP-Bench benchmark, outperforming both commercial and open-source parsers. System works on CPU, generates multiple output formats (JSON, Markdown, RAG-ready texts, knowledge graphs).

Conclusion: NovaLAD provides an accurate, fast, and cost-effective document parsing solution that balances performance with practical deployment considerations, making document extraction accessible without GPU requirements.

Abstract: Document extraction is an important step before retrieval-augmented generation (RAG), knowledge bases, and downstream generative AI can work. It turns unstructured documents like PDFs and scans into structured text and layout-aware representations. We introduce NovaLAD, a comprehensive document parsing system that integrates two concurrent YOLO object detection models - element detection and layout detection - with rule-based grouping and optional vision-language enhancement. When a page image is sent in, the first thing that happens is that it goes through both models at the same time. The element model finds semantic content like the title, header, text, table, image, and so on, and the layout model finds structural regions like layout_box, column_group, multi_column, row_group, and so on. A key design decision is to first send an image or figure through an image classifier (ViT) that decides whether it is relevant or not. Only useful images are then submitted to the Vision LLM for title, summary, and structured information, which cuts down on noise and costs. NovaLAD is built for speed: it works on CPU, employs parallel execution for detection, classification, OCR, and conversion, and generates several forms, including structured JSON, Markdown, RAG-ready texts, and knowledge graphs. We test on the DP-Bench benchmark (upstage/dp-bench) and get 96.49% TEDS and 98.51% NID, which is better than both commercial and open-source parsers. This paper explains how to extract data, how the architecture works, how data flows, and how to make NovaLAD both accurate and usable without needing a GPU.

Method: Propose CT-Flow framework using Model Context Protocol (MCP) for interoperable volumetric interpretation. Create CT-FlowBench benchmark for 3D CT tool-use and multi-step reasoning. Framework acts as clinical orchestrator decomposing queries into automated tool-use sequences.

Result: State-of-the-art performance on CT-FlowBench and standard 3D VQA datasets. 41% improvement in diagnostic accuracy over baselines. 95% success rate in autonomous tool invocation.

Conclusion: CT-Flow provides scalable foundation for integrating autonomous, agentic intelligence into real-world clinical radiology by shifting from closed-box inference to open, tool-aware paradigm.

Abstract: Recent advances in Large Vision-Language Models (LVLMs) have shown strong potential for multi-modal radiological reasoning, particularly in tasks like diagnostic visual question answering (VQA) and radiology report generation. However, most existing approaches for 3D CT analysis largely rely on static, single-pass inference. In practice, clinical interpretation is a dynamic, tool-mediated workflow where radiologists iteratively review slices and use measurement, radiomics, and segmentation tools to refine findings. To bridge this gap, we propose CT-Flow, an agentic framework designed for interoperable volumetric interpretation. By leveraging the Model Context Protocol (MCP), CT-Flow shifts from closed-box inference to an open, tool-aware paradigm. We curate CT-FlowBench, the first large-scale instruction-tuning benchmark tailored for 3D CT tool-use and multi-step reasoning. Built upon this, CT-Flow functions as a clinical orchestrator capable of decomposing complex natural language queries into automated tool-use sequences. Experimental evaluations on CT-FlowBench and standard 3D VQA datasets demonstrate that CT-Flow achieves state-of-the-art performance, surpassing baseline models by 41% in diagnostic accuracy and achieving a 95% success rate in autonomous tool invocation. This work provides a scalable foundation for integrating autonomous, agentic intelligence into real-world clinical radiology.

Method: Uses a unified end-to-end diffusion framework with Multi-Modal Transformer Blocks that explicitly model temporal correspondence between audio and video latent tokens via shared self-attention. Leverages pre-trained video generation model priors for visual fidelity and includes personalized voice cloning from brief audio references.

Result: Produces highly realistic talking portraits with superior performance over existing open-source approaches in lip-sync accuracy, audio naturalness, and overall perceptual quality.

Conclusion: UniTalking successfully bridges the accessibility gap in audio-video generation while achieving state-of-the-art performance in generating synchronized talking portraits with voice cloning capabilities.

Abstract: While state-of-the-art audio-video generation models like Veo3 and Sora2 demonstrate remarkable capabilities, their closed-source nature makes their architectures and training paradigms inaccessible. To bridge this gap in accessibility and performance, we introduce UniTalking, a unified, end-to-end diffusion framework for generating high-fidelity speech and lip-synchronized video. At its core, our framework employs Multi-Modal Transformer Blocks to explicitly model the fine-grained temporal correspondence between audio and video latent tokens via a shared self-attention mechanism. By leveraging powerful priors from a pre-trained video generation model, our framework ensures state-of-the-art visual fidelity while enabling efficient training. Furthermore, UniTalking incorporates a personalized voice cloning capability, allowing the generation of speech in a target style from a brief audio reference. Qualitative and quantitative results demonstrate that our method produces highly realistic talking portraits, achieving superior performance over existing open-source approaches in lip-sync accuracy, audio naturalness, and overall perceptual quality.

Method: Uses Dynamic Graph CNN trained on landmark-reconstructed point clouds from 3DTeethLand dataset, integrated with rule-based biomechanical engine grounded in orthodontic evidence. System decomposes per-tooth motion across six degrees of freedom and computes movement-specific predictability.

Result: Segmentation achieves 81.4% Tooth Identification Rate and 8.25% mIoU on surrogate point clouds. The end-to-end pipeline runs in <4s on consumer hardware. System provides alerts when biomechanical limits are exceeded and derives composite index.

Conclusion: Establishes baseline for future full-mesh training, highlights current perceptual limits, and provides open-source tools for reproducible research in geometric deep learning and digital orthodontics.

Abstract: Clear aligner therapy now dominates orthodontics, yet clinician review of digitally planned tooth movements-typically via ClinCheck (Align Technology)-remains slow and error-prone. We present OrthoAI, an open-source proof-of-concept decision-support system combining lightweight 3D dental segmentation with automated biomechanical analysis to assist treatment-plan evaluation. The framework uses a Dynamic Graph CNN trained on landmark-reconstructed point clouds from 3DTeethLand (MICCAI) and integrates a rule-based biomechanical engine grounded in orthodontic evidence (Kravitz et al 2009; Simon et al 2014). The system decomposes per-tooth motion across six degrees of freedom, computes movement-specific predictability, issues alerts when biomechanical limits are exceeded, and derives an exploratory composite index. With 60,705 trainable parameters, segmentation reaches a Tooth Identification Rate of $81.4%$ and mIoU of $8.25%$ on surrogate point clouds-reflecting sparse landmark supervision rather than dense meshes. Although spatial boundaries are coarse, downstream analysis depends mainly on tooth identity and approximate centroid/axis estimation. Results establish a baseline for future full-mesh training and highlight current perceptual limits. The end-to-end pipeline runs in $<4s$ on consumer hardware. Code, weights, and analysis tools are released to support reproducible research in geometric deep learning and digital orthodontics. The system has not been validated on real intraoral meshes and should not be assumed to generalize beyond landmark-derived representations.

Method: Proposes MM-Mem with hierarchical memory: Sensory Buffer (fine-grained perceptual traces), Episodic Stream, and Symbolic Schema (high-level semantics). Uses Semantic Information Bottleneck objective with SIB-GRPO for memory optimization, and entropy-driven top-down retrieval that drills down from abstract to detailed memory under uncertainty.

Result: Extensive experiments across 4 benchmarks show effectiveness on both offline and streaming tasks, demonstrating robust generalization and validating the cognition-inspired memory organization.

Conclusion: MM-Mem bridges the gap between vision-centric and text-centric approaches for long-horizon video understanding through hierarchical memory organization inspired by human cognition, achieving better efficiency and accuracy.

Abstract: While multimodal large language models have demonstrated impressive short-term reasoning, they struggle with long-horizon video understanding due to limited context windows and static memory mechanisms that fail to mirror human cognitive efficiency. Existing paradigms typically fall into two extremes: vision-centric methods that incur high latency and redundancy through dense visual accumulation, or text-centric approaches that suffer from detail loss and hallucination via aggressive captioning. To bridge this gap, we propose MM-Mem, a pyramidal multimodal memory architecture grounded in Fuzzy-Trace Theory. MM-Mem structures memory hierarchically into a Sensory Buffer, Episodic Stream, and Symbolic Schema, enabling the progressive distillation of fine-grained perceptual traces (verbatim) into high-level semantic schemas (gist). Furthermore, to govern the dynamic construction of memory, we derive a Semantic Information Bottleneck objective and introduce SIB-GRPO to optimize the trade-off between memory compression and task-relevant information retention. In inference, we design an entropy-driven top-down memory retrieval strategy, which first tries with the abstract Symbolic Schema and progressively “drills down” to the Sensory Buffer and Episodic Stream under high uncertainty. Extensive experiments across 4 benchmarks confirm the effectiveness of MM-Mem on both offline and streaming tasks, demonstrating robust generalization and validating the effectiveness of cognition-inspired memory organization. Code is available at https://github.com/EliSpectre/MM-Mem.

Method: Self-annotation technique using superpixel algorithms to identify perceptually similar local regions, relating them to global context via CNN receptive field to learn global-local relationships for semantic segmentation.

Result: Performance of unsupervised training methodology on XCT datasets is presented, with discussion of potential improvements.

Conclusion: Unsupervised self-annotation approach shows promise for semantic segmentation of challenging low-contrast XCT images without requiring labeled training data.

Abstract: This work tests a self-annotation-based unsupervised methodology for training a convolutional neural network (CNN) model for semantic segmentation of X-ray computed tomography (XCT) scans of concretes. Concrete poses a unique challenge for XCT imaging due to similar X-ray attenuation coefficients of aggregates and mortar, resulting in low-contrast between the two phases in the ensuing images. While CNN-based models are a proven technique for semantic segmentation in such challenging cases, they typically require labeled training data, which is often unavailable for new datasets or are costly to obtain. To counter that limitation, a self-annotation technique is used here which leverages superpixel algorithms to identify perceptually similar local regions in an image and relates them to the global context in the image by utilizing the receptive field of a CNN-based model. This enables the model to learn a global-local relationship in the images and enables identification of semantically similar structures. We therefore present the performance of the unsupervised training methodology on our XCT datasets and discuss potential avenues for further improvements.

Method: Calculated 321 2D morphometric attributes from building footprints and street networks at multiple spatial scales. Developed four classification schemes: 1) morphometric-based LCZ prediction, 2) baseline image-based prediction, and 3) two fusion techniques combining morphometrics with satellite imagery. Evaluated across five different sites.

Result: Morphometric-based prediction showed selective and inconsistent correspondence with LCZ types, being highly site-dependent. Fusion approaches yielded modest accuracy improvements at only two sites, with negligible or slightly negative gains at others. The study revealed that a broader range of urban form properties are relevant for distinguishing LCZ types than standard parameters suggest.

Conclusion: The relationship between LCZs and measurable, visible aspects of urban form is tenuous. The LCZ framework should be used cautiously in morphological studies, as morphometrics alone cannot reliably predict LCZ classifications across different urban contexts.

Abstract: The Local Climate Zone (LCZ) framework is commonly employed to represent urban form in morphological analyses despite its mapping predominantly relies on satellite imagery. Urban morphometrics, describing urban form via numerical measures of physical aspects and spatial relationships of its elements, offers another avenue. This study evaluates the ability of morphometric assessment to predict LCZs using a) a morphometric-based LCZ prediction, and b) a fusion-based LCZ prediction combining morphometrics with satellite imagery. We calculate 321 2D morphometric attributes from building footprints and street networks, covering their various properties at multiple spatial scales. Subsequently, we develop four classification schemes: morphometric-based prediction, baseline image-based prediction, and two techniques fusing morphometrics with imagery. We evaluate them across five sites. Results from the morphometric-based prediction indicate that the correspondence between 2D urban morphometrics and urban LCZ types is selective and inconsistent, rendering the efficacy of this method site-dependent. Nevertheless, it demonstrated that a much broader range of urban form properties is relevant for distinguishing LCZ types compared to standard parameters. Relative to the image-based baseline, the fusion yielded relatively distinct accuracy improvements for urban LCZ types at two sites; however, gains at the remaining sites were negligible or even slightly negative, suggesting that the benefits of fusion are modest and inconsistent. Collectively, these results indicate that the relationship between the LCZs and the measurable, visible aspects of urban form is tenuous, thus the LCZ framework should be used with caution in morphological studies.

Method: GUARD uses attractive-repulsive dynamics during image denoising. It includes a novel cross-attention attenuation mechanism that: (1) statistically identifies prompt positions where cross-attention must be attenuated, and (2) attenuates cross-attention in these per-prompt locations. This provides a surgical, dynamic per-prompt inference-time approach.

Result: GUARD produces state-of-the-art results for memorization mitigation across two architectures and for both verbatim and template memorization. It consistently outperforms other methods while improving or maintaining comparable image quality.

Conclusion: GUARD offers an effective inference-time solution for memorization mitigation in text-to-image diffusion models that is robust across architectures and memorization types while preserving image quality.

Abstract: Generative models have been shown to “memorize” certain training data, leading to verbatim or near-verbatim generating images, which may cause privacy concerns or copyright infringement. We introduce Guidance Using Attractive-Repulsive Dynamics (GUARD), a novel framework for memorization mitigation in text-to-image diffusion models. GUARD adjusts the image denoising process to guide the generation away from an original training image and towards one that is distinct from training data while remaining aligned with the prompt, guarding against reproducing training data, without hurting image generation quality. We propose a concrete instantiation of this framework, where the positive target that we steer towards is given by a novel method for (cross) attention attenuation based on (i) a novel statistical mechanism that automatically identifies the prompt positions where cross attention must be attenuated and (ii) attenuating cross-attention in these per-prompt locations. The resulting GUARD offers a surgical, dynamic per-prompt inference-time approach that, we find, is by far the most robust method in terms of consistently producing state-of-the-art results for memorization mitigation across two architectures and for both verbatim and template memorization, while also improving upon or yielding comparable results in terms of image quality.

Method: Three key innovations: (1) Decoupled architecture separating visual inference from text encoding, allowing precomputed class embeddings in flash memory; (2) Matryoshka distillation training nested embeddings at multiple dimensions (16-256) for flexible accuracy-memory trade-offs; (3) Quantized embedding storage reducing class prototype memory by 4x with minimal accuracy loss.

Result: Trained on CC3M, achieves competitive zero-shot accuracy on COCO, Flowers102, and Food101 with only 285KB RAM and 892KB flash memory. Real-time inference at 26 FPS on STM32H7 and >1,000 FPS on MAX78000 with CNN accelerator.

Conclusion: Enables practical zero-shot object detection on edge devices for the first time, overcoming memory constraints of MCUs while maintaining competitive accuracy.

Abstract: Zero-shot object detection enables recognising novel objects without task-specific training, but current approaches rely on large vision language models (VLMs) like CLIP that require hundreds of megabytes of memory - far exceeding the constraints of micro controller units (MCUs). We present TinyVLM, the first framework enabling zero-shot object detection on resource-constrained MCUs with less than 1MB of memory. Our approach introduces three key innovations: (1) a decoupled architecture that separates visual inference from text encoding, allowing precomputed class embeddings to be stored in flash memory; (2) Matryoshka distillation that trains nested embeddings at multiple dimensions (16-256), enabling flexible accuracy-memory trade-offs; and (3) quantized embedding storage that reduces class prototype memory by 4x with minimal accuracy loss. Trained on Conceptual Captions 3M (CC3M), TinyVLM achieves competitive zero-shot accuracy on COCO, Flowers102, and Food101 while requiring only 285KB of RAM and 892KB of flash memory for the deployed vision encoder. We demonstrate real-time inference at 26 FPS on STM32H7 and over 1,000 FPS on MAX78000 with its CNN accelerator, enabling practical zero-shot detection on edge devices for the first time.

Method: ADE-CoT incorporates three key strategies: (1) difficulty-aware resource allocation that assigns dynamic budgets based on estimated edit difficulty, (2) edit-specific verification using region localization and caption consistency for early pruning, and (3) depth-first opportunistic stopping guided by an instance-specific verifier.

Result: Extensive experiments on three state-of-the-art editing models (Step1X-Edit, BAGEL, FLUX.1 Kontext) across three benchmarks show ADE-CoT achieves superior performance-efficiency trade-offs, obtaining better performance with more than 2x speedup over Best-of-N with comparable sampling budgets.

Conclusion: ADE-CoT effectively addresses the challenges of applying Image-CoT to image editing by providing an adaptive framework that improves both efficiency and editing quality through intelligent resource allocation and verification mechanisms.

Abstract: Image Chain-of-Thought (Image-CoT) is a test-time scaling paradigm that improves image generation by extending inference time. Most Image-CoT methods focus on text-to-image (T2I) generation. Unlike T2I generation, image editing is goal-directed: the solution space is constrained by the source image and instruction. This mismatch causes three challenges when applying Image-CoT to editing: inefficient resource allocation with fixed sampling budgets, unreliable early-stage verification using general MLLM scores, and redundant edited results from large-scale sampling. To address this, we propose ADaptive Edit-CoT (ADE-CoT), an on-demand test-time scaling framework to enhance editing efficiency and performance. It incorporates three key strategies: (1) a difficulty-aware resource allocation that assigns dynamic budgets based on estimated edit difficulty; (2) edit-specific verification in early pruning that uses region localization and caption consistency to select promising candidates; and (3) depth-first opportunistic stopping, guided by an instance-specific verifier, that terminates when intent-aligned results are found. Extensive experiments on three SOTA editing models (Step1X-Edit, BAGEL, FLUX.1 Kontext) across three benchmarks show that ADE-CoT achieves superior performance-efficiency trade-offs. With comparable sampling budgets, ADE-CoT obtains better performance with more than 2x speedup over Best-of-N.

Method: Proposes a scalable approach to annotate videos as interleaved multimodal sequences. Designs a block-causal diffusion transformer trained on three proxy tasks: next-image prediction, current segmentation prediction, and next-segmentation prediction. Also introduces a novel multi-turn image editing benchmark.

Result: The model exhibits strong in-context image editing capabilities and achieves state-of-the-art results on two multi-turn image editing benchmarks. Despite video-only training, it shows promising abilities in multi-concept composition, story generation, and chain-of-editing applications.

Conclusion: In-context image editing models can be effectively learned directly from videos without task-specific pipelines, demonstrating strong performance and generalization to various editing applications.

Abstract: In-context image editing aims to modify images based on a contextual sequence comprising text and previously generated images. Existing methods typically depend on task-specific pipelines and expert models (e.g., segmentation and inpainting) to curate training data. In this work, we explore whether an in-context image editing model can be learned directly from videos. We introduce a scalable approach to annotate videos as interleaved multimodal sequences. To effectively learn from this data, we design a block-causal diffusion transformer trained on three proxy tasks: next-image prediction, current segmentation prediction, and next-segmentation prediction. Additionally, we propose a novel multi-turn image editing benchmark to advance research in this area. Extensive experiments demonstrate that our model exhibits strong in-context image editing capabilities and achieves state-of-the-art results on two multi-turn image editing benchmarks. Despite being trained exclusively on videos, our model also shows promising abilities in multi-concept composition, story generation, and chain-of-editing applications.

Method: 1. Task-adaptive compression using learnable FiLM conditioning to preserve discriminative features per task; 2. Spatial-diverse exemplar selection via farthest-point sampling in IoU space to prevent localization bias; 3. MCU-deployable system storing only 150 bytes per sample instead of >10KB for images.

Result: LRD achieves strong mAP@50 on CORe50 (50 classes, 5 tasks), maintains performance across tasks, and runs on STM32H753ZI, ESP32-S3, and MAX78000 MCUs with 4.9-97.5ms latency per inference within 64KB memory budget.

Conclusion: LRD enables practical continual object detection on edge devices for the first time by synergistically combining task-adaptive compression and spatial-diverse exemplar selection within strict MCU memory constraints.

Abstract: Deploying object detection on microcontrollers (MCUs) enables intelligent edge devices but current models cannot learn new object categories after deployment. Existing continual learning methods require storing raw images far exceeding MCU memory budgets of tens of kilobytes. We present Latent Replay Detection (LRD), the first framework for continual object detection under MCU memory constraints. Our key contributions are: 1. Task-Adaptive Compression: Unlike fixed PCA, we propose learnable compression with FiLM (Feature-wise Linear Modulation) conditioning, where task specific embeddings modulate the compression to preserve discriminative features for each task’s distribution; 2. Spatial-Diverse Exemplar Selection: Traditional sampling ignores spatial information critical for detection - we select exemplars maximizing bounding box diversity via farthest-point sampling in IoU space, preventing localization bias in replay; 3. MCU-Deployable System: Our latent replay stores 150 bytes per sample versus >10KB for images, enabling a 64KB buffer to hold 400+ exemplars. Experiments on CORe50 (50 classes, 5 tasks) demonstrate that LRD achieves mAP@50 on the initial task and maintains strong performance across subsequent tasks - a significant improvement over naive fine-tuning while operating within strict MCU constraints. Our task-adaptive FiLM compression and spatial diverse exemplar selection work synergistically to preserve detection capabilities. Deployed on STM32H753ZI, ESP32-S3, and MAX78000 MCUs, LRD achieves 4.9-97.5ms latency per inference within a 64KB memory budget-enabling practical continual detection on edge devices for the first time.

Method: Combines SAM2 for image segmentation, Florence2 and ChatGPT for labeling, with specialized ontology (ontoShip) and glossary (glosShip) of nautical architecture to enhance labeling accuracy for XVI-XVII century shipbuilding treatises

Result: Preliminary results show potential for marrying these technologies to improve curation and retrieval of historical documents, though challenges and limitations are acknowledged

Conclusion: The approach demonstrates feasibility for historical document analysis despite data scarcity, with ongoing work needed to address current limitations

Abstract: Image segmentation and image recognition are well established computational techniques in the broader discipline of image processing. Segmentation allows to locate areas in an image, while recognition identifies specific objects within an image. These techniques have shown remarkable accuracy with modern images, mainly because the amount of training data is vast. Achieving similar accuracy in digitized images of centuries-old documents is more challenging. This difficulty is due to two main reasons: first, the lack of sufficient training data, and second, because the degree of specialization in a given domain. Despite these limitations, the ability to segment and recognize objects in these collections is important for automating the curation, cataloging, and dissemination of knowledge, making the contents of priceless collections accessible to scholars and the general public. In this paper, we report on our ongoing work in segmenting and labeling images pertaining to shipbuilding treatises from the XVI and XVII centuries, a historical period known as the Age of Exploration. To this end, we leverage SAM2 for image segmentation; Florence2 and ChatGPT for labeling; and a specialized ontology ontoShip and glossary glosShip of nautical architecture for enhancing the labeling process. Preliminary results demonstrate the potential of marrying these technologies for improving curation and retrieval of priceless historical documents. We also discuss the challenges and limitations encountered in this approach and ideas on how to overcome them in the future.

Method: Neural network-based solution that considers spatio-temporal variability across different agricultural parcels. Uses real-world remote sensing dataset for training and validation.

Result: Experimental study validates the soundness of TerrAI on data-driven agricultural practices. The system effectively handles the complex variability in agricultural conditions.

Conclusion: TerrAI provides an effective neural network solution for targeted agricultural operations that can optimize resource use and promote environmental sustainability through precise application of inputs.

Abstract: The modernization of agriculture has motivated the development of advanced analytics and decision-support systems to improve resource utilization and reduce environmental impacts. Targeted Spraying and Fertilization (TSF) is a critical operation that enables farmers to apply inputs more precisely, optimizing resource use and promoting environmental sustainability. However, accurate TSF is a challenging problem, due to external factors such as crop type, fertilization phase, soil conditions, and weather dynamics. In this paper, we present TerrAI, a Neural Network-based solution for TSF, which considers the spatio-temporal variability across different parcels. Our experimental study over a real-world remote sensing dataset validates the soundness of TerrAI on data-driven agricultural practices.

Method: TTOM uses test-time optimization with a layout-attention objective to align VFM outputs with spatiotemporal layouts during inference. It integrates and optimizes new parameters rather than directly intervening in latents or attention. The framework operates in a streaming setting with a parametric memory mechanism that supports insert, read, update, and delete operations for maintaining historical optimization contexts.

Result: Experimental results on T2V-CompBench and Vbench benchmarks show TTOM effectively improves compositional video generation. The framework demonstrates powerful transferability and generalization, disentangling compositional world knowledge while being practical, scalable, and efficient.

Conclusion: TTOM provides an effective training-free framework for achieving cross-modal alignment in compositional video generation, addressing VFM limitations in compositional scenarios through test-time optimization and memory mechanisms.

Abstract: Video Foundation Models (VFMs) exhibit remarkable visual generation performance, but struggle in compositional scenarios (e.g., motion, numeracy, and spatial relation). In this work, we introduce Test-Time Optimization and Memorization (TTOM), a training-free framework that aligns VFM outputs with spatiotemporal layouts during inference for better text-image alignment. Rather than direct intervention to latents or attention per-sample in existing work, we integrate and optimize new parameters guided by a general layout-attention objective. Furthermore, we formulate video generation within a streaming setting, and maintain historical optimization contexts with a parametric memory mechanism that supports flexible operations, such as insert, read, update, and delete. Notably, we found that TTOM disentangles compositional world knowledge, showing powerful transferability and generalization. Experimental results on the T2V-CompBench and Vbench benchmarks establish TTOM as an effective, practical, scalable, and efficient framework to achieve cross-modal alignment for compositional video generation on the fly.

Method: Models diffusion denoising as a dynamical system, applies reachability analysis to approximate the “backward reachable tube” (states that evolve into memorized samples), and formulates mitigation as constrained RL where a policy learns to steer trajectories via minimal perturbations in caption embedding space.

Result: Achieves superior Pareto frontier between generation diversity (SSCD), quality (FID), and alignment (CLIP) compared to state-of-the-art baselines. Provides robust mitigation without modifying diffusion backbone, offering plug-and-play solution.

Conclusion: RADS effectively prevents memorization in text-to-image diffusion models while maintaining generation quality and alignment, offering a practical inference-time solution without model modifications.

Abstract: Text-to-image diffusion models often memorize training data, revealing a fundamental failure to generalize beyond the training set. Current mitigation strategies typically sacrifice image quality or prompt alignment to reduce memorization. To address this, we propose Reachability-Aware Diffusion Steering (RADS), an inference-time framework that prevents memorization while preserving generation fidelity. RADS models the diffusion denoising process as a dynamical system and applies concepts from reachability analysis to approximate the “backward reachable tube”–the set of intermediate states that inevitably evolve into memorized samples. We then formulate mitigation as a constrained reinforcement learning (RL) problem, where a policy learns to steer the trajectory away from memorization via minimal perturbations in the caption embedding space. Empirical evaluations show that RADS achieves a superior Pareto frontier between generation diversity (SSCD), quality (FID), and alignment (CLIP) compared to state-of-the-art baselines. Crucially, RADS provides robust mitigation without modifying the diffusion backbone, offering a plug-and-play solution for safe generation. Our website is available at: https://s-karnik.github.io/rads-memorization-project-page/.

Method: The paper describes an image processing technology approach for automatic number-plate recognition (ANPR), also known as automatic vehicle identification or optical character recognition for cars, though specific technical details are not provided in the abstract.

Result: The abstract presents a conceptual framework for implementing car license plate recognition systems in Malaysia, highlighting their potential applications across multiple domains including transportation, law enforcement, and other specialized fields.

Conclusion: Car license plate recognition systems have significant potential in Malaysia for various practical applications and could be combined with techniques from other fields like biology and aerospace to solve specialized problems.

Abstract: Car license plate recognition system is an image processing technology used to identify vehicles by capturing their Car License Plates. The car license plate recognition technology is also known as automatic number-plate recognition, automatic vehicle identification, car license plate recognition or optical character recognition for cars. In Malaysia, as the number of vehicle is increasing rapidly nowadays, a pretty great number of vehicle on the road has brought about the considerable demands of car license plate recognition system. Car license plate recognition system can be implemented in electronic parking payment system, highway toll-fee system, traffic surveillance system and as police enforcement tools. Additionally, car license plate recognition system technology also has potential to be combined with various techniques in other different fields like biology, aerospace and so on to achieve the goal of solving some specialized problems.

Method: GrapHist integrates masked autoencoders with heterophilic graph neural networks designed to capture tumor microenvironment heterogeneity. Pre-trained on 11 million cell graphs from breast tissues using self-supervised learning.

Result: Achieves competitive performance compared to vision-based counterparts in slide-, region-, and cell-level tasks with 4x fewer parameters. Drastically outperforms fully-supervised graph models on cancer subtyping tasks. Released five graph-based digital pathology datasets.

Conclusion: Graph-based self-supervised learning with biologically-informed cell graph modeling offers efficient representation learning for histopathology, enabling diverse downstream tasks with fewer parameters while establishing the first large-scale graph benchmark in digital pathology.

Abstract: Self-supervised vision models have achieved notable success in digital pathology. However, their domain-agnostic transformer architectures are not originally designed to account for fundamental biological elements of histopathology images, namely cells and their complex interactions. In this work, we hypothesize that a biologically-informed modeling of tissues as cell graphs offers a more efficient representation learning. Thus, we introduce GrapHist, a novel graph-based self-supervised learning framework for histopathology, which learns generalizable and structurally-informed embeddings that enable diverse downstream tasks. GrapHist integrates masked autoencoders and heterophilic graph neural networks that are explicitly designed to capture the heterogeneity of tumor microenvironments. We pre-train GrapHist on a large collection of 11 million cell graphs derived from breast tissues and evaluate its transferability across in- and out-of-domain benchmarks. Our results show that GrapHist achieves competitive performance compared to its vision-based counterparts in slide-, region-, and cell-level tasks, while requiring four times fewer parameters. It also drastically outperforms fully-supervised graph models on cancer subtyping tasks. Finally, we also release five graph-based digital pathology datasets used in our study at https://huggingface.co/ogutsevda/datasets , establishing the first large-scale graph benchmark in this field. Our code is available at https://github.com/ogutsevda/graphist .

Method: Proposes a two-branch network with task-aware attention for feature mixing, multi-stage training framework, task-driven perceptual loss, and automatically constructs a Task-Inspired UIE Dataset using various task-specific networks.

Result: DTI-UIE significantly improves task performance by generating preprocessed images beneficial for downstream tasks including semantic segmentation, object detection, and instance segmentation.

Conclusion: The framework successfully bridges the gap between image enhancement and downstream task performance by focusing on task-relevant features rather than just human visual perception.

Abstract: In real underwater environments, downstream image recognition tasks such as semantic segmentation and object detection often face challenges posed by problems like blurring and color inconsistencies. Underwater image enhancement (UIE) has emerged as a promising preprocessing approach, aiming to improve the recognizability of targets in underwater images. However, most existing UIE methods mainly focus on enhancing images for human visual perception, frequently failing to reconstruct high-frequency details that are critical for task-specific recognition. To address this issue, we propose a Downstream Task-Inspired Underwater Image Enhancement (DTI-UIE) framework, which leverages human visual perception model to enhance images effectively for underwater vision tasks. Specifically, we design an efficient two-branch network with task-aware attention module for feature mixing. The network benefits from a multi-stage training framework and a task-driven perceptual loss. Additionally, inspired by human perception, we automatically construct a Task-Inspired UIE Dataset (TI-UIED) using various task-specific networks. Experimental results demonstrate that DTI-UIE significantly improves task performance by generating preprocessed images that are beneficial for downstream tasks such as semantic segmentation, object detection, and instance segmentation. The codes are publicly available at https://github.com/oucailab/DTIUIE.

Method: NERFIFY uses six key innovations: (1) Context-free grammar constraining LLM synthesis to Nerfstudio architecture, (2) Graph-of-Thought code synthesis with specialized multi-file agents, (3) Compositional citation recovery integrating components from reference graphs, (4) Visual feedback through PSNR-minima ROI analysis and VLM-guided patching, (5) Knowledge enhancement for method improvements, and (6) Benchmarking framework for evaluation across 30 diverse papers.

Result: On papers without public implementations, NERFIFY achieves visual quality matching expert human code (+/-0.5 dB PSNR, +/-0.2 SSIM) while reducing implementation time from weeks to minutes. The framework demonstrates reliable conversion of NeRF papers into trainable plugins.

Conclusion: NERFIFY demonstrates that domain-aware design enables code translation for complex vision papers, potentiating accelerated and democratized reproducible research in the NeRF domain. The approach shows promise for automating implementation of vision research papers.

Abstract: The proliferation of neural radiance field (NeRF) research requires significant efforts to reimplement papers before building upon them. We introduce NERFIFY, a multi-agent framework that reliably converts NeRF research papers into trainable Nerfstudio plugins, in contrast to generic paper-to-code methods and frontier models like GPT-5 that usually fail to produce runnable code. NERFIFY achieves domain-specific executability through six key innovations: (1) Context-free grammar (CFG): LLM synthesis is constrained by Nerfstudio formalized as a CFG, ensuring generated code satisfies architectural invariants. (2) Graph-of-Thought code synthesis: Specialized multi-file-agents generate repositories in topological dependency order, validating contracts and errors at each node. (3) Compositional citation recovery: Agents automatically retrieve and integrate components (samplers, encoders, proposal networks) from citation graphs of references. (4) Visual feedback: Artifacts are diagnosed through PSNR-minima ROI analysis, cross-view geometric validation, and VLM-guided patching to iteratively improve quality. (5) Knowledge enhancement: Beyond reproduction, methods can be improved with novel optimizations. (6) Benchmarking: An evaluation framework is designed for NeRF paper-to-code synthesis across 30 diverse papers. On papers without public implementations, NERFIFY achieves visual quality matching expert human code (+/-0.5 dB PSNR, +/-0.2 SSIM) while reducing implementation time from weeks to minutes. NERFIFY demonstrates that a domain-aware design enables code translation for complex vision papers, potentiating accelerated and democratized reproducible research. Code, data and implementations will be publicly released.

Method: Proposes DHVAE with CoTransformer module to disentangle global interaction context and individual motion patterns into decoupled latent structure. Uses contrastive learning constraints for physically plausible latent space, and DDIM-based diffusion denoising in hierarchical latent space with skip-connected AdaLN-Transformer denoiser.

Result: DHVAE achieves superior motion fidelity, text alignment, and physical plausibility with greater computational efficiency compared to existing methods.

Conclusion: The disentangled hierarchical approach with latent diffusion effectively addresses limitations in 3D HHI generation, producing more realistic and physically plausible human-human interactions.

Abstract: Generating realistic 3D Human-Human Interaction (HHI) requires coherent modeling of the physical plausibility of the agents and their interaction semantics. Existing methods compress all motion information into a single latent representation, limiting their ability to capture fine-grained actions and inter-agent interactions. This often leads to semantic misalignment and physically implausible artifacts, such as penetration or missed contact. We propose Disentangled Hierarchical Variational Autoencoder (DHVAE) based latent diffusion for structured and controllable HHI generation. DHVAE explicitly disentangles the global interaction context and individual motion patterns into a decoupled latent structure by employing a CoTransformer module. To mitigate implausible and physically inconsistent contacts in HHI, we incorporate contrastive learning constraints with our DHVAE to promote a more discriminative and physically plausible latent interaction space. For high-fidelity interaction synthesis, DHVAE employs a DDIM-based diffusion denoising process in the hierarchical latent space, enhanced by a skip-connected AdaLN-Transformer denoiser. Extensive evaluations show that DHVAE achieves superior motion fidelity, text alignment, and physical plausibility with greater computational efficiency.

Method: A systematic framework with: 1) Geometry+Semantics Data Factory using geometric feature clustering for balanced dataset synthesis, 2) Three-Stage Progressive Training: multi-task pre-alignment, specialized SFT for Markdown output, and Format-Constrained GRPO reinforcement learning for syntactic validity.

Result: Achieves state-of-the-art performance on OmniDocBench v1.5 with 92.94% overall score, outperforming DeepSeek-OCR 2 and OCRVerse across text, formula, table, and reading order metrics.

Conclusion: FireRed-OCR successfully specializes general VLMs into high-performance OCR models for structural document parsing, demonstrating the viability of the “General VLM to Specialized Structural Expert” paradigm.

Abstract: We present FireRed-OCR, a systematic framework to specialize general VLMs into high-performance OCR models. Large Vision-Language Models (VLMs) have demonstrated impressive general capabilities but frequently suffer from structural hallucination'' when processing complex documents, limiting their utility in industrial OCR applications. In this paper, we introduce FireRed-OCR, a novel framework designed to transform general-purpose VLMs (based on Qwen3-VL) into pixel-precise structural document parsing experts. To address the scarcity of high-quality structured data, we construct a Geometry + Semantics’’ Data Factory. Unlike traditional random sampling, our pipeline leverages geometric feature clustering and multi-dimensional tagging to synthesize and curate a highly balanced dataset, effectively handling long-tail layouts and rare document types. Furthermore, we propose a Three-Stage Progressive Training strategy that guides the model from pixel-level perception to logical structure generation. This curriculum includes: (1) Multi-task Pre-alignment to ground the model’s understanding of document structure; (2) Specialized SFT for standardizing full-image Markdown output; and (3) Format-Constrained Group Relative Policy Optimization (GRPO), which utilizes reinforcement learning to enforce strict syntactic validity and structural integrity (e.g., table closure, formula syntax). Extensive evaluations on OmniDocBench v1.5 demonstrate that FireRed-OCR achieves state-of-the-art performance with an overall score of 92.94%, significantly outperforming strong baselines such as DeepSeek-OCR 2 and OCRVerse across text, formula, table, and reading order metrics. We open-source our code and model weights to facilitate the ``General VLM to Specialized Structural Expert’’ paradigm.

Method: TP-Blend uses two attention processors: 1) Cross-Attention Object Fusion (CAOF) averages head-wise attention to locate spatial tokens, solves entropy-regularized optimal transport to reassign multi-head feature vectors, preserving cross-head correlations. 2) Self-Attention Style Fusion (SASF) injects style through Detail-Sensitive Instance Normalization using Gaussian filtering to separate low/high frequencies, and swaps Key/Value matrices with style prompt.

Result: Extensive experiments show TP-Blend produces high-resolution, photo-realistic edits with precise control over content and appearance, surpassing recent baselines in quantitative fidelity, perceptual quality, and inference speed.

Conclusion: TP-Blend provides an effective lightweight training-free framework for simultaneous object replacement and style transfer in diffusion models, achieving superior performance compared to existing methods.

Abstract: Current text-conditioned diffusion editors handle single object replacement well but struggle when a new object and a new style must be introduced simultaneously. We present Twin-Prompt Attention Blend (TP-Blend), a lightweight training-free framework that receives two separate textual prompts, one specifying a blend object and the other defining a target style, and injects both into a single denoising trajectory. TP-Blend is driven by two complementary attention processors. Cross-Attention Object Fusion (CAOF) first averages head-wise attention to locate spatial tokens that respond strongly to either prompt, then solves an entropy-regularised optimal transport problem that reassigns complete multi-head feature vectors to those positions. CAOF updates feature vectors at the full combined dimensionality of all heads (e.g., 640 dimensions in SD-XL), preserving rich cross-head correlations while keeping memory low. Self-Attention Style Fusion (SASF) injects style at every self-attention layer through Detail-Sensitive Instance Normalization. A lightweight one-dimensional Gaussian filter separates low- and high-frequency components; only the high-frequency residual is blended back, imprinting brush-stroke-level texture without disrupting global geometry. SASF further swaps the Key and Value matrices with those derived from the style prompt, enforcing context-aware texture modulation that remains independent of object fusion. Extensive experiments show that TP-Blend produces high-resolution, photo-realistic edits with precise control over both content and appearance, surpassing recent baselines in quantitative fidelity, perceptual quality, and inference speed.

Method: Adapts 3D Gaussian Splatting to MRI reconstruction with three key contributions: (1) Magnetic Gaussian primitives with physics-consistent volumetric rendering, (2) neural residual field for high-frequency detail refinement, and (3) multi-resolution progressive training.

Result: Achieves 40.31 dB PSNR on the FeTA dataset while being 14 times faster than previous methods, representing the first successful adaptation of 3D Gaussian Splatting to multi-stack MRI reconstruction.

Conclusion: M-Gaussian achieves an optimal balance between quality and speed for MRI reconstruction, successfully adapting 3D Gaussian Splatting to medical imaging with physics-consistent rendering and efficient training.

Abstract: Magnetic Resonance Imaging (MRI) is a crucial non-invasive imaging modality. In routine clinical practice, multi-stack thick-slice acquisitions are widely used to reduce scan time and motion sensitivity, particularly in challenging scenarios such as fetal brain imaging. However, the resulting severe through-plane anisotropy compromises volumetric analysis and downstream quantitative assessment, necessitating robust reconstruction of isotropic high-resolution volumes. Implicit neural representation methods, while achieving high quality, suffer from computational inefficiency due to complex network structures. We present M-Gaussian, adapting 3D Gaussian Splatting to MRI reconstruction. Our contributions include: (1) Magnetic Gaussian primitives with physics-consistent volumetric rendering, (2) neural residual field for high-frequency detail refinement, and (3) multi-resolution progressive training. Our method achieves an optimal balance between quality and speed. On the FeTA dataset, M-Gaussian achieves 40.31 dB PSNR while being 14 times faster, representing the first successful adaptation of 3D Gaussian Splatting to multi-stack MRI reconstruction.

Method: Extracts propeller rotational speed directly from raw event data and fuses it within an RPM-aware Kalman filtering framework for event-only drone forecasting.

Result: Outperforms learning-based approaches and vanilla Kalman filter on FRED dataset in terms of average distance error and final distance error at 0.4s and 0.8s forecasting horizons.

Conclusion: Demonstrates robust and accurate short- and medium-horizon trajectory forecasting without reliance on RGB imagery or training data by exploiting propeller motion cues from event cameras.

Abstract: Event cameras provide high-temporal-resolution visual sensing that is well suited for observing fast-moving aerial objects; however, their use for drone trajectory prediction remains limited. This work introduces an event-only drone forecasting method that exploits propeller-induced motion cues. Propeller rotational speed are extracted directly from raw event data and fused within an RPM-aware Kalman filtering framework. Evaluations on the FRED dataset show that the proposed method outperforms learning-based approaches and vanilla kalman filter in terms of average distance error and final distance error at 0.4s and 0.8s forecasting horizons. The results demonstrate robust and accurate short- and medium-horizon trajectory forecasting without reliance on RGB imagery or training data.

Method: Used PSF-Med dataset for systematic consistency evaluation, validated transfer of Gemma Scope 2 Sparse Autoencoders to MedGemma, then fine-tuned Low-Rank Adaptation (LoRA) adapters with a combined loss balancing paraphrase consistency with answer accuracy to prevent mode collapse.

Result: Reduced flip rate from 14.6% to 4.4% (p=0.002) and margin difference from 1.63 to 0.33 logits (79.5% reduction) on MIMIC-CXR while maintaining accuracy (84.2% vs 82.3%). On PadChest Balanced, flip rate dropped from 13.6% to 7.8% and accuracy increased from 66.4% to 69.4%.

Conclusion: Combined consistency-accuracy training effectively reduces paraphrase inconsistency in medical VLMs without sacrificing accuracy, with early layers being more effective than middle layers for reducing margin differences.

Abstract: Medical Vision-Language Models can give different yes or no answers to rephrasings of the same clinical question. We study this in MedGemma-4B using PSF-Med Sadanandan and Behzadan (2025), which provides paraphrase pairs for systematic consistency evaluation on medical VQA. On MIMIC-CXR binary questions (n = 158), the baseline flip rate is 14.6% and mean margin difference is 1.63 logits. We validate that Gemma Scope 2 Sparse Autoencoders (SAEs) transfer to MedGemma activations, achieving R2 ~= 0.997 on both medical and general text (n = 100 prompts each, p < 0.001 for exceeding a 0.95 threshold). We then fine-tune Low-Rank Adaptation (LoRA) adapters with a combined loss that balances paraphrase consistency with answer accuracy. This combined approach prevents mode collapse that occurs with pure consistency training while reducing flip rate from 14.6% to 4.4% (p = 0.002, two-proportion z-test) and margin difference from 1.63 to 0.33 (79.5% reduction). Accuracy remains stable at 84.2% baseline versus 82.3% after training (-1.9pp, not significant). On PadChest Balanced (n = 250), flip rate drops from 13.6% to 7.8%, mean margin difference drops from 1.08 to 0.35 (67.9% reduction), and accuracy increases from 66.4% to 69.4%. A layer-range ablation shows that early layers reduce margin differences more than mechanistically selected middle layers.

Method: ReMD performs multigrid residual correction at each reverse diffusion step, coupling data consistency with lightweight physics cues and correcting residuals across scales using a multi-wavelet basis to capture both large structures and fine vortical details. This coarse-to-fine design accelerates convergence while remaining equation-free.

Result: Across atmospheric and oceanic benchmarks, ReMD improves accuracy and spectral fidelity, reduces divergence, and reaches comparable quality with markedly fewer sampling steps than diffusion baselines.

Conclusion: Enforcing physics consistency inside the diffusion process via multigrid residual correction and multi-wavelet multiscale modeling is an effective route to efficient fluid super-resolution.

Abstract: Existing image SR and generic diffusion models transfer poorly to fluid SR: they are sampling-intensive, ignore physical constraints, and often yield spectral mismatch and spurious divergence. We address fluid super-resolution (SR) with \textbf{ReMD} (\underline{Re}sidual-\underline{M}ultigrid \underline{D}iffusion), a physics-consistent diffusion framework. At each reverse step, ReMD performs a \emph{multigrid residual correction}: the update direction is obtained by coupling data consistency with lightweight physics cues and then correcting the residual across scales; the multiscale hierarchy is instantiated with a \emph{multi-wavelet} basis to capture both large structures and fine vortical details. This coarse-to-fine design accelerates convergence and preserves fine structures while remaining equation-free. Across atmospheric and oceanic benchmarks, ReMD improves accuracy and spectral fidelity, reduces divergence, and reaches comparable quality with markedly fewer sampling steps than diffusion baselines. Our results show that enforcing physics consistency \emph{inside} the diffusion process via multigrid residual correction and multi-wavelet multiscale modeling is an effective route to efficient fluid SR. Our code are available on https://github.com/lizhihao2022/ReMD.

Method: Uses “anchors and shims” approach with inverse latents as anchors and finds shim perturbations that gradually deviate anchor latents. Applies perturbations to cross-attention mechanisms at different timesteps to induce semantic modifications that bypass copyright detection without requiring additional training.

Result: Experiments on MS-COCO and real-world copyrighted images show diffusion models can successfully replicate copyrighted images while evading watermark and copyright detection systems.

Conclusion: Neural plagiarism poses a critical threat to copyright protection, and the demonstrated vulnerabilities in current watermarking techniques highlight the urgent need for more robust countermeasures against data plagiarism by neural models.

Abstract: In this paper, we highlight a critical threat posed by emerging neural models: data plagiarism. We demonstrate how modern neural models (e.g., diffusion models) can replicate copyrighted images, even when protected by advanced watermarking techniques. To expose vulnerabilities in copyright protection and facilitate future research, we propose a general approach to neural plagiarism that can either forge replicas of copyrighted data or introduce copyright ambiguity. Our method, based on “anchors and shims”, employs inverse latents as anchors and finds shim perturbations that gradually deviate the anchor latents, thereby evading watermark or copyright detection. By applying perturbations to the cross-attention mechanism at different timesteps, our approach induces varying degrees of semantic modification in copyrighted images, enabling it to bypass protections ranging from visible trademarks and signatures to invisible watermarks. Notably, our method is a purely gradient-based search that requires no additional training or fine-tuning. Experiments on MS-COCO and real-world copyrighted images show that diffusion models can replicate copyrighted images, underscoring the urgent need for countermeasures against neural plagiarism.

Method: Proposes Dr.Seg, a plug-and-play GRPO-based framework with two key components: 1) Look-to-Confirm mechanism for broader output space, and 2) Distribution-Ranked Reward module for fine-grained, stable rewards. The framework requires no architectural modifications and integrates with existing GRPO-based VLLMs.

Result: Extensive experiments show Dr.Seg improves performance in complex visual scenarios while maintaining strong generalization. The framework demonstrates effectiveness in reasoning segmentation tasks.

Conclusion: Direct transfer of language reasoning training paradigms to visual perception is flawed due to intrinsic differences. Dr.Seg provides an effective solution by addressing the need for broader output space and fine-grained rewards in visual perception tasks.

Abstract: Following the success of Group Relative Policy Optimization (GRPO) in foundation LLMs, an increasing number of works have sought to adapt GRPO to Visual Large Language Models (VLLMs) for visual perception tasks (e.g., detection and segmentation). However, much of this line of research rests on a long-standing yet unexamined assumption: training paradigms developed for language reasoning can be transferred seamlessly to visual perception. Our experiments show that this assumption is not valid, revealing intrinsic differences between reasoning-oriented and perception-oriented settings. Using reasoning segmentation as a representative case, we surface two overlooked factors: (i) the need for a broader output space, and (ii) the importance of fine-grained, stable rewards. Building on these observations, we propose Dr.~Seg, a simple, plug-and-play GRPO-based framework consisting of a Look-to-Confirm mechanism and a Distribution-Ranked Reward module, requiring no architectural modifications and integrating seamlessly with existing GRPO-based VLLMs. Extensive experiments demonstrate that Dr.~Seg improves performance in complex visual scenarios while maintaining strong generalization. Code and models will be available at https://github.com/xVI-group-SCU/Dr-Seg.

Method: Three core innovations: 1) Semantic-aware Key Information Retrieval (SKIR) using semantic contribution graphs to model inter-segment relationships, 2) Visual-based Context Compression (VCC) that renders text segments into images to reduce token usage, and 3) Agentless Layout Violation Detection (ALVD) using deterministic color-gradient algorithms for layout verification.

Result: EfficientPosterGen achieves substantial improvements in token efficiency and layout reliability while maintaining high poster quality, offering a scalable solution for automated academic poster generation.

Conclusion: The framework addresses critical limitations of existing MLLM approaches and provides an efficient, reliable solution for academic poster generation with practical scalability.

Abstract: Automated academic poster generation aims to distill lengthy research papers into concise, visually coherent presentations. Existing Multimodal Large Language Models (MLLMs) based approaches, however, suffer from three critical limitations: low information density in full-paper inputs, excessive token consumption, and unreliable layout verification. We present EfficientPosterGen, an end-to-end framework that addresses these challenges through semantic-aware retrieval and token-efficient multimodal generation. EfficientPosterGen introduces three core innovations: (1) Semantic-aware Key Information Retrieval (SKIR), which constructs a semantic contribution graph to model inter-segment relationships and selectively preserves important content; (2) Visual-based Context Compression (VCC), which renders selected text segments into images to shift textual information into the visual modality, significantly reducing token usage while generating poster-ready bullet points; and (3) Agentless Layout Violation Detection (ALVD), a deterministic color-gradient-based algorithm that reliably detects content overflow and spatial sparsity without auxiliary MLLMs. Extensive experiments demonstrate that EfficientPosterGen achieves substantial improvements in token efficiency and layout reliability while maintaining high poster quality, offering a scalable solution for automated academic poster generation. Our code is available at https://github.com/vinsontang1/EfficientPosterGen-Code.

Method: BiCLIP features a bidirectional multimodal fusion mechanism where visual features iteratively refine textual representations for superior semantic alignment. It also implements an augmentation consistency objective that regularizes intermediate representations against perturbed input views to stabilize learning.

Result: Evaluation on QaTa-COV19 and MosMedData+ benchmarks shows BiCLIP consistently surpasses state-of-the-art image-only and multimodal baselines. It maintains high performance with only 1% labeled data and exhibits significant resistance to clinical artifacts like motion blur and low-dose CT noise.

Conclusion: BiCLIP provides a robust framework for medical image segmentation that effectively handles real-world clinical challenges including limited annotations and hardware-induced degradations through bidirectional multimodal fusion and consistency regularization.

Abstract: Medical image segmentation is a cornerstone of computer-assisted diagnosis and treatment planning. While recent multimodal vision-language models have shown promise in enhancing semantic understanding through textual descriptions, their resilience in “in-the-wild” clinical settings-characterized by scarce annotations and hardware-induced image degradations-remains under-explored. We introduce BiCLIP (Bidirectional and Consistent Language-Image Processing), a framework engineered to bolster robustness in medical segmentation. BiCLIP features a bidirectional multimodal fusion mechanism that enables visual features to iteratively refine textual representations, ensuring superior semantic alignment. To further stabilize learning, we implement an augmentation consistency objective that regularizes intermediate representations against perturbed input views. Evaluation on the QaTa-COV19 and MosMedData+ benchmarks demonstrates that BiCLIP consistently surpasses state-of-the-art image-only and multimodal baselines. Notably, BiCLIP maintains high performance when trained on as little as 1% of labeled data and exhibits significant resistance to clinical artifacts, including motion blur and low-dose CT noise.

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 2508.16479: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.16479&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Late-fusion multimodal framework combining YOLO-based image features from webcam imagery with structured meteorological data (concurrent and forecasted weather conditions). Classifies visibility into five categories using image-derived probabilities fused with numerical weather features.

Result: Achieved ACC ~0.89 for same-day prediction and up to 84% for next-day forecasts. Vision features dominate short-term horizons (“nowcasting” and “samedaycasting”), while weather data becomes primary predictive signal beyond +1 day. Late fusion consistently yields highest accuracy.

Conclusion: FujiView establishes Scenic Visibility Forecasting as a new benchmark task for multimodal learning, demonstrating effective fusion of visual and meteorological data. The framework and dataset (100,000+ webcam images from 40+ cameras) will be released to support environmental forecasting research.

Abstract: Visibility of natural landmarks such as Mount Fuji is a defining factor in both tourism planning and visitor experience, yet it remains difficult to predict due to rapidly changing atmospheric conditions. We present FujiView, a multimodal learning framework and dataset for predicting scenic visibility by fusing webcam imagery with structured meteorological data. Our late-fusion approach combines image-derived class probabilities with numerical weather features to classify visibility into five categories. The dataset currently comprises over 100,000 webcam images paired with concurrent and forecasted weather conditions from more than 40 cameras around Mount Fuji, and continues to expand; it will be released to support further research in environmental forecasting. Experiments show that YOLO-based vision features dominate short-term horizons such as “nowcasting” and “samedaycasting”, while weather-driven forecasts increasingly take over as the primary predictive signal beyond $+1$d. Late fusion consistently yields the highest overall accuracy, achieving ACC of approx 0.89 for same-day prediction and up to 84% for next-day forecasts. These results position Scenic Visibility Forecasting (SVF) as a new benchmark task for multimodal learning.

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 2508.16557: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2508.16557&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Collected 618,642 crop-weed images, filtered to 199,388 for fine-tuning DINOv3 ViT-small via sequential curation. Integrated fine-tuned backbone into YOLO26 as primary or dual-backbone with feature alignment loss for enhanced fusion.

Result: DINOv3-finetuned YOLO26-large achieved +5.4% mAP50 on in-domain 2025 images, +14.0% on 2021-2023 dataset, and +11.9% on 2024 dataset. Maintains real-time performance at ~28.5 fps despite 45.6% more parameters and 2.9x latency increase.

Conclusion: The approach successfully addresses dataset scarcity through self-supervised learning and heterogeneous data integration, creating a robust foundational model for crop-weed detection with strong generalization capabilities.

Abstract: Developing robust models for precision vegetable weeding is currently constrained by the scarcity of large-scale, annotated weed-crop datasets. To address this limitation, this study proposes a foundational crop-weed detection model by integrating heterogeneous datasets and leveraging self-supervised learning. A total of 618,642 crop-weed images were initially collected and subsequently refined to 199,388 filtered images for fine-tuning a DINOv3 vision transformer (ViT-small) through a sequential curation strategy. The fine-tuned DINOv3 backbone was then integrated into YOLO26, serving either as a primary backbone or part of a dual-backbone architecture. A feature alignment loss was introduced in the dual backbone framework to enhance feature fusion with minimal computational overhead. Experimental results show that the proposed DINOv3-finetuned ViT-small-based YOLO26-large achieved up to a +5.4% mAP50 gain on in-domain images collected in the 2025 season. Moreover, it demonstrated strong cross-domain generalization with mAP50 improvements of +14.0% on the 2021-2023 season dataset and +11.9% on the 2024 season dataset, compared to the standard YOLO26-large. Although the DINOv3-YOLO26-large model has 45.6% more parameters and a 2.9x increase in inference latency, it maintains real-time performance at ~28.5 frames per second (fps). The curated dataset and software programs developed in this study will be made publicly available.

Method: Developed a React/FastAPI web application with five complementary screening modules: (1) redness quantification via LAB a* color-space normalization, (2) blink-rate estimation using MediaPipe FaceMesh Eye Aspect Ratio with adaptive thresholding, (3) pupil light reflex characterization through Pupil-to-Iris Ratio time-series analysis, (4) scleral color indexing for icterus/anemia proxies via LAB/HSV statistics, and (5) iris-landmark-calibrated lesion encroachment measurement with millimeter-scale estimates.

Result: The system demonstrates that multi-signal ophthalmic screening is feasible on unmodified smartphones without cloud-based AI inference, providing deterministic, privacy-preserving algorithms for non-diagnostic consumer triage.

Conclusion: SKINOPATHY AI provides a foundation for future clinically validated mobile ophthalmoscopy tools that can operate entirely on commodity smartphones in low-resource settings.

Abstract: Early ophthalmic screening in low-resource and remote settings is constrained by access to specialized equipment and trained practitioners. We present SKINOPATHY AI, a smartphone-first web application that delivers five complementary, explainable screening modules entirely through commodity mobile hardware: (1) redness quantification via LAB a* color-space normalization; (2) blink-rate estimation using MediaPipe FaceMesh Eye Aspect Ratio (EAR) with adaptive thresholding; (3) pupil light reflex characterization through Pupil-to-Iris Ratio (PIR) time-series analysis; (4) scleral color indexing foricterus and anemia proxies via LAB/HSV statistics; and (5) iris-landmark-calibrated lesion encroachment measurement with millimeter-scale estimates and longitudinal trend tracking. The system is implemented as a React/FastAPI stack with OpenCV and MediaPipe, MongoDB-backed session persistence, and PDF report generation. All algorithms are fully deterministic, privacy-preserving, and designed for non-diagnostic consumer triage. We detail system architecture, algorithm design, evaluation methodology, clinical context, and ethical boundaries of the platform. SKINOPATHY AI demonstrates that multi-signal ophthalmic screening is feasible on unmodified smartphones without cloud-based AI inference, providing a foundation for future clinically validated mobile ophthalmoscopy tools.

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 2509.22240: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2509.22240&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Proposes an evaluation protocol combining region metrics (F1, IoU), boundary metrics (BF1, B-IoU), core/thin-subset equity analysis, and per-image robustness statistics with seeded multi-run training and non-parametric significance testing. Tests five losses on DeepLabV3-MobileNetV3 model with three training runs each on 12 held-out images split into core and thin subsets.

Result: Overlap-based losses improve F1 by over 20 points compared to cross-entropy (0.663 vs 0.438, p < 0.001). Boundary metrics confirm gains extend to contour precision. Classical baselines like Sauvola binarization achieve higher mean F1 (0.787) but worse worst-case performance (F1 = 0.452 vs 0.565 for Tversky), revealing a consistency-accuracy trade-off. Doubling training resolution increases F1 by 12.7 points.

Conclusion: Comprehensive evaluation reveals hidden trade-offs in whiteboard stroke segmentation. While classical methods achieve higher mean accuracy, learned models provide better worst-case reliability. The proposed evaluation protocol with boundary metrics and thin-stroke equity analysis is crucial for proper assessment of segmentation methods in imbalanced scenarios.

Abstract: The binary segmentation of whiteboard strokes is hindered by extreme class imbalance, caused by stroke pixels that constitute only $1.79%$ of the image on average, and in addition, the thin-stroke subset averages $1.14% \pm 0.41%$ in the foreground. Standard region metrics (F1, IoU) can mask thin-stroke failures because the vast majority of the background dominates the score. In contrast, adding boundary-aware metrics and a thin-subset equity analysis changes how loss functions rank and exposes hidden trade-offs. We contribute an evaluation protocol that jointly examines region metrics, boundary metrics (BF1, B-IoU), a core/thin-subset equity analysis, and per-image robustness statistics (median, IQR, worst-case) under seeded, multi-run training with non-parametric significance testing. Five losses – cross-entropy, focal, Dice, Dice+focal, and Tversky – are trained three times each on a DeepLabV3-MobileNetV3 model and evaluated on 12 held-out images split into core and thin subsets. Overlap-based losses improve F1 by more than 20 points over cross-entropy ($0.663$ vs $0.438$, $p < 0.001$). In addition, the boundary metrics confirm that the gain extends to the precision of the contour. Adaptive thresholding and Sauvola binarization at native resolution achieve a higher mean F1 ($0.787$ for Sauvola) but with substantially worse worst-case performance (F1 $= 0.452$ vs $0.565$ for Tversky), exposing a consistency-accuracy trade-off: classical baselines lead on mean F1 while the learned model delivers higher worst-case reliability. Doubling training resolution further increases F1 by 12.7 points.

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 2602.23557: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2602.23557&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: ConFoThinking framework learns to aggregate attention into a designated intermediate layer, mines salient regions from this consolidated attention, and zooms in on them. Uses concise semantic cues for attention extraction instead of question-based extraction.

Result: Experiments across five VQA benchmarks demonstrate significant improvement in perception performance compared to existing methods.

Conclusion: Consolidating fragmented attention and using concise semantic cues for attention extraction effectively improves MLLM performance on fine-grained visual question answering tasks.

Abstract: Thinking with Images improves fine-grained VQA for MLLMs by emphasizing visual cues. However, tool-augmented methods depend on the capacity of grounding, which remains unreliable for MLLMs. In parallel, attention-driven methods to crop the Region of Interest (ROIs) are proposed but they are constrained by (1) fragmented attention signals scattered across layers, leading to suboptimal localization and (2) relying on question- or redundant-text-conditioned attention extraction. Our analysis reveals three patterns: MLLMs may attend to the correct region yet generate incorrect coordinates, where-to-look attention is often fragmented across layers, and attention extraction is query-sensitive. Motivated by these, We propose ConFoThinking, a Consolidated-Focused-Attention-Driven Thinking framework that learns to aggregate attention into a designated intermediate layer, from which we mine and zoom in salient regions for downstream visual understanding. Moreover, we extract attention using concise semantic cues of what to look into, which mitigates the semantic noise introduced by question- or redundant-text-based attention extraction. Experiments across five VQA benchmarks demonstrate ConFoThinking significantly improves perception performance. The code, checkpoints, and dataset will be released after being accepted.

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 2307.14025: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2307.14025&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Formalize Obedience as instruction alignment capability, establish hierarchical grading system from basic semantic to pixel-level precision, conduct case studies to identify obedience gaps, and create VIOLIN benchmark for evaluating pure color generation across six variants.

Result: Extensive experiments on state-of-the-art models reveal fundamental obedience limitations, showing how generative priors often override logical constraints, with exploratory insights into model behavior.

Conclusion: Establishes framework for AI Obedience to draw attention to this gap and encourage deeper exploration to bridge the paradox between complex generation capabilities and simple task failures.

