LongVideoAgent: Multi-Agent Reasoning with Long Videos

Runtao Liu*, Ziyi Liu*, Jiaqi Tang, Yue Ma, Renjie Pi, Jipeng Zhang, Qifeng Chen
Hong Kong University of Science and Technology
{rliuay@connect.ust.hk, ziyiliu0811@outlook.com}
* Equal Contribution. Work done during Ziyi's internship in HKUST.

Abstract

Recent advances in multimodal LLMs and systems that use tools for long-video QA point to the promise of reasoning over hour-long episodes. However, many methods still compress content into lossy summaries or rely on limited toolsets, weakening temporal grounding and missing fine-grained cues. We propose LongVideoAgent, a multi-agent framework in which a master LLM coordinates a grounding agent to localize question-relevant segments and a vision agent to extract targeted textual observations. The master agent plans with a step limit, and is trained with reinforcement learning to encourage concise, correct, and efficient multi-agent cooperation. This design helps the master agent focus on relevant clips via grounding, complements subtitles with visual detail, and yields interpretable trajectories. On our proposed LongTVQA and LongTVQA+ which are episode-level datasets aggregated from TVQA/TVQA+, our multi-agent system significantly outperforms strong non-agent baselines. Experiments also show reinforcement learning further strengthens reasoning and planning for the trained agent.

Overview

Traditional single-pass MLLMs that ingest entire long videos in one context—typically (may through heavy downsampling and compression) often miss crucial evidence and produce wrong answers, whereas LongVideoAgent conducts multi-agent, multi-round, and multimodal reasoning to extract sparse, task-relevant cues and answer correctly.

Overview Teaser

Method: Multi-Agent Framework

Architecture

Architecture of LongVideoAgent. A MasterAgent runs for up to \(K\) rounds, collaborating with a GroundingAgent to localize relevant clips from videos and a VisionAgent to read fine-grained cues from the localized frames. Evidence accumulates until the MasterAgent feels confident to answer the user.

Iterative Reasoning Loop

Unlike single-pass models, LongVideoAgent operates in a bounded loop (max \(K\) steps). At each step, the MasterAgent generates a "thinking" trace and emits a structured action token:

  • <request_grounding>: Calls the GroundingAgent to localize relevant video segments based on subtitles. The agent returns a symbolic tag <clip_X>.
  • <visual_query>: Calls the VisionAgent to extract specific visual details (objects, actions, text) from the localized clip. The agent returns textual observations.
  • <answer>: Terminates the loop and provides the final response when sufficient evidence is gathered.

Reinforcement Learning (GRPO)

We optimize the MasterAgent using Group Relative Policy Optimization (GRPO). The training objective includes: 1. Structural validity. 2. Answer Correctness: Rewarding the agent for reaching the correct final answer.

Experimental Results

We evaluate LongVideoAgent on LongTVQA and LongTVQA+, which are episode-level datasets.

Main Results

Main Results

Performance on LongTVQA and LongTVQA+. The left block lists model attributes (Agentic, Input, RL fine-tune); the right block reports validation accuracy (%). GPT-4o and Gemini-2.5 Pro are multimodal baselines that process and accept the full long video directly. Methods labeled Agentic indicate the model operates as the MasterAgent; methods labeled AgenticRL additionally denote RL fine-tuning. Parenthesized green numbers denote absolute gains over the immediately preceding (non-agentic or non-RL) setting. We observe that: (i) our multi-agent framework, LongVideoAgent, consistently outperforms the non-agentic counterparts; (ii) agentic RL yields additional gains, especially for smaller open-source models; (iii) using frames provides visual evidence beyond subtitles, and generally outperforms subtitle-only inputs; (iv) closed-source models remain strong, but the gap narrows much when open-source models adopt agentic designs and agentic RL.

Ablation Analysis

We conduct comprehensive ablation studies to validate our design choices. First, we observe that both grounding and vision agents are essential, with the full multi-agent system achieving the highest accuracy. Second, increasing the reasoning step limit \(K\) improves performance until saturation, confirming the value of iterative planning. Finally, stronger vision backbones and larger temporal windows provide richer context, further boosting the agent's reasoning capabilities.

Ablation Analysis

BibTeX

@misc{liu2025longvideoagentmultiagentreasoninglong,
      title={LongVideoAgent: Multi-Agent Reasoning with Long Videos}, 
      author={Runtao Liu and Ziyi Liu and Jiaqi Tang and Yue Ma and Renjie Pi and Jipeng Zhang and Qifeng Chen},
      year={2025},
      eprint={2512.20618},
      archivePrefix={arXiv},
      primaryClass={cs.AI},
      url={https://arxiv.org/abs/2512.20618}, 
}