Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent

Analysis

Why This Paper Matters

Video-DeepResearch (Video-DR) addresses a critical gap in AI: extending deepresearch agents from static images to continuous video streams. While existing agents like GPT-5 and Claude-4.5-Sonnet handle images and text, they struggle with video due to the need for dense spatiotemporal grounding and open-web exploration. This paper systematically identifies two key bottlenecks—modality bias and parametric knowledge leakage—that prevent current models from effectively using visual tools. By proposing a decoupled perception-exploration pipeline and a two-stage training recipe, Video-DR offers a practical solution that significantly improves performance on a new benchmark.

The introduction of Video-DR-Bench, a human-AI collaborative benchmark with 200 complex multi-hop VQA instances, is also significant. It provides a standardized evaluation for video-based deepresearch, which is currently lacking. The benchmark's design encourages genuine tool use and tests the ability to integrate visual and web information, making it a valuable resource for future research.

Technical Contributions

• Decoupled Perception-Exploration Pipeline: Video-DR separates visual grounding from web retrieval, forcing the model to exhaustively analyze video frames before searching the web. This reduces modality bias and ensures that visual evidence is considered first.
• Stage-wise Tool Unlocking: The pipeline progressively unlocks tools, preventing the model from skipping visual analysis and jumping to textual search. This is a novel mechanism to enforce tool usage.
• Two-Stage Training Recipe: The combination of supervised fine-tuning (SFT) and Group Relative Policy Optimization (GRPO) is a key innovation. GRPO enables autonomous exploration, allowing the model to learn beyond imitation and develop its own tool-use strategies.
• Video-DR-Bench: A new benchmark with 200 complex, multi-hop VQA instances that require both video understanding and web research. The human-AI collaborative curation ensures high quality and difficulty.
• Model Release: The authors release Video-DeepResearch-35B-A3B and 30B-A3B variants, with code available on GitHub, facilitating reproducibility and further research.

Results

Video-DeepResearch-35B-A3B achieves a state-of-the-art average accuracy of 64.0% on Video-DR-Bench, outperforming proprietary models like Claude-4.5-Sonnet (59.0%) by 5.0 points, GPT-5 (52.5%) by 11.5 points, and Gemini 2.5 Pro (57.5%) by 6.5 points. The smaller 30B-A3B variant achieves 59.3%, matching Claude-4.5-Sonnet's performance, demonstrating that the training paradigm is effective even at compact scale. These results highlight the importance of tool-augmented execution and the effectiveness of the decoupled pipeline.

Significance

Video-DR represents a significant step toward next-generation multimodal agents that can handle continuous video streams. By addressing modality bias and parametric knowledge leakage, it sets a new standard for tool-augmented reasoning. The training recipe and benchmark could influence future work in video understanding, embodied AI, and autonomous research agents. The open-source release of models and code will accelerate progress in this area, making it a valuable contribution to the AI community.