Selected work

Representative Papers

MoA-VR: A Mixture-of-Agents System Towards All-in-One Video Restoration

Oct 09, 2025

Real-world videos often suffer from complex, heterogeneous degradations—including noise, compression artifacts, and low-light distortions—posing significant challenges for generalization across diverse and compound degradation types. To address this, we propose the first hybrid agent system for video restoration, inspired by human expert collaboration, comprising three synergistic modules: degradation identification, adaptive routing-based restoration, and quality assessment. Our approach introduces a novel multi-agent architecture featuring a learnable routing mechanism that integrates vision-language models with large language models. Furthermore, we develop Res-VQ, the first restoration-oriented video quality assessment model, along with its dedicated benchmark dataset, Res-VQ-Bench. Extensive experiments demonstrate that our method achieves substantial improvements over state-of-the-art methods in both objective metrics and perceptual quality, particularly under challenging, composite degradation scenarios.

1 citationsRead paper

Video-MOPD: Multi-Teacher On-Policy Distillation for Video Understanding

Sep 08, 2026

为提升视频理解能力，通过多教师在线策略蒸馏和强化学习优化，结合可靠性感知信息采样方法，实现模型在多种视频理解任务上的性能提升。

0 citationsRead paper

Do Text-to-Music Models Really Follow Instructions? A Counterfactual Evaluation of Key and Beat Grouping

Aug 12, 2026

Current evaluations of instruction-following in text-to-music models are susceptible to output prior bias, making it difficult to discern whether generated attributes stem from the given instructions or the model’s inherent preferences. This work proposes a matched counterfactual evaluation framework that disentangles instruction controllability from output priors by comparing outputs from neutral prompts against those from swapped target prompts, under controlled conditions including shared random seeds, frozen adapters, and external discriminator validation. Integrating blind expert annotations with a multi-seed sentinel mechanism, this approach provides the first systematic quantification of model controllability over tonality and metrical grouping. Experiments reveal that ACE-Step 1.5 and Stable Audio 3 Medium exhibit significant tonal control, whereas LeVo2 does not; furthermore, the high consistency in quadruple-meter structures is primarily attributable to model priors rather than instruction adherence.

0 citationsRead paper

MemeMind: Reference-Guided Trace Construction for Offline Context Optimization

Aug 10, 2026

This work addresses the challenge in offline context optimization where only failed trajectories are available, particularly in complex multimodal tasks such as anime and gaming memes, which lack successful examples of tool usage. To overcome this limitation, the authors propose a reference-answer-guided trajectory reconstruction method. Their framework employs TraceBuilder to synthesize successful tool-invocation trajectories—encompassing text search, image retrieval, and visual grounding—by leveraging reference answers. ToolGuide then distills both shared guidelines and tool-specific instructions from these reconstructed trajectories to optimize reasoning contexts under frozen model settings. This approach is the first to exploit offline reference answers for reconstructing successful experiences, effectively complementing failed queries and yielding generalizable tool-use policies. On the MemeX benchmark, it boosts performance of Qwen3-VL-30B and 235B models by 22.0%/21.1% and 8.1%/8.0%, respectively (as evaluated by GPT-5), with ablation studies confirming that trajectory construction contributes the largest performance gain.

0 citationsRead paper

MemeBench: What LVLMs Miss When Interpreting Culture-Dependent Memes

Jul 30, 2026

This study addresses the limitations of current large vision-language models (LVLMs) in interpreting internet memes that rely on cultural context, where inadequate integration of visual content and external knowledge often leads to incomplete semantic understanding. To this end, the authors introduce MemeBench—the first fine-grained diagnostic benchmark specifically designed for Chinese and English subcultural memes—and propose the VIKR framework (Visual cues, Identity linkage, Knowledge units, Reasoning mechanisms) to systematically evaluate model capabilities. Leveraging a newly constructed CultureBase knowledge repository, they develop an entity-guided Knowledge-Augmented Retrieval (KAR) method. Experiments reveal a 22.6% average gap in visual-knowledge comprehension across 26 mainstream LVLMs; KAR improves VIKR success rates by 3.6–7.4% on four representative models, effectively mitigating knowledge deficits and demonstrating the efficacy of structured knowledge injection.

