Video World Models · 2026-10-06

VideoResearchAgent: Grounded Task Synthesis and Sim-to-Real RL for Open-Web Video Research

Yuhang Zhou, Fei Li, Yuxi Wu, Bin Zhu, Jingjing Chen

arXiv:2610.04911PDFCode

AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.

TL;DR

VideoResearchAgent trains agents for open-web video research, where an agent must discover relevant videos, navigate temporal content, and ground answers in visual evidence. The authors add (i) controllable multi-hop video-grounded task synthesis with filtering against text-only solutions, (ii) a field-aligned local video simulator that preserves search/watch interactions and speeds video search, and (iii) Retrieval-Domain-Randomized GRPO to reduce simulator-to-web overfitting. On Video-BrowseComp, Qwen3.5-4B reaches 40.48% OA and uses fewer API tokens.

Why it matters

The paper targets a gap between text/image deep-research agents and video systems that assume the right videos are already provided. By training on tasks that require inspecting video frames and by addressing sim-to-real differences in retrieval, the authors propose a recipe for scalable RL training that aims to transfer to open-web video search and watch/verify behavior.

Method

  • Controllable task-synthesis pipeline: builds multi-hop evidence graphs from timestamped visual evidence, generates grounded questions, and filters tasks by testing against a closed-book LLM and a text-based web-search agent.
  • Field-aligned local simulator: constructs a local corpus (43,751 videos, 1.39M keyframes) and provides local implementations of video search metadata retrieval and watchable video/keyframe acquisition while preserving live tool schema/URL structure.
  • Retrieval-Domain-Randomized GRPO (RDR-GRPO): perturbs candidate rankings, distractors, metadata, and result structure during simulated training to reduce overfitting to simulator retrieval patterns.

Limitation

RL rollouts on live video platforms are prohibitively expensive and unstable (latency, rate limits, HTTP errors), motivating a local simulator; the simulator retrieval distribution is cleaner and more predictable than the live web, so the authors add retrieval-domain randomization to reduce but address this mismatch.

Abstract (from arXiv)

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.

Related papers

Publication record

Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.

First public release: 2026-10-04 · Indexed by Pub-AI: 2026-10-06 UTC

preprint

VideoResearchAgent: Grounded Task Synthesis and Sim-to-Real RL for Open-Web Video Research

Yuhang Zhou, Fei Li, Yuxi Wu, Bin Zhu, Jingjing Chen

Preprint posted: 2026-10-04 · Original record ↗

arxiv
2610.04911
Explicit source versions (1)
  • v1 · 2026-10-04
  1. Preprint available

    Source date: 2026-10-04 · First detected: 2026-10-25 UTC · Historical baseline

    Source evidence (1)
Abstract · arxiv

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.