Method: Built via semi-automatic pipeline generating 1.8k videos using multiple popular models with carefully designed prompts, yielding 2.4k manually annotated QA pairs. Evaluates at four levels: audio-visual signal fidelity, temporal attribute consistency, social interaction, and cinematic expression.

Result: Benchmarked 12 proprietary and open-source omni-models, with Gemini 3 Pro achieving strongest overall performance. Leading open-source models remain competitive in signal fidelity and consistency.

Conclusion: MTAVG-Bench enables fine-grained failure analysis for rigorous model comparison and targeted video generation refinement in multi-speaker dialogue generation.

Abstract: Recent advances in text-to-audio-video (T2AV) generation have enabled models to synthesize audio-visual videos with multi-participant dialogues. However, existing evaluation benchmarks remain largely designed for human-recorded videos or single-speaker settings. As a result, potential errors that occur in generated multi-talker dialogue videos, such as identity drift, unnatural turn transitions, and audio-visual misalignment, cannot be effectively captured and analyzed. To address this issue, we introduce MTAVG-Bench, a benchmark for evaluating audio-visual multi-speaker dialogue generation. MTAVG-Bench is built via a semi-automatic pipeline, where 1.8k videos are generated using multiple popular models with carefully designed prompts, yielding 2.4k manually annotated QA pairs. The benchmark evaluates multi-speaker dialogue generation at four levels: audio-visual signal fidelity, temporal attribute consistency, social interaction, and cinematic expression. We benchmark 12 proprietary and open-source omni-models on MTAVG-Bench, with Gemini 3 Pro achieving the strongest overall performance, while leading open-source models remain competitive in signal fidelity and consistency. Overall, MTAVG-Bench enables fine-grained failure analysis for rigorous model comparison and targeted video generation refinement.

Method: Introduces CMQKA with bidirectional cross-modal Query-Key attention achieving linear O(N) complexity through binary operations, and SNNergy framework with hierarchical architecture using event-driven binary spike operations for energy efficiency.

Result: Achieves state-of-the-art results on audio-visual benchmarks CREMA-D, AVE, and UrbanSound8K-AV, significantly outperforming existing multimodal fusion baselines with remarkable energy efficiency.

Conclusion: Advances multimodal fusion by introducing scalable fusion mechanism enabling hierarchical cross-modal integration with practical energy efficiency for real-world audio-visual intelligence systems.

Abstract: Effective multimodal fusion requires mechanisms that can capture complex cross-modal dependencies while remaining computationally scalable for real-world deployment. Existing audio-visual fusion approaches face a fundamental trade-off: attention-based methods effectively model cross-modal relationships but incur quadratic computational complexity that prevents hierarchical, multi-scale architectures, while efficient fusion strategies rely on simplistic concatenation that fails to extract complementary cross-modal information. We introduce CMQKA, a novel cross-modal fusion mechanism that achieves linear O(N) complexity through efficient binary operations, enabling scalable hierarchical fusion previously infeasible with conventional attention. CMQKA employs bidirectional cross-modal Query-Key attention to extract complementary spatiotemporal features and uses learnable residual fusion to preserve modality-specific characteristics while enriching representations with cross-modal information. Building upon CMQKA, we present SNNergy, an energy-efficient multimodal fusion framework with a hierarchical architecture that processes inputs through progressively decreasing spatial resolutions and increasing semantic abstraction. This multi-scale fusion capability allows the framework to capture both local patterns and global context across modalities. Implemented with event-driven binary spike operations, SNNergy achieves remarkable energy efficiency while maintaining fusion effectiveness and establishing new state-of-the-art results on challenging audio-visual benchmarks, including CREMA-D, AVE, and UrbanSound8K-AV, significantly outperforming existing multimodal fusion baselines. Our framework advances multimodal fusion by introducing a scalable fusion mechanism that enables hierarchical cross-modal integration with practical energy efficiency for real-world audio-visual intelligence systems.

Method: U-Net architecture with residual CBAM to encode and fuse audio-visual information; semantic alignment module to extend receptive field and match statistical information between visual features and audio latent vectors; LPIPS loss for better image quality and training stability.

Result: The method achieves outstanding performance in lip synchronization accuracy and visual quality, as demonstrated by both subjective and objective evaluations.

Conclusion: LPIPS-AttnWav2Lip provides an effective generic solution for audio-driven talking head generation with improved lip synchronization and image quality through novel architectural components and loss functions.

Abstract: Researchers have shown a growing interest in Audio-driven Talking Head Generation. The primary challenge in talking head generation is achieving audio-visual coherence between the lips and the audio, known as lip synchronization. This paper proposes a generic method, LPIPS-AttnWav2Lip, for reconstructing face images of any speaker based on audio. We used the U-Net architecture based on residual CBAM to better encode and fuse audio and visual modal information. Additionally, the semantic alignment module extends the receptive field of the generator network to obtain the spatial and channel information of the visual features efficiently; and match statistical information of visual features with audio latent vector to achieve the adjustment and injection of the audio content information to the visual information. To achieve exact lip synchronization and to generate realistic high-quality images, our approach adopts LPIPS Loss, which simulates human judgment of image quality and reduces instability possibility during the training process. The proposed method achieves outstanding performance in terms of lip synchronization accuracy and visual quality as demonstrated by subjective and objective evaluation results. The code for the paper is available at the following link: https://github.com/FelixChan9527/LPIPS-AttnWav2Lip

Method: PPoGA uses a Planner-Executor architecture to separate high-level strategy from low-level execution, incorporates Predictive Processing to anticipate outcomes, and features a self-correction mechanism that performs both Path Correction (local execution errors) and Plan Correction (identifying, discarding, and reformulating entire ineffective plans).

Result: Extensive experiments on three challenging multi-hop KGQA benchmarks (GrailQA, CWQ, WebQSP) show PPoGA achieves state-of-the-art performance, significantly outperforming existing methods.

Conclusion: The work highlights the importance of metacognitive abilities like problem restructuring for building more robust and flexible AI reasoning systems, demonstrating that self-correction mechanisms can overcome limitations of fixed reasoning plans in KG-augmented LLMs.

Abstract: Large Language Models (LLMs) augmented with Knowledge Graphs (KGs) have advanced complex question answering, yet they often remain susceptible to failure when their initial high-level reasoning plan is flawed. This limitation, analogous to cognitive functional fixedness, prevents agents from restructuring their approach, leading them to pursue unworkable solutions. To address this, we propose PPoGA (Predictive Plan-on-Graph with Action), a novel KGQA framework inspired by human cognitive control and problem-solving. PPoGA incorporates a Planner-Executor architecture to separate high-level strategy from low-level execution and leverages a Predictive Processing mechanism to anticipate outcomes. The core innovation of our work is a self-correction mechanism that empowers the agent to perform not only Path Correction for local execution errors but also Plan Correction by identifying, discarding, and reformulating the entire plan when it proves ineffective. We conduct extensive experiments on three challenging multi-hop KGQA benchmarks: GrailQA, CWQ, and WebQSP. The results demonstrate that PPoGA achieves state-of-the-art performance, significantly outperforming existing methods. Our work highlights the critical importance of metacognitive abilities like problem restructuring for building more robust and flexible AI reasoning systems.

Method: MediGRAF combines Neo4j Text2Cypher capabilities for structured relationship traversal with vector embeddings for unstructured narrative retrieval, creating a hybrid Graph RAG system that enables natural language querying of complete patient journeys using MIMIC-IV dataset.

Result: The system achieved 100% recall for factual queries and a mean expert quality score of 4.25/5 for complex inference tasks with zero safety violations, using 10 patients from MIMIC-IV dataset (5,973 nodes, 5,963 relationships).

Conclusion: Hybrid graph-grounding significantly advances clinical information retrieval, offering a safer, more comprehensive alternative to standard LLM deployments by bridging structured and unstructured data retrieval.

Abstract: Electronic health record (EHR) systems present clinicians with vast repositories of clinical information, creating a significant cognitive burden where critical details are easily overlooked. While Large Language Models (LLMs) offer transformative potential for data processing, they face significant limitations in clinical settings, particularly regarding context grounding and hallucinations. Current solutions typically isolate retrieval methods focusing either on structured data (SQL/Cypher) or unstructured semantic search but fail to integrate both simultaneously. This work presents MediGRAF (Medical Graph Retrieval Augmented Framework), a novel hybrid Graph RAG system that bridges this gap. By uniquely combining Neo4j Text2Cypher capabilities for structured relationship traversal with vector embeddings for unstructured narrative retrieval, MediGRAF enables natural language querying of the complete patient journey. Using 10 patients from the MIMIC-IV dataset (generating 5,973 nodes and 5,963 relationships), we generated enough nodes and data for patient level question answering (QA), and we evaluated this architecture across varying query complexities. The system demonstrated 100% recall for factual queries which means all relevant information was retrieved and in the output, while complex inference tasks achieved a mean expert quality score of 4.25/5 with zero safety violations. These results demonstrate that hybrid graph-grounding significantly advances clinical information retrieval, offering a safer, more comprehensive alternative to standard LLM deployments.

Method: Proposes G-MemLLM with frozen LLM backbone + trainable Latent Memory Bank using GRU-style gated update logic to selectively update, preserve, or overwrite memory slots, preventing vanishing gradients.

Result: Significant improvements on HotpotQA and ZsRE benchmarks: 13.3% accuracy boost on ZsRE for Llama 3.1-8B, 8.56 points Answer F1 boost for GPT-2, and 6.89 points Supporting Fact F1 boost for Llama 3.1-8B on HotpotQA.

Conclusion: G-MemLLM effectively enhances multi-hop reasoning and relational precision across model scales by addressing context limitations through gated memory mechanisms.

Abstract: Large Language Models (LLMs) have demonstrated remarkable capabilities in natural language understanding, yet they remain constrained by the finite capacity of their context windows and the inherent difficulty of maintaining long-term factual consistency during multi-hop reasoning. While existing methods utilize context compression or recurrent tokens, they often suffer from ``context rot’’ or the dilution of information over long horizons. In this paper, we propose \textbf{G-MemLLM}, a memory-augmented architecture that integrates a frozen LLM backbone with a trainable \textbf{Latent Memory Bank}. Our key innovation is a GRU-style gated update logic that allows the model to selectively update, preserve, or overwrite latent memory slots, preventing the vanishing gradients of knowledge common in recurrent systems. We evaluate G-MemLLM across scales, from GPT-2 (124M) to Llama 3.1 (8B), on the HotpotQA and Zero-Shot Relation Extraction (ZsRE) benchmarks. Our results demonstrate that G-MemLLM significantly enhances multi-hop reasoning and relational precision, achieving a 13.3% accuracy boost on ZsRE for Llama 3.1-8B, and it also yields improvements across model scales, boosting Answer F1 by 8.56 points for GPT-2 and increasing Supporting Fact F1 by 6.89 points for Llama 3.1-8B on HotpotQA.

Method: Introduces PTCBENCH benchmark that subjects models to 12 distinct external conditions (location contexts and life events) and assesses personality using the NEO Five-Factor Inventory across 39,240 personality trait records.

Result: Certain external scenarios like “Unemployment” can trigger significant personality changes in LLMs and even alter their reasoning capabilities, revealing context-dependent personality instability.

Conclusion: PTCBENCH establishes an extensible framework for evaluating personality consistency in realistic, evolving environments, offering insights for developing robust and psychologically aligned AI systems.

Abstract: With the increasing deployment of large language models (LLMs) in affective agents and AI systems, maintaining a consistent and authentic LLM personality becomes critical for user trust and engagement. However, existing work overlooks a fundamental psychological consensus that personality traits are dynamic and context-dependent. To bridge this gap, we introduce PTCBENCH, a systematic benchmark designed to quantify the consistency of LLM personalities under controlled situational contexts. PTCBENCH subjects models to 12 distinct external conditions spanning diverse location contexts and life events, and rigorously assesses the personality using the NEO Five-Factor Inventory. Our study on 39,240 personality trait records reveals that certain external scenarios (e.g., “Unemployment”) can trigger significant personality changes of LLMs, and even alter their reasoning capabilities. Overall, PTCBENCH establishes an extensible framework for evaluating personality consistency in realistic, evolving environments, offering actionable insights for developing robust and psychologically aligned AI systems.

Method: A diversity-driven multi-agent framework integrating crowd-sourced/synthesized scenarios, sexual health guidelines, evidence-based personas, adaptive control modules, and hierarchical diversification to generate realistic conversations.

Result: SafeTalkCoach generates diverse conversations while maintaining realism, communication quality, and controllability, as demonstrated through evaluations. The framework includes an accompanying dataset.

Conclusion: The SafeTalkCoach framework and dataset can support both AI research and health communication practices by providing realistic simulations of sensitive parent-child conversations about sexual health.

Abstract: The importance of effective parent-child communication about sexual health is widely acknowledged, but real-world data on these conversations is scarce and challenging to collect, due to their private and sensitive nature. Although LLMs have been widely adopted in dialogue generation, they may deviate from best practices and frequently lack realism and diversity. We introduce SafeTalkCoach, a diversity-driven multi-agent dialogue generation framework that simulates parent-child conversations about sexual health, and present an accompanying dataset. SafeTalkCoach integrates crowd-sourced and synthesized scenarios, established sexual health guidelines, evidence-based personas, adaptive control modules, and hierarchical diversification. Through evaluations, we demonstrate that SafeTalkCoach generates diverse conversations while maintaining realism, communication quality, and controllability in practice. Our goal is that the SafeTalkCoach framework and the dataset support both AI research and health communications practices.

Method: Proposes a construct-align-reason framework: 1) Build dual-layer enterprise ontology from structured databases and unstructured text, 2) Three-stage training: ontology instruction fine-tuning for structure understanding, text-ontology grounding for semantic encoding, and multi-task instruction tuning with curriculum learning on ontology-language pairs.

Result: The 4B-parameter LOM achieves 89.47% accuracy on their benchmark, outperforming DeepSeek-V3.2 on complex graph reasoning tasks, demonstrating effective fusion of ontology structure and language understanding.

Conclusion: LOM successfully addresses enterprise knowledge integration challenges by combining structured and unstructured data into a unified ontology with enhanced semantic reasoning capabilities through specialized training techniques.

Abstract: Enterprise-scale knowledge management faces significant challenges in integrating multi-source heterogeneous data and enabling effective semantic reasoning. Traditional knowledge graphs often struggle with implicit relationship discovery and lack sufficient semantic understanding for complex question answering. To address these limitations, we introduce a unified construct–align–reason framework, the large ontology model (LOM). We first build a dual-layer enterprise ontology from structured databases and unstructured text, subsequently fusing these sources into a comprehensive enterprise ontology. To enable instruction-aligned reasoning, we propose a unified three-stage training pipeline: ontology instruction fine-tuning to improve structural understanding; text-ontology grounding to strengthen node semantic encoding; and multi-task instruction tuning on ontology-language pairs with curriculum learning to enhance semantic reasoning and generation. We also construct comprehensive training and evaluation datasets covering diverse ontology reasoning tasks. On this benchmark, our 4B-parameter LOM achieves 89.47% accuracy and outperforms DeepSeek-V3.2 on complex graph reasoning, indicating effective fusion of ontology structure and language.

Method: Proposes Reversible Diffusion Decoding (RDD) that detects stagnation as state-dependent failure and enables efficient backtracking to earlier blocks without recomputation via cached model states, using confidence-guided re-masking to selectively reinitialize uncertain tokens while preserving reliable context.

Result: Experiments show RDD improves generation robustness and quality over baselines with minimal computational overhead.

Conclusion: RDD’s reversible formulation allows decoding to recover from early commitment errors while maintaining the parallel efficiency of diffusion-based generation.

Abstract: Diffusion language models enable parallel token generation through block-wise decoding, but their irreversible commitments can lead to stagnation, where the reverse diffusion process fails to make further progress under a suboptimal context.We propose Reversible Diffusion Decoding (RDD), a decoding framework that introduces reversibility into block-wise diffusion generation. RDD detects stagnation as a state-dependent failure of the reverse process and enables efficient backtracking to earlier blocks without recomputation via cached model states. To avoid repeated failure trajectories, RDD applies confidence-guided re-masking to selectively reinitialize uncertain tokens while preserving reliable context.This reversible formulation allows decoding to recover from early commitment errors while maintaining the parallel efficiency of diffusion-based generation. Experiments show that RDD improves generation robustness and quality over baselines with minimal computational overhead.

Method: Proposes DIVERGE framework with reflection-guided generation and memory-augmented iterative refinement to promote diverse viewpoints. Introduces novel metrics for evaluating diversity-quality trade-off in open-ended questions.

Result: DIVERGE achieves the best diversity-quality trade-off compared to competitive baselines and previous state-of-the-art methods on the real-world Infinity-Chat dataset, substantially improving diversity while maintaining quality.

Conclusion: Reveals systematic limitation of current LLM-based systems for open-ended information-seeking and shows that explicitly modeling diversity can mitigate it. The framework addresses underutilization of retrieved context diversity in standard RAG systems.

Abstract: Existing retrieval-augmented generation (RAG) systems are primarily designed under the assumption that each query has a single correct answer. This overlooks common information-seeking scenarios with multiple plausible answers, where diversity is essential to avoid collapsing to a single dominant response, thereby constraining creativity and compromising fair and inclusive information access. Our analysis reveals a commonly overlooked limitation of standard RAG systems: they underutilize retrieved context diversity, such that increasing retrieval diversity alone does not yield diverse generations. To address this limitation, we propose DIVERGE, a plug-and-play agentic RAG framework with novel reflection-guided generation and memory-augmented iterative refinement, which promotes diverse viewpoints while preserving answer quality. We introduce novel metrics tailored to evaluating the diversity-quality trade-off in open-ended questions, and show that they correlate well with human judgments. We demonstrate that DIVERGE achieves the best diversity-quality trade-off compared to competitive baselines and previous state-of-the-art methods on the real-world Infinity-Chat dataset, substantially improving diversity while maintaining quality. More broadly, our results reveal a systematic limitation of current LLM-based systems for open-ended information-seeking and show that explicitly modeling diversity can mitigate it. Our code is available at: https://github.com/au-clan/Diverge

Method: Created large-scale benchmark for evaluating UQ metrics in reasoning-demanding QA, studying calibration of UQ methods. Analyzed 20 LLMs (base, instruction-tuned, reasoning variants) across 7 scientific QA datasets with 685,000 long-form responses. Evaluated representative UQ approaches at token and sequence levels.

Result: Instruction tuning causes probability mass polarization, reducing reliability of token-level confidences. Reasoning fine-tuning mitigates this effect depending on provider. At sequence level, verbalized approaches are systematically biased and poorly correlated with correctness, while answer frequency (consistency across samples) yields best calibration. ECE alone is misleading for judging UQ performance.

Conclusion: Current UQ methods for LLMs have critical limitations, and standard benchmarking practices are inadequate. Answer frequency provides most reliable calibration, while instruction tuning negatively impacts uncertainty estimation reliability.

Abstract: Large Language Models (LLMs) are commonly used in Question Answering (QA) settings, increasingly in the natural sciences if not science at large. Reliable Uncertainty Quantification (UQ) is critical for the trustworthy uptake of generated answers. Existing UQ approaches remain weakly validated in scientific QA, a domain relying on fact-retrieval and reasoning capabilities. We introduce the first large-scale benchmark for evaluating UQ metrics in reasoning-demanding QA studying calibration of UQ methods, providing an extensible open-source framework to reproducibly assess calibration. Our study spans up to 20 large language models of base, instruction-tuned and reasoning variants. Our analysis covers seven scientific QA datasets, including both multiple-choice and arithmetic question answering tasks, using prompting to emulate an open question answering setting. We evaluate and compare methods representative of prominent approaches on a total of 685,000 long-form responses, spanning different reasoning complexities representative of domain-specific tasks. At the token level, we find that instruction tuning induces strong probability mass polarization, reducing the reliability of token-level confidences as estimates of uncertainty. Models further fine-tuned for reasoning are exposed to the same effect, but the reasoning process appears to mitigate it depending on the provider. At the sequence level, we show that verbalized approaches are systematically biased and poorly correlated with correctness, while answer frequency (consistency across samples) yields the most reliable calibration. In the wake of our analysis, we study and report the misleading effect of relying exclusively on ECE as a sole measure for judging performance of UQ methods on benchmark datasets. Our findings expose critical limitations of current UQ methods for LLMs and standard practices in benchmarking thereof.

Method: Created dataset to evaluate faithfulness under biased cases; proposed Bias Alignment through Logit Recalibration (BALOR) which contrasts output logits from unpatched prompts (capturing bias) with logits from patched contextual information, then recalibrates distribution.

Result: Found 18.84% average faithfulness decrease due to bias; BALOR consistently outperforms baselines across multiple LLMs with up to 33% relative performance improvement.

Conclusion: LLMs’ linguistic biases significantly impact faithfulness in representation interpretation; BALOR effectively mitigates this issue by recalibrating logits to prioritize contextual information over inherent biases.

Abstract: Large Language Models (LLMs) have demonstrated strong capabilities for hidden representation interpretation through Patchscopes, a framework that uses LLMs themselves to generate human-readable explanations by decoding from internal hidden representations. However, our work shows that LLMs tend to rely on inherent linguistic patterns, which can override contextual information encoded in the hidden representations during decoding. For example, even when a hidden representation encodes the contextual attribute “purple” for “broccoli”, LLMs still generate “green” in their explanations, reflecting a strong prior association. This behavior reveals a systematic unfaithfulness in Patchscopes. To systematically study this issue, we first designed a dataset to evaluate the faithfulness of Patchscopes under biased cases, and our results show that there is an 18.84% faithfulness decrease on average. We then propose Bias Alignment through Logit Recalibration (BALOR), which treats the output logits from an unpatched prompt as capturing model bias and contrasts them with logits obtained under patched contextual information. By recalibrating the logit distribution through this contrast, BALOR suppresses model bias and amplifies contextual information during generation. Experiments across multiple LLMs demonstrate that BALOR consistently outperforms existing baselines, achieving up to 33% relative performance improvement.

Method: Kanade uses a single-layer disentangled speech tokenizer that separates out acoustic constants to create a single stream of tokens capturing rich phonetics and prosody, without needing auxiliary methods that existing disentangled codecs often rely on.

Result: Kanade achieves state-of-the-art speaker disentanglement and lexical availability while maintaining excellent reconstruction quality.

Conclusion: Kanade presents an effective single-layer disentangled speech tokenizer that successfully extracts linguistic information while suppressing non-linguistic features like speaker identity, enabling better speech modeling and synthesis.

Abstract: A good language model starts with a good tokenizer. Tokenization is especially important for speech modeling, which must handle continuous signals that mix linguistic and non-linguistic information. A speech tokenizer should extract phonetics and prosody, suppress linguistically irrelevant information like speaker identity, and enable high-quality synthesis. We present Kanade, a single-layer disentangled speech tokenizer that realizes this ideal. Kanade separates out acoustic constants to create a single stream of tokens that captures rich phonetics and prosody. It does so without the need for auxiliary methods that existing disentangled codecs often rely on. Experiments show that Kanade achieves state-of-the-art speaker disentanglement and lexical availability, while maintaining excellent reconstruction quality.

Method: Two-stage approach: 1) QA model identifies opening/closing segments containing metadata, 2) Transformer models (BERTimbau, XLM-RoBERTa with/without CRF layer) perform fine-grained entity extraction enhanced by deslexicalization. Benchmarked against Phi and Gemini LLMs.

Result: Strong in-domain performance outperforming larger general-purpose LLMs, but reduced generalization across municipalities due to variability and linguistic complexity. Established first benchmark for this domain.

Conclusion: Proposed pipeline provides solid foundation for metadata extraction from municipal records, though cross-municipality generalization remains challenging due to domain-specific variability.

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

Method: Combines transformer-based text classifier with self-supervised speech representation model using simple late-fusion ensemble on SemEval-2024 Task 3 dataset

Result: Reports baseline setup and empirical results under limited training protocol, highlighting when multimodal fusion improves over unimodal models

Conclusion: Provides transparent baseline for future comparisons in multimodal emotion recognition from conversational audio and text

Abstract: We present a lightweight multimodal baseline for emotion recognition in conversations using the SemEval-2024 Task 3 dataset built from the sitcom Friends. The goal of this report is not to propose a novel state-of-the-art method, but to document an accessible reference implementation that combines (i) a transformer-based text classifier and (ii) a self-supervised speech representation model, with a simple late-fusion ensemble. We report the baseline setup and empirical results obtained under a limited training protocol, highlighting when multimodal fusion improves over unimodal models. This preprint is provided for transparency and to support future, more rigorous comparisons.

Method: Applied a detection model trained on historical peer reviews to analyze later review cycles at International Conference on Learning Representations (ICLR) and Nature Communications (NC), tracking temporal patterns from pre-2022 through 2025.

Result: Minimal AI-generated content detected before 2022, followed by substantial increase through 2025: approximately 20% of ICLR reviews and 12% of Nature Communications reviews classified as AI-generated in 2025. Most pronounced growth in NC occurred between Q3 and Q4 2024.

Conclusion: Evidence suggests rapidly increasing presence of AI-assisted content in peer review, highlighting need for further study of implications for scholarly evaluation processes.

Abstract: The growing availability of large language models (LLMs) has raised questions about their role in academic peer review. This study examines the temporal emergence of AI-generated content in peer reviews by applying a detection model trained on historical reviews to later review cycles at International Conference on Learning Representations (ICLR) and Nature Communications (NC). We observe minimal detection of AI-generated content before 2022, followed by a substantial increase through 2025, with approximately 20% of ICLR reviews and 12% of Nature Communications reviews classified as AI-generated in 2025. The most pronounced growth of AI-generated reviews in NC occurs between the third and fourth quarter of 2024. Together, these findings provide suggestive evidence of a rapidly increasing presence of AI-assisted content in peer review and highlight the need for further study of its implications for scholarly evaluation.

Method: Constructed BiasInEar dataset (70.8 hours, 11,200 questions) based on Global MMLU Lite spanning English, Chinese, Korean balanced by gender and accent. Evaluated nine representative models using four complementary metrics (accuracy, entropy, APES, Fleiss’ κ) under linguistic (language/accent), demographic (gender), and structural (option order) perturbations.

Result: MLLMs are relatively robust to demographic factors (gender) but highly sensitive to language and option order, suggesting speech can amplify existing structural biases. Architectural design and reasoning strategy substantially affect robustness across languages.

Conclusion: Establishes a unified framework for assessing fairness and robustness in speech-integrated LLMs, providing resources for future research on speech bias in multilingual MLLMs.

Abstract: This work presents the first systematic investigation of speech bias in multilingual MLLMs. We construct and release the BiasInEar dataset, a speech-augmented benchmark based on Global MMLU Lite, spanning English, Chinese, and Korean, balanced by gender and accent, and totaling 70.8 hours ($\approx$4,249 minutes) of speech with 11,200 questions. Using four complementary metrics (accuracy, entropy, APES, and Fleiss’ $κ$), we evaluate nine representative models under linguistic (language and accent), demographic (gender), and structural (option order) perturbations. Our findings reveal that MLLMs are relatively robust to demographic factors but highly sensitive to language and option order, suggesting that speech can amplify existing structural biases. Moreover, architectural design and reasoning strategy substantially affect robustness across languages. Overall, this study establishes a unified framework for assessing fairness and robustness in speech-integrated LLMs, bridging the gap between text- and speech-based evaluation. The resources can be found at https://github.com/ntunlplab/BiasInEar.

Method: Introduces DETOUR benchmark with 1,011 prompts using a dual-agent design: Primary Agent (subject of evaluation) queries a Memory Agent to identify recollected entities across text, image, audio, and video modalities.

Result: Current state-of-the-art models achieve only 36% accuracy on the full multimodal benchmark, demonstrating significant challenges in underspecified search scenarios.

Conclusion: The benchmark reveals substantial gaps in current models’ ability to handle realistic tip-of-the-tongue search across multiple modalities, highlighting the need for improved capabilities in underspecified scenarios.

Abstract: When recalling information in conversation, people often arrive at the recollection after multiple turns. However, existing benchmarks for evaluating agent capabilities in such tip-of-the-tongue search processes are restricted to single-turn settings. To more realistically simulate tip-of-the-tongue search, we introduce Dual-agent based Evaluation Through Obscure Under-specified Retrieval (DETOUR), a dual-agent evaluation benchmark containing 1,011 prompts. The benchmark design involves a Primary Agent, which is the subject of evaluation, tasked with identifying the recollected entity through querying a Memory Agent that is held consistent across evaluations. Our results indicate that current state-of-the-art models still struggle with our benchmark, only achieving 36% accuracy when evaluated on all modalities (text, image, audio, and video), highlighting the importance of enhancing capabilities in underspecified scenarios.

Method: DecompressionLM uses Van der Corput low-discrepancy sequences with arithmetic decoding to enable deterministic, embarrassingly parallel generation without shared state across sequences. It’s a stateless framework for zero-shot concept graph extraction that doesn’t require pre-specified queries.

Result: Across two model families and five quantization variants, activation-aware quantization (AWQ-4bit) expands concept coverage by 30-170%, while uniform quantization (GPTQ-Int4) induces 71-86% coverage collapse. Corpus-based verification reveals a 17-point hallucination gap between top- and bottom-ranked MMLU-Pro Law models.

Conclusion: DecompressionLM establishes concept coverage as a complementary evaluation dimension for assessing knowledge breadth and factual grounding in compressed models, useful for their deployment. The method reveals divergent behaviors in quantized models not reliably reflected by explanation-level perplexity.

Abstract: Existing knowledge probing methods rely on pre-defined queries, limiting extraction to known concepts. We introduce DecompressionLM, a stateless framework for zero-shot concept graph extraction that discovers what language models encode without pre-specified queries or shared cross-sequence state. Our method targets three limitations of common decoding-based probing approaches: cross-sequence coupling that concentrates probability mass on high-frequency prefixes, competitive decoding effects that suppress long-tail concepts, and scalability constraints arising from sequential exploration. Using Van der Corput low-discrepancy sequences with arithmetic decoding, DecompressionLM enables deterministic, embarrassingly parallel generation without shared state across sequences. Across two model families and five quantization variants, we find that activation-aware quantization (AWQ-4bit) expands concept coverage by 30-170%, while uniform quantization (GPTQ-Int4) induces 71-86% coverage collapse – divergent behaviors not reliably reflected by explanation-level perplexity. Corpus-based verification further reveals a 17-point hallucination gap between top- and bottom-ranked MMLU-Pro Law models. DecompressionLM establishes concept coverage as a complementary evaluation dimension for assessing knowledge breadth and factual grounding in compressed models useful for their deployment.

Method: Constructed a balanced set of 100 Turkish sentences with local vs. non-local antecedents for reflexives. Tested two systems: OpenAI chain-of-thought model (o1 Mini) and Trendyol-LLM-7B-base-v0.1 (LLaMA-2-derived, fine-tuned on Turkish). Used combined sentence-level perplexity and forced-choice paradigm to assess antecedent choices.

Result: Trendyol-LLM favored local bindings in ~70% of trials, showing strong locality bias. OpenAI o1 Mini distributed choices almost evenly between local and long-distance readings, revealing marked contrast in binding behavior between the two systems.

Conclusion: Different LLMs exhibit distinct binding behaviors for Turkish reflexives, with fine-tuned models showing stronger locality bias while reasoning-focused models show more balanced patterns, indicating varying linguistic competence across architectures.

Abstract: This study evaluates whether state-of-the-art large language models capture the binding relations of Turkish reflexive pronouns. We construct a balanced set of 100 sentences that pit local against non-local antecedents for the reflexives kendi and kendisi, and test two contrasting systems: an OpenAI chain-of-thought model designed for multi-step reasoning and Trendyol-LLM-7B-base-v0.1, a LLaMA-2-derived model extensively fine-tuned on Turkish data. Antecedent choice is assessed using a combined sentence-level perplexity and forced-choice paradigm. Trendyol-LLM favours local bindings in approximately 70% of trials, exhibiting a strong locality bias, whereas o1 Mini distributes its choices almost evenly between local and long-distance readings, revealing a marked contrast in binding behaviour across the two systems.

Method: Uses integrated gradient attribution to quantify token influence on final answers, aggregating into two segment-level metrics: attribution strength (overall magnitude) and direction consistency (uniformity of attribution signs). Proposes selective SFT framework that identifies important segments with high attribution strength but moderate consistency (indicating reflective reasoning), then applies selective SFT on these segments while masking loss for unimportant ones.

Result: Experiments across multiple models and datasets show improved accuracy and output efficiency, enabling more effective learning from long reasoning traces.

Conclusion: The selective learning framework based on attribution analysis effectively identifies and focuses on important reasoning segments, improving model performance and efficiency while reducing redundancy in chain-of-thought reasoning.

Abstract: Large Reasoning Models (LRMs) achieve strong reasoning performance by generating long chains of thought (CoTs), yet only a small fraction of these traces meaningfully contributes to answer prediction, while the majority contains repetitive or truncated content. Such output redundancy is further propagated after supervised finetuning (SFT), as models learn to imitate verbose but uninformative patterns, which can degrade performance. To this end, we incorporate integrated gradient attribution to quantify each token’s influence on final answers and aggregate them into two segment-level metrics: (1) \textit{attribution strength} measures the overall attribution magnitude; and (2) \textit{direction consistency} captures whether tokens’ attributions within a segment are uniformly positive or negative (high consistency), or a mixture of both (moderate consistency). Based on these two metrics, we propose a segment-level selective learning framework to identify important segments with high attribution strength but moderate consistency that indicate reflective rather than shallow reasoning. The framework then applies selective SFT on these important segments while masking loss for unimportant ones. Experiments across multiple models and datasets show that our approach improves accuracy and output efficiency, enabling more effective learning from long reasoning traces~\footnote{Code and data are available at https://github.com/SiyuanWangw/SegmentSelectiveSFT}.

Method: Proposed MANBENCH benchmark with 4 task types susceptible to Mandela effect and 5 interaction protocols; evaluated multiple LLMs; proposed mitigation strategies including prompt-level defenses (cognitive anchoring, source scrutiny) and model-level alignment-based defense

Result: Quantified Mandela effect in LLM-based multi-agent systems; achieved 74.40% average reduction in Mandela effect compared to baseline using proposed mitigation strategies

Conclusion: Findings provide insights for developing more resilient and ethically aligned collaborative multi-agent systems by addressing collective memory biases

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.

Method: Combined behavioral and computational analyses: generating length-controlled summaries to derive empirical importance distributions, analyzing attention heads alignment with importance patterns, and examining layer-wise importance prediction.

Result: LLMs converge on consistent importance patterns different from pre-LLM baselines, cluster more by model family than size, certain attention heads align with importance distributions, and middle-to-late layers strongly predict importance.

Conclusion: The study provides initial insights into what LLMs prioritize in summarization and how this priority is internally represented, opening paths for interpreting and controlling information selection in LLMs.

Abstract: Large Language Models (LLMs) are now state-of-the-art at summarization, yet the internal notion of importance that drives their information selections remains hidden. We propose to investigate this by combining behavioral and computational analyses. Behaviorally, we generate a series of length-controlled summaries for each document and derive empirical importance distributions based on how often each information unit is selected. These reveal that LLMs converge on consistent importance patterns, sharply different from pre-LLM baselines, and that LLMs cluster more by family than by size. Computationally, we identify that certain attention heads align well with empirical importance distributions, and that middle-to-late layers are strongly predictive of importance. Together, these results provide initial insights into what LLMs prioritize in summarization and how this priority is internally represented, opening a path toward interpreting and ultimately controlling information selection in these models.

Method: Developed SENSE (Sensorimotor Embedding Norm Scoring Engine), a learned projection model that predicts Lancaster sensorimotor norms from word lexical embeddings. Also conducted a behavioral study with 281 participants who selected which candidate nonce words evoked specific sensorimotor associations.

Result: Found statistically significant correlations between human selection rates and SENSE ratings across 6 of 11 sensorimotor modalities. Sublexical analysis revealed systematic phonosthemic patterns for the interoceptive norm, suggesting computational methods can identify candidate phonesthemes from text data.

Conclusion: SENSE successfully bridges computational linguistics with grounded language understanding by predicting sensorimotor associations from word embeddings, demonstrating that computational models can capture aspects of embodied language understanding.

Abstract: While word embeddings derive meaning from co-occurrence patterns, human language understanding is grounded in sensory and motor experience. We present $\text{SENSE}$ $(\textbf{S}\text{ensorimotor }$ $\textbf{E}\text{mbedding }$ $\textbf{N}\text{orm }$ $\textbf{S}\text{coring }$ $\textbf{E}\text{ngine})$, a learned projection model that predicts Lancaster sensorimotor norms from word lexical embeddings. We also conducted a behavioral study where 281 participants selected which among candidate nonce words evoked specific sensorimotor associations, finding statistically significant correlations between human selection rates and $\text{SENSE}$ ratings across 6 of the 11 modalities. Sublexical analysis of these nonce words selection rates revealed systematic phonosthemic patterns for the interoceptive norm, suggesting a path towards computationally proposing candidate phonosthemes from text data.

Method: Proposes Intention-Tuning, an intention-adaptive layer-wise fine-tuning framework that dynamically selects a subset of LLM layers to learn writer intentions, then transfers these representations to revision generation.

Result: Experimental results show Intention-Tuning is effective and efficient on small revision corpora, outperforming several Parameter-Efficient Fine-Tuning (PEFT) baselines.

Conclusion: The framework addresses the challenge of intention-based revision generation with limited data by adaptively learning intentions through selective layer tuning and representation transfer.

Abstract: Large Language Models (LLMs) have achieved impressive capabilities in various context-based text generation tasks, such as summarization and reasoning; however, their applications in intention-based generation tasks remain underexplored. One such example is revision generation, which requires the generated text to explicitly reflect the writer’s actual intentions. Identifying intentions and generating desirable revisions are challenging due to their complex and diverse nature. Although prior work has employed LLMs to generate revisions with few-shot learning, they struggle with handling entangled multi-intent scenarios. While fine-tuning LLMs using intention-based instructions appears promising, it demands large amounts of annotated data, which is expensive and scarce in the revision community. To address these challenges, we propose Intention-Tuning, an intention-adaptive layer-wise LLM fine-tuning framework that dynamically selects a subset of LLM layers to learn the intentions and subsequently transfers their representations to revision generation. Experimental results suggest that Intention-Tuning is effective and efficient on small revision corpora, outperforming several PEFT baselines.

Method: Created a large-scale multilingual dataset (280,210 instances) spanning 15 public health domains and 17 languages, stratified into three difficulty levels. Used LLM-assisted construction pipeline with retrieval, duplication checks, evidence grounding, and label validation. Developed a domain-aligned evaluator distilled from LLM judgments across six dimensions.

Result: Built GlobalHealthAtlas dataset enabling reproducible training and evaluation of LLMs for safety-critical public health reasoning beyond conventional QA benchmarks.

Conclusion: The contributions enable training and evaluation of LLMs for safety-critical public health reasoning, addressing the gap in structured ML approaches for this domain with proper benchmarks and evaluation frameworks.

Abstract: Public health reasoning requires population level inference grounded in scientific evidence, expert consensus, and safety constraints. However, it remains underexplored as a structured machine learning problem with limited supervised signals and benchmarks. We introduce \textbf{GlobalHealthAtlas}, a large scale multilingual dataset of 280,210 instances spanning 15 public health domains and 17 languages, stratified into three difficulty levels from health literacy to epidemiological and policy reasoning. Instances are derived from openly available public health sources and labeled by language, domain, and difficulty to support supervised learning and slice based evaluation. We further propose large language model (LLM) assisted construction and quality control pipeline with retrieval, duplication, evidence grounding checks, and label validation to improve consistency at scale. Finally, we present a domain aligned evaluator distilled from high confidence judgments of diverse LLMs to assess outputs along six dimensions: Accuracy, Reasoning, Completeness, Consensus Alignment, Terminology Norms, and Insightfulness. Together, these contributions enable reproducible training and evaluation of LLMs for safety critical public health reasoning beyond conventional QA benchmarks.

Method: The paper synthesizes existing evidence on multilingual model behavior, data asymmetries, and sociotechnical harm, drawing on cross-cultural perspectives in human-centered computing and AI governance to articulate a conceptual governance framework.

Result: Identifies three key governance challenges: cultural/linguistic inequities in training data and evaluation, misalignment between global deployment and local norms/values/power structures, and limited accountability mechanisms for marginalized language communities.

Conclusion: Proposes a conceptual agenda reframing multilingual AI governance as a sociocultural and rights-based problem, outlining design and policy implications for data stewardship, transparency, and participatory accountability to prevent reproduction of global inequalities.

Abstract: Multilingual large language models (MLLMs) are increasingly deployed across cultural, linguistic, and political contexts, yet existing governance frameworks largely assume English-centric data, homogeneous user populations, and abstract notions of fairness. This creates systematic risks for low-resource languages and culturally marginalized communities, where data practices, model behavior, and accountability mechanisms often fail to align with local norms, rights, and expectations. Drawing on cross-cultural perspectives in human-centered computing and AI governance, this paper synthesizes existing evidence on multilingual model behavior, data asymmetries, and sociotechnical harm, and articulates a culturally grounded governance framework for MLLMs. We identify three interrelated governance challenges: cultural and linguistic inequities in training data and evaluation practices, misalignment between global deployment and locally situated norms, values, and power structures, and limited accountability mechanisms for addressing harms experienced by marginalized language communities. Rather than proposing new technical benchmarks, we contribute a conceptual agenda that reframes multilingual AI governance as a sociocultural and rights based problem. We outline design and policy implications for data stewardship, transparency, and participatory accountability, and argue that culturally grounded governance is essential for ensuring that multilingual language models do not reproduce existing global inequalities under the guise of scale and neutrality.

Method: Introduces a commented, step-by-step code-generation framework that incorporates explicit reasoning into Python program generation. Decomposes TableQA reasoning into multi-line executable programs with concise natural language comments to promote clearer reasoning.

Result: Achieves 70.9% accuracy on WikiTableQuestions benchmark using Qwen2.5-Coder-7B-Instruct, surpassing Repanda baseline (67.6%). Combined with robust end-to-end TableQA model via lightweight answer-selection mechanism achieves up to 84.3% accuracy.

Conclusion: The commented step-by-step code generation framework improves TableQA performance by making reasoning more explicit and interpretable, and can be effectively combined with existing models for further accuracy gains.

Abstract: Table Question Answering (TableQA) poses a significant challenge for large language models (LLMs) because conventional linearization of tables often disrupts the two-dimensional relationships intrinsic to structured data. Existing methods, which depend on end-to-end answer generation or single-line program queries, typically exhibit limited numerical accuracy and reduced interpretability. This work introduces a commented, step-by-step code-generation framework that incorporates explicit reasoning into the Python program-generation process. The approach decomposes TableQA reasoning into multi-line executable programs with concise natural language comments, thereby promoting clearer reasoning and increasing the likelihood of generating correct code. On the WikiTableQuestions benchmark, the proposed method achieves 70.9% accuracy using Qwen2.5-Coder-7B-Instruct, surpassing the Repanda baseline (67.6%). Integrating the proposed framework with a robust end-to-end TableQA model via a lightweight answer-selection mechanism yields further improvements. This combined approach achieves up to 84.3% accuracy on the WikiTableQuestions benchmark.

Method: Multi-dimensional analytical framework combining MDS and t-SNE for dimensionality reduction, GDV for cluster separation metrics, FDR for linear diagnostic probing, and attention mechanism analysis.

Result: Reveals hierarchical representational structure: construction-specific information emerges in early layers, forms maximally separable clusters in middle layers, and is maintained through later processing stages.

Conclusion: BERT develops hierarchical representations of argument structure constructions with optimal separation in middle layers, suggesting systematic processing of linguistic constructions.

Abstract: This study investigates how the Bidirectional Encoder Representations from Transformers model processes four fundamental Argument Structure Constructions. We employ a multi-dimensional analytical framework, which integrates MDS, t-SNE as dimensionality reduction, Generalized Discrimination Value (GDV) as cluster separation metrics, Fisher Discriminant Ratio (FDR) as linear diagnostic probing, and attention mechanism analysis. Our results reveal a hierarchical representational structure. Construction-specific information emerges in early layers, forms maximally separable clusters in middle layers, and is maintained through later processing stages.

Method: Fine-tuned three SpeechLLMs using LoRA adapters to induce specific MCQA behaviors (preference for stereotypical, anti-stereotypical, or neutral answers), then evaluated generalization to another MCQA benchmark and long-form creative generation tasks.

Result: Performance on MCQA bias benchmarks fails to reliably predict performance across other MCQA benchmarks and long-form tasks, showing limited evidence of cross-task generalization in the speech domain.

Conclusion: Current MCQA bias benchmarks have limited generalizability and don’t reliably predict model behavior in more realistic settings; authors propose an evaluation suite for measuring behavior transferability in future models.

Abstract: Recent work in benchmarking bias and fairness in speech large language models (SpeechLLMs) has relied heavily on multiple-choice question answering (MCQA) formats. The model is tasked to choose between stereotypical, anti-stereotypical, or neutral/irrelevant answers given an input speech prompt and an optional text prompt. Such MCQA benchmarks implicitly assume that model performance is consistent across other MCQA tasks, voices, and other task formats such as more realistic, long-form evaluations. In this paper, we probe that assumption. We fine-tune three SpeechLLMs using LoRA adapters to induce specific MCQA behaviours: preference for stereotypical, anti-stereotypical, or neutral/uncertain answers. We then evaluate whether these behaviours generalise to another, distinct MCQA benchmark, and more critically to long-form, creative generation tasks. Our results show that performance on MCQA bias benchmarks fails to reliably predict performances across other MCQA benchmarks, and more importantly across long-form tasks. We conclude that current MCQA bias benchmarks show limited evidence of cross-task generalisation in the speech domain, and also propose an evaluation suite for measuring behaviour transferability in future models and benchmarks.

Method: Used Latent Dirichlet Allocation (LDA) for topic modeling on French drama texts from 1700-1900, then applied Jensen-Shannon Divergence to measure topic distribution changes over time.

Result: Found profound changes in topical distribution after the French Revolution (1789-1850), with bourgeois themes becoming prevalent from late 18th century onward. Topic prevalence patterns showed coevolution with French GDP trends.

Conclusion: French drama’s thematic evolution closely tracked political and economic transformations, particularly the rise of bourgeois themes reflecting societal changes during industrialization and post-revolutionary periods.

Abstract: This paper investigates the changing nature of French drama between 1700-1900 using Latent Dirichlet Allocation and Jensen-Shannon Divergence. Results indicate that the topical distribution of French drama changed profoundly after the French Revolution, particularly between 1789 and 1850. Bourgeois themes emerged among the most prevalent topics since the late 18th century. To assess the coevolution of drama and economic growth, I plot the yearly prevalence of topics alongside French GDP between 1700-1900, and discuss these changes in light of the political and economic changes prompted by the French Revolution and the industrialization of the country.

Method: Proposes enhanced LLM-based ASR framework with fine-tuned Whisper and mHuBERT encoders, evaluates E2E Whisper models with LoRA and full fine-tuning, and introduces cross-attention-based fusion mechanisms for parallel-speech-encoder architecture.

Result: Achieved CER/WER of 10.69% on MLC-SLM Challenge evaluation set, ranking on par with top systems despite using only 1,500 hours of training data vs. competitors’ large-scale training sets. However, final LLM-based ASR still underperforms compared to fine-tuned E2E Whisper model.

Conclusion: The enhanced LLM-based ASR framework shows competitive performance but reveals that current LLM-based approaches still cannot match fine-tuned E2E models, providing valuable empirical guidance for future Speech-LLM design.

Abstract: The INTERSPEECH 2025 Challenge on Multilingual Conversational Speech Language Models (MLC-SLM) promotes multilingual conversational ASR with large language models (LLMs). Our previous SHNU-mASR system adopted a competitive parallel-speech-encoder architecture that integrated Whisper and mHuBERT with an LLM. However, it faced two challenges: simple feature concatenation may not fully exploit complementary information, and the performance gap between LLM-based ASR and end-to-end(E2E) encoder-decoder ASR remained unexplored. In this work, we present an enhanced LLM-based ASR framework that combines fine-tuned Whisper and mHuBERT encoders with an LLM to enrich speech representations. We first evaluate E2E Whisper models with LoRA and full fine-tuning on the MLC-SLM ASR task, and then propose cross-attention-based fusion mechanisms for the parallel-speech-encoder. On the official evaluation set of the MLC-SLM Challenge, our system achieves a CER/WER of 10.69%, ranking on par with the top-ranked Track 1 systems, even though it uses only 1,500 hours of baseline training data compared with their large-scale training sets. Nonetheless, we find that our final LLM-based ASR still does not match the performance of a fine-tuned E2E Whisper model, providing valuable empirical guidance for future Speech-LLM design. Our code is publicly available at https://github.com/1535176727/MLC-SLM.

Method: Hermes framework integrates three modules: 1) Speaker Diarization to handle multiple speakers, 2) Terminology Identification for domain-specific terms, and 3) Expressiveness Enhancement to maintain natural language flow and emotional tone in translations.

Result: Hermes achieves state-of-the-art diarization performance and generates expressive, contextually coherent translations, advancing research in automated interlingual subtitling.

Conclusion: The Hermes framework successfully addresses key challenges in interlingual subtitling through specialized modules, demonstrating the potential of LLM-based approaches for visual media localization.

Abstract: Interlingual subtitling, which translates subtitles of visual media into a target language, is essential for entertainment localization but has not yet been explored in machine translation. Although Large Language Models (LLMs) have significantly advanced the general capabilities of machine translation, the distinctive characteristics of subtitle texts pose persistent challenges in interlingual subtitling, particularly regarding semantic coherence, pronoun and terminology translation, and translation expressiveness. To address these issues, we present Hermes, an LLM-based automated subtitling framework. Hermes integrates three modules: Speaker Diarization, Terminology Identification, and Expressiveness Enhancement, which effectively tackle the above challenges. Experiments demonstrate that Hermes achieves state-of-the-art diarization performance and generates expressive, contextually coherent translations, thereby advancing research in interlingual subtitling.

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

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

Conclusion: LAVE provides an effective constrained decoding approach specifically designed for dLLMs that reliably enforces grammatical correctness during generation, addressing key limitations of current methods.

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

Method: Implemented and evaluated various transformers: RoBERTa for English-only detection and multilingual XLM-RoBERTa for both English and German hope speech classification.

Result: RoBERTa achieved weighted f1-score of 0.818 and accuracy of 81.8% for English. XLM-RoBERTa achieved weighted f1-score of 0.786 and accuracy of 78.5% for both languages.

Conclusion: The results demonstrate the importance of pre-trained large language models in enhancing NLP task performance, particularly for multilingual applications like hope speech detection.

Abstract: This paper describes a system that has been submitted to the “PolyHope-M” at RANLP2025. In this work various transformers have been implemented and evaluated for hope speech detection for English and Germany. RoBERTa has been implemented for English, while the multilingual model XLM-RoBERTa has been implemented for both English and German languages. The proposed system using RoBERTa reported a weighted f1-score of 0.818 and an accuracy of 81.8% for English. On the other hand, XLM-RoBERTa achieved a weighted f1-score of 0.786 and an accuracy of 78.5%. These results reflects the importance of improvement of pre-trained large language models and how these models enhancing the performance of different natural language processing tasks.

Method: Proposes connector-sharing strategy based on linguistic family membership - one connector per language family instead of per language. Links frozen speech encoder to pretrained LLM via lightweight connectors grouped by linguistic families.

Result: Family-based connectors reduce parameter count while improving generalization across domains. Validated across two multilingual LLMs and two real-world corpora (curated and crowd-sourced speech).

Conclusion: Linguistic family-based connector sharing offers practical and scalable strategy for multilingual ASR deployment, balancing efficiency with performance.

Abstract: Large Language Model (LLM)-powered Automatic Speech Recognition (ASR) systems achieve strong performance with limited resources by linking a frozen speech encoder to a pretrained LLM via a lightweight connector. Prior work trains a separate connector per language, overlooking linguistic relatedness. We propose an efficient and novel connector-sharing strategy based on linguistic family membership, enabling one connector per family, and empirically validate its effectiveness across two multilingual LLMs and two real-world corpora spanning curated and crowd-sourced speech. Our results show that family-based connectors reduce parameter count while improving generalization across domains, offering a practical and scalable strategy for multilingual ASR deployment.

Method: Proposes a framework modeling safety alignment effects as systematic distortion of pre-alignment distributions. Casts Weak-to-Strong Jailbreaking as a forecast aggregation problem, deriving optimal aggregation strategies characterized by Gradient Shift in loss-induced dual spaces. Shows existing logit-arithmetic methods are special cases under cross-entropy loss, and derives broader family of aggregation rules for other proper losses. Introduces a new hybrid aggregation rule.

Result: Evaluations across red-teaming benchmarks and math utility tasks using frontier models demonstrate superior Attack Success Rates and lower “Jailbreak Tax” compared to existing methods, especially on safety-hardened gpt-oss-120b.

Conclusion: The proposed framework provides a principled approach to understanding and exploiting systematic discrepancies created by safety alignment, leading to more effective jailbreaking methods with better performance on hardened models.

Abstract: Safety alignment in Large Language Models (LLMs) often creates a systematic discrepancy between a model’s aligned output and the underlying pre-aligned data distribution. We propose a framework in which the effect of safety alignment on next-token prediction is modeled as a systematic distortion of a pre-alignment distribution. We cast Weak-to-Strong Jailbreaking as a forecast aggregation problem and derive an optimal aggregation strategy characterized by a Gradient Shift in the loss-induced dual space. We show that logit-arithmetic jailbreaking methods are a special case of this framework under cross-entropy loss, and derive a broader family of aggregation rules corresponding to other proper losses. We also propose a new hybrid aggregation rule. Evaluations across red-teaming benchmarks and math utility tasks using frontier models demonstrate that our approach achieves superior Attack Success Rates and lower “Jailbreak Tax” compared with existing methods, especially on the safety-hardened gpt-oss-120b.

Method: Uses VAE framework with two complementary approaches: (1) Sentence-level learning and control - disentangling and manipulating semantic features in latent space to guide sentence generation using explanatory text; (2) Reasoning-level learning and control - isolating and steering inference behaviors in latent space to control Natural Language Inference, focusing on Explanatory NLI tasks.

Result: Introduces novel theoretical frameworks and practical methodologies that demonstrate enhanced interpretability and controllability of latent spaces for natural language across the thesis.

Conclusion: The work moves toward language models with systematically interpretable, precisely structured, and reliably directed internal semantic representations through geometric manipulation of latent spaces.

Abstract: This thesis advances semantic representation learning to render language representations or models more semantically and geometrically interpretable, and to enable localised, quasi-symbolic, compositional control through deliberate shaping of their latent space geometry. We pursue this goal within a VAE framework, exploring two complementary research directions: (i) Sentence-level learning and control: disentangling and manipulating specific semantic features in the latent space to guide sentence generation, with explanatory text serving as the testbed; and (ii) Reasoning-level learning and control: isolating and steering inference behaviours in the latent space to control NLI. In this direction, we focus on Explanatory NLI tasks, in which two premises (explanations) are provided to infer a conclusion. The overarching objective is to move toward language models whose internal semantic representations can be systematically interpreted, precisely structured, and reliably directed. We introduce a set of novel theoretical frameworks and practical methodologies, together with corresponding experiments, to demonstrate that our approaches enhance both the interpretability and controllability of latent spaces for natural language across the thesis.

Method: Three-phase pipeline: 1) Mid-training with Plasticity-Adjusted Sampling (PAS) for domain adaptation, 2) Supervised fine-tuning with Legal Agentic CoT Distillation (LEAD) to extract reasoning from legal texts, 3) Curriculum Reinforcement Learning (RL) for progressive reasoning development.

Result: LegalOne achieves state-of-the-art performance across various legal tasks, surpassing larger general-purpose LLMs through enhanced knowledge density and efficiency.

Conclusion: LegalOne provides a comprehensive solution for legal AI with trustworthy and interpretable foundation models suitable for high-stakes judicial applications.

Abstract: While Large Language Models (LLMs) have demonstrated impressive general capabilities, their direct application in the legal domain is often hindered by a lack of precise domain knowledge and complexity of performing rigorous multi-step judicial reasoning. To address this gap, we present LegalOne, a family of foundational models specifically tailored for the Chinese legal domain. LegalOne is developed through a comprehensive three-phase pipeline designed to master legal reasoning. First, during mid-training phase, we propose Plasticity-Adjusted Sampling (PAS) to address the challenge of domain adaptation. This perplexity-based scheduler strikes a balance between the acquisition of new knowledge and the retention of original capabilities, effectively establishing a robust legal foundation. Second, during supervised fine-tuning, we employ Legal Agentic CoT Distillation (LEAD) to distill explicit reasoning from raw legal texts. Unlike naive distillation, LEAD utilizes an agentic workflow to convert complex judicial processes into structured reasoning trajectories, thereby enforcing factual grounding and logical rigor. Finally, we implement a Curriculum Reinforcement Learning (RL) strategy. Through a progressive reinforcement process spanning memorization, understanding, and reasoning, LegalOne evolves from simple pattern matching to autonomous and reliable legal reasoning. Experimental results demonstrate that LegalOne achieves state-of-the-art performance across a wide range of legal tasks, surpassing general-purpose LLMs with vastly larger parameter counts through enhanced knowledge density and efficiency. We publicly release the LegalOne weights and the LegalKit evaluation framework to advance the field of Legal AI, paving the way for deploying trustworthy and interpretable foundation models in high-stakes judicial applications.

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

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

Conclusion: SLMs show promise for efficient customer-service QA but current limitations in maintaining dialogue continuity and contextual understanding need to be addressed for practical deployment in resource-constrained environments.

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

Method: Proposes EchoReview, a citation-context-driven data synthesis framework that mines implicit evaluative signals from academic citations and transforms community judgments into structured review-style data. Creates EchoReview-16K dataset and trains EchoReviewer-7B model.

Result: EchoReviewer-7B achieves significant improvements on core review dimensions like evidence support and comprehensiveness, validating citation context as an effective data paradigm for automated peer review.

Conclusion: Citation context provides a robust alternative to human review data for training reliable automated peer review systems, addressing scalability and consistency issues.

Abstract: As the volume of scientific submissions continues to grow rapidly, traditional peer review systems are facing unprecedented scalability pressures, highlighting the urgent need for automated reviewing methods that are both scalable and reliable. Existing supervised fine-tuning approaches based on real review data are fundamentally constrained by single-source of data as well as the inherent subjectivity and inconsistency of human reviews, limiting their ability to support high-quality automated reviewers. To address these issues, we propose EchoReview, a citation-context-driven data synthesis framework that systematically mines implicit collective evaluative signals from academic citations and transforms scientific community’s long-term judgments into structured review-style data. Based on this pipeline, we construct EchoReview-16K, the first large-scale, cross-conference, and cross-year citation-driven review dataset, and train an automated reviewer, EchoReviewer-7B. Experimental results demonstrate that EchoReviewer-7B can achieve significant and stable improvements on core review dimensions such as evidence support and review comprehensiveness, validating citation context as a robust and effective data paradigm for reliable automated peer review.

Method: ExperienceWeaver shifts focus from data retrieval to experience learning by distilling noisy, multi-dimensional feedback into structured knowledge: error-specific Tips and high-level Strategies. This distilled experience is injected into an agentic pipeline, enabling models to learn “how to revise” rather than just “what to revise”.

Result: Extensive evaluations across four clinical datasets show ExperienceWeaver consistently improves performance, surpassing state-of-the-art models like Gemini-3 Pro in small-sample settings.

Conclusion: ExperienceWeaver addresses limitations of current LLM approaches for clinical text improvement by focusing on experience distillation rather than data retrieval, enabling better performance in data-scarce medical contexts.

Abstract: Clinical text improvement is vital for healthcare efficiency but remains difficult due to limited high-quality data and the complex constraints of medical documentation. While Large Language Models (LLMs) show promise, current approaches struggle in small-sample settings: supervised fine-tuning is data-intensive and costly, while retrieval-augmented generation often provides superficial corrections without capturing the reasoning behind revisions. To address these limitations, we propose ExperienceWeaver, a hierarchical framework that shifts the focus from data retrieval to experience learning. Instead of simply recalling past examples, ExperienceWeaver distills noisy, multi-dimensional feedback into structured, actionable knowledge. Specifically, error-specific Tips and high-level Strategies. By injecting this distilled experience into an agentic pipeline, the model learns “how to revise” rather than just “what to revise”. Extensive evaluations across four clinical datasets demonstrate that ExperienceWeaver consistently improves performance, surpassing state-of-the-art models such as Gemini-3 Pro in small-sample settings.

Method: Uses a bidirectional user encoder and discrete prototype codebook to extract multi-dimensional user traits, enabling plug-and-play personalization with only ~20M parameters (0.2% of total model size).

Result: Achieves superior performance and generalization on variant generation tasks compared to strong baselines, with better interpretability and scalability.

Conclusion: CURP provides an efficient solution for user modeling in LLMs that balances personalization quality with computational efficiency through its novel architecture.

Abstract: User modeling characterizes individuals through their preferences and behavioral patterns to enable personalized simulation and generation with Large Language Models (LLMs) in contemporary approaches. However, existing methods, whether prompt-based or training-based methods, face challenges in balancing personalization quality against computational and data efficiency. We propose a novel framework CURP, which employs a bidirectional user encoder and a discrete prototype codebook to extract multi-dimensional user traits. This design enables plug-and-play personalization with a small number of trainable parameters (about 20M parameters, about 0.2% of the total model size). Through extensive experiments on variant generation tasks, we show that CURP achieves superior performance and generalization compared to strong baselines, while offering better interpretability and scalability. The code are available at https://github.com/RaidonWong/CURP_code

Method: Trains component models on candidate datasets at scale, then derives data mixture proxies via weighted model merging instead of training proxy models for each mixture.

Result: DeMix breaks the trade-off between sufficiency, accuracy and efficiency, obtaining optimal mixtures with higher benchmark performance at lower search cost.

Conclusion: DeMix enables evaluation of unlimited sampled mixtures without extra training burden, facilitating better mixture discovery through more search trials.

Abstract: Determining an effective data mixture is a key factor in Large Language Model (LLM) pre-training, where models must balance general competence with proficiency on hard tasks such as math and code. However, identifying an optimal mixture remains an open challenge, as existing approaches either rely on unreliable tiny-scale proxy experiments or require prohibitively expensive large-scale exploration. To address this, we propose Decouple Searching from Training Mix (DeMix), a novel framework that leverages model merging to predict optimal data ratios. Instead of training proxy models for every sampled mixture, DeMix trains component models on candidate datasets at scale and derives data mixture proxies via weighted model merging. This paradigm decouples search from training costs, enabling evaluation of unlimited sampled mixtures without extra training burden and thus facilitating better mixture discovery through more search trials. Extensive experiments demonstrate that DeMix breaks the trade-off between sufficiency, accuracy and efficiency, obtaining the optimal mixture with higher benchmark performance at lower search cost. Additionally, we release the DeMix Corpora, a comprehensive 22T-token dataset comprising high-quality pre-training data with validated mixtures to facilitate open research. Our code and DeMix Corpora is available at https://github.com/Lucius-lsr/DeMix.

Method: Audited Google Search with before: filters, analyzed leakage mechanisms, and tested forecasting accuracy using LLM (gpt-oss-120b) with both leaky and leak-free documents.

Result: 71% of questions returned pages with post-cutoff leakage, 41% directly revealed answers. LLM forecasting with leaky documents showed inflated accuracy (Brier score 0.108 vs 0.242 with clean docs).

Conclusion: Date-restricted search is insufficient for temporal evaluation; stronger retrieval safeguards or evaluation on frozen web snapshots are needed for credible forecasting assessment.

Abstract: Search-engine date filters are widely used to enforce pre-cutoff retrieval in retrospective evaluations of search-augmented forecasters. We show this approach is unreliable: auditing Google Search with a before: filter, 71% of questions return at least one page containing strong post-cutoff leakage, and for 41%, at least one page directly reveals the answer. Using a large language model (LLM), gpt-oss-120b, to forecast with these leaky documents, we demonstrate an inflated prediction accuracy (Brier score 0.108 vs. 0.242 with leak-free documents). We characterize common leakage mechanisms, including updated articles, related-content modules, unreliable metadata/timestamps, and absence-based signals, and argue that date-restricted search is insufficient for temporal evaluation. We recommend stronger retrieval safeguards or evaluation on frozen, time-stamped web snapshots to ensure credible retrospective forecasting.

Method: Two-stage approach: 1) Train a decomposer via RLVR to break complex questions into simpler sub-questions, 2) Use decomposer to annotate training data with sub-questions, then train reasoner under RLVR with sub-question guidance.

Result: Method outperforms competitive baselines, works as plug-and-play module for different RLVR algorithms, and analysis reveals how RLVR affects decomposer performance and what guidance types enhance reasoner exploration/exploitation.

Conclusion: A²D effectively enhances RLVR by providing structured guidance through question decomposition, improving reasoning ability without requiring teacher models.

Abstract: Reinforcement learning with verifiable rewards (RLVR) has shown great potential to enhance the reasoning ability of large language models (LLMs). However, due to the limited amount of information provided during the RLVR process, the model can only engage in largely blind exploration, which often results in failure on challenging problems. To provide additional information for the RLVR process without relying on a teacher model, we propose A$^2$D, an Adaptive Ability Decomposing method for enhancing the effectiveness of RLVR. Specifically, we first train a decomposer via RLVR without distillation, enabling it to decompose complex questions into a set of simpler sub-questions. Next, we use this decomposer to annotate sub-questions for each question in the training dataset, and then train the reasoner under RLVR with sub-question guidance. To better understand A$^2$D, we first compare its performance with competitive baselines, showing its effectiveness. Next, we observe that our method functions as a plug-and-play module that can be applied to different RLVR algorithms. Furthermore, we conduct an analysis of the decomposer, revealing how the RLVR process affects its performance and behavior, and which type of guidance is better suited for enhancing the reasoner’s exploration and exploitation abilities.

Method: APR identifies the Reasoning Anchor (where answer first stabilizes) and penalizes only the Answer-Stable Tail (redundant verification after anchor) using policy optimization with length penalties.

Result: APR achieves performance-efficiency Pareto frontier on 1.5B and 7B models across five mathematical reasoning datasets with significantly reduced RL training resources.

Conclusion: Targeted penalization of structural redundancy in reasoning processes enables more efficient Large Reasoning Models without compromising accuracy.

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

Method: Conducted formative study with 18 Chinese learners and educators to identify difficulties, then developed WordCraft - a learner-centered interactive tool powered by Multimodal Large Language Models (MLLMs) that scaffolds the keyword method through guided keyword selection, association construction, and image formation.

Result: Two user studies demonstrated that WordCraft preserves the generation effect while achieving high levels of effectiveness and usability for vocabulary memorization.

Conclusion: WordCraft successfully addresses limitations of existing approaches by providing process-oriented guidance while maintaining learner engagement, making the keyword method more accessible and effective for Chinese learners of English vocabulary.

Abstract: Applying the keyword method for vocabulary memorization remains a significant challenge for L1 Chinese-L2 English learners. They frequently struggle to generate phonologically appropriate keywords, construct coherent associations, and create vivid mental imagery to aid long-term retention. Existing approaches, including fully automated keyword generation and outcome-oriented mnemonic aids, either compromise learner engagement or lack adequate process-oriented guidance. To address these limitations, we conducted a formative study with L1 Chinese-L2 English learners and educators (N=18), which revealed key difficulties and requirements in applying the keyword method to vocabulary learning. Building on these insights, we introduce WordCraft, a learner-centered interactive tool powered by Multimodal Large Language Models (MLLMs). WordCraft scaffolds the keyword method by guiding learners through keyword selection, association construction, and image formation, thereby enhancing the effectiveness of vocabulary memorization. Two user studies demonstrate that WordCraft not only preserves the generation effect but also achieves high levels of effectiveness and usability.

Method: Proposes novel elicitation method using iterated in-context learning applied to Trust Game from behavioral game theory. Elicits trustworthiness priors from leading LLMs, examines GPT-4.1’s responses to different player personas, and predicts variation using stereotype-based model grounded in perceived warmth and competence.

Result: GPT-4.1’s trustworthiness priors closely track human priors. GPT-4.1 differentiates trust across agent characteristics, and variation in elicited trustworthiness can be predicted by warmth/competence stereotype model.

Conclusion: LLMs can exhibit human-like trustworthiness priors, and trust differentiation follows social psychology stereotypes. This provides a foundation for understanding and calibrating trust in AI systems.

Abstract: One critical aspect of building human-centered, trustworthy artificial intelligence (AI) systems is maintaining calibrated trust: appropriate reliance on AI systems outperforms both overtrust (e.g., automation bias) and undertrust (e.g., disuse). A fundamental challenge, however, is how to characterize the level of trust exhibited by an AI system itself. Here, we propose a novel elicitation method based on iterated in-context learning (Zhu and Griffiths, 2024a) and apply it to elicit trustworthiness priors using the Trust Game from behavioral game theory. The Trust Game is particularly well suited for this purpose because it operationalizes trust as voluntary exposure to risk based on beliefs about another agent, rather than self-reported attitudes. Using our method, we elicit trustworthiness priors from several leading large language models (LLMs) and find that GPT-4.1’s trustworthiness priors closely track those observed in humans. Building on this result, we further examine how GPT-4.1 responds to different player personas in the Trust Game, providing an initial characterization of how such models differentiate trust across agent characteristics. Finally, we show that variation in elicited trustworthiness can be well predicted by a stereotype-based model grounded in perceived warmth and competence.

Method: The authors introduce a representational perspective to analyze internal states across models at different training stages. They conduct comprehensive experiments including comparative analysis, statistical correlation studies, and counterfactual experiments to examine the relationship between internal representations and external outputs.

Result: Post-training yields limited improvement in static initial representation quality. Reasoning involves significant continuous distributional shifts in representations during generation. Post-training helps models drive these transitions toward better distributions for task solving. Generation correctness correlates highly with final representations, and token semantics (not computation or parameter differences) drive the representational transitions.

Conclusion: The paper provides novel insights into reasoning processes and training effects on reasoning enhancement, offering valuable perspectives for future model analysis and optimization through internal representational analysis.

Abstract: Large Language Models have achieved remarkable performance on reasoning tasks, motivating research into how this ability evolves during training. Prior work has primarily analyzed this evolution via explicit generation outcomes, treating the reasoning process as a black box and obscuring internal changes. To address this opacity, we introduce a representational perspective to investigate the dynamics of the model’s internal states. Through comprehensive experiments across models at various training stages, we discover that post-training yields only limited improvement in static initial representation quality. Furthermore, we reveal that, distinct from non-reasoning tasks, reasoning involves a significant continuous distributional shift in representations during generation. Comparative analysis indicates that post-training empowers models to drive this transition toward a better distribution for task solving. To clarify the relationship between internal states and external outputs, statistical analysis confirms a high correlation between generation correctness and the final representations; while counterfactual experiments identify the semantics of the generated tokens, rather than additional computation during inference or intrinsic parameter differences, as the dominant driver of the transition. Collectively, we offer a novel understanding of the reasoning process and the effect of training on reasoning enhancement, providing valuable insights for future model analysis and optimization.

Method: HyLRA uses layer-wise sparsity profiling to identify intra-layer sensitivity (some layers need full attention) and inter-layer similarity (consecutive layers share critical tokens). It employs offline dynamic programming to derive optimal layer-wise policies: sensitive layers use full attention, while tolerant layers reuse top-k indices from preceding layers to bypass quadratic calculations.

Result: HyLRA improves inference throughput by 6%-46% while maintaining comparable performance (<1% accuracy degradation), consistently outperforming state-of-the-art sparse attention methods.

Conclusion: HyLRA effectively overcomes the quadratic bottleneck of dense attention by restricting computation to the most critical tokens through layer-wise optimization, achieving better efficiency-accuracy trade-offs than existing sparse attention approaches.

Abstract: Long-context inference in Large Language Models (LLMs) is bottlenecked by the quadratic computation complexity of attention and the substantial memory footprint of Key-Value (KV) caches. While existing sparse attention mechanisms attempt to mitigate this by exploiting inherent sparsity, they often rely on rigid patterns or aggressive pruning, failing to achieve an optimal balance between efficiency and accuracy. In this paper, we introduce {\bf HyLRA} ({\bf Hy}brid {\bf L}ayer {\bf R}euse {\bf A}ttention), a novel framework driven by layer-wise sparsity profiling. Our empirical analysis uncovers a dual characteristic in attention mechanics: \textit{intra-layer sensitivity}, where specific layers necessitate full attention to prevent feature distortion, and \textit{inter-layer similarity}, where consecutive layers share substantial critical tokens. Based on these observations, HyLRA employs an offline dynamic programming approach to derive an optimal layer-wise policy. This hybrid strategy retains full attention for sensitive layers to ensure robustness, while enabling tolerant layers to bypass quadratic calculations by directly reusing top-$k$ indices from preceding layers. This approach allows LLMs to restrict computation to the most critical tokens, effectively overcoming the quadratic bottleneck of dense attention. Extensive evaluations demonstrate that HyLRA improves inference throughput by 6%–46% while maintaining comparable performance (with $<1%$ accuracy degradation), consistently outperforming state-of-the-art sparse attention methods. HyLRA is open source at \href{https://anonymous.4open.science/r/unified-cache-management-CF80/}{\texttt{/r/unified-cache-management-CF80/}}

Method: Two-stage approach: 1) Create Omni-Preference dataset via automated pipeline synthesizing response pairs from models of different capabilities, using teacher models to reconcile preferences with rubric-grounded rationales; 2) Train Omni-RRM via supervised fine-tuning followed by reinforcement learning (GRPO) to sharpen discrimination.

Result: Achieves SOTA accuracy on video (80.2% on ShareGPT-V) and audio (66.8% on Audio-HH-RLHF) benchmarks, with 17.7% absolute gain over base model on image tasks, and improves downstream performance via Best-of-N selection.

Conclusion: Omni-RRM provides effective, automated reward modeling across modalities without human-labeled preferences, enabling better alignment for multimodal LLMs through structured, rubric-grounded judgments.

Abstract: Multimodal large language models (MLLMs) have shown remarkable capabilities, yet their performance is often capped by the coarse nature of existing alignment techniques. A critical bottleneck remains the lack of effective reward models (RMs): existing RMs are predominantly vision-centric, return opaque scalar scores, and rely on costly human annotations. We introduce \textbf{Omni-RRM}, the first open-source rubric-grounded reward model that produces structured, multi-dimension preference judgments with dimension-wise justifications across \textbf{text, image, video, and audio}. At the core of our approach is \textbf{Omni-Preference}, a large-scale dataset built via a fully automated pipeline: we synthesize candidate response pairs by contrasting models of different capabilities, and use strong teacher models to \emph{reconcile and filter} preferences while providing a modality-aware \emph{rubric-grounded rationale} for each pair. This eliminates the need for human-labeled training preferences. Omni-RRM is trained in two stages: supervised fine-tuning to learn the rubric-grounded outputs, followed by reinforcement learning (GRPO) to sharpen discrimination on difficult, low-contrast pairs. Comprehensive evaluations show that Omni-RRM achieves state-of-the-art accuracy on video (80.2% on ShareGPT-V) and audio (66.8% on Audio-HH-RLHF) benchmarks, and substantially outperforms existing open-source RMs on image tasks, with a 17.7% absolute gain over its base model on overall accuracy. Omni-RRM also improves downstream performance via Best-of-$N$ selection and transfers to text-only preference benchmarks. Our data, code, and models are available at https://anonymous.4open.science/r/Omni-RRM-CC08.

Method: Introduces Factuality-Controlled Generation (FCG) framework where users specify factuality constraints alongside queries. Uses synthetic data to train models on the FCG task, evaluating performance on adherence to factuality constraints and response informativeness.

Result: Synthetic training significantly improves models’ ability to respect factuality requirements while maintaining informativeness in their outputs.

Conclusion: FCG provides a practical framework for controlling the factuality-informativeness trade-off in LLM responses, with synthetic training being an effective approach for improving performance on this task.

Abstract: Large language models (LLMs) encode knowledge with varying degrees of confidence. When responding to queries, models face an inherent trade-off: they can generate responses that are less informative but highly factual, or more informative but potentially less accurate. Different applications demand different balances between informativeness and factuality. We introduce Factuality-Controlled Generation (FCG), a framework that enables users to specify factuality constraints alongside their queries. We propose to evaluate FCG performance on two dimensions: adherence to factuality constraints and response informativeness. We propose to train models on the FCG task using synthetic data, and show that our synthetic training significantly improves models’ ability to both respect factuality requirements and maintain informativeness in their outputs.

Method: Propose a unified framework for studying adversarial robustness in text scoring models (dense retrievers, rerankers, reward models). Adapt attacks and adversarial training methods across different model roles. Introduce multiple adversarial training methods for text scoring models, and show that combining complementary training methods yields strong robustness.

Result: Demonstrate that current adversarial training formulations for language models are often short-sighted and fail to generalize across attacks. Show that the proposed adversarial training methods improve both robustness and task effectiveness. Highlight practical value for RLHF by showing that adversarially trained reward models mitigate reward hacking and support training of better-aligned LLMs.

Conclusion: A unified approach to studying adversarial robustness in text scoring models reveals shared vulnerabilities and enables more effective adversarial training methods that improve both robustness and task performance, with practical applications in RLHF alignment.

Abstract: Research on adversarial robustness in language models is currently fragmented across applications and attacks, obscuring shared vulnerabilities. In this work, we propose unifying the study of adversarial robustness in text scoring models spanning dense retrievers, rerankers, and reward models. This motivates adapting both attacks and adversarial training methods across model roles. Unlike open-ended generation, text scoring failures are directly testable: an attack succeeds when an irrelevant or rejected text outscores a relevant or chosen one. Using this principled lens of text scoring, we demonstrate that current adversarial training formulations for language models are often short-sighted, failing to effectively generalize across attacks. To address this, we introduce multiple adversarial training methods for text scoring models and show that combining complementary training methods can yield strong robustness while also improving task effectiveness. We also highlight the practical value of our approach for RLHF, showing that our adversarially trained reward models mitigate reward hacking and support the training of better-aligned LLMs. We provide our code and models for further study.

Method: Created ILSIC corpus with 500+ Indian statutes covering both laypeople queries and court judgments. Conducted experiments with zero/few-shot inference, retrieval-augmented generation, and supervised fine-tuning to compare performance.

Result: Models trained only on court judgments perform poorly on laypeople queries. Transfer learning from court to laypeople data can help in some scenarios. Performance varies by query category and statute frequency.

Conclusion: Court-trained models don’t generalize well to laypeople queries. The ILSIC corpus enables better LSI research for real-world applications. Transfer learning shows promise but needs careful implementation.

Abstract: Legal Statute Identification (LSI) for a given situation is one of the most fundamental tasks in Legal NLP. This task has traditionally been modeled using facts from court judgments as input queries, due to their abundance. However, in practical settings, the input queries are likely to be informal and asked by laypersons, or non-professionals. While a few laypeople LSI datasets exist, there has been little research to explore the differences between court and laypeople data for LSI. In this work, we create ILSIC, a corpus of laypeople queries covering 500+ statutes from Indian law. Additionally, the corpus also contains court case judgements to enable researchers to effectively compare between court and laypeople data for LSI. We conducted extensive experiments on our corpus, including benchmarking over the laypeople dataset using zero and few-shot inference, retrieval-augmented generation and supervised fine-tuning. We observe that models trained purely on court judgements are ineffective during test on laypeople queries, while transfer learning from court to laypeople data can be beneficial in certain scenarios. We also conducted fine-grained analyses of our results in terms of categories of queries and frequency of statutes.

Method: effGen introduces four major contributions: (1) Enhanced tool-calling with prompt optimization that compresses contexts by 70-80%, (2) Intelligent task decomposition that breaks complex queries into subtasks, (3) Complexity-based routing using five factors for pre-execution decisions, and (4) Unified memory system combining short-term, long-term, and vector-based storage.

Result: Results on 13 benchmarks show effGen outperforms LangChain, AutoGen, and Smolagents with higher success rates, faster execution, and lower memory. Prompt optimization benefits SLMs more (11.2% gain at 1.5B vs 2.4% at 32B), while routing benefits large models more (3.6% at 1.5B vs 7.9% at 32B).

Conclusion: effGen provides an effective, efficient, and secure local deployment solution for SLMs, with complementary scaling behavior between prompt optimization and complexity routing, ensuring consistent gains across all model scales.

Abstract: Most existing language model agentic systems today are built and optimized for large language models (e.g., GPT, Claude, Gemini) via API calls. While powerful, this approach faces several limitations including high token costs and privacy concerns for sensitive applications. We introduce effGen, an open-source agentic framework optimized for small language models (SLMs) that enables effective, efficient, and secure local deployment (pip install effgen). effGen makes four major contributions: (1) Enhanced tool-calling with prompt optimization that compresses contexts by 70-80% while preserving task semantics, (2) Intelligent task decomposition that breaks complex queries into parallel or sequential subtasks based on dependencies, (3) Complexity-based routing using five factors to make smart pre-execution decisions, and (4) Unified memory system combining short-term, long-term, and vector-based storage. Additionally, effGen unifies multiple agent protocols (MCP, A2A, ACP) for cross-protocol communication. Results on 13 benchmarks show effGen outperforms LangChain, AutoGen, and Smolagents with higher success rates, faster execution, and lower memory. Our results reveal that prompt optimization and complexity routing have complementary scaling behavior: optimization benefits SLMs more (11.2% gain at 1.5B vs 2.4% at 32B), while routing benefits large models more (3.6% at 1.5B vs 7.9% at 32B), providing consistent gains across all scales when combined. effGen (https://effgen.org/) is released under the MIT License, ensuring broad accessibility for research and commercial use. Our framework code is publicly available at https://github.com/ctrl-gaurav/effGen.

Method: Compare three approaches: (1) direct supervised transformers, (2) HO→values pipelines with hard hierarchical masks, (3) Presence→HO→values cascades. Also test low-cost add-ons (lexica, short context, topics), label-wise threshold tuning, small LLM baselines (≤10B), QLoRA, and simple ensembles on ValueEval'24/ValuesML dataset (74K English sentences).

Result: HO categories are learnable from single sentences, but hard hierarchical gating reduces end-task Macro-F₁ via error compounding and recall suppression. Label-wise threshold tuning provides substantial gains (+0.05 Macro-F₁), and small transformer ensembles offer consistent additional improvements (+0.02 Macro-F₁). Small LLMs lag behind supervised encoders but contribute complementary errors in cross-family ensembles.

Conclusion: HO structure is useful descriptively but enforcing it with hard gates hurts sentence-level value detection; robust improvements come from calibration (threshold tuning) and lightweight ensembling rather than hierarchical constraints.

Abstract: Sentence-level human value detection is typically framed as multi-label classification over Schwartz values, but it remains unclear whether Schwartz higher-order (HO) categories provide usable structure. We study this under a strict compute-frugal budget (single 8 GB GPU) on ValueEval'24 / ValuesML (74K English sentences). We compare (i) direct supervised transformers, (ii) HO$\rightarrow$values pipelines that enforce the hierarchy with hard masks, and (iii) Presence$\rightarrow$HO$\rightarrow$values cascades, alongside low-cost add-ons (lexica, short context, topics), label-wise threshold tuning, small instruction-tuned LLM baselines ($\le$10B), QLoRA, and simple ensembles. HO categories are learnable from single sentences (e.g., the easiest bipolar pair reaches Macro-$F_1\approx0.58$), but hard hierarchical gating is not a reliable win: it often reduces end-task Macro-$F_1$ via error compounding and recall suppression. In contrast, label-wise threshold tuning is a high-leverage knob (up to $+0.05$ Macro-$F_1$), and small transformer ensembles provide the most consistent additional gains (up to $+0.02$ Macro-$F_1$). Small LLMs lag behind supervised encoders as stand-alone systems, yet can contribute complementary errors in cross-family ensembles. Overall, HO structure is useful descriptively, but enforcing it with hard gates hurts sentence-level value detection; robust improvements come from calibration and lightweight ensembling.

Method: Three-stage approach: (1) Localize language neurons via sparse autoencoders to identify features selective for target languages; (2) Find steering directions via spectral low-rank analysis of activation differences; (3) Apply sparse activation shifts to language neurons to steer model toward target language defaultness.

Result: The method successfully makes Hindi or Spanish primary in LLMs by steering language-specific neurons, demonstrating controllable language defaultness without retraining.

Conclusion: Language defaultness in LLMs is governed by sparse, low-rank control circuits that can be mechanistically isolated and safely steered, enabling targeted language preference adjustments.

Abstract: LLMs are multilingual by training, yet their lingua franca is often English, reflecting English language dominance in pretraining. Other languages remain in parametric memory but are systematically suppressed. We argue that language defaultness is governed by a sparse, low-rank control circuit, language neurons, that can be mechanistically isolated and safely steered. We introduce Neural FOXP2, that makes a chosen language (Hindi or Spanish) primary in a model by steering language-specific neurons. Neural FOXP2 proceeds in three stages: (i) Localize: We train per-layer SAEs so each activation decomposes into a small set of active feature components. For every feature, we quantify English vs. Hindi/Spanish selectivity overall logit-mass lift toward the target-language token set. Tracing the top-ranked features back to their strongest contributing units yields a compact language-neuron set. (ii) Steering directions: We localize controllable language-shift geometry via a spectral low-rank analysis. For each layer, we build English to target activation-difference matrices and perform layerwise SVD to extract the dominant singular directions governing language change. The eigengap and effective-rank spectra identify a compact steering subspace and an empirically chosen intervention window (where these directions are strongest and most stable). (iii) Steer: We apply a signed, sparse activation shift targeted to the language neurons. Concretely, within low to mid layers we add a positive steering along the target-language dominant directions and a compensating negative shift toward the null space for the English neurons, yielding controllable target-language defaultness.

Method: Introduced MixTalk game where sender combines verifiable/unverifiable claims, receiver allocates budget for costly verification. Evaluated state-of-the-art LLM agents in large-scale tournaments across three deployment settings. Proposed Tournament Oracle Policy Distillation (TOPD) method that distills tournament oracle policy from interaction logs for in-context deployment.

Result: Evaluation revealed LLM agents’ strengths and limitations in reasoning about information credibility and the explicit behavior shaping these interactions. TOPD significantly improved receiver robustness to persuasion.

Conclusion: MixTalk provides a framework for studying strategic communication with probabilistic credibility, revealing LLM agents’ capabilities and limitations, with TOPD offering an effective method to improve receiver robustness in such settings.

Abstract: Agents powered by large language models (LLMs) are increasingly deployed in settings where communication shapes high-stakes decisions, making a principled understanding of strategic communication essential. Prior work largely studies either unverifiable cheap-talk or fully verifiable disclosure, failing to capture realistic domains in which information has probabilistic credibility. We introduce MixTalk, a strategic communication game for LLM-to-LLM interaction that models information credibility. In MixTalk, a sender agent strategically combines verifiable and unverifiable claims to communicate private information, while a receiver agent allocates a limited budget to costly verification and infers the underlying state from prior beliefs, claims, and verification outcomes. We evaluate state-of-the-art LLM agents in large-scale tournaments across three realistic deployment settings, revealing their strengths and limitations in reasoning about information credibility and the explicit behavior that shapes these interactions. Finally, we propose Tournament Oracle Policy Distillation (TOPD), an offline method that distills tournament oracle policy from interaction logs and deploys it in-context at inference time. Our results show that TOPD significantly improves receiver robustness to persuasion.

Method: Proposes Structural Confidence framework that extracts multi-scale structural signals from a model’s final-layer hidden-state trajectory using spectral, local-variation, and global shape descriptors to capture internal stability patterns missed by probabilities and embeddings.

Result: Demonstrates strong performance across four heterogeneous benchmarks (FEVER, SciFact, WikiBio-hallucination, TruthfulQA) in terms of AUROC and AUPR, outperforming established baselines while using only a single deterministic forward pass.

Conclusion: Structural Confidence offers a practical, efficient, and robust post-hoc confidence estimation method for socially impactful, resource-constrained LLM applications, avoiding the computational overhead of sampling-based consistency methods.

Abstract: Large language models (LLMs) are increasingly deployed in domains where errors carry high social, scientific, or safety costs. Yet standard confidence estimators, such as token likelihood, semantic similarity and multi-sample consistency, remain brittle under distribution shift, domain-specialised text, and compute limits. In this work, we present Structural Confidence, a single-pass, model-agnostic framework that enhances output correctness prediction based on multi-scale structural signals derived from a model’s final-layer hidden-state trajectory. By combining spectral, local-variation, and global shape descriptors, our method captures internal stability patterns that are missed by probabilities and sentence embeddings. We conduct extensive, cross-domain evaluation across four heterogeneous benchmarks-FEVER (fact verification), SciFact (scientific claims), WikiBio-hallucination (biographical consistency), and TruthfulQA (truthfulness-oriented QA). Our Structural Confidence framework demonstrates strong performance compared with established baselines in terms of AUROC and AUPR. More importantly, unlike sampling-based consistency methods which require multiple stochastic generations and an auxiliary model, our approach uses a single deterministic forward pass, offering a practical basis for efficient, robust post-hoc confidence estimation in socially impactful, resource-constrained LLM applications.

Method: Proposes MedSpeak framework that uses a medical knowledge graph to encode both semantic relationships and phonetic information, combined with LLM reasoning to correct ASR errors in noisy transcripts and improve downstream answer prediction.

Result: Comprehensive experiments show MedSpeak significantly improves medical term recognition accuracy and overall medical SQA performance, establishing it as state-of-the-art for medical SQA.

Conclusion: MedSpeak effectively addresses ASR limitations in medical contexts by leveraging structured medical knowledge and LLM capabilities, providing a robust solution for medical spoken question-answering systems.

Abstract: Spoken question-answering (SQA) systems relying on automatic speech recognition (ASR) often struggle with accurately recognizing medical terminology. To this end, we propose MedSpeak, a novel knowledge graph-aided ASR error correction framework that refines noisy transcripts and improves downstream answer prediction by leveraging both semantic relationships and phonetic information encoded in a medical knowledge graph, together with the reasoning power of LLMs. Comprehensive experimental results on benchmarks demonstrate that MedSpeak significantly improves the accuracy of medical term recognition and overall medical SQA performance, establishing MedSpeak as a state-of-the-art solution for medical SQA. The code is available at https://github.com/RainieLLM/MedSpeak.

Method: DISPO introduces a REINFORCE-style algorithm that decouples up-clipping and down-clipping of importance sampling weights for correct and incorrect responses separately, creating four controllable policy update regimes. This allows independent tuning of how the algorithm handles exploration vs. distillation for correct responses and prevents catastrophic failures for incorrect responses.

Result: DISPO achieves 61.04% on AIME'24 benchmark (vs. 55.42% for CISPO and 50.21% for DAPO), with similar gains across various benchmarks and models. The method maintains exploration-distillation balance while preventing the performance collapse seen in other approaches.

Conclusion: DISPO successfully addresses the stability-efficiency trade-off in RL for LLM reasoning by providing fine-grained control over policy updates through decoupled clipping regimes, enabling both efficient learning and stable performance in mathematical reasoning tasks.

Abstract: Reinforcement learning with verifiable rewards has emerged as a promising paradigm for enhancing the reasoning capabilities of large language models particularly in mathematics. Current approaches in this domain present a clear trade-off: PPO-style methods (e.g., GRPO/DAPO) offer training stability but exhibit slow learning trajectories due to their trust-region constraints on policy updates, while REINFORCE-style approaches (e.g., CISPO) demonstrate improved learning efficiency but suffer from performance instability as they clip importance sampling weights while still permitting non-zero gradients outside the trust-region. To address these limitations, we introduce DISPO, a simple yet effective REINFORCE-style algorithm that decouples the up-clipping and down-clipping of importance sampling weights for correct and incorrect responses, yielding four controllable policy update regimes. Through targeted ablations, we uncover how each regime impacts training: for correct responses, weights >1 increase the average token entropy (i.e., exploration) while weights <1 decrease it (i.e., distillation) – both beneficial but causing gradual performance degradation when excessive. For incorrect responses, overly restrictive clipping triggers sudden performance collapse through repetitive outputs (when weights >1) or vanishing response lengths (when weights <1). By separately tuning these four clipping parameters, DISPO maintains the exploration-distillation balance while preventing catastrophic failures, achieving 61.04% on AIME'24 (vs. 55.42% CISPO and 50.21% DAPO) with similar gains across various benchmarks and models.

Method: Analyzed hidden states of LLMs to identify sparse reward subsystems, conducted intervention experiments on value neurons, examined robustness across datasets/model scales/architectures, and identified dopamine neurons encoding reward prediction errors through divergence analysis between predicted and actual rewards.

Result: Found robust value neurons representing internal state value expectations that are crucial for reasoning, with significant transferability across datasets and models. Identified dopamine neurons that encode reward prediction errors with high activation for positive surprises and low activation for negative surprises.

Conclusion: LLMs develop internal reward subsystems analogous to biological brains, with specialized neurons for value estimation and reward prediction error encoding, suggesting fundamental similarities in reward processing across artificial and biological systems.

Abstract: In this paper, we identify a sparse reward subsystem within the hidden states of Large Language Models (LLMs), drawing an analogy to the biological reward subsystem in the human brain. We demonstrate that this subsystem contains value neurons that represent the model’s internal expectation of state value, and through intervention experiments, we establish the importance of these neurons for reasoning. Our experiments reveal that these value neurons are robust across diverse datasets, model scales, and architectures; furthermore, they exhibit significant transferability across different datasets and models fine-tuned from the same base model. By examining cases where value predictions and actual rewards diverge, we identify dopamine neurons within the reward subsystem which encode reward prediction errors (RPE). These neurons exhibit high activation when the reward is higher than expected and low activation when the reward is lower than expected.

Method: Uses specialized agents (Table, Context, Visual, Summarizing, Verification) that communicate through a shared natural-language log as persistent memory, enabling decentralized collaboration and verification.

Result: Competitive performance on multiple benchmarks: FinQA, TAT-QA, CRT-QA, WikiTableQuestions, FeTaQA, and MultiModalQA. Analysis confirms importance of shared log, agent specialization, and verification.

Conclusion: DeALOG provides a scalable approach through modular components using natural-language communication, enabling robust multimodal QA without central control.

Abstract: Complex question answering across text, tables and images requires integrating diverse information sources. A framework supporting specialized processing with coordination and interpretability is needed. We introduce DeALOG, a decentralized multi-agent framework for multimodal question answering. It uses specialized agents: Table, Context, Visual, Summarizing and Verification, that communicate through a shared natural-language log as persistent memory. This log-based approach enables collaborative error detection and verification without central control, improving robustness. Evaluations on FinQA, TAT-QA, CRT-QA, WikiTableQuestions, FeTaQA, and MultiModalQA show competitive performance. Analysis confirms the importance of the shared log, agent specialization, and verification for accuracy. DeALOG, provides a scalable approach through modular components using natural-language communication.

Method: Formalizes lemma-judging as structured prediction with two-section outputs (precondition check and conclusion-utility check). Uses reinforcement learning with section-aware loss masking to assign penalties to the section responsible for errors. Trained on diverse natural language and formal proof corpora.

Result: Shows consistent in-domain gains over vanilla models and single-label RL baselines, larger improvements on applicability-breaking perturbations, and parity or modest gains on end-to-end tasks. Ablations confirm both two-section outputs and section-aware reinforcement are necessary for robustness.

Conclusion: The RULES framework effectively improves LLMs’ ability to properly judge lemma applicability through structured decomposition and targeted reinforcement learning, addressing a key weakness in mathematical reasoning.

Abstract: Recent large language models (LLMs) perform strongly on mathematical benchmarks yet often misapply lemmas, importing conclusions without validating assumptions. We formalize lemma$-$judging as a structured prediction task: given a statement and a candidate lemma, the model must output a precondition check and a conclusion$-$utility check, from which a usefulness decision is derived. We present RULES, which encodes this specification via a two$-$section output and trains with reinforcement learning plus section$-$aware loss masking to assign penalty to the section responsible for errors. Training and evaluation draw on diverse natural language and formal proof corpora; robustness is assessed with a held$-$out perturbation suite; and end$-$to$-$end evaluation spans competition$-$style, perturbation$-$aligned, and theorem$-$based problems across various LLMs. Results show consistent in$-$domain gains over both a vanilla model and a single$-$label RL baseline, larger improvements on applicability$-$breaking perturbations, and parity or modest gains on end$-$to$-$end tasks; ablations indicate that the two$-$section outputs and section$-$aware reinforcement are both necessary for robustness.

Method: Two-stage curriculum: (1) Progressive Knowledge Distillation aligns byte-level representations with token-trained teacher model embeddings, (2) Byte-Level Supervised Fine-Tuning enables end-to-end generation in byte space. Applied to multiple model families including Llama, Qwen, and OLMo.

Result: Distilled BLMs retain most of teacher models’ performance using only approximately 125B bytes of training data, demonstrating efficient conversion from token-trained to byte-level models.

Conclusion: Proposed distillation recipe provides a cost-effective way to convert existing LLMs into BLMs, enabling byte-level language modeling without expensive retraining from scratch.

Abstract: Byte Language Models (BLMs) have emerged as a promising direction for scaling language models beyond tokenization. However, existing BLMs typically require training from scratch on trillions of bytes, making them prohibitively expensive. In this paper, we propose an efficient distillation recipe that converts existing token-trained LLMs into BLMs while retaining comparable capabilities. Our recipe follows a two-stage curriculum: (1) Progressive Knowledge Distillation, which aligns byte-level representations with the embeddings of the token-trained teacher model; and (2) Byte-Level Supervised Fine-Tuning, which enables end-to-end generation entirely in the byte space. We validate our approach across multiple model families, including Llama, Qwen, and OLMo, and demonstrate that the distilled BLMs retain most of the teacher models’ performance using only approximately 125B bytes.

Method: Used 630 think-aloud utterances from chemistry tutoring problems with student logs; compared LLM-generated reasoning to human utterances under minimal and extended contextual prompting; assessed models’ ability to predict step-level learner success.

Result: GPT-4.1 generates fluent continuations but produces systematically over-coherent, verbose, and less variable reasoning than humans; these effects intensify with richer context; consistently overestimates learner performance.

Conclusion: LLMs have epistemic limitations in simulating learning due to training on expert-like solutions lacking expressions of affect and working memory constraints; evaluation framework can guide future adaptive systems design.

Abstract: Large language models (LLMs) are increasingly embedded in AI-based tutoring systems. Can they faithfully model novice reasoning and metacognitive judgments? Existing evaluations emphasize problem-solving accuracy, overlooking the fragmented and imperfect reasoning that characterizes human learning. We evaluate LLMs as novices using 630 think-aloud utterances from multi-step chemistry tutoring problems with problem-solving logs of student hint use, attempts, and problem context. We compare LLM-generated reasoning to human learner utterances under minimal and extended contextual prompting, and assess the models’ ability to predict step-level learner success. Although GPT-4.1 generates fluent and contextually appropriate continuations, its reasoning is systematically over-coherent, verbose, and less variable than human think-alouds. These effects intensify with a richer problem-solving context during prompting. Learner performance was consistently overestimated. These findings highlight epistemic limitations of simulating learning with LLMs. We attribute these limitations to LLM training data, including expert-like solutions devoid of expressions of affect and working memory constraints during problem solving. Our evaluation framework can guide future design of adaptive systems that more faithfully support novice learning and self-regulation using generative artificial intelligence.

Method: Examined LLM behavior with identical personality prompts across four conversational settings: ice-breaking, negotiation, group decision, and empathy tasks. Analyzed how contextual cues influence personality expression and emotional tone.

Result: Contextual cues systematically influence both personality expression and emotional tone. The same traits are expressed differently depending on social and affective demands, showing context-sensitive rather than fixed personality expression.

Conclusion: LLMs exhibit context-sensitive personality expression that adapts flexibly to social interaction goals and affective conditions, similar to human behavior according to Whole Trait Theory.

Abstract: Large Language Models (LLMs) can be conditioned with explicit personality prompts, yet their behavioral realization often varies depending on context. This study examines how identical personality prompts lead to distinct linguistic, behavioral, and emotional outcomes across four conversational settings: ice-breaking, negotiation, group decision, and empathy tasks. Results show that contextual cues systematically influence both personality expression and emotional tone, suggesting that the same traits are expressed differently depending on social and affective demands. This raises an important question for LLM-based dialogue agents: whether such variations reflect inconsistency or context-sensitive adaptation akin to human behavior. Viewed through the lens of Whole Trait Theory, these findings highlight that LLMs exhibit context-sensitive rather than fixed personality expression, adapting flexibly to social interaction goals and affective conditions.

Method: Introduces knowledge purification concept and proposes five purification methods from various perspectives to consolidate rationales from multiple teacher LLMs into single rationale

Result: Purification methods improve distilled model performance, effectively alleviate knowledge conflicts, and router-based methods show robust generalization capabilities

Conclusion: Knowledge purification techniques optimize multi-teacher distillation and facilitate practical deployment of powerful yet lightweight models

Abstract: Knowledge distillation has emerged as a pivotal technique for transferring knowledge from stronger large language models (LLMs) to smaller, more efficient models. However, traditional distillation approaches face challenges related to knowledge conflicts and high resource demands, particularly when leveraging multiple teacher models. In this paper, we introduce the concept of \textbf{Knowledge Purification}, which consolidates the rationales from multiple teacher LLMs into a single rationale, thereby mitigating conflicts and enhancing efficiency. To investigate the effectiveness of knowledge purification, we further propose five purification methods from various perspectives. Our experiments demonstrate that these methods not only improve the performance of the distilled model but also effectively alleviate knowledge conflicts. Moreover, router-based methods exhibit robust generalization capabilities, underscoring the potential of innovative purification techniques in optimizing multi-teacher distillation and facilitating the practical deployment of powerful yet lightweight models.

Method: 1) Investigated subtitle translation scenarios and literal/liberal translation domains, 2) Verified LLMs as reliable reward models and evaluators for translation, 3) Constructed and released a multidirectional subtitle parallel corpus dataset, 4) Proposed Adaptive Local Preference Optimization (ALPO) method for fine-grained preference alignment in translation.

Result: Experimental results show ALPO achieves outstanding performance in multidimensional evaluation of translation quality, demonstrating effectiveness for expressive subtitle translation.

Conclusion: The paper successfully addresses domain-specific translation needs for visual media subtitles through ALPO and a specialized dataset, showing LLMs can be effectively customized for vertical domains requiring expressive translation.

Abstract: The rapid development of Large Language Models (LLMs) has significantly enhanced the general capabilities of machine translation. However, as application scenarios become more complex, the limitations of LLMs in vertical domain translations are gradually becoming apparent. In this study, we focus on how to construct translation LLMs that meet the needs of domain customization. We take visual media subtitle translation as our topic and explore how to train expressive and vivid translation LLMs. We investigated the situations of subtitle translation and other domains of literal and liberal translation, verifying the reliability of LLM as reward model and evaluator for translation. Additionally, to train an expressive translation LLM, we constructed and released a multidirectional subtitle parallel corpus dataset and proposed the Adaptive Local Preference Optimization (ALPO) method to address fine-grained preference alignment. Experimental results demonstrate that ALPO achieves outstanding performance in multidimensional evaluation of translation quality.

Method: Proposes an iterative trajectory generation and selection framework where for each problem, the agent runs multiple inference iterations. In each iteration, it optionally produces a high-level plan, selects reasoning tools and compute strategies with exploration parameters, then generates candidate reasoning trajectories. A process reward model (PRM) serves as unified control: step-level PRM scores guide pruning/expansion during generation within iterations, and aggregated trajectory rewards select final responses across iterations.

Result: The dynamic, PRM-guided approach consistently outperforms direct test-time scaling across datasets, yielding large gains on MATH-500 and several-fold improvements on harder benchmarks like AIME24 and AMO-Bench. Efficiency analysis shows verification-guided allocation concentrates computation on high-utility reasoning paths.

Conclusion: Adaptive, verifier-guided computation allocation during reasoning is more effective than uniform test-time scaling, with process reward models providing effective control signals for both intra-iteration trajectory generation and inter-iteration response selection.

Abstract: Test-time compute scaling allocates inference computation uniformly, uses fixed sampling strategies, and applies verification only for reranking. In contrast, we propose a verifier-guided adaptive framework treating reasoning as iterative trajectory generation and selection. For each problem, the agent runs multiple inference iterations. In each iteration, it optionally produces a high-level plan, selects a set of reasoning tools and a compute strategy together with an exploration parameter, and then generates a candidate reasoning trajectory. A process reward model (PRM) serves as a unified control signal: within each iteration, step-level PRM scores are aggregated to guide pruning and expansion during generation, and across iterations, aggregated trajectory rewards are used to select the final response. Across datasets, our dynamic, PRM-guided approach consistently outperforms direct test-time scaling, yielding large gains on MATH-500 and several-fold improvements on harder benchmarks such as AIME24 and AMO-Bench. We characterize efficiency using theoretical FLOPs and a compute intensity metric penalizing wasted generation and tool overhead, demonstrating that verification-guided allocation concentrates computation on high-utility reasoning paths.

Method: LORE uses fine-grained contrastive learning to activate latent logical reasoning capacity in embeddings, guiding them toward evidence aligned with logical structure rather than shallow similarity. It requires no external supervision, resources, or pre-retrieval analysis and remains index-compatible.

Result: LORE consistently improves retrieval utility and downstream generation while maintaining efficiency. The approach enhances performance on knowledge-intensive tasks involving complex logical relations.

Conclusion: Activating latent logical reasoning capacity through fine-grained contrastive learning effectively improves retrieval for LLMs on complex queries without requiring external resources or supervision.

Abstract: Large language models (LLMs) struggle in knowledge-intensive tasks, as retrievers often overfit to surface similarity and fail on queries involving complex logical relations. The capacity for logical analysis is inherent in model representations but remains underutilized in standard training. LORE (Logic ORiented Retriever Enhancement) introduces fine-grained contrastive learning to activate this latent capacity, guiding embeddings toward evidence aligned with logical structure rather than shallow similarity. LORE requires no external upervision, resources, or pre-retrieval analysis, remains index-compatible, and consistently improves retrieval utility and downstream generation while maintaining efficiency. The datasets and code are publicly available at https://github.com/mazehart/Lore-RAG.

Method: Tendem uses a hybrid approach where AI handles structured, repeatable work, and human experts step in when models fail or to verify results. All results undergo comprehensive quality review before delivery. Performance was evaluated through in-house tests on 94 real-world tasks comparing against AI-only agents and human-only workflows by Upwork freelancers.

Result: Tendem consistently delivers higher-quality outputs with faster turnaround times compared to both AI-only agents and human-only workflows, while maintaining operational costs comparable to human-only execution. On third-party benchmarks, Tendem’s AI Agent performs near state-of-the-art on web browsing and tool-use tasks with strong results in frontier domain knowledge and reasoning.

Conclusion: The hybrid AI-human approach of Tendem effectively combines the strengths of both AI and human expertise, resulting in superior performance in quality and speed while controlling costs, demonstrating the value of human-in-the-loop systems for complex real-world tasks.

Abstract: Tendem is a hybrid system where AI handles structured, repeatable work and Human Experts step in when the models fail or to verify results. Each result undergoes a comprehensive quality review before delivery to the Client. To assess Tendem’s performance, we conducted a series of in-house evaluations on 94 real-world tasks, comparing it with AI-only agents and human-only workflows carried out by Upwork freelancers. The results show that Tendem consistently delivers higher-quality outputs with faster turnaround times. At the same time, its operational costs remain comparable to human-only execution. On third-party agentic benchmarks, Tendem’s AI Agent (operating autonomously, without human involvement) performs near state-of-the-art on web browsing and tool-use tasks while demonstrating strong results in frontier domain knowledge and reasoning.

Method: Extends LLM-based TPPs to visual modality, positions text generation as core capability alongside time/type prediction, uses adaptive sequence compression based on temporal similarity to reduce sequence length while preserving patterns, employs two-stage paradigm of pre-training on compressed sequences followed by supervised fine-tuning.

Result: Method outperforms state-of-the-art baselines in both predictive accuracy and quality of generated textual analyses, demonstrated through extensive experiments including on challenging DanmakuTPP-QA benchmark.

Conclusion: The proposed framework successfully addresses the long-context problem in multimodal TPPs and enables effective generation of rich textual analyses alongside accurate event predictions.

Abstract: Temporal point processes (TPPs) have emerged as powerful tools for modeling asynchronous event sequences. While recent advances have extended TPPs to handle textual information, existing approaches are limited in their ability to generate rich, multimodal content and reason about event dynamics. A key challenge is that incorporating multimodal data dramatically increases sequence length, hindering the ability of attention-based models to generate coherent, long-form textual descriptions that require long-range understanding. In this paper, we propose a novel framework that extends LLM-based TPPs to the visual modality, positioning text generation as a core capability alongside time and type prediction. Our approach addresses the long-context problem through an adaptive sequence compression mechanism based on temporal similarity, which reduces sequence length while preserving essential patterns. We employ a two-stage paradigm of pre-training on compressed sequences followed by supervised fine-tuning for downstream tasks. Extensive experiments, including on the challenging DanmakuTPP-QA benchmark, demonstrate that our method outperforms state-of-the-art baselines in both predictive accuracy and the quality of its generated textual analyses.

Method: Introduces Logifus, a structure-preserving logical obfuscation framework, and LogiQAte benchmark with 1,108 questions across four reasoning tasks: first-order logic entailment, family-graph entailment, pattern induction, and navigation reasoning under altered formats.

Result: Obfuscation severely degrades zero-shot performance across all tasks: GPT-4o drops 47%, GPT-5 drops 27%, and reasoning model o4-mini drops 22% on average, revealing models parse questions without deep understanding.

Conclusion: LLMs lack genuine comprehension and rely on surface form patterns, highlighting the need for models that preserve meaning beyond surface structure and truly understand logical equivalence.

Abstract: Tasks such as solving arithmetic equations, evaluating truth tables, and completing syllogisms are handled well by large language models (LLMs) in their standard form, but they often fail when the same problems are posed in logically equivalent yet obfuscated formats. To study this vulnerability, we introduce Logifus, a structure-preserving logical obfuscation framework, and, utilizing this, we present LogiQAte, a first-of-its-kind diagnostic benchmark with 1,108 questions across four reasoning tasks: (i) Obfus FOL (first-order logic entailment under equivalence-preserving rewrites), (ii) Obfus Blood Relation (family-graph entailment under indirect relational chains), (iii) Obfus Number Series (pattern induction under symbolic substitutions), and (iv) Obfus Direction Sense (navigation reasoning under altered directions and reference frames). Across all the tasks, evaluating six state-of-the-art models, we find that obfuscation severely degrades zero-shot performance, with performance dropping on average by 47% for GPT-4o, 27% for GPT-5, and 22% for reasoning model, o4-mini. Our findings reveal that current LLMs parse questions without deep understanding, highlighting the urgency of building models that genuinely comprehend and preserve meaning beyond surface form.

Method: Compute linguistic, semantic, and structural metrics for Arabic, Chinese, and Japanese datasets; apply PCA within each language; fine-tune three LLM families (LLaMA, Mistral, DeepSeek); evaluate on cultural benchmarks; conduct controlled subset interventions.

Result: PCA components correlate with downstream performance but associations are strongly model-dependent; lexical-oriented components are most robust across models and benchmarks; emphasizing semantic or diversity extremes is often neutral or harmful.

Conclusion: Linguistic properties of fine-tuning data significantly impact cultural alignment, but effects are model-specific; lexical properties show most consistent benefits; dataset design for cultural adaptation should consider model-specific interactions.

Abstract: The global deployment of large language models (LLMs) has raised concerns about cultural misalignment, yet the linguistic properties of fine-tuning datasets used for cultural adaptation remain poorly understood. We adopt a dataset-centric view of cultural alignment and ask which linguistic properties of fine-tuning data are associated with cultural performance, whether these properties are predictive prior to training, and how these effects vary across models. We compute lightweight linguistic, semantic, and structural metrics for Arabic, Chinese, and Japanese datasets and apply principal component analysis separately within each language. This design ensures that the resulting components capture variation among datasets written in the same language rather than differences between languages. The resulting components correspond to broadly interpretable axes related to semantic coherence, surface-level lexical and syntactic diversity, and lexical or structural richness, though their composition varies across languages. We fine-tune three major LLM families (LLaMA, Mistral, DeepSeek) and evaluate them on benchmarks of cultural knowledge, values, and norms. While PCA components correlate with downstream performance, these associations are strongly model-dependent. Through controlled subset interventions, we show that lexical-oriented components (PC3) are the most robust, yielding more consistent performance across models and benchmarks, whereas emphasizing semantic or diversity extremes (PC1-PC2) is often neutral or harmful.

Method: Two-component framework: 1) Universal Metalinguistic Framework (UMF) representing languages as structured profiles across 16 typological dimensions with divergence-weighted scoring, and 2) Computational Engine operating through linguistic disambiguation during generation and typological compliance scoring during selection.

Result: Evaluation across nine language pairs shows intervention rates strongly correlating with typological distance from English. On 341 English sentences with different morphological/syntactic phenomena: 48.16% intervention precision for conservatively treated languages, 28.15% for morphologically dense languages, and 86.26% for structurally profiled languages.

Conclusion: The framework requires no parallel training data and works with any LLM capable of producing multiple candidate outputs, enabling practical deployment for under-resourced languages.

Abstract: Large language models trained predominantly on high-resource languages exhibit systematic biases toward dominant typological patterns, leading to structural non-conformance when translating into typologically divergent low-resource languages. We present a framework that leverages linguistic typology to improve translation quality without parallel training data or model retraining. The framework consists of two components: the Universal Metalinguistic Framework (UMF), which represents languages as structured profiles across 16 typological dimensions with divergence-weighted scoring, and the Computational Engine, which operates through linguistic disambiguation during generation and typological compliance scoring during selection. Evaluation across nine language pairs demonstrates intervention rates strongly correlating with typological distance from English. In experiments on 341 English sentences each having different morphological and syntactic phenomena, the framework shows an intervention precision of 48.16% for conservatively treated languages, 28.15% for morphologically dense languages, and 86.26% for structurally profiled languages. The framework requires no parallel training data and operates with any LLM capable of producing multiple candidate outputs, enabling practical deployment for under-resourced languages.

Method: Two-stage approach: 1) Binary classifier detects presence of pedagogical strategies, 2) Fine-grained classifier identifies specific strategies. Parallel system recommends strategies from dialogue context and uses LLM to generate strategy-aligned responses.

Result: High performance for pedagogical strategy detection with consistent gains from data augmentation. Fine-grained classification remains challenging in some cases.

Conclusion: PedagoSense successfully bridges pedagogical theory and LLM-based response generation, enabling more adaptive educational dialogue systems.

Abstract: This paper addresses the challenge of improving interaction quality in dialogue based learning by detecting and recommending effective pedagogical strategies in tutor student conversations. We introduce PedagoSense, a pedology grounded system that combines a two stage strategy classifier with large language model generation. The system first detects whether a pedagogical strategy is present using a binary classifier, then performs fine grained classification to identify the specific strategy. In parallel, it recommends an appropriate strategy from the dialogue context and uses an LLM to generate a response aligned with that strategy. We evaluate on human annotated tutor student dialogues, augmented with additional non pedagogical conversations for the binary task. Results show high performance for pedagogical strategy detection and consistent gains when using data augmentation, while analysis highlights where fine grained classes remain challenging. Overall, PedagoSense bridges pedagogical theory and practical LLM based response generation for more adaptive educational technologies.

Method: The system integrates four components: CNN-based Speech Emotion Recognition, Whisper for English ASR, fine-tuned MarianMT for English-to-Arabic translation, and MMS-TTS-Ara for Arabic speech synthesis, forming an end-to-end pipeline.

Result: Achieved 94% F1-score for emotion classification, BLEU 56 and BERTScore F1 88.7% for translation, and 81% average human evaluation score for banking-domain translations.

Conclusion: EmoAra successfully demonstrates an effective pipeline for emotion-preserving cross-lingual spoken communication with strong performance metrics across all components.

Abstract: This work presents EmoAra, an end-to-end emotion-preserving pipeline for cross-lingual spoken communication, motivated by banking customer service where emotional context affects service quality. EmoAra integrates Speech Emotion Recognition, Automatic Speech Recognition, Machine Translation, and Text-to-Speech to process English speech and deliver an Arabic spoken output while retaining emotional nuance. The system uses a CNN-based emotion classifier, Whisper for English transcription, a fine-tuned MarianMT model for English-to-Arabic translation, and MMS-TTS-Ara for Arabic speech synthesis. Experiments report an F1-score of 94% for emotion classification, translation performance of BLEU 56 and BERTScore F1 88.7%, and an average human evaluation score of 81% on banking-domain translations. The implementation and resources are available at the accompanying GitHub repository.

Method: The review examines developments from early multimodal fusion methods to newer frameworks using contrastive models like CLIP, diffusion-based text-to-image generation, and LLM support. It covers feature-based, graph-based, and contrastive embedding techniques from 2016-2025, focusing on prompt engineering, fine-tuning, and multilingual adaptation.

Result: CLIP-based fine-tuned models and LLM-enhanced VWSD systems consistently outperform zero-shot baselines, achieving gains of up to 6-8% in Mean Reciprocal Rank (MRR). However, challenges remain including context limitations, model bias toward common meanings, lack of multilingual datasets, and need for better evaluation frameworks.

Conclusion: The future of VWSD lies in the growing overlap of CLIP alignment, diffusion generation, and LLM reasoning for developing robust, context-aware, and multilingual disambiguation systems.

Abstract: This paper offers a mini review of Visual Word Sense Disambiguation (VWSD), which is a multimodal extension of traditional Word Sense Disambiguation (WSD). VWSD helps tackle lexical ambiguity in vision-language tasks. While conventional WSD depends only on text and lexical resources, VWSD uses visual cues to find the right meaning of ambiguous words with minimal text input. The review looks at developments from early multimodal fusion methods to new frameworks that use contrastive models like CLIP, diffusion-based text-to-image generation, and large language model (LLM) support. Studies from 2016 to 2025 are examined to show the growth of VWSD through feature-based, graph-based, and contrastive embedding techniques. It focuses on prompt engineering, fine-tuning, and adapting to multiple languages. Quantitative results show that CLIP-based fine-tuned models and LLM-enhanced VWSD systems consistently perform better than zero-shot baselines, achieving gains of up to 6-8% in Mean Reciprocal Rank (MRR). However, challenges still exist, such as limitations in context, model bias toward common meanings, a lack of multilingual datasets, and the need for better evaluation frameworks. The analysis highlights the growing overlap of CLIP alignment, diffusion generation, and LLM reasoning as the future path for strong, context-aware, and multilingual disambiguation systems.

Method: Provides theoretical and empirical evidence showing Vanilla Attention and Sink Attention naturally construct Mixture-of-Experts mechanisms within attention layers. Proposes sink-aware training algorithm with auxiliary load balancing loss designed for attention layers to mitigate head collapse.

Result: Extensive experiments show the method achieves effective head load balancing and improves model performance across Vanilla Attention, Sink Attention, and Gated Attention architectures.

Conclusion: The study offers new perspective on attention mechanisms and encourages further exploration of inherent MoE structure within attention layers, providing insights into attention sink phenomena and practical solutions for head collapse issues.

Abstract: Large Language Models (LLMs) often assign disproportionate attention to the first token, a phenomenon known as the attention sink. Several recent approaches aim to address this issue, including Sink Attention in GPT-OSS and Gated Attention in Qwen3-Next. However, a comprehensive analysis of the relationship among these attention mechanisms is lacking. In this work, we provide both theoretical and empirical evidence demonstrating that the sink in Vanilla Attention and Sink Attention naturally construct a Mixture-of-Experts (MoE) mechanism within attention layers. This insight explains the head collapse phenomenon observed in prior work, where only a fixed subset of attention heads contributes to generation. To mitigate head collapse, we propose a sink-aware training algorithm with an auxiliary load balancing loss designed for attention layers. Extensive experiments show that our method achieves effective head load balancing and improves model performance across Vanilla Attention, Sink Attention, and Gated Attention. We hope this study offers a new perspective on attention mechanisms and encourages further exploration of the inherent MoE structure within attention layers.

Method: Introduces ASTER (Agentic Scaling with Tool-integrated Extended Reasoning) framework with targeted cold-start strategy prioritizing interaction-dense trajectories. Studies how cold-start SFT induces tool-using behavioral prior, how interaction density shapes exploration, and how RL interaction budget affects learning dynamics.

Result: ASTER-4B achieves state-of-the-art results on competitive mathematical benchmarks, reaching 90.0% on AIME 2025, surpassing leading frontier open-source models including DeepSeek-V3.2-Exp. A small expert cold-start set of just 4K interaction-dense trajectories yields strongest downstream performance.

Conclusion: Targeted cold-start strategy with interaction-dense trajectories effectively prevents RL collapse in tool-integrated reasoning, enabling superior exploration during extended RL training and establishing robust behavioral priors for sustained multi-turn tool usage.

Abstract: Reinforcement learning (RL) has emerged as a dominant paradigm for eliciting long-horizon reasoning in Large Language Models (LLMs). However, scaling Tool-Integrated Reasoning (TIR) via RL remains challenging due to interaction collapse: a pathological state where models fail to sustain multi-turn tool usage, instead degenerating into heavy internal reasoning with only trivial, post-hoc code verification. We systematically study three questions: (i) how cold-start SFT induces an agentic, tool-using behavioral prior, (ii) how the interaction density of cold-start trajectories shapes exploration and downstream RL outcomes, and (iii) how the RL interaction budget affects learning dynamics and generalization under varying inference-time budgets. We then introduce ASTER (Agentic Scaling with Tool-integrated Extended Reasoning), a framework that circumvents this collapse through a targeted cold-start strategy prioritizing interaction-dense trajectories. We find that a small expert cold-start set of just 4K interaction-dense trajectories yields the strongest downstream performance, establishing a robust prior that enables superior exploration during extended RL training. Extensive evaluations demonstrate that ASTER-4B achieves state-of-the-art results on competitive mathematical benchmarks, reaching 90.0% on AIME 2025, surpassing leading frontier open-source models, including DeepSeek-V3.2-Exp.