Abstract: Recent advances in generative AI have demonstrated remarkable ability to produce high-quality content. However, these models often exhibit “Paradox of Simplicity”: while they can render intricate landscapes, they often fail at simple, deterministic tasks. To address this, we formalize Obedience as the ability to align with instructions and establish a hierarchical grading system ranging from basic semantic alignment to pixel-level systemic precision, which provides a unified paradigm for incorporating and categorizing existing literature. Then, we conduct case studies to identify common obedience gaps, revealing how generative priors often override logical constraints. To evaluate high-level obedience, we present VIOLIN (VIsual Obedience Level-4 EvaluatIoN), the first benchmark focused on pure color generation across six variants. Extensive experiments on SOTA models reveal fundamental obedience limitations and further exploratory insights. By establishing this framework, we aim to draw more attention on AI Obedience and encourage deeper exploration to bridge this gap.

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 2411.10492: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2411.10492&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Used stereo camera to capture images of five olive species, preprocessed images, employed CNN architectures (MobileNetV2 and EfficientNetB0) with transfer learning optimization for image classification

Result: EfficientNetB0 model achieved optimal performance with 94.5% accuracy, demonstrating deep learning’s effectiveness for olive species classification

Conclusion: Deep learning-based systems provide effective solutions for agricultural product classification with high accuracy, showing significant potential for automatic identification and quality control applications

Abstract: In this study, image processing and deep learning methodologies were employed to automatically classify local olive species cultivated in Turkiye. A stereo camera was utilized to capture images of five distinct olive species, which were then preprocessed to ensure their suitability for analysis. Convolutional Neural Network (CNN) architectures, specifically MobileNetV2 and EfficientNetB0, were employed for image classification. These models were optimized through a transfer learning approach. The training and testing results indicated that the EfficientNetB0 model exhibited the optimal performance, with an accuracy of 94.5%. The findings demonstrate that deep learning-based systems offer an effective solution for classifying olive species with high accuracy. The developed method has significant potential for application in areas such as automatic identification and quality control of agricultural products.

Method: No method information available - paper content inaccessible due to HTTP 429 error

Result: No results available - API request was rate limited

Conclusion: Unable to analyze paper due to technical limitations in accessing the content

Abstract: Failed to fetch summary for 2504.06027: Page request resulted in HTTP 429 (https://export.arxiv.org/api/query?search_query=&id_list=2504.06027&sortBy=relevance&sortOrder=descending&start=0&max_results=100)

Method: Lilium uses a 3D cone-based representation to model soft-tissue variability, optimized via Differential Evolution algorithm. It enforces anatomical plausibility through constraints: landmark matching, camera parameter consistency, head pose alignment, skull containment within facial boundaries, and region parallelism.

Result: Lilium outperforms state-of-the-art methods in both accuracy and robustness for skull-face overlay in forensic identification.

Conclusion: The proposed evolutionary approach successfully addresses soft-tissue variability challenges in craniofacial superimposition, providing a more accurate and robust automated solution for forensic identification.

Abstract: Craniofacial Superimposition is a forensic technique for identifying skeletal remains by comparing a post-mortem skull with ante-mortem facial photographs. A critical step in this process is Skull-Face Overlay (SFO). This stage involves aligning a 3D skull model with a 2D facial image, typically guided by cranial and facial landmarks’ correspondence. However, its accuracy is undermined by individual variability in soft-tissue thickness, introducing significant uncertainty into the overlay. This paper introduces Lilium, an automated evolutionary method to enhance the accuracy and robustness of SFO. Lilium explicitly models soft-tissue variability using a 3D cone-based representation whose parameters are optimized via a Differential Evolution algorithm. The method enforces anatomical, morphological, and photographic plausibility through a combination of constraints: landmark matching, camera parameter consistency, head pose alignment, skull containment within facial boundaries, and region parallelism. This emulation of the usual forensic practitioners’ approach leads Lilium to outperform the state-of-the-art method in terms of both accuracy and robustness.

Method: AdaFollows a two-stage pipeline: (1) a confidence-based module that decides WHEN to crop based on the model’s confidence in its current visual understanding, and (2) a semantic-guided localization module that determines WHERE to crop by aligning semantic intent with spatial attention to accurately localize user-focused regions.

Result: AdaFocus delivers substantial performance gains while achieving approximately 4.0× speedup in inference speed compared to the state-of-the-art method ZoomEyes, representing significant advances in both accuracy and efficiency.

Conclusion: AdaFocus provides an effective training-free framework for adaptive visual reasoning in MLLMs that addresses key limitations of existing methods, offering improved performance and efficiency through intelligent, adaptive cropping strategies.

Abstract: Multimodal Large Language Models (MLLMs) are shifting towards “Thinking with Images” by actively exploring image details. While effective, large-scale training is computationally expensive, which has spurred growing interest in lightweight, training-free solutions. However, existing training-free methods suffer from two flaws: perceptual redundancy from indiscriminate cropping, which adds overhead and noise; and a drift between semantic intent and spatial attention, which prevents accurate localization of user-focused regions. To address these challenges, we propose AdaFocus, a novel training-free framework designed for adaptive visual reasoning. AdaFocus follows a two-stage pipeline: a confidence-based module decides when to crop, and a semantic-guided localization module determines where to crop. This enables adaptive visual reasoning without additional training. Experimentally, AdaFocus delivers substantial performance gains while achieving approximately 4.0\times speedup inference speedup than the SOTA method ZoomEyes, representing a significant advance in both accuracy and efficiency.

Method: Combined metadata-driven dataset curation, multi-stage filtering, μP (muP) parameterization, and hypersphere-constrained optimization. Developed Lavender Data system for dataset management and made inference-aware architectural choices. Trained on approximately 50 million video clips.

Result: Successfully trained Summer-22B, a 22B parameter video foundation model. Key observations: dataset engineering consumed majority of effort, architectural variants showed smaller differences than expected, and μP hyperparameter transfer appeared effective even under geometric constraints.

Conclusion: The report provides valuable engineering insights for large-scale video model development, emphasizing the importance of dataset curation and sharing practical lessons that could benefit similar projects in the field.

Abstract: We describe our experience training Summer-22B, a video foundation model developed from scratch. This report documents the engineering challenges, design decisions, and lessons learned while scaling from raw footage collection to a functional model trained on approximately 50 million clips. We outline our approach combining metadata-driven dataset curation, multi-stage filtering, $μ$P parameterization, and hypersphere-constrained optimization. We developed the Lavender Data system for dataset management and adopted inference-aware architectural choices. We share observations on what worked in our setting: dataset engineering consumed the majority of effort, architectural variants showed smaller differences than we expected, and $μ$P hyperparameter transfer appeared effective even under geometric constraints. We hope this account proves useful to others undertaking similar projects.

Method: Introduces Infinite Self-Attention (InfSA) as a spectral reformulation treating attention layers as diffusion steps on content-adaptive token graphs, with Linear-InfSA variant approximating the principal eigenvector without forming full attention matrices.

Result: Achieves 84.7% top-1 on ImageNet-1K (+3.2 points over softmax ViT), runs at 231 images/s with 13x better throughput, and scales to 9216×9216 resolution without memory issues.

Conclusion: InfSA provides an efficient, scalable attention mechanism for vision transformers with strong performance, interpretability through graph centrality connections, and practical linear-time computation.

Abstract: The quadratic cost of softmax attention limits Transformer scalability in high-resolution vision. We introduce Infinite Self-Attention (InfSA), a spectral reformulation that treats each attention layer as a diffusion step on a content-adaptive token graph, accumulating multi-hop interactions through a discounted Neumann series over attention matrices. This links self-attention to classical graph centrality (Katz, PageRank, eigenvector centrality) for interpretable token weighting. We also show the Neumann kernel equals the fundamental matrix of an absorbing Markov chain, so a token’s centrality is its expected number of random-walk visits before absorption. We then propose Linear-InfSA, a linear-time variant that approximates the principal eigenvector of the implicit attention operator without forming the full attention matrix. It keeps an auxiliary state of fixed size proportional to per-head dimension dh (independent of sequence length N), is drop-in compatible with Vision Transformers, and supports stable training at 4096 by 4096 and inference at 9216 by 9216 (about 332k tokens). In a 4-layer ViT (53.5M parameters, 59 GFLOPs at 224 by 224), Linear-InfSA reaches 84.7% top-1 on ImageNet-1K, a +3.2 point architectural gain over an equal-depth softmax ViT trained with the same recipe. On ImageNet-V2, InfViT variants outperform all compared baselines (up to 79.8% vs 76.8%), indicating robustness under distribution shift. On an A100 40GB GPU, Linear-InfViT runs at 231 images/s and 0.87 J/image (13x better throughput and energy than equal-depth ViT) and is the only tested model to complete 9216 by 9216 inference without out-of-memory. The linear approximation closely matches the dominant eigenvector of the quadratic operator (cosine 0.985).

Method: Two pipelines built on frozen SAM 2.1 backbone: (1) Zero-shot pipeline uses Grounding DINO 1.5 to detect birds via text prompt “bird” then prompts SAM 2.1 with bounding boxes; (2) Supervised pipeline fine-tunes YOLOv11 on CUB-200-2011 for high-precision detection then prompts SAM 2.1. Segmentation model is never retrained for new species/domains.

Result: On CUB-200-2011 (11,788 images, 200 species): Supervised pipeline achieves IoU 0.912, Dice 0.954, F1 0.953, outperforming SegFormer-B2 (IoU 0.842) by +7.0 percentage points. Zero-shot pipeline achieves IoU 0.831 using only text prompt, first such result reported on this benchmark.

Conclusion: Prompt-based foundation model pipelines outperform task-specific end-to-end trained segmentation networks while requiring only lightweight detector fine-tuning (~1 hour) for domain adaptation. Approach demonstrates strong generalization without segmentation model retraining.

Abstract: Bird image segmentation remains a challenging task in computer vision due to extreme pose diversity, complex plumage patterns, and variable lighting conditions. This paper presents a dual-pipeline framework for binary bird image segmentation leveraging 2025 foundation models. We introduce two operating modes built upon Segment Anything Model 2.1 (SAM 2.1) as a shared frozen backbone: (1) a zero-shot pipeline using Grounding DINO 1.5 to detect birds via the text prompt “bird” before prompting SAM 2.1 with bounding boxes requiring no labelled bird data; and (2) a supervised pipeline that fine-tunes YOLOv11 on the CUB-200-2011 dataset for high-precision detection, again prompting SAM 2.1 for pixel-level masks. The segmentation model is never retrained for new species or domains. On CUB-200-2011 (11,788 images, 200 species), the supervised pipeline achieves IoU 0.912, Dice 0.954, and F1 0.953 outperforming all prior baselines including SegFormer-B2 (IoU 0.842) by +7.0 percentage points. The zero-shot pipeline achieves IoU 0.831 using only a text prompt, the first such result reported on this benchmark. We demonstrate that prompt-based foundation model pipelines outperform task specific end-to-end trained segmentation networks, while requiring only lightweight detector fine-tuning (~1 hour) for domain adaptation. Complete PyTorch implementation, dataset preparation scripts, and trained weights are publicly available.

Method: Proposes ST-Lite with dual-branch scoring: Component-centric Spatial Saliency (CSS) preserves UI element structural integrity by evaluating local neighborhood saliency, and Trajectory-aware Semantic Gating (TSG) filters visually repetitive KV pairs within interaction trajectories.

Result: With only 10-20% cache budget, ST-Lite achieves 2.45x decoding acceleration while maintaining comparable or superior performance to full-cache baselines.

Conclusion: ST-Lite offers a scalable, training-free solution for efficient GUI agents by addressing GUI-specific attention patterns and dynamic spatio-trajectory dependencies.

Abstract: Large Vision-Language Models (VLMs) have emerged as powerful engines for autonomous GUI agents, yet their deployment is severely constrained by the substantial memory footprint and latency of the Key-Value (KV) cache during long-horizon interactions. While existing cache compression methods have proven effective for LLMs, we empirically demonstrate that they suffer from suboptimal performance in GUI scenarios due to a fundamental misalignment: unlike general visual tasks where attention sparsity varies across layers, GUI attention patterns exhibit uniform high-sparsity across all transformer layers. Motivated by this insight, we propose ST-Lite, a training-free KV cache compression framework tailored for efficient GUI agents that explicitly addresses the dynamic spatio-trajectory dependencies within GUI data streams. ST-Lite introduces a novel dual-branch scoring policy incorporating Component-centric Spatial Saliency (CSS) and Trajectory-aware Semantic Gating (TSG). Specifically, CSS preserves the structural integrity of interactive UI elements by evaluating local neighborhood saliency, while TSG mitigates historical redundancy by dynamically filtering visually repetitive KV pairs within the interaction trajectory. Extensive evaluations demonstrate that with only a 10-20% cache budget, ST-Lite achieves a 2.45x decoding acceleration while maintaining comparable or even superior performance compared to full-cache baselines, offering a scalable solution for resource-constrained GUI agents.

Method: SKeDA has two components: 1) Shuffle-Key-based Distribution-preserving Sampling (SKe) uses a single base pseudo-random binary sequence for watermark encryption and derives frame-level sequences through permutation, making extraction tolerant to frame reordering. 2) Differential Attention (DA) computes inter-frame differences and dynamically adjusts attention weights during extraction to handle temporal distortions.

Result: Extensive experiments show that SKeDA preserves high video generation quality while maintaining strong watermark robustness against various attacks and distortions.

Conclusion: SKeDA provides an effective watermarking solution for text-to-video diffusion models that addresses the unique challenges of video content, offering improved robustness against frame-level and temporal distortions while maintaining generation quality.

Abstract: The rise of text-to-video generation models has raised growing concerns over content authenticity, copyright protection, and malicious misuse. Watermarking serves as an effective mechanism for regulating such AI-generated content, where high fidelity and strong robustness are particularly critical. Recent generative image watermarking methods provide a promising foundation by leveraging watermark information and pseudo-random keys to control the initial sampling noise, enabling lossless embedding. However, directly extending these techniques to videos introduces two key limitations: Existing designs implicitly rely on strict alignment between video frames and frame-dependent pseudo-random binary sequences used for watermark encryption. Once this alignment is disrupted, subsequent watermark extraction becomes unreliable; and Video-specific distortions, such as inter-frame compression, significantly degrade watermark reliability. To address these issues, we propose SKeDA, a generative watermarking framework tailored for text-to-video diffusion models. SKeDA consists of two components: (1) Shuffle-Key-based Distribution-preserving Sampling (SKe) employs a single base pseudo-random binary sequence for watermark encryption and derives frame-level encryption sequences through permutation. This design transforms watermark extraction from synchronization-sensitive sequence decoding into permutation-tolerant set-level aggregation, substantially improving robustness against frame reordering and loss; and (2) Differential Attention (DA), which computes inter-frame differences and dynamically adjusts attention weights during extraction, enhancing robustness against temporal distortions. Extensive experiments demonstrate that SKeDA preserves high video generation quality and watermark robustness.

Method: Applied the same Concept Induction workflow to SUN2012 dataset, assigning interpretable semantic labels to neurons and validating them through web-sourced images and statistical testing.

Result: The method successfully transfers to SUN2012, confirming its broader applicability for analyzing hidden neuron semantics in DNNs across different scene recognition datasets.

Conclusion: The Concept Induction framework for hidden neuron analysis generalizes well to SUN2012, demonstrating its robustness and broader applicability for understanding DNN internal representations.

Abstract: Deep Neural Networks (DNNs) have advanced applications in domains such as healthcare, autonomous systems, and scene understanding, yet the internal semantics of their hidden neurons remain poorly understood. Prior work introduced a Concept Induction-based framework for hidden neuron analysis and demonstrated its effectiveness on the ADE20K dataset. In this case study, we investigate whether the approach generalizes by applying it to the SUN2012 dataset, a large-scale scene recognition benchmark. Using the same workflow, we assign interpretable semantic labels to neurons and validate them through web-sourced images and statistical testing. Our findings confirm that the method transfers to SUN2012, showing its broader applicability.

Method: Analyzes token reduction behavior through layerwise sparsity and importance stability. Proposes low-to-high progressive reduction schedule and unified language-aware scoring mechanism for both attention and Mamba blocks (using implicit-attention proxy for Mamba). Enables all-layer token reduction in hybrid architectures.

Result: Under aggressive compression (retaining 25% of visual tokens), achieves 3.8-4.2x prefilling speedups with near-baseline accuracy. Light finetuning under reduction further improves performance on long-context video benchmarks.

Conclusion: Progressive token reduction with unified scoring effectively accelerates hybrid video VLMs while maintaining accuracy, addressing challenges of token importance instability across layers in mixed attention-Mamba architectures.

Abstract: Token reduction is an effective way to accelerate long-video vision-language models (VLMs), but most existing methods are designed for dense Transformers and do not directly account for hybrid architectures that interleave attention with linear-time state-space blocks (e.g., Mamba). We study query-conditioned token reduction for hybrid video VLMs and analyze reduction behavior through two properties: layerwise sparsity (how many tokens capture query-relevant information) and importance stability (whether token-importance rankings persist across depth). Although token importance is sparse within each layer, the set of important tokens changes across layers, so aggressive early pruning is unreliable. Motivated by this, we propose a low-to-high progressive reduction schedule and a unified language-aware scoring mechanism for both attention and Mamba blocks (using an implicit-attention proxy for Mamba), enabling all-layer token reduction in hybrids. Under an aggressive compression setting (retaining 25% of visual tokens), our approach delivers substantial prefilling speedups (3.8–4.2x) with near-baseline accuracy at test time, and light finetuning under reduction further improves performance on long-context video benchmarks.

Method: Proposes AdURA-Net with two key components: 1) Adaptive dilated convolution and multiscale deformable alignment integrated with DenseNet backbone to capture anatomical complexities, and 2) Dual Head Loss combining masked binary cross entropy with logit and Dirichlet evidential learning objective.

Result: The paper presents a framework for thoracic disease classification that can appropriately handle uncertain cases by expressing model uncertainty rather than forcing confident predictions when evidence is insufficient.

Conclusion: AdURA-Net provides a reliable uncertainty-aware framework for medical image classification that addresses the critical need for models to express uncertainty in clinical decision-making, particularly for high-risk applications.

Abstract: One of the common issues in clinical decision-making is the presence of uncertainty, which often arises due to ambiguity in radiology reports, which often reflect genuine diagnostic uncertainty or limitations of automated label extraction in various complex cases. Especially the case of multilabel datasets such as CheXpert, MIMIC-CXR, etc., which contain labels such as positive, negative, and uncertain. In clinical decision-making, the uncertain label plays a tricky role as the model should not be forced to provide a confident prediction in the absence of sufficient evidence. The ability of the model to say it does not understand whenever it is not confident is crucial, especially in the cases of clinical decision-making involving high risks. Here, we propose AdURA-Net, a geometry-driven adaptive uncertainty-aware framework for reliable thoracic disease classification. The key highlights of the proposed model are: a) Adaptive dilated convolution and multiscale deformable alignment coupled with the backbone Densenet architecture capturing the anatomical complexities of the medical images, and b) Dual Head Loss, which combines masked binary cross entropy with logit and a Dirichlet evidential learning objective.

Method: Created a benchmark with 10 tasks across 6 reasoning domains (spatial navigation, abstract pattern completion, causal simulation, logical constraint satisfaction, graph theory, topology). Provides dual-track evaluation: generative track where models produce solution images verified through deterministic computer-vision pipelines, and discriminative track with five-way multiple choice using structurally plausible near-miss distractors.

Result: Version 0.1.0 distributes 6,000 puzzles (108,000 PNG images across three resolutions) with fully deterministic seeded generation and reproducible verification. Dataset, generation code, and evaluation harness released under Apache 2.0 license on HuggingFace.

Conclusion: TACIT Benchmark provides a rigorous, objective framework for evaluating visual reasoning capabilities across diverse domains, addressing limitations of existing benchmarks through programmatic generation and deterministic verification.

Abstract: Existing visual reasoning benchmarks predominantly rely on natural language prompts, evaluate narrow reasoning modalities, or depend on subjective scoring procedures such as LLM-as-judge. We introduce the TACIT Benchmark, a programmatic visual reasoning benchmark comprising 10 tasks across 6 reasoning domains: spatial navigation, abstract pattern completion, causal simulation, logical constraint satisfaction, graph theory, and topology. The benchmark provides dual-track evaluation: a generative track in which models must produce solution images verified through deterministic computer-vision pipelines, and a discriminative track offering five-way multiple choice with structurally plausible near-miss distractors. Each distractor violates exactly one structural constraint, requiring models to reason about fine-grained visual differences rather than exploit superficial cues. Version 0.1.0 distributes 6,000 puzzles (108,000 PNG images across three resolutions) with fully deterministic seeded generation and reproducible verification. The dataset, generation code, and evaluation harness are released under the Apache 2.0 license on HuggingFace (DOI: 10.57967/hf/7904).

Method: VisRef actively guides reasoning by re-injecting a coreset of visual tokens that are semantically relevant to the reasoning context while remaining diverse and globally representative of the image. This enables more grounded multimodal reasoning without additional RL fine-tuning.

Result: Experiments on three visual reasoning benchmarks with state-of-the-art multimodal large reasoning models show that VisRef consistently outperforms existing test-time scaling approaches by up to 6.4% under fixed test-time compute budgets.

Conclusion: VisRef provides an effective framework for maintaining visual grounding during extended reasoning in multimodal models, offering performance improvements without the computational cost of RL-based approaches.

Abstract: Advances in large reasoning models have shown strong performance on complex reasoning tasks by scaling test-time compute through extended reasoning. However, recent studies observe that in vision-dependent tasks, extended textual reasoning at inference time can degrade performance as models progressively lose attention to visual tokens and increasingly rely on textual priors alone. To address this, prior works use reinforcement learning (RL)-based fine-tuning to route visual tokens or employ refocusing mechanisms during reasoning. While effective, these methods are computationally expensive, requiring large-scale data generation and policy optimization. To leverage the benefits of test-time compute without additional RL fine-tuning, we propose VisRef, a visually grounded test-time scaling framework. Our key idea is to actively guide the reasoning process by re-injecting a coreset of visual tokens that are semantically relevant to the reasoning context while remaining diverse and globally representative of the image, enabling more grounded multi-modal reasoning. Experiments on three visual reasoning benchmarks with state-of-the-art multi-modal large reasoning models demonstrate that, under fixed test-time compute budgets, VisRef consistently outperforms existing test-time scaling approaches by up to 6.4%.

Method: Created CompGTSRB dataset by pasting GTSRB traffic signs onto undistorted AV backgrounds, trained YOLOv5 detector, generated patches using GAN with latent space optimization, and conducted physical experiments on Quanser QCar testbed with varying distances, patch sizes, and placements.

Result: NAPs successfully reduced the detector’s STOP class confidence across different configurations, demonstrating the effectiveness of physical adversarial patches and the utility of the CompGTSRB dataset for credible evaluation.

Conclusion: The research shows adversarial patches can physically attack AV perception systems, highlights the need for defenses against localized patch corruption, and provides systematic protocols for evaluating physical adversarial attacks.

Abstract: This paper studies how well Naturalistic Adversarial Patches (NAPs) transfer to a physical traffic sign setting when the detector is trained on a customized dataset for an autonomous vehicle (AV) environment. We construct a composite dataset, CompGTSRB (which is customized dataset for AV environment), by pasting traffic sign instances from the German Traffic Sign Recognition Benchmark (GTSRB) onto undistorted backgrounds captured from the target platform. CompGTSRB is used to train a YOLOv5 model and generate patches using a Generative Adversarial Network (GAN) with latent space optimization, following existing NAP methods. We carried out a series of experiments on our Quanser QCar testbed utilizing the front CSI camera provided in QCar. Across configurations, NAPs reduce the detector’s STOP class confidence. Different configurations include distance, patch sizes, and patch placement. These results along with a detailed step-by-step methodology indicate the utility of CompGTSRB dataset and the proposed systematic physical protocols for credible patch evaluation. The research further motivate researching the defenses that address localized patch corruption in embedded perception pipelines.

Method: Associates each class with encoded mixed states and performs classification through a single POVM construction based on the Pretty Good Measurement (PGM), reformulating classification as discrimination of class-dependent density operators.

Result: PGM-based classifier is consistently competitive with and sometimes improves upon standard methods, performing especially well in NSCLC binary and three-class tasks, and remaining competitive in PCa risk stratification with clinically relevant sensitivity-specificity trade-offs.

Conclusion: Quantum-inspired PGM classification provides a viable multi-class approach for biomedical radiomics, demonstrating competitive performance and potential advantages in handling class overlap through geometric discrimination principles.

Abstract: We investigate a quantum-inspired approach to supervised multi-class classification based on the \emph{Pretty Good Measurement} (PGM), viewed as an operator-valued decision rule derived from quantum state discrimination. The method associates each class with an encoded mixed state and performs classification through a single POVM construction, thus providing a genuinely multi-class strategy without reduction to pairwise or one-vs-rest schemes. In this perspective, classification is reformulated as the discrimination of a finite ensemble of class-dependent density operators, with performance governed by the geometry induced by the encoding map and by the overlap structure among classes. To assess the practical scope of this framework, we apply the PGM-based classifier to two biomedical radiomics case studies: histopathological subtyping of non-small-cell lung carcinoma (NSCLC) and prostate cancer (PCa) risk stratification. The evaluation is conducted under protocols aligned with previously reported radiomics studies, enabling direct comparison with established classical baselines. The results show that the PGM-based classifier is consistently competitive and, in several settings, improves upon standard methods. In particular, the method performs especially well in the NSCLC binary and three-class tasks, while remaining competitive in the four-class case, where increased class overlap yields a more demanding discrimination geometry. In the PCa study, the PGM classifier remains close to the strongest ensemble baseline and exhibits clinically relevant sensitivity–specificity trade-offs across feature-selection scenarios.

Method: Proposes AP-PCO: a joint position-color optimization framework that simultaneously optimizes patch placement and color composition using a fitness function from model outputs. Introduces crossmodal color adaptation strategy to constrain patch appearance according to infrared grayscale characteristics while maintaining strong perturbations in visible domain, reducing cross-spectral saliency. Operates without internal model information (black-box attacks).

Result: Extensive experiments on visual-infrared dense prediction tasks show AP-PCO achieves consistently strong attack performance across multiple architectures, providing a practical benchmark for robustness evaluation in VI perception systems.

Conclusion: AP-PCO effectively addresses the challenges of multimodal adversarial attacks in visual-infrared dense prediction by jointly optimizing position and color while bridging spectral discrepancies, enabling effective black-box attacks on VI perception systems.

Abstract: Multimodal adversarial attacks for dense prediction remain largely underexplored. In particular, visual-infrared (VI) perception systems introduce unique challenges due to heterogeneous spectral characteristics and modality-specific intensity distributions. Existing adversarial patch methods are primarily designed for single-modal inputs and fail to account for crossspectral inconsistencies, leading to reduced attack effectiveness and poor stealthiness when applied to VI dense prediction models. To address these challenges, we propose a joint position-color optimization framework (AP-PCO) for generating adversarial patches in visual-infrared settings. The proposed method optimizes patch placement and color composition simultaneously using a fitness function derived from model outputs, enabling a single patch to perturb both visible and infrared modalities. To further bridge spectral discrepancies, we introduce a crossmodal color adaptation strategy that constrains patch appearance according to infrared grayscale characteristics while maintaining strong perturbations in the visible domain, thereby reducing cross-spectral saliency. The optimization procedure operates without requiring internal model information, supporting flexible black-box attacks. Extensive experiments on visual-infrared dense prediction tasks demonstrate that the proposed AP-PCO achieves consistently strong attack performance across multiple architectures, providing a practical benchmark for robustness evaluation in VI perception systems.

Method: Two methods: 1) Quadspectral - uses four narrowband measurements (two at water vapor absorption line, two at ozone absorption line) for closed-form range estimation. 2) Hyperspectral - uses broader spectral range to estimate temperature, emissivity profiles, and downwelling contributions from multiple zenith angles.

Result: Experimental results show dramatic improvement: error reduced from over 100m (when reflected light not modeled) to 6.8m with quadspectral method and 1.2m with hyperspectral method.

Conclusion: Using ozone absorption features to model reflected downwelling radiance significantly improves passive LWIR ranging accuracy, with hyperspectral method achieving sub-meter precision.

Abstract: Passive long-wave infrared (LWIR) absorption-based ranging relies on atmospheric absorption to estimate distances to objects from their emitted thermal radiation. First demonstrated decades ago for objects much hotter than the air and recently extended to scenes with low temperature variations, this ranging has depended on reflected radiance being negligible. Downwelling radiance is especially problematic, sometimes causing large inaccuracies. In two new ranging methods, we use characteristic features from ozone absorption to estimate the contribution of reflected downwelling radiance. The quadspectral method gives a simple closed-form range estimate from four narrowband measurements, two at a water vapor absorption line and two at an ozone absorption line. The hyperspectral method uses a broader spectral range to improve accuracy while also providing estimates of temperature, emissivity profiles, and contributions of downwelling from a collection of zenith angles. Experimental results demonstrate improved ranging accuracy, in one case reducing error from over 100 m when reflected light is not modeled to 6.8 m with the quadspectral method and 1.2 m with the hyperspectral method.

Method: Extends PNS to multimodal scenarios by decomposing representations into modality-invariant and modality-specific components, then deriving tractable PNS objectives for each component. Addresses violations of PNS estimation conditions in multimodal settings.

Result: Experiments on synthetic and real-world medical datasets demonstrate the method’s effectiveness in learning necessary and sufficient multimodal features.

Conclusion: The proposed approach successfully extends PNS to multimodal medical AI, enabling learning of features that improve model performance and robustness to missing modalities.

Abstract: Learning multimodal representations from medical images and other data sources can provide richer information for decision-making. While various multimodal models have been developed for this, they overlook learning features that are both necessary (must be present for the outcome to occur) and sufficient (enough to determine the outcome). We argue learning such features is crucial as they can improve model performance by capturing essential predictive information, and enhance model robustness to missing modalities as each modality can provide adequate predictive signals. Such features can be learned by leveraging the Probability of Necessity and Sufficiency (PNS) as a learning objective, an approach that has proven effective in unimodal settings. However, extending PNS to multimodal scenarios remains underexplored and is non-trivial as key conditions of PNS estimation are violated. We address this by decomposing multimodal representations into modality-invariant and modality-specific components, then deriving tractable PNS objectives for each. Experiments on synthetic and real-world medical datasets demonstrate our method’s effectiveness. Code will be available on GitHub.

Method: PoP casts multimodal reasoning as an executable graph where each perception or logic node outputs a conformal set with calibrated uncertainty. A lightweight controller uses these certificates to allocate compute under a budget, expanding with extra tool calls only when needed and stopping early otherwise.

Result: Across document, chart, and multi-image QA benchmarks, PoP improves performance and reliability over strong chain-of-thought, ReAct-style, and program-of-thought baselines while using computation more efficiently.

Conclusion: PoP provides a principled approach to multimodal reasoning with explicit reliability guarantees, reducing error compounding and hallucinations while enabling verifiable evidence-based answers and efficient compute allocation.

Abstract: We present Proof-of-Perception (PoP), a tool-using framework that casts multimodal reasoning as an executable graph with explicit reliability guarantees. Each perception or logic node outputs a conformal set, yielding calibrated, stepwise uncertainty; a lightweight controller uses these certificates to allocate compute under a budget, expanding with extra tool calls only when needed and stopping early otherwise. This grounds answers in verifiable evidence, reduces error compounding and hallucinations, and enables principled accuracy-compute trade-offs. Across document, chart, and multi-image QA benchmarks, PoP improves performance and reliability over strong chain-of-thought, ReAct-style, and program-of-thought baselines while using computation more efficiently.

Method: Proposes a conditional diffusion framework with Structured Control Embedding Module (SCEM) that decomposes low-light images into four informative components: illumination, illumination-invariant features, shadow priors, and color-invariant cues. These serve as control signals for a U-Net-based diffusion model trained with simplified noise-prediction loss.

Result: Achieves state-of-the-art performance in quantitative and perceptual metrics on LOLv2-real, LSRW, DICM, MEF, and LIME datasets, demonstrating strong generalization despite being trained only on LOLv1 dataset.

Conclusion: The SCEM-equipped diffusion method effectively enforces structured enhancement guided by physical priors, showing superior performance and generalization across multiple low-light image enhancement benchmarks.

Abstract: Low-light images often suffer from low contrast, noise, and color distortion, degrading visual quality and impairing downstream vision tasks. We propose a novel conditional diffusion framework for low-light image enhancement that incorporates a Structured Control Embedding Module (SCEM). SCEM decomposes a low-light image into four informative components including illumination, illumination-invariant features, shadow priors, and color-invariant cues. These components serve as control signals that condition a U-Net-based diffusion model trained with a simplified noise-prediction loss. Thus, the proposed SCEM equipped Diffusion method enforces structured enhancement guided by physical priors. In experiments, our model is trained only on the LOLv1 dataset and evaluated without fine-tuning on LOLv2-real, LSRW, DICM, MEF, and LIME. The method achieves state-of-the-art performance in quantitative and perceptual metrics, demonstrating strong generalization across benchmarks. https://casted.github.io/scem/.

Method: Formulates electrode placement as constrained optimization in anatomical space, treating electrode coordinates as learnable parameters optimized end-to-end using a differentiable forward model of prosthetic vision. Incorporates vascular avoidance and gray matter feasibility constraints while minimizing task-level perceptual error.

Result: Percept-aware optimization consistently improves reconstruction fidelity relative to coverage-based strategies on simulated reading and natural image tasks using realistic folded cortical geometry. Vascular safety constraints eliminate margin violations while preserving perceptual performance, and enables co-optimization of multi-electrode thread configurations.

Conclusion: Differentiable percept models can inform anatomically grounded, safety-aware computer-assisted planning for cortical neural interfaces, providing a foundation for optimizing next-generation visual prostheses.

Abstract: Cortical visual prostheses aim to restore sight by electrically stimulating neurons in early visual cortex (V1). With the emergence of high-density and flexible neural interfaces, electrode placement within three-dimensional cortex has become a critical surgical planning problem. Existing strategies emphasize visual field coverage and anatomical heuristics but do not directly optimize predicted perceptual outcomes under safety constraints. We present a percept-aware framework for surgical planning of cortical visual prostheses that formulates electrode placement as a constrained optimization problem in anatomical space. Electrode coordinates are treated as learnable parameters and optimized end-to-end using a differentiable forward model of prosthetic vision. The objective minimizes task-level perceptual error while incorporating vascular avoidance and gray matter feasibility constraints. Evaluated on simulated reading and natural image tasks using realistic folded cortical geometry (FreeSurfer fsaverage), percept-aware optimization consistently improves reconstruction fidelity relative to coverage-based placement strategies. Importantly, vascular safety constraints eliminate margin violations while preserving perceptual performance. The framework further enables co-optimization of multi-electrode thread configurations under fixed insertion budgets. These results demonstrate how differentiable percept models can inform anatomically grounded, safety-aware computer-assisted planning for cortical neural interfaces and provide a foundation for optimizing next-generation visual prostheses.

Method: Generate pseudo labels via clustering on voxel values, train segmentation model on these labels, then use Unbiased Teacher approach to self-correct labels for accurate final segmentations.

Result: Improves pixel-wise accuracy by 13.31% and mIoU by 15.94% over baseline pseudo labels on magnesium crystal SR-CT sample. Framework works well on additional samples.

Conclusion: The framework enables automatic segmentation of large SR-CT datasets without manual annotation, addressing a major bottleneck in CT analysis workflows.

Abstract: X-ray computed tomography (CT) is a widely used imaging technique that provides detailed examinations into the internal structure of an object with synchrotron CT (SR-CT) enabling improved data quality by using higher energy, monochromatic X-rays. While SR-CT allows for improved resolution, time-resolved experimentation, and reduced imaging artifacts, it also produces significantly larger datasets than conventional CT. Accurate and efficient evaluation of these datasets is a critical component of these workflows; yet is often done manually representing a major bottleneck in the analysis phase. While deep learning has emerged as a powerful tool capable of providing a wide range of purely data-driven solutions, it requires a substantial amount of labeled data for training and manual annotation of SR-CT datasets is impractical in practice. In this paper, we introduce a novel framework that enables automatic segmentation of large, high-resolution SR-CT datasets by eliminating the need to hand label images for deep learning training. First, we generate pseudo labels by clustering on the voxel values identifying regions in the volume with similar attenuation coefficients producing an initial semantic map. Afterwards, we train a segmentation model on the pseudo labels before utilizing the Unbiased Teacher approach to self-correct them ensuring accurate final segmentations. We find our approach improves pixel-wise accuracy and mIoU by 13.31% and 15.94%, respectively, over the baseline pseudo labels when using a magnesium crystal SR-CT sample. Additionally, we extensively evaluate the different components of our workflow including segmentation model, loss function, pseudo labeling strategy, and input type. Finally, we evaluate our approach on to two additional samples highlighting our frameworks ability to produce segmentations that are considerably better than the original pseudo labels.

Method: Proposes DiffSOS, a conditional diffusion model with specialized acoustic ControlNet to ground denoising in physical wave measurements. Uses hybrid loss function integrating noise prediction, spatial reconstruction, and noise frequency content. Employs stochastic DDIM sampling for accelerated inference (10 steps).

Result: Significantly outperforms state-of-the-art networks on OpenPros USCT benchmark, achieving average Multi-scale Structural Similarity of 0.957. Provides high-fidelity SoS maps with pixel-wise uncertainty estimation and near real-time reconstruction.

Conclusion: DiffSOS enables high-fidelity speed-of-sound reconstruction with principled confidence measures, facilitating safer and faster clinical interpretation in ultrasound computed tomography.

Abstract: Accurate Speed-of-Sound (SoS) reconstruction from acoustic waveforms is a cornerstone of ultrasound computed tomography (USCT), enabling quantitative velocity mapping that reveals subtle anatomical details and pathological variations often invisible in conventional imaging. However, practical utility is hindered by the limitations of existing algorithms; traditional Full Waveform Inversion (FWI) is computationally intensive, while current deep learning approaches tend to produce oversmoothed results lacking fine details. We propose DiffSOS, a conditional diffusion model that directly maps acoustic waveforms to SoS maps. Our framework employs a specialized acoustic ControlNet to strictly ground the denoising process in physical wave measurements. To ensure structural consistency, we optimize a hybrid loss function that integrates noise prediction, spatial reconstruction, and noise frequency content. To accelerate inference, we employ stochastic Denoising Diffusion Implicit Model (DDIM) sampling, achieving near real-time reconstruction with only 10 steps. Crucially, we exploit the stochastic generative nature of our framework to estimate pixel-wise uncertainty, providing a measure of reliability that is often absent in deterministic approaches. Evaluated on the OpenPros USCT benchmark, DiffSOS significantly outperforms state-of-the-art networks, achieving an average Multi-scale Structural Similarity of 0.957. Our approach provides high-fidelity SoS maps with a principled measure of confidence, facilitating safer and faster clinical interpretation.

Method: Lightweight alignment anchors 3D geometric features to pre-aligned 2D visual semantics via cross-modal addition and token interleaving. Novel scene graph generation pipeline represents global layouts as chain of independent local triplets with relative coordinates, complemented by incremental generation algorithm.

Result: At 7B parameter scale, SSR achieves state-of-the-art performance on multiple spatial intelligence benchmarks, scoring 73.9 on VSI-Bench, outperforming much larger models.

Conclusion: Efficient feature alignment and structured scene reasoning are cornerstones of authentic spatial intelligence in multimodal models.

Abstract: While Multimodal Large Language Models (MLLMs) excel in semantic tasks, they frequently lack the “spatial sense” essential for sophisticated geometric reasoning. Current models typically suffer from exorbitant modality-alignment costs and deficiency in fine-grained structural modeling precision.We introduce SSR, a framework designed for Structured Scene Reasoning that seamlessly integrates 2D and 3D representations via a lightweight alignment mechanism. To minimize training overhead, our framework anchors 3D geometric features to the large language model’s pre-aligned 2D visual semantics through cross-modal addition and token interleaving, effectively obviating the necessity for large-scale alignment pre-training. To underpin complex spatial reasoning, we propose a novel scene graph generation pipeline that represents global layouts as a chain of independent local triplets defined by relative coordinates. This is complemented by an incremental generation algorithm, enabling the model to construct “language-model-friendly” structural scaffolds for complex environments. Furthermore, we extend these capabilities to global-scale 3D global grounding task, achieving absolute metric precision across heterogeneous data sources. At a 7B parameter scale, SSR achieves state-of-the-art performance on multiple spatial intelligence benchmarks, notably scoring 73.9 on VSI-Bench. Our approach significantly outperforms much larger models, demonstrating that efficient feature alignment and structured scene reasoning are the cornerstones of authentic spatial intelligence.

Method: Proposes PointAlign with explicit supervision of intermediate point cloud tokens to preserve 3D geometric-semantic information via consistency loss between point cloud tokens and visual input tokens, using lightweight alignment projector and LoRA adapters.

Result: Achieves 2.08pp average improvement on classification tasks, 7.50pp gain on open-vocabulary Objaverse classification, and 4.88pp improvement on 3D object captioning evaluated by Qwen2-72B-Instruct.

Conclusion: PointAlign effectively prevents geometric degradation in 3D VLMs with minimal computational overhead, demonstrating significant performance improvements on 3D classification and captioning tasks.

Abstract: The development of 3D Vision-Language Models (VLMs), crucial for applications in robotics, autonomous driving, and augmented reality, is severely constrained by the scarcity of paired 3D-text data. Existing methods rely solely on next-token prediction loss, using only language tokens for supervision. This results in inefficient utilization of limited 3D data and leads to a significant degradation and loss of valuable geometric information in intermediate representations. To address these limitations, we propose {\mname}, a novel feature-level alignment regularization method. {\mname} explicitly supervises intermediate point cloud tokens to preserve fine-grained 3D geometric-semantic information throughout the language modeling process. Specifically, we constrain the intermediate point cloud tokens within the LLM to align with visual input tokens via a consistency loss. By training only a lightweight alignment projector and LoRA adapters, {\mname} achieves explicit feature-level supervision with minimal computational overhead, effectively preventing geometric degradation. Extensive experiments on ModelNet40 and Objaverse datasets demonstrate that our method achieves \textbf{2.08} pp improvement on average for classification tasks, with a substantial \textbf{7.50} pp gain on the challenging open-vocabulary Objaverse classification task and \textbf{4.88} pp improvement on 3D object captioning evaluated by Qwen2-72B-Instruct, validating the effectiveness of {\mname}. Code is publicly available at \href{https://github.com/yharoldsu0627/PointAlign}{https://github.com/yharoldsu0627/PointAlign}.

Method: 1) Use FlexiCubes with dilation and smoothness regularization for initial geometry reconstruction from multi-view silhouettes; 2) Employ environment light radiance field for scene environment recovery; 3) Design recursive differentiable ray tracer to simultaneously optimize geometry, index of refraction, and absorption rate in end-to-end manner with CUDA acceleration.

Result: Superior reconstruction performance on multiple benchmarks, especially in intricate scenes with transparent objects having diverse topology and complex texture. Significantly reduced computational cost through CUDA implementation.

Conclusion: DiffTrans enables accurate reconstruction of transparent objects in complex real-world scenes through efficient decomposition and reconstruction of geometry and materials using differentiable rendering.

Abstract: Reconstructing transparent objects from a set of multi-view images is a challenging task due to the complicated nature and indeterminate behavior of light propagation. Typical methods are primarily tailored to specific scenarios, such as objects following a uniform topology, exhibiting ideal transparency and surface specular reflections, or with only surface materials, which substantially constrains their practical applicability in real-world settings. In this work, we propose a differentiable rendering framework for transparent objects, dubbed DiffTrans, which allows for efficient decomposition and reconstruction of the geometry and materials of transparent objects, thereby reconstructing transparent objects accurately in intricate scenes with diverse topology and complex texture. Specifically, we first utilize FlexiCubes with dilation and smoothness regularization as the iso-surface representation to reconstruct an initial geometry efficiently from the multi-view object silhouette. Meanwhile, we employ the environment light radiance field to recover the environment of the scene. Then we devise a recursive differentiable ray tracer to further optimize the geometry, index of refraction and absorption rate simultaneously in a unified and end-to-end manner, leading to high-quality reconstruction of transparent objects in intricate scenes. A prominent advantage of the designed ray tracer is that it can be implemented in CUDA, enabling a significantly reduced computational cost. Extensive experiments on multiple benchmarks demonstrate the superior reconstruction performance of our DiffTrans compared with other methods, especially in intricate scenes involving transparent objects with diverse topology and complex texture. The code is available at https://github.com/lcp29/DiffTrans.

Method: Proposes Query-Conditioned Gaussian Splatting (QCGS) framework that combines a radar point proposal network to identify rainfall-support locations with an implicit neural representation network that predicts Gaussian parameters for each point. Unlike conventional 2D Gaussian splatting, QCGS selectively renders only queried precipitation regions.

Result: QCGS demonstrates over 50% improvement in RMSE compared to conventional gridded precipitation products, maintains high performance across multiple spatiotemporal scales, and enables efficient, resolution-flexible precipitation field generation in real time.

Conclusion: QCGS effectively fuses weather station observations with satellite imagery for precipitation field generation, overcoming limitations of individual data sources while achieving significant accuracy improvements and computational efficiency.

Abstract: Precipitation forecasting relies on heterogeneous data. Weather radar is accurate, but coverage is geographically limited and costly to maintain. Weather stations provide accurate but sparse point measurements, while satellites offer dense, high-resolution coverage without direct rainfall retrieval. To overcome these limitations, we propose Query-Conditioned Gaussian Splatting (QCGS), the first framework to fuse automatic weather station (AWS) observations with satellite imagery for generating precipitation fields. Unlike conventional 2D Gaussian splatting, which renders the entire image plane, QCGS selectively renders only queried precipitation regions, avoiding unnecessary computation in non-precipitating areas while preserving sharp precipitation structures. The framework combines a radar point proposal network that identifies rainfall-support locations with an implicit neural representation (INR) network that predicts Gaussian parameters for each point. QCGS enables efficient, resolution-flexible precipitation field generation in real time. Through extensive evaluation with benchmark precipitation products, QCGS demonstrates over 50% improvement in RMSE compared to conventional gridded precipitation products, and consistently maintains high performance across multiple spatiotemporal scales.

Method: Proposes InstructX2X with Region-Specific Editing that restricts modifications to specific regions to prevent unintended changes, and provides a Guidance Map for inherently interpretable visual explanations of the editing process. Also introduces MIMIC-EDIT-INSTRUCTION dataset from expert-verified medical VQA pairs.

Result: Achieves state-of-the-art performance across all major evaluation metrics. Successfully generates high-quality counterfactual chest X-ray images along with interpretable explanations.

Conclusion: InstructX2X addresses fundamental limitations in counterfactual medical image generation by providing region-specific editing to prevent unintended modifications and offering interpretable visual explanations through Guidance Maps, enhancing utility for real-world medical applications.

Abstract: Counterfactual medical image generation have emerged as a critical tool for enhancing AI-driven systems in medical domain by answering “what-if” questions. However, existing approaches face two fundamental limitations: First, they fail to prevent unintended modifications, resulting collateral changes in demographic attributes when only disease features should be affected. Second, they lack interpretability in their editing process, which significantly limits their utility in real-world medical applications. To address these limitations, we present InstructX2X, a novel interpretable local editing model for counterfactual medical image generation featuring Region-Specific Editing. This approach restricts modifications to specific regions, effectively preventing unintended changes while simultaneously providing a Guidance Map that offers inherently interpretable visual explanations of the editing process. Additionally, we introduce MIMIC-EDIT-INSTRUCTION, a dataset for counterfactual medical image generation derived from expert-verified medical VQA pairs. Through extensive experiments, InstructX2X achieve state-of-the-art performance across all major evaluation metrics. Our model successfully generates high-quality counterfactual chest X-ray images along with interpretable explanations.

Method: Proposes Taxonomy-Aware Representation Alignment (TARA) that: 1) Aligns intermediate visual representations with biology foundation models (BFMs) that encode rich biological relationships through hierarchical contrastive learning, 2) Aligns first answer token representations with ground-truth labels to bridge visual features and categories of varying granularity.

Result: TARA consistently enhances LMMs’ hierarchical consistency and leaf node accuracy, enabling reliable recognition of both known and novel categories within complex biological taxonomies.

Conclusion: TARA effectively injects taxonomic knowledge into LMMs, improving their hierarchical visual recognition capabilities for biological categories, including novel ones with limited training data.

Abstract: A high-performing, general-purpose visual understanding model should map visual inputs to a taxonomic tree of labels, identify novel categories beyond the training set for which few or no publicly available images exist. Large Multimodal Models (LMMs) have achieved remarkable progress in fine-grained visual recognition (FGVR) for known categories. However, they remain limited in hierarchical visual recognition (HVR) that aims at predicting consistent label paths from coarse to fine categories, especially for novel categories. To tackle these challenges, we propose Taxonomy-Aware Representation Alignment (TARA), a simple yet effective strategy to inject taxonomic knowledge into LMMs. TARA leverages representations from biology foundation models (BFMs) that encode rich biological relationships through hierarchical contrastive learning. By aligning the intermediate representations of visual features with those of BFMs, LMMs are encouraged to extract discriminative visual cues well structured in the taxonomy tree. Additionally, we align the representations of the first answer token with the ground-truth label, flexibly bridging the gap between contextualized visual features and categories of varying granularity according to user intent. Experiments demonstrate that TARA consistently enhances LMMs’ hierarchical consistency and leaf node accuracy, enabling reliable recognition of both known and novel categories within complex biological taxonomies. Code is available at https://github.com/PKU-ICST-MIPL/TARA_CVPR2026.

Method: Proposes TAP-SLF framework with task-aware soft prompts to encode task-specific priors into input tokens, and applies LoRA (Low-Rank Adaptation) to selected top layers of the encoder. Updates only a small fraction of VFM parameters while keeping pre-trained backbone frozen.

Result: Achieved fifth place in FMC_UIA 2026 Challenge test set. Evaluations on officially released training dataset with 8:2 train-test split demonstrate effectiveness of task-aware prompting and selective layer tuning for efficient VFM adaptation.

Conclusion: Task-aware prompting combined with selective high-layer fine-tuning enables efficient adaptation of Vision Foundation Models to diverse medical tasks within a shared backbone, addressing overfitting and computational cost issues.

Abstract: Executing multiple tasks simultaneously in medical image analysis, including segmentation, classification, detection, and regression, often introduces significant challenges regarding model generalizability and the optimization of shared feature representations. While Vision Foundation Models (VFMs) provide powerful general representations, full fine-tuning on limited medical data is prone to overfitting and incurs high computational costs. Moreover, existing parameter-efficient fine-tuning approaches typically adopt task-agnostic adaptation protocols, overlooking both task-specific mechanisms and the varying sensitivity of model layers during fine-tuning. In this work, we propose Task-Aware Prompting and Selective Layer Fine-Tuning (TAP-SLF), a unified framework for multi-task ultrasound image analysis. TAP-SLF incorporates task-aware soft prompts to encode task-specific priors into the input token sequence and applies LoRA to selected specific top layers of the encoder. This strategy updates only a small fraction of the VFM parameters while keeping the pre-trained backbone frozen. By combining task-aware prompts with selective high-layer fine-tuning, TAP-SLF enables efficient VFM adaptation to diverse medical tasks within a shared backbone. Results on the FMC_UIA 2026 Challenge test set, where TAP-SLF wins fifth place, combined with evaluations on the officially released training dataset using an 8:2 train-test split, demonstrate that task-aware prompting and selective layer tuning are effective strategies for efficient VFM adaptation.

Method: Proposes Internal self-Correction mechanism utilizing Layer Attention (ICLA) that operates directly on hidden states during generation. Each layer selectively retrieves information from all preceding layers through a diagonal cross-layer attention mechanism, enabling self-refinement without external correction signals.

Result: With only 0.2M and 0.1M additional parameters on LLaVA1.5-7B and Qwen2.5-VL-7B respectively, ICLA consistently improves visual grounding across multiple hallucination benchmarks.

Conclusion: ICLA demonstrates effectiveness for more advanced LVLMs by providing an efficient internal self-correction mechanism that reduces hallucinations through layer attention-based refinement of hidden states.

Abstract: Although Large Vision-Language Models (LVLMs) have made substantial progress, hallucination, where generated text is not grounded in the visual input, remains a challenge. As LVLMs become stronger, previously reported hallucination patterns, such as linguistic bias and overthinking phenomenon, become far less consistent, making the corresponding mitigation techniques substantially less effective. In this paper, we introduce an Internal self-Correction mechanism utilizing Layer Attention (ICLA) that operates directly on hidden states during generation. Each layer selectively retrieves information from all preceding layers through a diagonal cross-layer attention mechanism, enabling self-refinement without any external correction signals. With introducing and training only 0.2M and 0.1M additional parameters on LLaVA1.5-7B and Qwen2.5-VL-7B, \ours consistently improves visual grounding across multiple hallucination benchmarks, demonstrating its effectiveness for more advanced LVLMs.

Method: Uses Mamba architecture encoder-decoder with CAD reconstruction pre-training to learn latent representations, then employs GAN to generate fake representations that are decoded back to parametric CAD sequences.

Result: Demonstrates effectiveness through comprehensive experiments with new dataset of 77,078 CAD models, showing improved generation length of valid parametric CAD sequences.

Conclusion: Mamba-CAD successfully addresses the challenge of modeling longer parametric CAD sequences for complex industrial CAD models through self-supervised learning and Mamba architecture.

Abstract: Computer-Aided Design (CAD) generative modeling has a strong and long-term application in the industry. Recently, the parametric CAD sequence as the design logic of an object has been widely mined by sequence models. However, the industrial CAD models, especially in component objects, are fine-grained and complex, requiring a longer parametric CAD sequence to define. To address the problem, we introduce Mamba-CAD, a self-supervised generative modeling for complex CAD models in the industry, which can model on a longer parametric CAD sequence. Specifically, we first design an encoder-decoder framework based on a Mamba architecture and pair it with a CAD reconstruction task for pre-training to model the latent representation of CAD models; and then we utilize the learned representation to guide a generative adversarial network to produce the fake representation of CAD models, which would be finally recovered into parametric CAD sequences via the decoder of MambaCAD. To train Mamba-CAD, we further create a new dataset consisting of 77,078 CAD models with longer parametric CAD sequences. Comprehensive experiments are conducted to demonstrate the effectiveness of our model under various evaluation metrics, especially in the generation length of valid parametric CAD sequences. The code and dataset can be achieved from https://github.com/Sunny-Hack/Code-for-Mamba-CAD-AAAI-2025-.

Method: 1) Semantic alignment: Chain-of-Thought inference extracts human behavior semantics from VLM-generated captions to suppress irrelevant details. 2) Structural alignment: Hierarchical structural fusion integrates multi-granularity structural info with hand structure attention enhancement for better hand-body alignment.

Result: Method outperforms prior work in generation performance and improves 3D hand reconstruction using generated hand images.

Conclusion: SesaHand effectively enhances controllable hand image generation through semantic and structural alignment, benefiting 3D hand reconstruction tasks.

Abstract: Recent studies on 3D hand reconstruction have demonstrated the effectiveness of synthetic training data to improve estimation performance. However, most methods rely on game engines to synthesize hand images, which often lack diversity in textures and environments, and fail to include crucial components like arms or interacting objects. Generative models are promising alternatives to generate diverse hand images, but still suffer from misalignment issues. In this paper, we present SesaHand, which enhances controllable hand image generation from both semantic and structural alignment perspectives for 3D hand reconstruction. Specifically, for semantic alignment, we propose a pipeline with Chain-of-Thought inference to extract human behavior semantics from image captions generated by the Vision-Language Model. This semantics suppresses human-irrelevant environmental details and ensures sufficient human-centric contexts for hand image generation. For structural alignment, we introduce hierarchical structural fusion to integrate structural information with different granularity for feature refinement to better align the hand and the overall human body in generated images. We further propose a hand structure attention enhancement method to efficiently enhance the model’s attention on hand regions. Experiments demonstrate that our method not only outperforms prior work in generation performance but also improves 3D hand reconstruction with the generated hand images.

Method: Distills a large 3D DiT teacher (DOVE) into a pruned 2D Stable Diffusion-based backbone augmented with lightweight 1D temporal convolutions. Introduces dual-head adversarial distillation with discriminators in both pixel and feature domains to disentangle detail and consistency optimization.

Result: The compressed AdcVSR model reduces complexity by 95% in parameters and achieves 8× acceleration over its DiT teacher DOVE while maintaining competitive video quality and efficiency.

Conclusion: The proposed improved ADC method effectively balances efficiency and quality for real-world video super-resolution by combining architectural innovations with specialized adversarial training.

Abstract: While many diffusion models have achieved impressive results in real-world video super-resolution (Real-VSR) by generating rich and realistic details, their reliance on multi-step sampling leads to slow inference. One-step networks like SeedVR2, DOVE, and DLoRAL alleviate this through condensing generation into one single step, yet they remain heavy, with billions of parameters and multi-second latency. Recent adversarial diffusion compression (ADC) offers a promising path via pruning and distilling these models into a compact AdcSR network, but directly applying it to Real-VSR fails to balance spatial details and temporal consistency due to its lack of temporal awareness and the limitations of standard adversarial learning. To address these challenges, we propose an improved ADC method for Real-VSR. Our approach distills a large diffusion Transformer (DiT) teacher DOVE equipped with 3D spatio-temporal attentions, into a pruned 2D Stable Diffusion (SD)-based AdcSR backbone, augmented with lightweight 1D temporal convolutions, achieving significantly higher efficiency. In addition, we introduce a dual-head adversarial distillation scheme, in which discriminators in both pixel and feature domains explicitly disentangle the discrimination of details and consistency into two heads, enabling both objectives to be effectively optimized without sacrificing one for the other. Experiments demonstrate that the resulting compressed AdcVSR model reduces complexity by 95% in parameters and achieves an 8$\times$ acceleration over its DiT teacher DOVE, while maintaining competitive video quality and efficiency.

Method: The framework uses Local Self-Similarity (LSS) mechanism to extract dense, illumination-invariant texture descriptors that disentangle necrotic tissue from background artifacts. It also employs a Liquid Time-Constant (LTC) refinement module that treats boundary evolution as an ODE-governed dynamic system, iteratively refining masks over continuous time-steps.

Result: On the MICCAI FUSeg dataset, LSS-LTCNet achieves state-of-the-art boundary alignment with a peak Dice score of 86.96% and exceptional 95th percentile Hausdorff Distance (HD95) of 8.91 pixels. With only 25.70M parameters, it outperforms heavier U-Net and transformer baselines in efficiency.

Conclusion: LSS-LTCNet offers a robust and transparent solution for computer-aided diagnosis in mobile healthcare settings by providing inherent visual audit trails alongside high-fidelity predictions, addressing both accuracy and explainability needs.

Abstract: Accurate semantic segmentation of foot ulcers is essential for automated wound monitoring, yet boundary delineation remains challenging due to tissue heterogeneity and poor contrast with surrounding skin. To overcome the limitations of standard intensity-based networks, we present LSS-LTCNet:an ante-hoc explainable framework synergizing deterministic structural priors with continuous-time neural dynamics. Our architecture departs from traditional black-box models by employing a Local Self-Similarity (LSS) mechanism that extracts dense, illumination-invariant texture descriptors to explicitly disentangle necrotic tissue from background artifacts. To enforce topological precision, we introduce a Liquid Time-Constant (LTC) refinement module that treats boundary evolution as an ODEgoverned dynamic system, iteratively refining masks over continuous time-steps. Comprehensive evaluation on the MICCAI FUSeg dataset demonstrates that LSS-LTCNet achieves state-of-the-art boundary alignment, securing a peak Dice score of 86.96% and an exceptional 95th percentile Hausdorff Distance (HD95) of 8.91 pixels. Requiring merely 25.70M parameters, the model significantly outperforms heavier U-Net and transformer baselines in efficiency. By providing inherent visual audit trails alongside high-fidelity predictions, LSS-LTCNet offers a robust and transparent solution for computer-aided diagnosis in mobile healthcare (mHealth) settings.

Method: ReMoT integrates two core components: 1) A rule-based automatic framework that generates ReMoT-16K, a large-scale motion-contrast dataset (16.5K triplets) from video meta-annotations, avoiding costly manual/model-based generation. 2) Group Relative Policy Optimization, which optimizes learning of contrastive reasoning more effectively than standard Supervised Fine-Tuning.

Result: The approach achieves state-of-the-art performance on the new fine-grained motion contrast benchmark and multiple standard VLM benchmarks, with a remarkable 25.1% performance improvement on spatio-temporal reasoning tasks.

Conclusion: ReMoT provides a systematic solution to improve VLMs’ spatio-temporal consistency through automated dataset generation and optimized training, enabling better motion understanding for real-world applications.

Abstract: We present ReMoT, a unified training paradigm to systematically address the fundamental shortcomings of VLMs in spatio-temporal consistency – a critical failure point in navigation, robotics, and autonomous driving. ReMoT integrates two core components: (1) A rule-based automatic framework that generates ReMoT-16K, a large-scale (16.5K triplets) motion-contrast dataset derived from video meta-annotations, surpassing costly manual or model-based generation. (2) Group Relative Policy Optimization, which we empirically validate yields optimal performance and data efficiency for learning this contrastive reasoning, far exceeding standard Supervised Fine-Tuning. We also construct the first benchmark for fine-grained motion contrast triplets to measure a VLM’s discrimination of subtle motion attributes (e.g., opposing directions). The resulting model achieves state-of-the-art performance on our new benchmark and multiple standard VLM benchmarks, culminating in a remarkable 25.1% performance leap on spatio-temporal reasoning tasks.

Method: OPGAgent has three main components: (1) Hierarchical Evidence Gathering module that decomposes analysis into global, quadrant, and tooth-level phases with dynamic tool invocation; (2) Specialized Toolbox with spatial, detection, utility, and expert zoos; (3) Consensus Subagent that resolves conflicts through anatomical constraints. Also introduces OPG-Bench, a structured-report protocol based on (Location, Field, Value) triples.

Result: OPGAgent outperforms current dental VLMs and medical agent frameworks on both the proposed OPG-Bench benchmark and the public MMOral-OPG benchmark, across both structured-report and VQA evaluation metrics.

Conclusion: The agentic approach with specialized tool coordination provides superior performance for dental OPG analysis compared to end-to-end VLMs, offering both versatility and accuracy while enabling comprehensive auditability of findings.

Abstract: Orthopantomograms (OPGs) are the standard panoramic radiograph in dentistry, used for full-arch screening across multiple diagnostic tasks. While Vision Language Models (VLMs) now allow multi-task OPG analysis through natural language, they underperform task-specific models on most individual tasks. Agentic systems that orchestrate specialized tools offer a path to both versatility and accuracy, this approach remains unexplored in the field of dental imaging. To address this gap, we propose OPGAgent, a multi-tool agentic system for auditable OPG interpretation. OPGAgent coordinates specialized perception modules with a consensus mechanism through three components: (1) a Hierarchical Evidence Gathering module that decomposes OPG analysis into global, quadrant, and tooth-level phases with dynamically invoking tools, (2) a Specialized Toolbox encapsulating spatial, detection, utility, and expert zoos, and (3) a Consensus Subagent that resolves conflicts through anatomical constraints. We further propose OPG-Bench, a structured-report protocol based on (Location, Field, Value) triples derived from real clinical reports, which enables a comprehensive review of findings and hallucinations, extending beyond the limitations of VQA indicators. On our OPG-Bench and the public MMOral-OPG benchmark, OPGAgent outperforms current dental VLMs and medical agent frameworks across both structured-report and VQA evaluation. Code will be released upon acceptance.

