EMNLP 2026 · Main Conference

VideoSearcher: Empowering Video Deep Research with Multi-Tool Agentic Reasoning via Reinforcement Learning

Zhenkun Gao^*1,2, Yicheng Bao^*1, Jinlong Peng^*†✉2, Xueheng Li^*✉3, Suyuan Huang^*4,2

Bangwei Liu¹, Kunquan Li⁵, Zhenye Gan², Tao Hu³, Chengjun Xie³, Mingqian Yang¹

Xuanhua He^✉6, Zhizhong Zhang¹, Xin Tan¹, Chengjie Wang², Yuan Xie^✉1

^* Equal Contribution, ^† Project Leader, ✉ Corresponding Author

¹East China Normal University, ²Tencent Youtu Lab, ³HFIPS, Chinese Academy of Sciences, ⁴Wuhan University, ⁵Xiamen University, ⁶The Hong Kong University of Science and Technology

arXiv Paper (PDF) Code Model Data BibTeX

VideoSearcher studies Video Deep Research, where an agent must localize visual evidence in videos, invoke external search tools, and synthesize answers through a single coherent reasoning trajectory.

Demo

A full VideoSearcher rollout: the agent localizes key frames in the video, zooms into the visual clue, launches image and web search, and integrates the retrieved evidence into a final answer.

Abstract

Video understanding is moving beyond closed-context perception toward open-world evidence exploration, a paradigm formalized as Video Deep Research (VDR). Existing multimodal search agents primarily target static images, and current VDR benchmarks rely on text-centric retrieval that discards crucial visual information. We propose VideoSearcher, a closed-loop agentic framework that empowers Vision-Language Models with multi-tool reasoning for VDR. VideoSearcher unifies temporal localization, spatial focusing, and multimodal search within a single reasoning trajectory, enabling agents to progressively ground visual clues, retrieve relevant evidence, and synthesize answers.

To optimize knowledge-intensive reasoning trajectories, we introduce Bi-branch Sequence Policy Optimization (BiSPO), which decouples tool-invocation optimization from answer-accuracy optimization. We further construct VideoSearch-QA, a benchmark for open-world video information grounding and multimodal search-based reasoning. Experiments show that VideoSearcher substantially improves over prior open-source agentic baselines across search-oriented and multimodal understanding benchmarks.

Method

VideoSearcher training and reasoning pipeline — VideoSearcher combines video grounding, spatial inspection, image search, and web search in a closed-loop reasoning process.

Video Grounding

The agent narrows long videos to relevant segments and locks key frames before searching for external evidence.

Multi-Tool Search

Image search and web search connect localized visual clues with open-world knowledge.

BiSPO Training

Answer correctness and tool-use behavior are optimized through separated reward branches.

Experiment Results

VideoSearcher-8B reaches an average score of 57.66% across 8 search-oriented benchmarks, outperforming the Qwen3-VL-8B agentic baseline by 15.71% on average.

Table 1: performance on search-oriented benchmarks — **Table 1.** Performance on search-oriented benchmarks under Direct Answer and Agentic Model workflows. Zero-shot denotes tool-augmented inference without task-specific training.

Table 2: performance on general video understanding benchmarks — **Table 2.** Performance on general video understanding benchmarks.

Table 3: algorithmic ablation of RL — **Table 3.** Algorithmic ablation of RL. GRPO and GSPO are single-branch variants; BiSPO is the proposed bi-branch variant, with BiSPO Full using the correctness-gated bell-shaped tool reward.

Table 4: tool ablation study — **Table 4.** Tool ablation study. Locate denotes the video localization tools, Search the external web retrieval tools.

Case Study

Qualitative examples illustrate how VideoSearcher builds a trajectory from video localization to visual entity search, evidence retrieval, and final answer synthesis.

Successful case 1: identifying a humanoid robot through visual grounding and external search.

Successful case 2: tracing a robot traffic police deployment location.

Failed case — Failure case: noisy retrieval feedback propagates into the final answer.

Citation

@inproceedings{gao2026videosearcher,
  title     = {VideoSearcher: Empowering Video Deep Research with Multi-Tool Agentic Reasoning via Reinforcement Learning},
  author    = {Gao, Zhenkun and Bao, Yicheng and Peng, Jinlong and Li, Xueheng and Huang, Suyuan and Liu, Bangwei and Li, Kunquan and Gan, Zhenye and Hu, Tao and Xie, Chengjun and Yang, Mingqian and He, Xuanhua and Zhang, Zhizhong and Tan, Xin and Wang, Chengjie and Xie, Yuan},
  booktitle = {Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP)},
  year      = {2026},
  eprint    = {2607.02927},
  archivePrefix = {arXiv}
}