Selected work

Representative Papers

MathScape: Evaluating MLLMs in multimodal Math Scenarios through a Hierarchical Benchmark

Aug 14, 2024arXiv.org

Existing mathematical reasoning benchmarks heavily rely on synthetic images, failing to capture the complexity of real-world multimodal reasoning involving photographs and textual mathematics. Method: We introduce MathScape—the first hierarchical multimodal benchmark for photorealistic mathematical problems—featuring a novel “scene–semantics–task” three-level taxonomy that systematically integrates authentic images with formal mathematical semantics, thereby addressing the longstanding gap in joint vision-language mathematical reasoning evaluation. Contribution/Results: Leveraging 11 state-of-the-art multimodal large language models (MLLMs), we conduct dual-track evaluation assessing both theoretical understanding and practical application. Empirical results reveal that even top-performing models achieve sub-50% average accuracy, exposing critical weaknesses in cross-modal alignment, symbolic parsing, and multi-step reasoning. MathScape establishes a new, high-challenge, fine-grained, and interpretable evaluation paradigm for multimodal mathematical reasoning.

7 citationsRead paper

When Do Prompt-Side Agent Playbooks Transfer? Accuracy, Cost, and Runtime Shift in Agent Deployment

Aug 06, 2026

This study systematically evaluates the transferability of frozen prompt scripts—those deployed without retraining—across diverse environments, with a focus on their impact on accuracy, inference cost, and runtime behavior. Employing a unified distillation–validation–transfer protocol, experiments are conducted on benchmarks including ALFWorld, TAU2-Bench, and XBench-DeepSearch, enabling the first quantitative assessment of how runtime latency and cost shifts constrain transfer performance. The findings reveal that script transfer is highly condition-dependent, showing effectiveness only in specific scenarios—such as when distillation-based prompting outperforms fixed exemplars—while most transfer paths yield negligible gains and often trigger issues like redundant queries, premature termination, and cost surges. These results suggest that frozen scripts should be treated as conditional cold-start options rather than default reusable strategies, underscoring the necessity of rigorous validation in target environments.

0 citationsRead paper

Beyond Parameter Arithmetic: Sparse Complementary Fusion for Distribution-Aware Model Merging

Feb 12, 2026

Existing model merging approaches rely on heuristic operations in parameter space, which often induce functional interference, leading to degraded generalization and unstable generation. This work proposes SCF-RKL, a sparse complementary fusion framework that abandons the assumption of linear parameter interpolation and instead introduces a distribution alignment mechanism based on reverse Kullback-Leibler divergence to identify and merge complementary parameters. By enabling distribution-aware sparse updates, SCF-RKL effectively mitigates functional interference while preserving stable representations. The method achieves state-of-the-art performance across 24 benchmarks spanning reasoning, instruction following, knowledge retention, and safety, demonstrating superior generalization and generation stability compared to existing techniques.

0 citationsRead paper

SkinFlow: Efficient Information Transmission for Open Dermatological Diagnosis via Dynamic Visual Encoding and Staged RL

Jan 14, 2026

This work addresses the limited diagnostic accuracy of general-purpose large vision-language models in dermatology, which stems from their diffuse attention mechanisms struggling to distinguish subtle lesions from background noise. To overcome this, the study formulates skin disease diagnosis as an optimization problem of visual information transmission efficiency and proposes a Dynamic Visual Encoder with virtual width expansion (DVE) that effectively unfolds complex pathological manifolds without increasing parameter count. A two-stage reinforcement learning mechanism is further introduced to progressively align explicit medical descriptions with implicit diagnostic textures. Evaluated under a clinically safety-oriented protocol on Fitzpatrick17k, the proposed 7B model achieves a 12.06% improvement in Top-1 accuracy and a 28.57% gain in Top-6 accuracy, outperforming much larger models such as Qwen3VL-235B and GPT-5.2.

0 citationsRead paper

JointAVBench: A Benchmark for Joint Audio-Visual Reasoning Evaluation

Dec 14, 2025

Existing benchmarks inadequately evaluate omni-modal large language models (Omni-LLMs) due to insufficient coverage of multimodal dependencies, diverse audio modalities (e.g., speech, sound events, music, vocal traits), and single-/cross-/full-scenario spans. To address this, we introduce AV-Bench—the first rigorous benchmark mandating joint audiovisual understanding. Its design comprises three stringent dimensions: strong audiovisual coupling, five cognitive capability axes, and three scenario spans (single, cross-, and full-modality). We further develop an automated, multi-large-model-collaborative QA synthesis pipeline ensuring questions necessitate integrated audiovisual reasoning. High-quality annotation is achieved via multi-granularity scene segmentation and cross-modal alignment. Experiments reveal that state-of-the-art Omni-LLMs achieve only 62.6% average accuracy—substantially below human performance—highlighting cross-scenario joint reasoning as the critical bottleneck. AV-Bench thus provides a precise, challenging evaluation standard for advancing multimodal foundation models.