Method: Proposes DreamWorld with Joint World Modeling Paradigm that jointly predicts video pixels and features from foundation models to capture temporal dynamics, spatial geometry, and semantic consistency. Introduces Consistent Constraint Annealing (CCA) to progressively regulate world-level constraints during training, and Multi-Source Inner-Guidance to enforce learned world priors at inference.

Result: Extensive evaluations show DreamWorld improves world consistency, outperforming Wan2.1 by 2.26 points on VBench benchmark.

Conclusion: DreamWorld provides a unified framework for integrating complementary world knowledge into video generation, addressing the limitations of single-knowledge approaches through joint modeling of multiple world dimensions.

Abstract: Despite impressive progress in video generation, existing models remain limited to surface-level plausibility, lacking a coherent and unified understanding of the world. Prior approaches typically incorporate only a single form of world-related knowledge or rely on rigid alignment strategies to introduce additional knowledge. However, aligning the single world knowledge is insufficient to constitute a world model that requires jointly modeling multiple heterogeneous dimensions (e.g., physical commonsense, 3D and temporal consistency). To address this limitation, we introduce \textbf{DreamWorld}, a unified framework that integrates complementary world knowledge into video generators via a \textbf{Joint World Modeling Paradigm}, jointly predicting video pixels and features from foundation models to capture temporal dynamics, spatial geometry, and semantic consistency. However, naively optimizing these heterogeneous objectives can lead to visual instability and temporal flickering. To mitigate this issue, we propose \textit{Consistent Constraint Annealing (CCA)} to progressively regulate world-level constraints during training, and \textit{Multi-Source Inner-Guidance} to enforce learned world priors at inference. Extensive evaluations show that DreamWorld improves world consistency, outperforming Wan2.1 by 2.26 points on VBench. Code will be made publicly available at \href{https://github.com/ABU121111/DreamWorld}{\textcolor{mypink}{\textbf{Github}}}.

Method: 1) Hardware co-design: asymmetric dual-modality system integrating SVE micro-attenuation camera with event sensor. 2) Two-stage cross-modal alignment: feature-guided coarse homography estimation + multi-scale refinement with spatial pooling and frequency-domain filtering. 3) Cross-modal HDR reconstruction network: convolutional fusion, mutual-information regularization, and learnable fusion loss balancing intensity cues and event-derived structural constraints.

Result: Comprehensive experiments on synthetic benchmarks and real captures show consistent improvements in highlight recovery, edge fidelity, and robustness compared to frame-only or event-only HDR pipelines. The system effectively handles highly dynamic and radiometrically challenging environments.

Conclusion: Joint optimization of optical design, cross-modal alignment, and computational fusion provides an effective foundation for reliable HDR perception in challenging illumination conditions. The hardware-algorithm co-design approach enables superior performance over single-modality solutions.

Abstract: High dynamic range (HDR) imaging under extreme illumination remains challenging for conventional cameras due to overexposure. Event cameras provide microsecond temporal resolution and high dynamic range, while spatially varying exposure (SVE) sensors offer single-shot radiometric diversity.We present a hardware–algorithm co-designed HDR imaging system that tightly integrates an SVE micro-attenuation camera with an event sensor in an asymmetric dual-modality configuration. To handle non-coaxial geometry and heterogeneous optics, we develop a two-stage cross-modal alignment framework that combines feature-guided coarse homography estimation with a multi-scale refinement module based on spatial pooling and frequency-domain filtering. On top of aligned representations, we develop a cross-modal HDR reconstruction network with convolutional fusion, mutual-information regularization, and a learnable fusion loss that adaptively balances intensity cues and event-derived structural constraints. Comprehensive experiments on both synthetic benchmarks and real captures demonstrate that the proposed system consistently improves highlight recovery, edge fidelity, and robustness compared with frame-only or event-only HDR pipelines. The results indicate that jointly optimizing optical design, cross-modal alignment, and computational fusion provides an effective foundation for reliable HDR perception in highly dynamic and radiometrically challenging environments.

Method: Progressive training from segmentation to classification to report generation, plus multi-layer visual injection that routes U-Net encoder features to corresponding language model layers. Each stage can use different datasets without unified annotations.

Result: State-of-the-art performance on CT-RATE (F1: 0.414 vs 0.258, BLEU-mean: 0.349 vs 0.305) and AbdomenAtlas 3.0 (F1: 0.624 vs 0.518) using only a 0.1B decoder trained from scratch, showing well-designed vision encoder pretraining outweighs benefits of 7B+ pre-trained language models.

Conclusion: Hierarchical vision-language modeling with progressive training and multi-layer visual injection enables effective 3D medical report generation, demonstrating that specialized vision encoder design is more important than large pre-trained language models for this domain.

Abstract: Automated radiology report generation is key for reducing radiologist workload and improving diagnostic consistency, yet generating accurate reports for 3D medical imaging remains challenging. Existing vision-language models face two limitations: they do not leverage segmentation-pretrained encoders, and they inject visual features only at the input layer of language models, losing multi-scale information. We propose U-VLM, which enables hierarchical vision-language modeling in both training and architecture: (1) progressive training from segmentation to classification to report generation, and (2) multi-layer visual injection that routes U-Net encoder features to corresponding language model layers. Each training stage can leverage different datasets without unified annotations. U-VLM achieves state-of-the-art performance on CT-RATE (F1: 0.414 vs 0.258, BLEU-mean: 0.349 vs 0.305) and AbdomenAtlas 3.0 (F1: 0.624 vs 0.518 for segmentation-based detection) using only a 0.1B decoder trained from scratch, demonstrating that well-designed vision encoder pretraining outweighs the benefits of 7B+ pre-trained language models. Ablation studies show that progressive pretraining significantly improves F1, while multi-layer injection improves BLEU-mean. Code is available at https://github.com/yinghemedical/U-VLM.

Method: 1) Dual-visual tokenizer architecture processing both high- and low-resolution images; 2) Bilateral Attention Network (BAN) for intelligent fusion of multi-scale visual tokens; 3) Kolmogorov Arnold Network (KAN)-based modality projector with learnable activation functions for precise visual-to-text alignment; 4) Integration into multimodal multitask token communication system with joint VLM-channel coding.

Result: Experimental results validate TaiChi’s superior performance and demonstrate the feasibility and effectiveness of the TaiChi-driven token communication system.

Conclusion: TaiChi successfully addresses key limitations of VLMs in token communications through its novel architecture, achieving better visual understanding, compact token representation, and precise cross-modal alignment for intelligent communication systems.

Abstract: Visual-Language Models (VLMs), with their strong capabilities in image and text understanding, offer a solid foundation for intelligent communications. However, their effectiveness is constrained by limited token granularity, overlong visual token sequences, and inadequate cross-modal alignment. To overcome these challenges, we propose TaiChi, a novel VLM framework designed for token communications. TaiChi adopts a dual-visual tokenizer architecture that processes both high- and low-resolution images to collaboratively capture pixel-level details and global conceptual features. A Bilateral Attention Network (BAN) is introduced to intelligently fuse multi-scale visual tokens, thereby enhancing visual understanding and producing compact visual tokens. In addition, a Kolmogorov Arnold Network (KAN)-based modality projector with learnable activation functions is employed to achieve precise nonlinear alignment from visual features to the text semantic space, thus minimizing information loss. Finally, TaiChi is integrated into a multimodal and multitask token communication system equipped with a joint VLM-channel coding scheme. Experimental results validate the superior performance of TaiChi, as well as the feasibility and effectiveness of the TaiChi-driven token communication system.

Method: RAISE formulates image generation as requirement-driven adaptive scaling, evolving candidate images through refinement actions (prompt rewriting, noise resampling, instructional editing) and verifying each generation against structured requirement checklists to dynamically allocate computation where needed.

Result: Achieves state-of-the-art alignment (0.94 overall GenEval) while reducing generated samples by 30-40% and VLM calls by 80% compared to prior methods, demonstrating efficient and generalizable multi-round self-improvement.

Conclusion: RAISE provides an efficient, model-agnostic framework for adaptive text-to-image generation that aligns computational effort with semantic complexity without requiring training or fine-tuning.

Abstract: Recent text-to-image (T2I) diffusion models achieve remarkable realism, yet faithful prompt-image alignment remains challenging, particularly for complex prompts with multiple objects, relations, and fine-grained attributes. Existing training-free inference-time scaling methods rely on fixed iteration budgets that cannot adapt to prompt difficulty, while reflection-tuned models require carefully curated reflection datasets and extensive joint fine-tuning of diffusion and vision-language models, often overfitting to reflection paths data and lacking transferability across models. We introduce RAISE (Requirement-Adaptive Self-Improving Evolution), a training-free, requirement-driven evolutionary framework for adaptive T2I generation. RAISE formulates image generation as a requirement-driven adaptive scaling process, evolving a population of candidates at inference time through a diverse set of refinement actions-including prompt rewriting, noise resampling, and instructional editing. Each generation is verified against a structured checklist of requirements, enabling the system to dynamically identify unsatisfied items and allocate further computation only where needed. This achieves adaptive test-time scaling that aligns computational effort with semantic query complexity. On GenEval and DrawBench, RAISE attains state-of-the-art alignment (0.94 overall GenEval) while incurring fewer generated samples (reduced by 30-40%) and VLM calls (reduced by 80%) than prior scaling and reflection-tuned baselines, demonstrating efficient, generalizable, and model-agnostic multi-round self-improvement. Code is available at https://github.com/LiyaoJiang1998/RAISE.

Method: Proposes random grouping strategy for vision tokens - a simple and fast approach that randomly groups tokens for self-attention calculations, analyzed through multiple baselines and downstream tasks.

Result: Random grouping outperforms most carefully designed grouping methods on vision tasks, shows pronounced advantages in downstream tasks like object detection, and works across multiple modalities including point clouds and vision-language models.

Conclusion: Vision tokens only need an extremely simple grouping strategy (random grouping) that meets four key conditions: positional information, head feature diversity, global receptive field, and fixed grouping pattern, making complex designs unnecessary.

Abstract: Since Transformers are introduced into vision architectures, their quadratic complexity has always been a significant issue that many research efforts aim to address. A representative approach involves grouping tokens, performing self-attention calculations within each group, or pooling the tokens within each group into a single token. To this end, various carefully designed grouping strategies have been proposed to enhance the performance of Vision Transformers. Here, we pose the following questions: \textbf{Are these carefully designed grouping methods truly necessary? Is there a simpler and more unified token grouping method that can replace these diverse methods?} Therefore, we propose the random grouping strategy, which involves a simple and fast random grouping strategy for vision tokens. We validate this approach on multiple baselines, and experiments show that random grouping almost outperforms all other grouping methods. When transferred to downstream tasks, such as object detection, random grouping demonstrates even more pronounced advantages. In response to this phenomenon, we conduct a detailed analysis of the advantages of random grouping from multiple perspectives and identify several crucial elements for the design of grouping strategies: positional information, head feature diversity, global receptive field, and fixed grouping pattern. We demonstrate that as long as these four conditions are met, vision tokens require only an extremely simple grouping strategy to efficiently and effectively handle various visual tasks. We also validate the effectiveness of our proposed random method across multiple modalities, including visual tasks, point cloud processing, and vision-language models. Code will be available at https://github.com/qhfan/random.

Method: Two-stage pipeline: 1) Train bidirectional generative model with novel opacity mixing strategy for consistency with observations while extrapolating novel content; 2) Distill into causal auto-regressive model that generates hundreds of frames in single pass for novel view synthesis or pseudo-supervision.

Result: Outperforms all existing baselines by wide margin, exceeding prior state-of-the-art by 1-3 dB PSNR. Generates plausible reconstructions where existing approaches fail completely.

Conclusion: Proposed method solves scalability and quality issues in generative novel view synthesis, enabling consistent content generation in under-observed areas while maintaining high-quality results.

Abstract: Per-scene optimization methods such as 3D Gaussian Splatting provide state-of-the-art novel view synthesis quality but extrapolate poorly to under-observed areas. Methods that leverage generative priors to correct artifacts in these areas hold promise but currently suffer from two shortcomings. The first is scalability, as existing methods use image diffusion models or bidirectional video models that are limited in the number of views they can generate in a single pass (and thus require a costly iterative distillation process for consistency). The second is quality itself, as generators used in prior work tend to produce outputs that are inconsistent with existing scene content and fail entirely in completely unobserved regions. To solve these, we propose a two-stage pipeline that leverages two key insights. First, we train a powerful bidirectional generative model with a novel opacity mixing strategy that encourages consistency with existing observations while retaining the model’s ability to extrapolate novel content in unseen areas. Second, we distill it into a causal auto-regressive model that generates hundreds of frames in a single pass. This model can directly produce novel views or serve as pseudo-supervision to improve the underlying 3D representation in a simple and highly efficient manner. We evaluate our method extensively and demonstrate that it can generate plausible reconstructions in scenarios where existing approaches fail completely. When measured on commonly benchmarked datasets, we outperform existing all existing baselines by a wide margin, exceeding prior state-of-the-art methods by 1-3 dB PSNR.

Method: Proposes COG framework that formulates correspondence estimation as confidence-aware optimal transport problem. Predicts point-wise confidences and injects them as optimal transport marginals to suppress non-overlapping regions. Uses semantic priors from vision foundation models to regularize correspondences for stable pose estimation. Enables unsupervised learning by integrating confidence into correspondence finding and pose estimation pipeline.

Result: Unsupervised COG achieves comparable performance to supervised methods, and supervised COG outperforms existing supervised methods. The framework demonstrates robust pose estimation despite occlusions and viewpoint changes.

Conclusion: COG provides an effective unsupervised framework for 6DoF pose estimation by formulating correspondence as confidence-aware optimal transport, leveraging semantic priors, and achieving state-of-the-art performance in both unsupervised and supervised settings.

Abstract: Estimating the 6DoF pose of a novel object with a single reference view is challenging due to occlusions, view-point changes, and outliers. A core difficulty lies in finding robust cross-view correspondences, as existing methods often rely on discrete one-to-one matching that is non-differentiable and tends to collapse onto sparse key-points. We propose Confidence-aware Optimal Geometric Correspondence (COG), an unsupervised framework that formulates correspondence estimation as a confidence-aware optimal transport problem. COG produces balanced soft correspondences by predicting point-wise confidences and injecting them as optimal transport marginals, suppressing non-overlapping regions. Semantic priors from vision foundation models further regularize the correspondences, leading to stable pose estimation. This design integrates confidence into the correspondence finding and pose estimation pipeline, enabling unsupervised learning. Experiments show unsupervised COG achieves comparable performance to supervised methods, and supervised COG outperforms them.

Method: Proposes M² framework with dual-tier memory: 1) Internal Memory using Dynamic Trajectory Summarization to compress interaction history, and 2) External Memory using Insight Retrieval Augmentation to retrieve actionable guidelines from an offline insight bank.

Result: Achieves up to 19.6% success rate increase and 58.7% token reduction for Qwen3-VL-32B, with proprietary models like Claude achieving up to 12.5% accuracy gains and significantly lower computational overhead.

Conclusion: M² demonstrates that training-free, memory-augmented approaches can effectively enhance MLLM agents’ performance on long-horizon web navigation tasks while reducing computational costs.

Abstract: Multimodal Large Language Models (MLLMs) based agents have demonstrated remarkable potential in autonomous web navigation. However, handling long-horizon tasks remains a critical bottleneck. Prevailing strategies often rely heavily on extensive data collection and model training, yet still struggle with high computational costs and insufficient reasoning capabilities when facing complex, long-horizon scenarios. To address this, we propose M$^2$, a training-free, memory-augmented framework designed to optimize context efficiency and decision-making robustness. Our approach incorporates a dual-tier memory mechanism that synergizes Dynamic Trajectory Summarization (Internal Memory) to compress verbose interaction history into concise state updates, and Insight Retrieval Augmentation (External Memory) to guide the agent with actionable guidelines retrieved from an offline insight bank. Extensive evaluations across WebVoyager and OnlineMind2Web demonstrate that M$^2$ consistently surpasses baselines, yielding up to a 19.6% success rate increase and 58.7% token reduction for Qwen3-VL-32B, while proprietary models like Claude achieve accuracy gains up to 12.5% alongside significantly lower computational overhead.

Method: Extends multiple instance learning with bidirectional feature integration between coarse-grained and fine-grained representations. Introduces specialized loss functions: hierarchical consistency loss, intra- and inter-class distance loss, and group-wise cross-entropy loss to optimize hierarchical learning.

Result: Tested on gastric biopsy dataset with 4 coarse-grained and 14 fine-grained classes. Achieved superior classification performance for both coarse-grained and fine-grained classification compared to existing methods.

Conclusion: HiClass effectively improves whole-slide image classification by capturing hierarchical histopathological characteristics, demonstrating the value of modeling label hierarchies in pathology image analysis.

Abstract: Whole-slide image analysis is essential for diagnostic tasks in pathology, yet existing deep learning methods primarily rely on flat classification, ignoring hierarchical relationships among class labels. In this study, we propose HiClass, a hierarchical classification framework for improved histopathology image analysis, that enhances both coarse-grained and fine-grained WSI classification. Built based upon a multiple instance learning approach, HiClass extends it by introducing bidirectional feature integration that facilitates information exchange between coarse-grained and fine-grained feature representations, effectively learning hierarchical features. Moreover, we introduce tailored loss functions, including hierarchical consistency loss, intra- and inter-class distance loss, and group-wise cross-entropy loss, to further optimize hierarchical learning. We assess the performance of HiClass on a gastric biopsy dataset with 4 coarse-grained and 14 fine-grained classes, achieving superior classification performance for both coarse-grained classification and fine-grained classification. These results demonstrate the effectiveness of HiClass in improving WSI classification by capturing coarse-grained and fine-grained histopathological characteristics.

Method: Introduces EmbedLens, a two-fold analytical framework with a novel probing tool for fine-grained analysis of visual token processing. Uses targeted patch-compression benchmark to analyze token categories and their semantic content.

Result: Visual tokens consistently partition into sink, dead, and alive categories; only ~60% alive tokens carry image-specific meaning. Alive tokens already encode rich fine-grained visual cues (objects, colors, OCR) before LLM processing. Internal visual computations are redundant for most standard tasks, and for vision-centric tasks, alive tokens naturally align with intermediate LLM layers rather than initial embedding space.

Conclusion: Provides unified mechanistic view of visual token processing, enabling more efficient MLLM architectures through selective token pruning, minimized visual computation, and mid-layer injection strategies.

Abstract: Multimodal large language models (MLLMs) project visual tokens into the embedding space of language models, yet the internal structuring and processing of visual semantics remain poorly understood. In this work, we introduce a two-fold analytical framework featuring a novel probing tool, $\textbf{EmbedLens}$, to conduct a fine-grained analysis. We uncover a pronounced semantic sparsity at the input level: visual tokens consistently partition into sink, dead, and alive categories. Remarkably, only the alive tokens, comprising $\approx60%$ of the total input, carry image-specific meaning. Furthermore, using a targeted patch-compression benchmark, we demonstrate that these alive tokens already encode rich, fine-grained cues (e.g., objects, colors, and OCR) prior to entering the LLM. Internal visual computations (such as visual attention and feed-forward networks) are redundant for most standard tasks. For the small subset of highly vision-centric tasks that actually benefit from internal processing, we reveal that alive tokens naturally align with intermediate LLM layers rather than the initial embedding space, indicating that shallow-layer processing is unnecessary and that direct mid-layer injection is both sufficient. Ultimately, our findings provide a unified mechanistic view of visual token processing, paving the way for more efficient and interpretable MLLM architectures through selective token pruning, minimized visual computation, and mid-layer injection. The code is released at: https://github.com/EIT-NLP/EmbedLens.

Method: Proposes Multimodal Adaptive RAG (MMA-RAG) which dynamically assesses model confidence in internal knowledge to decide whether to incorporate retrieved external information. Uses a decision classifier trained through layer-wise analysis that leverages joint internal visual and textual representations to guide reverse image retrieval.

Result: Achieves significant improvement in response performance across three VQA datasets. Ablation studies highlight the importance of internal representations in adaptive retrieval decisions. Effectively balances external knowledge utilization and inference robustness in diverse multimodal scenarios.

Conclusion: MMA-RAG effectively addresses hallucination issues in VQA by adaptively deciding when to use external knowledge, improving reliability through better integration of multimodal internal representations with external retrieval.

Abstract: Visual Question Answering systems face reliability issues due to hallucinations, where models generate answers misaligned with visual input or factual knowledge. While Retrieval Augmented Generation frameworks mitigate this issue by incorporating external knowledge, static retrieval often introduces irrelevant or conflicting content, particularly in visual RAG settings where visually similar but semantically incorrect evidence may be retrieved. To address this, we propose Multimodal Adaptive RAG (MMA-RAG), which dynamically assesses the confidence in the internal knowledge of the model to decide whether to incorporate the retrieved external information into the generation process. Central to MMA-RAG is a decision classifier trained through a layer-wise analysis, which leverages joint internal visual and textual representations to guide the use of reverse image retrieval. Experiments demonstrated that the model achieves a significant improvement in response performance in three VQA datasets. Meanwhile, ablation studies highlighted the importance of internal representations in adaptive retrieval decisions. In general, the experimental results demonstrated that MMA-RAG effectively balances external knowledge utilization and inference robustness in diverse multimodal scenarios.

Method: 1) Use wavelet transform to decompose noisy query-frame similarity signals into multiple scales, extracting clean semantic change signals from coarsest scale; 2) Identify local extrema as semantic boundaries to segment video into coherent clips; 3) Two-stage selection: adaptively allocate frame budget to each clip based on composite importance score, then use Maximal Marginal Relevance within each clip for diverse yet relevant frame selection.

Result: Significant performance improvements: 5.5% accuracy boost on VideoMME, 9.5% on MLVU, and 6.2% on LongVideoBench, consistently outperforming state-of-the-art methods.

Conclusion: WFS-SB demonstrates that effective video understanding requires capturing semantic shifts, not just query relevance, and provides a training-free framework using wavelet analysis for robust semantic boundary detection and intelligent frame selection.

Abstract: Frame selection is crucial due to high frame redundancy and limited context windows when applying Large Vision-Language Models (LVLMs) to long videos. Current methods typically select frames with high relevance to a given query, resulting in a disjointed set of frames that disregard the narrative structure of video. In this paper, we introduce Wavelet-based Frame Selection by Detecting Semantic Boundary (WFS-SB), a training-free framework that presents a new perspective: effective video understanding hinges not only on high relevance but, more importantly, on capturing semantic shifts - pivotal moments of narrative change that are essential to comprehending the holistic storyline of video. However, direct detection of abrupt changes in the query-frame similarity signal is often unreliable due to high-frequency noise arising from model uncertainty and transient visual variations. To address this, we leverage the wavelet transform, which provides an ideal solution through its multi-resolution analysis in both time and frequency domains. By applying this transform, we decompose the noisy signal into multiple scales and extract a clean semantic change signal from the coarsest scale. We identify the local extrema of this signal as semantic boundaries, which segment the video into coherent clips. Building on this, WFS-SB comprises a two-stage strategy: first, adaptively allocating a frame budget to each clip based on a composite importance score; and second, within each clip, employing the Maximal Marginal Relevance approach to select a diverse yet relevant set of frames. Extensive experiments show that WFS-SB significantly boosts LVLM performance, e.g., improving accuracy by 5.5% on VideoMME, 9.5% on MLVU, and 6.2% on LongVideoBench, consistently outperforming state-of-the-art methods.

Method: 1) Cost-efficient data curation pipeline repurposing stereo video datasets into 4D spatiotemporal instructional data (MLLM4D-2M, MLLM4D-R1-30k datasets); 2) Post-training strategy with SFT for foundational understanding and Group Relative Policy Optimization with Spatiotemporal Chain of Thought prompting and reward functions for reasoning.

Result: MLLM-4D achieves state-of-the-art spatial-temporal understanding and reasoning capabilities from purely 2D RGB inputs, as demonstrated through extensive experiments.

Conclusion: The framework successfully bridges gaps in training data and model post-training for spatiotemporal understanding, advancing MLLMs’ 4D intelligence without architectural modifications.

Abstract: Humans are born with vision-based 4D spatial-temporal intelligence, which enables us to perceive and reason about the evolution of 3D space over time from purely visual inputs. Despite its importance, this capability remains a significant bottleneck for current multimodal large language models (MLLMs). To tackle this challenge, we introduce MLLM-4D, a comprehensive framework designed to bridge the gaps in training data curation and model post-training for spatiotemporal understanding and reasoning. On the data front, we develop a cost-efficient data curation pipeline that repurposes existing stereo video datasets into high-quality 4D spatiotemporal instructional data. This results in the MLLM4D-2M and MLLM4D-R1-30k datasets for Supervised Fine-Tuning (SFT) and Reinforcement Fine-Tuning (RFT), alongside MLLM4D-Bench for comprehensive evaluation. Regarding model training, our post-training strategy establishes a foundational 4D understanding via SFT and further catalyzes 4D reasoning capabilities by employing Group Relative Policy Optimization (GRPO) with specialized Spatiotemporal Chain of Thought (ST-CoT) prompting and Spatiotemporal reward functions (ST-reward) without involving the modification of architecture. Extensive experiments demonstrate that MLLM-4D achieves state-of-the-art spatial-temporal understanding and reasoning capabilities from purely 2D RGB inputs. Project page: https://github.com/GVCLab/MLLM-4D.

Method: Proposes Vision-TTT, which extends Test-Time Training (TTT) to vision by compressing visual token sequences in a self-supervised manner. Incorporates bidirectional scan strategy and Conv2d modules to model 2D visual correlations with global receptive fields while maintaining linear-time complexity.

Result: Vision-TTT achieves 77.3%, 81.2%, 82.5% Top-1 accuracy on ImageNet for different model sizes, outperforms counterparts on downstream tasks, reduces FLOPs by 79.4%, runs 4.38x faster with 88.9% less memory than DeiT-T at 1280x1280 resolution.

Conclusion: Vision-TTT demonstrates both expressiveness and efficiency as a next-generation visual backbone, offering a promising alternative to traditional Vision Transformers by addressing their computational limitations while maintaining strong performance.

Abstract: Learning efficient and expressive visual representation has long been the pursuit of computer vision research. While Vision Transformers (ViTs) gradually replace traditional Convolutional Neural Networks (CNNs) as more scalable vision learners, their applications are plagued by the quadratic complexity of the self-attention mechanism. To address the challenge, we introduce a new linear-time sequence modeling method Test-Time Training (TTT) into vision and propose Vision-TTT, which compresses the visual token sequence in a novel self-supervised learning manner. By incorporating bidirectional scan strategy and the Conv2d module, Vision-TTT effectively extends vanilla TTT to model 2D visual correlations with global receptive fields. Extensive experiments show that \texttt{Vittt-T/S/B} achieve 77.3%,81.2%,82.5% Top-1 accuracy on ImageNet classification and also greatly outperform their counterparts on downstream tasks. At 1280x1280 resolution, \texttt{Vittt-T} reduces FLOPs by 79.4% and runs 4.38x faster with 88.9% less memory than DeiT-T. These results demonstrate the expressiveness and efficiency of Vision-TTT as a strong candidate for the next-generation generic visual backbone.

Method: Jano introduces: 1) Early-stage complexity recognition algorithm that identifies regional convergence requirements within initial denoising steps, and 2) Adaptive token scheduling runtime that optimizes computational resource allocation based on regional complexity.

Result: Jano achieves substantial acceleration with average 2.0× speedup (up to 2.4×) while preserving generation quality, evaluated on state-of-the-art models.

Conclusion: The work challenges conventional uniform processing assumptions in diffusion models and provides a practical solution for accelerating large-scale content generation through region-aware optimization.

Abstract: Diffusion models have achieved remarkable success in generative AI, yet their computational efficiency remains a significant challenge, particularly for Diffusion Transformers (DiTs) requiring intensive full-attention computation. While existing acceleration approaches focus on content-agnostic uniform optimization strategies, we observe that different regions in generated content exhibit heterogeneous convergence patterns during the denoising process. We present Jano, a training-free framework that leverages this insight for efficient region-aware generation. Jano introduces an early-stage complexity recognition algorithm that accurately identifies regional convergence requirements within initial denoising steps, coupled with an adaptive token scheduling runtime that optimizes computational resource allocation. Through comprehensive evaluation on state-of-the-art models, Jano achieves substantial acceleration (average 2.0 times speedup, up to 2.4 times) while preserving generation quality. Our work challenges conventional uniform processing assumptions and provides a practical solution for accelerating large-scale content generation. The source code of our implementation is available at https://github.com/chen-yy20/Jano.

Method: 1) First asynchronous online RL framework for 3D mesh generation; 2) Advantage-guided Ranking Preference Optimization (ARPO) algorithm; 3) Diagonal-aware mixed triangular-quadrilateral tokenization for mesh representation; 4) Ray-based reward for geometric integrity.

Result: Asynchronous framework is 3.75× faster than synchronous RL. Mesh-Pro achieves state-of-the-art performance on both artistic and dense meshes with better training efficiency and generalization.

Conclusion: The proposed Mesh-Pro framework significantly advances RL-based 3D mesh generation through asynchronous training, novel ARPO algorithm, and specialized mesh representation, setting new state-of-the-art benchmarks.

Abstract: Reinforcement learning (RL) has demonstrated remarkable success in text and image generation, yet its potential in 3D generation remains largely unexplored. Existing attempts typically rely on offline direct preference optimization (DPO) method, which suffers from low training efficiency and limited generalization. In this work, we aim to enhance both the training efficiency and generation quality of RL in 3D mesh generation. Specifically, (1) we design the first asynchronous online RL framework tailored for 3D mesh generation post-training efficiency improvement, which is 3.75$\times$ faster than synchronous RL. (2) We propose Advantage-guided Ranking Preference Optimization (ARPO), a novel RL algorithm that achieves a better trade-off between training efficiency and generalization than current RL algorithms designed for 3D mesh generation, such as DPO and group relative policy optimization (GRPO). (3) Based on asynchronous ARPO, we propose Mesh-Pro, which additionally introduces a novel diagonal-aware mixed triangular-quadrilateral tokenization for mesh representation and a ray-based reward for geometric integrity. Mesh-Pro achieves state-of-the-art performance on artistic and dense meshes.

Method: Proposes TP-Spikformer with: 1) heuristic spatiotemporal information-retaining criterion to evaluate token importance, 2) information-retaining token pruning framework using block-level early stopping strategy for uninformative tokens (instead of removal), 3) training-free operation across diverse spiking transformer architectures.

Result: Demonstrated effectiveness, efficiency and scalability across diverse architectures (Spikformer, QKFormer, Spike-driven Transformer V1/V3) and tasks (image classification, object detection, semantic segmentation, event-based object tracking). Performs well in training-free manner.

Conclusion: TP-Spikformer offers an efficient and practical solution for deploying SNNs in real-world applications with limited computational resources, showing potential for resource-constrained deployment scenarios.

Abstract: Spiking neural networks (SNNs) offer an energy-efficient alternative to traditional neural networks due to their event-driven computing paradigm. However, recent advancements in spiking transformers have focused on improving accuracy with large-scale architectures, which require significant computational resources and limit deployment on resource-constrained devices. In this paper, we propose a simple yet effective token pruning method for spiking transformers, termed TP-Spikformer, that reduces storage and computational overhead while maintaining competitive performance. Specifically, we first introduce a heuristic spatiotemporal information-retaining criterion that comprehensively evaluates tokens’ importance, assigning higher scores to informative tokens for retention and lower scores to uninformative ones for pruning. Based on this criterion, we propose an information-retaining token pruning framework that employs a block-level early stopping strategy for uninformative tokens, instead of removing them outright. This also helps preserve more information during token pruning. We demonstrate the effectiveness, efficiency and scalability of TP-Spikformer through extensive experiments across diverse architectures, including Spikformer, QKFormer and Spike-driven Transformer V1 and V3, and a range of tasks such as image classification, object detection, semantic segmentation and event-based object tracking. Particularly, TP-Spikformer performs well in a training-free manner. These results reveal its potential as an efficient and practical solution for deploying SNNs in real-world applications with limited computational resources.

Method: Develops CaptionFool, a novel universal adversarial attack against transformer-based captioning models. The attack modifies only 7 out of 577 image patches (approximately 1.2% of the image) to generate arbitrary target captions. The method specifically designs perturbations to generate “slang” terms that can evade existing content moderation filters.

Result: Achieves 94-96% success rate in generating arbitrary target captions, including offensive content. Demonstrates ability to generate slang terms specifically designed to evade content moderation filters. Shows that minimal image perturbations (just 7 patches) can completely manipulate caption generation.

Conclusion: Exposes critical vulnerabilities in deployed vision-language models and underscores the urgent need for robust defenses against such universal adversarial attacks. The attack’s effectiveness with minimal modifications highlights significant security concerns for multimodal AI systems.

Abstract: Image captioning models are encoder-decoder architectures trained on large-scale image-text datasets, making them susceptible to adversarial attacks. We present CaptionFool, a novel universal (input-agnostic) adversarial attack against state-of-the-art transformer-based captioning models. By modifying only 7 out of 577 image patches (approximately 1.2% of the image), our attack achieves 94-96% success rate in generating arbitrary target captions, including offensive content. We further demonstrate that CaptionFool can generate “slang” terms specifically designed to evade existing content moderation filters. Our findings expose critical vulnerabilities in deployed vision-language models and underscore the urgent need for robust defenses against such attacks. Warning: This paper contains model outputs which are offensive in nature.

Method: Proposes Retrieval-Augmented Flow Matching (RAFM) that adapts rectified flow to medical imaging by using a frozen DINOv3 encoder and global CT memory bank to construct retrieval-guided pseudo pairs, improving coupling quality and stabilizing unpaired flow-based training.

Result: Experiments on SynthRAD2023 under strict subject-level true-unpaired protocol show RAFM outperforms existing methods across FID, MAE, SSIM, PSNR, and SegScore metrics.

Conclusion: RAFM effectively addresses the challenges of unpaired CBCT-to-CT translation in medical imaging by improving pseudo pair construction and stabilizing flow-based training, achieving state-of-the-art performance.

Abstract: Cone-beam CT (CBCT) is routinely acquired in radiotherapy but suffers from severe artifacts and unreliable Hounsfield Unit (HU) values, limiting its direct use for dose calculation. Synthetic CT (sCT) generation from CBCT is therefore an important task, yet paired CBCT–CT data are often unavailable or unreliable due to temporal gaps, anatomical variation, and registration errors. In this work, we introduce rectified flow (RF) into unpaired CBCT-to-CT translation in medical imaging. Although RF is theoretically compatible with unpaired learning through distribution-level coupling and deterministic transport, its practical effectiveness under small medical datasets and limited batch sizes remains underexplored. Direct application with random or batch-local pseudo pairing can produce unstable supervision due to semantically mismatched endpoint samples. To address this challenge, we propose Retrieval-Augmented Flow Matching (RAFM), which adapts RF to the medical setting by constructing retrieval-guided pseudo pairs using a frozen DINOv3 encoder and a global CT memory bank. This strategy improves empirical coupling quality and stabilizes unpaired flow-based training. Experiments on SynthRAD2023 under a strict subject-level true-unpaired protocol show that RAFM outperforms existing methods across FID, MAE, SSIM, PSNR, and SegScore. The code is available at https://github.com/HiLab-git/RAFM.git.

Method: Proposes a closed-loop optimization framework with two mechanisms: (1) task feedback loop that dynamically modulates dehazing outputs based on performance across multiple downstream tasks, and (2) text instruction interface allowing users to specify high-level task preferences for real-time adjustment.

Result: Extensive experiments across various vision tasks demonstrate strong effectiveness, robustness, and generalizability of the approach, establishing a new paradigm for interactive, task-adaptive dehazing.

Conclusion: The framework enables feedback-driven refinement and user instruction-guided adjustment during inference, allowing the model to satisfy specific requirements of multiple downstream tasks without retraining, representing an interactive and collaborative dehazing approach.

Abstract: In real-world vision systems,haze removal is required not only to enhance image visibility but also to meet the specific needs of diverse downstream tasks.To address this challenge,we propose a novel adaptive dynamic dehazing framework that incorporates a closed-loop optimization mechanism.It enables feedback-driven refinement based on downstream task performance and user instruction-guided adjustment during inference,allowing the model to satisfy the specific requirements of multiple downstream tasks without retraining.Technically,our framework integrates two complementary and innovative mechanisms: (1)a task feedback loop that dynamically modulates dehazing outputs based on performance across multiple downstream tasks,and (2) a text instruction interface that allows users to specify high-level task preferences.This dual-guidance strategy enables the model to adapt its dehazing behavior after training,tailoring outputs in real time to the evolving needs of multiple tasks.Extensive experiments across various vision tasks demonstrate the strong effectiveness,robustness,and generalizability of our approach.These results establish a new paradigm for interactive,task-adaptive dehazing that actively collaborates with downstream applications.

Method: Proposes ScaleFormer architecture that tokenizes images into patch sequences with variable lengths proportional to image scale. Uses Scale-Aware Patchify module for training from fixed-size crops, decouples intra-patch spatial feature learning from inter-patch sequential dependency modeling, and incorporates Rotary Positional Encoding for extrapolation to unseen scales.

Result: Extensive experiments show ScaleFormer outperforms state-of-the-art methods in both fusion quality and cross-scale generalization. The approach is validated on the newly introduced PanScale dataset and PanScale-Bench benchmark.

Conclusion: The proposed ScaleFormer successfully addresses cross-scale pansharpening challenges by reframing resolution generalization as sequence length generalization, demonstrating superior performance and generalization capabilities across varying resolutions.

Abstract: Pansharpening aims to generate high-resolution multi-spectral images by fusing the spatial detail of panchromatic images with the spectral richness of low-resolution MS data. However, most existing methods are evaluated under limited, low-resolution settings, limiting their generalization to real-world, high-resolution scenarios. To bridge this gap, we systematically investigate the data, algorithmic, and computational challenges of cross-scale pansharpening. We first introduce PanScale, the first large-scale, cross-scale pansharpening dataset, accompanied by PanScale-Bench, a comprehensive benchmark for evaluating generalization across varying resolutions and scales. To realize scale generalization, we propose ScaleFormer, a novel architecture designed for multi-scale pansharpening. ScaleFormer reframes generalization across image resolutions as generalization across sequence lengths: it tokenizes images into patch sequences of the same resolution but variable length proportional to image scale. A Scale-Aware Patchify module enables training for such variations from fixed-size crops. ScaleFormer then decouples intra-patch spatial feature learning from inter-patch sequential dependency modeling, incorporating Rotary Positional Encoding to enhance extrapolation to unseen scales. Extensive experiments show that our approach outperforms SOTA methods in fusion quality and cross-scale generalization. The datasets and source code are available upon acceptance.

Method: MIMD-3DVT (Multiple Inputs and Mixed Data 3D Vision Transformer) processes consecutive MRI slices to capture 3D spatial information, fuses multiple 3D ROI imaging data, and integrates mixed data including demographic factors and cognitive assessments.

Result: The method achieved 97.14% accuracy in distinguishing between Normal Cognition and Alzheimer’s Disease using combined datasets (ADNI, AIBL, OASIS), outperforming state-of-the-art methods.

Conclusion: The proposed 3D Vision Transformer approach with multimodal data integration provides superior Alzheimer’s Disease diagnosis by capturing comprehensive 3D brain information and leveraging diverse data sources.

Abstract: The current methods for diagnosing Alzheimer Disease using Magnetic Resonance Imaging (MRI) have significant limitations. Many previous studies used 2D Transformers to analyze individual brain slices independently, potentially losing critical 3D contextual information. Region of interest-based models often focus on only a few brain regions despite Alzheimer’s affecting multiple areas. Additionally, most classification models rely on a single test, whereas diagnosing Alzheimer’s requires a multifaceted approach integrating diverse data sources for a more accurate assessment. This study introduces a novel methodology called the Multiple Inputs and Mixed Data 3D Vision Transformer (MIMD-3DVT). This method processes consecutive slices together to capture the feature dimensions and spatial information, fuses multiple 3D ROI imaging data inputs, and integrates mixed data from demographic factors, cognitive assessments, and brain imaging. The proposed methodology was experimentally evaluated using a combined dataset that included the Alzheimer’s Disease Neuroimaging Initiative (ADNI), the Australian Imaging, Biomarker, and Lifestyle Flagship Study of Ageing (AIBL), and the Open Access Series of Imaging Studies (OASIS). Our MIMD-3DVT, utilizing single or multiple ROIs, achieved an accuracy of 97.14%, outperforming the state-of-the-art methods in distinguishing between Normal Cognition and Alzheimer’s Disease.

Method: Proposes LAS-VAD with three key mechanisms: 1) Anomaly-connected component mechanism groups video frames into semantic groups sharing identical information, 2) Intention awareness mechanism distinguishes between similar normal and abnormal behaviors, and 3) Incorporates anomaly attribute information (e.g., flames and smoke for explosions) to guide accurate detection.

Result: Extensive experiments on XD-Violence and UCF-Crime datasets show LAS-VAD outperforms current state-of-the-art methods with remarkable gains, demonstrating effectiveness in learning anomaly semantics under weak supervision.

Conclusion: LAS-VAD successfully addresses the limitation of learning anomaly semantics in weakly supervised VAD by integrating semantic grouping, intention awareness, and attribute information, achieving superior performance on benchmark datasets.

Abstract: Weakly supervised video anomaly detection (WS-VAD) involves identifying the temporal intervals that contain anomalous events in untrimmed videos, where only video-level annotations are provided as supervisory signals. However, a key limitation persists in WS-VAD, as dense frame-level annotations are absent, which often leaves existing methods struggling to learn anomaly semantics effectively. To address this issue, we propose a novel framework named LAS-VAD, short for Learning Anomaly Semantics for WS-VAD, which integrates anomaly-connected component mechanism and intention awareness mechanism. The former is designed to assign video frames into distinct semantic groups within a video, and frame segments within the same group are deemed to share identical semantic information. The latter leverages an intention-aware strategy to distinguish between similar normal and abnormal behaviors (e.g., taking items and stealing). To further model the semantic information of anomalies, as anomaly occurrence is accompanied by distinct characteristic attributes (i.e., explosions are characterized by flames and thick smoke), we additionally incorporate anomaly attribute information to guide accurate detection. Extensive experiments on two benchmark datasets, XD-Violence and UCF-Crime, demonstrate that our LAS-VAD outperforms current state-of-the-art methods with remarkable gains.

Method: Two-stage pipeline: 1) Multi-view Metric Geometry Rectification module that rectifies imprecise calibration into a scale-consistent and geometrically consistent camera setup with single global scale, 2) Occlusion-Robust 3D Point Tracking directly in unified OR world frame.

Result: On MM-OR benchmark, rectification front-end reduces cross-view depth disagreement by more than 30× compared to raw calibration. Ablation studies show relationship between calibration quality and tracking accuracy, demonstrating improved geometric consistency yields stronger world-frame tracking.

Conclusion: The Geometry OR Tracker addresses geometric inconsistency in clinical OR settings through calibration rectification, enabling more accurate multi-view 3D tracking for downstream applications like surgeon behavior recognition.

Abstract: In operating rooms (OR), world-scale multi-view 3D tracking supports downstream applications such as surgeon behavior recognition, where physically meaningful quantities such as distances and motion statistics must be measured in meters. However, real clinical deployments rarely satisfy the geometric prerequisites for stable multi-view fusion and tracking: camera calibration and RGB-D registration are always unreliable, leading to cross-view geometric inconsistency that produces “ghosting” during fusion and degrades 3D trajectories in a shared OR coordinate frame. To address this, we introduce Geometry OR Tracker, a two-stage pipeline that first rectifies imprecise calibration into a scaleconsistent and geometrically consistent camera setup with a single global scale via a Multi-view Metric Geometry Rectification module, and then performs Occlusion-Robust 3D Point Tracking directly in the unified OR world frame. On the MM-OR benchmark, improved geometric consistency translates into tracking gains: our rectification front-end reduces cross-view depth disagreement by more than 30$\times$ compared to raw calibration. Ablation studies further demonstrate the relationship between calibration quality and tracking accuracy, showing that improved geometric consistency yields stronger world-frame tracking.

Method: MIDAS decomposes harmful semantics into risk-bearing subunits, disperses them across multiple visual clues, and leverages cross-image reasoning to gradually reconstruct malicious intent. It enforces longer, structured multi-image chained reasoning to increase model reliance on visual cues while delaying exposure of malicious semantics.

Result: Extensive experiments show MIDAS outperforms state-of-the-art jailbreak attacks for MLLMs, achieving an average attack success rate of 81.46% across 4 closed-source MLLMs.

Conclusion: MIDAS demonstrates that multi-image dispersion and semantic reconstruction can effectively bypass MLLM safety mechanisms, highlighting vulnerabilities in current multimodal security approaches and the need for more robust defenses.

Abstract: Multimodal Large Language Models (MLLMs) have achieved remarkable performance but remain vulnerable to jailbreak attacks that can induce harmful content and undermine their secure deployment. Previous studies have shown that introducing additional inference steps, which disrupt security attention, can make MLLMs more susceptible to being misled into generating malicious content. However, these methods rely on single-image masking or isolated visual cues, which only modestly extend reasoning paths and thus achieve limited effectiveness, particularly against strongly aligned commercial closed-source models. To address this problem, in this paper, we propose Multi-Image Dispersion and Semantic Reconstruction (MIDAS), a multimodal jailbreak framework that decomposes harmful semantics into risk-bearing subunits, disperses them across multiple visual clues, and leverages cross-image reasoning to gradually reconstruct the malicious intent, thereby bypassing existing safety mechanisms. The proposed MIDAS enforces longer and more structured multi-image chained reasoning, substantially increases the model’s reliance on visual cues while delaying the exposure of malicious semantics and significantly reducing the model’s security attention, thereby improving the performance of jailbreak against advanced MLLMs. Extensive experiments across different datasets and MLLMs demonstrate that the proposed MIDAS outperforms state-of-the-art jailbreak attacks for MLLMs and achieves an average attack success rate of 81.46% across 4 closed-source MLLMs. Our code is available at this link.

Method: DASP uses a diagnose-then-mitigate framework: first analyzes interdimensional redundancy to identify biased vs unbiased modalities, then applies asymmetric adaptation with decoupled stable/plastic components per modality.

Result: Comprehensive evaluations on diverse multimodal benchmarks show DASP significantly outperforms state-of-the-art methods.

Conclusion: The asymmetric adaptation strategy effectively balances stability and plasticity for different modalities, enabling flexible domain adaptation while preserving generalizable knowledge.

Abstract: Adapting pretrained multi-modal models to evolving test-time distributions, known as multi-modal test-time adaptation, presents a significant challenge. Existing methods frequently encounter negative transfer in the unbiased modality and catastrophic forgetting in the biased modality. To address these challenges, we propose Decoupling Adaptation for Stability and Plasticity (DASP), a novel diagnose-then-mitigate framework. Our analysis reveals a critical discrepancy within the unified latent space: the biased modality exhibits substantially higher interdimensional redundancy (i.e., strong correlations across feature dimensions) compared to the unbiased modality. Leveraging this insight, DASP identifies the biased modality and implements an asymmetric adaptation strategy. This strategy employs a decoupled architecture where each modality-specific adapter is divided into stable and plastic components. The asymmetric mechanism works as follows: for the biased modality, which requires plasticity, the plastic component is activated and updated to capture domain-specific information, while the stable component remains fixed. Conversely, for the unbiased modality, which requires stability, the plastic component is bypassed, and the stable component is updated using KL regularization to prevent negative transfer. This asymmetric design enables the model to adapt flexibly to new domains while preserving generalizable knowledge. Comprehensive evaluations on diverse multi-modal benchmarks demonstrate that DASP significantly outperforms state-of-the-art methods.

Method: Proposes WildActor framework with Asymmetric Identity-Preserving Attention and Viewpoint-Adaptive Monte Carlo Sampling that iteratively re-weights reference conditions for balanced manifold coverage, trained on Actor-18M dataset (1.6M videos, 18M images).

Result: WildActor consistently preserves body identity under diverse shot compositions, large viewpoint transitions, and substantial motions, outperforming existing methods on the proposed Actor-Bench evaluation.

Conclusion: The approach successfully addresses the challenge of identity-consistent human video generation across unconstrained viewpoints and environments, advancing production-ready digital actor creation.

Abstract: Production-ready human video generation requires digital actors to maintain strictly consistent full-body identities across dynamic shots, viewpoints and motions, a setting that remains challenging for existing methods. Prior methods often suffer from face-centric behavior that neglects body-level consistency, or produce copy-paste artifacts where subjects appear rigid due to pose locking. We present Actor-18M, a large-scale human video dataset designed to capture identity consistency under unconstrained viewpoints and environments. Actor-18M comprises 1.6M videos with 18M corresponding human images, covering both arbitrary views and canonical three-view representations. Leveraging Actor-18M, we propose WildActor, a framework for any-view conditioned human video generation. We introduce an Asymmetric Identity-Preserving Attention mechanism coupled with a Viewpoint-Adaptive Monte Carlo Sampling strategy that iteratively re-weights reference conditions by marginal utility for balanced manifold coverage. Evaluated on the proposed Actor-Bench, WildActor consistently preserves body identity under diverse shot compositions, large viewpoint transitions, and substantial motions, surpassing existing methods in these challenging settings.

Method: Proposes AlignVAR with two key components: (1) Spatial Consistency Autoregression (SCA) uses adaptive mask to reweight attention toward structurally correlated regions, enhancing long-range dependencies; (2) Hierarchical Consistency Constraint (HCC) adds full reconstruction supervision at each scale to expose accumulated deviations early.

Result: AlignVAR enhances structural coherence and perceptual fidelity over existing generative methods, delivers over 10x faster inference with nearly 50% fewer parameters than leading diffusion-based approaches.

Conclusion: AlignVAR establishes a new paradigm for efficient image super-resolution by addressing key challenges in applying visual autoregressive models to this task.

Abstract: Visual autoregressive (VAR) models have recently emerged as a promising alternative for image generation, offering stable training, non-iterative inference, and high-fidelity synthesis through next-scale prediction. This encourages the exploration of VAR for image super-resolution (ISR), yet its application remains underexplored and faces two critical challenges: locality-biased attention, which fragments spatial structures, and residual-only supervision, which accumulates errors across scales, severely compromises global consistency of reconstructed images. To address these issues, we propose AlignVAR, a globally consistent visual autoregressive framework tailored for ISR, featuring two key components: (1) Spatial Consistency Autoregression (SCA), which applies an adaptive mask to reweight attention toward structurally correlated regions, thereby mitigating excessive locality and enhancing long-range dependencies; and (2) Hierarchical Consistency Constraint (HCC), which augments residual learning with full reconstruction supervision at each scale, exposing accumulated deviations early and stabilizing the coarse-to-fine refinement process. Extensive experiments demonstrate that AlignVAR consistently enhances structural coherence and perceptual fidelity over existing generative methods, while delivering over 10x faster inference with nearly 50% fewer parameters than leading diffusion-based approaches, establishing a new paradigm for efficient ISR.

Method: Created UNICBench with 5,300 images (5,508 QA), 872 documents (5,888 QA), and 2,069 audio clips (2,905 QA) annotated with three-level capability taxonomy and difficulty tags. Implemented standardized evaluation protocol with fixed splits/prompts/seeds and modality-specific matching rules.

Result: Evaluation of 45 state-of-the-art MLLMs shows strong performance on basic counting tasks but significant gaps on reasoning and hardest partitions, revealing long-tail errors and substantial room for improvement in general counting ability.

Conclusion: UNICBench provides a rigorous, comparable basis for measuring multimodal counting capabilities and offers a public toolkit to accelerate progress in this important area of MLLM research.

Abstract: Counting is a core capability for multimodal large language models (MLLMs), yet there is no unified counting dataset to rigorously evaluate this ability across image, text, and audio. We present UNICBench, a unified multimodal, multi level counting benchmark and evaluation toolkit with accurate ground truth, deterministic numeric parsing, and stratified reporting. The corpus comprises 5,300 images (5,508 QA), 872 documents (5,888 QA), and 2,069 audio clips (2,905 QA), annotated with a three level capability taxonomy and difficulty tags. Under a standardized protocol with fixed splits/prompts/seeds and modality specific matching rules, we evaluate 45 state-of-the-art MLLMs across modalities. Results show strong performance on some basic counting tasks but significant gaps on reasoning and the hardest partitions, highlighting long-tail errors and substantial headroom for improving general counting. UNICBench offers a rigorous and comparable basis for measurement and a public toolkit to accelerate progress.

Method: Introduces a novel benchmark with publicly available dataset and two techniques leveraging model uncertainty, prediction consistency, and representation analysis to identify, quantify, and rank training samples by label noise level.

Result: Proposed methods consistently outperform established baselines across various experimental settings.

Conclusion: Provides valuable tools for improving remote sensing segmentation by addressing label noise issues, with publicly available resources for the research community.

Abstract: High-quality pixel-level annotations are essential for the semantic segmentation of remote sensing imagery. However, such labels are expensive to obtain and often affected by noise due to the labor-intensive and time-consuming nature of pixel-wise annotation, which makes it challenging for human annotators to label every pixel accurately. Annotation errors can significantly degrade the performance and robustness of modern segmentation models, motivating the need for reliable mechanisms to identify and quantify noisy training samples. This paper introduces a novel Data-Centric benchmark, together with a novel, publicly available dataset and two techniques for identifying, quantifying, and ranking training samples according to their level of label noise in remote sensing semantic segmentation. Such proposed methods leverage complementary strategies based on model uncertainty, prediction consistency, and representation analysis, and consistently outperform established baselines across a range of experimental settings. The outcomes of this work are publicly available at https://github.com/keillernogueira/label_noise_segmentation.

Method: Two-stage framework: 1) Supervised fine-tuning stage with task-adaptive timestep scheduling (linear decay schedule and temporal gating mechanism) and VLM for prompt synthesis; 2) Fine-grained group-level DPO with weighted margin formulation to eliminate artifacts and improve harmony.

Result: IdGlow fundamentally mitigates the stability-plasticity conflict, achieving superior Pareto balance between state-of-the-art facial fidelity and commercial-grade aesthetic quality on two challenging benchmarks: direct multi-person fusion and age-transformed group generation.

Conclusion: The proposed mask-free, progressive two-stage framework successfully addresses complex multi-subject generation challenges, particularly identity-preserving age transformation, through innovative timestep scheduling and optimization techniques.

Abstract: Multi-subject image generation requires seamlessly harmonizing multiple reference identities within a coherent scene. However, existing methods relying on rigid spatial masks or localized attention often struggle with the “stability-plasticity dilemma,” particularly failing in tasks that require complex structural deformations, such as identity-preserving age transformation. To address this, we present IdGlow, a mask-free, progressive two-stage framework built upon Flow Matching diffusion models. In the supervised fine-tuning (SFT) stage, we introduce task-adaptive timestep scheduling aligned with diffusion generative dynamics: a linear decay schedule that progressively relaxes constraints for natural group composition, and a temporal gating mechanism that concentrates identity injection within a critical semantic window, successfully preserving adult facial semantics without overriding child-like anatomical structures. To resolve attribute leakage and semantic ambiguity without explicit layout inputs, we further integrate a badcase-driven Vision-Language Model (VLM) for precise, context-aware prompt synthesis. In the second stage, we design a Fine-Grained Group-Level Direct Preference Optimization (DPO) with a weighted margin formulation to simultaneously eliminate multi-subject artifacts, elevate texture harmony, and recalibrate identity fidelity towards real-world distributions. Extensive experiments on two challenging benchmarks – direct multi-person fusion and age-transformed group generation – demonstrate that IdGlow fundamentally mitigates the stability-plasticity conflict, achieving a superior Pareto balance between state-of-the-art facial fidelity and commercial-grade aesthetic quality.

Method: Proposes CodeAlign framework using cross-modal feature-code-feature (FCF) translation. Uses codebooks to regularize feature spaces into compact code spaces, then learns mappings between modality-specific feature spaces by translating features to corresponding codes of other modalities and decoding back into target feature spaces.

Result: Achieves state-of-the-art perception performance on OPV2V and DAIR-V2X datasets while requiring only 8% of training parameters compared to prior alignment methods and reducing communication load by 1024x.

Conclusion: CodeAlign provides an efficient, co-occurrence-free alignment framework that effectively addresses modality isolation in collaborative perception through codebook-based feature translation, enabling seamless cross-modal collaboration without spatial correspondence requirements.

Abstract: Collaborative perception leverages data exchange among multiple agents to enhance overall perception capabilities. However, heterogeneity across agents introduces domain gaps that hinder collaboration, and this is further exacerbated by an underexplored issue: modality isolation. It arises when multiple agents with different modalities never co-occur in any training data frame, enlarging cross-modal domain gaps. Existing alignment methods rely on supervision from spatially overlapping observations, thus fail to handle modality isolation. To address this challenge, we propose CodeAlign, the first efficient, co-occurrence-free alignment framework that smoothly aligns modalities via cross-modal feature-code-feature(FCF) translation. The key idea is to explicitly identify the representation consistency through codebook, and directly learn mappings between modality-specific feature spaces, thereby eliminating the need for spatial correspondence. Codebooks regularize feature spaces into code spaces, providing compact yet expressive representations. With a prepared code space for each modality, CodeAlign learns FCF translations that map features to the corresponding codes of other modalities, which are then decoded back into features in the target code space, enabling effective alignment. Experiments show that, when integrating three modalities, CodeAlign requires only 8% of the training parameters of prior alignment methods, reduces communication load by 1024x, and achieves state-of-the-art perception performance on both OPV2V and DAIR-V2X dataset. Code will be released on https://github.com/cxliu0314/CodeAlign.

Method: Proposed Propagation-Guided Spectral Video Reconstruction Transformer (PG-SVRT) with spatial-then-temporal attention to reconstruct spectral features from video information, using bridged tokens to reduce computational complexity. Also created DynaSpec dataset (first high-quality dynamic hyperspectral image dataset) and conducted experiments with four SCI systems including a DD-CASSI prototype.

Result: PG-SVRT achieves superior performance in reconstruction quality, spectral fidelity, and temporal consistency while maintaining minimal FLOPs. Extensive experiments demonstrate effectiveness across simulation and real-world data collection.

Conclusion: The paper successfully advances spectral reconstruction from image to video level by leveraging complementary features and temporal continuity across adjacent frames, addressing key limitations of existing methods through the proposed PG-SVRT architecture and comprehensive experimental validation.

Abstract: Recently, Spectral Compressive Imaging (SCI) has achieved remarkable success, unlocking significant potential for dynamic spectral vision. However, existing reconstruction methods, primarily image-based, suffer from two limitations: (i) Encoding process masks spatial-spectral features, leading to uncertainty in reconstructing missing information from single compressed measurements, and (ii) The frame-by-frame reconstruction paradigm fails to ensure temporal consistency, which is crucial in the video perception. To address these challenges, this paper seeks to advance spectral reconstruction from the image level to the video level, leveraging the complementary features and temporal continuity across adjacent frames in dynamic scenes. Initially, we construct the first high-quality dynamic hyperspectral image dataset (DynaSpec), comprising 30 sequences obtained through frame-scanning acquisition. Subsequently, we propose the Propagation-Guided Spectral Video Reconstruction Transformer (PG-SVRT), which employs a spatial-then-temporal attention to effectively reconstruct spectral features from abundant video information, while using a bridged token to reduce computational complexity. Finally, we conduct simulation experiments to assess the performance of four SCI systems, and construct a DD-CASSI prototype for real-world data collection and benchmarking. Extensive experiments demonstrate that PG-SVRT achieves superior performance in reconstruction quality, spectral fidelity, and temporal consistency, while maintaining minimal FLOPs. Project page: https://github.com/nju-cite/DynaSpec

Method: This is a position paper that presents arguments and calls for paradigm shifts rather than proposing specific technical methods. The approach is conceptual, advocating for rebalancing evaluation frameworks toward human-centered, context-aware, and fine-grained assessment.

Result: The paper establishes the need for evaluation reform in visual processing systems, highlighting the divergence between objective metrics and human perception, and provides a framework for more appropriate assessment approaches.

Conclusion: The conclusion calls for moving beyond single-metric IQA benchmarks toward more comprehensive evaluation paradigms that better align with human perception and user preferences, particularly for generative and perception-oriented visual processing systems.

Abstract: This position paper argues that the evaluation of modern visual processing systems should no longer be driven primarily by single-metric image quality assessment benchmarks, particularly in the era of generative and perception-oriented methods. Image restoration exemplifies this divergence: while objective IQA metrics enable reproducible, scalable evaluation, they have increasingly drifted apart from human perception and user preferences. We contend that this mismatch risks constraining innovation and misguiding research progress across visual processing tasks. Rather than rejecting metrics altogether, this paper calls for a rebalancing of evaluation paradigms, advocating a more human-centered, context-aware, and fine-grained approach to assessing the visual models’ outcomes.

Method: The authors formulate pruning as approximating full-data expected risk with a weighted subset, identifying two key errors: coverage error (insufficient representativeness) and prior-mismatch bias (inconsistency between subset-induced class weights and target metrics). They propose representation-aware subset selection with per-class retention quotas for long-tail coverage, and prior-invariant teacher supervision using calibrated soft labels and embedding-geometry distillation.

Result: Extensive experiments on 3D datasets show that the method can improve both OA and mAcc metrics across multiple settings while adapting to different downstream preferences. The retention quota serves as a switch to control the OA-mAcc trade-off.

Conclusion: The paper presents the first comprehensive study of dataset pruning for 3D data, addressing the unique challenges of long-tail distributions and conflicting evaluation metrics through novel theoretical formulation and practical methods that enable flexible trade-off control between different accuracy metrics.

Abstract: Dataset pruning has been widely studied for 2D images to remove redundancy and accelerate training, while particular pruning methods for 3D data remain largely unexplored. In this work, we study dataset pruning for 3D data, where its observed common long-tail class distribution nature make optimization under conventional evaluation metrics Overall Accuracy (OA) and Mean Accuracy (mAcc) inherently conflicting, and further make pruning particularly challenging. To address this, we formulate pruning as approximating the full-data expected risk with a weighted subset, which reveals two key errors: coverage error from insufficient representativeness and prior-mismatch bias from inconsistency between subset-induced class weights and target metrics. We propose representation-aware subset selection with per-class retention quotas for long-tail coverage, and prior-invariant teacher supervision using calibrated soft labels and embedding-geometry distillation. The retention quota also serves as a switch to control the OA-mAcc trade-off. Extensive experiments on 3D datasets show that our method can improve both metrics across multiple settings while adapting to different downstream preferences. Our code is available at https://github.com/XiaohanZhao123/3D-Dataset-Pruning.

Method: Three key components: 1) Geometric Structure Rectification (GSR) aligns visual semantics with radar-derived geometry, 2) Uncertainty-Aware Communication (UAC) selectively transmits informative features based on inter-agent disagreement, 3) Consensus-Driven Assembler (CDA) aggregates multi-agent information via shared geometric anchors.

Result: State-of-the-art performance on V2X-Radar and V2X-R benchmarks with significantly reduced communication overhead compared to existing methods.

Conclusion: RC-GeoCP successfully demonstrates the effectiveness of radar-camera fusion in collaborative perception, establishing a new benchmark and showing superior performance with efficient communication.