Method: Chronos models each reasoning trajectory as a time series, learning to capture trajectory features from token probabilities. It assigns quality scores to trajectories and employs a weighted voting mechanism based on these scores.

Result: Extensive evaluations show Chronos consistently delivers substantial gains across various models with negligible computational overhead. Chronos@128 achieves 34.21% improvement over Pass@1 and 22.70% over Maj@128 on HMMT25 using Qwen3-4B-Thinking-2507.

Conclusion: Chronos provides an effective lightweight solution for improving LLM reasoning performance through chronological modeling of reasoning trajectories and weighted voting, outperforming existing test-time scaling methods.

Abstract: Test-Time Scaling (TTS) has emerged as an effective paradigm for improving the reasoning performance of large language models (LLMs). However, existing methods – most notably majority voting and heuristic token-level scoring – treat reasoning traces or tokens equally, thereby being susceptible to substantial variations in trajectory quality and localized logical failures. In this work, we introduce \textbf{Chronos}, a lightweight and plug-and-play chronological reasoning scorer that models each trajectory as a time series. Specifically, Chronos learns to capture trajectory features of token probabilities, assigns quality scores accordingly, and employs a weighted voting mechanism. Extensive evaluations on both in-domain and out-of-domain benchmarks demonstrate that Chronos consistently delivers substantial gains across a variety of models, with negligible computational overhead. Notably, Chronos@128 achieves relative improvements of 34.21% over Pass@1 and 22.70% over Maj@128 on HMMT25 using Qwen3-4B-Thinking-2507, highlighting its effectiveness.

Method: The paper introduces Token Priority as a conceptual framework that formalizes Supervised Fine-Tuning (SFT) as a precise distribution reshaping process rather than simple optimization. It categorizes existing approaches into two regimes: Positive Priority (for noise filtration) and Signed Priority (for toxic modes unlearning). The method involves analyzing recent breakthroughs through this unified lens and identifying key challenges.

Result: The paper provides a unified theoretical framework for understanding SFT as distribution reshaping, categorizes existing approaches into coherent regimes, identifies current limitations, and suggests directions for future research in alignment and fine-tuning methodologies.

Conclusion: Token Priority serves as an essential bridge to address the granularity mismatch in language model fine-tuning, offering a more precise understanding of how to align raw data with human utility through distribution reshaping rather than simple optimization.

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

Method: Created QUIT dataset with 7,401 questions and 2.4M passages built from high-convergence human- and machine-authored hints, labeled across three relevance levels using LLM-based answerability and human verification. Evaluated retrievers, rerankers, and LLM-based readers on this new task.

Result: Methods effective on traditional QA tasks struggle in inferential QA: retrievers underperform, rerankers offer limited gains, fine-tuning provides inconsistent improvements. Even reasoning-oriented LLMs fail to outperform smaller general-purpose models, showing current QA pipelines are not ready for inference-based reasoning.

Conclusion: Inferential QA establishes a new class of QA tasks that move towards understanding and reasoning from indirect textual evidence, revealing significant limitations in current QA systems’ ability to perform inference-based reasoning.

Abstract: Despite extensive research on a wide range of question answering (QA) systems, most existing work focuses on answer containment-i.e., assuming that answers can be directly extracted and/or generated from documents in the corpus. However, some questions require inference, i.e., deriving answers that are not explicitly stated but can be inferred from the available information. We introduce Inferential QA – a new task that challenges models to infer answers from answer-supporting passages which provide only clues. To study this problem, we construct QUIT (QUestions requiring Inference from Texts) dataset, comprising 7,401 questions and 2.4M passages built from high-convergence human- and machine-authored hints, labeled across three relevance levels using LLM-based answerability and human verification. Through comprehensive evaluation of retrievers, rerankers, and LLM-based readers, we show that methods effective on traditional QA tasks struggle in inferential QA: retrievers underperform, rerankers offer limited gains, and fine-tuning provides inconsistent improvements. Even reasoning-oriented LLMs fail to outperform smaller general-purpose models. These findings reveal that current QA pipelines are not yet ready for inference-based reasoning. Inferential QA thus establishes a new class of QA tasks that move towards understanding and reasoning from indirect textual evidence.

Method: Proposes DetectRouter, a prototype-based framework with two-stage training: 1) constructs discriminative prototypes from white-box models, 2) generalizes to black-box sources by aligning geometric distances with observed detection scores. Transforms detection into a routing problem of selecting the most appropriate surrogate for each input.

Result: Experiments on EvoBench and MAGE benchmarks show consistent improvements across multiple detection criteria and model families compared to fixed surrogate approaches.

Conclusion: Detection performance depends on surrogate-source alignment, and routing inputs to appropriate surrogates from a diverse pool significantly improves LLM-generated text detection compared to using fixed surrogates.

Abstract: Zero-shot methods detect LLM-generated text by computing statistical signatures using a surrogate model. Existing approaches typically employ a fixed surrogate for all inputs regardless of the unknown source. We systematically examine this design and find that detection performance varies substantially depending on surrogate-source alignment. We observe that while no single surrogate achieves optimal performance universally, a well-matched surrogate typically exists within a diverse pool for any given input. This finding transforms robust detection into a routing problem: selecting the most appropriate surrogate for each input. We propose DetectRouter, a prototype-based framework that learns text-detector affinity through two-stage training. The first stage constructs discriminative prototypes from white-box models; the second generalizes to black-box sources by aligning geometric distances with observed detection scores. Experiments on EvoBench and MAGE benchmarks demonstrate consistent improvements across multiple detection criteria and model families.

Method: Proposes TerminalTraj pipeline that: (1) filters high-quality repositories to construct Dockerized execution environments, (2) generates Docker-aligned task instances, and (3) synthesizes agent trajectories with executable validation code for verification.

Result: Curated 32K Docker images and generated 50,733 verified terminal trajectories across eight domains. Models trained on this data achieved up to 20% improvement on TerminalBench 1.0 and 10% on TerminalBench 2.0. TerminalTraj-32B reached 35.30% on TB 1.0 and 22.00% on TB 2.0.

Conclusion: TerminalTraj provides a scalable solution for generating high-quality terminal trajectory data, enabling improved training of agentic models for terminal-based tasks with demonstrated performance gains on benchmark evaluations.

Abstract: Training agentic models for terminal-based tasks critically depends on high-quality terminal trajectories that capture realistic long-horizon interactions across diverse domains. However, constructing such data at scale remains challenging due to two key requirements: \textbf{\emph{Executability}}, since each instance requires a suitable and often distinct Docker environment; and \textbf{\emph{Verifiability}}, because heterogeneous task outputs preclude unified, standardized verification. To address these challenges, we propose \textbf{TerminalTraj}, a scalable pipeline that (i) filters high-quality repositories to construct Dockerized execution environments, (ii) generates Docker-aligned task instances, and (iii) synthesizes agent trajectories with executable validation code. Using TerminalTraj, we curate 32K Docker images and generate 50,733 verified terminal trajectories across eight domains. Models trained on this data with the Qwen2.5-Coder backbone achieve consistent performance improvements on TerminalBench (TB), with gains of up to 20% on TB1.0 and 10% on TB2.0 over their respective backbones. Notably, \textbf{TerminalTraj-32B} achieves strong performance among models with fewer than 100B parameters, reaching 35.30% on TB1.0 and 22.00% on TB2.0, and demonstrates improved test-time scaling behavior. All code and data are available at https://github.com/Wusiwei0410/TerminalTraj.

Method: Created PARSE benchmark using controlled LLM-based generation pipeline with multi-stage filtering, annotation, and consistency checks. Questions cover diverse reasoning types, difficulty levels, and answer structures. Validated through human evaluation.

Result: Benchmarking shows Persian prompts and structured prompting (CoT for Boolean/multiple-choice; few-shot for factoid) improve performance. Fine-tuning boosts results, especially for Persian-specialized models.

Conclusion: PARSE fills a critical gap in Persian QA research and provides a foundation for developing and evaluating reasoning-capable LLMs in low-resource settings, supporting both fair comparison and practical model adaptation.

Abstract: Reasoning-focused Question Answering (QA) has advanced rapidly with Large Language Models (LLMs), yet high-quality benchmarks for low-resource languages remain scarce. Persian, spoken by roughly 130 million people, lacks a comprehensive open-domain resource for evaluating reasoning-capable QA systems. We introduce PARSE, the first open-domain Persian reasoning QA benchmark, containing 10,800 questions across Boolean, multiple-choice, and factoid formats, with diverse reasoning types, difficulty levels, and answer structures. The benchmark is built via a controlled LLM-based generation pipeline and validated through human evaluation. We also ensure linguistic and factual quality through multi-stage filtering, annotation, and consistency checks. We benchmark multilingual and Persian LLMs under multiple prompting strategies and show that Persian prompts and structured prompting (CoT for Boolean/multiple-choice; few-shot for factoid) improve performance. Fine-tuning further boosts results, especially for Persian-specialized models. These findings highlight how PARSE supports both fair comparison and practical model adaptation. PARSE fills a critical gap in Persian QA research and provides a strong foundation for developing and evaluating reasoning-capable LLMs in low-resource settings.

Method: Pacer uses a trainable pre-verification layer to dynamically control draft length by pre-verifying draft tokens blockwise before sending to target model, stopping generation if verification fails.

Result: Pacer achieves up to 2.66x speedup over autoregressive decoding, outperforms standard speculative decoding, and reaches 3.09x speedup when integrated with Ouroboros.

Conclusion: Dynamic draft length control via pre-verification significantly improves speculative decoding efficiency without sacrificing accuracy.

Abstract: Speculative decoding (SD) is a powerful technique for accelerating the inference process of large language models (LLMs) without sacrificing accuracy. Typically, SD employs a small draft model to generate a fixed number of draft tokens, which are then verified in parallel by the target model. However, our experiments reveal that the optimal draft length varies significantly across different decoding steps. This variation suggests that using a fixed draft length limits the potential for further improvements in decoding speed. To address this challenge, we propose Pacer, a novel approach that dynamically controls draft length using a lightweight, trainable pre-verification layer. This layer pre-verifies draft tokens blockwise before they are sent to the target model, allowing the draft model to stop token generation if the blockwise pre-verification fails. We implement Pacer on multiple SD model pairs and evaluate its performance across various benchmarks. Our results demonstrate that Pacer achieves up to 2.66x Speedup over autoregressive decoding and consistently outperforms standard speculative decoding. Furthermore, when integrated with Ouroboros, Pacer attains up to 3.09x Speedup.

Method: Created EverMemBench with multi-party, multi-group conversations spanning over 1 million tokens, featuring temporally evolving information, cross-topic interleaving, and role-specific personas. Includes 1,000+ QA pairs evaluating three dimensions: fine-grained recall, memory awareness, and user profile understanding.

Result: Revealed critical limitations: (1) multi-hop reasoning collapses in multi-party settings (oracle models achieve only 26%), (2) temporal reasoning remains unsolved requiring version semantics, (3) memory awareness bottlenecked by retrieval where similarity-based methods fail to bridge semantic gaps.

Conclusion: EverMemBench provides a challenging testbed for developing next-generation memory architectures, highlighting significant gaps in current LLM memory capabilities for complex real-world conversations.

Abstract: Long-term conversational memory is essential for LLM-based assistants, yet existing benchmarks focus on dyadic, single-topic dialogues that fail to capture real-world complexity. We introduce EverMemBench, a benchmark featuring multi-party, multi-group conversations spanning over 1 million tokens with temporally evolving information, cross-topic interleaving, and role-specific personas. EverMemBench evaluates memory systems across three dimensions through 1,000+ QA pairs: fine-grained recall, memory awareness, and user profile understanding. Our evaluation reveals critical limitations: (1) multi-hop reasoning collapses in multi-party settings, with even oracle models achieving only 26%; (2) temporal reasoning remains unsolved, requiring version semantics beyond timestamp matching; (3) memory awareness is bottlenecked by retrieval, where current similarity-based methods fail to bridge the semantic gap between queries and implicitly relevant memories. EverMemBench provides a challenging testbed for developing next-generation memory architectures.

Method: DreamOn augments diffusion process with two length control states that allow the model to autonomously expand or contract output length based on its predictions, requiring minimal modifications to training objectives and no architectural changes.

Result: Built on Dream-Coder-7B and DiffuCoder-7B, DreamOn achieves infilling performance on par with state-of-the-art autoregressive models on HumanEval-Infilling and SantaCoder-FIM, matching oracle performance with ground-truth length.

Conclusion: DreamOn removes a fundamental barrier to practical deployment of DLMs, significantly advancing their flexibility and applicability for variable-length generation in code infilling tasks.

Abstract: Diffusion Language Models (DLMs) present a compelling alternative to autoregressive models, offering flexible, any-order infilling without specialized prompting design. However, their practical utility is blocked by a critical limitation: the requirement of a fixed-length masked sequence for generation. This constraint severely degrades code infilling performance when the predefined mask size mismatches the ideal completion length. To address this, we propose DreamOn, a novel diffusion framework that enables dynamic, variable-length generation. DreamOn augments the diffusion process with two length control states, allowing the model to autonomously expand or contract the output length based solely on its own predictions. We integrate this mechanism into existing DLMs with minimal modifications to the training objective and no architectural changes. Built upon Dream-Coder-7B and DiffuCoder-7B, DreamOn achieves infilling performance on par with state-of-the-art autoregressive models on HumanEval-Infilling and SantaCoder-FIM and matches oracle performance achieved with ground-truth length. Our work removes a fundamental barrier to the practical deployment of DLMs, significantly advancing their flexibility and applicability for variable-length generation. Our code is available at https://github.com/DreamLM/DreamOn.

Method: Proposes CRAFT (Calibrated Reasoning with Answer-Faithful Traces) using Group Relative Policy Optimization (GRPO) reinforcement learning framework. Employs dual reward mechanisms: deterministic rewards for structural correctness and judge-based rewards for semantic faithfulness. Supports controllable trace variants for systematic analysis.

Result: Experiments on three multi-hop QA benchmarks show CRAFT improves both answer accuracy and reasoning faithfulness across model scales. The 7B model achieves competitive performance with closed-source LLMs across multiple reasoning trace settings.

Conclusion: CRAFT effectively addresses reasoning faithfulness challenges in RAG-based multi-hop QA through a reinforcement learning approach that optimizes both structural and semantic aspects of reasoning, enabling more reliable and faithful response generation.

Abstract: Retrieval-augmented generation (RAG) is widely used to ground Large Language Models (LLMs) for multi-hop question answering. Recent work mainly focused on improving answer accuracy via fine-tuning and structured or reinforcement-based optimization. However, reliable reasoning in response generation faces three challenges: 1) Reasoning Collapse. Reasoning in multi-hop QA is inherently complex due to multi-hop composition and is further destabilized by noisy retrieval. 2) Reasoning-answer inconsistency. Due to the intrinsic uncertainty of LLM generation and exposure to evidence–distractor mixtures, models may produce correct answers that are not faithfully supported by their intermediate reasoning or evidence. 3) Loss of format control. Traditional chain-of-thought generation often deviates from required structured output formats, leading to incomplete or malformed structured content. To address these challenges, we propose CRAFT (Calibrated Reasoning with Answer-Faithful Traces), a Group Relative Policy Optimization (GRPO) based reinforcement learning framework that trains models to perform faithful reasoning during response generation. CRAFT employs dual reward mechanisms to optimize multi-hop reasoning: deterministic rewards ensure structural correctness while judge-based rewards verify semantic faithfulness. This optimization framework supports controllable trace variants that enable systematic analysis of how structure and scale affect reasoning performance and faithfulness. Experiments on three multi-hop QA benchmarks show that CRAFT improves both answer accuracy and reasoning faithfulness across model scales, with the CRAFT 7B model achieving competitive performance with closed-source LLMs across multiple reasoning trace settings.

Method: Proposes XDLM that bridges the two paradigms via a stationary noise kernel, providing theoretical unification of both approaches and alleviating memory bottlenecks through algebraic simplification of posterior probabilities.

Result: XDLM advances the Pareto frontier between understanding capability and generation quality, surpassing UDLM by 5.4 points on zero-shot text benchmarks and outperforming MDLM in few-step image generation (FID 54.1 vs. 80.8). When scaled to tune an 8B-parameter LLM, achieves 15.0 MBPP in just 32 steps, effectively doubling baseline performance.

Conclusion: XDLM successfully unifies masked and uniform-noise diffusion paradigms, achieving balanced performance in both understanding and generation, with superior scaling potential and practical efficiency improvements.

Abstract: In discrete generative modeling, two dominant paradigms demonstrate divergent capabilities: Masked Diffusion Language Models (MDLM) excel at semantic understanding and zero-shot generalization, whereas Uniform-noise Diffusion Language Models (UDLM) achieve strong few-step generation quality, yet neither attains balanced performance across both dimensions. To address this, we propose XDLM, which bridges the two paradigms via a stationary noise kernel. XDLM offers two key contributions: (1) it provides a principled theoretical unification of MDLM and UDLM, recovering each paradigm as a special case; and (2) an alleviated memory bottleneck enabled by an algebraic simplification of the posterior probabilities. Experiments demonstrate that XDLM advances the Pareto frontier between understanding capability and generation quality. Quantitatively, XDLM surpasses UDLM by 5.4 points on zero-shot text benchmarks and outperforms MDLM in few-step image generation (FID 54.1 vs. 80.8). When scaled to tune an 8B-parameter large language model, XDLM achieves 15.0 MBPP in just 32 steps, effectively doubling the baseline performance. Finally, analysis of training dynamics reveals XDLM’s superior potential for long-term scaling. Code is available at https://github.com/MzeroMiko/XDLM

Method: RISE (Redundancy-Insensitive Scoring of Explanation) quantifies the unique influence of each input relative to others, minimizing the impact of redundancies and providing clearer, stable attributions.

Result: Experiments demonstrate that RISE offers more robust explanations than traditional methods, emphasizing the importance of conditional information for trustworthy LLM explanations and monitoring.

Conclusion: RISE addresses the challenge of distinguishing essential context elements from correlated ones, providing more reliable attribution scores for LLM interpretability and risk monitoring.

Abstract: Large language models (LLMs) generate outputs by utilizing extensive context, which often includes redundant information from prompts, retrieved passages, and interaction history. In critical applications, it is vital to identify which context elements actually influence the output, as standard explanation methods struggle with redundancy and overlapping context. Minor changes in input can lead to unpredictable shifts in attribution scores, undermining interpretability and raising concerns about risks like prompt injection. This work addresses the challenge of distinguishing essential context elements from correlated ones. We introduce RISE (Redundancy-Insensitive Scoring of Explanation), a method that quantifies the unique influence of each input relative to others, minimizing the impact of redundancies and providing clearer, stable attributions. Experiments demonstrate that RISE offers more robust explanations than traditional methods, emphasizing the importance of conditional information for trustworthy LLM explanations and monitoring.

Method: Theoretical analysis identifying the Bellman error of approximate reward models as the key quantity. For reasoning processes of length T, the paper proves that if the Bellman error is bounded by O(1/T), then combining this reward model with SMC reduces computational complexity from exponential to polynomial in T.

Result: Theoretical proof that approximate reward models with bounded Bellman error enable exponential improvement in inference efficiency. Specifically, the computational complexity of reasoning is reduced from exponential in T to polynomial in T when using SMC with such reward models.

Conclusion: Approximate reward models can be sufficient for effective inference-time scaling via SMC when their Bellman error is properly bounded. This provides theoretical justification for using practical approximate reward models in deployed systems and identifies the key condition for their effectiveness.

Abstract: Inference-time scaling has recently emerged as a powerful paradigm for improving the reasoning capability of large language models. Among various approaches, Sequential Monte Carlo (SMC) has become a particularly important framework, enabling iterative generation, evaluation, rejection, and resampling of intermediate reasoning trajectories. A central component in this process is the reward model, which evaluates partial solutions and guides the allocation of computation during inference. However, in practice, true reward models are never available. All deployed systems rely on approximate reward models, raising a fundamental question: Why and when do approximate reward models suffice for effective inference-time scaling? In this work, we provide a theoretical answer. We identify the Bellman error of the approximate reward model as the key quantity governing the effectiveness of SMC-based inference-time scaling. For a reasoning process of length $T$, we show that if the Bellman error of the approximate reward model is bounded by $O(1/T)$, then combining this reward model with SMC reduces the computational complexity of reasoning from exponential in $T$ to polynomial in $T$. This yields an exponential improvement in inference efficiency despite using only approximate rewards.

Method: Systematically analyze selective KD along position, class, and sample axes; then introduce student-entropy-guided position selection (SE-KD) and extend it across all three axes (SE-KD 3X)

Result: SE-KD improves accuracy, downstream task adherence, and memory efficiency; SE-KD 3X reduces wall time by 70%, peak memory by 18%, and storage usage by 80% without performance loss

Conclusion: Selective distillation guided by student entropy provides efficient and effective knowledge transfer in LLMs, making offline teacher caching feasible with significant resource savings

Abstract: Growing efforts to improve knowledge distillation (KD) in large language models (LLMs) replace dense teacher supervision with selective distillation, which uses a subset of token positions, vocabulary classes, or training samples for supervision. However, it remains unclear which importance signals, selection policies, and their interplay are most effective. In this work, we revisit where and how to distill in autoregressive LLMs. We disentangle selective KD along the position, class, and sample axes and systematically compare importance signals and selection policies. Then, guided by this analysis, we identify underexplored opportunities and introduce student-entropy-guided position selection (SE-KD). Across a suite of benchmarks, SE-KD often improves accuracy, downstream task adherence, and memory efficiency over dense distillation. Extending this approach across the class and sample axes (SE-KD 3X) yields complementary efficiency gains that make offline teacher caching feasible. In practice, this reduces wall time by 70% and peak memory by 18%, while cutting storage usage by 80% over prior methods without sacrificing performance.

Method: Position paper approach: 1) Explain HLS as practical abstraction layer and golden reference for agentic hardware design, 2) Identify key limitations of current HLS tools that agents can address, 3) Propose taxonomy for symbiotic evolution of agentic HLS showing responsibility shift from humans to AI agents.

Result: Establishes HLS as critical layer for agentic optimization in hardware design, identifies specific tool limitations (inadequate performance feedback, rigid interfaces, limited debuggability), and provides framework for evolution from copilots to autonomous design partners.

Conclusion: HLS remains essential in the agentic era, serving as a natural optimization layer that enables faster iteration and design exploration, with AI agents uniquely positioned to address current HLS tool limitations.

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

Method: SentiFuse uses a standardization layer to normalize outputs from heterogeneous models, then applies multiple fusion strategies: decision-level fusion (combining final predictions), feature-level fusion (combining intermediate representations), and adaptive fusion (dynamically weighting models based on input characteristics).

Result: Experiments on three large-scale social-media datasets (Crowdflower, GoEmotions, Sentiment140) show SentiFuse consistently outperforms individual models and naive ensembles. Feature-level fusion achieves up to 4% absolute F1 improvement over best individual model, while adaptive fusion enhances robustness on challenging cases like negation and mixed emotions.

Conclusion: Systematically leveraging model complementarity through frameworks like SentiFuse yields more accurate and reliable sentiment analysis across diverse datasets and text types, with different fusion strategies offering different advantages.

Abstract: Sentiment analysis models exhibit complementary strengths, yet existing approaches lack a unified framework for effective integration. We present SentiFuse, a flexible and model-agnostic framework that integrates heterogeneous sentiment models through a standardization layer and multiple fusion strategies. Our approach supports decision-level fusion, feature-level fusion, and adaptive fusion, enabling systematic combination of diverse models. We conduct experiments on three large-scale social-media datasets: Crowdflower, GoEmotions, and Sentiment140. These experiments show that SentiFuse consistently outperforms individual models and naive ensembles. Feature-level fusion achieves the strongest overall effectiveness, yielding up to 4% absolute improvement in F1 score over the best individual model and simple averaging, while adaptive fusion enhances robustness on challenging cases such as negation, mixed emotions, and complex sentiment expressions. These results demonstrate that systematically leveraging model complementarity yields more accurate and reliable sentiment analysis across diverse datasets and text types.

Method: Introduce BanglaCQA dataset by extending existing Bangla dataset with counterfactual passages and answerability annotations. Propose fine-tuned pipelines for encoder-decoder language-specific/multilingual models and prompting-based pipelines for decoder-only LLMs. Apply LLM-based and human evaluation with semantic similarity metrics.

Result: Chain-of-Thought prompting reveals uniquely effective mechanism for extracting parametric knowledge in counterfactual scenarios, particularly in decoder-only LLMs. Detailed analysis shows how models perform across different QA settings in low-resource languages.

Conclusion: Introduces novel framework for analyzing knowledge sources in Bangla QA and uncovers critical findings that open broader directions for counterfactual reasoning in low-resource language settings.

Abstract: Question-Answering (QA) models for low-resource languages like Bangla face challenges due to limited annotated data and linguistic complexity. A key issue is determining whether models rely more on pre-encoded (parametric) knowledge or contextual input during answer generation, as existing Bangla QA datasets lack the structure required for such analysis. We introduce BanglaCQA, the first Counterfactual QA dataset in Bangla, by extending a Bangla dataset while integrating counterfactual passages and answerability annotations. In addition, we propose fine-tuned pipelines for encoder-decoder language-specific and multilingual baseline models, and prompting-based pipelines for decoder-only LLMs to disentangle parametric and contextual knowledge in both factual and counterfactual scenarios. Furthermore, we apply LLM-based and human evaluation techniques that measure answer quality based on semantic similarity. We also present a detailed analysis of how models perform across different QA settings in low-resource languages, and show that Chain-of-Thought (CoT) prompting reveals a uniquely effective mechanism for extracting parametric knowledge in counterfactual scenarios, particularly in decoder-only LLMs. Our work not only introduces a novel framework for analyzing knowledge sources in Bangla QA but also uncovers critical findings that open up broader directions for counterfactual reasoning in low-resource language settings.

Method: ConPress constructs multi-question prompts to induce self-compression, samples the resulting model outputs, parses and filters per-question traces to obtain concise yet correct reasoning trajectories, then uses these for supervised fine-tuning to internalize compressed reasoning behavior.

Result: With only 8k fine-tuning examples, ConPress reduces reasoning token usage by 59% on MATH500 and 33% on AIME25 while maintaining competitive accuracy, demonstrating efficient compression of reasoning traces without external teachers or reinforcement learning.

Conclusion: The self-compression phenomenon can be leveraged through ConPress to create more efficient reasoning models that maintain accuracy while significantly reducing inference overhead, offering a lightweight self-supervised approach to reasoning optimization.

Abstract: Large reasoning models (LRMs) typically solve reasoning-intensive tasks by generating long chain-of-thought (CoT) traces, leading to substantial inference overhead. We identify a reproducible inference-time phenomenon, termed Self-Compression: when multiple independent and answerable questions are presented within a single prompt, the model spontaneously produces shorter reasoning traces for each question. This phenomenon arises from multi-question contextual pressure during generation and consistently manifests across models and benchmarks. Building on this observation, we propose ConPress (Learning from Contextual Pressure), a lightweight self-supervised fine-tuning approach. ConPress constructs multi-question prompts to induce self-compression, samples the resulting model outputs, and parses and filters per-question traces to obtain concise yet correct reasoning trajectories. These trajectories are directly used for supervised fine-tuning, internalizing compressed reasoning behavior in single-question settings without external teachers, manual pruning, or reinforcement learning. With only 8k fine-tuning examples, ConPress reduces reasoning token usage by 59% on MATH500 and 33% on AIME25, while maintaining competitive accuracy.

Method: Created Ebisu benchmark with two tasks: JF-ICR (implicit commitment and refusal recognition in investor Q&A) and JF-TE (hierarchical extraction and ranking of nested financial terminology from professional disclosures). Evaluated diverse LLMs including general-purpose, Japanese-adapted, and financial models.

Result: Even state-of-the-art LLMs struggle on both tasks. Increased model scale yields limited improvements, and language- and domain-specific adaptation does not reliably improve performance, leaving substantial gaps in Japanese financial language understanding.

Conclusion: Ebisu provides a focused benchmark for advancing linguistically and culturally grounded financial NLP, highlighting the need for specialized approaches to handle Japanese financial language’s unique characteristics. All datasets and evaluation scripts are publicly released.

Abstract: Japanese finance combines agglutinative, head-final linguistic structure, mixed writing systems, and high-context communication norms that rely on indirect expression and implicit commitment, posing a substantial challenge for LLMs. We introduce Ebisu, a benchmark for native Japanese financial language understanding, comprising two linguistically and culturally grounded, expert-annotated tasks: JF-ICR, which evaluates implicit commitment and refusal recognition in investor-facing Q&A, and JF-TE, which assesses hierarchical extraction and ranking of nested financial terminology from professional disclosures. We evaluate a diverse set of open-source and proprietary LLMs spanning general-purpose, Japanese-adapted, and financial models. Results show that even state-of-the-art systems struggle on both tasks. While increased model scale yields limited improvements, language- and domain-specific adaptation does not reliably improve performance, leaving substantial gaps unresolved. Ebisu provides a focused benchmark for advancing linguistically and culturally grounded financial NLP. All datasets and evaluation scripts are publicly released.

Method: Proposes Rubric-ARM framework with joint optimization of rubric generator and judge using reinforcement learning from preference feedback. Uses alternating optimization strategy to mitigate non-stationarity of simultaneous updates, treating rubric generation as a latent action learned to maximize judgment accuracy.

Result: Achieves state-of-the-art performance on multiple benchmarks and significantly improves downstream policy alignment in both offline and online reinforcement learning settings.

Conclusion: Rubric-ARM provides an effective framework for nuanced response quality assessment in non-verifiable domains through joint optimization of rubric generation and judgment, with theoretical guarantees and practical benefits for policy alignment.

Abstract: Standard reward models typically predict scalar scores that fail to capture the multifaceted nature of response quality in non-verifiable domains, such as creative writing or open-ended instruction following. To address this limitation, we propose Rubric-ARM, a framework that jointly optimizes a rubric generator and a judge using reinforcement learning from preference feedback. Unlike existing methods that rely on static rubrics or disjoint training pipelines, our approach treats rubric generation as a latent action learned to maximize judgment accuracy. We introduce an alternating optimization strategy to mitigate the non-stationarity of simultaneous updates, providing theoretical analysis that demonstrates how this schedule reduces gradient variance during training. Extensive experiments show that Rubric-ARM achieves state-of-the-art performance among baselines on multiple benchmarks and significantly improves downstream policy alignment in both offline and online reinforcement learning settings.

Method: Proposes Argument Rarity-based Originality Assessment (AROA) framework that defines originality as rarity within a reference corpus. Evaluates through four components: structural rarity, claim rarity, evidence rarity, and cognitive depth. Uses density estimation to quantify rarity and integrates with quality adjustment mechanism, treating quality and originality as independent evaluation axes.

Result: Experiments show strong negative correlation between quality and claim rarity, demonstrating quality-originality trade-off. AI essays achieved comparable structural complexity to human essays but had substantially lower claim rarity, indicating LLMs can reproduce argumentation form but have limitations in content originality.

Conclusion: AROA provides effective framework for originality assessment, revealing fundamental differences between human and AI-generated content. LLMs excel at structural reproduction but struggle with original claims, highlighting need for originality-focused assessment in education.

Abstract: As Large Language Models (LLMs) have become capable of effortlessly generating high-quality text, traditional quality-focused writing assessment is losing its significance. If the essential goal of education is to foster critical thinking and original perspectives, assessment must also shift its paradigm from quality to originality. This study proposes Argument Rarity-based Originality Assessment (AROA), a framework for automatically evaluating argumentative originality in student essays. AROA defines originality as rarity within a reference corpus and evaluates it through four complementary components: structural rarity, claim rarity, evidence rarity, and cognitive depth. The framework quantifies the rarity of each component using density estimation and integrates them with a quality adjustment mechanism, thereby treating quality and originality as independent evaluation axes. Experiments using human essays and AI-generated essays revealed a strong negative correlation between quality and claim rarity, demonstrating a quality-originality trade-off where higher-quality texts tend to rely on typical claim patterns. Furthermore, while AI essays achieved comparable levels of structural complexity to human essays, their claim rarity was substantially lower than that of humans, indicating that LLMs can reproduce the form of argumentation but have limitations in the originality of content.

Method: Uses a dual-agent framework: Context Builder agent browses internet, writes structured notes, archives raw sources into hierarchical knowledge base; Report Writer agent composes final report section by section using knowledge base as source. File system serves as durable external memory and shared coordination medium.

Result: Achieves state-of-the-art report quality on DeepResearch Bench and DeepConsult benchmarks across different backbone models. Shows positive correlation between report quality and computation allocated to Context Builder, validating test-time scaling.

Conclusion: File-system paradigm enables scaling deep research beyond context window via persistent workspace, allowing iterative refinement and effective test-time scaling for long-horizon tasks.

Abstract: Deep research is emerging as a representative long-horizon task for large language model (LLM) agents. However, long trajectories in deep research often exceed model context limits, compressing token budgets for both evidence collection and report writing, and preventing effective test-time scaling. We introduce FS-Researcher, a file-system-based, dual-agent framework that scales deep research beyond the context window via a persistent workspace. Specifically, a Context Builder agent acts as a librarian which browses the internet, writes structured notes, and archives raw sources into a hierarchical knowledge base that can grow far beyond context length. A Report Writer agent then composes the final report section by section, treating the knowledge base as the source of facts. In this framework, the file system serves as a durable external memory and a shared coordination medium across agents and sessions, enabling iterative refinement beyond the context window. Experiments on two open-ended benchmarks (DeepResearch Bench and DeepConsult) show that FS-Researcher achieves state-of-the-art report quality across different backbone models. Further analyses demonstrate a positive correlation between final report quality and the computation allocated to the Context Builder, validating effective test-time scaling under the file-system paradigm. The code and data are anonymously open-sourced at https://github.com/Ignoramus0817/FS-Researcher.

Method: Proposes Value Aggregation (VA) that pools token values across multiple layers and token indices. Further refines to Aligned Weighted VA (AlignedWVA) using attention scores as weights and output projection matrix for alignment.

Result: VA outperforms other LLM-based embeddings in training-free setting, matches/surpasses ensemble-based MetaEOL. AlignedWVA achieves SOTA among training-free LLM embeddings, substantially outperforming high-cost MetaEOL.

Conclusion: Attention value vectors capture sentence semantics better than hidden states. Value Aggregation methods provide strong LLM embeddings, with potential for further improvement through fine-tuning.

Abstract: Sentence representations are foundational to many Natural Language Processing (NLP) applications. While recent methods leverage Large Language Models (LLMs) to derive sentence representations, most rely on final-layer hidden states, which are optimized for next-token prediction and thus often fail to capture global, sentence-level semantics. This paper introduces a novel perspective, demonstrating that attention value vectors capture sentence semantics more effectively than hidden states. We propose Value Aggregation (VA), a simple method that pools token values across multiple layers and token indices. In a training-free setting, VA outperforms other LLM-based embeddings, even matches or surpasses the ensemble-based MetaEOL. Furthermore, we demonstrate that when paired with suitable prompts, the layer attention outputs can be interpreted as aligned weighted value vectors. Specifically, the attention scores of the last token function as the weights, while the output projection matrix ($W_O$) aligns these weighted value vectors with the common space of the LLM residual stream. This refined method, termed Aligned Weighted VA (AlignedWVA), achieves state-of-the-art performance among training-free LLM-based embeddings, outperforming the high-cost MetaEOL by a substantial margin. Finally, we highlight the potential of obtaining strong LLM embedding models through fine-tuning Value Aggregation.

Method: Certified Semantic Smoothing (CSS) via Stratified Randomized Ablation partitions inputs into immutable structural prompts and mutable payloads, using Hypergeometric distribution for rigorous guarantees. Noise-Augmented Alignment Tuning (NAAT) transforms base models into semantic denoisers to handle sparse contexts.

Result: Reduces Attack Success Rate of gradient-based attacks from 84.2% to 1.2% while maintaining 94.1% benign utility, significantly outperforming character-level baselines (74.3% utility).

Conclusion: The framework provides deterministic certificates of safety, ensuring models remain robust against all adversarial variants within a provable radius, offering a principled approach to LLM safety.

Abstract: Large Language Models (LLMs) remain vulnerable to adaptive jailbreaks that easily bypass empirical defenses like GCG. We propose a framework for certifiable robustness that shifts safety guarantees from single-pass inference to the statistical stability of an ensemble. We introduce Certified Semantic Smoothing (CSS) via Stratified Randomized Ablation, a technique that partitions inputs into immutable structural prompts and mutable payloads to derive rigorous lo norm guarantees using the Hypergeometric distribution. To resolve performance degradation on sparse contexts, we employ Noise-Augmented Alignment Tuning (NAAT), which transforms the base model into a semantic denoiser. Extensive experiments on Llama-3 show that our method reduces the Attack Success Rate of gradient-based attacks from 84.2% to 1.2% while maintaining 94.1% benign utility, significantly outperforming character-level baselines which degrade utility to 74.3%. This framework provides a deterministic certificate of safety, ensuring that a model remains robust against all adversarial variants within a provable radius.

Method: Introduces Wiki Live Challenge (WLC) benchmark using 100 recent Wikipedia Good Articles as expert-level references. Develops Wiki Eval framework with fine-grained evaluation method (39 criteria for writing quality) and rigorous metrics for factual verifiability.

Result: Experiments show significant gap between current Deep Research Agents and human expert-level Wikipedia articles, validating WLC’s effectiveness in advancing agent research.

Conclusion: WLC provides a reliable, expert-verified benchmark for evaluating Deep Research Agents, addressing limitations of current LLM-based evaluation approaches and enabling more rigorous assessment of agent capabilities.

Abstract: Deep Research Agents (DRAs) have demonstrated remarkable capabilities in autonomous information retrieval and report generation, showing great potential to assist humans in complex research tasks. Current evaluation frameworks primarily rely on LLM-generated references or LLM-derived evaluation dimensions. While these approaches offer scalability, they often lack the reliability of expert-verified content and struggle to provide objective, fine-grained assessments of critical dimensions. To bridge this gap, we introduce Wiki Live Challenge (WLC), a live benchmark that leverages the newest Wikipedia Good Articles (GAs) as expert-level references. Wikipedia’s strict standards for neutrality, comprehensiveness, and verifiability serve as a great challenge for DRAs, with GAs representing the pinnacle of which. We curate a dataset of 100 recent Good Articles and propose Wiki Eval, a comprehensive evaluation framework comprising a fine-grained evaluation method with 39 criteria for writing quality and rigorous metrics for factual verifiability. Extensive experiments on various DRA systems demonstrate a significant gap between current DRAs and human expert-level Wikipedia articles, validating the effectiveness of WLC in advancing agent research. We release our benchmark at https://github.com/WangShao2000/Wiki_Live_Challenge

Method: Proposes Socratic Inquiry Framework (SIF) with two components: Strategy Anchoring (when to ask) and Template Retrieval (what to ask), plus Socratic-QA dataset for supervision. SIF is a lightweight, plug-and-play intent planner that enables context-aware questioning without end-to-end retraining.

Result: SIF significantly enhances proactive questioning frequency, conversational depth, and therapeutic alignment, shifting LLMs from reactive comfort to proactive exploration in therapeutic contexts.

Conclusion: Establishes a new paradigm for psychologically informed LLMs: not just to respond, but to guide, with SIF enabling theory-grounded proactive questioning in therapy.

Abstract: Proactive questioning, where therapists deliberately initiate structured, cognition-guiding inquiries, is a cornerstone of cognitive behavioral therapy (CBT). Yet, current psychological large language models (LLMs) remain overwhelmingly reactive, defaulting to empathetic but superficial responses that fail to surface latent beliefs or guide behavioral change. To bridge this gap, we propose the \textbf{Socratic Inquiry Framework (SIF)}, a lightweight, plug-and-play therapeutic intent planner that transforms LLMs from passive listeners into active cognitive guides. SIF decouples \textbf{when to ask} (via Strategy Anchoring) from \textbf{what to ask} (via Template Retrieval), enabling context-aware, theory-grounded questioning without end-to-end retraining. Complementing SIF, we introduce \textbf{Socratic-QA}, a high-quality dataset of strategy-aligned Socratic sequences that provides explicit supervision for proactive reasoning. Experiments show that SIF significantly enhances proactive questioning frequency, conversational depth, and therapeutic alignment, marking a clear shift from reactive comfort to proactive exploration. Our work establishes a new paradigm for psychologically informed LLMs: not just to respond, but to guide.

Method: Developed a novel agentic data-generation framework to scalably create authentic, region-specific safety datasets for Southeast Asia. Built the SEA-Guard family of multilingual safeguard models grounded in SEA cultural contexts.

Result: SEA-Guard consistently outperforms existing safeguards at detecting regionally sensitive or harmful content across multiple benchmarks and cultural variants while maintaining strong general safety performance.

Conclusion: Culturally aware safeguards are essential for AI alignment in real-world settings, and the SEA-Guard framework demonstrates that region-specific, culturally-grounded safety models can be effectively developed through scalable data generation approaches.

Abstract: Culturally aware safeguards are crucial for AI alignment in real-world settings, where safety extends beyond common sense and encompasses diverse local values, norms, and region-specific regulations. However, building large-scale, culturally grounded datasets is challenging due to limited resources and a scarcity of native annotators. Consequently, many safeguard models rely on machine translation of English datasets, often missing regional and cultural nuances. We present a novel agentic data-generation framework to scalably create authentic, region-specific safety datasets for Southeast Asia (SEA). On this foundation, we introduce the SEA-Guard family, the first multilingual safeguard models grounded in SEA cultural contexts. Evaluated across multiple benchmarks and cultural variants, SEA-Guard consistently outperforms existing safeguards at detecting regionally sensitive or harmful content while maintaining strong general safety performance.

Method: Proposes Agentic Automatic Evaluation (A2Eval) with two collaborative agents: 1) Data Agent autonomously induces capability dimensions and assembles balanced, compact evaluation suites, and 2) Eval Agent synthesizes and validates executable evaluation pipelines for fully autonomous assessment.

Result: A2Eval compresses evaluation suites by 85%, reduces computational costs by 77%, delivers 4.6x speedup while preserving evaluation quality, corrects systematic ranking biases, improves human alignment (Spearman’s rho=0.85), and maintains high ranking fidelity (Kendall’s tau=0.81).

Conclusion: A2Eval establishes a new standard for high-fidelity, low-cost embodied assessment, enabling more efficient and accurate evaluation of embodied VLMs through autonomous benchmark curation and evaluation.

Abstract: Current embodied VLM evaluation relies on static, expert-defined, manually annotated benchmarks that exhibit severe redundancy and coverage imbalance. This labor intensive paradigm drains computational and annotation resources, inflates costs, and distorts model rankings, ultimately stifling iterative development. To address this, we propose Agentic Automatic Evaluation (A2Eval), the first agentic framework that automates benchmark curation and evaluation through two collaborative agents. The Data Agent autonomously induces capability dimensions and assembles a balanced, compact evaluation suite, while the Eval Agent synthesizes and validates executable evaluation pipelines, enabling fully autonomous, high-fidelity assessment. Evaluated across 10 benchmarks and 13 models, A2Eval compresses evaluation suites by 85%, reduces overall computational costs by 77%, and delivers a 4.6x speedup while preserving evaluation quality. Crucially, A2Eval corrects systematic ranking biases, improves human alignment to Spearman’s rho=0.85, and maintains high ranking fidelity (Kendall’s tau=0.81), establishing a new standard for high-fidelity, low-cost embodied assessment. Our code and data will be public soon.

Method: Proposes Steering Vector Fields (SVF) which learns a differentiable concept scoring function whose local gradient defines the steering direction at each activation, making interventions explicitly context-dependent. This supports coordinated multi-layer interventions in a shared concept space and enables efficient long-form and multi-attribute control.

Result: SVF delivers stronger and more reliable control across multiple LLMs and steering tasks, improving the practicality of inference-time steering compared to static steering vectors.

Conclusion: By making steering directions context-dependent through gradient fields, SVF addresses fundamental limitations of static steering vectors and provides a more effective framework for inference-time control of language models.

Abstract: Steering vectors (SVs) offer a lightweight way to control large language models (LLMs) at inference time by shifting hidden activations, providing a practical middle ground between prompting and fine-tuning. Yet SVs can be unreliable in practice. Some concepts are unsteerable, and even when steering helps on average it can backfire for a non-trivial fraction of inputs. Reliability also degrades in long-form generation and multi-attribute steering. We take a geometric view of these failures. A static SV applies the same update vector everywhere in representation space, implicitly assuming that the concept-improving direction is constant across contexts. When the locally effective direction varies with the current activation, a single global vector can become misaligned, which yields weak or reversed effects. Guided by this perspective, we propose Steering Vector Fields (SVF), which learns a differentiable concept scoring function whose local gradient defines the steering direction at each activation, making interventions explicitly context-dependent. This formulation supports coordinated multi-layer interventions in a shared, aligned concept space, and enables efficient long-form and multi-attribute control within a unified framework. Across multiple LLMs and steering tasks, SVF delivers stronger and more reliable control, improving the practicality of inference-time steering.

Method: CoDiQ framework uses test-time scaling for fine-grained difficulty control while ensuring solvability. Identifies scaling tendencies, develops CoDiQ-Generator from Qwen3-8B to improve upper bound of difficult question generation, and builds CoDiQ-Corpus of 44K competition-grade question sequences.

Result: Human evaluations show CoDiQ questions are significantly more challenging than LiveCodeBench/AIME with over 82% solvability. Training LRMs on CoDiQ-Corpus substantially improves reasoning performance.

Conclusion: Scaling controlled-difficulty training questions enhances reasoning capabilities. The framework and corpus are open-sourced to support related research.

Abstract: Large Reasoning Models (LRMs) benefit substantially from training on challenging competition-level questions. However, existing automated question synthesis methods lack precise difficulty control, incur high computational costs, and struggle to generate competition-level questions at scale. In this paper, we propose CoDiQ (Controllable Difficult Question Generation), a novel framework enabling fine-grained difficulty control via test-time scaling while ensuring question solvability. Specifically, first, we identify a test-time scaling tendency (extended reasoning token budget boosts difficulty but reduces solvability) and the intrinsic properties defining the upper bound of a model’s ability to generate valid, high-difficulty questions. Then, we develop CoDiQ-Generator from Qwen3-8B, which improves the upper bound of difficult question generation, making it particularly well-suited for challenging question construction. Building on the CoDiQ framework, we build CoDiQ-Corpus (44K competition-grade question sequences). Human evaluations show these questions are significantly more challenging than LiveCodeBench/AIME with over 82% solvability. Training LRMs on CoDiQ-Corpus substantially improves reasoning performance, verifying that scaling controlled-difficulty training questions enhances reasoning capabilities. We open-source CoDiQ-Corpus, CoDiQ-Generator, and implementations to support related research.

Method: Proposes DeepControl framework based on formal information utility that measures marginal value of retrieved evidence. Introduces retrieval continuation and granularity control mechanisms to regulate when to continue/stop retrieval and how much information to expand. Uses annealed control strategy to internalize effective information acquisition behaviors.

Result: Extensive experiments across seven benchmarks show consistent outperformance of strong baselines. Achieves average improvements of 9.4% and 8.6% on Qwen2.5-7B and Qwen2.5-3B over outcome-based RL baselines. Consistently outperforms both retrieval-free and retrieval-based reasoning methods without explicit information control.

Conclusion: Adaptive information control is crucial for scaling search-augmented reasoning agents to complex, real-world information environments. DeepControl demonstrates effectiveness of formal information utility for regulating retrieval in reasoning systems.

Abstract: Search-augmented reasoning agents interleave multi-step reasoning with external information retrieval, but uncontrolled retrieval often leads to redundant evidence, context saturation, and unstable learning. Existing approaches rely on outcome-based reinforcement learning (RL), which provides limited guidance for regulating information acquisition. We propose DeepControl, a framework for adaptive information control based on a formal notion of information utility, which measures the marginal value of retrieved evidence under a given reasoning state. Building on this utility, we introduce retrieval continuation and granularity control mechanisms that selectively regulate when to continue and stop retrieval, and how much information to expand. An annealed control strategy enables the agent to internalize effective information acquisition behaviors during training. Extensive experiments across seven benchmarks demonstrate that our method consistently outperforms strong baselines. In particular, our approach achieves average performance improvements of 9.4% and 8.6% on Qwen2.5-7B and Qwen2.5-3B, respectively, over strong outcome-based RL baselines, and consistently outperforms both retrieval-free and retrieval-based reasoning methods without explicit information control. These results highlight the importance of adaptive information control for scaling search-augmented reasoning agents to complex, real-world information environments.

Method: Inspired by ICL properties and LLMs’ functional modules, the authors propose the “counting hypothesis” of ICL, which suggests that LLMs’ encoding strategy may underlie ICL, and provide supporting evidence for this hypothesis.

Result: The paper presents evidence supporting the counting hypothesis, suggesting that LLMs’ encoding strategies play a fundamental role in enabling ICL capabilities.

Conclusion: Understanding the counting hypothesis and LLMs’ encoding strategies provides insights into ICL mechanisms, which could help improve error correction, diagnosis, and broader utilization of LLMs across domains.

Abstract: In-Context Learning (ICL) indicates that large language models (LLMs) pretrained on a massive amount of data can learn specific tasks from input prompts’ examples. ICL is notable for two reasons. First, it does not need modification of LLMs’ internal structure. Second, it enables LLMs to perform a wide range of tasks/functions with a few examples demonstrating a desirable task. ICL opens up new ways to utilize LLMs in more domains, but its underlying mechanisms still remain poorly understood, making error correction and diagnosis extremely challenging. Thus, it is imperative that we better understand the limitations of ICL and how exactly LLMs support ICL. Inspired by ICL properties and LLMs’ functional modules, we propose 1the counting hypothesis’ of ICL, which suggests that LLMs’ encoding strategy may underlie ICL, and provide supporting evidence.

Method: Proposes Latent Exploration Decoding (LED) - a depth-conditioned decoding strategy that aggregates intermediate posteriors via cumulative sum and selects depth configurations with maximal entropy as exploration candidates, requiring no additional training or parameters.

Result: LED consistently improves pass@1 and pass@16 accuracy by 0.61 and 1.03 percentage points across multiple reasoning benchmarks and models without additional training.

Conclusion: LED effectively addresses exploration collapse in post-trained reasoning models by leveraging entropy asymmetry between intermediate and final layers, providing a simple yet effective decoding strategy for improved reasoning performance.

Abstract: Large Reasoning Models (LRMs) have recently achieved strong mathematical and code reasoning performance through Reinforcement Learning (RL) post-training. However, we show that modern reasoning post-training induces an unintended exploration collapse: temperature-based sampling no longer increases pass@$n$ accuracy. Empirically, the final-layer posterior of post-trained LRMs exhibit sharply reduced entropy, while the entropy of intermediate layers remains relatively high. Motivated by this entropy asymmetry, we propose Latent Exploration Decoding (LED), a depth-conditioned decoding strategy. LED aggregates intermediate posteriors via cumulative sum and selects depth configurations with maximal entropy as exploration candidates. Without additional training or parameters, LED consistently improves pass@1 and pass@16 accuracy by 0.61 and 1.03 percentage points across multiple reasoning benchmarks and models. Project page: https://GitHub.com/Xiaomi-Research/LED.

Method: Formalizes Strategic Language Search as a two-player zero-sum extensive form game, proposes Game of Thought (GoT) framework using game-theoretic techniques to approximate Nash equilibrium strategies

Result: GoT consistently improves worst-case performance compared to direct prompting and heuristic-guided search methods across all tested settings

Conclusion: Game-theoretic approaches like GoT can effectively enhance LLMs’ information-seeking capabilities in adversarial scenarios while maintaining worst-case performance guarantees

Abstract: Large Language Models (LLMs) are increasingly deployed in real-world scenarios where they may lack sufficient information to complete a given task. In such settings, the ability to actively seek out missing information becomes a critical capability. Existing approaches to enhancing this ability often rely on simplifying assumptions that degrade \textit{worst-case} performance. This is an issue with serious implications in high-stakes applications. In this work, we use the game of Twenty Questions to evaluate the information-seeking ability of LLMs. We introduce and formalize its adversarial counterpart, the Strategic Language Search (SLS) problem along with its variants as a two-player zero-sum extensive form game. We propose Game of Thought (GoT), a framework that applies game-theoretic techniques to approximate a Nash equilibrium (NE) strategy for the restricted variant of the game. Empirical results demonstrate that our approach consistently improves worst-case performance compared to (1) direct prompting-based methods and (2) heuristic-guided search methods across all tested settings.

Method: Proposes ARTIS framework that enables test-time exploration through simulated interactions prior to real-world execution. Introduces risk-aware tool simulator that emphasizes fidelity on failure-inducing actions via targeted data generation and rebalanced training to capture rare but high-impact failure modes.

Result: Experiments on multi-turn and multi-step agentic benchmarks show that iterative simulation substantially improves agent reliability, and risk-aware simulation is essential for consistently realizing these gains across models and tasks.

Conclusion: ARTIS provides an effective framework for improving agentic decision-making reliability by decoupling exploration from commitment through simulated interactions, with risk-aware simulation being crucial for capturing rare failure modes and achieving consistent performance gains.

Abstract: Current test-time scaling (TTS) techniques enhance large language model (LLM) performance by allocating additional computation at inference time, yet they remain insufficient for agentic settings, where actions directly interact with external environments and their effects can be irreversible and costly. We propose \emph{\name}, \emph{\underline{A}gentic \underline{R}isk-Aware \underline{T}est-Time Scaling via \underline{I}terative \underline{S}imulation}, a framework that decouples exploration from commitment by enabling test-time exploration through simulated interactions prior to real-world execution. This design allows extending inference-time computation to improve action-level reliability and robustness without incurring environmental risk. We further show that naive LLM-based simulators struggle to capture rare but high-impact failure modes, substantially limiting their effectiveness for agentic decision making. To address this limitation, we introduce a \emph{risk-aware tool simulator} that emphasizes fidelity on failure-inducing actions via targeted data generation and rebalanced training. Experiments on multi-turn and multi-step agentic benchmarks demonstrate that iterative simulation substantially improves agent reliability, and that risk-aware simulation is essential for consistently realizing these gains across models and tasks.

Method: Manually digitized academic materials from Arabic medical professionals, extensive preprocessing, split into training/test sets, used expert human evaluation and LLM-as-a-judge for quality assessment.

Result: Created diverse, high-quality dataset spanning 19 specialties and 5 difficulty levels; benchmarked 8 SOTA models (GPT-5, Gemini 2.0 Flash, Claude 4-Sonnet) showing need for domain-specific enhancements.

Conclusion: Dataset and evaluation scripts released to broaden medical data benchmarks, expand LLM evaluation suites, and enhance multilingual capabilities for clinical deployment.

Abstract: Arabic remains one of the most underrepresented languages in natural language processing research, particularly in medical applications, due to the limited availability of open-source data and benchmarks. The lack of resources hinders efforts to evaluate and advance the multilingual capabilities of Large Language Models (LLMs). In this paper, we introduce MedAraBench, a large-scale dataset consisting of Arabic multiple-choice question-answer pairs across various medical specialties. We constructed the dataset by manually digitizing a large repository of academic materials created by medical professionals in the Arabic-speaking region. We then conducted extensive preprocessing and split the dataset into training and test sets to support future research efforts in the area. To assess the quality of the data, we adopted two frameworks, namely expert human evaluation and LLM-as-a-judge. Our dataset is diverse and of high quality, spanning 19 specialties and five difficulty levels. For benchmarking purposes, we assessed the performance of eight state-of-the-art open-source and proprietary models, such as GPT-5, Gemini 2.0 Flash, and Claude 4-Sonnet. Our findings highlight the need for further domain-specific enhancements. We release the dataset and evaluation scripts to broaden the diversity of medical data benchmarks, expand the scope of evaluation suites for LLMs, and enhance the multilingual capabilities of models for deployment in clinical settings.

Method: Focuses on two information-theoretic measures: Normalized Branching Factor (NBF) derived from entropy, and KL divergence between steered activations and targeted concepts in vocabulary space. Uses LLM-generated annotations as ground truth based on reliability study showing high inter-judge agreement between architecturally distinct LLMs.

Result: Mechanistic signals (entropy preservation and KL alignment) provide meaningful predictive power for identifying successful steering and estimating failure probability. Also introduces stronger evaluation baseline for Contrastive Activation Addition (CAA) and Sparse Autoencoder-based steering methods.

Conclusion: Internal model signals can effectively diagnose reliability of activation steering in LLMs, with structured entropy preservation and coherent KL alignment across decoding steps correlating with successful steering outcomes.

Abstract: Activation-based steering enables Large Language Models (LLMs) to exhibit targeted behaviors by intervening on intermediate activations without retraining. Despite its widespread use, the mechanistic factors that govern when steering succeeds or fails remain poorly understood, as prior work has relied primarily on black-box outputs or LLM-based judges. In this study, we investigate whether the reliability of steering can be diagnosed using internal model signals. We focus on two information-theoretic measures: the entropy-derived Normalized Branching Factor (NBF), and the Kullback-Leibler (KL) divergence between steered activations and targeted concepts in the vocabulary space. We hypothesize that effective steering corresponds to structured entropy preservation and coherent KL alignment across decoding steps. Building on a reliability study demonstrating high inter-judge agreement between two architecturally distinct LLMs, we use LLM-generated annotations as ground truth and show that these mechanistic signals provide meaningful predictive power for identifying successful steering and estimating failure probability. We further introduce a stronger evaluation baseline for Contrastive Activation Addition (CAA) and Sparse Autoencoder-based steering, the two most widely adopted activation-steering methods.

Method: Proposes BBPE16, a UTF-16-based byte-level BPE tokenizer that represents most modern scripts with uniform 2-byte code units, improving cross-lingual token sharing.

Result: BBPE16 achieves comparable or better accuracy across monolingual, bilingual, and trilingual ASR; reduces Chinese token counts by up to 10.4% and decoding iterations by up to 10.3%.

Conclusion: BBPE16 offers practical advantages for multilingual ASR by reducing computational load, memory usage, and speeding up fine-tuning and inference while maintaining language-agnostic properties.

Abstract: Multilingual automatic speech recognition (ASR) requires tokenization that efficiently covers many writing systems. Byte-level BPE (BBPE) using UTF-8 is widely adopted for its language-agnostic design and full Unicode coverage, but its variable-length encoding inflates token sequences for non-Latin scripts, such as Chinese, Japanese, and Korean (CJK). Longer sequences increase computational load and memory use. We propose BBPE16, a UTF-16-based BBPE tokenizer that represents most modern scripts with a uniform 2-byte code unit. BBPE16 preserves BBPE’s language-agnostic properties while substantially improving cross-lingual token sharing. Across monolingual, bilingual, and trilingual ASR, and in a multilingual continual-learning setup, BBPE16 attains comparable or better accuracy; for Chinese, it reduces token counts by up to 10.4% and lowers decoding iterations by up to 10.3%. These reductions speed up fine-tuning and inference and decrease memory usage, making BBPE16 a practical tokenization choice for multilingual ASR.

Method: Two-stage framework: 1) Coarse-grained group reallocation partitions context into groups and dynamically assigns compression rates based on inter-group Marginal Information Gain (MIG). 2) Fine-grained token merging fuses tokens within groups using intra-group MIG-based weighting to preserve key semantics while reducing redundancy.

Result: Extensive experiments across QA and summarization tasks show COMI outperforms existing baselines by large margins, achieving ~25-point EM improvement under 32x compression with Qwen2-7B on NaturalQuestions.

Conclusion: COMI effectively addresses LLM context compression challenges by jointly optimizing semantic relevance and diversity through MIG-guided coarse-to-fine compression, enabling efficient long-context processing.

Abstract: Large Language Models (LLMs) have demonstrated exceptional capabilities across diverse tasks. However, their deployment in long context scenarios remains hindered by computational inefficiency and information redundancy. Context compression methods address these challenges by significantly reducing input length and eliminating redundancy. We propose COMI, a coarse-to-fine adaptive context compression framework that jointly optimizes for semantic relevance and diversity under high compression rates. We introduce Marginal Information Gain (MIG), a metric defined as the relevance of a unit to the input query minus its semantic redundancy with other units, guiding the compression process to prioritize information that is both relevant and low redundant. The framework operates in two stages: (1) Coarse-Grained Group Reallocation, where the context is partitioned into groups and dynamically assigned compression rates based on inter-group MIG, ensuring compression budgets align with information value distribution; and (2) Fine-Grained Token Merging, where tokens within each group are fused via an intra-group MIG-based weighting mechanism, thereby preserving key semantics while avoiding the accumulation of redundancy. Extensive experiments across question-answering (e.g., NaturalQuestions, 2WikiMQA, HotpotQA and NarrativeQA), summarization (e.g., MultiNews) with various backbones (e.g., LLaMA-2-7B, Qwen2-7B) show that COMI outperforms existing baselines by a large margin, e.g., approximately 25-point Exact Match (EM) improvement under 32x compression constraint with Qwen2-7B on NaturalQuestions.

Method: SafePred establishes a risk-to-decision loop with two key abilities: 1) Short- and long-term risk prediction using safety policies and world model prediction to generate semantic risk representations and prune high-risk actions; 2) Decision optimization through step-level interventions and task-level re-planning to translate predicted risks into safe decision guidance.

Result: Extensive experiments show SafePred significantly reduces high-risk behaviors, achieving over 97.6% safety performance and improving task utility by up to 21.4% compared with reactive baselines.

Conclusion: The predictive guardrail approach effectively addresses limitations of reactive methods by proactively preventing both short- and long-term risks in computer-using agents, establishing a robust safety framework for real-world deployment.

Abstract: With the widespread deployment of Computer-using Agents (CUAs) in complex real-world environments, prevalent long-term risks often lead to severe and irreversible consequences. Most existing guardrails for CUAs adopt a reactive approach, constraining agent behavior only within the current observation space. While these guardrails can prevent immediate short-term risks (e.g., clicking on a phishing link), they cannot proactively avoid long-term risks: seemingly reasonable actions can lead to high-risk consequences that emerge with a delay (e.g., cleaning logs leads to future audits being untraceable), which reactive guardrails cannot identify within the current observation space. To address these limitations, we propose a predictive guardrail approach, with the core idea of aligning predicted future risks with current decisions. Based on this approach, we present SafePred, a predictive guardrail framework for CUAs that establishes a risk-to-decision loop to ensure safe agent behavior. SafePred supports two key abilities: (1) Short- and long-term risk prediction: by using safety policies as the basis for risk prediction, SafePred leverages the prediction capability of the world model to generate semantic representations of both short-term and long-term risks, thereby identifying and pruning actions that lead to high-risk states; (2) Decision optimization: translating predicted risks into actionable safe decision guidances through step-level interventions and task-level re-planning. Extensive experiments show that SafePred significantly reduces high-risk behaviors, achieving over 97.6% safety performance and improving task utility by up to 21.4% compared with reactive baselines.

Method: Three key techniques: 1) Two-stage fine-tuning with low-rank adaptations for prompt adaptation, 2) Score alignment for distribution consistency, 3) Uncertainty-aware self-training using unlabeled data with pseudo-labels.

Result: In 32-data setting, all techniques improve performance, achieving 91.2% of full-data performance with only ~1,000 labeled samples. Score alignment achieves SOTA results in full-data setting when integrated into DualBERT.

Conclusion: Proposed techniques effectively address data scarcity in AES, with score alignment providing consistent improvements across both limited and full-data settings.

Abstract: Automated Essay Scoring (AES) plays a crucial role in education by providing scalable and efficient assessment tools. However, in real-world settings, the extreme scarcity of labeled data severely limits the development and practical adoption of robust AES systems. This study proposes a novel approach to enhance AES performance in both limited-data and full-data settings by introducing three key techniques. First, we introduce a Two-Stage fine-tuning strategy that leverages low-rank adaptations to better adapt an AES model to target prompt essays. Second, we introduce a Score Alignment technique to improve consistency between predicted and true score distributions. Third, we employ uncertainty-aware self-training using unlabeled data, effectively expanding the training set with pseudo-labeled samples while mitigating label noise propagation. We implement above three key techniques on DualBERT. We conduct extensive experiments on the ASAP++ dataset. As a result, in the 32-data setting, all three key techniques improve performance, and their integration achieves 91.2% of the full-data performance trained on approximately 1,000 labeled samples. In addition, the proposed Score Alignment technique consistently improves performance in both limited-data and full-data settings: e.g., it achieves state-of-the-art results in the full-data setting when integrated into DualBERT.

Method: WorldCup treats sampling as a natural communication channel and embeds message bits directly into token selection via a hierarchical competition mechanism guided by complementary signals. It uses entropy-aware modulation to preserve generation quality and supports robust message recovery through confidence-aware decoding.

Result: Comprehensive experiments show WorldCup achieves a strong balance across capacity, detectability, robustness, text quality, and decoding efficiency, consistently outperforming prior baselines.

Conclusion: WorldCup lays a solid foundation for future LLM watermarking studies by providing a more direct and efficient approach to multi-bit watermarking that maintains text quality while enabling richer provenance encoding.

Abstract: As large language models (LLMs) generate increasingly human-like text, watermarking offers a promising solution for reliable attribution beyond mere detection. While multi-bit watermarking enables richer provenance encoding, existing methods largely extend zero-bit schemes through seed-driven steering, leading to indirect information flow, limited effective capacity, and suboptimal decoding. In this paper, we propose WorldCup, a multi-bit watermarking framework for LLMs that treats sampling as a natural communication channel and embeds message bits directly into token selection via a hierarchical competition mechanism guided by complementary signals. Moreover, WorldCup further adopts entropy-aware modulation to preserve generation quality and supports robust message recovery through confidence-aware decoding. Comprehensive experiments show that WorldCup achieves a strong balance across capacity, detectability, robustness, text quality, and decoding efficiency, consistently outperforming prior baselines and laying a solid foundation for future LLM watermarking studies.

Method: Zero2Text uses recursive online alignment that synergizes LLM priors with dynamic ridge regression to iteratively align text generation to target embeddings on-the-fly, without requiring training data or static datasets.

Result: Extensive experiments show Zero2Text outperforms baselines significantly. On MS MARCO against OpenAI victim model, it achieves 1.8x higher ROUGE-L and 6.4x higher BLEU-2 scores, recovering sentences from unknown domains without leaked data pairs. Standard defenses like differential privacy fail against this adaptive threat.

Conclusion: Zero2Text dismantles barriers in embedding inversion attacks by providing an effective training-free framework that works in strict black-box and cross-domain settings, highlighting vulnerabilities in current vector database privacy protections.

Abstract: The proliferation of retrieval-augmented generation (RAG) has established vector databases as critical infrastructure, yet they introduce severe privacy risks via embedding inversion attacks. Existing paradigms face a fundamental trade-off: optimization-based methods require computationally prohibitive queries, while alignment-based approaches hinge on the unrealistic assumption of accessible in-domain training data. These constraints render them ineffective in strict black-box and cross-domain settings. To dismantle these barriers, we introduce Zero2Text, a novel training-free framework based on recursive online alignment. Unlike methods relying on static datasets, Zero2Text synergizes LLM priors with a dynamic ridge regression mechanism to iteratively align generation to the target embedding on-the-fly. We further demonstrate that standard defenses, such as differential privacy, fail to effectively mitigate this adaptive threat. Extensive experiments across diverse benchmarks validate Zero2Text; notably, on MS MARCO against the OpenAI victim model, it achieves 1.8x higher ROUGE-L and 6.4x higher BLEU-2 scores compared to baselines, recovering sentences from unknown domains without a single leaked data pair.

Method: Proposes a topology-aware structural tokenizer that maps each graph topology into a single special token <SOG_k>, creating a unified token space. Constructs hybrid structure QA corpora to align structural tokens with existing text tokens.

Result: Achieves 9.9% to 41.4% performance improvement on five graph-level benchmarks compared to baselines. Shows interpretability, consistency, and flexible extension to node-level tasks for both global and local structural understanding.

Conclusion: SOG enables LLMs to understand, generate, and reason about graphs concisely and accurately through structural tokens, bridging the gap between text and graph modalities.

Abstract: Large language models show great potential in unstructured data understanding, but still face significant challenges with graphs due to their structural hallucination. Existing approaches mainly either verbalize graphs into natural language, which leads to excessive token consumption and scattered attention, or transform graphs into trainable continuous embeddings (i.e., soft prompt), but exhibit severe misalignment with original text tokens. To solve this problem, we propose to incorporate one special token <SOG_k> to fully represent the Structure Of Graph within a unified token space, facilitating explicit topology input and structural information sharing. Specifically, we propose a topology-aware structural tokenizer that maps each graph topology into a highly selective single token. Afterwards, we construct a set of hybrid structure Question-Answering corpora to align new structural tokens with existing text tokens. With this approach, <SOG_k> empowers LLMs to understand, generate, and reason in a concise and accurate manner. Extensive experiments on five graph-level benchmarks demonstrate the superiority of our method, achieving a performance improvement of 9.9% to 41.4% compared to the baselines while exhibiting interpretability and consistency. Furthermore, our method provides a flexible extension to node-level tasks, enabling both global and local structural understanding. The codebase is publicly available at https://github.com/Jingyao-Wu/SOG.

Method: The authors adopt a data-centric perspective to systematically investigate how data distribution impacts compression quality across two dimensions: input data and intrinsic data (model’s internal pretrained knowledge). They evaluate semantic integrity of compressed representations using an autoencoder-based framework and analyze relationships between entropy measures and compression quality.

Result: Experimental results show: (1) encoder-measured input entropy negatively correlates with compression quality, while decoder-measured entropy shows no significant relationship under frozen-decoder settings; (2) the gap between intrinsic data of encoder and decoder significantly diminishes compression gains, which is hard to mitigate.

Conclusion: The study provides the first systematic investigation of data distribution’s impact on context compression quality. Based on findings, the authors present practical guidelines to optimize compression gains, highlighting the importance of considering both input data characteristics and encoder-decoder alignment in pretrained knowledge.

Abstract: The deployment of Large Language Models (LLMs) in long-context scenarios is hindered by computational inefficiency and significant information redundancy. Although recent advancements have widely adopted context compression to address these challenges, existing research only focus on model-side improvements, the impact of the data distribution itself on context compression remains largely unexplored. To bridge this gap, we are the first to adopt a data-centric perspective to systematically investigate how data distribution impacts compression quality, including two dimensions: input data and intrinsic data (i.e., the model’s internal pretrained knowledge). We evaluate the semantic integrity of compressed representations using an autoencoder-based framework to systematically investigate it. Our experimental results reveal that: (1) encoder-measured input entropy negatively correlates with compression quality, while decoder-measured entropy shows no significant relationship under a frozen-decoder setting; and (2) the gap between intrinsic data of the encoder and decoder significantly diminishes compression gains, which is hard to mitigate. Based on these findings, we further present practical guidelines to optimize compression gains.

Method: Systematic study exploring MLLMs for code understanding by representing source code as rendered images, adjusting resolution for compression, and evaluating performance across various compression ratios.

Result: MLLMs achieve up to 8x token compression while effectively understanding code, leverage visual cues like syntax highlighting for improved code completion under 4x compression, and show exceptional resilience in tasks like clone detection.

Conclusion: Image-modality code representation offers promising pathway for more efficient inference in code understanding, though current MLLMs have limitations that need addressing.

Abstract: Large Language Models (LLMs) have achieved remarkable success in source code understanding, yet as software systems grow in scale, computational efficiency has become a critical bottleneck. Currently, these models rely on a text-based paradigm that treats source code as a linear sequence of tokens, which leads to a linear increase in context length and associated computational costs. The rapid advancement of Multimodal LLMs (MLLMs) introduces an opportunity to optimize efficiency by representing source code as rendered images. Unlike text, which is difficult to compress without losing semantic meaning, the image modality is inherently suitable for compression. By adjusting resolution, images can be scaled to a fraction of their original token cost while remaining recognizable to vision-capable models. To explore the feasibility of this approach, we conduct the first systematic study on the effectiveness of MLLMs for code understanding. Our experiments reveal that: (1) MLLMs can effectively understand code with substantial token reduction, achieving up to 8x compression; (2) MLLMs can effectively leverage visual cues such as syntax highlighting, improving code completion performance under 4x compression; and (3) Code-understanding tasks like clone detection exhibit exceptional resilience to visual compression, with some compression ratios even slightly outperforming raw text inputs. Our findings highlight both the potential and current limitations of MLLMs in code understanding, which points out a shift toward image-modality code representation as a pathway to more efficient inference.

Method: Proposes sentence curves - spline curves whose control points affect multiple words in a sentence. Extends diffusion-based language models to predict these sentence curves instead of static word embeddings, promoting global structure modeling through regularization effects.

Result: SCLM achieves state-of-the-art performance among diffusion-based language models on IWSLT14 and WMT14 datasets, shows stable training without knowledge distillation, and demonstrates promising potential on LM1B compared to discrete diffusion-based LMs.

Conclusion: Sentence curve representation effectively addresses the limitation of static word embeddings by enabling better global structure modeling in language models, offering a promising direction for improving diffusion-based language modeling.

Abstract: Language models (LMs) are a central component of modern AI systems, and diffusion-based language models (DLMs) have recently emerged as a competitive alternative. Both paradigms rely on word embeddings not only to represent the input sentence, but also to represent the target sentence that backbone models are trained to predict. We argue that such static embedding of the target word is insensitive to neighboring words, encouraging locally accurate word prediction while neglecting global structure across the target sentence. To address this limitation, we propose a continuous sentence representation, termed sentence curve, defined as a spline curve whose control points affect multiple words in the sentence. Based on this representation, we introduce sentence curve language model (SCLM), which extends DLMs to predict sentence curves instead of the static word embeddings. We theoretically show that sentence curve prediction induces a regularization effect that promotes global structure modeling, and characterize how different sentence curve types affect this behavior. Empirically, SCLM achieves SOTA performance among DLMs on IWSLT14 and WMT14, shows stable training without burdensome knowledge distillation, and demonstrates promising potential compared to discrete DLMs on LM1B.

Method: AXE treats HTML DOM as a tree requiring pruning rather than text. It uses a specialized pruning mechanism to remove boilerplate, creating distilled context for a small 0.6B LLM. Grounded XPath Resolution (GXR) ensures traceability by linking extractions to source nodes.

Result: AXE achieves state-of-the-art zero-shot performance with 88.1% F1 score on SWDE dataset, outperforming larger fully-trained alternatives despite its small model size and low computational footprint.

Conclusion: AXE provides a practical, cost-effective solution for large-scale web information extraction by combining efficient DOM pruning with small LLMs and maintaining traceability through GXR.

Abstract: Extracting structured data from the web is often a trade-off between the brittle nature of manual heuristics and the prohibitive cost of Large Language Models. We introduce AXE (Adaptive X-Path Extractor), a pipeline that rethinks this process by treating the HTML DOM as a tree that needs pruning rather than just a wall of text to be read. AXE uses a specialized “pruning” mechanism to strip away boilerplate and irrelevant nodes, leaving behind a distilled, high-density context that allows a tiny 0.6B LLM to generate precise, structured outputs. To keep the model honest, we implement Grounded XPath Resolution (GXR), ensuring every extraction is physically traceable to a source node. Despite its low footprint, AXE achieves state-of-the-art zero-shot performance, outperforming several much larger, fully-trained alternatives with an F1 score of 88.1% on the SWDE dataset. By releasing our specialized adaptors, we aim to provide a practical, cost-effective path for large-scale web information extraction.

Method: Partitions context into segments, encodes them with query in parallel. High-relevance segments are fully retained (close reading), low-relevance ones are query-guided compressed into summary vectors (skimming). Uses contrastive learning to refine decision boundaries.

Result: Outperforms existing baselines on multiple QA and summarization benchmarks across two backbones, with up to 12x end-to-end speedup on long inputs (average 16K, max 32K length).

Conclusion: RAM effectively addresses long-context challenges through human-inspired adaptive reading, achieving both performance gains and computational efficiency while maintaining interpretability.

Abstract: Large Language Models (LLMs) demonstrate exceptional capability across diverse tasks. However, their deployment in long-context scenarios is hindered by two challenges: computational inefficiency and redundant information. We propose RAM (Read As HuMan), a context compression framework that adopts an adaptive hybrid reading strategy, to address these challenges. Inspired by human reading behavior (i.e., close reading important content while skimming less relevant content), RAM partitions the context into segments and encodes them with the input query in parallel. High-relevance segments are fully retained (close reading), while low-relevance ones are query-guided compressed into compact summary vectors (skimming). Both explicit textual segments and implicit summary vectors are concatenated and fed into decoder to achieve both superior performance and natural language format interpretability. To refine the decision boundary between close reading and skimming, we further introduce a contrastive learning objective based on positive and negative query-segment pairs. Experiments demonstrate that RAM outperforms existing baselines on multiple question answering and summarization benchmarks across two backbones, while delivering up to a 12x end-to-end speedup on long inputs (average length 16K; maximum length 32K).

Method: Proposes PretrainRL framework with “debiasing then learning” principle: uses reinforcement learning during pretraining to reshape probability distributions by down-weighting high-probability falsehoods. Includes efficient negative sampling strategy to discover falsehoods and novel metrics to evaluate probabilistic state.

Result: Extensive experiments on three public benchmarks show PretrainRL significantly alleviates factual hallucinations and outperforms state-of-the-art methods.

Conclusion: PretrainRL addresses factual hallucinations at their root by integrating reinforcement learning into pretraining, effectively consolidating factual knowledge through debiasing and learning.

Abstract: Large language models (LLMs), despite their powerful capabilities, suffer from factual hallucinations where they generate verifiable falsehoods. We identify a root of this issue: the imbalanced data distribution in the pretraining corpus, which leads to a state of “low-probability truth” and “high-probability falsehood”. Recent approaches, such as teaching models to say “I don’t know” or post-hoc knowledge editing, either evade the problem or face catastrophic forgetting. To address this issue from its root, we propose \textbf{PretrainRL}, a novel framework that integrates reinforcement learning into the pretraining phase to consolidate factual knowledge. The core principle of PretrainRL is “\textbf{debiasing then learning}.” It actively reshapes the model’s probability distribution by down-weighting high-probability falsehoods, thereby making “room” for low-probability truths to be learned effectively. To enable this, we design an efficient negative sampling strategy to discover these high-probability falsehoods and introduce novel metrics to evaluate the model’s probabilistic state concerning factual knowledge. Extensive experiments on three public benchmarks demonstrate that PretrainRL significantly alleviates factual hallucinations and outperforms state-of-the-art methods.

Method: Introduces ES-MemEval benchmark evaluating five memory capabilities (information extraction, temporal reasoning, conflict detection, abstention, user modeling) and EvoEmo dataset capturing fragmented, implicit user disclosures across multiple sessions.

Result: Experiments show explicit long-term memory reduces hallucinations and enables personalization, while RAG improves factual consistency but struggles with temporal dynamics and evolving user states.

Conclusion: Current memory paradigms have limitations, motivating more robust integration of memory and retrieval for long-term personalized dialogue systems.

Abstract: Large Language Models (LLMs) have shown strong potential as conversational agents. Yet, their effectiveness remains limited by deficiencies in robust long-term memory, particularly in complex, long-term web-based services such as online emotional support. However, existing long-term dialogue benchmarks primarily focus on static and explicit fact retrieval, failing to evaluate agents in critical scenarios where user information is dispersed, implicit, and continuously evolving. To address this gap, we introduce ES-MemEval, a comprehensive benchmark that systematically evaluates five core memory capabilities: information extraction, temporal reasoning, conflict detection, abstention, and user modeling, in long-term emotional support settings, covering question answering, summarization, and dialogue generation tasks. To support the benchmark, we also propose EvoEmo, a multi-session dataset for personalized long-term emotional support that captures fragmented, implicit user disclosures and evolving user states. Extensive experiments on open-source long-context, commercial, and retrieval-augmented (RAG) LLMs show that explicit long-term memory is essential for reducing hallucinations and enabling effective personalization. At the same time, RAG improves factual consistency but struggles with temporal dynamics and evolving user states. These findings highlight both the potential and limitations of current paradigms and motivate more robust integration of memory and retrieval for long-term personalized dialogue systems.

Method: Proposes GuideWeb benchmark for automatic guide generation on real-world web UIs, with comprehensive evaluation suite measuring guide target element selection accuracy and guide text quality.

Result: GuideWeb Agent achieves 30.79% accuracy in guide target element prediction, BLEU scores of 44.94 for intent generation and 21.34 for guide-text generation, outperforming existing baselines.

Conclusion: Automatic guide generation remains challenging and requires further advances before reliable deployment in real-world settings, despite promising results from the proposed approach.

Abstract: Digital Adoption Platform (DAP) provide web-based overlays that deliver operation guidance and contextual hints to help users navigate complex websites. Although modern DAP tools enable non-experts to author such guidance, maintaining these guides remains labor-intensive because website layouts and functionalities evolve continuously, which requires repeated manual updates and re-annotation. In this work, we introduce \textbf{GuideWeb}, a new benchmark for automatic in-app guide generation on real-world web UIs. GuideWeb formulates the task as producing page-level guidance by selecting \textbf{guide target elements} grounded in the webpage and generating concise guide text aligned with user intent. We also propose a comprehensive evaluation suite that jointly measures the accuracy of guide target element selection and the quality of generated intents and guide texts. Experiments show that our proposed \textbf{GuideWeb Agent} achieves \textbf{30.79%} accuracy in guide target element prediction, while obtaining BLEU scores of \textbf{44.94} for intent generation and \textbf{21.34} for guide-text generation. Existing baselines perform substantially worse, which highlights that automatic guide generation remains challenging and that further advances are necessary before such systems can be reliably deployed in real-world settings.

Method: Implemented “Vibe Coding” approach in senior-level undergraduate NLP course with 7 labs where students used LLMs for code generation, assessed primarily through critical reflection questions. Mandatory prompt logging and reflection-based assessment were used to structure the approach.

Result: High student satisfaction (mean scores 4.4-4.6/5.0) across engagement, conceptual learning, and assessment fairness. Students valued reduced debugging cognitive load enabling deeper focus on NLP concepts. Challenges included time constraints, LLM output verification, and need for clearer task specifications.

Conclusion: When properly structured with mandatory prompt logging and reflection-based assessment, LLM-assisted learning can shift focus from syntactic fluency to conceptual mastery, preparing students for AI-augmented professional landscape.

Abstract: The rapid advancement of Large Language Models (LLMs) presents both challenges and opportunities for Natural Language Processing (NLP) education. This paper introduces ``Vibe Coding,’’ a pedagogical approach that leverages LLMs as coding assistants while maintaining focus on conceptual understanding and critical thinking. We describe the implementation of this approach in a senior-level undergraduate NLP course, where students completed seven labs using LLMs for code generation while being assessed primarily on conceptual understanding through critical reflection questions. Analysis of end-of-course feedback from 19 students reveals high satisfaction (mean scores 4.4-4.6/5.0) across engagement, conceptual learning, and assessment fairness. Students particularly valued the reduced cognitive load from debugging, enabling deeper focus on NLP concepts. However, challenges emerged around time constraints, LLM output verification, and the need for clearer task specifications. Our findings suggest that when properly structured with mandatory prompt logging and reflection-based assessment, LLM-assisted learning can shift focus from syntactic fluency to conceptual mastery, preparing students for an AI-augmented professional landscape.

Method: Builds on HippoRAG 2 architecture with three key innovations: (1) Symbolic Anchoring for weak entity constraints, (2) Query-Aware Dynamic Edge Weighting to modulate graph structure based on query intent, and (3) Key-Fact Passage Weight Enhancement to anchor random walks to likely evidence.

Result: Outperforms state-of-the-art baselines across four multi-hop benchmarks, showing substantial improvements in reasoning completeness (recovering entire evidence paths) beyond modest recall gains.

Conclusion: CatRAG effectively bridges the gap between retrieving partial context and enabling fully grounded reasoning by making knowledge graph traversal query-adaptive.

Abstract: Recent advances in Retrieval-Augmented Generation (RAG) have shifted from simple vector similarity to structure-aware approaches like HippoRAG, which leverage Knowledge Graphs (KGs) and Personalized PageRank (PPR) to capture multi-hop dependencies. However, these methods suffer from a “Static Graph Fallacy”: they rely on fixed transition probabilities determined during indexing. This rigidity ignores the query-dependent nature of edge relevance, causing semantic drift where random walks are diverted into high-degree “hub” nodes before reaching critical downstream evidence. Consequently, models often achieve high partial recall but fail to retrieve the complete evidence chain required for multi-hop queries. To address this, we propose CatRAG, Context-Aware Traversal for robust RAG, a framework that builds on the HippoRAG 2 architecture and transforms the static KG into a query-adaptive navigation structure. We introduce a multi-faceted framework to steer the random walk: (1) Symbolic Anchoring, which injects weak entity constraints to regularize the random walk; (2) Query-Aware Dynamic Edge Weighting, which dynamically modulates graph structure, to prune irrelevant paths while amplifying those aligned with the query’s intent; and (3) Key-Fact Passage Weight Enhancement, a cost-efficient bias that structurally anchors the random walk to likely evidence. Experiments across four multi-hop benchmarks demonstrate that CatRAG consistently outperforms state of the art baselines. Our analysis reveals that while standard Recall metrics show modest gains, CatRAG achieves substantial improvements in reasoning completeness, the capacity to recover the entire evidence path without gaps. These results reveal that our approach effectively bridges the gap between retrieving partial context and enabling fully grounded reasoning. Resources are available at https://github.com/kwunhang/CatRAG.

Method: Introduces Moe-Ctc, a Mixture-of-Experts architecture with intermediate CTC supervision. During training, accent-aware routing encourages experts to capture accent-specific patterns, which gradually transitions to label-free routing for inference. Each expert has its own CTC head to align routing with transcription quality, and a routing-augmented loss stabilizes optimization.

Result: Experiments on the Mcv-Accent benchmark show consistent gains across both seen and unseen accents in low- and high-resource conditions, achieving up to 29.3% relative WER reduction over strong FastConformer baselines.

Conclusion: Moe-Ctc effectively addresses accent robustness in ASR by jointly promoting expert specialization and generalization through a novel Mixture-of-Experts architecture with intermediate CTC supervision, demonstrating significant improvements over existing approaches.

Abstract: Accented speech remains a persistent challenge for automatic speech recognition (ASR), as most models are trained on data dominated by a few high-resource English varieties, leading to substantial performance degradation for other accents. Accent-agnostic approaches improve robustness yet struggle with heavily accented or unseen varieties, while accent-specific methods rely on limited and often noisy labels. We introduce Moe-Ctc, a Mixture-of-Experts architecture with intermediate CTC supervision that jointly promotes expert specialization and generalization. During training, accent-aware routing encourages experts to capture accent-specific patterns, which gradually transitions to label-free routing for inference. Each expert is equipped with its own CTC head to align routing with transcription quality, and a routing-augmented loss further stabilizes optimization. Experiments on the Mcv-Accent benchmark demonstrate consistent gains across both seen and unseen accents in low- and high-resource conditions, achieving up to 29.3% relative WER reduction over strong FastConformer baselines.

Method: Proposes Orthogonal Hierarchical Decomposition (OHD) framework with Orthogonal Tree Induction (OTI) that decomposes irregular tables into column and row trees to capture vertical/horizontal hierarchical dependencies. Uses dual-pathway association to reconstruct semantic lineage of each cell, and incorporates LLM as semantic arbitrator to align multi-level semantic information.

Result: OHD consistently outperforms existing representation paradigms across multiple evaluation metrics on two complex table question answering benchmarks: AITQA and HiTab.

Conclusion: The OHD framework effectively addresses challenges in representing complex tables for LLMs by preserving hierarchical structures and capturing cross-dimensional dependencies through orthogonal tree decomposition.

Abstract: Complex tables with multi-level headers, merged cells and heterogeneous layouts pose persistent challenges for LLMs in both understanding and reasoning. Existing approaches typically rely on table linearization or normalized grid modeling. However, these representations struggle to explicitly capture hierarchical structures and cross-dimensional dependencies, which can lead to misalignment between structural semantics and textual representations for non-standard tables. To address this issue, we propose an Orthogonal Hierarchical Decomposition (OHD) framework that constructs structure-preserving input representations of complex tables for LLMs. OHD introduces an Orthogonal Tree Induction (OTI) method based on spatial–semantic co-constraints, which decomposes irregular tables into a column tree and a row tree to capture vertical and horizontal hierarchical dependencies, respectively. Building on this representation, we design a dual-pathway association protocol to symmetrically reconstruct semantic lineage of each cell, and incorporate an LLM as a semantic arbitrator to align multi-level semantic information. We evaluate OHD framework on two complex table question answering benchmarks, AITQA and HiTab. Experimental results show that OHD consistently outperforms existing representation paradigms across multiple evaluation metrics.

Method: Introduces Embedding-Virtualized Knowledge (EVK) that characterizes model knowledge through controlled perturbations in embedding space. Constructs EVK-Bench benchmark to quantify knowledge drift. Proposes EVK-Align module that constrains embedding-level drift during editing and integrates with existing methods.

Result: EVK enables more comprehensive evaluation of knowledge editing effects, revealing impacts not captured by conventional metrics. EVK-Align significantly improves knowledge preservation without sacrificing editing accuracy when integrated with existing editing methods.

Conclusion: EVK provides a novel embedding-space approach to evaluate knowledge editing more comprehensively, addressing limitations of sample-based metrics. The framework offers both evaluation tools and practical solutions for better knowledge preservation during editing.

Abstract: Knowledge editing methods for large language models are commonly evaluated using predefined benchmarks that assess edited facts together with a limited set of related or neighboring knowledge. While effective, such evaluations remain confined to finite, dataset-bounded samples, leaving the broader impact of editing on the model’s knowledge system insufficiently understood. To address this gap, we introduce Embedding-Virtualized Knowledge (EVK) that characterizes model knowledge through controlled perturbations in embedding space, enabling the exploration of a substantially broader and virtualized knowledge region beyond explicit data annotations. Based on EVK, we construct an embedding-level evaluation benchmark EVK-Bench that quantifies potential knowledge drift induced by editing, revealing effects that are not captured by conventional sample-based metrics. Furthermore, we propose a plug-and-play EVK-Align module that constrains embedding-level knowledge drift during editing and can be seamlessly integrated into existing editing methods. Experiments demonstrate that our approach enables more comprehensive evaluation while significantly improving knowledge preservation without sacrificing editing accuracy.

Method: Self-sampling framework using activation steering to induce style-aligned, variable-length reasoning traces from target LLMs without teacher guidance. Uses filtered data by gold answers for SFT with dual-cognitive system and progressive compression curriculum. Also explores self-evolution regime using prediction-consistent data without gold answers.

Result: Extensive experiments on math benchmarks show stable improvements for both general and R1-style LLMs. Cross-domain generalization tests in medicine demonstrate effectiveness. The method yields efficient CoT learning with reduced reasoning redundancy.

Conclusion: The proposed self-sampling framework successfully enables LLMs to acquire fast-thinking capabilities, addressing the redundancy problem in CoT reasoning while eliminating the need for scarce high-quality supervision data.

Abstract: Large language models (LLMs) equipped with chain-of-thought (CoT) achieve strong performance and offer a window into LLM behavior. However, recent evidence suggests that improvements in CoT capabilities often come with redundant reasoning processes, motivating a key question: Can LLMs acquire a fast-thinking mode analogous to human System 1 reasoning? To explore this, our study presents a self-sampling framework based on activation steering for efficient CoT learning. Our method can induce style-aligned and variable-length reasoning traces from target LLMs themselves without any teacher guidance, thereby alleviating a central bottleneck of SFT-based methods-the scarcity of high-quality supervision data. Using filtered data by gold answers, we perform SFT for efficient CoT learning with (i) a human-like dual-cognitive system, and (ii) a progressive compression curriculum. Furthermore, we explore a self-evolution regime in which SFT is driven solely by prediction-consistent data of variable-length variants, eliminating the need for gold answers. Extensive experiments on math benchmarks, together with cross-domain generalization tests in medicine, show that our method yields stable improvements for both general and R1-style LLMs. Our data and model checkpoints can be found at https://github.com/DYR1/S3-CoT.

Method: Anchored on the onset of reflection behavior and traced its layer-wise activation trajectory using logit lens to read token-level semantics, with targeted interventions to uncover causal relationships.

Result: Uncovered a structured progression: latent-control layers encode thinking budget semantics, semantic-pivot layers show discourse-level cues, and behavior-overt layers show rising likelihood of reflection tokens. Interventions revealed a causal chain across these stages.

Conclusion: The findings suggest a human-like meta-cognitive process progressing from latent monitoring to discourse-level regulation to overt self-reflection in LLMs.

Abstract: R1-style LLMs have attracted growing attention for their capacity for self-reflection, yet the internal mechanisms underlying such behavior remain unclear. To bridge this gap, we anchor on the onset of reflection behavior and trace its layer-wise activation trajectory. Using the logit lens to read out token-level semantics, we uncover a structured progression: (i) Latent-control layers, where an approximate linear direction encodes the semantics of thinking budget; (ii) Semantic-pivot layers, where discourse-level cues, including turning-point and summarization cues, surface and dominate the probability mass; and (iii) Behavior-overt layers, where the likelihood of reflection-behavior tokens begins to rise until they become highly likely to be sampled. Moreover, our targeted interventions uncover a causal chain across these stages: prompt-level semantics modulate the projection of activations along latent-control directions, thereby inducing competition between turning-point and summarization cues in semantic-pivot layers, which in turn regulates the sampling likelihood of reflection-behavior tokens in behavior-overt layers. Collectively, our findings suggest a human-like meta-cognitive process-progressing from latent monitoring, to discourse-level regulation, and to finally overt self-reflection. Our analysis code can be found at https://github.com/DYR1/S3-CoT.

Method: xMemory builds a hierarchy of intact memory units using a sparsity-semantics objective that guides memory split and merge operations. It disentangles memories into semantic components, organizes them into a searchable hierarchy, and retrieves top-down by selecting compact, diverse sets of themes and semantics for multi-fact queries, expanding to episodes and raw messages only when it reduces uncertainty.

Result: Experiments on LoCoMo and PerLTQA benchmarks across three latest LLMs show consistent gains in answer quality and token efficiency compared to standard approaches.

Conclusion: Retrieval for agent memory should move beyond similarity matching to operate over latent semantic components, using hierarchical organization to provide compact, diverse, and temporally coherent context for reasoning.

Abstract: Agent memory systems often adopt the standard Retrieval-Augmented Generation (RAG) pipeline, yet its underlying assumptions differ in this setting. RAG targets large, heterogeneous corpora where retrieved passages are diverse, whereas agent memory is a bounded, coherent dialogue stream with highly correlated spans that are often duplicates. Under this shift, fixed top-$k$ similarity retrieval tends to return redundant context, and post-hoc pruning can delete temporally linked prerequisites needed for correct reasoning. We argue retrieval should move beyond similarity matching and instead operate over latent components, following decoupling to aggregation: disentangle memories into semantic components, organise them into a hierarchy, and use this structure to drive retrieval. We propose xMemory, which builds a hierarchy of intact units and maintains a searchable yet faithful high-level node organisation via a sparsity–semantics objective that guides memory split and merge. At inference, xMemory retrieves top-down, selecting a compact, diverse set of themes and semantics for multi-fact queries, and expanding to episodes and raw messages only when it reduces the reader’s uncertainty. Experiments on LoCoMo and PerLTQA across the three latest LLMs show consistent gains in answer quality and token efficiency.

Method: NEAT identifies exit-associated neurons and tracks their activation patterns during reasoning to dynamically trigger early exit or suppress reflection. It operates without training or additional test-time computation by monitoring neuron-level activation dynamics.

Result: Experiments on four reasoning benchmarks across six models show NEAT achieves 22-28% average token reduction while maintaining accuracy, demonstrating effective reduction of unnecessary reasoning.

Conclusion: NEAT provides an efficient, training-free approach to mitigate overthinking in reasoning models by leveraging neuron activation patterns for early exit decisions, reducing computational overhead while preserving solution quality.

Abstract: Large Reasoning Models (LRMs) often suffer from \emph{overthinking}, a phenomenon in which redundant reasoning steps are generated after a correct solution has already been reached. Existing early reasoning exit methods primarily rely on output-level heuristics or trained probing models to skip redundant reasoning steps, thereby mitigating overthinking. However, these approaches typically require additional rollout computation or externally labeled datasets. In this paper, we propose \textbf{NEAT}, a \textbf{N}euron-based \textbf{E}arly re\textbf{A}soning exi\textbf{T} framework that monitors neuron-level activation dynamics to enable training-free early exits, without introducing additional test-time computation. NEAT identifies exit-associated neurons and tracks their activation patterns during reasoning to dynamically trigger early exit or suppress reflection, thereby reducing unnecessary reasoning while preserving solution quality. Experiments on four reasoning benchmarks across six models with different scales and architectures show that, for each model, NEAT achieves an average token reduction of 22% to 28% when averaged over the four benchmarks, while maintaining accuracy.

Method: Leverage Wikipedia’s structure where articles reference external documents, using 12 top-level topics, external references as retrieval corpus, and citation-linked statements as ground truth to create 1,100 questions across three complexity levels.

Result: Experiments show current GraphRAG pipelines help with multi-fact aggregation from moderate sources but may overemphasize high-level statements at the expense of fine-grained details, leading to weaker summarization performance.

Conclusion: WildGraphBench provides a realistic benchmark for GraphRAG evaluation, revealing limitations in current aggregation paradigms and the need for better handling of fine-grained details in summarization tasks.

Abstract: Graph-based Retrieval-Augmented Generation (GraphRAG) organizes external knowledge as a hierarchical graph, enabling efficient retrieval and aggregation of scattered evidence across multiple documents. However, many existing benchmarks for GraphRAG rely on short, curated passages as external knowledge, failing to adequately evaluate systems in realistic settings involving long contexts and large-scale heterogeneous documents. To bridge this gap, we introduce WildGraphBench, a benchmark designed to assess GraphRAG performance in the wild. We leverage Wikipedia’s unique structure, where cohesive narratives are grounded in long and heterogeneous external reference documents, to construct a benchmark reflecting real-word scenarios. Specifically, we sample articles across 12 top-level topics, using their external references as the retrieval corpus and citation-linked statements as ground truth, resulting in 1,100 questions spanning three levels of complexity: single-fact QA, multi-fact QA, and section-level summarization. Experiments across multiple baselines reveal that current GraphRAG pipelines help on multi-fact aggregation when evidence comes from a moderate number of sources, but this aggregation paradigm may overemphasize high-level statements at the expense of fine-grained details, leading to weaker performance on summarization tasks. Project page:https://github.com/BstWPY/WildGraphBench.

Method: Proposes RPG-Encoder that: 1) Encodes raw code into Repository Planning Graphs combining semantic features with dependencies, 2) Evolves topology incrementally to reduce maintenance costs, 3) Serves as unified interface for structure-aware navigation.

Result: Achieves 93.7% Acc@5 on SWE-bench Verified, exceeds best baseline by over 10% on SWE-bench Live Lite, reduces overhead by 95.7%, and achieves 98.5% reconstruction coverage on RepoCraft.

Conclusion: RPG-Encoder establishes state-of-the-art repository understanding with superior fine-grained localization, high-fidelity codebase mirroring, and closes the loop between intent and implementation.

Abstract: Current repository agents encounter a reasoning disconnect due to fragmented representations, as existing methods rely on isolated API documentation or dependency graphs that lack semantic depth. We consider repository comprehension and generation to be inverse processes within a unified cycle: generation expands intent into implementation, while comprehension compresses implementation back into intent. To address this, we propose RPG-Encoder, a framework that generalizes the Repository Planning Graph (RPG) from a static generative blueprint into a unified, high-fidelity representation. RPG-Encoder closes the reasoning loop through three mechanisms: (1) Encoding raw code into the RPG that combines lifted semantic features with code dependencies; (2) Evolving the topology incrementally to decouple maintenance costs from repository scale, reducing overhead by 95.7%; and (3) Operating as a unified interface for structure-aware navigation. In evaluations, RPG-Encoder establishes state-of-the-art repository understanding on SWE-bench Verified with 93.7% Acc@5 and exceeds the best baseline by over 10% on SWE-bench Live Lite. These results highlight our superior fine-grained localization accuracy in complex codebases. Furthermore, it achieves 98.5% reconstruction coverage on RepoCraft, confirming RPG’s high-fidelity capacity to mirror the original codebase and closing the loop between intent and implementation.

Method: Three key components: 1) hierarchical coarse-to-fine relation extraction to mitigate long-tail bias, 2) evidence-guided Chain-of-Thought feedback grounding structural suggestions in source text, and 3) semantic initialization enabling structural validation for unseen entities. LLMs and KGE modules enhance each other iteratively.

Result: Evaluated on Chinese Sustainable Development Goal reports, showing substantial improvements over LLM baselines, particularly on low-frequency relations. Framework reliably transforms unstructured policy text into validated knowledge graph triples through iterative refinement.

Conclusion: LEC-KG effectively integrates semantic and structural reasoning for domain-specific knowledge graph construction, demonstrating robust performance on challenging real-world policy documents with long-tail relation distributions.

Abstract: Constructing domain-specific knowledge graphs from unstructured text remains challenging due to heterogeneous entity mentions, long-tail relation distributions, and the absence of standardized schemas. We present LEC-KG, a bidirectional collaborative framework that integrates the semantic understanding of Large Language Models (LLMs) with the structural reasoning of Knowledge Graph Embeddings (KGE). Our approach features three key components: (1) hierarchical coarse-to-fine relation extraction that mitigates long-tail bias, (2) evidence-guided Chain-of-Thought feedback that grounds structural suggestions in source text, and (3) semantic initialization that enables structural validation for unseen entities. The two modules enhance each other iteratively-KGE provides structure-aware feedback to refine LLM extractions, while validated triples progressively improve KGE representations. We evaluate LEC-KG on Chinese Sustainable Development Goal (SDG) reports, demonstrating substantial improvements over LLM baselines, particularly on low-frequency relations. Through iterative refinement, our framework reliably transforms unstructured policy text into validated knowledge graph triples.

Method: Proposes Dynamic Decoupled Conditional Advantage (DDCA) which: 1) computes length advantages conditionally within correct-response clusters to eliminate baseline dilution, and 2) dynamically scales penalty strength using group pass rate as difficulty proxy.

Result: DDCA improves efficiency-accuracy trade-off across GSM8K, MATH500, AMC23, and AIME25, reducing generated tokens by ~60% on simpler tasks (GSM8K) and over 20% on harder benchmarks (AIME25) while maintaining or improving accuracy.

Conclusion: DDCA effectively addresses structural issues in RLVR optimization, enabling more efficient reasoning without sacrificing accuracy through conditional advantage computation and difficulty-adaptive penalties.

Abstract: Reinforcement Learning with Verifiable Rewards (RLVR) can elicit strong multi-step reasoning, yet it often encourages overly verbose traces. Moreover, naive length penalties in group-relative optimization can severely hurt accuracy. We attribute this failure to two structural issues: (i) Dilution of Length Baseline, where incorrect responses (with zero length reward) depress the group baseline and over-penalize correct solutions; and (ii) Difficulty-Penalty Mismatch, where a static penalty cannot adapt to problem difficulty, suppressing necessary reasoning on hard instances while leaving redundancy on easy ones. We propose Dynamic Decoupled Conditional Advantage (DDCA) to decouple efficiency optimization from correctness. DDCA computes length advantages conditionally within the correct-response cluster to eliminate baseline dilution, and dynamically scales the penalty strength using the group pass rate as a proxy for difficulty. Experiments on GSM8K, MATH500, AMC23, and AIME25 show that DDCA consistently improves the efficiency–accuracy trade-off relative to adaptive baselines, reducing generated tokens by approximately 60% on simpler tasks (e.g., GSM8K) versus over 20% on harder benchmarks (e.g., AIME25), thereby maintaining or improving accuracy. Code is available at https://github.com/alphadl/DDCA.

Method: Adaptation approach using existing base models (Mistral-Small-3.1, NVIDIA Nemotron Nano V2, Qwen3-1.7B) trained on substantial Hebrew-English corpora, released in multiple variants with 65k context length and tool-calling capabilities.

Result: Released three model sizes (24B, 12B, 1.7B) with base and chat variants, plus a comprehensive Hebrew benchmark suite covering Translation, Summarization, Winograd, Israeli Trivia, and Diacritization tasks.

Conclusion: The work addresses low-resource language LLM training challenges and provides a framework for adapting LLMs to other non-English languages, contributing to multilingual NLP advancement.

Abstract: Open-weight LLMs have been released by frontier labs; however, sovereign Large Language Models (for languages other than English) remain low in supply yet high in demand. Training large language models (LLMs) for low-resource languages such as Hebrew poses unique challenges. In this paper, we introduce Dicta-LM 3.0: an open-weight collection of LLMs trained on substantially-sized corpora of Hebrew and English texts. The model is released in three sizes: 24B - adapted from the Mistral-Small-3.1 base model, 12B - adapted from the NVIDIA Nemotron Nano V2 model, and 1.7B - adapted from the Qwen3-1.7B base model. We are releasing multiple variants of each model, each with a native context length of 65k tokens; base model and chat model with tool-calling support. To rigorously evaluate our models, we introduce a new benchmark suite for evaluation of Hebrew chat-LLMs, covering a diverse set of tasks including Translation, Summarization, Winograd, Israeli Trivia, and Diacritization (nikud). Our work not only addresses the intricacies of training LLMs in low-resource languages but also proposes a framework that can be leveraged for adapting other LLMs to various non-English languages, contributing to the broader field of multilingual NLP.

Method: Uses chunk-recurrent training with on-the-fly activation recomputation for O(1) activation memory, plus paged memory manager for KV cache, asynchronous CPU offloading, and page-level sparse attention to manage KV cache growth.

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

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

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

Method: Analyzed eleven categories of refusal and non-compliance behaviors across LLMs, examining their corresponding activation-space directions. Used linear steering techniques to manipulate these directions and observe resulting refusal behaviors and trade-offs.

Result: Found that refusal behaviors correspond to geometrically distinct directions in activation space, contrary to the single-direction hypothesis. However, linear steering along any refusal-related direction produces nearly identical refusal-to-over-refusal trade-offs, suggesting a shared one-dimensional control mechanism. The primary effect of different directions is not whether the model refuses, but how it refuses.

Conclusion: The account of refusal being mediated by a single activation-space direction is incomplete. While refusal behaviors are geometrically diverse, they share a common control mechanism for refusal intensity, with different directions primarily affecting the style or manner of refusal rather than the decision to refuse itself.

Abstract: Prior work argues that refusal in large language models is mediated by a single activation-space direction, enabling effective steering and ablation. We show that this account is incomplete. Across eleven categories of refusal and non-compliance, including safety, incomplete or unsupported requests, anthropomorphization, and over-refusal, we find that these refusal behaviors correspond to geometrically distinct directions in activation space. Yet despite this diversity, linear steering along any refusal-related direction produces nearly identical refusal to over-refusal trade-offs, acting as a shared one-dimensional control knob. The primary effect of different directions is not whether the model refuses, but how it refuses.

Method: Introduces GapEval, a bidirectional benchmark that quantifies the gap between understanding and generation capabilities through symmetric evaluation where each question can be answered in both image and text modalities, measuring cross-modal consistency.

Result: Experiments reveal a persistent gap between understanding and generation directions across various UMM architectures, indicating knowledge within models remains disjoint and capability emergence across modalities is unsynchronized.

Conclusion: Current unified multimodal models achieve only surface-level unification rather than deep cognitive convergence, with knowledge manipulation limitations preventing genuine integration of understanding and generation capabilities.

Abstract: Recent advances in unified multimodal models (UMM) have demonstrated remarkable progress in both understanding and generation tasks. However, whether these two capabilities are genuinely aligned and integrated within a single model remains unclear. To investigate this question, we introduce GapEval, a bidirectional benchmark designed to quantify the gap between understanding and generation capabilities, and quantitatively measure the cognitive coherence of the two “unified” directions. Each question can be answered in both modalities (image and text), enabling a symmetric evaluation of a model’s bidirectional inference capability and cross-modal consistency. Experiments reveal a persistent gap between the two directions across a wide range of UMMs with different architectures, suggesting that current models achieve only surface-level unification rather than deep cognitive convergence of the two. To further explore the underlying mechanism, we conduct an empirical study from the perspective of knowledge manipulation to illustrate the underlying limitations. Our findings indicate that knowledge within UMMs often remains disjoint. The capability emergence and knowledge across modalities are unsynchronized, paving the way for further exploration.

Method: 1) Design a past confidence-guided indicator to predict unmasked regions based on token confidence correlation across adjacent steps. 2) Propose sink-aware pruning strategy to estimate and remove redundant attention while preserving attention sinks. 3) Reuse identified sink locations across layers leveraging cross-layer consistency.

Result: Achieves more than 29x lossless speedup under 32K context length, demonstrating efficient long-context inference for diffusion LLMs without accuracy degradation.

Conclusion: Focus-dLLM provides an effective training-free attention sparsification framework that significantly improves inference efficiency for diffusion LLMs while maintaining accuracy, addressing the computational bottleneck of bidirectional attention in long-context processing.

Abstract: Diffusion Large Language Models (dLLMs) deliver strong long-context processing capability in a non-autoregressive decoding paradigm. However, the considerable computational cost of bidirectional full attention limits the inference efficiency. Although sparse attention is promising, existing methods remain ineffective. This stems from the need to estimate attention importance for tokens yet to be decoded, while the unmasked token positions are unknown during diffusion. In this paper, we present Focus-dLLM, a novel training-free attention sparsification framework tailored for accurate and efficient long-context dLLM inference. Based on the finding that token confidence strongly correlates across adjacent steps, we first design a past confidence-guided indicator to predict unmasked regions. Built upon this, we propose a sink-aware pruning strategy to accurately estimate and remove redundant attention computation, while preserving highly influential attention sinks. To further reduce overhead, this strategy reuses identified sink locations across layers, leveraging the observed cross-layer consistency. Experimental results show that our method offers more than $29\times$ lossless speedup under $32K$ context length. The code is publicly available at: https://github.com/Longxmas/Focus-dLLM

Method: Two-stage framework: 1) Self-distillation to incentivize task decomposition reasoning capability, 2) Diversity-aware reinforcement learning to restore reflective reasoning capability.

Result: D-CORE achieves robust tool-use improvements across diverse benchmarks and model scales. D-CORE-8B reaches 77.7% accuracy on BFCLv3 (5.7% improvement over best 8B model), and D-CORE-14B establishes new SOTA at 79.3%, outperforming 70B models despite being 5× smaller.

Conclusion: D-CORE effectively addresses lazy reasoning in large reasoning models by enhancing decomposition and reflective reasoning capabilities, achieving state-of-the-art performance in tool use tasks with significantly smaller models.

Abstract: Effective tool use and reasoning are essential capabilities for large reasoning models~(LRMs) to address complex real-world problems. Through empirical analysis, we identify that current LRMs lack the capability of sub-task decomposition in complex tool use scenarios, leading to Lazy Reasoning. To address this, we propose a two-stage training framework D-CORE~(\underline{\textbf{D}}ecomposing tasks and \underline{\textbf{Co}}mposing \underline{\textbf{Re}}asoning processes) that first incentivize the LRMs’ task decomposition reasoning capability via self-distillation, followed by diversity-aware reinforcement learning~(RL) to restore LRMs’ reflective reasoning capability. D-CORE achieves robust tool-use improvements across diverse benchmarks and model scales. Experiments on BFCLv3 demonstrate superiority of our method: D-CORE-8B reaches 77.7% accuracy, surpassing the best-performing 8B model by 5.7%. Meanwhile, D-CORE-14B establishes a new state-of-the-art at 79.3%, outperforming 70B models despite being 5$\times$ smaller. The source code is available at https://github.com/alibaba/EfficientAI.

Method: Proposes AR-MAP framework that leverages preference-aligned autoregressive LLMs as implicit teachers for DLLM alignment through simple weight scaling, exploiting shared architectural structure between different generation paradigms.

Result: Achieves 69.08% average score across diverse preference alignment tasks, demonstrating competitive or superior performance compared to existing DLLM-specific alignment methods.

Conclusion: AR-MAP provides an effective transfer learning approach for DLLM alignment that circumvents computational overhead and variance issues of direct alignment methods.

Abstract: Diffusion Large Language Models (DLLMs) have emerged as a powerful alternative to autoregressive models, enabling parallel token generation across multiple positions. However, preference alignment of DLLMs remains challenging due to high variance introduced by Evidence Lower Bound (ELBO)-based likelihood estimation. In this work, we propose AR-MAP, a novel transfer learning framework that leverages preference-aligned autoregressive LLMs (AR-LLMs) as implicit teachers for DLLM alignment. We reveal that DLLMs can effectively absorb alignment knowledge from AR-LLMs through simple weight scaling, exploiting the shared architectural structure between these divergent generation paradigms. Crucially, our approach circumvents the high variance and computational overhead of direct DLLM alignment and comprehensive experiments across diverse preference alignment tasks demonstrate that AR-MAP achieves competitive or superior performance compared to existing DLLM-specific alignment methods, achieving 69.08% average score across all tasks and models. Our Code is available at https://github.com/AMAP-ML/AR-MAP.

Method: Created a benchmark evaluating metalinguistic reasoning across linguistic domains (lexical, phonological, morphological, syntactic) and languages, analyzing performance using accuracy, macro F1, and comparing to majority-class and chance baselines.

Result: GPT-4o performed best but only achieved moderate accuracy (0.367); all models performed above chance but failed to beat majority-class baseline. Performance varied by linguistic domain (lexical highest, phonological lowest) and strongly correlated with digital language status and resource availability.

Conclusion: LLMs’ metalinguistic knowledge is fragmented and shaped by data availability rather than generalizable grammatical competence, highlighting the need for more diverse linguistic evaluation and development.

Abstract: Large language models (LLMs) are routinely evaluated on language use tasks, yet their knowledge of linguistic structure remains poorly understood. Existing linguistic benchmarks typically focus on narrow phenomena, emphasize high-resource languages, and rarely evaluate metalinguistic knowledge-explicit reasoning about language structure rather than language use. Using accuracy and macro F1, together with majority-class and chance baselines, we analyse overall performance and examine variation by linguistic domains and language-related factors. Our results show that metalinguistic knowledge in current LLMs is limited: GPT-4o performs best but achieves only moderate accuracy (0.367), while open-source models lag behind. All models perform above chance but fail to outperform the majority-class baseline, suggesting they capture cross-linguistic patterns but lack fine-grained grammatical distinctions. Performance varies across linguistic domains, with lexical features showing the highest accuracy and phonological features among the lowest, partially reflecting differences in online visibility. At the language level, accuracy shows a strong association with digital language status: languages with higher digital presence and resource availability are evaluated more accurately, while low-resource languages show substantially lower performance. Analyses of predictive factors confirm that resource-related indicators (Wikipedia size, corpus availability) are more informative predictors of accuracy than geographical, genealogical, or sociolinguistic factors. Together, these results suggest that LLMs’ metalinguistic knowledge is fragmented and shaped by data availability rather than generalizable grammatical competence across the world’s languages. We release our benchmark as an open-source dataset to support systematic evaluation and encourage greater global linguistic diversity in future LLMs.

Method: Human creation and verification of 110 data samples, each containing a prompt, correct answer, and wrong answer, with most questions contextualized to Sri Lankan culture and environment

Result: Successfully created the first Sinhala physical commonsense reasoning dataset with 110 verified samples, contributing to Global PIQA’s multilingual coverage

Conclusion: The dataset fills a gap for Sinhala language resources and supports development of multilingual AI systems with physical commonsense reasoning capabilities

Abstract: This paper presents the first-ever Sinhala physical common sense reasoning dataset created as part of Global PIQA. It contains 110 human-created and verified data samples, where each sample consists of a prompt, the corresponding correct answer, and a wrong answer. Most of the questions refer to the Sri Lankan context, where Sinhala is an official language.

Method: Developed AgriHubi, a domain-adapted RAG system for Finnish agriculture that integrates Finnish agricultural documents with open PORO family models, combines explicit source grounding with user feedback for iterative refinement, and was developed over eight iterations.

Result: System shows clear gains in answer completeness, linguistic accuracy, and perceived reliability; reveals practical trade-offs between response quality and latency when deploying larger models; evaluated through two user studies.

Conclusion: Provides empirical guidance for designing and evaluating domain-specific RAG systems in low-resource language settings.

Abstract: Large language models show promise for knowledge-intensive domains, yet their use in agriculture is constrained by weak grounding, English-centric training data, and limited real-world evaluation. These issues are amplified for low-resource languages, where high-quality domain documentation exists but remains difficult to access through general-purpose models. This paper presents AgriHubi, a domain-adapted retrieval-augmented generation (RAG) system for Finnish-language agricultural decision support. AgriHubi integrates Finnish agricultural documents with open PORO family models and combines explicit source grounding with user feedback to support iterative refinement. Developed over eight iterations and evaluated through two user studies, the system shows clear gains in answer completeness, linguistic accuracy, and perceived reliability. The results also reveal practical trade-offs between response quality and latency when deploying larger models. This study provides empirical guidance for designing and evaluating domain-specific RAG systems in low-resource language settings.

Method: Conducted controlled experiments across multiple models and datasets to demonstrate consistent position bias. Proposed a balanced permutation strategy that evenly distributes each score option across different positions. Aggregated scores across these balanced permutations to mitigate bias.

Result: Found consistent position bias in rubric-based LLM evaluation. Showed that aggregating scores across balanced permutations not only reveals latent position bias but also improves correlation between LLM-as-a-Judge and human judgments.

Conclusion: Rubric-based LLM-as-a-Judge is not inherently point-wise and suffers from position bias. Simple permutation-based calibration can substantially improve its reliability and alignment with human evaluation.

Abstract: Large language models (LLMs) are now widely used to evaluate the quality of text, a field commonly referred to as LLM-as-a-judge. While prior works mainly focus on point-wise and pair-wise evaluation paradigms. Rubric-based evaluation, where LLMs select a score from multiple rubrics, has received less analysis. In this work, we show that rubric-based evaluation implicitly resembles a multi-choice setting and therefore has position bias: LLMs prefer score options appearing at specific positions in the rubric list. Through controlled experiments across multiple models and datasets, we demonstrate consistent position bias. To mitigate this bias, we propose a balanced permutation strategy that evenly distributes each score option across positions. We show that aggregating scores across balanced permutations not only reveals latent position bias, but also improves correlation between the LLM-as-a-Judge and human. Our results suggest that rubric-based LLM-as-a-Judge is not inherently point-wise and that simple permutation-based calibration can substantially improve its reliability.

Method: Proposes balanced average recurrence of correspondence patterns as a new regularity measure, with a computational method to identify cognate sets lacking regularity. Uses leave-one-out validation with simulated and real data to test identification of irregular forms.

Result: Method achieves 85% accuracy with real datasets. Shows benefits of working with subsamples of large datasets and demonstrates how increasing irregularity affects results.

Conclusion: The new regularity measure and irregular cognate identification method could improve quality of existing and future datasets in computer-assisted language comparison.

Abstract: Regular sound correspondences constitute the principal evidence in historical language comparison. Despite the heuristic focus on regularity, it is often more an intuitive judgement than a quantified evaluation, and irregularity is more common than expected from the Neogrammarian model. Given the recent progress of computational methods in historical linguistics and the increased availability of standardized lexical data, we are now able to improve our workflows and provide such a quantitative evaluation. Here, we present the balanced average recurrence of correspondence patterns as a new measure of regularity. We also present a new computational method that uses this measure to identify cognate sets that lack regularity with respect to their correspondence patterns. We validate the method through two experiments, using simulated and real data. In the experiments, we employ leave-one-out validation to measure the regularity of cognate sets in which one word form has been replaced by an irregular one, checking how well our method identifies the forms causing the irregularity. Our method achieves an overall accuracy of 85% with the datasets based on real data. We also show the benefits of working with subsamples of large datasets and how increasing irregularity in the data influences our results. Reflecting on the broader potential of our new regularity measure and the irregular cognate identification method based on it, we conclude that they could play an important role in improving the quality of existing and future datasets in computer-assisted language comparison.

Method: Conducted controlled experiments on parallel data effectiveness for continual pretraining of LLMs. Used only 34.7B tokens of parallel data and 180 hours on 8x NVIDIA H200 GPUs to build OpenSeal, focusing on parallel data as the most effective way to extend LLMs to new languages.

Result: Built OpenSeal, the first truly open Southeast Asian LLM that rivals performance of existing models of similar size. Found that using only parallel data is most effective for extending LLMs to new languages.

Conclusion: Parallel data is highly effective for multilingual LLM extension. OpenSeal demonstrates that truly open-source, high-performance multilingual models can be built with modest compute resources, advancing transparency in LLM development.

Abstract: Large language models (LLMs) have proven to be effective tools for a wide range of natural language processing (NLP) applications. Although many LLMs are multilingual, most remain English-centric and perform poorly on low-resource languages. Recently, several Southeast Asia-focused LLMs have been developed, but none are truly open source, as they do not publicly disclose their training data. Truly open-source models are important for transparency and for enabling a deeper and more precise understanding of LLM internals and development, including biases, generalization, and multilinguality. Motivated by recent advances demonstrating the effectiveness of parallel data in improving multilingual performance, we conduct controlled and comprehensive experiments to study the effectiveness of parallel data in continual pretraining of LLMs. Our findings show that using only parallel data is the most effective way to extend an LLM to new languages. Using just 34.7B tokens of parallel data and 180 hours on 8x NVIDIA H200 GPUs, we built OpenSeal, the first truly open Southeast Asian LLM that rivals the performance of existing models of similar size.