0 citationsRead paper

Recent publications

Latest Papers

When Do Prompt-Side Agent Playbooks Transfer? Accuracy, Cost, and Runtime Shift in Agent Deployment

Aug 06, 2026

This study systematically evaluates the transferability of frozen prompt scripts—those deployed without retraining—across diverse environments, with a focus on their impact on accuracy, inference cost, and runtime behavior. Employing a unified distillation–validation–transfer protocol, experiments are conducted on benchmarks including ALFWorld, TAU2-Bench, and XBench-DeepSearch, enabling the first quantitative assessment of how runtime latency and cost shifts constrain transfer performance. The findings reveal that script transfer is highly condition-dependent, showing effectiveness only in specific scenarios—such as when distillation-based prompting outperforms fixed exemplars—while most transfer paths yield negligible gains and often trigger issues like redundant queries, premature termination, and cost surges. These results suggest that frozen scripts should be treated as conditional cold-start options rather than default reusable strategies, underscoring the necessity of rigorous validation in target environments.

0 citationsRead paper

Beyond Parameter Arithmetic: Sparse Complementary Fusion for Distribution-Aware Model Merging

Feb 12, 2026

Existing model merging approaches rely on heuristic operations in parameter space, which often induce functional interference, leading to degraded generalization and unstable generation. This work proposes SCF-RKL, a sparse complementary fusion framework that abandons the assumption of linear parameter interpolation and instead introduces a distribution alignment mechanism based on reverse Kullback-Leibler divergence to identify and merge complementary parameters. By enabling distribution-aware sparse updates, SCF-RKL effectively mitigates functional interference while preserving stable representations. The method achieves state-of-the-art performance across 24 benchmarks spanning reasoning, instruction following, knowledge retention, and safety, demonstrating superior generalization and generation stability compared to existing techniques.

0 citationsRead paper

SkinFlow: Efficient Information Transmission for Open Dermatological Diagnosis via Dynamic Visual Encoding and Staged RL

Jan 14, 2026

This work addresses the limited diagnostic accuracy of general-purpose large vision-language models in dermatology, which stems from their diffuse attention mechanisms struggling to distinguish subtle lesions from background noise. To overcome this, the study formulates skin disease diagnosis as an optimization problem of visual information transmission efficiency and proposes a Dynamic Visual Encoder with virtual width expansion (DVE) that effectively unfolds complex pathological manifolds without increasing parameter count. A two-stage reinforcement learning mechanism is further introduced to progressively align explicit medical descriptions with implicit diagnostic textures. Evaluated under a clinically safety-oriented protocol on Fitzpatrick17k, the proposed 7B model achieves a 12.06% improvement in Top-1 accuracy and a 28.57% gain in Top-6 accuracy, outperforming much larger models such as Qwen3VL-235B and GPT-5.2.

0 citationsRead paper

JointAVBench: A Benchmark for Joint Audio-Visual Reasoning Evaluation

Dec 14, 2025

Existing benchmarks inadequately evaluate omni-modal large language models (Omni-LLMs) due to insufficient coverage of multimodal dependencies, diverse audio modalities (e.g., speech, sound events, music, vocal traits), and single-/cross-/full-scenario spans. To address this, we introduce AV-Bench—the first rigorous benchmark mandating joint audiovisual understanding. Its design comprises three stringent dimensions: strong audiovisual coupling, five cognitive capability axes, and three scenario spans (single, cross-, and full-modality). We further develop an automated, multi-large-model-collaborative QA synthesis pipeline ensuring questions necessitate integrated audiovisual reasoning. High-quality annotation is achieved via multi-granularity scene segmentation and cross-modal alignment. Experiments reveal that state-of-the-art Omni-LLMs achieve only 62.6% average accuracy—substantially below human performance—highlighting cross-scenario joint reasoning as the critical bottleneck. AV-Bench thus provides a precise, challenging evaluation standard for advancing multimodal foundation models.

0 citationsRead paper

ChronusOmni: Improving Time Awareness of Omni Large Language Models

Dec 10, 2025

Multimodal large language models (MLLMs) suffer from insufficient explicit and implicit cross-modal temporal localization capabilities in long-video understanding and complex temporal reasoning: existing methods primarily focus on explicit vision-language grounding while neglecting audio modalities and implicit audio-visual temporal correlations (e.g., “what appears visually when a person speaks”). To address this, we propose a unified temporal modeling framework featuring: (i) novel text-based timestamp tokens and interleaved frame-level audio-visual representations; (ii) a temporal consistency reinforcement learning reward tailored for fine-grained temporal reasoning; and (iii) ChronusAV—the first high-precision, fully modality-aligned benchmark for audio-visual temporal understanding. Experiments demonstrate over 30% performance gain on ChronusAV, state-of-the-art results across multiple temporal localization metrics, and no degradation in general audio-visual comprehension capability.

0 citationsRead paper