VideoResearchAgent improves video search and reasoning efficiency online
VideoResearchAgent: Grounded Task Synthesis and Sim-to-Real RL for Open-Web Video Research
Artificial IntelligenceComputer Vision and Pattern Recognition
Summary
Searching and understanding videos on the open web is hard because videos are slow to watch and it's tricky to find the right moments. The authors designed a training system called VideoResearchAgent that can practice video searching and answering questions efficiently using simulated videos, avoiding shortcuts that don’t work on real web videos. Their system learns to find and navigate videos better and reduces the amount of data and cost needed to train these skills. This approach helps build AI that can handle real-time video research more accurately and quickly.
What this means in practice
- •For video platform developers: Train AI agents to autonomously find relevant videos and answer questions based on video content for improved search and recommendation systems.
- •For media monitoring teams: Deploy systems that efficiently search large video libraries and extract visual evidence to support real-time video-based research and reporting.
Authors
Yuhang Zhou, Fei Li, Yuxi Wu, Bin Zhu, Jingjing Chen
Abstract
Existing deep research agents are designed primarily for text- and image-based web sources, while video reasoning systems typically assume that relevant videos are provided in advance. We study open-web video research, where an agent must autonomously discover relevant videos, navigate their temporal content, and ground answers in visual evidence. Training such agents at scale is challenging as live video interaction is slow and unreliable, whereas fixed local simulation can induce retrieval-specific shortcuts that fail to transfer to the open web. We introduce VideoResearchAgent, a scalable training framework to address these challenges. First, we introduce controllable task synthesis pipeline to synthesize multi-hop research tasks from timestamped visual evidence while filtering text-only shortcuts. Second, we build a field-aligned local video simulator that preserves deployment-facing search and watch interactions while accelerating video search by a factor of 34.5-64.6. Third, we introduce Retrieval-Domain-Randomized GRPO (RDR-GRPO), which diversifies candidate rankings, distractors, metadata, and result structure during training to reduce overfitting to simulated retrieval. On Video-BrowseComp, the VideoResearchAgent trained using Qwen3.5-4B achieves 40.48% accuracy, comparable to Gemini-3-Flash-Preview, while reducing cumulative API-token consumption by 74.9% relative to the untrained model. Together, these results establish an accurate and efficient training recipe for open-web video research.