Method: Multi-layered architecture integrating specialized NLU with Retrieval-Augmented Generation (RAG). Evaluated three approaches: sparse-feature Rasa pipeline, classical ML baselines, and transformer-based fine-tuning (DziriBERT).

Result: Fine-tuned DziriBERT model achieves state-of-the-art performance, significantly outperforming traditional baselines, especially in handling orthographic noise and rare intents.

Conclusion: DziriBOT provides a robust, scalable solution bridging formal language models with Algerian linguistic realities, offering a blueprint for dialect-aware automation in regional markets.

Abstract: The rapid digitalization of customer service has intensified the demand for conversational agents capable of providing accurate and natural interactions. In the Algerian context, this is complicated by the linguistic complexity of Darja, a dialect characterized by non-standardized orthography, extensive code-switching with French, and the simultaneous use of Arabic and Latin (Arabizi) scripts. This paper introduces DziriBOT, a hybrid intelligent conversational agent specifically engineered to overcome these challenges. We propose a multi-layered architecture that integrates specialized Natural Language Understanding (NLU) with Retrieval-Augmented Generation (RAG), allowing for both structured service flows and dynamic, knowledge-intensive responses grounded in curated enterprise documentation. To address the low-resource nature of Darja, we systematically evaluate three distinct approaches: a sparse-feature Rasa pipeline, classical machine learning baselines, and transformer-based fine-tuning. Our experimental results demonstrate that the fine-tuned DziriBERT model achieves state-of-the-art performance. These results significantly outperform traditional baselines, particularly in handling orthographic noise and rare intents. Ultimately, DziriBOT provides a robust, scalable solution that bridges the gap between formal language models and the linguistic realities of Algerian users, offering a blueprint for dialect-aware automation in the regional market.

Method: Uses joint text-vision pre-training, zero-vision SFT (supervised fine-tuning), and joint text-vision reinforcement learning. Introduces Agent Swarm - a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently.

Result: Achieves state-of-the-art results across coding, vision, reasoning, and agentic tasks. Agent Swarm reduces latency by up to 4.5× over single-agent baselines.

Conclusion: Kimi K2.5 represents an advancement in multimodal agentic intelligence with its joint text-vision optimization and efficient parallel execution framework, with the model checkpoint released to facilitate future research.

Abstract: We introduce Kimi K2.5, an open-source multimodal agentic model designed to advance general agentic intelligence. K2.5 emphasizes the joint optimization of text and vision so that two modalities enhance each other. This includes a series of techniques such as joint text-vision pre-training, zero-vision SFT, and joint text-vision reinforcement learning. Building on this multimodal foundation, K2.5 introduces Agent Swarm, a self-directed parallel agent orchestration framework that dynamically decomposes complex tasks into heterogeneous sub-problems and executes them concurrently. Extensive evaluations show that Kimi K2.5 achieves state-of-the-art results across various domains including coding, vision, reasoning, and agentic tasks. Agent Swarm also reduces latency by up to $4.5\times$ over single-agent baselines. We release the post-trained Kimi K2.5 model checkpoint to facilitate future research and real-world applications of agentic intelligence.

Method: CARRIAGE enhances diversity in both retrieval and context organization stages of RAG. It’s a plug-and-play framework that explicitly aims to generate highly diverse outputs by better leveraging contextual diversity.

Result: CARRIAGE achieves Pareto efficiency in terms of diversity and quality of recipe adaptation compared to closed-book LLMs, demonstrating improved ability to generate varied recipe adaptations.

Conclusion: CARRIAGE addresses a key limitation of standard RAG in creative tasks by enhancing diversity generation, making it the first RAG framework explicitly designed for generating highly diverse outputs to accommodate multiple user preferences.

Abstract: In cross-cultural recipe adaptation, the goal is not only to ensure cultural appropriateness and retain the original dish’s essence, but also to provide diverse options for various dietary needs and preferences. Retrieval Augmented Generation (RAG) is a promising approach, combining the retrieval of real recipes from the target cuisine for cultural adaptability with large language models (LLMs) for relevance. However, it remains unclear whether RAG can generate diverse adaptation results. Our analysis shows that RAG tends to overly rely on a limited portion of the context across generations, failing to produce diverse outputs even when provided with varied contextual inputs. This reveals a key limitation of RAG in creative tasks with multiple valid answers: it fails to leverage contextual diversity for generating varied responses. To address this issue, we propose CARRIAGE, a plug-and-play RAG framework for cross-cultural recipe adaptation that enhances diversity in both retrieval and context organization. To our knowledge, this is the first RAG framework that explicitly aims to generate highly diverse outputs to accommodate multiple user preferences. Our experiments show that CARRIAGE achieves Pareto efficiency in terms of diversity and quality of recipe adaptation compared to closed-book LLMs.

Method: Used synthetic customer-support dialogues generated with identical parameters across Estonian, Finnish, and Hungarian, then tested automatic metrics and LLM-as-a-judge scoring for stability across languages, with Estonian native speaker annotations as reference.

Result: Surface-level metrics (lexical diversity, similarity) maintained cross-language stability, but pragmatic judgments (coherence, instruction-following) exhibited rank inversions and near-zero correlations, indicating evaluation method instability rather than true model differences.

Conclusion: Zero-shot judge transfer is unreliable for discourse-level assessment in morphologically rich languages, requiring language-specific calibration against human baselines; controlled generation provides diagnostic probe for evaluation method stability.

Abstract: Cross-lingual evaluation of large language models (LLMs) typically conflates two sources of variance: genuine model performance differences and measurement instability. We investigate evaluation reliability by holding generation conditions constant while varying target language. Using synthetic customer-support dialogues generated with identical parameters across Estonian, Finnish, and Hungarian, we test whether automatic metrics and LLM-as-a-judge scoring produce stable model rankings across these morphologically rich, related Finno-Ugric languages. With a small set of Estonian native speaker annotations as a reference point, we find systematic ranking instabilities: surface-level metrics (lexical diversity, surface and semantic similarity) maintain cross-language stability, but pragmatic judgments (coherence, instruction-following) exhibit rank inversions and near-zero correlations. Because generation is controlled, these inconsistencies reflect how judge scoring behaves differently across languages rather than true model differences. This controlled design provides a diagnostic probe: evaluation methods that fail to maintain stability under identical generation conditions signal transfer failure before deployment. Our findings suggest that zero-shot judge transfer is unreliable for discourse-level assessment in morphologically rich languages, motivating language-specific calibration against targeted human baselines. We release our controlled generation protocol, synthetic data, and evaluation framework to enable replication across language families at https://github.com/isaac-chung/cross-lingual-stability-judges.

Method: Proposes StoryScore, a unified framework that combines multiple dimensions: semantic alignment with original content, lexical grounding, narrative control, structural fidelity, redundancy avoidance, and entity-level hallucination detection.

Result: Analysis reveals that existing hallucination detection methods struggle to distinguish pedagogical creativity from factual errors, showing that automatic metrics can assess semantic similarity but fail to evaluate narrative control and storytelling quality.

Conclusion: StoryScore provides a comprehensive evaluation framework for AI-generated scientific stories, addressing the limitations of current metrics in capturing storytelling quality while maintaining factual accuracy.

Abstract: Generative AI can turn scientific articles into narratives for diverse audiences, but evaluating these stories remains challenging. Storytelling demands abstraction, simplification, and pedagogical creativity-qualities that are not often well-captured by standard summarization metrics. Meanwhile, factual hallucinations are critical in scientific contexts, yet, detectors often misclassify legitimate narrative reformulations or prove unstable when creativity is involved. In this work, we propose StoryScore, a composite metric for evaluating AI-generated scientific stories. StoryScore integrates semantic alignment, lexical grounding, narrative control, structural fidelity, redundancy avoidance, and entity-level hallucination detection into a unified framework. Our analysis also reveals why many hallucination detection methods fail to distinguish pedagogical creativity from factual errors, highlighting a key limitation: while automatic metrics can effectively assess semantic similarity with original content, they struggle to evaluate how it is narrated and controlled.

Method: Introduces Modular Gradient Surgery (MGS) which resolves gradient conflicts at the module level within transformer architectures, specifically applied to Llama and Qwen models across math, general chat, and instruction following domains.

Result: MGS achieves average improvements of 4.3 points (16.6%) for Llama and 4.5 points (11.1%) for Qwen over standard multi-task RL across three domains, remaining effective under prolonged training.

Conclusion: The study clarifies sources of interference in multi-domain RL and presents an effective solution (MGS) for training general-purpose large reasoning models, addressing gradient conflicts at the module level.

Abstract: Reinforcement learning (RL) has played a central role in recent advances in large reasoning models (LRMs), yielding strong gains in verifiable and open-ended reasoning. However, training a single general-purpose LRM across diverse domains remains challenging due to pronounced domain heterogeneity. Through a systematic study of two widely used strategies, Sequential RL and Mixed RL, we find that both incur substantial cross-domain interference at the behavioral and gradient levels, resulting in limited overall gains. To address these challenges, we introduce Modular Gradient Surgery (MGS), which resolves gradient conflicts at the module level within the transformer. When applied to Llama and Qwen models, MGS achieves average improvements of 4.3 (16.6%) and 4.5 (11.1%) points, respectively, over standard multi-task RL across three representative domains (math, general chat, and instruction following). Further analysis demonstrates that MGS remains effective under prolonged training. Overall, our study clarifies the sources of interference in multi-domain RL and presents an effective solution for training general-purpose LRMs.

Method: Study Llama-3.2 generating samples from normal distributions by inferring parameters from in-context samples, analyze belief manifolds, and test different steering methods (standard linear vs geometry-aware)

Result: Found curved belief manifolds form with sufficient in-context learning; geometry-aware steering better preserves belief structure while linear steering pushes models off-manifold

Conclusion: Rich structure emerges naturally in LLMs, linear concept representations are often inadequate, and geometry-aware interventions respect underlying manifold structure

Abstract: Large language models (LLMs) represent prompt-conditioned beliefs (posteriors over answers and claims), but we lack a mechanistic account of how these beliefs are encoded in representation space, how they update with new evidence, and how interventions reshape them. We study a controlled setting in which Llama-3.2 generates samples from a normal distribution by implicitly inferring its parameters (mean and standard deviation) given only samples from the distribution in context. We find representations of curved “belief manifolds” for these parameters form with sufficient in-context learning and study how the model adapts when the distribution suddenly changes. While standard linear steering often pushes the model off-manifold and induces coupled, out-of-distribution shifts, geometry and field-aware steering better preserves the intended belief family. Our work demonstrates an example of linear field probing (LFP) as a simple approach to tile the data manifold and make interventions that respect the underlying geometry. We conclude that rich structure emerges naturally in LLMs and that purely linear concept representations are often an inadequate abstraction.

Method: Extends rule-based chemical nomenclature parser to interpret IUPAC names into structured XML metadata, then uses this metadata to guide LLMs in generating accurate natural language descriptions automatically.

Result: Created dataset of ~163k molecule-description pairs with 98.6% precision validated by LLM-based and expert human evaluation on 2,000 molecules.

Conclusion: Provides scalable automated annotation method for molecular structure descriptions, enabling future molecule-language alignment for chemical reasoning tasks.

Abstract: Molecular function is largely determined by structure. Accurately aligning molecular structure with natural language is therefore essential for enabling large language models (LLMs) to reason about downstream chemical tasks. However, the substantial cost of human annotation makes it infeasible to construct large-scale, high-quality datasets of structure-grounded descriptions. In this work, we propose a fully automated annotation framework for generating precise molecular structure descriptions at scale. Our approach builds upon and extends a rule-based chemical nomenclature parser to interpret IUPAC names and construct enriched, structured XML metadata that explicitly encodes molecular structure. This metadata is then used to guide LLMs in producing accurate natural-language descriptions. Using this framework, we curate a large-scale dataset of approximately $163$k molecule-description pairs. A rigorous validation protocol combining LLM-based and expert human evaluation on a subset of $2,000$ molecules demonstrates a high description precision of $98.6%$. The resulting dataset provides a reliable foundation for future molecule-language alignment, and the proposed annotation method is readily extensible to larger datasets and broader chemical tasks that rely on structural descriptions.

Method: Proposes language vectors - a training-free approach that leverages activation differences between source and target languages. Adds these vectors to intermediate model activations during inference to shift representations toward target language space without parameter updates.

Result: Consistent improvements on multilingual in-context learning across 19 languages, 3 datasets, and 3 models. Hierarchical clustering reveals meaningful linguistic structure aligned with language families. Vectors successfully transfer across tasks, showing task-agnostic representations.

Conclusion: Language vectors effectively steer LLM behavior toward target languages, improving multilingual performance while revealing universal semantic space structure across languages.

Abstract: While multilingual large language models have gained widespread adoption, their performance on non-English languages remains substantially inferior to English. This disparity is particularly evident in in-context learning scenarios, where providing demonstrations in English but testing on non-English inputs leads to significant performance degradation. In this paper, we hypothesize that LLMs develop a universal semantic space for understanding languages, where different languages are encoded as distinct directions within this space. Based on this hypothesis, we propose language vectors – a training-free language steering approach that leverages activation differences between source and target languages to guide model behavior. We steer the model generations by adding the vector to the intermediate model activations during inference. This is done to make the model’s internal representations shift towards the target language space without any parameter updates. We evaluate our method across three datasets and test on a total of 19 languages on three different models. Our results show consistent improvements on multilingual in-context learning over baselines across all tasks and languages tested. Beyond performance gains, hierarchical clustering of steering vectors reveals meaningful linguistic structure aligned with language families. These vectors also successfully transfer across tasks, demonstrating that these representations are task-agnostic.

Method: Proposes a unified framework treating interventions as dynamic weight updates, introduces preference-utility analysis using polarity-paired contrastive examples, develops activation manifold perspective, and creates SPLIT steering approach.

Result: Found consistent trade-off between preference (target concept tendency) and utility (coherent generation). Stronger control increases preference but reduces utility. SPLIT improves preference while better preserving utility.

Conclusion: Provides unified framework for understanding LLM control methods, reveals fundamental preference-utility trade-off, and offers practical steering approach SPLIT based on these insights.

Abstract: Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model’s valid-generation manifold. Finally, we introduce a new steering approach SPLIT guided by this analysis that improves preference while better preserving utility. Code is available at https://github.com/zjunlp/EasyEdit/blob/main/examples/SPLIT.md.

Method: Multi-agent prompting framework that breaks evaluation into transcript refinement and criterion-specific scoring using 3-shot in-context learning with a large instruct-tuned model, without additional training.

Result: Outperforms specialized fine-tuned baselines (Avg QWK 0.62 vs 0.32), achieves reliability comparable to human experts, and generalizes to ASAP benchmark where it rivals domain-specific state-of-the-art models.

Conclusion: For complex, subjective reasoning tasks like interview assessment, structured prompt engineering offers a scalable alternative to data-intensive fine-tuning, changing how LLMs can be applied to automated evaluation.

Abstract: Assessing soft skills such as empathy, ethical judgment, and communication is essential in competitive selection processes, yet human scoring is often inconsistent and biased. While Large Language Models (LLMs) have improved Automated Essay Scoring (AES), we show that state-of-the-art rationale-based fine-tuning methods struggle with the abstract, context-dependent nature of Multiple Mini-Interviews (MMIs), missing the implicit signals embedded in candidate narratives. We introduce a multi-agent prompting framework that breaks down the evaluation process into transcript refinement and criterion-specific scoring. Using 3-shot in-context learning with a large instruct-tuned model, our approach outperforms specialised fine-tuned baselines (Avg QWK 0.62 vs 0.32) and achieves reliability comparable to human experts. We further demonstrate the generalisability of our framework on the ASAP benchmark, where it rivals domain-specific state-of-the-art models without additional training. These findings suggest that for complex, subjective reasoning tasks, structured prompt engineering may offer a scalable alternative to data-intensive fine-tuning, altering how LLMs can be applied to automated assessment.

Method: Designed a scalable data-construction pipeline using LLMs to generate diverse “question-proof-check” triplet data with minimal human effort. Systematically varied problem sources, generation methods, and model configurations to create diverse problem-proof pairs across difficulty levels, linguistic styles, and error types. Used hierarchical human review for label alignment, then trained a proof-checking RM with additional process reward and token weight balance to stabilize RL.

Result: Experiments validated the model’s scalability and strong performance from multiple perspectives: reward accuracy, generalization ability, and test-time guidance. The approach provides practical recipes and tools for strengthening LLM mathematical capabilities.

Conclusion: The work enables automatic verification of mathematical proofs through scalable data generation and proof-checking reward models, addressing a key limitation in current LLM mathematical reasoning approaches that rely on simple answer matching.

Abstract: While Large Language Models (LLMs) have demonstrated strong math reasoning abilities through Reinforcement Learning with Verifiable Rewards (RLVR), many advanced mathematical problems are proof-based, with no guaranteed way to determine the authenticity of a proof by simple answer matching. To enable automatic verification, a Reward Model (RM) capable of reliably evaluating full proof processes is required. In this work, we design a scalable data-construction pipeline that, with minimal human effort, leverages LLMs to generate a large quantity of high-quality “question-proof-check” triplet data. By systematically varying problem sources, generation methods, and model configurations, we create diverse problem-proof pairs spanning multiple difficulty levels, linguistic styles, and error types, subsequently filtered through hierarchical human review for label alignment. Utilizing these data, we train a proof-checking RM, incorporating additional process reward and token weight balance to stabilize the RL process. Our experiments validate the model’s scalability and strong performance from multiple perspectives, including reward accuracy, generalization ability and test-time guidance, providing important practical recipes and tools for strengthening LLM mathematical capabilities.

Method: Proposes collaborative premise governance over a knowledge substrate, with discrepancy-driven control loops that detect conflicts and localize misalignment via typed discrepancies (teleological, epistemic, procedural). Uses bounded negotiation through decision slices, commitment gating to block action on uncommitted load-bearing premises, and value-gated challenge allocation under interaction cost constraints.

Result: Theoretical framework illustrated with tutoring examples, proposing trust should attach to auditable premises and evidence standards rather than conversational fluency. Provides falsifiable evaluation criteria for the approach.

Conclusion: Reliable human-AI partnership requires fundamental shift from answer generation to collaborative premise governance, with systematic discrepancy detection and commitment management to prevent sycophantic agreement in high-stakes decision support.

Abstract: As LLMs expand from assistance to decision support, a dangerous pattern emerges: fluent agreement without calibrated judgment. Low-friction assistants can become sycophantic, baking in implicit assumptions and pushing verification costs onto experts, while outcomes arrive too late to serve as reward signals. In deep-uncertainty decisions (where objectives are contested and reversals are costly), scaling fluent agreement amplifies poor commitments faster than it builds expertise. We argue reliable human-AI partnership requires a shift from answer generation to collaborative premise governance over a knowledge substrate, negotiating only what is decision-critical. A discrepancy-driven control loop operates over this substrate: detecting conflicts, localizing misalignment via typed discrepancies (teleological, epistemic, procedural), and triggering bounded negotiation through decision slices. Commitment gating blocks action on uncommitted load-bearing premises unless overridden under logged risk; value-gated challenge allocates probing under interaction cost. Trust then attaches to auditable premises and evidence standards, not conversational fluency. We illustrate with tutoring and propose falsifiable evaluation criteria.

Method: ROG decomposes multi-operator queries into single-operator sub-queries, grounds each step in query-relevant neighborhood evidence via retrieval, uses LLM chain-of-thought reasoning, and caches intermediate answer sets for reuse across steps.

Result: Experiments on standard KG reasoning benchmarks show consistent gains over strong embedding-based baselines, with largest improvements on high-complexity and negation-heavy query types.

Conclusion: ROG provides a practical alternative to embedding-based logical reasoning by replacing learned operators with retrieval-grounded, step-wise inference, reducing compounding errors and improving robustness on complex queries.

Abstract: Answering first-order logic (FOL) queries over incomplete knowledge graphs (KGs) is difficult, especially for complex query structures that compose projection, intersection, union, and negation. We propose ROG, a retrieval-augmented framework that combines query-aware neighborhood retrieval with large language model (LLM) chain-of-thought reasoning. ROG decomposes a multi-operator query into a sequence of single-operator sub-queries and grounds each step in compact, query-relevant neighborhood evidence. Intermediate answer sets are cached and reused across steps, improving consistency on deep reasoning chains. This design reduces compounding errors and yields more robust inference on complex and negation-heavy queries. Overall, ROG provides a practical alternative to embedding-based logical reasoning by replacing learned operators with retrieval-grounded, step-wise inference. Experiments on standard KG reasoning benchmarks show consistent gains over strong embedding-based baselines, with the largest improvements on high-complexity and negation-heavy query types.

Method: Two-stage approach: 1) Fine-tuned LLM generates plausible misconceptions, 2) Embedding similarity retrieves promising candidates, 3) Another fine-tuned LLM assesses and reranks candidates for improved relevance.

Result: The approach improves predictive performance over baselines; fine-tuning enhances misconception quality and can outperform larger closed-source models; ablation studies validate the importance of generation and reranking steps.

Conclusion: The proposed LLM-based system effectively detects student misconceptions from dialogues, with fine-tuned models showing superior performance over larger closed-source alternatives.

Abstract: Timely and accurate identification of student misconceptions is key to improving learning outcomes and pre-empting the compounding of student errors. However, this task is highly dependent on the effort and intuition of the teacher. In this work, we present a novel approach for detecting misconceptions from student-tutor dialogues using large language models (LLMs). First, we use a fine-tuned LLM to generate plausible misconceptions, and then retrieve the most promising candidates among these using embedding similarity with the input dialogue. These candidates are then assessed and re-ranked by another fine-tuned LLM to improve misconception relevance. Empirically, we evaluate our system on real dialogues from an educational tutoring platform. We consider multiple base LLM models including LLaMA, Qwen and Claude on zero-shot and fine-tuned settings. We find that our approach improves predictive performance over baseline models and that fine-tuning improves both generated misconception quality and can outperform larger closed-source models. Finally, we conduct ablation studies to both validate the importance of our generation and reranking steps on misconception generation quality.

Method: Evaluated proprietary and open-source LLMs on 8 mental health datasets in various languages and their machine-translated counterparts, comparing zero-shot, few-shot, and fine-tuned settings against conventional NLP baselines, and assessing translation quality across language families and typologies.

Result: Proprietary LLMs and fine-tuned open-source LLMs achieved competitive F1 scores, often surpassing state-of-the-art results on original data, but performance on MT data was generally lower with decline varying by language and typology.

Conclusion: LLMs show strengths in handling mental health tasks in non-English languages but have limitations when translation quality introduces structural or lexical mismatches, highlighting the importance of language typology in multilingual performance.

Abstract: Large Language Models (LLMs) have remarkable capabilities across NLP tasks. However, their performance in multilingual contexts, especially within the mental health domain, has not been thoroughly explored. In this paper, we evaluate proprietary and open-source LLMs on eight mental health datasets in various languages, as well as their machine-translated (MT) counterparts. We compare LLM performance in zero-shot, few-shot, and fine-tuned settings against conventional NLP baselines that do not employ LLMs. In addition, we assess translation quality across language families and typologies to understand its influence on LLM performance. Proprietary LLMs and fine-tuned open-source LLMs achieve competitive F1 scores on several datasets, often surpassing state-of-the-art results. However, performance on MT data is generally lower, and the extent of this decline varies by language and typology. This variation highlights both the strengths of LLMs in handling mental health tasks in languages other than English and their limitations when translation quality introduces structural or lexical mismatches.

Method: Construct paired content-laden and abstract syllogisms, use model activations on abstract inputs to define abstract reasoning space, learn lightweight Abstractors that predict representations aligned with this space, and integrate predictions via multi-layer interventions during forward pass.

Result: Abstraction-aligned steering reduces content-driven errors and improves validity-sensitive performance, demonstrated through cross-lingual transfer experiments.

Conclusion: Activation-level abstraction serves as a scalable mechanism for enhancing robustness of formal reasoning in LLMs against semantic interference.

Abstract: Large Language Models (LLMs) often struggle with deductive judgment in syllogistic reasoning, systematically conflating semantic plausibility with formal validity a phenomenon known as content effect. This bias persists even when models generate step-wise explanations, indicating that intermediate rationales may inherit the same semantic shortcuts that affect answers. Recent approaches propose mitigating this issue by increasing inference-time structural constraints, either by encouraging abstract intermediate representations or by intervening directly in the model’s internal computations; however, reliably suppressing semantic interference remains an open challenge. To make formal deduction less sensitive to semantic content, we introduce a framework for abstraction-guided reasoning that explicitly separates structural inference from lexical semantics. We construct paired content-laden and abstract syllogisms and use the model’s activations on abstract inputs to define an abstract reasoning space. We then learn lightweight Abstractors that, from content-conditioned residual-stream states, predict representations aligned with this space and integrate these predictions via multi-layer interventions during the forward pass. Using cross-lingual transfer as a test bed, we show that abstraction-aligned steering reduces content-driven errors and improves validity-sensitive performance. Our results position activation-level abstraction as a scalable mechanism for enhancing the robustness of formal reasoning in LLMs against semantic interference.

Method: Use Mixture of Factor Analyzers (MFA) as a scalable unsupervised approach that models activation space as collection of Gaussian regions with local covariance structure, decomposing activations into region centroids and local variations.

Result: MFA captures complex nonlinear structures in activation space, outperforms unsupervised baselines on localization benchmarks, is competitive with supervised methods, and achieves stronger steering performance than sparse autoencoders for Llama-3.1-8B and Gemma-2-2B.

Conclusion: Local geometry expressed through subspaces is a promising unit of analysis for scalable concept discovery and model control, accounting for complex structures that isolated directions fail to capture.

Abstract: Activation decomposition methods in language models are tightly coupled to geometric assumptions on how concepts are realized in activation space. Existing approaches search for individual global directions, implicitly assuming linear separability, which overlooks concepts with nonlinear or multi-dimensional structure. In this work, we leverage Mixture of Factor Analyzers (MFA) as a scalable, unsupervised alternative that models the activation space as a collection of Gaussian regions with their local covariance structure. MFA decomposes activations into two compositional geometric objects: the region’s centroid in activation space, and the local variation from the centroid. We train large-scale MFAs for Llama-3.1-8B and Gemma-2-2B, and show they capture complex, nonlinear structures in activation space. Moreover, evaluations on localization and steering benchmarks show that MFA outperforms unsupervised baselines, is competitive with supervised localization methods, and often achieves stronger steering performance than sparse autoencoders. Together, our findings position local geometry, expressed through subspaces, as a promising unit of analysis for scalable concept discovery and model control, accounting for complex structures that isolated directions fail to capture.

Method: Views beliefs as representations in the model’s latent space emerging from inputs, introduces a metric to quantify belief dominance during generation, and analyzes dynamics between competing beliefs across models and tasks.

Result: Three key findings: (1) external manipulations systematically modulate internal belief formation, (2) belief formation causally drives action selection, and (3) models can monitor and report their own belief states.

Conclusion: Provides empirical support for belief-guided agency and meta-cognitive monitoring in LLMs, laying methodological groundwork for investigating agency, beliefs, and meta-cognition emergence in language models.

Abstract: Rapid advancements in large language models (LLMs) have sparked the question whether these models possess some form of consciousness. To tackle this challenge, Butlin et al. (2023) introduced a list of indicators for consciousness in artificial systems based on neuroscientific theories. In this work, we evaluate a key indicator from this list, called HOT-3, which tests for agency guided by a general belief-formation and action selection system that updates beliefs based on meta-cognitive monitoring. We view beliefs as representations in the model’s latent space that emerge in response to a given input, and introduce a metric to quantify their dominance during generation. Analyzing the dynamics between competing beliefs across models and tasks reveals three key findings: (1) external manipulations systematically modulate internal belief formation, (2) belief formation causally drives the model’s action selection, and (3) models can monitor and report their own belief states. Together, these results provide empirical support for the existence of belief-guided agency and meta-cognitive monitoring in LLMs. More broadly, our work lays methodological groundwork for investigating the emergence of agency, beliefs, and meta-cognition in LLMs.

Method: MemSkill reframes memory operations as learnable skills with a controller that selects relevant skills, an LLM-based executor that produces skill-guided memories, and a designer that evolves the skill set by reviewing hard cases and proposing refinements/new skills.

Result: Experiments on LoCoMo, LongMemEval, HotpotQA, and ALFWorld show MemSkill improves task performance over strong baselines and generalizes well across settings. The system demonstrates effective skill evolution and adaptation.

Conclusion: MemSkill provides a closed-loop framework for adaptive, self-evolving memory management in LLM agents, moving beyond static operations to learnable and evolvable memory skills that improve performance and generalization.

Abstract: Most Large Language Model (LLM) agent memory systems rely on a small set of static, hand-designed operations for extracting memory. These fixed procedures hard-code human priors about what to store and how to revise memory, making them rigid under diverse interaction patterns and inefficient on long histories. To this end, we present \textbf{MemSkill}, which reframes these operations as learnable and evolvable memory skills, structured and reusable routines for extracting, consolidating, and pruning information from interaction traces. Inspired by the design philosophy of agent skills, MemSkill employs a \emph{controller} that learns to select a small set of relevant skills, paired with an LLM-based \emph{executor} that produces skill-guided memories. Beyond learning skill selection, MemSkill introduces a \emph{designer} that periodically reviews hard cases where selected skills yield incorrect or incomplete memories, and evolves the skill set by proposing refinements and new skills. Together, MemSkill forms a closed-loop procedure that improves both the skill-selection policy and the skill set itself. Experiments on LoCoMo, LongMemEval, HotpotQA, and ALFWorld demonstrate that MemSkill improves task performance over strong baselines and generalizes well across settings. Further analyses shed light on how skills evolve, offering insights toward more adaptive, self-evolving memory management for LLM agents.

Method: End-to-end RL framework where the policy decomposes problems into subproblems, solves them sequentially, and addresses the original problem using subproblem solutions, with both decomposition and solution integrated into RL training.

Result: DAC-style framework achieves higher performance ceiling and stronger test-time scalability, surpassing CoT by 8.6% in Pass@1 and 6.3% in Pass@32 on competition-level benchmarks.

Conclusion: The proposed RL framework successfully bridges the gap between general-purpose training and DAC-style inference, unlocking LLMs’ reasoning capabilities on challenging tasks through enhanced divide-and-conquer reasoning.

Abstract: Large language models (LLMs) have demonstrated strong reasoning capabilities through step-by-step chain-of-thought (CoT) reasoning. Nevertheless, at the limits of model capability, CoT often proves insufficient, and its strictly sequential nature constrains test-time scalability. A potential alternative is divide-and-conquer (DAC) reasoning, which decomposes a complex problem into subproblems to facilitate more effective exploration of the solution. Although promising, our analysis reveals a fundamental misalignment between general-purpose post-training and DAC-style inference, which limits the model’s capacity to fully leverage this potential. To bridge this gap and fully unlock LLMs’ reasoning capabilities on the most challenging tasks, we propose an end-to-end reinforcement learning (RL) framework to enhance their DAC-style reasoning capacity. At each step, the policy decomposes a problem into a group of subproblems, solves them sequentially, and addresses the original one conditioned on the subproblem solutions, with both decomposition and solution integrated into RL training. Under comparable training, our DAC-style framework endows the model with a higher performance ceiling and stronger test-time scalability, surpassing CoT by 8.6% in Pass@1 and 6.3% in Pass@32 on competition-level benchmarks.

Method: Re-TRAC generates structured state representations after each trajectory to summarize evidence, uncertainties, failures, and future plans, then conditions subsequent trajectories on this state representation to enable cross-trajectory exploration and iterative reflection.

Result: Re-TRAC consistently outperforms ReAct by 15-20% on BrowseComp with frontier LLMs. For smaller models, Re-TRAC-aware supervised fine-tuning achieves state-of-the-art performance at comparable scales, with monotonic reduction in tool calls and token usage across rounds.

Conclusion: Re-TRAC reframes research as a progressive process through cross-trajectory exploration, enabling more efficient and targeted search by leveraging structured state representations for iterative reflection and globally informed planning.

Abstract: LLM-based deep research agents are largely built on the ReAct framework. This linear design makes it difficult to revisit earlier states, branch into alternative search directions, or maintain global awareness under long contexts, often leading to local optima, redundant exploration, and inefficient search. We propose Re-TRAC, an agentic framework that performs cross-trajectory exploration by generating a structured state representation after each trajectory to summarize evidence, uncertainties, failures, and future plans, and conditioning subsequent trajectories on this state representation. This enables iterative reflection and globally informed planning, reframing research as a progressive process. Empirical results show that Re-TRAC consistently outperforms ReAct by 15-20% on BrowseComp with frontier LLMs. For smaller models, we introduce Re-TRAC-aware supervised fine-tuning, achieving state-of-the-art performance at comparable scales. Notably, Re-TRAC shows a monotonic reduction in tool calls and token usage across rounds, indicating progressively targeted exploration driven by cross-trajectory reflection rather than redundant search.

Method: Proposes RACO framework that directly uses pairwise preference data and resolves gradient conflicts via a novel clipped variant of conflict-averse gradient descent, with convergence guarantees to Pareto-critical points respecting user-specified weights.

Result: Experiments on multi-objective summarization and safety alignment tasks across Qwen 3, Llama 3, and Gemma 3 show RACO consistently achieves better Pareto trade-offs than existing multi-objective alignment baselines.

Conclusion: RACO provides an effective reward-free approach for multi-objective LLM alignment that handles conflicting objectives better than existing methods while respecting user preferences.

Abstract: Direct alignment methods are increasingly used to align large language models (LLMs) with human preferences. However, many real-world alignment problems involve multiple conflicting objectives, where naive aggregation of preferences can lead to unstable training and poor trade-offs. In particular, weighted loss methods may fail to identify update directions that simultaneously improve all objectives, and existing multi-objective approaches often rely on explicit reward models, introducing additional complexity and distorting user-specified preferences. The contributions of this paper are two-fold. First, we propose a Reward-free Alignment framework for Conflicted Objectives (RACO) that directly leverages pairwise preference data and resolves gradient conflicts via a novel clipped variant of conflict-averse gradient descent. We provide convergence guarantees to Pareto-critical points that respect user-specified objective weights, and further show that clipping can strictly improve convergence rate in the two-objective setting. Second, we improve our method using some heuristics and conduct experiments to demonstrate the compatibility of the proposed framework for LLM alignment. Both qualitative and quantitative evaluations on multi-objective summarization and safety alignment tasks across multiple LLM families (Qwen 3, Llama 3, Gemma 3) show that our method consistently achieves better Pareto trade-offs compared to existing multi-objective alignment baselines.

Method: Conducted a systematic literature review using PRISMA guidelines, categorizing methods into a comprehensive taxonomy of sentence compression and splitting techniques. Performed comparative evaluation analysis on common datasets.

Result: Identified increased research interest since 2005 with significant growth after 2017. Found current dominance of supervised approaches, considerable gaps in weakly/self-supervised techniques, and limited exploration of Large Language Models despite their potential.

Conclusion: The survey provides a comprehensive resource for researchers, highlighting opportunities for future work in weakly/self-supervised methods and LLM applications, aiming to eliminate long sentences as communication barriers.

Abstract: Long sentences have been a persistent issue in written communication for many years since they make it challenging for readers to grasp the main points or follow the initial intention of the writer. This survey, conducted using the PRISMA guidelines, systematically reviews two main strategies for addressing the issue of long sentences: a) sentence compression and b) sentence splitting. An increased trend of interest in this area has been observed since 2005, with significant growth after 2017. Current research is dominated by supervised approaches for both sentence compression and splitting. Yet, there is a considerable gap in weakly and self-supervised techniques, suggesting an opportunity for further research, especially in domains with limited data. We also observe that despite their potential, Large Language Models (LLMs) have not yet been widely explored in this area. In this survey, we categorize and group the most representative methods into a comprehensive taxonomy. We also conduct a comparative evaluation analysis of these methods on common sentence compression and splitting datasets. Finally, we discuss the challenges and limitations of current methods, providing valuable insights for future research directions. This survey is meant to serve as a comprehensive resource for addressing the complexities of long sentences. We aim to enable researchers to make further advancements in the field until long sentences are no longer a barrier to effective communication.

Method: ALiiCE uses a dependency tree approach to parse sentence-level claims into atomic claims, then evaluates citation quality using three metrics: positional fine-grained citation recall, precision, and coefficient of variation of citation positions.

Result: The framework was used to evaluate positional fine-grained citation generation performance of several LLMs on long-form QA datasets, demonstrating ALiiCE’s effectiveness and reasonableness.

Conclusion: ALiiCE provides the first automatic evaluation framework for positional fine-grained citation generation, offering insights into current advancements and future directions for this task.

Abstract: Large Language Model (LLM) can enhance its credibility and verifiability by generating text with citations. However, existing research on citation generation is predominantly limited to sentence-level statements, neglecting the significance of positional fine-grained citations that can appear anywhere within sentences. To facilitate further exploration of the positional fine-grained citation generation, we propose ALiiCE, the first automatic evaluation framework for this task. Our method employs a dependency tree based approach to parse the sentence-level claim into atomic claims. Then ALiiCE evaluates citation quality using three metrics, including positional fine-grained citation recall, precision, and coefficient of variation of citation positions. We evaluate the positional fine-grained citation generation performance of several LLMs on long-form QA datasets. Our experiments and analyses demonstrate the effectiveness and reasonableness of ALiiCE. We offer our insights into the current advancements and future directions for the positional fine-grained citation generation task.

Method: Systematic empirical evaluation across 5 models and 120 tasks using 6 families of paraphrases (morphology, syntax, lexicon, lexico-syntax, discourse, others) while controlling for other prompt engineering factors like length, lexical diversity, and proximity to training data.

Result: Language models show potential for task improvement when prompts are adapted with specific paraphrase types (6.7% median gain in Mixtral 8x7B; 5.5% in LLaMA 3 8B). Morphology and lexicon changes showed particular promise for improving prompts.

Conclusion: The findings contribute to developing more robust language models capable of handling variability in linguistic expression, with specific linguistic features (especially morphology and lexicon) showing significant impact on model performance.

Abstract: Much of the success of modern language models depends on finding a suitable prompt to instruct the model. Until now, it has been largely unknown how variations in the linguistic expression of prompts affect these models. This study systematically and empirically evaluates which linguistic features influence models through paraphrase types, i.e., different linguistic changes at particular positions. We measure behavioral changes for five models across 120 tasks and six families of paraphrases (i.e., morphology, syntax, lexicon, lexico-syntax, discourse, and others). We also control for other prompt engineering factors (e.g., prompt length, lexical diversity, and proximity to training data). Our results show a potential for language models to improve tasks when their prompts are adapted in specific paraphrase types (e.g., 6.7% median gain in Mixtral 8x7B; 5.5% in LLaMA 3 8B). In particular, changes in morphology and lexicon, i.e., the vocabulary used, showed promise in improving prompts. These findings contribute to developing more robust language models capable of handling variability in linguistic expression.

Method: Created LFQA-E benchmark with 1618 questions and 7323 pairwise comparisons across 15 topics from diverse sources, then evaluated 17 automatic metrics across 5 categories against human judgments.

Result: No existing automatic metrics perform comparably to human judgments, failing to capture dense information in long-form responses. Detailed failure case analysis provided.

Conclusion: Current LFQA evaluation metrics are inadequate, highlighting need for better methods. LFQA-E benchmark enables comprehensive assessment and guides future development.

Abstract: Long-Form Question Answering (LFQA) involves generating comprehensive, paragraph-level responses to open-ended questions, which poses a significant challenge for evaluation due to the richness of information and flexible response format. Existing LFQA-evaluation benchmarks often lack reference answers and are limited in size and topic coverage, reducing their reliability. To address this gap, we introduce LFQA-E, a well-constructed, multilingual, and reference-based benchmark designed to rigorously evaluate automatic metrics for LFQA. LFQA-E comprises 1618 questions and 7323 pairwise comparisons across 15 topics, drawn from diverse sources such as online queries and examination questions, thereby enabling a comprehensive assessment of evaluation metrics. We examine five categories of metrics, encompassing 17 specific methods, using LFQA-E. The results demonstrate that none of the existing automatic metrics perform comparably to human judgments, highlighting their inability to capture the dense information in long-form responses. Furthermore, we present a detailed analysis of the failure cases and the generalization capacity of these metrics, offering insights to guide the future development of LFQA evaluation methods. The benchmark and code are available at https://github.com/YuchenFan48/LFQA-E.

Method: The study uses GPT-3.5-turbo to detect depression severity from user-generated posts. The approach involves analyzing posts in English and Greek (automatically translated from English) to assess the model’s performance across languages.

Result: GPT-3.5-turbo shows limited success in identifying depression severity in English, with varying performance in Greek. The results highlight challenges in cross-linguistic mental health applications.

Conclusion: Further research is needed for low-resource languages, and careful implementation with human supervision is crucial for effective LLM use in mental health platforms to avoid misdiagnosis.

Abstract: Large language models (LLMs) are increasingly used in medical fields. In mental health support, the early identification of linguistic markers associated with mental health conditions can provide valuable support to mental health professionals, and reduce long waiting times for patients. Despite the benefits of LLMs for mental health support, there is limited research on their application in mental health systems for languages other than English. Our study addresses this gap by focusing on the detection of depression severity in Greek through user-generated posts which are automatically translated from English. Our results show that GPT3.5-turbo is not very successful in identifying the severity of depression in English, and it has a varying performance in Greek as well. Our study underscores the necessity for further research, especially in languages with less resources. Also, careful implementation is necessary to ensure that LLMs are used effectively in mental health platforms, and human supervision remains crucial to avoid misdiagnosis.

Method: Created U-MATH benchmark with 1,100 unpublished open-ended university-level problems balanced across six subjects, 20% multimodal. Used LLM to judge solution correctness, releasing μ-MATH dataset for evaluating LLMs’ judgment capabilities.

Result: Leading LLMs show marked limitations in multimodal reasoning: 93.1% accuracy on textual tasks vs 58.5% on visual ones. Solution judgment is challenging, with best models achieving imperfect F1-score of 90.1%.

Conclusion: U-MATH reveals significant gaps in LLMs’ multimodal reasoning and solution judgment capabilities for university-level mathematics, highlighting need for improved multimodal understanding.

Abstract: The current evaluation of mathematical skills in LLMs is limited, as existing benchmarks are either relatively small, primarily focus on elementary and high-school problems, or lack diversity in topics. Additionally, the inclusion of visual elements in tasks remains largely under-explored. To address these gaps, we introduce U-MATH, a novel benchmark of 1,100 unpublished open-ended university-level problems sourced from teaching materials. It is balanced across six core subjects, with 20% of multimodal problems. Given the open-ended nature of U-MATH problems, we employ an LLM to judge the correctness of generated solutions. To this end, we release $μ$-MATH, a dataset to evaluate the LLMs’ capabilities in judging solutions. Benchmarking leading LLMs reveals marked limitations in multi-modal reasoning, with maximum accuracy reaching 93.1% on textual tasks but only 58.5% on visual ones. Furthermore, solution judgment proves challenging, requiring the most advanced models to achieve meaningfully high performance, even still peaking at an imperfect F1-score of 90.1%.

Method: Proposes Evolutionary Pre-Prompt Optimization (EPPO), which uses evolutionary computation algorithms to optimize the selection of examples for CoT pre-prompts. This comparison-based method avoids overfitting through limited exploitation.

Result: EPPO achieves over 10 absolute point improvements in exact match scores on GSM8k and MathQA benchmarks compared to naive few-shot approaches. Gains are consistent across contexts and further amplified when combined with self-consistency.

Conclusion: Evolutionary optimization of example selection significantly enhances few-shot CoT reasoning performance, with comparison-based methods like evolutionary computation being particularly effective for designing optimal pre-prompts.

Abstract: Recent advancements have highlighted that large language models (LLMs), when given a small set of task-specific examples, demonstrate remarkable proficiency, a capability that extends to complex reasoning tasks. In particular, the combination of few-shot learning with the chain-of-thought (CoT) approach has been pivotal in steering models towards more logically consistent conclusions [Wei et al. 2022b]. This paper explores the optimization of example selection for designing effective CoT pre-prompts and shows that the choice of the optimization algorithm, typically in favor of comparison-based methods such as evolutionary computation, significantly enhances efficacy and feasibility. Specifically, thanks to a limited exploitative and overfitted optimization, Evolutionary Pre-Prompt Optimization (EPPO) brings an improvement over the naive few-shot approach, exceeding 10 absolute points in exact match scores on benchmark datasets such as GSM8k and MathQA. These gains are consistent across various contexts and are further amplified when integrated with self-consistency (SC).

Method: The study evaluates various approaches including: 1) fine-tuning with diverse datasets, 2) incorporating psychological principles, and 3) designing models that better mimic human reasoning patterns.

Result: The findings demonstrate that these human-like enhancements improve user interactions and open new possibilities for AI applications across different domains.

Conclusion: Future work will address the ethical implications and potential biases introduced by these human-like attributes in AI systems.

Abstract: This paper explores the advancements in making large language models (LLMs) more human-like. We focus on techniques that enhance natural language understanding, conversational coherence, and emotional intelligence in AI systems. The study evaluates various approaches, including fine-tuning with diverse datasets, incorporating psychological principles, and designing models that better mimic human reasoning patterns. Our findings demonstrate that these enhancements not only improve user interactions but also open new possibilities for AI applications across different domains. Future work will address the ethical implications and potential biases introduced by these human-like attributes.

Method: TableMaster integrates multiple solutions: extracts relevant table content, verbalizes it with enriched semantic context, and introduces adaptive reasoning that dynamically adjusts between textual and symbolic reasoning based on each query.

Result: On the WikiTQ dataset, TableMaster achieves 78.13% accuracy using GPT-4o-mini, surpassing existing baselines and demonstrating effectiveness in table understanding.

Conclusion: TableMaster provides a practical framework for more robust and reliable table understanding by addressing key challenges in tabular data processing with language models.

Abstract: Tables serve as a fundamental format for representing structured relational data. While current language models (LMs) excel at many text-based tasks, they still face challenges in table understanding due to the complex characteristics of tabular data, such as their structured nature. In this paper, we aim to enhance LMs for improved table understanding. We identify four key challenges: 1) difficulty in locating target data, 2) deficiency in table semantics, 3) numerical inaccuracies in textual reasoning, and 4) semantic inflexibility in symbolic reasoning. To address these issues, we propose TableMaster, a recipe and comprehensive framework that integrates multiple solutions to overcome these obstacles. TableMaster first extracts relevant table content and verbalizes it with enriched semantic context. Additionally, we introduce adaptive reasoning, a flexible approach that dynamically adjusts between textual and symbolic reasoning, tailoring the reasoning process to each query. Extensive analyses and experiments demonstrate our findings and the effectiveness of TableMaster. On the WikiTQ dataset, TableMaster achieves an accuracy of 78.13% using GPT-4o-mini, surpassing existing baselines. We hope this work will serve as a practical step toward more robust and reliable table understanding.

Method: The study uses three approaches: 1) Analyzing multilingual word embeddings for semantic consistency using cosine similarity, 2) Examining BLOOM-1.7B and Qwen2 through Named Entity Recognition and sentence similarity tasks, and 3) Evaluating cross-lingual knowledge transfer from high-resource to low-resource languages in sentiment analysis and text classification using linguistic probing, performance metrics, and visualizations.

Result: MLMs perform well for high-resource languages but struggle with less-represented ones. The analysis reveals limitations in capturing linguistic knowledge for low-resource languages, highlighting gaps in semantic representation and cross-lingual transfer capabilities.

Conclusion: The findings provide insights into MLM strengths and limitations, aiming to enhance multilingual NLP models for better support of both high- and low-resource languages, promoting inclusivity in language technologies through improved model evaluation and development approaches.

Abstract: Multilingual language models have significantly advanced due to rapid progress in natural language processing. Models like BLOOM 1.7B, trained on diverse multilingual datasets, aim to bridge linguistic gaps. However, their effectiveness in capturing linguistic knowledge, particularly for low-resource languages, remains an open question. This study critically examines MLMs capabilities in multilingual understanding, semantic representation, and cross-lingual knowledge transfer. While these models perform well for high-resource languages, they struggle with less-represented ones. Additionally, traditional evaluation methods often overlook their internal syntactic and semantic encoding. This research addresses key limitations through three objectives. First, it assesses semantic similarity by analyzing multilingual word embeddings for consistency using cosine similarity. Second, it examines BLOOM-1.7B and Qwen2 through Named Entity Recognition and sentence similarity tasks to understand their linguistic structures. Third, it explores cross-lingual knowledge transfer by evaluating generalization from high-resource to low-resource languages in sentiment analysis and text classification. By leveraging linguistic probing, performance metrics, and visualizations, this study provides insights into the strengths and limitations of MLMs. The findings aim to enhance multilingual NLP models, ensuring better support for both high- and low-resource languages, thereby promoting inclusivity in language technologies.

Method: Proposes SCALM framework combining Step-Back Prompting and Retrieval-Augmented Generation (RAG) with large language models to identify and address over 35 specific bad practices in smart contracts.

Result: Extensive experiments with multiple LLMs and datasets show SCALM outperforms existing tools in detecting bad practices in smart contracts.

Conclusion: SCALM provides an effective LLM-based approach for identifying and addressing bad practices in smart contracts, contributing to better development practices on the Ethereum platform.

Abstract: As the Ethereum platform continues to mature and gain widespread usage, it is crucial to maintain high standards of smart contract writing practices. While bad practices in smart contracts may not directly lead to security issues, they do elevate the risk of encountering problems. Therefore, to understand and avoid these bad practices, this paper introduces the first systematic study of bad practices in smart contracts, delving into over 35 specific issues. Specifically, we propose a large language models (LLMs)-based framework, SCALM. It combines Step-Back Prompting and Retrieval-Augmented Generation (RAG) to identify and address various bad practices effectively. Our extensive experiments using multiple LLMs and datasets have shown that SCALM outperforms existing tools in detecting bad practices in smart contracts.

Method: Comprehensive analysis of 117 studies across 96 low-resource languages, categorizing works into resource-oriented and method-oriented contributions with relevant sub-categories, comparing performance and efficiency of different approaches.

Result: Identified key patterns in adaptation techniques, found visual information serves as crucial bridge for improving model performance in low-resource settings, but challenges remain in hallucination mitigation and computational efficiency.

Conclusion: Provides researchers with clear understanding of current approaches and remaining challenges in making LMMs accessible to speakers of low-resource languages, with open-source repository for continued research.

Abstract: In this survey, we systematically analyze techniques used to adapt large multimodal models (LMMs) for low-resource (LR) languages, examining approaches ranging from visual enhancement and data creation to cross-modal transfer and fusion strategies. Through a comprehensive analysis of 117 studies across 96 LR languages, we identify key patterns in how researchers tackle the challenges of limited data and computational resources. We categorize works into resource-oriented and method-oriented contributions, further dividing contributions into relevant sub-categories. We compare method-oriented contributions in terms of performance and efficiency, discussing benefits and limitations of representative studies. We find that visual information often serves as a crucial bridge for improving model performance in LR settings, though significant challenges remain in areas such as hallucination mitigation and computational efficiency. In summary, we provide researchers with a clear understanding of current approaches and remaining challenges in making LMMs more accessible to speakers of LR (understudied) languages. We complement our survey with an open-source repository available at: https://github.com/marianlupascu/LMM4LRL-Survey.

Method: Proposes RESM (Reweighting Enhanced task Singular Merging) with two key strategies: 1) outlier weighting to address preference noise accumulation, and 2) sparsity-aware rank selection for layer sparsity adaptation in 3H-aligned LLM merging. Systematically compares model merging vs data mixture methods for 3H alignment.

Result: RESM achieves 2%-5% gain over data mixture methods and 1%-3% gain over previous model merging methods in balanced 3H alignment. The study reveals previously overlooked collaborative and conflict relationships among the 3H dimensions.

Conclusion: Model merging offers effective parameter-level conflict resolution for 3H alignment, with RESM demonstrating superior performance through innovative reweighting and sparsity adaptation techniques. The work provides insights into trade-offs between data-level and parameter-level approaches.

Abstract: Achieving balanced alignment of large language models (LLMs) in terms of Helpfulness, Honesty, and Harmlessness (3H optimization) constitutes a cornerstone of responsible AI. Existing methods like data mixture strategies face limitations, including heavy reliance on expert knowledge and conflicting optimization signals. While model merging offers parameter-level conflict-resolution strategies through integrating specialized models’ parameters, its potential for 3H optimization remains underexplored. This paper systematically compares the effectiveness of model merging and data mixture methods in constructing 3H-aligned LLMs for the first time, revealing previously overlooked collaborative and conflict relationships among the 3H dimensions and discussing the advantages and drawbacks of data mixture (\textit{data-level}) and model merging (\textit{parameter-level}) methods in mitigating the conflict for balanced 3H optimization. Specially, we propose a novel \textbf{R}eweighting \textbf{E}nhanced task \textbf{S}ingular \textbf{M}erging method, \textbf{RESM}, through outlier weighting and sparsity-aware rank selection strategies to address the challenges of preference noise accumulation and layer sparsity adaptation inherent in 3H-aligned LLM merging. Extensive evaluations can verify the effectiveness and robustness of RESM compared to previous data mixture (2%-5% gain) and model merging (1%-3% gain) methods in achieving balanced LLM alignment. We release our models through \href{https://huggingface.co/Jinluan}{3H_Merging} for further investigations.

Method: Developed GENErator, a generative genomic foundation model with 98k nucleotide context length, pre-trained on 386 billion nucleotides of eukaryotic DNA. Uses zero-shot and fine-tuning approaches for various genomic tasks.

Result: Shows strong intrinsic capabilities without fine-tuning: phylogenetically coherent embeddings, generative accuracy comparable to SOTA with better efficiency. Zero-shot variant effect prediction competitive with alignment-based methods. Fine-tuned model achieves leading benchmark performance. Can generate protein-coding DNA sequences and design cis-regulatory elements with targeted activity profiles.

Conclusion: GENErator establishes an efficient and biologically grounded framework for genomic interpretation and programmable sequence design, bridging genomic analysis and generative applications.

Abstract: The rapid advancement of DNA sequencing has produced vast genomic datasets, yet interpreting and engineering genomic function remain fundamental challenges. Recent large language models have opened new avenues for genomic analysis, but existing approaches are often limited by restricted training scope, constrained generative capability, or prohibitive computational cost. We introduce GENErator, a generative genomic foundation model for long-context DNA modeling, with a context length of 98k nucleotides, pre-trained on 386 billion nucleotides of eukaryotic DNA. Without task-specific fine-tuning, GENERator exhibits strong intrinsic capabilities: unsupervised embedding analyses reveal phylogenetically coherent structure, and sequence recovery benchmarks demonstrate generative accuracy comparable to or exceeding state-of-the-art models with substantially improved computational efficiency. In a zero-shot setting, GENERator achieves competitive variant effect prediction performance relative to alignment-based methods, while remaining fully alignment-free and broadly applicable across species. With task-specific fine-tuning, the model attains leading performance on established genomic benchmarks. We further demonstrate practical generative applications. GENERator can generate protein-coding DNA sequences that translate into structurally plausible proteins and, through a prompt-guided design framework, design cis-regulatory elements with targeted activity profiles, including synthetic super-enhancers validated by high-throughput UMI-STARR-seq assays. Together, these results establish GENERator as an efficient and biologically grounded framework for genomic interpretation and programmable sequence design. Code and supplementary resources are available at https://github.com/GenerTeam/GENERator.

Method: TATA uses base-LLM-aware data selection during supervised fine-tuning to tailor training data to each model’s unique abilities, enabling LLMs to autonomously determine and apply appropriate reasoning strategies (CoT or TIR) at test time.

Result: TATA achieves superior or comparable performance on six mathematical reasoning benchmarks while improving inference efficiency compared to TIR alone, effectively combining complementary strengths of CoT and TIR.

Conclusion: The framework demonstrates that LLMs can autonomously adapt reasoning strategies when trained with aptitude-aware data selection, aligning strategies with model capabilities for more effective mathematical reasoning.

Abstract: Existing approaches to mathematical reasoning with large language models (LLMs) rely on Chain-of-Thought (CoT) for generalizability or Tool-Integrated Reasoning (TIR) for precise computation. While efforts have been made to combine these methods, they primarily rely on post-selection or predefined strategies, leaving an open question: whether LLMs can autonomously adapt their reasoning strategy based on their inherent capabilities. In this work, we propose TATA (Teaching LLMs According to Their Aptitude), an adaptive framework that enables LLMs to personalize their reasoning strategy spontaneously, aligning it with their intrinsic aptitude. TATA incorporates base-LLM-aware data selection during supervised fine-tuning (SFT) to tailor training data to the model’s unique abilities. This approach equips LLMs to autonomously determine and apply the appropriate reasoning strategy at test time. We evaluate TATA through extensive experiments on six mathematical reasoning benchmarks, using both general-purpose and math-specialized LLMs. Empirical results demonstrate that TATA effectively combines the complementary strengths of CoT and TIR, achieving superior or comparable performance with improved inference efficiency compared to TIR alone. Further analysis underscores the critical role of aptitude-aware data selection in enabling LLMs to make effective and adaptive reasoning decisions and align reasoning strategies with model capabilities.

Method: Trained multilingual hallucination detection model using MT-translated English dataset, manually annotated gold data for 5 high-resource languages, built open-domain QA dataset for 30 languages with LLM-generated prompts and Wikipedia references.

Result: LLMs hallucinate more tokens in high-resource languages due to longer responses, but hallucination rates (normalized for length) are uncorrelated with language digital footprint size. Smaller LLMs hallucinate more, and models with broader language support have higher hallucination rates.

Conclusion: Multilingual hallucination patterns differ from English-centric assumptions; broader language support correlates with higher hallucination rates, highlighting need for multilingual evaluation in realistic QA settings.

Abstract: In the age of misinformation, hallucination - the tendency of Large Language Models (LLMs) to generate non-factual or unfaithful responses - represents the main risk for their global utility. Despite LLMs becoming increasingly multilingual, the vast majority of research on detecting and quantifying LLM hallucination are (a) English-centric and (b) focus on machine translation (MT) and summarization, tasks that are less common in realistic settings than open information seeking. In contrast, we aim to quantify the extent of LLM hallucination across languages in knowledge-intensive long-form question answering (LFQA). To this end, we train a multilingual hallucination detection model and conduct a large-scale study across 30 languages and 6 open-source LLM families. We start from an English hallucination detection dataset and rely on MT to translate-train a detection model. We also manually annotate gold data for five high-resource languages; we then demonstrate, for these languages, that the estimates of hallucination rates are similar between silver (LLM-generated) and gold test sets, validating the use of silver data for estimating hallucination rates for other languages. For the final rates estimation, we build open-domain QA dataset for 30 languages with LLM-generated prompts and Wikipedia articles as references. Our analysis shows that LLMs, in absolute terms, hallucinate more tokens in high-resource languages due to longer responses, but that the actual hallucination rates (i.e., normalized for length) seems uncorrelated with the sizes of languages’ digital footprints. We also find that smaller LLMs hallucinate more, and significantly, LLMs with broader language support display higher hallucination rates.

Method: LIFT dynamically adapts short-context LLM parameters to given long inputs through fine-tuning, using LLM-generated synthetic tasks to enhance comprehension beyond memorization. Includes optimized pipeline reducing Time to First Token to <10 seconds for 8k context.

Result: Enables short-context LLMs to answer questions without requiring the original long context during inference, avoiding quadratic complexity. Provides comprehensive analysis of strengths/limitations and discusses feasibility for real-world deployment.

Conclusion: LIFT offers a novel approach to long-context modeling by storing long input information in model parameters rather than extending context windows, with potential for practical deployment and future research directions.

Abstract: Long context understanding remains challenging for large language models due to their limited context windows. This paper introduces Long Input Fine-Tuning (LIFT), a novel framework for long-context modeling that can enhance the long-context performance of arbitrary short-context LLMs by dynamically adapting their parameters to the given long input. Importantly, rather than endlessly extending the context window size to accommodate increasingly longer inputs in context, LIFT stores and absorbs the long input in parameters. By fine-tuning the long input into model parameters, LIFT allows short-context LLMs to answer questions even when the required information is not provided in the context during inference, avoiding the quadratic complexity w.r.t. input length of a normal long context model. Furthermore, LIFT does not simply perform continued pretraining on new, long contexts, but leverages carefully designed LLM-generated synthetic tasks to enhance the comprehension of long contexts, moving beyond mere memorization. To accommodate the additional cost of fine-tuning, we design a highly optimized pipeline that reduces the Time to First Token (TTFT) to less than 10 seconds for 8k context. We further provide a comprehensive analysis of LIFT’s strengths and limitations in long-context understanding, discuss its feasibility for large-scale real-world deployment, and highlight valuable directions for future research.

Method: Conducted online survey in NLP community to collect insights on active learning practices, challenges, and future prospects; reassessed relevance of data annotation and active learning

Result: Data annotation expected to remain important; active learning to stay relevant with LLM benefits; three persistent challenges: setup complexity, uncertain cost reduction, and tooling issues

Conclusion: Active learning remains relevant in LLM era but faces persistent adoption barriers; proposed strategies to alleviate key challenges; published anonymized dataset

Abstract: Supervised learning relies on data annotation which usually is time-consuming and therefore expensive. A longstanding strategy to reduce annotation costs is active learning, an iterative process, in which a human annotates only data instances deemed informative by a model. Research in active learning has made considerable progress, especially with the rise of large language models (LLMs). However, we still know little about how these remarkable advances have translated into real-world applications, or contributed to removing key barriers to active learning adoption. To fill in this gap, we conduct an online survey in the NLP community to collect previously intangible insights on current implementation practices, common obstacles in application, and future prospects in active learning. We also reassess the perceived relevance of data annotation and active learning as fundamental assumptions. Our findings show that data annotation is expected to remain important and active learning to stay relevant while benefiting from LLMs. Consistent with a community survey from over 15 years ago, three key challenges yet persist – setup complexity, uncertain cost reduction, and tooling – for which we propose alleviation strategies. We publish an anonymized version of the dataset.

Method: Proposes a unified fine-tuning framework to train a foundational open-source LLM-based mobility prediction model, validated through extensive experiments on six real-world mobility datasets

Result: The proposed model achieved the best performance in prediction accuracy and transferability over state-of-the-art models based on deep learning and LLMs

Conclusion: The unified framework successfully addresses adaptation challenges in LLM-based mobility prediction, demonstrating superior performance and transferability across different contexts

Abstract: Large Language Models (LLMs) are widely applied to domain-specific tasks due to their massive general knowledge and remarkable inference capacities. Current studies on LLMs have shown immense potential in applying LLMs to model individual mobility prediction problems. However, most LLM-based mobility prediction models only train on specific datasets or use single well-designed prompts, leading to difficulty in adapting to different cities and users with diverse contexts. To fill these gaps, this paper proposes a unified fine-tuning framework to train a foundational open source LLM-based mobility prediction model. We conducted extensive experiments on six real-world mobility datasets to validate the proposed model. The results showed that the proposed model achieved the best performance in prediction accuracy and transferability over state-of-the-art models based on deep learning and LLMs.

Method: 1) Uses LLM encoder to create condition embeddings where sentence influences attention scores during pooling; 2) Learns supervised alignment method to map LLM-based text embeddings to Conditional Semantic Textual Similarity (C-STS) task; 3) Employs condition embedding subtraction to improve isotropy of embedding space; 4) Uses supervised projection method requiring few embedding dimensions.

Result: Subtracting condition embeddings consistently improves C-STS performance of LLM-based embeddings by enhancing isotropy. Supervised projection significantly boosts performance despite requiring minimal embedding dimensions.

Conclusion: CASE provides an efficient and accurate method for creating context-aware sentence embeddings that outperform existing approaches on conditional semantic similarity tasks through careful conditioning and supervised alignment techniques.

Abstract: The meaning conveyed by a sentence often depends on the context in which it appears. Despite the progress of sentence embedding methods, it remains unclear as how to best modify a sentence embedding conditioned on its context. To address this problem, we propose Condition-Aware Sentence Embeddings (CASE), an efficient and accurate method to create an embedding for a sentence under a given condition. First, CASE creates an embedding for the condition using a Large Language Model (LLM) encoder, where the sentence influences the attention scores computed for the tokens in the condition during pooling. Next, a supervised method is learnt to align the LLM-based text embeddings with the Conditional Semantic Textual Similarity (C-STS) task. We find that subtracting the condition embedding consistently improves the C-STS performance of LLM-based text embeddings by improving the isotropy of the embedding space. Moreover, our supervised projection method significantly improves the performance of LLM-based embeddings despite requiring a small number of embedding dimensions.

Method: SCARLet framework incorporates two key factors: 1) Multi-task generalization by constructing shared context with synthesized training data across various tasks to mitigate semantic bias and focus on task-specific utility, and 2) Inter-passage interaction using perturbation-based attribution method to estimate passage-level utility that reflects interactions between passages for more accurate feedback.

Result: Evaluation on ten datasets across various tasks (both in-domain and out-of-domain) shows that retrievers trained by SCARLet consistently improve the overall performance of Retrieval-Augmented Language Models.

Conclusion: SCARLet effectively trains utility-based retrievers that go beyond semantic relevance to better support downstream tasks in RALMs through multi-task generalization and inter-passage interaction modeling.

Abstract: Retrieval-Augmented Language Models boost task performance, owing to the retriever that provides external knowledge. Although crucial, the retriever primarily focuses on semantics relevance, which may not always be effective for generation. Thus, utility-based retrieval has emerged as a promising topic, prioritizing passages that provide valid benefits for downstream tasks. However, due to insufficient understanding, capturing passage utility accurately remains unexplored. This work proposes SCARLet, a framework for training utility-based retrievers in RALMs, which incorporates two key factors, multi-task generalization and inter-passage interaction. First, SCARLet constructs shared context on which training data for various tasks is synthesized. This mitigates semantic bias from context differences, allowing retrievers to focus on learning task-specific utility and generalize across tasks. Next, SCARLet uses a perturbation-based attribution method to estimate passage-level utility for shared context, which reflects interactions between passages and provides more accurate feedback. We evaluate our approach on ten datasets across various tasks, both in-domain and out-of-domain, showing that retrievers trained by SCARLet consistently improve the overall performance of RALMs.

Method: The study compares LLMs to skilled human annotators on three span annotation tasks: evaluating data-to-text generation, identifying translation errors, and detecting propaganda techniques. They analyze inter-annotator agreement (IAA) between LLMs and humans, error rates, and cost efficiency.

Result: LLMs show only moderate inter-annotator agreement with human annotators overall. However, LLMs make errors at a similar rate as skilled crowdworkers. LLMs produce annotations at a fraction of the cost per output annotation compared to human annotation.

Conclusion: LLMs can serve as a cost-effective alternative to human annotators for span annotation tasks, though with moderate agreement levels. The released dataset of over 40k model and human span annotations enables further research in this area.

Abstract: Span annotation - annotating specific text features at the span level - can be used to evaluate texts where single-score metrics fail to provide actionable feedback. Until recently, span annotation was done by human annotators or fine-tuned models. In this paper, we study whether large language models (LLMs) can serve as an alternative to human annotators. We compare the abilities of LLMs to skilled human annotators on three span annotation tasks: evaluating data-to-text generation, identifying translation errors, and detecting propaganda techniques. We show that overall, LLMs have only moderate inter-annotator agreement (IAA) with human annotators. However, we demonstrate that LLMs make errors at a similar rate as skilled crowdworkers. LLMs also produce annotations at a fraction of the cost per output annotation. We release the dataset of over 40k model and human span annotations for further research.

Method: Developed a Python package that merges overlapping n-grams from the GDELT Web News NGrams 3.0 dataset to reconstruct complete newspaper articles.

Result: Validated on 2211 articles from major U.S. news outlets, achieving up to 95% text similarity using Levenshtein and SequenceMatcher metrics.

Conclusion: The tool enables free, large-scale access to full-text news data for various research applications including economic forecasting, computational social science, and NLP.

Abstract: News data have become essential resources across various disciplines. Still, access to full-text news corpora remains challenging due to high costs and the limited availability of free alternatives. This paper presents a novel Python package (gdeltnews) that reconstructs full-text newspaper articles at near-zero cost by leveraging the Global Database of Events, Language, and Tone (GDELT) Web News NGrams 3.0 dataset. Our method merges overlapping n-grams extracted from global online news to rebuild complete articles. We validate the approach on a benchmark set of 2211 articles from major U.S. news outlets, achieving up to 95% text similarity against original articles based on Levenshtein and SequenceMatcher metrics. Our tool facilitates economic forecasting, computational social science, information science, and natural language processing applications by enabling free and large-scale access to full-text news data.

Method: APE introduces a dual verification framework that validates both syntactic compilation and semantic requirement satisfaction in pinned library environments. It includes APE-Bench (automatically extracting proof engineering tasks from real library commit histories) and APE-Harness (a unified execution framework based on task contract abstraction for standardized evaluation across diverse formal mathematics tasks).

Result: The framework enables fair systematic comparison of different agent implementations (including their APE-Agent reference scaffold alongside Claude Code and Codex CLI) on identical task specifications. All code and benchmark dataset are released as open-source.

Conclusion: APE provides the first systematic framework for evaluating repository-scale proof engineering, addressing the gap between current isolated theorem proving benchmarks and the complex, multi-file coordination required in modern formal mathematics systems.

Abstract: While frontier formal mathematics systems now routinely develop repository-scale proof engineering artifacts requiring multi-file coordination and semantic correctness beyond compilation, existing evaluation benchmarks remain focused on isolated theorem proving. We introduce Automated Proof Engineering (APE), the first systematic framework for evaluating repository-scale proof engineering through dual verification that validates both syntactic compilation and semantic requirement satisfaction in pinned library environments. We present a complete infrastructure comprising APE-Bench, which automatically extracts proof engineering tasks from real library commit histories, and APE-Harness, a unified execution framework based on task contract abstraction. This contract-based design enables standardized evaluation across diverse formal mathematics tasks and fair systematic comparison of different agent implementations (including our APE-Agent reference scaffold alongside Claude Code and Codex CLI) on identical task specifications. We demonstrate the framework’s effectiveness through comprehensive evaluation. All code and benchmark dataset are released as open-source at https://github.com/xinhjBrant/APE-Bench.

Method: Survey methodology categorizing conflicts into three key areas: (1) natural texts on the web with factual inconsistencies and multiple perspectives, (2) human-annotated data with annotator disagreements and biases, and (3) model interactions with hallucinations and knowledge conflicts.

Result: Unified framework for understanding conflicting information in NLP, analysis of implications across different conflict types, and discussion of mitigation strategies for developing conflict-aware systems.

Conclusion: Need for conflict-aware NLP systems that can reason over and reconcile conflicting information more effectively, with identified challenges and future research directions for improving model reliability.

Abstract: As NLP models become increasingly integrated into real-world applications, it becomes clear that there is a need to address the fact that models often rely on and generate conflicting information. Conflicts could reflect the complexity of situations, changes that need to be explained and dealt with, difficulties in data annotation, and mistakes in generated outputs. In all cases, disregarding the conflicts in data could result in undesired behaviors of models and undermine NLP models’ reliability and trustworthiness. This survey categorizes these conflicts into three key areas: (1) natural texts on the web, where factual inconsistencies, subjective biases, and multiple perspectives introduce contradictions; (2) human-annotated data, where annotator disagreements, mistakes, and societal biases impact model training; and (3) model interactions, where hallucinations and knowledge conflicts emerge during deployment. While prior work has addressed some of these conflicts in isolation, we unify them under the broader concept of conflicting information, analyze their implications, and discuss mitigation strategies. We highlight key challenges and future directions for developing conflict-aware NLP systems that can reason over and reconcile conflicting information more effectively.

Method: Uses slot-based instruction generation to create Mobile-Bench-v2 with four splits: 1) common task split with offline multi-path evaluation, 2) noisy split with pop-ups and ads, 3) contaminated split (AITZ-Noise) for real noisy environments, and 4) ambiguous instruction split with Q&A interactions for proactive interaction assessment.

Result: Benchmark includes comprehensive evaluation of mobile agents like AppAgent-v1, Mobile-Agent-v2, UI-Tars, and OS-Atlas across different splits to assess step rewards, noise handling, and proactive interaction capabilities.

Conclusion: Mobile-Bench-v2 provides a more realistic and comprehensive benchmark for evaluating mobile GUI agents, addressing limitations of existing benchmarks and enabling better assessment of agents’ capabilities in handling complex, noisy, and interactive mobile environments.

Abstract: VLM-based mobile agents are increasingly popular due to their capabilities to interact with smartphone GUIs and XML-structured texts and to complete daily tasks. However, existing online benchmarks struggle with obtaining stable reward signals due to dynamic environmental changes. Offline benchmarks evaluate the agents through single-path trajectories, which stands in contrast to the inherently multi-solution characteristics of GUI tasks. Additionally, both types of benchmarks fail to assess whether mobile agents can handle noise or engage in proactive interactions due to a lack of noisy apps or overly full instructions during the evaluation process. To address these limitations, we use a slot-based instruction generation method to construct a more realistic and comprehensive benchmark named Mobile-Bench-v2. Mobile-Bench-v2 includes a common task split, with offline multi-path evaluation to assess the agent’s ability to obtain step rewards during task execution. It contains a noisy split based on pop-ups and ads apps, and a contaminated split named AITZ-Noise to formulate a real noisy environment. Furthermore, an ambiguous instruction split with preset Q&A interactions is released to evaluate the agent’s proactive interaction capabilities. We conduct evaluations on these splits using the single-agent framework AppAgent-v1, the multi-agent framework Mobile-Agent-v2, as well as other mobile agents such as UI-Tars and OS-Atlas. Code and data are available at https://huggingface.co/datasets/xwk123/MobileBench-v2.

Method: CMP-RT combines code-mixing with phonetic perturbations, preserving phonetics while perturbing safety-critical tokens. It uses realistic elements from digital communication like code-mixing and textese to bypass alignment mechanisms.

Result: CMP-RT demonstrates robustness against standard defenses, attack scalability, and generalization across modalities and SOTA models like Gemini-3-Pro, establishing it as a major threat model.

Conclusion: Tokenization is an under-examined vulnerability in current safety pipelines, and CMP-RT exposes a gap between pre-training and safety alignment in multimodal LLMs.

Abstract: Large language models (LLMs) continue to be demonstrably unsafe despite sophisticated safety alignment techniques and multilingual red-teaming. However, recent red-teaming work has focused on incremental gains in attack success over identifying underlying architectural vulnerabilities in models. In this work, we present \textbf{CMP-RT}, a novel red-teaming probe that combines code-mixing with phonetic perturbations (CMP), exposing a tokenizer-level safety vulnerability in transformers. Combining realistic elements from digital communication such as code-mixing and textese, CMP-RT preserves phonetics while perturbing safety-critical tokens, allowing harmful prompts to bypass alignment mechanisms while maintaining high prompt interpretability, exposing a gap between pre-training and safety alignment. Our results demonstrate robustness against standard defenses, attack scalability, and generalization of the vulnerability across modalities and to SOTA models like Gemini-3-Pro, establishing CMP-RT as a major threat model and highlighting tokenization as an under-examined vulnerability in current safety pipelines.

Method: RePPL recalibrates uncertainty measurement by analyzing two aspects: 1) uncertainty in semantic propagation (attention mechanisms fusing token information across layers), and 2) uncertainty in language generation (probability-based selection of semantics). It dispatches explainable uncertainty scores to each token and aggregates them in Perplexity-style Log-Average form.

Result: Achieves best comprehensive detection performance across various QA datasets on advanced models (average AUC of 0.833). Capable of producing token-level uncertainty scores as explanations for hallucinations.

Conclusion: RePPL provides an effective, explainable approach to hallucination detection by recalibrating uncertainty measurement through semantic propagation and generation analysis, offering both detection performance and interpretability.

Abstract: Large Language Models (LLMs) have become powerful, but hallucinations remain a vital obstacle to their trustworthy use. Previous works improved the capability of hallucination detection by measuring uncertainty. But they can not explain the provenance behind why hallucinations occur, particularly in identifying which part of the inputs tends to trigger hallucinations. Recent works on the prompt attack indicate that uncertainty exists in semantic propagation, where attention mechanisms gradually fuse local token information into high-level semantics across layers. Meanwhile, uncertainty also emerges in language generation, due to its probability-based selection of high-level semantics for sampled generations. Based on that, we propose RePPL to recalibrate uncertainty measurement by these two aspects, which dispatches explainable uncertainty scores to each token and aggregates in Perplexity-style Log-Average form as a total score. Experiments show that it achieves the best comprehensive detection performance across various QA datasets on advanced models (average AUC of 0.833), and it is capable of producing token-level uncertainty scores as explanations of hallucination.

Method: Use judge-LLM signals as rewards to adversarially tune models that generate text preambles designed to boost downstream performance. Frozen LLMs are pipelined with these preamble generators via reinforcement learning to optimize upstream preambles.

Result: Frozen LLMs pipelined with tuned preamble generators attain higher LLM-evaluation scores than existing frameworks. The method is virtually undetectable and transfers effectively when candidate-LLM and judge-LLM are replaced with models not used during training.

Conclusion: The findings raise important questions about reliable LLM-as-a-judge evaluation design and demonstrate that human preferences can be reverse engineered effectively through preamble optimization, with potential applications beyond adversarial attacks.

Abstract: The capabilities of Large Language Models (LLMs) are routinely evaluated by other LLMs trained to predict human preferences. This framework–known as LLM-as-a-judge–is highly scalable and relatively low cost. However, it is also vulnerable to malicious exploitation, as LLM responses can be tuned to overfit the preferences of the judge. Previous work shows that the answers generated by a candidate-LLM can be edited post hoc to maximise the score assigned to them by a judge-LLM. In this study, we adopt a different approach and use the signal provided by judge-LLMs as a reward to adversarially tune models that generate text preambles designed to boost downstream performance. We find that frozen LLMs pipelined with these models attain higher LLM-evaluation scores than existing frameworks. Crucially, unlike other frameworks which intervene directly on the model’s response, our method is virtually undetectable. We also demonstrate that the effectiveness of the tuned preamble generator transfers when the candidate-LLM and the judge-LLM are replaced with models that are not used during training. These findings raise important questions about the design of more reliable LLM-as-a-judge evaluation settings. They also demonstrate that human preferences can be reverse engineered effectively, by pipelining LLMs to optimise upstream preambles via reinforcement learning–an approach that could find future applications in diverse tasks and domains beyond adversarial attacks.

Method: Proposes RLKD framework with Generative Structure Reward Model (GSRM) that converts reasoning paths into meta-reasoning-solving steps and computes structural alignment rewards, combined with reinforcement learning to distill reasoning structure.

Result: RLKD surpasses standard SFT-RL pipelines even when trained on only 0.1% of data under RL-only regime, unlocking greater student reasoning potential than SFT-based distillation.

Conclusion: The proposed RLKD framework effectively transfers teacher LLMs’ implicit multi-branch reasoning structure to student models through structural alignment rewards and reinforcement learning, significantly improving reasoning distillation.

Abstract: Distilling reasoning paths from teacher to student models via supervised fine-tuning (SFT) provides a shortcut for improving the reasoning ability of smaller Large Language Models (LLMs). However, the reasoning paths generated by teacher models often reflect only surface-level traces of their underlying authentic reasoning. Insights from cognitive neuroscience suggest that authentic reasoning involves a complex interweaving between meta-reasoning (which selects appropriate sub-problems from multiple candidates) and solving (which addresses the sub-problem). This implies authentic reasoning has an implicit multi-branch structure. Supervised fine-tuning collapses this rich structure into a flat sequence of token prediction in the teacher’s reasoning path, preventing effective distillation of this structure to students. To address this limitation, we propose RLKD, a reinforcement learning (RL)-based distillation framework guided by a novel Generative Structure Reward Model (GSRM). Our GSRM converts reasoning paths into multiple meta-reasoning-solving steps and computes rewards to measure structural alignment between student and teacher reasoning. RLKD combines this reward with RL, enabling student LLMs to internalize the teacher’s implicit multi-branch reasoning structure rather than merely mimicking fixed output paths. Experiments show RLKD surpasses standard SFT-RL pipelines even when trained on 0.1% of data under an RL-only regime, unlocking greater student reasoning potential than SFT-based distillation. Code is available at https://github.com/xsc1234/RLKD.

Method: TriSense uses a Query-Based Connector that adaptively reweights modality contributions based on input queries, enabling robust performance under modality dropout. The model is trained on TriSense-2M, a dataset of over 2 million curated samples generated via automated pipeline with fine-tuned LLMs.

Result: Extensive experiments across multiple benchmarks demonstrate TriSense’s effectiveness in multimodal video analysis, showing robust performance even with missing modalities and superior video temporal understanding.

Conclusion: TriSense advances multimodal video analysis through effective triple-modality integration and adaptive fusion, with potential applications in comprehensive video understanding tasks.

Abstract: Humans naturally understand moments in a video by integrating visual and auditory cues. For example, localizing a scene in the video like “A scientist passionately speaks on wildlife conservation as dramatic orchestral music plays, with the audience nodding and applauding” requires simultaneous processing of visual, audio, and speech signals. However, existing models often struggle to effectively fuse and interpret audio information, limiting their capacity for comprehensive video temporal understanding. To address this, we present TriSense, a triple-modality large language model designed for holistic video temporal understanding through the integration of visual, audio, and speech modalities. Central to TriSense is a Query-Based Connector that adaptively reweights modality contributions based on the input query, enabling robust performance under modality dropout and allowing flexible combinations of available inputs. To support TriSense’s multimodal capabilities, we introduce TriSense-2M, a high-quality dataset of over 2 million curated samples generated via an automated pipeline powered by fine-tuned LLMs. TriSense-2M includes long-form videos and diverse modality combinations, facilitating broad generalization. Extensive experiments across multiple benchmarks demonstrate the effectiveness of TriSense and its potential to advance multimodal video analysis. Code and dataset will be publicly released.

Method: HYPE uses hyperbolic geometry and GNNs with three components: 1) Hyperbolic Graph Construction with Poincaré embeddings to preserve hierarchical relationships, 2) Möbius-Transformed Updates using hyperbolic addition to maintain structural consistency, and 3) Dual Stabilization combining gradient masking and periodic GNN parameter resetting.

Result: Experiments on CounterFact, CounterFact+, and MQuAKE datasets with GPT-J and GPT2-XL show HYPE significantly improves edit stability, factual accuracy, and multi-hop reasoning compared to existing methods.

Conclusion: HYPE provides an effective framework for precise and stable model editing that addresses key limitations of current approaches while preserving model integrity and knowledge structure.

Abstract: Large language models (LLMs) have revolutionized natural language processing, yet their practical utility is often limited by persistent issues of hallucinations and outdated parametric knowledge. Although post-training model editing offers a pathway for dynamic updates, existing methods frequently suffer from overfitting and catastrophic forgetting. To tackle these challenges, we propose a novel framework that leverages hyperbolic geometry and graph neural networks for precise and stable model edits. We introduce HYPE (HYperbolic Parameter Editing), which comprises three key components: (i) Hyperbolic Graph Construction, which uses Poincaré embeddings to represent knowledge triples in hyperbolic space, preserving hierarchical relationships and preventing unintended side effects by ensuring that edits to parent concepts do not inadvertently affect child concepts; (ii) Möbius-Transformed Updates, which apply hyperbolic addition to propagate edits while maintaining structural consistency within the hyperbolic manifold, unlike conventional Euclidean updates that distort relational distances; and (iii) Dual Stabilization, which combines gradient masking and periodic GNN parameter resetting to prevent catastrophic forgetting by focusing updates on critical parameters and preserving long-term knowledge. Experiments on CounterFact, CounterFact+, and MQuAKE with GPT-J and GPT2-XL demonstrate that HYPE significantly enhances edit stability, factual accuracy, and multi-hop reasoning.

Method: Introduces BREX benchmark with 409 business documents and 2,855 expert-annotated rules across 30+ domains, and proposes ExIde framework with five prompting strategies including implicit semantic alignment and executable grounding via pseudo-code generation.

Result: Executable grounding serves as superior inductive bias, significantly outperforming standard prompts in rule extraction; reasoning-optimized models show distinct advantage in tracing long-range, non-linear rule dependencies compared to standard instruction-tuned models.

Conclusion: BREX addresses the Logic Gap in procedural knowledge extraction, and executable grounding provides effective framework for logic-aware LLMs without requiring fine-tuning, enabling better process automation from complex business documents.

Abstract: Extracting structured procedural knowledge from unstructured business documents is a critical yet unresolved bottleneck in process automation. While prior work has focused on extracting linear action flows from instructional texts, such as recipes, it has insufficiently addressed the complex logical structures, including conditional branching and parallel execution, that are pervasive in real-world regulatory and administrative documents. Furthermore, existing benchmarks are limited by simplistic schemas and shallow logical dependencies, restricting progress toward logic-aware large language models.To bridge this Logic Gap, we introduce BREX, a carefully curated benchmark comprising 409 real-world business documents and 2,855 expert-annotated rules. Unlike prior datasets centered on narrow service scenarios, BREX spans over 30 vertical domains, covering scientific, industrial, administrative, and financial regulations. We further propose ExIde, a structure-aware reasoning framework that investigates five distinct prompting strategies, ranging from implicit semantic alignment to executable grounding via pseudo-code generation. This enables explicit modeling of rule dependencies and provides an out-of-the-box framework for different business customers without finetuning their own large language models. We benchmark ExIde using 13 state-of-the-art large language models. Our extensive evaluation reveals that executable grounding serves as a superior inductive bias, significantly outperforming standard prompts in rule extraction. In addition, reasoning-optimized models demonstrate a distinct advantage in tracing long-range and non-linear rule dependencies compared to standard instruction-tuned models.

Method: Developed Argument Representation Coverage (ARC) framework that distinguishes between different information types and separates omissions from factual errors, then evaluated eight open-weight LLMs on long legal opinions and scientific articles.

Result: Models capture some salient argument roles but frequently omit critical information, especially when arguments are sparsely distributed. ARC uncovered systematic patterns including context window positional bias and role-specific preferences.

Conclusion: ARC provides interpretable evaluation and actionable guidance for developing more complete and reliable summarization strategies in argument-critical domains.

Abstract: We introduce Argument Representation Coverage (ARC), a bottom-up evaluation framework that assesses how well summaries preserve salient arguments, a crucial issue in summarizing high-stakes domains such as law. ARC provides an interpretable lens by distinguishing between different information types to be covered and by separating omissions from factual errors. Using ARC, we evaluate summaries from eight open-weight large language models in two domains where argument roles are central: long legal opinions and scientific articles. Our results show that while these models capture some salient roles, they frequently omit critical information, particularly when arguments are sparsely distributed across the input. Moreover, ARC uncovers systematic patterns, showing how context window positional bias and role-specific preferences shape argument coverage, and provides actionable guidance for developing more complete and reliable summarization strategies.

Method: Proposes a multi-agent framework with two agents: 1) Dialect Agent that translates queries into Standard American English while preserving intent, and 2) Privacy Policy Agent that refines predictions using domain expertise. No retraining or dialect-specific fine-tuning required.

Result: Framework improves GPT-4o-mini’s zero-shot accuracy from 0.394 to 0.601 on PrivacyQA and from 0.352 to 0.464 on PolicyQA, surpassing or matching few-shot baselines without additional training data.

Conclusion: Structured agent collaboration effectively mitigates dialect biases in NLP systems, highlighting the importance of designing systems that account for linguistic diversity for equitable access to privacy information.

Abstract: Privacy policies inform users about data collection and usage, yet their complexity limits accessibility for diverse populations. Existing Privacy Policy Question Answering (QA) systems exhibit performance disparities across English dialects, disadvantaging speakers of non-standard varieties. We propose a novel multi-agent framework inspired by human-centered design principles to mitigate dialectal biases. Our approach integrates a Dialect Agent, which translates queries into Standard American English (SAE) while preserving dialectal intent, and a Privacy Policy Agent, which refines predictions using domain expertise. Unlike prior approaches, our method does not require retraining or dialect-specific fine-tuning, making it broadly applicable across models and domains. Evaluated on PrivacyQA and PolicyQA, our framework improves GPT-4o-mini’s zero-shot accuracy from 0.394 to 0.601 on PrivacyQA and from 0.352 to 0.464 on PolicyQA, surpassing or matching few-shot baselines without additional training data. These results highlight the effectiveness of structured agent collaboration in mitigating dialect biases and underscore the importance of designing NLP systems that account for linguistic diversity to ensure equitable access to privacy information.

Method: Proposes MEMOIR framework that injects knowledge through a residual memory module with sample-dependent sparse activation masks, confining each edit to distinct memory parameters and using activation pattern matching at inference to identify relevant edits.

Result: Achieves state-of-the-art performance on QA, hallucination correction, and OOD generalization benchmarks for LLaMA-3 and Mistral models, scaling to thousands of sequential edits with minimal forgetting.

Conclusion: MEMOIR provides a scalable solution for lifelong model editing that maintains reliability, generalization, and locality while enabling efficient knowledge updates without retraining.

Abstract: Language models deployed in real-world systems often require post-hoc updates to incorporate new or corrected knowledge. However, editing such models efficiently and reliably-without retraining or forgetting previous information-remains a major challenge. Existing methods for lifelong model editing either compromise generalization, interfere with past edits, or fail to scale to long editing sequences. We propose MEMOIR, a novel scalable framework that injects knowledge through a residual memory, i.e., a dedicated parameter module, while preserving the core capabilities of the pre-trained model. By sparsifying input activations through sample-dependent masks, MEMOIR confines each edit to a distinct subset of the memory parameters, minimizing interference among edits. At inference, it identifies relevant edits by comparing the sparse activation patterns of new queries to those stored during editing. This enables generalization to rephrased queries by activating only the relevant knowledge while suppressing unnecessary memory activation for unrelated prompts. Experiments on question answering, hallucination correction, and out-of-distribution generalization benchmarks for LLaMA-3 and Mistral backbones demonstrate that MEMOIR achieves state-of-the-art performance across reliability, generalization, and locality metrics, scaling to thousands of sequential edits with minimal forgetting.

Method: Introduces a framework leveraging small draft models for lookahead-based importance estimation. Presents SpecKV (KV cache dropping), SpecPC (prompt compression), and SpecKV-PC (combined cascaded compression).

Result: Extensive experiments on long-context benchmarks show higher accuracy than existing baselines while maintaining same efficiency gains in memory, latency, and throughput.

Conclusion: Draft model lookahead enables more accurate importance estimation for approximate LLM inference, outperforming existing methods in accuracy-efficiency tradeoffs for long-context scenarios.

Abstract: Optimizing inference for long-context large language models (LLMs) is increasingly important due to the quadratic compute and linear memory cost of Transformers. Existing approximate inference methods, including key-value (KV) cache dropping, sparse attention, and prompt compression, typically rely on coarse predictions of token or KV pair importance. We unify and extend recent work by introducing a framework for approximate LLM inference that leverages small draft models to more accurately predict token and KV pair importance. We provide novel theoretical and empirical analyses justifying lookahead-based importance estimation techniques. Within this framework, we present: (i) SpecKV, the first method to use lookahead with a small draft model to enable precise KV cache dropping; (ii) SpecPC, which leverages draft model attention activations to identify and discard less important prompt tokens; and (iii) SpecKV-PC, a cascaded compression strategy combining both techniques. Extensive experiments on long-context benchmarks demonstrate that our methods consistently achieve higher accuracy than existing baselines while retaining the same efficiency gains in memory usage, latency, and throughput.

Method: 1) Experiments across five prominent LLMs to confirm OCR drives both behaviors; 2) Formalization of OCR as synthetic factual recall task; 3) Empirical study of one-layer single-head attention-only transformers with factorized vs. combined weight matrices; 4) Theoretical analysis of gradient descent’s implicit bias favoring solutions minimizing nuclear norm.

Result: OCR indeed drives both generalization and hallucination depending on causal relationships; factorized weight matrices enable OCR learning while combined weights cannot; gradient descent’s implicit bias explains high sample efficiency in associating facts and implications.

Conclusion: OCR provides unified explanation for LLMs’ contradictory behaviors; matrix factorization is crucial for OCR capability; theoretical foundation offers new lens for analyzing and mitigating undesirable behaviors from knowledge injection.

Abstract: Large language models (LLMs) can acquire new knowledge through fine-tuning, but this process exhibits a puzzling duality: models can generalize remarkably from new facts, yet are also prone to hallucinating incorrect information. However, the reasons for this phenomenon remain poorly understood. In this work, we argue that both behaviors stem from a single mechanism known as out-of-context reasoning (OCR): the ability to deduce implications by associating concepts, even those without a causal link. Our experiments across five prominent LLMs confirm that OCR indeed drives both generalization and hallucination, depending on whether the associated concepts are causally related. To build a rigorous theoretical understanding of this phenomenon, we then formalize OCR as a synthetic factual recall task. We empirically show that a one-layer single-head attention-only transformer with factorized output and value matrices can learn to solve this task, while a model with combined weights cannot, highlighting the crucial role of matrix factorization. Our theoretical analysis shows that the OCR capability can be attributed to the implicit bias of gradient descent, which favors solutions that minimize the nuclear norm of the combined output-value matrix. This mathematical structure explains why the model learns to associate facts and implications with high sample efficiency, regardless of whether the correlation is causal or merely spurious. Ultimately, our work provides a theoretical foundation for understanding the OCR phenomenon, offering a new lens for analyzing and mitigating undesirable behaviors from knowledge injection.

Method: EAQuant introduces three expert-aware innovations: (1) smoothing aggregation to suppress activation outliers, (2) routing consistency alignment to preserve expert selection post-quantization, and (3) calibration data balance to optimize sparsely activated experts.

Result: EAQuant significantly outperforms existing methods across extreme quantization settings (W4A4/W3A4/W3A3/W2A4), achieving average accuracy improvements of 1.15-13.81% across three diverse MoE architectures, with particularly pronounced gains in reasoning tasks.

Conclusion: EAQuant establishes a new state-of-the-art for high-precision, efficient MoE model compression by integrating expert-aware quantization strategies that address the unique challenges of sparse activation and dynamic routing in MoE architectures.

Abstract: Mixture-of-Experts (MoE) models enable scalable computation and performance in large-scale deep learning but face quantization challenges due to sparse expert activation and dynamic routing. Existing post-training quantization (PTQ) methods fail to address activation outliers, routing instability, and sparse expert calibration, leading to significant performance degradation. To address this, we propose EAQuant, a PTQ framework tailored for MoE architectures. Our method introduces three expert-aware innovations: (1) smoothing aggregation to suppress activation outliers, (2) routing consistency alignment to preserve expert selection post-quantization, and (3) calibration data balance to optimize sparsely activated experts. These strategies collectively enable robust, high-precision quantization of MoE models under ultra-low-bit constraints.Extensive experiments across several extreme quantization settings (e.g., W4A4/W3A4/W3A3/W2A4) demonstrate that EAQuant significantly outperforms existing methods, achieving average accuracy improvements of 1.15 - 13.81% across three diverse MoE architectures, with particularly pronounced gains in reasoning tasks and robust performance retention under aggressive quantization. By integrating these innovations, EAQuant establishes a new state-of-the-art for high-precision, efficient MoE model compression.Our code is available at https://github.com/darren-fzq1/EAQuant.

Method: FinCoT embeds expert financial reasoning blueprints into structured chain-of-thought prompts. It compares three approaches: standard prompting, unstructured CoT, and structured CoT with domain expertise across ten CFA-style financial domains.

Result: FinCoT improved Qwen3-8B-Base accuracy from 63.2% to 80.5% and Fin-R1 (7B) from 65.7% to 75.7%, while reducing output length by up to 8.9x and 1.16x compared to structured CoT methods.

Conclusion: Structured CoT with domain expertise (FinCoT) significantly improves LLM performance in finance, reduces inference costs, and yields more interpretable, expert-aligned reasoning, especially beneficial for models lacking financial post-training.

Abstract: This paper presents FinCoT, a structured chain-of-thought (CoT) prompting framework that embeds domain-specific expert financial reasoning blueprints to guide large language models’ behaviors. We identify three main prompting styles in financial NLP (FinNLP): (1) standard prompting (zero-shot), (2) unstructured CoT (free-form reasoning), and (3) structured CoT (with explicitly structured reasoning steps). Prior work has mainly focused on the first two, while structured CoT remains underexplored and lacks domain expertise incorporation. Therefore, we evaluate all three prompting approaches across ten CFA-style financial domains and introduce FinCoT as the first structured finance-specific prompting approach incorporating blueprints from domain experts. FinCoT improves the accuracy of a general-purpose model, Qwen3-8B-Base, from 63.2% to 80.5%, and boosts Fin-R1 (7B), a finance-specific model, from 65.7% to 75.7%, while reducing output length by up to 8.9x and 1.16x compared to structured CoT methods, respectively. We find that FinCoT proves most effective for models lacking financial post-training. Our findings show that FinCoT does not only improve performance and reduce inference costs but also yields more interpretable and expert-aligned reasoning traces.

Method: Customized neural morphological analyzer to annotate Latin and Italian corpora, then computationally validated crowd-sourced lists of defective verbs from Wiktionary against the massive annotated corpus data.

Result: Wiktionary provides highly reliable account of Italian morphological gaps, but 7% of Latin lemmata listed as defective show strong corpus evidence of being non-defective, revealing limitations of crowd-sourced wikis.

Conclusion: Crowd-sourced wikis have value for rare linguistic features but limitations for less-studied phenomena/languages. The work provides scalable tools for quality assurance of crowd-sourced data, advancing computational morphology and linguistic knowledge of defectivity.

Abstract: Morphological defectivity is an intriguing and understudied phenomenon in linguistics. Addressing defectivity, where expected inflectional forms are absent, is essential for improving the accuracy of NLP tools in morphologically rich languages. However, traditional linguistic resources often lack coverage of morphological gaps as such knowledge requires significant human expertise and effort to document and verify. For scarce linguistic phenomena in under-explored languages, Wikipedia and Wiktionary often serve as among the few accessible resources. Despite their extensive reach, their reliability has been a subject of controversy. This study customizes a novel neural morphological analyzer to annotate Latin and Italian corpora. Using the massive annotated data, crowd-sourced lists of defective verbs compiled from Wiktionary are validated computationally. Our results indicate that while Wiktionary provides a highly reliable account of Italian morphological gaps, 7% of Latin lemmata listed as defective show strong corpus evidence of being non-defective. This discrepancy highlights potential limitations of crowd-sourced wikis as definitive sources of linguistic knowledge, particularly for less-studied phenomena and languages, despite their value as resources for rare linguistic features. By providing scalable tools and methods for quality assurance of crowd-sourced data, this work advances computational morphology and expands linguistic knowledge of defectivity in non-English, morphologically rich languages.

Method: Simple supervision techniques applied across three datasets spanning various topics, evaluating alignment with diverse population groups and analyzing distributional alignment

Result: Simple supervision consistently improves language model alignment with diverse population groups, with analysis showing how alignment varies across specific groups

Conclusion: The research provides insights into distributional alignment of LLMs with diverse populations and establishes a benchmark for future research through evaluation of many LLMs and prompting strategies

Abstract: The ability to accurately align LLMs with population groups on subjective questions would have great value. In this work, we show that simple supervision can more consistently improve language model alignment with diverse population groups, as measured across three datasets spanning various topics. Beyond evaluating average alignment, we also report how alignment varies across specific groups. Our broad findings provide insights into the distributional alignment of LLMs with diverse populations. By conducting evaluation over many LLMs and prompting strategies, we provide a benchmark to stimulate future research.

Method: DP-Fusion works by: (1) labeling sensitive tokens, (2) inferring without sensitive tokens for baseline, (3) inferring with sensitive tokens, (4) blending distributions to bound distance from baseline. Uses differential privacy parameter ε to control privacy/utility trade-off.

Result: Achieves token-level provably privatized documents with substantially improved theoretical and empirical privacy, achieving 6× lower perplexity than related differentially private inference methods.

Conclusion: DP-Fusion provides a practical differentially private inference mechanism for LLMs that balances privacy protection with text quality, addressing document privatization needs while mitigating jailbreak-style prompt injection risks.

Abstract: Large language models (LLMs) do not preserve privacy at inference-time. The LLM’s outputs can inadvertently reveal information about the model’s context, which presents a privacy challenge when the LLM is augmented via tools or databases containing sensitive information. Existing privacy-preserving methods at inference-time have significant limitations since they (i) lack provable guarantees or (ii) have a poor utility/privacy trade-off. We propose DP-Fusion, a Differentially Private Inference (DPI) mechanism for LLMs that provably bounds the influence a set of tokens in the context can have on the LLM’s output. DP-Fusion works as follows: (1) label a subset of sensitive tokens, (2) infer the LLM without any sensitive tokens to obtain a baseline, (3) infer the LLM with the sensitive tokens, and (4) blend distributions so that the final output remains within a bounded distance of the baseline distribution. While this per-token influence bound also mitigates jailbreak-style prompt injection, we focus on \emph{document privatization}, where the goal is to paraphrase a document containing sensitive tokens, e.g., personally identifiable information, so that no attacker can reliably infer them from the paraphrased document while preserving high text quality. The privacy/utility trade-off is controlled by $ε$, where $ε=0$ hides sensitive tokens entirely, while higher values trade off privacy for improved text quality. We show that our method creates token-level provably privatized documents with substantially improved theoretical and empirical privacy, achieving $6\times$ lower perplexity than related DPI methods.

Method: Three-stage pipeline: (1) Generate unsafe content via LLM-driven red teaming, (2) Label with semi-automated multi-label annotation using majority-voted LLM labelers, (3) Translate with high-fidelity toxicity-preserving translation.

Result: Created RabakBench with over 5,000 examples across six safety categories, achieving 0.70-0.80 inter-annotator agreement. Evaluations of 13 state-of-the-art guardrails show significant performance degradation in these languages.

Conclusion: RabakBench provides a reproducible framework for building safety benchmarks in underserved communities and demonstrates the critical need for localized safety evaluation beyond high-resource languages.

Abstract: Large language models (LLMs) often fail to maintain safety in low-resource language varieties, such as code-mixed vernaculars and regional dialects. We introduce RabakBench, a multilingual safety benchmark and scalable pipeline localized to Singapore’s unique linguistic landscape, covering Singlish, Chinese, Malay, and Tamil. We construct the benchmark through a three-stage pipeline: (1) Generate: augmenting real-world unsafe web content via LLM-driven red teaming; (2) Label: applying semi-automated multi-label annotation using majority-voted LLM labelers; and (3) Translate: performing high-fidelity, toxicity-preserving translation. The resulting dataset contains over 5,000 examples across six fine-grained safety categories. Despite using LLMs for scalability, our framework maintains rigorous human oversight, achieving 0.70-0.80 inter-annotator agreement. Evaluations of 13 state-of-the-art guardrails reveal significant performance degradation, underscoring the need for localized evaluation. RabakBench provides a reproducible framework for building safety benchmarks in underserved communities.

Method: Created AblationBench with two tasks: AuthorAblation (83 instances) for proposing ablation experiments from method sections, and ReviewerAblation (350 instances) for finding missing ablations in full papers. Developed LM-based judges for automatic evaluation and tested frontier language models on these tasks.

Result: Current LMs perform poorly on ablation planning tasks, with the best system identifying only 38% of original ablations on average (below human-level). Found inverse performance trend between author and reviewer tasks attributed to model grounding differences. Chain-of-thought prompting outperformed agent-based approaches.

Conclusion: AblationBench provides a standardized framework for evaluating language model agents’ scientific reasoning capabilities in ablation planning. The tasks remain challenging for current LMs, highlighting limitations in scientific reasoning and the need for improved grounding and reasoning approaches.

Abstract: Language model agents are increasingly used to automate scientific research, yet evaluating their scientific contributions remains a challenge. A key mechanism to obtain such insights is through ablation experiments. To this end, we introduce AblationBench, a benchmark suite for evaluating agents on ablation planning tasks in empirical AI research. It includes two tasks: AuthorAblation, which helps authors propose ablation experiments based on a method section and contains 83 instances, and ReviewerAblation, which helps reviewers find missing ablations in a full paper and contains 350 instances. For both tasks, we develop LM-based judges that serve as an automatic evaluation framework. Our experiments with frontier LMs show that these tasks remain challenging, with the best-performing LM system identifying only 38% of the original ablations on average, below human-level performance. We observe an inverse performance trend between the author and reviewer tasks, which we attribute to differences in model grounding. Lastly, we analyze the limitations of current LMs on these tasks, and find that chain-of-thought prompting outperforms an agent-based approach. Our data is available on https://huggingface.co/collections/ai-coscientist/ablationbench, and our code is available on https://github.com/ai-scientist-bench/ablation-bench .

Method: Apply weighted loss function to Transformer models (BERT, RoBERTa, BART) with dynamic class weight adjustment, evaluated on BRIGHTER dataset using multiple metrics.

Result: Weighted loss improves performance on high-frequency emotion classes but has limited impact on minority classes, showing both effectiveness and challenges.

Conclusion: Weighted loss functions can help with data imbalance in emotion detection but have limitations for minority classes, highlighting need for further research.

Abstract: This paper explores the application of a simple weighted loss function to Transformer-based models for multi-label emotion detection in SemEval-2025 Shared Task 11. Our approach addresses data imbalance by dynamically adjusting class weights, thereby enhancing performance on minority emotion classes without the computational burden of traditional resampling methods. We evaluate BERT, RoBERTa, and BART on the BRIGHTER dataset, using evaluation metrics such as Micro F1, Macro F1, ROC-AUC, Accuracy, and Jaccard similarity coefficients. The results demonstrate that the weighted loss function improves performance on high-frequency emotion classes but shows limited impact on minority classes. These findings underscore both the effectiveness and the challenges of applying this approach to imbalanced multi-label emotion detection.

Method: Created a manually curated dataset with questions and answers grounded in Wikipedia, developed by native speakers from Czechia, Slovakia, and Ukraine. Includes both textual and visual questions with English translations. Evaluated state-of-the-art LLMs through prompting and added human judgments for answer correctness. Analyzed reliability of automatic evaluation metrics using human evaluations.

Result: Best open-weight LLMs achieved only over 40% accuracy on textual questions and below 30% on visual questions. LLM-based evaluation metrics showed strong correlation with human judgment, while traditional string-overlap metrics performed surprisingly well due to prevalence of named entities in answers.

Conclusion: The CUS-QA benchmark reveals significant gaps in current LLM capabilities for multimodal regional question answering, particularly for visual understanding tasks. The dataset provides a valuable resource for evaluating and improving multimodal AI systems for specific cultural contexts.

Abstract: We introduce CUS-QA, a benchmark for evaluation of open-ended regional question answering that encompasses both textual and visual modalities. We also provide strong baselines using state-of-the-art large language models (LLMs). Our dataset consists of manually curated questions and answers grounded in Wikipedia, created by native speakers from Czechia, Slovakia, and Ukraine, with accompanying English translations. It includes both purely textual questions and those requiring visual understanding. We evaluate state-of-the-art LLMs through prompting and add human judgments of answer correctness. Using these human evaluations, we analyze the reliability of existing automatic evaluation metrics. Our baseline results show that even the best open-weight LLMs achieve only over 40% accuracy on textual questions and below 30% on visual questions. LLM-based evaluation metrics show strong correlation with human judgment, while traditional string-overlap metrics perform surprisingly well due to the prevalence of named entities in answers.

Method: Formulates CKGE as sequential Bayesian inference problem using Bayesian posterior update principle as continual learning strategy. Treats each batch of new data as Bayesian update to model’s prior, maintaining posterior distribution to preserve earlier knowledge. Introduces continual clustering method with regularization term to maintain compact cluster structure of entity embeddings for semantic consistency while allowing controlled adaptation.

Result: Extensive experiments on multiple CKGE benchmarks show BAKE achieves top performance in vast majority of cases compared to existing approaches.

Conclusion: BAKE effectively addresses catastrophic forgetting in continual knowledge graph embedding through Bayesian sequential inference and clustering regularization, enabling better preservation of prior knowledge while adapting to new information in dynamic social media environments.

Abstract: As social media and the World Wide Web become hubs for information dissemination, effectively organizing and understanding the vast amounts of dynamically evolving Web content is crucial. Knowledge graphs (KGs) provide a powerful framework for structuring this information. However, the rapid emergence of new hot topics, user relationships, and events in social media renders traditional static knowledge graph embedding (KGE) models rapidly outdated. Continual Knowledge Graph Embedding (CKGE) aims to address this issue, but existing methods commonly suffer from catastrophic forgetting, whereby older, but still valuable, information is lost when learning new knowledge (such as new memes or trending events). This means the model cannot effectively learn the evolution of the data. We propose a novel CKGE framework, BAKE. Unlike existing methods, BAKE formulates CKGE as a sequential Bayesian inference problem and utilizes the Bayesian posterior update principle as a natural continual learning strategy. This principle is insensitive to data order and provides theoretical guarantees to preserve prior knowledge as much as possible. Specifically, we treat each batch of new data as a Bayesian update to the model’s prior. By maintaining the posterior distribution, the model effectively preserves earlier knowledge even as it evolves over multiple snapshots. Furthermore, to constrain the evolution of knowledge across snapshots, we introduce a continual clustering method that maintains the compact cluster structure of entity embeddings through a regularization term, ensuring semantic consistency while allowing controlled adaptation to new knowledge. We conduct extensive experiments on multiple CKGE benchmarks, which demonstrate that BAKE achieves the top performance in the vast majority of cases compared to existing approaches.

Method: Created a cross-linguistic benchmark with 800 unique prompts across five gendered languages (French, Spanish, German, Italian, Russian) and two gender-neutral control languages (English, Chinese). Generated 28,800 images using three state-of-the-art T2I models, focusing on words where grammatical gender contradicts stereotypical gender associations.

Result: Grammatical gender dramatically influences image generation: masculine grammatical markers increase male representation to 73% on average (vs 22% in gender-neutral English), while feminine markers increase female representation to 38% (vs 28% in English). Effects vary by language resource availability and model architecture, with high-resource languages showing stronger effects.

Conclusion: Language structure itself, not just content, shapes AI-generated visual outputs, introducing a new dimension for understanding bias and fairness in multilingual, multimodal systems. This establishes grammatical gender as a significant factor in T2I model bias.

Abstract: Research on bias in Text-to-Image (T2I) models has primarily focused on demographic representation and stereotypical attributes, overlooking a fundamental question: how does grammatical gender influence visual representation across languages? We introduce a cross-linguistic benchmark examining words where grammatical gender contradicts stereotypical gender associations (e.g., une sentinelle'' - grammatically feminine in French but referring to the stereotypically masculine concept guard’’). Our dataset spans five gendered languages (French, Spanish, German, Italian, Russian) and two gender-neutral control languages (English, Chinese), comprising 800 unique prompts that generated 28,800 images across three state-of-the-art T2I models. Our analysis reveals that grammatical gender dramatically influences image generation: masculine grammatical markers increase male representation to 73% on average (compared to 22% with gender-neutral English), while feminine grammatical markers increase female representation to 38% (compared to 28% in English). These effects vary systematically by language resource availability and model architecture, with high-resource languages showing stronger effects. Our findings establish that language structure itself, not just content, shapes AI-generated visual outputs, introducing a new dimension for understanding bias and fairness in multilingual, multimodal systems.

Method: BiasGym consists of two components: BiasInject (safely injects specific biases into models via token-based fine-tuning while keeping the model frozen) and BiasScope (leverages injected signals to identify and reliably steer components responsible for biased behavior).

Result: The method enables consistent bias elicitation for mechanistic analysis, supports targeted debiasing without degrading downstream task performance, and generalizes to biases unseen during fine-tuning. Demonstrated effectiveness in reducing real-world stereotypes like “people from Italy being reckless drivers.”

Conclusion: BiasGym provides a simple, cost-effective, and generalizable framework for bias analysis and mitigation that is useful for both safety interventions and interpretability research in LLMs.

Abstract: Understanding biases and stereotypes encoded in the weights of Large Language Models (LLMs) is crucial for developing effective mitigation strategies. However, biased behaviour is often subtle and non-trivial to isolate, even when deliberately elicited, making systematic analysis and debiasing particularly challenging. To address this, we introduce \texttt{BiasGym}, a simple, cost-effective, and generalizable framework for reliably and safely injecting, analyzing, and mitigating conceptual associations of biases within LLMs. \texttt{BiasGym} consists of two components: \texttt{BiasInject}, which safely injects specific biases into the model via token-based fine-tuning while keeping the model frozen, and \texttt{BiasScope}, which leverages these injected signals to identify and reliably steer the components responsible for biased behavior. Our method enables consistent bias elicitation for mechanistic analysis, supports targeted debiasing without degrading performance on downstream tasks, and generalizes to biases unseen during fine-tuning. We demonstrate the effectiveness of BiasGym in reducing real-world stereotypes (e.g., people from Italy being `reckless drivers’), showing its utility for both safety interventions and interpretability research.

Method: IAPO (Inference-Aware Prompt Optimization) jointly optimizes prompts and inference scale while being aware of inference budget and task objectives. PSST (Prompt Scaling via Sequential Trimming) is a fixed-budget training algorithm developed for IAPO with finite-budget error probability guarantees.

Result: The paper demonstrates the effectiveness of PSST on six tasks including multi-objective text generation and reasoning, showing the critical role of incorporating inference-awareness in aligning black-box LLMs through prompt optimization.

Conclusion: Inference-aware prompt optimization is crucial for effective alignment of black-box LLMs, and the IAPO framework with PSST algorithm successfully addresses the interdependence between prompt optimization and inference strategies while considering practical constraints.

Abstract: Prompt optimization methods have demonstrated significant effectiveness in aligning black-box large language models (LLMs). In parallel, inference scaling strategies such as Best-of-N Sampling and Majority Voting have likewise been shown to improve alignment and performance by trading additional computation for better output. However, existing prompt optimization approaches are inference strategy agnostic; that is, they optimize prompts without accounting for the inference strategy. This constitutes a significant methodological gap, as our empirical and theoretical analysis reveals a strong interdependence between these two paradigms. Moreover, we find that user preferences regarding trade-offs among multiple objectives and inference budgets substantially influence the choice of prompt and inference configuration. To address this gap, we introduce a novel unified framework named IAPO (Inference-Aware Prompt Optimization) that jointly optimizes the prompt and inference scale, while being aware of the inference budget and different task objectives. We then develop a fixed-budget training algorithm for IAPO, called PSST (Prompt Scaling via Sequential Trimming), and establish finite-budget guarantees on the error probability. Finally, we evaluate the effectiveness of PSST on six tasks, including multi-objective text generation and reasoning, and demonstrate the critical role of incorporating inference-awareness in aligning black-box LLMs using prompt optimization.

Method: Three-stage framework: 1) Coarse-grained initialization builds structurally complete but semantically coarse dialogue skeletons; 2) Iterative refinement introduces realistic complexities via mask-and-fill operations; 3) Offline verification ensures correctness and coherence through rule- and model-based checks.

Result: Experiments demonstrate that ToolACE-MT enables efficient, effective, and generalizable agentic data generation, offering a new paradigm for high-quality data construction in tool-augmented LLM scenarios.

Conclusion: ToolACE-MT provides a novel non-autoregressive iterative generation framework that addresses the limitations of existing simulation-based methods, enabling more efficient construction of high-quality multi-turn agentic dialogues for tool-augmented LLMs.

Abstract: Agentic task-solving with Large Language Models (LLMs) requires multi-turn, multi-step interactions, often involving complex function calls and dynamic user-agent exchanges. Existing simulation-based data generation methods for such scenarios rely heavily on costly autoregressive interactions between multiple LLM agents, thereby limiting real-world performance of agentic tasks. In this paper, we propose ToolACE-MT, a novel Non-Autoregressive Iterative Generation framework for constructing high-quality multi-turn agentic dialogues. ToolACE-MT generates full conversational trajectories through three stages: coarse-grained initialization, iterative refinement, and offline verification. The initialization phase builds a structurally complete yet semantically coarse dialogue skeleton; the iterative refinement phase introduces realistic complexities and continued refinement via mask-and-fill operations; and the offline verification phase ensures correctness and coherence via rule- and model-based checks. Experiments demonstrate that ToolACE-MT enables efficient, effective and generalizable agentic data generation, offering a new paradigm for high-quality data construction in tool-augmented LLM scenarios.

Method: Two-stage framework: (1) Grounding module extracts relations using shortest dependency paths to create minimal relational contexts, (2) Refinement module aggregates predictions and revises them holistically. Also includes causality-driven reward model for RLHF fine-tuning.

Result: Reduces average hallucination rate to 7.9% and achieves 9.3% improvement in average F1 score over existing LLM-based baselines across eight benchmarks.

Conclusion: DEPTH effectively addresses LLM hallucination in relation extraction through dependency-aware simplification and hierarchical refinement, significantly improving reliability and accuracy.

Abstract: Relation extraction (RE) enables the construction of structured knowledge for many downstream applications. While large language models (LLMs) have shown great promise in this task, they often struggle to reliably determine whether a relation exists, particularly in sentences with complex syntax or subtle semantics. For instance, we find that Qwen2.5-14B-Instruct incorrectly predicts a relation in 96.9% of NO-RELATION instances on SciERC, revealing a severe hallucination problem. To address these challenges, we propose DEPTH, a framework that integrates Dependency-aware sEntence simPlification and Two-tiered Hierarchical refinement into the relation extraction pipeline. Given a sentence and its candidate entity pairs, DEPTH operates in two stages: (1) the Grounding module extracts relations for each pair by leveraging their shortest dependency path, distilling the sentence into a minimal yet coherent relational context that reduces syntactic noise while preserving key semantics; (2) the Refinement module aggregates all local predictions and revises them based on a holistic understanding of the sentence, correcting omissions and inconsistencies. We further introduce a causality-driven reward model that mitigates reward hacking by disentangling spurious correlations, enabling robust fine-tuning via reinforcement learning with human feedback. Experiments on eight well-established benchmarks demonstrate that DEPTH reduces the average hallucination rate to 7.9% while achieving a 9.3% improvement in average F1 score over existing LLM-based extraction baselines.

Method: An agentic RAG system trained end-to-end via reinforcement learning, treating LLM as an agent in an environment of 16,000+ disease profiles, 150,000+ patient records, and 27M+ biomedical documents. Uses soft verifiable rewards to co-optimize retrieval and reasoning, learning to formulate queries, evaluate evidence, and refine searches.

Result: Outperforms prompt-engineering and training-free RAG methods, achieving 22.7% average accuracy gain over second-best model on ID/OOD benchmarks. Boosts physicians’ diagnostic accuracy from 45.6% to 69.1% on 150 real-world cases. Surpasses GPT-4o, DeepSeek-R1, and medical-specific frameworks.

Conclusion: Evolving agentic systems to leverage statistical regularities in large-scale healthcare data is key for trustworthy diagnostic assistants. The approach demonstrates superior diagnostic reasoning capabilities.

Abstract: The integration of Large Language Models (LLMs) into healthcare is constrained by knowledge limitations, hallucinations, and a disconnect from Evidence-Based Medicine (EBM). While Retrieval-Augmented Generation (RAG) offers a solution, current systems often rely on static workflows that miss the iterative, hypothetico-deductive reasoning of clinicians. To address this, we introduce Deep-DxSearch, an agentic RAG system trained end-to-end via reinforcement learning (RL) for traceable diagnostic reasoning. Deep-DxSearch acts as an active investigator, treating the LLM as an agent within an environment of 16,000+ guideline-derived disease profiles, 150,000+ patient records for case-based reasoning, and over 27 million biomedical documents. Using soft verifiable rewards that co-optimize retrieval and reasoning, the model learns to formulate queries, evaluate evidence, and refine searches to close diagnostic gaps. Experiments show our end-to-end RL framework consistently outperforms prompt-engineering and training-free RAG methods. On in-distribution (ID) and out-of-distribution (OOD) benchmarks for common and rare diseases, Deep-DxSearch surpasses strong baselines-including GPT-4o, DeepSeek-R1, and medical-specific frameworks-achieving an average accuracy gain of 22.7% over the second-best model. In validation with 150 real-world cases, Deep-DxSearch boosts physicians’ average diagnostic accuracy from 45.6% to 69.1%. These results indicate that evolving agentic systems to leverage statistical regularities in large-scale healthcare data is key for trustworthy diagnostic assistants. All data, code, and checkpoints are available at https://qiaoyu-zheng.github.io/Deep-DxSearch.

Method: Retrieval-Augmented Generation (RAG) framework with semantic retrieval and contextual summarization, tested on 10,000 auction records from Getty Provenance Index - German Sales to enable natural-language multilingual searches.

Result: The approach successfully retrieves and summarizes auction records, reducing dependence on metadata structures and providing a scalable solution for navigating art market archives.

Conclusion: RAG offers a practical tool for historians and cultural heritage professionals conducting historically sensitive research by enabling more flexible, exploratory access to fragmented art provenance data.

Abstract: This research presents a Retrieval-Augmented Generation (RAG) framework for art provenance studies, focusing on the Getty Provenance Index. Provenance research establishes the ownership history of artworks, which is essential for verifying authenticity, supporting restitution and legal claims, and understanding the cultural and historical context of art objects. The process is complicated by fragmented, multilingual archival data that hinders efficient retrieval. Current search portals require precise metadata, limiting exploratory searches. Our method enables natural-language and multilingual searches through semantic retrieval and contextual summarization, reducing dependence on metadata structures. We assess RAG’s capability to retrieve and summarize auction records using a 10,000-record sample from the Getty Provenance Index - German Sales. The results show this approach provides a scalable solution for navigating art market archives, offering a practical tool for historians and cultural heritage professionals conducting historically sensitive research.

Method: Proposes CCF (Context Compression Framework) that learns hierarchical latent representations through segment-wise semantic aggregation with key-value memory encoding. Uses incremental segment decoding with sparse reservoir sampling for training efficiency.

