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

TL;DR - Video-DeepResearch is a multimodal agent framework that combines cross-frame video grounding with open-web exploration. Its 35B-A3B model achieves 64.0% accuracy on a new multi-hop video QA benchmark, outperforming the cited proprietary models.

• Uses stage-wise tool unlocking to require visual analysis before web retrieval, reducing modality bias and reliance on memorized knowledge.
• Separates perception from exploration to support dense spatiotemporal grounding across continuous video.
• Trains with supervised fine-tuning followed by GRPO to improve autonomous exploration beyond imitation learning.
• Introduces Video-DR-Bench, comprising 200 complex, multi-hop video question-answering examples.