Abstract: Collaborative perception (CP) enhances scene understanding through multi-agent information sharing. While LiDAR-centric systems offer precise geometry, high costs and performance degradation in adverse weather necessitate multi-modal alternatives. Despite dense visual semantics and robust spatial measurements, the synergy between cameras and 4D radar remains underexplored in collaborative settings. This work introduces RC-GeoCP, the first framework to explore the fusion of 4D radar and images in CP. To resolve misalignment caused by depth ambiguity and spatial dispersion across agents, RC-GeoCP establishes a radar-anchored geometric consensus. Specifically, Geometric Structure Rectification (GSR) aligns visual semantics with geometry derived from radar to generate spatially grounded, geometry-consistent representations. Uncertainty-Aware Communication (UAC) formulates selective transmission as a conditional entropy reduction process to prioritize informative features based on inter-agent disagreement. Finally, the Consensus-Driven Assembler (CDA) aggregates multi-agent information via shared geometric anchors to form a globally coherent representation. We establish the first unified radar-camera CP benchmark on V2X-Radar and V2X-R, demonstrating state-of-the-art performance with significantly reduced communication overhead. Code will be released soon.

Method: Proposes SCVM with: 1) recursively updated cross-layer memory state inside vision encoder to model long-range inter-layer dependencies; 2) layer-wise feedback modulation that refreshes token representations at each layer based on accumulated memory; 3) auxiliary semantic alignment objective supervising final memory state for progressive compression and reinforcement of task-relevant information.

Result: Experimental results on multiple visual question answering and hallucination evaluation benchmarks show consistent performance improvements without expanding visual tokens, adding vision encoders, or modifying/fine-tuning the language model.

Conclusion: SCVM effectively addresses visual representation limitations in MLLMs through representation evolution control, achieving better performance while maintaining model compatibility and efficiency.

Abstract: Recent multimodal large language models (MLLMs) widely adopt multi-layer visual feature fusion to enhance visual representation. However, existing approaches typically perform static concatenation or weighted aggregation after visual encoding, without intervening in the representation formation process itself. As a result, fine-grained details from early layers may be progressively suppressed during hierarchical abstraction. Moreover, directly introducing shallow-layer features into the language model often leads to semantic distribution mismatch with the visual feature space that the LLM’s cross-attention layers were pretrained on, which typically requires additional adaptation or fine-tuning of the LLM. To address these limitations, we revisit visual representation learning from the perspective of representation evolution control and propose a cross-layer memory-modulated vision framework(SCVM). Specifically, we introduce a recursively updated cross-layer memory state inside the vision encoder to model long-range inter-layer dependencies. We further design a layer-wise feedback modulation mechanism that refreshes token representations at each layer based on the accumulated memory, thereby structurally regulating the representation evolution trajectory. In addition, we incorporate an auxiliary semantic alignment objective that explicitly supervises the final memory state, encouraging progressive compression and reinforcement of task-relevant information. Experimental results on multiple visual question answering and hallucination evaluation benchmarks demonstrate that SCVM achieves consistent performance improvements without expanding visual tokens, introducing additional vision encoders, or modifying or fine-tuning the language model.

Method: Proposes HistoSelect with two components: 1) group sampler identifies question-relevant tissue regions, 2) patch selector retrieves most informative patches within those regions. Uses question-guided, tissue-aware, coarse-to-fine retrieval to mimic human slide examination patterns.

Result: Reduces visual token usage by 70% on average while improving accuracy across three pathology QA tasks. Evaluated on 356,000 question-answer pairs, outperforms existing methods and produces answers grounded in interpretable, pathologist-consistent regions.

Conclusion: Bringing human-like search and attention patterns into whole slide image reasoning is promising for building practical and reliable pathology vision-language models. Selective attention to relevant regions improves efficiency and accuracy.

Abstract: Computational pathology has advanced rapidly in recent years, driven by domain-specific image encoders and growing interest in using vision-language models to answer natural-language questions about diseases. Yet, the core problem behind pathology question-answering remains unsolved, considering that a gigapixel slide contains far more information than necessary for a given question. Pathologists naturally navigate tissue and morphology complexity by scanning broadly, and zooming in selectively according to the clinical questions. Current models, in contrast, rely on uniform patch sampling or broad attention maps, often attending equally to irrelevant regions while overlooking key visual evidence. In this work, we try to bring models closer to how humans actually examine slides. We propose a question-guided, tissue-aware, and coarse-to-fine retrieval framework, HistoSelect, that consists of two key components: a group sampler that identifies question-relevant tissue regions, followed by a patch selector that retrieves the most informative patches within those regions. By selecting only the most informative patches, our method becomes significantly more efficient: reducing visual token usage by 70% on average, while improving accuracy across three pathology QA tasks. Evaluated on 356,000 question-answer pairs, our approach outperforms existing methods and produces answers grounded in interpretable, pathologist-consistent regions. Our results suggest that bringing human-like search and attention patterns into WSI reasoning is a promising direction for building practical and reliable pathology VLMs.

Method: Proposes a k-space dual channel U-Net that processes real and imaginary components of undersampled low-field k-space data to restore missing frequency content, enabling direct reconstruction of high-field-like images from undersampled acquisitions.

Result: The k-space-driven image enhancement consistently outperforms spatial-domain methods, and reconstructions from undersampled k-space achieve image quality comparable to full k-space acquisitions on low-field brain MRI data.

Conclusion: This is the first work investigating low-field MRI super-resolution/image quality transfer directly from undersampled k-space, demonstrating superior performance over spatial-domain approaches for accelerated low-field MRI reconstruction.

Abstract: Low-field magnetic resonance imaging (MRI) provides affordable access to diagnostic imaging but suffers from prolonged acquisition and limited image quality. Accelerated imaging can be achieved with k-space undersampling, while super-resolution (SR) and image quality transfer (IQT) methods typically rely on spatial-domain post-processing. In this work, we propose a novel framework for reconstructing high-field MR like images directly from undersampled low-field k-space. Our approach employs a k-space dual channel U-Net that processes the real and imaginary components of undersampled k-space to restore missing frequency content. Experiments on low-field brain MRI demonstrate that our k-space-driven image enhancement consistently outperforms the counterpart spatial-domain method. Furthermore, reconstructions from undersampled k-space achieve image quality comparable to full k-space acquisitions. To the best of our knowledge, this is the first work that investigates low-field MRI SR/IQT directly from undersampled k-space.

Method: Proposes Mixture of Low-Rank Experts (MoLRE) framework that extends LoRA with multiple specialized low-rank adapters and unsupervised soft routing, enabling conditional feature adaptation with less than 0.5% additional parameters and without explicit pathology supervision.

Result: Consistent performance improvements across all tested models on 70,000+ head CT scans with 75 annotated findings. Gains varied by model type: general-purpose and medical-domain models showed largest improvements (DINOv3-Base: +4.6%; MedGemma: +4.3%), while 3D CT-specialized or very large models showed more modest gains (+0.2-1.3%). MoLRE + MedGemma achieved highest average detection AUC of 0.917.

Conclusion: MoLRE effectively adapts foundation models to complex multi-label medical tasks, highlighting the importance of systematic benchmarking as pretraining domain, architecture, and model scale interact in non-obvious ways for clinical applications.

Abstract: Foundation models pre-trained on large-scale datasets demonstrate strong transfer learning capabilities; however, their adaptation to complex multi-label diagnostic tasks-such as comprehensive head CT finding detection-remains understudied. Standard parameter-efficient fine-tuning methods such as LoRA apply uniform adaptations across pathology types, which may limit performance for diverse medical findings. We propose a Mixture of Low-Rank Experts (MoLRE) framework that extends LoRA with multiple specialized low-rank adapters and unsupervised soft routing. This approach enables conditional feature adaptation with less than 0.5% additional parameters and without explicit pathology supervision. We present a comprehensive benchmark of MoLRE across six state-of-the-art medical imaging foundation models spanning 2D and 3D architectures, general-domain, medical-domain, and head CT-specific pretraining, and model sizes ranging from 7M to 431M parameters. Using over 70,000 non-contrast head CT scans with 75 annotated findings-including hemorrhage, infarction, trauma, mass lesions, structural abnormalities, and chronic changes-our experiments demonstrate consistent performance improvements across all models. Gains vary substantially: general-purpose and medical-domain models show the largest improvements (DINOv3-Base: +4.6%; MedGemma: +4.3%), whereas 3D CT-specialized or very large models show more modest gains (+0.2-1.3%). The combination of MoLRE and MedGemma achieves the highest average detection AUC of 0.917. These findings highlight the importance of systematic benchmarking on target clinical tasks, as pretraining domain, architecture, and model scale interact in non-obvious ways.

Method: Three complementary designs: 1) Foreground-Aware Point Sampling (FAPS) for selective transmission, 2) Completion-Enhanced Early Fusion (CEEF) to reconstruct dense pillars from sparse inputs, 3) Dense-Guided Dual Alignment (DGDA) for semantic/geometric consistency during training

Result: Achieves superior perception-communication trade-offs and remains robust under heterogeneous model settings on both simulated and real-world datasets

Conclusion: CoLC enables practical deployment of early fusion collaborative perception by addressing communication efficiency while maintaining perceptual benefits

Abstract: Collaborative perception empowers autonomous agents to share complementary information and overcome perception limitations. While early fusion offers more perceptual complementarity and is inherently robust to model heterogeneity, its high communication cost has limited its practical deployment, prompting most existing works to favor intermediate or late fusion. To address this, we propose a communication-efficient early Collaborative perception framework that incorporates LiDAR Completion to restore scene completeness under sparse transmission, dubbed as CoLC. Specifically, the CoLC integrates three complementary designs. First, each neighbor agent applies Foreground-Aware Point Sampling (FAPS) to selectively transmit informative points that retain essential structural and contextual cues under bandwidth constraints. The ego agent then employs Completion-Enhanced Early Fusion (CEEF) to reconstruct dense pillars from the received sparse inputs and adaptively fuse them with its own observations, thereby restoring spatial completeness. Finally, the Dense-Guided Dual Alignment (DGDA) strategy enforces semantic and geometric consistency between the enhanced and dense pillars during training, ensuring consistent and robust feature learning. Experiments on both simulated and real-world datasets demonstrate that CoLC achieves superior perception-communication trade-offs and remains robust under heterogeneous model settings. The code is available at https://github.com/CatOneTwo/CoLC.

Method: Proposes a reconstruction method that uses spatial nonlocal similarity and conjugate properties of the projection domain to generate pseudo-3D data for self-supervised training, requiring no external data and avoiding lengthy pre-training.

Result: Extensive experiments show the method mitigates detector-induced ring artifacts and exhibits unprecedented capabilities in detail recovery, achieving high-fidelity results in very short time.

Conclusion: The method provides a new paradigm for research using unlabeled raw projection data in medical imaging reconstruction.

Abstract: Noise and artifacts during computed tomography (CT) scans are a fundamental challenge affecting disease diagnosis. However, current methods either involve excessively long reconstruction times or rely on data-driven models for optimization, failing to adequately consider the valuable information inherent in the data itself, especially medical 3D data. This work proposes a reconstruction method under ultra-low raw data conditions, requiring no external data and avoiding lengthy pre-training processes. By leveraging spatial nonlocal similarity and the conjugate properties of the projection domain to generate pseudo-3D data for self-supervised training, high-fidelity results can be achieved in a very short time. Extensive experiments demonstrate that this method not only mitigates detector-induced ring artifacts but also exhibits unprecedented capabilities in detail recovery. This method provides a new paradigm for research using unlabeled raw projection data. Code is available at https://github.com/yqx7150/SCOUT.

Method: Three-component framework: 1) Segmentation-Guided Feature Modulation (SFM) uses SAM-generated masks for foreground enhancement and background suppression via learnable attention modulation; 2) Semantic Token Reallocation (STR) employs learnable query tokens with adaptive reallocation to extract compact representations without token discarding; 3) Cross-Modal Hypergraph Interaction (CHI) constructs unified hypergraph across modalities to capture high-order semantic relationships.

Result: Extensive experiments on RGBNT201, RGBNT100, and MSVR310 benchmarks demonstrate the effectiveness and robustness of STMI in multi-modal ReID scenarios.

Conclusion: STMI provides an effective solution for multi-modal object ReID by addressing limitations of existing methods through integrated segmentation guidance, token reallocation, and cross-modal hypergraph modeling.

Abstract: Multi-modal object Re-Identification (ReID) aims to exploit complementary information from different modalities to retrieve specific objects. However, existing methods often rely on hard token filtering or simple fusion strategies, which can lead to the loss of discriminative cues and increased background interference. To address these challenges, we propose STMI, a novel multi-modal learning framework consisting of three key components: (1) Segmentation-Guided Feature Modulation (SFM) module leverages SAM-generated masks to enhance foreground representations and suppress background noise through learnable attention modulation; (2) Semantic Token Reallocation (STR) module employs learnable query tokens and an adaptive reallocation mechanism to extract compact and informative representations without discarding any tokens; (3) Cross-Modal Hypergraph Interaction (CHI) module constructs a unified hypergraph across modalities to capture high-order semantic relationships. Extensive experiments on public benchmarks (i.e., RGBNT201, RGBNT100, and MSVR310) demonstrate the effectiveness and robustness of our proposed STMI framework in multi-modal ReID scenarios.

Method: Introduces Token-aligned Gaussian Prediction module that aligns semantically corresponding information across views in feature space, guided by coarse token positions and fusion confidence. Uses learnable camera tokens and Asymmetric Dual-Flow Decoder (ADF-Decoder) to enforce directionally constrained communication between camera and image tokens, maintaining clean factorization.

Result: Achieves higher reconstruction fidelity and novel-view synthesis quality in pose-free settings, and significantly improves pose estimation accuracy compared to prior pose-free methods.

Conclusion: TokenSplat demonstrates that joint 3D reconstruction and camera pose estimation can be effectively achieved through a feed-forward architecture with proper token alignment and factorization mechanisms, without requiring iterative refinement.

Abstract: We present TokenSplat, a feed-forward framework for joint 3D Gaussian reconstruction and camera pose estimation from unposed multi-view images. At its core, TokenSplat introduces a Token-aligned Gaussian Prediction module that aligns semantically corresponding information across views directly in the feature space. Guided by coarse token positions and fusion confidence, it aggregates multi-scale contextual features to enable long-range cross-view reasoning and reduce redundancy from overlapping Gaussians. To further enhance pose robustness and disentangle viewpoint cues from scene semantics, TokenSplat employs learnable camera tokens and an Asymmetric Dual-Flow Decoder (ADF-Decoder) that enforces directionally constrained communication between camera and image tokens. This maintains clean factorization within a feed-forward architecture, enabling coherent reconstruction and stable pose estimation without iterative refinement. Extensive experiments demonstrate that TokenSplat achieves higher reconstruction fidelity and novel-view synthesis quality in pose-free settings, and significantly improves pose estimation accuracy compared to prior pose-free methods. Project page: https://kidleyh.github.io/tokensplat/.

Method: Proposes a universal Khmer text recognition (UKTR) framework with modality-aware adaptive feature selection (MAFS) that adapts visual features according to input image modality to enhance recognition robustness across different text modalities.

Result: The model achieves state-of-the-art performance and the authors introduce the first comprehensive benchmark for universal Khmer text recognition, releasing datasets and models to the community.

Conclusion: The proposed universal framework effectively addresses modality-specific challenges in Khmer text recognition through adaptive feature selection, enabling robust performance across diverse text modalities while facilitating future research with released benchmarks.

Abstract: Khmer is a low-resource language characterized by a complex script, presenting significant challenges for optical character recognition (OCR). While document printed text recognition has advanced because of available datasets, performance on other modalities, such as handwritten and scene text, remains limited by data scarcity. Training modality-specific models for each modality does not allow cross-modality transfer learning, from which modalities with limited data could otherwise benefit. Moreover, deploying many modality-specific models results in significant memory overhead and requires error-prone routing each input image to the appropriate model. On the other hand, simply training on a combined dataset with a non-uniform data distribution across different modalities often leads to degraded performance on underrepresented modalities. To address these, we propose a universal Khmer text recognition (UKTR) framework capable of handling diverse text modalities. Central to our method is a novel modality-aware adaptive feature selection (MAFS) technique designed to adapt visual features according to a particular input image modality and enhance recognition robustness across modalities. Extensive experiments demonstrate that our model achieves state-of-the-art (SoTA) performance. Furthermore, we introduce the first comprehensive benchmark for universal Khmer text recognition, which we release to the community to facilitate future research. Our datasets and models can be accessible via this gated repository\footnote{in review}.

Method: Three key elements: (1) robust training/benchmarking dataset for Khmer scene layouts, (2) open-source document augmentation tool for synthesizing realistic scene documents to scale training data, (3) layout detection baselines using YOLO-based architectures with oriented bounding boxes (OBB) to handle geometric distortions.

Result: The paper presents the first comprehensive framework for Khmer scene document layout detection, releasing models, code, and datasets to foster research in Khmer document analysis and recognition community.

Conclusion: Addresses critical gap in Khmer document layout analysis through a complete framework combining dataset creation, data augmentation, and detection models, enabling future research in this underserved language domain.

Abstract: While document layout analysis for Latin scripts has advanced significantly, driven by the advent of large multimodal models (LMMs), progress for the Khmer language remains constrained because of the scarcity of annotated training data. This gap is particularly acute for scene documents, where perspective distortions and complex backgrounds challenge traditional methods. Given the structural complexities of Khmer script, such as diacritics and multi-layer character stacking, existing Latin-based layout analysis models fail to accurately delineate semantic layout units, particularly for dense text regions (e.g., list items). In this paper, we present the first comprehensive study on Khmer scene document layout detection. We contribute a novel framework comprising three key elements: (1) a robust training and benchmarking dataset specifically for Khmer scene layouts; (2) an open-source document augmentation tool capable of synthesizing realistic scene documents to scale training data; and (3) layout detection baselines utilizing YOLO-based architectures with oriented bounding boxes (OBB) to handle geometric distortions. To foster further research in the Khmer document analysis and recognition (DAR) community, we release our models, code, and datasets in this gated repository (in review).

Method: System extracts key frames from endoscope video, applies polar coordinate transformation to unwrap annular frames, and uses image stitching techniques to create planar panoramic images of pipeline inner walls.

Result: The method enables efficient processing of industrial endoscope videos, generating panoramic stitched images that preserve all detailed features of pipeline inner walls for defect detection.

Conclusion: The approach significantly improves pipeline inner wall reconstruction efficiency compared to traditional methods and has considerable engineering application value for industrial inspection.

Abstract: Visual analysis and reconstruction of pipeline inner walls remain challenging in industrial inspection scenarios. This paper presents a dedicated reconstruction system for pipeline inner walls via industrial endoscopes, which is built on panoramic image stitching technology. Equipped with a custom graphical user interface (GUI), the system extracts key frames from endoscope video footage, and integrates polar coordinate transformation with image stitching techniques to unwrap annular video frames of pipeline inner walls into planar panoramic images. Experimental results demonstrate that the proposed method enables efficient processing of industrial endoscope videos, and the generated panoramic stitched images preserve all detailed features of pipeline inner walls in their entirety. This provides intuitive and accurate visual support for defect detection and condition assessment of pipeline inner walls. In comparison with the traditional frame-by-frame video review method, the proposed approach significantly elevates the efficiency of pipeline inner wall reconstruction and exhibits considerable engineering application value.

Method: Proposes anti-feature-collapse learning framework that filters task-irrelevant components and suppresses excessive overlap among different forgery cues in representation space to prevent discriminative information collapse.

Result: Significantly outperforms state-of-the-art in cross-model scenarios with 5.02% accuracy improvement, demonstrating superior generalization and detection reliability on multiple public benchmarks.

Conclusion: The framework maintains diverse complementary evidence, reduces reliance on limited salient cues, and enhances robustness under unseen generative settings for reliable generated image detection.

Abstract: With the rapid advancement of generative models, generated image detection has become an important task in visual forensics. Although existing methods have achieved remarkable progress, they often rely, after training, on only a small subset of highly salient forgery cues, which limits their ability to generalize to unseen generative mechanisms. We argue that reliably generated image detection should not depend on a single decision path but should preserve multiple judgment perspectives, enabling the model to understand the differences between real and generated images from diverse viewpoints. Based on this idea, we propose an anti-feature-collapse learning framework that filters task-irrelevant components and suppresses excessive overlap among different forgery cues in the representation space, preventing discriminative information from collapsing into a few dominant feature directions. This design maintains diverse and complementary evidence within the model, reduces reliance on a small set of salient cues, and enhances robustness under unseen generative settings. Extensive experiments on multiple public benchmarks demonstrate that the proposed method significantly outperforms the state-of-the-art approaches in cross-model scenarios, achieving an accuracy improvement of 5.02% and exhibiting superior generalization and detection reliability. The source code is available at https://github.com/Yanmou-Hui/DoU.

Method: Proposes BornoViT, a novel lightweight Vision Transformer that uses a simplified deep convolutional neural network (DCNN) architecture to reduce computational burden while maintaining effectiveness for Bengali character classification.

Result: Achieved 95.77% accuracy on BanglaLekha Isolated dataset and 91.51% on self-collected Bornomala dataset (222 samples). Model has only 0.65M parameters, 0.62MB size, and 0.16 GFLOPs.

Conclusion: BornoViT is significantly lighter than state-of-the-art models while maintaining competitive accuracy, making it suitable for resource-limited environments essential for Bengali handwritten character classification.

Abstract: Handwritten character classification in the Bengali script is a significant challenge due to the complexity and variability of the characters. The models commonly used for classification are often computationally expensive and data-hungry, making them unsuitable for resource-limited languages such as Bengali. In this experiment, we propose a novel, efficient, and lightweight Vision Transformer model that effectively classifies Bengali handwritten basic characters and digits, addressing several shortcomings of traditional methods. The proposed solution utilizes a deep convolutional neural network (DCNN) in a more simplified manner compared to traditional DCNN architectures, with the aim of reducing computational burden. With only 0.65 million parameters, a model size of 0.62 MB, and 0.16 GFLOPs, our model, BornoViT, is significantly lighter than current state-of-the-art models, making it more suitable for resource-limited environments, which is essential for Bengali handwritten character classification. BornoViT was evaluated on the BanglaLekha Isolated dataset, achieving an accuracy of 95.77%, and demonstrating superior efficiency compared to existing state-of-the-art approaches. Furthermore, the model was evaluated on our self-collected dataset, Bornomala, consisting of approximately 222 samples from different age groups, where it achieved an accuracy of 91.51%.

Method: Apply diffusion probabilistic models to generate self-supervised stroke representations from CT images. Extend to incorporate longitudinal images and time-from-stroke-onset information for improved representations.

Result: Achieves best performance for predicting next-day severity and functional outcome at discharge on a dataset of 5,824 CT images from 3,573 patients across two medical centers with minimal labels.

Conclusion: Diffusion models with longitudinal and temporal extensions provide effective self-supervised stroke representations that outperform existing methods for clinical outcome prediction.

Abstract: Stroke is a major cause of death and disability worldwide. Accurate outcome and evolution prediction has the potential to revolutionize stroke care by individualizing clinical decision-making leading to better outcomes. However, despite a plethora of attempts and the rich data provided by neuroimaging, modelling the ultimate fate of brain tissue remains a challenging task. In this work, we apply recent ideas in the field of diffusion probabilistic models to generate a self-supervised semantically meaningful stroke representation from Computed Tomography (CT) images. We then improve this representation by extending the method to accommodate longitudinal images and the time from stroke onset. The effectiveness of our approach is evaluated on a dataset consisting of 5,824 CT images from 3,573 patients across two medical centers with minimal labels. Comparative experiments show that our method achieves the best performance for predicting next-day severity and functional outcome at discharge.

Method: Systematically elucidates design space of sampling acceleration methods, identifies sampling time schedule as most pivotal factor. Inspired by geometric properties revealed through Frenet-Serret formulas, proposes constant total rotation schedule (TORS) that ensures uniform geometric variation along sampling trajectory.

Result: TORS outperforms previous training-free acceleration methods and produces high-quality images with only 10 sampling steps on Flux.1-Dev and Stable Diffusion 3.5. Shows adaptability to unseen models, hyperparameters, and downstream applications.

Conclusion: The sampling time schedule is crucial for diffusion model acceleration. TORS provides an effective geometric scheduling strategy that enables high-quality image generation with significantly fewer sampling steps while maintaining compatibility across different models and applications.

Abstract: Text-to-image diffusion models have achieved unprecedented success but still struggle to produce high-quality results under limited sampling budgets. Existing training-free sampling acceleration methods are typically developed independently, leaving the overall performance and compatibility among these methods unexplored. In this paper, we bridge this gap by systematically elucidating the design space, and our comprehensive experiments identify the sampling time schedule as the most pivotal factor. Inspired by the geometric properties of diffusion models revealed through the Frenet-Serret formulas, we propose constant total rotation schedule (TORS), a scheduling strategy that ensures uniform geometric variation along the sampling trajectory. TORS outperforms previous training-free acceleration methods and produces high-quality images with 10 sampling steps on Flux.1-Dev and Stable Diffusion 3.5. Extensive experiments underscore the adaptability of our method to unseen models, hyperparameters, and downstream applications.

Method: Introduces DUCK (Decomposing Unfairness in Chest X-ray agents) with a stage-wise fairness decomposition that separates end-to-end bias from three agent-specific sources: tool exposure bias (utility gaps conditioned on tool presence), tool transition bias (subgroup differences in tool-routing patterns), and model reasoning bias (subgroup differences in synthesis behaviors). Uses MedRAX framework with five driver backbones.

Result: Demographic gaps persist in end-to-end performance (equalized odds up to 20.79%, lowest fairness-utility tradeoff down to 28.65%). Intermediate behaviors, tool usage, transition patterns, and reasoning traces exhibit distinct subgroup disparities not predictable from end-to-end evaluation alone (e.g., conditioned on segmentation-tool availability, subgroup utility gap reaches 50%).

Conclusion: Findings underscore the need for process-level fairness auditing and debiasing to ensure equitable deployment of clinical agentic systems, as agent-specific biases emerge beyond standalone model evaluation.

Abstract: Tool-using medical agents can improve chest X-ray question answering by orchestrating specialized vision and language modules, but this added pipeline complexity also creates new pathways for demographic bias beyond standalone models. We present ours (Decomposing Unfairness in Chest X-ray agents), a systematic audit of chest X-ray agents instantiated with MedRAX. To localize where disparities arise, we introduce a stage-wise fairness decomposition that separates end-to-end bias from three agent-specific sources: tool exposure bias (utility gaps conditioned on tool presence), tool transition bias (subgroup differences in tool-routing patterns), and model reasoning bias (subgroup differences in synthesis behaviors). Extensive experiments on tool-used based agentic frameworks across five driver backbones reveal that (i) demographic gaps persist in end-to-end performance, with equalized odds up to 20.79%, and the lowest fairness-utility tradeoff down to 28.65%, and (ii) intermediate behaviors, tool usage, transition patterns, and reasoning traces exhibit distinct subgroup disparities that are not predictable from end-to-end evaluation alone (e.g., conditioned on segmentation-tool availability, the subgroup utility gap reaches as high as 50%). Our findings underscore the need for process-level fairness auditing and debiasing to ensure the equitable deployment of clinical agentic systems. Code is available here: https://anonymous.4open.science/r/DUCK-E5FE/README.md

Method: Models layer-wise embeddings as depth-wise dynamical trajectories, applies Dynamic Mode Decomposition to extract stable modes, projects representations into biologically anchored coordinate system defined by distributed neural responses, and introduces Signal-to-Noise Consistency Index for cross-model consistency quantification.

Result: Across 45 pretrained models spanning vision, audio, and language, NFAS reveals structured organization including modality-specific clustering and cross-modal convergence in integrative cortical systems.

Conclusion: Representation dynamics provide a principled basis for functional alignment of artificial neural networks, enabling cross-modal comparison in brain-referenced space.

Abstract: We propose the Neural Functional Alignment Space (NFAS), a brain-referenced representational framework for characterizing artificial neural networks on equal functional grounds. NFAS departs from conventional alignment approaches that rely on layer-wise features or task-specific activations by modeling the intrinsic dynamical evolution of stimulus representations across network depth. Specifically, we model layer-wise embeddings as a depth-wise dynamical trajectory and apply Dynamic Mode Decomposition (DMD) to extract the stable mode. This representation is then projected into a biologically anchored coordinate system defined by distributed neural responses. We also introduce the Signal-to-Noise Consistency Index (SNCI) to quantify cross-model consistency at the modality level. Across 45 pretrained models spanning vision, audio, and language, NFAS reveals structured organization within this brain-referenced space, including modality-specific clustering and cross-modal convergence in integrative cortical systems. Our findings suggest that representation dynamics provide a principled basis for

Method: Uses a 1.2B parameter Diffusion Transformer conditioned on latent representations from a deep compression autoencoder. Employs state-of-the-art techniques including causal distillation and diffusion forcing for real-time inference. Trained solely on single-player inputs without explicit opponent policy supervision, learning from partially observed data.

Result: Successfully demonstrates that diffusion models can simulate dynamic opponents that react to player actions, with emergent sophisticated agent behavior. Achieves real-time inference and learns effectively from partially observed data to generate responsive behaviors for controllable Player 1.

Conclusion: COMBAT establishes a foundation for training interactive agents within diffusion-based world models, showing that diffusion models can learn emergent agent behavior without explicit supervision, opening new possibilities for interactive AI systems.

Abstract: Recent advances in video generation have spurred the development of world models capable of simulating 3D-consistent environments and interactions with static objects. However, a significant limitation remains in their ability to model dynamic, reactive agents that can intelligently influence and interact with the world. To address this gap, we introduce COMBAT, a real-time, action-controlled world model trained on the complex 1v1 fighting game Tekken 3. Our work demonstrates that diffusion models can successfully simulate a dynamic opponent that reacts to player actions, learning its behavior implicitly. Our approach utilizes a 1.2 billion parameter Diffusion Transformer, conditioned on latent representations from a deep compression autoencoder. We employ state-of-the-art techniques, including causal distillation and diffusion forcing, to achieve real-time inference. Crucially, we observe the emergence of sophisticated agent behavior by training the model solely on single-player inputs, without any explicit supervision for the opponent’s policy. Unlike traditional imitation learning methods, which require complete action labels, COMBAT learns effectively from partially observed data to generate responsive behaviors for a controllable Player 1. We present an extensive study and introduce novel evaluation methods to benchmark this emergent agent behavior, establishing a strong foundation for training interactive agents within diffusion-based world models.

Method: Proposes a Mixture of Experts framework with: (1) novel gate architecture using random walks to capture regions experts attend to, (2) attention mechanism for focusing on informative areas, and (3) dynamic loss balancing scheme adjusting trade-off between diversity and similarity losses during training.

Result: Achieves state-of-the-art results in mesh classification, retrieval, and semantic segmentation tasks, demonstrating the effectiveness of combining multiple mesh analysis approaches.

Conclusion: The proposed MoE framework successfully harnesses complementary strengths of diverse mesh analysis methods through specialized gating and balanced training, advancing 3D mesh understanding capabilities.

Abstract: In recent years, various methods have been proposed for mesh analysis, each offering distinct advantages and often excelling on different object classes. We present a novel Mixture of Experts (MoE) framework designed to harness the complementary strengths of these diverse approaches. We propose a new gate architecture that encourages each expert to specialise in the classes it excels in. Our design is guided by two key ideas: (1) random walks over the mesh surface effectively capture the regions that individual experts attend to, and (2) an attention mechanism that enables the gate to focus on the areas most informative for each expert’s decision-making. To further enhance performance, we introduce a dynamic loss balancing scheme that adjusts a trade-off between diversity and similarity losses throughout the training, where diversity prompts expert specialization, and similarity enables knowledge sharing among the experts. Our framework achieves state-of-the-art results in mesh classification, retrieval, and semantic segmentation tasks. Our code is available at: https://github.com/amirbelder/MME-Mixture-of-Mesh-Experts.

Method: Proposes Neural Discrimination Priors (NDP) to measure feature differences, Neural Discrimination-Prompted Attention (NDPA) to incorporate NDP into attention mechanisms, Neural Discrimination-Prompted Network (NDPN) with continuous gating guided by NDP, and a super-resolution-guided reconstruction approach.

Result: UHDPromer achieves state-of-the-art performance on three UHD image restoration tasks (low-light enhancement, dehazing, deblurring) while maintaining the best computational efficiency compared to other methods.

Conclusion: The proposed UHDPromer effectively leverages neural discrimination priors for UHD image restoration, demonstrating superior performance and efficiency across multiple enhancement tasks.

Abstract: We propose a simple yet effective UHDPromer, a neural discrimination-prompted Transformer, for Ultra-High-Definition (UHD) image restoration and enhancement. Our UHDPromer is inspired by an interesting observation that there implicitly exist neural differences between high-resolution and low-resolution features, and exploring such differences can facilitate low-resolution feature representation. To this end, we first introduce Neural Discrimination Priors (NDP) to measure the differences and then integrate NDP into the proposed Neural Discrimination-Prompted Attention (NDPA) and Neural Discrimination-Prompted Network (NDPN). The proposed NDPA re-formulates the attention by incorporating NDP to globally perceive useful discrimination information, while the NDPN explores a continuous gating mechanism guided by NDP to selectively permit the passage of beneficial content. To enhance the quality of restored images, we propose a super-resolution-guided reconstruction approach, which is guided by super-resolving low-resolution features to facilitate final UHD image restoration. Experiments show that UHDPromer achieves the best computational efficiency while still maintaining state-of-the-art performance on $3$ UHD image restoration and enhancement tasks, including low-light image enhancement, image dehazing, and image deblurring. The source codes and pre-trained models will be made available at https://github.com/supersupercong/uhdpromer.

Method: Proposes post-hoc alignment pipeline to map vision encoder representations into SONAR space, then extends to V-LCM which encodes vision/language inputs into unified latent embeddings and trains with latent diffusion objective for next-embedding prediction.

Result: V-SONAR achieves competitive text-to-video retrieval and surpasses SOTA on video captioning (DREAM-1K: BLEU 23.9 vs 19.6; PE-VIDEO: BLEU 39.0 vs 30.0). V-LCM matches SOTA on English tasks while significantly outperforming across 61 out of 62 tested languages.

Conclusion: The approach successfully extends multilingual text capabilities to vision, demonstrating effective zero-shot multimodal understanding and strong performance across diverse languages, highlighting the potential of unified multimodal embedding spaces.

Abstract: We introduce V-SONAR, a vision-language embedding space extended from the text-only embedding space SONAR (Omnilingual Embeddings Team et al., 2026), which supports 1500 text languages and 177 speech languages. To construct V-SONAR, we propose a post-hoc alignment pipeline that maps the representations of an existing vision encoder into the SONAR space. We thoroughly evaluate V-SONAR and show that its embeddings achieve competitive performance on text-to-video retrieval. Equipped with the OMNISONAR text decoder, V-SONAR further surpasses state-of-the-art vision-language models on video captioning tasks, including DREAM-1K (BLEU 23.9 vs. 19.6) and PE-VIDEO (BLEU 39.0 vs. 30.0). Leveraging V-SONAR, we first demonstrate that the Large Concept Model (LCM; LCM team et al. 2024) operating in SONAR and trained with English text only, can perform both single- and multi-visual concept understanding in a zero-shot manner. Finally, we introduce V-LCM, which extends the LCM with vision-language instruction tuning. V-LCM encodes vision and language inputs into an unified sequence of latent embeddings via V-SONAR and SONAR, and it is trained with the same latent diffusion objective for next-embedding prediction as in LCM’s text-only pre-training. Experiments on a large-scale multilingual and -modal instruction-tuning data mixture highlight the potential of V-LCM: V-LCM matches state-of-the-art vision-language models on tasks covering image/video captioning and question answering, while significantly outperforming them across 61 rich- to low-resource languages out of all 62 tested languages.

Method: Proposes PPC-MT, a parallel framework using hybrid Mamba-Transformer architecture. Introduces PCA-guided parallel completion strategy that imposes geometric structure on unordered point clouds, transforming them into ordered sets decomposed into multiple subsets. These subsets are reconstructed in parallel using multi-head reconstructor. Combines Mamba’s linear complexity for efficient encoding with Transformer’s fine-grained multi-sequence modeling for decoding.

Result: Extensive experiments on PCN, ShapeNet-55/34, and KITTI datasets show PPC-MT outperforms state-of-the-art methods across multiple metrics. Achieves better uniformity of point distribution and detail fidelity while preserving computational efficiency.

Conclusion: PPC-MT effectively balances efficiency and reconstruction accuracy through its parallel hybrid architecture and PCA-guided structured synthesis, offering a promising solution for point cloud completion tasks.

Abstract: Existing point cloud completion methods struggle to balance high-quality reconstruction with computational efficiency. To address this, we propose PPC-MT, a novel parallel framework for point cloud completion leveraging a hybrid Mamba-Transformer architecture. Our approach introduces an innovative parallel completion strategy guided by Principal Component Analysis (PCA), which imposes a geometrically meaningful structure on unordered point clouds, transforming them into ordered sets and decomposing them into multiple subsets. These subsets are reconstructed in parallel using a multi-head reconstructor. This structured parallel synthesis paradigm significantly enhances the uniformity of point distribution and detail fidelity, while preserving computational efficiency. By integrating Mamba’s linear complexity for efficient feature extraction during encoding with the Transformer’s capability to model fine-grained multi-sequence relationships during decoding, PPC-MT effectively balances efficiency and reconstruction accuracy. Extensive quantitative and qualitative experiments on benchmark datasets, including PCN, ShapeNet-55/34, and KITTI, demonstrate that PPC-MT outperforms state-of-the-art methods across multiple metrics, validating the efficacy of our proposed framework.

Method: Multi-Membership Temporal Attention (MMTA) allows each frame to attend to multiple locally normalized temporal attention windows within the same layer, fusing concurrent temporal views via feature-space overlap resolution to preserve competing local contexts near transitions.

Result: MMTA consistently improves over Global Attention transformer, boosting Edit Score by +1.3 (Video) and +1.6 (IMU) on StrokeRehab dataset, and further improving 50Salads by +3.3.

Conclusion: MMTA offers a practical solution for resource-constrained rehabilitation assessment by improving boundary sensitivity without additional depth or multi-stage refinement, supporting both video and wearable IMU inputs in a unified architecture.

Abstract: To empower the iterative assessments involved during a person’s rehabilitation, automated assessment of a person’s abilities during daily activities requires temporally precise segmentation of fine-grained actions in therapy videos. Existing temporal action segmentation (TAS) models struggle to capture sub-second micro-movements while retaining exercise context, blurring rapid phase transitions and limiting reliable downstream assessment of motor recovery. We introduce Multi-Membership Temporal Attention (MMTA), a high-resolution temporal transformer for fine-grained rehabilitation assessment. Unlike standard temporal attention, which assigns each frame a single attention context per layer, MMTA lets each frame attend to multiple locally normalized temporal attention windows within the same layer. We fuse these concurrent temporal views via feature-space overlap resolution, preserving competing local contexts near transitions while enabling longer-range reasoning through layer-wise propagation. This increases boundary sensitivity without additional depth or multi-stage refinement. MMTA supports both video and wearable IMU inputs within a unified single-stage architecture, making it applicable to both clinical and home settings. MMTA consistently improves over the Global Attention transformer, boosting Edit Score by +1.3 (Video) and +1.6 (IMU) on StrokeRehab while further improving 50Salads by +3.3. Ablations confirm that performance gains stem from multi-membership temporal views rather than architectural complexity, offering a practical solution for resource-constrained rehabilitation assessment.

Method: Proposes SMART: 1) Uses SAM3’s promptable concept segmentation design in a teacher-student framework, 2) Integrates vessel mask warping and motion consistency loss to model complex vessel dynamics, 3) Implements progressive confidence-aware consistency regularization to mitigate unreliable teacher predictions caused by blurred boundaries and minimal contrast.

Result: Extensive experiments on three XCA datasets from different institutions show SMART achieves state-of-the-art performance while requiring significantly fewer annotations, making it valuable for clinical applications with scarce labeled data.

Conclusion: SMART effectively addresses challenges in XCA vessel segmentation through its innovative SAM3-based teacher-student framework with motion-aware consistency and progressive confidence regularization, demonstrating strong performance with limited annotations for real-world clinical use.

Abstract: Segmentation of the main coronary artery from X-ray coronary angiography (XCA) sequences is crucial for the diagnosis of coronary artery diseases. However, this task is challenging due to issues such as blurred boundaries, inconsistent radiation contrast, complex motion patterns, and a lack of annotated images for training. Although Semi-Supervised Learning (SSL) can alleviate the annotation burden, conventional methods struggle with complicated temporal dynamics and unreliable uncertainty quantification. To address these challenges, we propose SAM3-based Teacher-student framework with Motion-Aware consistency and Progressive Confidence Regularization (SMART), a semi-supervised vessel segmentation approach for X-ray angiography videos. First, our method utilizes SAM3’s unique promptable concept segmentation design and innovates a SAM3-based teacher-student framework to maximize the performance potential of both the teacher and the student. Second, we enhance segmentation by integrating the vessel mask warping technique and motion consistency loss to model complex vessel dynamics. To address the issue of unreliable teacher predictions caused by blurred boundaries and minimal contrast, we further propose a progressive confidence-aware consistency regularization to mitigate the risk of unreliable outputs. Extensive experiments on three datasets of XCA sequences from different institutions demonstrate that SMART achieves state-of-the-art performance while requiring significantly fewer annotations, making it particularly valuable for real-world clinical applications where labeled data is scarce. Our code is available at: https://github.com/qimingfan10/SMART.

Method: Proposes VEMamba with: 1) 3D Dependency Reordering paradigm via Axial-Lateral Chunking Selective Scan Module (ALCSSM) that remaps 3D spatial dependencies into optimized 1D sequences for Mamba-based modeling, 2) Dynamic Weights Aggregation Module (DWAM) for adaptive output aggregation, and 3) realistic degradation simulation with Momentum Contrast (MoCo) for degradation-aware training.

Result: Extensive experiments on simulated and real-world anisotropic VEM datasets show VEMamba achieves highly competitive performance across various metrics while maintaining lower computational footprint compared to existing methods.

Conclusion: VEMamba provides an efficient solution for isotropic reconstruction of anisotropic VEM data through innovative 3D dependency modeling and degradation-aware training, demonstrating superior performance with reduced computational requirements.

Abstract: Volume Electron Microscopy (VEM) is crucial for 3D tissue imaging but often produces anisotropic data with poor axial resolution, hindering visualization and downstream analysis. Existing methods for isotropic reconstruction often suffer from neglecting abundant axial information and employing simple downsampling to simulate anisotropic data. To address these limitations, we propose VEMamba, an efficient framework for isotropic reconstruction. The core of VEMamba is a novel 3D Dependency Reordering paradigm, implemented via two key components: an Axial-Lateral Chunking Selective Scan Module (ALCSSM), which intelligently re-maps complex 3D spatial dependencies (both axial and lateral) into optimized 1D sequences for efficient Mamba-based modeling, explicitly enforcing axial-lateral consistency; and a Dynamic Weights Aggregation Module (DWAM) to adaptively aggregate these reordered sequence outputs for enhanced representational power. Furthermore, we introduce a realistic degradation simulation and then leverage Momentum Contrast (MoCo) to integrate this degradation-aware knowledge into the network for superior reconstruction. Extensive experiments on both simulated and real-world anisotropic VEM datasets demonstrate that VEMamba achieves highly competitive performance across various metrics while maintaining a lower computational footprint. The source code is available on GitHub: https://github.com/I2-Multimedia-Lab/VEMamba

Method: A visual programming framework where MLLMs generate Python code to interface with spatial tools (3D reconstruction, camera-pose recovery, novel-view rendering) that convert 2D images into explorable 3D scenes for explicit spatial reasoning.

Result: Outperforms GPT-4.1-mini by 12.94% on MindCube benchmark, achieves strong results on Omni3D-Bench, and enables successful real-world indoor navigation in complex environments.

Conclusion: pySpatial effectively enhances MLLMs’ 3D spatial reasoning capabilities through code-based tool interfacing without fine-tuning, demonstrating practical applications in robotics and spatial understanding tasks.

Abstract: Multi-modal Large Language Models (MLLMs) have demonstrated strong capabilities in general-purpose perception and reasoning, but they still struggle with tasks that require spatial understanding of the 3D world. To address this, we introduce pySpatial, a visual programming framework that equips MLLMs with the ability to interface with spatial tools via Python code generation. Given an image sequence and a natural-language query, the model composes function calls to spatial tools including 3D reconstruction, camera-pose recovery, novel-view rendering, etc. These operations convert raw 2D inputs into an explorable 3D scene, enabling MLLMs to reason explicitly over structured spatial representations. Notably, pySpatial requires no gradient-based fine-tuning and operates in a fully zero-shot setting. Experimental evaluations on the challenging MindCube and Omni3D-Bench benchmarks demonstrate that our framework pySpatial consistently surpasses strong MLLM baselines; for instance, it outperforms GPT-4.1-mini by 12.94% on MindCube. Furthermore, we conduct real-world indoor navigation experiments where the robot can successfully traverse complex environments using route plans generated by pySpatial, highlighting the practical effectiveness of our approach.

Method: Three complementary components: 1) Learnable Spatial Shift (LSS) module applies channel-wise spatial offsets to expand receptive field, 2) Asymmetric dual-branch architecture allocates more computation to information-dense branch for efficiency, 3) Error-bounded Adaptive Sampling (EAS) compresses feature-level LUTs to minimize storage.

Result: Achieves 3.8× larger receptive field than previous SOTA TinyLUT, improves average PSNR by over 0.21 dB across multiple benchmarks, while maintaining small storage size and inference time.

Conclusion: ShiftLUT demonstrates that efficient LUT-based image restoration with large receptive fields is achievable through complementary design of spatial shifting, asymmetric computation allocation, and feature compression.

Abstract: Look-Up Table based methods have emerged as a promising direction for efficient image restoration tasks. Recent LUT-based methods focus on improving their performance by expanding the receptive field. However, they inevitably introduce extra computational and storage overhead, which hinders their deployment in edge devices. To address this issue, we propose ShiftLUT, a novel framework that attains the largest receptive field among all LUT-based methods while maintaining high efficiency. Our key insight lies in three complementary components. First, Learnable Spatial Shift module (LSS) is introduced to expand the receptive field by applying learnable, channel-wise spatial offsets on feature maps. Second, we propose an asymmetric dual-branch architecture that allocates more computation to the information-dense branch, substantially reducing inference latency without compromising restoration quality. Finally, we incorporate a feature-level LUT compression strategy called Error-bounded Adaptive Sampling (EAS) to minimize the storage overhead. Compared to the previous state-of-the-art method TinyLUT, ShiftLUT achieves a 3.8$\times$ larger receptive field and improves an average PSNR by over 0.21 dB across multiple standard benchmarks, while maintaining a small storage size and inference time.

Method: Proposes UD-SfPNet, a unified network that jointly models polarization-based image descattering and shape-from-polarization normal estimation. Includes a color embedding module to enhance geometric consistency by relating color encodings to surface orientation, and a detail enhancement convolution module to preserve high-frequency geometric details lost under scattering.

Result: Achieves state-of-the-art performance on the MuS-Polar3D dataset with a mean surface normal angular error of 15.12°, the lowest among compared methods. Demonstrates significant improvement in reconstruction accuracy.

Conclusion: The unified approach combining descattering with polarization-based shape inference is effective for underwater 3D imaging. UD-SfPNet shows practical significance for optical 3D imaging in challenging underwater environments.

Abstract: Underwater optical imaging is severely hindered by scattering, but polarization imaging offers the unique dual advantages of descattering and shape-from-polarization (SfP) 3D reconstruction. To exploit these advantages, this paper proposes UD-SfPNet, an underwater descattering shape-from-polarization network that leverages polarization cues for improved 3D surface normal prediction. The framework jointly models polarization-based image descattering and SfP normal estimation in a unified pipeline, avoiding error accumulation from sequential processing and enabling global optimization across both tasks. UD-SfPNet further incorporates a novel color embedding module to enhance geometric consistency by exploiting the relationship between color encodings and surface orientation. A detail enhancement convolution module is also included to better preserve high-frequency geometric details that are lost under scattering. Experiments on the MuS-Polar3D dataset show that the proposed method significantly improves reconstruction accuracy, achieving a mean surface normal angular error of 15.12$^\circ$ (the lowest among compared methods). These results confirm the efficacy of combining descattering with polarization-based shape inference, and highlight the practical significance and potential applications of UD-SfPNet for optical 3D imaging in challenging underwater environments. The code is available at https://github.com/WangPuyun/UD-SfPNet.

Method: Hierarchical algorithm with composite discovery (dual inspection and breadth-first pruning) to identify rectangular regions with internal repetition, normalization to minimal representative form, and prime extraction (selective duplication and hierarchical memoization) to handle irregular dimensions efficiently.

Result: Scalable performance on grid sizes from 2x2 to 32x32: simple repeating tiles process in under 1ms, while complex patterns requiring exhaustive search show exponential growth. Provides deterministic behavior for exact, axis-aligned, rectangular tessellations.

Conclusion: The algorithm addresses a critical gap in symbolic grid analysis techniques, applicable to puzzle solving reasoning tasks and identification of exact repeating structures in discrete symbolic domains.

Abstract: The identification of repeating patterns in discrete grids is rudimentary within symbolic reasoning, algorithm synthesis and structural optimization across diverse computational domains. Although statistical approaches targeting noisy data can approximately recognize patterns, symbolic analysis utilizing deterministic extraction of periodic structures is underdeveloped. This paper aims to fill this gap by employing a hierarchical algorithm that discovers exact tessellations in finite planar grids, addressing the problem where multiple independent patterns may coexist within a hierarchical structure. The proposed method utilizes composite discovery (dual inspection and breadth-first pruning) for identifying rectangular regions with internal repetition, normalization to a minimal representative form, and prime extraction (selective duplication and hierarchical memoization) to account for irregular dimensions and to achieve efficient computation time. We evaluate scalability on grid sizes from 2x2 to 32x32, showing overlap detection on simple repeating tiles exhibits processing time under 1ms, while complex patterns which require exhaustive search and systematic exploration shows exponential growth. This algorithm provides deterministic behavior for exact, axis-aligned, rectangular tessellations, addressing a critical gap in symbolic grid analysis techniques, applicable to puzzle solving reasoning tasks and identification of exact repeating structures in discrete symbolic domains.

Method: Integrates VGGT encoder into transformer-based pipeline with two novel components: Attention-Guided Query Generation (uses VGGT attention maps as semantic priors to initialize object queries) and Query-Driven Feature Aggregation (learnable See-Query interacts with object queries to dynamically aggregate multi-level geometric features across VGGT layers).

Result: Significantly surpasses best-performing SG-Free method by 4.4 mAP@0.25 on ScanNet and 8.6 mAP@0.25 on ARKitScenes. Ablation shows VGGT’s learned semantic and geometric priors can be effectively leveraged.

Conclusion: VGGT-Det demonstrates effective sensor-geometry-free 3D object detection by integrating VGGT’s visual geometry understanding into a transformer pipeline with novel attention-guided and query-driven components.

Abstract: Current multi-view indoor 3D object detectors rely on sensor geometry that is costly to obtain (i.e., precisely calibrated multi-view camera poses) to fuse multi-view information into a global scene representation, limiting deployment in real-world scenes. We target a more practical setting: Sensor-Geometry-Free (SG-Free) multi-view indoor 3D object detection, where there are no sensor-provided geometric inputs (multi-view poses or depth). Recent Visual Geometry Grounded Transformer (VGGT) shows that strong 3D cues can be inferred directly from images. Building on this insight, we present VGGT-Det, the first framework tailored for SG-Free multi-view indoor 3D object detection. Rather than merely consuming VGGT predictions, our method integrates VGGT encoder into a transformer-based pipeline. To effectively leverage both the semantic and geometric priors from inside VGGT, we introduce two novel key components: (i) Attention-Guided Query Generation (AG): exploits VGGT attention maps as semantic priors to initialize object queries, improving localization by focusing on object regions while preserving global spatial structure; (ii) Query-Driven Feature Aggregation (QD): a learnable See-Query interacts with object queries to ‘see’ what they need, and then dynamically aggregates multi-level geometric features across VGGT layers that progressively lift 2D features into 3D. Experiments show that VGGT-Det significantly surpasses the best-performing method in the SG-Free setting by 4.4 and 8.6 mAP@0.25 on ScanNet and ARKitScenes, respectively. Ablation study shows that VGGT’s internally learned semantic and geometric priors can be effectively leveraged by our AG and QD.

Method: Train on 98k report-volume pairs from single hospital plus public data using SigLIP-style contrastive pretraining with prompt-based disease supervision. Automatically mine 262k snippet-slice pairs and introduce intra-scan snippet localization task to predict axial depth from text snippets.

Result: Achieves SOTA text-to-image retrieval (R@10 31.5 vs 22.2) and competitive disease classification (AUC 83.8 vs 83.8). Intra-scan localization reduces mean absolute error to 36.3mm at 12mm resolution vs 67.0mm baseline, with retrieval/classification unchanged.

Conclusion: The unified model effectively combines retrieval, classification, and intra-scan grounding capabilities, demonstrating that precise localization can be integrated without compromising other vision-language tasks in medical imaging.

Abstract: Recent 3D CT vision-language models align volumes with reports via contrastive pretraining, but typically rely on limited public data and provide only coarse global supervision. We train a 3D CT vision-language model on 98k report-volume pairs (50k patients) collected at a single hospital, combined with public datasets, using SigLIP-style contrastive pretraining together with prompt-based disease supervision in the shared vision-text embedding space. On CT-RATE, our model achieves state-of-the-art text-to-image retrieval (R@10 31.5 vs. 22.2) and competitive disease classification (AUC 83.8 vs. 83.8), with consistent results on Rad-ChestCT (AUC 77.0 vs. 77.3). We further observe that radiologists routinely reference specific images within their reports (e.g., ``series X, image Y’’), linking textual descriptions to precise axial locations. We automatically mine 262k such snippet-slice pairs and introduce the task of intra-scan snippet localization – predicting the axial depth referred to by a text snippet – reducing mean absolute error to 36.3 mm at 12 mm feature resolution, compared with 67.0 mm for the best baseline. Adding this localization objective leaves retrieval and classification broadly unchanged within confidence bounds, yielding a single unified model for retrieval, classification, and intra-scan grounding.

Method: ARC uses self-denoising probes to evaluate how accurately the model recovers injected noise, converting this intrinsic self-confidence signal into scalar rewards. The framework reinforces high-confidence generations without external supervision, and can also be integrated with external rewards for complementary improvements.

Result: ARC delivers consistent gains in compositional generation, text rendering, and text-image alignment over baseline models. When integrated with external rewards, it shows complementary improvement with alleviated reward hacking issues.

Conclusion: ARC provides an effective unsupervised post-training approach for text-to-image models that leverages internal self-confidence signals, offering improvements in key generation metrics and compatibility with external reward systems.

Abstract: Text-to-image generation powers content creation across design, media, and data augmentation. Post-training of text-to-image generative models is a promising path to better match human preferences, factuality, and improved aesthetics. We introduce ARC (Adaptive Rewarding by self-Confidence), a post-training framework that replaces external reward supervision with an internal self-confidence signal, obtained by evaluating how accurately the model recovers injected noise under self-denoising probes. ARC converts this intrinsic signal into scalar rewards, enabling fully unsupervised optimization without additional datasets, annotators, or reward models. Empirically, by reinforcing high-confidence generations, ARC delivers consistent gains in compositional generation, text rendering and text-image alignment over the baseline. We also find that integrating ARC with external rewards results in a complementary improvement, with alleviated reward hacking.

Method: DriveCode uses a number projector to map numbers into the language model’s hidden space as dedicated embeddings rather than discrete text tokens. This enables seamless integration with visual and textual features in a unified multimodal sequence for autonomous driving applications.

Result: Evaluated on OmniDrive, DriveGPT4, and DriveGPT4-V2 datasets, DriveCode demonstrates superior performance in trajectory prediction and control signal generation compared to traditional token-based numerical encoding methods.

Conclusion: DriveCode effectively addresses numerical precision limitations in LLM-based autonomous driving systems by representing numbers as dedicated embeddings, confirming its effectiveness for trajectory prediction and control signal generation tasks.

Abstract: Large language models (LLMs) have shown great promise for autonomous driving. However, discretizing numbers into tokens limits precise numerical reasoning, fails to reflect the positional significance of digits in the training objective, and makes it difficult to achieve both decoding efficiency and numerical precision. These limitations affect both the processing of sensor measurements and the generation of precise control commands, creating a fundamental barrier for deploying LLM-based autonomous driving systems. In this paper, we introduce DriveCode, a novel numerical encoding method that represents numbers as dedicated embeddings rather than discrete text tokens. DriveCode employs a number projector to map numbers into the language model’s hidden space, enabling seamless integration with visual and textual features in a unified multimodal sequence. Evaluated on OmniDrive, DriveGPT4, and DriveGPT4-V2 datasets, DriveCode demonstrates superior performance in trajectory prediction and control signal generation, confirming its effectiveness for LLM-based autonomous driving systems.

Method: Proposes Multimodal Weight Predictor (MWP) combining RGB images with physics metadata using Vision Transformer for visual features, dedicated metadata encoder, and Stacked Mutual Attention Fusion to integrate visual and physical cues. Trained with Mean Squared Logarithmic Error for wide weight range stability.

Result: Achieves 88.06 kg MAE, 6.39% MAPE, and R² of 0.9548 on Waste-Weight-10K dataset. Strong accuracy for light objects (0-100 kg: 2.38 kg MAE, 3.1% MAPE) and reliable for heavy waste (1000-2000 kg: 11.1% MAPE). Includes explainable AI module with SHAP and LLM.

Conclusion: The multimodal approach effectively addresses perspective and density challenges in waste weight estimation, providing accurate predictions across wide weight ranges with interpretable explanations.

Abstract: Accurate weight estimation of commercial and industrial waste is important for efficient operations, yet image-based estimation remains difficult because similar-looking objects may have different densities, and the visible size changes with camera distance. Addressing this problem, we propose Multimodal Weight Predictor (MWP) framework that estimates waste weight by combining RGB images with physics-informed metadata, including object dimensions, camera distance, and camera height. We also introduce Waste-Weight-10K, a real-world dataset containing 10,421 synchronized image-metadata collected from logistics and recycling sites. The dataset covers 11 waste categories and a wide weight range from 3.5 to 3,450 kg. Our model uses a Vision Transformer for visual features and a dedicated metadata encoder for geometric and category information, combining them with Stacked Mutual Attention Fusion that allows visual and physical cues guide each other. This helps the model manage perspective effects and link objects to material properties. To ensure stable performance across the wide weight range, we train the model using Mean Squared Logarithmic Error. On the test set, the proposed method achieves 88.06 kg Mean Absolute Error (MAE), 6.39% Mean Absolute Percentage Error (MAPE), and an R2 coefficient of 0.9548. The model shows strong accuracy for light objects in the 0-100 kg range with 2.38 kg MAE and 3.1% MAPE, maintaining reliable performance for heavy waste in the 1000-2000 kg range with 11.1% MAPE. Finally, we incorporate a physically grounded explanation module using Shapley Additive Explanations (SHAP) and a large language model to provide clear, human-readable explanations for each prediction.

Method: 1) Created Beyond8Bits dataset with 44K HDR videos and 1.5M ratings; 2) Developed HDR-Q MLLM with HDR-aware vision encoder for HDR-sensitive embeddings; 3) Introduced HDR-Aware Policy Optimization (HAPO) RL framework with HDR-SDR contrastive KL loss and Gaussian weighted regression reward.

Result: HDR-Q achieves state-of-the-art performance on Beyond8Bits dataset and public HDR-VQA benchmarks, demonstrating superior HDR video quality assessment capabilities.

Conclusion: The paper presents a comprehensive solution for HDR video quality assessment, addressing the gap in existing SDR-focused VQA systems through dataset creation, novel architecture, and specialized training techniques.

Abstract: High Dynamic Range (HDR) user-generated (UGC) videos are rapidly proliferating across social platforms, yet most perceptual video quality assessment (VQA) systems remain tailored to Standard Dynamic Range (SDR). HDR has a higher bit depth, wide color gamut, and elevated luminance range, exposing distortions such as near-black crushing, highlight clipping, banding, and exposure flicker that amplify UGC artifacts and challenge SDR models. To catalyze progress, we curate Beyond8Bits, a large-scale subjective dataset of 44K videos from 6.5K sources with over 1.5M crowd ratings, spanning diverse scenes, capture conditions, and compression settings. We further introduce HDR-Q, the first Multimodal Large Language Model (MLLM) for HDR-UGC VQA. We propose (i) a novel HDR-aware vision encoder to produce HDR-sensitive embeddings, and (ii) HDR-Aware Policy Optimization (HAPO), an RL finetuning framework that anchors reasoning to HDR cues. HAPO augments GRPO via an HDR-SDR contrastive KL that encourages token reliance on HDR inputs and a Gaussian weighted regression reward for fine-grained MOS calibration. Across Beyond8Bits and public HDR-VQA benchmarks, HDR-Q delivers state-of-the-art performance.

Method: Proposes a modular TeacherNet-GarmentNet-TryonNet (TGT) architecture with Feature-Guided Adversarial (FGA) Distillation, trajectory-consistency loss for garment preservation, and lightweight cross-modal conditioning for garment alignment.

Result: Achieves high-fidelity generation matching or outperforming server-based baselines on VITON-HD and DressCode datasets at 1024×768 resolution while running entirely offline on mobile devices.

Conclusion: High-quality virtual try-on is feasible and practical on-device, offering a secure, privacy-preserving solution for real-world applications.

Abstract: Virtual try-on (VTON) has recently achieved impressive visual fidelity, but most existing systems require uploading personal photos to cloud-based GPUs, raising privacy concerns and limiting on-device deployment. To address this, we present \textsc{Mobile-VTON}, a high-quality, privacy-preserving framework that enables fully offline virtual try-on on commodity mobile devices using only a single user image and a garment image. \textsc{Mobile-VTON} introduces a modular TeacherNet–GarmentNet–TryonNet (TGT) architecture that integrates knowledge distillation, garment-conditioned generation, and garment alignment into a unified pipeline optimized for on-device efficiency. Within this framework, we propose a Feature-Guided Adversarial (FGA) Distillation strategy that combines teacher supervision with adversarial learning to better match real-world image distributions. GarmentNet is trained with a trajectory-consistency loss to preserve garment semantics across diffusion steps, while TryonNet uses latent concatenation and lightweight cross-modal conditioning to enable robust garment-to-person alignment without large-scale pretraining. By combining these components, \textsc{Mobile-VTON} achieves high-fidelity generation with low computational overhead. Experiments on VITON-HD and DressCode at $1024{\times}768$ show that it matches or outperforms strong server-based baselines while running entirely offline. These results demonstrate that high-quality VTON is not only feasible but also practical on-device, offering a secure solution for real-world applications.

Method: Uses Instant-NGP’s hash encoding function as a key-controlled scene switcher. Associates a default key with a cover scene and a secret key with a hidden scene, training a single model to interweave both representations within the same network weights. Also introduces an enhanced Multi-Key scheme that assigns multiple independent keys across hash levels to expand key space and improve robustness against partial key disclosure attacks.

Result: StegoNGP can hide a complete high-quality 3D scene with strong imperceptibility and security. The resulting model is indistinguishable from a standard Instant-NGP in architecture and parameter count, providing a new paradigm for high-capacity, undetectable information hiding in neural fields.

Conclusion: StegoNGP offers a novel approach to 3D cryptographic steganography that overcomes limitations of existing methods by leveraging Instant-NGP’s hash encoding as a key-controlled scene switcher, enabling secure embedding of entire 3D scenes within neural field representations without architectural modifications.