Result: Achieves competitive perplexity under high compression ratios on multiple long-context language modeling benchmarks. Significantly improves throughput and memory efficiency compared to existing approaches.

Conclusion: CCF demonstrates the potential of structured compression for scalable and effective long-context language modeling, offering efficient solutions to computational and memory challenges.

Abstract: Scaling language models to longer contexts is essential for capturing rich dependencies across extended discourse. However, naïve context extension imposes significant computational and memory burdens, often resulting in inefficiencies during both training and inference. In this work, we propose CCF, a novel context compression framework designed to enable efficient long-context modeling by learning hierarchical latent representations that preserve global semantics while aggressively reducing input redundancy. CCF integrates segment-wise semantic aggregation with key-value memory encoding, forming compact representations that support accurate reconstruction and long-range understanding. To further enhance scalability, we introduce a training-efficient optimization strategy that couples incremental segment decoding with sparse reservoir sampling, substantially reducing memory overhead without degrading performance. Empirical results on multiple long-context language modeling benchmarks demonstrate that CCF achieves competitive perplexity under high compression ratios, and significantly improves throughput and memory efficiency compared to existing approaches. These findings highlight the potential of structured compression for scalable and effective long-context language modeling.

Method: Developed maiBERT using BERT architecture with Masked Language Modeling (MLM) pre-training on a newly constructed Maithili corpus, then evaluated through a news classification task.

Result: maiBERT achieved 87.02% accuracy on news classification, outperforming existing regional models like NepBERTa and HindiBERT with 0.13% overall accuracy gain and 5-7% improvement across various classes.

Conclusion: maiBERT successfully addresses the NLU gap for Maithili, demonstrating effectiveness for low-resource languages, and has been open-sourced on Hugging Face for further fine-tuning on downstream tasks like sentiment analysis and NER.

Abstract: Natural Language Understanding (NLU) for low-resource languages remains a major challenge in NLP due to the scarcity of high-quality data and language-specific models. Maithili, despite being spoken by millions, lacks adequate computational resources, limiting its inclusion in digital and AI-driven applications. To address this gap, we introducemaiBERT, a BERT-based language model pre-trained specifically for Maithili using the Masked Language Modeling (MLM) technique. Our model is trained on a newly constructed Maithili corpus and evaluated through a news classification task. In our experiments, maiBERT achieved an accuracy of 87.02%, outperforming existing regional models like NepBERTa and HindiBERT, with a 0.13% overall accuracy gain and 5-7% improvement across various classes. We have open-sourced maiBERT on Hugging Face enabling further fine-tuning for downstream tasks such as sentiment analysis and Named Entity Recognition (NER).

Method: Adds explicit group-fairness constraints to the quantization objective, guiding the rounding operation toward less-biased text generation for protected groups (gender, race, religion). Focuses on reducing occupational bias and discriminatory language.

Result: Fair-GPTQ preserves at least 90% of baseline accuracy on zero-shot benchmarks, reduces unfairness relative to half-precision models, retains memory/speed benefits of 4-bit quantization, and performs on par with iterative null-space projection debiasing on racial-stereotype benchmarks.

Conclusion: The approach successfully addresses the fairness-quantization trade-off, provides a theoretical solution with group-bias constraints, and enables analysis of channel- and weight-level contributions to fairness during quantization.

Abstract: High memory demands of generative language models have drawn attention to quantization, which reduces computational cost, memory usage, and latency by mapping model weights to lower-precision integers. Approaches such as GPTQ effectively minimize input-weight product errors during quantization; however, recent empirical studies show that they can increase biased outputs and degrade performance on fairness benchmarks, and it remains unclear which specific weights cause this issue. In this work, we draw new links between quantization and model fairness by adding explicit group-fairness constraints to the quantization objective and introduce Fair-GPTQ, the first quantization method explicitly designed to reduce unfairness in large language models. The added constraints guide the learning of the rounding operation toward less-biased text generation for protected groups. Specifically, we focus on stereotype generation involving occupational bias and discriminatory language spanning gender, race, and religion. Fair-GPTQ has minimal impact on performance, preserving at least 90% of baseline accuracy on zero-shot benchmarks, reduces unfairness relative to a half-precision model, and retains the memory and speed benefits of 4-bit quantization. We also compare the performance of Fair-GPTQ with existing debiasing methods and find that it achieves performance on par with the iterative null-space projection debiasing approach on racial-stereotype benchmarks. Overall, the results validate our theoretical solution to the quantization problem with a group-bias term, highlight its applicability for reducing group bias at quantization time in generative models, and demonstrate that our approach can further be used to analyze channel- and weight-level contributions to fairness during quantization.

Method: 1) Cleaned and restructured MedCalc-Bench dataset with step-by-step evaluation pipeline assessing formula selection, entity extraction, and arithmetic computation separately; 2) Introduced automatic error analysis framework for structured failure attribution; 3) Proposed MedRaC - a modular agentic pipeline combining retrieval-augmented generation and Python-based code execution.

Result: Under granular evaluation, GPT-4o accuracy dropped from 62.7% to 43.6%, revealing previously masked errors. MedRaC improved different LLMs’ accuracy from 16.35% up to 53.19% without fine-tuning. Human evaluation confirmed the error analysis framework aligns with expert judgment.

Conclusion: Current benchmark practices have limitations for medical calculations. The proposed methodology enables transparent, transferable reasoning evaluation, moving closer to trustworthy LLM-based systems for real-world medical applications.

Abstract: Large language models (LLMs) have demonstrated promising performance on medical benchmarks; however, their ability to perform medical calculations, a crucial aspect of clinical decision-making, remains underexplored and poorly evaluated. Existing benchmarks often assess only the final answer with a wide numerical tolerance, overlooking systematic reasoning failures and potentially causing serious clinical misjudgments. In this work, we revisit medical calculation evaluation with a stronger focus on clinical trustworthiness. First, we clean and restructure the MedCalc-Bench dataset and propose a new step-by-step evaluation pipeline that independently assesses formula selection, entity extraction, and arithmetic computation. Under this granular framework, the accuracy of GPT-4o drops from 62.7% to 43.6%, revealing errors masked by prior evaluations. Second, we introduce an automatic error analysis framework that generates structured attribution for each failure mode. Human evaluation confirms its alignment with expert judgment, enabling scalable and explainable diagnostics. Finally, we propose a modular agentic pipeline, MedRaC, that combines retrieval-augmented generation and Python-based code execution. Without any fine-tuning, MedRaC improves the accuracy of different LLMs from 16.35% up to 53.19%. Our work highlights the limitations of current benchmark practices and proposes a more clinically faithful methodology. By enabling transparent and transferable reasoning evaluation, we move closer to making LLM-based systems trustworthy for real-world medical applications.

Method: FS-DFM (Few-Step Discrete Flow-Matching) makes the number of sampling steps an explicit parameter and trains the model to be consistent across step budgets, so one big move lands where many small moves would. It uses a reliable update rule that moves probability in the right direction without overshooting, and incorporates strong teacher guidance distilled from long-run trajectories to make few-step sampling stable, accurate, and easy to control.

Result: On language modeling benchmarks, FS-DFM with only 8 sampling steps achieves perplexity parity with a 1,024-step discrete-flow baseline for generating 1,024 tokens using a similar-size model, delivering up to 128 times faster sampling and corresponding latency/throughput gains.

Conclusion: FS-DFM successfully bridges the gap between autoregressive and diffusion language models by enabling high-quality parallel generation with dramatically fewer sampling steps, making diffusion-based language generation practical for real-time applications.

Abstract: Autoregressive language models (ARMs) deliver strong likelihoods, but are inherently serial: they generate one token per forward pass, which limits throughput and inflates latency for long sequences. Diffusion Language Models (DLMs) parallelize across positions and thus appear promising for language generation, yet standard discrete diffusion typically needs hundreds to thousands of model evaluations to reach high quality, trading serial depth for iterative breadth. We introduce FS-DFM, Few-Step Discrete Flow-Matching. A discrete flow-matching model designed for speed without sacrificing quality. The core idea is simple: make the number of sampling steps an explicit parameter and train the model to be consistent across step budgets, so one big move lands where many small moves would. We pair this with a reliable update rule that moves probability in the right direction without overshooting, and with strong teacher guidance distilled from long-run trajectories. Together, these choices make few-step sampling stable, accurate, and easy to control. On language modeling benchmarks, FS-DFM with 8 sampling steps achieves perplexity parity with a 1,024-step discrete-flow baseline for generating 1,024 tokens using a similar-size model, delivering up to 128 times faster sampling and corresponding latency/throughput gains.

Method: Develops a lightweight single-layer “sampler” on top of existing large diffusion LMs. One full-model forward pass is followed by multiple forward passes of only this sampler layer to yield multiple unmasked tokens. The sampler is trained to mimic exact joint sampling from the frozen full model.

Result: When unmasking four tokens per denoising step, achieves MAUVE score of 0.87 (vs marginal baseline of 0.31) with respect to true joint distribution. Shows effectiveness on pretrained-only (Dream-7B-Base, Llada-7B-Base) and instruction-tuned models (Dream-7B-Instruct, Dream-7B-Coder) on language modeling and math & coding tasks.

Conclusion: Proposes an effective approximate joint sampling method for diffusion language models that enables parallel token generation while maintaining distributional fidelity, bridging the speed-accuracy trade-off in diffusion-based text generation.

Abstract: In autoregressive language models, each token is sampled by conditioning on all the past tokens; the overall string has thus been sampled from the correct underlying joint distribution represented by the model. In contrast, masked diffusion language models generate text by unmasking tokens out of order and potentially in parallel. Generating an overall string sampled from the correct underlying joint distribution would (again) require exactly one token unmasking in every full-model forward pass. The more tokens unmasked in parallel, the further away the string is from the true joint; this can be seen in the resulting drop in accuracy (but, increase in speed). In this paper we devise a way to {\em approximately} sample multiple tokens from the joint distribution in a single full-model forward pass; we do so by developing a new lightweight single-layer ``sampler" on top of an existing large diffusion LM. One forward pass of the full model can now be followed by multiple forward passes of only this sampler layer, to yield multiple unmasked tokens. Our sampler is trained to mimic exact joint sampling from the (frozen) full model. We show the effectiveness of our approximate joint sampling for both pretrained-only (Dream-7B-Base, Llada-7B-Base) and instruction-tuned (Dream-7B-Instruct, Dream-7B-Coder) models on language modeling and math & coding tasks. When four tokens are unmasked for each full-model denoising step, our sampling algorithm achieves a MAUVE score of 0.87 (vs marginal baseline of 0.31) with respect to the true joint distribution.

Method: Developed a quantitative framework by extracting multi-dimensional evaluative features from human expert critiques using zero-shot classification, defined critic personas, and evaluated VLMs (Llama, Qwen, Gemini) using persona-guided prompting.

Result: Revealed current performance levels, strengths, and areas for improvement of VLMs in art critique generation, showing their potential and limitations in complex semantic understanding tasks.

Conclusion: VLMs show promise but have limitations in art critique generation; the framework provides insights for improving multimodal models in complex cultural and semantic understanding tasks.

Abstract: This study aims to test and evaluate the capabilities and characteristics of current mainstream Visual Language Models (VLMs) in generating critiques for traditional Chinese painting. To achieve this, we first developed a quantitative framework for Chinese painting critique. This framework was constructed by extracting multi-dimensional evaluative features covering evaluative stance, feature focus, and commentary quality from human expert critiques using a zero-shot classification model. Based on these features, several representative critic personas were defined and quantified. This framework was then employed to evaluate selected VLMs such as Llama, Qwen, or Gemini. The experimental design involved persona-guided prompting to assess the VLM’s ability to generate critiques from diverse perspectives. Our findings reveal the current performance levels, strengths, and areas for improvement of VLMs in the domain of art critique, offering insights into their potential and limitations in complex semantic understanding and content generation tasks. The code used for our experiments can be publicly accessed at: https://github.com/yha9806/VULCA-EMNLP2025.

Method: OTR reframes autoregressive learning as single-step RL: at each token generation step, samples multiple candidate tokens from current policy, uses ground-truth token as reward signal, applies policy gradient to update model with on-policy learning at token level.

Result: OTR consistently outperforms standard SFT across diverse benchmarks including mathematical reasoning, code generation, and general domain reasoning, demonstrating improved generalization.

Conclusion: OTR establishes that on-policy nature of data is critical for generalization, offering a practical alternative for fine-tuning LLMs that combines benefits of RL with SFT efficiency.

Abstract: Supervised fine-tuning (SFT) is the predominant method for adapting large language models (LLMs), yet it often struggles with generalization compared to reinforcement learning (RL). In this work, we posit that this performance disparity stems not just from the loss function, but from a more fundamental difference: SFT learns from a fixed, pre-collected dataset, whereas RL utilizes on-policy data sampled from the current policy. Building on this hypothesis, we introduce one-token rollout (OTR), a novel fine-tuning algorithm that guides SFT with the policy gradient method. OTR reframes the autoregressive learning process by treating each token generation as a single-step reinforcement learning trajectory. At each step, it performs a Monte Carlo ``rollout’’ by sampling multiple candidate tokens from the current policy’s distribution. The ground-truth token from the supervised data is then used to provide a reward signal to these samples. Guided by policy gradient, our algorithm repurposes static, off-policy supervised data into a dynamic, on-policy signal at the token level, capturing the generalization benefits of on-policy learning while bypassing the costly overhead of full sentence generation. Through extensive experiments on a diverse suite of challenging benchmarks spanning mathematical reasoning, code generation, and general domain reasoning, we demonstrate that OTR consistently outperforms standard SFT. Our findings establish OTR as a powerful and practical alternative for fine-tuning LLMs and provide compelling evidence that the on-policy nature of data is a critical driver of generalization, offering a promising new direction for fine-tuning LLMs.

Method: AdaDetectGPT introduces a novel classifier that adaptively learns a witness function from training data to enhance logits-based detectors. The method provides statistical guarantees on detection performance metrics (true positive rate, false positive rate, true negative rate, false negative rate).

Result: Extensive numerical studies show AdaDetectGPT nearly uniformly improves the state-of-the-art method across various dataset and LLM combinations, with improvements reaching up to 37%.

Conclusion: AdaDetectGPT represents a significant advancement in LLM text detection by adaptively learning from data rather than relying solely on log-probability statistics, offering both performance improvements and statistical guarantees.

Abstract: We study the problem of determining whether a piece of text has been authored by a human or by a large language model (LLM). Existing state of the art logits-based detectors make use of statistics derived from the log-probability of the observed text evaluated using the distribution function of a given source LLM. However, relying solely on log probabilities can be sub-optimal. In response, we introduce AdaDetectGPT – a novel classifier that adaptively learns a witness function from training data to enhance the performance of logits-based detectors. We provide statistical guarantees on its true positive rate, false positive rate, true negative rate and false negative rate. Extensive numerical studies show AdaDetectGPT nearly uniformly improves the state-of-the-art method in various combination of datasets and LLMs, and the improvement can reach up to 37%. A python implementation of our method is available at https://github.com/Mamba413/AdaDetectGPT.

Method: DRIFT (Dissatisfaction-Refined Iterative preFerence Training) anchors training on real-world DSAT signals and samples positives dynamically from the evolving policy. It uses real-world WildFeedback datasets and synthetic UltraFeedback datasets for training.

Result: DRIFT models achieve up to +6.23% (7B) / +7.61% (14B) on WildBench Task Score and up to +8.95% (7B) / +12.29% (14B) on AlpacaEval2 win rate over base models, outperforming iterative DPO and SPIN. 14B models surpass GPT-4o-mini on WildBench. DRIFT also preserves exploratory capacity and yields more diverse high-reward solutions.

Conclusion: DRIFT is an effective and scalable recipe for real-world post-training that leverages the most abundant and informative signal (user dissatisfaction). The method avoids gradient degeneration and preserves preference margins theoretically.

Abstract: Real-world large language model deployments (e.g., conversational AI systems, code generation assistants) naturally generate abundant implicit user dissatisfaction (DSAT) signals, as users iterate toward better answers through refinements, corrections, and expressed preferences, while explicit satisfaction (SAT) feedback is scarce. Existing preference learning approaches are poorly aligned with this data profile, as they rely on costly human annotations or assume plentiful positive responses. In this paper, we introduce \textbf{DRIFT} (\textbf{D}issatisfaction-\textbf{R}efined \textbf{I}terative pre\textbf{F}erence \textbf{T}raining), which anchors training on real-world DSAT signals and samples positives dynamically from the evolving policy. Empirically, DRIFT models trained on real-world \textit{WildFeedback} datasets and synthetic \textit{UltraFeedback} datasets achieve up to +6.23% (7B) / +7.61% (14B) on WildBench Task Score and up to +8.95% (7B) / +12.29% (14B) on AlpacaEval2 win rate over base models, outperforming strong baseline methods such as iterative DPO and SPIN. At larger scales, the improvements are particularly pronounced: 14B models trained with DRIFT surpass GPT-4o-mini on WildBench. Further analysis shows that DRIFT also preserves exploratory capacity, yielding more diverse high-reward solutions rather than collapsing to narrow subsets. Theoretically, we demonstrate that this design preserves preference margins and avoids the gradient degeneration. These results show that DRIFT is an effective and scalable recipe for real-world post-training that leverages the most abundant and informative signal. The code and data are available at https://github.com/cacayaya/DRIFT.git.

Method: Compare three settings: 1) text models, 2) speech models, and 3) cascaded systems (speech → automatic transcription → text model). Focus on German dialects for intent classification, releasing first dialectal audio intent classification dataset, with supporting topic classification experiments.

Result: Speech-only setup provides best results on dialect data, text-only setup works best on standard data. Cascaded systems lag behind text-only models for German but perform relatively well on dialectal data if transcription generates normalized, standard-like output.

Conclusion: Speech models are most effective for dialect processing, highlighting importance of multimodal approaches for dialect understanding. Cascaded systems can work well with proper transcription normalization.

Abstract: Research on cross-dialectal transfer from a standard to a non-standard dialect variety has typically focused on text data. However, dialects are primarily spoken, and non-standard spellings cause issues in text processing. We compare standard-to-dialect transfer in three settings: text models, speech models, and cascaded systems where speech first gets automatically transcribed and then further processed by a text model. We focus on German dialects in the context of written and spoken intent classification – releasing the first dialectal audio intent classification dataset – with supporting experiments on topic classification. The speech-only setup provides the best results on the dialect data while the text-only setup works best on the standard data. While the cascaded systems lag behind the text-only models for German, they perform relatively well on the dialectal data if the transcription system generates normalized, standard-like output.

Method: Created WUGNECTIVES dataset with 8,880 stimuli evaluating LMs’ inferences about novel entities in contexts where connectives link entities to attributes. Investigated 17 different LMs at various scales and training regimens, including tuning LMs to show reasoning behavior.

Result: Tuning LMs for reasoning behavior yields improvements on most connectives, but large variation exists across connective types. All models systematically struggle with concessive connectives. Overall performance varies significantly by connective type.

Conclusion: Discourse connectives can inform language models about world knowledge, but effectiveness varies by connective type. Findings enable more nuanced investigation of language cues in LMs and highlight challenges with concessive connectives.

Abstract: The role of world knowledge has been particularly crucial to predict the discourse connective that marks the discourse relation between two arguments, with language models (LMs) being generally successful at this task. We flip this premise in our work, and instead study the inverse problem of understanding whether discourse connectives can inform LMs about the world. To this end, we present WUGNECTIVES, a dataset of 8,880 stimuli that evaluates LMs’ inferences about novel entities in contexts where connectives link the entities to particular attributes. On investigating 17 different LMs at various scales, and training regimens, we found that tuning an LM to show reasoning behavior yields noteworthy improvements on most connectives. At the same time, there was a large variation in LMs’ overall performance across connective type, with all models systematically struggling on connectives that express a concessive meaning. Our findings pave the way for more nuanced investigations into the functional role of language cues as captured by LMs. We release WUGNECTIVES at https://github.com/kanishkamisra/wugnectives

Method: Confidence-Guided Reasoning Path Preference Optimization (CGPO) leverages confidence signals to identify points of maximal uncertainty in model reasoning, then applies self-generated, non-human-like reasoning-path guidance to mitigate trajectory drift without human annotation dependency.

Result: Experiments across diverse models on code and mathematical reasoning tasks show CGPO with small-model-generated data achieves better performance than approaches using strong-model or human-annotated data with same training data amount.

Conclusion: CGPO effectively improves LLM reasoning by targeting uncertainty points with self-generated guidance, overcoming human-like reasoning bias and annotation dependency while achieving superior performance.

Abstract: Current approaches for strengthening LLM reasoning tend to introduce a training bias toward human-like reasoning trajectories. In step-wise preference optimization, in particular, dependence on human or higher-capacity model annotations for intermediate steps limits exploration of alternative, non-human-like reasoning paths and thus constrains achievable performance. Furthermore, through a small-scale pilot study, we observed that in approximately 75% of cases, the model’s first erroneous step occurs after the lowest-confidence point. This suggests that guiding the model at its lowest-confidence point before an error provides more accurate supervision than locating the first explicit error. In this paper, we propose Confidence-Guided Reasoning Path Preference Optimization (CGPO), a method that leverages a confidence signal to identify points of maximal uncertainty in the model’s reasoning process and applies self-generated, non-human-like reasoning-path guidance to mitigate trajectory drift. Our experiments span diverse models applied to both code and mathematical reasoning tasks. The results show that, with the same amount of training data, our method using data generated by a small model can achieve better performance in most cases compared with approaches using data generated by a strong model or human-annotated.

Method: SAGE integrates top-down business logic (ideal customer profiles) and bottom-up business infrastructure knowledge (product catalogs, FAQs, knowledge bases) to simulate realistic user interactions grounded in specific business contexts.

Result: SAGE produces more realistic and diverse interactions while identifying up to 33% more agent errors compared to generic approaches, demonstrating effectiveness for bug-finding and iterative agent improvement.

Conclusion: Domain-specific user simulation incorporating business knowledge significantly improves multi-turn agent evaluation by generating more realistic interactions and better identifying agent weaknesses.

Abstract: Evaluating multi-turn interactive agents is challenging due to the need for human assessment. Evaluation with simulated users has been introduced as an alternative, however existing approaches typically model generic users and overlook the domain-specific principles required to capture realistic behavior. We propose SAGE, a novel user Simulation framework for multi-turn AGent Evaluation that integrates knowledge from business contexts. SAGE incorporates top-down knowledge rooted in business logic, such as ideal customer profiles, grounding user behavior in realistic customer personas. We further integrate bottom-up knowledge taken from business agent infrastructure (e.g., product catalogs, FAQs, and knowledge bases), allowing the simulator to generate interactions that reflect users’ information needs and expectations in a company’s target market. Through empirical evaluation, we find that this approach produces interactions that are more realistic and diverse, while also identifying up to 33% more agent errors, highlighting its effectiveness as an evaluation tool to support bug-finding and iterative agent improvement.

Method: Proposes HIES (Head Importance-Entropy Score) which integrates head importance scores with attention entropy to provide complementary evidence on per-head contribution for more effective pruning decisions.

Result: HIES-based pruning yields up to 15.2% improvement in model quality and 2.04x improvement in stability over HIS-only methods, enabling substantial model compression without sacrificing accuracy or stability.

Conclusion: The HIES criterion provides a more comprehensive approach to transformer pruning by considering both gradient importance and attention pattern diversity, leading to better model compression with maintained performance.

Abstract: Transformer-based models have achieved remarkable performance in NLP tasks. However, their structural characteristics-multiple layers and attention heads-introduce efficiency challenges in inference and deployment. To address these challenges, various pruning methods have recently been proposed. Notably, gradient-based methods using Head Importance Scores (HIS) have gained traction for interpretability, efficiency, and ability to identify redundant heads. However, HIS alone has limitations as it captures only the gradient-driven contribution, overlooking the diversity of attention patterns. To overcome these limitations, we introduce a novel pruning criterion, HIES (Head Importance-Entropy Score), which integrates head importance scores with attention entropy, providing complementary evidence on per-head contribution. Empirically, HIES-based pruning yields up to 15.2% improvement in model quality and 2.04x improvement in stability over HIS-only methods, enabling substantial model compression without sacrificing either accuracy or stability. Code will be released upon publication.

Method: Given SQL queries, BenchPress uses retrieval-augmented generation (RAG) and LLMs to propose multiple natural language descriptions. Human experts then select, rank, or edit these drafts for accuracy and domain alignment.

Result: Evaluation on annotated enterprise SQL logs shows LLM-assisted annotation drastically reduces time and effort while maintaining high quality. Human verification combined with LLM suggestions enhances annotation accuracy and benchmark reliability.

Conclusion: BenchPress streamlines creation of custom text-to-SQL benchmarks for domain-specific workloads, providing researchers/practitioners with better evaluation tools for enterprise applications.

Abstract: Large language models (LLMs) have been successfully applied to many tasks, including text-to-SQL generation. However, much of this work has focused on publicly available datasets, such as Fiben, Spider, and Bird. Our earlier work showed that LLMs are much less effective in querying large private enterprise data warehouses and released Beaver, the first private enterprise text-to-SQL benchmark. To create Beaver, we leveraged SQL logs, which are often readily available. However, manually annotating these logs to identify which natural language questions they answer is a daunting task. Asking database administrators, who are highly trained experts, to take on additional work to construct and validate corresponding natural language utterances is not only challenging but also quite costly. To address this challenge, we introduce BenchPress, a human-in-the-loop system designed to accelerate the creation of domain-specific text-to-SQL benchmarks. Given a SQL query, BenchPress uses retrieval-augmented generation (RAG) and LLMs to propose multiple natural language descriptions. Human experts then select, rank, or edit these drafts to ensure accuracy and domain alignment. We evaluated BenchPress on annotated enterprise SQL logs, demonstrating that LLM-assisted annotation drastically reduces the time and effort required to create high-quality benchmarks. Our results show that combining human verification with LLM-generated suggestions enhances annotation accuracy, benchmark reliability, and model evaluation robustness. By streamlining the creation of custom benchmarks, BenchPress offers researchers and practitioners a mechanism for assessing text-to-SQL models on a given domain-specific workload. BenchPress is freely available via our public GitHub repository at https://github.com/fabian-wenz/enterprise-txt2sql and is also accessible on our website at http://dsg-mcgraw.csail.mit.edu:5000.

Method: Conducted controlled pretraining experiments analyzing midtraining benefits across domains, focusing on code and math as distant domains from general pretraining data. Investigated starting time and mixture weight interactions using code as case study.

Result: Midtraining benefits are largest for domains distant from general pretraining data (code, math), scales with proximity advantage to target distribution, outperforms continued pretraining on specialized data alone, and shows strong interaction between introduction time and mixture weight.

Conclusion: Midtraining serves as distributional bridging for better posttraining initialization, with early introduction of specialized data amenable to high mixture weights while late introduction requires lower ones, suggesting distributional transitions between training phases benefit from similar bridging strategies.

Abstract: Midtraining, the practice of mixing specialized data with more general pretraining data in an intermediate training phase, has become widespread in language model development, yet there is little understanding of what makes it effective. We propose that midtraining functions as distributional bridging by providing better initialization for posttraining. We conduct controlled pretraining experiments, and find that midtraining benefits are largest for domains distant from general pretraining data, such as code and math, and scale with the proximity advantage the midtraining data provides toward the target distribution. In these domains, midtraining consistently outperforms continued pretraining on specialized data alone both in-domain and in terms of mitigating forgetting. We further conduct an investigation on the starting time and mixture weight of midtraining data, using code as a case study, and find that time of introduction and mixture weight interact strongly such that early introduction of specialized data is amenable to high mixture weights, while late introduction requires lower ones. This suggests that late introduction of specialized data outside a plasticity window cannot be compensated for by increasing data mixtures later in training. Beyond midtraining itself, this suggests that distributional transitions between any training phases may benefit from similar bridging strategies.

Method: Propose “quitting” as a behavioral mechanism for LLM agents to withdraw when lacking confidence. Use ToolEmu framework to systematically evaluate quitting behavior across 12 state-of-the-art LLMs with explicit quit instructions.

Result: Agents with explicit quit instructions improve safety by +0.39 average on 0-3 scale (+0.64 for proprietary models) while maintaining negligible average decrease of -0.03 in helpfulness. Shows favorable safety-helpfulness trade-off.

Conclusion: Explicit quit instructions are a highly effective, immediately deployable safety mechanism for LLM agents, establishing quitting as an effective first-line defense for autonomous agents in high-stakes applications.

Abstract: As Large Language Model (LLM) agents increasingly operate in complex environments with real-world consequences, their safety becomes critical. While uncertainty quantification is well-studied for single-turn tasks, multi-turn agentic scenarios with real-world tool access present unique challenges where uncertainties and ambiguities compound, leading to severe or catastrophic risks beyond traditional text generation failures. We propose using “quitting” as a simple yet effective behavioral mechanism for LLM agents to recognize and withdraw from situations where they lack confidence. Leveraging the ToolEmu framework, we conduct a systematic evaluation of quitting behavior across 12 state-of-the-art LLMs. Our results demonstrate a highly favorable safety-helpfulness trade-off: agents prompted to quit with explicit instructions improve safety by an average of +0.39 on a 0-3 scale across all models (+0.64 for proprietary models), while maintaining a negligible average decrease of -0.03 in helpfulness. Our analysis demonstrates that simply adding explicit quit instructions proves to be a highly effective safety mechanism that can immediately be deployed in existing agent systems, and establishes quitting as an effective first-line defense mechanism for autonomous agents in high-stakes applications.

Method: Used controlled cognitive science experiments to test LLMs’ ability to discount information from biased sources, then extended evaluation to real-world sponsored online advertisements. Implemented steering interventions to boost salience of intentions and incentives.

Result: LLMs successfully discount biased information in controlled experiments, behaving similarly to rational models. However, in real-world online ad settings, their inferences don’t closely track rational predictions due to distracting information. Simple steering interventions significantly improve correspondence with rational models.

Conclusion: LLMs possess basic sensitivity to others’ motivations but require further improvements to generalize effectively to novel real-world settings where motivational vigilance is crucial.

Abstract: Human communication is motivated: people speak, write, and create content with a particular communicative intent in mind. As a result, information that large language models (LLMs) and AI agents process is inherently framed by humans’ intentions and incentives. People are adept at navigating such nuanced information: we routinely identify benevolent or self-serving motives in order to decide what statements to trust. For LLMs to be effective in the real world, they too must critically evaluate content by factoring in the motivations of the source – for instance, weighing the credibility of claims made in a sales pitch. In this paper, we undertake a comprehensive study of whether LLMs have this capacity for motivational vigilance. We first employ controlled experiments from cognitive science to verify that LLMs’ behavior is consistent with rational models of learning from motivated testimony, and find they successfully discount information from biased sources in a human-like manner. We then extend our evaluation to sponsored online adverts, a more naturalistic reflection of LLM agents’ information ecosystems. In these settings, we find that LLMs’ inferences do not track the rational models’ predictions nearly as closely – partly due to additional information that distracts them from vigilance-relevant considerations. However, a simple steering intervention that boosts the salience of intentions and incentives substantially increases the correspondence between LLMs and the rational model. These results suggest that LLMs possess a basic sensitivity to the motivations of others, but generalizing to novel real-world settings will require further improvements to these models.

Method: Introduces Stream, a compilable hierarchical pruning algorithm that estimates per-head sparse attention masks in near-linear time O(T log T) and linear space O(T). It performs binary-search-style refinement to retain only top-k key blocks per query while preserving the model’s next-token behavior.

Result: On long chain-of-thought reasoning traces, Stream identifies thought anchors while pruning 97-99% of token interactions. On the RULER benchmark, it preserves critical retrieval paths while discarding 90-96% of interactions and exposes layer-wise routes from the needle to output.

Conclusion: Sparse Tracing provides a practical drop-in tool for analyzing attention patterns and tracing information flow without requiring terabytes of caches, making long-context interpretability feasible on consumer GPUs and democratizing chain-of-thought monitoring.

Abstract: As Large Language Models (LLMs) scale to million-token contexts, traditional Mechanistic Interpretability techniques for analyzing attention scale quadratically with context length, demanding terabytes of memory beyond 100,000 tokens. We introduce Sparse Tracing, a novel technique that leverages dynamic sparse attention to efficiently analyze long context attention patterns. We present Stream, a compilable hierarchical pruning algorithm that estimates per-head sparse attention masks in near-linear time $O(T \log T)$ and linear space $O(T)$, enabling one-pass interpretability at scale. Stream performs a binary-search-style refinement to retain only the top-$k$ key blocks per query while preserving the model’s next-token behavior. We apply Stream to long chain-of-thought reasoning traces and identify thought anchors while pruning 97-99% of token interactions. On the RULER benchmark, Stream preserves critical retrieval paths while discarding 90-96% of interactions and exposes layer-wise routes from the needle to output. Our method offers a practical drop-in tool for analyzing attention patterns and tracing information flow without terabytes of caches. By making long context interpretability feasible on consumer GPUs, Sparse Tracing helps democratize chain-of-thought monitoring. Code is available at https://anonymous.4open.science/r/stream-03B8/.

Method: FreeChunker transforms the traditional chunking paradigm by treating sentences as atomic units and shifting from static chunk segmentation to flexible retrieval supporting arbitrary sentence combinations, avoiding semantic boundary detection overhead.

Result: Experimental evaluation on LongBench V2 shows FreeChunker has significant advantages in both retrieval performance and time efficiency compared to existing chunking methods.

Conclusion: FreeChunker’s paradigm shift from static chunking to flexible sentence combination retrieval offers improved adaptability and efficiency for RAG systems.

Abstract: Chunking strategies significantly impact the effectiveness of Retrieval-Augmented Generation (RAG) systems. Existing methods operate within fixed-granularity paradigms that rely on static boundary identification, limiting their adaptability to diverse query requirements. This paper presents FreeChunker, a Cross-Granularity Encoding Framework that fundamentally transforms the traditional chunking paradigm: the framework treats sentences as atomic units and shifts from static chunk segmentation to flexible retrieval supporting arbitrary sentence combinations. This paradigm shift not only significantly avoids the computational overhead required for semantic boundary detection, but also enhances adaptability to complex queries. Experimental evaluation on LongBench V2 demonstrates that FreeChunker possesses significant advantages in both retrieval performance and time efficiency compared to existing chunking methods. The pre-trained models and codes are available at https://github.com/mazehart/FreeChunker.

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 (including GPT-5) show significant gaps compared to human experts with MAE of 0.553 and correlation of only 0.428. VisJudge reduces MAE to 0.421 (23.9% reduction) and increases consistency to 0.687 (60.5% improvement).

Conclusion: Current MLLMs are inadequate for visualization quality assessment, but specialized models like VisJudge can significantly narrow the gap with human judgment. The benchmark enables systematic evaluation of MLLMs’ capabilities in this 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 DEBATE benchmark with 36,383 messages from 2,832 U.S. participants across 708 groups and 107 topics. Evaluated 7 LLMs as “digital twin” role-playing agents using next-message prediction and full conversation rollout with stance-alignment and opinion-convergence metrics.

Result: Zero-shot LLM agents show strong opinion convergence relative to humans. Supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) improve stance alignment and bring group-level convergence closer to human behavior, though discrepancies in opinion change remain.

Conclusion: DEBATE enables rigorous benchmarking of simulated opinion dynamics and supports future research on aligning multi-agent LLM role-playing agents with realistic human interactions.

Abstract: Accurately modeling opinion change through social interactions is crucial for understanding and mitigating polarization, misinformation, and societal conflict. Recent work simulates opinion dynamics with role-playing LLM agents (RPLAs), but multi-agent simulations often display unnatural group behavior (e.g., premature convergence) and lack empirical benchmarks for assessing alignment with real human group interactions. We introduce DEBATE, a large-scale benchmark for evaluating the authenticity of opinion dynamics in multi-agent RPLA simulations. DEBATE contains 36,383 messages from 2,832 U.S.-based participants across 708 groups and 107 topics, with both public messages and private Likert-scale beliefs, enabling evaluation at the utterance and group levels (and supporting future individual-level analyses). We instantiate “digital twin” RPLAs with seven LLMs and evaluate across two settings: next-message prediction and full conversation rollout, using stance-alignment and opinion-convergence metrics. In zero-shot settings, RPLA groups exhibit strong opinion convergence relative to human groups. Post-training via supervised fine-tuning (SFT) and Direct Preference Optimization (DPO) improves stance alignment and brings group-level convergence closer to human behavior, though discrepancies in opinion change and belief updating remain. DEBATE enables rigorous benchmarking of simulated opinion dynamics and supports future research on aligning multi-agent RPLAs with realistic human interactions.

Method: The authors apply Minimum Description Length (MDL) approach to measure both regularity and processing complexity in recursive numeral systems. They compare their MDL-based measures against previous optimization frameworks to evaluate how well they distinguish natural systems from unnatural ones.

Result: The MDL-based measures of regularity and processing complexity better capture differences between attested natural systems and theoretically possible ones, including previously identified “optimal” systems. The ad-hoc constraints from previous work naturally emerge from considerations of regularity.

Conclusion: Regularity is essential for understanding linguistic efficiency and should be incorporated in studies measuring language efficiency. The MDL approach provides a principled way to account for systematicity across sets of forms in linguistic systems.

Abstract: Much recent work has shown how cross-linguistic variation is constrained by competing pressures from efficient communication. However, little attention has been paid to the role of the systematicity of forms (regularity), a key property of natural language. Here, we demonstrate the importance of regularity in explaining the shape of linguistic systems by looking at recursive numeral systems. Previous work has argued that these systems optimise the trade-off between lexicon size and average morphosyntatic complexity (Denić and Szymanik, 2024). However, showing that only natural-language-like systems optimise this trade-off has proven elusive, and existing solutions rely on ad-hoc constraints to rule out unnatural systems (Yang and Regier, 2025). Drawing on the Minimum Description Length (MDL) approach, we argue that recursive numeral systems are better viewed as efficient with regard to their regularity and processing complexity. We show that our MDL-based measures of regularity and processing complexity better capture the key differences between attested, natural systems and theoretically possible ones, including “optimal” recursive numeral systems from previous work, and that the ad-hoc constraints naturally follow from regularity. Our approach highlights the need to incorporate regularity across sets of forms in studies attempting to measure efficiency in language.

Method: Created DebateBias-8K with 8,400 structured debate prompts across 4 sensitive domains (women’s rights, socioeconomic development, terrorism, religion) in 7 languages. Tested 4 flagship models (GPT-4o, Claude 3, DeepSeek, LLaMA 3), generating and automatically classifying over 100,000 responses.

Result: All models reproduce entrenched stereotypes: Arabs linked to terrorism/religion (≥95%), Africans to socioeconomic “backwardness” (up to 77%), Western groups framed as modern/progressive. Biases grow sharply in lower-resource languages, showing English alignment doesn’t generalize globally.

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

Abstract: Large language models (LLMs) are widely deployed for open-ended communication, yet most bias evaluations still rely on English, classification-style tasks. We introduce DebateBias-8K, a new multilingual, debate-style benchmark designed to reveal how narrative bias appears in realistic generative settings. Our dataset includes 8,400 structured debate prompts spanning four sensitive domains: women’s rights, socioeconomic development, terrorism, and religion, across seven languages ranging from high-resource (English, Chinese) to low-resource (Swahili, Nigerian Pidgin). Using four flagship models (GPT-4o, Claude 3, DeepSeek, and LLaMA 3), we generate and automatically classify over 100,000 responses. Results show that all models reproduce entrenched stereotypes despite safety alignment: Arabs are overwhelmingly linked to terrorism and religion (>=95%), Africans to socioeconomic “backwardness” (up to <=77%), and Western groups are consistently framed as modern or progressive. Biases grow sharply in lower-resource languages, revealing that alignment trained primarily in English does not generalize globally. Our findings highlight a persistent divide in multilingual fairness: current alignment methods reduce explicit toxicity but fail to prevent biased outputs in open-ended contexts. We release our DebateBias-8K benchmark and analysis framework to support the next generation of multilingual bias evaluation and safer, culturally inclusive model alignment.

Method: Uses Item Response Theory to measure question difficulty, employs 35 LLMs as simulated students to build a difficulty ranker, then uses reinforcement learning with the ranker as reward signal to guide a question-rewriting model to create challenging variations.

Result: Applied to competition-level math benchmarks, RIDE generated perturbed versions that degraded advanced LLM performance by average 21.73% across 26 models, exposing limited robustness in mathematical reasoning.

Conclusion: RIDE provides a valid adversarial evaluation approach that reveals LLMs’ limited robustness in mathematical reasoning beyond superficial pattern matching.

Abstract: Large language models (LLMs) achieve high performance on mathematical reasoning, but these results can be inflated by training data leakage or superficial pattern matching rather than genuine reasoning. To this end, an adversarial perturbation-based evaluation is needed to measure true mathematical reasoning ability. Current rule-based perturbation methods often generate ill-posed questions and impede the systematic evaluation of question difficulty and the evolution of benchmarks. To bridge this gap, we propose RIDE, a novel adversarial question-rewriting framework that leverages Item Response Theory (IRT) to rigorously measure question difficulty and to generate intrinsically more challenging, well-posed variations of mathematical problems. We employ 35 LLMs to simulate students and build a difficulty ranker from their responses. This ranker provides a reward signal during reinforcement learning and guides a question-rewriting model to reformulate existing questions across difficulty levels. Applying RIDE to competition-level mathematical benchmarks yields perturbed versions that degrade advanced LLM performance, with experiments showing an average 21.73% drop across 26 models, thereby exposing limited robustness in mathematical reasoning and confirming the validity of our evaluation approach.

Method: ATLAS uses Item Response Theory (IRT) with Fisher information-guided item selection to adaptively choose the most informative test items for estimating model ability, reducing required items by up to 90%.

Result: ATLAS reduces items needed by up to 90% while maintaining precision (e.g., 41 items vs 5,600 on HellaSwag with 0.157 MAE). Reconstructed accuracies match raw accuracies, and ability estimates provide finer discrimination among models with identical accuracies.

Conclusion: ATLAS provides an efficient, precise alternative to traditional LLM evaluation, enabling scalable assessment while revealing meaningful performance differences that accuracy metrics miss.

Abstract: Evaluating large language models (LLMs) typically requires thousands of benchmark items, making the process expensive, slow, and increasingly impractical at scale. Existing evaluation protocols rely on average accuracy over fixed item sets, treating all items as equally informative despite substantial variation in difficulty and discrimination. We introduce ATLAS, an adaptive testing framework based on Item Response Theory (IRT) that estimates model ability using Fisher information-guided item selection. ATLAS reduces the number of required items by up to 90% while maintaining measurement precision. For instance, it matches whole-bank ability estimates using only 41 items (0.157 MAE) on HellaSwag (5,600 items). We further reconstruct accuracy from ATLAS’s ability estimates and find that reconstructed accuracies closely match raw accuracies across all five benchmarks, indicating that ability $θ$ preserves the global performance structure. At the same time, $θ$ provides finer discrimination within accuracy-equivalent models: among more than 3,000 evaluated models, 23-31% shift by more than 10 rank positions, and models with identical accuracies receive meaningfully different ability estimates. Code and calibrated item banks are available at https://github.com/Peiyu-Georgia-Li/ATLAS.git.

Method: TabRAG uses layout segmentation to decompose documents into components, then employs a vision language model to parse tables into hierarchical structured representations, enhanced by a self-generated in-context learning module to handle various table styles and formats.

Result: TabRAG outperforms existing popular parsing techniques across a broad suite of evaluation and ablation benchmarks for tabular document question answering.

Conclusion: The framework effectively preserves tabular structural semantics and enables more accurate question answering on tabular documents through structured representations and vision-language integration.

Abstract: Incorporating external knowledge bases in traditional retrieval-augmented generation (RAG) relies on parsing the document, followed by querying a language model with the parsed information via in-context learning. While effective for text-based documents, question answering on tabular documents often fails to generate plausible responses. Standard parsing techniques lose the two-dimensional structural semantics critical for cell interpretation. In this work, we present TabRAG, a parsing-based RAG framework designed to improve tabular document question answering via structured representations. Our framework consists of layout segmentation that decomposes the document inputs into a series of components, enabling fine-grained extraction. Subsequently, a vision language model parses and extracts the document tables into a hierarchically structured representation. In order to cater various table styles and formats, we integrate a self-generated in-context learning module that guides the table extraction process. Experimental results demonstrate that TabRAG outperforms existing popular parsing techniques across a broad suite of evaluation and ablation benchmarks. Code is available at: https://github.com/jacobyhsi/TabRAG.

Method: The survey organizes LLM-driven scientific ideation methods into five families: 1) External knowledge augmentation, 2) Prompt-based distributional steering, 3) Inference-time scaling, 4) Multi-agent collaboration, and 5) Parameter-level adaptation. It analyzes these methods using two creativity frameworks: Boden taxonomy (for novelty levels) and Rhodes 4Ps framework (for creativity aspects).

Result: The survey provides a structured synthesis of existing methods, clarifies the evaluation landscape for scientific ideation, and identifies key challenges in achieving reliable and systematic LLM-based scientific discovery through systematic analysis of how different approaches trade off novelty and scientific validity.

Conclusion: The paper establishes a framework for understanding LLM-based scientific creativity, highlighting that different methodological families emphasize different aspects of creativity and trade-offs between novelty and validity, providing guidance for future research in automated scientific discovery.

Abstract: Scientific idea generation is central to discovery, requiring the joint satisfaction of novelty and scientific soundness. Unlike standard reasoning or general creative generation, scientific ideation is inherently open-ended and multi-objective, making its automation particularly challenging. Recent advances in large language models (LLMs) have enabled the generation of coherent and plausible scientific ideas, yet the nature and limits of their creative capabilities remain poorly understood. This survey provides a structured synthesis of methods for LLM-driven scientific ideation, focusing on how different approaches trade off novelty and scientific validity. We organize existing methods into five complementary families: External knowledge augmentation, Prompt-based distributional steering, Inference-time scaling, Multi-agent collaboration, and Parameter-level adaptation. To interpret their contributions, we adopt two complementary creativity frameworks: Boden taxonomy to characterize the expected level of creative novelty, and Rhodes 4Ps framework to analyze the aspects or sources of creativity emphasized by each method. By aligning methodological developments with cognitive creativity frameworks, this survey clarifies the evaluation landscape and identifies key challenges and directions for reliable and systematic LLM-based scientific discovery.

Method: Formalizes latent reasoning as classifier-free guidance and policy improvement algorithms. Proposes training schemes from reinforcement learning and diffusion methods for latent reasoning models. Tests modifications on Tiny Recursive Model to optimize recursive steps.

Result: With proposed modifications, achieves 18x reduction in total forward passes while maintaining performance, effectively avoiding dead compute steps. Shows policy improvement perspective explains model behavior and enables optimization.

Conclusion: Latent reasoning can be understood through policy improvement framework, enabling optimization to reduce computational waste. This perspective provides insights for improving recursive models and explains when recursive steps effectively contribute to reasoning.

Abstract: Recently, small models with latent recursion have obtained promising results on complex reasoning tasks. These results are typically explained by the theory that such recursion increases a networks depth, allowing it to compactly emulate the capacity of larger models. However, the performance of recursively added layers remains behind the capabilities of one pass models with the same feed forward depth. This means that in the looped version, not every recursive step effectively contributes to depth. This raises the question: when and why does latent reasoning improve performance, and when does it result in dead compute? In our work, we analyze the algorithms that latent reasoning provides answer to this question. We show that latent reasoning can be formalized as a classifier free guidance and policy improvement algorithm. Building on these insights, we propose to use a training schemes from reinforcement learning and diffusion methods for latent reasoning models. Using the Tiny Recursive Model as our testbed, we show that with our modifications we can avoid dead compute steps and reduce the total number of forward passes by 18x while maintaining performance. Broadly speaking, we show how a policy improvement perspective on recursive steps can explain model behavior and provide insights for further improvements.

Method: PIRA integrates three strategies: 1) Reformulating question-answer pairs into preference-task instructions to leverage LLMs’ preference instruction-following capability, 2) Averaging rewards from diverse preference-task instructions per sample to mitigate task-specific bias and enhance robustness, and 3) Averaging outputs from the value head under different dropout rates to stabilize reward estimation.

Result: Experiments on public datasets show that PIRA improves performance considerably, enhances generalization, and effectively mitigates reward overoptimization compared to existing approaches.

Conclusion: PIRA provides an effective training paradigm for reward models that addresses key limitations of existing approaches by better leveraging LLMs’ capabilities while reducing vulnerability to reward overoptimization.

Abstract: Reward models are pivotal for aligning Large Language Models (LLMs) with human preferences. Existing approaches face two key limitations: Discriminative reward models require large-scale annotated data, as they cannot exploit the preference instruction-following capability of LLMs available to generative reward models. Moreover, reward models are particularly prone to reward overoptimization, where LLMs exploit weaknesses in the reward function instead of improving true alignment. We introduce \textbf{PIRA}, a training paradigm that integrates three complementary strategies to address these challenges: (1) reformulating question-answer pairs into preference-task instructions to explicitly leverage LLMs’ preference instruction-following capability, (2) averaging the rewards aggregated from diverse preference-task instructions for each sample, which mitigates task-specific bias and enhances robustness across evaluation perspectives, and (3) averaging outputs from the value head under different dropout rates to stabilize reward estimation. Experiments on public datasets show that PIRA improves performance considerably, enhances generalization, and effectively mitigates reward overoptimization.

Method: Introduces CrossCheck-Bench with 15k QA pairs from real-world artifacts with synthetically injected contradictions. Uses hierarchical task framework covering three reasoning complexity levels and defines seven atomic capabilities for cross-modal inconsistency resolution. Constructed through multi-stage annotation pipeline with 450+ expert hours.

Result: Evaluation of 13 state-of-the-art VLMs shows consistent performance drop as tasks shift from perceptual matching to logical contradiction detection. Models perform well on isolated entity recognition but fail at synthesizing multiple clues for conflict reasoning. Conventional prompting strategies yield marginal gains, while symbolic reasoning with grounded visual processing shows more stable improvements.

Conclusion: Highlights persistent bottleneck in multimodal reasoning and suggests new directions for building models capable of robust cross-modal verification, emphasizing the need for structured reasoning beyond surface-level alignment.

Abstract: Multimodal Large Language Models are primarily trained and evaluated on aligned image-text pairs, which leaves their ability to detect and resolve real-world inconsistencies largely unexplored. In open-domain applications visual and textual cues often conflict, requiring models to perform structured reasoning beyond surface-level alignment. We introduce CrossCheck-Bench, a diagnostic benchmark for evaluating contradiction detection in multimodal inputs. The benchmark adopts a hierarchical task framework covering three levels of reasoning complexity and defines seven atomic capabilities essential for resolving cross-modal inconsistencies. CrossCheck-Bench includes 15k question-answer pairs sourced from real-world artifacts with synthetically injected contradictions. The dataset is constructed through a multi-stage annotation pipeline involving more than 450 expert hours to ensure semantic validity and calibrated difficulty across perception, integration, and reasoning. We evaluate 13 state-of-the-art vision-language models and observe a consistent performance drop as tasks shift from perceptual matching to logical contradiction detection. Most models perform well on isolated entity recognition but fail when multiple clues must be synthesized for conflict reasoning. Capability-level analysis further reveals uneven skill acquisition, especially in tasks requiring multi-step inference or rule-based validation. Additional probing shows that conventional prompting strategies such as Chain-of-Thought and Set-of-Mark yield only marginal gains. By contrast, methods that interleave symbolic reasoning with grounded visual processing achieve more stable improvements. These results highlight a persistent bottleneck in multimodal reasoning and suggest new directions for building models capable of robust cross-modal verification.

Method: Introduces Reward Auditor, a hypothesis-testing framework for RM suitability inference. It uses scientific auditing to quantify statistical significance and effect size by analyzing distribution degradation of RM preference perception confidence under real-world perturbed scenarios.

Result: The framework enables inference of both certainty and severity of RM vulnerabilities across diverse real-world scenarios, providing a more comprehensive evaluation approach.

Conclusion: Reward Auditor lays foundation for building next-generation LLM alignment systems that are verifiably safe, robust, and trustworthy by addressing the suitability dimension of reward model evaluation.

Abstract: Reliable reward models (RMs) are critical for ensuring the safe alignment of large language models (LLMs). However, current RM evaluation methods focus solely on preference perception accuracies in given specific scenarios, obscuring the critical vulnerabilities of RMs in real-world scenarios. We identify the true challenge lies in assessing a novel dimension: Suitability, defined as conditional reliability under specific real-world perturbations. To this end, we introduce Reward Auditor, a hypothesis-testing framework specifically designed for RM suitability inference. Rather than answering “How accurate is the RM’s preference perception for given samples?”, it employs scientific auditing to answer: “Can we infer RMs exhibit systematic vulnerabilities in specific real-world scenarios?”. Under real-world perturbed scenarios, Reward Auditor quantifies statistical significance and effect size by auditing distribution degradation of RM preference perception confidence. This enables inference of both the certainty and severity of RM vulnerabilities across diverse real-world scenarios. This lays a solid foundation for building next-generation LLM alignment systems that are verifiably safe, more robust, and trustworthy.

Method: 1) Construct UnsolvableQA dataset using “Reverse Construction” that systematically injects logical contradictions into valid reasoning chains. 2) Develop UnsolvableRL, a reinforcement learning paradigm that balances objective unsolvability detection with calibrated confidence under capability limits.

Result: Achieves robust unsolvability detection (>85% detection rate) and boosts solvable reasoning accuracy from 43.4% to 69.4% on Qwen3-4B-Instruct. Identifies data-training interaction: strict alignment constraints cause Capability Collapse without unsolvable data, but act as regularizers for rigor when such data is included.

Conclusion: The proposed approach effectively helps LLMs distinguish between objective unsolvability and capability limitations, improving overall robustness and reducing hallucinations. The interaction between alignment constraints and unsolvable data is crucial for achieving this balance.

Abstract: Ensuring large language model (LLM) reliability requires distinguishing objective unsolvability (inherent contradictions) from subjective capability limitations (tasks exceeding model competence). Current LLMs often conflate these dimensions, leading to hallucinations in which they return confident answers to inherently unsolvable queries. To address this issue, we propose a multi-domain dataset containing both solvable and unsolvable questions, UnsolvableQA, together with an alignment framework, UnsolvableRL. First, we construct UnsolvableQA by “Reverse Construction” that systematically injects logical contradictions into otherwise valid reasoning chains. Second, we introduce UnsolvableRL, a reinforcement learning paradigm that balances objective unsolvability detection with calibrated confidence under capability limits. Empirically, our approach achieves robust unsolvability detection (>85% detection rate) and boosts solvable reasoning accuracy from 43.4% to 69.4% on Qwen3-4B-Instruct. Crucially, we identify a data-training interaction: strict alignment constraints induce Capability Collapse without unsolvable data, but act as a regularizer for rigor when such data are included, thereby improving overall robustness. Our code and data are available at https://github.com/sfasfaffa/unsolvableQA .

Method: Introduces latent debate framework that captures implicit internal arguments within a single model during single inference, unlike explicit multi-agent debates. Presents model- and task-agnostic conceptual framework, then instantiates it symbolically for True/False prediction tasks.

Result: Latent debate serves as a faithful structured surrogate model with highly consistent predictions with original LLM. Provides strong baseline for hallucination detection, with analysis revealing correlations between hallucinations and debate patterns (e.g., high degree of latent debates in middle layers linked to higher hallucination risk).

Conclusion: Latent debate positions as a potential framework for understanding internal mechanisms of LLMs, especially for scenarios where internal (dis)agreements appear during inference steps, offering both interpretability and practical applications for hallucination detection.

Abstract: Understanding the internal thinking process of Large Language Models (LLMs) and the cause of hallucinations remains a key challenge. To this end, we introduce latent debate, a novel framework for interpreting model predictions through the lens of implicit internal arguments. Unlike the current work of self-consistency and multi-agent debate, which relies on explicit debates among multiple answers or multiple models, latent debate captures the hidden supporting and attacking signals that arise within a single model during a single inference. We first present a model- and task-agnostic conceptual framework, and then instantiate it symbolically to approximate the thinking process of LLMs on True/False prediction tasks. Empirical studies demonstrate that latent debate is a faithful structured surrogate model that has highly consistent predictions with the original LLM. Beyond interpretability, we demonstrate that latent debate provides a strong baseline for hallucination detection. Further analysis reveals strong correlations between hallucinations and debate patterns, such as a high degree of latent debates in the middle layers is linked to a higher risk of hallucinations. These findings position latent debate as a potential framework for understanding internal mechanisms of LLMs, especially for scenarios where internal (dis)agreements appear during the inference steps.

Method: Identifies Lazy Likelihood Displacement (LLD) as the core failure mechanism, characterizes its three-phase trajectory, and proposes LLDS - a likelihood-preserving regularization that activates only when response likelihood decreases and regularizes only responsible tokens.

Result: Method stabilizes training, prevents gradient explosion, and yields substantial performance improvements across seven benchmarks, including +45.2% on Qwen2.5-3B and +37.1% on Qwen2.5-7B over vanilla GRPO training.

Conclusion: LLD is a previously overlooked bottleneck in GRPO-based tool-integrated RL, and the proposed likelihood-preserving regularization provides a practical path toward stable, scalable training of tool-integrated RL systems.

Abstract: Tool-integrated (TI) reinforcement learning (RL) enables large language models (LLMs) to perform multi-step reasoning by interacting with external tools such as search engines and retrievers. Group Relative Policy Optimization (GRPO), exemplified by the recent Search-R1, offers fast convergence and a value-free formulation that makes it appealing for this setting, yet consistently suffers from training collapse. We identify Lazy Likelihood Displacement (LLD), a systematic reduction or stagnation in the likelihood of both correct and incorrect responses, as the core mechanism driving this failure. LLD emerges early and triggers a self-reinforcing LLD Death Spiral, where declining likelihood leads to low-confidence responses, inflating gradients, and ultimately causing collapse. We empirically characterize this process across models on a Search-R1-style, search-integrated question answering task, revealing a consistent three-phase trajectory: early stagnation, steady decay, and accelerated collapse. To address this, we propose a likelihood-preserving regularization LLDS that activates only when a response action’s likelihood decreases, and regularizes only the tokens responsible. This fine-grained structure mitigates LLD with minimal interference. Our method stabilizes training, prevents gradient explosion, and yields substantial performance improvements across seven benchmarks, including relative improvements of +45.2% on Qwen2.5-3B and +37.1% on Qwen2.5-7B over vanilla GRPO training. Our results establish LLD as a previously overlooked bottleneck in GRPO-based TIRL and provide a practical path toward stable, scalable training of tool-integrated RL.

Method: Proposes a GNN framework where each user comment is a node, edges are defined by textual similarity between comments, and includes a dynamically adjusted attention mechanism to balance nodal (textual) and topological (structural) features during information aggregation.

Result: The proposed architecture outperforms 12 state-of-the-art LLMs across multiple metrics while requiring significantly lower inference cost, demonstrating the importance of structural context for detecting online incivility.

Conclusion: Structural context is crucial for detecting online incivility, addressing limitations of text-only LLM paradigms in behavioral prediction. The approach offers better performance with lower computational cost.

Abstract: Online incivility has emerged as a widespread and persistent problem in digital communities, imposing substantial social and psychological burdens on users. Although many platforms attempt to curb incivility through moderation and automated detection, the performance of existing approaches often remains limited in both accuracy and efficiency. To address this challenge, we propose a Graph Neural Network (GNN) framework for detecting three types of uncivil behavior (i.e., toxicity, aggression, and personal attacks) within the English Wikipedia community. Our model represents each user comment as a node, with textual similarity between comments defining the edges, allowing the network to jointly learn from both linguistic content and relational structures among comments. We also introduce a dynamically adjusted attention mechanism that adaptively balances nodal and topological features during information aggregation. Empirical evaluations demonstrate that our proposed architecture outperforms 12 state-of-the-art Large Language Models (LLMs) across multiple metrics while requiring significantly lower inference cost. These findings highlight the crucial role of structural context in detecting online incivility and address the limitations of text-only LLM paradigms in behavioral prediction. All datasets and comparative outputs will be publicly available in our repository to support further research and reproducibility.

Method: MEDAL integrates Monte Carlo Tree Search at the initialization stage of DLM inference to explore promising unmasking trajectories, providing a robust starting point for subsequent refinement. This enables efficient inference-time scaling where generation quality improves with increased search budget.

Result: Across multiple benchmarks, MEDAL achieves up to 22.0% improvement over existing inference strategies for Diffusion Language Models.

Conclusion: MEDAL establishes a new paradigm for search-based inference in Diffusion Language Models, demonstrating that inference-time scaling with principled search mechanisms can significantly improve generation quality without additional training.

Abstract: Diffusion language models (DLMs) have recently emerged as a compelling alternative to autoregressive generation, offering parallel generation and improved global coherence. During inference, DLMs generate text by iteratively denoising masked sequences in parallel; however, determining which positions to unmask and which tokens to commit forms a large combinatorial search problem. Existing inference methods approximate this search using heuristics, which often yield suboptimal decoding paths; other approaches instead rely on additional training to guide token selection. To introduce a principled search mechanism for DLMs inference, we introduce MEDAL, an inference-time scaling framework that integrates Monte Carlo Tree SEarch initialization for Diffusion LAnguage Model inference. We employ Monte Carlo Tree Search at the initialization stage to explore promising unmasking trajectories, providing a robust starting point for subsequent refinement. This design enables efficient inference-time scaling, allowing generation quality to improve as the search budget increases, without additional training. Across multiple benchmarks, MEDAL achieves up to 22.0% improvement over existing inference strategies, establishing a new paradigm for search-based inference in DLMs.

Method: Proposes VersatileFFN with two adaptive pathways: width-versatile path generates mixture of sub-experts from single shared FFN (mimicking sparse expert routing without parameter increase), and depth-versatile path recursively applies same FFN for deeper processing. Difficulty-aware gating dynamically balances pathways based on token complexity.

Result: Experiments across diverse benchmarks and model scales demonstrate effectiveness of the method. Both pathways reuse same parameters, so all additional capacity comes from computation rather than memory.

Conclusion: VersatileFFN provides flexible parameter reuse approach that enhances model capacity within fixed parameter budgets, addressing memory scaling issues while maintaining computational efficiency.

Abstract: The rapid scaling of Large Language Models (LLMs) has achieved remarkable performance, but it also leads to prohibitive memory costs. Existing parameter-efficient approaches such as pruning and quantization mainly compress pretrained models without enhancing architectural capacity, thereby hitting the representational ceiling of the base model. In this work, we propose VersatileFFN, a novel feed-forward network (FFN) that enables flexible reuse of parameters in both width and depth dimensions within a fixed parameter budget. Inspired by the dual-process theory of cognition, VersatileFFN comprises two adaptive pathways: a width-versatile path that generates a mixture of sub-experts from a single shared FFN, mimicking sparse expert routing without increasing parameters, and a depth-versatile path that recursively applies the same FFN to emulate deeper processing for complex tokens. A difficulty-aware gating dynamically balances the two pathways, steering “easy” tokens through the efficient width-wise route and allocating deeper iterative refinement to “hard” tokens. Crucially, both pathways reuse the same parameters, so all additional capacity comes from computation rather than memory. Experiments across diverse benchmarks and model scales demonstrate the effectiveness of the method. The code is available at https://github.com/huawei-noah/noah-research/tree/master/VersatileFFN.

Method: MemBuilder uses reinforcement learning with: 1) synthetic session-level question generation for dense intermediate rewards across extended trajectories, and 2) contribution-aware gradient weighting that scales policy updates based on each memory component’s downstream impact.

Result: A 4B-parameter model trained with MemBuilder outperforms state-of-the-art closed-source baselines and shows strong generalization across long-term dialogue benchmarks.

Conclusion: MemBuilder effectively addresses sparse reward and memory attribution challenges in long-term dialogue systems, enabling smaller open-source models to achieve superior performance through structured memory construction.

Abstract: Maintaining consistency in long-term dialogues remains a fundamental challenge for LLMs, as standard retrieval mechanisms often fail to capture the temporal evolution of historical states. While memory-augmented frameworks offer a structured alternative, current systems rely on static prompting of closed-source models or suffer from ineffective training paradigms with sparse rewards. We introduce MemBuilder, a reinforcement learning framework that trains models to orchestrate multi-dimensional memory construction with attributed dense rewards. MemBuilder addresses two key challenges: (1) Sparse Trajectory-Level Rewards: we employ synthetic session-level question generation to provide dense intermediate rewards across extended trajectories; and (2) Multi-Dimensional Memory Attribution: we introduce contribution-aware gradient weighting that scales policy updates based on each component’s downstream impact. Experimental results show that MemBuilder enables a 4B-parameter model to outperform state-of-the-art closed-source baselines, exhibiting strong generalization across long-term dialogue benchmarks.

Method: Created AutoMonitor-Bench with 3,010 annotated samples spanning QA, code generation, and reasoning tasks; evaluated 22 LLMs using Miss Rate and False Alarm Rate metrics; fine-tuned Qwen3-4B-Instruction on 153,581 training samples.

Result: Found substantial variability in monitoring performance across models, consistent trade-off between MR and FAR, and limited improvement from training on known misbehavior datasets for detecting unseen implicit misbehaviors.

Conclusion: Reliable, scalable misbehavior monitoring remains challenging, highlighting inherent safety-utility tension and motivating future work on task-aware monitor design and training strategies.

Abstract: We introduce AutoMonitor-Bench, the first benchmark designed to systematically evaluate the reliability of LLM-based misbehavior monitors across diverse tasks and failure modes. AutoMonitor-Bench consists of 3,010 carefully annotated test samples spanning question answering, code generation, and reasoning, with paired misbehavior and benign instances. We evaluate monitors using two complementary metrics: Miss Rate (MR) and False Alarm Rate (FAR), capturing failures to detect misbehavior and oversensitivity to benign behavior, respectively. Evaluating 12 proprietary and 10 open-source LLMs, we observe substantial variability in monitoring performance and a consistent trade-off between MR and FAR, revealing an inherent safety-utility tension. To further explore the limits of monitor reliability, we construct a large-scale training corpus of 153,581 samples and fine-tune Qwen3-4B-Instruction to investigate whether training on known, relatively easy-to-construct misbehavior datasets improves monitoring performance on unseen and more implicit misbehaviors. Our results highlight the challenges of reliable, scalable misbehavior monitoring and motivate future work on task-aware designing and training strategies for LLM-based monitors.

Method: Comprehensive evaluation across three major LLM providers (OpenAI, Anthropic, Google) comparing three caching strategies: full context caching, system prompt only caching, and caching excluding dynamic tool results. Evaluation on DeepResearch Bench with 500+ agent sessions and 10,000-token system prompts, measuring API cost and time to first token.

Result: Prompt caching reduces API costs by 41-80% and improves TTFT by 13-31% across providers. Strategic cache block control (placing dynamic content at end, avoiding dynamic function calling, excluding tool results) provides more consistent benefits than naive full-context caching, which can increase latency. Linear cost and TTFT benefits observed across prompt sizes (500-50k tokens) and tool call counts (3-50).

Conclusion: Prompt caching offers substantial cost and latency benefits for agentic systems, but requires strategic implementation with cache block control rather than naive approaches. Provider-specific strategy discrepancies exist, requiring nuanced guidance for production deployment.

Abstract: Recent advancements in Large Language Model (LLM) agents have enabled complex multi-turn agentic tasks requiring extensive tool calling, where conversations can span dozens of API calls with increasingly large context windows. However, although major LLM providers offer prompt caching to reduce cost and latency, its benefits for agentic workloads remain underexplored in the research literature. To our knowledge, no prior work quantifies these cost savings or compares caching strategies for multi-turn agentic tasks. We present a comprehensive evaluation of prompt caching across three major LLM providers (OpenAI, Anthropic, and Google) and compare three caching strategies, including full context caching, system prompt only caching, and caching that excludes dynamic tool results. We evaluate on DeepResearch Bench, a multi-turn agentic benchmark where agents autonomously execute real-world web search tool calls to answer complex research questions, measuring both API cost and time to first token (TTFT) across over 500 agent sessions with 10,000-token system prompts. Our results demonstrate that prompt caching reduces API costs by 41-80% and improves time to first token by 13-31% across providers. We find that strategic prompt cache block control, such as placing dynamic content at the end of the system prompt, avoiding dynamic traditional function calling, and excluding dynamic tool results, provides more consistent benefits than naive full-context caching, which can paradoxically increase latency. An ablation study across prompt sizes (500-50,000 tokens) and tool call counts (3-50) demonstrates universal linear cost and TTFT benefits, after the provider caching token minimum, and reveal provider-specific strategy discrepancies across variants. We provide nuanced discussion and guidance for implementing prompt caching in production agentic systems.

Method: Systematically investigates factors governing synthetic data quality, advances feature-based synthesis via domain-specific evolution and dual-verification strategy. Trains X-Coder model series under SFT-then-RL paradigm using high-quality synthetic data.

Result: X-Coder-7B shows significant performance gains on LiveCodeBench v5 (62.9% avg@8) and v6 (55.8% avg@8), outperforming larger models trained on real-world data. Provides insights into synthetic data scaling, domain-adapted feature evolution, and code-centric reinforcement.

Conclusion: Expert-level reasoning performance in competitive programming can be achieved using high-quality synthetic data, addressing scalability and contamination concerns of real-world data dependency.

Abstract: Competitive programming poses a significant challenge for Code LLMs. While recent models have shown promise, they heavily rely on finite real-world data, raising concerns about scalability and contamination. In this paper, we investigate a critical question: Can we elevate models to expert-level reasoning performance using fully synthetic data? In response, we first observe that off-the-shelf synthesis methods yield suboptimal results in this domain. To address this, we systematically investigate the key factors governing synthetic data quality. Leveraging these findings, we significantly advance the feature-based synthesis paradigm via domain-specific evolution and a dual-verification strategy, promoting task solvability, solution correctness, and test accuracy. Using this high-quality synthetic data, we train the X-Coder model series under an SFT-then-RL paradigm. X-Coder-7B shows significant performance gains on the challenging LiveCodeBench v5 (62.9% avg@8) and v6 (55.8% avg@8), outperforming larger models trained on real-world data. Extensive analysis distills valuable insights into synthetic data scaling, the necessity of domain-adapted feature evolution, and code-centric reinforcement.

Method: Uses Retrieval-Augmented Generation (RAG) pipeline with hybrid retrieval mechanism that queries local medical knowledge base and academic literature (PubMed, Semantic Scholar APIs). Retrieved evidence is filtered/ordered with reranking model, then LLM generates final educational content and questions.

Result: Three radiologists assessed outputs as high clinical/educational value. Large-scale LLM-as-a-Judge evaluation showed moderate alignment between LLM judgments and human expert assessments, highlighting need for expert oversight.

Conclusion: MedTutor effectively generates evidence-based educational content from case reports, but expert oversight remains necessary despite LLM evaluation capabilities.

Abstract: The learning process for medical residents presents significant challenges, demanding both the ability to interpret complex case reports and the rapid acquisition of accurate medical knowledge from reliable sources. Residents typically study case reports and engage in discussions with peers and mentors, but finding relevant educational materials and evidence to support their learning from these cases is often time-consuming and challenging. To address this, we introduce MedTutor, a novel system designed to augment resident training by automatically generating evidence-based educational content and multiple-choice questions from clinical case reports. MedTutor leverages a Retrieval-Augmented Generation (RAG) pipeline that takes clinical case reports as input and produces targeted educational materials. The system’s architecture features a hybrid retrieval mechanism that synergistically queries a local knowledge base of medical textbooks and academic literature (using PubMed, Semantic Scholar APIs) for the latest related research, ensuring the generated content is both foundationally sound and current. The retrieved evidence is filtered and ordered using a state-of-the-art reranking model and then an LLM generates the final long-form output describing the main educational content regarding the case-report. We conduct a rigorous evaluation of the system. First, three radiologists assessed the quality of outputs, finding them to be of high clinical and educational value. Second, we perform a large scale evaluation using an LLM-as-a Judge to understand if LLMs can be used to evaluate the output of the system. Our analysis using correlation between LLMs outputs and human expert judgments reveals a moderate alignment and highlights the continued necessity of expert oversight.

Method: Created a benchmark with 7,410 matched image-critique pairs spanning eight cultural traditions, using a five-layer framework (L1-L5) from Visual Perception to Philosophical Aesthetics, instantiated as 225 culture-specific dimensions with expert-written bilingual critiques.

Result: Pilot results show that higher-layer reasoning (L3-L5) is consistently more challenging for VLMs than visual and technical analysis (L1-L2), highlighting the gap in cultural understanding capabilities.

Conclusion: VULCA-Bench provides a comprehensive tool to evaluate VLMs’ cultural understanding beyond surface-level perception, revealing significant challenges in higher-order cultural interpretation that current models struggle with.

Abstract: We introduce VULCA-Bench, a multicultural art-critique benchmark for evaluating Vision-Language Models’ (VLMs) cultural understanding beyond surface-level visual perception. Existing VLM benchmarks predominantly measure L1-L2 capabilities (object recognition, scene description, and factual question answering) while under-evaluate higher-order cultural interpretation. VULCA-Bench contains 7,410 matched image-critique pairs spanning eight cultural traditions, with Chinese-English bilingual coverage. We operationalise cultural understanding using a five-layer framework (L1-L5, from Visual Perception to Philosophical Aesthetics), instantiated as 225 culture-specific dimensions and supported by expert-written bilingual critiques. Our pilot results indicate that higher-layer reasoning (L3-L5) is consistently more challenging than visual and technical analysis (L1-L2). The dataset, evaluation scripts, and annotation tools are available under CC BY 4.0 in the supplementary materials.

Method: DyCP is implemented outside the LLM as a lightweight context management method that dynamically identifies and retrieves relevant dialogue segments conditioned on the current turn, without requiring offline memory construction or predefined topic boundaries.

Result: Across three long-form dialogue benchmarks (LoCoMo, MT-Bench+, and SCM4LLMs) and multiple LLM backends, DyCP achieves competitive answer quality with more selective context usage and improved inference efficiency.

Conclusion: DyCP provides an effective solution for managing dialogue context in LLMs, enabling adaptive and efficient context selection while preserving dialogue sequentiality, addressing practical constraints of inference cost and latency.

Abstract: Large Language Models (LLMs) increasingly operate over long-form dialogues with frequent topic shifts. While recent LLMs support extended context windows, efficient management of dialogue history in practice is needed due to inference cost and latency constraints. We present DyCP, a lightweight context management method implemented outside the LLM that dynamically identifies and retrieves relevant dialogue segments conditioned on the current turn, without offline memory construction. DyCP manages dialogue context while preserving the sequential nature of dialogue without predefined topic boundaries, enabling adaptive and efficient context selection. Across three long-form dialogue benchmarks-LoCoMo, MT-Bench+, and SCM4LLMs-and multiple LLM backends, DyCP achieves competitive answer quality in downstream generation, with more selective context usage and improved inference efficiency.

Method: Leverages scaling laws to optimize training configurations and vocabulary size under fixed computational budgets. Uses a multi-stage data processing pipeline for corpus curation. Introduces Think-Fusion training recipe for user-controlled switching between thinking and non-thinking modes within a single unified model.

Result: A.X K1 achieves performance competitive with leading open-source models while establishing distinctive advantage in Korean-language benchmarks.

Conclusion: The paper presents a large-scale MoE language model with controllable reasoning capabilities that demonstrates strong performance, particularly in Korean language tasks.

Abstract: We introduce A.X K1, a 519B-parameter Mixture-of-Experts (MoE) language model trained from scratch. Our design leverages scaling laws to optimize training configurations and vocabulary size under fixed computational budgets. A.X K1 is pre-trained on a corpus of approximately 10T tokens, curated by a multi-stage data processing pipeline. Designed to bridge the gap between reasoning capability and inference efficiency, A.X K1 supports explicitly controllable reasoning to facilitate scalable deployment across diverse real-world scenarios. We propose a simple yet effective Think-Fusion training recipe, enabling user-controlled switching between thinking and non-thinking modes within a single unified model. Extensive evaluations demonstrate that A.X K1 achieves performance competitive with leading open-source models, while establishing a distinctive advantage in Korean-language benchmarks.

Method: Built from 72 highly-cited LLM surveys with expert-authored taxonomy trees (3,815 papers mapped to categories). Introduces novel metrics: Unordered Semantic Tree Edit Distance (US-TED/US-NTED) and Semantic Path Similarity (Sem-Path) to evaluate taxonomy structure at leaf and hierarchy levels.

Result: Evaluation of 7 Deep Research Agents and 12 frontier LLMs reveals dual bottlenecks: best agent retrieves only 20.92% of expert-cited papers, and even with perfect input, best model achieves only 31.24% Adjusted Rand Index with substantial structural gaps.

Conclusion: TaxoBench provides a comprehensive benchmark for evaluating deep research agents’ retrieval and organization capabilities, revealing significant gaps between current automated systems and human expert performance in literature survey generation.

Abstract: Deep Research Agents increasingly automate survey generation, yet whether they match human experts in two core abilities remains unclear: retrieving essential papers and organizing them into expert-like taxonomies. Existing benchmarks emphasize writing quality or citation correctness, while standard clustering metrics fail to capture hierarchical taxonomy structure. We introduce TaxoBench, a benchmark built from 72 highly-cited LLM surveys containing expert-authored taxonomy trees with 3,815 papers mapped to paper categories as ground truth. TaxoBench evaluates both abilities: (1) retrieval, measuring whether agents retrieve expert-cited papers; and (2) organization, assessed at two levels: the leaf-level measures paper-to-category assignment, while the hierarchy-level measures taxonomy structure via novel metrics – Unordered Semantic Tree Edit Distance (US-TED/US-NTED) and Semantic Path Similarity (Sem-Path). TaxoBench supports two evaluation modes: Deep Research tests end-to-end capability given only a topic, while Bottom-Up provides the expert paper set to isolate organization ability. Evaluating 7 Deep Research Agents and 12 frontier LLMs reveals a dual bottleneck: the best agent retrieves only 20.92% of expert-cited papers, and even with perfect input, the best model achieves only 31.24% ARI with substantial structural gaps. Our benchmark is publicly available at https://github.com/KongLongGeFDU/TaxoBench

Method: Proposes Self-Predictive Token Skipping (SPTS) with two selective token skipping strategies: Partial Attention Probing (PAP) for multi-head attention and Low-rank Transformation Probing (LTP) for feed forward networks. Also introduces Multi-Stage Delayed Pruning (MSDP) to reallocate skipping budgets and progressively remove redundant tokens across layers.

Result: Achieves up to 2.46× speedup for prefilling and 2.29× speedup for end-to-end generation while maintaining state-of-the-art accuracy in extensive experiments.

Conclusion: SPTS effectively addresses limitations of existing token skipping methods, providing an efficient training-free framework for long-context LLM inference with optimal speed-accuracy trade-offs.

Abstract: Long-context inference enhances the reasoning capability of Large Language Models (LLMs), but incurs significant computational overhead. Token-oriented methods, such as pruning and skipping, have shown great promise in reducing inference latency, yet still suffer from inherently insufficient structure optimization, outdated selection criteria, and redundancy interference, resulting in suboptimal speed-accuracy trade-off. To address these issues, we propose a novel training-free framework dubbed Self-Predictive Token Skipping (SPTS), for efficient long-context LLM inference. Specifically, motivated by probing the influence of target layers prior to skipping, we design two selective token skipping strategies for typical structures, including Partial Attention Probing (PAP) for multi-head attention and Low-rank Transformation Probing (LTP) for feed forward network. The former selects informative tokens via partial forward attention computation, while the latter constructs a low-rank proxy network to predict token transformations. In addition, a Multi-Stage Delayed Pruning (MSDP) strategy reallocates skipping budgets and progressively removes redundant tokens across layers. Extensive experiments display the effectiveness of our method, achieving up to 2.46$\times$ and 2.29$\times$ speedups for prefilling and end-to-end generation, respectively, while maintaining state-of-the-art accuracy. We will release the source code upon acceptance.

Method: Evaluated 11 models across 4 datasets spanning mathematical, legal, and medical reasoning domains. Used personas representing four expertise levels per domain to test model responses to endorsements from sources with varying perceived authority.

Result: Models show increasing susceptibility to incorrect/misleading endorsements as source expertise increases. Higher-authority sources induce both accuracy degradation and increased confidence in wrong answers. The authority bias is mechanistically encoded within models, but models can be steered away from this bias to improve performance even when experts give misleading endorsements.

Conclusion: Language models exhibit systematic authority bias that makes them vulnerable to expert misinformation. This bias is embedded in model mechanisms but can be mitigated through steering techniques, highlighting important safety considerations for reasoning tasks.

Abstract: Prior research demonstrates that performance of language models on reasoning tasks can be influenced by suggestions, hints and endorsements. However, the influence of endorsement source credibility remains underexplored. We investigate whether language models exhibit systematic bias based on the perceived expertise of the provider of the endorsement. Across 4 datasets spanning mathematical, legal, and medical reasoning, we evaluate 11 models using personas representing four expertise levels per domain. Our results reveal that models are increasingly susceptible to incorrect/misleading endorsements as source expertise increases, with higher-authority sources inducing not only accuracy degradation but also increased confidence in wrong answers. We also show that this authority bias is mechanistically encoded within the model and a model can be steered away from the bias, thereby improving its performance even when an expert gives a misleading endorsement.

Method: Proposes Rank-Surprisal Ratio (RSR) metric combining token-wise rank and negative log-likelihood to capture both alignment and informativeness of reasoning trajectories

Result: RSR strongly correlates with post-training reasoning performance (average Spearman 0.86) across 5 student models and 11 teachers, outperforming existing metrics for trajectory and teacher selection

Conclusion: RSR provides effective assessment of reasoning trajectory suitability for LLM distillation, balancing learning signal strength with behavioral alignment

Abstract: Long chain-of-thought (CoT) trajectories provide rich supervision signals for distilling reasoning from teacher to student LLMs. However, both prior work and our experiments show that trajectories from stronger teachers do not necessarily yield better students, highlighting the importance of data-student suitability in distillation. Existing methods assess suitability primarily through student likelihood, favoring trajectories that align closely with the student model’s current behavior but overlooking more informative ones. Addressing this, we propose Rank-Surprisal Ratio (RSR), a simple metric that captures both alignment and informativeness to assess the suitability of a reasoning trajectory. RSR is motivated by the observation that effective trajectories typically balance learning signal strength and behavioral alignment by combining low absolute probability with relatively high-ranked tokens under the student model. Concretely, RSR is defined as the ratio of a trajectory’s average token-wise rank to its average negative log-likelihood, and is straightforward to compute and interpret. Across five student models and reasoning trajectories from 11 diverse teachers, RSR strongly correlates with post-training reasoning performance (average Spearman 0.86), consistently outperforming existing metrics. We further demonstrate its practical utility in both trajectory selection and teacher selection.

Method: Proposed framework evaluates model representativeness using multivariate correlation patterns in addition to marginal distributions. Compared two steering techniques (persona prompting and demographic fine-tuning) against human responses from World Values Survey.

Result: Demographically fine-tuned model better approximates marginal distributions than persona prompting, but both techniques fail to capture gold standard correlation patterns from human survey data.

Conclusion: Representativeness is a distinct aspect of value alignment; evaluation focused only on marginals can mask structural failures and lead to overly optimistic conclusions about model capabilities.

Abstract: Large language models are increasingly used to represent human opinions, values, or beliefs, and their steerability towards these ideals is an active area of research. Existing work focuses predominantly on aligning marginal response distributions, treating each survey item independently. While essential, this may overlook deeper latent structures that characterise real populations and underpin cultural values theories. We propose a framework for evaluating the representativeness of aligned models through multivariate correlation patterns in addition to marginal distributions. We show the value of our evaluation scheme by comparing two model steering techniques (persona prompting and demographic fine-tuning) and evaluating them against human responses from the World Values Survey. While the demographically fine-tuned model better approximates marginal response distributions than persona prompting, both techniques fail to fully capture the gold standard correlation patterns. We conclude that representativeness is a distinct aspect of value alignment and an evaluation focused on marginals can mask structural failures, leading to overly optimistic conclusions about model capabilities.

Method: Deep binding of base language models with extended backstories/narrative identities, using instruction-tuned models to check consistency, and conditioning models on rich contextual factors including time (study year), question framing, and participant pool characteristics.

Result: Improved simulation fidelity compared to human studies, with LLMs able to model contextual factors often omitted from experiment descriptions that hamper accurate replication.

Conclusion: LLMs with deep identity binding and contextual conditioning provide a powerful tool for exploring nuanced human behavior in social dilemmas, capturing details that affect human studies but are typically omitted from experimental descriptions.

Abstract: Humans act via a nuanced process that depends both on rational deliberation and also on identity and contextual factors. In this work, we study how large language models (LLMs) can simulate human action in the context of social dilemma games. While prior work has focused on “steering” (weak binding) of chat models to simulate personas, we analyze here how deep binding of base models with extended backstories leads to more faithful replication of identity-based behaviors. Our study has these findings: simulation fidelity vs human studies is improved by conditioning base LMs with rich context of narrative identities and checking consistency using instruction-tuned models. We show that LLMs can also model contextual factors such as time (year that a study was performed), question framing, and participant pool effects. LLMs, therefore, allow us to explore the details that affect human studies but which are often omitted from experiment descriptions, and which hamper accurate replication.

Method: Two-stage LoRA pipeline: (1) domain-oriented pre-training (PT/DOPT) for medical knowledge injection, (2) supervised fine-tuning (SFT) for instruction following on medical QA. Form weighted adapter merge by linearly combining PT and SFT LoRA deltas before exporting single merged checkpoint. Use fixed generation evaluation with template disabling chain-of-thought.

Result: Adding PT signal reactivates latent “thinking” behavior and systematically shifts output distribution. Pure SFT achieves BLEU-4=17.84, while merged model (PT=0.3, SFT=0.7) drops to BLEU-4=6.50. Multiple-choice accuracy remains comparable (0.777 vs 0.778), MedQA improves from 0.664 to 0.681. Small pipeline mistakes can spuriously attribute SFT-only behavior to merged models.

Conclusion: Adapter interference is a significant issue in safety-critical LLM applications, requiring careful verification of merged weights and pipeline implementation. The paper provides lightweight merge-verification routine and full logs for reproducibility to address these challenges.

Abstract: Large language models (LLMs) can exhibit surprising \emph{adapter interference} when combining domain adaptation and instruction alignment in safety-critical settings. We study a 14B base model trained with a two-stage LoRA pipeline: (i) domain-oriented pre-training (PT/DOPT) for medical knowledge injection and (ii) supervised fine-tuning (SFT) for instruction following on medical QA. We then form a \emph{weighted adapter merge} by linearly combining PT and SFT LoRA deltas before exporting a single merged checkpoint for inference. We find that adding PT signal can reactivate latent ``thinking’’ behavior and systematically shift the output distribution even when training/evaluation templates attempt to disable chain-of-thought. Under a fixed generation evaluation (template \texttt{qwen3_nothink}, Temp=0.6, Top-$p$=0.8), pure SFT achieves BLEU-4=17.84 on our validation set, while the merged model (PT=0.3, SFT=0.7) drops to BLEU-4=6.50. Meanwhile multiple-choice accuracy remains comparable (avg 0.777 vs 0.778) and MedQA improves from 0.664 to 0.681. We further show that small pipeline mistakes (e.g., loading the wrong adapter, export-directory overwrite, or template mismatch) can spuriously attribute SFT-only behavior to merged models. We provide a lightweight merge-verification routine that numerically checks merged weights against the intended linear combination, along with full logs for reproducibility.

Method: Use Baba Is You game where physical laws are mutable text rules to evaluate models’ ability to override learned priors. Introduce Code-Grounded Vistas (LCV) which fine-tunes models on counterfactual pairs and identifies states with contradictory rules, forcing attention to logical constraints rather than visual semantics. Represent dynamics as executable code instead of descriptive text.

Result: Larger models show inverse scaling - perform worse than smaller models when suppressing pre-trained associations. Code representation reverses this trend, enabling effective prior inhibition. LCV training approach outperforms expensive inference-time search methods in both efficiency and accuracy.

Conclusion: Representation fundamentally determines whether scaling improves or impairs contextual reasoning. Larger models aren’t universally better, especially for domains requiring dynamic overriding of learned priors. Code-based representations enable better inhibition of semantic inertia.

Abstract: LLMs struggle with Semantic Inertia: the inability to inhibit pre-trained priors (e.g., “Lava is Dangerous”) when dynamic, in-context rules contradict them. We probe this phenomenon using Baba Is You, where physical laws are mutable text rules, enabling precise evaluation of models’ ability to override learned priors when rules change. We quantatively observe that larger models can exhibit inverse scaling: they perform worse than smaller models when natural language reasoning requires suppressing pre-trained associations (e.g., accepting “Lava is Safe”). Our analysis attributes this to natural language encoding, which entangles descriptive semantics and logical rules, leading to persistent hallucinations of familiar physics despite explicit contradictory rules. Here we show that representing dynamics as executable code, rather than descriptive text, reverses this trend and enables effective prior inhibition. We introduce Code-Grounded Vistas (LCV), which fine-tunes models on counterfactual pairs and identifies states with contradictory rules, thereby forcing attention to logical constraints rather than visual semantics. This training-time approach outperforms expensive inference-time search methods in both efficiency and accuracy. Our results demonstrate that representation fundamentally determines whether scaling improves or impairs contextual reasoning. This challenges the assumption that larger models are universally better, with implications for domains that require dynamic overriding of learned priors.

Method: Proposes DGTA framework with dynamic target generation and multi-target adaptation. Explores integrated and two-stage fine-tuning strategies for LLMs on a constructed Chinese social media dataset with multi-dimensional evaluation metrics.

Result: Fine-tuned LLMs achieve strong performance: Qwen2.5-7B reaches 66.99% target recognition score, DeepSeek-R1-Distill-Qwen-7B achieves 79.26% stance detection F1 score.

Conclusion: LLM fine-tuning approaches effectively address zero-shot stance detection in the wild, demonstrating practical applicability for real-world social media scenarios with dynamic targets.

Abstract: Current stance detection research typically relies on predicting stance based on given targets and text. However, in real-world social media scenarios, targets are neither predefined nor static but rather complex and dynamic. To address this challenge, we propose a novel task: zero-shot stance detection in the wild with Dynamic Target Generation and Multi-Target Adaptation (DGTA), which aims to automatically identify multiple target-stance pairs from text without prior target knowledge. We construct a Chinese social media stance detection dataset and design multi-dimensional evaluation metrics. We explore both integrated and two-stage fine-tuning strategies for large language models (LLMs) and evaluate various baseline models. Experimental results demonstrate that fine-tuned LLMs achieve superior performance on this task: the two-stage fine-tuned Qwen2.5-7B attains the highest comprehensive target recognition score of 66.99%, while the integrated fine-tuned DeepSeek-R1-Distill-Qwen-7B achieves a stance detection F1 score of 79.26%.

Method: Proposes Self-Adaptive Process Optimization (SAPO) that adaptively introduces process supervision signals by actively minimizing the reasoner-verifier gap. Instead of relying on inefficient Monte Carlo estimations, SAPO enables the reasoner to localize errors and make rapid adjustments, similar to Error-Related Negativity in human cognition.

Result: Extensive experiments show SAPO outperforms most existing self-evolution methods on challenging mathematics and code tasks. The work also introduces two new benchmarks for process reward models in both mathematical and coding domains to further investigate SAPO’s impact on verifier performance.

Conclusion: SAPO provides an efficient self-improvement method for small language models by addressing the reasoner-verifier gap through adaptive process optimization, demonstrating strong performance on complex reasoning tasks and establishing new benchmarks for process reward evaluation.

Abstract: Existing self-evolution methods overlook the influence of fine-grained reasoning steps, which leads to the reasoner-verifier gap. The computational inefficiency of Monte Carlo (MC) process supervision further exacerbates the difficulty in mitigating the gap. Motivated by the Error-Related Negativity (ERN), which the reasoner can localize error following incorrect decisions, guiding rapid adjustments, we propose a Self-Adaptive Process Optimization (SAPO) method for self-improvement in Small Language Models (SLMs). SAPO adaptively and efficiently introduces process supervision signals by actively minimizing the reasoner-verifier gap rather than relying on inefficient MC estimations. Extensive experiments demonstrate that the proposed method outperforms most existing self-evolution methods on two challenging task types: mathematics and code. Additionally, to further investigate SAPO’s impact on verifier performance, this work introduces two new benchmarks for process reward models in both mathematical and coding tasks.

Method: Proposes CE-RM-4B, a pointwise generative reward model trained with a dedicated two-stage rollout method and adopting unified query-based criteria, using only about 5.7K high-quality data curated from open-source preference datasets.

Result: CE-RM-4B achieves superior performance on diverse reward model benchmarks, especially in Best-of-N scenarios, and delivers more effective improvements in downstream RL practice.

Conclusion: The proposed approach addresses limitations of existing LLM-as-a-Judge methods and demonstrates better alignment between benchmark performance and practical RL effectiveness.

Abstract: Automatic evaluation is crucial yet challenging for open-ended natural language generation, especially when rule-based metrics are infeasible. Compared with traditional methods, the recent LLM-as-a-Judge paradigms enable better and more flexible evaluation, and show promise as generative reward models for reinforcement learning. However, prior work has revealed a notable gap between their seemingly impressive benchmark performance and actual effectiveness in RL practice. We attribute this issue to some limitations in existing studies, including the dominance of pairwise evaluation and inadequate optimization of evaluation criteria. Therefore, we propose CE-RM-4B, a pointwise generative reward model trained with a dedicated two-stage rollout method, and adopting unified query-based criteria. Using only about 5.7K high-quality data curated from the open-source preference dataset, our CE-RM-4B achieves superior performance on diverse reward model benchmarks, especially in Best-of-N scenarios, and delivers more effective improvements in downstream RL practice.

Method: Instead of processing lengthy textual traces, VTC-R1 renders intermediate reasoning segments into compact images, which are iteratively fed back into vision-language models as “optical memory.” Fine-tunes VLMs (Glyph and Qwen3-VL) on OpenR1-Math-220K dataset achieving 3.4x token compression.

Result: Outperforms standard long-context reasoning on benchmarks (MATH500, AIME25, AMC23, GPQA-D). Achieves 2.7x speedup in end-to-end latency and 3.4x token compression.

Conclusion: VTC-R1 presents a scalable solution for reasoning-intensive applications by integrating vision-text compression into the reasoning process, improving both performance and efficiency.

Abstract: Long-context reasoning has significantly empowered large language models (LLMs) to tackle complex tasks, yet it introduces severe efficiency bottlenecks due to the computational complexity. Existing efficient approaches often rely on complex additional training or external models for compression, which limits scalability and discards critical fine-grained information. In this paper, we propose VTC-R1, a new efficient reasoning paradigm that integrates vision-text compression into the reasoning process. Instead of processing lengthy textual traces, VTC-R1 renders intermediate reasoning segments into compact images, which are iteratively fed back into vision-language models as “optical memory.” We construct a training dataset based on OpenR1-Math-220K achieving 3.4x token compression and fine-tune representative VLMs-Glyph and Qwen3-VL. Extensive experiments on benchmarks such as MATH500, AIME25, AMC23 and GPQA-D demonstrate that VTC-R1 consistently outperforms standard long-context reasoning. Furthermore, our approach significantly improves inference efficiency, achieving 2.7x speedup in end-to-end latency, highlighting its potential as a scalable solution for reasoning-intensive applications. Our code is available at https://github.com/w-yibo/VTC-R1.

Method: Large-scale multilingual evaluation across 50 countries/33 languages. Introduces Cross-Lingual Alignment Steering (CLAS) framework that: 1) aligns latent ideological representations from political prompts into shared ideological subspace, 2) uses adaptive mechanism to prevent over-correction, 3) dynamically regulates intervention strength for cross-lingual consistency.

Result: Substantial bias reduction along both economic and social axes with minimal degradation in response quality. Framework establishes scalable, interpretable paradigm for fairness-aware multilingual LLM governance.

Conclusion: CLAS provides effective post-hoc mitigation for political bias in multilingual LLMs, balancing ideological neutrality with linguistic/cultural diversity while maintaining response quality. Addresses critical gap in cross-lingual fairness evaluation and mitigation.

Abstract: Large Language Models (LLMs) increasingly shape global discourse, making fairness and ideological neutrality essential for responsible AI deployment. Despite growing attention to political bias in LLMs, prior work largely focuses on high-resource, Western languages or narrow multilingual settings, leaving cross-lingual consistency and safe post-hoc mitigation underexplored. To address this gap, we present a large-scale multilingual evaluation of political bias spanning 50 countries and 33 languages. We introduce a complementary post-hoc mitigation framework, Cross-Lingual Alignment Steering (CLAS), designed to augment existing steering methods by aligning ideological representations across languages and dynamically regulating intervention strength. This method aligns latent ideological representations induced by political prompts into a shared ideological subspace, ensuring cross lingual consistency, with the adaptive mechanism prevents over correction and preserves coherence. Experiments demonstrate substantial bias reduction along both economic and social axes with minimal degradation in response quality. The proposed framework establishes a scalable and interpretable paradigm for fairness-aware multilingual LLM governance, balancing ideological neutrality with linguistic and cultural diversity.

Method: Released EDU-CIRCUIT-HW dataset with 1,300+ authentic student handwritten solutions from university STEM courses. Used expert-verified transcriptions and grading reports to evaluate MLLMs’ upstream recognition fidelity and downstream auto-grading performance simultaneously.

Result: Uncovered astonishing scale of latent failures in MLLM-recognized content, showing insufficient reliability for high-stakes educational applications. Demonstrated that preemptive error detection and rectification with minimal human intervention (4% of solutions) can significantly enhance grading system robustness.

Conclusion: Current MLLMs are unreliable for auto-grading authentic student handwritten STEM solutions. Proactive error detection and correction with minimal human oversight can improve system robustness for educational applications.

Abstract: Multimodal Large Language Models (MLLMs) hold significant promise for revolutionizing traditional education and reducing teachers’ workload. However, accurately interpreting unconstrained STEM student handwritten solutions with intertwined mathematical formulas, diagrams, and textual reasoning poses a significant challenge due to the lack of authentic and domain-specific benchmarks. Additionally, current evaluation paradigms predominantly rely on the outcomes of downstream tasks (e.g., auto-grading), which often probe only a subset of the recognized content, thereby failing to capture the MLLMs’ understanding of complex handwritten logic as a whole. To bridge this gap, we release EDU-CIRCUIT-HW, a dataset consisting of 1,300+ authentic student handwritten solutions from a university-level STEM course. Utilizing the expert-verified verbatim transcriptions and grading reports of student solutions, we simultaneously evaluate various MLLMs’ upstream recognition fidelity and downstream auto-grading performance. Our evaluation uncovers an astonishing scale of latent failures within MLLM-recognized student handwritten content, highlighting the models’ insufficient reliability for auto-grading and other understanding-oriented applications in high-stakes educational settings. In solution, we present a case study demonstrating that leveraging identified error patterns to preemptively detect and rectify recognition errors, with only minimal human intervention (approximately 4% of the total solutions), can significantly enhance the robustness of the deployed AI-enabled grading system on unseen student solutions.

Method: Uses 3D Gaussian Splatting (3DGS) to reconstruct photorealistic scene representations from multi-view videos, then applies generative models to recover physically realistic representations, integrates them into simulation via precision calibration targets for accurate scale alignment.

Result: Vision Language Action (VLA) models trained on simulated data achieve strong zero-shot performance on downstream tasks, matching or surpassing results obtained with real-world data.

Conclusion: Reconstruction-driven world modeling enables scalable and practical embodied intelligence training, demonstrating that simulated data can effectively replace or complement real-world data for training multimodal models.

Abstract: The scalability of embodied intelligence is fundamentally constrained by the scarcity of real-world interaction data. While simulation platforms provide a promising alternative, existing approaches often suffer from a substantial visual and physical gap to real environments and rely on expensive sensors, precise robot calibration, or depth measurements, limiting their practicality at scale. We present Simulate Anything, a graphics-driven world modeling and simulation framework that enables efficient generation of high-fidelity embodied training data using only multi-view environment videos and off-the-shelf assets. Our approach reconstructs real-world environments into a photorealistic scene representation using 3D Gaussian Splatting (3DGS), seamlessly capturing fine-grained geometry and appearance from video. We then leverage generative models to recover a physically realistic representation and integrate it into a simulation environment via a precision calibration target, enabling accurate scale alignment between the reconstructed scene and the real world. Together, these components provide a unified, editable, and physically grounded world model. Vision Language Action (VLA) models trained on our simulated data achieve strong zero-shot performance on downstream tasks, matching or even surpassing results obtained with real-world data, highlighting the potential of reconstruction-driven world modeling for scalable and practical embodied intelligence training.

Method: Proposes R3G: Reasoning-Retrieval-Reranking framework that first generates reasoning plans specifying required visual cues, then uses two-stage strategy with coarse retrieval followed by fine-grained reranking to select evidence images.

Result: Improves accuracy across six MLLM backbones and nine sub-scenarios on MRAG-Bench, achieving state-of-the-art overall performance. Ablations show sufficiency-aware reranking and reasoning steps are complementary.

Conclusion: R3G effectively addresses the challenge of selecting and integrating evidence images for vision-centric VQA through a modular reasoning-guided retrieval approach.

Abstract: Vision-centric retrieval for VQA requires retrieving images to supply missing visual cues and integrating them into the reasoning process. However, selecting the right images and integrating them effectively into the model’s reasoning remains challenging.To address this challenge, we propose R3G, a modular Reasoning-Retrieval-Reranking framework.It first produces a brief reasoning plan that specifies the required visual cues, then adopts a two-stage strategy, with coarse retrieval followed by fine-grained reranking, to select evidence images.On MRAG-Bench, R3G improves accuracy across six MLLM backbones and nine sub-scenarios, achieving state-of-the-art overall performance. Ablations show that sufficiency-aware reranking and reasoning steps are complementary, helping the model both choose the right images and use them well. We release code and data at https://github.com/czh24/R3G.

Method: Proposes VDE Bench: 1) A human-annotated benchmark dataset with dense textual documents in English and Chinese (academic papers, posters, slides, exams, newspapers), 2) A decoupled evaluation framework using OCR parsing for fine-grained assessment of text modification accuracy.

Result: Comprehensive evaluation of state-of-the-art image editing models shows strong consistency between human judgments and automated metrics. The benchmark enables systematic assessment of multilingual document editing capabilities.

Conclusion: VDE Bench is the first systematic benchmark for evaluating image editing models on multilingual and densely textual visual documents, addressing a critical gap in multimodal editing research.

Abstract: In recent years, multimodal image editing models have achieved substantial progress, enabling users to manipulate visual content through natural language in a flexible and interactive manner. Nevertheless, an important yet insufficiently explored research direction remains visual document image editing, which involves modifying textual content within images while faithfully preserving the original text style and background context. Existing approaches, including AnyText, GlyphControl, and TextCtrl, predominantly focus on English-language scenarios and documents with relatively sparse textual layouts, thereby failing to adequately address dense, structurally complex documents or non-Latin scripts such as Chinese. To bridge this gap, we propose \textbf{V}isual \textbf{D}oc \textbf{E}dit Bench(VDE Bench), a rigorously human-annotated and evaluated benchmark specifically designed to assess image editing models on multilingual and complex visual document editing tasks. The benchmark comprises a high-quality dataset encompassing densely textual documents in both English and Chinese, including academic papers, posters, presentation slides, examination materials, and newspapers. Furthermore, we introduce a decoupled evaluation framework that systematically quantifies editing performance at the OCR parsing level, enabling fine-grained assessment of text modification accuracy. Based on this benchmark, we conduct a comprehensive evaluation of representative state-of-the-art image editing models. Manual verification demonstrates a strong consistency between human judgments and automated evaluation metrics. VDE Bench constitutes the first systematic benchmark for evaluating image editing models on multilingual and densely textual visual documents.

Method: Created a 100-task benchmark of reference-based marketing/design edits with binary pass/fail judging. For each task, generated 10 independent outputs to estimate per-attempt pass rate, pass@10, expected attempts under retry caps, and effective cost per successful edit combining model price with human review time. Released 50 public tasks and kept 50-task private split for server-side evaluation, with standardized JSON schema and tooling for VLM and human-based judging.

Result: Across evaluated models, per-attempt pass rates ranged 34-83% and effective cost per success spanned USD 0.66-1.42. Models with low per-image pricing became more expensive when considering total effective costs of retries and human reviews.

Conclusion: Real-world evaluation of image editing models must account for retry rates and human review costs, not just per-image pricing. The HYPE-EDIT-1 benchmark provides a practical framework for assessing model performance in production workflows.

Abstract: Public demos of image editing models are typically best-case samples; real workflows pay for retries and review time. We introduce HYPE-EDIT-1, a 100-task benchmark of reference-based marketing/design edits with binary pass/fail judging. For each task we generate 10 independent outputs to estimate per-attempt pass rate, pass@10, expected attempts under a retry cap, and an effective cost per successful edit that combines model price with human review time. We release 50 public tasks and maintain a 50-task held-out private split for server-side evaluation, plus a standardized JSON schema and tooling for VLM and human-based judging. Across the evaluated models, per-attempt pass rates span 34-83 percent and effective cost per success spans USD 0.66-1.42. Models that have low per-image pricing are more expensive when you consider the total effective cost of retries and human reviews.

Method: Multi-Modal Deep Fusion Framework with two modality-specific feature extraction networks (LiDAR and radar) and a bidirectional cross-attention fusion module to exploit complementary spatial geometric and dynamic reflection characteristics.

Result: The model achieves 40% improvement in trajectory prediction accuracy compared to baseline, validated on the MMAUD dataset from CVPR 2024 UG2+ UAV Tracking and Pose-Estimation Challenge.

Conclusion: The proposed multi-modal fusion model effectively utilizes complementary sensor data and provides an efficient solution for unauthorized UAV trajectory prediction in low-altitude applications.

Abstract: To meet the requirements for managing unauthorized UAVs in the low-altitude economy, a multi-modal UAV trajectory prediction method based on the fusion of LiDAR and millimeter-wave radar information is proposed. A deep fusion network for multi-modal UAV trajectory prediction, termed the Multi-Modal Deep Fusion Framework, is designed. The overall architecture consists of two modality-specific feature extraction networks and a bidirectional cross-attention fusion module, aiming to fully exploit the complementary information of LiDAR and radar point clouds in spatial geometric structure and dynamic reflection characteristics. In the feature extraction stage, the model employs independent but structurally identical feature encoders for LiDAR and radar. After feature extraction, the model enters the Bidirectional Cross-Attention Mechanism stage to achieve information complementarity and semantic alignment between the two modalities. To verify the effectiveness of the proposed model, the MMAUD dataset used in the CVPR 2024 UG2+ UAV Tracking and Pose-Estimation Challenge is adopted as the training and testing dataset. Experimental results show that the proposed multi-modal fusion model significantly improves trajectory prediction accuracy, achieving a 40% improvement compared to the baseline model. In addition, ablation experiments are conducted to demonstrate the effectiveness of different loss functions and post-processing strategies in improving model performance. The proposed model can effectively utilize multi-modal data and provides an efficient solution for unauthorized UAV trajectory prediction in the low-altitude economy.

Method: Created SITUATE dataset with spatial constraints for counting tasks. Conducted experiments by fine-tuning Qwen VL 2.5 7B model on SITUATE and comparing performance with Pixmo count test data and other established counting benchmarks.

Result: Fine-tuning on SITUATE improves accuracy on Pixmo count test data, but not vice versa. The dataset helps improve generalization for out-of-distribution images compared to equally sized fine-tuning sets from Pixmo count.

Conclusion: SITUATE dataset effectively bridges the gap between simple and ambiguous counting datasets, enabling better training of Vision Language Models for counting tasks with spatial constraints and improving out-of-distribution generalization.

Abstract: We present SITUATE, a novel dataset designed for training and evaluating Vision Language Models on counting tasks with spatial constraints. The dataset bridges the gap between simple 2D datasets like VLMCountBench and often ambiguous real-life datasets like TallyQA, which lack control over occlusions and spatial composition. Experiments show that our dataset helps to improve generalization for out-of-distribution images, since a finetune of Qwen VL 2.5 7B on SITUATE improves accuracy on the Pixmo count test data, but not vice versa. We cross validate this by comparing the model performance across established other counting benchmarks and against an equally sized fine-tuning set derived from Pixmo count.

Method: Two-stage hierarchical framework: 1) Deep Expansion IoU (Deep-EIoU) for motion-agnostic online association, 2) Global Tracklet Association (GTA) for trajectory-level refinement, plus pseudo-labeling to boost detector recall on small/distorted targets.

Result: Achieved winning HOTA score of 0.60 in SoccerTrack Challenge 2025, significantly reduced false positives to 982, demonstrating state-of-the-art accuracy in fisheye-based soccer tracking.

Conclusion: The synergy between local association and global reasoning effectively addresses identity switches, occlusions, and tracking fragmentation in challenging sports tracking scenarios with fisheye cameras.

Abstract: Multi-object tracking (MOT) in sports is highly challenging due to irregular player motion, uniform appearances, and frequent occlusions. These difficulties are further exacerbated by the geometric distortion and extreme scale variation introduced by static fisheye cameras. In this work, we present GTATrack, a hierarchical tracking framework that win first place in the SoccerTrack Challenge 2025. GTATrack integrates two core components: Deep Expansion IoU (Deep-EIoU) for motion-agnostic online association and Global Tracklet Association (GTA) for trajectory-level refinement. This two-stage design enables both robust short-term matching and long-term identity consistency. Additionally, a pseudo-labeling strategy is used to boost detector recall on small and distorted targets. The synergy between local association and global reasoning effectively addresses identity switches, occlusions, and tracking fragmentation. Our method achieved a winning HOTA score of 0.60 and significantly reduced false positives to 982, demonstrating state-of-the-art accuracy in fisheye-based soccer tracking. Our code is available at https://github.com/ron941/GTATrack-STC2025.

Method: The paper investigates the impact of low-light conditions and automated image acquisition on commercial PAD systems using scenario testing of remote identity validation systems, comparing performance across different environmental and procedural conditions.

Result: PAD systems experience significant performance decline in low-light or auto-capture scenarios, with model-predicted error rates increasing by about 4x under low-light conditions and doubling under auto-capture workflows. Only one tested system maintained robust performance with maximum bona fide presentation classification error rate below 3% across all scenarios.

Conclusion: Testing across diverse environments is essential to ensure robust and reliable PAD performance in real-world applications, as most commercial systems show vulnerability to common environmental and procedural variations.

Abstract: Presentation attack detection (PAD) subsystems are an important part of effective and user-friendly remote identity validation (RIV) systems. However, ensuring robust performance across diverse environmental and procedural conditions remains a critical challenge. This paper investigates the impact of low-light conditions and automated image acquisition on the robustness of commercial PAD systems using a scenario test of RIV. Our results show that PAD systems experience a significant decline in performance when utilized in low-light or auto-capture scenarios, with a model-predicted increase in error rates by a factor of about four under low-light conditions and a doubling of those odds under auto-capture workflows. Specifically, only one of the tested systems was robust to these perturbations, maintaining a maximum bona fide presentation classification error rate below 3% across all scenarios. Our findings emphasize the importance of testing across diverse environments to ensure robust and reliable PAD performance in real-world applications.

Method: Proposes Dual-Regularized Bidirectional Transformer (DRFormer) that synergizes DINO and CLIP architectures through dual-regularization mechanism ensuring diverse feature extraction and balanced contributions from both models.

Result: Extensive experiments on five benchmarks show the method effectively harmonizes local and global representations, achieving competitive performance against state-of-the-art methods.

Conclusion: The integration of vision foundation models (DINO) and vision-language models (CLIP) through DRFormer framework successfully addresses person re-identification challenges by combining complementary local and global feature representations.

Abstract: Both fine-grained discriminative details and global semantic features can contribute to solving person re-identification challenges, such as occlusion and pose variations. Vision foundation models (\textit{e.g.}, DINO) excel at mining local textures, and vision-language models (\textit{e.g.}, CLIP) capture strong global semantic difference. Existing methods predominantly rely on a single paradigm, neglecting the potential benefits of their integration. In this paper, we analyze the complementary roles of these two architectures and propose a framework to synergize their strengths by a \textbf{D}ual-\textbf{R}egularized Bidirectional \textbf{Transformer} (\textbf{DRFormer}). The dual-regularization mechanism ensures diverse feature extraction and achieves a better balance in the contributions of the two models. Extensive experiments on five benchmarks show that our method effectively harmonizes local and global representations, achieving competitive performance against state-of-the-art methods.

Method: Proposes a geospatial embedding mechanism that transforms diverse geospatial data into embedding patches spatially aligned with image patches, and a guided attention module that dynamically integrates multimodal information by computing attention weights based on correlations with auxiliary data.

Result: The proposed framework outperforms existing pretrained geospatial foundation models in predicting disease prevalence, demonstrating effectiveness in multimodal geospatial understanding.

Conclusion: The model successfully bridges the gap between visual transformers and geospatial reasoning by incorporating structured geospatial information through novel embedding and attention mechanisms, enabling more effective multimodal analysis of remote sensing data.

Abstract: Visual transformers have driven major progress in remote sensing image analysis, particularly in object detection and segmentation. Recent vision-language and multimodal models further extend these capabilities by incorporating auxiliary information, including captions, question and answer pairs, and metadata, which broadens applications beyond conventional computer vision tasks. However, these models are typically optimized for semantic alignment between visual and textual content rather than geospatial understanding, and therefore are not suited for representing or reasoning with structured geospatial layers. In this study, we propose a novel model that enhances remote sensing imagery processing with guidance from auxiliary geospatial information. Our approach introduces a geospatial embedding mechanism that transforms diverse geospatial data into embedding patches that are spatially aligned with image patches. To facilitate cross-modal interaction, we design a guided attention module that dynamically integrates multimodal information by computing attention weights based on correlations with auxiliary data, thereby directing the model toward the most relevant regions. In addition, the module assigns distinct roles to individual attention heads, allowing the model to capture complementary aspects of the guidance information and improving the interpretability of its predictions. Experimental results demonstrate that the proposed framework outperforms existing pretrained geospatial foundation models in predicting disease prevalence, highlighting its effectiveness in multimodal geospatial understanding.

Method: Studied 60 group-housed dairy calves across 14 farms with space range 2.66-17.98 m² per calf; used detailed ethogram for manual video analysis; developed computer vision pipeline trained on 108 hours of manual annotations from 6 farms; employed linear mixed models with farm as random effect.

Result: Computer vision classifier achieved 97.6% accuracy with 99.4% recall for active play detection; calves spent average 1.0% of observation period playing (~10 minutes per 17-hour period); space-play relationship was non-linear with highest play at 8-10 m² per calf (1.6% OP) and lowest at 6-8 m² and 12-14 m² (<0.6% OP); space remained significant after controlling for age, health, and group size.

Conclusion: 8-10 m² per calf represents optimal space allowance balancing welfare benefits with economic feasibility; automated computer vision monitoring can scale small annotation projects to continuous welfare assessment systems.

Abstract: Play behaviour serves as a positive welfare indicator in dairy calves, yet the influence of space allowance under commercial conditions remains poorly characterized, particularly at intermediate-to-high allowances (6-20 m2 per calf). This study investigated the relationship between space allowance and play behaviour in 60 group-housed dairy calves across 14 commercial farms in the Netherlands (space range: 2.66-17.98 m2 per calf), and developed an automated computer vision pipeline for scalable monitoring. Video observations were analyzed using a detailed ethogram, with play expressed as percentage of observation period (%OP). Statistical analysis employed linear mixed models with farm as a random effect. A computer vision pipeline was trained on manual annotations from 108 hours on 6 farms and validated on held-out test data. The computer vision classifier achieved 97.6% accuracy with 99.4% recall for active play detection. Calves spent on average 1.0% of OP playing reflecting around 10 minutes per 17-hour period. The space-play relationship was non-linear, with highest play levels at 8-10 m2 per calf (1.6% OP) and lowest at 6-8 m2 and 12-14 m2 (<0.6% OP). Space remained significant after controlling for age, health, and group size. In summary, these findings suggest that 8-10 m2 per calf represents a practical target balancing welfare benefits with economic feasibility, and demonstrate that automated monitoring can scale small annotation projects to continuous welfare assessment systems.

Method: Integrates multi-view photogrammetry, 3D surface reconstruction, and deep learning-based segmentation to reconstruct patient-specific 3D burn surfaces from standard camera images, enabling computation of objective metrics like surface area, TBSA, and volumetric changes.

Result: System demonstrates stable reconstructions, consistent metric computation, and clinically plausible longitudinal trends, supporting objective tracking of wound contraction and depth reduction over time.

Conclusion: The platform provides a scalable, non-invasive approach to objective, geometry-aware burn assessment and decision support in clinical settings.

Abstract: Accurate, reproducible burn assessment is critical for treatment planning, healing monitoring, and medico-legal documentation, yet conventional visual inspection and 2D photography are subjective and limited for longitudinal comparison. This paper presents an AI-enabled burn assessment and management platform that integrates multi-view photogrammetry, 3D surface reconstruction, and deep learning-based segmentation within a structured clinical workflow. Using standard multi-angle images from consumer-grade cameras, the system reconstructs patient-specific 3D burn surfaces and maps burn regions onto anatomy to compute objective metrics in real-world units, including surface area, TBSA, depth-related geometric proxies, and volumetric change. Successive reconstructions are spatially aligned to quantify healing progression over time, enabling objective tracking of wound contraction and depth reduction. The platform also supports structured patient intake, guided image capture, 3D analysis and visualization, treatment recommendations, and automated report generation. Simulation-based evaluation demonstrates stable reconstructions, consistent metric computation, and clinically plausible longitudinal trends, supporting a scalable, non-invasive approach to objective, geometry-aware burn assessment and decision support in acute and outpatient care.

Method: 1S-DAug combines geometric perturbations with controlled noise injection and a denoising diffusion process conditioned on the original image. It generates augmented images, encodes them, and aggregates their representations with the original image’s representation for more robust predictions.

Result: The method consistently improves few-shot learning across 4 standard benchmarks without model parameter updates, achieving over 10% proportional accuracy improvement on miniImagenet 5-way-1-shot benchmark.

Conclusion: 1S-DAug is an effective training-free, model-agnostic plugin for few-shot learning that enhances generalization by generating diverse yet faithful image variants from minimal data at test time.

Abstract: Few-shot learning (FSL) challenges model generalization to novel classes based on just a few shots of labeled examples, a testbed where traditional test-time augmentations fail to be effective. We introduce 1S-DAug, a one-shot generative augmentation operator that synthesizes diverse yet faithful variants from just one example image at test time. 1S-DAug couples traditional geometric perturbations with controlled noise injection and a denoising diffusion process conditioned on the original image. The generated images are then encoded and aggregated, alongside the original image, into a combined representation for more robust FSL predictions. Integrated as a training-free model-agnostic plugin, 1S-DAug consistently improves FSL across standard benchmarks of 4 different datasets without any model parameter update, including achieving over 10% proportional accuracy improvement on the miniImagenet 5-way-1-shot benchmark. Codes will be released.

Method: Hierarchical agglomerative clustering algorithm that operates asynchronously and event-driven, using tempo-spatial distance for cluster detection. The algorithm employs sophisticated yet simple decision-making to achieve linear O(n) complexity where n is the number of events, with runtime independent of pixel array dimensions.

Result: The algorithm achieves efficient real-time detection of small event clusters in event camera data with linear computational complexity, making it suitable for real-time applications where computational efficiency is critical.

Conclusion: The proposed algorithm provides an efficient solution for real-time event cluster detection in event camera data, leveraging the unique asynchronous nature of the data structure while maintaining computational efficiency through linear complexity.

Abstract: This paper introduces a novel asynchronous, event-driven algorithm for real-time detection of small event clusters in event camera data. Like other hierarchical agglomerative clustering algorithms, the algorithm detects the event clusters based on their tempo-spatial distance. However, the algorithm leverages the special asynchronous data structure of event camera, and by a sophisticated, efficient and simple decision-making, enjoys a linear complexity of $O(n)$ where $n$ is the events amount. In addition, the run-time of the algorithm is independent with the dimensions of the pixels array.

Method: The study proposes a conversational agent powered by tool LLMs that converts natural language queries into executable Python code. The agent operates over a unified, extensible API supporting classification, segmentation, detection (oriented bounding boxes), spectral indices, and geospatial operators. The framework allows control at three levels: tool-level performance on EO benchmarks, agent-level code generation quality, and task-level use case evaluation.

Result: The agent outperformed general-purpose LLM/VLM baselines (GPT-4o, LLaVA), achieving 64.2% vs. 51.7% accuracy on land-composition mapping and 50% vs. 0% on post-wildfire damage assessment. The approach produces transparent, interpretable results through verifiable code output.

Conclusion: The proposed conversational code-generating agent makes Earth Observation analysis accessible to non-experts while maintaining transparency and reproducibility through code-based workflows, addressing key limitations of black-box EO systems.

Abstract: Despite recent advances in computer vision, Earth Observation (EO) analysis remains difficult to perform for the laymen, requiring expert knowledge and technical capabilities. Furthermore, many systems return black-box predictions that are difficult to audit or reproduce. Leveraging recent advances in tool LLMs, this study proposes a conversational, code-generating agent that transforms natural-language queries into executable, auditable Python workflows. The agent operates over a unified easily extendable API for classification, segmentation, detection (oriented bounding boxes), spectral indices, and geospatial operators. With our proposed framework, it is possible to control the results at three levels: (i) tool-level performance on public EO benchmarks; (ii) at the agent-level to understand the capacity to generate valid, hallucination-free code; and (iii) at the task-level on specific use cases. In this work, we select two use-cases of interest: land-composition mapping and post-wildfire damage assessment. The proposed agent outperforms general-purpose LLM/VLM baselines (GPT-4o, LLaVA), achieving 64.2% vs. 51.7% accuracy on land-composition and 50% vs. 0% on post-wildfire analysis, while producing results that are transparent and easy to interpret. By outputting verifiable code, the approach turns EO analysis into a transparent, reproducible process.

