Video-DeepResearch: Advancing Multimodal Agents for Continuous Video Streams

A new framework called Video-DeepResearch (Video-DR) has been developed by researchers to transition multimodal agents from handling static images to managing continuous video streams, necessitating intricate spatiotemporal grounding and exploration of the open web. Initial assessments revealed two major limitations in existing models: modality bias, where agents favor textual searches over visual tools, and parametric knowledge leakage, which leads to reliance on internal memory instead of authentic tool-based execution. To overcome these issues, Video-DR incorporates a separate perception-exploration pipeline with phased tool unlocking, ensuring thorough cross-frame visual grounding prior to web retrieval. The framework employs a two-step training approach: supervised fine-tuning followed by Group Relative Policy Optimization (GRPO), facilitating independent exploration that surpasses imitation-learning limitations. Furthermore, the team created Video-DR-Bench, a benchmark for assessing human-AI performance. This research is documented in a paper on arXiv (2608.03979v1) and marks progress toward advanced multimodal deep research agents.