Abstract: Recently, Instant Neural Graphics Primitives (Instant-NGP) has achieved significant success in rapid 3D scene reconstruction, but securely embedding high-capacity hidden data, such as an entire 3D scene, remains a challenge. Existing methods rely on external decoders, require architectural modifications, and suffer from limited capacity, which makes them easily detectable. We propose a novel parameter-free 3D Cryptographic Steganography using Instant-NGP (StegoNGP), which leverages the Instant-NGP hash encoding function as a key-controlled scene switcher. By associating a default key with a cover scene and a secret key with a hidden scene, our method trains a single model to interweave both representations within the same network weights. The resulting model is indistinguishable from a standard Instant-NGP in architecture and parameter count. We also introduce an enhanced Multi-Key scheme, which assigns multiple independent keys across hash levels, dramatically expanding the key space and providing high robustness against partial key disclosure attacks. Experimental results demonstrated that StegoNGP can hide a complete high-quality 3D scene with strong imperceptibility and security, providing a new paradigm for high-capacity, undetectable information hiding in neural fields. The code can be found at https://github.com/jiang-wenxiang/StegoNGP.

Method: Proposes VeGaS with two key innovations: 1) Galilean shearing matrix that explicitly incorporates time-varying velocity to model complex non-linear motions while isolating motion from geometry, and 2) Geometric Deformation Network that refines Gaussian shapes/orientations using spatio-temporal context and velocity cues.

Result: Extensive experiments on public datasets demonstrate state-of-the-art performance in dynamic scene reconstruction compared to existing methods.

Conclusion: VeGaS successfully addresses limitations of previous 4DGS approaches by decoupling motion and geometry, enabling more expressive modeling of complex dynamic scenes with reduced artifacts.

Abstract: High-fidelity reconstruction of dynamic scenes is an important yet challenging problem. While recent 4D Gaussian Splatting (4DGS) has demonstrated the ability to model temporal dynamics, it couples Gaussian motion and geometric attributes within a single covariance formulation, which limits its expressiveness for complex motions and often leads to visual artifacts. To address this, we propose VeGaS, a novel velocity-based 4D Gaussian Splatting framework that decouples Gaussian motion and geometry. Specifically, we introduce a Galilean shearing matrix that explicitly incorporates time-varying velocity to flexibly model complex non-linear motions, while strictly isolating the effects of Gaussian motion from the geometry-related conditional Gaussian covariance. Furthermore, a Geometric Deformation Network is introduced to refine Gaussian shapes and orientations using spatio-temporal context and velocity cues, enhancing temporal geometric modeling. Extensive experiments on public datasets demonstrate that VeGaS achieves state-of-the-art performance.

Method: Proposes PreciseCache with two components: 1) LFCache for step-wise caching using Low-Frequency Difference (LFD) to detect redundant steps, and 2) BlockCache for block-wise caching to skip redundant computations within network blocks. Both work together to precisely identify and skip only truly redundant computations.

Result: Extensive experiments on various backbones show PreciseCache achieves average 2.6x speedup without noticeable quality loss, outperforming prior feature caching methods that suffer from quality degradation.

Conclusion: PreciseCache effectively accelerates video generation inference by precisely detecting and skipping truly redundant computations, enabling practical applications without sacrificing quality.

Abstract: High computational costs and slow inference hinder the practical application of video generation models. While prior works accelerate the generation process through feature caching, they often suffer from notable quality degradation. In this work, we reveal that this issue arises from their inability to distinguish truly redundant features, which leads to the unintended skipping of computations on important features. To address this, we propose \textbf{PreciseCache}, a plug-and-play framework that precisely detects and skips truly redundant computations, thereby accelerating inference without sacrificing quality. Specifically, PreciseCache contains two components: LFCache for step-wise caching and BlockCache for block-wise caching. For LFCache, we compute the Low-Frequency Difference (LFD) between the prediction features of the current step and those from the previous cached step. Empirically, we observe that LFD serves as an effective measure of step-wise redundancy, accurately detecting highly redundant steps whose computation can be skipped through reusing cached features. To further accelerate generation within each non-skipped step, we propose BlockCache, which precisely detects and skips redundant computations at the block level within the network. Extensive experiments on various backbones demonstrate the effectiveness of our PreciseCache, which achieves an average of 2.6x speedup without noticeable quality loss. Source code will be released.

Method: Formulates concept erasure as constrained multi-objective optimization, uses implicit gradient surgery for utility-preserving unlearning, integrates LoRA-based parameter tuning with attention-level regularization, and employs anchor-and-propagate mechanism for temporal consistency in videos.

Result: Substantially outperforms prior methods in erasure effectiveness, generative fidelity, and temporal consistency across both image and video benchmarks, establishing new state-of-the-art for concept erasure in next-generation diffusion models.

Conclusion: EraseAnything++ provides an effective unified solution for concept erasure in modern flow-matching diffusion models, successfully addressing the limitations of previous methods while maintaining generative quality and temporal consistency.

Abstract: Removing undesired concepts from large-scale text-to-image (T2I) and text-to-video (T2V) diffusion models while preserving overall generative quality remains a major challenge, particularly as modern models such as Stable Diffusion v3, Flux, and OpenSora employ flow-matching and transformer-based architectures and extend to long-horizon video generation. Existing concept erasure methods, designed for earlier T2I/T2V models, often fail to generalize to these paradigms. To address this issue, we propose EraseAnything++, a unified framework for concept erasure in both image and video diffusion models with flow-matching objectives. Central to our approach is formulating concept erasure as a constrained multi-objective optimization problem that explicitly balances concept removal with preservation of generative utility. To solve the resulting conflicting objectives, we introduce an efficient utility-preserving unlearning strategy based on implicit gradient surgery. Furthermore, by integrating LoRA-based parameter tuning with attention-level regularization, our method anchors erasure on key visual representations and propagates it consistently across spatial and temporal dimensions. In the video setting, we further enhance consistency through an anchor-and-propagate mechanism that initializes erasure on reference frames and enforces it throughout subsequent transformer layers, thereby mitigating temporal drift. Extensive experiments on both image and video benchmarks demonstrate that EraseAnything++ substantially outperforms prior methods in erasure effectiveness, generative fidelity, and temporal consistency, establishing a new state of the art for concept erasure in next-generation diffusion models.

Method: Replace basic mathematical primitives with lightweight shape bank from de-identified segmentation masks of 5 subjects. Introduce structure-aware sequential placement strategy using spatial anchors for localization and topological graph to manage inter-organ interactions during patch synthesis.

Result: Outperforms state-of-the-art FDSL baselines by 1.74% and SSL methods by up to 1.66% on BTCV and MSD datasets. Shows robust scaling effect where performance improves with increased synthetic data volume.

Conclusion: Provides data-efficient, privacy-compliant solution for medical segmentation by unifying infinite scalability of synthetic data with anatomical realism through anatomy-informed synthetic supervision.

Abstract: Vision Transformers (ViTs) excel in 3D medical segmentation but require massive annotated datasets. While Self-Supervised Learning (SSL) mitigates this using unlabeled data, it still faces strict privacy and logistical barriers. Formula-Driven Supervised Learning (FDSL) offers a privacy-preserving alternative by pre-training on synthetic mathematical primitives. However, a critical semantic gap limits its efficacy: generic shapes lack the morphological fidelity, fixed spatial layouts, and inter-organ relationships of real anatomy, preventing models from learning essential global structural priors. To bridge this gap, we propose an Anatomy-Informed Synthetic Supervised Pre-training framework unifying FDSL’s infinite scalability with anatomical realism. We replace basic primitives with a lightweight shape bank with de-identified, label-only segmentation masks from 5 subjects. Furthermore, we introduce a structure-aware sequential placement strategy to govern the patch synthesis process. Instead of random placement, we enforce physiological plausibility using spatial anchors for correct localization and a topological graph to manage inter-organ interactions (e.g., preventing impossible overlaps). Extensive experiments on BTCV and MSD datasets demonstrate that our method significantly outperforms state-of-the-art FDSL baselines and SSL methods by 1.74% and up to 1.66%, while exhibiting a robust scaling effect where performance improves with increased synthetic data volume. This provides a data-efficient, privacy-compliant solution for medical segmentation. The code will be made publicly available upon acceptance.

Method: EFS uses self-supervised DINO embeddings to partition videos into visually homogeneous temporal segments (event proxies), selects query-relevant frames within each event as anchors, then applies adaptive Maximal Marginal Relevance for global refinement to optimize event coverage, query relevance, and visual diversity.

Result: When applied to LLaVA-Video-7B, EFS improves accuracy by 4.7% on VideoMME, 4.9% on LongVideoBench, and 8.8% on MLVU benchmarks.

Conclusion: EFS is an effective training-free, plug-and-play module for LVLMs that leverages event structure for better long-video understanding through hierarchical frame selection.

Abstract: Massive frame redundancy and limited context window make efficient frame selection crucial for long-video understanding with large vision-language models (LVLMs). Prevailing approaches, however, adopt a flat sampling paradigm which treats the video as an unstructured collection of frames. In this paper, we introduce Event-Anchored Frame Selection (EFS), a hierarchical, event-aware pipeline. Leveraging self-supervised DINO embeddings, EFS first partitions the video stream into visually homogeneous temporal segments, which serve as proxies for semantic events. Within each event, it then selects the most query-relevant frame as an anchor. These anchors act as structural priors that guide a global refinement stage using an adaptive Maximal Marginal Relevance (MMR) scheme. This pipeline ensures the final keyframe set jointly optimizes for event coverage, query relevance, and visual diversity. As a training-free, plug-and-play module, EFS can be seamlessly integrated into off-the-shelf LVLMs, yielding substantial gains on challenging video understanding benchmarks. Specifically, when applied to LLaVA-Video-7B, EFS improves accuracy by 4.7%, 4.9%, and 8.8% on VideoMME, LongVideoBench, and MLVU, respectively.

Method: Proposes a Physics-inspired Spatially-Decoupled Synthesis framework that orthogonalizes synthesis: 1) creates gradient-shielded buffer zones around boundaries to preserve shape learning, and 2) injects physics-driven spectral textures into object cores. This addresses boundary aliasing where high-frequency textures corrupt gradient signals needed for learning structural boundaries.

Result: Outperforms previous FDSL methods and SSL methods trained on real-world medical datasets by 1.43% on BTCV and up to 1.51% on MSD tasks, demonstrating superior performance in medical image segmentation.

Conclusion: The framework offers a scalable, annotation-free foundation for medical Vision Transformers by effectively reconciling robust shape representation learning with invariance to acquisition noise through spatially-decoupled synthesis.

Abstract: Vision Transformers (ViTs) have revolutionized medical image analysis, yet their data-hungry nature clashes with the scarcity and privacy constraints of clinical archives. Formula-Driven Supervised Learning (FDSL) has emerged as a promising solution to this bottleneck, synthesizing infinite annotated samples from mathematical formulas without utilizing real patient data. However, existing FDSL paradigms rely on simple geometric shapes with homogeneous intensities, creating a substantial gap by neglecting tissue textures and noise patterns inherent in modalities like CT and MRI. In this paper, we identify a critical optimization conflict termed boundary aliasing: when high-frequency synthetic textures are naively added, they corrupt the image gradient signals necessary for learning structural boundaries, causing the model to fail in delineating real anatomical margins. To bridge this gap, we propose a novel Physics-inspired Spatially-Decoupled Synthesis framework. Our approach orthogonalizes the synthesis process: it first constructs a gradient-shielded buffer zone based on boundary distance to ensure stable shape learning, and subsequently injects physics-driven spectral textures into the object core. This design effectively reconciles robust shape representation learning with invariance to acquisition noise. Extensive experiments on the BTCV and MSD datasets demonstrate that our method significantly outperforms previous FDSL, as well as SSL methods trained on real-world medical datasets, by 1.43% on BTCV and up to 1.51% on MSD task, offering a scalable, annotation-free foundation for medical ViTs. The code will be made publicly available upon acceptance.

Method: Technical survey methodology that categorizes existing foundation models into unimodal and multimodal approaches, provides background/motivation analysis, and includes tutorial-like guidance for training and applying these models to real-world remote sensing tasks.

Result: A comprehensive resource that addresses three key questions: defining foundation models in remote sensing, explaining their necessity, and providing practical guidance for junior researchers to understand and apply these models effectively.

Conclusion: Foundation models represent a transformative opportunity for remote sensing, and this survey serves as a valuable entry point for researchers to understand their evolution from unimodal to multimodal approaches and apply them across various applications.

Abstract: Remote sensing (RS) techniques are increasingly crucial for deepening our understanding of the planet. As the volume and diversity of RS data continue to grow exponentially, there is an urgent need for advanced data modeling and understanding capabilities to manage and interpret these vast datasets effectively. Foundation models present significant new growth opportunities and immense potential to revolutionize the RS field. In this paper, we conduct a comprehensive technical survey on foundation models in RS, offering a brand-new perspective by exploring their evolution from unimodality to multimodality. We hope this work serves as a valuable entry point for researchers interested in both foundation models and RS and helps them launch new projects or explore new research topics in this rapidly evolving area. This survey addresses the following three key questions: What are foundation models in RS? Why are foundation models needed in RS? How can we effectively guide junior researchers in gaining a comprehensive and practical understanding of foundation models in RS applications? More specifically, we begin by outlining the background and motivation, emphasizing the importance of foundation models in RS. We then review existing foundation models in RS, systematically categorizing them into unimodal and multimodal approaches. Additionally, we provide a tutorial-like section to guide researchers, especially beginners, on how to train foundation models in RS and apply them to real-world tasks. The survey aims to equip researchers in RS with a deeper and more efficient understanding of foundation models, enabling them to get started easily and effectively apply these models across various RS applications.

Method: Uses a single commodity RGB-D camera with vision foundation models for continuous markerless probe tracking. Includes a vision-guided divergence detector for tracking integrity monitoring and failure recovery. Proposes a dual-stage pose refinement network that disentangles high-frequency jitter from low-frequency bias to denoise trajectories while maintaining kinematic fidelity.

Result: Outperforms competing sensorless and sensor-aided methods, achieving average position errors as low as 0.88 mm on complex trajectories. Produces high-quality 3D reconstructions with sub-millimeter mean surface accuracy.

Conclusion: Establishes a new benchmark for low-cost, accessible volumetric ultrasound imaging in resource-limited clinical settings by overcoming the limitations of existing tracking paradigms.

Abstract: Freehand 3D ultrasound (US) reconstruction promises volumetric imaging with the flexibility of standard 2D probes, yet existing tracking paradigms face a restrictive trilemma: marker-based systems demand prohibitive costs, inside-out methods require intrusive sensor attachment, and sensorless approaches suffer from severe cumulative drift. To overcome these limitations, we present MLRecon, a robust markerless 3D US reconstruction framework delivering drift-resilient 6D probe pose tracking using a single commodity RGB-D camera. Leveraging the generalization power of vision foundation models, our pipeline enables continuous markerless tracking of the probe, augmented by a vision-guided divergence detector that autonomously monitors tracking integrity and triggers failure recovery to ensure uninterrupted scanning. Crucially, we further propose a dual-stage pose refinement network that explicitly disentangles high-frequency jitter from low-frequency bias, effectively denoising the trajectory while maintaining the kinematic fidelity of operator maneuvers. Experiments demonstrate that MLRecon significantly outperforms competing sensorless and sensor-aided methods, achieving average position errors as low as 0.88 mm on complex trajectories and yielding high-quality 3D reconstructions with sub-millimeter mean surface accuracy. This establishes a new benchmark for low-cost, accessible volumetric US imaging in resource-limited clinical settings.

Method: Introduces Motion Decoupled Mask Attention Mechanism using object-specific masks to constrain attention, and Differentiated Mask Propagation Mechanism that derives masks from diffusion attention and propagates them across frames efficiently.

Result: Extensive experiments demonstrate precise, compositional, and state-of-the-art performance in I2V-based multi-object multi-motion transfer.

Conclusion: FlexiMMT successfully addresses cross-object motion entanglement and enables flexible multi-object motion transfer with accurate motion-to-object mappings.

Abstract: Motion transfer has emerged as a promising direction for controllable video generation, yet existing methods largely focus on single-object scenarios and struggle when multiple objects require distinct motion patterns. In this work, we present FlexiMMT, the first implicit image-to-video (I2V) motion transfer framework that explicitly enables multi-object, multi-motion transfer. Given a static multi-object image and multiple reference videos, FlexiMMT independently extracts motion representations and accurately assigns them to different objects, supporting flexible recombination and arbitrary motion-to-object mappings. To address the core challenge of cross-object motion entanglement, we introduce a Motion Decoupled Mask Attention Mechanism that uses object-specific masks to constrain attention, ensuring that motion and text tokens only influence their designated regions. We further propose a Differentiated Mask Propagation Mechanism that derives object-specific masks directly from diffusion attention and progressively propagates them across frames efficiently. Extensive experiments demonstrate that FlexiMMT achieves precise, compositional, and state-of-the-art performance in I2V-based multi-object multi-motion transfer.

Method: Proposes Dr.Occ with two key components: 1) Depth-guided 2D-to-3D View Transformer (D²-VFormer) that leverages high-quality dense depth cues from MoGe-2 to construct reliable geometric priors for precise voxel feature alignment, and 2) Region-guided Expert Transformer (R/R²-EFormer) inspired by Mixture-of-Experts framework that adaptively allocates region-specific experts to handle spatial semantic variations.

Result: On the Occ3D-nuScenes benchmark, Dr.Occ improves the strong baseline BEVDet4D by 7.43% mIoU and 3.09% IoU under the full vision-only setting, demonstrating significant performance gains in 3D semantic occupancy prediction.

Conclusion: Dr.Occ effectively addresses geometric misalignment through depth guidance and spatial class imbalance through region-guided experts, providing complementary contributions that enhance 3D semantic occupancy prediction for autonomous driving perception.

Abstract: 3D semantic occupancy prediction is crucial for autonomous driving perception, offering comprehensive geometric scene understanding and semantic recognition. However, existing methods struggle with geometric misalignment in view transformation due to the lack of pixel-level accurate depth estimation, and severe spatial class imbalance where semantic categories exhibit strong spatial anisotropy. To address these challenges, we propose Dr.Occ, a depth- and region-guided occupancy prediction framework. Specifically, we introduce a depth-guided 2D-to-3D View Transformer (D$^2$-VFormer) that effectively leverages high-quality dense depth cues from MoGe-2 to construct reliable geometric priors, thereby enabling precise geometric alignment of voxel features. Moreover, inspired by the Mixture-of-Experts (MoE) framework, we propose a region-guided Expert Transformer (R/R$^2$-EFormer) that adaptively allocates region-specific experts to focus on different spatial regions, effectively addressing spatial semantic variations. Thus, the two components make complementary contributions: depth guidance ensures geometric alignment, while region experts enhance semantic learning. Experiments on the Occ3D-nuScenes benchmark demonstrate that \textbf{Dr.Occ} improves the strong baseline BEVDet4D by 7.43% mIoU and 3.09% IoU under the full vision-only setting.

Method: Proposes Data-to-Data Flow Matching framework that learns deterministic transformations directly between paired views. Introduces Probability Density Geodesic Flow Matching (PDG-FM) which constrains flow trajectories using geodesic interpolants derived from probability density metrics of pretrained diffusion models, aligning with high-density regions of the data manifold.

Result: Empirically surpasses diffusion-based NVS baselines, demonstrating improved structural coherence and smoother transitions across views.

Conclusion: Incorporating data-dependent geometric regularization into deterministic flow matching provides advantages for consistent novel view generation.

Abstract: Recent advances in generative modeling have substantially enhanced novel view synthesis, yet maintaining consistency across viewpoints remains challenging. Diffusion-based models rely on stochastic noise-to-data transitions, which obscure deterministic structures and yield inconsistent view predictions. We propose a Data-to-Data Flow Matching framework that learns deterministic transformations directly between paired views, enhancing view-consistent synthesis through explicit data coupling. To further enhance geometric coherence, we introduce Probability Density Geodesic Flow Matching (PDG-FM), which constrains flow trajectories using geodesic interpolants derived from probability density metrics of pretrained diffusion models. Such alignment with high-density regions of the data manifold promotes more realistic interpolants between samples. Empirically, our method surpasses diffusion-based NVS baselines, demonstrating improved structural coherence and smoother transitions across views. These results highlight the advantages of incorporating data-dependent geometric regularization into deterministic flow matching for consistent novel view generation.

Method: Proposes RaUF framework with: 1) anisotropic probabilistic model to learn fine-grained uncertainty and resolve conflicting feature-to-label mapping, 2) Bidirectional Domain Attention mechanism that exploits mutual complementarity between spatial structure and Doppler consistency to suppress spurious/multipath reflections.

Result: Extensive experiments on public benchmarks and real-world datasets show RaUF delivers highly reliable spatial detections with well-calibrated uncertainty. Downstream case studies validate enhanced reliability and scalability under challenging real-world driving scenarios.

Conclusion: RaUF provides a robust framework for radar-based perception by modeling spatial uncertainty fields, addressing fundamental challenges in radar sensing through physically grounded anisotropic properties and bidirectional attention mechanisms.

Abstract: Millimeter-wave radar offers unique advantages in adverse weather but suffers from low spatial fidelity, severe azimuth ambiguity, and clutter-induced spurious returns. Existing methods mainly focus on improving spatial perception effectiveness via coarse-to-fine cross-modal supervision, yet often overlook the ambiguous feature-to-label mapping, which may lead to ill-posed geometric inference and pose fundamental challenges to downstream perception tasks. In this work, we propose RaUF, a spatial uncertainty field learning framework that models radar measurements through their physically grounded anisotropic properties. To resolve conflicting feature-to-label mapping, we design an anisotropic probabilistic model that learns fine-grained uncertainty. To further enhance reliability, we propose a Bidirectional Domain Attention mechanism that exploits the mutual complementarity between spatial structure and Doppler consistency, effectively suppressing spurious or multipath-induced reflections. Extensive experiments on public benchmarks and real-world datasets demonstrate that RaUF delivers highly reliable spatial detections with well-calibrated uncertainty. Moreover, downstream case studies further validate the enhanced reliability and scalability of RaUF under challenging real-world driving scenarios.

Method: Proposes Content-Aware Frequency Encoding (CAFE) which builds on Fourier features using multiple parallel linear layers combined via Hadamard product to explicitly synthesize broader frequency bases with learned weights for task-relevant frequency selection. Extends to CAFE+ by incorporating Chebyshev features as complementary components to Fourier bases for stronger, more stable frequency representation.

Result: Extensive experiments across multiple benchmarks validate effectiveness and efficiency, consistently achieving superior performance over existing methods.

Conclusion: CAFE and CAFE+ address spectral bias limitations in INRs by enabling explicit and efficient synthesis of broader frequency bases with learned task-relevant frequency selection, outperforming existing approaches.

Abstract: Implicit Neural Representations (INRs) have emerged as a powerful paradigm for various signal processing tasks, but their inherent spectral bias limits the ability to capture high-frequency details. Existing methods partially mitigate this issue by using Fourier-based features, which usually rely on fixed frequency bases. This forces multi-layer perceptrons (MLPs) to inefficiently compose the required frequencies, thereby constraining their representational capacity. To address this limitation, we propose Content-Aware Frequency Encoding (CAFE), which builds upon Fourier features through multiple parallel linear layers combined via a Hadamard product. CAFE can explicitly and efficiently synthesize a broader range of frequency bases, while the learned weights enable the selection of task-relevant frequencies. Furthermore, we extend this framework to CAFE+, which incorporates Chebyshev features as a complementary component to Fourier bases. This combination provides a stronger and more stable frequency representation. Extensive experiments across multiple benchmarks validate the effectiveness and efficiency of our approach, consistently achieving superior performance over existing methods. Our code is available at https://github.com/JunboKe0619/CAFE.

Method: Proposes VL-Anomaly with: 1) prompt learning-driven alignment module that adapts Mask2Former visual features to CLIP text embeddings of known categories to suppress spurious anomaly responses; 2) multi-source inference strategy integrating text-guided similarity, CLIP-based image-text similarity, and detector confidence for more reliable anomaly prediction.

Result: Achieves state-of-the-art performance on benchmark datasets including RoadAnomaly, SMIYC and Fishyscapes, demonstrating improved anomaly segmentation with reduced false positives and better recall.

Conclusion: VL-Anomaly effectively incorporates semantic priors from pre-trained VLMs to address limitations of pixel-statistics-based approaches, providing more reliable road anomaly segmentation crucial for safe autonomous systems.

Abstract: Safe autonomous systems in complex environments require robust road anomaly segmentation to identify unknown obstacles. However, existing approaches often rely on pixel-level statistics to determine whether a region appears anomalous. This reliance leads to high false-positive rates on semantically normal background regions such as sky or vegetation, and poor recall of true Out-of-distribution (OOD) instances, thereby posing safety risks for robotic perception and decision-making. To address these challenges, we propose VL-Anomaly, a vision-language anomaly segmentation framework that incorporates semantic priors from pre-trained Vision-Language Models (VLMs). Specifically, we design a prompt learning-driven alignment module that adapts Mask2Forme’s visual features to CLIP text embeddings of known categories, effectively suppressing spurious anomaly responses in background regions. At inference time, we further introduce a multi-source inference strategy that integrates text-guided similarity, CLIP-based image-text similarity and detector confidence, enabling more reliable anomaly prediction by leveraging complementary information sources. Extensive experiments demonstrate that VL-Anomaly achieves state-of-the-art performance on benchmark datasets including RoadAnomaly, SMIYC and Fishyscapes.Code is released on https://github.com/NickHezhuolin/VL-aligner-Road-anomaly-segment.

Method: Proposes a TR functional decomposition with factors parameterized by Implicit Neural Representations (INRs). Through frequency-domain analysis, identifies spectral limitations and introduces a reparameterized version where each TR factor is a structured combination of learnable latent tensor and fixed basis to improve training dynamics and high-frequency modeling.

Result: Demonstrates superior performance on image inpainting, denoising, super-resolution, and point cloud recovery tasks compared to existing approaches. Theoretically proves improved training dynamics, Lipschitz continuity, and provides principled initialization scheme.

Conclusion: The reparameterized TR functional decomposition effectively addresses limitations of traditional TR methods for continuous data, enabling better high-frequency modeling and superior performance across various vision tasks.

Abstract: Tensor Ring (TR) decomposition is a powerful tool for high-order data modeling, but is inherently restricted to discrete forms defined on fixed meshgrids. In this work, we propose a TR functional decomposition for both meshgrid and non-meshgrid data, where factors are parameterized by Implicit Neural Representations (INRs). However, optimizing this continuous framework to capture fine-scale details is intrinsically difficult. Through a frequency-domain analysis, we demonstrate that the spectral structure of TR factors determines the frequency composition of the reconstructed tensor and limits the high-frequency modeling capacity. To mitigate this, we propose a reparameterized TR functional decomposition, in which each TR factor is a structured combination of a learnable latent tensor and a fixed basis. This reparameterization is theoretically shown to improve the training dynamics of TR factor learning. We further derive a principled initialization scheme for the fixed basis and prove the Lipschitz continuity of our proposed model. Extensive experiments on image inpainting, denoising, super-resolution, and point cloud recovery demonstrate that our method achieves consistently superior performance over existing approaches. Code is available at https://github.com/YangyangXu2002/RepTRFD.

Method: Proposes SMR-Net with attention-enhanced multi-scale feature fusion: attention-embedded feature extractor encodes raw sensor data, parallel multi-scale feature processing with standard/dilated convolutions, and adaptive reweighting network for dynamic feature weighting.

Result: On Type A and Type B snap datasets, SMR-Net outperforms Faster R-CNN with IoU improvements of 6.52% and 5.8%, and mAP increases of 2.8% and 1.5% respectively.

Conclusion: The method demonstrates superiority in complex snap detection and localization tasks for robot automated assembly, addressing limitations of traditional visual approaches.

Abstract: In robot automated assembly, snap assembly precision and efficiency directly determine overall production quality. As a core prerequisite, snap detection and localization critically affect subsequent assembly success. Traditional visual methods suffer from poor robustness and large localization errors when handling complex scenarios (e.g., transparent or low-contrast snaps), failing to meet high-precision assembly demands. To address this, this paper designs a dedicated sensor and proposes SMR-Net, an self-attention-based multi-scale object detection algorithm, to synergistically enhance detection and localization performance. SMR-Net adopts an attention-enhanced multi-scale feature fusion architecture: raw sensor data is encoded via an attention-embedded feature extractor to strengthen key snap features and suppress noise; three multi-scale feature maps are processed in parallel with standard and dilated convolution for dimension unification while preserving resolution; an adaptive reweighting network dynamically assigns weights to fused features, generating fine representations integrating details and global semantics. Experimental results on Type A and Type B snap datasets show SMR-Net outperforms traditional Faster R-CNN significantly: Intersection over Union (IoU) improves by 6.52% and 5.8%, and mean Average Precision (mAP) increases by 2.8% and 1.5% respectively. This fully demonstrates the method’s superiority in complex snap detection and localization tasks.

Method: Proposes TAR-FAS framework with CoT-VT paradigm, tool-augmented data annotation pipeline creating ToolFAS-16K dataset, and Diverse-Tool Group Relative Policy Optimization for autonomous tool learning.

Result: Achieves state-of-the-art performance under challenging one-to-eleven cross-domain protocol while providing fine-grained visual investigation for trustworthy spoof detection.

Conclusion: Incorporating external visual tools into MLLMs significantly improves FAS performance by enabling deeper investigation of subtle spoof clues beyond intuitive semantic cues.

Abstract: Face recognition remains vulnerable to presentation attacks, calling for robust Face Anti-Spoofing (FAS) solutions. Recent MLLM-based FAS methods reformulate the binary classification task as the generation of brief textual descriptions to improve cross-domain generalization. However, their generalizability is still limited, as such descriptions mainly capture intuitive semantic cues (e.g., mask contours) while struggling to perceive fine-grained visual patterns. To address this limitation, we incorporate external visual tools into MLLMs to encourage deeper investigation of subtle spoof clues. Specifically, we propose the Tool-Augmented Reasoning FAS (TAR-FAS) framework, which reformulates the FAS task as a Chain-of-Thought with Visual Tools (CoT-VT) paradigm, allowing MLLMs to begin with intuitive observations and adaptively invoke external visual tools for fine-grained investigation. To this end, we design a tool-augmented data annotation pipeline and construct the ToolFAS-16K dataset, which contains multi-turn tool-use reasoning trajectories. Furthermore, we introduce a tool-aware FAS training pipeline, where Diverse-Tool Group Relative Policy Optimization (DT-GRPO) enables the model to autonomously learn efficient tool use. Extensive experiments under a challenging one-to-eleven cross-domain protocol demonstrate that TAR-FAS achieves SOTA performance while providing fine-grained visual investigation for trustworthy spoof detection.

Method: Three key designs: 1) Hyper-Search (hypergraph-based QA generation connecting visual/textual nodes), 2) DR-TTS (decomposes tasks by tool type, optimizes specialized experts, recomposes via tree search), 3) Offline search engine supporting multiple tools for cost-effective training.

Result: Developed MM-DeepResearch, a powerful multimodal deep research agent showing superiority across benchmarks, with code publicly available.

Conclusion: The proposed three-pronged approach successfully addresses key challenges in developing multimodal research agents, enabling effective search-intensive multimodal QA through innovative data generation, tool optimization, and cost-effective training methods.

Abstract: We aim to develop a multimodal research agent capable of explicit reasoning and planning, multi-tool invocation, and cross-modal information synthesis, enabling it to conduct deep research tasks. However, we observe three main challenges in developing such agents: (1) scarcity of search-intensive multimodal QA data, (2) lack of effective search trajectories, and (3) prohibitive cost of training with online search APIs. To tackle them, we first propose Hyper-Search, a hypergraph-based QA generation method that models and connects visual and textual nodes within and across modalities, enabling to generate search-intensive multimodal QA pairs that require invoking various search tools to solve. Second, we introduce DR-TTS, which first decomposes search-involved tasks into several categories according to search tool types, and respectively optimize specialized search tool experts for each tool. It then recomposes tool experts to jointly explore search trajectories via tree search, producing trajectories that successfully solve complex tasks using various search tools. Third, we build an offline search engine supporting multiple search tools, enabling agentic reinforcement learning without using costly online search APIs. With the three designs, we develop MM-DeepResearch, a powerful multimodal deep research agent, and extensive results shows its superiority across benchmarks. Code is available at https://github.com/HJYao00/MM-DeepResearch

Method: ELF-VLA augments RL with structured diagnostic feedback that produces detailed, interpretable reports identifying specific failure modes. The VLA policy uses this explicit feedback to generate Feedback-Guided Refinement, injecting corrected high-reward samples back into RL training batches for targeted gradient updates.

Result: Extensive experiments show the method unlocks latent capabilities of VLA models, achieving state-of-the-art performance on the public NAVSIM benchmark for overall PDMS, EPDMS score and high-level planning accuracy.

Conclusion: Explicit learning from failures through structured diagnostic feedback enables VLA models to overcome RL optimization plateaus by providing targeted guidance for critical scenarios that unguided exploration cannot solve.

Abstract: Vision-Language-Action (VLA) models for autonomous driving often hit a performance plateau during Reinforcement Learning (RL) optimization. This stagnation arises from exploration capabilities constrained by previous Supervised Fine-Tuning (SFT), leading to persistent failures in long-tail scenarios. In these critical situations, all explored actions yield a zero-value driving score. This information-sparse reward signals a failure, yet fails to identify its root cause – whether it is due to incorrect planning, flawed reasoning, or poor trajectory execution. To address this limitation, we propose VLA with Explicit Learning from Failures (ELF-VLA), a framework that augments RL with structured diagnostic feedback. Instead of relying on a vague scalar reward, our method produces detailed, interpretable reports that identify the specific failure mode. The VLA policy then leverages this explicit feedback to generate a Feedback-Guided Refinement. By injecting these corrected, high-reward samples back into the RL training batch, our approach provides a targeted gradient, which enables the policy to solve critical scenarios that unguided exploration cannot. Extensive experiments demonstrate that our method unlocks the latent capabilities of VLA models, achieving state-of-the-art (SOTA) performance on the public NAVSIM benchmark for overall PDMS, EPDMS score and high-level planning accuracy.

Method: Uses Mixture of Diffusion (MoD) framework that decouples discrete masked diffusion for text understanding and continuous diffusion for visual generation, coupled through a shared attention backbone. Introduces data-centric length adaptation for flexible-length decoding without architectural changes

Result: Achieves state-of-the-art performance among omni-diffusion models on multimodal understanding and generation benchmarks, with 87.04 on DPG-Bench for text-to-image generation

Conclusion: LLaDA-o demonstrates the effectiveness of unified omni diffusion modeling for multimodal tasks, providing an efficient and flexible solution for both understanding and generation

Abstract: We present \textbf{LLaDA-o}, an effective and length-adaptive omni diffusion model for multimodal understanding and generation. LLaDA-o is built on a Mixture of Diffusion (MoD) framework that decouples discrete masked diffusion for text understanding and continuous diffusion for visual generation, while coupling them through a shared, simple, and efficient attention backbone that reduces redundant computation for fixed conditions. Building on MoD, we further introduce a data-centric length adaptation strategy that enables flexible-length decoding in multimodal settings without architectural changes. Extensive experiments show that LLaDA-o achieves state-of-the-art performance among omni-diffusion models on multimodal understanding and generation benchmarks, and reaches 87.04 on DPG-Bench for text-to-image generation, supporting the effectiveness of unified omni diffusion modeling. Code is available at https://github.com/ML-GSAI/LLaDA-o.

Method: Proposes FlowReg, a flow-matching framework in displacement field space with warmup-reflow training (single-step network as teacher, student learns to refine from arbitrary states) and Initial Guess strategy (feeding back predictions as next starting point).

Result: Outperforms state-of-the-art on 5 out of 6 tasks (+0.6% mean Dice score, largest gain in left ventricle +1.09%), reduces LVEF estimation error on all six tasks (-2.58 percentage points), using only 0.7% extra parameters and no segmentation labels.

Conclusion: FlowReg enables efficient and accurate cardiac image registration with minimal inference steps, addressing practical limitations of existing diffusion-based methods while maintaining strong performance.

Abstract: Diffusion-based unsupervised image registration has been explored for cardiac cine MR, but expensive multi-step inference limits practical use. We propose FlowReg, a flow-matching framework in displacement field space that achieves strong registration in as few as two steps and supports further refinement with more steps. FlowReg uses warmup-reflow training: a single-step network first acts as a teacher, then a student learns to refine from arbitrary intermediate states, removing the need for a pre-trained model as in existing methods. An Initial Guess strategy feeds back the model prediction as the next starting point, improving refinement from step two onward. On ACDC and MM2 across six tasks (including cross-dataset generalization), FlowReg outperforms the state of the art on five tasks (+0.6% mean Dice score on average), with the largest gain in the left ventricle (+1.09%), and reduces LVEF estimation error on all six tasks (-2.58 percentage points), using only 0.7% extra parameters and no segmentation labels. Anonymized code is available at https://github.com/mathpluscode/FlowReg.

Method: Proposes A3Point with two key components: 1) Semantic Confusion Prior (SCP) latent learning that captures model’s inherent semantic confusion information, and 2) Semantic Shift Region (SSR) localization that decouples semantic confusion from semantic shift, enabling adaptive optimization strategies for different disturbance levels.

Result: Extensive experiments on multiple standard generalized LiDAR segmentation benchmarks under adverse weather demonstrate effectiveness, achieving new state-of-the-art results.

Conclusion: A3Point effectively addresses the augmentation trade-off problem in adverse weather LiDAR segmentation by adaptively handling semantic shifts through latent learning and region localization techniques.

Abstract: Adverse weather conditions significantly degrade the performance of LiDAR point cloud semantic segmentation networks by introducing large distribution shifts. Existing augmentation-based methods attempt to enhance robustness by simulating weather interference during training. However, they struggle to fully exploit the potential of augmentations due to the trade-off between minor and aggressive augmentations. To address this, we propose A3Point, an adaptive augmentation-aware latent learning framework that effectively utilizes a diverse range of augmentations while mitigating the semantic shift, which refers to the change in the semantic meaning caused by augmentations. A3Point consists of two key components: semantic confusion prior (SCP) latent learning, which captures the model’s inherent semantic confusion information, and semantic shift region (SSR) localization, which decouples semantic confusion and semantic shift, enabling adaptive optimization strategies for different disturbance levels. Extensive experiments on multiple standard generalized LiDAR segmentation benchmarks under adverse weather demonstrate the effectiveness of our method, setting new state-of-the-art results.

Method: Created MCMR benchmark spanning five product domains (clothing, jewelry, shoes, furniture) with rich long-form metadata. Each query integrates complementary visual and textual attributes requiring joint satisfaction of all conditions. Benchmarked diverse MLLM-based multimodal retrievers and vision-language rerankers.

Result: Revealed: (1) distinct modality asymmetries across models; (2) visual cues dominate early-rank precision while textual metadata stabilizes long-tail ordering; (3) MLLM-based pointwise rerankers significantly improve fine-grained matching by explicitly verifying query-candidate consistency.

Conclusion: MCMR establishes a challenging diagnostic benchmark for advancing multimodal retrieval toward compositional, constraint-aware, and interpretable understanding, addressing limitations of existing benchmarks.

Abstract: Recent advances in multimodal large language models (MLLMs) have substantially expanded the capabilities of multimodal retrieval, enabling systems to align and retrieve information across visual and textual modalities. Yet, existing benchmarks largely focus on coarse-grained or single-condition alignment, overlooking real-world scenarios where user queries specify multiple interdependent constraints across modalities. To bridge this gap, we introduce MCMR (Multi-Conditional Multimodal Retrieval): a large-scale benchmark designed to evaluate fine-grained, multi-condition cross-modal retrieval under natural-language queries. MCMR spans five product domains: upper and bottom clothing, jewelry, shoes, and furniture. It also preserves rich long-form metadata essential for compositional matching. Each query integrates complementary visual and textual attributes, requiring models to jointly satisfy all specified conditions for relevance. We benchmark a diverse suite of MLLM-based multimodal retrievers and vision-language rerankers to assess their condition-aware reasoning abilities. Experimental results reveal: (i) distinct modality asymmetries across models; (ii) visual cues dominate early-rank precision, while textual metadata stabilizes long-tail ordering; and (iii) MLLM-based pointwise rerankers markedly improve fine-grained matching by explicitly verifying query-candidate consistency. Overall, MCMR establishes a challenging and diagnostic benchmark for advancing multimodal retrieval toward compositional, constraint-aware, and interpretable understanding. Our code and dataset is available at https://github.com/EIT-NLP/MCMR

Method: Introduces AesEval-Bench with 4 dimensions, 12 indicators, and 3 quantifiable tasks (aesthetic judgment, region selection, precise localization). Systematically evaluates proprietary, open-source, and reasoning-augmented VLMs. Creates training dataset using human-guided VLM labeling to produce task labels at scale and indicator-grounded reasoning to link abstract indicators to concrete design regions.

Result: Reveals clear performance gaps between VLMs and nuanced demands of aesthetic assessment. Fine-tuned VLMs show improved performance on aesthetic quality assessment tasks. Establishes first systematic framework for aesthetic quality assessment in graphic design.

Conclusion: The work provides comprehensive benchmark and training framework for improving VLMs’ ability to assess graphic design aesthetics, addressing key limitations in current research and enabling better visual communication evaluation.

Abstract: Assessing the aesthetic quality of graphic design is central to visual communication, yet remains underexplored in vision language models (VLMs). We investigate whether VLMs can evaluate design aesthetics in ways comparable to humans. Prior work faces three key limitations: benchmarks restricted to narrow principles and coarse evaluation protocols, a lack of systematic VLM comparisons, and limited training data for model improvement. In this work, we introduce AesEval-Bench, a comprehensive benchmark spanning four dimensions, twelve indicators, and three fully quantifiable tasks: aesthetic judgment, region selection, and precise localization. Then, we systematically evaluate proprietary, open-source, and reasoning-augmented VLMs, revealing clear performance gaps against the nuanced demands of aesthetic assessment. Moreover, we construct a training dataset to fine-tune VLMs for this domain, leveraging human-guided VLM labeling to produce task labels at scale and indicator-grounded reasoning to tie abstract indicators to concrete design regions.Together, our work establishes the first systematic framework for aesthetic quality assessment in graphic design. Our code and dataset will be released at: \href{https://github.com/arctanxarc/AesEval-Bench}{https://github.com/arctanxarc/AesEval-Bench}

Method: Introduces DP-RGMI framework that interprets DP as structured transformation of representation space. Decomposes performance degradation into encoder geometry (quantified by representation displacement from initialization and spectral effective dimension) and task-head utilization (measured as gap between linear-probe and end-to-end utility). Evaluated across 594,000+ images from four chest X-ray datasets with multiple pretrained initializations.

Result: DP consistently associated with utilization gap even when linear separability is preserved. Representation displacement and spectral dimension show non-monotonic, initialization- and dataset-dependent reshaping, indicating DP alters representation anisotropy rather than uniformly collapsing features. Correlation analysis shows end-to-end performance and utilization association is robust across datasets but varies by initialization, while geometric quantities capture additional prior- and dataset-conditioned variation.

Conclusion: DP-RGMI provides reproducible framework for diagnosing privacy-induced failure modes and informing privacy model selection in medical imaging. Reveals DP’s complex effects on representation spaces beyond simple performance metrics.

Abstract: Differential privacy (DP)’s effect in medical imaging is typically evaluated only through end-to-end performance, leaving the mechanism of privacy-induced utility loss unclear. We introduce Differential Privacy Representation Geometry for Medical Imaging (DP-RGMI), a framework that interprets DP as a structured transformation of representation space and decomposes performance degradation into encoder geometry and task-head utilization. Geometry is quantified by representation displacement from initialization and spectral effective dimension, while utilization is measured as the gap between linear-probe and end-to-end utility. Across over 594,000 images from four chest X-ray datasets and multiple pretrained initializations, we show that DP is consistently associated with a utilization gap even when linear separability is largely preserved. At the same time, displacement and spectral dimension exhibit non-monotonic, initialization- and dataset-dependent reshaping, indicating that DP alters representation anisotropy rather than uniformly collapsing features. Correlation analysis reveals that the association between end-to-end performance and utilization is robust across datasets but can vary by initialization, while geometric quantities capture additional prior- and dataset-conditioned variation. These findings position DP-RGMI as a reproducible framework for diagnosing privacy-induced failure modes and informing privacy model selection.

Method: HeroGS establishes hierarchical guidance across three levels: 1) Image level converts sparse supervision into pseudo-dense guidance for global regularization, 2) Feature level uses Feature-Adaptive Densification and Pruning (FADP) to refine high-frequency details and adaptively densify background Gaussians, 3) Parameter level employs Co-Pruned Geometry Consistency (CPG) to guide geometric consistency through parameter freezing and co-pruning.

Result: Extensive experiments demonstrate that HeroGS achieves high-fidelity reconstructions and consistently surpasses state-of-the-art baselines under sparse-view conditions, effectively addressing the limitations of irregular Gaussian distributions.

Conclusion: The hierarchical guidance strategy effectively constrains and optimizes overall Gaussian distributions, enhancing both structural fidelity and rendering quality for 3DGS in sparse-view scenarios.

Abstract: 3D Gaussian Splatting (3DGS) has recently emerged as a promising approach in novel view synthesis, combining photorealistic rendering with real-time efficiency. However, its success heavily relies on dense camera coverage; under sparse-view conditions, insufficient supervision leads to irregular Gaussian distributions, characterized by globally sparse coverage, blurred background, and distorted high-frequency areas. To address this, we propose HeroGS, Hierarchical Guidance for Robust 3D Gaussian Splatting, a unified framework that establishes hierarchical guidance across the image, feature, and parameter levels. At the image level, sparse supervision is converted into pseudo-dense guidance, globally regularizing the Gaussian distributions and forming a consistent foundation for subsequent optimization. Building upon this, Feature-Adaptive Densification and Pruning (FADP) at the feature level leverages low-level features to refine high-frequency details and adaptively densifies Gaussians in background regions. The optimized distributions then support Co-Pruned Geometry Consistency (CPG) at parameter level, which guides geometric consistency through parameter freezing and co-pruning, effectively removing inconsistent splats. The hierarchical guidance strategy effectively constrains and optimizes the overall Gaussian distributions, thereby enhancing both structural fidelity and rendering quality. Extensive experiments demonstrate that HeroGS achieves high-fidelity reconstructions and consistently surpasses state-of-the-art baselines under sparse-view conditions.

Method: Proposes StrokeDiff, a diffusion-based framework with Smooth Regularization (SmR) that injects stochastic visual priors during training to stabilize diffusion models under sparse supervision. Includes Bézier-based conditioning module for controllability and integrates into a complete stroke-based painting pipeline.

Result: Produces diverse and structurally coherent brushstrokes, enables paintings with richer texture and layering. Validated by both automatic metrics and human evaluation using only 470 hand-drawn samples.

Conclusion: Demonstrates how data-efficient primitive modeling can support expressive and structured multimedia content creation through stable diffusion training with limited data and controllable primitive generation.

Abstract: Many creative multimedia systems are built upon visual primitives such as strokes or textures, which are difficult to collect at scale and fundamentally different from natural image data. This data scarcity makes it challenging for modern generative models to learn expressive and controllable primitives, limiting their use in process-aware content creation. In this work, we study the problem of learning human-like brushstroke generation from a small set of hand-drawn samples (n=470) and propose StrokeDiff, a diffusion-based framework with Smooth Regularization (SmR). SmR injects stochastic visual priors during training, providing a simple mechanism to stabilize diffusion models under sparse supervision without altering the inference process. We further show how the learned primitives can be made controllable through a Bézier-based conditioning module and integrated into a complete stroke-based painting pipeline, including prediction, generation, ordering, and compositing. This demonstrates how data-efficient primitive modeling can support expressive and structured multimedia content creation. Experiments indicate that the proposed approach produces diverse and structurally coherent brushstrokes and enables paintings with richer texture and layering, validated by both automatic metrics and human evaluation.

Method: Introduces GroundedSurg benchmark with surgical images paired with natural-language descriptions targeting single instruments, accompanied by structured spatial grounding annotations (bounding boxes, point-level anchors). Dataset spans ophthalmic, laparoscopic, robotic, and open procedures with diverse instrument types and imaging conditions.

Result: Extensive experiments show substantial performance gaps across modern segmentation and vision-language models, highlighting the need for clinically grounded vision-language reasoning in surgical AI systems.

Conclusion: GroundedSurg enables systematic and realistic evaluation of vision-language models in clinically realistic multi-instrument surgical scenes, addressing limitations of current benchmarks and advancing intelligent, context-aware intraoperative assistance.

Abstract: Clinically reliable perception of surgical scenes is essential for advancing intelligent, context-aware intraoperative assistance such as instrument handoff guidance, collision avoidance, and workflow-aware robotic support. Existing surgical tool benchmarks primarily evaluate category-level segmentation, requiring models to detect all instances of predefined instrument classes. However, real-world clinical decisions often require resolving references to a specific instrument instance based on its functional role, spatial relation, or anatomical interaction capabilities not captured by current evaluation paradigms. We introduce GroundedSurg, the first language-conditioned, instance-level surgical grounding benchmark. Each instance pairs a surgical image with a natural-language description targeting a single instrument, accompanied by structured spatial grounding annotations including bounding boxes and point-level anchors. The dataset spans ophthalmic, laparoscopic, robotic, and open procedures, encompassing diverse instrument types, imaging conditions, and operative complexities. By jointly evaluating linguistic reference resolution and pixel-level localization, GroundedSurg enables a systematic and realistic evaluation of vision-language models in clinically realistic multi-instrument scenes. Extensive experiments demonstrate substantial performance gaps across modern segmentation and VLMs, highlighting the urgent need for clinically grounded vision-language reasoning in surgical AI systems. Code and data are publicly available at https://github.com/gaash-lab/GroundedSurg

Method: Proposes DeAR framework that decomposes attention head roles using Concept Entropy metric to classify heads into Attribute, Generalization, and Mixed roles. Uses specialized attribute tokens and Role-Based Attention Mask mechanism to control information flow, plus Task-Adaptive Fusion Strategy for inference.

Result: Extensive experiments on fifteen datasets show DeAR achieves strong balance between task adaptation and generalization, outperforming previous methods across various tasks.

Conclusion: Functional specialization in VLMs occurs at the finer-grained level of individual attention heads rather than between layers, and controlling information flow based on head roles enables better adaptation while preserving generalization capabilities.

Abstract: Prompt learning is a dominant paradigm for adapting pre-trained Vision-Language Models (VLMs) to downstream tasks. However, existing methods often rely on a simplistic, layer-centric view, assuming shallow layers capture general features while deep layers handle task-specific knowledge. This assumption results in uncontrolled interactions between learnable tokens and original tokens. Task-specific knowledge could degrades the model’s core generalization and creates a trade-off between task adaptation and the preservation of zero-shot generalization. To address this, we challenge the layer-centric view and propose \textbf{DeAR}, a framework that achieves fine-grained VLM adaptation by \textbf{De}composing \textbf{A}ttention head \textbf{R}oles. We posit that the functional specialization within VLMs occurs not between layers, but at the finer-grained level of individual attention heads in the deeper layers. Based on this insight, we introduce a novel metric, Concept Entropy, to systematically classify attention heads into distinct functional roles: \textit{Attribute}, \textit{Generalization}, and \textit{Mixed}. Guided by these roles, we introduce specialized attribute tokens and a Role-Based Attention Mask mechanism to precisely control information flow, ensuring generalization heads remain isolated from task-specific knowledge. We further incorporate a Task-Adaptive Fusion Strategy for inference. Extensive experiments on fifteen datasets show that DeAR achieves a strong balance between task adaptation and generalization, outperforming previous methods across various tasks.

Method: Extracts visual features from DINOv3 and converts them into spatial guide masks via a lightweight TokenBook mechanism that aggregates token-prototype similarities. These guide masks gate feature activations in multiple segmentation backbones. Uses guide supervision loss to align masks with ground truth, optionally with boundary-focused hinge loss. Supports parameter-efficient adaptation via LoRA on DINOv3.

Result: Consistently improves segmentation quality and boundary robustness across diverse medical datasets with nnUNet-style inference, offering better performance than fine-tuning alternatives.

Conclusion: GuiDINO provides a practical alternative to fine-tuning foundation models for medical vision, demonstrating how foundation models can serve as guidance generators rather than direct feature extractors, preserving domain-specific architectural benefits.

Abstract: Foundation vision models are increasingly adopted in medical image analysis. Due to domain shift, these pretrained models misalign with medical image segmentation needs without being fully fine-tuned or lightly adapted. We introduce GuiDINO, a framework that repositions native foundation model to acting as a visual guidance generator for downstream segmentation. GuiDINO extracts visual feature representation from DINOv3 and converts them into a spatial guide mask via a lightweight TokenBook mechanism, which aggregates token-prototype similarities. This guide mask gates feature activations in multiple segmentation backbones, thereby injecting foundation-model priors while preserving the inductive biases and efficiency of medical dedicated architectures. Training relies on a guide supervision objective loss that aligns the guide mask to ground-truth regions, optionally augmented by a boundary-focused hinge loss to sharpen fine structures. GuiDINO also supports parameter-efficient adaptation through LoRA on the DINOv3 guide backbone. Across diverse medical datasets and nnUNet-style inference, GuiDINO consistently improves segmentation quality and boundary robustness, suggesting a practical alternative to fine-tuning and offering a new perspective on how foundation models can best serve medical vision. Code is available at https://github.com/Hi-FishU/GuiDINO

Method: Enhances MambaBDA (ChangeMamba architecture) with three modular components: (1) Focal Loss to address class imbalance, (2) lightweight Attention Gates to suppress irrelevant context, and (3) compact Alignment Module to spatially warp pre-event features toward post-event content before decoding

Result: Consistent improvements over baseline with 0.8% to 5% performance gains in-domain, and up to 27% improvement on unseen disasters across multiple datasets (xBD, Pakistan Flooding, Turkey Earthquake, Ida Hurricane)

Conclusion: The proposed modular enhancements significantly improve building damage assessment performance, with particularly strong benefits for generalization capability to unseen disaster scenarios

Abstract: Reliable post-disaster building damage assessment (BDA) from satellite imagery is hindered by severe class imbalance, background clutter, and domain shift across disaster types and geographies. In this work, we address these problems and explore ways to improve the MambaBDA, the BDA network of ChangeMamba architecture, one of the most successful BDA models. The approach enhances the MambaBDA with three modular components: (i) Focal Loss to mitigate class imbalance damage classification, (ii) lightweight Attention Gates to suppress irrelevant context, and (iii) a compact Alignment Module to spatially warp pre-event features toward post-event content before decoding. We experiment on multiple satellite imagery datasets, including xBD, Pakistan Flooding, Turkey Earthquake, and Ida Hurricane, and conduct in-domain and crossdataset tests. The proposed modular enhancements yield consistent improvements over the baseline model, with 0.8% to 5% performance gains in-domain, and up to 27% on unseen disasters. This indicates that the proposed enhancements are especially beneficial for the generalization capability of the system.

Method: Proposes ClinCoT framework with automatic data generation pipeline constructing clinically grounded preference pairs through reasoning with hypotheses-driven region proposals. Uses multiple Med-LLMs evaluators to rank and score responses, with rankings as supervision. Introduces scoring-based margin-aware optimization incorporating preference ranking and score difference to refine region-level reasoning trajectories, with iterative learning scheme that dynamically regenerates preference data.

Result: Extensive experiments on three medical VQA and report generation benchmarks demonstrate that ClinCoT consistently improves factual grounding and achieves superior performance compared with existing preference-based alignment methods.

Conclusion: ClinCoT effectively addresses the gap in medical vision-language models by transforming preference optimization from response-level correction to visual-driven reasoning, enhancing factual grounding through region-level reasoning integration.

Abstract: Medical Vision-Language Models have shown promising potential in clinical decision support, yet they remain prone to factual hallucinations due to insufficient grounding in localized pathological evidence. Existing medical alignment methods primarily operate at the response level through preference optimization, improving output correctness but leaving intermediate reasoning weakly connected to visual regions. Although chain-of-thought (CoT) enhances multimodal reasoning, it remains largely text-centric, limiting effective integration of clinical visual cues. To address this gap, we propose ClinCoT, a clinical-aware visual chain-of-thought framework that transforms preference optimization from response-level correction to visual-driven reasoning. We introduce an automatic data generation pipeline that constructs clinically grounded preference pairs through reasoning with hypotheses-driven region proposals. Multiple Med-LLMs evaluators rank and assign scores to each response, and these rankings serve as supervision to train the target model. We further introduce a scoring-based margin-aware optimization strategy that incorporates both preference ranking and score difference to refine region-level reasoning trajectories. To maintain alignment as the model’s policy evolves during training, we adopt an iterative learning scheme that dynamically regenerates preference data. Extensive experiments on three medical VQA and report generation benchmarks demonstrate that ClinCoT consistently improves factual grounding and achieves superior performance compared with existing preference-based alignment methods.

Method: Proposes Predictive Reasoning with Augmented Anomaly Contrastive Learning (PR-A²CL) with two key components: 1) Augmented Anomaly Contrastive Learning to distill discriminative features by maximizing similarity among normal instances while minimizing similarity between normal and anomalous outliers, and 2) A predict-and-verify paradigm using Predictive Anomaly Reasoning Blocks (PARBs) that iteratively leverage features from three images to predict the fourth and progressively pinpoint discrepancies.

Result: Experimental results on SVRT, CVR, and MC²R datasets show that PR-A²CL significantly outperforms state-of-the-art reasoning models.

Conclusion: The proposed PR-A²CL framework effectively addresses compositional visual reasoning tasks through anomaly contrastive learning and predictive verification mechanisms, demonstrating superior performance over existing methods.

Abstract: While visual reasoning for simple analogies has received significant attention, compositional visual relations (CVR) remain relatively unexplored due to their greater complexity. To solve CVR tasks, we propose Predictive Reasoning with Augmented Anomaly Contrastive Learning (PR-A$^2$CL), \ie, to identify an outlier image given three other images that follow the same compositional rules. To address the challenge of modelling abundant compositional rules, an Augmented Anomaly Contrastive Learning is designed to distil discriminative and generalizable features by maximizing similarity among normal instances while minimizing similarity between normal and anomalous outliers. More importantly, a predict-and-verify paradigm is introduced for rule-based reasoning, in which a series of Predictive Anomaly Reasoning Blocks (PARBs) iteratively leverage features from three out of the four images to predict those of the remaining one. Throughout the subsequent verification stage, the PARBs progressively pinpoint the specific discrepancies attributable to the underlying rules. Experimental results on SVRT, CVR and MC$^2$R datasets show that PR-A$^2$CL significantly outperforms state-of-the-art reasoning models.

Method: Proposes TC-SSA (Token Compression via Semantic Slot Aggregation) with a gated routing module that assigns patches to semantic slots using sparse Top-2 routing, followed by weighted aggregation to achieve global slide coverage under strict token budgets.

Result: Achieves 78.34% overall accuracy on SlideBench(TCGA) and 77.14% on diagnosis subset, outperforming sampling-based baselines. Generalizes to MIL classification with AUCs of 95.83% on TCGA-BRCA, 98.27% on TCGA-NSCLC, and 79.80% on PANDA.

Conclusion: Learnable semantic aggregation provides an effective trade-off between efficiency and diagnostic performance for gigapixel pathology reasoning, retaining diagnostically relevant information while dramatically reducing computational requirements.

Abstract: The application of large vision-language models to computational pathology holds great promise for diagnostic assistants but faces a critical computational bottleneck: the gigapixel scale of Whole Slide Images (WSIs). A single WSI typically contains over 105 patches, creating sequence lengths that exceed the constraints of standard Transformer architectures. Existing solutions often resort to spatial sampling, which risks discarding diagnostically critical evidence. To address this, we propose TC-SSA (Token Compression via Semantic Slot Aggregation), a learnable token compression framework that aggregates patch features into a fixed number of semantic slots. A gated routing module assigns patches to slots using sparse Top-2 routing, followed by weighted aggregation, enabling global slide coverage under a strict token budget. The resulting representation retains diagnostically relevant information while reducing the number of visual tokens to 1.7% of the original sequence. On the SlideBench(TCGA), our model achieves 78.34% overall accuracy and 77.14% on the diagnosis subset, outperforming sampling-based baselines under comparable token budgets. The method also generalizes to MIL classification, reaching AUC of 95.83% on TCGA-BRCA, 98.27% on TCGA-NSCLC and 79.80% on PANDA. These results suggest that learnable semantic aggregation provides an effective trade-off between efficiency and diagnostic performance for gigapixel pathology reasoning.

Method: Teacher-Guided Causal Disentanglement Network (TCD-Net) with three components: 1) Environmental Bias Adjustment module for de-confounding, 2) dual-branch disentanglement head with orthogonality constraint for strict content-noise separation, 3) Nano Banana Pro (Google’s reasoning-guided AI image generation model) to provide causal priors and guide content representations back to natural-image manifold.

Result: Outperforms mainstream methods across multiple benchmarks in both fidelity and efficiency, achieving 104.2 FPS real-time speed on a single RTX 5090 GPU.

Conclusion: Causal intervention approach effectively disentangles content from noise, improving robustness under distribution shifts while preserving subtle textures and eliminating residual artifacts.

Abstract: Conventional image denoising models often inadvertently learn spurious correlations between environmental factors and noise patterns. Moreover, due to high-frequency ambiguity, they struggle to reliably distinguish subtle textures from stochastic noise, resulting in over-removed details or residual noise artifacts. We therefore revisit denoising via causal intervention, arguing that purely correlational fitting entangles intrinsic content with extrinsic noise, which directly degrades robustness under distribution shifts. Motivated by this, we propose the Teacher-Guided Causal Disentanglement Network (TCD-Net), which explicitly decomposes the generative mechanism via structured interventions on feature spaces within a Vision Transformer framework. Specifically, our method integrates three key components: (1) An Environmental Bias Adjustment (EBA) module projects features into a stable, de-centered subspace to suppress global environmental bias (de-confounding). (2) A dual-branch disentanglement head employs an orthogonality constraint to force a strict separation between content and noise representations, preventing information leakage. (3) To resolve structural ambiguity, we leverage Nano Banana Pro, Google’s reasoning-guided AI image generation model, to guide a causal prior, effectively pulling content representations back onto the natural-image manifold. Extensive experiments demonstrate that TCD-Net outperforms mainstream methods across multiple benchmarks in both fidelity and efficiency, achieving a real-time speed of 104.2 FPS on a single RTX 5090 GPU.

Method: Proposes GRAD-Former with Adaptive Feature Relevance and Refinement (AFRAR) module containing Selective Embedding Amplification (SEA) and Global-Local Feature Refinement (GLFR) components. SEA uses gating mechanisms and GLFR uses differential attention with multiple softmax heaps to capture important features while minimizing irrelevant ones.

Result: Outperforms current state-of-the-art models across all metrics on three challenging CD datasets (LEVIR-CD, CDD, DSIFN-CD) while using fewer parameters, establishing a new benchmark for remote sensing change detection performance.

Conclusion: GRAD-Former effectively addresses computational complexity and feature utilization challenges in remote sensing change detection, providing superior performance with reduced model parameters.