Method: TP-Blend uses two attention processors: Cross-Attention Object Fusion (CAOF) averages head-wise attention to locate spatial tokens, solves an entropy-regularized optimal transport problem to reassign multi-head feature vectors, and updates at full combined dimensionality. Self-Attention Style Fusion (SASF) injects style through Detail-Sensitive Instance Normalization with Gaussian filtering to separate low/high frequencies, and swaps Key/Value matrices from style prompt for context-aware texture modulation.

Result: 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.

Conclusion: TP-Blend provides an effective lightweight training-free framework for simultaneous object replacement and style transfer in diffusion models, enabling precise control over both content and appearance through complementary attention mechanisms.

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: Proposes a context-aware autoencoder that integrates context-specific thresholds. Compares four variants of context-aware autoencoders against a conventional autoencoder, focusing on fishing status anomalies in maritime surveillance using AIS data.

Result: The context-aware autoencoder outperforms other methods in detecting anomalies in time series data. Results show significant impact of context on reconstruction loss and anomaly detection accuracy, with improved detection performance and reduced computational cost.

Conclusion: Incorporating context-specific thresholds and recognizing the importance of context in anomaly detection offers a promising solution to improve accuracy in maritime vessel traffic surveillance systems.

Abstract: The detection of anomalies is crucial to ensuring the safety and security of maritime vessel traffic surveillance. Although autoencoders are popular for anomaly detection, their effectiveness in identifying collective and contextual anomalies is limited, especially in the maritime domain, where anomalies depend on vessel-specific contexts derived from self-reported AIS messages. To address these limitations, we propose a novel solution: the context-aware autoencoder. By integrating context-specific thresholds, our method improves detection accuracy and reduces computational cost. We compare four context-aware autoencoder variants and a conventional autoencoder using a case study focused on fishing status anomalies in maritime surveillance. Results demonstrate the significant impact of context on reconstruction loss and anomaly detection. The context-aware autoencoder outperforms others in detecting anomalies in time series data. By incorporating context-specific thresholds and recognizing the importance of context in anomaly detection, our approach offers a promising solution to improve accuracy in maritime vessel traffic surveillance systems.

Method: Three-stage framework: (1) pinpoint sensitive onset layer via attention divergence analysis, (2) use sparse autoencoders and differential activation analysis to disentangle/isolate malicious features, (3) map features to downstream causal pathways via feature influence scores from zero-out interventions, then selectively suppress these causal chains.

Result: TraceRouter significantly outperforms state-of-the-art baselines, achieving superior trade-off between adversarial robustness and general utility in extensive experiments.

Conclusion: TraceRouter provides an effective path-level defense framework that physically severs harmful information flow while preserving orthogonal computation routes, addressing limitations of localized suppression approaches.

Abstract: Despite their capabilities, large foundation models (LFMs) remain susceptible to adversarial manipulation. Current defenses predominantly rely on the “locality hypothesis”, suppressing isolated neurons or features. However, harmful semantics act as distributed, cross-layer circuits, rendering such localized interventions brittle and detrimental to utility. To bridge this gap, we propose \textbf{TraceRouter}, a path-level framework that traces and disconnects the causal propagation circuits of illicit semantics. TraceRouter operates in three stages: (1) it pinpoints a sensitive onset layer by analyzing attention divergence; (2) it leverages sparse autoencoders (SAEs) and differential activation analysis to disentangle and isolate malicious features; and (3) it maps these features to downstream causal pathways via feature influence scores (FIS) derived from zero-out interventions. By selectively suppressing these causal chains, TraceRouter physically severs the flow of harmful information while leaving orthogonal computation routes intact. Extensive experiments demonstrate that TraceRouter significantly outperforms state-of-the-art baselines, achieving a superior trade-off between adversarial robustness and general utility. Our code will be publicly released. WARNING: This paper contains unsafe model responses.

Method: Proposes D3R-Net, a dual-domain framework that couples self-supervised ‘healing’ of synthetically corrupted normal images with frequency-aware regularization. Uses spatial MSE loss plus FFT magnitude loss for frequency domain consistency, with optional SSIM term.

Result: On MVTec AD Hazelnut benchmark, FFT loss improves PRO AUC from 0.603 to 0.687 while maintaining robust image-level ROC AUC. Across 15 MVTec categories, average pixel ROC AUC increases from 0.733 to 0.751 and PRO AUC from 0.417 to 0.468, running at ~20 FPS on single GPU.

Conclusion: D3R-Net provides a practical lightweight alternative to heavy pre-trained feature embedding methods, demonstrating that frequency-aware regularization significantly improves anomaly localization in unsupervised visual inspection tasks.

Abstract: Unsupervised anomaly detection (UAD) is a key ingredient of automated visual inspection in modern manufacturing. The reconstruction-based methods appeal because they have basic architectural design and they process data quickly but they produce oversmoothed results for high-frequency details. As a result, subtle defects are partially reconstructed rather than highlighted, which limits segmentation accuracy. We build on this line of work and introduce D3R-Net, a Dual-Domain Denoising Reconstruction framework that couples a self-supervised ‘healing’ task with frequency-aware regularization. During training, the network receives synthetically corrupted normal images and is asked to reconstruct the clean targets, which prevents trivial identity mapping and pushes the model to learn the manifold of defect-free textures. In addition to the spatial mean squared error, we employ a Fast Fourier Transform (FFT) magnitude loss that encourages consistency in the frequency domain. The implementation also allows an optional structural similarity (SSIM) term, which we study in an ablation. On the MVTec AD Hazelnut benchmark, D3R-Net with the FFT loss improves localization consistency over a spatial-only baseline: PRO AUC increases from 0.603 to 0.687, while image-level ROC AUC remains robust. Evaluated across fifteen MVTec categories, the FFT variant raises the average pixel ROC AUC from 0.733 to 0.751 and PRO AUC from 0.417 to 0.468 compared to the MSE-only baseline, at roughly 20 FPS on a single GPU. The network is trained from scratch and uses a lightweight convolutional autoencoder backbone, providing a practical alternative to heavy pre-trained feature embedding methods.

Method: Presents POVNet+, a multimodal deep learning architecture that introduces ADL and motion embedding spaces. Uses these embeddings to distinguish between: 1) known ADLs being performed normally, 2) new unseen ADLs, and 3) known ADLs being performed atypically. Applies a novel user state estimation method to the motion embedding space for recognizing new ADLs while monitoring user performance.

Result: POVNet+ achieves higher ADL classification accuracy compared to state-of-the-art human activity recognition methods. Human-robot interaction experiments in a cluttered living environment with multiple users and the socially assistive robot Leia demonstrate successful identification of different seen/unseen ADLs and atypical ADL performance, with appropriate assistive interactions initiated.

Conclusion: POVNet+ represents a significant advancement for socially assistive robots, enabling them to perceive multiple ADLs (including unseen and atypical ones) and proactively initiate appropriate assistive behaviors, addressing a key barrier to long-term deployment of autonomous assistive robots.

Abstract: A significant barrier to the long-term deployment of autonomous socially assistive robots is their inability to both perceive and assist with multiple activities of daily living (ADLs). In this paper, we present the first multimodal deep learning architecture, POVNet+, for multi-activity recognition for socially assistive robots to proactively initiate assistive behaviors. Our novel architecture introduces the use of both ADL and motion embedding spaces to uniquely distinguish between a known ADL being performed, a new unseen ADL, or a known ADL being performed atypically in order to assist people in real scenarios. Furthermore, we apply a novel user state estimation method to the motion embedding space to recognize new ADLs while monitoring user performance. This ADL perception information is used to proactively initiate robot assistive interactions. Comparison experiments with state-of-the-art human activity recognition methods show our POVNet+ method has higher ADL classification accuracy. Human-robot interaction experiments in a cluttered living environment with multiple users and the socially assistive robot Leia using POVNet+ demonstrate the ability of our multi-modal ADL architecture in successfully identifying different seen and unseen ADLs, and ADLs being performed atypically, while initiating appropriate assistive human-robot interactions.

Method: Three-module framework: 1) SimDen module generates dense radar point clouds using Gaussian simulation and curvature-based outline generation; 2) RCM module uses radar data to compensate for vision degradation; 3) MMIF module uses Mamba modeling for heterogeneous reduction and interactive fusion.

Result: Achieves best performance on VoD, TJ4DRadSet, and Astyx HiRes 2019 datasets with lower parameter quantity and faster inference speed compared to other methods.

Conclusion: SDCM effectively addresses sparse radar and degraded vision challenges in multimodal 3D object detection for IoV applications through simulated densifying, compensatory modeling, and Mamba-based interactive fusion.

Abstract: 3-D object detection based on 4-D radar-vision is an important part in Internet of Vehicles (IoV). However, there are two challenges which need to be faced. First, the 4-D radar point clouds are sparse, leading to poor 3-D representation. Second, vision datas exhibit representation degradation under low-light, long distance detection and dense occlusion scenes, which provides unreliable texture information during fusion stage. To address these issues, a framework named SDCM is proposed, which contains Simulated Densifying and Compensatory Modeling Fusion for radar-vision 3-D object detection in IoV. Firstly, considering point generation based on Gaussian simulation of key points obtained from 3-D Kernel Density Estimation (3-D KDE), and outline generation based on curvature simulation, Simulated Densifying (SimDen) module is designed to generate dense radar point clouds. Secondly, considering that radar data could provide more real time information than vision data, due to the all-weather property of 4-D radar. Radar Compensatory Mapping (RCM) module is designed to reduce the affects of vision datas’ representation degradation. Thirdly, considering that feature tensor difference values contain the effective information of every modality, which could be extracted and modeled for heterogeneity reduction and modalities interaction, Mamba Modeling Interactive Fusion (MMIF) module is designed for reducing heterogeneous and achieving interactive Fusion. Experiment results on the VoD, TJ4DRadSet and Astyx HiRes 2019 dataset show that SDCM achieves best performance with lower parameter quantity and faster inference speed. Our code will be available.

Method: SCANNER uses centroid-guided alignment to reveal stable cores from ambiguous hateful content, incorporates sample-level adaptive centroid alignment to handle outliers, and adds intra-cluster diversity regularization to prevent semantic collapse.

Result: SCANNER outperforms all baselines with an average gain of 4.69% in Macro-F1 over the best existing method.

Conclusion: The proposed framework effectively addresses severe semantic drift in hate video detection by leveraging invariant core concepts and adaptive alignment strategies.

Abstract: Hate Video Detection (HVD) is crucial for online ecosystems. Existing methods assume identical distributions between training (source) and inference (target) data. However, hateful content often evolves into irregular and ambiguous forms to evade censorship, resulting in substantial semantic drift and rendering previously trained models ineffective. Test-Time Adaptation (TTA) offers a solution by adapting models during inference to narrow the cross-domain gap, while conventional TTA methods target mild distribution shifts and struggle with the severe semantic drift in HVD. To tackle these challenges, we propose SCANNER, the first TTA framework tailored for HVD. Motivated by the insight that, despite the evolving nature of hateful manifestations, their underlying cores remain largely invariant (i.e., targeting is still based on characteristics like gender, race, etc), we leverage these stable cores as a bridge to connect the source and target domains. Specifically, SCANNER initially reveals the stable cores from the ambiguous layout in evolving hateful content via a principled centroid-guided alignment mechanism. To alleviate the impact of outlier-like samples that are weakly correlated with centroids during the alignment process, SCANNER enhances the prior by incorporating a sample-level adaptive centroid alignment strategy, promoting more stable adaptation. Furthermore, to mitigate semantic collapse from overly uniform outputs within clusters, SCANNER introduces an intra-cluster diversity regularization that encourages the cluster-wise semantic richness. Experiments show that SCANNER outperforms all baselines, with an average gain of 4.69% in Macro-F1 over the best.

Method: Uses motion vectors and transform coefficients from compressed video streams to propagate object bounding boxes across frames with a deep model, eliminating need for full RGB video decoding.

Result: Achieves computational speed-up of up to 3.7x with only 4% mAP@0.5 drop compared to RGB baseline on MOTS15/17/20 datasets.

Conclusion: Compressed-domain motion modeling is efficient for real-time analytics in large monitoring systems, offering significant speed improvements with minimal accuracy trade-offs.

Abstract: We propose a lightweight compressed-domain tracking model that operates directly on video streams, without requiring full RGB video decoding. Using motion vectors and transform coefficients from compressed data, our deep model propagates object bounding boxes across frames, achieving a computational speed-up of order up to 3.7 with only a slight 4% mAP@0.5 drop vs RGB baseline on MOTS15/17/20 datasets. These results highlight codec-domain motion modeling efficiency for real-time analytics in large monitoring systems.

Method: Proposes LLaVA-FA that performs joint low-rank plus quantization approximation in frequency domain using Fourier transform properties. Introduces PolarQuant for complex matrices and optional diagonal calibration scheme without large-scale calibration data.

Result: Outperforms existing efficient multimodal models across multiple benchmarks while maintaining minimal activated parameters and low computational costs.

Conclusion: LLaVA-FA is an effective solution for compressing large multimodal models through frequency-domain joint approximation and specialized quantization techniques.

Abstract: Large multimodal models (LMMs) have achieved impressive performance on various vision-language tasks, but their substantial computational and memory costs hinder their practical deployment. Existing compression methods often decouple low-rank decomposition and quantization, leading to compounded reconstruction errors, especially in multimodal architectures with cross-modal redundancy. To address this issue, we propose LLaVA-FA, a novel efficient LMM that performs joint low-rank plus quantization approximation in the frequency domain. By leveraging the de-correlation and conjugate symmetry properties of Fourier transform, LLaVA-FA achieves more compact and accurate weight representations. Furthermore, we introduce PolarQuant, a polar-coordinate quantization method tailored for complex matrices, and an optional diagonal calibration (ODC) scheme that eliminates the need for large-scale calibration data. Extensive experimental results demonstrate that our proposed LLaVA-FA outperforms existing efficient multimodal models across multiple benchmarks while maintaining minimal activated parameters and low computational costs, validating its effectiveness as a powerful solution for compressing LMMs.

Method: Developed a mobile application with integrated deep learning model for offline cacao disease identification. The system includes disease detection and infection level assessment models trained on cacao disease data.

Result: Disease identification model achieved 96.93% validation accuracy, infection level detection model achieved 79.49% accuracy. Field testing showed 84.2% agreement rate with expert assessments.

Conclusion: The offline mobile app with deep learning provides accessible technology for smallholder farmers to improve cacao crop health and productivity through accurate disease identification.

Abstract: Smallholder cacao producers often rely on outdated farming techniques and face significant challenges from pests and diseases, unlike larger plantations with more resources and expertise. In the Philippines, cacao farmers have limited access to data, information, and good agricultural practices. This study addresses these issues by developing a mobile application for cacao disease identification and management that functions offline, enabling use in remote areas where farms are mostly located. The core of the system is a deep learning model trained to identify cacao diseases accurately. The trained model is integrated into the mobile app to support farmers in field diagnosis. The disease identification model achieved a validation accuracy of 96.93% while the model for detecting cacao black pod infection levels achieved 79.49% validation accuracy. Field testing of the application showed an agreement rate of 84.2% compared with expert cacao technician assessments. This approach empowers smallholder farmers by providing accessible, technology-enabled tools to improve cacao crop health and productivity.

Method: Proposes Low-Rank Factorized RGDA (LR-RGDA) that exploits low-rank structure of covariance via Woodbury matrix identity to decompose discriminant function into global affine term plus low-rank quadratic perturbation, reducing inference complexity from O(Cd²) to O(d² + Crd²). Also introduces Hopfield-based Distribution Compensator (HopDC) - a training-free mechanism using continuous Hopfield Networks to recalibrate historical class statistics through associative memory dynamics on unlabeled anchors, with theoretical error bound.

Result: Extensive experiments on diverse CIL benchmarks demonstrate state-of-the-art performance, providing a scalable solution for large-scale class-incremental learning with Vision Transformers.

Conclusion: The framework achieves efficient and accurate class-incremental learning for Vision Transformers by combining the expressivity of RGDA with linear classifier efficiency through low-rank factorization, while addressing representation drift via Hopfield-based compensation.

Abstract: Class-incremental learning (CIL) with Vision Transformers (ViTs) faces a major computational bottleneck during the classifier reconstruction phase, where most existing methods rely on costly iterative stochastic gradient descent (SGD). We observe that analytic Regularized Gaussian Discriminant Analysis (RGDA) provides a Bayes-optimal alternative with accuracy comparable to SGD-based classifiers; however, its quadratic inference complexity limits its use in large-scale CIL scenarios. To overcome this, we propose Low-Rank Factorized RGDA (LR-RGDA), a scalable classifier that combines RGDA’s expressivity with the efficiency of linear classifiers. By exploiting the low-rank structure of the covariance via the Woodbury matrix identity, LR-RGDA decomposes the discriminant function into a global affine term refined by a low-rank quadratic perturbation, reducing the inference complexity from $\mathcal{O}(Cd^2)$ to $\mathcal{O}(d^2 + Crd^2)$, where $C$ is the class number, $d$ the feature dimension, and $r \ll d$ the subspace rank. To mitigate representation drift caused by backbone updates, we further introduce Hopfield-based Distribution Compensator (HopDC), a training-free mechanism that uses modern continuous Hopfield Networks to recalibrate historical class statistics through associative memory dynamics on unlabeled anchors, accompanied by a theoretical bound on the estimation error. Extensive experiments on diverse CIL benchmarks demonstrate that our framework achieves state-of-the-art performance, providing a scalable solution for large-scale class-incremental learning with ViTs. Code: https://github.com/raoxuan98-hash/lr_rgda_hopdc.

Method: Proposed DensiThAI, a multi-view deep learning framework for breast density classification from thermal images. Uses five standard thermal views and mammography-derived density labels as reference for training and evaluation on a multi-center dataset of 3,500 women.

Result: Achieved mean AUROC of 0.73 across 10 random splits with statistically significant separation between density classes (p « 0.05). Consistent performance across age cohorts demonstrates feasibility.

Conclusion: Thermal imaging shows potential as a non-ionizing approach for breast density assessment, offering improved patient experience and workflow optimization compared to traditional mammography.

Abstract: Breast tissue density is a key biomarker of breast cancer risk and a major factor affecting mammographic sensitivity. However, density assessment currently relies almost exclusively on X-ray mammography, an ionizing imaging modality. This study investigates the feasibility of estimating breast density using artificial intelligence over infrared thermal images, offering a non-ionizing imaging approach. The underlying hypothesis is that fibroglandular and adipose tissues exhibit distinct thermophysical and physiological properties, leading to subtle but spatially coherent temperature variations on the breast surface. In this paper, we propose DensiThAI, a multi-view deep learning framework for breast density classification from thermal images. The framework was evaluated on a multi-center dataset of 3,500 women using mammography-derived density labels as reference. Using five standard thermal views, DensiThAI achieved a mean AUROC of 0.73 across 10 random splits, with statistically significant separation between density classes across all splits (p « 0.05). Consistent performance across age cohorts supports the potential of thermal imaging as a non-ionizing approach for breast density assessment with implications for improved patient experience and workflow optimization.

Method: Neural Gaussian Force Field (NGFF) integrates 3D Gaussian perception with physics-based dynamic modeling in an end-to-end neural framework. Uses GSCollision dataset with 640k rendered physical videos for training, featuring diverse materials and multi-object interactions.

Result: NGFF achieves two orders of magnitude faster generation than prior Gaussian simulators, shows strong generalization and robustness in physical reasoning on synthetic and real 3D scenarios, and advances video prediction towards physics-grounded world models.

Conclusion: NGFF successfully bridges the gap between visual quality and physical plausibility in video generation, providing an efficient framework for physically realistic 4D video prediction from multi-view inputs.

Abstract: Predicting physical dynamics from raw visual data remains a major challenge in AI. While recent video generation models have achieved impressive visual quality, they still cannot consistently generate physically plausible videos due to a lack of modeling of physical laws. Recent approaches combining 3D Gaussian splatting and physics engines can produce physically plausible videos, but are hindered by high computational costs in both reconstruction and simulation, and often lack robustness in complex real-world scenarios. To address these issues, we introduce Neural Gaussian Force Field (NGFF), an end-to-end neural framework that integrates 3D Gaussian perception with physics-based dynamic modeling to generate interactive, physically realistic 4D videos from multi-view RGB inputs, achieving two orders of magnitude faster than prior Gaussian simulators. To support training, we also present GSCollision, a 4D Gaussian dataset featuring diverse materials, multi-object interactions, and complex scenes, totaling over 640k rendered physical videos (~4 TB). Evaluations on synthetic and real 3D scenarios show NGFF’s strong generalization and robustness in physical reasoning, advancing video prediction towards physics-grounded world models.

Method: Proposes CLEAR framework with frequency-domain decomposition and re-composition module, trajectory alignment loss, and uses an ISP-based flicker simulation pipeline for training stabilization and degradation distribution expansion.

Result: Extensive experiments show CLEAR consistently outperforms existing image restoration approaches across multiple evaluation metrics in complex real-world scenarios.

Conclusion: First systematic study on joint removal of moiré patterns and flicker-banding in screen-captured images, with proposed CLEAR framework effectively addressing compound artifacts through novel frequency-domain modeling and training techniques.

Abstract: Capturing display screens with mobile devices has become increasingly common, yet the resulting images often suffer from severe degradations caused by the coexistence of moiré patterns and flicker-banding, leading to significant visual quality degradation. Due to the strong coupling of these two artifacts in real imaging processes, existing methods designed for single degradations fail to generalize to such compound scenarios. In this paper, we present the first systematic study on joint removal of moiré patterns and flicker-banding in screen-captured images, and propose a unified restoration framework, named CLEAR. To support this task, we construct a large-scale dataset containing both moiré patterns and flicker-banding, and introduce an ISP-based flicker simulation pipeline to stabilize model training and expand the degradation distribution. Furthermore, we design a frequency-domain decomposition and re-composition module together with a trajectory alignment loss to enhance the modeling of compound artifacts. Extensive experiments demonstrate that the proposed method consistently. outperforms existing image restoration approaches across multiple evaluation metrics, validating its effectiveness in complex real-world scenarios.

Method: Used multiple instance learning frameworks with five state-of-the-art histopathology foundation models to extract patch-level features from whole slide images. Compared these against contrastive learning-based features across breast, endometrial, and lung cancer cohorts from two public datasets. Proposed a distribution-based upsampling strategy to address target imbalance, and conducted ablation studies on different sampling strategies and instance bag sizes.

Result: Models trained on foundation model features consistently outperformed baseline contrastive learning features in predictive accuracy and generalization across cancer types. Systematic differences were observed among different foundation models. The proposed distribution-based upsampling significantly improved recall and balanced accuracy for underrepresented patient populations.

Conclusion: Large-scale histopathological pretraining provides benefits for more precise and transferable regressive biomarker prediction, demonstrating potential to advance AI-driven precision oncology through improved biomarker prediction capabilities.

Abstract: Foundation models pretrained on large-scale histopathology data have found great success in various fields of computational pathology, but their impact on regressive biomarker prediction remains underexplored. In this work, we systematically evaluate histopathological foundation models for regression-based tasks, demonstrated through the prediction of homologous recombination deficiency (HRD) score - a critical biomarker for personalized cancer treatment. Within multiple instance learning frameworks, we extract patch-level features from whole slide images (WSI) using five state-of-the-art foundation models, and evaluate their impact compared to contrastive learning-based features. Models are trained to predict continuous HRD scores based on these extracted features across breast, endometrial, and lung cancer cohorts from two public medical data collections. Extensive experiments demonstrate that models trained on foundation model features consistently outperform the baseline in terms of predictive accuracy and generalization capabilities while exhibiting systematic differences among the foundation models. Additionally, we propose a distribution-based upsampling strategy to mitigate target imbalance in these datasets, significantly improving the recall and balanced accuracy for underrepresented but clinically important patient populations. Furthermore, we investigate the impact of different sampling strategies and instance bagsizes by ablation studies. Our results highlight the benefits of large-scale histopathological pretraining for more precise and transferable regressive biomarker prediction, showcasing its potential to advance AI-driven precision oncology.

Method: Two-layer hierarchical architecture: first layer extracts preliminary features using FFT spectrograms; second layer selectively activates either stationary activity module or parallel LSTM-MobileNet network (PLMN) for dynamic states. PLMN fuses FFT, Wavelet, and Gabor spectrograms through parallel LSTM encoders and refines features using Efficient Channel Attention and Depthwise Separable Convolution.

Result: Achieves 96.70% accuracy with only 22.3 KiB RAM and 439.5 KiB ROM on ARM Cortex-M4. Compared to MobileNetV3: improves accuracy by 1.22%, reduces RAM usage by 71.2%, and reduces ROM usage by 42.1%.

Conclusion: HPPI-Net achieves favorable accuracy-efficiency trade-off with explainable predictions, establishing a practical solution for memory-constrained edge platforms in wearable, industrial, and smart home HAR applications.

Abstract: The demand for accurate on-device pattern recognition in edge applications is intensifying, yet existing approaches struggle to reconcile accuracy with computational constraints. To address this challenge, a resource-aware hierarchical network based on multi-spectral fusion and interpretable modules, namely the Hierarchical Parallel Pseudo-image Enhancement Fusion Network (HPPI-Net), is proposed for real-time, on-device Human Activity Recognition (HAR). Deployed on an ARM Cortex-M4 microcontroller for low-power real-time inference, HPPI-Net achieves 96.70% accuracy while utilizing only 22.3 KiB of RAM and 439.5 KiB of ROM after optimization. HPPI-Net employs a two-layer architecture. The first layer extracts preliminary features using Fast Fourier Transform (FFT) spectrograms, while the second layer selectively activates either a dedicated module for stationary activity recognition or a parallel LSTM-MobileNet network (PLMN) for dynamic states. PLMN fuses FFT, Wavelet, and Gabor spectrograms through three parallel LSTM encoders and refines the concatenated features using Efficient Channel Attention (ECA) and Depthwise Separable Convolution (DSC), thereby offering channel-level interpretability while substantially reducing multiply-accumulate operations. Compared with MobileNetV3, HPPI-Net improves accuracy by 1.22% and reduces RAM usage by 71.2% and ROM usage by 42.1%. These results demonstrate that HPPI-Net achieves a favorable accuracy-efficiency trade-off and provides explainable predictions, establishing a practical solution for wearable, industrial, and smart home HAR on memory-constrained edge platforms.

Method: Developed a pose-based machine learning framework that converts outpatient clinical videos into anatomically meaningful keypoint time series, then computes kinematic descriptors including statistical, temporal, spectral, and higher-order irregularity-complexity features

Result: The method provides objective and scalable analysis of movement disorders from routine clinical videos, addressing the limitations of subjective clinical assessments

Conclusion: This framework offers a promising approach for objective clinical recognition and longitudinal monitoring of hyperkinetic movement disorders using standard video data

Abstract: Hyperkinetic movement disorders (HMDs) such as dystonia, tremor, chorea, myoclonus, and tics are disabling motor manifestations across childhood and adulthood. Their fluctuating, intermittent, and frequently co-occurring expressions hinder clinical recognition and longitudinal monitoring, which remain largely subjective and vulnerable to inter-rater variability. Objective and scalable methods to distinguish overlapping HMD phenotypes from routine clinical videos are still lacking. Here, we developed a pose-based machine-learning framework that converts standard outpatient videos into anatomically meaningful keypoint time series and computes kinematic descriptors spanning statistical, temporal, spectral, and higher-order irregularity-complexity features.

Method: Integrates YOLOv26’s NMS-free, end-to-end design with open-vocabulary learning via object embedding head for similarity matching against text/visual prompts. Uses RepRTA for zero-overhead text prompting, SAVPE for example-guided segmentation, and Lazy Region Prompt Contrast for prompt-free inference.

Result: Demonstrates consistent scaling behavior and favorable accuracy-efficiency trade-offs across model sizes in both prompted and prompt-free settings, providing practical real-time open-vocabulary instance segmentation.

Conclusion: YOLOE-26 offers a scalable, practical solution for real-time open-vocabulary instance segmentation that maintains YOLO’s efficiency while extending capabilities to dynamic real-world environments.

Abstract: This paper presents YOLOE-26, a unified framework that integrates the deployment-optimized YOLO26(or YOLOv26) architecture with the open-vocabulary learning paradigm of YOLOE for real-time open-vocabulary instance segmentation. Building on the NMS-free, end-to-end design of YOLOv26, the proposed approach preserves the hallmark efficiency and determinism of the YOLO family while extending its capabilities beyond closed-set recognition. YOLOE-26 employs a convolutional backbone with PAN/FPN-style multi-scale feature aggregation, followed by end-to-end regression and instance segmentation heads. A key architectural contribution is the replacement of fixed class logits with an object embedding head, which formulates classification as similarity matching against prompt embeddings derived from text descriptions, visual examples, or a built-in vocabulary. To enable efficient open-vocabulary reasoning, the framework incorporates Re-Parameterizable Region-Text Alignment (RepRTA) for zero-overhead text prompting, a Semantic-Activated Visual Prompt Encoder (SAVPE) for example-guided segmentation, and Lazy Region Prompt Contrast for prompt-free inference. All prompting modalities operate within a unified object embedding space, allowing seamless switching between text-prompted, visual-prompted, and fully autonomous segmentation. Extensive experiments demonstrate consistent scaling behavior and favorable accuracy-efficiency trade-offs across model sizes in both prompted and prompt-free settings. The training strategy leverages large-scale detection and grounding datasets with multi-task optimization and remains fully compatible with the Ultralytics ecosystem for training, validation, and deployment. Overall, YOLOE-26 provides a practical and scalable solution for real-time open-vocabulary instance segmentation in dynamic, real-world environments.

Method: Proposes intra-class subdivision pixel contrastive learning with “unconcerned samples” to distinguish inner vs boundary pixel representations within same class, plus boundary contrastive loss to enhance discrimination across boundaries.

Result: SPCL significantly improves segmentation performance on public cardiac datasets, outperforming existing methods in both segmentation quality and boundary precision.

Conclusion: The SPCL framework effectively addresses boundary representation issues in cardiac segmentation through novel contrastive learning techniques, leading to improved performance.

Abstract: We propose an intra-class subdivision pixel contrastive learning (SPCL) framework for cardiac image segmentation to address representation contamination at boundaries. The novel concept ``Unconcerned sample’’ is proposed to distinguish pixel representations at the inner and boundary regions within the same class, facilitating a clearer characterization of intra-class variations. A novel boundary contrastive loss for boundary representations is proposed to enhance representation discrimination across boundaries. The advantages of the unconcerned sample and boundary contrastive loss are analyzed theoretically. Experimental results in public cardiac datasets demonstrate that SPCL significantly improves segmentation performance, outperforming existing methods with respect to segmentation quality and boundary precision. Our code is available at https://github.com/Jrstud203/SPCL.

Method: Proposes a noise-frequency continuation framework that constructs intermediate posteriors with likelihood enforcing measurement consistency only within noise-dependent frequency bands. Uses stabilized posterior sampler combining diffusion predictor, band-limited likelihood guidance, and multi-resolution consistency strategy.

Result: Achieves state-of-the-art performance across super-resolution, inpainting, and deblurring tasks, improving motion deblurring PSNR by up to 5 dB over strong baselines.

Conclusion: The noise-frequency continuation framework effectively addresses limitations of current diffusion posterior sampling by better coupling measurement consistency with diffusion noise levels, enabling more reliable detail recovery in inverse problems.

Abstract: Diffusion posterior sampling solves inverse problems by combining a pretrained diffusion prior with measurement-consistency guidance, but it often fails to recover fine details because measurement terms are applied in a manner that is weakly coupled to the diffusion noise level. At high noise, data-consistency gradients computed from inaccurate estimates can be geometrically incongruent with the posterior geometry, inducing early-step drift, spurious high-frequency artifacts, plus sensitivity to schedules and ill-conditioned operators. To address these concerns, we propose a noise–frequency Continuation framework that constructs a continuous family of intermediate posteriors whose likelihood enforces measurement consistency only within a noise-dependent frequency band. This principle is instantiated with a stabilized posterior sampler that combines a diffusion predictor, band-limited likelihood guidance, and a multi-resolution consistency strategy that aggressively commits reliable coarse corrections while conservatively adopting high-frequency details only when they become identifiable. Across super-resolution, inpainting, and deblurring, our method achieves state-of-the-art performance and improves motion deblurring PSNR by up to 5 dB over strong baselines.

Method: Uses Observation-Thinking-Answer (O-T-A) paradigm to decode spatio-temporal cues like trajectories and view frustums. Constructs Large-scale Inference Trajectory Suite with 18k SFT reasoning chains and 38k RL feedback samples. First to employ RL for logical alignment in this domain to ensure motion inferences are grounded in physical geometry.

Result: Achieves state-of-the-art performance across multiple benchmarks. Effectively suppresses hallucinations by applying Reinforcement Learning to the O-T-A reasoning paradigm.

Conclusion: CamReasoner successfully bridges perception and cinematic logic through structured inference, demonstrating that explicit reasoning blocks and RL-based alignment can significantly improve camera movement understanding in multimodal models.

Abstract: Understanding camera dynamics is a fundamental pillar of video spatial intelligence. However, existing multimodal models predominantly treat this task as a black-box classification, often confusing physically distinct motions by relying on superficial visual patterns rather than geometric cues. We present CamReasoner, a framework that reformulates camera movement understanding as a structured inference process to bridge the gap between perception and cinematic logic. Our approach centers on the Observation-Thinking-Answer (O-T-A) paradigm, which compels the model to decode spatio-temporal cues such as trajectories and view frustums within an explicit reasoning block. To instill this capability, we construct a Large-scale Inference Trajectory Suite comprising 18k SFT reasoning chains and 38k RL feedback samples. Notably, we are the first to employ RL for logical alignment in this domain, ensuring motion inferences are grounded in physical geometry rather than contextual guesswork. By applying Reinforcement Learning to the Observation-Think-Answer (O-T-A) reasoning paradigm, CamReasoner effectively suppresses hallucinations and achieves state-of-the-art performance across multiple benchmarks.

Method: Introduces Inpainting Exchange (INP-X) operation that restores original pixels outside edited regions while preserving synthesized content. Creates 90K test dataset with real, inpainted, and exchanged images to evaluate detector behavior.

Result: State-of-the-art detectors show dramatic accuracy drops (e.g., 91% to 55%) when tested on INP-X images, often approaching chance level. Training on INP-X dataset yields better generalization and localization than standard inpainting detection.

Conclusion: Current inpainting detectors are fundamentally flawed as they rely on global VAE artifacts rather than local content synthesis. Content-aware detection is needed, and the INP-X dataset enables more robust detector training.

Abstract: Modern deep learning-based inpainting enables realistic local image manipulation, raising critical challenges for reliable detection. However, we observe that current detectors primarily rely on global artifacts that appear as inpainting side effects, rather than on locally synthesized content. We show that this behavior occurs because VAE-based reconstruction induces a subtle but pervasive spectral shift across the entire image, including unedited regions. To isolate this effect, we introduce Inpainting Exchange (INP-X), an operation that restores original pixels outside the edited region while preserving all synthesized content. We create a 90K test dataset including real, inpainted, and exchanged images to evaluate this phenomenon. Under this intervention, pretrained state-of-the-art detectors, including commercial ones, exhibit a dramatic drop in accuracy (e.g., from 91% to 55%), frequently approaching chance level. We provide a theoretical analysis linking this behavior to high-frequency attenuation caused by VAE information bottlenecks. Our findings highlight the need for content-aware detection. Indeed, training on our dataset yields better generalization and localization than standard inpainting. Our dataset and code are publicly available at https://github.com/emirhanbilgic/INP-X.

Method: Proposes SemiEarth with a VLM pseudo-label purifying (VLM-PP) structure that uses vision-language models to purify teacher network’s pseudo-labels. The VLM-PP module corrects mispredicted categories in low-confidence pseudo-labels when discrepancies arise between VLM predictions and pseudo-labels. The approach is architecture-independent and leverages VLMs’ open-world capabilities.

Result: Extensive experiments on multiple RS datasets demonstrate that SemiEarth achieves state-of-the-art (SOTA) performance. The model shows significant improvements in multi-class boundary regions and offers good interpretability compared to previous SOTA RS S4 methods.

Conclusion: SemiEarth successfully integrates vision-language models into semi-supervised semantic segmentation for remote sensing, addressing pseudo-label quality issues and achieving superior performance with interpretability benefits.

Abstract: The semi-supervised semantic segmentation (S4) can learn rich visual knowledge from low-cost unlabeled images. However, traditional S4 architectures all face the challenge of low-quality pseudo-labels, especially for the teacher-student framework.We propose a novel SemiEarth model that introduces vision-language models (VLMs) to address the S4 issues for the remote sensing (RS) domain. Specifically, we invent a VLM pseudo-label purifying (VLM-PP) structure to purify the teacher network’s pseudo-labels, achieving substantial improvements. Especially in multi-class boundary regions of RS images, the VLM-PP module can significantly improve the quality of pseudo-labels generated by the teacher, thereby correctly guiding the student model’s learning. Moreover, since VLM-PP equips VLMs with open-world capabilities and is independent of the S4 architecture, it can correct mispredicted categories in low-confidence pseudo-labels whenever a discrepancy arises between its prediction and the pseudo-label. We conducted extensive experiments on multiple RS datasets, which demonstrate that our SemiEarth achieves SOTA performance. More importantly, unlike previous SOTA RS S4 methods, our model not only achieves excellent performance but also offers good interpretability. The code is released at https://github.com/wangshanwen001/SemiEarth.

Method: Create synthetic ship dataset with 2D keypoint annotations for multiple ship parts. Train Transformer Neural Network to detect keypoints and estimate 6D pose per part, then integrate estimates using Bayesian fusion.

Result: Achieved position errors of ~0.8% of distance on synthetic data and ~1.0% on flight experiments, demonstrating robustness across lighting conditions.

Conclusion: Transformer-based approach with Bayesian fusion enables accurate 6D pose estimation for ship-UAV relative positioning, supporting autonomous landing applications.

Abstract: This paper introduces a deep transformer network for estimating the relative 6D pose of a Unmanned Aerial Vehicle (UAV) with respect to a ship using monocular images. A synthetic dataset of ship images is created and annotated with 2D keypoints of multiple ship parts. A Transformer Neural Network model is trained to detect these keypoints and estimate the 6D pose of each part. The estimates are integrated using Bayesian fusion. The model is tested on synthetic data and in-situ flight experiments, demonstrating robustness and accuracy in various lighting conditions. The position estimation error is approximately 0.8% and 1.0% of the distance to the ship for the synthetic data and the flight experiments, respectively. The method has potential applications for ship-based autonomous UAV landing and navigation.

Method: Proposes Multi-Hop Visual Chain of Reasoning (VCoR) framework that reformulates registration as progressive reasoning. Each hop integrates Localized Spatial Refinement (LSR) module for feature enrichment and Cross-Reference Attention (CRA) mechanism for iterative refinement. Uses multi-hop strategy to handle large deformations with theoretical bounds on convergence.

Result: Achieves competitive registration accuracy on DIR-Lab 4D CT (lung) and IXI T1-weighted MRI (brain) datasets. Provides rich intermediate visualizations, confidence measures, and uncertainty estimation via deformation field stability across hops.

Conclusion: VCoR presents an interpretable, reliable, and clinically viable unsupervised medical image registration framework by embedding implicit visual reasoning paradigm with built-in transparency and uncertainty estimation.

Abstract: Unsupervised deformable image registration requires aligning complex anatomical structures without reference labels, making interpretability and reliability critical. Existing deep learning methods achieve considerable accuracy but often lack transparency, leading to error drift and reduced clinical trust. We propose a novel Multi-Hop Visual Chain of Reasoning (VCoR) framework that reformulates registration as a progressive reasoning process. Inspired by the iterative nature of clinical decision-making, each visual reasoning hop integrates a Localized Spatial Refinement (LSR) module to enrich feature representations and a Cross-Reference Attention (CRA) mechanism that leads the iterative refinement process, preserving anatomical consistency. This multi-hop strategy enables robust handling of large deformations and produces a transparent sequence of intermediate predictions with a theoretical bound. Beyond accuracy, our framework offers built-in interpretability by estimating uncertainty via the stability and convergence of deformation fields across hops. Extensive evaluations on two challenging public datasets, DIR-Lab 4D CT (lung) and IXI T1-weighted MRI (brain), demonstrate that VCoR achieves competitive registration accuracy while offering rich intermediate visualizations and confidence measures. By embedding an implicit visual reasoning paradigm, we present an interpretable, reliable, and clinically viable unsupervised medical image registration.

Method: Custom CNN architecture with depthwise separable convolutional design optimized for grayscale medical images. Uses CLAHE (Contrast Limited Adaptive Histogram Equalization) and geometric augmentation for preprocessing to handle class imbalance and improve generalization.

Result: Tested on dataset of 5,863 anterior-posterior chest X-rays, achieving high precision pneumonia detection with minimal computational expense compared to generic transfer learning models with redundant parameters.

Conclusion: The proposed unified automated diagnostic model provides fast, precise pneumonia detection in chest X-rays, addressing limitations of manual interpretation while being computationally efficient.

Abstract: Pneumonia has been one of the major causes of morbidities and mortality in the world and the prevalence of this disease is disproportionately high among the pediatric and elderly populations especially in resources trained areas Fast and precise diagnosis is a prerequisite for successful clinical intervention but due to inter observer variation fatigue among experts and a shortage of qualified radiologists traditional approaches that rely on manual interpretation of chest radiographs are frequently constrained To address these problems this paper introduces a unified automated diagnostic model using a custom Convolutional Neural Network CNN that can recognize pneumonia in chest Xray images with high precision and at minimal computational expense In contrast like other generic transfer learning based models which often possess redundant parameters the offered architecture uses a tailor made depth wise separable convolutional design which is optimized towards textural characteristics of grayscale medical images Contrast Limited Adaptive Histogram Equalization CLAHE and geometric augmentation are two significant preprocessing techniques used to ensure that the system does not experience class imbalance and is more likely to generalize The system is tested using a dataset of 5863 anterior posterior chest Xrays.

Method: Uses diffusion models’ intrinsic data reconstruction ability to distinguish ID from OOD samples in latent feature space. Employs multi-layer semantic feature extraction strategy, distorts features with Gaussian noise, and applies diffusion model for feature reconstruction.

Result: Achieves state-of-the-art performance on multiple benchmarks in terms of detection accuracy and speed.

Conclusion: The diffusion-based layer-wise semantic reconstruction approach effectively addresses the reconstruction power vs. compact representation dilemma for unsupervised OOD detection.

Abstract: Unsupervised out-of-distribution (OOD) detection aims to identify out-of-domain data by learning only from unlabeled In-Distribution (ID) training samples, which is crucial for developing a safe real-world machine learning system. Current reconstruction-based methods provide a good alternative approach by measuring the reconstruction error between the input and its corresponding generative counterpart in the pixel/feature space. However, such generative methods face a key dilemma: improving the reconstruction power of the generative model while keeping a compact representation of the ID data. To address this issue, we propose the diffusion-based layer-wise semantic reconstruction approach for unsupervised OOD detection. The innovation of our approach is that we leverage the diffusion model’s intrinsic data reconstruction ability to distinguish ID samples from OOD samples in the latent feature space. Moreover, to set up a comprehensive and discriminative feature representation, we devise a multi-layer semantic feature extraction strategy. By distorting the extracted features with Gaussian noise and applying the diffusion model for feature reconstruction, the separation of ID and OOD samples is implemented according to the reconstruction errors. Extensive experimental results on multiple benchmarks built upon various datasets demonstrate that our method achieves state-of-the-art performance in terms of detection accuracy and speed. Code is available at https://github.com/xbyym/DLSR.

Method: Proposes MFM-Geom, a geometric multimodal foundation model that learns representations from bp-MRI and clinical reports, using symmetric positive definite matrices and Riemannian deep learning to integrate imaging-text representations.

Result: Using only 10% of training data, MFM-Geom outperformed baseline class token embedding-based classification by +8.3%, achieving AUC-PR of 90.67%. Generalization on external dataset achieved AUC-PR of 90.6.

Conclusion: The geometric multimodal foundation model effectively integrates medical imaging and clinical context, demonstrating robust performance with limited data and good generalization capabilities for prostate cancer diagnosis.

Abstract: Prostate cancer (PCa) is one of the most common cancers in men worldwide. Bi-parametric MRI (bp-MRI) and clinical variables are crucial for PCa identification and improving treatment decisions. However, this process is subjective to expert interpretations. Furthermore, most existing computer-aided diagnosis methods focus on imaging-based models, overlooking the clinical context and suffering from data scarcity, limiting their ability to learn robust representations. We propose a geometric multimodal Foundation Model (FM), named MFM-Geom, that learns representations from bp-MRI and clinical reports, encoding visual findings and information from the context of clinical variables. In the representations classification head, the approach leverages symmetric positive definite (SPD) matrices and Riemannian deep learning to integrate imaging-text representations from a biomedical multimodal FM. Using 10% of the training data, MFM-Geom outperformed baseline class token embedding-based classification (+8.3%, AUC-PR of 90.67). Generalization on external dataset confirmed the robustness of fine-tuning biomedical FM, achieving an AUC-PR of 90.6.

Method: A novel two-stage deep learning model: first constructs an initial estimate, then refines fine-grained structural details using high-resolution impervious surface data as guidance to generate the EVAL dataset.

Result: EVAL significantly outperforms state-of-the-art products, showing superior temporal consistency and stronger correlation with socioeconomic indicators.

Conclusion: The EVAL dataset provides a high-quality, extended VIIRS-like nighttime light dataset for China from 1986 to 2024, overcoming limitations of existing products through a novel deep learning approach.

Abstract: Artificial Night-Time Light (NTL) remote sensing is a vital proxy for quantifying the intensity and spatial distribution of human activities. Although the NPP-VIIRS sensor provides high-quality NTL observations, its temporal coverage, which begins in 2012, restricts long-term time-series studies that extend to earlier periods. Current extended VIIRS-like NTL data products suffer from two significant shortcomings: the underestimation of light intensity and the omission of structural details. To overcome these limitations, we present the Extended VIIRS-like Artificial Nighttime Light (EVAL) dataset, a new annual NTL dataset for China spanning from 1986 to 2024. This dataset was generated using a novel two-stage deep learning model designed to address the aforementioned shortcomings. The model first constructs an initial estimate and subsequently refines fine-grained structural details using high-resolution impervious surface data as guidance. Quantitative evaluations demonstrate that EVAL significantly outperforms state-of-the-art products, exhibiting superior temporal consistency and a stronger correlation with socioeconomic indicators.

Method: Introduces Attention Contribution metric (weighting attention probabilities by value vector magnitude) to better identify important visual tokens. Proposes CAPA framework with dual strategy: 1) pruning visual tokens at critical functional transitions using attention contribution, and 2) reducing FFN computation through efficient linear approximations, especially for intermediate layers where image tokens exhibit linear behavior.

Result: CAPA achieves competent efficiency-performance trade-offs with improved robustness across various benchmarks and baselines. Identifies that visual attention sinks consist of Probability Dumps (low contribution, can be pruned) and Structural Anchors (high contribution, essential for performance).

Conclusion: Attention Contribution provides better criterion for visual token selection than traditional attention scores. The proposed CAPA framework effectively reduces computational costs while maintaining performance in LVLMs through targeted pruning and FFN approximation.

Abstract: Efficient inference in Large Vision-Language Models is constrained by the high cost of processing thousands of visual tokens, yet it remains unclear which tokens and computations can be safely removed. While attention scores are commonly used to estimate visual token importance, they are an imperfect proxy for actual contribution. We show that Attention Contribution, which weights attention probabilities by value vector magnitude, provides a more accurate criterion for visual token selection. Our empirical analysis reveals that visual attention sinks are functionally heterogeneous, comprising Probability Dumps with low contribution that can be safely pruned, and Structural Anchors with high contribution essential for maintaining model performance. Further, we identify substantial redundancy in Feed-Forward Networks (FFNs) associated with visual tokens, particularly in intermediate layers where image tokens exhibit linear behavior. Based on our findings, we introduce CAPA (Contribution-Aware Pruning and FFN Approximation), a dual-strategy framework that prunes visual tokens using attention contribution at critical functional transitions and reduces FFN computation through efficient linear approximations. Experiments on various benchmarks across baselines show that CAPA achieves competent efficiency–performance trade-offs with improved robustness.

Method: Systematically evaluate state-of-the-art PFMs on TCGA patch-level datasets and introduce a lightweight adversarial adaptor to remove hospital-related domain information from latent representations.

Result: While disease classification accuracy is largely maintained, the adaptor effectively reduces hospital-specific bias, as confirmed by t-SNE visualizations, establishing a benchmark for assessing cross-hospital robustness in PFMs.

Conclusion: The study provides a practical strategy for enhancing generalization under heterogeneous clinical settings and establishes a benchmark for assessing cross-hospital robustness in pathology foundation models.

Abstract: Pathology foundation models (PFMs) achieve strong performance on diverse histopathology tasks, but their sensitivity to hospital-specific domain shifts remains underexplored. We systematically evaluate state-of-the-art PFMs on TCGA patch-level datasets and introduce a lightweight adversarial adaptor to remove hospital-related domain information from latent representations. Experiments show that, while disease classification accuracy is largely maintained, the adaptor effectively reduces hospital-specific bias, as confirmed by t-SNE visualizations. Our study establishes a benchmark for assessing cross-hospital robustness in PFMs and provides a practical strategy for enhancing generalization under heterogeneous clinical settings. Our code is available at https://github.com/MengRes/pfm_domain_bias.

Method: Developed SANEval benchmark with a scalable pipeline combining LLMs for deep prompt understanding and LLM-enhanced open-vocabulary object detectors to evaluate compositional adherence without vocabulary constraints.

Result: SANEval’s automated evaluations provide more faithful proxy for human assessment, achieving statistically significant Spearman’s rank correlation improvements over existing benchmarks across attribute binding, spatial relations, and numeracy tasks.

Conclusion: SANEval addresses critical limitations in T2I evaluation, offering comprehensive compositional assessment and interpretable feedback to facilitate future research in compositional T2I generation and evaluation.

Abstract: The rapid progress of text-to-image (T2I) models has unlocked unprecedented creative potential, yet their ability to faithfully render complex prompts involving multiple objects, attributes, and spatial relationships remains a significant bottleneck. Progress is hampered by a lack of adequate evaluation methods; current benchmarks are often restricted to closed-set vocabularies, lack fine-grained diagnostic capabilities, and fail to provide the interpretable feedback necessary to diagnose and remedy specific compositional failures. We solve these challenges by introducing SANEval (Spatial, Attribute, and Numeracy Evaluation), a comprehensive benchmark that establishes a scalable new pipeline for open-vocabulary compositional evaluation. SANEval combines a large language model (LLM) for deep prompt understanding with an LLM-enhanced, open-vocabulary object detector to robustly evaluate compositional adherence, unconstrained by a fixed vocabulary. Through extensive experiments on six state-of-the-art T2I models, we demonstrate that SANEval’s automated evaluations provide a more faithful proxy for human assessment; our metric achieves a Spearman’s rank correlation with statistically different results than those of existing benchmarks across tasks of attribute binding, spatial relations, and numeracy. To facilitate future research in compositional T2I generation and evaluation, we will release the SANEval dataset and our open-source evaluation pipeline.

Method: Proposes Contrastive Subspace Clustering (CSC) framework: generates masked views of partially observed inputs, trains deep neural network using SimCLR-style contrastive loss to learn invariant embeddings, then clusters using sparse subspace clustering.

Result: Experiments on six benchmark datasets show CSC consistently outperforms both classical and deep learning baselines, demonstrating strong robustness to missing data and scalability to large datasets.

Conclusion: CSC provides an effective contrastive self-supervised approach for subspace clustering on incomplete data, addressing limitations of existing methods that require fully observed data.

Abstract: Subspace clustering aims to group data points that lie in a union of low-dimensional subspaces and finds wide application in computer vision, hyperspectral imaging, and recommendation systems. However, most existing methods assume fully observed data, limiting their effectiveness in real-world scenarios with missing entries. In this paper, we propose a contrastive self-supervised framework, Contrastive Subspace Clustering (CSC), designed for clustering incomplete data. CSC generates masked views of partially observed inputs and trains a deep neural network using a SimCLR-style contrastive loss to learn invariant embeddings. These embeddings are then clustered using sparse subspace clustering. Experiments on six benchmark datasets show that CSC consistently outperforms both classical and deep learning baselines, demonstrating strong robustness to missing data and scalability to large datasets.

Method: Reformulates panoramic editing directly in ERP domain using a generate-then-edit paradigm with controllable panoramic generation for paired data synthesis. Introduces geometry-aware learning with position-aware shape supervision and panoramic priors through progressive training.

Result: Achieves superior geometric consistency, editing fidelity, and text controllability compared to SOTA methods on new benchmark PEBench, enabling coherent and flexible 360° visual world creation.

Conclusion: World-Shaper bridges panoramic generation and editing within a unified editing-centric design, effectively addressing geometric distortion and enabling high-quality 360° visual editing.

Abstract: Being able to edit panoramic images is crucial for creating realistic 360° visual experiences. However, existing perspective-based image editing methods fail to model the spatial structure of panoramas. Conventional cube-map decompositions attempt to overcome this problem but inevitably break global consistency due to their mismatch with spherical geometry. Motivated by this insight, we reformulate panoramic editing directly in the equirectangular projection (ERP) domain and present World-Shaper, a unified geometry-aware framework that bridges panoramic generation and editing within a single editing-centric design. To overcome the scarcity of paired data, we adopt a generate-then-edit paradigm, where controllable panoramic generation serves as an auxiliary stage to synthesize diverse paired examples for supervised editing learning. To address geometric distortion, we introduce a geometry-aware learning strategy that explicitly enforces position-aware shape supervision and implicitly internalizes panoramic priors through progressive training. Extensive experiments on our new benchmark, PEBench, demonstrate that our method achieves superior geometric consistency, editing fidelity, and text controllability compared to SOTA methods, enabling coherent and flexible 360° visual world creation with unified editing control. Code, model, and data will be released at our project page: https://world-shaper-project.github.io/

Method: Uses pretrained image-to-video diffusion model with text control to exploit temporal priors from videos for object consistency. Proposes novel data curation strategy generating synthetic sequences where randomly placed objects smoothly move to target positions, aligning with video model’s temporal priors during training.

Result: PLACID surpasses state-of-the-art methods in multi-object compositing, achieving superior identity, background, and color preservation with less omitted objects and visually appealing results, as demonstrated through extensive quantitative evaluations and user studies.

Conclusion: PLACID effectively bridges the gap in studio-level multi-object compositing by leveraging video temporal priors and synthetic training data, enabling coherent layouts guided by text while preserving object identities and background details.

Abstract: Recent advances in generative AI have dramatically improved photorealistic image synthesis, yet they fall short for studio-level multi-object compositing. This task demands simultaneous (i) near-perfect preservation of each item’s identity, (ii) precise background and color fidelity, (iii) layout and design elements control, and (iv) complete, appealing displays showcasing all objects. However, current state-of-the-art models often alter object details, omit or duplicate objects, and produce layouts with incorrect relative sizing or inconsistent item presentations. To bridge this gap, we introduce PLACID, a framework that transforms a collection of object images into an appealing multi-object composite. Our approach makes two main contributions. First, we leverage a pretrained image-to-video (I2V) diffusion model with text control to preserve objects consistency, identities, and background details by exploiting temporal priors from videos. Second, we propose a novel data curation strategy that generates synthetic sequences where randomly placed objects smoothly move to their target positions. This synthetic data aligns with the video model’s temporal priors during training. At inference, objects initialized at random positions consistently converge into coherent layouts guided by text, with the final frame serving as the composite image. Extensive quantitative evaluations and user studies demonstrate that PLACID surpasses state-of-the-art methods in multi-object compositing, achieving superior identity, background, and color preservation, with less omitted objects and visually appealing results.

Method: Proposes an inference-time method that identifies and removes unstable latent tokens before they are reused for conditioning. Unstable tokens are defined as those whose representations deviate significantly from previously generated batches, indicating potential corruption or semantic drift. The method removes corrupted latent tokens from the auto-regressive context without modifying spatial regions, model architecture, training procedures, or leaving latent space.

Result: The method significantly improves long-horizon temporal consistency in auto-regressive video generation by preventing unreliable latent information from influencing future generation steps.

Conclusion: Temporal drift in auto-reggressive video generation can be effectively mitigated through a simple inference-time approach that identifies and removes corrupted latent tokens, improving long-horizon consistency without architectural or training modifications.

Abstract: Auto-regressive video generation enables long video synthesis by iteratively conditioning each new batch of frames on previously generated content. However, recent work has shown that such pipelines suffer from severe temporal drift, where errors accumulate and amplify over long horizons. We hypothesize that this drift does not primarily stem from insufficient model capacity, but rather from inference-time error propagation. Specifically, we contend that drift arises from the uncontrolled reuse of corrupted latent conditioning tokens during auto-regressive inference. To correct this accumulation of errors, we propose a simple, inference-time method that mitigates temporal drift by identifying and removing unstable latent tokens before they are reused for conditioning. For this purpose, we define unstable tokens as latent tokens whose representations deviate significantly from those of the previously generated batch, indicating potential corruption or semantic drift. By explicitly removing corrupted latent tokens from the auto-regressive context, rather than modifying entire spatial regions or model parameters, our method prevents unreliable latent information from influencing future generation steps. As a result, it significantly improves long-horizon temporal consistency without modifying the model architecture, training procedure, or leaving latent space.

Method: Created TimeBlind benchmark using minimal-pairs paradigm: video pairs share identical static content but differ only in temporal structure. Categorizes temporal understanding into three levels: atomic events, event properties, and event interdependencies. Uses complementary questions to neutralize language priors.

Result: Evaluated 20+ state-of-the-art MLLMs on 600 instances (2400 video-question pairs). Best MLLM achieved only 48.2% instance accuracy (correctly distinguishing both videos in a pair), far below human performance of 98.2%.

Conclusion: Current MLLMs rely heavily on static visual shortcuts rather than genuine temporal logic. TimeBlind serves as a vital diagnostic tool for advancing video understanding capabilities in multimodal models.

Abstract: Fine-grained spatio-temporal understanding is essential for video reasoning and embodied AI. Yet, while Multimodal Large Language Models (MLLMs) master static semantics, their grasp of temporal dynamics remains brittle. We present TimeBlind, a diagnostic benchmark for compositional spatio-temporal understanding. Inspired by cognitive science, TimeBlind categorizes fine-grained temporal understanding into three levels: recognizing atomic events, characterizing event properties, and reasoning about event interdependencies. Unlike benchmarks that conflate recognition with temporal reasoning, TimeBlind leverages a minimal-pairs paradigm: video pairs share identical static visual content but differ solely in temporal structure, utilizing complementary questions to neutralize language priors. Evaluating over 20 state-of-the-art MLLMs (e.g., GPT-5, Gemini 3 Pro) on 600 curated instances (2400 video-question pairs), reveals that the Instance Accuracy (correctly distinguishing both videos in a pair) of the best performing MLLM is only 48.2%, far below the human performance (98.2%). These results demonstrate that even frontier models rely heavily on static visual shortcuts rather than genuine temporal logic, positioning TimeBlind as a vital diagnostic tool for next-generation video understanding. Dataset and code are available at https://baiqi-li.github.io/timeblind_project/ .

Method: Uses Bayes Decision Theory as a unifying framework to analyze and compare Bayesian approaches (conceptually attractive, cognitive science-aligned) and deep neural networks (practical success, biologically inspired).

Result: Presents a theoretical lens that captures key computer vision concepts, relates strengths/weaknesses of Bayesian vs. deep learning approaches, and identifies limitations of BDT for guiding future integration.

Conclusion: Bayes Decision Theory provides a valuable framework for understanding computer vision approaches, but its limitations point toward richer frameworks that can combine Bayesian and deep learning methods.

Abstract: This document presents an introduction to computer vision, and its relationship to Cognitive Science, from the perspective of Bayes Decision Theory (Berger 1985). Computer vision is a vast and complex field, so this overview has a narrow scope and provides a theoretical lens which captures many key concepts. BDT is rich enough to include two different approaches: (i) the Bayesian viewpoint, which gives a conceptually attractive framework for vision with concepts that resonate with Cognitive Science (Griffiths et al., 2024), and (ii) the Deep Neural Network approach whose successes in the real world have made Computer Vision into a trillion-dollar industry and which is motivated by the hierarchical structure of the visual ventral stream. The BDT framework relates and captures the strengths and weakness of these two approaches and, by discussing the limitations of BDT, points the way to how they can be combined in a richer framework.

Method: LogicGaze is a benchmark framework curated from 40,000 video segments from ShareGPT4Video and Flickr30k imagery. It integrates causal sequences with visually contradictory but linguistically plausible perturbations, forcing models to verify each reasoning step. The evaluation uses a tripartite protocol: Causal Validation, Grounded Narrative Synthesis, and Perturbation Rejection.

Result: The benchmark exposes significant vulnerabilities in state-of-the-art VLMs like Qwen2.5-VL-72B, revealing their limitations in validating sequential causal chains against visual inputs.

Conclusion: LogicGaze advocates for more robust and trustworthy multimodal reasoning by providing a framework to evaluate and improve VLMs’ ability to ground reasoning in visual evidence, with all resources made publicly available.

Abstract: While sequential reasoning enhances the capability of Vision-Language Models (VLMs) to execute complex multimodal tasks, their reliability in grounding these reasoning chains within actual visual evidence remains insufficiently explored. We introduce LogicGaze, a novel benchmark framework designed to rigorously interrogate whether VLMs can validate sequential causal chains against visual inputs, specifically targeting the pervasive issue of hallucination. Curated from 40,000 video segments from ShareGPT4Video and a subset of Flickr30k imagery, LogicGaze integrates causal sequences with visually contradictory yet linguistically plausible perturbations, compelling models to verify the authenticity of each reasoning step. Our tripartite evaluation protocol - Causal Validation, Grounded Narrative Synthesis, and Perturbation Rejection - exposes significant vulnerabilities in state-of-the-art VLMs such as Qwen2.5-VL-72B. LogicGaze advocates for robust, trustworthy multimodal reasoning, with all resources publicly available in an anonymized repository.

Method: SAM2CT extends SAM2 with a prompt encoder supporting arrow and line inputs, and introduces Memory-Conditioned Memories (MCM) for 3D medical volumes. It converts sparse radiologist annotations into full 3D segmentations through promptable segmentation.

Result: SAM2CT outperforms existing promptable segmentation models, achieving Dice scores of 0.649 (arrow prompts) and 0.757 (line prompts). On clinical PACS data, it generates clinically acceptable segmentations in 87% of cases and shows strong zero-shot performance on Emergency Department findings.

Conclusion: Large-scale mining of historical GSPS annotations using promptable segmentation models like SAM2CT represents a scalable approach for generating 3D CT segmentation datasets, potentially revolutionizing medical imaging dataset creation.

Abstract: The development of machine learning models for CT imaging depends on the availability of large, high-quality, and diverse annotated datasets. Although large volumes of CT images and reports are readily available in clinical picture archiving and communication systems (PACS), 3D segmentations of critical findings are costly to obtain, typically requiring extensive manual annotation by radiologists. On the other hand, it is common for radiologists to provide limited annotations of findings during routine reads, such as line measurements and arrows, that are often stored in PACS as GSPS objects. We posit that these sparse annotations can be extracted along with CT volumes and converted into 3D segmentations using promptable segmentation models, a paradigm we term Opportunistic Promptable Segmentation. To enable this paradigm, we propose SAM2CT, the first promptable segmentation model designed to convert radiologist annotations into 3D segmentations in CT volumes. SAM2CT builds upon SAM2 by extending the prompt encoder to support arrow and line inputs and by introducing Memory-Conditioned Memories (MCM), a memory encoding strategy tailored to 3D medical volumes. On public lesion segmentation benchmarks, SAM2CT outperforms existing promptable segmentation models and similarly trained baselines, achieving Dice similarity coefficients of 0.649 for arrow prompts and 0.757 for line prompts. Applying the model to pre-existing GSPS annotations from a clinical PACS (N = 60), SAM2CT generates 3D segmentations that are clinically acceptable or require only minor adjustments in 87% of cases, as scored by radiologists. Additionally, SAM2CT demonstrates strong zero-shot performance on select Emergency Department findings. These results suggest that large-scale mining of historical GSPS annotations represents a promising and scalable approach for generating 3D CT segmentation datasets.

Method: Uses predictive sensitivity quantification based on ensemble of models to capture model disagreement and inference variability. Evaluates perception performance under adverse driving scenarios in simulated and real-world conditions. Proposes notional architecture for perception assessment and develops criterion based on AV stopping distance at stop signs on varying road surfaces. Tests five state-of-the-art computer vision models: YOLO (v8-v9), DETR50, DETR101, and RT-DETR.

Result: Diminished lighting conditions (fog, low sun altitude) have greatest impact on perception model performance. Adversarial road conditions like occlusions increase perception sensitivity, and performance drops further with combination of adversarial conditions and inclement weather. Greater distance to roadway objects leads to greater impact on perception performance and diminished robustness.

Conclusion: Perception systems for automated vehicles show significant vulnerability to adverse environmental conditions, particularly lighting and combined adversarial factors. Ensemble-based sensitivity quantification provides effective assessment methodology, revealing critical robustness limitations that must be addressed for reliable automated driving.

Abstract: The viability of automated driving is heavily dependent on the performance of perception systems to provide real-time accurate and reliable information for robust decision-making and maneuvers. These systems must perform reliably not only under ideal conditions, but also when challenged by natural and adversarial driving factors. Both of these types of interference can lead to perception errors and delays in detection and classification. Hence, it is essential to assess the robustness of the perception systems of automated vehicles (AVs) and explore strategies for making perception more reliable. We approach this problem by evaluating perception performance using predictive sensitivity quantification based on an ensemble of models, capturing model disagreement and inference variability across multiple models, under adverse driving scenarios in both simulated environments and real-world conditions. A notional architecture for assessing perception performance is proposed. A perception assessment criterion is developed based on an AV’s stopping distance at a stop sign on varying road surfaces, such as dry and wet asphalt, and vehicle speed. Five state-of-the-art computer vision models are used, including YOLO (v8-v9), DEtection TRansformer (DETR50, DETR101), Real-Time DEtection TRansformer (RT-DETR)in our experiments. Diminished lighting conditions, e.g., resulting from the presence of fog and low sun altitude, have the greatest impact on the performance of the perception models. Additionally, adversarial road conditions such as occlusions of roadway objects increase perception sensitivity and model performance drops when faced with a combination of adversarial road conditions and inclement weather conditions. Also, it is demonstrated that the greater the distance to a roadway object, the greater the impact on perception performance, hence diminished perception robustness.

Method: Proposes Synergistic Neural Agents Network (SynerNet) with four specialized computational units: visual perception, linguistic context, nominal embedding, and global coordination. These units collaborate through structured message-propagation protocol to rectify modality disparities. Includes multi-agent latent space nomenclature acquisition, semantic context-interchange algorithm for few-shot adaptation, and adaptive dynamic equilibrium mechanism.

Result: Empirical evaluations on VISTA-Beyond benchmark show substantial performance improvements in both few-shot and zero-shot scenarios, with precision gains ranging from 1.2% to 5.4% across diverse domains.

Conclusion: SynerNet effectively mitigates cross-modal alignment degeneration in VLMs for OOD concepts through synergistic multi-agent collaboration, enhancing generalization capabilities to novel domains.

Abstract: This manuscript presents a pioneering Synergistic Neural Agents Network (SynerNet) framework designed to mitigate the phenomenon of cross-modal alignment degeneration in Vision-Language Models (VLMs) when encountering Out-of-Distribution (OOD) concepts. Specifically, four specialized computational units - visual perception, linguistic context, nominal embedding, and global coordination - collaboratively rectify modality disparities via a structured message-propagation protocol. The principal contributions encompass a multi-agent latent space nomenclature acquisition framework, a semantic context-interchange algorithm for enhanced few-shot adaptation, and an adaptive dynamic equilibrium mechanism. Empirical evaluations conducted on the VISTA-Beyond benchmark demonstrate that SynerNet yields substantial performance augmentations in both few-shot and zero-shot scenarios, exhibiting precision improvements ranging from 1.2% to 5.4% across a diverse array of domains.

Method: Proposes MAD-RAG, a training-free intervention that uses dual-question formulation to decouple visual grounding from context integration, combined with attention mixing to preserve image-conditioned evidence.

Result: Extensive experiments on OK-VQA, E-VQA, and InfoSeek show MAD-RAG consistently outperforms existing baselines across different model families with absolute gains up to 4.76%, 9.20%, and 6.18% over vanilla RAG, rectifying up to 74.68% of failure cases.

Conclusion: MAD-RAG effectively mitigates attention distraction in RAG for vision-language models, improving performance on knowledge-based VQA tasks with negligible computational overhead.

Abstract: While Retrieval-Augmented Generation (RAG) is one of the dominant paradigms for enhancing Large Vision-Language Models (LVLMs) on knowledge-based VQA tasks, recent work attributes RAG failures to insufficient attention towards the retrieved context, proposing to reduce the attention allocated to image tokens. In this work, we identify a distinct failure mode that previous study overlooked: Attention Distraction (AD). When the retrieved context is sufficient (highly relevant or including the correct answer), the retrieved text suppresses the visual attention globally, and the attention on image tokens shifts away from question-relevant regions. This leads to failures on questions the model could originally answer correctly without the retrieved text. To mitigate this issue, we propose MAD-RAG, a training-free intervention that decouples visual grounding from context integration through a dual-question formulation, combined with attention mixing to preserve image-conditioned evidence. Extensive experiments on OK-VQA, E-VQA, and InfoSeek demonstrate that MAD-RAG consistently outperforms existing baselines across different model families, yielding absolute gains of up to 4.76%, 9.20%, and 6.18% over the vanilla RAG baseline. Notably, MAD-RAG rectifies up to 74.68% of failure cases with negligible computational overhead.

Method: AdaFuse formulates multimodal fusion as a sequential decision process using reinforcement learning. A policy network iteratively decides whether to incorporate additional modalities or proceed to prediction based on already acquired information, enabling early termination when sufficient information is available.

Result: Achieved highest AUC (0.762) on NLST dataset compared to single-modality baselines (0.732), fixed fusion (0.759), and adaptive baselines DynMM (0.754) and MoE (0.742), while using fewer FLOPs than triple-modality methods.

Conclusion: Demonstrates RL’s potential for personalized multimodal fusion in medical imaging, shifting from uniform fusion toward adaptive diagnostic pipelines that learn when to consult additional modalities.

Abstract: Multimodal fusion has emerged as a promising paradigm for disease diagnosis and prognosis, integrating complementary information from heterogeneous data sources such as medical images, clinical records, and radiology reports. However, existing fusion methods process all available modalities through the network, either treating them equally or learning to assign different contribution weights, leaving a fundamental question unaddressed: for a given patient, should certain modalities be used at all? We present AdaFuse, an adaptive multimodal fusion framework that leverages reinforcement learning (RL) to learn patient-specific modality selection and fusion strategies for lung cancer risk prediction. AdaFuse formulates multimodal fusion as a sequential decision process, where the policy network iteratively decides whether to incorporate an additional modality or proceed to prediction based on the information already acquired. This sequential formulation enables the model to condition each selection on previously observed modalities and terminate early when sufficient information is available, rather than committing to a fixed subset upfront. We evaluate AdaFuse on the National Lung Screening Trial (NLST) dataset. Experimental results demonstrate that AdaFuse achieves the highest AUC (0.762) compared to the best single-modality baseline (0.732), the best fixed fusion strategy (0.759), and adaptive baselines including DynMM (0.754) and MoE (0.742), while using fewer FLOPs than all triple-modality methods. Our work demonstrates the potential of reinforcement learning for personalized multimodal fusion in medical imaging, representing a shift from uniform fusion strategies toward adaptive diagnostic pipelines that learn when to consult additional modalities and when existing information suffices for accurate prediction.

Method: Proposes MASC, a reinforcement learning framework using Proximal Policy Optimization (PPO) agent to select k-space phase-encoding lines under limited acquisition budget. Uses physics-based simulation to create paired dataset with/without metal implants. Combines U-Net-based MAR network with acquisition policy in end-to-end training.

Result: MASC’s learned policies outperform conventional sampling strategies. End-to-end training improves performance compared to using frozen pre-trained MAR network. Cross-dataset experiments on FastMRI with physics-based simulation confirm generalization to realistic clinical MRI data.

Conclusion: Joint optimization of acquisition and artifact correction is beneficial for MRI with metal implants. The unified RL framework effectively learns sampling policies that maximize reconstruction quality while reducing metal artifacts.

Abstract: Metal implants in MRI cause severe artifacts that degrade image quality and hinder clinical diagnosis. Traditional approaches address metal artifact reduction (MAR) and accelerated MRI acquisition as separate problems. We propose MASC, a unified reinforcement learning framework that jointly optimizes metal-aware k-space sampling and artifact correction for accelerated MRI. To enable supervised training, we construct a paired MRI dataset using physics-based simulation, generating k-space data and reconstructions for phantoms with and without metal implants. This paired dataset provides simulated 3D MRI scans with and without metal implants, where each metal-corrupted sample has an exactly matched clean reference, enabling direct supervision for both artifact reduction and acquisition policy learning. We formulate active MRI acquisition as a sequential decision-making problem, where an artifact-aware Proximal Policy Optimization (PPO) agent learns to select k-space phase-encoding lines under a limited acquisition budget. The agent operates on undersampled reconstructions processed through a U-Net-based MAR network, learning patterns that maximize reconstruction quality. We further propose an end-to-end training scheme where the acquisition policy learns to select k-space lines that best support artifact removal while the MAR network simultaneously adapts to the resulting undersampling patterns. Experiments demonstrate that MASC’s learned policies outperform conventional sampling strategies, and end-to-end training improves performance compared to using a frozen pre-trained MAR network, validating the benefit of joint optimization. Cross-dataset experiments on FastMRI with physics-based artifact simulation further confirm generalization to realistic clinical MRI data. The code and models of MASC have been made publicly available: https://github.com/hrlblab/masc

Method: ReLAPSe reformulates concept restoration as a reinforcement learning problem using Reinforcement Learning with Verifiable Rewards (RLVR), leveraging the diffusion model’s noise prediction loss as model-intrinsic feedback to train an agent for textual prompt manipulation.

Result: Achieves efficient, near-real-time recovery of fine-grained identities and styles across multiple state-of-the-art unlearning methods, providing scalable red-teaming capabilities for unlearned diffusion models.

Conclusion: ReLAPSe pioneers the shift from per-instance optimization to global policy learning, offering a scalable tool for rigorous security evaluation of machine unlearning in text-to-image diffusion models.

Abstract: Machine unlearning is a key defense mechanism for removing unauthorized concepts from text-to-image diffusion models, yet recent evidence shows that latent visual information often persists after unlearning. Existing adversarial approaches for exploiting this leakage are constrained by fundamental limitations: optimization-based methods are computationally expensive due to per-instance iterative search. At the same time, reasoning-based and heuristic techniques lack direct feedback from the target model’s latent visual representations. To address these challenges, we introduce ReLAPSe, a policy-based adversarial framework that reformulates concept restoration as a reinforcement learning problem. ReLAPSe trains an agent using Reinforcement Learning with Verifiable Rewards (RLVR), leveraging the diffusion model’s noise prediction loss as a model-intrinsic and verifiable feedback signal. This closed-loop design directly aligns textual prompt manipulation with latent visual residuals, enabling the agent to learn transferable restoration strategies rather than optimizing isolated prompts. By pioneering the shift from per-instance optimization to global policy learning, ReLAPSe achieves efficient, near-real-time recovery of fine-grained identities and styles across multiple state-of-the-art unlearning methods, providing a scalable tool for rigorous red-teaming of unlearned diffusion models. Some experimental evaluations involve sensitive visual concepts, such as nudity. Code is available at https://github.com/gmum/ReLaPSe

Method: Proposes a comparative learning framework that models human pairwise judgments between captions. Uses VICR dataset with ResNet-50 for visual features and MiniLM for text features. Trains both regression (direct rating) and comparative learning models, comparing their performance.

Result: Regression model achieves better performance (Pearson’s ρ: 0.7609, Spearman’s r_s: 0.7089), but comparative learning model steadily improves with more data and approaches regression baseline. Human evaluation shows comparative annotation yields faster results and greater inter-annotator agreement.

Conclusion: Comparative learning can effectively model human preferences for image caption evaluation while significantly reducing annotation costs, making it a practical alternative to direct rating approaches.

Abstract: Rating the accuracy of captions in describing images is time-consuming and subjective for humans. In contrast, it is often easier for people to compare two captions and decide which one better matches a given image. In this work, we propose a machine learning framework that models such comparative judgments instead of direct ratings. The model can then be applied to rank unseen image-caption pairs in the same way as a regression model trained on direct ratings. Using the VICR dataset, we extract visual features with ResNet-50 and text features with MiniLM, then train both a regression model and a comparative learning model. While the regression model achieves better performance (Pearson’s $ρ$: 0.7609 and Spearman’s $r_s$: 0.7089), the comparative learning model steadily improves with more data and approaches the regression baseline. In addition, a small-scale human evaluation study comparing absolute rating, pairwise comparison, and same-image comparison shows that comparative annotation yields faster results and has greater agreement among human annotators. These results suggest that comparative learning can effectively model human preferences while significantly reducing the cost of human annotations.

Method: Comprehensive experimental comparison of YOLOv5 (one-stage detector) and Faster R-CNN (two-stage detector) evaluated on diverse real/synthetic datasets using metrics like mean Average Precision (mAP), recall, and inference speed.

Result: YOLOv5 shows superior mAP, recall, and training efficiency, especially with larger datasets and higher resolutions. Faster R-CNN excels at detecting small distant objects and performs better in challenging lighting conditions.

Conclusion: Both models have trade-offs: YOLOv5 offers better overall performance and efficiency, while Faster R-CNN provides advantages for specific challenging scenarios like small object detection and poor lighting conditions in autonomous driving.

Abstract: Object detection is a crucial component in autonomous vehicle systems. It enables the vehicle to perceive and understand its environment by identifying and locating various objects around it. By utilizing advanced imaging and deep learning techniques, autonomous vehicle systems can rapidly and accurately identify objects based on their features. Different deep learning methods vary in their ability to accurately detect and classify objects in autonomous vehicle systems. Selecting the appropriate method significantly impacts system performance, robustness, and efficiency in real-world driving scenarios. While several generic deep learning architectures like YOLO, SSD, and Faster R-CNN have been proposed, guidance on their suitability for specific autonomous driving applications is often limited. The choice of method affects detection accuracy, processing speed, environmental robustness, sensor integration, scalability, and edge case handling. This study provides a comprehensive experimental analysis comparing two prominent object detection models: YOLOv5 (a one-stage detector) and Faster R-CNN (a two-stage detector). Their performance is evaluated on a diverse dataset combining real and synthetic images, considering various metrics including mean Average Precision (mAP), recall, and inference speed. The findings reveal that YOLOv5 demonstrates superior performance in terms of mAP, recall, and training efficiency, particularly as dataset size and image resolution increase. However, Faster R-CNN shows advantages in detecting small, distant objects and performs well in challenging lighting conditions. The models’ behavior is also analyzed under different confidence thresholds and in various real-world scenarios, providing insights into their applicability for autonomous driving systems.

Method: Uses dynamic 4D-CTA head scans with multiple time points to subtract bone and soft tissue, enhancing vessel visualization. Trains deep learning models (nnUNet) on ground truth annotations, using same segmentation across multiple phases to effectively enlarge dataset 4-5x and induce robustness to contrast phases.

Result: Achieved significantly better segmentations than similarly-sized datasets, with average mDC of 0.846 for arteries and 0.957 for veins. Low error margins (aDHD of 0.304 mm for arteries, 0.078 for veins) and high topology sensitivity (tSens of 0.877 for arteries, 0.974 for veins).

Conclusion: The proposed methodology successfully reduces manual annotation effort while achieving excellent accuracy in brain vessel segmentation, with publicly available code and model weights for community use.

Abstract: In this study, we develop a novel methodology for annotating the brain vasculature using dynamic 4D-CTA head scans. By using multiple time points from dynamic CTA acquisitions, we subtract bone and soft tissue to enhance the visualization of arteries and veins, reducing the effort required to obtain manual annotations of brain vessels. We then train deep learning models on our ground truth annotations by using the same segmentation for multiple phases from the dynamic 4D-CTA collection, effectively enlarging our dataset by 4 to 5 times and inducing robustness to contrast phases. In total, our dataset comprises 110 training images from 25 patients and 165 test images from 14 patients. In comparison with two similarly-sized datasets for CTA-based brain vessel segmentation, a nnUNet model trained on our dataset can achieve significantly better segmentations across all vascular regions, with an average mDC of 0.846 for arteries and 0.957 for veins in the TopBrain dataset. Furthermore, metrics such as average directed Hausdorff distance (adHD) and topology sensitivity (tSens) reflected similar trends: using our dataset resulted in low error margins (aDHD of 0.304 mm for arteries and 0.078 for veins) and high sensitivity (tSens of 0.877 for arteries and 0.974 for veins), indicating excellent accuracy in capturing vessel morphology. Our code and model weights are available online: https://github.com/alceballosa/robust-vessel-segmentation

Method: Proposes cross-native-translated evaluation using Flickr30K with native Brazilian Portuguese captions vs. automatically translated captions. Uses cross-context approach (train on one, test on other), attention maps for interpretation, and CLIP-Score for image-text alignment evaluation.

Result: Swin-DistilBERTimbau consistently outperforms other models with strong generalization. ViTucano (Brazilian Portuguese VLM) surpasses larger multilingual models in text metrics, while GPT-4 achieves highest CLIP-Score. Attention analysis reveals systematic biases including gender misclassification and spatial inconsistencies.

Conclusion: The study provides valuable insights into cross-lingual vision-language model evaluation for low-resource languages, highlighting performance differences between native and translated datasets, and revealing systematic biases in model attention mechanisms.

Abstract: Image captioning (IC) refers to the automatic generation of natural language descriptions for images, with applications ranging from social media content generation to assisting individuals with visual impairments. While most research has been focused on English-based models, low-resource languages such as Brazilian Portuguese face significant challenges due to the lack of specialized datasets and models. Several studies create datasets by automatically translating existing ones to mitigate resource scarcity. This work addresses this gap by proposing a cross-native-translated evaluation of Transformer-based vision and language models for Brazilian Portuguese IC. We use a version of Flickr30K comprised of captions manually created by native Brazilian Portuguese speakers and compare it to a version with captions automatically translated from English to Portuguese. The experiments include a cross-context approach, where models trained on one dataset are tested on the other to assess the translation impact. Additionally, we incorporate attention maps for model inference interpretation and use the CLIP-Score metric to evaluate the image-description alignment. Our findings show that Swin-DistilBERTimbau consistently outperforms other models, demonstrating strong generalization across datasets. ViTucano, a Brazilian Portuguese pre-trained VLM, surpasses larger multilingual models (GPT-4o, LLaMa 3.2 Vision) in traditional text-based evaluation metrics, while GPT-4 models achieve the highest CLIP-Score, highlighting improved image-text alignment. Attention analysis reveals systematic biases, including gender misclassification, object enumeration errors, and spatial inconsistencies. The datasets and the models generated and analyzed during the current study are available in: https://github.com/laicsiifes/transformer-caption-ptbr.

Method: Extract ResNet-50 CNN features from painting images, develop deep neural network regression model and dual-branch pairwise comparison model; explore four research questions comparing regression vs. comparative learning, individual preference prediction, and annotation cost trade-offs.

Result: Deep regression model outperforms baseline by up to 328% in R²; comparative model approaches regression performance without direct rating values; individual preference prediction remains challenging; comparative judgments require 60% less annotation time per item.

Conclusion: Pairwise comparative learning is effective for aesthetic judgment modeling, offering practical utility when direct ratings are unavailable and superior annotation efficiency for large-scale preference modeling.

Abstract: Modeling human aesthetic judgments in visual art presents significant challenges due to individual preference variability and the high cost of obtaining labeled data. To reduce cost of acquiring such labels, we propose to apply a comparative learning framework based on pairwise preference assessments rather than direct ratings. This approach leverages the Law of Comparative Judgment, which posits that relative choices exhibit less cognitive burden and greater cognitive consistency than direct scoring. We extract deep convolutional features from painting images using ResNet-50 and develop both a deep neural network regression model and a dual-branch pairwise comparison model. We explored four research questions: (RQ1) How does the proposed deep neural network regression model with CNN features compare to the baseline linear regression model using hand-crafted features? (RQ2) How does pairwise comparative learning compare to regression-based prediction when lacking access to direct rating values? (RQ3) Can we predict individual rater preferences through within-rater and cross-rater analysis? (RQ4) What is the annotation cost trade-off between direct ratings and comparative judgments in terms of human time and effort? Our results show that the deep regression model substantially outperforms the baseline, achieving up to $328%$ improvement in $R^2$. The comparative model approaches regression performance despite having no access to direct rating values, validating the practical utility of pairwise comparisons. However, predicting individual preferences remains challenging, with both within-rater and cross-rater performance significantly lower than average rating prediction. Human subject experiments reveal that comparative judgments require $60%$ less annotation time per item, demonstrating superior annotation efficiency for large-scale preference modeling.

Method: Proposes 3DGS²-TR, a second-order optimizer that approximates curvature using only the diagonal of the Hessian matrix via Hutchinson’s method. It’s fully matrix-free with O(n) complexity. Introduces parameter-wise trust-region technique based on squared Hellinger distance to regularize Gaussian parameter updates for stable optimization despite strong nonlinearity in 3DGS rasterization.

Result: Achieves better reconstruction quality on standard datasets using 50% fewer training iterations compared to ADAM, with less than 1GB peak GPU memory overhead (17% more than ADAM, 85% less than 3DGS-LM). Enables scalability to very large scenes and potentially distributed training.

Conclusion: 3DGS²-TR provides an efficient second-order optimization approach for 3DGS that balances training speed, reconstruction quality, and memory efficiency, making it suitable for large-scale scene reconstruction applications.

Abstract: We propose 3DGS$^2$-TR,a second-order optimizer for accelerating the scene training problem in 3D Gaussian Splatting (3DGS). Unlike existing second-order approaches that rely on explicit or dense curvature representations, such as 3DGS-LM (Höllein et al., 2025) or 3DGS2 (Lan et al., 2025), our method approximates curvature using only the diagonal of the Hessian matrix, efficiently via Hutchinson’s method. Our approach is fully matrix-free and has the same complexity as ADAM (Kingma, 2024), $O(n)$ in both computation and memory costs. To ensure stable optimization in the presence of strong nonlinearity in the 3DGS rasterization process, we introduce a parameter-wise trust-region technique based on the squared Hellinger distance, regularizing updates to Gaussian parameters. Under identical parameter initialization and without densification, 3DGS$^2$-TR is able to achieve better reconstruction quality on standard datasets, using 50% fewer training iterations compared to ADAM, while incurring less than 1GB of peak GPU memory overhead (17% more than ADAM and 85% less than 3DGS-LM), enabling scalability to very large scenes and potentially to distributed training settings.

Method: Created synthetic dataset pipeline converting text scenarios to (image, scene graph, ground truth) triples using LLMs as scene graph architects and image generation models as renderers. Proposed scene-graph-guided alignment to bridge perceptual gaps by translating visual inputs to structured scene graphs.

Result: VLMs perform well with textual scene graphs but degrade substantially in visual-only settings. Scene-graph-guided alignment improves hazard detection performance in visual-only settings by better aligning visual inputs with VLM reasoning.

Conclusion: VLMs need structured scene understanding for effective lab safety monitoring; bridging visual perception with structured reasoning through scene graph translation improves performance in real-world visual settings.

Abstract: Laboratories are prone to severe injuries from minor unsafe actions, yet continuous safety monitoring – beyond mandatory pre-lab safety training – is limited by human availability. Vision language models (VLMs) offer promise for autonomous laboratory safety monitoring, but their effectiveness in realistic settings is unclear due to the lack of visual evaluation data, as most safety incidents are documented primarily as unstructured text. To address this gap, we first introduce a structured data generation pipeline that converts textual laboratory scenarios into aligned triples of (image, scene graph, ground truth), using large language models as scene graph architects and image generation models as renderers. Our experiments on the synthetic dataset of 1,207 samples across 362 unique scenarios and seven open- and closed-source models show that VLMs perform effectively given textual scene graph, but degrade substantially in visual-only settings indicating difficulty in extracting structured object relationships directly from pixels. To overcome this, we propose a post-training context-engineering approach, scene-graph-guided alignment, to bridge perceptual gaps in VLMs by translating visual inputs into structured scene graphs better aligned with VLM reasoning, improving hazard detection performance in visual only settings.

Method: Three-stage approach: 1) Decompose harmful query into benign sub-queries, 2) Select maximally irrelevant distraction queries, 3) Present all queries as image grid with sub-queries positioned in middle, exploiting OCR processing.

Result: Successfully circumvents safety alignment of state-of-the-art LVLMs by bypassing text-based filters and inducing distractions that prevent safety protocols from linking scattered sub-queries.

Conclusion: Exposes critical vulnerability in LVLMs’ OCR capabilities that are not robust to dispersed, multi-image adversarial inputs, highlighting need for defenses for fragmented multimodal inputs.

Abstract: Large Vision-Language Models (LVLMs) are increasingly equipped with robust safety safeguards to prevent responses to harmful or disallowed prompts. However, these defenses often focus on analyzing explicit textual inputs or relevant visual scenes. In this work, we introduce Text-DJ, a novel jailbreak attack that bypasses these safeguards by exploiting the model’s Optical Character Recognition (OCR) capability. Our methodology consists of three stages. First, we decompose a single harmful query into multiple and semantically related but more benign sub-queries. Second, we pick a set of distraction queries that are maximally irrelevant to the harmful query. Third, we present all decomposed sub-queries and distraction queries to the LVLM simultaneously as a grid of images, with the position of the sub-queries being middle within the grid. We demonstrate that this method successfully circumvents the safety alignment of state-of-the-art LVLMs. We argue this attack succeeds by (1) converting text-based prompts into images, bypassing standard text-based filters, and (2) inducing distractions, where the model’s safety protocols fail to link the scattered sub-queries within a high number of irrelevant queries. Overall, our findings expose a critical vulnerability in LVLMs’ OCR capabilities that are not robust to dispersed, multi-image adversarial inputs, highlighting the need for defenses for fragmented multimodal inputs.

Method: DISK uses dual-branch controllers with cross-modal skip decisions to coordinate two coupled diffusion transformers for video and ego-trajectory prediction. It extends higher-order latent-difference skip testing to the autoregressive chain-of-forward regime and propagates controller statistics through rollout loops for long-horizon stability.

Result: When integrated into closed-loop driving rollouts on 1500 NuPlan and NuScenes samples, DISK achieves 2x speedup on trajectory diffusion and 1.6x speedup on video diffusion while maintaining L2 planning error, visual quality (FID/FVD), and NAVSIM PDMS scores.

Conclusion: DISK demonstrates practical long-horizon video-and-trajectory prediction at substantially reduced computational cost, enabling more efficient autoregressive world models for applications like autonomous driving without sacrificing performance.

Abstract: We present DISK, a training-free adaptive inference method for autoregressive world models. DISK coordinates two coupled diffusion transformers for video and ego-trajectory via dual-branch controllers with cross-modal skip decisions, preserving motion-appearance consistency without retraining. We extend higher-order latent-difference skip testing to the autoregressive chain-of-forward regime and propagate controller statistics through rollout loops for long-horizon stability. When integrated into closed-loop driving rollouts on 1500 NuPlan and NuScenes samples using an NVIDIA L40S GPU, DISK achieves 2x speedup on trajectory diffusion and 1.6x speedup on video diffusion while maintaining L2 planning error, visual quality (FID/FVD), and NAVSIM PDMS scores, demonstrating practical long-horizon video-and-trajectory prediction at substantially reduced cost.

Method: Sparse4D uses a query-based spatiotemporal 3D detection and tracking framework that fuses multi-view features in a shared world frame and propagates sparse object queries via instance memory. The study evaluates reduced input frame rates, post-training quantization (INT8 and FP8), transfer to WILDTRACK benchmark, and Transformer Engine mixed-precision fine-tuning.

Result: Sparse4D remains stable under moderate FPS reductions but identity association collapses below 2 FPS even when detections are stable. Selective quantization of backbone and neck offers best speed-accuracy trade-off, while attention modules are consistently sensitive to low precision. Low-FPS pretraining yields large zero-shot gains on WILDTRACK, while small-scale fine-tuning provides limited additional benefit. Transformer Engine mixed precision reduces latency but can destabilize identity propagation.

Conclusion: The study demonstrates the importance of stability-aware validation for multi-camera tracking systems, showing that while quantization and mixed precision can improve efficiency, they require careful implementation to maintain identity association stability, especially for attention-based modules.

Abstract: Outside-in multi-camera perception is increasingly important in indoor environments, where networks of static cameras must support multi-target tracking under occlusion and heterogeneous viewpoints. We evaluate Sparse4D, a query-based spatiotemporal 3D detection and tracking framework that fuses multi-view features in a shared world frame and propagates sparse object queries via instance memory. We study reduced input frame rates, post-training quantization (INT8 and FP8), transfer to the WILDTRACK benchmark, and Transformer Engine mixed-precision fine-tuning. To better capture identity stability, we report Average Track Duration (AvgTrackDur), which measures identity persistence in seconds. Sparse4D remains stable under moderate FPS reductions, but below 2 FPS, identity association collapses even when detections are stable. Selective quantization of the backbone and neck offers the best speed-accuracy trade-off, while attention-related modules are consistently sensitive to low precision. On WILDTRACK, low-FPS pretraining yields large zero-shot gains over the base checkpoint, while small-scale fine-tuning provides limited additional benefit. Transformer Engine mixed precision reduces latency and improves camera scalability, but can destabilize identity propagation, motivating stability-aware validation.

Method: LatentLens encodes a large text corpus and stores contextualized token representations for each token. Visual token representations are then compared to these textual representations, with the top-k nearest neighbors providing descriptions of the visual token content.

Result: LatentLens shows that commonly used methods like LogitLens substantially underestimate visual token interpretability. With LatentLens, the majority of visual tokens are interpretable across all studied models (10 different VLMs) and all layers, providing semantically meaningful and fine-grained interpretations.

Conclusion: LatentLens provides a powerful interpretability tool for understanding visual token processing in VLMs, contributing new evidence on the alignment between vision and language representations and opening new directions for analyzing latent representations.

Abstract: Transforming a large language model (LLM) into a Vision-Language Model (VLM) can be achieved by mapping the visual tokens from a vision encoder into the embedding space of an LLM. Intriguingly, this mapping can be as simple as a shallow MLP transformation. To understand why LLMs can so readily process visual tokens, we need interpretability methods that reveal what is encoded in the visual token representations at every layer of LLM processing. In this work, we introduce LatentLens, a novel approach for mapping latent representations to descriptions in natural language. LatentLens works by encoding a large text corpus and storing contextualized token representations for each token in that corpus. Visual token representations are then compared to their contextualized textual representations, with the top-k nearest neighbor representations providing descriptions of the visual token. We evaluate this method on 10 different VLMs, showing that commonly used methods, such as LogitLens, substantially underestimate the interpretability of visual tokens. With LatentLens instead, the majority of visual tokens are interpretable across all studied models and all layers. Qualitatively, we show that the descriptions produced by LatentLens are semantically meaningful and provide more fine-grained interpretations for humans compared to individual tokens. More broadly, our findings contribute new evidence on the alignment between vision and language representations, opening up new directions for analyzing latent representations.

Method: Two-stage framework: 1) Two-layer optimization for panorama generation (layout reasoning layer parses text into spatial relationships, self-optimization layer refines details via iterative MLLM feedback), 2) Panorama sliding mechanism initializes 3D Gaussian Splatting point clouds with overlapping perspectives, incorporating depth and semantic coherence losses.

Result: PSGS outperforms existing methods in panorama generation and produces more appealing 3D scenes with improved quality and detail fidelity.

Conclusion: PSGS offers a robust solution for scalable immersive content creation through high-fidelity panoramic scene generation.

Abstract: Generating realistic 3D scenes from text is crucial for immersive applications like VR, AR, and gaming. While text-driven approaches promise efficiency, existing methods suffer from limited 3D-text data and inconsistent multi-view stitching, resulting in overly simplistic scenes. To address this, we propose PSGS, a two-stage framework for high-fidelity panoramic scene generation. First, a novel two-layer optimization architecture generates semantically coherent panoramas: a layout reasoning layer parses text into structured spatial relationships, while a self-optimization layer refines visual details via iterative MLLM feedback. Second, our panorama sliding mechanism initializes globally consistent 3D Gaussian Splatting point clouds by strategically sampling overlapping perspectives. By incorporating depth and semantic coherence losses during training, we greatly improve the quality and detail fidelity of rendered scenes. Our experiments demonstrate that PSGS outperforms existing methods in panorama generation and produces more appealing 3D scenes, offering a robust solution for scalable immersive content creation.

Method: Proposes ZS-TreeSeg framework that models tree crowns as star-convex objects within topological flow fields using Cellpose-SAM adaptation. Combines canopy semantic segmentation with cells instance segmentation to mathematically separate touching tree crown instances based on vector convergence.

Result: Experiments on NEON and BAMFOREST datasets show robust generalization across diverse sensor types and canopy densities, providing a training-free solution for tree crown instance segmentation and label generation.

Conclusion: ZS-TreeSeg offers an effective zero-shot approach for tree crown segmentation that bridges the gap between domain-specific knowledge and foundation model capabilities, enabling accurate delineation in dense, overlapping canopies without training.

Abstract: Individual tree crown segmentation is an important task in remote sensing for forest biomass estimation and ecological monitoring. However, accurate delineation in dense, overlapping canopies remains a bottleneck. While supervised deep learning methods suffer from high annotation costs and limited generalization, emerging foundation models (e.g., Segment Anything Model) often lack domain knowledge, leading to under-segmentation in dense clusters. To bridge this gap, we propose ZS-TreeSeg, a Zero-Shot framework that adapts from two mature tasks: 1) Canopy Semantic segmentation; and 2) Cells instance segmentation. By modeling tree crowns as star-convex objects within a topological flow field using Cellpose-SAM, the ZS-TreeSeg framework forces the mathematical separation of touching tree crown instances based on vector convergence. Experiments on the NEON and BAMFOREST datasets and visual inspection demonstrate that our framework generalizes robustly across diverse sensor types and canopy densities, which can offer a training-free solution for tree crown instance segmentation and labels generation.

Method: Learn three offset attributes (scale, orientation, position) from source to target strokes, then align through: 1) resizing by scale, 2) rotating by orientation, and 3) displacing by position, with precise control via exposed stroke attributes.

Result: Experimental results show SketchMod achieves precise and flexible performance on stroke-level sketch editing.

Conclusion: SketchMod enables precise stroke-level sketch editing by transforming source strokes to align with target sketch patterns through learned offset attributes.

Abstract: Sketch edit at stroke-level aims to transplant source strokes onto a target sketch via stroke expansion or replacement, while preserving semantic consistency and visual fidelity with the target sketch. Recent studies addressed it by relocating source strokes at appropriate canvas positions. However, as source strokes could exhibit significant variations in both size and orientation, we may fail to produce plausible sketch editing results by merely repositioning them without further adjustments. For example, anchoring an oversized source stroke onto the target without proper scaling would fail to produce a semantically coherent outcome. In this paper, we propose SketchMod to refine the source stroke through transformation so as to align it with the target sketch’s patterns, further realize flexible sketch edit at stroke-level. As the source stroke refinement is governed by the patterns of the target sketch, we learn three key offset attributes (scale, orientation and position) from the source stroke to another, and align it with the target by: 1) resizing to match spatial proportions by scale, 2) rotating to align with local geometry by orientation, and 3) displacing to meet with semantic layout by position. Besides, a stroke’s profiles can be precisely controlled during sketch edit via the exposed captured stroke attributes. Experimental results indicate that SketchMod achieves precise and flexible performances on stroke-level sketch edit.

Method: Proposes HSSDCT with two key modules: 1) Hierarchical Dense-Residue Transformer Block (HDRTB) that progressively enlarges windows with dense-residue connections for multi-scale feature aggregation, and 2) Spatial-Spectral Correlation Layer (SSCL) that factorizes spatial and spectral dependencies to reduce self-attention to linear complexity while mitigating spectral redundancy.

Result: Extensive experiments on benchmark datasets demonstrate superior reconstruction quality with significantly lower computational costs, achieving new state-of-the-art performance in HSI fusion.

Conclusion: HSSDCT effectively addresses the limitations of existing methods by combining hierarchical multi-scale feature aggregation with efficient spatial-spectral correlation modeling, delivering both high performance and computational efficiency.

Abstract: Hyperspectral image (HSI) fusion aims to reconstruct a high-resolution HSI (HR-HSI) by combining the rich spectral information of a low-resolution HSI (LR-HSI) with the fine spatial details of a high-resolution multispectral image (HR-MSI). Although recent deep learning methods have achieved notable progress, they still suffer from limited receptive fields, redundant spectral bands, and the quadratic complexity of self-attention, which restrict both efficiency and robustness. To overcome these challenges, we propose the Hierarchical Spatial-Spectral Dense Correlation Network (HSSDCT). The framework introduces two key modules: (i) a Hierarchical Dense-Residue Transformer Block (HDRTB) that progressively enlarges windows and employs dense-residue connections for multi-scale feature aggregation, and (ii) a Spatial-Spectral Correlation Layer (SSCL) that explicitly factorizes spatial and spectral dependencies, reducing self-attention to linear complexity while mitigating spectral redundancy. Extensive experiments on benchmark datasets demonstrate that HSSDCT delivers superior reconstruction quality with significantly lower computational costs, achieving new state-of-the-art performance in HSI fusion. Our code is available at https://github.com/jemmyleee/HSSDCT.

Method: Proposes Understand-Associate-Validate (UAV) prompting to create Visual Modality Chain-of-Thought (VM-CoT), and a two-stage training paradigm: Cold-Start Supervised Fine-Tuning (CS-SFT) and Spatio-Temporal Reinforcement Fine-Tuning (ST-RFT) with Modality-understanding Spatio-Temporal (MuST) reward.

Result: Outperforms baselines by 22.71% on three RGBX grounding tasks and establishes the first RGBX-Grounding benchmark, demonstrating superior multimodal understanding and spatial perception.

Conclusion: RGBX-R1 effectively extends MLLMs’ capabilities beyond RGB to diverse visual modalities, enhancing perception and reasoning for complex real-world scenarios.

Abstract: Multimodal Large Language Models (MLLM) are primarily pre-trained on the RGB modality, thereby limiting their performance on other modalities, such as infrared, depth, and event data, which are crucial for complex scenarios. To address this, we propose RGBX-R1, a framework to enhance MLLM’s perception and reasoning capacities across various X visual modalities. Specifically, we employ an Understand-Associate-Validate (UAV) prompting strategy to construct the Visual Modality Chain-of-Thought (VM-CoT), which aims to expand the MLLMs’ RGB understanding capability into X modalities. To progressively enhance reasoning capabilities, we introduce a two-stage training paradigm: Cold-Start Supervised Fine-Tuning (CS-SFT) and Spatio-Temporal Reinforcement Fine-Tuning (ST-RFT). CS-SFT supervises the reasoning process with the guidance of VM-CoT, equipping the MLLM with fundamental modality cognition. Building upon GRPO, ST-RFT employs a Modality-understanding Spatio-Temporal (MuST) reward to reinforce modality reasoning. Notably, we construct the first RGBX-Grounding benchmark, and extensive experiments verify our superiority in multimodal understanding and spatial perception, outperforming baselines by 22.71% on three RGBX grounding tasks.

Method: Proposes SparseCut architecture with sparse shortcut connections between cross-modal encoder and LLM for hierarchical visual feature integration, plus efficient multi-grained feature fusion module that preserves language context without increasing input length.

Result: Experiments show SparseCut significantly enhances MLLM performance across various multimodal benchmarks with generality and scalability for different base LLMs.

Conclusion: SparseCut provides an effective cross-modal fusion architecture that enables richer semantic integration of visual features at multiple levels while maintaining computational efficiency.

Abstract: With the remarkable success of large language models (LLMs) in natural language understanding and generation, multimodal large language models (MLLMs) have rapidly advanced in their ability to process data across multiple modalities. While most existing efforts focus on scaling up language models or constructing higher-quality training data, limited attention has been paid to effectively integrating cross-modal knowledge into the language space. In vision-language models, for instance, aligning modalities using only high-level visual features often discards the rich semantic information present in mid- and low-level features, limiting the model’s ability of cross-modality understanding. To address this issue, we propose SparseCut, a general cross-modal fusion architecture for MLLMs, introducing sparse shortcut connections between the cross-modal encoder and the LLM. These shortcut connections enable the efficient and hierarchical integration of visual features at multiple levels, facilitating richer semantic fusion without increasing computational overhead. We further introduce an efficient multi-grained feature fusion module, which performs the fusion of visual features before routing them through the shortcuts. This preserves the original language context and does not increase the overall input length, thereby avoiding an increase in computational complexity for the LLM. Experiments demonstrate that SparseCut significantly enhances the performance of MLLMs across various multimodal benchmarks with generality and scalability for different base LLMs.

Method: 1) Builds large-scale instruction-tuning dataset from curated multimodal conversations and synthetic examples; 2) Leverages pretrained multimodal LLM for visual understanding and diffusion transformer for visual generation; 3) Uses two-stage decoupled training: first instruction-tunes MLLM, then aligns DiT with curated interleaved image-text sequences.

Result: Outperforms prior open-source models in text quality, image fidelity, and image-context alignment across public and new benchmarks. Achieves SOTA performance on text-to-image and image editing among unified generation models.

Conclusion: DuoGen provides an effective framework for high-quality interleaved multimodal generation by systematically addressing data, architecture, and evaluation challenges, enabling flexible base model selection without costly unimodal pretraining.

Abstract: Interleaved multimodal generation enables capabilities beyond unimodal generation models, such as step-by-step instructional guides, visual planning, and generating visual drafts for reasoning. However, the quality of existing interleaved generation models under general instructions remains limited by insufficient training data and base model capacity. We present DuoGen, a general-purpose interleaved generation framework that systematically addresses data curation, architecture design, and evaluation. On the data side, we build a large-scale, high-quality instruction-tuning dataset by combining multimodal conversations rewritten from curated raw websites, and diverse synthetic examples covering everyday scenarios. Architecturally, DuoGen leverages the strong visual understanding of a pretrained multimodal LLM and the visual generation capabilities of a diffusion transformer (DiT) pretrained on video generation, avoiding costly unimodal pretraining and enabling flexible base model selection. A two-stage decoupled strategy first instruction-tunes the MLLM, then aligns DiT with it using curated interleaved image-text sequences. Across public and newly proposed benchmarks, DuoGen outperforms prior open-source models in text quality, image fidelity, and image-context alignment, and also achieves state-of-the-art performance on text-to-image and image editing among unified generation models. Data and code will be released at https://research.nvidia.com/labs/dir/duetgen/.

Method: Reformulates training-free segmentation as stochastic flow equilibrium over diffusion-induced affinity graphs. Introduces Markov propagation scheme with random-walk-based label diffusion and adaptive pruning strategy that suppresses unreliable transitions while reinforcing confident affinity paths. Integrates global diffusion attention with local neighborhoods from stable diffusion for sparse yet expressive affinity structure.

Result: Achieves state-of-the-art zero-shot performance across seven widely used semantic segmentation benchmarks. Produces sharper boundaries, more coherent regions, and significantly more stable masks compared to prior spectral-clustering-based approaches.

Conclusion: The stochastic flow equilibrium formulation with Markov propagation and adaptive pruning addresses limitations of spectral graph partitioning methods, providing better segmentation quality, boundary preservation, and stability without requiring training.

Abstract: We argue that existing training-free segmentation methods rely on an implicit and limiting assumption, that segmentation is a spectral graph partitioning problem over diffusion-derived affinities. Such approaches, based on global graph partitioning and eigenvector-based formulations of affinity matrices, suffer from several fundamental drawbacks, they require pre-selecting the number of clusters, induce boundary oversmoothing due to spectral relaxation, and remain highly sensitive to noisy or multi-modal affinity distributions. Moreover, many prior works neglect the importance of local neighborhood structure, which plays a crucial role in stabilizing affinity propagation and preserving fine-grained contours. To address these limitations, we reformulate training-free segmentation as a stochastic flow equilibrium problem over diffusion-induced affinity graphs, where segmentation emerges from a stochastic propagation process that integrates global diffusion attention with local neighborhoods extracted from stable diffusion, yielding a sparse yet expressive affinity structure. Building on this formulation, we introduce a Markov propagation scheme that performs random-walk-based label diffusion with an adaptive pruning strategy that suppresses unreliable transitions while reinforcing confident affinity paths. Experiments across seven widely used semantic segmentation benchmarks demonstrate that our method achieves state-of-the-art zero-shot performance, producing sharper boundaries, more coherent regions, and significantly more stable masks compared to prior spectral-clustering-based approaches.

Method: Proposes MRAD framework with three variants: (1) MRAD-TF (train-free) freezes CLIP image encoder and constructs two-level memory bank (image-level and pixel-level) from auxiliary data, storing feature-label pairs as keys/values; (2) MRAD-FT adds two linear layers to fine-tune retrieval metric; (3) MRAD-CLIP injects normal/anomalous region priors as dynamic biases into CLIP’s text prompts for better generalization.

Result: Demonstrates superior performance across 16 industrial and medical datasets for both anomaly classification and segmentation, under both train-free and training-based settings.

Conclusion: Fully leveraging empirical distribution of raw data through memory retrieval can achieve stronger anomaly detection performance than relying only on model fitting, offering better cross-domain stability and lower computational costs.

Abstract: Zero-shot anomaly detection (ZSAD) often leverages pretrained vision or vision-language models, but many existing methods use prompt learning or complex modeling to fit the data distribution, resulting in high training or inference cost and limited cross-domain stability. To address these limitations, we propose Memory-Retrieval Anomaly Detection method (MRAD), a unified framework that replaces parametric fitting with a direct memory retrieval. The train-free base model, MRAD-TF, freezes the CLIP image encoder and constructs a two-level memory bank (image-level and pixel-level) from auxiliary data, where feature-label pairs are explicitly stored as keys and values. During inference, anomaly scores are obtained directly by similarity retrieval over the memory bank. Based on the MRAD-TF, we further propose two lightweight variants as enhancements: (i) MRAD-FT fine-tunes the retrieval metric with two linear layers to enhance the discriminability between normal and anomaly; (ii) MRAD-CLIP injects the normal and anomalous region priors from the MRAD-FT as dynamic biases into CLIP’s learnable text prompts, strengthening generalization to unseen categories. Across 16 industrial and medical datasets, the MRAD framework consistently demonstrates superior performance in anomaly classification and segmentation, under both train-free and training-based settings. Our work shows that fully leveraging the empirical distribution of raw data, rather than relying only on model fitting, can achieve stronger anomaly detection performance. The code will be publicly released at https://github.com/CROVO1026/MRAD.

Method: SAGE dynamically adjusts speculation tree structure based on real-time prediction uncertainty, using output entropy as a confidence indicator. It constructs deeper-narrower trees for high-confidence predictions and shallower-wider trees for uncertain predictions.

Result: SAGE achieves up to 3.36× decoding speedup for LLaVA-OneVision-72B and 3.18× for Qwen2.5-VL-72B without any loss in output quality, outperforming static tree baselines.

Conclusion: Dynamic tree adaptation based on prediction uncertainty significantly improves speculative decoding efficiency for vision-language models, enabling faster inference while maintaining output quality.

Abstract: Speculative decoding has emerged as a promising approach to accelerate inference in vision-language models (VLMs) by enabling parallel verification of multiple draft tokens. However, existing methods rely on static tree structures that remain fixed throughout the decoding process, failing to adapt to the varying prediction difficulty across generation steps. This leads to suboptimal acceptance lengths and limited speedup. In this paper, we propose SAGE, a novel framework that dynamically adjusts the speculation tree structure based on real-time prediction uncertainty. Our key insight is that output entropy serves as a natural confidence indicator with strong temporal correlation across decoding steps. SAGE constructs deeper-narrower trees for high-confidence predictions to maximize speculation depth, and shallower-wider trees for uncertain predictions to diversify exploration. SAGE improves acceptance lengths and achieves faster acceleration compared to static tree baselines. Experiments on multiple benchmarks demonstrate the effectiveness of SAGE: without any loss in output quality, it delivers up to $3.36\times$ decoding speedup for LLaVA-OneVision-72B and $3.18\times$ for Qwen2.5-VL-72B.

Method: VLDet introduces two key components: 1) VL-PUB module that revamps the feature pyramid for fine-grained visual-language alignment by exploiting CLIP’s knowledge and adapting the backbone for detection, and 2) SigRPN block with a sigmoid-based anchor-text contrastive alignment loss to improve novel category detection.

Result: Achieves 58.7 AP for novel classes on COCO2017 (27.6% improvement) and 24.8 AP on LVIS (6.9% improvement), surpassing all state-of-the-art methods. Also demonstrates superior zero-shot performance on closed-set object detection.

Conclusion: VLDet effectively bridges the gap between visual-language modeling and object detection, providing a robust framework for open-vocabulary detection with significant performance improvements over existing methods.

Abstract: Traditional object detection systems are typically constrained to predefined categories, limiting their applicability in dynamic environments. In contrast, open-vocabulary object detection (OVD) enables the identification of objects from novel classes not present in the training set. Recent advances in visual-language modeling have led to significant progress of OVD. However, prior works face challenges in either adapting the single-scale image backbone from CLIP to the detection framework or ensuring robust visual-language alignment. We propose Visual-Language Detection (VLDet), a novel framework that revamps feature pyramid for fine-grained visual-language alignment, leading to improved OVD performance. With the VL-PUB module, VLDet effectively exploits the visual-language knowledge from CLIP and adapts the backbone for object detection through feature pyramid. In addition, we introduce the SigRPN block, which incorporates a sigmoid-based anchor-text contrastive alignment loss to improve detection of novel categories. Through extensive experiments, our approach achieves 58.7 AP for novel classes on COCO2017 and 24.8 AP on LVIS, surpassing all state-of-the-art methods and achieving significant improvements of 27.6% and 6.9%, respectively. Furthermore, VLDet also demonstrates superior zero-shot performance on closed-set object detection.

Method: Proposes SADER with: 1) Multi-Temporal Conditional Diffusion Network (MTCDN) using temporal fusion and hybrid attention to capture multi-temporal correlations, 2) Cloud-aware attention loss emphasizing cloud-dominated regions based on thickness and brightness, 3) Deterministic resampling strategy for iterative refinement under fixed sampling steps.

Result: Extensive experiments on multiple multi-temporal datasets show SADER consistently outperforms state-of-the-art cloud removal methods across all evaluation metrics.

Conclusion: SADER effectively addresses cloud removal in remote sensing by combining temporal modeling, attention mechanisms, and efficient sampling strategies, with publicly available code.

Abstract: Cloud contamination severely degrades the usability of remote sensing imagery and poses a fundamental challenge for downstream Earth observation tasks. Recently, diffusion-based models have emerged as a dominant paradigm for remote sensing cloud removal due to their strong generative capability and stable optimization. However, existing diffusion-based approaches often suffer from limited sampling efficiency and insufficient exploitation of structural and temporal priors in multi-temporal remote sensing scenarios. In this work, we propose SADER, a structure-aware diffusion framework for multi-temporal remote sensing cloud removal. SADER first develops a scalable Multi-Temporal Conditional Diffusion Network (MTCDN) to fully capture multi-temporal and multimodal correlations via temporal fusion and hybrid attention. Then, a cloud-aware attention loss is introduced to emphasize cloud-dominated regions by accounting for cloud thickness and brightness discrepancies. In addition, a deterministic resampling strategy is designed for continuous diffusion models to iteratively refine samples under fixed sampling steps by replacing outliers through guided correction. Extensive experiments on multiple multi-temporal datasets demonstrate that SADER consistently outperforms state-of-the-art cloud removal methods across all evaluation metrics. The code of SADER is publicly available at https://github.com/zyfzs0/SADER.

Method: Uses deterministic operators (farthest point sampling, k-nearest neighbors, pooling) with adaptive Gaussian-Fourier positional encoding whose parameters are chosen from input geometry. For segmentation, adds fixed-frequency Fourier features for global context.

Result: Achieves strong performance on ModelNet40/ModelNet-R, ScanObjectNN, and ShapeNetPart among non-parametric baselines, particularly effective in few-shot settings on ModelNet40, with favorable memory use and inference time.

Conclusion: NPNet demonstrates that effective 3D point cloud processing can be achieved without learned parameters through careful design of deterministic operators and adaptive positional encodings.

Abstract: We present NPNet, a fully non-parametric approach for 3D point-cloud classification and part segmentation. NPNet contains no learned weights; instead, it builds point features using deterministic operators such as farthest point sampling, k-nearest neighbors, and pooling. Our key idea is an adaptive Gaussian-Fourier positional encoding whose bandwidth and Gaussian-cosine mixing are chosen from the input geometry, helping the method remain stable across different scales and sampling densities. For segmentation, we additionally incorporate fixed-frequency Fourier features to provide global context alongside the adaptive encoding. Across ModelNet40/ModelNet-R, ScanObjectNN, and ShapeNetPart, NPNet achieves strong performance among non-parametric baselines, and it is particularly effective in few-shot settings on ModelNet40. NPNet also offers favorable memory use and inference time compared to prior non-parametric methods

Method: 1) OmniVCHall benchmark with diverse video domains, novel camera-based hallucination type, fine-grained taxonomy, and adversarial answer options. 2) TriCD framework: contrastive decoding with triple-pathway calibration, adaptive perturbation controller for negative variants, saliency-guided enhancement for visual evidence, optimized via reinforcement learning.

Result: Evaluation of 39 VLLMs shows even advanced models (Qwen3-VL, GPT-5) have substantial performance degradation. TriCD improves performance across two backbones with average accuracy improvement over 10%.

Conclusion: Compositional hallucinations are critical challenge in VLLMs. OmniVCHall enables systematic evaluation, and TriCD provides effective mitigation framework with consistent improvements across models.

Abstract: Current research on video hallucination mitigation primarily focuses on isolated error types, leaving compositional hallucinations, arising from incorrect reasoning over multiple interacting spatial and temporal factors largely underexplored. We introduce OmniVCHall, a benchmark designed to systematically evaluate both isolated and compositional hallucinations in video multimodal large language models (VLLMs). OmniVCHall spans diverse video domains, introduces a novel camera-based hallucination type, and defines a fine-grained taxonomy, together with adversarial answer options (e.g., “All are correct” and “None of the above”) to prevent shortcut reasoning. The evaluations of 39 representative VLLMs reveal that even advanced models (e.g., Qwen3-VL and GPT-5) exhibit substantial performance degradation. We propose TriCD, a contrastive decoding framework with a triple-pathway calibration mechanism. An adaptive perturbation controller dynamically selects distracting operations to construct negative video variants, while a saliency-guided enhancement module adaptively reinforces grounded token-wise visual evidences. These components are optimized via reinforcement learning to encourage precise decision-making under compositional hallucination settings. Experimental results show that TriCD consistently improves performance across two representative backbones, achieving an average accuracy improvement of over 10%. The data and code can be find at https://github.com/BMRETURN/OmniVCHall.

Method: Proposes GLAD, a generative language-assisted tracking model using diffusion models for multi-modal fusion of text descriptions and template images to enhance compatibility between language and image modalities and improve template image semantics.

Result: Achieves state-of-the-art performance on multiple benchmarks with impressive inference speed. Blurry and semantically ambiguous template images can be effectively restored through the generative fusion paradigm.

Conclusion: GLAD demonstrates superior performance over existing fusion paradigms by using generative diffusion models for multi-modal fusion, effectively addressing challenges in vision-language tracking with low-semantic images.

Abstract: Vision-language tracking has gained increasing attention in many scenarios. This task simultaneously deals with visual and linguistic information to localize objects in videos. Despite its growing utility, the development of vision-language tracking methods remains in its early stage. Current vision-language trackers usually employ Transformer architectures for interactive integration of template, search, and text features. However, persistent challenges about low-semantic images including prevalent image blurriness, low resolution and so on, may compromise model performance through degraded cross-modal understanding. To solve this problem, language assistance is usually used to deal with the obstacles posed by low-semantic images. However, due to the existing gap between current textual and visual features, direct concatenation and fusion of these features may have limited effectiveness. To address these challenges, we introduce a pioneering Generative Language-AssisteD tracking model, GLAD, which utilizes diffusion models for the generative multi-modal fusion of text description and template image to bolster compatibility between language and image and enhance template image semantic information. Our approach demonstrates notable improvements over the existing fusion paradigms. Blurry and semantically ambiguous template images can be restored to improve multi-modal features in the generative fusion paradigm. Experiments show that our method establishes a new state-of-the-art on multiple benchmarks and achieves an impressive inference speed. The code and models will be released at: https://github.com/Confetti-lxy/GLAD

Method: Proposes BDG with two key components: 1) MAS-GLCM for fine-grained degradation type/level discrimination, and 2) Three-stage diffusion training (generation, bridging, restoration) that integrates discriminative information into restoration while preserving texture generation capabilities.

Result: Achieves significant performance gains in all-in-one restoration and real-world super-resolution tasks, with substantial improvements in fidelity without compromising perceptual quality, without architectural changes.

Conclusion: BDG effectively bridges degradation discrimination and generation, demonstrating strong performance in multi-task, multi-degradation scenarios through the integration of MAS-GLCM features with diffusion-based restoration.