0 citationsRead paper

Recent publications

Latest Papers

Video-MOPD: Multi-Teacher On-Policy Distillation for Video Understanding

Sep 08, 2026

为提升视频理解能力，通过多教师在线策略蒸馏和强化学习优化，结合可靠性感知信息采样方法，实现模型在多种视频理解任务上的性能提升。

0 citationsRead paper

Do Text-to-Music Models Really Follow Instructions? A Counterfactual Evaluation of Key and Beat Grouping

Aug 12, 2026

Current evaluations of instruction-following in text-to-music models are susceptible to output prior bias, making it difficult to discern whether generated attributes stem from the given instructions or the model’s inherent preferences. This work proposes a matched counterfactual evaluation framework that disentangles instruction controllability from output priors by comparing outputs from neutral prompts against those from swapped target prompts, under controlled conditions including shared random seeds, frozen adapters, and external discriminator validation. Integrating blind expert annotations with a multi-seed sentinel mechanism, this approach provides the first systematic quantification of model controllability over tonality and metrical grouping. Experiments reveal that ACE-Step 1.5 and Stable Audio 3 Medium exhibit significant tonal control, whereas LeVo2 does not; furthermore, the high consistency in quadruple-meter structures is primarily attributable to model priors rather than instruction adherence.

0 citationsRead paper

MemeMind: Reference-Guided Trace Construction for Offline Context Optimization

Aug 10, 2026

This work addresses the challenge in offline context optimization where only failed trajectories are available, particularly in complex multimodal tasks such as anime and gaming memes, which lack successful examples of tool usage. To overcome this limitation, the authors propose a reference-answer-guided trajectory reconstruction method. Their framework employs TraceBuilder to synthesize successful tool-invocation trajectories—encompassing text search, image retrieval, and visual grounding—by leveraging reference answers. ToolGuide then distills both shared guidelines and tool-specific instructions from these reconstructed trajectories to optimize reasoning contexts under frozen model settings. This approach is the first to exploit offline reference answers for reconstructing successful experiences, effectively complementing failed queries and yielding generalizable tool-use policies. On the MemeX benchmark, it boosts performance of Qwen3-VL-30B and 235B models by 22.0%/21.1% and 8.1%/8.0%, respectively (as evaluated by GPT-5), with ablation studies confirming that trajectory construction contributes the largest performance gain.

0 citationsRead paper

MemeBench: What LVLMs Miss When Interpreting Culture-Dependent Memes

Jul 30, 2026

This study addresses the limitations of current large vision-language models (LVLMs) in interpreting internet memes that rely on cultural context, where inadequate integration of visual content and external knowledge often leads to incomplete semantic understanding. To this end, the authors introduce MemeBench—the first fine-grained diagnostic benchmark specifically designed for Chinese and English subcultural memes—and propose the VIKR framework (Visual cues, Identity linkage, Knowledge units, Reasoning mechanisms) to systematically evaluate model capabilities. Leveraging a newly constructed CultureBase knowledge repository, they develop an entity-guided Knowledge-Augmented Retrieval (KAR) method. Experiments reveal a 22.6% average gap in visual-knowledge comprehension across 26 mainstream LVLMs; KAR improves VIKR success rates by 3.6–7.4% on four representative models, effectively mitigating knowledge deficits and demonstrating the efficacy of structured knowledge injection.

0 citationsRead paper

SPEED: One-Step Pixel Diffusion for High-quality Video Frame Interpolation

Jul 16, 2026

This work addresses the high memory consumption, inference latency, and detail degradation inherent in existing diffusion-based video frame interpolation methods that rely on multi-step sampling. To overcome these limitations, we propose SPEED—a single-step, pixel-level diffusion framework that jointly models multi-scale motion, structure, and appearance through a progressive multi-stage architecture with dynamic block scaling, directly predicting intermediate frames in pixel space. Key innovations include a Noise-Update-Only Attention mechanism that preserves semantic fidelity of conditioning frames while reducing computational overhead by nearly 50%, and a Drift-aware Timestep Sampling strategy that enhances single-step generation quality. Experiments demonstrate that SPEED achieves an 8.8% lower LPIPS on SNU-FILM, 63.3% faster inference, and 10.6% less memory usage; on 4K benchmarks, it improves LPIPS by up to 51.5% over prior methods.

0 citationsRead paper