Abstract: Change detection (CD) in remote sensing aims to identify semantic differences between satellite images captured at different times. While deep learning has significantly advanced this field, existing approaches based on convolutional neural networks (CNNs), transformers and Selective State Space Models (SSMs) still struggle to precisely delineate change regions. In particular, traditional transformer-based methods suffer from quadratic computational complexity when applied to very high-resolution (VHR) satellite images and often perform poorly with limited training data, leading to under-utilization of the rich spatial information available in VHR imagery. We present GRAD-Former, a novel framework that enhances contextual understanding while maintaining efficiency through reduced model size. The proposed framework consists of a novel encoder with Adaptive Feature Relevance and Refinement (AFRAR) module, fusion and decoder blocks. AFRAR integrates global-local contextual awareness through two proposed components: the Selective Embedding Amplification (SEA) module and the Global-Local Feature Refinement (GLFR) module. SEA and GLFR leverage gating mechanisms and differential attention, respectively, which generates multiple softmax heaps to capture important features while minimizing the captured irreverent features. Multiple experiments across three challenging CD datasets (LEVIR-CD, CDD, DSIFN-CD) demonstrate GRAD-Former’s superior performance compared to existing approaches. Notably, GRAD-Former outperforms the current state-of-the-art models across all the metrics and all the datasets while using fewer parameters. Our framework establishes a new benchmark for remote sensing change detection performance. Our code will be released at: https://github.com/Ujjwal238/GRAD-Former

Method: Train 3D multimodal large language model on large-scale articulation dataset (existing datasets + procedurally generated objects). Model autoregressively predicts variable number of parts and joints from object point clouds, then conditions 3D generative model to synthesize part geometries.

Result: Significantly outperforms state-of-the-art methods on PartNet-Mobility dataset in both part layout accuracy and joint prediction. Generalizes robustly to real-world objects and demonstrates utility in constructing digital twins for robot learning.

Conclusion: ArtLLM provides scalable framework for generating high-quality articulated assets, addressing limitations of existing methods and showing potential for applications in robotics and interactive environments.

Abstract: Creating interactive digital environments for gaming, robotics, and simulation relies on articulated 3D objects whose functionality emerges from their part geometry and kinematic structure. However, existing approaches remain fundamentally limited: optimization-based reconstruction methods require slow, per-object joint fitting and typically handle only simple, single-joint objects, while retrieval-based methods assemble parts from a fixed library, leading to repetitive geometry and poor generalization. To address these challenges, we introduce ArtLLM, a novel framework for generating high-quality articulated assets directly from complete 3D meshes. At its core is a 3D multimodal large language model trained on a large-scale articulation dataset curated from both existing articulation datasets and procedurally generated objects. Unlike prior work, ArtLLM autoregressively predicts a variable number of parts and joints, inferring their kinematic structure in a unified manner from the object’s point cloud. This articulation-aware layout then conditions a 3D generative model to synthesize high-fidelity part geometries. Experiments on the PartNet-Mobility dataset show that ArtLLM significantly outperforms state-of-the-art methods in both part layout accuracy and joint prediction, while generalizing robustly to real-world objects. Finally, we demonstrate its utility in constructing digital twins, highlighting its potential for scalable robot learning.

Method: Proposes TripleSumm architecture with adaptive weighting and fusion of visual, text, and audio modalities at frame level. Also introduces MoSu, the first large-scale benchmark providing all three modalities for multimodal video summarization research.

Result: TripleSumm achieves state-of-the-art performance, outperforming existing methods by significant margin on four benchmarks including the new MoSu dataset.

Conclusion: The adaptive multimodal fusion approach effectively addresses limitations of static fusion strategies, and the MoSu benchmark addresses the critical bottleneck in multimodal video summarization research.

Abstract: The exponential growth of video content necessitates effective video summarization to efficiently extract key information from long videos. However, current approaches struggle to fully comprehend complex videos, primarily because they employ static or modality-agnostic fusion strategies. These methods fail to account for the dynamic, frame-dependent variations in modality saliency inherent in video data. To overcome these limitations, we propose TripleSumm, a novel architecture that adaptively weights and fuses the contributions of visual, text, and audio modalities at the frame level. Furthermore, a significant bottleneck for research into multimodal video summarization has been the lack of comprehensive benchmarks. Addressing this bottleneck, we introduce MoSu (Most Replayed Multimodal Video Summarization), the first large-scale benchmark that provides all three modalities. Extensive experiments demonstrate that TripleSumm achieves state-of-the-art performance, outperforming existing methods by a significant margin on four benchmarks, including MoSu. Our code and dataset are available at https://github.com/smkim37/TripleSumm.

Method: ConVibNet extends VibNet to detect low-visibility needles by leveraging temporal dependencies across ultrasound frames. Uses novel intersection-and-difference loss for motion correlation awareness and curated dataset for development.

Result: Achieved tip error of 2.80±2.42 mm and angle error of 1.69±2.00 deg, with 0.75 mm improvement over best baseline while maintaining real-time inference.

Conclusion: ConVibNet advances real-time needle detection through temporal correlation modeling and novel loss, improving accuracy and robustness for potential integration into autonomous insertion systems.

Abstract: Purpose: Ultrasound-guided needle interventions are widely used in clinical practice, but their success critically depends on accurate needle placement, which is frequently hindered by the poor and intermittent visibility of needles in ultrasound images. Existing approaches remain limited by artifacts, occlusions, and low contrast, and often fail to support real-time continuous insertion. To overcome these challenges, this study introduces a robust real-time framework for continuous needle detection. Methods: We present ConVibNet, an extension of VibNet for detecting needles with significantly reduced visibility, addressing real-time, continuous needle tracking during insertion. ConVibNet leverages temporal dependencies across successive ultrasound frames to enable continuous estimation of both needle tip position and shaft angle in dynamic scenarios. To strengthen temporal awareness of needle-tip motion, we introduce a novel intersection-and-difference loss that explicitly leverages motion correlations across consecutive frames. In addition, we curated a dedicated dataset for model development and evaluation. Results: The performance of the proposed ConVibNet model was evaluated on our dataset, demonstrating superior accuracy compared to the baseline VibNet and UNet-LSTM models. Specifically, ConVibNet achieved a tip error of 2.80+-2.42 mm and an angle error of 1.69+-2.00 deg. These results represent a 0.75 mm improvement in tip localization accuracy over the best-performing baseline, while preserving real-time inference capability. Conclusion: ConVibNet advances real-time needle detection in ultrasound-guided interventions by integrating temporal correlation modeling with a novel intersection-and-difference loss, thereby improving accuracy and robustness and demonstrating high potential for integration into autonomous insertion systems.

Method: Proposes BeautyGRPO with FRPref-10K dataset covering five retouching dimensions, trains a specialized reward model, and introduces Dynamic Path Guidance (DPG) to stabilize stochastic sampling by computing anchor-based ODE paths and replanning guided trajectories.

Result: Extensive experiments show BeautyGRPO outperforms specialized face retouching methods and general image editing models, achieving superior texture quality, more accurate blemish removal, and better alignment with human aesthetic preferences.

Conclusion: BeautyGRPO successfully reconciles the exploration-fidelity trade-off in RL for face retouching, demonstrating effective alignment with human aesthetic preferences through dynamic path guidance and fine-grained preference modeling.

Abstract: Face retouching requires removing subtle imperfections while preserving unique facial identity features, in order to enhance overall aesthetic appeal. However, existing methods suffer from a fundamental trade-off. Supervised learning on labeled data is constrained to pixel-level label mimicry, failing to capture complex subjective human aesthetic preferences. Conversely, while online reinforcement learning (RL) excels at preference alignment, its stochastic exploration paradigm conflicts with the high-fidelity demands of face retouching and often introduces noticeable noise artifacts due to accumulated stochastic drift. To address these limitations, we propose BeautyGRPO, a reinforcement learning framework that aligns face retouching with human aesthetic preferences. We construct FRPref-10K, a fine-grained preference dataset covering five key retouching dimensions, and train a specialized reward model capable of evaluating subtle perceptual differences. To reconcile exploration and fidelity, we introduce Dynamic Path Guidance (DPG). DPG stabilizes the stochastic sampling trajectory by dynamically computing an anchor-based ODE path and replanning a guided trajectory at each sampling timestep, effectively correcting stochastic drift while maintaining controlled exploration. Extensive experiments show that BeautyGRPO outperforms both specialized face retouching methods and general image editing models, achieving superior texture quality, more accurate blemish removal, and overall results that better align with human aesthetic preferences.

Method: Proposes VisNec (Visual Necessity Score) that measures the marginal contribution of visual input by comparing predictive loss with and without visual context. Combines with semantic clustering to preserve task diversity while selecting high-necessity samples.

Result: Training on only 15% of LLaVA-665K dataset selected by VisNec achieves 100.2% of full-data performance. On Vision-Flan-186K, selection reduces data size further and surpasses full-data training by 15.8%.

Conclusion: Measuring and leveraging visual necessity provides an effective solution for both efficient and robust multimodal instruction tuning, demonstrating that data quality (visual necessity) is more important than quantity.

Abstract: The effectiveness of multimodal instruction tuning depends not only on dataset scale, but critically on whether training samples genuinely require visual reasoning. However, existing instruction datasets often contain a substantial portion of visually redundant samples (solvable from text alone), as well as multimodally misaligned supervision that can degrade learning. To address this, we propose VisNec (Visual Necessity Score), a principled data selection framework that measures the marginal contribution of visual input during instruction tuning. By comparing predictive loss with and without visual context, VisNec identifies whether a training instance is vision-critical, redundant, or misaligned. To preserve task diversity, we combine VisNec with semantic clustering and select high-necessity samples within each cluster. Across 10 downstream benchmarks, training on only 15% of the LLaVA-665K dataset selected by VisNec achieves 100.2% of full-data performance. On the smaller Vision-Flan-186K dataset, our selection not only further reduces data size but also surpasses full-data training by 15.8%. These results demonstrate that measuring and leveraging visual necessity provides an effective solution for both efficient and robust multimodal instruction tuning. Codes and selected subsets will be released upon acceptance.

Method: Proposes Editing-awaRE (REE) injection method that modulates injection intensity per token based on editing masks. Uses optical flow to track editing areas throughout video, then generates editing-aware feature injection where injection is not conducted on editing areas. Built on rectified-Flow models for zero-shot editing.

Result: FREE-Edit demonstrates effectiveness in various image-driven video editing scenarios without fine-tuning or training, producing higher-quality outputs compared to existing techniques.

Conclusion: The proposed REE injection method and FREE-Edit framework provide an effective zero-shot solution for image-driven video editing that better balances preservation of source motion/layout with edit propagation.

Abstract: Image-driven video editing aims to propagate edit contents from the modified first frame to the rest frames. The existing methods usually invert the source video to noise using a pre-trained image-to-video (I2V) model and then guide the sampling process using the edited first frame. Generally, a popular choice for maintaining motion and layout from the source video is intervening in the denoising process by injecting attention during reconstruction. However, such injection often leads to unsatisfactory results, where excessive injection leads to conflicting semantics from the source video while insufficient injection brings limited source representation. Recognizing this, we propose an Editing-awaRE (REE) injection method to modulate injection intensity of each token. Specifically, we first compute the pixel difference between the source and edited first frame to form a corresponding editing mask. Next, we track the editing area throughout the entire video by using optical flow to warp the first-frame mask. Then, editing-aware feature injection intensity for each token is generated accordingly, where injection is not conducted on editing areas. Building upon REE injection, we further propose a zero-shot image-driven video editing framework with recent-emerging rectified-Flow models, dubbed FREE-Edit. Without fine-tuning or training, our FREE-Edit demonstrates effectiveness in various image-driven video editing scenarios, showing its capability to produce higher-quality outputs compared with existing techniques. Project page: https://free-edit.github.io/page/.

Method: Unifies linear-time State-Space Models (SSMs) with Transformers in a hybrid architecture, using 3D-CNN spectral front-end, Hybrid Mamba-Transformer backbone, and dual-modal Visual and Textual Prompts for context-aware guidance.

Result: Achieves state-of-the-art performance with only 2% training data: 99.69% OA on Salinas dataset and 99.45% on Longkou dataset, demonstrating superior sample efficiency.

Conclusion: Hybrid sequence modeling combined with multi-modal prompting provides a robust path for high-performance, sample-efficient remote sensing classification.

Abstract: Accurate classification of hyperspectral imagery (HSI) is often frustrated by the tension between high-dimensional spectral data and the extreme scarcity of labeled training samples. While hierarchical models like LoLA-SpecViT have demonstrated the power of local windowed attention and parameter-efficient fine-tuning, the quadratic complexity of standard Transformers remains a barrier to scaling. We introduce VP-Hype, a framework that rethinks HSI classification by unifying the linear-time efficiency of State-Space Models (SSMs) with the relational modeling of Transformers in a novel hybrid architecture. Building on a robust 3D-CNN spectral front-end, VP-Hype replaces conventional attention blocks with a Hybrid Mamba-Transformer backbone to capture long-range dependencies with significantly reduced computational overhead. Furthermore, we address the label-scarcity problem by integrating dual-modal Visual and Textual Prompts that provide context-aware guidance for the feature extraction process. Our experimental evaluation demonstrates that VP-Hype establishes a new state of the art in low-data regimes. Specifically, with a training sample distribution of only 2%, the model achieves Overall Accuracy (OA) of 99.69% on the Salinas dataset and 99.45% on the Longkou dataset. These results suggest that the convergence of hybrid sequence modeling and multi-modal prompting provides a robust path forward for high-performance, sample-efficient remote sensing.

Method: Collected and curated a heterogeneous dataset of 100,000+ images from wrist-mounted cameras. Fine-tuned a VLM using QLoRA with a custom regression head for 3D position prediction. Implemented conditional routing to maintain general visual query capabilities while adding specialized 3D estimation.

Result: Achieved robust predictive performance with median MAE of 13mm on test set, representing a five-fold improvement over baseline without fine-tuning. In about 25% of cases, predictions were within acceptable range for robot-object interaction.

Conclusion: The approach successfully extends VLMs to 3D position estimation tasks while preserving their general visual understanding capabilities, demonstrating potential for enhanced human-machine interactions in robotics applications.

Abstract: Pre-trained general-purpose Vision-Language Models (VLM) hold the potential to enhance intuitive human-machine interactions due to their rich world knowledge and 2D object detection capabilities. However, VLMs for 3D coordinates detection tasks are rare. In this work, we investigate interactive abilities of VLMs by returning 3D object positions given a monocular RGB image from a wrist-mounted camera, natural language input, and robot states. We collected and curated a heterogeneous dataset of more than 100,000 images and finetuned a VLM using QLoRA with a custom regression head. By implementing conditional routing, our model maintains its ability to process general visual queries while adding specialized 3D position estimation capabilities. Our results demonstrate robust predictive performance with a median MAE of 13 mm on the test set and a five-fold improvement over a simpler baseline without finetuning. In about 25% of the cases, predictions are within a range considered acceptable for the robot to interact with objects.

Method: Proposes RnG with reconstruction-guided causal attention that separates reconstruction and generation at the attention level. Treats KV-cache as implicit 3D representation, allowing arbitrary poses to query this cache for novel-view RGBD rendering.

Result: Achieves state-of-the-art performance in both generalizable 3D reconstruction and novel view generation, while operating efficiently enough for real-time interactive applications.

Conclusion: RnG successfully unifies reconstruction and generation, accurately reconstructing visible geometry while generating plausible, coherent unseen geometry and appearance from limited viewpoints.

Abstract: Human perceive the 3D world through 2D observations from limited viewpoints. While recent feed-forward generalizable 3D reconstruction models excel at recovering 3D structures from sparse images, their representations are often confined to observed regions, leaving unseen geometry un-modeled. This raises a key, fundamental challenge: Can we infer a complete 3D structure from partial 2D observations? We present RnG (Reconstruction and Generation), a novel feed-forward Transformer that unifies these two tasks by predicting an implicit, complete 3D representation. At the core of RnG, we propose a reconstruction-guided causal attention mechanism that separates reconstruction and generation at the attention level, and treats the KV-cache as an implicit 3D representation. Then, arbitrary poses can efficiently query this cache to render high-fidelity, novel-view RGBD outputs. As a result, RnG not only accurately reconstructs visible geometry but also generates plausible, coherent unseen geometry and appearance. Our method achieves state-of-the-art performance in both generalizable 3D reconstruction and novel view generation, while operating efficiently enough for real-time interactive applications. Project page: https://npucvr.github.io/RnG

Method: Proposes Canonical3D with two key components: 1) Construction: Creates a unified canonical dataset through LLM-guided intra- and cross-category alignment across 200 categories, exposing canonical spatial regularities. 2) Induction: Implements a dual-branch architecture with canonical map anchoring and canonical box calibration to collapse pose variation and symmetry into stable canonical embeddings, shifting from input pose space to canonical embedding space.

Result: Establishes new state-of-the-art in open-world promptable 3D segmentation, demonstrating that the shift from input pose space to canonical embedding yields far more stable and transferable part semantics.

Conclusion: Learning latent canonical reference frames directly from data enables canonical space perception for 3D segmentation, overcoming brittleness of sensor-coordinate semantics and achieving more human-like interpretation of parts via functional roles in canonical spaces.

Abstract: Open-world promptable 3D semantic segmentation remains brittle as semantics are inferred in the input sensor coordinates. Yet, humans, in contrast, interpret parts via functional roles in a canonical space – wings extend laterally, handles protrude to the side, and legs support from below. Psychophysical evidence shows that we mentally rotate objects into canonical frames to reveal these roles. To fill this gap, we propose \methodName{}, which attains canonical space perception by inducing a latent canonical reference frame learned directly from data. By construction, we create a unified canonical dataset through LLM-guided intra- and cross-category alignment, exposing canonical spatial regularities across 200 categories. By induction, we realize canonicality inside the model through a dual-branch architecture with canonical map anchoring and canonical box calibration, collapsing pose variation and symmetry into a stable canonical embedding. This shift from input pose space to canonical embedding yields far more stable and transferable part semantics. Experimental results show that \methodName{} establishes new state of the art in open-world promptable 3D segmentation.

Method: Two contributions: 1) SafeEditBench evaluation suite using image-editing models to create policy-aligned safe/unsafe image pairs with human labels under 5 distinct policies; 2) SafeGuard-VL method using reinforcement learning with verifiable rewards (RLVR) instead of supervised fine-tuning to optimize models across evolving policies.

Result: Extensive experiments verify SafeGuard-VL’s effectiveness for unsafe image guardrails across various policies, addressing the cross-policy generalization problem that existing VLM-based methods fail to handle.

Conclusion: The proposed SafeGuard-VL with RLVR provides a robust solution for adaptive harmful image detection that can evolve with changing safety policies, overcoming limitations of traditional classifiers and overfitted VLM-based methods.

Abstract: Accurate rejection of sensitive or harmful visual content, i.e., harmful image guardrail, is critical in many application scenarios. This task must continuously adapt to the evolving safety policies and content across various domains and over time. However, traditional classifiers, confined to fixed categories, require frequent retraining when new policies are introduced. Vision-language models (VLMs) offer a more adaptable and generalizable foundation for dynamic safety guardrails. Despite this potential, existing VLM-based safeguarding methods are typically trained and evaluated under only a fixed safety policy. We find that these models are heavily overfitted to the seen policy, fail to generalize to unseen policies, and even lose the basic instruction-following ability and general knowledge. To address this issue, in this paper we make two key contributions. First, we benchmark the cross-policy generalization performance of existing VLMs with SafeEditBench, a new evaluation suite. SafeEditBench leverages image-editing models to convert unsafe images into safe counterparts, producing policy-aligned datasets where each safe-unsafe image pair remains visually similar except for localized regions violating specific safety rules. Human annotators then provide accurate safe/unsafe labels under five distinct policies, enabling fine-grained assessment of policy-aware generalization. Second, we introduce SafeGuard-VL, a reinforcement learning-based method with verifiable rewards (RLVR) for robust unsafe-image guardrails. Instead of relying solely on supervised fine-tuning (SFT) under fixed policies, SafeGuard-VL explicitly optimizes the model with policy-grounded rewards, promoting verifiable adaptation across evolving policies. Extensive experiments verify the effectiveness of our method for unsafe image guardrails across various policies.

Method: Conduct empirical analysis using effective rank (erank) to measure feature diversity and attention score entropy to investigate visual token processing mechanisms; analyze strengths/weaknesses of each approach; propose image-aware adjustments to hybrid pruning strategies and a simple adaptive pruning mechanism.

Result: Diversity-based pruning preserves less feature diversity than intended and correlates with increased hallucination; attention-based methods work better on simple images while diversity-based methods handle complex images better; adaptive pruning achieves strong performance across benchmarks and hallucination evaluations.

Conclusion: Empirical insights reveal limitations of current pruning approaches; image-aware hybrid pruning and adaptive mechanisms can improve performance and reliability in LVLMs.

Abstract: Large Vision-Language Models (LVLMs) have adopted visual token pruning strategies to mitigate substantial computational overhead incurred by extensive visual token sequences. While prior works primarily focus on either attention-based or diversity-based pruning methods, in-depth analysis of these approaches’ characteristics and limitations remains largely unexplored. In this work, we conduct thorough empirical analysis using effective rank (erank) as a measure of feature diversity and attention score entropy to investigate visual token processing mechanisms and analyze the strengths and weaknesses of each approach. Our analysis reveals two insights: (1) Our erank-based quantitative analysis shows that many diversity-oriented pruning methods preserve substantially less feature diversity than intended; moreover, analysis using the CHAIR dataset reveals that the diversity they do retain is closely tied to increased hallucination frequency compared to attention-based pruning. (2) We further observe that attention-based approaches are more effective on simple images where visual evidence is concentrated, while diversity-based methods better handle complex images with distributed features. Building on these empirical insights, we show that incorporating image-aware adjustments into existing hybrid pruning strategies consistently improves their performance. We also provide a minimal instantiation of our empirical findings through a simple adaptive pruning mechanism, which achieves strong and reliable performance across standard benchmarks as well as hallucination-specific evaluations. Our project page available at https://cvsp-lab.github.io/AgilePruner.

Method: Created a large-scale benchmark challenge with training cohort of 1,506 patients from multiple US institutions and external test set of 574 patients from three independent European centers. Used unified scoring framework combining predictive performance with subgroup consistency across age, menopausal status, and breast density.

Result: Twenty-six international teams participated, revealing substantial performance variability under external testing and trade-offs between overall accuracy and subgroup fairness. Demonstrated challenges in cross-continental and cross-institutional generalization.

Conclusion: The challenge provides standardized datasets, evaluation protocols, and public resources to promote development of robust and equitable AI systems for breast cancer imaging, highlighting the importance of external validation and subgroup fairness assessment.

Abstract: Breast cancer is the most frequently diagnosed malignancy among women worldwide and a leading cause of cancer-related mortality. Dynamic contrast-enhanced magnetic resonance imaging plays a central role in tumor characterization and treatment monitoring, particularly in patients receiving neoadjuvant chemotherapy. However, existing artificial intelligence models for breast magnetic resonance imaging are often developed using single-center data and evaluated using aggregate performance metrics, limiting their generalizability and obscuring potential performance disparities across demographic subgroups. The MAMA-MIA Challenge was designed to address these limitations by introducing a large-scale benchmark that jointly evaluates primary tumor segmentation and prediction of pathologic complete response using pre-treatment magnetic resonance imaging only. The training cohort comprised 1,506 patients from multiple institutions in the United States, while evaluation was conducted on an external test set of 574 patients from three independent European centers to assess cross-continental and cross-institutional generalization. A unified scoring framework combined predictive performance with subgroup consistency across age, menopausal status, and breast density. Twenty-six international teams participated in the final evaluation phase. Results demonstrate substantial performance variability under external testing and reveal trade-offs between overall accuracy and subgroup fairness. The challenge provides standardized datasets, evaluation protocols, and public resources to promote the development of robust and equitable artificial intelligence systems for breast cancer imaging.

Method: Proposes incorporating additional imaging modalities (like X-ray CT) as guidance for diffusion-based reconstruction of sparse-view neutron CT, without retraining the diffusion prior

Result: Substantially improved reconstruction quality for sparse-view neutron CT by incorporating X-ray CT of the same samples, even with imperfect side modalities

Conclusion: Cross-modal guidance enables accelerated imaging of costly modalities by leveraging complementary imaging data without retraining diffusion priors

Abstract: Diffusion models have emerged as powerful priors for solving inverse problems in computed tomography (CT). In certain applications, such as neutron CT, it can be expensive to collect large amounts of measurements even for a single scan, leading to sparse data sets from which it is challenging to obtain high quality reconstructions even with diffusion models. One strategy to mitigate this challenge is to leverage a complementary, easily available imaging modality; however, such approaches typically require retraining the diffusion model with large datasets. In this work, we propose incorporating an additional modality without retraining the diffusion prior, enabling accelerated imaging of costly modalities. We further examine the impact of imperfect side modalities on cross-modal guidance. Our method is evaluated on sparse-view neutron computed tomography, where reconstruction quality is substantially improved by incorporating X-ray computed tomography of the same samples.

Method: Uses multi-level decoder interaction with Task Interaction Modules at all decoder levels for bidirectional segmentation-classification communication via attention weighted pooling and multiplicative modulation. Includes Uncertainty-Proxy Attention for adaptive feature weighting and multi-scale context fusion for varying lesion sizes.

Result: Achieves 74.5% lesion IoU and 90.6% classification accuracy on BUSI dataset, with competitive performance on multiple breast ultrasound datasets. Ablation studies show multi-level task interaction provides significant gains.

Conclusion: Decoder-level bidirectional communication is more effective than conventional encoder-only parameter sharing for multi-task medical image analysis, with multi-level interaction capturing scale-specific task synergies.

Abstract: Breast ultrasound interpretation requires simultaneous lesion segmentation and tissue classification. However, conventional multi-task learning approaches suffer from task interference and rigid coordination strategies that fail to adapt to instance-specific prediction difficulty. We propose a multi-task framework addressing these limitations through multi-level decoder interaction and uncertainty-aware adaptive coordination. Task Interaction Modules operate at all decoder levels, establishing bidirectional segmentation-classification communication during spatial reconstruction through attention weighted pooling and multiplicative modulation. Unlike prior single-level or encoder-only approaches, this multi-level design captures scale specific task synergies across semantic-to-spatial scales, producing complementary task interaction streams. Uncertainty-Proxy Attention adaptively weights base versus enhanced features at each level using feature activation variance, enabling per-level and per-sample task balancing without heuristic tuning. To support instance-adaptive prediction, multi-scale context fusion captures morphological cues across varying lesion sizes. Evaluation on multiple publicly available breast ultrasound datasets demonstrates competitive performance, including 74.5% lesion IoU and 90.6% classification accuracy on BUSI dataset. Ablation studies confirm that multi-level task interaction provides significant performance gains, validating that decoder-level bidirectional communication is more effective than conventional encoder-only parameter sharing. The code is available at: https://github.com/C-loud-Nine/Uncertainty-Aware-Multi-Level-Decoder-Interaction.

Method: Proposes FoSS, a dual-branch framework with: 1) Frequency-domain branch using discrete Fourier transform to decompose trajectories into amplitude (global intent) and phase (local variations) components, with progressive helix reordering and selective state-space modules (Coarse2Fine-SSM and SpecEvolve-SSM) for O(N) complexity refinement; 2) Time-domain dynamic selective SSM that reconstructs self-attention behavior in linear time; 3) Cross-attention layer to fuse temporal and spectral representations; 4) Learnable queries for multiple candidate trajectories; 5) Weighted fusion head to express motion uncertainty.

Result: Achieves state-of-the-art accuracy on Argoverse 1 and Argoverse 2 benchmarks while reducing computation by 22.5% and parameters by over 40%. Comprehensive ablations confirm the necessity of each component.

Conclusion: FoSS successfully balances modeling power and computational efficiency for trajectory prediction by unifying frequency-domain reasoning with linear-time sequence modeling, offering a promising direction for efficient multimodal understanding in autonomous driving applications.

Abstract: Accurate trajectory prediction is vital for safe autonomous driving, yet existing approaches struggle to balance modeling power and computational efficiency. Attention-based architectures incur quadratic complexity with increasing agents, while recurrent models struggle to capture long-range dependencies and fine-grained local dynamics. Building upon this, we present FoSS, a dual-branch framework that unifies frequency-domain reasoning with linear-time sequence modeling. The frequency-domain branch performs a discrete Fourier transform to decompose trajectories into amplitude components encoding global intent and phase components capturing local variations, followed by a progressive helix reordering module that preserves spectral order; two selective state-space submodules, Coarse2Fine-SSM and SpecEvolve-SSM, refine spectral features with O(N) complexity. In parallel, a time-domain dynamic selective SSM reconstructs self-attention behavior in linear time to retain long-range temporal context. A cross-attention layer fuses temporal and spectral representations, while learnable queries generate multiple candidate trajectories, and a weighted fusion head expresses motion uncertainty. Experiments on Argoverse 1 and Argoverse 2 benchmarks demonstrate that FoSS achieves state-of-the-art accuracy while reducing computation by 22.5% and parameters by over 40%. Comprehensive ablations confirm the necessity of each component.

Method: Introduces three learnable semantic anchor tokens ([SEG], [NOR], [ANO]), Semantic-Pixel Alignment Module (SPAM) for cross-modal alignment, Anchor-Guided Mask Decoder (AGMD), and Anomaly-Instruct20K dataset with structured anomaly descriptions.

Result: Achieves state-of-the-art performance on six industrial and medical benchmarks in zero-shot setting.

Conclusion: AG-VAS provides an effective framework for zero-shot visual anomaly segmentation by addressing fundamental limitations of LMM-based approaches through anchor-guided segmentation and improved cross-modal alignment.

Abstract: Large multimodal models (LMMs) exhibit strong task generalization capabilities, offering new opportunities for zero-shot visual anomaly segmentation (ZSAS). However, existing LMM-based segmentation approaches still face fundamental limitations: anomaly concepts are inherently abstract and context-dependent, lacking stable visual prototypes, and the weak alignment between high-level semantic embeddings and pixel-level spatial features hinders precise anomaly localization. To address these challenges, we present AG-VAS (Anchor-Guided Visual Anomaly Segmentation), a new framework that expands the LMM vocabulary with three learnable semantic anchor tokens-[SEG], [NOR], and [ANO], establishing a unified anchor-guided segmentation paradigm. Specifically, [SEG] serves as an absolute semantic anchor that translates abstract anomaly semantics into explicit, spatially grounded visual entities (e.g., holes or scratches), while [NOR] and [ANO] act as relative anchors that model the contextual contrast between normal and abnormal patterns across categories. To further enhance cross-modal alignment, we introduce a Semantic-Pixel Alignment Module (SPAM) that aligns language-level semantic embeddings with high-resolution visual features, along with an Anchor-Guided Mask Decoder (AGMD) that performs anchor-conditioned mask prediction for precise anomaly localization. In addition, we curate Anomaly-Instruct20K, a large-scale instruction dataset that organizes anomaly knowledge into structured descriptions of appearance, shape, and spatial attributes, facilitating effective learning and integration of the proposed semantic anchors. Extensive experiments on six industrial and medical benchmarks demonstrate that AG-VAS achieves consistent state-of-the-art performance in the zero-shot setting.

Method: Controlled study using MedMNIST as multi-modality testbed, benchmarking VLM vision towers against vision-only baselines, quantifying reasoning support via Accuracy@1 vs Pass@K metrics, and evaluating when RL closes support gaps across modalities.

Result: RL is most effective when models already have non-trivial support (high Pass@K) - it primarily sharpens output distributions, improving accuracy and sampling efficiency, while SFT expands support and enables RL effectiveness.

Conclusion: Proposed boundary-aware recipe for RL post-training, demonstrated by achieving strong performance across six medical VQA benchmarks using OctoMed-initialized model trained on small balanced PMC VQA subset.

Abstract: Reinforcement learning (RL) is increasingly used to post-train medical Vision-Language Models (VLMs), yet it remains unclear whether RL improves medical visual reasoning or mainly sharpens behaviors already induced by supervised fine-tuning (SFT). We present a controlled study that disentangles these effects along three axes: vision, SFT, and RL. Using MedMNIST as a multi-modality testbed, we probe visual perception by benchmarking VLM vision towers against vision-only baselines, quantify reasoning support and sampling efficiency via Accuracy@1 versus Pass@K, and evaluate when RL closes the support gap and how gains transfer across modalities. We find that RL is most effective when the model already has non-trivial support (high Pass@K): it primarily sharpens the output distribution, improving Acc@1 and sampling efficiency, while SFT expands support and makes RL effective. Based on these findings, we propose a boundary-aware recipe and instantiate it by RL post-training an OctoMed-initialized model on a small, balanced subset of PMC multiple-choice VQA, achieving strong average performance across six medical VQA benchmarks.

Method: Extract single-view features directly from blind face images, use a feature manipulator to transform these into 3D-aware multi-view latent representations, and leverage diffusion models to synthesize high-quality consistent novel-view face images.

Result: Experimental results show the method significantly outperforms traditional two-stage approaches in both consistency and fidelity of novel-view face image generation.

Conclusion: NVB-Face provides an effective one-stage solution for novel-view synthesis from blind face images, overcoming limitations of conventional two-stage approaches by directly processing degraded inputs.

Abstract: We propose a novel one-stage method, NVB-Face, for generating consistent Novel-View images directly from a single Blind Face image. Existing approaches to novel-view synthesis for objects or faces typically require a high-resolution RGB image as input. When dealing with degraded images, the conventional pipeline follows a two-stage process: first restoring the image to high resolution, then synthesizing novel views from the restored result. However, this approach is highly dependent on the quality of the restored image, often leading to inaccuracies and inconsistencies in the final output. To address this limitation, we extract single-view features directly from the blind face image and introduce a feature manipulator that transforms these features into 3D-aware, multi-view latent representations. Leveraging the powerful generative capacity of a diffusion model, our framework synthesizes high-quality, consistent novel-view face images. Experimental results show that our method significantly outperforms traditional two-stage approaches in both consistency and fidelity.

Method: Comparative evaluation of supervised learning and open-vocabulary vision models for semantic segmentation and object detection across multiple disaster datasets (FloodNet+, RescueNet, DFire, LADD). Analysis of performance trends, failure modes, and practical trade-offs.

Result: Supervised training remains the most reliable approach when label space is fixed and annotations are available, especially for small objects and fine boundary delineation in cluttered scenes.

Conclusion: While open-vocabulary models offer advantages in annotation efficiency and flexibility, supervised approaches still outperform for precise disaster scene understanding tasks where fixed labels and annotations exist.

Abstract: Aerial imagery is critical for large-scale post-disaster damage assessment. Automated interpretation remains challenging due to clutter, visual variability, and strong cross-event domain shift, while supervised approaches still rely on costly, task-specific annotations with limited coverage across disaster types and regions. Recent open-vocabulary and foundation vision models offer an appealing alternative, by reducing dependence on fixed label sets and extensive task-specific annotations. Instead, they leverage large-scale pretraining and vision-language representations. These properties are particularly relevant for post-disaster domains, where visual concepts are ambiguous and data availability is constrained. In this work, we present a comparative evaluation of supervised learning and open-vocabulary vision models for post-disaster scene understanding, focusing on semantic segmentation and object detection across multiple datasets, including FloodNet+, RescueNet, DFire, and LADD. We examine performance trends, failure modes, and practical trade-offs between different learning paradigms, providing insight into their applicability for real-world disaster response. The most notable remark across all evaluated benchmarks is that supervised training remains the most reliable approach (i.e., when the label space is fixed and annotations are available), especially for small objects and fine boundary delineation in cluttered scenes.

Method: PEFD exploits projective geometry of camera systems to leverage richer group structure for null-space recovery, and efficiently learns without ground truth by fine-tuning pretrained foundation models designed for 1-3 channel imaging.

Result: On intraoperative and automotive datasets, PEFD recovers fine details like blood vessels and preserves spectral fidelity, substantially outperforming recent approaches and nearing supervised performance.

Conclusion: PEFD provides an effective framework for multispectral demosaicing that eliminates the need for costly ground truth data while achieving performance close to supervised methods.

Abstract: Multispectral demosaicing is crucial to reconstruct full-resolution spectral images from snapshot mosaiced measurements, enabling real-time imaging from neurosurgery to autonomous driving. Classical methods are blurry, while supervised learning requires costly ground truth (GT) obtained from slow line-scanning systems. We propose Perspective-Equivariant Fine-tuning for Demosaicing (PEFD), a framework that learns multispectral demosaicing from mosaiced measurements alone. PEFD a) exploits the projective geometry of camera-based imaging systems to leverage a richer group structure than previous demosaicing methods to recover more null-space information, and b) learns efficiently without GT by adapting pretrained foundation models designed for 1-3 channel imaging. On intraoperative and automotive datasets, PEFD recovers fine details such as blood vessels and preserves spectral fidelity, substantially outperforming recent approaches, nearing supervised performance.

Method: MixerCSeg architecture with coordinated specialist pathways: CNN-like for local textures, Transformer-style for global dependencies, Mamba-inspired for sequential context. Core TransMixer explores Mamba’s latent attention with dedicated pathways. Spatial block processing strategy and Direction-guided Edge Gated Convolution (DEGConv) enhance edge sensitivity. Spatial Refinement Multi-Level Fusion (SRF) refines multi-scale details.

Result: State-of-the-art performance on multiple crack segmentation benchmarks with only 2.05 GFLOPs and 2.54M parameters, demonstrating both efficiency and strong representational capability.

Conclusion: MixerCSeg successfully integrates complementary strengths of CNN, Transformer, and Mamba architectures for crack segmentation, achieving superior performance with minimal computational overhead through coordinated specialist pathways and edge-aware enhancements.

Abstract: Feature encoders play a key role in pixel-level crack segmentation by shaping the representation of fine textures and thin structures. Existing CNN-, Transformer-, and Mamba-based models each capture only part of the required spatial or structural information, leaving clear gaps in modeling complex crack patterns. To address this, we present MixerCSeg, a mixer architecture designed like a coordinated team of specialists, where CNN-like pathways focus on local textures, Transformer-style paths capture global dependencies, and Mamba-inspired flows model sequential context within a single encoder. At the core of MixerCSeg is the TransMixer, which explores Mamba’s latent attention behavior while establishing dedicated pathways that naturally express both locality and global awareness. To further enhance structural fidelity, we introduce a spatial block processing strategy and a Direction-guided Edge Gated Convolution (DEGConv) that strengthens edge sensitivity under irregular crack geometries with minimal computational overhead. A Spatial Refinement Multi-Level Fusion (SRF) module is then employed to refine multi-scale details without increasing complexity. Extensive experiments on multiple crack segmentation benchmarks show that MixerCSeg achieves state-of-the-art performance with only 2.05 GFLOPs and 2.54 M parameters, demonstrating both efficiency and strong representational capability. The code is available at https://github.com/spiderforest/MixerCSeg.

Method: TIMI introduces two key modules: 1) Instance-aware Separation Guidance (ISG) for instance disentanglement during early denoising, and 2) Spatial-stabilized Geometry-adaptive Update (SGU) to preserve geometric characteristics and relative relationships while stabilizing the ISG guidance.

Result: Extensive experiments show TIMI outperforms existing multi-instance methods in both global layout and distinct local instances, without requiring additional training and with faster inference speed.

Conclusion: TIMI demonstrates that training-free approaches can effectively leverage existing spatial priors in pre-trained models for high-fidelity multi-instance 3D generation, offering computational efficiency and improved spatial fidelity.

Abstract: Precise spatial fidelity in Image-to-3D multi-instance generation is critical for downstream real-world applications. Recent work attempts to address this by fine-tuning pre-trained Image-to-3D (I23D) models on multi-instance datasets, which incurs substantial training overhead and struggles to guarantee spatial fidelity. In fact, we observe that pre-trained I23D models already possess meaningful spatial priors, which remain underutilized as evidenced by instance entanglement issues. Motivated by this, we propose TIMI, a novel Training-free framework for Image-to-3D Multi-Instance generation that achieves high spatial fidelity. Specifically, we first introduce an Instance-aware Separation Guidance (ISG) module, which facilitates instance disentanglement during the early denoising stage. Next, to stabilize the guidance introduced by ISG, we devise a Spatial-stabilized Geometry-adaptive Update (SGU) module that promotes the preservation of the geometric characteristics of instances while maintaining their relative relationships. Extensive experiments demonstrate that our method yields better performance in terms of both global layout and distinct local instances compared to existing multi-instance methods, without requiring additional training and with faster inference speed.

Method: VidDoS uses universal optimization to create instance-agnostic triggers without inference-time gradient calculations. It employs masked teacher forcing to steer models toward expensive target sequences, combined with refusal penalty and early-termination suppression to override conciseness priors.

Result: Testing across three mainstream Video-LLMs and three video datasets (including video QA and autonomous driving) shows extreme degradation: token expansion >205× and inference latency inflation >15× relative to clean baselines. Simulations reveal induced latency leads to critical safety violations in autonomous driving.

Conclusion: The community needs to recognize and mitigate these high-hazard Energy-Latency Attacks in Video-LLMs, especially for safety-critical applications like autonomous driving.

Abstract: Video-LLMs are increasingly deployed in safety-critical applications but are vulnerable to Energy-Latency Attacks (ELAs) that exhaust computational resources. Current image-centric methods fail because temporal aggregation mechanisms dilute individual frame perturbations. Additionally, real-time demands make instance-wise optimization impractical for continuous video streams. We introduce VidDoS, which is the first universal ELA framework tailored for Video-LLMs. Our method leverages universal optimization to create instance-agnostic triggers that require no inference-time gradient calculation. We achieve this through $\textit{masked teacher forcing}$ to steer models toward expensive target sequences, combined with a $\textit{refusal penalty}$ and $\textit{early-termination suppression}$ to override conciseness priors. Testing across three mainstream Video-LLMs and three video datasets, which include video question answering and autonomous driving scenarios, shows extreme degradation. VidDoS induces a token expansion of more than 205$\times$ and inflates the inference latency by more than 15$\times$ relative to clean baselines. Simulations of real-time autonomous driving streams further reveal that this induced latency leads to critical safety violations. We urge the community to recognize and mitigate these high-hazard ELA in Video-LLMs.

Method: Revisits modulation transfer function (MTF) to propose Exposure-Time-dependent MTF (ET-MTF) modeling blur as continuous function of exposure-time. Derives tilt-invariant PSF from ET-MTF, integrates with spatially varying blur-width field for comprehensive turbulence blur synthesis.

Result: Constructs ET-Turb dataset with 5,083 videos (2,005,835 frames) across diverse optical/atmospheric conditions with continuous exposure-time modeling. Models trained on ET-Turb produce more realistic restorations and achieve superior generalization on real-world turbulence data.

Conclusion: ET-MTF provides physically accurate turbulence blur modeling, and ET-Turb dataset enables training models that generalize better to real turbulence restoration tasks.

Abstract: Atmospheric turbulence significantly degrades long-range imaging by introducing geometric warping and exposure-time-dependent blur, which adversely affects both visual quality and the performance of high-level vision tasks. Existing methods for synthesizing turbulence effects often oversimplify the relationship between blur and exposure-time, typically assuming fixed or binary exposure settings. This leads to unrealistic synthetic data and limited generalization capability of trained models. To address this gap, we revisit the modulation transfer function (MTF) formulation and propose a novel Exposure-Time-dependent MTF (ET-MTF) that models blur as a continuous function of exposure-time. For blur synthesis, we derive a tilt-invariant point spread function (PSF) from the ET-MTF, which, when integrated with a spatially varying blur-width field, provides a comprehensive and physically accurate characterization of turbulence-induced blur. Building on this synthesis pipeline, we construct ET-Turb, a large-scale synthetic turbulence dataset that explicitly incorporates continuous exposure-time modeling across diverse optical and atmospheric conditions. The dataset comprises 5,083 videos (2,005,835 frames), partitioned into 3,988 training and 1,095 test videos. Extensive experiments demonstrate that models trained on ET-Turb produce more realistic restorations and achieve superior generalization on real-world turbulence data compared to those trained on other datasets. The dataset is publicly available at: github.com/Jun-Wei-Zeng/ET-Turb.

Method: Proposes AOT (Anchors via Optimal Transport) that establishes local- and global-aware token anchors within frames using attention guidance, then uses optimal transport to aggregate informative contexts from pruned tokens. For temporal reduction, uses keyframe anchors to ensemble similar information from consecutive frames via optimal transport while keeping distinct tokens for temporal dynamics.

Result: Extensive evaluations show competitive performance across various short- and long-video benchmarks on leading video LLMs, obtaining substantial computational efficiency while preserving temporal and visual fidelity.

Conclusion: AOT provides an effective training-free token reduction method for Video LLMs that comprehensively addresses spatiotemporal redundancy while preserving informative contexts through optimal transport-based aggregation.

Abstract: Video Large Language Models (VLLMs) demonstrate strong video understanding but suffer from inefficiency due to redundant visual tokens. Existing pruning primary targets intra-frame spatial redundancy or prunes inside the LLM with shallow-layer overhead, yielding suboptimal spatiotemporal reduction and underutilizing long-context compressibility. All of them often discard subtle yet informative context from merged or pruned tokens. In this paper, we propose a new perspective that elaborates token \textbf{A}nchors within intra-frame and inter-frame to comprehensively aggregate the informative contexts via local-global \textbf{O}ptimal \textbf{T}ransport (\textbf{AOT}). Specifically, we first establish local- and global-aware token anchors within each frame under the attention guidance, which then optimal transport aggregates the informative contexts from pruned tokens, constructing intra-frame token anchors. Then, building on the temporal frame clips, the first frame within each clip will be considered as the keyframe anchors to ensemble similar information from consecutive frames through optimal transport, while keeping distinct tokens to represent temporal dynamics, leading to efficient token reduction in a training-free manner. Extensive evaluations show that our proposed AOT obtains competitive performances across various short- and long-video benchmarks on leading video LLMs, obtaining substantial computational efficiency while preserving temporal and visual fidelity. Project webpage: \href{https://tyroneli.github.io/AOT}{AOT}.

Method: Introduces two key components: 1) Token-Pooling-based Mixture-of-Experts mechanism for feature aggregation and expert specialization, and 2) Target-aware Adaptive Distillation strategy that selectively performs distillation based on sample characteristics to reduce redundant supervision.

Result: Achieves 69.2% AUC on LaSOT benchmark and runs at 163/56/60 FPS on GPU/CPU/AGX platforms, demonstrating superior speed-accuracy trade-off compared to previous methods across 12 benchmarks on 3 hardware platforms.

Conclusion: UETrack addresses the gap in efficient multi-modal tracking with high practicality and versatility, efficiently handling multiple modalities while maintaining strong performance across various hardware platforms.

Abstract: With growing real-world demands, efficient tracking has received increasing attention. However, most existing methods are limited to RGB inputs and struggle in multi-modal scenarios. Moreover, current multi-modal tracking approaches typically use complex designs, making them too heavy and slow for resource-constrained deployment. To tackle these limitations, we propose UETrack, an efficient framework for single object tracking. UETrack demonstrates high practicality and versatility, efficiently handling multiple modalities including RGB, Depth, Thermal, Event, and Language, and addresses the gap in efficient multi-modal tracking. It introduces two key components: a Token-Pooling-based Mixture-of-Experts mechanism that enhances modeling capacity through feature aggregation and expert specialization, and a Target-aware Adaptive Distillation strategy that selectively performs distillation based on sample characteristics, reducing redundant supervision and improving performance. Extensive experiments on 12 benchmarks across 3 hardware platforms show that UETrack achieves a superior speed-accuracy trade-off compared to previous methods. For instance, UETrack-B achieves 69.2% AUC on LaSOT and runs at 163/56/60 FPS on GPU/CPU/AGX, demonstrating strong practicality and versatility. Code is available at https://github.com/kangben258/UETrack.

Method: ATA uses complementary attention-guided and action-guided strategies to formulate implicit reasoning by integrating attention maps with action-based regions of interest, adaptively refining visual inputs without training or annotations.

Result: Extensive experiments show ATA consistently improves task success and robustness while preserving and even enhancing inference efficiency compared to explicit reasoning methods.

Conclusion: ATA provides a lightweight, plug-and-play implicit reasoning approach for VLA models that addresses limitations of explicit reasoning methods without requiring additional training or annotations.

Abstract: Vision-Language-Action (VLA) models rely on current observations, including images, language instructions, and robot states, to predict actions and complete tasks. While accurate visual perception is crucial for precise action prediction and execution, recent work has attempted to further improve performance by introducing explicit reasoning during inference. However, such approaches face significant limitations. They often depend on data-intensive resources such as Chain-of-Thought (CoT) style annotations to decompose tasks into step-by-step reasoning, and in many cases require additional visual grounding annotations (e.g., bounding boxes or masks) to highlight relevant image regions. Moreover, they involve time-consuming dataset construction, labeling, and retraining, which ultimately results in longer inference sequences and reduced efficiency. To address these challenges, we propose ATA, a novel training-free framework that introduces implicit reasoning into VLA inference through complementary attention-guided and action-guided strategies. Unlike CoT or explicit visual-grounding methods, ATA formulates reasoning implicitly by integrating attention maps with an action-based region of interest (RoI), thereby adaptively refining visual inputs without requiring extra training or annotations. ATA is a plug-and-play implicit reasoning approach for VLA models, lightweight yet effective. Extensive experiments show that it consistently improves task success and robustness while preserving, and even enhancing, inference efficiency.

Method: SeaVIS uses Causal Cross Attention Fusion (CCAF) for online processing, integrating visual features from current frames with entire audio history under causal constraints. It employs Audio-Guided Contrastive Learning (AGCL) to generate instance prototypes encoding both visual appearance and sounding activity, enabling suppression of silent objects during instance association.

Result: SeaVIS surpasses existing state-of-the-art models on the AVISeg dataset across multiple evaluation metrics while maintaining competitive inference speed suitable for real-time processing.

Conclusion: SeaVIS is the first online framework for audio-visual instance segmentation that effectively handles continuous video streams and distinguishes between sounding and silent object states, demonstrating superior performance and real-time capability.

Abstract: Recently, an audio-visual instance segmentation (AVIS) task has been introduced, aiming to identify, segment and track individual sounding instances in videos. However, prevailing methods primarily adopt the offline paradigm, that cannot associate detected instances across consecutive clips, making them unsuitable for real-world scenarios that involve continuous video streams. To address this limitation, we introduce SeaVIS, the first online framework designed for audio-visual instance segmentation. SeaVIS leverages the Causal Cross Attention Fusion (CCAF) module to enable efficient online processing, which integrates visual features from the current frame with the entire audio history under strict causal constraints. A major challenge for conventional VIS methods is that appearance-based instance association fails to distinguish between an object’s sounding and silent states, resulting in the incorrect segmentation of silent objects. To tackle this, we employ an Audio-Guided Contrastive Learning (AGCL) strategy to generate instance prototypes that encode not only visual appearance but also sounding activity. In this way, instances preserved during per-frame prediction that do not emit sound can be effectively suppressed during instance association process, thereby significantly enhancing the audio-following capability of SeaVIS. Extensive experiments conducted on the AVISeg dataset demonstrate that SeaVIS surpasses existing state-of-the-art models across multiple evaluation metrics while maintaining a competitive inference speed suitable for real-time processing.

Method: Created DOCFORGE-BENCH as a unified zero-shot benchmark evaluating 14 methods across 8 datasets covering text tampering, receipt forgery, and identity document manipulation. All methods were applied with published pretrained weights without domain adaptation. Used Pixel-AUC and Pixel-F1 metrics with analysis of score-distribution shifts and calibration issues.

Result: Found pervasive calibration failure: methods achieve moderate Pixel-AUC (≥0.76) but near-zero Pixel-F1 due to score-distribution shift (tampered regions occupy only 0.27-4.17% of pixels). Oracle-F1 is 2-10x higher than fixed-threshold Pixel-F1, showing calibration is the bottleneck. Simple threshold adaptation on N=10 domain images recovers 39-55% of Oracle-F1 gap.

Conclusion: No evaluated method works reliably out-of-the-box on diverse document types, highlighting document forgery detection as an unsolved problem. Current datasets predate generative AI editing, creating a critical gap for modern attack surfaces involving diffusion- and LLM-based forgeries.

Abstract: We present DOCFORGE-BENCH, the first unified zero-shot benchmark for document forgery detection, evaluating 14 methods across eight datasets spanning text tampering, receipt forgery, and identity document manipulation. Unlike fine-tuning-oriented evaluations such as ForensicHub [Du et al., 2025], DOCFORGE-BENCH applies all methods with their published pretrained weights and no domain adaptation – a deliberate design choice that reflects the realistic deployment scenario where practitioners lack labeled document training data. Our central finding is a pervasive calibration failure invisible under single-threshold protocols: methods achieve moderate Pixel-AUC (>=0.76) yet near-zero Pixel-F1. This AUC-F1 gap is not a discrimination failure but a score-distribution shift: tampered regions occupy only 0.27-4.17% of pixels in document images – an order of magnitude less than in natural image benchmarks – making the standard tau=0.5 threshold catastrophically miscalibrated. Oracle-F1 is 2-10x higher than fixed-threshold Pixel-F1, confirming that calibration, not representation, is the bottleneck. A controlled calibration experiment validates this: adapting a single threshold on N=10 domain images recovers 39-55% of the Oracle-F1 gap, demonstrating that threshold adaptation – not retraining – is the key missing step for practical deployment. Overall, no evaluated method works reliably out-of-the-box on diverse document types, underscoring that document forgery detection remains an unsolved problem. We further note that all eight datasets predate the era of generative AI editing; benchmarks covering diffusion- and LLM-based document forgeries represent a critical open gap on the modern attack surface.

Method: Three key innovations: 1) Structural link via shared discrete codebook for language and action tokens, 2) Deep semantic link through auxiliary action understanding objective (generating captions from trajectories), 3) Two-step coarse-to-fine generation (C2F) to replace slow auto-regressive decoding.

Result: Experiments on closed-loop driving benchmarks show consistent gains in instruction following accuracy and driving performance, with 86% reduction in inference time compared to auto-regressive methods.

Conclusion: LinkVLA successfully addresses both alignment and efficiency challenges in VLA models for autonomous driving through structural and semantic linking mechanisms combined with efficient generation techniques.

Abstract: Vision-Language-Action (VLA) models are emerging as a promising paradigm for end-to-end autonomous driving, valued for their potential to leverage world knowledge and reason about complex driving scenes. However, existing methods suffer from two critical limitations: a persistent misalignment between language instructions and action outputs, and the inherent inefficiency of typical auto-regressive action generation. In this paper, we introduce LinkVLA, a novel architecture that directly addresses these challenges to enhance both alignment and efficiency. First, we establish a structural link by unifying language and action tokens into a shared discrete codebook, processed within a single multi-modal model. This structurally enforces cross-modal consistency from the ground up. Second, to create a deep semantic link, we introduce an auxiliary action understanding objective that trains the model to generate descriptive captions from trajectories, fostering a bidirectional language-action mapping. Finally, we replace the slow, step-by-step generation with a two-step coarse-to-fine generation method C2F that efficiently decodes the action sequence, saving 86% inference time. Experiments on closed-loop driving benchmarks show consistent gains in instruction following accuracy and driving performance, alongside reduced inference latency.

Method: Proposes Deepfake Forensics Adapter (DFA) - a dual-stream framework that integrates pre-trained CLIP with three components: 1) Global Feature Adapter for global inconsistency detection, 2) Local Anomaly Stream for local facial forgery cues using facial structure priors, and 3) Interactive Fusion Classifier using transformer encoder for feature fusion.

Result: Achieves state-of-the-art performance on DFDC dataset with frame-level AUC/EER of 0.816/0.256 and video-level AUC/EER of 0.836/0.251, representing 4.8% video AUC improvement over previous methods. Demonstrates superior generalization capabilities across benchmarks.

Conclusion: DFA provides an effective framework for robust deepfake detection with enhanced generalization against evolving threats, pointing to a feasible direction for developing practical detection systems using vision-language foundation models.

Abstract: The rapid advancement of deepfake generation techniques poses significant threats to public safety and causes societal harm through the creation of highly realistic synthetic facial media. While existing detection methods demonstrate limitations in generalizing to emerging forgery patterns, this paper presents Deepfake Forensics Adapter (DFA), a novel dual-stream framework that synergizes vision-language foundation models with targeted forensics analysis. Our approach integrates a pre-trained CLIP model with three core components to achieve specialized deepfake detection by leveraging the powerful general capabilities of CLIP without changing CLIP parameters: 1) A Global Feature Adapter is used to identify global inconsistencies in image content that may indicate forgery, 2) A Local Anomaly Stream enhances the model’s ability to perceive local facial forgery cues by explicitly leveraging facial structure priors, and 3) An Interactive Fusion Classifier promotes deep interaction and fusion between global and local features using a transformer encoder. Extensive evaluations of frame-level and video-level benchmarks demonstrate the superior generalization capabilities of DFA, particularly achieving state-of-the-art performance in the challenging DFDC dataset with frame-level AUC/EER of 0.816/0.256 and video-level AUC/EER of 0.836/0.251, representing a 4.8% video AUC improvement over previous methods. Our framework not only demonstrates state-of-the-art performance, but also points out a feasible and effective direction for developing a robust deepfake detection system with enhanced generalization capabilities against the evolving deepfake threats. Our code is available at https://github.com/Liao330/DFA.git

Method: 3R uses three strategies: 1) RAG-based modifiers extraction for enriched contextual grounding, 2) diffusion-based Preference Optimization to align outputs with human preferences, and 3) temporal frame interpolation for producing temporally consistent visual content. The framework works with any T2V model without training.

Result: Experimental results show 3R enhances static fidelity and dynamic coherence of generated videos, demonstrating the importance of optimizing user prompts for better text-to-video generation.

Conclusion: 3R provides an effective, scalable prompt optimization framework that improves text-to-video generation quality without requiring model training, making it accessible and applicable to various T2V models.

Abstract: While large-scale datasets have driven significant progress in Text-to-Video (T2V) generative models, these models remain highly sensitive to input prompts, demonstrating that prompt design is critical to generation quality. Current methods for improving video output often fall short: they either depend on complex, post-editing models, risking the introduction of artifacts, or require expensive fine-tuning of the core generator, which severely limits both scalability and accessibility. In this work, we introduce 3R, a novel RAG based prompt optimization framework. 3R utilizes the power of current state-of-the-art T2V diffusion model and vision language model. It can be used with any T2V model without any kind of model training. The framework leverages three key strategies: RAG-based modifiers extraction for enriched contextual grounding, diffusion-based Preference Optimization for aligning outputs with human preferences, and temporal frame interpolation for producing temporally consistent visual contents. Together, these components enable more accurate, efficient, and contextually aligned text-to-video generation. Experimental results demonstrate the efficacy of 3R in enhancing the static fidelity and dynamic coherence of generated videos, underscoring the importance of optimizing user prompts.

Method: Proposes UltraStar which reformulates probe navigation as anchor-based global localization. Establishes a Star Graph treating historical keyframes as spatial anchors connected directly to current view, modeling geometric constraints. Uses semantic-aware sampling to select representative landmarks from history logs.

Result: Extensive experiments on dataset with over 1.31 million samples show UltraStar outperforms baselines and scales better with longer input lengths, revealing more effective topology for history modeling under noisy exploration.

Conclusion: Star Graph approach provides robust navigation by better leveraging historical scanning information, addressing limitations of sequential chain modeling for noisy trajectory data.

Abstract: Echocardiography is critical for diagnosing cardiovascular diseases, yet the shortage of skilled sonographers hinders timely patient care, due to high operational difficulties. Consequently, research on automated probe navigation has significant clinical potential. To achieve robust navigation, it is essential to leverage historical scanning information, mimicking how experts rely on past feedback to adjust subsequent maneuvers. Practical scanning data collected from sonographers typically consists of noisy trajectories inherently generated through trial-and-error exploration. However, existing methods typically model this history as a sequential chain, forcing models to overfit these noisy paths, leading to performance degradation on long sequences. In this paper, we propose UltraStar, which reformulates probe navigation from path regression to anchor-based global localization. By establishing a Star Graph, UltraStar treats historical keyframes as spatial anchors connected directly to the current view, explicitly modeling geometric constraints for precise positioning. We further enhance the Star Graph with a semantic-aware sampling strategy that actively selects the representative landmarks from massive history logs, reducing redundancy for accurate anchoring. Extensive experiments on a dataset with over 1.31 million samples demonstrate that UltraStar outperforms baselines and scales better with longer input lengths, revealing a more effective topology for history modeling under noisy exploration.

Method: Adds a lightweight point track head to VLA models during training that predicts 3D point tracks. Injects VLA features into this head to jointly predict future 3D trajectories, learning evolving scene geometry within the shared representation space.

Result: Significantly improves performance on challenging manipulation tasks: +10% gain on LIBERO-Long and +40% gain on RoboCasa. Shows 3D point track prediction is effective supervision for learning action-world dynamics.

Conclusion: Pri4R successfully endows VLA models with implicit understanding of world dynamics through privileged 4D information during training, enabling more physically aware context for precise control without inference overhead.

Abstract: Humans learn not only how their bodies move, but also how the surrounding world responds to their actions. In contrast, while recent Vision-Language-Action (VLA) models exhibit impressive semantic understanding, they often fail to capture the spatiotemporal dynamics governing physical interaction. In this paper, we introduce Pri4R, a simple yet effective approach that endows VLA models with an implicit understanding of world dynamics by leveraging privileged 4D information during training. Specifically, Pri4R augments VLAs with a lightweight point track head that predicts 3D point tracks. By injecting VLA features into this head to jointly predict future 3D trajectories, the model learns to incorporate evolving scene geometry within its shared representation space, enabling more physically aware context for precise control. Due to its architectural simplicity, Pri4R is compatible with dominant VLA design patterns with minimal changes. During inference, we run the model using the original VLA architecture unchanged; Pri4R adds no extra inputs, outputs, or computational overhead. Across simulation and real-world evaluations, Pri4R significantly improves performance on challenging manipulation tasks, including a +10% gain on LIBERO-Long and a +40% gain on RoboCasa. We further show that 3D point track prediction is an effective supervision target for learning action-world dynamics, and validate our design choices through extensive ablations.

Method: The authors propose WildCross, a cross-modal benchmark comprising over 476K sequential RGB frames with semi-dense depth and surface normal annotations, each aligned with accurate 6DoF poses and synchronized dense lidar submaps.

Result: Comprehensive experiments on visual, lidar, and cross-modal place recognition, as well as metric depth estimation, demonstrate the value of WildCross as a challenging benchmark for multi-modal robotic perception tasks.

Conclusion: WildCross addresses the critical gap in robotics datasets for natural environments and provides a valuable resource for advancing multi-modal robotic perception research in unstructured settings.

Abstract: Recent years have seen a significant increase in demand for robotic solutions in unstructured natural environments, alongside growing interest in bridging 2D and 3D scene understanding. However, existing robotics datasets are predominantly captured in structured urban environments, making them inadequate for addressing the challenges posed by complex, unstructured natural settings. To address this gap, we propose WildCross, a cross-modal benchmark for place recognition and metric depth estimation in large-scale natural environments. WildCross comprises over 476K sequential RGB frames with semi-dense depth and surface normal annotations, each aligned with accurate 6DoF poses and synchronized dense lidar submaps. We conduct comprehensive experiments on visual, lidar, and cross-modal place recognition, as well as metric depth estimation, demonstrating the value of WildCross as a challenging benchmark for multi-modal robotic perception tasks. We provide access to the code repository and dataset at https://csiro-robotics.github.io/WildCross.

Method: SCATR introduces two architecture-agnostic training strategies: 1) Second Chance Assignment - concatenates unassigned track queries to proposal queries before bipartite matching, giving them another chance to be assigned to ground truth objects; 2) Track Query Dropout - diversifies supervised object query configurations to train the decoder to handle different track query sets, enhancing robustness to missing or newborn tracks.