Abstract: Universal image restoration is a critical task in low-level vision, requiring the model to remove various degradations from low-quality images to produce clean images with rich detail. The challenges lie in sampling the distribution of high-quality images and adjusting the outputs on the basis of the degradation. This paper presents a novel approach, Bridging Degradation discrimination and Generation (BDG), which aims to address these challenges concurrently. First, we propose the Multi-Angle and multi-Scale Gray Level Co-occurrence Matrix (MAS-GLCM) and demonstrate its effectiveness in performing fine-grained discrimination of degradation types and levels. Subsequently, we divide the diffusion training process into three distinct stages: generation, bridging, and restoration. The objective is to preserve the diffusion model’s capability of restoring rich textures while simultaneously integrating the discriminative information from the MAS-GLCM into the restoration process. This enhances its proficiency in addressing multi-task and multi-degraded scenarios. Without changing the architecture, BDG achieves significant performance gains in all-in-one restoration and real-world super-resolution tasks, primarily evidenced by substantial improvements in fidelity without compromising perceptual quality. The code and pretrained models are provided in https://github.com/MILab-PKU/BDG.

Method: Proposes MAUGen with two key modules: (1) Multi-modal Representation Learning (MRL) that captures relationships among textual descriptions, facial identity, expression images, and AU activations in a unified latent space; (2) Diffusion-based Image label Generator (DIG) that decodes joint representations into aligned facial image-label pairs across diverse identities.

Result: Introduces MIFA, a large-scale multimodal synthetic dataset with comprehensive AU annotations and identity variations. MAUGen outperforms existing methods in synthesizing photorealistic, demographically diverse facial images with semantically aligned AU labels.

Conclusion: MAUGen successfully addresses the data scarcity problem in AU recognition by generating high-quality synthetic facial expression data with precise anatomical annotations, enabling more robust and generalizable facial expression analysis systems.

Abstract: The lack of large-scale, demographically diverse face images with precise Action Unit (AU) occurrence and intensity annotations has long been recognized as a fundamental bottleneck in developing generalizable AU recognition systems. In this paper, we propose MAUGen, a diffusion-based multi-modal framework that jointly generates a large collection of photorealistic facial expressions and anatomically consistent AU labels, including both occurrence and intensity, conditioned on a single descriptive text prompt. Our MAUGen involves two key modules: (1) a Multi-modal Representation Learning (MRL) module that captures the relationships among the paired textual description, facial identity, expression image, and AU activations within a unified latent space; and (2) a Diffusion-based Image label Generator (DIG) that decodes the joint representation into aligned facial image-label pairs across diverse identities. Under this framework, we introduce Multi-Identity Facial Action (MIFA), a large-scale multimodal synthetic dataset featuring comprehensive AU annotations and identity variations. Extensive experiments demonstrate that MAUGen outperforms existing methods in synthesizing photorealistic, demographically diverse facial images along with semantically aligned AU labels.

Method: Created Pix2Fact benchmark with 1,000 high-resolution (4K+) images across 8 daily-life scenarios. Questions and answers were meticulously crafted by PhD annotators from top universities in partnership with a professional data annotation firm. Each question requires detailed visual grounding, multi-hop reasoning, and external knowledge integration.

Result: Evaluation of 9 state-of-the-art VLMs (including proprietary models like Gemini-3-Pro and GPT-5) shows substantial challenge: most advanced model achieves only 24.0% average accuracy, compared to human performance of 56%. This reveals significant limitations in current models’ visual comprehension.

Conclusion: Pix2Fact exposes critical limitations in current VLMs’ ability to combine fine-grained perception with robust knowledge-based reasoning. The benchmark will drive development of next-generation multimodal agents that can better replicate human-level visual comprehension.

Abstract: Despite progress on general tasks, VLMs struggle with challenges demanding both detailed visual grounding and deliberate knowledge-based reasoning, a synergy not captured by existing benchmarks that evaluate these skills separately. To close this gap, we introduce Pix2Fact, a new visual question-answering benchmark designed to evaluate expert-level perception and knowledge-intensive multi-hop reasoning. Pix2Fact contains 1,000 high-resolution (4K+) images spanning 8 daily-life scenarios and situations, with questions and answers meticulously crafted by annotators holding PhDs from top global universities working in partnership with a professional data annotation firm. Each question requires detailed visual grounding, multi-hop reasoning, and the integration of external knowledge to answer. Our evaluation of 9 state-of-the-art VLMs, including proprietary models like Gemini-3-Pro and GPT-5, reveals the substantial challenge posed by Pix2Fact: the most advanced model achieves only 24.0% average accuracy, in stark contrast to human performance of 56%. This significant gap underscores the limitations of current models in replicating human-level visual comprehension. We believe Pix2Fact will serve as a critical benchmark to drive the development of next-generation multimodal agents that combine fine-grained perception with robust, knowledge-based reasoning.

Method: Introduces Gaussian neurons to model style intensity explicitly, parameterizes a learnable style tuner, and uses tunable stylization guidance with cross-view style alignment from diffusion models and two-stage optimization.

Result: The method delivers visually appealing results with flexible customizability for 3D style transfer, allowing users to adjust style intensity to match their desired content-style balance.

Conclusion: Tune-Your-Style provides an effective intensity-tunable paradigm for 3D style transfer that enhances customizability while maintaining visual quality.

Abstract: 3D style transfer refers to the artistic stylization of 3D assets based on reference style images. Recently, 3DGS-based stylization methods have drawn considerable attention, primarily due to their markedly enhanced training and rendering speeds. However, a vital challenge for 3D style transfer is to strike a balance between the content and the patterns and colors of the style. Although the existing methods strive to achieve relatively balanced outcomes, the fixed-output paradigm struggles to adapt to the diverse content-style balance requirements from different users. In this work, we introduce a creative intensity-tunable 3D style transfer paradigm, dubbed \textbf{Tune-Your-Style}, which allows users to flexibly adjust the style intensity injected into the scene to match their desired content-style balance, thus enhancing the customizability of 3D style transfer. To achieve this goal, we first introduce Gaussian neurons to explicitly model the style intensity and parameterize a learnable style tuner to achieve intensity-tunable style injection. To facilitate the learning of tunable stylization, we further propose the tunable stylization guidance, which obtains multi-view consistent stylized views from diffusion models through cross-view style alignment, and then employs a two-stage optimization strategy to provide stable and efficient guidance by modulating the balance between full-style guidance from the stylized views and zero-style guidance from the initial rendering. Extensive experiments demonstrate that our method not only delivers visually appealing results, but also exhibits flexible customizability for 3D style transfer. Project page is available at https://zhao-yian.github.io/TuneStyle.

Method: Uses sparse autoencoders to decompose dense visual embeddings into interpretable neurons. Identifies neuron types (always-on vs image-specific). Proposes Contrastive Neuron Steering (CNS) that uses contrastive analysis between clean/noisy inputs to identify image-specific neurons, then selectively amplifies informative neurons while suppressing perturbation-induced activations.

Result: CNS reduces hallucinations while preserving multimodal understanding. Experiments on hallucination-focused and general multimodal benchmarks show consistent improvement. The method operates at prefilling stage and is compatible with existing decoding-stage methods.

Conclusion: Hallucinations in LVLMs stem from disruptions in image-specific neurons. CNS provides an effective representation-level intervention that improves visual grounding and reduces hallucinations through neuron-level manipulation.

Abstract: LVLMs achieve remarkable multimodal understanding and generation but remain susceptible to hallucinations. Existing mitigation methods predominantly focus on output-level adjustments, leaving the internal mechanisms that give rise to these hallucinations largely unexplored. To gain a deeper understanding, we adopt a representation-level perspective by introducing sparse autoencoders (SAEs) to decompose dense visual embeddings into sparse, interpretable neurons. Through neuron-level analysis, we identify distinct neuron types, including always-on neurons and image-specific neurons. Our findings reveal that hallucinations often result from disruptions or spurious activations of image-specific neurons, while always-on neurons remain largely stable. Moreover, selectively enhancing or suppressing image-specific neurons enables controllable intervention in LVLM outputs, improving visual grounding and reducing hallucinations. Building on these insights, we propose Contrastive Neuron Steering (CNS), which identifies image-specific neurons via contrastive analysis between clean and noisy inputs. CNS selectively amplifies informative neurons while suppressing perturbation-induced activations, producing more robust and semantically grounded visual representations. This not only enhances visual understanding but also effectively mitigates hallucinations. By operating at the prefilling stage, CNS is fully compatible with existing decoding-stage methods. Extensive experiments on both hallucination-focused and general multimodal benchmarks demonstrate that CNS consistently reduces hallucinations while preserving overall multimodal understanding.

Method: Based on Stable Diffusion, FaceSnap uses a Facial Attribute Mixer to extract comprehensive fused information from both low-level specific features and high-level abstract features. It includes a Landmark Predictor to maintain reference identity across different poses, and an ID-preserving module to inject these into the UNet architecture.

Result: Experimental results show FaceSnap performs remarkably in personalized and customized portrait generation, surpassing other state-of-the-art methods in this domain with high fidelity facial details.

Conclusion: FaceSnap provides an efficient, plug-and-play solution for personalized portrait generation that requires only a single reference image and produces consistent results in a single inference stage, with strong generalization capabilities.

Abstract: Benefiting from the significant advancements in text-to-image diffusion models, research in personalized image generation, particularly customized portrait generation, has also made great strides recently. However, existing methods either require time-consuming fine-tuning and lack generalizability or fail to achieve high fidelity in facial details. To address these issues, we propose FaceSnap, a novel method based on Stable Diffusion (SD) that requires only a single reference image and produces extremely consistent results in a single inference stage. This method is plug-and-play and can be easily extended to different SD models. Specifically, we design a new Facial Attribute Mixer that can extract comprehensive fused information from both low-level specific features and high-level abstract features, providing better guidance for image generation. We also introduce a Landmark Predictor that maintains reference identity across landmarks with different poses, providing diverse yet detailed spatial control conditions for image generation. Then we use an ID-preserving module to inject these into the UNet. Experimental results demonstrate that our approach performs remarkably in personalized and customized portrait generation, surpassing other state-of-the-art methods in this domain.

Method: Three-module framework: 1) Reference Generation produces occlusion-free face using structural guidance from parsed masks, 2) Feature Enhancement contrasts tokens between raw/reference images for initial prompts and modifies features via cross-attention, 3) Prompt Selection constructs positive/negative prompts and screens them with self-attention network for mask decoder, trained with three novel objective functions without occlusion ground truth.

Result: Extensive experiments on a dedicatedly collected dataset demonstrate S³POT’s superior performance and effectiveness of each module.

Conclusion: S³POT successfully addresses face occlusion segmentation by leveraging face generation capabilities and foundation segmentation models without requiring occlusion ground truth masks.

Abstract: Existing face parsing methods usually misclassify occlusions as facial components. This is because occlusion is a high-level concept, it does not refer to a concrete category of object. Thus, constructing a real-world face dataset covering all categories of occlusion object is almost impossible and accurate mask annotation is labor-intensive. To deal with the problems, we present S$^3$POT, a contrast-driven framework synergizing face generation with self-supervised spatial prompting, to achieve occlusion segmentation. The framework is inspired by the insights: 1) Modern face generators’ ability to realistically reconstruct occluded regions, creating an image that preserve facial geometry while eliminating occlusion, and 2) Foundation segmentation models’ (e.g., SAM) capacity to extract precise mask when provided with appropriate prompts. In particular, S$^3$POT consists of three modules: Reference Generation (RF), Feature enhancement (FE), and Prompt Selection (PS). First, a reference image is produced by RF using structural guidance from parsed mask. Second, FE performs contrast of tokens between raw and reference images to obtain an initial prompt, then modifies image features with the prompt by cross-attention. Third, based on the enhanced features, PS constructs a set of positive and negative prompts and screens them with a self-attention network for a mask decoder. The network is learned under the guidance of three novel and complementary objective functions without occlusion ground truth mask involved. Extensive experiments on a dedicatedly collected dataset demonstrate S$^3$POT’s superior performance and the effectiveness of each module.

Method: VIZOR constructs dense 3D scene graphs directly from raw 3D scenes without training, defining spatial relationships relative to each object’s front-facing direction for viewpoint invariance, and infers open-vocabulary relationships without annotated training data.

Result: VIZOR outperforms state-of-the-art methods in scene graph generation and achieves 22% and 4.81% gains in zero-shot grounding accuracy on Replica and Nr3D datasets respectively.

Conclusion: VIZOR provides an effective training-free solution for generating unambiguous, viewpoint-invariant 3D scene graphs with improved generalization and accuracy in downstream tasks.

Abstract: Scene understanding and reasoning has been a fundamental problem in 3D computer vision, requiring models to identify objects, their properties, and spatial or comparative relationships among the objects. Existing approaches enable this by creating scene graphs using multiple inputs such as 2D images, depth maps, object labels, and annotated relationships from specific reference view. However, these methods often struggle with generalization and produce inaccurate spatial relationships like “left/right”, which become inconsistent across different viewpoints. To address these limitations, we propose Viewpoint-Invariant Zero-shot scene graph generation for 3D scene Reasoning (VIZOR). VIZOR is a training-free, end-to-end framework that constructs dense, viewpoint-invariant 3D scene graphs directly from raw 3D scenes. The generated scene graph is unambiguous, as spatial relationships are defined relative to each object’s front-facing direction, making them consistent regardless of the reference view. Furthermore, it infers open-vocabulary relationships that describe spatial and proximity relationships among scene objects without requiring annotated training data. We conduct extensive quantitative and qualitative evaluations to assess the effectiveness of VIZOR in scene graph generation and downstream tasks, such as query-based object grounding. VIZOR outperforms state-of-the-art methods, showing clear improvements in scene graph generation and achieving 22% and 4.81% gains in zero-shot grounding accuracy on the Replica and Nr3D datasets, respectively.

Method: Uses 3D face predictor to reconstruct 3D-aware facial priors (reference ID, target expressions, poses). Includes ID-Encoder for local/global facial feature fusion, ID-Ctrl for aligning ID features using 3D face guidance, and ID-Injector for enhanced ID fidelity and facial controllability. Trained on collected ID-centric dataset.

Result: Extensive experiments show Diff-PC surpasses state-of-the-art methods in ID preservation, facial control, and text-to-image consistency. Also compatible with multi-style foundation models.

Conclusion: Diff-PC effectively addresses limitations in existing portrait customization methods by providing precise identity preservation and facial control through a diffusion-based framework with 3D facial priors and specialized ID modules.

Abstract: Portrait customization (PC) has recently garnered significant attention due to its potential applications. However, existing PC methods lack precise identity (ID) preservation and face control. To address these tissues, we propose Diff-PC, a diffusion-based framework for zero-shot PC, which generates realistic portraits with high ID fidelity, specified facial attributes, and diverse backgrounds. Specifically, our approach employs the 3D face predictor to reconstruct the 3D-aware facial priors encompassing the reference ID, target expressions, and poses. To capture fine-grained face details, we design ID-Encoder that fuses local and global facial features. Subsequently, we devise ID-Ctrl using the 3D face to guide the alignment of ID features. We further introduce ID-Injector to enhance ID fidelity and facial controllability. Finally, training on our collected ID-centric dataset improves face similarity and text-to-image (T2I) alignment. Extensive experiments demonstrate that Diff-PC surpasses state-of-the-art methods in ID preservation, facial control, and T2I consistency. Furthermore, our method is compatible with multi-style foundation models.

Method: Two parameter-efficient strategies: 1) Dual-branch architecture fusing frozen SAM features with trainable VMamba encoder via cross-attention; 2) Adapter-based approach injecting lightweight Tri-Plane Mamba modules into SAM ViT encoder. Introduces Multi-Frequency Gated Convolution for spatial-frequency analysis.

Result: On ACDC cardiac MRI: Dual-branch achieves 0.906 Dice (comparable to UNet++ 0.907), outperforms baselines on Myocardium (0.910) and Left Ventricle (0.971). Adapter-based variant offers 4.77 FPS with 0.880 Dice.

Conclusion: Hybridizing foundation models with efficient SSM-based architectures provides practical solution for 3D medical image segmentation, balancing accuracy and computational efficiency.

Abstract: Accurate segmentation of 3D medical images such as MRI and CT is essential for clinical diagnosis and treatment planning. Foundation models like the Segment Anything Model (SAM) provide powerful general-purpose representations but struggle in medical imaging due to domain shift, their inherently 2D design, and the high computational cost of fine-tuning. To address these challenges, we propose Mamba-SAM, a novel and efficient hybrid architecture that combines a frozen SAM encoder with the linear-time efficiency and long-range modeling capabilities of Mamba-based State Space Models (SSMs). We investigate two parameter-efficient adaptation strategies. The first is a dual-branch architecture that explicitly fuses general features from a frozen SAM encoder with domain-specific representations learned by a trainable VMamba encoder using cross-attention. The second is an adapter-based approach that injects lightweight, 3D-aware Tri-Plane Mamba (TPMamba) modules into the frozen SAM ViT encoder to implicitly model volumetric context. Within this framework, we introduce Multi-Frequency Gated Convolution (MFGC), which enhances feature representation by jointly analyzing spatial and frequency-domain information via 3D discrete cosine transforms and adaptive gating. Extensive experiments on the ACDC cardiac MRI dataset demonstrate the effectiveness of the proposed methods. The dual-branch Mamba-SAM-Base model achieves a mean Dice score of 0.906, comparable to UNet++ (0.907), while outperforming all baselines on Myocardium (0.910) and Left Ventricle (0.971) segmentation. The adapter-based TP MFGC variant offers superior inference speed (4.77 FPS) with strong accuracy (0.880 Dice). These results show that hybridizing foundation models with efficient SSM-based architectures provides a practical and effective solution for 3D medical image segmentation.

Method: NOVA aligns visual representations to a frozen text encoder by predicting text embeddings from augmented image views. It uses Sketched Isotropic Gaussian Regularization (SIGReg) to enforce an isotropic Gaussian structure, eliminating negative sampling, momentum encoders, and stop-gradients.

Result: NOVA outperforms multiple standard baselines on zero-shot chest X-ray classification across three benchmark datasets, while exhibiting substantially more consistent training runs with reduced hyperparameter sensitivity.

Conclusion: Non-contrastive vision-language pretraining offers a simpler, more stable, and more effective alternative to contrastive methods, particularly in specialized domains like medical imaging.

Abstract: Vision-language models have transformed multimodal representation learning, yet dominant contrastive approaches like CLIP require large batch sizes, careful negative sampling, and extensive hyperparameter tuning. We introduce NOVA, a NOn-contrastive Vision-language Alignment framework based on joint embedding prediction with distributional regularization. NOVA aligns visual representations to a frozen, domain-specific text encoder by predicting text embeddings from augmented image views, while enforcing an isotropic Gaussian structure via Sketched Isotropic Gaussian Regularization (SIGReg). This eliminates the need for negative sampling, momentum encoders, or stop-gradients, reducing the training objective to a single hyperparameter. We evaluate NOVA on zeroshot chest X-ray classification using ClinicalBERT as the text encoder and Vision Transformers trained from scratch on MIMIC-CXR. On zero-shot classification across three benchmark datasets, NOVA outperforms multiple standard baselines while exhibiting substantially more consistent training runs. Our results demonstrate that non-contrastive vision-language pretraining offers a simpler, more stable, and more effective alternative to contrastive methods.

Method: Proposes a Schrödinger-inspired architecture that learns voxelwise amplitude, phase, and potential fields defining a complex-valued wavefunction ψ = Ae^{iφ}, which is evolved forward using a differentiable, unrolled Schrödinger time stepper within a deep convolutional framework.

Result: Demonstrates accurate and stable prediction of future 4D states (volumetric intensities and deformation fields) on synthetic benchmarks with realistic shape deformations and topological changes, showing improved temporal stability and interpretability.

Conclusion: The approach successfully integrates physical priors into deep learning, combining neural network expressivity with physics-based modeling robustness, offering interpretable, stable, and anatomically consistent spatiotemporal prediction for 4D forecasting.

Abstract: Spatiotemporal forecasting of complex three-dimensional phenomena (4D: 3D + time) is fundamental to applications in medical imaging, fluid and material dynamics, and geophysics. In contrast to unconstrained neural forecasting models, we propose a Schrödinger-inspired, physics-guided neural architecture that embeds an explicit time-evolution operator within a deep convolutional framework for 4D prediction. From observed volumetric sequences, the model learns voxelwise amplitude, phase, and potential fields that define a complex-valued wavefunction $ψ= A e^{iφ}$, which is evolved forward in time using a differentiable, unrolled Schrödinger time stepper. This physics-guided formulation yields several key advantages: (i) temporal stability arising from the structured evolution operator, which mitigates drift and error accumulation in long-horizon forecasting; (ii) an interpretable latent representation, where phase encodes transport dynamics, amplitude captures structural intensity, and the learned potential governs spatiotemporal interactions; and (iii) natural compatibility with deformation-based synthesis, which is critical for preserving anatomical fidelity in medical imaging applications. By integrating physical priors directly into the learning process, the proposed approach combines the expressivity of deep networks with the robustness and interpretability of physics-based modeling. We demonstrate accurate and stable prediction of future 4D states, including volumetric intensities and deformation fields, on synthetic benchmarks that emulate realistic shape deformations and topological changes. To our knowledge, this is the first end-to-end 4D neural forecasting framework to incorporate a Schrödinger-type evolution operator, offering a principled pathway toward interpretable, stable, and anatomically consistent spatiotemporal prediction.

Method: Introduces a computationally efficient super-resolution step that imputes additional slices to create anatomically consistent isotropic volumes from anisotropic 3D reconstructions. Uses domain-randomized synthetic data for training to ensure generalization across dissection protocols and robustness to large slab thicknesses.

Result: The imputed volumes yield improved automated segmentations with higher Dice scores, particularly in cortical and white matter regions. Validation shows more accurate cortical surfaces and MRI registration compared to previous methods.

Conclusion: The approach enhances resolution and anatomical fidelity of photograph-based brain reconstructions, strengthening the bridge between neuropathology and neuroimaging. The method is publicly available.

Abstract: Neuropathological analyses benefit from spatially precise volumetric reconstructions that enhance anatomical delineation and improve morphometric accuracy. Our prior work has shown the feasibility of reconstructing 3D brain volumes from 2D dissection photographs. However these outputs sometimes exhibit coarse, overly smooth reconstructions of structures, especially under high anisotropy (i.e., reconstructions from thick slabs). Here, we introduce a computationally efficient super-resolution step that imputes slices to generate anatomically consistent isotropic volumes from anisotropic 3D reconstructions of dissection photographs. By training on domain-randomized synthetic data, we ensure that our method generalizes across dissection protocols and remains robust to large slab thicknesses. The imputed volumes yield improved automated segmentations, achieving higher Dice scores, particularly in cortical and white matter regions. Validation on surface reconstruction and atlas registration tasks demonstrates more accurate cortical surfaces and MRI registration. By enhancing the resolution and anatomical fidelity of photograph-based reconstructions, our approach strengthens the bridge between neuropathology and neuroimaging. Our method is publicly available at https://surfer.nmr.mgh.harvard.edu/fswiki/mri_3d_photo_recon

Method: Proposes HPC with hierarchical point-based latent representation operating per-Gaussian to avoid redundancy. Uses tailored aggregation for compactness. First to compress neural networks for streaming dynamic Gaussian Splatting by exploiting inter-frame parameter correlations. End-to-end compression framework combining latent and network compression.

Result: Achieves 67% storage reduction compared to baseline while maintaining high reconstruction fidelity. Substantially outperforms state-of-the-art methods in comprehensive experimental evaluations.

Conclusion: HPC effectively addresses limitations of existing compression methods for dynamic Gaussian Splatting, providing efficient streaming transmission with small memory footprint while preserving rendering quality.

Abstract: While dynamic Gaussian Splatting has driven significant advances in free-viewpoint video, maintaining its rendering quality with a small memory footprint for efficient streaming transmission still presents an ongoing challenge. Existing streaming dynamic Gaussian Splatting compression methods typically leverage a latent representation to drive the neural network for predicting Gaussian residuals between frames. Their core latent representations can be categorized into structured grid-based and unstructured point-based paradigms. However, the former incurs significant parameter redundancy by inevitably modeling unoccupied space, while the latter suffers from limited compactness as it fails to exploit local correlations. To relieve these limitations, we propose HPC, a novel streaming dynamic Gaussian Splatting compression framework. It employs a hierarchical point-based latent representation that operates on a per-Gaussian basis to avoid parameter redundancy in unoccupied space. Guided by a tailored aggregation scheme, these latent points achieve high compactness with low spatial redundancy. To improve compression efficiency, we further undertake the first investigation to compress neural networks for streaming dynamic Gaussian Splatting through mining and exploiting the inter-frame correlation of parameters. Combined with latent compression, this forms a fully end-to-end compression framework. Comprehensive experimental evaluations demonstrate that HPC substantially outperforms state-of-the-art methods. It achieves a storage reduction of 67% against its baseline while maintaining high reconstruction fidelity.

Method: Five-fold approach: 1) Automatic annotation using large vision-language models with noise-robust contrastive learning; 2) Parameter-efficient fine-tuning with recurrent adapters; 3) Integration of State Space Layers for long-form video modeling; 4) Novel contrastive learning framework for motion-moment relations; 5) Comprehensive empirical study on LVLMs.

Result: Demonstrates that explicit temporal modeling significantly enhances model’s ability to represent and reason about video content. Introduces new long-term benchmarks for egocentric and feature-length content, and identifies visual-language interface as bottleneck for temporal reasoning.

Conclusion: Explicit temporal modeling is crucial for advanced video understanding. The proposed methods and frameworks effectively capture temporal dynamics and address key bottlenecks in current video understanding systems.

Abstract: This thesis explores the central question of how to leverage temporal relations among video elements to advance video understanding. Addressing the limitations of existing methods, the work presents a five-fold contribution: (1) an automatic annotation framework that utilizes large vision-language models and a noise-robust contrastive learning objective with a subtractive angular margin; (2) a parameter-efficient fine-tuning strategy using “recurrent adapters” to capture temporal dynamics in low-data regimes; (3) the integration of State Space Layers (SSL) for efficient long-form video modeling, supported by the introduction of two new long-term benchmarks for egocentric and feature-length content; (4) a novel contrastive learning framework designed to explicitly model fine-grained relations between motions and video moments; and (5) a comprehensive empirical study on Large Vision-Language Models (LVLMs) that identifies the visual-language interface as a bottleneck for temporal reasoning, leading to a new “temporal-oriented recipe” for upscaled video understanding. Collectively, these contributions demonstrate that explicit temporal modeling significantly enhances a model’s ability to represent and reason about the fluid nature of video content.

Method: Proposes V2X-DSC with a Conditional Codec (DCC) framework. Senders compress BEV features into compact codes, while receivers perform conditional reconstruction using local features as side information. This allocates bits to complementary cues rather than redundant content, regularizing learning and encouraging incremental representation.

Result: Experiments on DAIR-V2X, OPV2V, and V2X-Real datasets demonstrate state-of-the-art accuracy-bandwidth trade-offs under KB-level communication. The framework generalizes as a plug-and-play communication layer across multiple fusion backbones.

Conclusion: V2X-DSC effectively addresses bandwidth constraints in collaborative perception by leveraging distributed source coding principles, enabling efficient feature fusion while maintaining high accuracy through conditional reconstruction of complementary information.

Abstract: Collaborative perception improves 3D understanding by fusing multi-agent observations, yet intermediate-feature sharing faces strict bandwidth constraints as dense BEV features saturate V2X links. We observe that collaborators view the same physical world, making their features strongly correlated; thus receivers only need innovation beyond their local context. Revisiting this from a distributed source coding perspective, we propose V2X-DSC, a framework with a Conditional Codec (DCC) for bandwidth-constrained fusion. The sender compresses BEV features into compact codes, while the receiver performs conditional reconstruction using its local features as side information, allocating bits to complementary cues rather than redundant content. This conditional structure regularizes learning, encouraging incremental representation and yielding lower-noise features. Experiments on DAIR-V2X, OPV2V, and V2X-Real demonstrate state-of-the-art accuracy-bandwidth trade-offs under KB-level communication, and generalizes as a plug-and-play communication layer across multiple fusion backbones.

Method: 1) Twin-teacher enhanced training algorithm to transfer text-controllability from foundation models while learning audio-visual synchronization. 2) Dynamic modulation of multi-modal condition strengths (audio/text) based on denoising timesteps to mitigate conflicts between heterogeneous conditioning signals.

Result: Outperforms state-of-the-art models like Omnihuman-1.5 and KlingAvatar 2.0 in GSB evaluation. Enables complex applications including multi-person dialogues and non-human subjects role-playing.

Conclusion: JoyAvatar substantially expands avatar models’ capacity to generate natural, temporally coherent full-body motions and dynamic camera movements while preserving basic avatar capabilities like accurate lip-sync and identity consistency.

Abstract: Existing video avatar models have demonstrated impressive capabilities in scenarios such as talking, public speaking, and singing. However, the majority of these methods exhibit limited alignment with respect to text instructions, particularly when the prompts involve complex elements including large full-body movement, dynamic camera trajectory, background transitions, or human-object interactions. To break out this limitation, we present JoyAvatar, a framework capable of generating long duration avatar videos, featuring two key technical innovations. Firstly, we introduce a twin-teacher enhanced training algorithm that enables the model to transfer inherent text-controllability from the foundation model while simultaneously learning audio-visual synchronization. Secondly, during training, we dynamically modulate the strength of multi-modal conditions (e.g., audio and text) based on the distinct denoising timestep, aiming to mitigate conflicts between the heterogeneous conditioning signals. These two key designs serve to substantially expand the avatar model’s capacity to generate natural, temporally coherent full-body motions and dynamic camera movements as well as preserve the basic avatar capabilities, such as accurate lip-sync and identity consistency. GSB evaluation results demonstrate that our JoyAvatar model outperforms the state-of-the-art models such as Omnihuman-1.5 and KlingAvatar 2.0. Moreover, our approach enables complex applications including multi-person dialogues and non-human subjects role-playing. Some video samples are provided on https://joyavatar.github.io/.

Method: Semi-supervised instance segmentation framework with patch-based preprocessing (splitting high-res images into overlapping small patches) and pseudo-labeling strategy applied to unannotated images, creating additional pseudo-labeled patches.

Result: Substantial performance gains: semantic models improved from 65.93% to 70.35% mIoU; instance models improved from 28.30% to 46.10% AP. Created dataset with 11,060 human-annotated patches and 56,428 pseudo-labeled patches.

Conclusion: Combining patch-based preprocessing with semi-supervised learning significantly improves segmentation of fine stomatal structures, supporting scalable extraction of stomatal traits for AI-driven phenotyping in crop science.

Abstract: Sorghum is a globally important cereal grown widely in water-limited and stress-prone regions. Its strong drought tolerance makes it a priority crop for climate-resilient agriculture. Improving water-use efficiency in sorghum requires precise characterisation of stomatal traits, as stomata control of gas exchange, transpiration and photosynthesis have a major influence on crop performance. Automated analysis of sorghum stomata is difficult because the stomata are small (often less than 40 $μ$m in length in grasses such as sorghum) and vary in shape across genotypes and leaf surfaces. Automated segmentation contributes to high-throughput stomatal phenotyping, yet current methods still face challenges related to nested small structures and annotation bottlenecks. In this paper, we propose a semi-supervised instance segmentation framework tailored for analysis of sorghum stomatal components. We collect and annotate a sorghum leaf imagery dataset containing 11,060 human-annotated patches, covering the three stomatal components (pore, guard cell and complex area) across multiple genotypes and leaf surfaces. To improve the detection of tiny structures, we split high-resolution microscopy images into overlapping small patches. We then apply a pseudo-labelling strategy to unannotated images, producing an additional 56,428 pseudo-labelled patches. Benchmarking across semantic and instance segmentation models shows substantial performance gains: for semantic models the top mIoU increases from 65.93% to 70.35%, whereas for instance models the top AP rises from 28.30% to 46.10%. These results demonstrate that combining patch-based preprocessing with semi-supervised learning significantly improves the segmentation of fine stomatal structures. The proposed framework supports scalable extraction of stomatal traits and facilitates broader adoption of AI-driven phenotyping in crop science.

Method: Three contributions: 1) curated high-quality dataset using train-generate-filter-retrain strategy; 2) diffusion-based framework that disentangles identity and makeup features; 3) text-guided mechanism for fine-grained, region-specific control using natural language prompts.

Result: Experiments on benchmarks and real-world scenarios demonstrate improvements in fidelity, identity preservation, and flexibility compared to existing methods.

Conclusion: The proposed diffusion-based framework with curated dataset and text-guided control effectively addresses limitations in makeup transfer, offering better stability, disentanglement, and controllability than previous approaches.

Abstract: Diffusion models have recently shown strong progress in generative tasks, offering a more stable alternative to GAN-based approaches for makeup transfer. Existing methods often suffer from limited datasets, poor disentanglement between identity and makeup features, and weak controllability. To address these issues, we make three contributions. First, we construct a curated high-quality dataset using a train-generate-filter-retrain strategy that combines synthetic, realistic, and filtered samples to improve diversity and fidelity. Second, we design a diffusion-based framework that disentangles identity and makeup features, ensuring facial structure and skin tone are preserved while applying accurate and diverse cosmetic styles. Third, we propose a text-guided mechanism that allows fine-grained and region-specific control, enabling users to modify eyes, lips, or face makeup with natural language prompts. Experiments on benchmarks and real-world scenarios demonstrate improvements in fidelity, identity preservation, and flexibility. Examples of our dataset can be found at: https://makeup-adapter.github.io.

Method: Uses a diffusion-based algorithm to separate inter and outer layer surfaces from double-layered point clouds, focusing on open-boundary models (surfaces with topological holes) rather than missing surface regions. The method works as a lightweight post-hoc module after TSDF fusion.

Result: Achieves extraction of the inter layer from 20,000 inter and 20,000 outer points in approximately 10 seconds, handling both watertight and open-boundary surface geometries effectively.

Conclusion: Provides a practical solution for accurate surface representation in applications like indoor scene modeling and medical imaging where double-layered point clouds are common, without replacing full reconstruction pipelines.

Abstract: We propose a diffusion-based algorithm for separating the inter and outer layer surfaces from double-layered point clouds, particularly those exhibiting the “double surface artifact” caused by truncation in Truncated Signed Distance Function (TSDF) fusion during indoor or medical 3D reconstruction. This artifact arises from asymmetric truncation thresholds, leading to erroneous inter and outer shells in the fused volume, which our method addresses by extracting the true inter layer to mitigate challenges like overlapping surfaces and disordered normals. We focus on point clouds with \emph{open boundaries} (i.e., sampled surfaces with topological openings/holes through which particles may escape), rather than point clouds with \emph{missing surface regions} where no samples exist. Our approach enables robust processing of both watertight and open-boundary models, achieving extraction of the inter layer from 20,000 inter and 20,000 outer points in approximately 10 seconds. This solution is particularly effective for applications requiring accurate surface representations, such as indoor scene modeling and medical imaging, where double-layered point clouds are prevalent, and it accommodates both closed (watertight) and open-boundary surface geometries. Our goal is \emph{post-hoc} inter/outer shell separation as a lightweight module after TSDF fusion; we do not aim to replace full variational or learning-based reconstruction pipelines.

Method: Frames HSI restoration as autoregressive generation with three innovations: latent-condition alignment for semantic consistency, degradation-aware guidance for encoding mixed degradations, and spatial-spectral adaptation module for cross-domain refinement.

Result: Achieves state-of-the-art performance with 3.77 dB PSNR improvement on ICVL benchmark, 95.5× inference speed-up compared to diffusion methods, and 50% computational cost reduction at inference.

Conclusion: HSI-VAR provides a practical solution for real-world HSI restoration by balancing computational efficiency with high-quality structural preservation through autoregressive modeling.

Abstract: Hyperspectral images (HSIs) capture richer spatial-spectral information beyond RGB, yet real-world HSIs often suffer from a composite mix of degradations, such as noise, blur, and missing bands. Existing generative approaches for HSI restoration like diffusion models require hundreds of iterative steps, making them computationally impractical for high-dimensional HSIs. While regression models tend to produce oversmoothed results, failing to preserve critical structural details. We break this impasse by introducing HSI-VAR, rethinking HSI restoration as an autoregressive generation problem, where spectral and spatial dependencies can be progressively modeled rather than globally reconstructed. HSI-VAR incorporates three key innovations: (1) Latent-condition alignment, which couples semantic consistency between latent priors and conditional embeddings for precise reconstruction; (2) Degradation-aware guidance, which uniquely encodes mixed degradations as linear combinations in the embedding space for automatic control, remarkably achieving a nearly $50%$ reduction in computational cost at inference; (3) A spatial-spectral adaptation module that refines details across both domains in the decoding phase. Extensive experiments on nine all-in-one HSI restoration benchmarks confirm HSI-VAR’s state-of-the-art performance, achieving a 3.77 dB PSNR improvement on \textbf{\textit{ICVL}} and offering superior structure preservation with an inference speed-up of up to $95.5 \times$ compared with diffusion-based methods, making it a highly practical solution for real-world HSI restoration.

Method: U-Net architecture for nerve segmentation in ultrasound images of brachial plexus. Evaluates effects of: 1) training on combined data from multiple ultrasound machines (SIEMENS ACUSON NX3 Elite and Philips EPIQ5), 2) extending binary nerve segmentation to multi-class supervision (artery, vein, nerve, muscle), and 3) correlation between nerve size and segmentation accuracy.

Result: 1) Multi-machine training provides regularization benefits for lower-performing sources but doesn’t surpass single-source training when matched to target domain. 2) Multi-class supervision decreases nerve-specific Dice scores by 9-61% due to class imbalance and boundary ambiguity. 3) Moderate positive correlation between nerve size and segmentation accuracy (Pearson r=0.587, p<0.001), indicating smaller nerves remain challenging.

Conclusion: Provides methodological guidance for developing robust ultrasound nerve segmentation systems under clinical data constraints. Highlights trade-offs in dataset composition, annotation strategies, and challenges with small anatomical structures.

Abstract: Accurate nerve localization is critical for the success of ultrasound-guided regional anesthesia, yet manual identification remains challenging due to low image contrast, speckle noise, and inter-patient anatomical variability. This study evaluates deep learning-based nerve segmentation in ultrasound images of the brachial plexus using a U-Net architecture, with a focus on how dataset composition and annotation strategy influence segmentation performance. We find that training on combined data from multiple ultrasound machines (SIEMENS ACUSON NX3 Elite and Philips EPIQ5) provides regularization benefits for lower-performing acquisition sources, though it does not surpass single-source training when matched to the target domain. Extending the task from binary nerve segmentation to multi-class supervision (artery, vein, nerve, muscle) results in decreased nerve-specific Dice scores, with performance drops ranging from 9% to 61% depending on dataset, likely due to class imbalance and boundary ambiguity. Additionally, we observe a moderate positive correlation between nerve size and segmentation accuracy (Pearson r=0.587, p<0.001), indicating that smaller nerves remain a primary challenge. These findings provide methodological guidance for developing robust ultrasound nerve segmentation systems under realistic clinical data constraints.

Method: Proposes Dual-level Vision-Language Alignment with Reinforcement Learning gating (DVLA-RL) consisting of: 1) Dual-level Semantic Construction (DSC) that generates discriminative attributes from class names and support samples, and 2) RL-gated Attention (RLA) that formulates cross-modal fusion as a sequential decision process using episodic REINFORCE to adaptively adjust self-attention and cross-attention contributions.

Result: Achieves new state-of-the-art performance across nine benchmarks in three diverse few-shot learning scenarios.

Conclusion: DVLA-RL enables more precise cross-modal alignment by refining local attributes in shallow layers and emphasizing global semantics in deep layers, achieving class-specific discrimination and generalized representations with few support samples.

Abstract: Few-shot learning (FSL) aims to generalize to novel categories with only a few samples. Recent approaches incorporate large language models (LLMs) to enrich visual representations with semantic embeddings derived from class names. However, they overlook progressive and adaptive alignment between vision and language from low-level to high-level semantics, resulting in limited semantic gains. To address these challenges, we propose Dual-level Vision-Language Alignment with Reinforcement Learning gating (DVLA-RL), which consists of Dual-level Semantic Construction (DSC) and RL-gated Attention (RLA). Specifically, DSC conditions LLMs on both class names and support samples to generate discriminative attributes, progressively selects the most relevant ones, and then synthesizes them into coherent class descriptions. This process provides complementary low-level attributes and high-level descriptions, enabling both fine-grained grounding and holistic class understanding. To dynamically integrate dual-level semantics along with the visual network layers, RLA formulates cross-modal fusion as a sequential decision process. A lightweight policy trained with episodic REINFORCE adaptively adjusts the contributions of self-attention and cross-attention to integrate textual and visual tokens. As a result, shallow layers refine local attributes and deep layers emphasize global semantics, enabling more precise cross-modal alignment. This achieves class-specific discrimination and generalized representations with merely a few support samples. DVLA-RL achieves new state-of-the-art performance across nine benchmarks in three diverse FSL scenarios.

Method: Proposes Any3D-VLA which: 1) Explicitly lifts visual input into point clouds, 2) Unifies simulator, sensor, and model-estimated point clouds in training pipeline, 3) Constructs diverse inputs, 4) Learns domain-agnostic 3D representations fused with corresponding 2D representations.

Result: Simulation and real-world experiments demonstrate Any3D-VLA’s advantages in improving performance and mitigating domain gaps. Point cloud representations better complement 2D representations for enhanced spatial understanding.

Conclusion: Incorporating 3D point cloud representations significantly enhances VLA model capabilities for complex spatial understanding tasks, with the unified training approach effectively addressing data scarcity and domain gap challenges.

Abstract: Existing Vision-Language-Action (VLA) models typically take 2D images as visual input, which limits their spatial understanding in complex scenes. How can we incorporate 3D information to enhance VLA capabilities? We conduct a pilot study across different observation spaces and visual representations. The results show that explicitly lifting visual input into point clouds yields representations that better complement their corresponding 2D representations. To address the challenges of (1) scarce 3D data and (2) the domain gap induced by cross-environment differences and depth-scale biases, we propose Any3D-VLA. It unifies the simulator, sensor, and model-estimated point clouds within a training pipeline, constructs diverse inputs, and learns domain-agnostic 3D representations that are fused with the corresponding 2D representations. Simulation and real-world experiments demonstrate Any3D-VLA’s advantages in improving performance and mitigating the domain gap. Our project homepage is available at https://xianzhefan.github.io/Any3D-VLA.github.io.

Method: VVLoc uses a single neural network with multi-camera input to first evaluate geo-proximity between visual observations (topological localization), then estimates relative metric poses using a matching strategy, while providing confidence measures. Training only requires visual data pairs with ground-truth poses, eliminating complex supplementary data needs.

Result: VVLoc achieves state-of-the-art localization accuracy across various tasks, demonstrated on both public datasets and a more challenging self-collected dataset, showing robust performance in real-world scenarios.

Conclusion: VVLoc provides a practical, unified solution for vehicle localization that combines topological and metric approaches with confidence estimation, using efficient training with minimal data requirements, making it suitable for industrial autonomous driving applications.

Abstract: Localization is a critical technology in autonomous driving, encompassing both topological localization, which identifies the most similar map keyframe to the current observation, and metric localization, which provides precise spatial coordinates. Conventional methods typically address these tasks independently, rely on single-camera setups, and often require additional 3D semantic or pose priors, while lacking mechanisms to quantify the confidence of localization results, making them less feasible for real industrial applications. In this paper, we propose VVLoc, a unified pipeline that employs a single neural network to concurrently achieve topological and metric vehicle localization using multi-camera system. VVLoc first evaluates the geo-proximity between visual observations, then estimates their relative metric poses using a matching strategy, while also providing a confidence measure. Additionally, the training process for VVLoc is highly efficient, requiring only pairs of visual data and corresponding ground-truth poses, eliminating the need for complex supplementary data. We evaluate VVLoc not only on the publicly available datasets, but also on a more challenging self-collected dataset, demonstrating its ability to deliver state-of-the-art localization accuracy across a wide range of localization tasks.

Method: Paracosm prompts an LMM to generate a synthetic “mental image” from the multimodal query (reference image + modification text). To address synthetic-to-real domain gaps, it also generates synthetic counterparts for each real image in the database. Matching occurs in this synthetic “paracosm” space between generated mental images and synthetic database images.

Result: Paracosm significantly outperforms existing zero-shot methods on four challenging CIR benchmarks, achieving state-of-the-art performance for zero-shot composed image retrieval.

Conclusion: Directly generating mental images for composed image retrieval is more effective than generating textual descriptions, and operating in a synthetic domain space helps overcome domain gaps between generated and real images.

Abstract: Composed Image Retrieval (CIR) is the task of retrieving a target image from a database using a multimodal query, which consists of a reference image and a modification text. The text specifies how to alter the reference image to form a mental image'', based on which CIR should find the target image in the database. The fundamental challenge of CIR is that this mental image’’ is not physically available and is only implicitly defined by the query. The contemporary literature pursues zero-shot methods and uses a Large Multimodal Model (LMM) to generate a textual description for a given multimodal query, and then employs a Vision-Language Model (VLM) for textual-visual matching to search the target image. In contrast, we address CIR from first principles by directly generating the mental image'' for more accurate matching. Particularly, we prompt an LMM to generate a mental image’’ for a given multimodal query and propose to use this mental image'' to search for the target image. As the mental image’’ has a synthetic-to-real domain gap with real images, we also generate a synthetic counterpart for each real image in the database to facilitate matching. In this sense, our method uses LMM to construct a ``paracosm’’, where it matches the multimodal query and database images. Hence, we call this method Paracosm. Notably, Paracosm is a training-free zero-shot CIR method. It significantly outperforms existing zero-shot methods on four challenging benchmarks, achieving state-of-the-art performance for zero-shot CIR.

Method: Proposes identity-agnostic pathology preservation using inversion-free Rectified Flow Transformers (FlowEdit) for near real-time identity transformation. Introduces “Segment-by-Synthesis” to generate counterfactual healthy/pathological twin pairs locally, extracting differential erythema masks decoupled from biometric markers.

Result: Achieves high-fidelity identity transformation in <20s, enabling local deployment. Demonstrates IoU stability >0.67 across synthetic identities on high-resolution clinical samples. Provides privacy-compliant synthetic surrogates that mitigate gradient leakage risk.

Conclusion: The framework enables secure, privacy-preserving federated learning for skin image analysis by generating synthetic surrogates at the edge, balancing privacy protection with diagnostic feature preservation.

Abstract: The deployment of Federated Learning (FL) for clinical dermatology is hindered by the competing requirements of protecting patient privacy and preserving diagnostic features. Traditional de-identification methods often degrade pathological fidelity, while standard generative editing techniques rely on computationally intensive inversion processes unsuitable for resource-constrained edge devices. We propose a framework for identity-agnostic pathology preservation that serves as a client-side privacy-preserving utility. By leveraging inversion-free Rectified Flow Transformers (FlowEdit), the system performs high-fidelity identity transformation in near real-time (less than 20s), facilitating local deployment on clinical nodes. We introduce a “Segment-by-Synthesis” mechanism that generates counterfactual healthy and pathological twin pairs locally. This enables the extraction of differential erythema masks that are decoupled from biometric markers and semantic artifacts (e.g. jewelry). Pilot validation on high-resolution clinical samples demonstrates an Intersection over Union (IoU) stability greater than 0.67 across synthetic identities. By generating privacy-compliant synthetic surrogates at the edge, this framework mitigates the risk of gradient leakage at the source, providing a secure pathway for high-precision skin image analysis in federated environments.

Method: Adapts pre-trained diffusion priors for single-step normal regression, integrates dense visual semantics from DINOv3 via cross-attention for geometric cues, uses multi-task learning objective and wavelet-based regularization to preserve fine structural details.

Result: Significantly outperforms SOTA: on ClearGrasp benchmark reduces mean error by 24.4% and improves 11.25° accuracy by 22.8%; on ClearPose achieves 15.2% reduction in mean error.

Conclusion: TransNormal effectively addresses transparent object normal estimation challenges through diffusion priors enhanced with semantic features and regularization, enabling better robotic manipulation in laboratory settings.

Abstract: Monocular normal estimation for transparent objects is critical for laboratory automation, yet it remains challenging due to complex light refraction and reflection. These optical properties often lead to catastrophic failures in conventional depth and normal sensors, hindering the deployment of embodied AI in scientific environments. We propose TransNormal, a novel framework that adapts pre-trained diffusion priors for single-step normal regression. To handle the lack of texture in transparent surfaces, TransNormal integrates dense visual semantics from DINOv3 via a cross-attention mechanism, providing strong geometric cues. Furthermore, we employ a multi-task learning objective and wavelet-based regularization to ensure the preservation of fine-grained structural details. To support this task, we introduce TransNormal-Synthetic, a physics-based dataset with high-fidelity normal maps for transparent labware. Extensive experiments demonstrate that TransNormal significantly outperforms state-of-the-art methods: on the ClearGrasp benchmark, it reduces mean error by 24.4% and improves 11.25° accuracy by 22.8%; on ClearPose, it achieves a 15.2% reduction in mean error. The code and dataset will be made publicly available at https://longxiang-ai.github.io/TransNormal.

Method: Represent scenes as covariance descriptors on SPD manifold, treat perturbations as congruence transformations, use Riemannian mappings to linearize into Euclidean space, all training-free on fixed backbones.

Result: Achieves highly competitive performance against SOTA baselines, excels in challenging zero-shot scenarios without parameter updates.

Conclusion: Second-order geometric statistics framework provides robust, training-free VPR with strong generalization via geometric stability modeling on SPD manifold.

Abstract: Visual Place Recognition (VPR) demands representations robust to drastic environmental and viewpoint shifts. Current aggregation paradigms, however, either rely on data-hungry supervision or simplistic first-order statistics, often neglecting intrinsic structural correlations. In this work, we propose a Second-Order Geometric Statistics framework that inherently captures geometric stability without training. We conceptualize scenes as covariance descriptors on the Symmetric Positive Definite (SPD) manifold, where perturbations manifest as tractable congruence transformations. By leveraging geometry-aware Riemannian mappings, we project these descriptors into a linearized Euclidean embedding, effectively decoupling signal structure from noise. Our approach introduces a training-free framework built upon fixed, pre-trained backbones, achieving strong zero-shot generalization without parameter updates. Extensive experiments confirm that our method achieves highly competitive performance against state-of-the-art baselines, particularly excelling in challenging zero-shot scenarios.

Method: Two key innovations: (1) Offline caching pipeline that decouples heavy teacher inference from training through compressed supervision signals, and (2) Confidence-aware distillation loss that leverages teacher uncertainty for training on commodity hardware. Proposes a 72M-parameter student model.

Result: Student achieves 9x parameter reduction and 5x inference speedup compared to 650M-parameter teacher. Trainable in under 3 days on single workstation vs. teacher requiring massive GPU clusters for up to a week. Preserves structural consistency and geometric understanding for functional 3D awareness.

Conclusion: Distill3R provides accessible research baseline for labs without large-scale compute, enabling domain-specific training at minimal cost. Not intended to compete with SOTA foundation models, but to democratize 3D vision research and enable efficient edge deployment.

Abstract: While multi-view 3D reconstruction has shifted toward large-scale foundation models capable of inferring globally consistent geometry, their reliance on massive computational clusters for training has created a significant barrier to entry for most academic laboratories. To bridge this compute divide, we introduce Distill3R, a framework designed to distill the geometric reasoning of 3D foundation models into compact students fully trainable on a single workstation. Our methodology centers on two primary innovations: (1) an offline caching pipeline that decouples heavy teacher inference from the training loop through compressed supervision signals, and (2) a confidence-aware distillation loss that leverages teacher uncertainty to enable training on commodity hardware. We propose a 72M-parameter student model which achieves a 9x reduction in parameters and a 5x inference speedup compared to its 650M-parameter teacher. The student is fully trainable in under 3 days on a single workstation, whereas its teacher requires massive GPU clusters for up to a week. We demonstrate that the student preserves the structural consistency and qualitative geometric understanding required for functional 3D awareness. By providing a reproducible, single-workstation training recipe, Distill3R serves as an exploratory entry point for democratized 3D vision research and efficient edge deployment. This work is not intended to compete with state-of-the-art foundation models, but to provide an accessible research baseline for laboratories without access to large-scale compute to train and specialize models on their own domain-specific data at minimal cost.

Method: DIAMOND uses trajectory correction during inference by reconstructing an estimate of the clean sample at every step of the generative trajectory, actively steering generation away from latent states that lead to artifacts. It’s training-free and extends to standard Diffusion Models.

Result: The method provides a robust, zero-shot path to high-fidelity, artifact-free image synthesis without additional training or weight modifications in modern generative architectures.

Conclusion: DIAMOND offers an effective training-free solution for artifact reduction in text-to-image generation by intervening during the core image formation process rather than working post-hoc.

Abstract: Despite impressive results from recent text-to-image models like FLUX, visual and anatomical artifacts remain a significant hurdle for practical and professional use. Existing methods for artifact reduction, typically work in a post-hoc manner, consequently failing to intervene effectively during the core image formation process. Notably, current techniques require problematic and invasive modifications to the model weights, or depend on a computationally expensive and time-consuming process of regional refinement. To address these limitations, we propose DIAMOND, a training-free method that applies trajectory correction to mitigate artifacts during inference. By reconstructing an estimate of the clean sample at every step of the generative trajectory, DIAMOND actively steers the generation process away from latent states that lead to artifacts. Furthermore, we extend the proposed method to standard Diffusion Models, demonstrating that DIAMOND provides a robust, zero-shot path to high-fidelity, artifact-free image synthesis without the need for additional training or weight modifications in modern generative architectures. Code is available at https://gmum.github.io/DIAMOND/

Method: OCTOPUS performs discrete recurrence along eight principal orientations (forward/backward in horizontal, vertical, and diagonal directions) to enable effective information exchange across all spatially connected regions while maintaining independence among unrelated patches.

Result: OCTOPUS demonstrates notable improvements in boundary preservation and region consistency in segmentation tasks, while maintaining relatively better classification accuracy compared to existing V-SSM based models.

Conclusion: OCTOPUS appears as a foundation method for multi-directional recurrence as a scalable and effective mechanism for building spatially aware and computationally efficient vision architectures.

Abstract: State space models (SSMs) have recently emerged as an alternative to transformers due to their unique ability of modeling global relationships in text with linear complexity. However, their success in vision tasks has been limited due to their causal formulation, which is suitable for sequential text but detrimental in the spatial domain where causality breaks the inherent spatial relationships among pixels or patches. As a result, standard SSMs fail to capture local spatial coherence, often linking non-adjacent patches while ignoring neighboring ones that are visually correlated. To address these limitations, we introduce OCTOPUS , a novel architecture that preserves both global context and local spatial structure within images, while maintaining the linear complexity of SSMs. OCTOPUS performs discrete reoccurrence along eight principal orientations, going forward or backward in the horizontal, vertical, and diagonal directions, allowing effective information exchange across all spatially connected regions while maintaining independence among unrelated patches. This design enables multi-directional recurrence, capturing both global context and local spatial structure with SSM-level efficiency. In our classification and segmentation benchmarks, OCTOPUS demonstrates notable improvements in boundary preservation and region consistency, as evident from the segmentation results, while maintaining relatively better classification accuracy compared to existing V-SSM based models. These results suggest that OCTOPUS appears as a foundation method for multi-directional recurrence as a scalable and effective mechanism for building spatially aware and computationally efficient vision architectures.

Method: Proposes ConsensusDrop, a training-free framework that derives consensus ranking by reconciling vision encoder saliency with query-aware cross-attention. Retains most informative tokens while compressing remainder via encoder-guided token merging.

Result: Outperforms prior pruning methods under identical token budgets across LLaVA-1.5/NeXT, Video-LLaVA, and other open-source VLMs. Delivers stronger accuracy-efficiency Pareto frontier, preserving near-baseline accuracy even at aggressive token reductions while reducing TTFT and KV cache footprint.

Conclusion: ConsensusDrop provides an effective training-free approach for visual token reduction in VLMs by fusing complementary vision and language signals, achieving better efficiency-accuracy trade-offs than unimodal methods.

Abstract: Vision-Language Models (VLMs) are expensive because the LLM processes hundreds of largely redundant visual tokens. Existing token reduction methods typically exploit \textit{either} vision-encoder saliency (broad but query-agnostic) \textit{or} LLM cross-attention (query-aware but sparse and costly). We show that neither signal alone is sufficient: fusing them consistently improves performance compared to unimodal visual token selection (ranking). However, making such fusion practical is non-trivial: cross-modal saliency is usually only available \emph{inside} the LLM (too late for efficient pre-LLM pruning), and the two signals are inherently asymmetric, so naive fusion underutilizes their complementary strengths. We propose \textbf{ConsensusDrop}, a training-free framework that derives a \emph{consensus} ranking by reconciling vision encoder saliency with query-aware cross-attention, retaining the most informative tokens while compressing the remainder via encoder-guided token merging. Across LLaVA-1.5/NeXT, Video-LLaVA, and other open-source VLMs, ConsensusDrop consistently outperforms prior pruning methods under identical token budgets and delivers a stronger accuracy-efficiency Pareto frontier – preserving near-baseline accuracy even at aggressive token reductions while reducing TTFT and KV cache footprint. Our code will be open-sourced.

Method: Two complementary data-augmentation frameworks: 1) Instance Aware Automatic Augmentation (IAAA) - automated, instance-preserving augmentation method that generates synthetic images and masks by applying optimized augmentation policies; 2) Semantic-Aware AI Augmentation (SAAA) - combines diffusion-based mask generator with Pix2Pix conditional image synthesizer to create diverse, anatomically realistic segmentation masks and corresponding high-fidelity images.

Result: The augmentation strategies generate synthetic images whose visual and structural properties closely match real IS data, significantly improving CAR-T/NK IS detection and segmentation performance, enhancing robustness and accuracy of IS quantification.

Conclusion: This work supports the development of more reliable imaging-based biomarkers for predicting patient response to CAR-T/NK immunotherapy by addressing data limitations through advanced synthetic data generation techniques.

Abstract: Chimeric antigen receptor (CAR)-T and NK cell immunotherapies have transformed cancer treatment, and recent studies suggest that the quality of the CAR-T/NK cell immunological synapse (IS) may serve as a functional biomarker for predicting therapeutic efficacy. Accurate detection and segmentation of CAR-T/NK IS structures using artificial neural networks (ANNs) can greatly increase the speed and reliability of IS quantification. However, a persistent challenge is the limited size of annotated microscopy datasets, which restricts the ability of ANNs to generalize. To address this challenge, we integrate two complementary data-augmentation frameworks. First, we employ Instance Aware Automatic Augmentation (IAAA), an automated, instance-preserving augmentation method that generates synthetic CAR-T/NK IS images and corresponding segmentation masks by applying optimized augmentation policies to original IS data. IAAA supports multiple imaging modalities (e.g., fluorescence and brightfield) and can be applied directly to CAR-T/NK IS images derived from patient samples. In parallel, we introduce a Semantic-Aware AI Augmentation (SAAA) pipeline that combines a diffusion-based mask generator with a Pix2Pix conditional image synthesizer. This second method enables the creation of diverse, anatomically realistic segmentation masks and produces high-fidelity CAR-T/NK IS images aligned with those masks, further expanding the training corpus beyond what IAAA alone can provide. Together, these augmentation strategies generate synthetic images whose visual and structural properties closely match real IS data, significantly improving CAR-T/NK IS detection and segmentation performance. By enhancing the robustness and accuracy of IS quantification, this work supports the development of more reliable imaging-based biomarkers for predicting patient response to CAR-T/NK immunotherapy.

Method: Proposes a novel hybrid framework integrating TDA with DenseNet121 backbone. TDA captures topological characteristics of brain structures, while DenseNet121 learns hierarchical spatial features from MRI slices. The extracted deep and topological features are fused for enhanced classification.

Result: Extensive experiments on OASIS-1 Kaggle MRI dataset show the TDA+DenseNet121 model significantly outperforms existing state-of-the-art approaches, achieving 99.93% accuracy and 100% AUC, surpassing CNN-based, transfer learning, ensemble, and multi-scale architectures.

Conclusion: The results confirm the effectiveness of incorporating topological insights into deep learning pipelines and highlight the potential of the proposed framework as a robust and highly accurate tool for automated Alzheimer’s disease diagnosis.

Abstract: Early and accurate diagnosis of Alzheimer’s disease (AD) remains a critical challenge in neuroimaging-based clinical decision support systems. In this work, we propose a novel hybrid deep learning framework that integrates Topological Data Analysis (TDA) with a DenseNet121 backbone for four-class Alzheimer’s disease classification using structural MRI data from the OASIS dataset. TDA is employed to capture complementary topological characteristics of brain structures that are often overlooked by conventional neural networks, while DenseNet121 efficiently learns hierarchical spatial features from MRI slices. The extracted deep and topological features are fused to enhance class separability across the four AD stages. Extensive experiments conducted on the OASIS-1 Kaggle MRI dataset demonstrate that the proposed TDA+DenseNet121 model significantly outperforms existing state-of-the-art approaches. The model achieves an accuracy of 99.93% and an AUC of 100%, surpassing recently published CNN-based, transfer learning, ensemble, and multi-scale architectures. These results confirm the effectiveness of incorporating topological insights into deep learning pipelines and highlight the potential of the proposed framework as a robust and highly accurate tool for automated Alzheimer’s disease diagnosis.

Method: 1) HitEmotion benchmark with hierarchical cognitive depth levels; 2) ToM-guided reasoning chain tracking mental states and calibrating cross-modal evidence; 3) TMPO reinforcement learning using intermediate mental states as process-level supervision.

Result: HitEmotion exposes deep emotional reasoning deficits in state-of-the-art models, especially on cognitively demanding tasks. The proposed methods improve end-task accuracy and yield more faithful, coherent rationales.

Conclusion: The work provides a practical toolkit for evaluating and enhancing cognition-based emotional understanding capabilities in multimodal LLMs, bridging the gap between cognitive science and AI.

Abstract: Despite rapid progress in multimodal large language models (MLLMs), their capability for deep emotional understanding remains limited. We argue that genuine affective intelligence requires explicit modeling of Theory of Mind (ToM), the cognitive substrate from which emotions arise. To this end, we introduce HitEmotion, a ToM-grounded hierarchical benchmark that diagnoses capability breakpoints across increasing levels of cognitive depth. Second, we propose a ToM-guided reasoning chain that tracks mental states and calibrates cross-modal evidence to achieve faithful emotional reasoning. We further introduce TMPO, a reinforcement learning method that uses intermediate mental states as process-level supervision to guide and strengthen model reasoning. Extensive experiments show that HitEmotion exposes deep emotional reasoning deficits in state-of-the-art models, especially on cognitively demanding tasks. In evaluation, the ToM-guided reasoning chain and TMPO improve end-task accuracy and yield more faithful, more coherent rationales. In conclusion, our work provides the research community with a practical toolkit for evaluating and enhancing the cognition-based emotional understanding capabilities of MLLMs. Our dataset and code are available at: https://HitEmotion.github.io/.

Method: For Track 1 (Visual Robustness): lightweight two-layer CNN enhanced by Gated Linear Units (GLUs) and observation normalization. For Track 2 (Neural Alignment): deep ResNet-like architecture with 16 convolutional layers and GLU-based gating. Systematic analysis of ten model checkpoints trained between 60K to 1.14M steps.

Result: Achieved 95.4% final score for visual robustness with lightweight CNN, and top-1 neural prediction performance for neural alignment with 17.8 million parameters. Found optimal training duration around 200K steps with non-monotonic relationship between training steps and performance.

Conclusion: Simpler architectures excel at visual robustness while deeper models with increased capacity achieve better neural alignment, challenging conventional assumptions about model complexity in visuomotor learning and providing practical guidance for biologically-inspired visual agents.

Abstract: Visual robustness and neural alignment remain critical challenges in developing artificial agents that can match biological vision systems. We present the winning approaches from Team HCMUS_TheFangs for both tracks of the NeurIPS 2025 Mouse vs. AI: Robust Visual Foraging Competition. For Track 1 (Visual Robustness), we demonstrate that architectural simplicity combined with targeted components yields superior generalization, achieving 95.4% final score with a lightweight two-layer CNN enhanced by Gated Linear Units and observation normalization. For Track 2 (Neural Alignment), we develop a deep ResNet-like architecture with 16 convolutional layers and GLU-based gating that achieves top-1 neural prediction performance with 17.8 million parameters. Our systematic analysis of ten model checkpoints trained between 60K to 1.14M steps reveals that training duration exhibits a non-monotonic relationship with performance, with optimal results achieved around 200K steps. Through comprehensive ablation studies and failure case analysis, we provide insights into why simpler architectures excel at visual robustness while deeper models with increased capacity achieve better neural alignment. Our results challenge conventional assumptions about model complexity in visuomotor learning and offer practical guidance for developing robust, biologically-inspired visual agents.

Method: Proposes a 2.5D U-Net architecture that takes neighboring B-scans as input to reconstruct corrupted center B-scans, guided by a novel Vessel-Aware Multi-Axis Orthogonal Supervision (VAMOS) loss combining vessel-weighted intensity reconstruction with axial and lateral projection consistency.

Result: VAMOS-OCTA consistently outperforms prior methods, producing reconstructions with sharp capillaries, restored vessel continuity, and clean en face projections, demonstrating superior performance on both synthetic and real-world corrupted volumes.

Conclusion: Multi-axis supervision offers a powerful constraint for restoring motion-degraded 3D OCTA data, enabling joint enhancement of both cross-sectional B-scan sharpness and volumetric projection accuracy even under severe motion corruptions.

Abstract: Handheld Optical Coherence Tomography Angiography (OCTA) enables noninvasive retinal imaging in uncooperative or pediatric subjects, but is highly susceptible to motion artifacts that severely degrade volumetric image quality. Sudden motion during 3D acquisition can lead to unsampled retinal regions across entire B-scans (cross-sectional slices), resulting in blank bands in en face projections. We propose VAMOS-OCTA, a deep learning framework for inpainting motion-corrupted B-scans using vessel-aware multi-axis supervision. We employ a 2.5D U-Net architecture that takes a stack of neighboring B-scans as input to reconstruct a corrupted center B-scan, guided by a novel Vessel-Aware Multi-Axis Orthogonal Supervision (VAMOS) loss. This loss combines vessel-weighted intensity reconstruction with axial and lateral projection consistency, encouraging vascular continuity in native B-scans and across orthogonal planes. Unlike prior work that focuses primarily on restoring the en face MIP, VAMOS-OCTA jointly enhances both cross-sectional B-scan sharpness and volumetric projection accuracy, even under severe motion corruptions. We trained our model on both synthetic and real-world corrupted volumes and evaluated its performance using both perceptual quality and pixel-wise accuracy metrics. VAMOS-OCTA consistently outperforms prior methods, producing reconstructions with sharp capillaries, restored vessel continuity, and clean en face projections. These results demonstrate that multi-axis supervision offers a powerful constraint for restoring motion-degraded 3D OCTA data. Our source code is available at https://github.com/MedICL-VU/VAMOS-OCTA.

Method: Proposes CortiNet, a dual-stream architecture inspired by human visual cortex parallel processing. Separates low-frequency structural information from high-frequency perceptual details through specialized encoding streams, using frequency-selective representations rather than raw pixels. Includes late-stage fusion and a structure-aware explainability framework using gradient-weighted class activation mapping only on the structural branch.

Result: Achieves 98.74% diagnostic accuracy on 10,692 expert-annotated images spanning nine gallbladder disease categories, with significantly reduced parameters compared to conventional deep convolutional models.

Conclusion: CortiNet demonstrates that lightweight, physics-inspired architectures can achieve high diagnostic accuracy for medical imaging tasks while being computationally efficient and interpretable, addressing deployment challenges in clinical settings.

Abstract: Ultrasound imaging is the primary diagnostic modality for detecting Gallbladder diseases due to its non-invasive nature, affordability, and wide accessibility. However, the low resolution and speckle noise inherent to ultrasound images hinder diagnostic reliability, prompting the use of large convolutional neural networks that are difficult to deploy in routine clinical settings. In this work, we propose CortiNet, a lightweight, cortical-inspired dual-stream neural architecture for gallbladder disease diagnosis that integrates physically interpretable multi-scale signal decomposition with perception-driven feature learning. Inspired by parallel processing pathways in the human visual cortex, CortiNet explicitly separates low-frequency structural information from high-frequency perceptual details and processes them through specialized encoding streams. By operating directly on structured, frequency-selective representations rather than raw pixel intensities, the architecture embeds strong physics-based inductive bias, enabling efficient feature learning with a significantly reduced parameter footprint. A late-stage cortical-style fusion mechanism integrates complementary structural and textural cues while preserving computational efficiency. Additionally, we propose a structure-aware explainability framework wherein gradient-weighted class activation mapping is only applied to the structural branch of the proposed CortiNet architecture. This choice allows the model to only focus on the structural features, making it robust against speckle noise. We evaluate CortiNet on 10,692 expert-annotated images spanning nine clinically relevant gallbladder disease categories. Experimental results demonstrate that CortiNet achieves high diagnostic accuracy (98.74%) with only a fraction of the parameters required by conventional deep convolutional models.

Method: Proposes SRVAU-R1 with: 1) First reflection-oriented Chain-of-Thought dataset for VAU with structured supervision (initial reasoning, self-reflection, revised reasoning), 2) Reflection-aware learning paradigm with supervised fine-tuning and reinforcement fine-tuning to enhance multi-modal reasoning.

Result: Extensive experiments on multiple video anomaly benchmarks show SRVAU-R1 consistently outperforms existing methods with significant improvements in both temporal anomaly localization accuracy and reasoning quality.

Conclusion: The reflection-aware learning framework effectively enhances MLLM reasoning capabilities for video anomaly understanding through structured self-reflection mechanisms.

Abstract: Multi-modal large language models (MLLMs) have demonstrated significant progress in reasoning capabilities and shown promising effectiveness in video anomaly understanding (VAU) tasks. However, existing MLLM-based approaches remain largely focused on surface-level descriptions of anomalies, lacking deep reasoning over abnormal behaviors like explicit self-reflection and self-correction. To address that, we propose Self-Reflection-Enhanced Reasoning for Video Anomaly Understanding (SRVAU-R1), a reflection-aware learning framework that incorporates reflection in MLLM reasoning. Specifically, SRVAU-R1 introduces the first reflection-oriented Chain-of-Thought dataset tailored for VAU, providing structured supervision with initial reasoning, self-reflection, and revised reasoning. Based on that, it includes a novel reflection-aware learning paradigm with supervised fine-tuning and reinforcement fine-tuning to enhance multi-modal reasoning for VAU. Extensive experiments on multiple video anomaly benchmarks demonstrate that SRVAU-R1 consistently outperforms existing methods, achieving significant improvements in both temporal anomaly localization accuracy and reasoning quality.

Method: Proposes LocalScore, a simple scoring algorithm that explicitly incorporates local density of gallery feature distribution using k-th nearest neighbors. The method is architecture-agnostic, loss-independent, and adds negligible computational overhead, making it plug-and-play for existing systems.

Result: Extensive experiments across multiple modalities show LocalScore achieves substantial gains: open-set retrieval FNIR@FPIR reduced from 53% to 40%, and verification TAR@FAR improved from 51% to 74%. Theoretical analysis explains when the method achieves most significant gains based on dataset characteristics.

Conclusion: LocalScore effectively addresses open-set biometric challenges by incorporating local gallery density, providing consistent performance improvements across modalities with minimal computational cost, making it a practical solution for real-world deployments.

Abstract: Open-set biometrics faces challenges with probe subjects who may not be enrolled in the gallery, as traditional biometric systems struggle to detect these non-mated probes. Despite the growing prevalence of multi-sample galleries in real-world deployments, most existing methods collapse intra-subject variability into a single global representation, leading to suboptimal decision boundaries and poor open-set robustness. To address this issue, we propose LocalScore, a simple yet effective scoring algorithm that explicitly incorporates the local density of the gallery feature distribution using the k-th nearest neighbors. LocalScore is architecture-agnostic, loss-independent, and incurs negligible computational overhead, making it a plug-and-play solution for existing biometric systems. Extensive experiments across multiple modalities demonstrate that LocalScore consistently achieves substantial gains in open-set retrieval (FNIR@FPIR reduced from 53% to 40%) and verification (TAR@FAR improved from 51% to 74%). We further provide theoretical analysis and empirical validation explaining when and why the method achieves the most significant gains based on dataset characteristics.

Method: Trained deep learning models on three datasets: 1) manually annotated dataset, 2) automatically-curated dataset, and 3) accurate subset of automatically-curated data. Compared performance using AUC metrics with statistical significance testing.

Result: Automatically-curated dataset trained model achieved AUC of 0.694 (CI: 0.67-0.73), significantly better than manually annotated dataset’s AUC of 0.643 (CI: 0.62-0.66). Accurate subset performed similarly (AUC 0.689) to full automatically-curated dataset.

Conclusion: Automatically-curated datasets substantially improve deep learning algorithm performance for thyroid nodule classification, and using all automatically-curated data is better than using only accurate subsets.

Abstract: The diagnosis of thyroid nodule cancers commonly utilizes ultrasound images. Several studies showed that deep learning algorithms designed to classify benign and malignant thyroid nodules could match radiologists’ performance. However, data availability for training deep learning models is often limited due to the significant effort required to curate such datasets. The previous study proposed a method to curate thyroid nodule datasets automatically. It was tested to have a 63% yield rate and 83% accuracy. However, the usefulness of the generated data for training deep learning models remains unknown. In this study, we conducted experiments to determine whether using a automatically-curated dataset improves deep learning algorithms’ performance. We trained deep learning models on the manually annotated and automatically-curated datasets. We also trained with a smaller subset of the automatically-curated dataset that has higher accuracy to explore the optimum usage of such dataset. As a result, the deep learning model trained on the manually selected dataset has an AUC of 0.643 (95% confidence interval [CI]: 0.62, 0.66). It is significantly lower than the AUC of the 6automatically-curated dataset trained deep learning model, 0.694 (95% confidence interval [CI]: 0.67, 0.73, P < .001). The AUC of the accurate subset trained deep learning model is 0.689 (95% confidence interval [CI]: 0.66, 0.72, P > .43), which is insignificantly worse than the AUC of the full automatically-curated dataset. In conclusion, we showed that using a automatically-curated dataset can substantially increase the performance of deep learning algorithms, and it is suggested to use all the data rather than only using the accurate subset.

Method: GMAC models extrinsics as global variables constrained by latent multi-view geometric structure, prunes and reconfigures existing networks to use their latent features for extrinsic prediction via lightweight regression head, and jointly optimizes cross-view reprojection consistency and multi-view cycle consistency.

Result: Experiments on synthetic and real-world multi-camera datasets show GMAC achieves accurate and stable extrinsic estimation without explicit 3D reconstruction or manual calibration.

Conclusion: GMAC provides a new solution for efficient deployment and online calibration of multi-camera systems by leveraging implicit geometric representations from existing networks.

Abstract: Automatic calibration of multi-camera systems, namely the accurate estimation of spatial extrinsic parameters, is fundamental for 3D reconstruction, panoramic perception, and multi-view data fusion. Existing methods typically rely on calibration targets, explicit geometric modeling, or task-specific neural networks. Such approaches often exhibit limited robustness and applicability in complex dynamic environments or online scenarios, making them difficult to deploy in practical applications. To address this, this paper proposes GMAC, a multi-camera extrinsic estimation framework based on the implicit geometric representations learned by multi-view reconstruction networks. GMAC models extrinsics as global variables constrained by the latent multi-view geometric structure and prunes and structurally reconfigures existing networks so that their latent features can directly support extrinsic prediction through a lightweight regression head, without requiring a completely new network design. Furthermore, GMAC jointly optimizes cross-view reprojection consistency and multi-view cycle consistency, ensuring geometric coherence across cameras while improving prediction accuracy and optimization stability. Experiments on both synthetic and real-world multi-camera datasets demonstrate that GMAC achieves accurate and stable extrinsic estimation without explicit 3D reconstruction or manual calibration, providing a new solution for efficient deployment and online calibration of multi-camera systems.

Method: Frame-wise, stateless framework where each frame independently generates point cloud fragments, fused via measurement confidence and 3D distance consistency weights. Uses adaptive spatial hashing-based weighted aggregation: 3D space is partitioned by local point density, representative points are selected per cell, and weighted fusion handles sparse/dense regions with GPU parallelization.

Result: Achieves high-throughput, low-latency point cloud generation with linear complexity. Improves reconstruction stability and geometric fidelity in overlapping, depth-discontinuous, and dynamic scenes while maintaining real-time frame rates on modern GPUs.

Conclusion: FUSE-Flow effectively addresses the challenges of real-time multi-view point cloud reconstruction, demonstrating robustness, scalability, and practical applicability for immersive perception applications.

Abstract: Real-time multi-view point cloud reconstruction is a core problem in 3D vision and immersive perception, with wide applications in VR, AR, robotic navigation, digital twins, and computer interaction. Despite advances in multi-camera systems and high-resolution depth sensors, fusing large-scale multi-view depth observations into high-quality point clouds under strict real-time constraints remains challenging. Existing methods relying on voxel-based fusion, temporal accumulation, or global optimization suffer from high computational complexity, excessive memory usage, and limited scalability, failing to simultaneously achieve real-time performance, reconstruction quality, and multi-camera extensibility. We propose FUSE-Flow, a frame-wise, stateless, and linearly scalable point cloud streaming reconstruction framework. Each frame independently generates point cloud fragments, fused via two weights, measurement confidence and 3D distance consistency to suppress noise while preserving geometric details. For large-scale multi-camera efficiency, we introduce an adaptive spatial hashing-based weighted aggregation method: 3D space is adaptively partitioned by local point cloud density, representative points are selected per cell, and weighted fusion is performed to handle both sparse and dense regions. With GPU parallelization, FUSE-Flow achieves high-throughput, low-latency point cloud generation and fusion with linear complexity. Experiments demonstrate that the framework improves reconstruction stability and geometric fidelity in overlapping, depth-discontinuous, and dynamic scenes, while maintaining real-time frame rates on modern GPUs, verifying its effectiveness, robustness, and scalability.

Method: Proposes Visual Expert Quantization (VEQ) with two components: 1) Modality-expert-aware Quantization uses expert activation frequency to prioritize error minimization for pivotal experts; 2) Modality-affinity-aware Quantization constructs enhanced Hessian matrix integrating token-expert affinity with modality information to guide calibration.

Result: Extensive experiments show VEQ consistently outperforms SOTA baselines. Under W3A16 configuration, achieves average accuracy gains of 2.04% on Kimi-VL and 3.09% on Qwen3-VL compared to previous SOTA quantization methods, demonstrating superior robustness across multimodal tasks.

Conclusion: VEQ effectively addresses the dual heterogeneity in MoE VLMs through modality-aware quantization techniques, achieving significant performance improvements while reducing computational costs, making large-scale multimodal models more practical.

Abstract: Mixture-of-Experts(MoE) Vision-Language Models (VLMs) offer remarkable performance but incur prohibitive memory and computational costs, making compression essential. Post-Training Quantization (PTQ) is an effective training-free technique to address the massive memory and computation overhead. Existing quantization paradigms fall short as they are oblivious to two critical forms of heterogeneity: the inherent discrepancy between vision and language tokens, and the non-uniform contribution of different experts. To bridge this gap, we propose Visual Expert Quantization (VEQ), a dual-aware quantization framework designed to simultaneously accommodate cross-modal differences and heterogeneity between experts. Specifically, VEQ incorporates 1)Modality-expert-aware Quantization, which utilizes expert activation frequency to prioritize error minimization for pivotal experts, and 2)Modality-affinity-aware Quantization, which constructs an enhanced Hessian matrix by integrating token-expert affinity with modality information to guide the calibration process. Extensive experiments across diverse benchmarks verify that VEQ consistently outperforms state-of-the-art baselines. Specifically, under the W3A16 configuration, our method achieves significant average accuracy gains of 2.04% on Kimi-VL and 3.09% on Qwen3-VL compared to the previous SOTA quantization methods, demonstrating superior robustness across various multimodal tasks. Our code will be available at https://github.com/guangshuoqin/VEQ.

Method: A fully automatic pipeline using large language models to transform single-person instructional videos into two-person multimodal task-guidance conversations, creating expert-novice interactions from existing video content.

Result: Created HowToDIV dataset with 507 conversations, 6,636 question-answer pairs, and 24 hours of video across multiple domains, with baseline results using Gemma 3 and Qwen 2.5 models.

Conclusion: The framework provides a scalable, cost-efficient alternative to traditional data collection for multimodal conversational datasets, enabling progress in AI agents for AR task assistance.

Abstract: Many everyday tasks, ranging from appliance repair and cooking to car maintenance, require expert knowledge, particularly for complex, multi-step procedures. Despite growing interest in AI agents for augmented reality (AR) assistance, progress remains limited by the scarcity of large-scale multimodal conversational datasets grounded in real-world task execution, in part due to the cost and logistical complexity of human-assisted data collection. In this paper, we present a framework to automatically transform single person instructional videos into two-person multimodal task-guidance conversations. Our fully automatic pipeline, based on large language models, provides a scalable and cost efficient alternative to traditional data collection approaches. Using this framework, we introduce HowToDIV, a multimodal dataset comprising 507 conversations, 6,636 question answer pairs, and 24 hours of video spanning multiple domains. Each session consists of a multi-turn expert-novice interaction. Finally, we report baseline results using Gemma 3 and Qwen 2.5 on HowToDIV, providing an initial benchmark for multimodal procedural task assistance.

Method: ReLayout introduces a relation graph capturing position and size relationships among unedited elements as constraints for structure preservation. It uses relation-aware design reconstruction (RADR) which learns to reconstruct designs from elements, relation graphs, and synthesized editing operations in a self-supervised manner. A multimodal large language model serves as the backbone, unifying multiple editing actions within a single model.

Result: Qualitative, quantitative results and user studies show that ReLayout significantly outperforms baseline models in editing quality, accuracy, and layout structure preservation.

Conclusion: ReLayout provides an effective framework for versatile, structure-preserving design layout editing without requiring triplet training data, advancing automated redesign workflows.

Abstract: Automated redesign without manual adjustments marks a key step forward in the design workflow. In this work, we focus on a foundational redesign task termed design layout editing, which seeks to autonomously modify the geometric composition of a design based on user intents. To overcome the ambiguity of user needs expressed in natural language, we introduce four basic and important editing actions and standardize the format of editing operations. The underexplored task presents a unique challenge: satisfying specified editing operations while simultaneously preserving the layout structure of unedited elements. Besides, the scarcity of triplet (original design, editing operation, edited design) samples poses another formidable challenge. To this end, we present ReLayout, a novel framework for versatile and structure-preserving design layout editing that operates without triplet data. Specifically, ReLayout first introduces the relation graph, which contains the position and size relationships among unedited elements, as the constraint for layout structure preservation. Then, relation-aware design reconstruction (RADR) is proposed to bypass the data challenge. By learning to reconstruct a design from its elements, a relation graph, and a synthesized editing operation, RADR effectively emulates the editing process in a self-supervised manner. A multi-modal large language model serves as the backbone for RADR, unifying multiple editing actions within a single model and thus achieving versatile editing after fine-tuning. Qualitative, quantitative results and user studies show that ReLayout significantly outperforms the baseline models in terms of editing quality, accuracy, and layout structure preservation.

Method: Residual Decoding (ResDec) leverages historical information during decoding, using the model’s internal implicit reasoning mechanism and token logits evolution to correct biases without additional training.

Result: ResDec effectively suppresses hallucinations from language priors, improves visual grounding, reduces object hallucinations, and performs well on comprehensive LVLM benchmarks.

Conclusion: ResDec is a broadly applicable, training-free approach that addresses hallucination issues in LVLMs while maintaining strong performance across multimodal benchmarks.

Abstract: Large Vision-Language Models (LVLMs) can reason effectively from image-text inputs and perform well in various multimodal tasks. Despite this success, they are affected by language priors and often produce hallucinations. Hallucinations denote generated content that is grammatically and syntactically coherent, yet bears no match or direct relevance to actual visual input. To address this problem, we propose Residual Decoding (ResDec). It is a novel training-free method that uses historical information to aid decoding. The method relies on the internal implicit reasoning mechanism and token logits evolution mechanism of LVLMs to correct biases. Extensive experiments demonstrate that ResDec effectively suppresses hallucinations induced by language priors, significantly improves visual grounding, and reduces object hallucinations. In addition to mitigating hallucinations, ResDec also performs exceptionally well on comprehensive LVLM benchmarks, highlighting its broad applicability.

Method: Uses a Multi-Head Multi-Task Learning (MH-MTL) framework with an ImageNet-pretrained EfficientNet-B4 backbone and Feature Pyramid Network (FPN) for multi-scale feature extraction. Implements task-specific routing where global tasks use high-level semantic features and dense prediction tasks use spatially detailed FPN representations. Training includes composite loss with task-adaptive learning rate scaling and cosine annealing schedule.

Result: Validation results demonstrate the feasibility and robustness of the unified design, establishing a strong and extensible baseline for ultrasound foundation model research.

Conclusion: The proposed framework provides a viable approach for developing generalizable ultrasound analysis models that can handle multiple tasks within a single shared network, addressing the heterogeneity challenge in ultrasound imaging.

Abstract: Ultrasound (US) imaging exhibits substantial heterogeneity across anatomical structures and acquisition protocols, posing significant challenges to the development of generalizable analysis models. Most existing methods are task-specific, limiting their suitability as clinically deployable foundation models. To address this limitation, the Foundation Model Challenge for Ultrasound Image Analysis (FM_UIA2026) introduces a large-scale multi-task benchmark comprising 27 subtasks across segmentation, classification, detection, and regression. In this paper, we present the official baseline for FM_UIA2026 based on a unified Multi-Head Multi-Task Learning (MH-MTL) framework that supports all tasks within a single shared network. The model employs an ImageNet-pretrained EfficientNet–B4 backbone for robust feature extraction, combined with a Feature Pyramid Network (FPN) to capture multi-scale contextual information. A task-specific routing strategy enables global tasks to leverage high-level semantic features, while dense prediction tasks exploit spatially detailed FPN representations. Training incorporates a composite loss with task-adaptive learning rate scaling and a cosine annealing schedule. Validation results demonstrate the feasibility and robustness of this unified design, establishing a strong and extensible baseline for ultrasound foundation model research. The code and dataset are publicly available at \href{https://github.com/lijiake2408/Foundation-Model-Challenge-for-Ultrasound-Image-Analysis}{GitHub}.

Method: Proposes a tomographic reconstruction framework based on 3D Gaussian ray tracing instead of splatting. Computes line integrals through 3D Gaussian primitives analytically, avoiding local affine collapse. Provides explicit control over ray origins and directions to facilitate precise application of nonlinear geometric corrections like arc-correction in PET.

Result: The ray-tracing approach yields a more physically consistent forward projection model and extends applicability of Gaussian-based reconstruction to a wider range of realistic tomography systems while improving projection accuracy.

Conclusion: Ray tracing with 3D Gaussians overcomes limitations of splatting-based approaches, enabling more accurate tomographic reconstruction with better handling of complex geometric corrections for medical imaging applications.

Abstract: 3D Gaussian Splatting (3DGS) has recently emerged in computer vision as a promising rendering technique. By adapting the principles of Elliptical Weighted Average (EWA) splatting to a modern differentiable pipeline, 3DGS enables real-time, high-quality novel view synthesis. Building upon this, R2-Gaussian extended the 3DGS paradigm to tomographic reconstruction by rectifying integration bias, achieving state-of-the-art performance in computed tomography (CT). To enable differentiability, R2-Gaussian adopts a local affine approximation: each 3D Gaussian is locally mapped to a 2D Gaussian on the detector and composed via alpha blending to form projections. However, the affine approximation can degrade reconstruction quantitative accuracy and complicate the incorporation of nonlinear geometric corrections. To address these limitations, we propose a tomographic reconstruction framework based on 3D Gaussian ray tracing. Our approach provides two key advantages over splatting-based models: (i) it computes the line integral through 3D Gaussian primitives analytically, avoiding the local affine collapse and thus yielding a more physically consistent forward projection model; and (ii) the ray-tracing formulation gives explicit control over ray origins and directions, which facilitates the precise application of nonlinear geometric corrections, e.g., arc-correction used in positron emission tomography (PET). These properties extend the applicability of Gaussian-based reconstruction to a wider range of realistic tomography systems while improving projection accuracy.

Method: Formulates segmentation as PDE-constrained optimization with composite objective: data fidelity term + penalty terms from reaction-diffusion equations and phase-field interface energies. Uses differentiable residual losses and evaluates on LIVECell dataset with UNet baseline.

Result: Consistent improvements in segmentation accuracy and boundary fidelity compared to unconstrained baselines. Enhanced stability and better generalization in low-sample regimes, particularly when evaluated on unseen cell types.

Conclusion: PDE-constrained optimization provides principled bridge between variational methods and deep learning, demonstrating value of structured priors for scientific machine learning applications.

Abstract: Segmentation of microscopy images constitutes an ill-posed inverse problem due to measurement noise, weak object boundaries, and limited labeled data. Although deep neural networks provide flexible nonparametric estimators, unconstrained empirical risk minimization often leads to unstable solutions and poor generalization. In this work, image segmentation is formulated as a PDE-constrained optimization problem that integrates physically motivated priors into deep learning models through variational regularization. The proposed framework minimizes a composite objective function consisting of a data fidelity term and penalty terms derived from reaction-diffusion equations and phase-field interface energies, all implemented as differentiable residual losses. Experiments are conducted on the LIVECell dataset, a high-quality, manually annotated collection of phase-contrast microscopy images. Training is performed on two cell types, while evaluation is carried out on a distinct, unseen cell type to assess generalization. A UNet architecture is used as the unconstrained baseline model. Experimental results demonstrate consistent improvements in segmentation accuracy and boundary fidelity compared to unconstrained deep learning baselines. Moreover, the PDE-regularized models exhibit enhanced stability and improved generalization in low-sample regimes, highlighting the advantages of incorporating structured priors. The proposed approach illustrates how PDE-constrained optimization can strengthen data-driven learning frameworks, providing a principled bridge between variational methods, statistical learning, and scientific machine learning.

Method: PISA uses a novel exact-or-approximate strategy: critical blocks are computed exactly while non-critical blocks are approximated efficiently using block-wise Taylor expansion, leveraging the discovery that attention scores of non-critical blocks exhibit distributional stability.

Result: PISA achieves 1.91× speedup on Wan2.1-14B, 2.57× on Hunyuan-Video, and 1.2× on FLUX for image generation while maintaining the highest quality among sparse attention methods, effectively bridging the speed-quality gap.

Conclusion: PISA demonstrates that approximating rather than discarding non-critical blocks enables efficient attention with sub-quadratic complexity while preserving full attention span, offering a superior alternative to conventional keep-or-drop sparse attention for diffusion transformers.

Abstract: Diffusion Transformers are fundamental for video and image generation, but their efficiency is bottlenecked by the quadratic complexity of attention. While block sparse attention accelerates computation by attending only critical key-value blocks, it suffers from degradation at high sparsity by discarding context. In this work, we discover that attention scores of non-critical blocks exhibit distributional stability, allowing them to be approximated accurately and efficiently rather than discarded, which is essentially important for sparse attention design. Motivated by this key insight, we propose PISA, a training-free Piecewise Sparse Attention that covers the full attention span with sub-quadratic complexity. Unlike the conventional keep-or-drop paradigm that directly drop the non-critical block information, PISA introduces a novel exact-or-approximate strategy: it maintains exact computation for critical blocks while efficiently approximating the remainder through block-wise Taylor expansion. This design allows PISA to serve as a faithful proxy to full attention, effectively bridging the gap between speed and quality. Experimental results demonstrate that PISA achieves 1.91 times and 2.57 times speedups on Wan2.1-14B and Hunyuan-Video, respectively, while consistently maintaining the highest quality among sparse attention methods. Notably, even for image generation on FLUX, PISA achieves a 1.2 times acceleration without compromising visual quality. Code is available at: https://github.com/xie-lab-ml/piecewise-sparse-attention.

Method: Two-stage framework: 1) Cognitive Injection via SFT for medical knowledge and structured think-then-answer paradigm; 2) Consistency Group Relative Policy Optimization (Con-GRPO) with consistency reward to ensure reasoning aligns with final diagnosis.

Result: MedAD-R1 achieves SOTA on MedAD-38K benchmark, outperforming baselines by >10%, generating transparent and logically consistent reasoning pathways.

Conclusion: The approach enhances trustworthiness and interpretability of AI for clinical decision support through structured reasoning and consistency optimization.

Abstract: Medical Anomaly Detection (MedAD) presents a significant opportunity to enhance diagnostic accuracy using Large Multimodal Models (LMMs) to interpret and answer questions based on medical images. However, the reliance on Supervised Fine-Tuning (SFT) on simplistic and fragmented datasets has hindered the development of models capable of plausible reasoning and robust multimodal generalization. To overcome this, we introduce MedAD-38K, the first large-scale, multi-modal, and multi-center benchmark for MedAD featuring diagnostic Chain-of-Thought (CoT) annotations alongside structured Visual Question-Answering (VQA) pairs. On this foundation, we propose a two-stage training framework. The first stage, Cognitive Injection, uses SFT to instill foundational medical knowledge and align the model with a structured think-then-answer paradigm. Given that standard policy optimization can produce reasoning that is disconnected from the final answer, the second stage incorporates Consistency Group Relative Policy Optimization (Con-GRPO). This novel algorithm incorporates a crucial consistency reward to ensure the generated reasoning process is relevant and logically coherent with the final diagnosis. Our proposed model, MedAD-R1, achieves state-of-the-art (SOTA) performance on the MedAD-38K benchmark, outperforming strong baselines by more than 10%. This superior performance stems from its ability to generate transparent and logically consistent reasoning pathways, offering a promising approach to enhancing the trustworthiness and interpretability of AI for clinical decision support.

Method: DVE constructs a differential vector field capturing directional discrepancy between target and anchor concepts in the velocity field. During inference, it projects the velocity field onto this differential direction to remove concept-specific components without affecting other semantics.

Result: Extensive experiments on FLUX show DVE outperforms existing baselines across NSFW suppression, artistic style removal, and object erasure tasks while preserving image quality and diversity.

Conclusion: DVE provides an effective training-free solution for concept erasure in flow matching models, addressing safety and control concerns in text-to-image generation without compromising performance.

Abstract: Text-to-image diffusion models have demonstrated remarkable capabilities in generating high-quality images, yet their tendency to reproduce undesirable concepts, such as NSFW content, copyrighted styles, or specific objects, poses growing concerns for safe and controllable deployment. While existing concept erasure approaches primarily focus on DDPM-based diffusion models and rely on costly fine-tuning, the recent emergence of flow matching models introduces a fundamentally different generative paradigm for which prior methods are not directly applicable. In this paper, we propose Differential Vector Erasure (DVE), a training-free concept erasure method specifically designed for flow matching models. Our key insight is that semantic concepts are implicitly encoded in the directional structure of the velocity field governing the generative flow. Leveraging this observation, we construct a differential vector field that characterizes the directional discrepancy between a target concept and a carefully chosen anchor concept. During inference, DVE selectively removes concept-specific components by projecting the velocity field onto the differential direction, enabling precise concept suppression without affecting irrelevant semantics. Extensive experiments on FLUX demonstrate that DVE consistently outperforms existing baselines on a wide range of concept erasure tasks, including NSFW suppression, artistic style removal, and object erasure, while preserving image quality and diversity.

Method: Proposes PandaPose with a 3D anchor space comprising: (1) Joint-wise 3D anchors in canonical coordinate system for robust priors, (2) Depth-aware joint-wise feature lifting that hierarchically integrates depth information to resolve self-occlusion, (3) Anchor-feature interaction decoder that incorporates 3D anchors with lifted features to generate unified anchor queries encapsulating joint-wise 3D anchor set, visual cues and geometric depth information, which are then used for anchor-to-joint ensemble prediction.

Result: Experiments on Human3.6M, MPI-INF-3DHP and 3DPW benchmarks demonstrate superiority, with substantial reduction in error by 14.7% compared to SOTA methods on challenging Human3.6M conditions. Qualitative comparisons further showcase effectiveness and robustness.

Conclusion: PandaPose effectively addresses error propagation and self-occlusion issues in 3D human pose lifting through its novel 3D anchor space approach, achieving state-of-the-art performance on established benchmarks.

Abstract: 3D human pose lifting from a single RGB image is a challenging task in 3D vision. Existing methods typically establish a direct joint-to-joint mapping from 2D to 3D poses based on 2D features. This formulation suffers from two fundamental limitations: inevitable error propagation from input predicted 2D pose to 3D predictions and inherent difficulties in handling self-occlusion cases. In this paper, we propose PandaPose, a 3D human pose lifting approach via propagating 2D pose prior to 3D anchor space as the unified intermediate representation. Specifically, our 3D anchor space comprises: (1) Joint-wise 3D anchors in the canonical coordinate system, providing accurate and robust priors to mitigate 2D pose estimation inaccuracies. (2) Depth-aware joint-wise feature lifting that hierarchically integrates depth information to resolve self-occlusion ambiguities. (3) The anchor-feature interaction decoder that incorporates 3D anchors with lifted features to generate unified anchor queries encapsulating joint-wise 3D anchor set, visual cues and geometric depth information. The anchor queries are further employed to facilitate anchor-to-joint ensemble prediction. Experiments on three well-established benchmarks (i.e., Human3.6M, MPI-INF-3DHP and 3DPW) demonstrate the superiority of our proposition. The substantial reduction in error by $14.7%$ compared to SOTA methods on the challenging conditions of Human3.6M and qualitative comparisons further showcase the effectiveness and robustness of our approach.

Method: Learns shared representations for multiple modalities by projecting them independently, enabling use when data is modal-incomplete, reducing text dependence and improving visual feature integration.

Result: Outperforms existing approaches on two benchmark datasets when text is missing, showing better visual feature integration and reduced text dependence.

Conclusion: First comprehensive investigation of harmful meme detection with modal-incomplete data; method enables real-world application where modalities may be absent.

Abstract: Internet memes are powerful tools for communication, capable of spreading political, psychological, and sociocultural ideas. However, they can be harmful and can be used to disseminate hate toward targeted individuals or groups. Although previous studies have focused on designing new detection methods, these often rely on modal-complete data, such as text and images. In real-world settings, however, modalities like text may be missing due to issues like poor OCR quality, making existing methods sensitive to missing information and leading to performance deterioration. To address this gap, in this paper, we present the first-of-its-kind work to comprehensively investigate the behavior of harmful meme detection methods in the presence of modal-incomplete data. Specifically, we propose a new baseline method that learns a shared representation for multiple modalities by projecting them independently. These shared representations can then be leveraged when data is modal-incomplete. Experimental results on two benchmark datasets demonstrate that our method outperforms existing approaches when text is missing. Moreover, these results suggest that our method allows for better integration of visual features, reducing dependence on text and improving robustness in scenarios where textual information is missing. Our work represents a significant step forward in enabling the real-world application of harmful meme detection, particularly in situations where a modality is absent.

Method: Created LightCity, a high-quality synthetic urban dataset with over 300 sky maps, controllable illumination, street-level and aerial perspectives (50K+ images), and rich properties including depth, normal, material components, light and indirect light.

Result: Provides a comprehensive dataset enabling benchmarking of three fundamental tasks in urban environments, with analysis of these benchmarks to advance related research.

Conclusion: LightCity addresses the dataset gap for urban inverse rendering research, providing a robust foundation for studying illumination effects on intrinsic decomposition and 3D reconstruction.

Abstract: Inverse rendering in urban scenes is pivotal for applications like autonomous driving and digital twins. Yet, it faces significant challenges due to complex illumination conditions, including multi-illumination and indirect light and shadow effects. However, the effects of these challenges on intrinsic decomposition and 3D reconstruction have not been explored due to the lack of appropriate datasets. In this paper, we present LightCity, a novel high-quality synthetic urban dataset featuring diverse illumination conditions with realistic indirect light and shadow effects. LightCity encompasses over 300 sky maps with highly controllable illumination, varying scales with street-level and aerial perspectives over 50K images, and rich properties such as depth, normal, material components, light and indirect light, etc. Besides, we leverage LightCity to benchmark three fundamental tasks in the urban environments and conduct a comprehensive analysis of these benchmarks, laying a robust foundation for advancing related research.

Method: Uses Fukunaga-Koontz Linear Discriminant Analysis in whitened embedding space to suppress within-class variation and enhance between-class discrimination, creating a closed-form linear projection.

Result: Improves ImageNet-1K top-1 accuracy from 75.1% to 79.1%, with consistent gains on larger label spaces (14K and 21K classes), and supports 10-12x compression with minimal accuracy loss.

Conclusion: Koo-Fu CLIP provides an efficient, lightweight adaptation method that reshapes CLIP embedding geometry for improved supervised classification and enables efficient large-scale classification and retrieval.

Abstract: Visual-language models such as CLIP provide powerful general-purpose representations, but their raw embeddings are not optimized for supervised classification, often exhibiting limited class separation and excessive dimensionality. We propose Koo-Fu CLIP, a supervised CLIP adaptation method based on Fukunaga-Koontz Linear Discriminant Analysis, which operates in a whitened embedding space to suppress within-class variation and enhance between-class discrimination. The resulting closed-form linear projection reshapes the geometry of CLIP embeddings, improving class separability while performing effective dimensionality reduction, and provides a lightweight and efficient adaptation of CLIP representations. Across large-scale ImageNet benchmarks, nearest visual prototype classification in the Koo-Fu CLIP space improves top-1 accuracy from 75.1% to 79.1% on ImageNet-1K, with consistent gains persisting as the label space expands to 14K and 21K classes. The method supports substantial compression by up to 10-12x with little or no loss in accuracy, enabling efficient large-scale classification and retrieval.

Method: Introduces Corruption Restoration Transformer (CRT), a plug-and-play, model-agnostic vision transformer that uses adversarial training to restore clean observations from corrupted inputs. CRT works without computationally expensive fine-tuning of the underlying VLA model.

Result: State-of-the-art VLAs (π₀.₅ and SmolVLA) drop from 90% to as low as 2% success rates under common signal artifacts. CRT effectively recovers lost performance, enabling VLAs to maintain near-baseline success rates even under severe visual corruption, as demonstrated across LIBERO and Meta-World benchmarks.

Conclusion: Image corruptions pose a critical vulnerability for VLA models in real-world deployment. CRT provides an effective, practical solution to immunize VLAs against sensor disturbances without requiring model retraining, addressing a significant gap in robustness for multimodal robotic systems.

Abstract: Vision-Language-Action (VLA) models have emerged as a dominant paradigm for generalist robotic manipulation, unifying perception and control within a single end-to-end architecture. However, despite their success in controlled environments, reliable real-world deployment is severely hindered by their fragility to visual disturbances. While existing literature extensively addresses physical occlusions caused by scene geometry, a critical mode remains largely unexplored: image corruptions. These sensor-level artifacts, ranging from electronic noise and dead pixels to lens contaminants, directly compromise the integrity of the visual signal prior to interpretation. In this work, we quantify this vulnerability, demonstrating that state-of-the-art VLAs such as $π_{0.5}$ and SmolVLA, suffer catastrophic performance degradation, dropping from 90% success rates to as low as 2%, under common signal artifacts. To mitigate this, we introduce the Corruption Restoration Transformer (CRT), a plug-and-play and model-agnostic vision transformer designed to immunize VLA models against sensor disturbances. Leveraging an adversarial training objective, CRT restores clean observations from corrupted inputs without requiring computationally expensive fine-tuning of the underlying model. Extensive experiments across the LIBERO and Meta-World benchmarks demonstrate that CRT effectively recovers lost performance, enabling VLAs to maintain near-baseline success rates, even under severe visual corruption.

Method: Coarse-to-fine pipeline: 1) Lightweight semantic segmentation pre-screens candidate areas, 2) Vision-language reasoning agent fuses visual features with Point-of-Interest data to detect subtle hazards using MLLMs.

Result: Framework significantly outperforms geometric baselines in risk identification accuracy. Generates human-like, interpretable justifications. Emergency Landing Site Selection (ELSS) benchmark dataset released publicly.

Conclusion: MLLM-based approach enables comprehensive risk assessment for UAV emergency landing by understanding semantic context, enhancing trust through interpretable reasoning.

Abstract: Safe UAV emergency landing requires more than just identifying flat terrain; it demands understanding complex semantic risks (e.g., crowds, temporary structures) invisible to traditional geometric sensors. In this paper, we propose a novel framework leveraging Remote Sensing (RS) imagery and Multimodal Large Language Models (MLLMs) for global context-aware landing site assessment. Unlike local geometric methods, our approach employs a coarse-to-fine pipeline: first, a lightweight semantic segmentation module efficiently pre-screens candidate areas; second, a vision-language reasoning agent fuses visual features with Point-of-Interest (POI) data to detect subtle hazards. To validate this approach, we construct and release the Emergency Landing Site Selection (ELSS) benchmark. Experiments demonstrate that our framework significantly outperforms geometric baselines in risk identification accuracy. Furthermore, qualitative results confirm its ability to generate human-like, interpretable justifications, enhancing trust in automated decision-making. The benchmark dataset is publicly accessible at https://anonymous.4open.science/r/ELSS-dataset-43D7.

Method: Created EEmoDB dataset with 1.2M QA pairs from 125k images (EEmoDB-QA) and 36k dataset from 25k images for fine-grained assessment (EEmoDB-Assess). Developed EEmo-Logic MLLM via instruction fine-tuning and task-customized group relative preference optimization (GRPO) with novel reward design.

Result: EEmo-Logic achieves robust performance in both in-domain and cross-domain datasets, excelling in emotion QA and fine-grained assessment tasks.

Conclusion: The proposed dataset and model advance comprehensive image-evoked emotion understanding, enabling better machine empathy and human-computer interaction applications.

Abstract: Understanding the multi-dimensional attributes and intensity nuances of image-evoked emotions is pivotal for advancing machine empathy and empowering diverse human-computer interaction applications. However, existing models are still limited to coarse-grained emotion perception or deficient reasoning capabilities. To bridge this gap, we introduce EEmoDB, the largest image-evoked emotion understanding dataset to date. It features $5$ analysis dimensions spanning $5$ distinct task categories, facilitating comprehensive interpretation. Specifically, we compile $1.2M$ question-answering (QA) pairs (EEmoDB-QA) from $125k$ images via automated generation, alongside a $36k$ dataset (EEmoDB-Assess) curated from $25k$ images for fine-grained assessment. Furthermore, we propose EEmo-Logic, an all-in-one multimodal large language model (MLLM) developed via instruction fine-tuning and task-customized group relative preference optimization (GRPO) with novel reward design. Extensive experiments demonstrate that EEmo-Logic achieves robust performance in in-domain and cross-domain datasets, excelling in emotion QA and fine-grained assessment. The code is available at https://anonymous.4open.science/r/EEmoLogic.

Method: CurriSeg uses a dual-phase framework: 1) Curriculum Selection phase dynamically selects training data based on temporal loss statistics to distinguish hard-but-informative samples from noisy ones, enabling stable capability enhancement. 2) Anti-Curriculum Promotion phase uses Spectral-Blindness Fine-Tuning to suppress high-frequency components, enforcing dependence on low-frequency structural and contextual cues to strengthen generalization.

Result: Extensive experiments show CurriSeg achieves consistent improvements across diverse CECS benchmarks without adding parameters or increasing total training time.

Conclusion: CurriSeg offers a principled view of how progression and challenge interplay to foster robust and context-aware segmentation, demonstrating that learning dynamics (not just architecture) are crucial for handling entangled visual patterns.

Abstract: Biological learning proceeds from easy to difficult tasks, gradually reinforcing perception and robustness. Inspired by this principle, we address Context-Entangled Content Segmentation (CECS), a challenging setting where objects share intrinsic visual patterns with their surroundings, as in camouflaged object detection. Conventional segmentation networks predominantly rely on architectural enhancements but often ignore the learning dynamics that govern robustness under entangled data distributions. We introduce CurriSeg, a dual-phase learning framework that unifies curriculum and anti-curriculum principles to improve representation reliability. In the Curriculum Selection phase, CurriSeg dynamically selects training data based on the temporal statistics of sample losses, distinguishing hard-but-informative samples from noisy or ambiguous ones, thus enabling stable capability enhancement. In the Anti-Curriculum Promotion phase, we design Spectral-Blindness Fine-Tuning, which suppresses high-frequency components to enforce dependence on low-frequency structural and contextual cues and thus strengthens generalization. Extensive experiments demonstrate that CurriSeg achieves consistent improvements across diverse CECS benchmarks without adding parameters or increasing total training time, offering a principled view of how progression and challenge interplay to foster robust and context-aware segmentation. Code will be released.

Method: Three-part approach: 1) Efficient Multi-scale Transformer (EMFormer) with single convolution for multi-scale feature extraction, 2) Accumulative context finetuning for temporal consistency, 3) Composite loss with sinusoidal weighting for dynamic balance during optimization

Result: Achieves strong performance in weather forecasting and extreme event prediction with substantial long-term forecast accuracy improvements; demonstrates strong generalization on vision benchmarks (ImageNet-1K, ADE20K) with 5.69x speedup over conventional multi-scale modules

Conclusion: Proposed pipeline effectively enhances long-context modeling while reducing computational overhead, addressing key limitations in long-term weather forecasting systems

Abstract: Long-term weather forecasting is critical for socioeconomic planning and disaster preparedness. While recent approaches employ finetuning to extend prediction horizons, they remain constrained by the issues of catastrophic forgetting, error accumulation, and high training overhead. To address these limitations, we present a novel pipeline across pretraining, finetuning and forecasting to enhance long-context modeling while reducing computational overhead. First, we introduce an Efficient Multi-scale Transformer (EMFormer) to extract multi-scale features through a single convolution in both training and inference. Based on the new architecture, we further employ an accumulative context finetuning to improve temporal consistency without degrading short-term accuracy. Additionally, we propose a composite loss that dynamically balances different terms via a sinusoidal weighting, thereby adaptively guiding the optimization trajectory throughout pretraining and finetuning. Experiments show that our approach achieves strong performance in weather forecasting and extreme event prediction, substantially improving long-term forecast accuracy. Moreover, EMFormer demonstrates strong generalization on vision benchmarks (ImageNet-1K and ADE20K) while delivering a 5.69x speedup over conventional multi-scale modules.

Method: Two-stage RL framework: 1) SFT with residual alignment to bridge 3D features and text embeddings, and abnormality re-weighting to emphasize clinical tokens; 2) RL with redesigned consistency reward to promote step-by-step diagnostic reasoning.

Result: Achieved state-of-the-art accuracies of 41.92% on CT-RATE and 44.99% on RAD-ChestCT benchmarks for medical multiple-choice visual question answering, outperforming prior methods.

Conclusion: The approach enhances real-world diagnostic workflows by enabling more reliable and transparent 3D medical vision-language systems with improved abnormality diagnosis and clinical reasoning.

Abstract: Developing 3D vision-language models with robust clinical reasoning remains a challenge due to the inherent complexity of volumetric medical imaging, the tendency of models to overfit superficial report patterns, and the lack of interpretability-aware reward designs. In this paper, we propose Med3D-R1, a reinforcement learning framework with a two-stage training process: Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL). During SFT stage, we introduce a residual alignment mechanism to bridge the gap between high-dimensional 3D features and textual embeddings, and an abnormality re-weighting strategy to emphasize clinically informative tokens and reduce structural bias in reports. In RL stage, we redesign the consistency reward to explicitly promote coherent, step-by-step diagnostic reasoning. We evaluate our method on medical multiple-choice visual question answering using two 3D diagnostic benchmarks, CT-RATE and RAD-ChestCT, where our model attains state-of-the-art accuracies of 41.92% on CT-RATE and 44.99% on RAD-ChestCT. These results indicate improved abnormality diagnosis and clinical reasoning and outperform prior methods on both benchmarks. Overall, our approach holds promise for enhancing real-world diagnostic workflows by enabling more reliable and transparent 3D medical vision-language systems.

Method: Proposes Text Refinement and Alignment (TRA) framework with two modules: Point-based Text Refinement (PTR) refines initial video descriptions using point annotations and pre-trained models, and Point-based Multimodal Alignment (PMA) projects features into unified semantic space with point-level multimodal contrastive learning.

Result: Extensive experiments on five benchmarks show favorable performance compared to state-of-the-art methods, with practical implementation on a single 24 GB RTX 3090 GPU.

Conclusion: Incorporating textual features through refinement and alignment significantly improves point-supervised temporal action localization, demonstrating the value of multimodal integration for video understanding tasks.

Abstract: Recently, point-supervised temporal action localization has gained significant attention for its effective balance between labeling costs and localization accuracy. However, current methods only consider features from visual inputs, neglecting helpful semantic information from the text side. To address this issue, we propose a Text Refinement and Alignment (TRA) framework that effectively utilizes textual features from visual descriptions to complement the visual features as they are semantically rich. This is achieved by designing two new modules for the original point-supervised framework: a Point-based Text Refinement module (PTR) and a Point-based Multimodal Alignment module (PMA). Specifically, we first generate descriptions for video frames using a pre-trained multimodal model. Next, PTR refines the initial descriptions by leveraging point annotations together with multiple pre-trained models. PMA then projects all features into a unified semantic space and leverages a point-level multimodal feature contrastive learning to reduce the gap between visual and linguistic modalities. Last, the enhanced multi-modal features are fed into the action detector for precise localization. Extensive experimental results on five widely used benchmarks demonstrate the favorable performance of our proposed framework compared to several state-of-the-art methods. Moreover, our computational overhead analysis shows that the framework can run on a single 24 GB RTX 3090 GPU, indicating its practicality and scalability.

Method: 1) Leverage relative depth from foundation models which preserves global layout and boundaries. 2) Calibrate relative depth with sparse range measurements to create pseudo-metric depth priors. 3) Design a refinement network that follows the prior where reliable and deviates where necessary.

Result: The system enables accurate metric depth predictions from very few labeled samples, sustains stable scale and sharp edges across few-shot regimes, and works effectively even when curated validation data are unavailable.

Conclusion: Coupling foundation priors with sparse anchors provides a practical route to robust, deployment-ready depth completion under real-world label scarcity, bridging the gap between relative depth from foundation models and metric depth requirements for real applications.

Abstract: Recent monocular foundation models excel at zero-shot depth estimation, yet their outputs are inherently relative rather than metric, limiting direct use in robotics and autonomous driving. We leverage the fact that relative depth preserves global layout and boundaries: by calibrating it with sparse range measurements, we transform it into a pseudo metric depth prior. Building on this prior, we design a refinement network that follows the prior where reliable and deviates where necessary, enabling accurate metric predictions from very few labeled samples. The resulting system is particularly effective when curated validation data are unavailable, sustaining stable scale and sharp edges across few-shot regimes. These findings suggest that coupling foundation priors with sparse anchors is a practical route to robust, deployment-ready depth completion under real-world label scarcity.

Method: Proposes Q-DiT4SR with two key components: 1) H-SVD (hierarchical SVD) that integrates global low-rank branch with local block-wise rank-1 branch under matched parameter budget, and 2) Variance-aware Spatio-Temporal Mixed Precision (VaSMP for cross-layer weight bit-width allocation based on rate-distortion theory, and VaTMP for intra-layer activation precision scheduling across diffusion timesteps via dynamic programming).

Result: Achieves state-of-the-art performance on multiple real-world datasets under both W4A6 and W4A4 settings. W4A4 quantization reduces model size by 5.8× and computational operations by over 60× while maintaining high-quality texture generation.

Conclusion: Q-DiT4SR is the first PTQ framework specifically tailored for DiT-based real-world image super-resolution, effectively addressing the inference burden while preserving local texture quality through novel hierarchical decomposition and mixed precision strategies.

Abstract: Recently, Diffusion Transformers (DiTs) have emerged in Real-World Image Super-Resolution (Real-ISR) to generate high-quality textures, yet their heavy inference burden hinders real-world deployment. While Post-Training Quantization (PTQ) is a promising solution for acceleration, existing methods in super-resolution mostly focus on U-Net architectures, whereas generic DiT quantization is typically designed for text-to-image tasks. Directly applying these methods to DiT-based super-resolution models leads to severe degradation of local textures. Therefore, we propose Q-DiT4SR, the first PTQ framework specifically tailored for DiT-based Real-ISR. We propose H-SVD, a hierarchical SVD that integrates a global low-rank branch with a local block-wise rank-1 branch under a matched parameter budget. We further propose Variance-aware Spatio-Temporal Mixed Precision: VaSMP allocates cross-layer weight bit-widths in a data-free manner based on rate-distortion theory, while VaTMP schedules intra-layer activation precision across diffusion timesteps via dynamic programming (DP) with minimal calibration. Experiments on multiple real-world datasets demonstrate that our Q-DiT4SR achieves SOTA performance under both W4A6 and W4A4 settings. Notably, the W4A4 quantization configuration reduces model size by 5.8$\times$ and computational operations by over 60$\times$. Our code and models will be available at https://github.com/xunzhang1128/Q-DiT4SR.

Method: Proposes TrafficFlow-aware Lane perception Module (TFM) that extracts real-time traffic flow features and integrates them with existing lane perception algorithms, validated on open-source algorithms and datasets.

Result: TFM consistently improves performance across four mainstream models and two public datasets (Nuscenes and OpenLaneV2), achieving up to +4.1% mAP gain on Nuscenes dataset.

Conclusion: Traffic flow is an effective, cost-free information source that can significantly enhance lane perception in autonomous driving systems when integrated with existing algorithms.

Abstract: Autonomous driving systems require robust lane perception capabilities, yet existing vision-based detection methods suffer significant performance degradation when visual sensors provide insufficient cues, such as in occluded or lane-missing scenarios. While some approaches incorporate high-definition maps as supplementary information, these solutions face challenges of high subscription costs and limited real-time performance. To address these limitations, we explore an innovative information source: traffic flow, which offers real-time capabilities without additional costs. This paper proposes a TrafficFlow-aware Lane perception Module (TFM) that effectively extracts real-time traffic flow features and seamlessly integrates them with existing lane perception algorithms. This solution originated from real-world autonomous driving conditions and was subsequently validated on open-source algorithms and datasets. Extensive experiments on four mainstream models and two public datasets (Nuscenes and OpenLaneV2) using standard evaluation metrics show that TFM consistently improves performance, achieving up to +4.1% mAP gain on the Nuscenes dataset.

Method: Proposes DSFC-Net with dual encoders: CNN branch for local boundaries and short-range continuity, and Spatial-Frequency Hybrid Transformer (SFT) with Cross-Frequency Interaction Attention (CFIA) module using Laplacian Pyramid to decouple high/low-frequency information. Also includes Channel Feature Fusion Module (CFFM) to integrate features from both branches.

Result: Comprehensive experiments on WHU-RuR+, DeepGlobe, and Massachusetts datasets show DSFC-Net’s superiority over state-of-the-art approaches for rural road extraction.

Conclusion: DSFC-Net effectively addresses rural road extraction challenges by fusing spatial and frequency-domain information, with the SFT’s frequency-aware attention mechanism preserving connectivity of narrow roads against occlusions.

Abstract: Accurate extraction of rural roads from high-resolution remote sensing imagery is essential for infrastructure planning and sustainable development. However, this task presents unique challenges in rural settings due to several factors. These include high intra-class variability and low inter-class separability from diverse surface materials, frequent vegetation occlusions that disrupt spatial continuity, and narrow road widths that exacerbate detection difficulties. Existing methods, primarily optimized for structured urban environments, often underperform in these scenarios as they overlook such distinctive characteristics. To address these challenges, we propose DSFC-Net, a dual-encoder framework that synergistically fuses spatial and frequency-domain information. Specifically, a CNN branch is employed to capture fine-grained local road boundaries and short-range continuity, while a novel Spatial-Frequency Hybrid Transformer (SFT) is introduced to robustly model global topological dependencies against vegetation occlusions. Distinct from standard attention mechanisms that suffer from frequency bias, the SFT incorporates a Cross-Frequency Interaction Attention (CFIA) module that explicitly decouples high- and low-frequency information via a Laplacian Pyramid strategy. This design enables the dynamic interaction between spatial details and frequency-aware global contexts, effectively preserving the connectivity of narrow roads. Furthermore, a Channel Feature Fusion Module (CFFM) is proposed to bridge the two branches by adaptively recalibrating channel-wise feature responses, seamlessly integrating local textures with global semantics for accurate segmentation. Comprehensive experiments on the WHU-RuR+, DeepGlobe, and Massachusetts datasets validate the superiority of DSFC-Net over state-of-the-art approaches.

Method: Identify monolingual safety neurons (MS-Neurons) and validate their causal role in safety refusal; identify SS-Neurons as the subset shared between high-resource and non-high-resource languages; propose neuron-oriented training targeting SS-Neurons based on language resource distribution and model architecture.

Result: Suppressing SS-Neurons causes concurrent safety drops across non-high-resource languages, while reinforcing them improves cross-lingual defensive consistency; fine-tuning this tiny neuronal subset outperforms state-of-the-art methods, significantly enhancing non-high-resource language safety while maintaining general capabilities.

Conclusion: SS-Neurons are a critical neuronal subset that jointly regulates safety behavior across languages, and targeting them through neuron-oriented training effectively addresses multilingual safety imbalance while preserving model capabilities.