Result: On the nuScenes tracking benchmark, SCATR achieves state-of-the-art performance among LiDAR-based TBA methods, outperforming previous works by 7.6% AMOTA and successfully bridging the performance gap between LiDAR-based TBA and TBD methods.

Conclusion: The proposed Second Chance Assignment and Track Query Dropout strategies effectively address the false negative problem in LiDAR-based tracking-by-attention frameworks, demonstrating their effectiveness and generalization through ablation studies.

Abstract: LiDAR-based tracking-by-attention (TBA) frameworks inherently suffer from high false negative errors, leading to a significant performance gap compared to traditional LiDAR-based tracking-by-detection (TBD) methods. This paper introduces SCATR, a novel LiDAR-based TBA model designed to address this fundamental challenge systematically. SCATR leverages recent progress in vision-based tracking and incorporates targeted training strategies specifically adapted for LiDAR. Our work’s core innovations are two architecture-agnostic training strategies for TBA methods: Second Chance Assignment and Track Query Dropout. Second Chance Assignment is a novel ground truth assignment that concatenates unassigned track queries to the proposal queries before bipartite matching, giving these track queries a second chance to be assigned to a ground truth object and effectively mitigating the conflict between detection and tracking tasks inherent in tracking-by-attention. Track Query Dropout is a training method that diversifies supervised object query configurations to efficiently train the decoder to handle different track query sets, enhancing robustness to missing or newborn tracks. Experiments on the nuScenes tracking benchmark demonstrate that SCATR achieves state-of-the-art performance among LiDAR-based TBA methods, outperforming previous works by 7.6% AMOTA and successfully bridging the long-standing performance gap between LiDAR-based TBA and TBD methods. Ablation studies further validate the effectiveness and generalization of Second Chance Assignment and Track Query Dropout. Code can be found at the following link: \href{https://github.com/TRAILab/SCATR}{https://github.com/TRAILab/SCATR}

Method: Introduces SkeleGuide framework with explicit skeletal reasoning, joint training of reasoning and rendering stages to produce internal pose as structural prior, and PoseInverter module for decoding latent pose into editable format for user control.

Result: SkeleGuide significantly outperforms both specialized and general-purpose models in generating high-fidelity, contextually-aware human images with better structural integrity.

Conclusion: Explicitly modeling skeletal structure is fundamental for robust and plausible human image synthesis, and SkeleGuide demonstrates the effectiveness of this approach.

Abstract: Generating realistic and structurally plausible human images into existing scenes remains a significant challenge for current generative models, which often produce artifacts like distorted limbs and unnatural poses. We attribute this systemic failure to an inability to perform explicit reasoning over human skeletal structure. To address this, we introduce SkeleGuide, a novel framework built upon explicit skeletal reasoning. Through joint training of its reasoning and rendering stages, SkeleGuide learns to produce an internal pose that acts as a strong structural prior, guiding the synthesis towards high structural integrity. For fine-grained user control, we introduce PoseInverter, a module that decodes this internal latent pose into an explicit and editable format. Extensive experiments demonstrate that SkeleGuide significantly outperforms both specialized and general-purpose models in generating high-fidelity, contextually-aware human images. Our work provides compelling evidence that explicitly modeling skeletal structure is a fundamental step towards robust and plausible human image synthesis.

Method: Proposes radiometric consistency - a physically-based constraint that minimizes residual between each Gaussian primitive’s learned radiance and its physically-based rendered counterpart for unobserved views. This creates a self-correcting feedback loop combining physically-based rendering and novel-view synthesis supervision. Implements RadioGS framework using Gaussian surfels and 2D Gaussian ray tracing, plus a finetuning-based relighting strategy.

Result: RadioGS outperforms existing Gaussian-based methods in inverse rendering benchmarks while maintaining computational efficiency (<10ms rendering cost). The finetuning-based relighting adapts Gaussian surfel radiances to new illuminations within minutes.

Conclusion: Radiometric consistency effectively addresses the indirect illumination modeling challenge in inverse rendering with Gaussian splatting, enabling accurate material disentanglement while preserving computational efficiency.

Abstract: Inverse rendering with Gaussian Splatting has advanced rapidly, but accurately disentangling material properties from complex global illumination effects, particularly indirect illumination, remains a major challenge. Existing methods often query indirect radiance from Gaussian primitives pre-trained for novel-view synthesis. However, these pre-trained Gaussian primitives are supervised only towards limited training viewpoints, thus lack supervision for modeling indirect radiances from unobserved views. To address this issue, we introduce radiometric consistency, a novel physically-based constraint that provides supervision towards unobserved views by minimizing the residual between each Gaussian primitive’s learned radiance and its physically-based rendered counterpart. Minimizing the residual for unobserved views establishes a self-correcting feedback loop that provides supervision from both physically-based rendering and novel-view synthesis, enabling accurate modeling of inter-reflection. We then propose Radiometrically Consistent Gaussian Surfels (RadioGS), an inverse rendering framework built upon our principle by efficiently integrating radiometric consistency by utilizing Gaussian surfels and 2D Gaussian ray tracing. We further propose a finetuning-based relighting strategy that adapts Gaussian surfel radiances to new illuminations within minutes, achieving low rendering cost (<10ms). Extensive experiments on existing inverse rendering benchmarks show that RadioGS outperforms existing Gaussian-based methods in inverse rendering, while retaining the computational efficiency.

Method: YCDa (chrominance-luminance decoupling and dynamic attention) separates color and luminance information at the input stage, then dynamically allocates attention across channels to amplify discriminative cues while suppressing misleading color noise. It’s a plug-and-play strategy that can be integrated into existing detectors by replacing the first downsampling layer.

Result: YCDa consistently improves performance with negligible overhead across multiple baselines. YCDa-YOLO12s achieves 112% improvement in mAP over baseline on COD10K-D and sets new state-of-the-art results for real-time camouflaged object detection across COD-D datasets.

Conclusion: The biologically-inspired YCDa strategy effectively enhances real-time camouflaged object detection by mimicking human visual adaptation mechanisms, demonstrating significant performance gains with minimal computational cost.

Abstract: Human vision exhibits remarkable adaptability in perceiving objects under camouflage. When color cues become unreliable, the visual system instinctively shifts its reliance from chrominance (color) to luminance (brightness and texture), enabling more robust perception in visually confusing environments. Drawing inspiration from this biological mechanism, we propose YCDa, an efficient early-stage feature processing strategy that embeds this “chrominance-luminance decoupling and dynamic attention” principle into modern real-time detectors. Specifically, YCDa separates color and luminance information in the input stage and dynamically allocates attention across channels to amplify discriminative cues while suppressing misleading color noise. The strategy is plug-and-play and can be integrated into existing detectors by simply replacing the first downsampling layer. Extensive experiments on multiple baselines demonstrate that YCDa consistently improves performance with negligible overhead as shown in Fig. Notably, YCDa-YOLO12s achieves a 112% improvement in mAP over the baseline on COD10K-D and sets new state-of-the-art results for real-time camouflaged object detection across COD-D datasets.

Method: Three-path architecture: main path with DINOv3 backbone for coarse features, auxiliary siamese path for fine-grained features, and multi-scale attention decoder with parallel convolutions for contextual enhancement

Result: Achieves optimal performance on Gaza facility damage assessment dataset and SECOND dataset; GradCAM shows main path focuses on semantic changes while auxiliary path captures structural details

Conclusion: Tripath DINO provides robust, interpretable solution for advanced change detection with synergistic complementarity between coarse and fine-grained feature learning

Abstract: In remote sensing imagery, multi class change detection (MCD) is crucial for fine grained monitoring, yet it has long been constrained by complex scene variations and the scarcity of detailed annotations. To address this, we propose the Tripath DINO architecture, which adopts a three path complementary feature learning strategy to facilitate the rapid adaptation of pre trained foundation models to complex vertical domains. Specifically, we employ the DINOv3 pre trained model as the backbone feature extraction network to learn coarse grained features. An auxiliary path also adopts a siamese structure, progressively aggregating intermediate features from the siamese encoder to enhance the learning of fine grained features. Finally, a multi scale attention mechanism is introduced to augment the decoder network, where parallel convolutions adaptively capture and enhance contextual information under different receptive fields. The proposed method achieves optimal performance on the MCD task on both the Gaza facility damage assessment dataset (Gaza change) and the classic SECOND dataset. GradCAM visualizations further confirm that the main and auxiliary paths naturally focus on coarse grained semantic changes and fine grained structural details, respectively. This synergistic complementarity provides a robust and interpretable solution for advanced change detection tasks, offering a basis for rapid and accurate damage assessment.

Method: BiCAM (Bidirectional Class Activation Mapping) preserves signed attributions to capture both positive and negative contributions. It introduces a Positive-to-Negative Ratio (PNR) that summarizes attribution balance. The method works with multiple ViT variants including DeiT and Swin Transformers, and enables lightweight adversarial example detection without retraining.

Result: BiCAM improves localization and faithfulness across ImageNet, VOC, and COCO datasets while remaining computationally efficient. It generalizes to multiple ViT variants and enables effective adversarial example detection through the PNR metric.

Conclusion: Modeling both supportive and suppressive evidence is important for interpreting transformer-based vision models. BiCAM provides more complete and contrastive explanations while maintaining computational efficiency.

Abstract: Vision Transformers (ViTs) achieve strong performance in visual recognition, yet their decision-making remains difficult to interpret. We propose BiCAM, a bidirectional class activation mapping method that captures both supportive (positive) and suppressive (negative) contributions to model predictions. Unlike prior CAM-based approaches that discard negative signals, BiCAM preserves signed attributions to produce more complete and contrastive explanations. BiCAM further introduces a Positive-to-Negative Ratio (PNR) that summarizes attribution balance and enables lightweight detection of adversarial examples without retraining. Across ImageNet, VOC, and COCO, BiCAM improves localization and faithfulness while remaining computationally efficient. It generalizes to multiple ViT variants, including DeiT and Swin. These results suggest the importance of modeling both supportive and suppressive evidence for interpreting transformer-based vision models.

Method: Uses Multi-LOD Gaussian representation with transformer-based global feature extraction and projection-based local feature sampling. Features fused with depth buffer guidance. Employs coarse-to-fine learning for LOD functionality and multi-region decomposition separating head and shoulder prediction then combining them.

Result: Outperforms state-of-the-art methods in reconstruction quality, reenactment performance, and computational efficiency (0.2s inference time). Supports Level-of-Detail functionality for diverse hardware capabilities.

Conclusion: OMG-Avatar provides fast, high-quality animatable 3D head reconstruction from single images with LOD support and better handling of non-head regions, making it practical for real-time applications.

Abstract: We propose OMG-Avatar, a novel One-shot method that leverages a Multi-LOD (Level-of-Detail) Gaussian representation for animatable 3D head reconstruction from a single image in 0.2s. Our method enables LOD head avatar modeling using a unified model that accommodates diverse hardware capabilities and inference speed requirements. To capture both global and local facial characteristics, we employ a transformer-based architecture for global feature extraction and projection-based sampling for local feature acquisition. These features are effectively fused under the guidance of a depth buffer, ensuring occlusion plausibility. We further introduce a coarse-to-fine learning paradigm to support Level-of-Detail functionality and enhance the perception of hierarchical details. To address the limitations of 3DMMs in modeling non-head regions such as the shoulders, we introduce a multi-region decomposition scheme in which the head and shoulders are predicted separately and then integrated through cross-region combination. Extensive experiments demonstrate that OMG-Avatar outperforms state-of-the-art methods in reconstruction quality, reenactment performance, and computational efficiency.

Method: Introduces FACE, an Autoregressive Autoencoder (ARAE) framework that generates meshes at face level using a one-face-one-token strategy. Each triangle face is treated as a single unified token, reducing sequence length by a factor of nine. Pairs a face-level decoder with a powerful VecSet encoder for efficient representation.

Result: Achieves unprecedented compression ratio of 0.11 (halving previous SOTA), reduces sequence length by 9x, achieves state-of-the-art reconstruction quality on standard benchmarks. The learned latent space enables training a latent diffusion model for high-fidelity single-image-to-mesh generation.

Conclusion: FACE provides a simple, scalable, and powerful paradigm that lowers the barrier to high-quality structured 3D content creation by operating at the appropriate semantic level (face instead of vertex).

Abstract: Autoregressive models for 3D mesh generation suffer from a fundamental limitation: they flatten meshes into long vertex-coordinate sequences. This results in prohibitive computational costs, hindering the efficient synthesis of high-fidelity geometry. We argue this bottleneck stems from operating at the wrong semantic level. We introduce FACE, a novel Autoregressive Autoencoder (ARAE) framework that reconceptualizes the task by generating meshes at the face level. Our one-face-one-token strategy treats each triangle face, the fundamental building block of a mesh, as a single, unified token. This simple yet powerful design reduces the sequence length by a factor of nine, leading to an unprecedented compression ratio of 0.11, halving the previous state-of-the-art. This dramatic efficiency gain does not compromise quality; by pairing our face-level decoder with a powerful VecSet encoder, FACE achieves state-of-the-art reconstruction quality on standard benchmarks. The versatility of the learned latent space is further demonstrated by training a latent diffusion model that achieves high-fidelity, single-image-to-mesh generation. FACE provides a simple, scalable, and powerful paradigm that lowers the barrier to high-quality structured 3D content creation.

Method: Multi-view video reward shaping (MVR) uses videos from multiple viewpoints with video-text similarity from frozen VLMs to learn state relevance functions, plus state-dependent reward shaping that reduces VLM influence once desired motion is achieved.

Result: Extensive experiments on HumanoidBench locomotion and MetaWorld manipulation tasks demonstrate efficacy, with ablation studies verifying design choices.

Conclusion: MVR effectively addresses limitations of image-based VLM reward methods for complex dynamic tasks through multi-view video analysis and adaptive reward shaping.

Abstract: Reward design is of great importance for solving complex tasks with reinforcement learning. Recent studies have explored using image-text similarity produced by vision-language models (VLMs) to augment rewards of a task with visual feedback. A common practice linearly adds VLM scores to task or success rewards without explicit shaping, potentially altering the optimal policy. Moreover, such approaches, often relying on single static images, struggle with tasks whose desired behavior involves complex, dynamic motions spanning multiple visually different states. Furthermore, single viewpoints can occlude critical aspects of an agent’s behavior. To address these issues, this paper presents Multi-View Video Reward Shaping (MVR), a framework that models the relevance of states regarding the target task using videos captured from multiple viewpoints. MVR leverages video-text similarity from a frozen pre-trained VLM to learn a state relevance function that mitigates the bias towards specific static poses inherent in image-based methods. Additionally, we introduce a state-dependent reward shaping formulation that integrates task-specific rewards and VLM-based guidance, automatically reducing the influence of VLM guidance once the desired motion pattern is achieved. We confirm the efficacy of the proposed framework with extensive experiments on challenging humanoid locomotion tasks from HumanoidBench and manipulation tasks from MetaWorld, verifying the design choices through ablation studies.

Method: Proposes Quality-guided Min-Cost Max-Flow (Q-MCMF) matcher that builds a flow graph, prunes implausible matches based on geometric quality, and optimizes for final matching that minimizes cost while maximizing valid assignments.

Result: Extensive experiments on COCO dataset under various incremental settings demonstrate consistent outperformance over existing state-of-the-art approaches.

Conclusion: Q-MCMF effectively addresses background foregrounding in DETR-like architectures for incremental object detection, eliminating harmful supervision while maximizing foreground learning signals.

Abstract: Incremental Object Detection (IOD) aims to continuously learn new object classes without forgetting previously learned ones. A persistent challenge is catastrophic forgetting, primarily attributed to background shift in conventional detectors. While pseudo-labeling mitigates this in dense detectors, we identify a novel, distinct source of forgetting specific to DETR-like architectures: background foregrounding. This arises from the exhaustiveness constraint of the Hungarian matcher, which forcibly assigns every ground truth target to one prediction, even when predictions primarily cover background regions (i.e., low IoU). This erroneous supervision compels the model to misclassify background features as specific foreground classes, disrupting learned representations and accelerating forgetting. To address this, we propose a Quality-guided Min-Cost Max-Flow (Q-MCMF) matcher. To avoid forced assignments, Q-MCMF builds a flow graph and prunes implausible matches based on geometric quality. It then optimizes for the final matching that minimizes cost and maximizes valid assignments. This strategy eliminates harmful supervision from background foregrounding while maximizing foreground learning signals. Extensive experiments on the COCO dataset under various incremental settings demonstrate that our method consistently outperforms existing state-of-the-art approaches.

Method: Proposes Cross-modal Identity Mapping (CIM), a reinforcement learning framework that evaluates information loss from two perspectives: Gallery Representation Consistency (how well captions represent image content) and Query-gallery Image Relevance (similarity between retrieved and original images). Uses these metrics as supervision signals without additional annotations.

Result: CIM demonstrates superior performance in image captioning, even outperforming Supervised Fine-Tuning. On COCO-LN500 benchmark, achieves 20% improvement in relation reasoning on Qwen2.5-VL-7B model.

Conclusion: The proposed CIM framework effectively enhances LVLM image captioning by minimizing information loss through cross-modal identity mapping, using image retrieval similarity as a novel proxy for measuring caption quality without requiring additional annotations.

Abstract: Large Vision-Language Models (LVLMs) often omit or misrepresent critical visual content in generated image captions. Minimizing such information loss will force LVLMs to focus on image details to generate precise descriptions. However, measuring information loss during modality conversion is inherently challenging due to the modal gap between visual content and text output. In this paper, we argue that the quality of an image caption is positively correlated with the similarity between images retrieved via text search using that caption. Based on this insight, we further propose Cross-modal Identity Mapping (CIM), a reinforcement learning framework that enhances image captioning without requiring additional annotations. Specifically, the method quantitatively evaluates the information loss from two perspectives: Gallery Representation Consistency and Query-gallery Image Relevance. Supervised under these metrics, LVLM minimizes information loss and aims to achieve identity mapping from images to captions. The experimental results demonstrate the superior performance of our method in image captioning, even when compared with Supervised Fine-Tuning. Particularly, on the COCO-LN500 benchmark, CIM achieves a 20% improvement in relation reasoning on Qwen2.5-VL-7B.The code will be released when the paper is accepted.

Method: Proposes a streaming inference pipeline that integrates an object detection model with a Finite State Machine (FSM). The FSM encodes knowledge of operational scenarios to guide and correct AI predictions on streaming video data for excavator workload counting.

Result: Tested on real-world dataset of over 7,000 images from 12 site videos with 300+ excavator workloads. The method demonstrated superior performance and greater robustness compared to original manual heuristic rules solution.

Conclusion: The proposed approach successfully enhances industrial AI applications by combining data-driven models with prior knowledge through FSM, improving robustness in real-world scenarios. Code will be released publicly.

Abstract: The widespread adoption of AI in industry is often hampered by its limited robustness when faced with scenarios absent from training data, leading to prediction bias and vulnerabilities. To address this, we propose a novel streaming inference pipeline that enhances data-driven models by explicitly incorporating prior knowledge. This paper presents the work on an industrial AI application that automatically counts excavator workloads from surveillance videos. Our approach integrates an object detection model with a Finite State Machine (FSM), which encodes knowledge of operational scenarios to guide and correct the AI’s predictions on streaming data. In experiments on a real-world dataset of over 7,000 images from 12 site videos, encompassing more than 300 excavator workloads, our method demonstrates superior performance and greater robustness compared to the original solution based on manual heuristic rules. We will release the code at https://github.com/thulab/video-streamling-inference-pipeline.

Method: Proposes Class-Aware Spectral Distribution Matching (CSDM) that reformulates distribution alignment via kernel spectrum analysis, creating Spectral Distribution Distance (SDD), and uses amplitude-phase decomposition to adaptively prioritize tail class realism.

Result: On CIFAR-10-LT, CSDM achieves 14.0% improvement over SOTA DD methods with only 5.7% performance drop when tail class images decrease from 500 to 25.

Conclusion: CSDM effectively addresses long-tailed dataset distillation challenges through spectral analysis and adaptive class balancing, demonstrating strong stability on imbalanced data.

Abstract: Dataset distillation (DD) aims to compress large-scale datasets into compact synthetic counterparts for efficient model training. However, existing DD methods exhibit substantial performance degradation on long-tailed datasets. We identify two fundamental challenges: heuristic design choices for distribution discrepancy measure and uniform treatment of imbalanced classes. To address these limitations, we propose Class-Aware Spectral Distribution Matching (CSDM), which reformulates distribution alignment via the spectrum of a well-behaved kernel function. This technique maps the original samples into frequency space, resulting in the Spectral Distribution Distance (SDD). To mitigate class imbalance, we exploit the unified form of SDD to perform amplitude-phase decomposition, which adaptively prioritizes the realism in tail classes. On CIFAR-10-LT, with 10 images per class, CSDM achieves a 14.0% improvement over state-of-the-art DD methods, with only a 5.7% performance drop when the number of images in tail classes decreases from 500 to 25, demonstrating strong stability on long-tailed data.

Method: Uses diffusion models to edit visual attributes (color, material, size, position at object level; weather, style at image level) of existing real images while preserving structural information, enabling reuse of segmentation labels to reduce annotation costs.

Result: Created Pascal-EA and COCO-EA benchmarks; found that open-vocabulary models aren’t more robust to geometric variations, data augmentation techniques are limited against appearance variations, and the pipeline can improve both in-distribution and out-of-distribution performance when used for data augmentation.

Conclusion: Generative models show potential as effective tools for automatically analyzing segmentation models, and the findings can help develop more robust and reliable segmentation models.

Abstract: Semantic segmentation takes pivotal roles in various applications such as autonomous driving and medical image analysis. When deploying segmentation models in practice, it is critical to test their behaviors in varied and complex scenes in advance. In this paper, we construct an automatic data generation pipeline Gen4Seg to stress-test semantic segmentation models by generating various challenging samples with different attribute changes. Beyond previous evaluation paradigms focusing solely on global weather and style transfer, we investigate variations in both appearance and geometry attributes at the object and image level. These include object color, material, size, position, as well as image-level variations such as weather and style. To achieve this, we propose to edit visual attributes of existing real images with precise control of structural information, empowered by diffusion models. In this way, the existing segmentation labels can be reused for the edited images, which greatly reduces the labor costs. Using our pipeline, we construct two new benchmarks, Pascal-EA and COCO-EA. We benchmark a wide variety of semantic segmentation models, spanning from closed-set models to open-vocabulary large models. We have several key findings: 1) advanced open-vocabulary models do not exhibit greater robustness compared to closed-set methods under geometric variations; 2) data augmentation techniques, such as CutOut and CutMix, are limited in enhancing robustness against appearance variations; 3) our pipeline can also be employed as a data augmentation tool and improve both in-distribution and out-of-distribution performances. Our work suggests the potential of generative models as effective tools for automatically analyzing segmentation models, and we hope our findings will assist practitioners and researchers in developing more robust and reliable segmentation models.

Method: Introduces Robustness Asymmetry Detection (RA-Det): 1) Identifies that natural images preserve stable semantic representations under small structured perturbations, while generated images exhibit larger feature drift. 2) Provides theoretical analysis connecting this asymmetry to memorization tendencies in generative models. 3) Converts robustness asymmetry into a detection signal using a behavior-driven framework that is data- and model-agnostic.

Result: Evaluated across 14 diverse generative models and against more than 10 strong detectors, RA-Det achieves superior performance, improving average performance by 7.81%. The method transfers across unseen generators and requires no generator fingerprints.

Conclusion: Robustness asymmetry is a stable, general cue for synthetic-image detection. Carefully designed probing can turn this behavioral signal into a practical, universal detector for AI-generated images.

Abstract: Recent image generators produce photo-realistic content that undermines the reliability of downstream recognition systems. As visual appearance cues become less pronounced, appearance-driven detectors that rely on forensic cues or high-level representations lose stability. This motivates a shift from appearance to behavior, focusing on how images respond to controlled perturbations rather than how they look. In this work, we identify a simple and universal behavioral signal. Natural images preserve stable semantic representations under small, structured perturbations, whereas generated images exhibit markedly larger feature drift. We refer to this phenomenon as robustness asymmetry and provide a theoretical analysis that establishes a lower bound connecting this asymmetry to memorization tendencies in generative models, explaining its prevalence across architectures. Building on this insight, we introduce Robustness Asymmetry Detection (RA-Det), a behavior-driven detection framework that converts robustness asymmetry into a reliable decision signal. Evaluated across 14 diverse generative models and against more than 10 strong detectors, RA-Det achieves superior performance, improving the average performance by 7.81 percent. The method is data- and model-agnostic, requires no generator fingerprints, and transfers across unseen generators. Together, these results indicate that robustness asymmetry is a stable, general cue for synthetic-image detection and that carefully designed probing can turn this cue into a practical, universal detector. The source code is publicly available at Github.

Method: Analyzed multi-task learning for hierarchical multi-label classification (car make and model). Evaluated parallel and cascaded multi-task architectures with different deep learning classifiers (CNNs, Transformers), varying dropout rates and loss weighting. Tested on StanfordCars and CompCars benchmarks.

Result: Multi-task learning improved performance on both datasets. CNNs showed improvements in almost all scenarios. Both CNN and Transformer models showed significant improvements on the CompCars dataset.

Conclusion: Multi-task learning is effective for hierarchical classification tasks like car make and model recognition, demonstrating that leveraging hierarchical structure can improve deep learning model performance.

Abstract: Most information in our world is organized hierarchically; however, many Deep Learning approaches do not leverage this semantically rich structure. Research suggests that human learning benefits from exploiting the hierarchical structure of information, and intelligent models could similarly take advantage of this through multi-task learning. In this work, we analyze the advantages and limitations of multi-task learning in a hierarchical multi-label classification problem: car make and model classification. Considering both parallel and cascaded multi-task architectures, we evaluate their impact on different Deep Learning classifiers (CNNs, Transformers) while varying key factors such as dropout rate and loss weighting to gain deeper insight into the effectiveness of this approach. The tests are conducted on two established benchmarks: StanfordCars and CompCars. We observe the effectiveness of the multi-task paradigm on both datasets, improving the performance of the investigated CNN in almost all scenarios. Furthermore, the approach yields significant improvements on the CompCars dataset for both types of models.

Method: Proposes SDAM with three key components: 1) Training-free framework using only pre-trained models, 2) Adaptive Object Memory module that selects and memorizes key objects based on motion cues, and 3) Spatio-temporal Decoupling approach where spatial domain handles precise localization/segmentation while temporal domain uses key object information for stable cross-frame propagation.

Result: Achieves excellent results on five benchmark datasets: Ref-YouTubeVOS, Ref-DAVIS17, MeViS, ReasonVOS, and ReVOS, outperforming existing methods that require fine-tuning.

Conclusion: SDAM demonstrates that training-free approaches can surpass fine-tuned methods for reasoning video object segmentation through effective spatio-temporal decoupling and adaptive memory mechanisms.

Abstract: Reasoning Video Object Segmentation (ReasonVOS) is a challenging task that requires stable object segmentation across video sequences using implicit and complex textual inputs. Previous methods fine-tune Multimodal Large Language Models (MLLMs) to produce segmentation outputs, which demand substantial resources. Additionally, some existing methods are coupled in the processing of spatio-temporal information, which affects the temporal stability of the model to some extent. To address these issues, we propose Training-Free \textbf{S}patio-temporal \textbf{D}ecoupled Reasoning Video Segmentation with \textbf{A}daptive Object \textbf{M}emory (SDAM). We aim to design a training-free reasoning video segmentation framework that outperforms existing methods requiring fine-tuning, using only pre-trained models. Meanwhile, we propose an Adaptive Object Memory module that selects and memorizes key objects based on motion cues in different video sequences. Finally, we propose Spatio-temporal Decoupling for stable temporal propagation. In the spatial domain, we achieve precise localization and segmentation of target objects, while in the temporal domain, we leverage key object temporal information to drive stable cross-frame propagation. Our method achieves excellent results on five benchmark datasets, including Ref-YouTubeVOS, Ref-DAVIS17, MeViS, ReasonVOS, and ReVOS.

Method: Proposes PathMoE, an interpretable multimodal framework that integrates H&E slides, pathology reports, and nuclei-level cell graphs via an interaction-aware mixture-of-experts architecture built on state-of-the-art foundation models for each modality. Uses specialized experts to capture modality uniqueness, redundancy, and synergy with an input-dependent gating mechanism.

Result: PathMoE improves macro-F1 from 0.762 to 0.799 (+0.037) on internal pediatric brain tumor dataset when integrating WSI, text, and graph modalities. On TCGA, augmenting WSI with graph knowledge improves macro-F1 from 0.668 to 0.709 (+0.041).

Conclusion: The framework demonstrates significant performance gains over image-only baselines while providing sample-level interpretability, which is critical for rare tumor subtypes and clinical trust.

Abstract: Accurate classification of pediatric central nervous system tumors remains challenging due to histological complexity and limited training data. While pathology foundation models have advanced whole-slide image (WSI) analysis, they often fail to leverage the rich, complementary information found in clinical text and tissue microarchitecture. To this end, we propose PathMoE, an interpretable multimodal framework that integrates H&E slides, pathology reports, and nuclei-level cell graphs via an interaction-aware mixture-of-experts architecture built on state-of-the-art foundation models for each modality. By training specialized experts to capture modality uniqueness, redundancy, and synergy, PathMoE employs an input-dependent gating mechanism that dynamically weights these interactions, providing sample-level interpretability. We evaluate our framework on two dataset-specific classification tasks on an internal pediatric brain tumor dataset (PBT) and external TCGA datasets. PathMoE improves macro-F1 from 0.762 to 0.799 (+0.037) on PBT when integrating WSI, text, and graph modalities; on TCGA, augmenting WSI with graph knowledge improves macro-F1 from 0.668 to 0.709 (+0.041). These results demonstrate significant performance gains over state-of-the-art image-only baselines while revealing the specific modality interactions driving individual predictions. This interpretability is particularly critical for rare tumor subtypes, where transparent model reasoning is essential for clinical trust and diagnostic validation.

Method: Proposes a diffusion autoencoder framework with explicit alignment objective to focus on progression-related regions, and structured latent subspaces separating progression conditions from subject identity information.

Result: The approach demonstrates more anatomically precise modeling of Alzheimer’s disease progression through better alignment and latent space structuring.

Conclusion: Enforcing alignment between modalities and structuring latent representational space improves controllability and anatomical precision in disease progression modeling using diffusion autoencoders.

Abstract: Generative AI framework-based modeling and prediction of longitudinal human brain images offer an efficient mechanism to track neurodegenerative progression, essential for the assessment of diseases like Alzheimer’s. Among the existing generative approaches, recent diffusion-based models have emerged as an effective alternative to generate disease progression images. Incorporating multi-modal and non-imaging attributes as conditional information into diffusion frameworks has been shown to improve controllability during such generations. However, existing methods do not explicitly ensure that information from non-imaging conditioning modalities is meaningfully aligned with image features to introduce desirable changes in the generated images, such as modulation of progression-specific regions. Further, more precise control over the generation process can be achieved by introducing progression-relevant structure into the internal representations of the model, lacking in the existing approaches. To address these limitations, we propose a diffusion autoencoder-based framework for disease progression modeling that explicitly enforces alignment between different modalities. The alignment is enforced by introducing an explicit objective function that enables the model to focus on the regions exhibiting progression-related changes. Further, we devise a mechanism to better structure the latent representational space of the diffusion auto-encoding framework. Specifically, we assign separate latent subspaces for integrating progression-related conditions and retaining subject-specific identity information, allowing better-controlled image generation. These results demonstrate that enforcing alignment and better structuring of the latent representational space of diffusion auto-encoding framework leads to more anatomically precise modeling of Alzheimer’s disease progression.

Method: TopoMaskV3 advances mask-based pipeline to 3D prediction via two novel dense prediction heads: dense offset field for sub-grid discretization correction and dense height map for direct 3D estimation. Introduces geographically distinct splits and long-range (+/-100m) benchmark to prevent memorization.

Result: Achieves state-of-the-art 28.5 OLS on geographically disjoint benchmark, surpassing all prior methods. Mask representation shows more robustness to geographic overfitting than Bezier methods. LiDAR fusion benefits most at long range and shows larger gains on overlapping splits, suggesting overlap-induced memorization effects.

Conclusion: TopoMaskV3 provides a robust standalone 3D predictor for road topology understanding that addresses geographic data leakage issues and demonstrates superior performance on fair evaluation benchmarks.

Abstract: Mask-based paradigms for road topology understanding, such as TopoMaskV2, offer a complementary alternative to query-based methods by generating centerlines via a dense rasterized intermediate representation. However, prior work was limited to 2D predictions and suffered from severe discretization artifacts, necessitating fusion with parametric heads. We introduce TopoMaskV3, which advances this pipeline into a robust, standalone 3D predictor via two novel dense prediction heads: a dense offset field for sub-grid discretization correction within the existing BEV resolution, and a dense height map for direct 3D estimation. Beyond the architecture, we are the first to address geographic data leakage in road topology evaluation by introducing (1) geographically distinct splits to prevent memorization and ensure fair generalization, and (2) a long-range (+/-100 m) benchmark. TopoMaskV3 achieves state-of-the-art 28.5 OLS on this geographically disjoint benchmark, surpassing all prior methods. Our analysis shows that the mask representation is more robust to geographic overfitting than Bezier, while LiDAR fusion is most beneficial at long range and exhibits larger relative gains on the overlapping original split, suggesting overlap-induced memorization effects.

Method: The authors introduce the first framework for generating natural non-verbal interactions between humans and AI in real-time from 2D body keypoints, using four lightweight architectures that run at up to 100 FPS on an NVIDIA Orin Nano. They trained on 437 human video clips and used pretraining on synthetically-generated sequences.

Result: Pretraining on synthetic sequences significantly reduces motion errors without sacrificing speed, but a measurable reality gap persists. When evaluated on keypoints from SORA and VEO, performance drops more on SORA-generated clips than VEO, suggesting temporal coherence (not image fidelity) drives real-world performance.

Conclusion: Statistically distinguishable differences persist between human and AI motion, indicating that current generative models haven’t fully captured the nuances of human non-verbal communication despite advances in motion generation.

Abstract: We study the ongoing debate regarding the statistical fidelity of AI-generated data compared to human-generated data in the context of non-verbal communication using full body motion. Concretely, we ask if contemporary generative models move beyond surface mimicry to participate in the silent, but expressive dialogue of body language. We tackle this question by introducing the first framework that generates a natural non-verbal interaction between Human and AI in real-time from 2D body keypoints. Our experiments utilize four lightweight architectures which run at up to 100 FPS on an NVIDIA Orin Nano, effectively closing the perception-action loop needed for natural Human-AI interaction. We trained on 437 human video clips and demonstrated that pretraining on synthetically-generated sequences reduces motion errors significantly, without sacrificing speed. Yet, a measurable reality gap persists. When the best model is evaluated on keypoints extracted from cutting-edge text-to-video systems, such as SORA and VEO, we observe that performance drops on SORA-generated clips. However, it degrades far less on VEO, suggesting that temporal coherence, not image fidelity, drives real-world performance. Our results demonstrate that statistically distinguishable differences persist between Human and AI motion.

Method: Created Cryo-Bench benchmark with datasets covering debris-covered glaciers, glacial lakes, sea ice, and calving fronts across multiple sensors and regions. Evaluated 14 GFMs alongside UNet and ViT baselines under frozen encoder, few-shot (10% data), and full fine-tuning settings.

Result: UNet achieved highest average mIoU (66.38) with frozen encoder. In few-shot settings, GFMs like DOFA and TerraMind outperformed UNet (59.53, 56.62 vs 56.60). Full fine-tuning showed inconsistent performance but improved 12.77% with learning rate tuning. GFMs demonstrated strong domain adaptation despite minimal cryosphere pretraining data.

Conclusion: GFMs show promise for cryospheric applications with notable domain adaptation. Recommended encoder fine-tuning with hyperparameter optimization for best performance, or frozen encoders for quick results without extensive experimentation.

Abstract: Geo-Foundation Models (GFMs) have been evaluated across diverse Earth observation task including multiple domains and have demonstrated strong potential of producing reliable maps even with sparse labels. However, benchmarking GFMs for Cryosphere applications has remained limited, primarily due to the lack of suitable evaluation datasets. To address this gap, we introduce \textbf{Cryo-Bench}, a benchmark compiled to evaluate GFM performance across key Cryospheric components. Cryo-Bench includes debris-covered glaciers, glacial lakes, sea ice, and calving fronts, spanning multiple sensors and broad geographic regions. We evaluate 14 GFMs alongside UNet and ViT baselines to assess their advantages, limitations, and optimal usage strategies. With a frozen encoder, UNet achieves the highest average mIoU of \textbf{66.38}, followed by TerraMind at \textbf{64.02} across five evluation dataset included in Cryo-Bench. In the few-shot setting (10% input data), GFMs such as DOFA and TerraMind outperform UNet, achieving mIoU scores of \textbf{59.53}, \textbf{56.62}, and \textbf{56.60}, respectively, comapred to U-Net’s 56.60. When fully finetuning GFMs, we observe inconsistent performance across datasets and models. However, tuning learning rate along with finetuning substantially improves GFM performance. For example, evaluation on two representative datasets (GLID and CaFFe) shows an average relative improvement of \textbf{12.77%}. Despite having minimal Cryosphere representation in their pretraining data, GFMs exhibit notable domain adaptation capabilities and produce meaningful results across tasks. Based on our findings, We recommend encoder fine-tuning with hyperparameter optimization optimization to achieve the best possible performance, while using frozen encoders when users need quick results without extensive experimentation.(\href{https://github.com/Sk-2103/Cryo-Bench}{GitHub}).

Method: Proposes InterCoG framework with three-step reasoning: 1) text-only object position reasoning with spatial relations, 2) visual grounding via bounding boxes and masks, 3) rewriting editing descriptions. Includes two auxiliary training modules: multimodal grounding reconstruction supervision and multimodal grounding reasoning alignment. Also introduces GroundEdit-45K dataset and GroundEdit-Bench for evaluation.

Result: Extensive experiments demonstrate superiority in highly precise edits under spatially intricate and multi-entity scenes. The framework shows strong capabilities in fine-grained editing where targets require spatial reasoning.

Conclusion: InterCoG effectively addresses the challenge of fine-grained editing in complex scenes through interleaved text-vision reasoning, achieving precise spatial localization and interpretable reasoning for multi-entity editing tasks.

Abstract: Emerging unified editing models have demonstrated strong capabilities in general object editing tasks. However, it remains a significant challenge to perform fine-grained editing in complex multi-entity scenes, particularly those where targets are not visually salient and require spatial reasoning. To this end, we propose InterCoG, a novel text-vision Interleaved Chain-of-Grounding reasoning framework for fine-grained image editing in complex real-world scenes. The key insight of InterCoG is to first perform object position reasoning solely within text that includes spatial relation details to explicitly deduce the location and identity of the edited target. It then conducts visual grounding via highlighting the editing targets with generated bounding boxes and masks in pixel space, and finally rewrites the editing description to specify the intended outcomes. To further facilitate this paradigm, we propose two auxiliary training modules: multimodal grounding reconstruction supervision and multimodal grounding reasoning alignment to enforce spatial localization accuracy and reasoning interpretability, respectively. We also construct GroundEdit-45K, a dataset comprising 45K grounding-oriented editing samples with detailed reasoning annotations, and GroundEdit-Bench for grounding-aware editing evaluation. Extensive experiments substantiate the superiority of our approach in highly precise edits under spatially intricate and multi-entity scenes.

Method: Proposes PPEDCRF with three components: 1) dynamic CRF for temporal consistency in tracking location-sensitive regions, 2) normalized control penalty for hierarchical sensitivity-based perturbation allocation, and 3) utility-preserving noise injection that minimizes interference with object detection/segmentation.

Result: Significantly reduces location-retrieval attack success (Top-k accuracy) while maintaining competitive detection performance (mAP and segmentation metrics) compared to baselines like global noise, white-noise masking, and feature-based anonymization.

Conclusion: PPEDCRF effectively balances location privacy protection with preservation of foreground detection utility in dashcam videos, addressing the tension between privacy and utility in autonomous driving data sharing.

Abstract: Dashcam videos collected by autonomous or assisted-driving systems are increasingly shared for safety auditing and model improvement. Even when explicit GPS metadata are removed, an attacker can still infer the recording location by matching background visual cues (e.g., buildings and road layouts) against large-scale street-view imagery. This paper studies location-privacy leakage under a background-based retrieval attacker, and proposes PPEDCRF, a privacy-preserving enhanced dynamic conditional random field framework that injects calibrated perturbations only into inferred location-sensitive background regions while preserving foreground detection utility. PPEDCRF consists of three components: (i) a dynamic CRF that enforces temporal consistency to discover and track location sensitive regions across frames, (ii) a normalized control penalty (NCP) that allocates perturbation strength according to a hierarchical sensitivity model, and (iii) a utility-preserving noise injection module that minimizes interference to object detection and segmentation. Experiments on public driving datasets demonstrate that PPEDCRF significantly reduces location-retrieval attack success (e.g., Top-k retrieval accuracy) while maintaining competitive detection performance (e.g., mAP and segmentation metrics) compared with common baselines such as global noise, white-noise masking, and feature-based anonymization. The source code is in https://github.com/mabo1215/PPEDCRF.git

Method: Proposes Preference Score Distillation (PSD), an optimization-based framework that: 1) Uses pretrained 2D reward models for human-aligned text-to-3D synthesis, 2) Reformulates preference alignment as a classifier-free guidance mechanism through implicit reward modeling, 3) Introduces adaptive co-optimization of preference scores and negative text embeddings, 4) Incorporates CFG during optimization for dynamic enhancement of alignment.

Result: PSD demonstrates superiority in aesthetic metrics, seamless integration with diverse pipelines, and strong extensibility. It successfully bridges human preference alignment with CFG theory under score distillation framework.

Conclusion: PSD provides an effective solution for human preference alignment in text-to-3D generation without requiring 3D training data, offering a novel connection between preference alignment and classifier-free guidance theory.

Abstract: Human preference alignment presents a critical yet underexplored challenge for diffusion models in text-to-3D generation. Existing solutions typically require task-specific fine-tuning, posing significant hurdles in data-scarce 3D domains. To address this, we propose Preference Score Distillation (PSD), an optimization-based framework that leverages pretrained 2D reward models for human-aligned text-to-3D synthesis without 3D training data. Our key insight stems from the incompatibility of pixel-level gradients: due to the absence of noisy samples during reward model training, direct application of 2D reward gradients disturbs the denoising process. Noticing that similar issue occurs in the naive classifier guidance in conditioned diffusion models, we fundamentally rethink preference alignment as a classifier-free guidance (CFG)-style mechanism through our implicit reward model. Furthermore, recognizing that frozen pretrained diffusion models constrain performance, we introduce an adaptive strategy to co-optimize preference scores and negative text embeddings. By incorporating CFG during optimization, online refinement of negative text embeddings dynamically enhances alignment. To our knowledge, we are the first to bridge human preference alignment with CFG theory under score distillation framework. Experiments demonstrate the superiority of PSD in aesthetic metrics, seamless integration with diverse pipelines, and strong extensibility.

Method: Proposes Dehallu3D with a plug-and-play optimization module using two constraints: (1) adjacent consistency for geometric continuity across views, and (2) adaptive smoothness to retain fine details. Also introduces Outlier Risk Measure (ORM) metric to quantify geometric fidelity from outlier perspective.

Result: Extensive experiments show Dehallu3D achieves high-fidelity 3D generation by effectively preserving structural details while removing hallucinated outliers.

Conclusion: Dehallu3D successfully mitigates hallucinations in 3D reconstruction models through balanced multi-view continuity constraints, enabling more reliable 3D content generation for applications like VR and 3D printing.

Abstract: Large 3D reconstruction models have revolutionized the 3D content generation field, enabling broad applications in virtual reality and gaming. Just like other large models, large 3D reconstruction models suffer from hallucinations as well, introducing structural outliers (e.g., odd holes or protrusions) that deviate from the input data. However, unlike other large models, hallucinations in large 3D reconstruction models remain severely underexplored, leading to malformed 3D-printed objects or insufficient immersion in virtual scenes. Such hallucinations majorly originate from that existing methods reconstruct 3D content from sparsely generated multi-view images which suffer from large viewpoint gaps and discontinuities. To mitigate hallucinations by eliminating the outliers, we propose Dehallu3D for 3D mesh generation. Our key idea is to design a balanced multi-view continuity constraint to enforce smooth transitions across dense intermediate viewpoints, while avoiding over-smoothing that could erase sharp geometric features. Therefore, Dehallu3D employs a plug-and-play optimization module with two key constraints: (i) adjacent consistency to ensure geometric continuity across views, and (ii) adaptive smoothness to retain fine details.We further propose the Outlier Risk Measure (ORM) metric to quantify geometric fidelity in 3D generation from the perspective of outliers. Extensive experiments show that Dehallu3D achieves high-fidelity 3D generation by effectively preserving structural details while removing hallucinated outliers.

Method: Three main components: (1) Physics-Informed Radar Encoder (PIR Encoder) with RCS Mapper and Quality Gate for transforming radar attributes into scattering priors and predicting point-wise reliability; (2) Radar-guided Interactive Fusion Module (RIFM) for query-level radar-image fusion using a dual-stream backbone with point-based local stream and transformer-based global stream with Scattering-Aware Self-Attention (SASA); (3) Temporal Query Aggregation module for aggregating frame-wise fused queries over temporal windows.

Result: Achieves 59.7% mAP50:95 and 90.3% mAP50 on WaterScenes (with T=5 radar history) using 5.6M parameters and 12.5G FLOPs, and reaches 94.8% mAP50 and 46.2% mAP50:95 on FLOW dataset under radar+camera setting. Ablation studies confirm contributions of PIR Encoder, SASA-based global reasoning, and RIFM.

Conclusion: PhysFusion effectively addresses the challenges of water-surface perception by integrating physics-informed radar encoding with multimodal fusion, demonstrating superior performance in maritime target detection through specialized handling of sparse radar data and temporal consistency.

Abstract: Detecting water-surface targets for Unmanned Surface Vehicles (USVs) is challenging due to wave clutter, specular reflections, and weak appearance cues in long-range observations. Although 4D millimeter-wave radar complements cameras under degraded illumination, maritime radar point clouds are sparse and intermittent, with reflectivity attributes exhibiting heavy-tailed variations under scattering and multipath, making conventional fusion designs struggle to exploit radar cues effectively. We propose PhysFusion, a physics-informed radar-image detection framework for water-surface perception. The framework integrates: (1) a Physics-Informed Radar Encoder (PIR Encoder) with an RCS Mapper and Quality Gate, transforming per-point radar attributes into compact scattering priors and predicting point-wise reliability for robust feature learning under clutter; (2) a Radar-guided Interactive Fusion Module (RIFM) performing query-level radar-image fusion between semantically enriched radar features and multi-scale visual features, with the radar branch modeled by a dual-stream backbone including a point-based local stream and a transformer-based global stream using Scattering-Aware Self-Attention (SASA); and (3) a Temporal Query Aggregation module (TQA) aggregating frame-wise fused queries over a short temporal window for temporally consistent representations. Experiments on WaterScenes and FLOW demonstrate that PhysFusion achieves 59.7% mAP50:95 and 90.3% mAP50 on WaterScenes (T=5 radar history) using 5.6M parameters and 12.5G FLOPs, and reaches 94.8% mAP50 and 46.2% mAP50:95 on FLOW under radar+camera setting. Ablation studies quantify the contributions of PIR Encoder, SASA-based global reasoning, and RIFM.

Method: Uses VGGT geometry foundation model for camera parameter estimation and initial 3D points, leverages VGGT attention maps for semantic entity matching, and employs Vision-Language Models to identify and preserve large static scene regions.

Result: Extensive experiments confirm effectiveness and robustness in mitigating transient distractors for sparse-view 3DGS training.

Conclusion: The proposed framework successfully addresses sparse-view limitations in distractor-free 3DGS by integrating rich prior information from foundation models.

Abstract: 3D Gaussian Splatting (3DGS) enables efficient training and fast novel view synthesis in static environments. To address challenges posed by transient objects, distractor-free 3DGS methods have emerged and shown promising results when dense image captures are available. However, their performance degrades significantly under sparse input conditions. This limitation primarily stems from the reliance on the color residual heuristics to guide the training, which becomes unreliable with limited observations. In this work, we propose a framework to enhance distractor-free 3DGS under sparse-view conditions by incorporating rich prior information. Specifically, we first adopt the geometry foundation model VGGT to estimate camera parameters and generate a dense set of initial 3D points. Then, we harness the attention maps from VGGT for efficient and accurate semantic entity matching. Additionally, we utilize Vision-Language Models (VLMs) to further identify and preserve the large static regions in the scene. We also demonstrate how these priors can be seamlessly integrated into existing distractor-free 3DGS methods. Extensive experiments confirm the effectiveness and robustness of our approach in mitigating transient distractors for sparse-view 3DGS training.

Method: 1) Uses DINO-ViT’s semantic awareness capability to deconstruct visual elements and establish semantic relationships with OSM. 2) Implements a coarse-to-fine search paradigm to replace global dense matching, enabling efficient progressive refinement.

Result: Extensive experiments show significant improvements in both localization accuracy and speed. When trained on a single dataset, the 3° orientation recall outperforms the 5° recall of state-of-the-art methods.

Conclusion: The proposed hierarchical search framework with semantic alignment effectively addresses cross-modal discrepancies and computational challenges in OSM-based localization, offering a scalable and privacy-preserving solution for monocular re-localization.

Abstract: Monocular re-localization plays a crucial role in enabling intelligent agents to achieve human-like perception. However, traditional methods rely on dense maps, which face scalability limitations and privacy risks. OpenStreetMap (OSM), as a lightweight map that protects privacy, offers semantic and geometric information with global scalability. Nonetheless, there are still challenges in using OSM for localization: the inherent cross-modal discrepancies between natural images and OSM, as well as the high computational cost of global map-based localization. In this paper, we propose a hierarchical search framework with semantic alignment for localization in OSM. First, the semantic awareness capability of DINO-ViT is utilised to deconstruct visual elements to establish semantic relationships with OSM. Second, a coarse-to-fine search paradigm is designed to replace global dense matching, enabling efficient progressive refinement. Extensive experiments demonstrate that our method significantly improves both localization accuracy and speed. When trained on a single dataset, the 3° orientation recall of our method even outperforms the 5° recall of state-of-the-art methods.

Method: Introduces clinically observed intermediate-dose scans as trajectory anchors, enforces timestep-dependent supervision to regularize the reverse process toward dose-aligned intermediate states, calibrates anchor timesteps via degradation matching between simulated diffusion corruption and real multi-dose PET pairs, and uses timestep-weighted anchor loss for stable stage-wise learning.

Result: Consistent improvements over strong CNN-, Transformer-, GAN-, and diffusion-based baselines. On internal dataset: PSNR improved from 42.48 dB to 43.71 dB (+1.23 dB), SSIM increased to 0.986, NMAE reduced from 0.115 to 0.103 (-0.012). Performance gains generalize across scanners with 34.42 dB PSNR and 0.141 NMAE on external cohort.

Conclusion: MAP-Diff effectively regularizes diffusion reverse trajectories using multi-dose anchors, enabling progressive, dose-consistent intermediate restoration from ultra-low-dose PET inputs while achieving state-of-the-art denoising performance.

Abstract: Low-dose Positron Emission Tomography (PET) reduces radiation exposure but suffers from severe noise and quantitative degradation. Diffusion-based denoising models achieve strong final reconstructions, yet their reverse trajectories are typically unconstrained and not aligned with the progressive nature of PET dose formation. We propose MAP-Diff, a multi-anchor guided diffusion framework for progressive 3D whole-body PET denoising. MAP-Diff introduces clinically observed intermediate-dose scans as trajectory anchors and enforces timestep-dependent supervision to regularize the reverse process toward dose-aligned intermediate states. Anchor timesteps are calibrated via degradation matching between simulated diffusion corruption and real multi-dose PET pairs, and a timestep-weighted anchor loss stabilizes stage-wise learning. At inference, the model requires only ultra-low-dose input while enabling progressive, dose-consistent intermediate restoration. Experiments on internal (Siemens Biograph Vision Quadra) and cross-scanner (United Imaging uEXPLORER) datasets show consistent improvements over strong CNN-, Transformer-, GAN-, and diffusion-based baselines. On the internal dataset, MAP-Diff improves PSNR from 42.48 dB to 43.71 dB (+1.23 dB), increases SSIM to 0.986, and reduces NMAE from 0.115 to 0.103 (-0.012) compared to 3D DDPM. Performance gains generalize across scanners, achieving 34.42 dB PSNR and 0.141 NMAE on the external cohort, outperforming all competing methods.

Method: Treats latent features as functions over time and approximates them with Chebyshev polynomials. Uses ridge regression to fit coefficients for each basis, then forecasts features at multiple future diffusion steps. This provides global, long-range feature reuse with controlled error.

Result: Achieves up to 4.79× speedup on FLUX.1 and 4.67× speedup on Wan2.1-14B while maintaining higher sample quality compared to baselines. The approach works with various state-of-the-art image and video diffusion models.

Conclusion: Spectrum enables significant inference speedups for diffusion models through global feature forecasting with theoretical error guarantees, outperforming local approximation methods.

Abstract: Diffusion models have become the dominant tool for high-fidelity image and video generation, yet are critically bottlenecked by their inference speed due to the numerous iterative passes of Diffusion Transformers. To reduce the exhaustive compute, recent works resort to the feature caching and reusing scheme that skips network evaluations at selected diffusion steps by using cached features in previous steps. However, their preliminary design solely relies on local approximation, causing errors to grow rapidly with large skips and leading to degraded sample quality at high speedups. In this work, we propose spectral diffusion feature forecaster (Spectrum), a training-free approach that enables global, long-range feature reuse with tightly controlled error. In particular, we view the latent features of the denoiser as functions over time and approximate them with Chebyshev polynomials. Specifically, we fit the coefficient for each basis via ridge regression, which is then leveraged to forecast features at multiple future diffusion steps. We theoretically reveal that our approach admits more favorable long-horizon behavior and yields an error bound that does not compound with the step size. Extensive experiments on various state-of-the-art image and video diffusion models consistently verify the superiority of our approach. Notably, we achieve up to 4.79$\times$ speedup on FLUX.1 and 4.67$\times$ speedup on Wan2.1-14B, while maintaining much higher sample quality compared with the baselines.

Method: 1) Proposed DriveCombo benchmark with text and vision-based compositional traffic rule reasoning. 2) Introduced a Five-Level Cognitive Ladder inspired by human drivers’ cognitive development, evaluating reasoning from single-rule understanding to multi-rule integration and conflict resolution. 3) Developed a Rule2Scene Agent that maps language-based traffic rules to dynamic driving scenes through rule crafting and scene generation for scene-level traffic rule visual reasoning.

Result: Evaluation of 14 mainstream MLLMs showed performance drops as task complexity increased, particularly during rule conflicts. After dataset splitting and fine-tuning on the training set, substantial improvements were observed in both traffic rule reasoning and downstream planning capabilities.

Conclusion: DriveCombo effectively advances compliant and intelligent autonomous driving systems by providing a comprehensive benchmark for evaluating MLLMs’ compositional reasoning abilities in complex traffic scenarios, highlighting the importance of multi-rule understanding and conflict resolution.

Abstract: Multimodal Large Language Models (MLLMs) are rapidly becoming the intelligence brain of end-to-end autonomous driving systems. A key challenge is to assess whether MLLMs can truly understand and follow complex real-world traffic rules. However, existing benchmarks mainly focus on single-rule scenarios like traffic sign recognition, neglecting the complexity of multi-rule concurrency and conflicts in real driving. Consequently, models perform well on simple tasks but often fail or violate rules in real world complex situations. To bridge this gap, we propose DriveCombo, a text and vision-based benchmark for compositional traffic rule reasoning. Inspired by human drivers’ cognitive development, we propose a systematic Five-Level Cognitive Ladder that evaluates reasoning from single-rule understanding to multi-rule integration and conflict resolution, enabling quantitative assessment across cognitive stages. We further propose a Rule2Scene Agent that maps language-based traffic rules to dynamic driving scenes through rule crafting and scene generation, enabling scene-level traffic rule visual reasoning. Evaluations of 14 mainstream MLLMs reveal performance drops as task complexity grows, particularly during rule conflicts. After splitting the dataset and fine-tuning on the training set, we further observe substantial improvements in both traffic rule reasoning and downstream planning capabilities. These results highlight the effectiveness of DriveCombo in advancing compliant and intelligent autonomous driving systems.

Method: Proposes MSP framework with three components: 1) Hairstyle-Oriented Augmentation (HSOA) generates intra-identity hairstyle diversity, 2) Cloth-Preserved Random Erasing (CPRE) performs ratio-controlled erasing within clothing regions to suppress texture bias while retaining body shape, 3) Region-based Parsing Attention (RPA) uses parsing-guided priors to highlight face/limb regions while suppressing hair features.

Result: Extensive experiments on multiple CC-ReID benchmarks demonstrate that MSP achieves state-of-the-art performance, providing robust and practical solution for long-term person re-identification.

Conclusion: MSP effectively addresses hairstyle distraction in cloth-changing person re-identification by reducing hairstyle dependence while preserving structural information, leading to improved robustness and performance.

Abstract: Cloth-Changing Person Re-Identification (CC-ReID) aims to match the same individual across cameras under varying clothing conditions. Existing approaches often remove apparel and focus on the head region to reduce clothing bias. However, treating the head holistically without distinguishing between face and hair leads to over-reliance on volatile hairstyle cues, causing performance degradation under hairstyle changes. To address this issue, we propose the Mitigating Hairstyle Distraction and Structural Preservation (MSP) framework. Specifically, MSP introduces Hairstyle-Oriented Augmentation (HSOA), which generates intra-identity hairstyle diversity to reduce hairstyle dependence and enhance attention to stable facial and body cues. To prevent the loss of structural information, we design Cloth-Preserved Random Erasing (CPRE), which performs ratio-controlled erasing within clothing regions to suppress texture bias while retaining body shape and context. Furthermore, we employ Region-based Parsing Attention (RPA) to incorporate parsing-guided priors that highlight face and limb regions while suppressing hair features. Extensive experiments on multiple CC-ReID benchmarks demonstrate that MSP achieves state-of-the-art performance, providing a robust and practical solution for long-term person re-identification.

Method: QCAgent incorporates a customized critique mechanism guided by user-defined checklists specifying required diagnostic details and constraints. It re-identifies informative regions in WSIs based on critique feedback and text-patch semantic retrieval, iteratively enriching and reconciling reports.

Result: Experiments show that by making report requirements explicitly prompt-defined, constraint-aware, and verifiable through evidence-grounded refinement, QCAgent enables controllable generation of clinically meaningful and high-coverage pathology reports from WSIs.

Conclusion: The proposed agentic framework successfully addresses limitations of current methods by enabling quality-controllable, evidence-grounded pathology report generation that mimics pathologists’ diagnostic workflow.

Abstract: Recent methods for pathology report generation from whole-slide image (WSI) are capable of producing slide-level diagnostic descriptions but fail to ground fine-grained statements in localized visual evidence. Furthermore, they lack control over which diagnostic details to include and how to verify them. Inspired by emerging agentic analysis paradigms and the diagnostic workflow of pathologists,who selectively examine multiple fields of view, we propose QCAgent, an agentic framework for quality-controllable WSI report generation. The core innovations of this framework are as follows: (i) it incorporates a customized critique mechanism guided by a user-defined checklist specifying required diagnostic details and constraints; (ii) it re-identifies informative regions in the WSI based on the critique feedback and text-patch semantic retrieval, a process that iteratively enriches and reconciles the report. Experiments demonstrate that by making report requirements explicitly prompt-defined, constraint-aware, and verifiable through evidence-grounded refinement, QCAgent enables controllable generation of clinically meaningful and high-coverage pathology reports from WSI.

Method: Proposed a detection-gated pipeline integrating YOLOv8-based detector with U-Net segmenter, with temporal consistency wrapper to suppress false positives during glottal closure and instrument occlusion. Trained on limited GIRAFE dataset (600 frames) and evaluated via zero-shot transfer on BAGLS dataset.

Result: Achieved state-of-the-art performance on GIRAFE benchmark (DSC 0.81) and superior generalizability on BAGLS (DSC 0.85) without institutional fine-tuning. Automated kinematic features remained consistent with clinical benchmarks, with glottal area coefficient of variation being significant marker for distinguishing healthy from pathological vocal function (p=0.006).

Conclusion: Detection-gated architecture provides lightweight, computationally efficient solution (~35 frames/s) for real-time clinical use, enabling robust zero-shot transfer and standardized extraction of clinical biomarkers across diverse endoscopy platforms.

Abstract: Background: Accurate glottal segmentation in high-speed videoendoscopy (HSV) is essential for extracting kinematic biomarkers of laryngeal function. However, existing deep learning models often produce spurious artifacts in non-glottal frames and fail to generalize across different clinical settings. Methods: We propose a detection-gated pipeline that integrates a YOLOv8-based detector with a U-Net segmenter. A temporal consistency wrapper ensures robustness by suppressing false positives during glottal closure and instrument occlusion. The model was trained on a limited subset of the GIRAFE dataset (600 frames) and evaluated via zero-shot transfer on the large-scale BAGLS dataset. Results: The pipeline achieved state-of-the-art performance on the GIRAFE benchmark (DSC 0.81) and demonstrated superior generalizability on BAGLS (DSC 0.85, in-distribution) without institutional fine-tuning. Downstream validation on a 65-subject clinical cohort confirmed that automated kinematic features (Open Quotient, coefficient of variation) remained consistent with established clinical benchmarks. The coefficient of variation (CV) of the glottal area was found to be a significant marker for distinguishing healthy from pathological vocal function (p=0.006). Conclusions: The detection-gated architecture provides a lightweight, computationally efficient solution (~35 frames/s) for real-time clinical use. By enabling robust zero-shot transfer, this framework facilitates the standardized, large-scale extraction of clinical biomarkers across diverse endoscopy platforms. Code, trained weights, and evaluation scripts are released at https://github.com/hari-krishnan/openglottal.

Method: Proposes Prompt Recurrent Unit (PRU), an iterative refinement module based on monocular depth foundation model decoders. Integrates monocular structure and stereo motion cues as prompts to enrich latent representations with stereo-scale information while preserving monocular depth priors.

Result: PromptStereo achieves state-of-the-art zero-shot generalization performance across multiple datasets while maintaining comparable or faster inference speed compared to existing methods.

Conclusion: Prompt-guided iterative refinement is a promising direction for zero-shot stereo matching, showing that effectively integrating monocular depth priors with stereo information during refinement significantly improves generalization.

Abstract: Modern stereo matching methods have leveraged monocular depth foundation models to achieve superior zero-shot generalization performance. However, most existing methods primarily focus on extracting robust features for cost volume construction or disparity initialization. At the same time, the iterative refinement stage, which is also crucial for zero-shot generalization, remains underexplored. Some methods treat monocular depth priors as guidance for iteration, but conventional GRU-based architectures struggle to exploit them due to the limited representation capacity. In this paper, we propose Prompt Recurrent Unit (PRU), a novel iterative refinement module based on the decoder of monocular depth foundation models. By integrating monocular structure and stereo motion cues as prompts into the decoder, PRU enriches the latent representations of monocular depth foundation models with absolute stereo-scale information while preserving their inherent monocular depth priors. Experiments demonstrate that our PromptStereo achieves state-of-the-art zero-shot generalization performance across multiple datasets, while maintaining comparable or faster inference speed. Our findings highlight prompt-guided iterative refinement as a promising direction for zero-shot stereo matching.