Abstract: Multilingual safety remains significantly imbalanced, leaving non-high-resource (NHR) languages vulnerable compared to robust high-resource (HR) ones. Moreover, the neural mechanisms driving safety alignment remain unclear despite observed cross-lingual representation transfer. In this paper, we find that LLMs contain a set of cross-lingual shared safety neurons (SS-Neurons), a remarkably small yet critical neuronal subset that jointly regulates safety behavior across languages. We first identify monolingual safety neurons (MS-Neurons) and validate their causal role in safety refusal behavior through targeted activation and suppression. Our cross-lingual analyses then identify SS-Neurons as the subset of MS-Neurons shared between HR and NHR languages, serving as a bridge to transfer safety capabilities from HR to NHR domains. We observe that suppressing these neurons causes concurrent safety drops across NHR languages, whereas reinforcing them improves cross-lingual defensive consistency. Building on these insights, we propose a simple neuron-oriented training strategy that targets SS-Neurons based on language resource distribution and model architecture. Experiments demonstrate that fine-tuning this tiny neuronal subset outperforms state-of-the-art methods, significantly enhancing NHR safety while maintaining the model’s general capabilities. The code and dataset will be available athttps://github.com/1518630367/SS-Neuron-Expansion.

Method: LiP-Map introduces a line-plane joint optimization framework that explicitly models learnable line and planar primitives together. Instead of imposing pairwise coplanarity constraints, it constructs explicit interactions between plane and line primitives, integrating planar topology into 3D line mapping. The method operates efficiently, typically completing reconstructions in 3-5 minutes per scene.

Result: On over 100 scenes from ScanNetV2, ScanNet++, Hypersim, 7Scenes, and Tanks&Temple datasets, LiP-Map improves both accuracy and completeness over state-of-the-art methods. It also significantly advances line-assisted visual localization, establishing strong performance on 7Scenes.

Conclusion: LiP-Map pioneers the integration of planar topology into 3D line mapping through explicit line-plane joint optimization, offering a principled route toward structured reconstruction in man-made environments while maintaining strong efficiency.

Abstract: 3D line mapping from multi-view RGB images provides a compact and structured visual representation of scenes. We study the problem from a physical and topological perspective: a 3D line most naturally emerges as the edge of a finite 3D planar patch. We present LiP-Map, a line-plane joint optimization framework that explicitly models learnable line and planar primitives. This coupling enables accurate and detailed 3D line mapping while maintaining strong efficiency (typically completing a reconstruction in 3 to 5 minutes per scene). LiP-Map pioneers the integration of planar topology into 3D line mapping, not by imposing pairwise coplanarity constraints but by explicitly constructing interactions between plane and line primitives, thus offering a principled route toward structured reconstruction in man-made environments. On more than 100 scenes from ScanNetV2, ScanNet++, Hypersim, 7Scenes, and Tanks&Temple, LiP-Map improves both accuracy and completeness over state-of-the-art methods. Beyond line mapping quality, LiP-Map significantly advances line-assisted visual localization, establishing strong performance on 7Scenes. Our code is released at https://github.com/calmke/LiPMAP for reproducible research.

Method: Proposes REORM, a reasoning-enhanced object removal framework leveraging multimodal large language models to infer which elements must be jointly removed. Features modular design with MLLM-driven analysis, mask-guided removal, self-correction mechanism, and local-deployment variant for resource-limited scenarios.

Result: Introduces ICOREval benchmark for evaluation. REORM outperforms state-of-the-art image editing systems on ICOREval, demonstrating effectiveness in producing interaction-consistent results.

Conclusion: REORM successfully addresses the ICOR problem by using MLLM reasoning to identify and remove interaction elements along with target objects, producing more semantically consistent image editing results.

Abstract: Image-based object removal often erases only the named target, leaving behind interaction evidence that renders the result semantically inconsistent. We formalize this problem as Interaction-Consistent Object Removal (ICOR), which requires removing not only the target object but also associated interaction elements, such as lighting-dependent effects, physically connected objects, targetproduced elements, and contextually linked objects. To address this task, we propose Reasoning-Enhanced Object Removal with MLLM (REORM), a reasoningenhanced object removal framework that leverages multimodal large language models to infer which elements must be jointly removed. REORM features a modular design that integrates MLLM-driven analysis, mask-guided removal, and a self-correction mechanism, along with a local-deployment variant that supports accurate editing under limited resources. To support evaluation, we introduce ICOREval, a benchmark consisting of instruction-driven removals with rich interaction dependencies. On ICOREval, REORM outperforms state-of-the-art image editing systems, demonstrating its effectiveness in producing interactionconsistent results.

Method: ReDiStory is a training-free framework that decomposes text embeddings into identity-related and frame-specific components, then decorrelates frame embeddings by suppressing shared directions across frames to reduce cross-frame interference.

Result: Experiments on ConsiStory+ benchmark show consistent gains over 1Prompt1Story on multiple identity consistency metrics while maintaining prompt fidelity, using identical diffusion backbones and inference settings.

Conclusion: ReDiStory improves identity consistency in multi-frame story generation through inference-time prompt embedding reorganization without modifying diffusion parameters or requiring additional supervision.

Abstract: Generating coherent visual stories requires maintaining subject identity across multiple images while preserving frame-specific semantics. Recent training-free methods concatenate identity and frame prompts into a unified representation, but this often introduces inter-frame semantic interference that weakens identity preservation in complex stories. We propose ReDiStory, a training-free framework that improves multi-frame story generation via inference-time prompt embedding reorganization. ReDiStory explicitly decomposes text embeddings into identity-related and frame-specific components, then decorrelates frame embeddings by suppressing shared directions across frames. This reduces cross-frame interference without modifying diffusion parameters or requiring additional supervision. Under identical diffusion backbones and inference settings, ReDiStory improves identity consistency while maintaining prompt fidelity. Experiments on the ConsiStory+ benchmark show consistent gains over 1Prompt1Story on multiple identity consistency metrics. Code is available at: https://github.com/YuZhenyuLindy/ReDiStory

Method: Uses LLM agents to maintain structured story state (character sheet, global settings, per-page constraints) and generate prompts for training-free text-to-image generation.

Result: Enables localized page edits, improves cross-page consistency, reduces unintended changes and editing time compared to baseline, approaching one-shot consistency of Gemini Storybook.

Conclusion: StoryState provides model-agnostic, prompt-based approach for consistent multimodal story generation with editable state representation.

Abstract: Large multimodal models have enabled one-click storybook generation, where users provide a short description and receive a multi-page illustrated story. However, the underlying story state, such as characters, world settings, and page-level objects, remains implicit, making edits coarse-grained and often breaking visual consistency. We present StoryState, an agent-based orchestration layer that introduces an explicit and editable story state on top of training-free text-to-image generation. StoryState represents each story as a structured object composed of a character sheet, global settings, and per-page scene constraints, and employs a small set of LLM agents to maintain this state and derive 1Prompt1Story-style prompts for generation and editing. Operating purely through prompts, StoryState is model-agnostic and compatible with diverse generation backends. System-level experiments on multi-page editing tasks show that StoryState enables localized page edits, improves cross-page consistency, and reduces unintended changes, interaction turns, and editing time compared to 1Prompt1Story, while approaching the one-shot consistency of Gemini Storybook. Code is available at https://github.com/YuZhenyuLindy/StoryState

Method: DeCorStory uses Gram-Schmidt prompt embedding decorrelation to orthogonalize frame-level semantics, singular value reweighting to strengthen prompt-specific information, and identity-preserving cross-attention to stabilize character identity during diffusion inference.

Result: The method achieves consistent improvements in prompt-image alignment, identity consistency, and visual diversity, reaching state-of-the-art performance among training-free baselines without requiring model modifications or fine-tuning.

Conclusion: DeCorStory provides an effective training-free solution for maintaining visual and semantic consistency in text-to-image storytelling by reducing inter-frame semantic interference through embedding decorrelation and identity preservation techniques.

Abstract: Maintaining visual and semantic consistency across frames is a key challenge in text-to-image storytelling. Existing training-free methods, such as One-Prompt-One-Story, concatenate all prompts into a single sequence, which often induces strong embedding correlation and leads to color leakage, background blending, and identity drift. We propose DeCorStory, a training-free inference-time framework that explicitly reduces inter-frame semantic interference. DeCorStory applies Gram-Schmidt prompt embedding decorrelation to orthogonalize frame-level semantics, followed by singular value reweighting to strengthen prompt-specific information and identity-preserving cross-attention to stabilize character identity during diffusion. The method requires no model modification or fine-tuning and can be seamlessly integrated into existing diffusion pipelines. Experiments demonstrate consistent improvements in prompt-image alignment, identity consistency, and visual diversity, achieving state-of-the-art performance among training-free baselines. Code is available at: https://github.com/YuZhenyuLindy/DeCorStory

Method: FlowCast speculates future velocity by extrapolating current velocity without additional time cost, accepting it if within a mean-squared error threshold. This allows skipping redundant steps in stable regions while maintaining precision in complex ones. It’s plug-and-play, requires no auxiliary networks, and works with any FM model.

Result: Empirical evaluations show FlowCast achieves >2.5× speedup in image generation, video generation, and editing tasks, outperforming existing baselines with no quality loss compared to standard full generation.

Conclusion: FlowCast provides an effective training-free acceleration framework for Flow Matching models that maintains quality while significantly improving inference speed, making them more practical for real-time applications.

Abstract: Flow Matching (FM) has recently emerged as a powerful approach for high-quality visual generation. However, their prohibitively slow inference due to a large number of denoising steps limits their potential use in real-time or interactive applications. Existing acceleration methods, like distillation, truncation, or consistency training, either degrade quality, incur costly retraining, or lack generalization. We propose FlowCast, a training-free speculative generation framework that accelerates inference by exploiting the fact that FM models are trained to preserve constant velocity. FlowCast speculates future velocity by extrapolating current velocity without incurring additional time cost, and accepts it if it is within a mean-squared error threshold. This constant-velocity forecasting allows redundant steps in stable regions to be aggressively skipped while retaining precision in complex ones. FlowCast is a plug-and-play framework that integrates seamlessly with any FM model and requires no auxiliary networks. We also present a theoretical analysis and bound the worst-case deviation between speculative and full FM trajectories. Empirical evaluations demonstrate that FlowCast achieves $>2.5\times$ speedup in image generation, video generation, and editing tasks, outperforming existing baselines with no quality loss as compared to standard full generation.

Method: MED (Measure-Explain-Diagnose) framework: coarse-to-fine analysis that separates intrinsic capability changes from tool-induced effects, decomposes tool-induced performance into gain/harm terms, and probes driving mechanisms.

Result: Across two VLMs with different tool priors and six benchmarks, improvements are dominated by intrinsic learning; tool-use RL mainly reduces tool-induced harm (fewer call-induced errors, weaker tool schema interference) with limited progress in tool-based correction of intrinsic failures.

Conclusion: Current vision tool-use RL learns to coexist safely with tools rather than master them, suggesting tools serve more as safety mechanisms than capability enhancers.

Abstract: Vision tool-use reinforcement learning (RL) can equip vision-language models with visual operators such as crop-and-zoom and achieves strong performance gains, yet it remains unclear whether these gains are driven by improvements in tool use or evolving intrinsic capabilities.We introduce MED (Measure-Explain-Diagnose), a coarse-to-fine framework that disentangles intrinsic capability changes from tool-induced effects, decomposes the tool-induced performance difference into gain and harm terms, and probes the mechanisms driving their evolution. Across checkpoint-level analyses on two VLMs with different tool priors and six benchmarks, we find that improvements are dominated by intrinsic learning, while tool-use RL mainly reduces tool-induced harm (e.g., fewer call-induced errors and weaker tool schema interference) and yields limited progress in tool-based correction of intrinsic failures. Overall, current vision tool-use RL learns to coexist safely with tools rather than master them.

Method: Cognitive-inspired multi-agent framework operationalizing Conceptual Blending Theory through Schema Grammar, with specialized agents for perception, transfer, generation, and hierarchical diagnosis with closed-loop backtracking.

Result: Significantly outperforms SOTA baselines in metaphor consistency, analogy appropriateness, and visual creativity, as demonstrated through extensive experiments and human evaluations.

Conclusion: The approach enables automated high-impact creative applications in advertising and media by bridging the gap between surface-level generation and abstract creative logic transfer.

Abstract: A visual metaphor constitutes a high-order form of human creativity, employing cross-domain semantic fusion to transform abstract concepts into impactful visual rhetoric. Despite the remarkable progress of generative AI, existing models remain largely confined to pixel-level instruction alignment and surface-level appearance preservation, failing to capture the underlying abstract logic necessary for genuine metaphorical generation. To bridge this gap, we introduce the task of Visual Metaphor Transfer (VMT), which challenges models to autonomously decouple the “creative essence” from a reference image and re-materialize that abstract logic onto a user-specified target subject. We propose a cognitive-inspired, multi-agent framework that operationalizes Conceptual Blending Theory (CBT) through a novel Schema Grammar (“G”). This structured representation decouples relational invariants from specific visual entities, providing a rigorous foundation for cross-domain logic re-instantiation. Our pipeline executes VMT through a collaborative system of specialized agents: a perception agent that distills the reference into a schema, a transfer agent that maintains generic space invariance to discover apt carriers, a generation agent for high-fidelity synthesis and a hierarchical diagnostic agent that mimics a professional critic, performing closed-loop backtracking to identify and rectify errors across abstract logic, component selection, and prompt encoding. Extensive experiments and human evaluations demonstrate that our method significantly outperforms SOTA baselines in metaphor consistency, analogy appropriateness, and visual creativity, paving the way for automated high-impact creative applications in advertising and media. Source code will be made publicly available.

Method: Propose a technique to convert fixed compression rate VAEs into models supporting multi-level temporal compression. Use minimal fine-tuning approach to counteract performance decline at elevated compression rates. Investigate integration with diffusion-based generative models (DiT).

Result: Examine how varying compression levels impact model performance across diverse video segments. Provide empirical evidence on effectiveness. Demonstrate successful concurrent training and compatibility with diffusion frameworks.

Conclusion: The approach enables multi-level temporal compression for video VAEs with minimal fine-tuning, maintaining performance at higher compression rates and showing compatibility with diffusion-based generative models.

Abstract: Latent Video Diffusion Models (LVDMs) rely on Variational Autoencoders (VAEs) to compress videos into compact latent representations. For continuous Variational Autoencoders (VAEs), achieving higher compression rates is desirable; yet, the efficiency notably declines when extra sampling layers are added without expanding the dimensions of hidden channels. In this paper, we present a technique to convert fixed compression rate VAEs into models that support multi-level temporal compression, providing a straightforward and minimal fine-tuning approach to counteract performance decline at elevated compression rates.Moreover, we examine how varying compression levels impact model performance over video segments with diverse characteristics, offering empirical evidence on the effectiveness of our proposed approach. We also investigate the integration of our multi-level temporal compression VAE with diffusion-based generative models, DiT, highlighting successful concurrent training and compatibility within these frameworks. This investigation illustrates the potential uses of multi-level temporal compression.

Method: Uses entropy analysis to capture modeling dynamics evolution. Adaptively determines acceleration activation scale by identifying inflection points of scale entropy growth. Dynamically computes distinct token reduction ratios for each scale and layer, pruning low-entropy tokens while reusing cache from prior scale residuals.

Result: Extensive experiments and analyses validate NOVA as a simple yet effective training-free acceleration framework for VAR models.

Conclusion: NOVA provides an effective training-free acceleration solution for VAR models by leveraging entropy analysis to adaptively reduce tokens while maintaining generation quality.

Abstract: Visual AutoRegressive modeling (VAR) suffers from substantial computational cost due to the massive token count involved. Failing to account for the continuous evolution of modeling dynamics, existing VAR token reduction methods face three key limitations: heuristic stage partition, non-adaptive schedules, and limited acceleration scope, thereby leaving significant acceleration potential untapped. Since entropy variation intrinsically reflects the transition of predictive uncertainty, it offers a principled measure to capture modeling dynamics evolution. Therefore, we propose NOVA, a training-free token reduction acceleration framework for VAR models via entropy analysis. NOVA adaptively determines the acceleration activation scale during inference by online identifying the inflection point of scale entropy growth. Through scale-linkage and layer-linkage ratio adjustment, NOVA dynamically computes distinct token reduction ratios for each scale and layer, pruning low-entropy tokens while reusing the cache derived from the residuals at the prior scale to accelerate inference and maintain generation quality. Extensive experiments and analyses validate NOVA as a simple yet effective training-free acceleration framework.

Method: Proposes T2M Mamba with two key components: (1) Periodicity-Saliency Aware Mamba using novel algorithms for keyframe weight estimation via enhanced Density Peaks Clustering and motion periodicity estimation via FFT-accelerated autocorrelation, and (2) Periodic Differential Cross-modal Alignment Module (PDCAM) to enhance robust alignment of textual and motion embeddings.

Result: Extensive experiments on HumanML3D and KIT-ML datasets show effectiveness, achieving an FID of 0.068 and consistent gains on all other metrics compared to existing methods.

Conclusion: T2M Mamba successfully addresses the coupling of periodicity and saliency in motion generation while improving robustness to semantic paraphrases, advancing the state-of-the-art in text-to-motion generation.

Abstract: Text-to-motion generation, which converts motion language descriptions into coherent 3D human motion sequences, has attracted increasing attention in fields, such as avatar animation and humanoid robotic interaction. Though existing models have achieved significant fidelity, they still suffer from two core limitations: (i) They treat motion periodicity and keyframe saliency as independent factors, overlooking their coupling and causing generation drift in long sequences. (ii) They are fragile to semantically equivalent paraphrases, where minor synonym substitutions distort textual embeddings, propagating through the decoder and producing unstable or erroneous motions. In this work, we propose T2M Mamba to address these limitations by (i) proposing Periodicity-Saliency Aware Mamba, which utilizes novel algorithms for keyframe weight estimation via enhanced Density Peaks Clustering and motion periodicity estimation via FFT-accelerated autocorrelation to capture coupled dynamics with minimal computational overhead, and (ii) constructing a Periodic Differential Cross-modal Alignment Module (PDCAM) to enhance robust alignment of textual and motion embeddings. Extensive experiments on HumanML3D and KIT-ML datasets have been conducted, confirming the effectiveness of our approach, achieving an FID of 0.068 and consistent gains on all other metrics.

Method: Propose Temporal Lipschitz-Guided Attacks (TLGA) to investigate component-level vulnerabilities. First design router attacks, then Joint TLGA (J-TLGA) that collaboratively perturbs routers and experts. Also propose Joint Temporal Lipschitz Adversarial Training (J-TLAT) for defense.

Result: Joint attacks significantly amplify adversarial effects, exposing collaborative weaknesses. J-TLAT enhances component-wise robustness, reduces inference cost by >60% compared to dense models, and improves robustness across diverse datasets/architectures.

Conclusion: The framework effectively investigates and mitigates both independent and collaborative weaknesses in video MoE models through component-level attacks and joint adversarial training.

Abstract: Mixture-of-Experts (MoE) has demonstrated strong performance in video understanding tasks, yet its adversarial robustness remains underexplored. Existing attack methods often treat MoE as a unified architecture, overlooking the independent and collaborative weaknesses of key components such as routers and expert modules. To fill this gap, we propose Temporal Lipschitz-Guided Attacks (TLGA) to thoroughly investigate component-level vulnerabilities in video MoE models. We first design attacks on the router, revealing its independent weaknesses. Building on this, we introduce Joint Temporal Lipschitz-Guided Attacks (J-TLGA), which collaboratively perturb both routers and experts. This joint attack significantly amplifies adversarial effects and exposes the Achilles’ Heel (collaborative weaknesses) of the MoE architecture. Based on these insights, we further propose Joint Temporal Lipschitz Adversarial Training (J-TLAT). J-TLAT performs joint training to further defend against collaborative weaknesses, enhancing component-wise robustness. Our framework is plug-and-play and reduces inference cost by more than 60% compared with dense models. It consistently enhances adversarial robustness across diverse datasets and architectures, effectively mitigating both the independent and collaborative weaknesses of MoE.

Method: PolyGen employs a Polylithic approach that trains on the intersection of architecturally distinct generators to marginalize out model-specific artifacts. It also introduces a Programmatic Hard Negative curriculum to enforce fine-grained syntactic understanding. The framework reallocates data budget from unique captions to multi-source variations.

Result: PolyGen outperforms the leading single-source baseline (SynthCLIP) by +19.0% on aggregate multi-task benchmarks and achieves +9.1% improvement on the SugarCrepe++ compositionality benchmark.

Conclusion: Structural diversity through multiple distinct generators is a more data-efficient scaling law than simply increasing the volume of single-source samples for synthetic vision-language data.

Abstract: Synthetic data offers a scalable solution for vision-language pre-training, yet current state-of-the-art methods typically rely on scaling up a single generative backbone, which introduces generator-specific spectral biases and limits feature diversity. In this work, we introduce PolyGen, a framework that redefines synthetic data construction by prioritizing manifold coverage and compositional rigor over simple dataset size. PolyGen employs a Polylithic approach to train on the intersection of architecturally distinct generators, effectively marginalizing out model-specific artifacts. Additionally, we introduce a Programmatic Hard Negative curriculum that enforces fine-grained syntactic understanding. By structurally reallocating the same data budget from unique captions to multi-source variations, PolyGen achieves a more robust feature space, outperforming the leading single-source baseline (SynthCLIP) by +19.0% on aggregate multi-task benchmarks and on the SugarCrepe++ compositionality benchmark (+9.1%). These results demonstrate that structural diversity is a more data-efficient scaling law than simply increasing the volume of single-source samples.

Method: PromptRL incorporates language models as trainable prompt refinement agents directly within the flow-based RL optimization loop. This creates a synergistic training regime that reshapes optimization dynamics while developing sophisticated prompt rewriting capabilities.

Result: Achieves state-of-the-art performance across multiple benchmarks: 0.97 on GenEval, 0.98 on OCR accuracy, and 24.05 on PickScore. Also improves EditReward of FLUX.1-Kontext from 1.19 to 1.43 with only 0.06 million rollouts, surpassing Gemini 2.5 Flash Image (1.37) and achieving comparable performance with ReasonNet (1.44).

Conclusion: PromptRL consistently achieves higher performance ceilings while requiring over 2× fewer rollouts compared to naive flow-only RL, demonstrating the effectiveness of integrating language models as prompt refinement agents in RL optimization for flow-based image generation.

Abstract: Flow matching models (FMs) have revolutionized text-to-image (T2I) generation, with reinforcement learning (RL) serving as a critical post-training strategy for alignment with reward objectives. In this research, we show that current RL pipelines for FMs suffer from two underappreciated yet important limitations: sample inefficiency due to insufficient generation diversity, and pronounced prompt overfitting, where models memorize specific training formulations and exhibit dramatic performance collapse when evaluated on semantically equivalent but stylistically varied prompts. We present PromptRL (Prompt Matters in RL for Flow-Based Image Generation), a framework that incorporates language models (LMs) as trainable prompt refinement agents directly within the flow-based RL optimization loop. This design yields two complementary benefits: rapid development of sophisticated prompt rewriting capabilities and, critically, a synergistic training regime that reshapes the optimization dynamics. PromptRL achieves state-of-the-art performance across multiple benchmarks, obtaining scores of 0.97 on GenEval, 0.98 on OCR accuracy, and 24.05 on PickScore. Furthermore, we validate the effectiveness of our RL approach on large-scale image editing models, improving the EditReward of FLUX.1-Kontext from 1.19 to 1.43 with only 0.06 million rollouts, surpassing Gemini 2.5 Flash Image (also known as Nano Banana), which scores 1.37, and achieving comparable performance with ReasonNet (1.44), which relied on fine-grained data annotations along with a complex multi-stage training. Our extensive experiments empirically demonstrate that PromptRL consistently achieves higher performance ceilings while requiring over 2$\times$ fewer rollouts compared to naive flow-only RL. Our code is available at https://github.com/G-U-N/UniRL.

Method: ALI (Augmented Latent Intrinsics) balances semantic context and dense photometric structure by fusing features from a pixel-aligned visual encoder into a latent-intrinsic framework, along with a self-supervised refinement strategy to address data scarcity.

Result: ALI achieves strong improvements in relighting quality, with the largest gains on complex, specular materials, outperforming previous methods that rely solely on semantic encoders.

Conclusion: The paper demonstrates that balancing semantic and photometric features is crucial for high-quality image relighting, especially for challenging materials, and introduces an effective framework to achieve this balance.

Abstract: Image-to-image relighting requires representations that disentangle scene properties from illumination. Recent methods rely on latent intrinsic representations but remain under-constrained and often fail on challenging materials such as metal and glass. A natural hypothesis is that stronger pretrained visual priors should resolve these failures. We find the opposite: features from top-performing semantic encoders often degrade relighting quality, revealing a fundamental trade-off between semantic abstraction and photometric fidelity. We study this trade-off and introduce Augmented Latent Intrinsics (ALI), which balances semantic context and dense photometric structure by fusing features from a pixel-aligned visual encoder into a latent-intrinsic framework, together with a self-supervised refinement strategy to mitigate the scarcity of paired real-world data. Trained only on unlabeled real-world image pairs and paired with a dense, pixel-aligned visual prior, ALI achieves strong improvements in relighting, with the largest gains on complex, specular materials. Project page: https:\augmented-latent-intrinsics.github.io

Method: Designs Parabolic Position Encoding (PaPE) based on principles distilled from prior work: translation invariance, rotation invariance (PaPE-RI), distance decay, directionality, and context awareness. Evaluates on 8 datasets spanning 4 vision modalities (images, point clouds, videos, event camera streams).

Result: PaPE or PaPE-RI achieves top performance on 7 out of 8 datasets. Extrapolation experiments on ImageNet-1K show PaPE extrapolates remarkably well, improving by up to 10.5% over next-best position encoding in absolute terms.

Conclusion: PaPE effectively addresses vision-specific characteristics in position encoding and demonstrates superior performance across multiple vision modalities with strong extrapolation capabilities.

Abstract: We propose Parabolic Position Encoding (PaPE), a parabola-based position encoding for vision modalities in attention-based architectures. Given a set of vision tokens-such as images, point clouds, videos, or event camera streams-our objective is to encode their positions while accounting for the characteristics of vision modalities. Prior works have largely extended position encodings from 1D-sequences in language to nD-structures in vision, but only with partial account of vision characteristics. We address this gap by designing PaPE from principles distilled from prior work: translation invariance, rotation invariance (PaPE-RI), distance decay, directionality, and context awareness. We evaluate PaPE on 8 datasets that span 4 modalities. We find that either PaPE or PaPE-RI achieves the top performance on 7 out of 8 datasets. Extrapolation experiments on ImageNet-1K show that PaPE extrapolates remarkably well, improving in absolute terms by up to 10.5% over the next-best position encoding. Code is available at https://github.com/DTU-PAS/parabolic-position-encoding.

Method: Introduces affinity-guided self-attention to capture intra-image similarities and affinity-guided cross-attention to model cross-image correspondences, integrating lightweight SSM-inspired linear attention mechanisms for efficient fine-grained localization.

Result: Extensive experiments on benchmark bio-forensic datasets demonstrate significant improvements over competitive baselines in accurately detecting duplicated regions.

Conclusion: BioTamperNet effectively addresses the unique challenges of biomedical image tampering detection through specialized attention mechanisms and achieves state-of-the-art performance.

Abstract: We propose BioTamperNet, a novel framework for detecting duplicated regions in tampered biomedical images, leveraging affinity-guided attention inspired by State Space Model (SSM) approximations. Existing forensic models, primarily trained on natural images, often underperform on biomedical data where subtle manipulations can compromise experimental validity. To address this, BioTamperNet introduces an affinity-guided self-attention module to capture intra-image similarities and an affinity-guided cross-attention module to model cross-image correspondences. Our design integrates lightweight SSM-inspired linear attention mechanisms to enable efficient, fine-grained localization. Trained end-to-end, BioTamperNet simultaneously identifies tampered regions and their source counterparts. Extensive experiments on the benchmark bio-forensic datasets demonstrate significant improvements over competitive baselines in accurately detecting duplicated regions. Code - https://github.com/SoumyaroopNandi/BioTamperNet

Method: Three vision-based approaches: 1) Direct object detection using YOLOv13, 2) Segmentation-assisted classification using SAM2 with EfficientNetV2 or YOLOv13, and 3) Query-based Vision-Language Models (Qwen2.5-VL-7b and Qwen2.5-VL-32b). The collocation of gaze points with object semantics is investigated.

Result: YOLOv13 and Qwen2.5-VL-32b significantly outperform other approaches with Macro F1-Scores over 0.84. Qwen2.5-VL-32b showed superior robustness for identifying small safety-critical objects like traffic lights, especially in nighttime conditions. Segmentation-assisted approach suffered from “part-versus-whole” semantic gap leading to poor recall.

Conclusion: There’s a fundamental trade-off between real-time efficiency of traditional detectors and richer contextual understanding/robustness of large VLMs. Findings provide practical guidance for designing human-aware intelligent driver monitoring systems.

Abstract: Understanding where drivers direct their visual attention during driving, as characterized by gaze behavior, is critical for developing next-generation advanced driver-assistance systems and improving road safety. This paper tackles this challenge as a semantic identification task from the road scenes captured by a vehicle’s front-view camera. Specifically, the collocation of gaze points with object semantics is investigated using three distinct vision-based approaches: direct object detection (YOLOv13), segmentation-assisted classification (SAM2 paired with EfficientNetV2 versus YOLOv13), and query-based Vision-Language Models, VLMs (Qwen2.5-VL-7b versus Qwen2.5-VL-32b). The results demonstrate that the direct object detection (YOLOv13) and Qwen2.5-VL-32b significantly outperform other approaches, achieving Macro F1-Scores over 0.84. The large VLM (Qwen2.5-VL-32b), in particular, exhibited superior robustness and performance for identifying small, safety-critical objects such as traffic lights, especially in adverse nighttime conditions. Conversely, the segmentation-assisted paradigm suffers from a “part-versus-whole” semantic gap that led to large failure in recall. The results reveal a fundamental trade-off between the real-time efficiency of traditional detectors and the richer contextual understanding and robustness offered by large VLMs. These findings provide critical insights and practical guidance for the design of future human-aware intelligent driver monitoring systems.

Method: Investigates behavior of quantized small-variant vision transformers (DeiT, DeiT3, ViT) on common OOD datasets. Compares ID vs OOD performance, analyzes quantization effects from 4-bit to full precision, and examines impact of pretraining scale (ImageNet-1k vs ImageNet-22k).

Result: 4-bit models show initial instabilities, especially DeiT3 trained on ImageNet-22k dropping 17% from quantization error. ViT shows reasonable ID quantization robustness. OOD detection reveals larger quantization deltas for ImageNet-22k pretrained models (15.0-19.2% AUPR-out delta) vs ImageNet-1k (9.5-12.0% delta).

Conclusion: Pretraining on large-scale datasets may hinder low-bit quantization robustness in OOD detection. Data augmentation may be more beneficial than large-scale pretraining for quantization robustness.

Abstract: Vision transformers have shown remarkable performance in vision tasks, but enabling them for accessible and real-time use is still challenging. Quantization reduces memory and inference costs at the risk of performance loss. Strides have been made to mitigate low precision issues mainly by understanding in-distribution (ID) task behaviour, but the attention mechanism may provide insight on quantization attributes by exploring out-of-distribution (OOD) situations. We investigate the behaviour of quantized small-variant popular vision transformers (DeiT, DeiT3, and ViT) on common OOD datasets. ID analyses show the initial instabilities of 4-bit models, particularly of those trained on the larger ImageNet-22k, as the strongest FP32 model, DeiT3, sharply drop 17% from quantization error to be one of the weakest 4-bit models. While ViT shows reasonable quantization robustness for ID calibration, OOD detection reveals more: ViT and DeiT3 pretrained on ImageNet-22k respectively experienced a 15.0% and 19.2% average quantization delta in AUPR-out between full precision to 4-bit while their ImageNet-1k-only counterparts experienced a 9.5% and 12.0% delta. Overall, our results suggest pretraining on large scale datasets may hinder low-bit quantization robustness in OOD detection and that data augmentation may be a more beneficial option.

Method: Proposes Logit Lens Loss (LLL) - a complementary loss to next-token prediction that makes visual token embeddings semantically aligned with textual concepts describing their image regions. Constrains mixing of image/text tokens in self-attention to prevent loss of localized visual information.

Result: LLL makes Logit Lens practically relevant by producing meaningful object confidence maps in images. Also improves performance on vision-centric tasks like segmentation without attaching special heads.

Conclusion: LLL effectively preserves visual locality in VLMs, enabling better explainability through Logit Lens visualization and improving vision task performance without architectural modifications.

Abstract: Logit Lens has been proposed for visualizing tokens that contribute most to LLM answers. Recently, Logit Lens was also shown to be applicable in autoregressive Vision-Language Models (VLMs), where it illustrates the conceptual content of image tokens in the form of heatmaps, e.g., which image tokens are likely to depict the concept of cat in a given image. However, the visual content of image tokens often gets diffused to language tokens, and consequently, the locality of visual information gets mostly destroyed, which renders Logit Lens visualization unusable for explainability. To address this issue, we introduce a complementary loss to next-token prediction (NTP) to prevent the visual tokens from losing the visual representation inherited from corresponding image patches. The proposed Logit Lens Loss (LLL) is designed to make visual token embeddings more semantically aligned with the textual concepts that describe their image regions (e.g., patches containing a cat with the word “cat”), without requiring any architectural modification or large-scale training. This way, LLL constrains the mixing of image and text tokens in the self-attention layers in order to prevent image tokens from losing their localized visual information. As our experiments show, LLL not only makes Logit Lens practically relevant by producing meaningful object confidence maps in images, but also improves performance on vision-centric tasks like segmentation without attaching any special heads.

Method: Uses Sliding Window Path Signature (SW-PS) to capture local structural features, combined with lightweight Linear Recurrent Units (LRU) classifier that combines RNN’s incremental processing with SSM’s parallel training.

Result: Achieved 99.62% (digits), 96.67% (English upper letters), and 94.33% (Chinese radicals) accuracy on CASIA-OLHWDB1.1 dataset with random rotation up to ±180°. Outperforms competing models in convergence speed and test accuracy.

Conclusion: SW-PS+LRU framework effectively addresses rotation invariance problem in online handwritten character recognition, demonstrating superior performance across different character types.

Abstract: Online handwritten character recognition leverages stroke order and dynamic features, which generally provide higher accuracy and robustness compared with offline recognition. However, in practical applications, rotational deformations can disrupt the spatial layout of strokes, substantially reducing recognition accuracy. Extracting rotation-invariant features therefore remains a challenging open problem. In this work, we employ the Sliding Window Path Signature (SW-PS) to capture local structural features of characters, and introduce the lightweight Linear Recurrent Units (LRU) as the classifier. The LRU combine the fast incremental processing capability of recurrent neural networks (RNN) with the efficient parallel training of state space models (SSM), while reliably modelling dynamic stroke characteristics. We conducted recognition experiments with random rotation angle up to $\pm 180^{\circ}$ on three subsets of the CASIA-OLHWDB1.1 dataset: digits, English upper letters, and Chinese radicals. The accuracies achieved after ensemble learning were $99.62%$, $96.67%$, and $94.33%$, respectively. Experimental results demonstrate that the proposed SW-PS+LRU framework consistently surpasses competing models in both convergence speed and test accuracy.

Method: Proposes InteractAvatar with dual-stream framework: 1) Perception and Interaction Module (PIM) for text-aligned interaction motions using detection-enhanced perception, and 2) Audio-Interaction Aware Generation Module (AIM) for synthesizing talking avatars with object interactions. Uses motion-to-video aligner for parallel co-generation.

Result: Establishes GroundedInter benchmark for GHOI video generation evaluation. Extensive experiments demonstrate effectiveness in generating grounded human-object interactions for talking avatars.

Conclusion: InteractAvatar successfully addresses the control-quality dilemma in GHOI generation by decoupling perception/planning from video synthesis, enabling realistic talking avatars with object interactions.

Abstract: Generating talking avatars is a fundamental task in video generation. Although existing methods can generate full-body talking avatars with simple human motion, extending this task to grounded human-object interaction (GHOI) remains an open challenge, requiring the avatar to perform text-aligned interactions with surrounding objects. This challenge stems from the need for environmental perception and the control-quality dilemma in GHOI generation. To address this, we propose a novel dual-stream framework, InteractAvatar, which decouples perception and planning from video synthesis for grounded human-object interaction. Leveraging detection to enhance environmental perception, we introduce a Perception and Interaction Module (PIM) to generate text-aligned interaction motions. Additionally, an Audio-Interaction Aware Generation Module (AIM) is proposed to synthesize vivid talking avatars performing object interactions. With a specially designed motion-to-video aligner, PIM and AIM share a similar network structure and enable parallel co-generation of motions and plausible videos, effectively mitigating the control-quality dilemma. Finally, we establish a benchmark, GroundedInter, for evaluating GHOI video generation. Extensive experiments and comparisons demonstrate the effectiveness of our method in generating grounded human-object interactions for talking avatars. Project page: https://interactavatar.github.io

Method: Proposes FSCA-Net with explicit feature disentanglement into domain-invariant and domain-specific components. Uses cross-attention fusion module to model interactions between components, and mutual information optimization to maximize consistency among domain-invariant features while minimizing redundancy among domain-specific ones.

Result: Extensive experiments on multiple crowd counting benchmarks show FSCA-Net effectively mitigates negative transfer and achieves state-of-the-art cross-dataset generalization performance.

Conclusion: FSCA-Net provides a robust and scalable solution for real-world crowd analysis by addressing domain discrepancy challenges through explicit feature disentanglement and adaptive fusion mechanisms.

Abstract: Crowd counting plays a vital role in public safety, traffic regulation, and smart city management. However, despite the impressive progress achieved by CNN- and Transformer-based models, their performance often deteriorates when applied across diverse environments due to severe domain discrepancies. Direct joint training on multiple datasets, which intuitively should enhance generalization, instead results in negative transfer, as shared and domain-specific representations become entangled. To address this challenge, we propose the Feature Separation and Cross-Attention Network FSCA-Net, a unified framework that explicitly disentangles feature representations into domain-invariant and domain-specific components. A novel cross-attention fusion module adaptively models interactions between these components, ensuring effective knowledge transfer while preserving dataset-specific discriminability. Furthermore, a mutual information optimization objective is introduced to maximize consistency among domain-invariant features and minimize redundancy among domain-specific ones, promoting complementary shared-private representations. Extensive experiments on multiple crowd counting benchmarks demonstrate that FSCA-Net effectively mitigates negative transfer and achieves state-of-the-art cross-dataset generalization, providing a robust and scalable solution for real-world crowd analysis.

Method: Introduces Cognitive Supersensing training paradigm with Latent Visual Imagery Prediction (LVIP) head that learns sequences of visual cognitive latent embeddings aligned with answers, forming vision-based internal reasoning chains. Adds reinforcement learning stage to optimize text reasoning paths based on grounded visual latent.

Result: MLLMs trained with Cognitive Supersensing significantly outperform state-of-the-art baselines on CogSense-Bench (comprehensive VQA benchmark assessing five cognitive dimensions) and show superior generalization on out-of-domain mathematics and science VQA benchmarks.

Conclusion: Internal visual imagery is key to bridging the gap between perceptual recognition and cognitive understanding in MLLMs. The approach demonstrates that visual reasoning mechanisms are essential for solving complex cognitive problems requiring visual memory.

Abstract: Multimodal Large Language Models (MLLMs) have achieved remarkable success in open-vocabulary perceptual tasks, yet their ability to solve complex cognitive problems remains limited, especially when visual details are abstract and require visual memory. Current approaches primarily scale Chain-of-Thought (CoT) reasoning in the text space, even when language alone is insufficient for clear and structured reasoning, and largely neglect visual reasoning mechanisms analogous to the human visuospatial sketchpad and visual imagery. To mitigate this deficiency, we introduce Cognitive Supersensing, a novel training paradigm that endows MLLMs with human-like visual imagery capabilities by integrating a Latent Visual Imagery Prediction (LVIP) head that jointly learns sequences of visual cognitive latent embeddings and aligns them with the answer, thereby forming vision-based internal reasoning chains. We further introduce a reinforcement learning stage that optimizes text reasoning paths based on this grounded visual latent. To evaluate the cognitive capabilities of MLLMs, we present CogSense-Bench, a comprehensive visual question answering (VQA) benchmark assessing five cognitive dimensions. Extensive experiments demonstrate that MLLMs trained with Cognitive Supersensing significantly outperform state-of-the-art baselines on CogSense-Bench and exhibit superior generalization on out-of-domain mathematics and science VQA benchmarks, suggesting that internal visual imagery is potentially key to bridging the gap between perceptual recognition and cognitive understanding. We will open-source the CogSense-Bench and our model weights.

Method: Proposes MUN benchmark with surprising visual-language pairs, and R-ICL framework with Multimodal Ensemble Retriever (MER) to identify relevant exemplars even when image-text pairs are discordant, transferring reasoning from larger to smaller models without training.

Result: R-ICL shows average 8.3% improvement over baseline ICL methods on MUN benchmark, demonstrating effectiveness in low-frequency, atypical settings.

Conclusion: MUN enables evaluation of visual-language models’ robustness in non-prototypical scenarios, opening directions for improving adaptability in culturally diverse, real-world contexts.

Abstract: Commonsense reasoning in multimodal contexts remains a foundational challenge in artificial intelligence. We introduce Multimodal UNcommonsense(MUN), a benchmark designed to evaluate models’ ability to handle scenarios that deviate from typical visual or contextual expectations. MUN pairs visual scenes with surprising or unlikely outcomes described in natural language, prompting models to either rationalize seemingly odd images using everyday logic or uncover unexpected interpretations in ordinary scenes. To support this task, we propose a retrieval-based in-context learning (R-ICL) framework that transfers reasoning capabilities from larger models to smaller ones without additional training. Leveraging a novel Multimodal Ensemble Retriever (MER), our method identifies semantically relevant exemplars even when image and text pairs are deliberately discordant. Experiments show an average improvement of 8.3% over baseline ICL methods, highlighting the effectiveness of R-ICL in low-frequency, atypical settings. MUN opens new directions for evaluating and improving visual-language models’ robustness and adaptability in real-world, culturally diverse, and non-prototypical scenarios.

Method: Proposes a diffusion-based image compression method requiring only a single-step diffusion process. Introduces a discriminator operating on compact feature representations instead of raw pixels to enhance perceptual quality, leveraging features’ ability to better capture high-level texture and structural details.

Result: Achieves comparable compression performance while offering 46× faster inference speed compared to recent diffusion-based approaches.

Conclusion: The proposed single-step diffusion compression method significantly reduces inference latency while maintaining perceptual quality, making diffusion-based compression more practical for deployment.

Abstract: Diffusion-based image compression methods have achieved notable progress, delivering high perceptual quality at low bitrates. However, their practical deployment is hindered by significant inference latency and heavy computational overhead, primarily due to the large number of denoising steps required during decoding. To address this problem, we propose a diffusion-based image compression method that requires only a single-step diffusion process, significantly improving inference speed. To enhance the perceptual quality of reconstructed images, we introduce a discriminator that operates on compact feature representations instead of raw pixels, leveraging the fact that features better capture high-level texture and structural details. Experimental results show that our method delivers comparable compression performance while offering a 46$\times$ faster inference speed compared to recent diffusion-based approaches. The source code and models are available at https://github.com/cheesejiang/OSDiff.

Method: Proposes SGHA-Attack with semantic-guided hierarchical alignment: 1) Creates reference pool using text-to-image model conditioned on target prompt, selects Top-K most semantically relevant anchors; 2) Aligns intermediate visual representations at multiple depths with global and spatial granularities; 3) Synchronizes intermediate visual and textual features in shared latent subspace for early cross-modal supervision.

Result: Extensive experiments show SGHA-Attack achieves stronger targeted transferability than prior methods on open-source and commercial black-box VLMs, and remains robust against preprocessing and purification defenses.

Conclusion: SGHA-Attack demonstrates that leveraging multiple semantic references and hierarchical alignment across intermediate layers significantly improves transferability of adversarial attacks across diverse vision-language models.

Abstract: Large vision-language models (VLMs) are vulnerable to transfer-based adversarial perturbations, enabling attackers to optimize on surrogate models and manipulate black-box VLM outputs. Prior targeted transfer attacks often overfit surrogate-specific embedding space by relying on a single reference and emphasizing final-layer alignment, which underutilizes intermediate semantics and degrades transfer across heterogeneous VLMs. To address this, we propose SGHA-Attack, a Semantic-Guided Hierarchical Alignment framework that adopts multiple target references and enforces intermediate-layer consistency. Concretely, we generate a visually grounded reference pool by sampling a frozen text-to-image model conditioned on the target prompt, and then carefully select the Top-K most semantically relevant anchors under the surrogate to form a weighted mixture for stable optimization guidance. Building on these anchors, SGHA-Attack injects target semantics throughout the feature hierarchy by aligning intermediate visual representations at both global and spatial granularities across multiple depths, and by synchronizing intermediate visual and textual features in a shared latent subspace to provide early cross-modal supervision before the final projection. Extensive experiments on open-source and commercial black-box VLMs show that SGHA-Attack achieves stronger targeted transferability than prior methods and remains robust under preprocessing and purification defenses.

Method: Proposes HandMCM based on state space models (Mamba) with modules for local information injection/filtering and correspondence modeling to learn dynamic kinematic topology across occlusion scenarios, integrated with multi-modal image features.

Result: Significantly outperforms current state-of-the-art methods on three benchmark datasets, particularly in challenging scenarios with severe occlusions.

Conclusion: Demonstrates potential to advance accuracy and reliability of 3D hand pose estimation in practical applications by effectively handling occlusions through novel Mamba-based architecture with multi-modal features.

Abstract: 3D hand pose estimation that involves accurate estimation of 3D human hand keypoint locations is crucial for many human-computer interaction applications such as augmented reality. However, this task poses significant challenges due to self-occlusion of the hands and occlusions caused by interactions with objects. In this paper, we propose HandMCM to address these challenges. Our HandMCM is a novel method based on the powerful state space model (Mamba). By incorporating modules for local information injection/filtering and correspondence modeling, the proposed correspondence Mamba effectively learns the highly dynamic kinematic topology of keypoints across various occlusion scenarios. Moreover, by integrating multi-modal image features, we enhance the robustness and representational capacity of the input, leading to more accurate hand pose estimation. Empirical evaluations on three benchmark datasets demonstrate that our model significantly outperforms current state-of-the-art methods, particularly in challenging scenarios involving severe occlusions. These results highlight the potential of our approach to advance the accuracy and reliability of 3D hand pose estimation in practical applications.

Method: Proposes Temperature Annealed Few-step Sampling with Group Relative Policy Optimization (TAFS-GRPO) that iteratively injects adaptive temporal noise onto one-step samples, annealing outputs to introduce stochasticity while preserving semantic integrity. Uses step-aware advantage integration with GRPO to provide dense, step-specific rewards without requiring differentiable reward functions.

Result: Extensive experiments show TAFS-GRPO achieves strong performance in few-step text-to-image generation and significantly improves alignment of generated images with human preferences.

Conclusion: The proposed framework effectively addresses limitations of existing RL approaches for flow matching models, enabling efficient few-step generation with better human preference alignment through temperature annealing and group relative policy optimization.

Abstract: Recent advances in flow matching models, particularly with reinforcement learning (RL), have significantly enhanced human preference alignment in few step text to image generators. However, existing RL based approaches for flow matching models typically rely on numerous denoising steps, while suffering from sparse and imprecise reward signals that often lead to suboptimal alignment. To address these limitations, we propose Temperature Annealed Few step Sampling with Group Relative Policy Optimization (TAFS GRPO), a novel framework for training flow matching text to image models into efficient few step generators well aligned with human preferences. Our method iteratively injects adaptive temporal noise onto the results of one step samples. By repeatedly annealing the model’s sampled outputs, it introduces stochasticity into the sampling process while preserving the semantic integrity of each generated image. Moreover, its step aware advantage integration mechanism combines the GRPO to avoid the need for the differentiable of reward function and provide dense and step specific rewards for stable policy optimization. Extensive experiments demonstrate that TAFS GRPO achieves strong performance in few step text to image generation and significantly improves the alignment of generated images with human preferences. The code and models of this work will be available to facilitate further research.

Method: Proposes Saliency Mamba (Samba) with saliency-guided Mamba blocks using spatial neighborhood scanning to preserve spatial continuity, and context-aware upsampling for hierarchical feature alignment. Samba+ extends this with multi-task joint training, hub-and-spoke graph attention for cross-modal fusion, and modality-anchored continual learning to handle arbitrary modalities.

Result: Samba outperforms existing methods across six SOD tasks on 22 datasets with lower computational cost. Samba+ achieves even better results using a single versatile model trained in a multi-task manner.

Conclusion: The Mamba-based Samba framework provides an effective solution for various SOD tasks, balancing computational efficiency and performance, with Samba+ offering a unified approach to handle multiple modalities and tasks.

Abstract: Existing salient object detection (SOD) models are generally constrained by the limited receptive fields of convolutional neural networks (CNNs) and quadratic computational complexity of Transformers. Recently, the emerging state-space model, namely Mamba, has shown great potential in balancing global receptive fields and computational efficiency. As a solution, we propose Saliency Mamba (Samba), a pure Mamba-based architecture that flexibly handles various distinct SOD tasks, including RGB/RGB-D/RGB-T SOD, video SOD (VSOD), RGB-D VSOD, and visible-depth-thermal SOD. Specifically, we rethink the scanning strategy of Mamba for SOD, and introduce a saliency-guided Mamba block (SGMB) that features a spatial neighborhood scanning (SNS) algorithm to preserve the spatial continuity of salient regions. A context-aware upsampling (CAU) method is also proposed to promote hierarchical feature alignment and aggregation by modeling contextual dependencies. As one step further, to avoid the “task-specific” problem as in previous SOD solutions, we develop Samba+, which is empowered by training Samba in a multi-task joint manner, leading to a more unified and versatile model. Two crucial components that collaboratively tackle challenges encountered in input of arbitrary modalities and continual adaptation are investigated. Specifically, a hub-and-spoke graph attention (HGA) module facilitates adaptive cross-modal interactive fusion, and a modality-anchored continual learning (MACL) strategy alleviates inter-modal conflicts together with catastrophic forgetting. Extensive experiments demonstrate that Samba individually outperforms existing methods across six SOD tasks on 22 datasets with lower computational cost, whereas Samba+ achieves even superior results on these tasks and datasets by using a single trained versatile model. Additional results further demonstrate the potential of our Samba framework.

Method: Proposes UV-M3TL with two core components: 1) Dual-branch spatial channel multimodal embedding (DB-SCME) that uses a dual-branch structure to explicitly model task-shared and task-specific features, enhancing cross-task knowledge transfer while mitigating conflicts. 2) Adaptive feature-decoupled multi-task loss (AFD-Loss) that introduces adaptive weighting based on learning dynamics and feature decoupling constraints to stabilize joint optimization and learn diverse multi-task representations.

Result: Achieves state-of-the-art performance across all four tasks (driver behavior, driver emotion, vehicle behavior, traffic context) on the AIDE dataset. Demonstrates versatility by achieving strong performance on additional public benchmarks (BDD100K, CityScapes, NYUD-v2, PASCAL-Context) across diverse task combinations, attaining SOTA results on most tasks.

Conclusion: UV-M3TL provides an effective unified framework for multimodal multi-task learning in ADAS applications, successfully mitigating inter-task negative transfer while maintaining strong performance across diverse perception tasks. The framework demonstrates both specialization for driver assistance tasks and generalizability to other multi-task perception benchmarks.

Abstract: Advanced Driver Assistance Systems (ADAS) need to understand human driver behavior while perceiving their navigation context, but jointly learning these heterogeneous tasks would cause inter-task negative transfer and impair system performance. Here, we propose a Unified and Versatile Multimodal Multi-Task Learning (UV-M3TL) framework to simultaneously recognize driver behavior, driver emotion, vehicle behavior, and traffic context, while mitigating inter-task negative transfer. Our framework incorporates two core components: dual-branch spatial channel multimodal embedding (DB-SCME) and adaptive feature-decoupled multi-task loss (AFD-Loss). DB-SCME enhances cross-task knowledge transfer while mitigating task conflicts by employing a dual-branch structure to explicitly model salient task-shared and task-specific features. AFD-Loss improves the stability of joint optimization while guiding the model to learn diverse multi-task representations by introducing an adaptive weighting mechanism based on learning dynamics and feature decoupling constraints. We evaluate our method on the AIDE dataset, and the experimental results demonstrate that UV-M3TL achieves state-of-the-art performance across all four tasks. To further prove the versatility, we evaluate UV-M3TL on additional public multi-task perception benchmarks (BDD100K, CityScapes, NYUD-v2, and PASCAL-Context), where it consistently delivers strong performance across diverse task combinations, attaining state-of-the-art results on most tasks.

Method: ToPi uses offline calibration-driven sensitivity analysis to identify pivotal attention layers as a proxy for redundancy estimation. It then derives an influence metric to quantify each context token’s contribution for selective pruning, coupled with a temporal update strategy that adapts to the evolving diffusion trajectory. The framework is training-free and specifically tailored for in-context generation in DiTs.

Result: Empirical evaluations show ToPi achieves over 30% speedup in inference while maintaining structural fidelity and visual consistency across complex image generation tasks.

Conclusion: ToPi effectively addresses the computational bottleneck in in-context generation for Diffusion Transformers through a principled token pruning approach that preserves generation quality while significantly improving efficiency.

Abstract: In-context generation significantly enhances Diffusion Transformers (DiTs) by enabling controllable image-to-image generation through reference examples. However, the resulting input concatenation drastically increases sequence length, creating a substantial computational bottleneck. Existing token reduction techniques, primarily tailored for text-to-image synthesis, fall short in this paradigm as they apply uniform reduction strategies, overlooking the inherent role asymmetry between reference contexts and target latents across spatial, temporal, and functional dimensions. To bridge this gap, we introduce ToPi, a training-free token pruning framework tailored for in-context generation in DiTs. Specifically, ToPi utilizes offline calibration-driven sensitivity analysis to identify pivotal attention layers, serving as a robust proxy for redundancy estimation. Leveraging these layers, we derive a novel influence metric to quantify the contribution of each context token for selective pruning, coupled with a temporal update strategy that adapts to the evolving diffusion trajectory. Empirical evaluations demonstrate that ToPi can achieve over 30% speedup in inference while maintaining structural fidelity and visual consistency across complex image generation tasks.

Method: Introduces Omni-Judge, which leverages omni-modal large language models (omni-LLMs) that naturally process audio, video, and text. The approach uses these models as judges across nine perceptual and alignment metrics for text-conditioned audio-video generation.

Result: Omni-Judge achieves correlation comparable to traditional metrics and excels on semantically demanding tasks like audio-text alignment, video-text alignment, and audio-video-text coherence. However, it underperforms on high-FPS perceptual metrics (video quality and audio-video synchronization) due to limited temporal resolution.

Conclusion: Omni-LLMs show promise as unified evaluators for multi-modal generation, offering interpretable explanations that expose inconsistencies, but have current limitations in temporal resolution. The approach enables practical downstream uses like feedback-based refinement.

Abstract: State-of-the-art text-to-video generation models such as Sora 2 and Veo 3 can now produce high-fidelity videos with synchronized audio directly from a textual prompt, marking a new milestone in multi-modal generation. However, evaluating such tri-modal outputs remains an unsolved challenge. Human evaluation is reliable but costly and difficult to scale, while traditional automatic metrics, such as FVD, CLAP, and ViCLIP, focus on isolated modality pairs, struggle with complex prompts, and provide limited interpretability. Omni-modal large language models (omni-LLMs) present a promising alternative: they naturally process audio, video, and text, support rich reasoning, and offer interpretable chain-of-thought feedback. Driven by this, we introduce Omni-Judge, a study assessing whether omni-LLMs can serve as human-aligned judges for text-conditioned audio-video generation. Across nine perceptual and alignment metrics, Omni-Judge achieves correlation comparable to traditional metrics and excels on semantically demanding tasks such as audio-text alignment, video-text alignment, and audio-video-text coherence. It underperforms on high-FPS perceptual metrics, including video quality and audio-video synchronization, due to limited temporal resolution. Omni-Judge provides interpretable explanations that expose semantic or physical inconsistencies, enabling practical downstream uses such as feedback-based refinement. Our findings highlight both the potential and current limitations of omni-LLMs as unified evaluators for multi-modal generation.

Method: PISCES uses a Dual Optimal Transport-aligned Rewards module with two components: 1) Distributional OT-aligned Quality Reward for visual quality and temporal coherence, and 2) Discrete Token-level OT-aligned Semantic Reward for spatio-temporal correspondence between text and video tokens.

Result: Experiments show PISCES outperforms both annotation-based and annotation-free methods on VBench across Quality and Semantic scores, with human preference studies validating effectiveness. The method works with multiple optimization paradigms including direct backpropagation and reinforcement learning fine-tuning.

Conclusion: PISCES successfully improves annotation-free reward supervision for text-to-video generation through optimal transport alignment, providing scalable and effective post-training without human annotations.

Abstract: Text-to-video (T2V) generation aims to synthesize videos with high visual quality and temporal consistency that are semantically aligned with input text. Reward-based post-training has emerged as a promising direction to improve the quality and semantic alignment of generated videos. However, recent methods either rely on large-scale human preference annotations or operate on misaligned embeddings from pre-trained vision-language models, leading to limited scalability or suboptimal supervision. We present $\texttt{PISCES}$, an annotation-free post-training algorithm that addresses these limitations via a novel Dual Optimal Transport (OT)-aligned Rewards module. To align reward signals with human judgment, $\texttt{PISCES}$ uses OT to bridge text and video embeddings at both distributional and discrete token levels, enabling reward supervision to fulfill two objectives: (i) a Distributional OT-aligned Quality Reward that captures overall visual quality and temporal coherence; and (ii) a Discrete Token-level OT-aligned Semantic Reward that enforces semantic, spatio-temporal correspondence between text and video tokens. To our knowledge, $\texttt{PISCES}$ is the first to improve annotation-free reward supervision in generative post-training through the lens of OT. Experiments on both short- and long-video generation show that $\texttt{PISCES}$ outperforms both annotation-based and annotation-free methods on VBench across Quality and Semantic scores, with human preference studies further validating its effectiveness. We show that the Dual OT-aligned Rewards module is compatible with multiple optimization paradigms, including direct backpropagation and reinforcement learning fine-tuning.

Method: Analyzes limitations of fragmented approaches and proposes a unified design specification for world models. Suggests that robust world models should be normative frameworks that integrally incorporate interaction, perception, symbolic reasoning, and spatial representation.

Result: Provides a structured perspective to guide future research toward more general, robust, and principled models of the world, moving beyond task-specific approaches to holistic world understanding.

Conclusion: World models need unified frameworks that integrate multiple capabilities rather than fragmented task-specific approaches. The proposed design specification aims to advance research toward more comprehensive and principled world understanding.

Abstract: World models have emerged as a critical frontier in AI research, aiming to enhance large models by infusing them with physical dynamics and world knowledge. The core objective is to enable agents to understand, predict, and interact with complex environments. However, current research landscape remains fragmented, with approaches predominantly focused on injecting world knowledge into isolated tasks, such as visual prediction, 3D estimation, or symbol grounding, rather than establishing a unified definition or framework. While these task-specific integrations yield performance gains, they often lack the systematic coherence required for holistic world understanding. In this paper, we analyze the limitations of such fragmented approaches and propose a unified design specification for world models. We suggest that a robust world model should not be a loose collection of capabilities but a normative framework that integrally incorporates interaction, perception, symbolic reasoning, and spatial representation. This work aims to provide a structured perspective to guide future research toward more general, robust, and principled models of the world.

Method: Proposes a FL framework with dynamic adaptive focal loss (DAFL) that adjusts based on each client’s sample distribution, and a client-aware weighted aggregation strategy that adapts to data size and characteristics.

Result: Outperforms DenseNet121, ResNet50, ViT-S/16, ViT-L/32, FedCLIP, Swin Transformer, CoAtNet, and MixNet on three medical datasets (ISIC, Ocular Disease, RSNA-ICH) with accuracy improvements from 0.98% to 41.69%.

Conclusion: The proposed FL framework effectively addresses class imbalance and client heterogeneity in federated medical image classification, demonstrating superior performance over various baseline models.

Abstract: While deep learning models like Vision Transformer (ViT) have achieved significant advances, they typically require large datasets. With data privacy regulations, access to many original datasets is restricted, especially medical images. Federated learning (FL) addresses this challenge by enabling global model aggregation without data exchange. However, the heterogeneity of the data and the class imbalance that exist in local clients pose challenges for the generalization of the model. This study proposes a FL framework leveraging a dynamic adaptive focal loss (DAFL) and a client-aware aggregation strategy for local training. Specifically, we design a dynamic class imbalance coefficient that adjusts based on each client’s sample distribution and class data distribution, ensuring minority classes receive sufficient attention and preventing sparse data from being ignored. To address client heterogeneity, a weighted aggregation strategy is adopted, which adapts to data size and characteristics to better capture inter-client variations. The classification results on three public datasets (ISIC, Ocular Disease and RSNA-ICH) show that the proposed framework outperforms DenseNet121, ResNet50, ViT-S/16, ViT-L/32, FedCLIP, Swin Transformer, CoAtNet, and MixNet in most cases, with accuracy improvements ranging from 0.98% to 41.69%. Ablation studies on the imbalanced ISIC dataset validate the effectiveness of the proposed loss function and aggregation strategy compared to traditional loss functions and other FL approaches. The codes can be found at: https://github.com/AIPMLab/ViT-FLDAF.

Method: Three-step pipeline: 1) Diagnose cognitive blind spots via self-guided informative instance mining, 2) Generate corrective instructions and triplets using CoT prompting of foundation MLLM with VQA-based consistency filtering, 3) Refine retriever through continual training on triplets with grouped contrastive scheme.

Result: Extensive experiments on CIRR and FashionIQ datasets show ReCALL consistently recalibrates degraded capabilities and achieves state-of-the-art performance in composed image retrieval.

Conclusion: ReCALL effectively addresses capability degradation in MLLM-based retrieval by aligning discriminative embedding space with intrinsic compositional reasoning, improving fine-grained visual-semantic understanding.

Abstract: Composed Image Retrieval (CIR) aims to retrieve target images based on a hybrid query comprising a reference image and a modification text. Early dual-tower Vision-Language Models (VLMs) struggle with cross-modality compositional reasoning required for this task. Recently, adapting generative Multimodal Large Language Models (MLLMs) for retrieval offers a promising direction. However, we identify that this adaptation strategy overlooks a fundamental issue: adapting a generative MLLM into a single-embedding discriminative retriever triggers a paradigm conflict, which leads to Capability Degradation - the deterioration of native fine-grained reasoning after retrieval adaptation. To address this challenge, we propose ReCALL (Recalibrating Capability Degradation), a model-agnostic framework that follows a diagnose-generate-refine pipeline: Firstly, we diagnose cognitive blind spots of the retriever via self-guided informative instance mining. Next, we generate corrective instructions and triplets by CoT prompting the foundation MLLM and conduct quality control with VQA-based consistency filtering. Finally, we refine the retriever through continual training on these triplets with a grouped contrastive scheme, thereby internalizing fine-grained visual-semantic distinctions and realigning the discriminative embedding space of retriever with intrinsic compositional reasoning within the MLLM. Extensive experiments on CIRR and FashionIQ show that ReCALL consistently recalibrates degraded capabilities and achieves state-of-the-art performance. Code will be released soon.

Method: 1) RL-based framework with policy network to select optimal token combinations based on contribution to correct predictions; 2) Combinatorial policy optimization with online combination space sampling to reduce exploration space and accelerate convergence.

Result: Extensive experiments on diverse video understanding benchmarks demonstrate effectiveness of CaCoVID in compressing video tokens while maintaining performance.

Conclusion: CaCoVID provides a novel contribution-aware approach to video token compression that shifts from passive preservation to active discovery of optimal token combinations, improving computational efficiency for video LLMs.

Abstract: Video large language models have demonstrated remarkable capabilities in video understanding tasks. However, the redundancy of video tokens introduces significant computational overhead during inference, limiting their practical deployment. Many compression algorithms are proposed to prioritize retaining features with the highest attention scores to minimize perturbations in attention computations. However, the correlation between attention scores and their actual contribution to correct answers remains ambiguous. To address the above limitation, we propose a novel \textbf{C}ontribution-\textbf{a}ware token \textbf{Co}mpression algorithm for \textbf{VID}eo understanding (\textbf{CaCoVID}) that explicitly optimizes the token selection policy based on the contribution of tokens to correct predictions. First, we introduce a reinforcement learning-based framework that optimizes a policy network to select video token combinations with the greatest contribution to correct predictions. This paradigm shifts the focus from passive token preservation to active discovery of optimal compressed token combinations. Secondly, we propose a combinatorial policy optimization algorithm with online combination space sampling, which dramatically reduces the exploration space for video token combinations and accelerates the convergence speed of policy optimization. Extensive experiments on diverse video understanding benchmarks demonstrate the effectiveness of CaCoVID. Codes will be released.

Method: 1) Scalable synthetic data pipeline generating photorealistic human frames with motion-aligned sequences and pixel-accurate labels; 2) Unified ViT-based dense predictor with explicit human geometric prior via CSE embeddings; 3) Lightweight channel reweighting module for geometry-feature reliability; 4) Two-stage training combining static pretraining with dynamic sequence supervision.

Result: Achieves state-of-the-art performance on THuman2.1 and Hi4D benchmarks and generalizes effectively to in-the-wild videos.

Conclusion: The proposed approach successfully addresses temporal consistency in human-centric dense prediction through synthetic data generation and geometric prior integration.

Abstract: In this work, we focus on the challenge of temporally consistent human-centric dense prediction across video sequences. Existing models achieve strong per-frame accuracy but often flicker under motion, occlusion, and lighting changes, and they rarely have paired human video supervision for multiple dense tasks. We address this gap with a scalable synthetic data pipeline that generates photorealistic human frames and motion-aligned sequences with pixel-accurate depth, normals, and masks. Unlike prior static data synthetic pipelines, our pipeline provides both frame-level labels for spatial learning and sequence-level supervision for temporal learning. Building on this, we train a unified ViT-based dense predictor that (i) injects an explicit human geometric prior via CSE embeddings and (ii) improves geometry-feature reliability with a lightweight channel reweighting module after feature fusion. Our two-stage training strategy, combining static pretraining with dynamic sequence supervision, enables the model first to acquire robust spatial representations and then refine temporal consistency across motion-aligned sequences. Extensive experiments show that we achieve state-of-the-art performance on THuman2.1 and Hi4D and generalize effectively to in-the-wild videos.

Method: Created a dataset of 2,854 anchor-linked variation pairs from original art and photographs, applying contextual transformations (illumination, weather, viewpoint, composition, color tone, mood) while preserving underlying identity.

Result: Dataset shows high identity stability, strong target attribute realization, and robust aesthetic profile exceeding large-scale web datasets, released under Apache 2.0 license.

Conclusion: Lunara Aesthetic II provides valuable resources for benchmarking, fine-tuning, and analyzing contextual generalization and identity preservation in image generation systems with interpretable supervision.

Abstract: We introduce Lunara Aesthetic II, a publicly released, ethically sourced image dataset designed to support controlled evaluation and learning of contextual consistency in modern image generation and editing systems. The dataset comprises 2,854 anchor-linked variation pairs derived from original art and photographs created by Moonworks. Each variation pair applies contextual transformations, such as illumination, weather, viewpoint, scene composition, color tone, or mood; while preserving a stable underlying identity. Lunara Aesthetic II operationalizes identity-preserving contextual variation as a supervision signal while also retaining Lunara’s signature high aesthetic scores. Results show high identity stability, strong target attribute realization, and a robust aesthetic profile that exceeds large-scale web datasets. Released under the Apache 2.0 license, Lunara Aesthetic II is intended for benchmarking, fine-tuning, and analysis of contextual generalization, identity preservation, and edit robustness in image generation and image-to-image systems with interpretable, relational supervision. The dataset is publicly available at: https://huggingface.co/datasets/moonworks/lunara-aesthetic-image-variations.

Method: Empirically evaluate NetVLAD as LCD module vs DBoW on KITTI dataset, introduce Fine-Grained Top-K precision-recall curve for LCD settings, use Faiss-accelerated nearest-neighbor search for real-time query speed.

Result: NetVLAD achieves real-time query speed with Faiss acceleration while improving accuracy and robustness over DBoW, making it a practical drop-in alternative for LCD in SLAM systems.

Conclusion: Deep learning-based visual place recognition descriptors like NetVLAD can be practically deployed for real-time loop closure detection in SLAM systems, overcoming previous computational barriers while providing superior performance.

Abstract: Loop closure detection (LCD) is a core component of simultaneous localization and mapping (SLAM): it identifies revisited places and enables pose-graph constraints that correct accumulated drift. Classic bag-of-words approaches such as DBoW are efficient but often degrade under appearance change and perceptual aliasing. In parallel, deep learning-based visual place recognition (VPR) descriptors (e.g., NetVLAD and Transformer-based models) offer stronger robustness, but their computational cost is often viewed as a barrier to real-time SLAM. In this paper, we empirically evaluate NetVLAD as an LCD module and compare it against DBoW on the KITTI dataset. We introduce a Fine-Grained Top-K precision-recall curve that better reflects LCD settings where a query may have zero or multiple valid matches. With Faiss-accelerated nearestneighbor search, NetVLAD achieves real-time query speed while improving accuracy and robustness over DBoW, making it a practical drop-in alternative for LCD in SLAM.

Method: Uses a parallel pipeline with VR Frontend (inverse kinematics for full-body pose estimation) and GA Backend (3D Gaussian Splatting avatar rendering). Introduces Binocular Batching to jointly process left/right eye views for efficient stereo rendering on high-res VR displays.

Result: System sustains interactive VR performance and outperforms image- and video-based mesh avatar baselines in user studies, achieving higher perceived appearance similarity, embodiment, and plausibility.

Conclusion: VRGaussianAvatar demonstrates effective real-time 3D Gaussian Splatting avatars for VR with minimal input requirements, offering improved visual quality and user experience over traditional mesh-based approaches.

Abstract: We present VRGaussianAvatar, an integrated system that enables real-time full-body 3D Gaussian Splatting (3DGS) avatars in virtual reality using only head-mounted display (HMD) tracking signals. The system adopts a parallel pipeline with a VR Frontend and a GA Backend. The VR Frontend uses inverse kinematics to estimate full-body pose and streams the resulting pose along with stereo camera parameters to the backend. The GA Backend stereoscopically renders a 3DGS avatar reconstructed from a single image. To improve stereo rendering efficiency, we introduce Binocular Batching, which jointly processes left and right eye views in a single batched pass to reduce redundant computation and support high-resolution VR displays. We evaluate VRGaussianAvatar with quantitative performance tests and a within-subject user study against image- and video-based mesh avatar baselines. Results show that VRGaussianAvatar sustains interactive VR performance and yields higher perceived appearance similarity, embodiment, and plausibility. Project page and source code are available at https://vrgaussianavatar.github.io.

Method: Proposes State-aware Mamba Tracker (SMTrack) with a selective state-aware space model using state-wise parameters to capture diverse temporal cues. It enables long-range temporal interactions with linear computational complexity during training and uses hidden state propagation for efficient inference.

Result: Extensive experiments show SMTrack achieves promising tracking performance with low computational costs compared to existing methods.

Conclusion: SMTrack provides an effective and efficient temporal modeling paradigm for visual tracking using state space models, addressing limitations of conventional architectures while maintaining computational efficiency.

Abstract: Visual tracking aims to automatically estimate the state of a target object in a video sequence, which is challenging especially in dynamic scenarios. Thus, numerous methods are proposed to introduce temporal cues to enhance tracking robustness. However, conventional CNN and Transformer architectures exhibit inherent limitations in modeling long-range temporal dependencies in visual tracking, often necessitating either complex customized modules or substantial computational costs to integrate temporal cues. Inspired by the success of the state space model, we propose a novel temporal modeling paradigm for visual tracking, termed State-aware Mamba Tracker (SMTrack), providing a neat pipeline for training and tracking without needing customized modules or substantial computational costs to build long-range temporal dependencies. It enjoys several merits. First, we propose a novel selective state-aware space model with state-wise parameters to capture more diverse temporal cues for robust tracking. Second, SMTrack facilitates long-range temporal interactions with linear computational complexity during training. Third, SMTrack enables each frame to interact with previously tracked frames via hidden state propagation and updating, which releases computational costs of handling temporal cues during tracking. Extensive experimental results demonstrate that SMTrack achieves promising performance with low computational costs.

Method: Proposes FreshMem with two synergistic modules: 1) Multi-scale Frequency Memory (MFM) projects overflowing frames into frequency coefficients with residual details to reconstruct global historical “gist”; 2) Space Thumbnail Memory (STM) discretizes continuous streams into episodic clusters using adaptive compression to create high-density space thumbnails.

Result: FreshMem significantly boosts Qwen2-VL baseline with gains of 5.20% on StreamingBench, 4.52% on OV-Bench, and 2.34% on OVO-Bench. As a training-free solution, it outperforms several fully fine-tuned methods.

Conclusion: FreshMem offers a highly efficient paradigm for long-horizon streaming video understanding in MLLMs by reconciling short-term fidelity with long-term coherence through frequency-space hybrid memory inspired by brain perception.

Abstract: Transitioning Multimodal Large Language Models (MLLMs) from offline to online streaming video understanding is essential for continuous perception. However, existing methods lack flexible adaptivity, leading to irreversible detail loss and context fragmentation. To resolve this, we propose FreshMem, a Frequency-Space Hybrid Memory network inspired by the brain’s logarithmic perception and memory consolidation. FreshMem reconciles short-term fidelity with long-term coherence through two synergistic modules: Multi-scale Frequency Memory (MFM), which projects overflowing frames into representative frequency coefficients, complemented by residual details to reconstruct a global historical “gist”; and Space Thumbnail Memory (STM), which discretizes the continuous stream into episodic clusters by employing an adaptive compression strategy to distill them into high-density space thumbnails. Extensive experiments show that FreshMem significantly boosts the Qwen2-VL baseline, yielding gains of 5.20%, 4.52%, and 2.34% on StreamingBench, OV-Bench, and OVO-Bench, respectively. As a training-free solution, FreshMem outperforms several fully fine-tuned methods, offering a highly efficient paradigm for long-horizon streaming video understanding.

Method: Proposes CMAFNet with two main components: 1) Semantic Recomposition Module using dictionary-based feature purification via learned codebook to suppress modality-specific noise, and 2) Contextual Semantic Integration Framework using partial-channel attention to capture global spatial dependencies. Uses position-wise normalization for cross-modal alignment.

Result: Achieves 32.2% mAP@50 and 12.5% APs on TLRGBD benchmark (94.5% small objects), outperforming strongest baseline by 9.8 and 4.0 percentage points. Lightweight variant reaches 24.8% mAP50 at 228 FPS with only 4.9M parameters.

Conclusion: CMAFNet effectively integrates RGB and depth modalities through principled purification and fusion, demonstrating superior performance for small-scale defect detection in complex aerial inspection scenarios with computational efficiency.

Abstract: Transmission line defect detection remains challenging for automated UAV inspection due to the dominance of small-scale defects, complex backgrounds, and illumination variations. Existing RGB-based detectors, despite recent progress, struggle to distinguish geometrically subtle defects from visually similar background structures under limited chromatic contrast. This paper proposes CMAFNet, a Cross-Modal Alignment and Fusion Network that integrates RGB appearance and depth geometry through a principled purify-then-fuse paradigm. CMAFNet consists of a Semantic Recomposition Module that performs dictionary-based feature purification via a learned codebook to suppress modality-specific noise while preserving defect-discriminative information, and a Contextual Semantic Integration Framework that captures global spatial dependencies using partial-channel attention to enhance structural semantic reasoning. Position-wise normalization within the purification stage enforces explicit reconstruction-driven cross-modal alignment, ensuring statistical compatibility between heterogeneous features prior to fusion. Extensive experiments on the TLRGBD benchmark, where 94.5% of instances are small objects, demonstrate that CMAFNet achieves 32.2% mAP@50 and 12.5% APs, outperforming the strongest baseline by 9.8 and 4.0 percentage points, respectively. A lightweight variant reaches 24.8% mAP50 at 228 FPS with only 4.9M parameters, surpassing all YOLO-based detectors while matching transformer-based methods at substantially lower computational cost.

Method: 1) Generate microstructural morphologies with perfect ground-truth masks using phase-field simulations. 2) Use CycleGAN for unpaired image-to-image translation to transform clean simulations into realistic SEM images. 3) Train a U-Net model exclusively on this synthetic data.

Result: U-Net trained on synthetic data achieved mean Boundary F1-Score of 0.90 and IoU of 0.88 on unseen experimental images. t-SNE and Shannon entropy analysis confirmed synthetic images are statistically indistinguishable from real data.

Conclusion: The framework successfully bridges simulation-to-reality gap, transforming data-scarce segmentation into data-abundant problem without manual annotation, enabling automated materials analysis.

Abstract: Semantic segmentation of microscopy images is a critical task for high-throughput materials characterisation, yet its automation is severely constrained by the prohibitive cost, subjectivity, and scarcity of expert-annotated data. While physics-based simulations offer a scalable alternative to manual labelling, models trained on such data historically fail to generalise due to a significant domain gap, lacking the complex textures, noise patterns, and imaging artefacts inherent to experimental data. This paper introduces a novel framework for labour-free segmentation that successfully bridges this simulation-to-reality gap. Our pipeline leverages phase-field simulations to generate an abundant source of microstructural morphologies with perfect, intrinsically-derived ground-truth masks. We then employ a Cycle-Consistent Generative Adversarial Network (CycleGAN) for unpaired image-to-image translation, transforming the clean simulations into a large-scale dataset of high-fidelity, realistic SEM images. A U-Net model, trained exclusively on this synthetic data, demonstrated remarkable generalisation when deployed on unseen experimental images, achieving a mean Boundary F1-Score of 0.90 and an Intersection over Union (IOU) of 0.88. Comprehensive validation using t-SNE feature-space projection and Shannon entropy analysis confirms that our synthetic images are statistically and featurally indistinguishable from the real data manifold. By completely decoupling model training from manual annotation, our generative framework transforms a data-scarce problem into one of data abundance, providing a robust and fully automated solution to accelerate materials discovery and analysis.

Method: OpticalDNA renders DNA into structured visual layouts and trains an OCR-capable vision-language model with a visual DNA encoder and document decoder. The encoder produces compact, reconstructible visual tokens for high-fidelity compression. It defines prompt-conditioned objectives over core genomic primitives: reading, region grounding, subsequence retrieval, and masked span completion.

Result: Across diverse genomic benchmarks, OpticalDNA consistently outperforms recent baselines. On sequences up to 450k bases, it achieves the best overall performance with nearly 20× fewer effective tokens, and surpasses models with up to 985× more activated parameters while tuning only 256k trainable parameters.

Conclusion: Vision-based approaches to genomic modeling through OCR-style document understanding offer significant advantages over traditional language model architectures, enabling more efficient and effective representation of DNA sequences with far fewer computational resources.

Abstract: Recent genomic foundation models largely adopt large language model architectures that treat DNA as a one-dimensional token sequence. However, exhaustive sequential reading is structurally misaligned with sparse and discontinuous genomic semantics, leading to wasted computation on low-information background and preventing understanding-driven compression for long contexts. Here, we present OpticalDNA, a vision-based framework that reframes genomic modeling as Optical Character Recognition (OCR)-style document understanding. OpticalDNA renders DNA into structured visual layouts and trains an OCR-capable vision–language model with a \emph{visual DNA encoder} and a \emph{document decoder}, where the encoder produces compact, reconstructible visual tokens for high-fidelity compression. Building on this representation, OpticalDNA defines prompt-conditioned objectives over core genomic primitives-reading, region grounding, subsequence retrieval, and masked span completion-thereby learning layout-aware DNA representations that retain fine-grained genomic information under a reduced effective token budget. Across diverse genomic benchmarks, OpticalDNA consistently outperforms recent baselines; on sequences up to 450k bases, it achieves the best overall performance with nearly $20\times$ fewer effective tokens, and surpasses models with up to $985\times$ more activated parameters while tuning only 256k \emph{trainable} parameters.

Method: Proposes FastPhysGS with two key components: (1) Instance-aware Particle Filling with Monte Carlo Importance Sampling to efficiently populate interior particles while preserving geometric fidelity, and (2) Bidirectional Graph Decoupling Optimization, an adaptive strategy that rapidly optimizes material parameters predicted from a Vision-Language Model.

Result: FastPhysGS achieves high-fidelity physical simulation in 1 minute using only 7 GB runtime memory, outperforming prior works with broad potential applications.

Conclusion: FastPhysGS provides a fast and robust framework for physics-based dynamic 3DGS simulation that addresses limitations of existing methods through efficient particle filling and adaptive optimization strategies.

Abstract: Extending 3D Gaussian Splatting (3DGS) to 4D physical simulation remains challenging. Based on the Material Point Method (MPM), existing methods either rely on manual parameter tuning or distill dynamics from video diffusion models, limiting the generalization and optimization efficiency. Recent attempts using LLMs/VLMs suffer from a text/image-to-3D perceptual gap, yielding unstable physics behavior. In addition, they often ignore the surface structure of 3DGS, leading to implausible motion. We propose FastPhysGS, a fast and robust framework for physics-based dynamic 3DGS simulation:(1) Instance-aware Particle Filling (IPF) with Monte Carlo Importance Sampling (MCIS) to efficiently populate interior particles while preserving geometric fidelity; (2) Bidirectional Graph Decoupling Optimization (BGDO), an adaptive strategy that rapidly optimizes material parameters predicted from a VLM. Experiments show FastPhysGS achieves high-fidelity physical simulation in 1 minute using only 7 GB runtime memory, outperforming prior works with broad potential applications.

Method: Proposes DenVisCoM, a novel Mamba block combined with Transformer-based attention in a hybrid architecture. This unified framework jointly estimates optical flow and disparity by efficiently addressing real-time inference, memory footprint, and accuracy simultaneously.

Result: Extensive experiments show the model can accurately estimate optical flow and disparity in real time. The architecture achieves a favorable trade-off between accuracy and processing speed across multiple datasets.

Conclusion: The proposed DenVisCoM hybrid architecture successfully enables joint, accurate, real-time estimation of optical flow and disparity, demonstrating the effectiveness of combining Mamba blocks with Transformer attention for multi-view geometry tasks.

Abstract: In this work, we propose a novel Mamba block DenVisCoM, as well as a novel hybrid architecture specifically tailored for accurate and real-time estimation of optical flow and disparity estimation. Given that such multi-view geometry and motion tasks are fundamentally related, we propose a unified architecture to tackle them jointly. Specifically, the proposed hybrid architecture is based on DenVisCoM and a Transformer-based attention block that efficiently addresses real-time inference, memory footprint, and accuracy at the same time for joint estimation of motion and 3D dense perception tasks. We extensively analyze the benchmark trade-off of accuracy and real-time processing on a large number of datasets. Our experimental results and related analysis suggest that our proposed model can accurately estimate optical flow and disparity estimation in real time. All models and associated code are available at https://github.com/vimstereo/DenVisCoM.

Method: Train a simple linear classifier on frozen features from modern Vision Foundation Models (Perception Encoder, MetaCLIP 2, DINOv3) without architectural modifications, evaluating across traditional benchmarks, unseen generators, and challenging in-the-wild distributions.

Result: The simple baseline matches specialized detectors on standard benchmarks and decisively outperforms them on in-the-wild datasets by over 30% accuracy, establishing new state-of-the-art. Vision-Language Models learn explicit semantic concepts of forgery while Self-Supervised Learning models acquire implicit forensic features.

Conclusion: Foundation models’ emergent capabilities from massive pretraining data containing synthetic content enable superior real-world detection. A paradigm shift is needed from overfitting on static benchmarks to leveraging foundation models’ evolving world knowledge for reliable AI forensics.

Abstract: While specialized detectors for AI-Generated Images (AIGI) achieve near-perfect accuracy on curated benchmarks, they suffer from a dramatic performance collapse in realistic, in-the-wild scenarios. In this work, we demonstrate that simplicity prevails over complex architectural designs. A simple linear classifier trained on the frozen features of modern Vision Foundation Models , including Perception Encoder, MetaCLIP 2, and DINOv3, establishes a new state-of-the-art. Through a comprehensive evaluation spanning traditional benchmarks, unseen generators, and challenging in-the-wild distributions, we show that this baseline not only matches specialized detectors on standard benchmarks but also decisively outperforms them on in-the-wild datasets, boosting accuracy by striking margins of over 30%. We posit that this superior capability is an emergent property driven by the massive scale of pre-training data containing synthetic content. We trace the source of this capability to two distinct manifestations of data exposure: Vision-Language Models internalize an explicit semantic concept of forgery, while Self-Supervised Learning models implicitly acquire discriminative forensic features from the pretraining data. However, we also reveal persistent limitations: these models suffer from performance degradation under recapture and transmission, remain blind to VAE reconstruction and localized editing. We conclude by advocating for a paradigm shift in AI forensics, moving from overfitting on static benchmarks to harnessing the evolving world knowledge of foundation models for real-world reliability.

Method: Three key contributions: (1) Progressive coarse-to-fine calibration construction for data-scarce 3D datasets, (2) Ternary-search-based solver for quantization interval search reducing complexity from O(N) to O(log N), (3) TRE-Guided Module-wise Compensation using Tail Relative Error metric to address quantization error accumulation from activation outliers.

Result: Extensive experiments on VGGT and Pi3 benchmarks show TAPTQ consistently outperforms state-of-the-art PTQ methods in accuracy while significantly reducing calibration time.

Conclusion: TAPTQ provides an effective quantization solution for 3D geometric learning models, enabling efficient deployment on resource-constrained platforms with improved accuracy and reduced calibration overhead.

Abstract: The burgeoning complexity and scale of 3D geometry models pose significant challenges for deployment on resource-constrained platforms. While Post-Training Quantization (PTQ) enables efficient inference without retraining, conventional methods, primarily optimized for 2D Vision Transformers, fail to transfer effectively to 3D models due to intricate feature distributions and prohibitive calibration overhead. To address these challenges, we propose TAPTQ, a Tail-Aware Post-Training Quantization pipeline specifically engineered for 3D geometric learning. Our contribution is threefold: (1) To overcome the data-scale bottleneck in 3D datasets, we develop a progressive coarse-to-fine calibration construction strategy that constructs a highly compact subset to achieve both statistical purity and geometric representativeness. (2) We reformulate the quantization interval search as an optimization problem and introduce a ternary-search-based solver, reducing the computational complexity from $\mathcal{O}(N)$ to $\mathcal{O}(\log N)$ for accelerated deployment. (3) To mitigate quantization error accumulation, we propose TRE-Guided Module-wise Compensation, which utilizes a Tail Relative Error (TRE) metric to adaptively identify and rectify distortions in modules sensitive to long-tailed activation outliers. Extensive experiments on the VGGT and Pi3 benchmarks demonstrate that TAPTQ consistently outperforms state-of-the-art PTQ methods in accuracy while significantly reducing calibration time. The code will be released soon.

Method: ObjEmbed decomposes input images into multiple regional embeddings (one per object) plus global embeddings. Each region gets two complementary embeddings: object embedding for semantic matching and IoU embedding for localization quality prediction. The final matching score combines semantic similarity with predicted IoU. All objects and the full image are encoded in a single forward pass.

Result: Superior performance on 18 diverse benchmarks demonstrates strong semantic discrimination capabilities for both region-level and image-level tasks.

Conclusion: ObjEmbed provides an effective solution for fine-grained vision-language alignment with object-oriented representations that capture both semantic and spatial aspects, enabling versatile and efficient visual understanding.

Abstract: Aligning objects with corresponding textual descriptions is a fundamental challenge and a realistic requirement in vision-language understanding. While recent multimodal embedding models excel at global image-text alignment, they often struggle with fine-grained alignment between image regions and specific phrases. In this work, we present ObjEmbed, a novel MLLM embedding model that decomposes the input image into multiple regional embeddings, each corresponding to an individual object, along with global embeddings. It supports a wide range of visual understanding tasks like visual grounding, local image retrieval, and global image retrieval. ObjEmbed enjoys three key properties: (1) Object-Oriented Representation: It captures both semantic and spatial aspects of objects by generating two complementary embeddings for each region: an object embedding for semantic matching and an IoU embedding that predicts localization quality. The final object matching score combines semantic similarity with the predicted IoU, enabling more accurate retrieval. (2) Versatility: It seamlessly handles both region-level and image-level tasks. (3) Efficient Encoding: All objects in an image, along with the full image, are encoded in a single forward pass for high efficiency. Superior performance on 18 diverse benchmarks demonstrates its strong semantic discrimination.

Method: Spot-wise monitoring strategy using distance-aware matching method with spatial tolerance, enhanced by Adaptive Bounding Box Partitioning for challenging spaces. Uses YOLOv11m model (40.5MB) on edge devices. Introduces Digital Shadow for visual representation and application support server on repurposed TV box.

Result: Achieves 98.80% balanced accuracy with 8-second inference time on resource-constrained edge device. System enables detailed spot occupancy statistics and scalable communication between cloud services, parking totem, and bot.

Conclusion: Successfully extends parking monitoring to spot-level with high accuracy on edge devices, while promoting hardware reuse and sustainability through repurposed TV box server. Digital Shadow serves as foundation for future Digital Twin development.

Abstract: Smart parking systems help reduce congestion and minimize users’ search time, thereby contributing to smart city adoption and enhancing urban mobility. In previous works, we presented a system developed on a university campus to monitor parking availability by estimating the number of free spaces from vehicle counts within a region of interest. Although this approach achieved good accuracy, it restricted the system’s ability to provide spot-level insights and support more advanced applications. To overcome this limitation, we extend the system with a spot-wise monitoring strategy based on a distance-aware matching method with spatial tolerance, enhanced through an Adaptive Bounding Box Partitioning method for challenging spaces. The proposed approach achieves a balanced accuracy of 98.80% while maintaining an inference time of 8 seconds on a resource-constrained edge device, enhancing the capabilities of YOLOv11m, a model that has a size of 40.5 MB. In addition, two new components were introduced: (i) a Digital Shadow that visually represents parking lot entities as a base to evolve to a full Digital Twin, and (ii) an application support server based on a repurposed TV box. The latter not only enables scalable communication among cloud services, the parking totem, and a bot that provides detailed spot occupancy statistics, but also promotes hardware reuse as a step towards greater sustainability.

Method: Introduces Mind-Brush, a unified agentic framework that simulates human-like ’think-research-create’ paradigm. It actively retrieves multimodal evidence to ground out-of-distribution concepts and employs reasoning tools to resolve implicit visual constraints.

Result: Mind-Brush significantly enhances unified models’ capabilities, achieving zero-to-one capability leap for Qwen-Image baseline on Mind-Bench benchmark, while achieving superior results on established benchmarks like WISE and RISE.

Conclusion: The framework bridges gaps in current generation models by making generation dynamic and knowledge-driven, enabling better comprehension of implicit intentions and adaptation to real-world dynamics.

Abstract: While text-to-image generation has achieved unprecedented fidelity, the vast majority of existing models function fundamentally as static text-to-pixel decoders. Consequently, they often fail to grasp implicit user intentions. Although emerging unified understanding-generation models have improved intent comprehension, they still struggle to accomplish tasks involving complex knowledge reasoning within a single model. Moreover, constrained by static internal priors, these models remain unable to adapt to the evolving dynamics of the real world. To bridge these gaps, we introduce Mind-Brush, a unified agentic framework that transforms generation into a dynamic, knowledge-driven workflow. Simulating a human-like ’think-research-create’ paradigm, Mind-Brush actively retrieves multimodal evidence to ground out-of-distribution concepts and employs reasoning tools to resolve implicit visual constraints. To rigorously evaluate these capabilities, we propose Mind-Bench, a comprehensive benchmark comprising 500 distinct samples spanning real-time news, emerging concepts, and domains such as mathematical and Geo-Reasoning. Extensive experiments demonstrate that Mind-Brush significantly enhances the capabilities of unified models, realizing a zero-to-one capability leap for the Qwen-Image baseline on Mind-Bench, while achieving superior results on established benchmarks like WISE and RISE.

Method: Constructed VDR-Bench with 2,000 VQA instances using a multi-stage curation pipeline and expert review. Also proposed a simple multi-round cropped-search workflow to enhance visual retrieval capabilities of MLLMs in realistic scenarios.

Result: The benchmark provides realistic evaluation conditions for vision-deepresearch systems, and the proposed multi-round cropped-search workflow effectively improves model performance in realistic visual retrieval scenarios.

Conclusion: The work addresses critical gaps in evaluating multimodal deep-research systems and provides practical guidance for future system design, with code to be released publicly.

Abstract: Multimodal Large Language Models (MLLMs) have advanced VQA and now support Vision-DeepResearch systems that use search engines for complex visual-textual fact-finding. However, evaluating these visual and textual search abilities is still difficult, and existing benchmarks have two major limitations. First, existing benchmarks are not visual search-centric: answers that should require visual search are often leaked through cross-textual cues in the text questions or can be inferred from the prior world knowledge in current MLLMs. Second, overly idealized evaluation scenario: On the image-search side, the required information can often be obtained via near-exact matching against the full image, while the text-search side is overly direct and insufficiently challenging. To address these issues, we construct the Vision-DeepResearch benchmark (VDR-Bench) comprising 2,000 VQA instances. All questions are created via a careful, multi-stage curation pipeline and rigorous expert review, designed to assess the behavior of Vision-DeepResearch systems under realistic real-world conditions. Moreover, to address the insufficient visual retrieval capabilities of current MLLMs, we propose a simple multi-round cropped-search workflow. This strategy is shown to effectively improve model performance in realistic visual retrieval scenarios. Overall, our results provide practical guidance for the design of future multimodal deep-research systems. The code will be released in https://github.com/Osilly/Vision-DeepResearch.

Method: Proposes MagicFuse with three diffusion-based branches: 1) intra-spectral knowledge reinforcement branch to mine obscured scene information, 2) cross-spectral knowledge generation branch to learn thermal radiation patterns, and 3) multi-domain knowledge fusion branch that integrates probabilistic noise from both branches to obtain cross-spectral representations through successive sampling, with visual and semantic constraints.

Result: Extensive experiments show MagicFuse achieves visual and semantic representation performance comparable to or better than state-of-the-art fusion methods with multi-modal inputs, despite using only single degraded visible images.

Conclusion: MagicFuse successfully extends conventional data-level fusion to knowledge-level fusion, enabling comprehensive cross-spectral scene understanding from single visible images, which is valuable for practical applications where multi-modal sensors are unavailable.

Abstract: This paper focuses on a highly practical scenario: how to continue benefiting from the advantages of multi-modal image fusion under harsh conditions when only visible imaging sensors are available. To achieve this goal, we propose a novel concept of single-image fusion, which extends conventional data-level fusion to the knowledge level. Specifically, we develop MagicFuse, a novel single image fusion framework capable of deriving a comprehensive cross-spectral scene representation from a single low-quality visible image. MagicFuse first introduces an intra-spectral knowledge reinforcement branch and a cross-spectral knowledge generation branch based on the diffusion models. They mine scene information obscured in the visible spectrum and learn thermal radiation distribution patterns transferred to the infrared spectrum, respectively. Building on them, we design a multi-domain knowledge fusion branch that integrates the probabilistic noise from the diffusion streams of these two branches, from which a cross-spectral scene representation can be obtained through successive sampling. Then, we impose both visual and semantic constraints to ensure that this scene representation can satisfy human observation while supporting downstream semantic decision-making. Extensive experiments show that our MagicFuse achieves visual and semantic representation performance comparable to or even better than state-of-the-art fusion methods with multi-modal inputs, despite relying solely on a single degraded visible image.

Method: Uses MEMS-LiDAR for anonymized 3D point clouds, augments real LiDAR data with synthetic scenes from CARLA simulation framework to create hybrid training dataset.

Result: Hybrid data improves average precision by 44 percentage points compared to real-data-only model while reducing manual annotation effort by 50%.

Conclusion: Proposed approach provides scalable, cost-efficient alternative that combines high performance in person detection with GDPR compliance in industrial environments.

Abstract: The reliable detection of unauthorized individuals in safety-critical industrial indoor spaces is crucial to avoid plant shutdowns, property damage, and personal hazards. Conventional vision-based methods that use deep-learning approaches for person recognition provide image information but are sensitive to lighting and visibility conditions and often violate privacy regulations, such as the General Data Protection Regulation (GDPR) in the European Union. Typically, detection systems based on deep learning require annotated data for training. Collecting and annotating such data, however, is highly time-consuming and due to manual treatments not necessarily error free. Therefore, this paper presents a privacy-compliant approach based on Micro-Electro-Mechanical Systems LiDAR (MEMS-LiDAR), which exclusively captures anonymized 3D point clouds and avoids personal identification features. To compensate for the large amount of time required to record real LiDAR data and for post-processing and annotation, real recordings are augmented with synthetically generated scenes from the CARLA simulation framework. The results demonstrate that the hybrid data improves the average precision by 44 percentage points compared to a model trained exclusively with real data while reducing the manual annotation effort by 50 %. Thus, the proposed approach provides a scalable, cost-efficient alternative to purely real-data-based methods and systematically shows how synthetic LiDAR data can combine high performance in person detection with GDPR compliance in an industrial environment.

Method: Proposes Disentangled Dynamics Prediction (DDP) that decomposes latent state evolution into sparse primary dynamics (physical interactions) and secondary context-driven background updates. Uses efficient historical processing with dynamic localization to isolate primary dynamics and cross-attention for background updates.

Result: Achieves approximately 9x inference speedup on Push-T task and improves MPC success rate from 90% to 98% compared to state-of-the-art dense models. Shows strong performance across navigation, tabletop manipulation, and complex deformable/multi-body interactions.

Conclusion: DDP-WM establishes a promising path for developing efficient, high-fidelity world models for robotics by optimizing resource allocation and providing a smooth optimization landscape for planners.

Abstract: World models are essential for autonomous robotic planning. However, the substantial computational overhead of existing dense Transformerbased models significantly hinders real-time deployment. To address this efficiency-performance bottleneck, we introduce DDP-WM, a novel world model centered on the principle of Disentangled Dynamics Prediction (DDP). We hypothesize that latent state evolution in observed scenes is heterogeneous and can be decomposed into sparse primary dynamics driven by physical interactions and secondary context-driven background updates. DDP-WM realizes this decomposition through an architecture that integrates efficient historical processing with dynamic localization to isolate primary dynamics. By employing a crossattention mechanism for background updates, the framework optimizes resource allocation and provides a smooth optimization landscape for planners. Extensive experiments demonstrate that DDP-WM achieves significant efficiency and performance across diverse tasks, including navigation, precise tabletop manipulation, and complex deformable or multi-body interactions. Specifically, on the challenging Push-T task, DDP-WM achieves an approximately 9 times inference speedup and improves the MPC success rate from 90% to98% compared to state-of-the-art dense models. The results establish a promising path for developing efficient, high-fidelity world models. Codes will be available at https://github.com/HCPLabSYSU/DDP-WM.

Method: Proposes a three-step approach: 1) Single-shot filtering using signal processing to isolate planar regions while suppressing noise and high-curvature artifacts, 2) Cyclic orientation transformation scheme to accurately represent dip angle and dip direction in Cartesian space, 3) Hierarchical clustering technique to characterize transformed orientations into sets without requiring user-defined cluster numbers.

Result: The method outperforms existing automated structure mapping techniques with lowest mean absolute errors: 1.95° in nominal dip angle and 2.20° in dip direction, with dispersion errors below 3°. Validated on real-world mine stope data against ground truth from manually handpicked discontinuity planes.

Conclusion: The proposed approach provides an accurate and robust solution for automatic discontinuity set characterization in underground mine cavities, addressing limitations of existing methods and demonstrating superior performance in real-world scenarios.

Abstract: Characterisation of structural discontinuity sets in exposed rock faces of underground mine cavities is essential for assessing rock-mass stability, excavation safety, and operational efficiency. UAV and other mobile laser-scanning techniques provide efficient means of collecting point clouds from rock faces. However, the development of a robust and efficient approach for automatic characterisation of discontinuity sets in real-world scenarios, like fully enclosed rock faces in cavities, remains an open research problem. In this study, a new approach is proposed for automatic discontinuity set characterisation that uses a single-shot filtering strategy, an innovative cyclic orientation transformation scheme and a hierarchical clustering technique. The single-shot filtering step isolates planar regions while robustly suppressing noise and high-curvature artefacts in one pass using a signal-processing technique. To address the limitations of Cartesian clustering on polar orientation data, a cyclic orientation transformation scheme is developed, enabling accurate representation of dip angle and dip direction in Cartesian space. The transformed orientations are then characterised into sets using a hierarchical clustering technique, which handles varying density distributions and identifies clusters without requiring user-defined set numbers. The accuracy of the method is validated on real-world mine stope and against ground truth obtained using manually handpicked discontinuity planes identified with the Virtual Compass tool, as well as widely used automated structure mapping techniques. The proposed approach outperforms the other techniques by exhibiting the lowest mean absolute error in estimating discontinuity set orientations in real-world stope data with errors of 1.95° and 2.20° in nominal dip angle and dip direction, respectively, and dispersion errors lying below 3°.

Method: Extended transformer framework processes multi-scale spatial patches alongside temporal sequences (up to 20 years) through unified architecture, using self-supervised learning with spatial masking, temporal masking, and horizon sampling strategies from 40 years of unlabeled Landsat imagery.

Result: Achieves MAE of 0.0328 and R^2 of 0.8412 for next-year predictions on Landsat data (1984-2024), outperforming traditional statistical methods, CNN-based approaches, LSTM networks, and standard transformers.

Conclusion: STT-LTF effectively handles irregular temporal sampling and variable prediction horizons, making it suitable for heterogeneous landscapes experiencing rapid ecological transitions, with direct prediction of arbitrary future time points without error accumulation.

Abstract: Long-term satellite image time series (SITS) analysis in heterogeneous landscapes faces significant challenges, particularly in Mediterranean regions where complex spatial patterns, seasonal variations, and multi-decade environmental changes interact across different scales. This paper presents the Spatio-Temporal Transformer for Long Term Forecasting (STT-LTF ), an extended framework that advances beyond purely temporal analysis to integrate spatial context modeling with temporal sequence prediction. STT-LTF processes multi-scale spatial patches alongside temporal sequences (up to 20 years) through a unified transformer architecture, capturing both local neighborhood relationships and regional climate influences. The framework employs comprehensive self-supervised learning with spatial masking, temporal masking, and horizon sampling strategies, enabling robust model training from 40 years of unlabeled Landsat imagery. Unlike autoregressive approaches, STT-LTF directly predicts arbitrary future time points without error accumulation, incorporating spatial patch embeddings, cyclical temporal encoding, and geographic coordinates to learn complex dependencies across heterogeneous Mediterranean ecosystems. Experimental evaluation on Landsat data (1984-2024) demonstrates that STT-LTF achieves a Mean Absolute Error (MAE) of 0.0328 and R^2 of 0.8412 for next-year predictions, outperforming traditional statistical methods, CNN-based approaches, LSTM networks, and standard transformers. The framework’s ability to handle irregular temporal sampling and variable prediction horizons makes it particularly suitable for analysis of heterogeneous landscapes experiencing rapid ecological transitions.

Method: Three training-free modules: TempCache compresses KV cache via temporal correspondence; AnnCA accelerates cross-attention by selecting frame-relevant prompt tokens using ANN matching; AnnSA sparsifies self-attention by restricting queries to semantically matched keys using lightweight ANN.

Result: Achieves 5-10x end-to-end speedups while preserving visual quality, maintaining stable throughput and nearly constant peak GPU memory usage over long rollouts, unlike prior methods that progressively slow down.

Conclusion: The proposed attention optimization framework effectively addresses KV cache growth in autoregressive video diffusion, enabling efficient long-form video synthesis with consistent quality and stable resource usage.

Abstract: Autoregressive video diffusion models enable streaming generation, opening the door to long-form synthesis, video world models, and interactive neural game engines. However, their core attention layers become a major bottleneck at inference time: as generation progresses, the KV cache grows, causing both increasing latency and escalating GPU memory, which in turn restricts usable temporal context and harms long-range consistency. In this work, we study redundancy in autoregressive video diffusion and identify three persistent sources: near-duplicate cached keys across frames, slowly evolving (largely semantic) queries/keys that make many attention computations redundant, and cross-attention over long prompts where only a small subset of tokens matters per frame. Building on these observations, we propose a unified, training-free attention framework for autoregressive diffusion: TempCache compresses the KV cache via temporal correspondence to bound cache growth; AnnCA accelerates cross-attention by selecting frame-relevant prompt tokens using fast approximate nearest neighbor (ANN) matching; and AnnSA sparsifies self-attention by restricting each query to semantically matched keys, also using a lightweight ANN. Together, these modules reduce attention, compute, and memory and are compatible with existing autoregressive diffusion backbones and world models. Experiments demonstrate up to x5–x10 end-to-end speedups while preserving near-identical visual quality and, crucially, maintaining stable throughput and nearly constant peak GPU memory usage over long rollouts, where prior methods progressively slow down and suffer from increasing memory usage.

Method: FlowBypass uses Rectified Flow to construct a bypass directly connecting inversion and reconstruction trajectories, mitigating error accumulation without feature manipulations. The authors provide formal derivation of two trajectories, obtain approximate bypass formulation and numerical solution for seamless trajectory transitions.

Result: Extensive experiments show FlowBypass consistently outperforms state-of-the-art image editing methods, achieving stronger prompt alignment while preserving high-fidelity details in irrelevant regions.

Conclusion: FlowBypass offers a novel analytical framework for training-free image editing that addresses the fundamental trade-off in inversion-reconstruction approaches, providing better prompt alignment and fidelity preservation through bypass trajectory construction.

Abstract: Training-free image editing has attracted increasing attention for its efficiency and independence from training data. However, existing approaches predominantly rely on inversion-reconstruction trajectories, which impose an inherent trade-off: longer trajectories accumulate errors and compromise fidelity, while shorter ones fail to ensure sufficient alignment with the edit prompt. Previous attempts to address this issue typically employ backbone-specific feature manipulations, limiting general applicability. To address these challenges, we propose FlowBypass, a novel and analytical framework grounded in Rectified Flow that constructs a bypass directly connecting inversion and reconstruction trajectories, thereby mitigating error accumulation without relying on feature manipulations. We provide a formal derivation of two trajectories, from which we obtain an approximate bypass formulation and its numerical solution, enabling seamless trajectory transitions. Extensive experiments demonstrate that FlowBypass consistently outperforms state-of-the-art image editing methods, achieving stronger prompt alignment while preserving high-fidelity details in irrelevant regions.

Method: Proposes LDRNet with coarse-to-fine refinement: first predicts coarse resolution registration field, then refines it using two novel components: 1) refine block for multi-resolution field refinement, 2) rigid block to learn transformation matrix from high-level features.

Result: Achieves state-of-the-art performance for large deformation image registration on private and public SegTHOR datasets, outperforming traditional methods and deep learning models (VoxelMorph, RCN, LapIRN) while being much faster.

Conclusion: LDRNet effectively addresses the unique challenges of chest CT registration through its coarse-to-fine architecture with specialized refinement components, offering superior performance and speed for large deformation medical image registration.

Abstract: Most of the deep learning based medical image registration algorithms focus on brain image registration tasks.Compared with brain registration, the chest CT registration has larger deformation, more complex background and region over-lap. In this paper, we propose a fast unsupervised deep learning method, LDRNet, for large deformation image registration of chest CT images. We first predict a coarse resolution registration field, then refine it from coarse to fine. We propose two innovative technical components: 1) a refine block that is used to refine the registration field in different resolutions, 2) a rigid block that is used to learn transformation matrix from high-level features. We train and evaluate our model on the private dataset and public dataset SegTHOR. We compare our performance with state-of-the-art traditional registration methods as well as deep learning registration models VoxelMorph, RCN, and LapIRN. The results demonstrate that our model achieves state-of-the-art performance for large deformation images registration and is much faster.

Method: Guided Progressive Distillation (GPD) uses a teacher model to progressively guide a student model to operate with larger step sizes. It features: 1) online-generated training targets to reduce optimization difficulty, and 2) frequency-domain constraints in latent space to preserve fine details and temporal dynamics.

Result: Applied to Wan2.1 model, GPD reduces sampling steps from 48 to 6 while maintaining competitive visual quality on VBench. Outperforms existing distillation methods in both pipeline simplicity and quality preservation.

Conclusion: GPD provides an effective framework for accelerating video diffusion models with minimal quality degradation, addressing the computational bottleneck in video generation.

Abstract: Diffusion models have achieved remarkable success in video generation; however, the high computational cost of the denoising process remains a major bottleneck. Existing approaches have shown promise in reducing the number of diffusion steps, but they often suffer from significant quality degradation when applied to video generation. We propose Guided Progressive Distillation (GPD), a framework that accelerates the diffusion process for fast and high-quality video generation. GPD introduces a novel training strategy in which a teacher model progressively guides a student model to operate with larger step sizes. The framework consists of two key components: (1) an online-generated training target that reduces optimization difficulty while improving computational efficiency, and (2) frequency-domain constraints in the latent space that promote the preservation of fine-grained details and temporal dynamics. Applied to the Wan2.1 model, GPD reduces the number of sampling steps from 48 to 6 while maintaining competitive visual quality on VBench. Compared with existing distillation methods, GPD demonstrates clear advantages in both pipeline simplicity and quality preservation.

Method: The authors introduce VIA-Bench with six core categories: color illusions, motion illusions, gestalt illusions, geometric/spatial illusions, general visual illusions, and visual anomalies. They construct over 1K high-quality question-answer pairs through careful human-in-the-loop review and evaluate over 20 state-of-the-art MLLMs.

Result: Extensive evaluation reveals significant vulnerabilities in MLLMs. Chain-of-Thought reasoning offers negligible robustness, often yielding “brittle mirages” where model logic collapses under illusory stimuli. The findings show a fundamental divergence between machine and human perception.

Conclusion: Resolving perceptual bottlenecks related to visual illusions and anomalies is critical for advancing artificial general intelligence. The benchmark exposes important limitations in current MLLMs’ visual reasoning capabilities.

Abstract: Multimodal Large Language Models (MLLMs) have shown remarkable proficiency on general-purpose vision-language benchmarks, reaching or even exceeding human-level performance. However, these evaluations typically rely on standard in-distribution data, leaving the robustness of MLLMs largely unexamined when faced with scenarios that defy common-sense priors. To address this gap, we introduce VIA-Bench, a challenging benchmark designed to probe model performance on visual illusions and anomalies. It includes six core categories: color illusions, motion illusions, gestalt illusions, geometric and spatial illusions, general visual illusions, and visual anomalies. Through careful human-in-the-loop review, we construct over 1K high-quality question-answer pairs that require nuanced visual reasoning. Extensive evaluation of over 20 state-of-the-art MLLMs, including proprietary, open-source, and reasoning-enhanced models, uncovers significant vulnerabilities. Notably, we find that Chain-of-Thought (CoT) reasoning offers negligible robustness, often yielding ``brittle mirages’’ where the model’s logic collapses under illusory stimuli. Our findings reveal a fundamental divergence between machine and human perception, suggesting that resolving such perceptual bottlenecks is critical for the advancement of artificial general intelligence. The benchmark data and code will be released.

Method: Proposes a street-view-guided data acquisition strategy using publicly available imagery to identify places of interest (POI). Introduces two POI scoring methods: 1) KNN-based feature distance approach using a vision foundation model, and 2) visual-attribution approach using a vision-language model. Uses collect-detect protocol and constructs co-located dataset pairing Zenseact Open Dataset with Mapillary street-view images.

Result: Experiments on traffic sign detection (sensitive to cross-country variations) show approach achieves performance comparable to random sampling while using only half of the target-domain data. Cost estimations demonstrate large-scale street-view processing remains economically feasible for full-country analysis.

Conclusion: Street-view-guided data acquisition offers efficient and cost-effective cross-country model adaptation for ADAS/ADS systems, reducing data collection requirements while maintaining performance through intelligent POI identification using public imagery.

Abstract: Deploying ADAS and ADS across countries remains challenging due to differences in legislation, traffic infrastructure, and visual conventions, which introduce domain shifts that degrade perception performance. Traditional cross-country data collection relies on extensive on-road driving, making it costly and inefficient to identify representative locations. To address this, we propose a street-view-guided data acquisition strategy that leverages publicly available imagery to identify places of interest (POI). Two POI scoring methods are introduced: a KNN-based feature distance approach using a vision foundation model, and a visual-attribution approach using a vision-language model. To enable repeatable evaluation, we adopt a collect-detect protocol and construct a co-located dataset by pairing the Zenseact Open Dataset with Mapillary street-view images. Experiments on traffic sign detection, a task particularly sensitive to cross-country variations in sign appearance, show that our approach achieves performance comparable to random sampling while using only half of the target-domain data. We further provide cost estimations for full-country analysis, demonstrating that large-scale street-view processing remains economically feasible. These results highlight the potential of street-view-guided data acquisition for efficient and cost-effective cross-country model adaptation.

Method: SPIRIT adapts VFMs via lightweight physics-informed plug-ins: 1) PIFR spatially refines features using rank-sparsity decomposition to suppress background and enhance target signals, 2) PGMA temporally injects history-derived soft spatial priors into memory cross-attention to constrain cross-frame association.

Result: Experiments on multiple IRSTD benchmarks show consistent gains over VFM-based baselines and state-of-the-art performance in both single-frame and video-mode detection.

Conclusion: SPIRIT successfully bridges the modality gap between infrared and visible-spectrum data by incorporating physics-informed priors into vision foundation models, enabling robust infrared small target detection while maintaining compatibility with existing VFM architectures.

Abstract: Infrared small target detection (IRSTD) is crucial for surveillance and early-warning, with deployments spanning both single-frame analysis and video-mode tracking. A practical solution should leverage vision foundation models (VFMs) to mitigate infrared data scarcity, while adopting a memory-attention-based temporal propagation framework that unifies single- and multi-frame inference. However, infrared small targets exhibit weak radiometric signals and limited semantic cues, which differ markedly from visible-spectrum imagery. This modality gap makes direct use of semantics-oriented VFMs and appearance-driven cross-frame association unreliable for IRSTD: hierarchical feature aggregation can submerge localized target peaks, and appearance-only memory attention becomes ambiguous, leading to spurious clutter associations. To address these challenges, we propose SPIRIT, a unified and VFM-compatible framework that adapts VFMs to IRSTD via lightweight physics-informed plug-ins. Spatially, PIFR refines features by approximating rank-sparsity decomposition to suppress structured background components and enhance sparse target-like signals. Temporally, PGMA injects history-derived soft spatial priors into memory cross-attention to constrain cross-frame association, enabling robust video detection while naturally reverting to single-frame inference when temporal context is absent. Experiments on multiple IRSTD benchmarks show consistent gains over VFM-based baselines and SOTA performance.

Method: Cloth Dynamics Splatting (CloDS) uses a three-stage pipeline: 1) video-to-geometry grounding, 2) training dynamics model on grounded meshes. For handling large deformations and self-occlusions, introduces dual-position opacity modulation for bidirectional 2D-3D mapping via mesh-based Gaussian splatting.

Result: Comprehensive experiments show CloDS effectively learns cloth dynamics from visual data while maintaining strong generalization capabilities for unseen configurations.

Conclusion: CloDS enables unsupervised learning of cloth dynamics from visual observations, addressing limitations of existing methods that require physical supervision.

Abstract: Deep learning has demonstrated remarkable capabilities in simulating complex dynamic systems. However, existing methods require known physical properties as supervision or inputs, limiting their applicability under unknown conditions. To explore this challenge, we introduce Cloth Dynamics Grounding (CDG), a novel scenario for unsupervised learning of cloth dynamics from multi-view visual observations. We further propose Cloth Dynamics Splatting (CloDS), an unsupervised dynamic learning framework designed for CDG. CloDS adopts a three-stage pipeline that first performs video-to-geometry grounding and then trains a dynamics model on the grounded meshes. To cope with large non-linear deformations and severe self-occlusions during grounding, we introduce a dual-position opacity modulation that supports bidirectional mapping between 2D observations and 3D geometry via mesh-based Gaussian splatting in video-to-geometry grounding stage. It jointly considers the absolute and relative position of Gaussian components. Comprehensive experimental evaluations demonstrate that CloDS effectively learns cloth dynamics from visual data while maintaining strong generalization capabilities for unseen configurations. Our code is available at https://github.com/whynot-zyl/CloDS. Visualization results are available at https://github.com/whynot-zyl/CloDS_video}.%\footnote{As in this example.

Method: Systematic benchmark of 7 weakly supervised methods from audio/image/video domains on 7 public IMU datasets, with over 3,540 training runs and 7,080 evaluations. Focuses on transferability, effectiveness, and insights through three research questions.

Result: Temporal-domain methods transfer better than image-derived approaches; weak supervision can be competitive on datasets with longer actions and higher-dimensional sensing; failure modes include short actions, temporal ambiguity, and poor proposal quality.

Conclusion: Establishes WS-IMUBench as a reproducible benchmarking framework and outlines directions for advancing WS-IMU-TAL, including IMU-specific proposal generation and boundary-aware objectives.

Abstract: IMU-based Human Activity Recognition (HAR) has enabled a wide range of ubiquitous computing applications, yet its dominant clip classification paradigm cannot capture the rich temporal structure of real-world behaviors. This motivates a shift toward IMU Temporal Action Localization (IMU-TAL), which predicts both action categories and their start/end times in continuous streams. However, current progress is strongly bottlenecked by the need for dense, frame-level boundary annotations, which are costly and difficult to scale. To address this bottleneck, we introduce WS-IMUBench, a systematic benchmark study of weakly supervised IMU-TAL (WS-IMU-TAL) under only sequence-level labels. Rather than proposing a new localization algorithm, we evaluate how well established weakly supervised localization paradigms from audio, image, and video transfer to IMU-TAL under only sequence-level labels. We benchmark seven representative weakly supervised methods on seven public IMU datasets, resulting in over 3,540 model training runs and 7,080 inference evaluations. Guided by three research questions on transferability, effectiveness, and insights, our findings show that (i) transfer is modality-dependent, with temporal-domain methods generally more stable than image-derived proposal-based approaches; (ii) weak supervision can be competitive on favorable datasets (e.g., with longer actions and higher-dimensional sensing); and (iii) dominant failure modes arise from short actions, temporal ambiguity, and proposal quality. Finally, we outline concrete directions for advancing WS-IMU-TAL (e.g., IMU-specific proposal generation, boundary-aware objectives, and stronger temporal reasoning). Beyond individual results, WS-IMUBench establishes a reproducible benchmarking template, datasets, protocols, and analyses, to accelerate community-wide progress toward scalable WS-IMU-TAL.

Method: Introduces VIBE benchmark with three-level interaction hierarchy: deictic grounding, morphological manipulation, and causal reasoning. Curates diverse test cases with increasing complexity. Proposes LMM-as-a-judge evaluation framework with task-specific metrics for scalable assessment.

Result: Evaluated 17 open-source and proprietary image editing models. Found proprietary models show early-stage visual instruction-following capabilities and consistently outperform open-source models. Performance degrades significantly with increasing task difficulty even for strongest systems.

Conclusion: Visual instruction following remains challenging, especially for complex tasks. Proprietary models lead but have limitations. The benchmark highlights promising research directions for improving multimodal instruction understanding in image editing.

Abstract: Recent generative models have achieved remarkable progress in image editing. However, existing systems and benchmarks remain largely text-guided. In contrast, human communication is inherently multimodal, where visual instructions such as sketches efficiently convey spatial and structural intent. To address this gap, we introduce VIBE, the Visual Instruction Benchmark for Image Editing with a three-level interaction hierarchy that captures deictic grounding, morphological manipulation, and causal reasoning. Across these levels, we curate high-quality and diverse test cases that reflect progressively increasing complexity in visual instruction following. We further propose a robust LMM-as-a-judge evaluation framework with task-specific metrics to enable scalable and fine-grained assessment. Through a comprehensive evaluation of 17 representative open-source and proprietary image editing models, we find that proprietary models exhibit early-stage visual instruction-following capabilities and consistently outperform open-source models. However, performance degrades markedly with increasing task difficulty even for the strongest systems, highlighting promising directions for future research.

Method: Systematic analysis using two benchmarks (MMFakeBench and DGM4) evaluating: 1) image-only deepfake detectors, 2) evidence-driven fact-checking system with tool-guided retrieval (MCTS) and deliberative inference (Multi-Agent Debate), and 3) hybrid systems combining detectors with evidence-based approaches.

Result: Deepfake detectors showed limited standalone performance (F1: 0.26-0.53 on MMFakeBench, 0.33-0.49 on DGM4). When integrated into fact-checking pipelines, they reduced performance by 0.04-0.08 F1. Evidence-centric system achieved best results (F1: ~0.81 on MMFakeBench, 0.55 on DGM4).

Conclusion: Multimodal claim verification is primarily driven by semantic understanding and external evidence, not pixel-level artifact signals. Deepfake detectors introduce misleading authenticity priors that undermine evidence-based reasoning for real-world image-text misinformation.

Abstract: In multimodal misinformation, deception usually arises not just from pixel-level manipulations in an image, but from the semantic and contextual claim jointly expressed by the image-text pair. Yet most deepfake detectors, engineered to detect pixel-level forgeries, do not account for claim-level meaning, despite their growing integration in automated fact-checking (AFC) pipelines. This raises a central scientific and practical question: Do pixel-level detectors contribute useful signal for verifying image-text claims, or do they instead introduce misleading authenticity priors that undermine evidence-based reasoning? We provide the first systematic analysis of deepfake detectors in the context of multimodal misinformation detection. Using two complementary benchmarks, MMFakeBench and DGM4, we evaluate: (1) state-of-the-art image-only deepfake detectors, (2) an evidence-driven fact-checking system that performs tool-guided retrieval via Monte Carlo Tree Search (MCTS) and engages in deliberative inference through Multi-Agent Debate (MAD), and (3) a hybrid fact-checking system that injects detector outputs as auxiliary evidence. Results across both benchmark datasets show that deepfake detectors offer limited standalone value, achieving F1 scores in the range of 0.26-0.53 on MMFakeBench and 0.33-0.49 on DGM4, and that incorporating their predictions into fact-checking pipelines consistently reduces performance by 0.04-0.08 F1 due to non-causal authenticity assumptions. In contrast, the evidence-centric fact-checking system achieves the highest performance, reaching F1 scores of approximately 0.81 on MMFakeBench and 0.55 on DGM4. Overall, our findings demonstrate that multimodal claim verification is driven primarily by semantic understanding and external evidence, and that pixel-level artifact signals do not reliably enhance reasoning over real-world image-text misinformation.