Method: Two-stage hierarchical compression: 1) Temporal Adjacency Selection (TAS) removes redundant tokens across adjacent frames, 2) Spatial Domain Consolidation (SDC) merges spatially repetitive regions within frames. Both use self-adaptive token compression based on scene statistics.

Result: State-of-the-art on streaming benchmarks: 76.4 on StreamingBench, 67.2 on OVO-Bench with real-time settings. Reduces latency by 69.9% and GPU memory by 34.5% on OVO-Bench. Maintains strong offline performance (73.1 on MLVU) with 65% fewer tokens.

Conclusion: FluxMem provides an efficient, training-free solution for streaming video understanding through adaptive token compression, achieving significant performance gains while reducing computational and memory overhead.

Abstract: This paper presents FluxMem, a training-free framework for efficient streaming video understanding. FluxMem adaptively compresses redundant visual memory through a hierarchical, two-stage design: (1) a Temporal Adjacency Selection (TAS) module removes redundant visual tokens across adjacent frames, and (2) a Spatial Domain Consolidation (SDC) module further merges spatially repetitive regions within each frame into compact representations. To adapt effectively to dynamic scenes, we introduce a self-adaptive token compression mechanism in both TAS and SDC, which automatically determines the compression rate based on intrinsic scene statistics rather than manual tuning. Extensive experiments demonstrate that FluxMem achieves new state-of-the-art results on existing online video benchmarks, reaching 76.4 on StreamingBench and 67.2 on OVO-Bench under real-time settings, while reducing latency by 69.9% and peak GPU memory by 34.5% on OVO-Bench. Furthermore, it maintains strong offline performance, achieving 73.1 on MLVU while using 65% fewer visual tokens.

Method: Proposes diffusion-based framework with LoRA adapters: 1) Base diffusion model trained with nodule mask conditioning for size/shape control; 2) Separate LoRA modules for specific radiological features; 3) Dynamic LoRA composability with novel orthogonality loss to address overlapping attention regions and non-orthogonal parameter spaces.

Result: Extensive experiments on in-house and public datasets show improved downstream nodule detection. Radiologist evaluations confirm fine-grained controllability. Quantitative metrics surpass existing nodule generation approaches for CXRs.

Conclusion: The proposed framework enables characteristic-controlled synthetic nodule generation, addressing data scarcity in medical imaging and improving CAD system performance through realistic, controllable synthetic data.

Abstract: Early detection of lung cancer in chest radiographs (CXRs) is crucial for improving patient outcomes, yet nodule detection remains challenging due to their subtle appearance and variability in radiological characteristics like size, texture, and boundary. For robust analysis, this diversity must be well represented in training datasets for deep learning based Computer-Assisted Diagnosis (CAD) systems. However, assembling such datasets is costly and often impractical, motivating the need for realistic synthetic data generation. Existing methods lack fine-grained control over synthetic nodule generation, limiting their utility in addressing data scarcity. This paper proposes a novel diffusion-based framework with low-rank adaptation (LoRA) adapters for characteristic controlled nodule synthesis on CXRs. We begin by addressing size and shape control through nodule mask conditioned training of the base diffusion model. To achieve individual characteristic control, we train separate LoRA modules, each dedicated to a specific radiological feature. However, since nodules rarely exhibit isolated characteristics, effective multi-characteristic control requires a balanced integration of features. We address this by leveraging the dynamic composability of LoRAs and revisiting existing merging strategies. Building on this, we identify two key issues, overlapping attention regions and non-orthogonal parameter spaces. To overcome these limitations, we introduce a novel orthogonality loss term during LoRA composition training. Extensive experiments on both in-house and public datasets demonstrate improved downstream nodule detection. Radiologist evaluations confirm the fine-grained controllability of our generated nodules, and across multiple quantitative metrics, our method surpasses existing nodule generation approaches for CXRs.

Method: Proposes FastLightGen algorithm that transforms large models into fast lightweight counterparts by constructing an optimal teacher model designed to maximize student performance within a synergistic framework for distilling both model size and inference steps.

Result: Experiments on HunyuanVideo-ATI2V and WanX-TI2V show optimal visual quality achieved with 4-step sampling and 30% parameter pruning under constrained inference budget. FastLightGen consistently outperforms all competing methods, establishing new SOTA in efficient video generation.

Conclusion: FastLightGen successfully addresses computational bottlenecks in video generation by simultaneous model compression and sampling step reduction, enabling practical deployment of powerful video generation models.

Abstract: The recent advent of powerful video generation models, such as Hunyuan, WanX, Veo3, and Kling, has inaugurated a new era in the field. However, the practical deployment of these models is severely impeded by their substantial computational overhead, which stems from enormous parameter counts and the iterative, multi-step sampling process required during inference. Prior research on accelerating generative models has predominantly followed two distinct trajectories: reducing the number of sampling steps (e.g., LCM, DMD, and MagicDistillation) or compressing the model size for more efficient inference (e.g., ICMD). The potential of simultaneously compressing both to create a fast and lightweight model remains an unexplored avenue. In this paper, we propose FastLightGen, an algorithm that transforms large, computationally expensive models into fast, lightweight counterparts. The core idea is to construct an optimal teacher model, one engineered to maximize student performance, within a synergistic framework for distilling both model size and inference steps. Our extensive experiments on HunyuanVideo-ATI2V and WanX-TI2V reveal that a generator using 4-step sampling and 30% parameter pruning achieves optimal visual quality under a constrained inference budget. Furthermore, FastLightGen consistently outperforms all competing methods, establishing a new state-of-the-art in efficient video generation.

Method: Uses a fine-grained, expression-controllable large-scale video diffusion Transformer that lifts incomplete input into richer kinematic representation. Features: (1) multi-granularity expression control combining shading maps with expression coefficients, (2) multi-reference conditioning mechanism aggregating cues from multiple frames for 3D consistency.

Result: Extensive experiments show LiftAvatar consistently boosts animation quality and quantitative metrics of state-of-the-art 3D avatar methods, especially under extreme, unseen expressions. Enables effective prior distillation from large-scale video generative models into 3D pipelines.

Conclusion: LiftAvatar provides a plug-and-play enhancer that addresses sparse kinematic data limitations in monocular videos, improving 3D avatar reconstruction and animation through kinematic space completion and video diffusion-based synthesis.

Abstract: We present LiftAvatar, a new paradigm that completes sparse monocular observations in kinematic space (e.g., facial expressions and head pose) and uses the completed signals to drive high-fidelity avatar animation. LiftAvatar is a fine-grained, expression-controllable large-scale video diffusion Transformer that synthesizes high-quality, temporally coherent expression sequences conditioned on single or multiple reference images. The key idea is to lift incomplete input data into a richer kinematic representation, thereby strengthening both reconstruction and animation in downstream 3D avatar pipelines. To this end, we introduce (i) a multi-granularity expression control scheme that combines shading maps with expression coefficients for precise and stable driving, and (ii) a multi-reference conditioning mechanism that aggregates complementary cues from multiple frames, enabling strong 3D consistency and controllability. As a plug-and-play enhancer, LiftAvatar directly addresses the limited expressiveness and reconstruction artifacts of 3D Gaussian Splatting-based avatars caused by sparse kinematic cues in everyday monocular videos. By expanding incomplete observations into diverse pose-expression variations, LiftAvatar also enables effective prior distillation from large-scale video generative models into 3D pipelines, leading to substantial gains. Extensive experiments show that LiftAvatar consistently boosts animation quality and quantitative metrics of state-of-the-art 3D avatar methods, especially under extreme, unseen expressions.

Method: Two-stage training process: 1) Investigate DDPM-LQ and GAN-based MUNIT-LQ approaches to generate low-quality CXRs as a style transfer problem, 2) Train a DDPM-based model on paired low-quality and high-quality images to learn X-ray image restoration.

Result: Promising results in enhancing image clarity, contrast, and overall diagnostic value of chest X-rays while preserving clinically significant artifacts, validated by quantitative metrics and expert radiological assessment.

Conclusion: DiffusionXRay addresses the data scarcity problem in medical imaging by improving synthetic X-ray quality through a novel diffusion-based restoration pipeline, potentially enhancing automated lung cancer diagnosis.

Abstract: Deep learning-based automated diagnosis of lung cancer has emerged as a crucial advancement that enables healthcare professionals to detect and initiate treatment earlier. However, these models require extensive training datasets with diverse case-specific properties. High-quality annotated data is particularly challenging to obtain, especially for cases with subtle pulmonary nodules that are difficult to detect even for experienced radiologists. This scarcity of well-labeled datasets can limit model performance and generalization across different patient populations. Digitally reconstructed radiographs (DRR) using CT-Scan to generate synthetic frontal chest X-rays with artificially inserted lung nodules offers one potential solution. However, this approach suffers from significant image quality degradation, particularly in the form of blurred anatomical features and loss of fine lung field structures. To overcome this, we introduce DiffusionXRay, a novel image restoration pipeline for Chest X-ray images that synergistically leverages denoising diffusion probabilistic models (DDPMs) and generative adversarial networks (GANs). DiffusionXRay incorporates a unique two-stage training process: First, we investigate two independent approaches, DDPM-LQ and GAN-based MUNIT-LQ, to generate low-quality CXRs, addressing the challenge of training data scarcity, posing this as a style transfer problem. Subsequently, we train a DDPM-based model on paired low-quality and high-quality images, enabling it to learn the nuances of X-ray image restoration. Our method demonstrates promising results in enhancing image clarity, contrast, and overall diagnostic value of chest X-rays while preserving subtle yet clinically significant artifacts, validated by both quantitative metrics and expert radiological assessment.

Method: Uses teacher-student paradigm: Quality-Aware Teacher performs voxel-level early fusion with Quality of Interest weighting and semantic guidance, produces clean supervision via diffusion denoiser. Dual-Branch Diffusion Student separates ego/cooperative streams for encoding, uses Ego-Guided Cross-Attention for balanced decoding under degradation.

Result: Outperforms prior methods across all degradation types on OPV2Vn and DAIR-V2Xn benchmarks with six corruption types, lowers relative corruption error, offers tunable balance between precision and inference efficiency.

Conclusion: CoopDiff effectively addresses corruption challenges in cooperative perception through diffusion-based denoising, demonstrating superior robustness and generalization across diverse degradation scenarios.

Abstract: Cooperative perception lets agents share information to expand coverage and improve scene understanding. However, in real-world scenarios, diverse and unpredictable corruptions undermine its robustness and generalization. To address these challenges, we introduce CoopDiff, a diffusion-based cooperative perception framework that mitigates corruptions via a denoising mechanism. CoopDiff adopts a teacher-student paradigm: the Quality-Aware Teacher performs voxel-level early fusion with Quality of Interest weighting and semantic guidance, then produces clean supervision features via a diffusion denoiser. The Dual-Branch Diffusion Student first separates ego and cooperative streams in encoding to reconstruct the teacher’s clean targets. And then, an Ego-Guided Cross-Attention mechanism facilitates balanced decoding under degradation by adaptively integrating ego and cooperative features. We evaluate CoopDiff on two constructed multi-degradation benchmarks, OPV2Vn and DAIR-V2Xn, each incorporating six corruption types, including environmental and sensor-level distortions. Benefiting from the inherent denoising properties of diffusion, CoopDiff consistently outperforms prior methods across all degradation types and lowers the relative corruption error. Furthermore, it offers a tunable balance between precision and inference efficiency.

Method: 1) Introduces a scalable data generation pipeline that transforms existing video editing pairs into high-fidelity training quadruplets using image generative models to create synthesized reference scaffolds. 2) Constructs RefVIE dataset for instruction-reference-following tasks and RefVIE-Bench for evaluation. 3) Proposes Kiwi-Edit architecture that synergizes learnable queries and latent visual features for reference semantic guidance with progressive multi-stage training.

Result: The model achieves significant gains in instruction following and reference fidelity. Extensive experiments demonstrate that the data and architecture establish a new state-of-the-art in controllable video editing.

Conclusion: The proposed scalable data generation pipeline, RefVIE dataset, and Kiwi-Edit architecture collectively advance the field of controllable video editing by addressing the data scarcity problem and improving reference-guided editing capabilities.

Abstract: Instruction-based video editing has witnessed rapid progress, yet current methods often struggle with precise visual control, as natural language is inherently limited in describing complex visual nuances. Although reference-guided editing offers a robust solution, its potential is currently bottlenecked by the scarcity of high-quality paired training data. To bridge this gap, we introduce a scalable data generation pipeline that transforms existing video editing pairs into high-fidelity training quadruplets, leveraging image generative models to create synthesized reference scaffolds. Using this pipeline, we construct RefVIE, a large-scale dataset tailored for instruction-reference-following tasks, and establish RefVIE-Bench for comprehensive evaluation. Furthermore, we propose a unified editing architecture, Kiwi-Edit, that synergizes learnable queries and latent visual features for reference semantic guidance. Our model achieves significant gains in instruction following and reference fidelity via a progressive multi-stage training curriculum. Extensive experiments demonstrate that our data and architecture establish a new state-of-the-art in controllable video editing. All datasets, models, and code is released at https://github.com/showlab/Kiwi-Edit.

Method: Redesign Siamese neck with MLP-based fusion module for pixel-level interaction with minimal overhead. Use hierarchical search space of MLP modules with customized relaxation strategy for differentiable NAS to decouple channel-width optimization from architectural choices.

Result: Achieves state-of-the-art accuracy-efficiency trade-offs, top performance on 4 general-purpose and 3 aerial tracking benchmarks, with real-time performance on constrained GPUs and NPUs.

Conclusion: The MLP-based fusion with targeted NAS optimization effectively addresses the accuracy-efficiency imbalance in visual tracking, enabling efficient pixel-level interactions on resource-constrained hardware.

Abstract: Siamese visual trackers have recently advanced through increasingly sophisticated fusion mechanisms built on convolutional or Transformer architectures. However, both struggle to deliver pixel-level interactions efficiently on resource-constrained hardware, leading to a persistent accuracy-efficiency imbalance. Motivated by this limitation, we redesign the Siamese neck with a simple yet effective Multilayer Perception (MLP)-based fusion module that enables pixel-level interaction with minimal structural overhead. Nevertheless, naively stacking MLP blocks introduces a new challenge: computational cost can scale quadratically with channel width. To overcome this, we construct a hierarchical search space of carefully designed MLP modules and introduce a customized relaxation strategy that enables differentiable neural architecture search (DNAS) to decouple channel-width optimization from other architectural choices. This targeted decoupling automatically balances channel width and depth, yielding a low-complexity architecture. The resulting tracker achieves state-of-the-art accuracy-efficiency trade-offs. It ranks among the top performers on four general-purpose and three aerial tracking benchmarks, while maintaining real-time performance on both resource-constrained Graphics Processing Units (GPUs) and Neural Processing Units (NPUs).

Method: Integrates hybrid CoAtNet architecture (combining convolution and self-attention) with model soups - a lightweight weight-space ensembling technique that averages checkpoints from a single training trajectory without increasing inference cost. Uses greedy and uniform soup strategies to combine diverse checkpoints.

Result: Achieves state-of-the-art results on ICH-17 dataset (7,406 images across 17 classes): 72.36% top-1 accuracy and 69.28% macro F1-score, outperforming ResNet-50, DenseNet-121, and ViT baselines. Model soups reduce variance while introducing minimal bias.

Conclusion: Diversity-aware checkpoint averaging provides a principled and efficient way to reduce variance and enhance generalization in culturally rich, data-scarce classification tasks. Model soups select geometrically diverse checkpoints unlike traditional ensemble methods.

Abstract: The classification of Intangible Cultural Heritage (ICH) images in the Mekong Delta poses unique challenges due to limited annotated data, high visual similarity among classes, and domain heterogeneity. In such low-resource settings, conventional deep learning models often suffer from high variance or overfit to spurious correlations, leading to poor generalization. To address these limitations, we propose a robust framework that integrates the hybrid CoAtNet architecture with model soups, a lightweight weight-space ensembling technique that averages checkpoints from a single training trajectory without increasing inference cost. CoAtNet captures both local and global patterns through stage-wise fusion of convolution and self-attention. We apply two ensembling strategies - greedy and uniform soup - to selectively combine diverse checkpoints into a final model. Beyond performance improvements, we analyze the ensembling effect through the lens of bias-variance decomposition. Our findings show that model soups reduces variance by stabilizing predictions across diverse model snapshots, while introducing minimal additional bias. Furthermore, using cross-entropy-based distance metrics and Multidimensional Scaling (MDS), we show that model soups selects geometrically diverse checkpoints, unlike Soft Voting, which blends redundant models centered in output space. Evaluated on the ICH-17 dataset (7,406 images across 17 classes), our approach achieves state-of-the-art results with 72.36% top-1 accuracy and 69.28% macro F1-score, outperforming strong baselines including ResNet-50, DenseNet-121, and ViT. These results underscore that diversity-aware checkpoint averaging provides a principled and efficient way to reduce variance and enhance generalization in culturally rich, data-scarce classification tasks.

Method: Proposes a receiver-centric paradigm where senders generate lightweight saliency metadata, receivers formulate global request plans to dynamically budget feature contributions across agents, and a collaborative feature routing module aligns messages before fusion for structural consistency.

Result: Achieves state-of-the-art performance, improving AP@0.7 on OPV2V by 2.4% with only 0.5% of communication cost. As plug-and-play component, boosts strong baselines with 5% of full bandwidth while maintaining robustness under localization noise.

Conclusion: Globally-coordinated allocation across what and where to share is key to efficient collaborative perception. The receiver-centric approach with dynamic feature budgeting enables superior performance with minimal bandwidth.

Abstract: Collaborative perception is vital for autonomous driving yet remains constrained by tight communication budgets. Earlier work reduced bandwidth by compressing full feature maps with fixed-rate encoders, which adapts poorly to a changing environment, and it further evolved into spatial selection methods that improve efficiency by focusing on salient regions, but this object-centric approach often sacrifices global context, weakening holistic scene understanding. To overcome these limitations, we introduce \textit{WhisperNet}, a bandwidth-aware framework that proposes a novel, receiver-centric paradigm for global coordination across agents. Senders generate lightweight saliency metadata, while the receiver formulates a global request plan that dynamically budgets feature contributions across agents and features, retrieving only the most informative features. A collaborative feature routing module then aligns related messages before fusion to ensure structural consistency. Extensive experiments show that WhisperNet achieves state-of-the-art performance, improving AP@0.7 on OPV2V by 2.4% with only 0.5% of the communication cost. As a plug-and-play component, it boosts strong baselines with merely 5% of full bandwidth while maintaining robustness under localization noise. These results demonstrate that globally-coordinated allocation across \textit{what} and \textit{where} to share is the key to achieving efficient collaborative perception.

Method: 1) Learn sketch-driven diffusion prior, 2) Distill into efficient rectified-flow student in latent space, 3) Apply differentiable energies over keyframes, trajectories, and physics-based constraints to shape transport field, 4) Augment with continuous-time Markov chain (CTMC) planner for scheduling discrete events like touches, grasps, and handoffs.

Result: Achieves state-of-the-art constraint adherence and perceptual quality on CORE4D and InterHuman datasets while offering significantly faster inference than diffusion-only baselines.

Conclusion: Sketch2Colab enables precise, physically plausible 3D multi-human motion generation from 2D sketches with fine-grained control, addressing limitations of conventional diffusion approaches through efficient rectified-flow distillation and CTMC-based event planning.

Abstract: We present Sketch2Colab, which turns storyboard-style 2D sketches into coherent, object-aware 3D multi-human motion with fine-grained control over agents, joints, timing, and contacts. Conventional diffusion-based motion generators have advanced realism; however, achieving precise adherence to rich interaction constraints typically demands extensive training and/or costly posterior guidance, and performance can degrade under strong multi-entity conditioning. Sketch2Colab instead first learns a sketch-driven diffusion prior and then distills it into an efficient rectified-flow student operating in latent space for fast, stable sampling. Differentiable energies over keyframes, trajectories, and physics-based constraints directly shape the student’s transport field, steering samples toward motions that faithfully satisfy the storyboard while remaining physically plausible. To capture coordinated interaction, we augment the continuous flow with a continuous-time Markov chain (CTMC) planner that schedules discrete events such as touches, grasps, and handoffs, modulating the dynamics to produce crisp, well-phased human-object-human collaborations. Experiments on CORE4D and InterHuman show that Sketch2Colab achieves state-of-the-art constraint adherence and perceptual quality while offering significantly faster inference than diffusion-only baselines.

Method: Uses a pre-trained encoder as teacher network to extract multi-scale features from support and query images, with a student decoder that distills knowledge from teacher on query images and self-distills on support images. Includes a learn-to-weight mechanism that dynamically assesses reference value of each support image conditioned on the query.

Result: Outperforms existing approaches on a comprehensive benchmark dataset of 13,084 images across four organs, four imaging modalities, and five disease categories, establishing new state-of-the-art in few-shot medical anomaly detection.

Conclusion: D²4FAD provides an effective solution for few-shot anomaly detection in medical imaging, enabling pathology identification with minimal normal reference data and better generalization across anatomical contexts.

Abstract: Anomaly detection is a critical task in computer vision with profound implications for medical imaging, where identifying pathologies early can directly impact patient outcomes. While recent unsupervised anomaly detection approaches show promise, they require substantial normal training data and struggle to generalize across anatomical contexts. We introduce D$^2$4FAD, a novel dual distillation framework for few-shot anomaly detection that identifies anomalies in previously unseen tasks using only a small number of normal reference images. Our approach leverages a pre-trained encoder as a teacher network to extract multi-scale features from both support and query images, while a student decoder learns to distill knowledge from the teacher on query images and self-distill on support images. We further propose a learn-to-weight mechanism that dynamically assesses the reference value of each support image conditioned on the query, optimizing anomaly detection performance. To evaluate our method, we curate a comprehensive benchmark dataset comprising 13,084 images across four organs, four imaging modalities, and five disease categories. Extensive experiments demonstrate that D$^2$4FAD significantly outperforms existing approaches, establishing a new state-of-the-art in few-shot medical anomaly detection. Code is available at https://github.com/ttttqz/D24FAD.

Method: Proposes Adaptive Confidence Regularization (ACR) with two components: 1) Adaptive Confidence Loss that penalizes when multimodal confidence is lower than unimodal branches (confidence degradation), 2) Multimodal Feature Swapping for outlier synthesis to generate challenging failure-aware training examples.

Result: Extensive experiments across four datasets, three modalities, and multiple evaluation settings show ACR achieves consistent and robust gains in failure detection performance.

Conclusion: ACR provides an effective framework for multimodal failure detection that improves reliability by addressing confidence degradation and using synthetic failure training.

Abstract: The deployment of multimodal models in high-stakes domains, such as self-driving vehicles and medical diagnostics, demands not only strong predictive performance but also reliable mechanisms for detecting failures. In this work, we address the largely unexplored problem of failure detection in multimodal contexts. We propose Adaptive Confidence Regularization (ACR), a novel framework specifically designed to detect multimodal failures. Our approach is driven by a key observation: in most failure cases, the confidence of the multimodal prediction is significantly lower than that of at least one unimodal branch, a phenomenon we term confidence degradation. To mitigate this, we introduce an Adaptive Confidence Loss that penalizes such degradations during training. In addition, we propose Multimodal Feature Swapping, a novel outlier synthesis technique that generates challenging, failure-aware training examples. By training with these synthetic failures, ACR learns to more effectively recognize and reject uncertain predictions, thereby improving overall reliability. Extensive experiments across four datasets, three modalities, and multiple evaluation settings demonstrate that ACR achieves consistent and robust gains. The source code will be available at https://github.com/mona4399/ACR.

Method: The framework includes: 1) Cross-modal Latent Alignment module for rigid registration, 2) Uncertainty-enhanced Overlap Landmark Detector with similarity matching for robust 2D-3D correspondence estimation, and 3) For non-rigid registration, a shape-constrained supervision strategy with reprojection consistency, local-isometric regularization, and rendered-mask alignment.

Result: Experimental results on the P2ILF dataset demonstrate superiority in both rigid pose estimation and non-rigid deformation compared to existing methods.

Conclusion: Land-Reg provides an interpretable, correspondence-driven approach for deformable registration in laparoscopic liver surgery, addressing limitations of existing methods through explicit landmark correspondence learning and robust geometric constraints.

Abstract: In laparoscopic liver surgery, augmented reality technology enhances intraoperative anatomical guidance by overlaying 3D liver models from preoperative CT/MRI onto laparoscopic 2D views. However, existing registration methods lack explicit modeling of reliable 2D-3D geometric correspondences supported by latent evidence, leading to limited interpretability and potentially unstable alignment in clinical scenarios. In this work, we introduce Land-Reg, a correspondence-driven deformable registration framework that explicitly learns latent-grounded 2D-3D landmark correspondences as an interpretable intermediate representation to bridge cross-modal alignment. For rigid registration, Land-Reg embraces a Cross-modal Latent Alignment module to map multi-modal features into a unified latent space. Further, an Uncertainty-enhanced Overlap Landmark Detector with similarity matching is proposed to robustly estimate explicit 2D-3D landmark correspondences. For non-rigid registration, we design a novel shape-constrained supervision strategy that anchors shape deformation to matched landmarks through reprojection consistency and incorporates local-isometric regularization to alleviate inherent 2D-3D depth ambiguity, while a rendered-mask alignment enforces global shape consistency. Experimental results on the P2ILF dataset demonstrate the superiority of our method on both rigid pose estimation and non-rigid deformation. Our code will be available at https://github.com/cuiruize/Land-Reg.

Method: Proposes Domain-Aware Task Prompt Representation Learning with two prompt pools: task prompts (implicitly encode task knowledge) and domain prompts (distill domain priors from multimodal LLMs). Uses Prompt Composition Mechanism to adaptively select and compose prompts into instance-level representations, then fuses them for restoration guidance.

Result: Extensive experiments show DATPRL-IR significantly outperforms existing SOTA image restoration methods and exhibits strong generalization capabilities.

Conclusion: The proposed multi-domain all-in-one image restoration approach successfully leverages both task-specific and shared knowledge across domains through domain-aware prompt representation learning.

Abstract: Recently, significant breakthroughs have been made in all-in-one image restoration (AiOIR), which can handle multiple restoration tasks with a single model. However, existing methods typically focus on a specific image domain, such as natural scene, medical imaging, or remote sensing. In this work, we aim to extend AiOIR to multiple domains and propose the first multi-domain all-in-one image restoration method, DATPRL-IR, based on our proposed Domain-Aware Task Prompt Representation Learning. Specifically, we first construct a task prompt pool containing multiple task prompts, in which task-related knowledge is implicitly encoded. For each input image, the model adaptively selects the most relevant task prompts and composes them into an instance-level task representation via a prompt composition mechanism (PCM). Furthermore, to endow the model with domain awareness, we introduce another domain prompt pool and distill domain priors from multimodal large language models into the domain prompts. PCM is utilized to combine the adaptively selected domain prompts into a domain representation for each input image. Finally, the two representations are fused to form a domain-aware task prompt representation which can make full use of both specific and shared knowledge across tasks and domains to guide the subsequent restoration process. Extensive experiments demonstrate that our DATPRL-IR significantly outperforms existing SOTA image restoration methods, while exhibiting strong generalization capabilities. Code is available at https://github.com/GuangluDong0728/DATPRL-IR.

Method: Introduces Action-Guided Attention (AGA) that explicitly leverages predicted action sequences as queries and keys to guide sequence modeling. Uses a gating function to combine attention-weighted past information with current frame embeddings, enabling post-training analysis of learned knowledge.

Result: Demonstrates strong generalization from validation to unseen test sets on EPIC-Kitchens-100 benchmark. Post-training analysis reveals captured action dependencies and counterfactual evidence, providing transparent insights into anticipative predictions.

Conclusion: AGA improves action anticipation by focusing on relevant past moments based on upcoming activities, offering better generalization and interpretable insights through post-training analysis of learned knowledge.

Abstract: Anticipating future actions in videos is challenging, as the observed frames provide only evidence of past activities, requiring the inference of latent intentions to predict upcoming actions. Existing transformer-based approaches, which rely on dot-product attention over pixel representations, often lack the high-level semantics necessary to model video sequences for effective action anticipation. As a result, these methods tend to overfit to explicit visual cues present in the past frames, limiting their ability to capture underlying intentions and degrading generalization to unseen samples. To address this, we propose Action-Guided Attention (AGA), an attention mechanism that explicitly leverages predicted action sequences as queries and keys to guide sequence modeling. Our approach fosters the attention module to emphasize relevant moments from the past based on the upcoming activity and combine this information with the current frame embedding via a dedicated gating function. The design of AGA enables post-training analysis of the knowledge discovered from the training set. Experiments on the widely adopted EPIC-Kitchens-100 benchmark demonstrate that AGA generalizes well from validation to unseen test sets. Post-training analysis can further examine the action dependencies captured by the model and the counterfactual evidence it has internalized, offering transparent and interpretable insights into its anticipative predictions.

Method: Unified framework that maps image features to probabilistic clinical concepts, composes differentiable logic-based reasoning chains, decodes into templated clauses, and refines via retrieval and constrained language-model editing with active sampling guided by rule-level uncertainty.

Result: Experiments on standard benchmarks show consistent improvements in factual consistency and standard language metrics compared to representative baselines.

Conclusion: NeuroSymb-MRG effectively addresses key limitations in radiology report generation through neurosymbolic reasoning and active uncertainty minimization, producing more clinically grounded and factually consistent reports.

Abstract: Automatic generation of radiology reports seeks to reduce clinician workload while improving documentation consistency. Existing methods that adopt encoder-decoder or retrieval-augmented pipelines achieve progress in fluency but remain vulnerable to visual-linguistic biases, factual inconsistency, and lack of explicit multi-hop clinical reasoning. We present NeuroSymb-MRG, a unified framework that integrates NeuroSymbolic abductive reasoning with active uncertainty minimization to produce structured, clinically grounded reports. The system maps image features to probabilistic clinical concepts, composes differentiable logic-based reasoning chains, decodes those chains into templated clauses, and refines the textual output via retrieval and constrained language-model editing. An active sampling loop driven by rule-level uncertainty and diversity guides clinician-in-the-loop adjudication and promptbook refinement. Experiments on standard benchmarks demonstrate consistent improvements in factual consistency and standard language metrics compared to representative baselines.

Method: Proposes StepVAR with dual-criterion token pruning: 1) lightweight high-pass filter to capture local texture details, 2) Principal Component Analysis (PCA) to preserve global structural information. Uses nearest neighbor feature propagation to reconstruct dense feature maps from pruned representations for valid next-scale prediction.

Result: Extensive experiments on state-of-the-art text-to-image and text-to-video VAR models show StepVAR achieves substantial inference speedups while maintaining generation quality. Outperforms existing acceleration approaches in both quantitative and qualitative evaluations.

Conclusion: StepVAR effectively accelerates VAR inference by jointly considering structural and textural importance, demonstrating general applicability across diverse VAR architectures while maintaining generation quality.

Abstract: Visual AutoRegressive (VAR) models based on next-scale prediction enable efficient hierarchical generation, yet the inference cost grows quadratically at high resolutions. We observe that the computationally intensive later scales predominantly refine high-frequency textures and exhibit substantial spatial redundancy, in contrast to earlier scales that determine the global structural layout. Existing pruning methods primarily focus on high-frequency detection for token selection, often overlooking structural coherence and consequently degrading global semantics. To address this limitation, we propose StepVAR, a training-free token pruning framework that accelerates VAR inference by jointly considering structural and textural importance. Specifically, we employ a lightweight high-pass filter to capture local texture details, while leveraging Principal Component Analysis (PCA) to preserve global structural information. This dual-criterion design enables the model to retain tokens critical for both fine-grained fidelity and overall composition. To maintain valid next-scale prediction under sparse tokens, we further introduce a nearest neighbor feature propagation strategy to reconstruct dense feature maps from pruned representations. Extensive experiments on state-of-the-art text-to-image and text-to-video VAR models demonstrate that StepVAR achieves substantial inference speedups while maintaining generation quality. Quantitative and qualitative evaluations consistently show that our method outperforms existing acceleration approaches, validating its effectiveness and general applicability across diverse VAR architectures.

Method: Proposes BabelRS with two key components: 1) Concept-Shared Instruction Aligning (CSIA) aligns each sensor modality to shared linguistic concepts using language as semantic pivot, 2) Layerwise Visual-Semantic Annealing (LVSA) progressively aggregates multi-scale visual features to provide fine-grained semantic guidance.

Result: Extensive experiments show BabelRS stabilizes training and consistently outperforms state-of-the-art methods without additional enhancements.

Conclusion: BabelRS provides an effective unified framework for heterogeneous multi-modal remote sensing object detection by decoupling modality alignment from downstream tasks through language-pivoted pretraining.

Abstract: Heterogeneous multi-modal remote sensing object detection aims to accurately detect objects from diverse sensors (e.g., RGB, SAR, Infrared). Existing approaches largely adopt a late alignment paradigm, in which modality alignment and task-specific optimization are entangled during downstream fine-tuning. This tight coupling complicates optimization and often results in unstable training and suboptimal generalization. To address these limitations, we propose BabelRS, a unified language-pivoted pretraining framework that explicitly decouples modality alignment from downstream task learning. BabelRS comprises two key components: Concept-Shared Instruction Aligning (CSIA) and Layerwise Visual-Semantic Annealing (LVSA). CSIA aligns each sensor modality to a shared set of linguistic concepts, using language as a semantic pivot to bridge heterogeneous visual representations. To further mitigate the granularity mismatch between high-level language representations and dense detection objectives, LVSA progressively aggregates multi-scale visual features to provide fine-grained semantic guidance. Extensive experiments demonstrate that BabelRS stabilizes training and consistently outperforms state-of-the-art methods without bells and whistles. Code: https://github.com/zcablii/SM3Det.

Method: The paper proposes adapting only the decoder for test-time optimization, based on the insight that depth foundation models concentrate depth-relevant information within a low-dimensional decoder subspace. The method updates only this low-dimensional subspace using sparse depth supervision.

Result: Achieves state-of-the-art performance on five indoor and outdoor datasets, establishing a new Pareto frontier between accuracy and efficiency for test-time adaptation. Demonstrates consistent improvements over prior methods.

Conclusion: The proposed lightweight test-time adaptation method enables fast zero-shot depth completion by efficiently updating only the decoder subspace, making it practical for real-world applications while maintaining high accuracy.

Abstract: Zero-shot depth completion has gained attention for its ability to generalize across environments without sensor-specific datasets or retraining. However, most existing approaches rely on diffusion-based test-time optimization, which is computationally expensive due to iterative denoising. Recent visual-prompt-based methods reduce training cost but still require repeated forward–backward passes through the full frozen network to optimize input-level prompts, resulting in slow inference. In this work, we show that adapting only the decoder is sufficient for effective test-time optimization, as depth foundation models concentrate depth-relevant information within a low-dimensional decoder subspace. Based on this insight, we propose a lightweight test-time adaptation method that updates only this low-dimensional subspace using sparse depth supervision. Our approach achieves state-of-the-art performance, establishing a new Pareto frontier between accuracy and efficiency for test-time adaptation. Extensive experiments on five indoor and outdoor datasets demonstrate consistent improvements over prior methods, highlighting the practicality of fast zero-shot depth completion.

Method: Propose neural operator-grounded mode-n operators as continuous, nonlinear alternatives to discrete, linear mode-n products. These operators directly map continuous core tensor functions to continuous target tensor functions. Build NO-CTR (neural operator-grounded continuous tensor function representation) using these operators.

Result: Theoretical proof that any continuous tensor function can be approximated by NO-CTR. Experimental validation through a multi-dimensional data completion model shows superiority across various data types: multi-spectral images, color videos, Sentinel-2 images (different resolutions), and point clouds.

Conclusion: NO-CTR provides a more faithful representation of complex real-world data compared to classic discrete tensor representations and existing continuous tensor function representations, effectively handling data on regular grids, grids with different resolutions, and beyond mesh grids.

Abstract: Recently, continuous tensor functions have attracted increasing attention, because they can unifiedly represent data both on mesh grids and beyond mesh grids. However, since mode-$n$ product is essentially discrete and linear, the potential of current continuous tensor function representations is still locked. To break this bottleneck, we suggest neural operator-grounded mode-$n$ operators as a continuous and nonlinear alternative of discrete and linear mode-$n$ product. Instead of mapping the discrete core tensor to the discrete target tensor, proposed mode-$n$ operator directly maps the continuous core tensor function to the continuous target tensor function, which provides a genuine continuous representation of real-world data and can ameliorate discretization artifacts. Empowering with continuous and nonlinear mode-$n$ operators, we propose a neural operator-grounded continuous tensor function representation (abbreviated as NO-CTR), which can more faithfully represent complex real-world data compared with classic discrete tensor representations and continuous tensor function representations. Theoretically, we also prove that any continuous tensor function can be approximated by NO-CTR. To examine the capability of NO-CTR, we suggest an NO-CTR-based multi-dimensional data completion model. Extensive experiments across various data on regular mesh grids (multi-spectral images and color videos), on mesh girds with different resolutions (Sentinel-2 images) and beyond mesh grids (point clouds) demonstrate the superiority of NO-CTR.

Method: 1) Overviews fundamental statements for affine transformations and epipolar geometry. 2) Investigates how transformation accuracy influences 3D reconstruction quality. 3) Proposes novel techniques for estimating local affine transformation from corresponding image directions. 4) Exploits fundamental matrix related to image pairs. 5) Conducts synthetic and real quantitative evaluations using a specially constructed object with three perpendicular planes with chessboard patterns.

Result: Quantitative evaluations based on reconstructed surface normal accuracy show estimation accuracy around a few degrees for realistic test cases. Special stereo poses and plane orientations are evaluated in detail.

Conclusion: Affine transformations provide effective means for stereo vision applications with good accuracy in surface normal estimation (few degrees error). The relationship between transformation accuracy and 3D reconstruction quality is established, with proposed methods showing practical utility.

Abstract: Affine transformations have been recently used for stereo vision. They can be exploited in various computer vision application, e.g., when estimating surface normals, homographies, fundamental and essential matrices. Even full 3D reconstruction can be obtained by using affine correspondences. First, this paper overviews the fundamental statements for affine transformations and epipolar geometry. Then it is investigated how the transformation accuracy influences the quality of the 3D reconstruction. Besides, we propose novel techniques for estimating the local affine transformation from corresponding image directions; moreover, the fundamental matrix, related to the processed image pair, can also be exploited. Both synthetic and real quantitative evaluations are implemented based on the accuracy of the reconstructed surface normals. For the latter one, a special object, containing three perpendicular planes with chessboard patterns, is constructed. The quantitative evaluations are based on the accuracy of the reconstructed surface normals and it is concluded that the estimation accuracy is around a few degrees for realistic test cases. Special stereo poses and plane orientations are also evaluated in detail.

Method: Proposes LEAR, a dual-task learning framework that jointly estimates edge structures and dense event-depth flow fields. Uses cross-modal fusion to inject modality-invariant geometric cues into motion representation, and iterative refinement to enforce mutual consistency between edge and flow estimation tasks over multiple update steps.

Result: Achieves superior performance over prior methods on several popular and challenging datasets. Produces edge-aware, depth-aligned flow fields that enable more robust and accurate pose recovery via Perspective-n-Point (PnP) solvers.

Conclusion: LEAR effectively bridges the sensing-modality divide between event cameras and LiDAR through synergistic edge and flow estimation, enabling reliable localization in GPS-denied and visually degraded environments.

Abstract: Event cameras offer high-temporal-resolution sensing that remains reliable under high-speed motion and challenging lighting, making them promising for localization from LiDAR point clouds in GPS-denied and visually degraded environments. However, aligning sparse, asynchronous events with dense LiDAR maps is fundamentally ill-posed, as direct correspondence estimation suffers from modality gaps. We propose LEAR, a dual-task learning framework that jointly estimates edge structures and dense event-depth flow fields to bridge the sensing-modality divide. Instead of treating edges as a post-hoc aid, LEAR couples them with flow estimation through a cross-modal fusion mechanism that injects modality-invariant geometric cues into the motion representation, and an iterative refinement strategy that enforces mutual consistency between the two tasks over multiple update steps. This synergy produces edge-aware, depth-aligned flow fields that enable more robust and accurate pose recovery via Perspective-n-Point (PnP) solvers. On several popular and challenging datasets, LEAR achieves superior performance over the best prior method. The source code, trained models, and demo videos are made publicly available online.

Method: Adds multiple diverse groups of object queries to transformer decoder with attention mask to prevent inter-group interactions. Each group predicts independent detection sets in parallel during single forward pass.

Result: Hybrid approach (MC-Dropout + GroupEnsemble) outperforms Deep Ensembles on several metrics at fraction of cost, validated on Cityscapes and COCO datasets.

Conclusion: GroupEnsemble provides efficient uncertainty estimation for DETR models, addressing both semantic and spatial uncertainty with single-pass inference.

Abstract: Detection Transformer (DETR) and its variants show strong performance on object detection, a key task for autonomous systems. However, a critical limitation of these models is that their confidence scores only reflect semantic uncertainty, failing to capture the equally important spatial uncertainty. This results in an incomplete assessment of the detection reliability. On the other hand, Deep Ensembles can tackle this by providing high-quality spatial uncertainty estimates. However, their immense memory consumption makes them impractical for real-world applications. A cheaper alternative, Monte Carlo (MC) Dropout, suffers from high latency due to the need of multiple forward passes during inference to estimate uncertainty. To address these limitations, we introduce GroupEnsemble, an efficient and effective uncertainty estimation method for DETR-like models. GroupEnsemble simultaneously predicts multiple individual detection sets by feeding additional diverse groups of object queries to the transformer decoder during inference. Each query group is transformed by the shared decoder in isolation and predicts a complete detection set for the same input. An attention mask is applied to the decoder to prevent inter-group query interactions, ensuring each group detects independently to achieve reliable ensemble-based uncertainty estimation. By leveraging the decoder’s inherent parallelism, GroupEnsemble efficiently estimates uncertainty in a single forward pass without sequential repetition. We validated our method under autonomous driving scenes and common daily scenes using the Cityscapes and COCO datasets, respectively. The results show that a hybrid approach combining MC-Dropout and GroupEnsemble outperforms Deep Ensembles on several metrics at a fraction of the cost. The code is available at https://github.com/yutongy98/GroupEnsemble.

Method: Proposes a lightweight streaming approach that integrates information from previous predictions using endpoint-aware modeling. Uses trajectory endpoints from previous forecasts as anchors to extract targeted scenario context encodings, efficiently guiding scene encoder without refinement iterations.

Result: Achieves state-of-the-art streaming trajectory prediction results on Argoverse 2 multi-agent and single-agent benchmarks while requiring substantially fewer resources and significantly reducing inference latency.

Conclusion: The approach effectively relays information across consecutive timesteps, making it well-suited for real-world autonomous driving deployment by providing consistent predictions with low latency.

Abstract: Future trajectories of neighboring traffic agents have a significant influence on the path planning and decision-making of autonomous vehicles. While trajectory forecasting is a well-studied field, research mainly focuses on snapshot-based prediction, where each scenario is treated independently of its global temporal context. However, real-world autonomous driving systems need to operate in a continuous setting, requiring real-time processing of data streams with low latency and consistent predictions over successive timesteps. We leverage this continuous setting to propose a lightweight yet highly accurate streaming-based trajectory forecasting approach. We integrate valuable information from previous predictions with a novel endpoint-aware modeling scheme. Our temporal context propagation uses the trajectory endpoints of the previous forecasts as anchors to extract targeted scenario context encodings. Our approach efficiently guides its scene encoder to extract highly relevant context information without needing refinement iterations or segment-wise decoding. Our experiments highlight that our approach effectively relays information across consecutive timesteps. Unlike methods using multi-stage refinement processing, our approach significantly reduces inference latency, making it well-suited for real-world deployment. We achieve state-of-the-art streaming trajectory prediction results on the Argoverse~2 multi-agent and single-agent benchmarks, while requiring substantially fewer resources.

Method: Proposes CTForensics dataset with 10 diverse CT generative methods, and ESF-CTFD network that extracts features across wavelet, spatial, and frequency domains using Wavelet-Enhanced Central Stem, Spatial Process Blocks, and Frequency Process Blocks.

Result: ESF-CTFD consistently outperforms existing methods and shows superior generalization across different CT generative models.

Conclusion: The comprehensive dataset and specialized detector effectively address CT forgery detection challenges, improving security in medical imaging applications.

Abstract: With the rapid development of generative AI in medical imaging, synthetic Computed Tomography (CT) images have demonstrated great potential in applications such as data augmentation and clinical diagnosis, but they also introduce serious security risks. Despite the increasing security concerns, existing studies on CT forgery detection are still limited and fail to adequately address real-world challenges. These limitations are mainly reflected in two aspects: the absence of datasets that can effectively evaluate model generalization to reflect the real-world application requirements, and the reliance on detection methods designed for natural images that are insensitive to CT-specific forgery artifacts. In this view, we propose CTForensics, a comprehensive dataset designed to systematically evaluate the generalization capability of CT forgery detection methods, which includes ten diverse CT generative methods. Moreover, we introduce the Enhanced Spatial-Frequency CT Forgery Detector (ESF-CTFD), an efficient CNN-based neural network that captures forgery cues across the wavelet, spatial, and frequency domains. First, it transforms the input CT image into three scales and extracts features at each scale via the Wavelet-Enhanced Central Stem. Then, starting from the largest-scale features, the Spatial Process Block gradually performs feature fusion with the smaller-scale ones. Finally, the Frequency Process Block learns frequency-domain information for predicting the final results. Experiments demonstrate that ESF-CTFD consistently outperforms existing methods and exhibits superior generalization across different CT generative models.

Method: Proposes CaMB-Diff: identifies monotonicity as physical invariant, introduces cascaded monotonic Bernstein (CaMB) operator to parameterize unknown forward model with hard architectural inductive bias, integrates with plug-and-play diffusion framework for structural/semantic recovery.

Result: Significantly outperforms state-of-the-art zero-shot baselines on low-light enhancement, low-field MRI enhancement, and HDR reconstruction in signal fidelity and physical consistency. Validates CaMB’s effectiveness in modeling unknown forward operators.

Conclusion: CaMB-Diff successfully addresses blind inverse UDRC problems by combining diffusion geometric priors with physically-grounded monotonic operators, enabling robust radiometric recovery across diverse applications.

Abstract: Recovering radiometric fidelity from unknown dynamic range compression (UDRC), such as low-light enhancement and HDR reconstruction, is a challenging blind inverse problem, due to the unknown forward model and irreversible information loss introduced by compression. To address this challenge, we first identify monotonicity as the fundamental physical invariant shared across UDRC tasks. Leveraging this insight, we introduce the \textbf{cascaded monotonic Bernstein} (CaMB) operator to parameterize the unknown forward model. CaMB enforces monotonicity as a hard architectural inductive bias, constraining optimization to physically consistent mappings and enabling robust and stable operator estimation. We further integrate CaMB with a plug-and-play diffusion framework, proposing \textbf{CaMB-Diff}. Within this framework, the diffusion model serves as a powerful geometric prior for structural and semantic recovery, while CaMB explicitly models and corrects radiometric distortions through a physically grounded forward operator. Extensive experiments on a variety of zero-shot UDRC tasks, including low-light enhancement, low-field MRI enhancement, and HDR reconstruction, demonstrate that CaMB-Diff significantly outperforms state-of-the-art zero-shot baselines in terms of both signal fidelity and physical consistency. Moreover, we empirically validate the effectiveness of the proposed CaMB parameterization in accurately modeling the unknown forward operator.

Method: Proposes Generative Visual Chain-of-Thought (GVCoT) that performs native visual reasoning by first generating spatial cues to localize target regions, then executing edits. Uses end-to-end joint optimization of visual tokens across reasoning and editing phases. Constructs GVCoT-Edit-Instruct dataset (1.8M samples, 19 tasks) and employs progressive training: supervised fine-tuning for foundational localization, followed by reinforcement learning.

Result: GVCoT consistently outperforms state-of-the-art models on SREdit-Bench (new benchmark for sophisticated scenes and fine-grained referring expressions) and ImgEdit benchmarks.

Conclusion: GVCoT enables more interpretable and precise image editing through innate spatial reasoning ability and effective utilization of visual-domain cues, inspiring future research in this direction.

Abstract: Existing image editing methods struggle to perceive where to edit, especially under complex scenes and nuanced spatial instructions. To address this issue, we propose Generative Visual Chain-of-Thought (GVCoT), a unified framework that performs native visual reasoning by first generating spatial cues to localize the target region and then executing the edit. Unlike prior text-only CoT or tool-dependent visual CoT paradigms, GVCoT jointly optimizes visual tokens generated during the reasoning and editing phases in an end-to-end manner. This way fosters the emergence of innate spatial reasoning ability and enables more effective utilization of visual-domain cues. The main challenge of training GCVoT lies in the scarcity of large-scale editing data with precise edit region annotations; to this end, we construct GVCoT-Edit-Instruct, a dataset of 1.8M high-quality samples spanning 19 tasks. We adopt a progressive training strategy: supervised fine-tuning to build foundational localization ability in reasoning trace before final editing, followed by reinforcement learning to further improve reasoning and editing quality. Finally, we introduce SREdit-Bench, a new benchmark designed to comprehensively stress-test models under sophisticated scenes and fine-grained referring expressions. Experiments demonstrate that GVCoT consistently outperforms state-of-the-art models on SREdit-Bench and ImgEdit. We hope our GVCoT will inspire future research toward interpretable and precise image editing.

Method: DACT combines a pre-trained high-field diffusion prior with a physically-informed adaptive forward model. It uses a differentiable Sinkhorn optimal transport module to explicitly model and correct intensity distribution shifts between low-field and high-field domains during reverse diffusion.

Result: Extensive experiments on simulated and real clinical low-field datasets show DACT achieves state-of-the-art performance, producing reconstructions with superior structural detail and correct tissue contrast.

Conclusion: DACT provides an effective zero-shot solution for high-quality MRI reconstruction from low-field data without paired supervision, addressing the fundamental limitations of low-field MRI through diffusion modeling and optimal transport.

Abstract: Low-field (LF) magnetic resonance imaging (MRI) democratizes access to diagnostic imaging but is fundamentally limited by low signal-to-noise ratio and significant tissue contrast distortion due to field-dependent relaxation dynamics. Reconstructing high-field (HF) quality images from LF data is a blind inverse problem, severely challenged by the scarcity of paired training data and the unknown, non-linear contrast transformation operator. Existing zero-shot methods, which assume simplified linear degradation, often fail to recover authentic tissue contrast. In this paper, we propose DACT(Diffusion-Based Adaptive Contrast Transport), a novel zero-shot framework that restores HF-quality images without paired supervision. DACT synergizes a pre-trained HF diffusion prior to ensure anatomical fidelity with a physically-informed adaptive forward model. Specifically, we introduce a differentiable Sinkhorn optimal transport module that explicitly models and corrects the intensity distribution shift between LF and HF domains during the reverse diffusion process. This allows the framework to dynamically learn the intractable contrast mapping while preserving topological consistency. Extensive experiments on simulated and real clinical LF datasets demonstrate that DACT achieves state-of-the-art performance, yielding reconstructions with superior structural detail and correct tissue contrast.

Method: Proposes Latent Spatio-Temporal VLA (LaST-VLA) with dual-feature alignment to distill geometric constraints from 3D foundation models and dynamic foresight from world models into latent space. Uses progressive SFT training transitioning from feature alignment to trajectory generation, refined via Reinforcement Learning with Group Relative Policy Optimization for safety.

Result: Sets new records on NAVSIM v1 (91.3 PDMS) and NAVSIM v2 (87.1 EPDMS), and excels in spatial-temporal reasoning on SURDS and NuDynamics benchmarks.

Conclusion: LaST-VLA successfully shifts reasoning from discrete symbolic processing to physically grounded latent spatio-temporal Chain-of-Thought, improving autonomous driving performance through better integration of perception and planning.

Abstract: While Vision-Language-Action (VLA) models have revolutionized autonomous driving by unifying perception and planning, their reliance on explicit textual Chain-of-Thought (CoT) leads to semantic-perceptual decoupling and perceptual-symbolic conflicts. Recent shifts toward latent reasoning attempt to bypass these bottlenecks by thinking in continuous hidden space. However, without explicit intermediate constraints, standard latent CoT often operates as a physics-agnostic representation. To address this, we propose the Latent Spatio-Temporal VLA (LaST-VLA), a framework shifting the reasoning paradigm from discrete symbolic processing into a physically grounded Latent Spatio-Temporal CoT. By implementing a dual-feature alignment mechanism, we distill geometric constraints from 3D foundation models and dynamic foresight from world models directly into the latent space. Coupled with a progressive SFT training strategy that transitions from feature alignment to trajectory generation, and refined via Reinforcement Learning with Group Relative Policy Optimization (GRPO) to ensure safety and rule compliance. \method~setting a new record on NAVSIM v1 (91.3 PDMS) and NAVSIM v2 (87.1 EPDMS), while excelling in spatial-temporal reasoning on SURDS and NuDynamics benchmarks.

Method: Proposes VISA, a two-stream segmentation network that decouples radiance and vegetation-index cues. The radiance stream learns from calibrated five-band reflectance using residual spectral-spatial attention. The index stream operates on vegetation-index maps with windowed self-attention, state-space layers for field-scale context, and Slot Attention for stable region descriptors. The network fuses streams at native resolution.

Result: On the BAWSeg dataset, VISA achieves 75.6% mIoU and 63.5% weed IoU with 22.8M parameters, outperforming multispectral SegFormer-B1 by 1.2 mIoU and 1.9 weed IoU. Maintains 71.2% and 69.2% mIoU under cross-plot and cross-year protocols respectively.

Conclusion: VISA effectively addresses limitations of existing weed mapping methods by decoupling radiance and index cues, achieving robust performance across fields and seasons. The BAWSeg dataset supports supervised training and deployment-oriented evaluation.

Abstract: Accurate weed mapping in cereal fields requires pixel-level segmentation from UAV imagery that remains reliable across fields, seasons, and illumination. Existing multispectral pipelines often depend on thresholded vegetation indices, which are brittle under radiometric drift and mixed crop–weed pixels, or on single-stream CNN and Transformer backbones that ingest stacked bands and indices, where radiance cues and normalized index cues interfere and reduce sensitivity to small weed clusters embedded in crop canopies. We propose VISA (Vegetation-Index and Spectral Attention), a two-stream segmentation network that decouples these cues and fuses them at native resolution. The radiance stream learns from calibrated five-band reflectance using residual spectral-spatial attention to preserve fine textures and row boundaries that are attenuated by ratio indices. The index stream operates on vegetation-index maps with windowed self-attention to model local structure efficiently, state-space layers to propagate field-scale context without quadratic attention cost, and Slot Attention to form stable region descriptors that improve discrimination of sparse weeds under canopy mixing. To support supervised training and deployment-oriented evaluation, we introduce BAWSeg, a four-year UAV multispectral dataset collected over commercial barley paddocks in Western Australia, providing radiometrically calibrated blue, green, red, red edge, and near-infrared orthomosaics, derived vegetation indices, and dense crop, weed, and other labels with leakage-free block splits. On BAWSeg, VISA achieves 75.6% mIoU and 63.5% weed IoU with 22.8M parameters, outperforming a multispectral SegFormer-B1 baseline by 1.2 mIoU and 1.9 weed IoU. Under cross-plot and cross-year protocols, VISA maintains 71.2% and 69.2% mIoU, respectively. The BAWSeg data, VISA code, and trained models will be released upon publication.

Method: Collected 4,941 handheld microscopy images across 11 food types using 4 different smartphones and 3 clip-on microscopes under diverse real-world conditions. Established baselines for mold detection and food classification using pretrained deep learning architectures with various augmentation strategies.

Result: Achieved near-ceiling performance with accuracy = 0.9954, F1 = 0.9954, MCC = 0.9907 for mold detection. Also provided saliency-based visual explanations highlighting mold regions associated with predictions.

Conclusion: MobileMold dataset enables effective food spoilage detection using smartphone microscopy and contributes to research on accessible food-safety sensing, mobile imaging, and smartphone-enhanced attachments.

Abstract: Smartphone clip-on microscopes turn everyday devices into low-cost, portable imaging systems that can even reveal fungal structures at the microscopic level, enabling mold inspection beyond unaided visual checks. In this paper, we introduce MobileMold, an open smartphone-based microscopy dataset for food mold detection and food classification. MobileMold contains 4,941 handheld microscopy images spanning 11 food types, 4 smartphones, 3 microscopes, and diverse real-world conditions. Beyond the dataset release, we establish baselines for (i) mold detection and (ii) food-type classification, including a multi-task setting that predicts both attributes. Across multiple pretrained deep learning architectures and augmentation strategies, we obtain near-ceiling performance (accuracy = 0.9954, F1 = 0.9954, MCC = 0.9907), validating the utility of our dataset for detecting food spoilage. To increase transparency, we complement our evaluation with saliency-based visual explanations highlighting mold regions associated with the model’s predictions. MobileMold aims to contribute to research on accessible food-safety sensing, mobile imaging, and exploring the potential of smartphones enhanced with attachments.

Method: PreSight combines clinical priors with preoperative MRI and deformation-based morphometry (DBM) features. It uses a patient-specific weighting module to adapt regional importance, producing end-to-end calibrated predictions with patient-level explanations.

Result: Achieved 88.89% accuracy on internal validation and 85.29% on external-center test for responder classification. Outperformed clinical, imaging-only, and multimodal baselines with better probability calibration and higher decision-curve net benefit.

Conclusion: Integrating clinical prior knowledge with region-adaptive morphometry enables reliable presurgical decision support for Parkinson’s disease surgery, demonstrating the value of patient-specific multimodal fusion.

Abstract: Preoperative improvement rate prediction for Parkinson’s disease surgery is clinically important yet difficult because imaging signals are subtle and patients are heterogeneous. We address this setting, where only information available before surgery is used, and the goal is to predict patient-specific postoperative motor benefit. We present PreSight, a presurgical outcome model that fuses clinical priors with preoperative MRI and deformation-based morphometry (DBM) and adapts regional importance through a patient-specific weighting module. The model produces end-to-end, calibrated, decision-ready predictions with patient-level explanations. We evaluate PreSight on a real-world two-center cohort of 400 subjects with multimodal presurgical inputs and postoperative improvement labels. PreSight outperforms strong clinical, imaging-only, and multimodal baselines. It attains 88.89% accuracy on internal validation and 85.29% on an external-center test for responder classification and shows better probability calibration and higher decision-curve net benefit. Ablations and analyses confirm the contribution of DBM and the patient-specific weighting module and indicate that the model emphasizes disease-relevant regions in a patient-specific manner. These results demonstrate that integrating clinical prior knowledge with region-adaptive morphometry enables reliable presurgical decision support in routine practice.

Method: Proposes a stain-normalized, decoupled training framework: 1) learns transferable representations using instance-balanced sampling, 2) rebalances classifier with class-aware sampling and hybrid loss (effective-number weighting + focal modulation), 3) enhances robustness with ensemble of trained backbones and test-time augmentation.

Result: Achieved top rank on the leaderboard of the WBCBench 2026: Robust White Blood Cell Classification Challenge at ISBI 2026.

Conclusion: The proposed framework effectively addresses appearance variations and class imbalance in WBC classification, demonstrating state-of-the-art performance on benchmark challenges.

Abstract: White blood cell (WBC) classification is fundamental for hematology applications such as infection assessment, leukemia screening, and treatment monitoring. However, real-world WBC datasets present substantial appearance variations caused by staining and scanning conditions, as well as severe class imbalance in which common cell types dominate while rare but clinically important categories are underrepresented. To address these challenges, we propose a stain-normalized, decoupled training framework that first learns transferable representations using instance-balanced sampling, and then rebalances the classifier with class-aware sampling and a hybrid loss combining effective-number weighting and focal modulation. In inference stage, we further enhance robustness by ensembling various trained backbones with test-time augmentation. Our approach achieved the top rank on the leaderboard of the WBCBench 2026: Robust White Blood Cell Classification Challenge at ISBI 2026.

Method: REFORM uses a three-stage curriculum: 1) induces forensic rationales, 2) aligns reasoning with final judgments, and 3) refines logical consistency via reinforcement learning. The framework is supported by ROM, a large-scale dataset with rich reasoning annotations.

Result: REFORM establishes new state-of-the-art performance with superior generalization: 81.52% ACC on ROM, 76.65% ACC on DGM4, and 74.9 F1 on MMFakeBench.

Conclusion: The reasoning-driven approach to manipulation detection provides better interpretability and generalization compared to traditional classification-based methods, addressing limitations of existing approaches in multimodal media manipulation detection.

Abstract: Recent advances in generative AI have significantly enhanced the realism of multimodal media manipulation, thereby posing substantial challenges to manipulation detection. Existing manipulation detection and grounding approaches predominantly focus on manipulation type classification under result-oriented supervision, which not only lacks interpretability but also tends to overfit superficial artifacts. In this paper, we argue that generalizable detection requires incorporating explicit forensic reasoning, rather than merely classifying a limited set of manipulation types, which fails to generalize to unseen manipulation patterns. To this end, we propose REFORM, a reasoning-driven framework that shifts learning from outcome fitting to process modeling. REFORM adopts a three-stage curriculum that first induces forensic rationales, then aligns reasoning with final judgments, and finally refines logical consistency via reinforcement learning. To support this paradigm, we introduce ROM, a large-scale dataset with rich reasoning annotations. Extensive experiments show that REFORM establishes new state-of-the-art performance with superior generalization, achieving 81.52% ACC on ROM, 76.65% ACC on DGM4, and 74.9 F1 on MMFakeBench.

Method: Created NICO Dataset with 200K+ multimodal samples (images + text) across 55 brands. Developed NICO-RAG framework that organizes image/text entities into hypergraphs for retrieval, using image features directly instead of processing image tokens through language models.

Result: NICO-RAG performs comparably to state-of-the-art image-adapted RAG methods on over 100 questions without needing to process additional image tokens, demonstrating efficient multimodal retrieval capability.

Conclusion: The NICO Dataset and NICO-RAG framework provide valuable resources for public health research on nicotine products, enabling efficient multimodal analysis without the computational costs of processing image tokens through large language models.

Abstract: The nicotine addiction public health crisis continues to be pervasive. In this century alone, the tobacco industry has released and marketed new products in an aggressive effort to lure new and young customers for life. Such innovations and product development, namely flavored nicotine or tobacco such as nicotine pouches, have undone years of anti-tobacco campaign work. Past work is limited both in scope and in its ability to connect large-scale data points. Thus, we introduce the Nicotine Innovation Counter-Offensive (NICO) Dataset to provide public health researchers with over 200,000 multimodal samples, including images and text descriptions, on 55 tobacco and nicotine product brands. In addition, to provide public health researchers with factual connections across a large-scale dataset, we propose NICO-RAG, a retrieval-augmented generation (RAG) framework that can retrieve image features without incurring the high-cost of language models, as well as the added cost of processing image tokens with large-scale datasets such as NICO. At construction time, NICO-RAG organizes image- and text-extracted entities and relations into hypergraphs to produce as factual responses as possible. This joint multimodal knowledge representation enables NICO-RAG to retrieve images for query answering not only by visual similarity but also by the semantic similarity of image descriptions. Experimentals show that without needing to process additional tokens from images for over 100 questions, NICO-RAG performs comparably to the state-of-the-art RAG method adapted for images.