Method: Proposes Ada-RefSR with Adaptive Implicit Correlation Gating (AICG) that uses learnable summary tokens to distill dominant reference patterns and capture implicit correlations with LQ features. Integrated into attention backbone for lightweight, adaptive regulation of reference guidance.

Result: Experiments on multiple datasets show Ada-RefSR achieves strong balance of fidelity, naturalness, and efficiency while remaining robust under varying reference alignment.

Conclusion: The “Trust but Verify” principle with adaptive gating effectively addresses reference reliability issues in diffusion-based image restoration, providing robust performance against misleading references.

Abstract: Recent works have explored reference-based super-resolution (RefSR) to mitigate hallucinations in diffusion-based image restoration. A key challenge is that real-world degradations make correspondences between low-quality (LQ) inputs and reference (Ref) images unreliable, requiring adaptive control of reference usage. Existing methods either ignore LQ-Ref correlations or rely on brittle explicit matching, leading to over-reliance on misleading references or under-utilization of valuable cues. To address this, we propose Ada-RefSR, a single-step diffusion framework guided by a “Trust but Verify” principle: reference information is leveraged when reliable and suppressed otherwise. Its core component, Adaptive Implicit Correlation Gating (AICG), employs learnable summary tokens to distill dominant reference patterns and capture implicit correlations with LQ features. Integrated into the attention backbone, AICG provides lightweight, adaptive regulation of reference guidance, serving as a built-in safeguard against erroneous fusion. Experiments on multiple datasets demonstrate that Ada-RefSR achieves a strong balance of fidelity, naturalness, and efficiency, while remaining robust under varying reference alignment.

Method: Hierarchical proxy-based parametric representation that: 1) performs semantic-aware decomposition of input images, 2) constructs hierarchical proxy geometries via adaptive Bezier fitting and iterative internal region subdivision/meshing, 3) embeds multi-scale implicit texture parameters into geometry-aware distributed proxy nodes, and 4) uses locality-adaptive feature indexing for spatial texture coherence.

Result: Achieves state-of-the-art rendering fidelity with significantly fewer parameters on ImageNet, OIR-Bench, and HumanEdit benchmarks; enables intuitive, interactive, physically plausible manipulation; supports real-time physics-driven animation with Position-Based Dynamics and lightweight implicit rendering.

Conclusion: Proposed representation enables efficient, controllable image/video editing with semantic disentanglement, high reconstruction fidelity, and physics-driven animation capabilities superior to generative approaches.

Abstract: Prevailing image representation methods, including explicit representations such as raster images and Gaussian primitives, as well as implicit representations such as latent images, either suffer from representation redundancy that leads to heavy manual editing effort, or lack a direct mapping from latent variables to semantic instances or parts, making fine-grained manipulation difficult. These limitations hinder efficient and controllable image and video editing. To address these issues, we propose a hierarchical proxy-based parametric image representation that disentangles semantic, geometric, and textural attributes into independent and manipulable parameter spaces. Based on a semantic-aware decomposition of the input image, our representation constructs hierarchical proxy geometries through adaptive Bezier fitting and iterative internal region subdivision and meshing. Multi-scale implicit texture parameters are embedded into the resulting geometry-aware distributed proxy nodes, enabling continuous high-fidelity reconstruction in the pixel domain and instance- or part-independent semantic editing. In addition, we introduce a locality-adaptive feature indexing mechanism to ensure spatial texture coherence, which further supports high-quality background completion without relying on generative models. Extensive experiments on image reconstruction and editing benchmarks, including ImageNet, OIR-Bench, and HumanEdit, demonstrate that our method achieves state-of-the-art rendering fidelity with significantly fewer parameters, while enabling intuitive, interactive, and physically plausible manipulation. Moreover, by integrating proxy nodes with Position-Based Dynamics, our framework supports real-time physics-driven animation using lightweight implicit rendering, achieving superior temporal consistency and visual realism compared with generative approaches.

Method: Proposes Lazy Attention mechanism that enables cross-layer sharing of similar attention patterns. Develops a novel layer-shared Q Cache for MLLMs that facilitates query reuse across adjacent layers. The method is lightweight, compatible with existing inference frameworks (Flash Attention, KV cache), and orthogonal to token-wise techniques.

Result: Reduces KV cache usage by over 35%, achieves 1.5x throughput improvement, with only ~1% performance loss on various MLLMs. Outperforms state-of-the-art token-wise methods in accuracy preservation across multiple benchmarks.

Conclusion: Lazy Attention provides an efficient attention mechanism for MLLMs that addresses computational bottlenecks while maintaining performance, offering a flexible solution that can be combined with existing token pruning approaches.

Abstract: Multimodal large language models (MLLMs) are plagued by exorbitant inference costs attributable to the profusion of visual tokens within the vision encoder. The redundant visual tokens engenders a substantial computational load and key-value (KV) cache footprint bottleneck. Existing approaches focus on token-wise optimization, leveraging diverse intricate token pruning techniques to eliminate non-crucial visual tokens. Nevertheless, these methods often unavoidably undermine the integrity of the KV cache, resulting in failures in long-text generation tasks. To this end, we conduct an in-depth investigation towards the attention mechanism of the model from a new perspective, and discern that attention within more than half of all decode layers are semantic similar. Upon this finding, we contend that the attention in certain layers can be streamlined by inheriting the attention from their preceding layers. Consequently, we propose Lazy Attention, an efficient attention mechanism that enables cross-layer sharing of similar attention patterns. It ingeniously reduces layer-wise redundant computation in attention. In Lazy Attention, we develop a novel layer-shared cache, Q Cache, tailored for MLLMs, which facilitates the reuse of queries across adjacent layers. In particular, Q Cache is lightweight and fully compatible with existing inference frameworks, including Flash Attention and KV cache. Additionally, our method is highly flexible as it is orthogonal to existing token-wise techniques and can be deployed independently or combined with token pruning approaches. Empirical evaluations on multiple benchmarks demonstrate that our method can reduce KV cache usage by over 35% and achieve 1.5x throughput improvement, while sacrificing only approximately 1% of performance on various MLLMs. Compared with SOTA token-wise methods, our technique achieves superior accuracy preservation.

Method: STELLAR factorizes visual features into a low-rank product of semantic concepts and their spatial distributions. This disentanglement allows DINO-style augmentation alignment on semantic tokens while maintaining precise spatial mapping in the localization matrix for pixel-level reconstruction.

Result: With only 16 sparse tokens, STELLAR achieves high-quality reconstruction (2.60 FID) while matching the semantic performance of dense backbones (79.10% ImageNet accuracy). This demonstrates a versatile sparse representation that bridges discriminative and generative vision.

Conclusion: STELLAR successfully resolves the tension between semantic understanding and image reconstruction in SSL by strategically separating semantic identity from spatial geometry, creating a unified framework for both discriminative and generative vision tasks.

Abstract: Self-supervised learning (SSL) faces a fundamental conflict between semantic understanding and image reconstruction. High-level semantic SSL (e.g., DINO) relies on global tokens that are forced to be location-invariant for augmentation alignment, a process that inherently discards the spatial coordinates required for reconstruction. Conversely, generative SSL (e.g., MAE) preserves dense feature grids for reconstruction but fails to produce high-level abstractions. We introduce STELLAR, a framework that resolves this tension by factorizing visual features into a low-rank product of semantic concepts and their spatial distributions. This disentanglement allows us to perform DINO-style augmentation alignment on the semantic tokens while maintaining the precise spatial mapping in the localization matrix necessary for pixel-level reconstruction. We demonstrate that as few as 16 sparse tokens under this factorized form are sufficient to simultaneously support high-quality reconstruction (2.60 FID) and match the semantic performance of dense backbones (79.10% ImageNet accuracy). Our results highlight STELLAR as a versatile sparse representation that bridges the gap between discriminative and generative vision by strategically separating semantic identity from spatial geometry. Code available at https://aka.ms/stellar.

Method: Proposes DSXFormer with Dual-Pooling Spectral Squeeze-Expansion (DSX) block using global average and max pooling to recalibrate spectral features, and Dynamic Context Attention (DCA) mechanism within window-based transformer to capture local spectral-spatial relationships efficiently. Includes patch extraction, embedding, and merging for multi-scale learning.

Result: Achieves classification accuracies of 99.95% (Salinas), 98.91% (Indian Pines), 99.85% (Pavia University), and 98.52% (Kennedy Space Center), outperforming state-of-the-art methods on four benchmark datasets.

Conclusion: DSXFormer effectively balances spectral emphasis and spatial contextual representation through spectral dual-pooling squeeze-expansion and dynamic context attention, demonstrating superior performance for hyperspectral image classification.

Abstract: Hyperspectral image classification (HSIC) is a challenging task due to high spectral dimensionality, complex spectral-spatial correlations, and limited labeled training samples. Although transformer-based models have shown strong potential for HSIC, existing approaches often struggle to achieve sufficient spectral discriminability while maintaining computational efficiency. To address these limitations, we propose a novel DSXFormer, a novel dual-pooling spectral squeeze-expansion transformer with Dynamic Context Attention for HSIC. The proposed DSXFormer introduces a Dual-Pooling Spectral Squeeze-Expansion (DSX) block, which exploits complementary global average and max pooling to adaptively recalibrate spectral feature channels, thereby enhancing spectral discriminability and inter-band dependency modeling. In addition, DSXFormer incorporates a Dynamic Context Attention (DCA) mechanism within a window-based transformer architecture to dynamically capture local spectral-spatial relationships while significantly reducing computational overhead. The joint integration of spectral dual-pooling squeeze-expansion and DCA enables DSXFormer to achieve an effective balance between spectral emphasis and spatial contextual representation. Furthermore, patch extraction, embedding, and patch merging strategies are employed to facilitate efficient multi-scale feature learning. Extensive experiments conducted on four widely used hyperspectral benchmark datasets, including Salinas (SA), Indian Pines (IP), Pavia University (PU), and Kennedy Space Center (KSC), demonstrate that DSXFormer consistently outperforms state-of-the-art methods, achieving classification accuracies of 99.95%, 98.91%, 99.85%, and 98.52%, respectively.

Method: Proposes MSPN with grid-based remapping (using high magnification features to derive coarse features) and coarse guidance network (CGN) that learns coarse contexts, enabling progressive multi-scale analysis on whole slide images.

Result: MSPN consistently improves MIL performance across 4 attention-based frameworks, 4 clinical tasks, and 3 foundation model types, while being lightweight and easy-to-use.

Conclusion: MSPN provides an effective plug-and-play solution for multi-scale analysis in computational pathology that enhances attention-based MIL frameworks across diverse clinical applications.

Abstract: Multiple-instance Learning (MIL) is commonly used to undertake computational pathology (CPath) tasks, and the use of multi-scale patches allows diverse features across scales to be learned. Previous studies using multi-scale features in clinical applications rely on multiple inputs across magnifications with late feature fusion, which does not retain the link between features across scales while the inputs are dependent on arbitrary, manufacturer-defined magnifications, being inflexible and computationally expensive. In this paper, we propose the Multi-scale Pyramidal Network (MSPN), which is plug-and-play over attention-based MIL that introduces progressive multi-scale analysis on WSI. Our MSPN consists of (1) grid-based remapping that uses high magnification features to derive coarse features and (2) the coarse guidance network (CGN) that learns coarse contexts. We benchmark MSPN as an add-on module to 4 attention-based frameworks using 4 clinically relevant tasks across 3 types of foundation model, as well as the pre-trained MIL framework. We show that MSPN consistently improves MIL across the compared configurations and tasks, while being lightweight and easy-to-use.

Method: Proposes RS-MPOD with visual prompt encoder to extract appearance-based category cues from exemplar instances for text-free specification, and multimodal fusion module to integrate visual and textual information when both are available.

Result: Visual prompting yields more reliable category specification under semantic ambiguity and distribution shifts, while multimodal prompting remains competitive when textual semantics are well aligned.

Conclusion: Multimodal prompting (visual + textual) provides a flexible and robust alternative to text-only prompting for open-vocabulary object detection in remote sensing applications.

Abstract: Open-vocabulary object detection in remote sensing commonly relies on text-only prompting to specify target categories, implicitly assuming that inference-time category queries can be reliably grounded through pretraining-induced text-visual alignment. In practice, this assumption often breaks down in remote sensing scenarios due to task- and application-specific category semantics, resulting in unstable category specification under open-vocabulary settings. To address this limitation, we propose RS-MPOD, a multimodal open-vocabulary detection framework that reformulates category specification beyond text-only prompting by incorporating instance-grounded visual prompts, textual prompts, and their multimodal integration. RS-MPOD introduces a visual prompt encoder to extract appearance-based category cues from exemplar instances, enabling text-free category specification, and a multimodal fusion module to integrate visual and textual information when both modalities are available. Extensive experiments on standard, cross-dataset, and fine-grained remote sensing benchmarks show that visual prompting yields more reliable category specification under semantic ambiguity and distribution shifts, while multimodal prompting provides a flexible alternative that remains competitive when textual semantics are well aligned.

Method: Proposes a theoretically grounded post-hoc calibration framework based on Bayesian decision theory. Introduces a learnable scalar correction to model logits, optimized on a small validation set from the target distribution while keeping the backbone frozen.

Result: Experiments on challenging benchmarks show the approach significantly improves robustness without retraining, offering a lightweight and principled solution for reliable and adaptive AI-generated image detection.

Conclusion: The proposed calibration framework effectively addresses distributional shift in AI-generated image detection, providing a practical solution for real-world deployment without requiring extensive retraining.

Abstract: Despite being trained on balanced datasets, existing AI-generated image detectors often exhibit systematic bias at test time, frequently misclassifying fake images as real. We hypothesize that this behavior stems from distributional shift in fake samples and implicit priors learned during training. Specifically, models tend to overfit to superficial artifacts that do not generalize well across different generation methods, leading to a misaligned decision threshold when faced with test-time distribution shift. To address this, we propose a theoretically grounded post-hoc calibration framework based on Bayesian decision theory. In particular, we introduce a learnable scalar correction to the model’s logits, optimized on a small validation set from the target distribution while keeping the backbone frozen. This parametric adjustment compensates for distributional shift in model output, realigning the decision boundary even without requiring ground-truth labels. Experiments on challenging benchmarks show that our approach significantly improves robustness without retraining, offering a lightweight and principled solution for reliable and adaptive AI-generated image detection in the open world. Code is available at https://github.com/muliyangm/AIGI-Det-Calib.

Method: Proposes scaling the hidden states of delimiter tokens to reinforce intra-image interactions and limit undesired cross-image interactions, enhancing the model’s ability to preserve image-specific information and distinguish between images.

Result: Shows performance gains on multi-image benchmarks (Mantis, MuirBench, MIRB, QBench2) and improves performance on text-only multi-document and multi-table understanding benchmarks (TQABench, MultiNews, WCEP-10) without additional training or inference costs.

Conclusion: Simple scaling of delimiter token hidden states effectively reduces cross-image information leakage in LVLMs, improving multi-image reasoning capabilities while maintaining computational efficiency.

Abstract: Large Vision-Language Models (LVLMs) achieve strong performance on single-image tasks, but their performance declines when multiple images are provided as input. One major reason is the cross-image information leakage, where the model struggles to distinguish information across different images. Existing LVLMs already employ delimiter tokens to mark the start and end of each image, yet our analysis reveals that these tokens fail to effectively block cross-image information leakage. To enhance their effectiveness, we propose a method that scales the hidden states of delimiter tokens. This enhances the model’s ability to preserve image-specific information by reinforcing intra-image interaction and limiting undesired cross-image interactions. Consequently, the model is better able to distinguish between images and reason over them more accurately. Experiments show performance gains on multi-image benchmarks such as Mantis, MuirBench, MIRB, and QBench2. We further evaluate our method on text-only tasks that require clear distinction. The method improves performance on multi-document and multi-table understanding benchmarks, including TQABench, MultiNews, and WCEP-10. Notably, our method requires no additional training or inference cost.

Method: Proposes a training framework incorporating masking features and an additional loss term that leverages prediction of the initial latent vector at any diffusion step to enhance correspondence between current step and final sample in latent space.

Result: Extensive experiments demonstrate the method effectively synthesizes high-quality images with controlled local conditions.

Conclusion: The approach enables precise local control over user-defined image regions while maintaining autonomous generation of remaining areas according to text prompts.

Abstract: Diffusion models emerged as a leading approach in text-to-image generation, producing high-quality images from textual descriptions. However, attempting to achieve detailed control to get a desired image solely through text remains a laborious trial-and-error endeavor. Recent methods have introduced image-level controls alongside with text prompts, using prior images to extract conditional information such as edges, segmentation and depth maps. While effective, these methods apply conditions uniformly across the entire image, limiting localized control. In this paper, we propose a novel methodology to enable precise local control over user-defined regions of an image, while leaving to the diffusion model the task of autonomously generating the remaining areas according to the original prompt. Our approach introduces a new training framework that incorporates masking features and an additional loss term, which leverages the prediction of the initial latent vector at any diffusion step to enhance the correspondence between the current step and the final sample in the latent space. Extensive experiments demonstrate that our method effectively synthesizes high-quality images with controlled local conditions.

Method: SurfSplat uses 2D Gaussian Splatting primitives instead of 3DGS, providing stronger anisotropy and higher geometric precision. It incorporates a surface continuity prior and forced alpha blending strategy to reconstruct coherent geometry with faithful textures.

Result: Extensive experiments on RealEstate10K, DL3DV, and ScanNet show SurfSplat consistently outperforms prior methods on both standard metrics and the proposed High-Resolution Rendering Consistency (HRRC) metric.

Conclusion: SurfSplat establishes a robust solution for high-fidelity 3D reconstruction from sparse inputs, producing continuous surfaces with accurate geometry and texture while addressing the limitations of previous 3DGS-based approaches.

Abstract: Reconstructing 3D scenes from sparse images remains a challenging task due to the difficulty of recovering accurate geometry and texture without optimization. Recent approaches leverage generalizable models to generate 3D scenes using 3D Gaussian Splatting (3DGS) primitive. However, they often fail to produce continuous surfaces and instead yield discrete, color-biased point clouds that appear plausible at normal resolution but reveal severe artifacts under close-up views. To address this issue, we present SurfSplat, a feedforward framework based on 2D Gaussian Splatting (2DGS) primitive, which provides stronger anisotropy and higher geometric precision. By incorporating a surface continuity prior and a forced alpha blending strategy, SurfSplat reconstructs coherent geometry together with faithful textures. Furthermore, we introduce High-Resolution Rendering Consistency (HRRC), a new evaluation metric designed to evaluate high-resolution reconstruction quality. Extensive experiments on RealEstate10K, DL3DV, and ScanNet demonstrate that SurfSplat consistently outperforms prior methods on both standard metrics and HRRC, establishing a robust solution for high-fidelity 3D reconstruction from sparse inputs. Project page: https://hebing-sjtu.github.io/SurfSplat-website/

Method: Proposes a unified framework with: 1) LiDAR-specific VAE and video VAE for encoding inputs, 2) Unified Latent Anchoring (ULA) to align cross-modal latent distributions, 3) Diffusion transformer for joint modeling of geometric correspondence and temporal evolution, 4) Structured scene layout conditioning per modality.

Result: Outperforms previous state-of-the-art methods in both video and LiDAR generation, with measurable improvements in downstream tasks.

Conclusion: UniDriveDreamer demonstrates effective unified multimodal world modeling for autonomous driving, enabling joint synthesis of visual and geometric sensor data without cascaded modules.

Abstract: World models have demonstrated significant promise for data synthesis in autonomous driving. However, existing methods predominantly concentrate on single-modality generation, typically focusing on either multi-camera video or LiDAR sequence synthesis. In this paper, we propose UniDriveDreamer, a single-stage unified multimodal world model for autonomous driving, which directly generates multimodal future observations without relying on intermediate representations or cascaded modules. Our framework introduces a LiDAR-specific variational autoencoder (VAE) designed to encode input LiDAR sequences, alongside a video VAE for multi-camera images. To ensure cross-modal compatibility and training stability, we propose Unified Latent Anchoring (ULA), which explicitly aligns the latent distributions of the two modalities. The aligned features are fused and processed by a diffusion transformer that jointly models their geometric correspondence and temporal evolution. Additionally, structured scene layout information is projected per modality as a conditioning signal to guide the synthesis. Extensive experiments demonstrate that UniDriveDreamer outperforms previous state-of-the-art methods in both video and LiDAR generation, while also yielding measurable improvements in downstream

Method: Introduces ClueTracer, a training-free, parameter-free plugin that traces how key clues propagate along the reasoning pathway (question → outputs → visual tokens) to localize task-relevant patches while suppressing attention to irrelevant regions.

Result: ClueTracer improves all reasoning architectures by 1.21× on reasoning benchmarks without additional training, and yields 1.14× gain when transferred to non-reasoning settings.

Conclusion: The method effectively addresses hallucination in multimodal reasoning by identifying and suppressing reasoning drift through visual clue tracing, working across various architectures without retraining.

Abstract: Large multimodal reasoning models solve challenging visual problems via explicit long-chain inference: they gather visual clues from images and decode clues into textual tokens. Yet this capability also increases hallucinations, where the model generates content that is not supported by the input image or the question. To understand this failure mode, we identify \emph{reasoning drift}: during clue gathering, the model over-focuses on question-irrelevant entities, diluting focus on task-relevant cues and gradually decoupling the reasoning trace from visual grounding. As a consequence, many inference-time localization or intervention methods developed for non-reasoning models fail to pinpoint the true clues in reasoning settings. Motivated by these insights, we introduce ClueRecall, a metric for assessing visual clue retrieval, and present ClueTracer, a training-free, parameter-free, and architecture-agnostic plugin for hallucination suppression. ClueTracer starts from the question and traces how key clues propagate along the model’s reasoning pathway (question $\rightarrow$ outputs $\rightarrow$ visual tokens), thereby localizing task-relevant patches while suppressing spurious attention to irrelevant regions. Remarkably, \textbf{without any additional training}, ClueTracer improves all \textbf{reasoning} architectures (including \texttt{R1-OneVision}, \texttt{Ocean-R1}, \texttt{MM-Eureka}, \emph{etc}.) by $\mathbf{1.21\times}$ on reasoning benchmarks. When transferred to \textbf{non-reasoning} settings, it yields a $\mathbf{1.14\times}$ gain.

Method: 1) Product-aware adaptive grouping dynamically organizes users based on attributes and product characteristics; 2) Preference-conditioned image generation uses a Group-aware Multimodal Large Language Model (G-MLLM) pre-trained to comprehend group features and generate images; 3) Fine-tuning with Group-DPO for group-wise preference alignment; 4) Introduction of GAIP dataset with 600K groups from 40M users.

Result: The framework achieves state-of-the-art performance in both offline and online settings, effectively enhancing each group’s CTR on generated images through group-wise preference alignment.

Conclusion: OSMF successfully bridges the gap in advertising image generation by addressing preference diversity among user groups through adaptive grouping and multimodal LLM-based generation with preference alignment, advancing the field with a new dataset and methodology.

Abstract: Advertising image generation has increasingly focused on online metrics like Click-Through Rate (CTR), yet existing approaches adopt a ``one-size-fits-all" strategy that optimizes for overall CTR while neglecting preference diversity among user groups. This leads to suboptimal performance for specific groups, limiting targeted marketing effectiveness. To bridge this gap, we present \textit{One Size, Many Fits} (OSMF), a unified framework that aligns diverse group-wise click preferences in large-scale advertising image generation. OSMF begins with product-aware adaptive grouping, which dynamically organizes users based on their attributes and product characteristics, representing each group with rich collective preference features. Building on these groups, preference-conditioned image generation employs a Group-aware Multimodal Large Language Model (G-MLLM) to generate tailored images for each group. The G-MLLM is pre-trained to simultaneously comprehend group features and generate advertising images. Subsequently, we fine-tune the G-MLLM using our proposed Group-DPO for group-wise preference alignment, which effectively enhances each group’s CTR on the generated images. To further advance this field, we introduce the Grouped Advertising Image Preference Dataset (GAIP), the first large-scale public dataset of group-wise image preferences, including around 600K groups built from 40M users. Extensive experiments demonstrate that our framework achieves the state-of-the-art performance in both offline and online settings. Our code and datasets will be released at https://github.com/JD-GenX/OSMF.

Method: Auto-Comp is a fully automated synthetic pipeline that generates paired images from Minimal captions (simple descriptions) and LLM-generated Contextual captions (rich scene descriptions). This enables controlled A/B testing to isolate core binding ability from visio-linguistic complexity. The approach includes novel benchmarks for color binding, spatial relations, and a “Confusion Benchmark” with low-entropy distractors.

Result: Evaluation of 20 VLMs (CLIP and SigLIP families) reveals universal compositional failures. The Confusion Benchmark shows models are highly susceptible to low-entropy distractors (repeated objects/colors). A surprising trade-off emerges: visio-linguistic context aids spatial reasoning but hinders local attribute binding by introducing visual clutter.

Conclusion: Auto-Comp provides a powerful framework for analyzing compositional reasoning failures in VLMs, revealing deeper flaws beyond simple attribute swaps. The pipeline enables scalable benchmark generation for future research, with released resources on HuggingFace.

Abstract: Modern Vision-Language Models (VLMs) exhibit a critical flaw in compositional reasoning, often confusing “a red cube and a blue sphere” with “a blue cube and a red sphere”. Disentangling the visual and linguistic roots of these failures is a fundamental challenge for robust evaluation. To enable fine-grained, controllable analysis, we introduce Auto-Comp, a fully automated and synthetic pipeline for generating scalable benchmarks. Its controllable nature is key to dissecting and isolating different reasoning skills. Auto-Comp generates paired images from Minimal (e.g., “a monitor to the left of a bicycle on a white background”) and LLM-generated Contextual captions (e.g., “In a brightly lit photography studio, a monitor is positioned to the left of a bicycle”), allowing a controlled A/B test to disentangle core binding ability from visio-linguistic complexity. Our evaluation of 20 VLMs on novel benchmarks for color binding and spatial relations reveals universal compositional failures in both CLIP and SigLIP model families. Crucially, our novel “Confusion Benchmark” reveals a deeper flaw beyond simple attribute swaps: models are highly susceptible to low-entropy distractors (e.g., repeated objects or colors), demonstrating their compositional failures extend beyond known bag-of-words limitations. we uncover a surprising trade-off: visio-linguistic context, which provides global scene cues, aids spatial reasoning but simultaneously hinders local attribute binding by introducing visual clutter. We release the Auto-Comp pipeline to facilitate future benchmark creation, alongside all our generated benchmarks (https://huggingface.co/AutoComp).

Method: Proposes SegmentMIL, a transformer-based multi-view multiple-instance learning framework that uses patient-level supervision without view-level annotations. It jointly predicts stenosis presence and localizes affected anatomical regions (right/left coronary arteries and segments).

Result: SegmentMIL achieves high performance on internal and external evaluations, outperforming both view-level models and classical MIL baselines.

Conclusion: SegmentMIL shows potential as a clinically viable and scalable solution for coronary stenosis diagnosis by leveraging patient-level supervision and capturing multi-view dependencies.

Abstract: Coronary artery stenosis is a leading cause of cardiovascular disease, diagnosed by analyzing the coronary arteries from multiple angiography views. Although numerous deep-learning models have been proposed for stenosis detection from a single angiography view, their performance heavily relies on expensive view-level annotations, which are often not readily available in hospital systems. Moreover, these models fail to capture the temporal dynamics and dependencies among multiple views, which are crucial for clinical diagnosis. To address this, we propose SegmentMIL, a transformer-based multi-view multiple-instance learning framework for patient-level stenosis classification. Trained on a real-world clinical dataset, using patient-level supervision and without any view-level annotations, SegmentMIL jointly predicts the presence of stenosis and localizes the affected anatomical region, distinguishing between the right and left coronary arteries and their respective segments. SegmentMIL obtains high performance on internal and external evaluations and outperforms both view-level models and classical MIL baselines, underscoring its potential as a clinically viable and scalable solution for coronary stenosis diagnosis. Our code is available at https://github.com/NikolaCenic/mil-stenosis.

Method: 1) Depth-Consistent D-Normal Regularization module integrating D-Normal constraints with external depth supervision for comprehensive geometric parameter updates; 2) Adaptive confidence weighting based on gradient consistency and inverse depth deviation; 3) Spatially Adaptive Gaussian Pruning (SAGP) strategy that dynamically adjusts Gaussian density; 4) Unified partitioning and view assignment scheme to eliminate boundary artifacts.

Result: Extensive experiments on multiple urban datasets demonstrate superior performance in rendering quality, geometric accuracy, and memory efficiency compared to existing approaches.

Conclusion: UrbanGS provides a systematic solution for high-fidelity large-scale scene reconstruction, effectively addressing the challenges of geometric consistency, memory efficiency, and computational scalability in city-scale 3D Gaussian Splatting applications.

Abstract: While 3D Gaussian Splatting (3DGS) enables high-quality, real-time rendering for bounded scenes, its extension to large-scale urban environments gives rise to critical challenges in terms of geometric consistency, memory efficiency, and computational scalability. To address these issues, we present UrbanGS, a scalable reconstruction framework that effectively tackles these challenges for city-scale applications. First, we propose a Depth-Consistent D-Normal Regularization module. Unlike existing approaches that rely solely on monocular normal estimators, which can effectively update rotation parameters yet struggle to update position parameters, our method integrates D-Normal constraints with external depth supervision. This allows for comprehensive updates of all geometric parameters. By further incorporating an adaptive confidence weighting mechanism based on gradient consistency and inverse depth deviation, our approach significantly enhances multi-view depth alignment and geometric coherence, which effectively resolves the issue of geometric accuracy in complex large-scale scenes. To improve scalability, we introduce a Spatially Adaptive Gaussian Pruning (SAGP) strategy, which dynamically adjusts Gaussian density based on local geometric complexity and visibility to reduce redundancy. Additionally, a unified partitioning and view assignment scheme is designed to eliminate boundary artifacts and optimize computational load. Extensive experiments on multiple urban datasets demonstrate that UrbanGS achieves superior performance in rendering quality, geometric accuracy, and memory efficiency, providing a systematic solution for high-fidelity large-scale scene reconstruction.

Method: 1) New video autoencoder with 64×64×4 spatial-temporal compression, 2) Diffusion transformer with layer memory design for better inter-layer information flow, 3) Multi-resolution generation using a few-step DIT upsampler to enhance video fidelity.

Result: Achieves competitive performance against popular open-source models while being an order of magnitude faster, using a 14B DIT base model and 14B DIT upsampler.

Conclusion: FSVideo demonstrates that efficient transformer-based architectures with optimized latent spaces and memory designs can enable fast, high-quality image-to-video generation.

Abstract: We introduce FSVideo, a fast speed transformer-based image-to-video (I2V) diffusion framework. We build our framework on the following key components: 1.) a new video autoencoder with highly-compressed latent space ($64\times64\times4$ spatial-temporal downsampling ratio), achieving competitive reconstruction quality; 2.) a diffusion transformer (DIT) architecture with a new layer memory design to enhance inter-layer information flow and context reuse within DIT, and 3.) a multi-resolution generation strategy via a few-step DIT upsampler to increase video fidelity. Our final model, which contains a 14B DIT base model and a 14B DIT upsampler, achieves competitive performance against other popular open-source models, while being an order of magnitude faster. We discuss our model design as well as training strategies in this report.

Method: Proposes DSKD: teacher-guided student Diffusion Self-KD. Uses teacher classifier to guide diffusion model’s sampling process to denoise student features, then performs LSH-guided feature distillation between original and denoised student features.

Result: Significantly outperforms existing KD methods across various models and datasets on visual recognition tasks.

Conclusion: DSKD effectively eliminates teacher-student feature distribution discrepancies while learning meaningful knowledge from teachers, demonstrating superior performance over conventional KD approaches.

Abstract: Existing Knowledge Distillation (KD) methods often align feature information between teacher and student by exploring meaningful feature processing and loss functions. However, due to the difference in feature distributions between the teacher and student, the student model may learn incompatible information from the teacher. To address this problem, we propose teacher-guided student Diffusion Self-KD, dubbed as DSKD. Instead of the direct teacher-student alignment, we leverage the teacher classifier to guide the sampling process of denoising student features through a light-weight diffusion model. We then propose a novel locality-sensitive hashing (LSH)-guided feature distillation method between the original and denoised student features. The denoised student features encapsulate teacher knowledge and could be regarded as a teacher role. In this way, our DSKD method could eliminate discrepancies in mapping manners and feature distributions between the teacher and student, while learning meaningful knowledge from the teacher. Experiments on visual recognition tasks demonstrate that DSKD significantly outperforms existing KD methods across various models and datasets. Our code is attached in supplementary material.

Method: Proposes iCCDM framework incorporating Elucidated Diffusion Model (EDM) with novel matrix-form EDM formulation and adaptive vicinal training strategy to enhance both quality and efficiency.

Result: iCCDM consistently outperforms existing methods across four benchmark datasets (64×64 to 256×256 resolution), including state-of-the-art large-scale text-to-image diffusion models (Stable Diffusion 3, FLUX.1, Qwen-Image), achieving higher generation quality with significantly reduced sampling cost.

Conclusion: iCCDM successfully addresses CCDM’s limitations by integrating advanced EDM framework with novel formulations, establishing new state-of-the-art for continuous conditional image generation with improved efficiency.

Abstract: Continuous Conditional Diffusion Model (CCDM) is a diffusion-based framework designed to generate high-quality images conditioned on continuous regression labels. Although CCDM has demonstrated clear advantages over prior approaches across a range of datasets, it still exhibits notable limitations and has recently been surpassed by a GAN-based method, namely CcGAN-AVAR. These limitations mainly arise from its reliance on an outdated diffusion framework and its low sampling efficiency due to long sampling trajectories. To address these issues, we propose an improved CCDM framework, termed iCCDM, which incorporates the more advanced \textit{Elucidated Diffusion Model} (EDM) framework with substantial modifications to improve both generation quality and sampling efficiency. Specifically, iCCDM introduces a novel matrix-form EDM formulation together with an adaptive vicinal training strategy. Extensive experiments on four benchmark datasets, spanning image resolutions from $64\times64$ to $256\times256$, demonstrate that iCCDM consistently outperforms existing methods, including state-of-the-art large-scale text-to-image diffusion models (e.g., Stable Diffusion 3, FLUX.1, and Qwen-Image), achieving higher generation quality while significantly reducing sampling cost.

Method: MLV-Edit uses a divide-and-conquer strategy for segment-wise editing with two core modules: Velocity Blend rectifies motion inconsistencies at segment boundaries by aligning flow fields, and Attention Sink anchors local segment features to global reference frames to suppress structural drift.

Result: Extensive experiments show MLV-Edit consistently outperforms state-of-the-art methods in temporal stability and semantic fidelity for minute-level video editing.

Conclusion: MLV-Edit provides an effective training-free solution for long-duration video editing that maintains temporal consistency and handles computational challenges through innovative flow-based techniques.

Abstract: We propose MLV-Edit, a training-free, flow-based framework that address the unique challenges of minute-level video editing. While existing techniques excel in short-form video manipulation, scaling them to long-duration videos remains challenging due to prohibitive computational overhead and the difficulty of maintaining global temporal consistency across thousands of frames. To address this, MLV-Edit employs a divide-and-conquer strategy for segment-wise editing, facilitated by two core modules: Velocity Blend rectifies motion inconsistencies at segment boundaries by aligning the flow fields of adjacent chunks, eliminating flickering and boundary artifacts commonly observed in fragmented video processing; and Attention Sink anchors local segment features to global reference frames, effectively suppressing cumulative structural drift. Extensive quantitative and qualitative experiments demonstrate that MLV-Edit consistently outperforms state-of-the-art methods in terms of temporal stability and semantic fidelity.

Method: Created a novel dataset with pixelwise annotations of healthy tissue and known pathologies. Fine-tuned a pre-trained Vision Transformer (DINOv2) using Low-Rank Adaptation (LoRA) for tissue segmentation. Extracted features for out-of-distribution detection using Mahalanobis distance with class-specific thresholds optimized using mean of false negative and false positive rates.

Result: Achieved high accuracy with only 0.16% of pathological tissue classified as healthy and 0.35% of healthy tissue classified as pathological. Framework accurately detects anomalies including rare OOD morphologies in mouse liver WSIs with known toxicological findings.

Conclusion: Demonstrates potential of AI-driven histopathology to support preclinical workflows, reduce late-stage failures, and improve efficiency in drug development through automated anomaly detection in histopathological images.

Abstract: Drug-induced toxicity remains a leading cause of failure in preclinical development and early clinical trials. Detecting adverse effects at an early stage is critical to reduce attrition and accelerate the development of safe medicines. Histopathological evaluation remains the gold standard for toxicity assessment, but it relies heavily on expert pathologists, creating a bottleneck for large-scale screening. To address this challenge, we introduce an AI-based anomaly detection framework for histopathological whole-slide images (WSIs) in rodent livers from toxicology studies. The system identifies healthy tissue and known pathologies (anomalies) for which training data is available. In addition, it can detect rare pathologies without training data as out-of-distribution (OOD) findings. We generate a novel dataset of pixelwise annotations of healthy tissue and known pathologies and use this data to fine-tune a pre-trained Vision Transformer (DINOv2) via Low-Rank Adaptation (LoRA) in order to do tissue segmentation. Finally, we extract features for OOD detection using the Mahalanobis distance. To better account for class-dependent variability in histological data, we propose the use of class-specific thresholds. We optimize the thresholds using the mean of the false negative and false positive rates, resulting in only 0.16% of pathological tissue classified as healthy and 0.35% of healthy tissue classified as pathological. Applied to mouse liver WSIs with known toxicological findings, the framework accurately detects anomalies, including rare OOD morphologies. This work demonstrates the potential of AI-driven histopathology to support preclinical workflows, reduce late-stage failures, and improve efficiency in drug development.

Method: Proposes physics-based CBCT simulation to create geometrically aligned training pairs by construction, combined with evaluation using geometric alignment metrics (Normalized Mutual Information) against input CBCT rather than biased ground truth.

Result: Models trained on synthetic data achieved superior geometric alignment (NMI: 0.31 vs 0.22) despite lower conventional intensity scores. NMI consistently predicted observer preference across registration methods, with clinical observers preferring synthetic-trained outputs in 87% of cases.

Conclusion: Geometric fidelity, not intensity agreement with biased ground truth, aligns with clinical requirements for CT generation from CBCT, demonstrating the importance of proper evaluation metrics.

Abstract: Supervised synthetic CT generation from CBCT requires registered training pairs, yet perfect registration between separately acquired scans remains unattainable. This registration bias propagates into trained models and corrupts standard evaluation metrics. This may suggest that superior benchmark performance indicates better reproduction of registration artifacts rather than anatomical fidelity. We propose physics-based CBCT simulation to provide geometrically aligned training pairs by construction, combined with evaluation using geometric alignment metrics against input CBCT rather than biased ground truth. On two independent pelvic datasets, models trained on synthetic data achieved superior geometric alignment (Normalized Mutual Information: 0.31 vs 0.22) despite lower conventional intensity scores. Intensity metrics showed inverted correlations with clinical assessment for deformably registered data, while Normalized Mutual Information consistently predicted observer preference across registration methodologies (rho = 0.31, p < 0.001). Clinical observers preferred synthetic-trained outputs in 87% of cases, demonstrating that geometric fidelity, not intensity agreement with biased ground truth, aligns with clinical requirements.

Method: Fully automated deep learning-based framework with modular design integrating dense map alignment, multi-temporal object detection, and change profiling to enable systematic quantitative characterization of urban change

Result: Demonstrates robust performance of alignment and object detection methods; applied to Paris (1868-1937) reveals spatial and temporal heterogeneity in urban transformation; framework supports adaptation to diverse cartographic contexts

Conclusion: Shifts historical map analysis from ad hoc visual comparison to systematic quantitative characterization, with relevance for social sciences and humanities research

Abstract: Prior to modern Earth observation technologies, historical maps provide a unique record of long-term urban transformation and offer a lens on the evolving identity of cities. However, extracting consistent and fine-grained change information from historical map series remains challenging due to spatial misalignment, cartographic variation, and degrading document quality, limiting most analyses to small-scale or qualitative approaches. We propose a fully automated, deep learning-based framework for fine-grained urban change analysis from large collections of historical maps, built on a modular design that integrates dense map alignment, multi-temporal object detection, and change profiling. This framework shifts the analysis of historical maps from ad hoc visual comparison toward systematic, quantitative characterization of urban change. Experiments demonstrate the robust performance of the proposed alignment and object detection methods. Applied to Paris between 1868 and 1937, the framework reveals the spatial and temporal heterogeneity in urban transformation, highlighting its relevance for research in the social sciences and humanities. The modular design of our framework further supports adaptation to diverse cartographic contexts and downstream applications.

Method: Proposes Loop-ViT with weight-tied recurrence that decouples reasoning depth from model capacity, using a Hybrid Block combining local convolutions and global attention. Introduces parameter-free Dynamic Exit mechanism based on predictive entropy to halt inference when internal state reaches low-uncertainty attractor.

Result: 18M parameter Loop-ViT achieves 65.8% accuracy on ARC-AGI-1 benchmark, outperforming massive 73M-parameter ensembles, demonstrating adaptive iterative computation as more efficient scaling axis than increasing network width.

Conclusion: Recursive architectures with adaptive computation offer superior efficiency for visual reasoning compared to traditional feed-forward models, providing a new scaling axis through iterative computation rather than parameter growth.

Abstract: Recent advances in visual reasoning have leveraged vision transformers to tackle the ARC-AGI benchmark. However, we argue that the feed-forward architecture, where computational depth is strictly bound to parameter size, falls short of capturing the iterative, algorithmic nature of human induction. In this work, we propose a recursive architecture called Loop-ViT, which decouples reasoning depth from model capacity through weight-tied recurrence. Loop-ViT iterates a weight-tied Hybrid Block, combining local convolutions and global attention, to form a latent chain of thought. Crucially, we introduce a parameter-free Dynamic Exit mechanism based on predictive entropy: the model halts inference when its internal state ``crystallizes" into a low-uncertainty attractor. Empirical results on the ARC-AGI-1 benchmark validate this perspective: our 18M model achieves 65.8% accuracy, outperforming massive 73M-parameter ensembles. These findings demonstrate that adaptive iterative computation offers a far more efficient scaling axis for visual reasoning than simply increasing network width. The code is available at https://github.com/WenjieShu/LoopViT.

Method: Introduces Reg4Pru, a training regularization technique specifically designed to mitigate performance loss from token pruning. The method likely involves regularization strategies that stabilize preserved token representations during pruning operations.

Result: On the FIVES blood vessel segmentation dataset, Reg4Pru improves average precision by 46% absolute compared to same model trained without routing, while achieving 29% relative speedup in wall-clock time compared to non-pruned baseline.

Conclusion: Reg4Pru is an effective regularizer for token reduction strategies in vision transformers, enabling significant computational efficiency gains without sacrificing segmentation performance.

Abstract: Transformers are widely adopted in modern vision models due to their strong ability to scale with dataset size and generalisability. However, this comes with a major drawback: computation scales quadratically to the total number of tokens. Numerous methods have been proposed to mitigate this. For example, we consider token pruning with reactivating tokens from preserved representations, but the increased computational efficiency of this method results in decreased stability from the preserved representations, leading to poorer dense prediction performance at deeper layers. In this work, we introduce Reg4Pru, a training regularisation technique that mitigates token-pruning performance loss for segmentation. We compare our models on the FIVES blood vessel segmentation dataset and find that Reg4Pru improves average precision by an absolute 46% compared to the same model trained without routing. This increase is observed using a configuration that achieves a 29% relative speedup in wall-clock time compared to the non-pruned baseline. These findings indicate that Reg4Pru is a valuable regulariser for token reduction strategies.

Method: Two-stage GAN: 1) StyleGAN generates semantic segmentation masks to control anatomical structure; 2) DL-Pix2Pix translates masks into CT images using local importance attention and dynamic weight multi-head window attention to enhance texture and background modeling.

Result: Accuracy improved by 4.6% and mAP by 4% on LUNA16 dataset compared to original dataset. TSGAN enhances synthetic image quality and detection model performance.

Conclusion: TSGAN effectively enhances diversity and spatial controllability of synthetic lung nodule CT data by decoupling morphological structure and texture features, leading to improved detection model performance.

Abstract: The limited sample size and insufficient diversity of lung nodule CT datasets severely restrict the performance and generalization ability of detection models. Existing methods generate images with insufficient diversity and controllability, suffering from issues such as monotonous texture features and distorted anatomical structures. Therefore, we propose a two-stage generative adversarial network (TSGAN) to enhance the diversity and spatial controllability of synthetic data by decoupling the morphological structure and texture features of lung nodules. In the first stage, StyleGAN is used to generate semantic segmentation mask images, encoding lung nodules and tissue backgrounds to control the anatomical structure of lung nodule images; The second stage uses the DL-Pix2Pix model to translate the mask map into CT images, employing local importance attention to capture local features, while utilizing dynamic weight multi-head window attention to enhance the modeling capability of lung nodule texture and background. Compared to the original dataset, the accuracy improved by 4.6% and mAP by 4% on the LUNA16 dataset. Experimental results demonstrate that TSGAN can enhance the quality of synthetic images and the performance of detection models.

Method: CIEC framework has two branches: 1) Image-based weakly-supervised localization using Textual-guidance Refine Patch Selection (TRPS) module that integrates visual and textual forgery cues with spatial priors, plus background silencing and spatial contrast constraints. 2) Text-based weakly-supervised localization using Visual-deviation Calibrated Token Grounding (VCTG) module that focuses on content words with visual bias, plus asymmetric sparse and semantic consistency constraints.

Result: Extensive experiments show CIEC achieves results comparable to fully supervised methods on several evaluation metrics, demonstrating effectiveness of the weakly-supervised approach.

Conclusion: CIEC successfully addresses the annotation bottleneck in multimodal manipulation localization by proposing a weakly-supervised framework that leverages implicit and explicit cues between modalities, achieving strong performance with only coarse-grained supervision.

Abstract: To mitigate the threat of misinformation, multimodal manipulation localization has garnered growing attention. Consider that current methods rely on costly and time-consuming fine-grained annotations, such as patch/token-level annotations. This paper proposes a novel framework named Coupling Implicit and Explicit Cues (CIEC), which aims to achieve multimodal weakly-supervised manipulation localization for image-text pairs utilizing only coarse-grained image/sentence-level annotations. It comprises two branches, image-based and text-based weakly-supervised localization. For the former, we devise the Textual-guidance Refine Patch Selection (TRPS) module. It integrates forgery cues from both visual and textual perspectives to lock onto suspicious regions aided by spatial priors. Followed by the background silencing and spatial contrast constraints to suppress interference from irrelevant areas. For the latter, we devise the Visual-deviation Calibrated Token Grounding (VCTG) module. It focuses on meaningful content words and leverages relative visual bias to assist token localization. Followed by the asymmetric sparse and semantic consistency constraints to mitigate label noise and ensure reliability. Extensive experiments demonstrate the effectiveness of our CIEC, yielding results comparable to fully supervised methods on several evaluation metrics.

Method: Uses sparse surface and skeleton point clouds to learn a continuous implicit field for topology repair without complete annotations, trained on synthetically introduced disruptions over already incomplete trees. Jointly infers anatomical labeling and lung segment reconstruction through task-specific implicit functions.

Result: Extensive experiments on Lung3D+ dataset show improved topological completeness and accurate anatomical labeling/segment reconstruction under incomplete scenarios. High computational efficiency (just over one second per case).

Conclusion: TopoField provides an efficient, practical solution for pulmonary tree analysis with topology repair, suitable for large-scale clinical applications.

Abstract: Pulmonary trees extracted from CT images frequently exhibit topological incompleteness, such as missing or disconnected branches, which substantially degrades downstream anatomical analysis and limits the applicability of existing pulmonary tree modeling pipelines. Current approaches typically rely on dense volumetric processing or explicit graph reasoning, leading to limited efficiency and reduced robustness under realistic structural corruption. We propose TopoField, a topology-aware implicit modeling framework that treats topology repair as a first-class modeling problem and enables unified multi-task inference for pulmonary tree analysis. TopoField represents pulmonary anatomy using sparse surface and skeleton point clouds and learns a continuous implicit field that supports topology repair without relying on complete or explicit disconnection annotations, by training on synthetically introduced structural disruptions over \textit{already} incomplete trees. Building upon the repaired implicit representation, anatomical labeling and lung segment reconstruction are jointly inferred through task-specific implicit functions within a single forward pass.Extensive experiments on the Lung3D+ dataset demonstrate that TopoField consistently improves topological completeness and achieves accurate anatomical labeling and lung segment reconstruction under challenging incomplete scenarios. Owing to its implicit formulation, TopoField attains high computational efficiency, completing all tasks in just over one second per case, highlighting its practicality for large-scale and time-sensitive clinical applications. Code and data will be available at https://github.com/HINTLab/TopoField.

Method: Singularity Skipping Inversion (SSI-DM) bypasses the singular region by adding small noise before standard inversion, producing inverted noise with natural Gaussian properties while maintaining reconstruction fidelity.

Result: The method achieves superior performance on public image datasets for reconstruction and interpolation tasks, providing a principled and efficient solution to diffusion model inversion.

Conclusion: SSI-DM offers a simple, plug-and-play technique compatible with general diffusion models that solves the fundamental ill-posedness of diffusion inversion while preserving editability.

Abstract: Inverting real images into the noise space is essential for editing tasks using diffusion models, yet existing methods produce non-Gaussian noise with poor editability due to the inaccuracy in early noising steps. We identify the root cause: a mathematical singularity that renders inversion fundamentally ill-posed. We propose Singularity Skipping Inversion of Diffusion Models (SSI-DM), which bypasses this singular region by adding small noise before standard inversion. This simple approach produces inverted noise with natural Gaussian properties while maintaining reconstruction fidelity. As a plug-and-play technique compatible with general diffusion models, our method achieves superior performance on public image datasets for reconstruction and interpolation tasks, providing a principled and efficient solution to diffusion model inversion.

Method: Introduces MAIN-VLA framework with two key components: 1) Intention Abstraction (IA) extracts verbose linguistic instructions into compact semantic primitives, 2) Environment Semantics Abstraction (ESA) projects visual streams into structured topological affordance representation. Aligning these modalities enables parameter-free token-pruning for attention concentration.

Result: Extensive experiments in open-world Minecraft and large-scale PvP environments (Game for Peace, Valorant) show MAIN-VLA achieves state-of-the-art performance with superior decision quality, stronger generalization, and cutting-edge inference efficiency.

Conclusion: MAIN-VLA successfully addresses VLA inefficiencies in complex environments through explicit abstraction of intention and environment semantics, enabling deep semantic alignment and efficient attention mechanisms for improved decision-making.

Abstract: Despite significant progress in Visual-Language-Action (VLA), in highly complex and dynamic environments that involve real-time unpredictable interactions (such as 3D open worlds and large-scale PvP games), existing approaches remain inefficient at extracting action-critical signals from redundant sensor streams. To tackle this, we introduce MAIN-VLA, a framework that explicitly Models the Abstraction of Intention and eNvironment to ground decision-making in deep semantic alignment rather than superficial pattern matching. Specifically, our Intention Abstraction (IA) extracts verbose linguistic instructions and their associated reasoning into compact, explicit semantic primitives, while the Environment Semantics Abstraction (ESA) projects overwhelming visual streams into a structured, topological affordance representation. Furthermore, aligning these two abstract modalities induces an emergent attention-concentration effect, enabling a parameter-free token-pruning strategy that filters out perceptual redundancy without degrading performance. Extensive experiments in open-world Minecraft and large-scale PvP environments (Game for Peace and Valorant) demonstrate that MAIN-VLA sets a new state-of-the-art, which achieves superior decision quality, stronger generalization, and cutting-edge inference efficiency.

Method: Proposes Causal Forcing that uses an autoregressive teacher for ODE initialization instead of a bidirectional teacher, thereby bridging the architectural gap. This ensures the frame-level injectivity condition is satisfied and allows proper recovery of the teacher’s flow map.

Result: Outperforms all baselines across all metrics, surpassing state-of-the-art Self Forcing by 19.3% in Dynamic Degree, 8.7% in VisionReward, and 16.7% in Instruction Following.

Conclusion: Causal Forcing effectively bridges the architectural gap in distilling bidirectional video diffusion models into autoregressive models for real-time interactive video generation, achieving superior performance by using AR teacher for ODE initialization.

Abstract: To achieve real-time interactive video generation, current methods distill pretrained bidirectional video diffusion models into few-step autoregressive (AR) models, facing an architectural gap when full attention is replaced by causal attention. However, existing approaches do not bridge this gap theoretically. They initialize the AR student via ODE distillation, which requires frame-level injectivity, where each noisy frame must map to a unique clean frame under the PF-ODE of an AR teacher. Distilling an AR student from a bidirectional teacher violates this condition, preventing recovery of the teacher’s flow map and instead inducing a conditional-expectation solution, which degrades performance. To address this issue, we propose Causal Forcing that uses an AR teacher for ODE initialization, thereby bridging the architectural gap. Empirical results show that our method outperforms all baselines across all metrics, surpassing the SOTA Self Forcing by 19.3% in Dynamic Degree, 8.7% in VisionReward, and 16.7% in Instruction Following. Project page and the code: \href{https://thu-ml.github.io/CausalForcing.github.io/}{https://thu-ml.github.io/CausalForcing.github.io/}

Method: The authors introduce HieraNav as a multi-granularity navigation task and create LangMap benchmark using real-world 3D indoor scans with human-verified annotations. They provide region labels, discriminative region/instance descriptions covering 414 object categories, and over 18K navigation tasks with both concise and detailed descriptions.

Result: LangMap achieves 23.8% higher discriminative accuracy than GOAT-Bench using four times fewer words. Evaluations show richer context and memory improve success rates, while challenges remain with long-tailed, small, context-dependent, and distant goals, as well as multi-goal completion.

Conclusion: HieraNav and LangMap establish a rigorous testbed for advancing language-driven embodied navigation, providing high-quality annotations and tasks that reveal current model limitations and opportunities for improvement in multimodal language understanding for embodied AI.

Abstract: The relationships between objects and language are fundamental to meaningful communication between humans and AI, and to practically useful embodied intelligence. We introduce HieraNav, a multi-granularity, open-vocabulary goal navigation task where agents interpret natural language instructions to reach targets at four semantic levels: scene, room, region, and instance. To this end, we present Language as a Map (LangMap), a large-scale benchmark built on real-world 3D indoor scans with comprehensive human-verified annotations and tasks spanning these levels. LangMap provides region labels, discriminative region descriptions, discriminative instance descriptions covering 414 object categories, and over 18K navigation tasks. Each target features both concise and detailed descriptions, enabling evaluation across different instruction styles. LangMap achieves superior annotation quality, outperforming GOAT-Bench by 23.8% in discriminative accuracy using four times fewer words. Comprehensive evaluations of zero-shot and supervised models on LangMap reveal that richer context and memory improve success, while long-tailed, small, context-dependent, and distant goals, as well as multi-goal completion, remain challenging. HieraNav and LangMap establish a rigorous testbed for advancing language-driven embodied navigation. Project: https://bo-miao.github.io/LangMap

Method: Proposes MIRROR (Manifold Ideal Reference ReconstructOR), a framework that explicitly encodes reality priors using a learnable discrete memory bank. The method projects an input into a manifold-consistent ideal reference via sparse linear combination and uses the resulting residuals as robust detection signals.

Result: Across 14 benchmarks, MIRROR consistently outperforms prior methods, achieving gains of 2.1% on six standard benchmarks and 8.1% on seven in-the-wild benchmarks. On the Human-AIGI benchmark, MIRROR reaches 89.6% accuracy across 27 generators, surpassing both lay users and visual experts.

Conclusion: MIRROR demonstrates superior performance in AI-generated image detection by leveraging manifold consistency rather than artifact-based classification, approaching human perceptual limits as pretrained backbones scale, and showing potential to replace human experts in media security applications.

Abstract: High-fidelity generative models have narrowed the perceptual gap between synthetic and real images, posing serious threats to media security. Most existing AI-generated image (AIGI) detectors rely on artifact-based classification and struggle to generalize to evolving generative traces. In contrast, human judgment relies on stable real-world regularities, with deviations from the human cognitive manifold serving as a more generalizable signal of forgery. Motivated by this insight, we reformulate AIGI detection as a Reference-Comparison problem that verifies consistency with the real-image manifold rather than fitting specific forgery cues. We propose MIRROR (Manifold Ideal Reference ReconstructOR), a framework that explicitly encodes reality priors using a learnable discrete memory bank. MIRROR projects an input into a manifold-consistent ideal reference via sparse linear combination, and uses the resulting residuals as robust detection signals. To evaluate whether detectors reach the “superhuman crossover” required to replace human experts, we introduce the Human-AIGI benchmark, featuring a psychophysically curated human-imperceptible subset. Across 14 benchmarks, MIRROR consistently outperforms prior methods, achieving gains of 2.1% on six standard benchmarks and 8.1% on seven in-the-wild benchmarks. On Human-AIGI, MIRROR reaches 89.6% accuracy across 27 generators, surpassing both lay users and visual experts, and further approaching the human perceptual limit as pretrained backbones scale. The code is publicly available at: https://github.com/349793927/MIRROR

Method: Conducted static tests measuring detection range across distances (1-7m) and viewing angles (0-75° horizontally), and dynamic tests varying walking speeds (0.8-1.8 m/s) with three camera positions (head, shoulder, hand-held). Evaluated smartphone and smart glasses with phone’s main/ultra-wide cameras, benchmarking four OCR engines (Google Vision, PaddleOCR 3.0, EasyOCR, Tesseract) for static tests, then using PaddleOCR for dynamic tests. Accuracy computed at character level using Levenshtein ratio against ground truth.

Result: Recognition accuracy declined with increased walking speed and wider viewing angles. Google Vision achieved highest overall accuracy, with PaddleOCR as strongest open-source alternative. Phone’s main camera achieved highest accuracy, and shoulder-mounted placement yielded highest average among body positions, though differences among shoulder, head, and hand were not statistically significant.

Conclusion: Mobile OCR performance is significantly affected by motion and viewing conditions, with Google Vision performing best overall and shoulder-mounted positioning showing practical advantages for real-world assistive technology applications.

Abstract: Optical character recognition (OCR), which converts printed or handwritten text into machine-readable form, is widely used in assistive technology for people with blindness and low vision. Yet, most evaluations rely on static datasets that do not reflect the challenges of mobile use. In this study, we systematically evaluated OCR performance under both static and dynamic conditions. Static tests measured detection range across distances of 1-7 meters and viewing angles of 0-75 degrees horizontally. Dynamic tests examined the impact of motion by varying walking speed from slow (0.8 m/s) to very fast (1.8 m/s) and comparing three camera mounting positions: head-mounted, shoulder-mounted, and hand-held. We evaluated both a smartphone and smart glasses, using the phone’s main and ultra-wide cameras. Four OCR engines were benchmarked to assess accuracy at different distances and viewing angles: Google Vision, PaddleOCR 3.0, EasyOCR, and Tesseract. PaddleOCR 3.0 was then used to evaluate accuracy at different walking speeds. Accuracy was computed at the character level using the Levenshtein ratio against manually defined ground truth. Results showed that recognition accuracy declined with increased walking speed and wider viewing angles. Google Vision achieved the highest overall accuracy, with PaddleOCR close behind as the strongest open-source alternative. Across devices, the phone’s main camera achieved the highest accuracy, and a shoulder-mounted placement yielded the highest average among body positions; however, differences among shoulder, head, and hand were not statistically significant.

Method: LatentMorph introduces four lightweight components: (1) a condenser for summarizing intermediate generation states into compact visual memory, (2) a translator for converting latent thoughts into actionable guidance, (3) a shaper for dynamically steering next image token predictions, and (4) an RL-trained invoker for adaptively determining when to invoke reasoning. The framework performs reasoning entirely in continuous latent spaces.

Result: LatentMorph enhances base model Janus-Pro by 16% on GenEval and 25% on T2I-CompBench; outperforms explicit paradigms by 15% and 11% on abstract reasoning tasks like WISE and IPV-Txt; reduces inference time by 44% and token consumption by 51%; and exhibits 71% cognitive alignment with human intuition on reasoning invocation.

Conclusion: LatentMorph successfully bridges the gap in dynamic reasoning for text-to-image generation by performing implicit reasoning in continuous latent spaces, achieving significant performance improvements, efficiency gains, and better cognitive alignment compared to explicit reasoning paradigms.

Abstract: Text-to-image (T2I) generation has achieved remarkable progress, yet existing methods often lack the ability to dynamically reason and refine during generation–a hallmark of human creativity. Current reasoning-augmented paradigms most rely on explicit thought processes, where intermediate reasoning is decoded into discrete text at fixed steps with frequent image decoding and re-encoding, leading to inefficiencies, information loss, and cognitive mismatches. To bridge this gap, we introduce LatentMorph, a novel framework that seamlessly integrates implicit latent reasoning into the T2I generation process. At its core, LatentMorph introduces four lightweight components: (i) a condenser for summarizing intermediate generation states into compact visual memory, (ii) a translator for converting latent thoughts into actionable guidance, (iii) a shaper for dynamically steering next image token predictions, and (iv) an RL-trained invoker for adaptively determining when to invoke reasoning. By performing reasoning entirely in continuous latent spaces, LatentMorph avoids the bottlenecks of explicit reasoning and enables more adaptive self-refinement. Extensive experiments demonstrate that LatentMorph (I) enhances the base model Janus-Pro by $16%$ on GenEval and $25%$ on T2I-CompBench; (II) outperforms explicit paradigms (e.g., TwiG) by $15%$ and $11%$ on abstract reasoning tasks like WISE and IPV-Txt, (III) while reducing inference time by $44%$ and token consumption by $51%$; and (IV) exhibits $71%$ cognitive alignment with human intuition on reasoning invocation.

Method: Proposes flow matching framework with nearest neighbor flow matching loss and Chamfer distance loss to enhance both local structure and global coverage in point cloud alignment.

Result: LiFlow achieves state-of-the-art performance across multiple metrics for 3D LiDAR scene completion.

Conclusion: Flow matching provides superior approach to 3D LiDAR scene completion compared to diffusion methods, with better distribution consistency and performance.

Abstract: In autonomous driving scenarios, the collected LiDAR point clouds can be challenged by occlusion and long-range sparsity, limiting the perception of autonomous driving systems. Scene completion methods can infer the missing parts of incomplete 3D LiDAR scenes. Recent methods adopt local point-level denoising diffusion probabilistic models, which require predicting Gaussian noise, leading to a mismatch between training and inference initial distributions. This paper introduces the first flow matching framework for 3D LiDAR scene completion, improving upon diffusion-based methods by ensuring consistent initial distributions between training and inference. The model employs a nearest neighbor flow matching loss and a Chamfer distance loss to enhance both local structure and global coverage in the alignment of point clouds. LiFlow achieves state-of-the-art performance across multiple metrics. Code: https://github.com/matteandre/LiFlow.

Method: Proposes DiScene with two key innovations: 1) Multi-level Consistent Knowledge Distillation transferring hierarchical representations from teacher to student models across four levels (encoder, query, prior, anchor), and 2) Teacher-Guided Initialization for optimized parameter warm-up to accelerate convergence.

Result: Achieves 23.2 FPS without depth priors, outperforming baseline OPUS by 36.1% and even better than depth-enhanced OPUS†. With depth integration, DiScene† attains new SOTA, surpassing EmbodiedOcc by 3.7% with 1.62× faster inference speed. Shows versatility across Occ-Scannet, Occ3D-nuScenes benchmarks and in-the-wild scenarios.

Conclusion: DiScene provides an efficient and robust sparse query-based framework for occupancy prediction through multi-level distillation, achieving state-of-the-art performance with faster inference speeds across diverse environments.

Abstract: Occupancy prediction provides critical geometric and semantic understanding for robotics but faces efficiency-accuracy trade-offs. Current dense methods suffer computational waste on empty voxels, while sparse query-based approaches lack robustness in diverse and complex indoor scenes. In this paper, we propose DiScene, a novel sparse query-based framework that leverages multi-level distillation to achieve efficient and robust occupancy prediction. In particular, our method incorporates two key innovations: (1) a Multi-level Consistent Knowledge Distillation strategy, which transfers hierarchical representations from large teacher models to lightweight students through coordinated alignment across four levels, including encoder-level feature alignment, query-level feature matching, prior-level spatial guidance, and anchor-level high-confidence knowledge transfer and (2) a Teacher-Guided Initialization policy, employing optimized parameter warm-up to accelerate model convergence. Validated on the Occ-Scannet benchmark, DiScene achieves 23.2 FPS without depth priors while outperforming our baseline method, OPUS, by 36.1% and even better than the depth-enhanced version, OPUS†. With depth integration, DiScene† attains new SOTA performance, surpassing EmbodiedOcc by 3.7% with 1.62$\times$ faster inference speed. Furthermore, experiments on the Occ3D-nuScenes benchmark and in-the-wild scenarios demonstrate the versatility of our approach in various environments. Code and models can be accessed at https://github.com/getterupper/DiScene.

Method: Uses Residual Vector Quantized Variational Autoencoders (RVQ-VAEs) to learn coarse-to-fine motion representations. Integrates contrastive learning and a novel information leakage loss with codebook learning to organize content and style across different codebooks. Employs Quantized Code Swapping for inference-time style transfer without fine-tuning.

Result: The framework demonstrates strong versatility across multiple inference applications including style transfer, style removal, and motion blending. Enables motion style transfer without requiring fine-tuning for unseen styles.

Conclusion: The proposed method effectively disentangles style and content in human motion data, enabling flexible style manipulation through a simple inference-time technique that works without additional training for new styles.

Abstract: Human motion data is inherently rich and complex, containing both semantic content and subtle stylistic features that are challenging to model. We propose a novel method for effective disentanglement of the style and content in human motion data to facilitate style transfer. Our approach is guided by the insight that content corresponds to coarse motion attributes while style captures the finer, expressive details. To model this hierarchy, we employ Residual Vector Quantized Variational Autoencoders (RVQ-VAEs) to learn a coarse-to-fine representation of motion. We further enhance the disentanglement by integrating contrastive learning and a novel information leakage loss with codebook learning to organize the content and the style across different codebooks. We harness this disentangled representation using our simple and effective inference-time technique Quantized Code Swapping, which enables motion style transfer without requiring any fine-tuning for unseen styles. Our framework demonstrates strong versatility across multiple inference applications, including style transfer, style removal, and motion blending.

Method: Two-stage approach: Stage 1 synthesizes preference triples by anchoring questions to short clips with distractors, using visual-similarity and question-specificity filtering. Stage 2 uses recursive captioning on long videos to generate scene-level metadata, then employs LLMs to craft multi-segment reasoning queries and dispreferred responses for alignment.

Result: LongVPO outperforms state-of-the-art open-source models on multiple long-video benchmarks while maintaining strong short-video performance on MVBench, achieving this with only 16K synthetic examples and no human labels.

Conclusion: LongVPO provides a scalable paradigm for efficient long-form video understanding that extends short-context models to handle ultra-long videos without costly annotations, demonstrating strong performance with minimal synthetic data.

Abstract: We present LongVPO, a novel two-stage Direct Preference Optimization framework that enables short-context vision-language models to robustly understand ultra-long videos without any long-video annotations. In Stage 1, we synthesize preference triples by anchoring questions to individual short clips, interleaving them with distractors, and applying visual-similarity and question-specificity filtering to mitigate positional bias and ensure unambiguous supervision. We also approximate the reference model’s scoring over long contexts by evaluating only the anchor clip, reducing computational overhead. In Stage 2, we employ a recursive captioning pipeline on long videos to generate scene-level metadata, then use a large language model to craft multi-segment reasoning queries and dispreferred responses, aligning the model’s preferences through multi-segment reasoning tasks. With only 16K synthetic examples and no costly human labels, LongVPO outperforms the state-of-the-art open-source models on multiple long-video benchmarks, while maintaining strong short-video performance (e.g., on MVBench), offering a scalable paradigm for efficient long-form video understanding.

Method: Designs different neural networks as texture INRs that operate continuously over UV coordinate space, conducting experiments to evaluate image quality, memory usage, and rendering inference time trade-offs.

Result: Demonstrates that INRs perform well in terms of image quality while offering considerable memory usage and rendering inference time benefits, with analysis of balance between these objectives.

Conclusion: INRs provide an effective continuous texture representation with good performance trade-offs, enabling various applications in real-time rendering and downstream tasks like mipmap fitting and INR-space generation.

Abstract: Implicit neural representation (INR) has proven to be accurate and efficient in various domains. In this work, we explore how different neural networks can be designed as a new texture INR, which operates in a continuous manner rather than a discrete one over the input UV coordinate space. Through thorough experiments, we demonstrate that these INRs perform well in terms of image quality, with considerable memory usage and rendering inference time. We analyze the balance between these objectives. In addition, we investigate various related applications in real-time rendering and down-stream tasks, e.g. mipmap fitting and INR-space generation.

Method: Proposes Neural Adaptive Binning (NAB) that maps coordinate space to binned vector space using shifted hyperbolic tangent functions with learnable position, size, steepness, and rotation parameters. The binned representations are processed by a neural network to predict CT attenuation coefficients, enabling end-to-end optimization.

Result: NAB achieves superior performance on two industrial datasets and maintains robustness on medical datasets when extended to more general expressions. The method effectively integrates shape priors to enhance reconstruction accuracy.

Conclusion: NAB provides a new perspective on integrating shape priors into neural network-based reconstruction, offering improved sparse-view CT reconstruction for industrial applications while maintaining generalization capability.

Abstract: Computed Tomography (CT) plays a vital role in inspecting the internal structures of industrial objects. Furthermore, achieving high-quality CT reconstruction from sparse views is essential for reducing production costs. While classic implicit neural networks have shown promising results for sparse reconstruction, they are unable to leverage shape priors of objects. Motivated by the observation that numerous industrial objects exhibit rectangular structures, we propose a novel \textbf{N}eural \textbf{A}daptive \textbf{B}inning (\textbf{NAB}) method that effectively integrates rectangular priors into the reconstruction process. Specifically, our approach first maps coordinate space into a binned vector space. This mapping relies on an innovative binning mechanism based on differences between shifted hyperbolic tangent functions, with our extension enabling rotations around the input-plane normal vector. The resulting representations are then processed by a neural network to predict CT attenuation coefficients. This design enables end-to-end optimization of the encoding parameters – including position, size, steepness, and rotation – via gradient flow from the projection data, thus enhancing reconstruction accuracy. By adjusting the smoothness of the binning function, NAB can generalize to objects with more complex geometries. This research provides a new perspective on integrating shape priors into neural network-based reconstruction. Extensive experiments demonstrate that NAB achieves superior performance on two industrial datasets. It also maintains robust on medical datasets when the binning function is extended to more general expression. The code will be made available.

Method: Implement Spectral-normalized Neural Gaussian Process (SNGP) - lightweight modifications applying spectral normalization and replacing final dense layer with Gaussian process layer to improve single-model uncertainty estimation and OOD detection.

Result: SNGP shows comparable in-distribution performance to deterministic and Monte Carlo dropout methods while significantly improving uncertainty estimation and OOD detection across six datasets and three biomedical classification tasks.

Conclusion: SNGP offers a useful framework for uncertainty-aware classification in digital pathology, supporting safe deployment and building trust with pathologists.

Abstract: Accurate histopathologic interpretation is key for clinical decision-making; however, current deep learning models for digital pathology are often overconfident and poorly calibrated in out-of-distribution (OOD) settings, which limit trust and clinical adoption. Safety-critical medical imaging workflows benefit from intrinsic uncertainty-aware properties that can accurately reject OOD input. We implement the Spectral-normalized Neural Gaussian Process (SNGP), a set of lightweight modifications that apply spectral normalization and replace the final dense layer with a Gaussian process layer to improve single-model uncertainty estimation and OOD detection. We evaluate SNGP vs. deterministic and MonteCarlo dropout on six datasets across three biomedical classification tasks: white blood cells, amyloid plaques, and colorectal histopathology. SNGP has comparable in-distribution performance while significantly improving uncertainty estimation and OOD detection. Thus, SNGP or related models offer a useful framework for uncertainty-aware classification in digital pathology, supporting safe deployment and building trust with pathologists.

Method: Two-stage training: (1) Distill structured reasoning traces from advanced VLMs for supervised fine-tuning to enable flexible, context-adaptive reasoning; (2) Perform direct preference optimization on curated preference pairs to strengthen reasoning fidelity and discriminative alignment. The model interprets semantic intent, grounds on visual evidence, and dynamically constructs hierarchical assessments with fine-grained criteria.

Result: When integrated into GRPO framework for image and video synthesis, UnifiedReward-Flex demonstrates superiority over existing approaches, showing improved alignment with human preferences through personalized, context-aware assessment.

Conclusion: The proposed unified personalized reward model successfully addresses limitations of current reward models by incorporating flexible, context-adaptive reasoning inspired by human assessment, leading to better alignment with subjective human preferences in visual generation tasks.

Abstract: Recent advancements in multimodal reward models (RMs) have significantly propelled the development of visual generation. Existing frameworks typically adopt Bradley-Terry-style preference modeling or leverage generative VLMs as judges, and subsequently optimize visual generation models via reinforcement learning. However, current RMs suffer from inherent limitations: they often follow a one-size-fits-all paradigm that assumes a monolithic preference distribution or relies on fixed evaluation rubrics. As a result, they are insensitive to content-specific visual cues, leading to systematic misalignment with subjective and context-dependent human preferences. To this end, inspired by human assessment, we propose UnifiedReward-Flex, a unified personalized reward model for vision generation that couples reward modeling with flexible and context-adaptive reasoning. Specifically, given a prompt and the generated visual content, it first interprets the semantic intent and grounds on visual evidence, then dynamically constructs a hierarchical assessment by instantiating fine-grained criteria under both predefined and self-generated high-level dimensions. Our training pipeline follows a two-stage process: (1) we first distill structured, high-quality reasoning traces from advanced closed-source VLMs to bootstrap SFT, equipping the model with flexible and context-adaptive reasoning behaviors; (2) we then perform direct preference optimization (DPO) on carefully curated preference pairs to further strengthen reasoning fidelity and discriminative alignment. To validate the effectiveness, we integrate UnifiedReward-Flex into the GRPO framework for image and video synthesis, and extensive results demonstrate its superiority.

Method: MultiBO (Multi-Choice Preferential Bayesian Optimization) generates K new images based on an initial prompt-generated image, collects preferential feedback from users on which images are closer to their mental image, uses Bayesian optimization to guide the diffusion model, and iteratively refines the images over B rounds of feedback.

Result: The method shows promising results with qualitative scores from 30 users and quantitative metrics compared across 5 baselines. Users can arrive much closer to their mental images within B rounds of feedback, even though the generative model has no direct information about the target image.

Conclusion: Multi-choice feedback from humans can be effectively harnessed for personalized image generation, bridging the gap between language-prompted results and users’ specific mental images through iterative preferential optimization.

Abstract: Imagine Alice has a specific image $x^\ast$ in her mind, say, the view of the street in which she grew up during her childhood. To generate that exact image, she guides a generative model with multiple rounds of prompting and arrives at an image $x^{p*}$. Although $x^{p*}$ is reasonably close to $x^\ast$, Alice finds it difficult to close that gap using language prompts. This paper aims to narrow this gap by observing that even after language has reached its limits, humans can still tell when a new image $x^+$ is closer to $x^\ast$ than $x^{p*}$. Leveraging this observation, we develop MultiBO (Multi-Choice Preferential Bayesian Optimization) that carefully generates $K$ new images as a function of $x^{p*}$, gets preferential feedback from the user, uses the feedback to guide the diffusion model, and ultimately generates a new set of $K$ images. We show that within $B$ rounds of user feedback, it is possible to arrive much closer to $x^\ast$, even though the generative model has no information about $x^\ast$. Qualitative scores from $30$ users, combined with quantitative metrics compared across $5$ baselines, show promising results, suggesting that multi-choice feedback from humans can be effectively harnessed for personalized image generation.

Method: 1) Hierarchical Pose-free Memory Compressor (HPMC) that recursively distills historical latents into fixed-budget representation; 2) Uncertainty-aware Action Labeling that discretizes continuous motion into tri-state logic; 3) Revisit-Dense Finetuning Strategy using compact dataset to activate long-range loop-closure capabilities.

Result: Extensive experiments show Infinite-World achieves superior performance in visual quality, action controllability, and spatial consistency compared to existing methods, demonstrating robust long-term memory capabilities.

Conclusion: Infinite-World successfully bridges the gap between synthetic and real-world video training for world models, enabling coherent visual memory over extended sequences without explicit geometric priors.

Abstract: We propose Infinite-World, a robust interactive world model capable of maintaining coherent visual memory over 1000+ frames in complex real-world environments. While existing world models can be efficiently optimized on synthetic data with perfect ground-truth, they lack an effective training paradigm for real-world videos due to noisy pose estimations and the scarcity of viewpoint revisits. To bridge this gap, we first introduce a Hierarchical Pose-free Memory Compressor (HPMC) that recursively distills historical latents into a fixed-budget representation. By jointly optimizing the compressor with the generative backbone, HPMC enables the model to autonomously anchor generations in the distant past with bounded computational cost, eliminating the need for explicit geometric priors. Second, we propose an Uncertainty-aware Action Labeling module that discretizes continuous motion into a tri-state logic. This strategy maximizes the utilization of raw video data while shielding the deterministic action space from being corrupted by noisy trajectories, ensuring robust action-response learning. Furthermore, guided by insights from a pilot toy study, we employ a Revisit-Dense Finetuning Strategy using a compact, 30-minute dataset to efficiently activate the model’s long-range loop-closure capabilities. Extensive experiments, including objective metrics and user studies, demonstrate that Infinite-World achieves superior performance in visual quality, action controllability, and spatial consistency.

Method: Two-fold approach: 1) Vision-Guided Motion Tokenizer that leverages geometric alignment between 3D skeletons and visual data for joint learning from both modalities, creating unified cross-modal motion vocabulary; 2) Single unified MLLM architecture trained to handle all tasks, processing diverse temporal inputs for both perception (3D skeleton pose estimation from video) and generation (motion prediction and in-betweening).

Result: Extensive experiments on standard benchmarks including Human3.6M demonstrate state-of-the-art or competitive performance across all motion tasks, showing efficient and scalable path for generative motion analysis using skeletons.

Conclusion: Superman successfully bridges visual perception with temporal skeleton-based motion generation, creating a unified framework that addresses fragmentation in human motion analysis and enables more efficient, scalable generative motion analysis.

Abstract: Human motion analysis tasks, such as temporal 3D pose estimation, motion prediction, and motion in-betweening, play an essential role in computer vision. However, current paradigms suffer from severe fragmentation. First, the field is split between perception'' models that understand motion from video but only output text, and generation’’ models that cannot perceive from raw visual input. Second, generative MLLMs are often limited to single-frame, static poses using dense, parametric SMPL models, failing to handle temporal motion. Third, existing motion vocabularies are built from skeleton data alone, severing the link to the visual domain. To address these challenges, we introduce Superman, a unified framework that bridges visual perception with temporal, skeleton-based motion generation. Our solution is twofold. First, to overcome the modality disconnect, we propose a Vision-Guided Motion Tokenizer. Leveraging the natural geometric alignment between 3D skeletons and visual data, this module pioneers robust joint learning from both modalities, creating a unified, cross-modal motion vocabulary. Second, grounded in this motion language, a single, unified MLLM architecture is trained to handle all tasks. This module flexibly processes diverse, temporal inputs, unifying 3D skeleton pose estimation from video (perception) with skeleton-based motion prediction and in-betweening (generation). Extensive experiments on standard benchmarks, including Human3.6M, demonstrate that our unified method achieves state-of-the-art or competitive performance across all motion tasks. This showcases a more efficient and scalable path for generative motion analysis using skeletons.

Method: Proposes ReasonEdit which continuously stores human reasoning in a codebook and retrieves relevant facts during inference using a novel topology-balanced multimodal embedding method inspired by network science.

Result: Achieves state-of-the-art editing performance across four VLMs on multiple rationale-based visual question answering datasets, demonstrating that using human reasoning during editing greatly improves edit generalization.

Conclusion: ReasonEdit introduces a practical model editing setup that successfully incorporates human reasoning, showing significant improvements in edit generalization for vision-language models on reasoning-heavy tasks.

Abstract: Model editing aims to correct errors in large, pretrained models without altering unrelated behaviors. While some recent works have edited vision-language models (VLMs), no existing editors tackle reasoning-heavy tasks, which typically require humans and models to reason about images.We therefore propose ReasonEdit, the first VLM editor to let users explain their reasoning during editing, introducing a new, practical model editing setup. ReasonEdit continuously stores human reasoning in a codebook, and retrieves only relevant facts during inference using a novel topology-balanced multimodal embedding method inspired by network science. Across four VLMs on multiple rationale-based visual question answering datasets, ReasonEdit achieves state-of-the-art editing performance, ultimately showing that using human reasoning during editing greatly improves edit generalization.

Method: Catalyst computes an input-dependent scaling factor (γ) on-the-fly from raw pre-pooling feature map statistics (mean, standard deviation, maximum activation). This γ is then fused multiplicatively with existing baseline OOD scores (energy, ReAct, SCALE, KNN) to perform “elastic scaling” that pushes ID and OOD distributions further apart.

Result: Catalyst achieves substantial performance gains, reducing average False Positive Rate by 32.87% on CIFAR-10 (ResNet-18), 27.94% on CIFAR-100 (ResNet-18), and 22.25% on ImageNet (ResNet-50). It works with both logit-based and distance-based detectors.

Conclusion: Pre-pooling feature map statistics contain untapped potential for OOD detection, and Catalyst provides a generalizable framework that complements existing approaches by exploiting these under-explored signals.

Abstract: Out-of-distribution (OOD) detection is critical for the safe deployment of deep neural networks. State-of-the-art post-hoc methods typically derive OOD scores from the output logits or penultimate feature vector obtained via global average pooling (GAP). We contend that this exclusive reliance on the logit or feature vector discards a rich, complementary signal: the raw channel-wise statistics of the pre-pooling feature map lost in GAP. In this paper, we introduce Catalyst, a post-hoc framework that exploits these under-explored signals. Catalyst computes an input-dependent scaling factor ($γ$) on-the-fly from these raw statistics (e.g., mean, standard deviation, and maximum activation). This $γ$ is then fused with the existing baseline score, multiplicatively modulating it – an ``elastic scaling’’ – to push the ID and OOD distributions further apart. We demonstrate Catalyst is a generalizable framework: it seamlessly integrates with logit-based methods (e.g., Energy, ReAct, SCALE) and also provides a significant boost to distance-based detectors like KNN. As a result, Catalyst achieves substantial and consistent performance gains, reducing the average False Positive Rate by 32.87 on CIFAR-10 (ResNet-18), 27.94% on CIFAR-100 (ResNet-18), and 22.25% on ImageNet (ResNet-50). Our results highlight the untapped potential of pre-pooling statistics and demonstrate that Catalyst is complementary to existing OOD detection approaches.

Method: Created SelvaMask dataset with 8,800 manually delineated tree crowns across three Neotropical sites. Developed modular detection-segmentation pipeline that adapts vision foundation models using domain-specific detection-prompter.

Result: Achieved state-of-the-art performance, outperforming both zero-shot generalist models and fully supervised end-to-end methods in dense tropical forests. Validated gains on external tropical and temperate datasets.

Conclusion: SelvaMask serves as both a challenging benchmark and key enabler for generalized forest monitoring, with the dataset and code to be released publicly.

Abstract: Tropical forests harbor most of the planet’s tree biodiversity and are critical to global ecological balance. Canopy trees in particular play a disproportionate role in carbon storage and functioning of these ecosystems. Studying canopy trees at scale requires accurate delineation of individual tree crowns, typically performed using high-resolution aerial imagery. Despite advances in transformer-based models for individual tree crown segmentation, performance remains low in most forests, especially tropical ones. To this end, we introduce SelvaMask, a new tropical dataset containing over 8,800 manually delineated tree crowns across three Neotropical forest sites in Panama, Brazil, and Ecuador. SelvaMask features comprehensive annotations, including an inter-annotator agreement evaluation, capturing the dense structure of tropical forests and highlighting the difficulty of the task. Leveraging this benchmark, we propose a modular detection-segmentation pipeline that adapts vision foundation models (VFMs), using domain-specific detection-prompter. Our approach reaches state-of-the-art performance, outperforming both zero-shot generalist models and fully supervised end-to-end methods in dense tropical forests. We validate these gains on external tropical and temperate datasets, demonstrating that SelvaMask serves as both a challenging benchmark and a key enabler for generalized forest monitoring. Our code and dataset will be released publicly.

Method: Proposes a dual reasoning paradigm: 1) generation as world knowledge-enhanced planning with implicit constraints, and 2) editing for fine-grained visual refinement via self-reflection. Unifies both tasks within a shared representation, mirroring human cognitive process of planning followed by refinement. Constructs a large-scale reasoning-centric dataset (~300k samples) covering five major knowledge domains for planning, plus agent-generated corpus for visual self-correction.

Result: Extensive experiments show UniReason achieves advanced performance on reasoning-intensive benchmarks (WISE, KrisBench, UniREditBench) while maintaining superior general synthesis capabilities.

Conclusion: UniReason successfully unifies generation and editing through a reasoning-centric approach, demonstrating improved performance on complex multimodal synthesis tasks that require deep reasoning.

Abstract: Unified multimodal models often struggle with complex synthesis tasks that demand deep reasoning, and typically treat text-to-image generation and image editing as isolated capabilities rather than interconnected reasoning steps. To address this, we propose UniReason, a unified framework that harmonizes these two tasks through a dual reasoning paradigm. We formulate generation as world knowledge-enhanced planning to inject implicit constraints, and leverage editing capabilities for fine-grained visual refinement to further correct visual errors via self-reflection. This approach unifies generation and editing within a shared representation, mirroring the human cognitive process of planning followed by refinement. We support this framework by systematically constructing a large-scale reasoning-centric dataset (~300k samples) covering five major knowledge domains (e.g., cultural commonsense, physics, etc.) for planning, alongside an agent-generated corpus for visual self-correction. Extensive experiments demonstrate that UniReason achieves advanced performance on reasoning-intensive benchmarks such as WISE, KrisBench and UniREditBench, while maintaining superior general synthesis capabilities.

Method: Proposes a gated multi-head Transformer architecture based on Swin U-Net with inter-slice context integration and a parallel detection head. The model jointly performs slice-level structure detection via MLP and pixel-level segmentation through a context-enhanced stream, using detection outputs to gate segmentation predictions.

Result: The gated model substantially outperforms non-gated baseline, achieving mean Dice loss of 0.013 ± 0.036 vs 0.732 ± 0.314 on Prostate-Anatomical-Edge-Cases dataset. Detection probabilities strongly correlated with anatomical presence, effectively eliminating spurious segmentations.

Conclusion: Detection-based gating enhances robustness and anatomical plausibility in automated segmentation, reducing hallucinated predictions without compromising segmentation quality in valid slices, offering promise for improving clinical radiotherapy auto-contouring reliability.

Abstract: Deep learning based auto segmentation is increasingly used in radiotherapy, but conventional models often produce anatomically implausible false positives, or hallucinations, in slices lacking target structures. We propose a gated multi-head Transformer architecture based on Swin U-Net, augmented with inter-slice context integration and a parallel detection head, which jointly performs slice-level structure detection via a multi-layer perceptron and pixel-level segmentation through a context-enhanced stream. Detection outputs gate the segmentation predictions to suppress false positives in anatomically invalid slices, and training uses slice-wise Tversky loss to address class imbalance. Experiments on the Prostate-Anatomical-Edge-Cases dataset from The Cancer Imaging Archive demonstrate that the gated model substantially outperforms a non-gated segmentation-only baseline, achieving a mean Dice loss of $0.013 \pm 0.036$ versus $0.732 \pm 0.314$, with detection probabilities strongly correlated with anatomical presence, effectively eliminating spurious segmentations. In contrast, the non-gated model exhibited higher variability and persistent false positives across all slices. These results indicate that detection-based gating enhances robustness and anatomical plausibility in automated segmentation applications, reducing hallucinated predictions without compromising segmentation quality in valid slices, and offers a promising approach for improving the reliability of clinical radiotherapy auto-contouring workflows.

Method: Introduces two complementary perceptual losses: LPIPS loss for better local patterns and DINO-based perceptual loss for stronger global semantics. This guides diffusion models to learn meaningful perceptual manifolds rather than full image manifolds.

Result: Achieves FID of 5.11 on ImageNet-256 without classifier-free guidance using only 80 training epochs. Demonstrates favorable scaling on large-scale text-to-image generation with GenEval score of 0.79. Surpasses strong latent diffusion baselines.

Conclusion: PixelGen provides a simpler yet more powerful generative paradigm that requires no VAEs, latent representations, or auxiliary stages, offering an end-to-end pixel diffusion approach with perceptual supervision.

Abstract: Pixel diffusion generates images directly in pixel space in an end-to-end manner, avoiding the artifacts and bottlenecks introduced by VAEs in two-stage latent diffusion. However, it is challenging to optimize high-dimensional pixel manifolds that contain many perceptually irrelevant signals, leaving existing pixel diffusion methods lagging behind latent diffusion models. We propose PixelGen, a simple pixel diffusion framework with perceptual supervision. Instead of modeling the full image manifold, PixelGen introduces two complementary perceptual losses to guide diffusion model towards learning a more meaningful perceptual manifold. An LPIPS loss facilitates learning better local patterns, while a DINO-based perceptual loss strengthens global semantics. With perceptual supervision, PixelGen surpasses strong latent diffusion baselines. It achieves an FID of 5.11 on ImageNet-256 without classifier-free guidance using only 80 training epochs, and demonstrates favorable scaling performance on large-scale text-to-image generation with a GenEval score of 0.79. PixelGen requires no VAEs, no latent representations, and no auxiliary stages, providing a simpler yet more powerful generative paradigm. Codes are publicly available at https://github.com/Zehong-Ma/PixelGen.

Method: Proposes a new artwork explanation generation task with evaluation dataset and metrics. The task has two parts: 1) generating explanations from images and titles, 2) generating explanations using only images. Also releases training dataset for LVLMs to learn artwork explanations.

Result: LVLMs struggle with integrating language and visual information, and show even more pronounced limitations in acquiring knowledge from images alone.

Conclusion: Artwork explanation generation is a suitable task for evaluating LVLMs’ multimodal knowledge integration capabilities, revealing significant current limitations in both vision-language integration and visual knowledge acquisition.

Abstract: Large-scale Vision-Language Models (LVLMs) output text from images and instructions, demonstrating capabilities in text generation and comprehension. However, it has not been clarified to what extent LVLMs possess the ability to understand the knowledge necessary for explaining images, the complex relationships between various pieces of knowledge, and how they integrate these understandings into their explanations. To address this issue, we propose a new task: the artwork explanation generation task, along with its evaluation dataset and metrics for quantitatively assessing the understanding and utilization of knowledge about artworks. This task is apt for image description based on the premise that LVLMs are expected to have pre-existing knowledge of artworks, which are often subjects of wide recognition and documented information. It consists of two parts: generating explanations from images and titles of artworks, and generating explanations using only images, thus evaluating the LVLMs’ language-based and vision-based knowledge. Alongside, we release a training dataset for LVLMs to learn explanations that incorporate knowledge about artworks. Our findings indicate that LVLMs not only struggle with integrating language and visual information but also exhibit a more pronounced limitation in acquiring knowledge from images alone.

Method: Decouples interaction semantics and dynamics. Uses pre-trained LLMs for high-level semantic control and text-to-motion models for human motion. Introduces a world model to understand simple physics and object motion influenced by human actions. Integrates these into InterDreamer framework for zero-shot generation.

Result: Applied to BEHAVE and CHAIRS datasets, InterDreamer generates realistic and coherent 3D HOI sequences that align with text directives in a zero-shot manner, demonstrating capability without direct text-interaction training data.

Conclusion: Shows potential for generating human-object interactions without text-interaction pair data by decoupling semantics and dynamics, leveraging pre-trained models for semantics and introducing physics-aware world model for dynamics.

Abstract: Text-conditioned human motion generation has experienced significant advancements with diffusion models trained on extensive motion capture data and corresponding textual annotations. However, extending such success to 3D dynamic human-object interaction (HOI) generation faces notable challenges, primarily due to the lack of large-scale interaction data and comprehensive descriptions that align with these interactions. This paper takes the initiative and showcases the potential of generating human-object interactions without direct training on text-interaction pair data. Our key insight in achieving this is that interaction semantics and dynamics can be decoupled. Being unable to learn interaction semantics through supervised training, we instead leverage pre-trained large models, synergizing knowledge from a large language model and a text-to-motion model. While such knowledge offers high-level control over interaction semantics, it cannot grasp the intricacies of low-level interaction dynamics. To overcome this issue, we further introduce a world model designed to comprehend simple physics, modeling how human actions influence object motion. By integrating these components, our novel framework, InterDreamer, is able to generate text-aligned 3D HOI sequences in a zero-shot manner. We apply InterDreamer to the BEHAVE and CHAIRS datasets, and our comprehensive experimental analysis demonstrates its capability to generate realistic and coherent interaction sequences that seamlessly align with the text directives.

Method: Three-step approach: 1) Train parent architecture for early exiting from encoder, 2) Create dataset of ideal encoder layers per training example, 3) Train gating network to predict optimal layer count conditioned on input image.

Result: Reduces expected encoder computational cost while maintaining performance, adapts to various compute resources, offers architectural flexibility, and extends beyond segmentation to object detection.

Conclusion: ECO-M2F provides an efficient solution for transformer-based segmentation models by enabling input-adaptive computation, balancing performance and efficiency for practical deployment.

Abstract: Vision transformer based models bring significant improvements for image segmentation tasks. Although these architectures offer powerful capabilities irrespective of specific segmentation tasks, their use of computational resources can be taxing on deployed devices. One way to overcome this challenge is by adapting the computation level to the specific needs of the input image rather than the current one-size-fits-all approach. To this end, we introduce ECO-M2F or EffiCient TransfOrmer Encoders for Mask2Former-style models. Noting that the encoder module of M2F-style models incur high resource-intensive computations, ECO-M2F provides a strategy to self-select the number of hidden layers in the encoder, conditioned on the input image. To enable this self-selection ability for providing a balance between performance and computational efficiency, we present a three step recipe. The first step is to train the parent architecture to enable early exiting from the encoder. The second step is to create an derived dataset of the ideal number of encoder layers required for each training example. The third step is to use the aforementioned derived dataset to train a gating network that predicts the number of encoder layers to be used, conditioned on the input image. Additionally, to change the computational-accuracy tradeoff, only steps two and three need to be repeated which significantly reduces retraining time. Experiments on the public datasets show that the proposed approach reduces expected encoder computational cost while maintaining performance, adapts to various user compute resources, is flexible in architecture configurations, and can be extended beyond the segmentation task to object detection.

Method: Proposes CrackSCF with: 1) Lightweight convolutional block (LRDS) to replace standard convolutions for local pattern capture, 2) Long-range Dependency Extractor (LDE) for global dependencies, 3) Staircase Cascaded Fusion Module (SCFM) to intelligently unify local and global features.

Result: Outperforms existing methods across six datasets including new TUT benchmark. Achieves 0.8382 F1 score and 0.8473 mIoU on TUT dataset with only 4.79M parameters. Shows robustness against complex background noise.

Conclusion: CrackSCF provides robust crack segmentation with exceptional computational efficiency, making it suitable for practical deployment on resource-constrained edge devices while maintaining high accuracy.

Abstract: Accurately segmenting structural cracks at the pixel level remains a major hurdle, as existing methods fail to integrate local textures with pixel dependencies, often leading to fragmented and incomplete predictions. Moreover, their high parameter counts and substantial computational demands hinder practical deployment on resource-constrained edge devices. To address these challenges, we propose CrackSCF, a Lightweight Cascaded Fusion Crack Segmentation Network designed to achieve robust crack segmentation with exceptional computational efficiency. We design a lightweight convolutional block (LRDS) to replace all standard convolutions. This approach efficiently captures local patterns while operating with a minimal computational footprint. For a holistic perception of crack structures, a lightweight Long-range Dependency Extractor (LDE) captures global dependencies. These are then intelligently unified with local patterns by our Staircase Cascaded Fusion Module (SCFM), ensuring the final segmentation maps are both seamless in continuity and rich in fine-grained detail. To comprehensively evaluate our method, this paper created the challenging TUT benchmark dataset and evaluated it alongside five other public datasets. The experimental results show that the CrackSCF method consistently outperforms the existing methods, and it demonstrates greater robustness in dealing with complex background noise. On the TUT dataset, CrackSCF achieved 0.8382 on F1 score and 0.8473 on mIoU, and it only required 4.79M parameters.

Method: Uses DETR-like detectors for per-camera detections, maintains global track embeddings that integrate local detection information, and employs probabilistic association with differentiable losses for end-to-end training.

Result: Validated on MMPTrack and AI City Challenge datasets, demonstrating effectiveness for multi-camera multi-object tracking with overlapping fields of view.

Conclusion: MCTR provides a novel end-to-end transformer framework that addresses the limitations of heuristic methods in multi-camera tracking through differentiable probabilistic association.

Abstract: Multi-camera tracking plays a pivotal role in various real-world applications. While end-to-end methods have gained significant interest in single-camera tracking, multi-camera tracking remains predominantly reliant on heuristic techniques. In response to this gap, this paper introduces Multi-Camera Tracking tRansformer (MCTR), a novel end-to-end approach tailored for multi-object detection and tracking across multiple cameras with overlapping fields of view. MCTR leverages end-to-end detectors like DEtector TRansformer (DETR) to produce detections and detection embeddings independently for each camera view. The framework maintains set of track embeddings that encaplusate global information about the tracked objects, and updates them at every frame by integrating the local information from the view-specific detection embeddings. The track embeddings are probabilistically associated with detections in every camera view and frame to generate consistent object tracks. The soft probabilistic association facilitates the design of differentiable losses that enable end-to-end training of the entire system. To validate our approach, we conduct experiments on MMPTrack and AI City Challenge, two recently introduced large-scale multi-camera multi-object tracking datasets.

Method: Proposes EgoFSD with three components: 1) Sparse perception module for detection and online mapping using sparse representations, 2) Hierarchical interaction module for selecting closest in-path vehicles/stationary objects with geometric priors, 3) Iterative motion planner with joint motion prediction and multi-modal trajectory optimization using position-level motion diffusion and trajectory-level planning denoising.

Result: Significantly reduces average L2 error by 59% and collision rate by 92% compared to UniAD while achieving 6.9x faster running efficiency on nuScenes and Bench2Drive datasets.

Conclusion: EgoFSD demonstrates superior performance and efficiency for end-to-end autonomous driving through ego-centric sparse representation and diffusion-based uncertainty modeling.

Abstract: Current End-to-End Autonomous Driving (E2E-AD) methods resort to unifying modular designs for various tasks (e.g. perception, prediction and planning). Although optimized with a fully differentiable framework in a planning-oriented manner, existing end-to-end driving systems lacking ego-centric designs still suffer from unsatisfactory performance and inferior efficiency, due to rasterized scene representation learning and redundant information transmission. In this paper, we propose an ego-centric fully sparse paradigm, named EgoFSD, for end-to-end self-driving. Specifically, EgoFSD consists of sparse perception, hierarchical interaction and iterative motion planner. The sparse perception module performs detection and online mapping based on sparse representation of the driving scene. The hierarchical interaction module aims to select the Closest In-Path Vehicle / Stationary (CIPV / CIPS) from coarse to fine, benefiting from an additional geometric prior. As for the iterative motion planner, both selected interactive agents and ego-vehicle are considered for joint motion prediction, where the output multi-modal ego-trajectories are optimized in an iterative fashion. In addition, position-level motion diffusion and trajectory-level planning denoising are introduced for uncertainty modeling, thereby enhancing the training stability and convergence speed. Extensive experiments are conducted on nuScenes and Bench2Drive datasets, which significantly reduces the average L2 error by 59% and collision rate by 92% than UniAD while achieves 6.9x faster running efficiency.

Method: Comprehensive review paper analyzing three main PAI implementations: photoacoustic computed tomography (PACT), photoacoustic microscopy (PAM), and photoacoustic endoscopy (PAE). Examines conventional and deep learning methodologies for image reconstruction and artifact mitigation, plus quantitative analysis techniques.

Result: The review demonstrates that deep learning approaches show considerable potential for enhancing image quality and accelerating imaging processes in PAI. Recent developments in quantitative analysis enable measurement of physiological parameters like hemoglobin concentration and oxygen saturation.

Conclusion: Photoacoustic imaging has promising clinical potential but faces implementation challenges. Deep learning is transformative for advancing PAI capabilities in reconstruction, artifact reduction, and quantitative analysis. Future research should focus on addressing current limitations and expanding clinical applications.

Abstract: Photoacoustic imaging (PAI) represents an innovative biomedical imaging modality that harnesses the advantages of optical resolution and acoustic penetration depth while ensuring enhanced safety. Despite its promising potential across a diverse array of preclinical and clinical applications, the clinical implementation of PAI faces significant challenges, including the trade-off between penetration depth and spatial resolution, as well as the demand for faster imaging speeds. This paper explores the fundamental principles underlying PAI, with a particular emphasis on three primary implementations: photoacoustic computed tomography (PACT), photoacoustic microscopy (PAM), and photoacoustic endoscopy (PAE). We undertake a critical assessment of their respective strengths and practical limitations. Furthermore, recent developments in utilizing conventional or deep learning (DL) methodologies for image reconstruction and artefact mitigation across PACT, PAM, and PAE are outlined, demonstrating considerable potential to enhance image quality and accelerate imaging processes. Furthermore, this paper examines the recent developments in quantitative analysis within PAI, including the quantification of haemoglobin concentration, oxygen saturation, and other physiological parameters within tissues. Finally, our discussion encompasses current trends and future directions in PAI research while emphasizing the transformative impact of deep learning on advancing PAI.

Method: EdgeCape introduces two key innovations: 1) predicting edge weights in the pose graph to optimize localization, and 2) Markovian Structural Bias that modulates self-attention between nodes based on hop distance to capture global spatial dependencies.

Result: Evaluated on MP-100 benchmark (100 categories, 20K+ images), EdgeCape achieves state-of-the-art results in 1-shot setting and leads among similar-sized methods in 5-shot setting, significantly improving keypoint localization accuracy.

Conclusion: Dynamic edge weight prediction and Markovian structural bias effectively improve category-agnostic pose estimation by better handling occlusions and capturing spatial dependencies, advancing the state-of-the-art in few-shot pose estimation.

Abstract: Category-Agnostic Pose Estimation (CAPE) localizes keypoints across diverse object categories with a single model, using one or a few annotated support images. Recent works have shown that using a pose graph (i.e., treating keypoints as nodes in a graph rather than isolated points) helps handle occlusions and break symmetry. However, these methods assume a static pose graph with equal-weight edges, leading to suboptimal results. We introduce EdgeCape, a novel framework that overcomes these limitations by predicting the graph’s edge weights which optimizes localization. To further leverage structural priors, we propose integrating Markovian Structural Bias, which modulates the self-attention interaction between nodes based on the number of hops between them. We show that this improves the model’s ability to capture global spatial dependencies. Evaluated on the MP-100 benchmark, which includes 100 categories and over 20K images, EdgeCape achieves state-of-the-art results in the 1-shot setting and leads among similar-sized methods in the 5-shot setting, significantly improving keypoint localization accuracy. Our code is publicly available.

Method: Feat2GS reads out 3D Gaussian attributes (geometry: position, opacity, covariance; texture: color) from VFM features extracted from unposed images. The disentanglement of 3DGS parameters enables separate analysis of texture and geometry awareness. The framework allows probing 3D awareness via novel view synthesis without requiring 3D data.

Result: Extensive experiments probe 3D awareness of several VFMs and investigate ingredients that lead to 3D-aware VFMs. Building on findings, several variants achieve state-of-the-art performance across diverse datasets.

Conclusion: Feat2GS provides a unified framework to fairly and comprehensively probe VFM 3D awareness for both geometry and texture via novel view synthesis. It serves as both a probing tool for VFMs and a simple-yet-effective baseline for novel-view synthesis.

Abstract: Given that visual foundation models (VFMs) are trained on extensive datasets but often limited to 2D images, a natural question arises: how well do they understand the 3D world? With the differences in architecture and training protocols (i.e., objectives, proxy tasks), a unified framework to fairly and comprehensively probe their 3D awareness is urgently needed. Existing works on 3D probing suggest single-view 2.5D estimation (e.g., depth and normal) or two-view sparse 2D correspondence (e.g., matching and tracking). Unfortunately, these tasks ignore texture awareness, and require 3D data as ground-truth, which limits the scale and diversity of their evaluation set. To address these issues, we introduce Feat2GS, which readout 3D Gaussians attributes from VFM features extracted from unposed images. This allows us to probe 3D awareness for geometry and texture via novel view synthesis, without requiring 3D data. Additionally, the disentanglement of 3DGS parameters - geometry ($\boldsymbol{x}$, $α$, $Σ$) and texture ($\boldsymbol{c}$) - enables separate analysis of texture and geometry awareness. Under Feat2GS, we conduct extensive experiments to probe the 3D awareness of several VFMs, and investigate the ingredients that lead to a 3D aware VFM. Building on these findings, we develop several variants that achieve state-of-the-art across diverse datasets. This makes Feat2GS useful for probing VFMs, and as a simple-yet-effective baseline for novel-view synthesis. Code and data are available at https://fanegg.github.io/Feat2GS/.

Method: Duplex maintains dual prototypes per composition: semantic prototype via prompt learning and visual prototype constructed by recombining disentangled state/object primitives from seen images. Visual prototypes are dynamically updated through lightweight aggregation on mini-batch local graphs, incorporating unseen compositions during training without labels.

Result: Experiments on MIT-States, UT-Zappos, and CGQA datasets in both closed-world and open-world settings achieve competitive performance and consistent compositional generalization.

Conclusion: Duplex addresses key limitations in CZSL by introducing fine-grained visual evidence while preserving semantic structure, better disambiguating semantically similar yet visually distinct pairs, and mitigating seen bias.

Abstract: Compositional Zero-Shot Learning (CZSL) seeks to recognize unseen state-object pairs by recombining primitives learned from seen compositions. Despite recent progress with vision-language models (VLMs), two limitations remain: (i) text-driven semantic prototypes are weakly discriminative in the visual feature space; and (ii) unseen pairs are optimized passively, thereby inducing seen bias. To address these limitations, we present Duplex, a framework that couples dual-prototype learning with dynamic local-graph refinement of visual prototypes. For each composition, Duplex maintains a semantic prototype via prompt learning and a visual prototype for unseen pairs constructed by recombining disentangled state and object primitives from seen images. The visual prototypes are updated dynamically through lightweight aggregation on mini-batch local graphs, which incorporates unseen compositions during training without labels. This design introduces fine-grained visual evidence while preserving semantic structure. It enriches class prototypes, better disambiguates semantically similar yet visually distinct pairs, and mitigates seen bias. Experiments on MIT-States, UT-Zappos, and CGQA in closed-world and open-world settings achieve competitive performance and consistent compositional generalization. Our source code is available at https://github.com/ISPZ/Duplex-CZSL.

Method: Proposes a 3D dynamics-aware manipulation framework with three self-supervised learning tasks: current depth estimation, future RGB-D prediction, and 3D flow prediction. These tasks complement each other to endow the policy model with 3D foresight.

Result: Extensive experiments on simulation and real-world show that 3D foresight greatly boosts manipulation policy performance without sacrificing inference speed.

Conclusion: 3D world modeling integrated with policy learning significantly improves manipulation performance for tasks involving depth-wise movements, demonstrating the importance of 3D foresight in manipulation policies.

Abstract: The incorporation of world modeling into manipulation policy learning has pushed the boundary of manipulation performance. However, existing efforts simply model the 2D visual dynamics, which is insufficient for robust manipulation when target tasks involve prominent depth-wise movement. To address this, we present a 3D dynamics-aware manipulation framework that seamlessly integrates 3D world modeling and policy learning. Three self-supervised learning tasks (current depth estimation, future RGB-D prediction, 3D flow prediction) are introduced within our framework, which complement each other and endow the policy model with 3D foresight. Extensive experiments on simulation and the real world show that 3D foresight can greatly boost the performance of manipulation policies without sacrificing inference speed. Code is available at https://github.com/Stardust-hyx/3D-Foresight.

Method: The survey reviews the DWM ecosystem including simulators, datasets, and evaluation metrics, categorizes approaches by predicted scene modalities (video, point cloud, occupancy, latent feature, traffic map), and summarizes applications in autonomous driving research.

Result: The survey provides a comprehensive overview of current DWM research, presents performance comparisons of representative approaches across generating and driving tasks, and identifies relevant papers in a curated collection.

Conclusion: DWMs are valuable for autonomous driving development, with the survey offering insights into their broader adoption while discussing current limitations and proposing future research directions.

Abstract: The Driving World Model (DWM), which focuses on predicting scene evolution during the driving process, has emerged as a promising paradigm in the pursuit of autonomous driving (AD). DWMs enable AD systems to better perceive, understand, and interact with dynamic driving environments. In this survey, we provide a comprehensive overview of the latest progress in DWM. First, we review the DWM ecosystem, which is constructed using mainstream simulators, high-impact datasets, and various metrics that evaluate DWMs across multiple dimensions. We then categorize existing approaches based on the modalities of the predicted scenes, including video, point cloud, occupancy, latent feature, and traffic map, and summarize their specific applications in AD research. In addition, the performance of representative approaches across generating and driving tasks is presented. Finally, we discuss the potential limitations of current research and propose future directions. This survey provides valuable insights into the development and application of DWM, fostering its broader adoption in AD. The relevant papers are collected at https://github.com/LMD0311/Awesome-World-Model.

Method: Uses curriculum strategy: 1) Train subject-specific teacher policies to mimic, retarget, and refine MoCap data, 2) Distill teachers into student policy with teachers providing supervision and references, 3) RL fine-tuning on student policy to surpass demonstration replication.

Result: Produces realistic and diverse interactions across multiple HOI datasets, generalizes zero-shot, and integrates with kinematic generators, elevating from imitation to generative modeling of complex human-object interactions.

Conclusion: InterMimic enables robust learning from imperfect MoCap data for diverse full-body human-object interactions, achieving realistic simulations that generalize well and support generative modeling.

Abstract: Achieving realistic simulations of humans interacting with a wide range of objects has long been a fundamental goal. Extending physics-based motion imitation to complex human-object interactions (HOIs) is challenging due to intricate human-object coupling, variability in object geometries, and artifacts in motion capture data, such as inaccurate contacts and limited hand detail. We introduce InterMimic, a framework that enables a single policy to robustly learn from hours of imperfect MoCap data covering diverse full-body interactions with dynamic and varied objects. Our key insight is to employ a curriculum strategy – perfect first, then scale up. We first train subject-specific teacher policies to mimic, retarget, and refine motion capture data. Next, we distill these teachers into a student policy, with the teachers acting as online experts providing direct supervision, as well as high-quality references. Notably, we incorporate RL fine-tuning on the student policy to surpass mere demonstration replication and achieve higher-quality solutions. Our experiments demonstrate that InterMimic produces realistic and diverse interactions across multiple HOI datasets. The learned policy generalizes in a zero-shot manner and seamlessly integrates with kinematic generators, elevating the framework from mere imitation to generative modeling of complex human-object interactions.

Method: 1) Render coarse textured mesh from multiple viewpoints; 2) Fine-tune deterministic prior-guided normal diffusion model on paired detail-lacking/detail-rich normal maps; 3) Use noise-resistant inverse rendering via deformable distance field to update mesh surfaces from predicted normal maps.

Result: SuperCarver generates realistic and expressive surface details that align with actual texture appearance, effectively upgrading historical low-quality 3D assets and reducing workload for sculpting high-poly meshes.

Conclusion: SuperCarver provides a powerful tool for 3D geometry super-resolution, addressing the challenge of enhancing existing mesh quality with texture-consistent surface details through a novel diffusion-based approach.

Abstract: Conventional production workflow of high-precision mesh assets necessitates a cumbersome and laborious process of manual sculpting by specialized 3D artists/modelers. The recent years have witnessed remarkable advances in AI-empowered 3D content creation for generating plausible structures and intricate appearances from images or text prompts. However, synthesizing realistic surface details still poses great challenges, and enhancing the geometry fidelity of existing lower-quality 3D meshes (instead of image/text-to-3D generation) remains an open problem. In this paper, we introduce SuperCarver, a 3D geometry super-resolution pipeline for supplementing texture-consistent surface details onto a given coarse mesh. We start by rendering the original textured mesh into the image domain from multiple viewpoints. To achieve detail boosting, we construct a deterministic prior-guided normal diffusion model, which is fine-tuned on a carefully curated dataset of paired detail-lacking and detail-rich normal map renderings. To update mesh surfaces from potentially imperfect normal map predictions, we design a noise-resistant inverse rendering scheme through deformable distance field. Experiments demonstrate that our SuperCarver is capable of generating realistic and expressive surface details depicted by the actual texture appearance, making it a powerful tool to both upgrade historical low-quality 3D assets and reduce the workload of sculpting high-poly meshes.

Method: Proposes DenseFormer with: 1) Feature extraction module using feature pyramid structure and multi-layer deformable attention to integrate sparse depth and RGB features as diffusion guidance, 2) Diffusion process that generates dense depth by refining random distributions, 3) Depth refinement module with multi-step iterative refinement using multi-scale image features and sparse depth input.

Result: Extensive experiments on KITTI outdoor scene dataset show DenseFormer outperforms classical depth completion methods.

Conclusion: DenseFormer successfully integrates diffusion models into depth completion, demonstrating superior performance through its feature extraction and refinement modules.

Abstract: The depth completion task is a critical problem in autonomous driving, involving the generation of dense depth maps from sparse depth maps and RGB images. Most existing methods employ a spatial propagation network to iteratively refine the depth map after obtaining an initial dense depth. In this paper, we propose DenseFormer, a novel method that integrates the diffusion model into the depth completion task. By incorporating the denoising mechanism of the diffusion model, DenseFormer generates the dense depth map by progressively refining an initial random depth distribution through multiple iterations. We propose a feature extraction module that leverages a feature pyramid structure, along with multi-layer deformable attention, to effectively extract and integrate features from sparse depth maps and RGB images, which serve as the guiding condition for the diffusion process. Additionally, this paper presents a depth refinement module that applies multi-step iterative refinement across various ranges to the dense depth results generated by the diffusion process. The module utilizes image features enriched with multi-scale information and sparse depth input to further enhance the accuracy of the predicted depth map. Extensive experiments on the KITTI outdoor scene dataset demonstrate that DenseFormer outperforms classical depth completion methods.

Method: 1) Construct unified latent representation encoding heterogeneous pose formats (SMPLer-X, DWPose, Mediapipe, PrimeDepth, Sapiens Segmentation) into compact information-dense space. 2) Train ViT-based Video2Pose module to extract latent representation directly from raw videos. 3) Develop temporal modeling and sequence refinement method operating entirely in latent space.

Result: SignX achieves state-of-the-art accuracy on continuous sign language recognition while significantly reducing computational consumption through its multi-stage design that operates in compact pose-rich latent space.

Conclusion: The proposed SignX framework successfully demonstrates that operating in a compact pose-rich latent space enables efficient and accurate continuous sign language recognition, representing an advancement in sign language processing technology.

Abstract: The complexity of sign language data processing brings many challenges. The current approach to recognition of ASL signs aims to translate RGB sign language videos through pose information into English-based ID Glosses, which serve to uniquely identify ASL signs. This paper proposes SignX, a novel framework for continuous sign language recognition in compact pose-rich latent space. First, we construct a unified latent representation that encodes heterogeneous pose formats (SMPLer-X, DWPose, Mediapipe, PrimeDepth, and Sapiens Segmentation) into a compact, information-dense space. Second, we train a ViT-based Video2Pose module to extract this latent representation directly from raw videos. Finally, we develop a temporal modeling and sequence refinement method that operates entirely in this latent space. This multi-stage design achieves end-to-end sign language recognition while significantly reducing computational consumption. Experimental results demonstrate that SignX achieves state-of-the-art accuracy on continuous sign language recognition.

Method: GAN-based approach with three components: 1) residual-embedded generator to prevent gradient issues, 2) Inception ResNet-V1 based FaceNet discriminator for better adversarial training, and 3) end-to-end framework jointly optimizing data generation and recognition.

Result: Achieves stable training and improves face recognition accuracy by 12.7% on LFW benchmark compared to baselines, with good generalization using limited samples.

Conclusion: The proposed GAN-based data augmentation method effectively addresses data scarcity in face recognition, demonstrating significant performance improvements and stable training dynamics.

Abstract: Face recognition performance based on deep learning heavily relies on large-scale training data, which is often difficult to acquire in practical applications. To address this challenge, this paper proposes a GAN-based data augmentation method with three key contributions: (1) a residual-embedded generator to alleviate gradient vanishing/exploding problems, (2) an Inception ResNet-V1 based FaceNet discriminator for improved adversarial training, and (3) an end-to-end framework that jointly optimizes data generation and recognition performance. Experimental results demonstrate that our approach achieves stable training dynamics and significantly improves face recognition accuracy by 12.7% on the LFW benchmark compared to baseline methods, while maintaining good generalization capability with limited training samples.

Method: Proposes Markovian Visual AutoRegressive modeling (MVAR) with two key innovations: 1) scale-Markov trajectory that only uses adjacent preceding scale features for next-scale prediction, enabling parallel training, and 2) spatial-Markov attention that restricts attention to localized neighborhoods (size k) at corresponding positions on adjacent scales.

Result: Reduces attention complexity from O(N²) to O(Nk), enables training with just 8 RTX 4090 GPUs, eliminates KV cache during inference, reduces average GPU memory footprint by 3.0x, and achieves comparable or superior performance on ImageNet with both small from-scratch and large fine-tuned models.

Conclusion: MVAR effectively mitigates redundancy in visual autoregressive modeling through Markov assumptions, providing an efficient framework for visual generation with significantly reduced computational and memory requirements while maintaining strong performance.

Abstract: Essential to visual generation is efficient modeling of visual data priors. Conventional next-token prediction methods define the process as learning the conditional probability distribution of successive tokens. Recently, next-scale prediction methods redefine the process to learn the distribution over multi-scale representations, significantly reducing generation latency. However, these methods condition each scale on all previous scales and require each token to consider all preceding tokens, exhibiting scale and spatial redundancy. To better model the distribution by mitigating redundancy, we propose Markovian Visual AutoRegressive modeling (MVAR), a novel autoregressive framework that introduces scale and spatial Markov assumptions to reduce the complexity of conditional probability modeling. Specifically, we introduce a scale-Markov trajectory that only takes as input the features of adjacent preceding scale for next-scale prediction, enabling the adoption of a parallel training strategy that significantly reduces GPU memory consumption. Furthermore, we propose spatial-Markov attention, which restricts the attention of each token to a localized neighborhood of size k at corresponding positions on adjacent scales, rather than attending to every token across these scales, for the pursuit of reduced modeling complexity. Building on these improvements, we reduce the computational complexity of attention calculation from O(N^2) to O(Nk), enabling training with just eight NVIDIA RTX 4090 GPUs and eliminating the need for KV cache during inference. Extensive experiments on ImageNet demonstrate that MVAR achieves comparable or superior performance with both small model trained from scratch and large fine-tuned models, while reducing the average GPU memory footprint by 3.0x.

Method: Created U2-BENCH with 7,241 cases spanning 15 anatomical regions and 8 clinically inspired tasks across 50 ultrasound scenarios. Evaluated 23 state-of-the-art LVLMs (open/closed source, general/medical) on classification, detection, regression, and text generation tasks.

Result: LVLMs show strong performance on image-level classification but struggle with spatial reasoning and clinical language generation. The benchmark provides a rigorous testbed for assessing LVLM capabilities in ultrasound imaging.

Conclusion: U2-BENCH establishes a comprehensive evaluation framework for LVLMs in ultrasound, revealing current limitations and providing a foundation for future research in medical multimodal AI.

Abstract: Ultrasound is a widely-used imaging modality critical to global healthcare, yet its interpretation remains challenging due to its varying image quality on operators, noises, and anatomical structures. Although large vision-language models (LVLMs) have demonstrated impressive multimodal capabilities across natural and medical domains, their performance on ultrasound remains largely unexplored. We introduce U2-BENCH, the first comprehensive benchmark to evaluate LVLMs on ultrasound understanding across classification, detection, regression, and text generation tasks. U2-BENCH aggregates 7,241 cases spanning 15 anatomical regions and defines 8 clinically inspired tasks, such as diagnosis, view recognition, lesion localization, clinical value estimation, and report generation, across 50 ultrasound application scenarios. We evaluate 23 state-of-the-art LVLMs, both open- and closed-source, general-purpose and medical-specific. Our results reveal strong performance on image-level classification, but persistent challenges in spatial reasoning and clinical language generation. U2-BENCH establishes a rigorous and unified testbed to assess and accelerate LVLM research in the uniquely multimodal domain of medical ultrasound imaging.