VideoSeek：具备工具引导搜索能力的长时程视频智能体
VideoSeek: Long-Horizon Video Agent with Tool-Guided Seeking
把「视频逻辑流」当作主动搜索线索，用 think-act-observe 循环加多粒度工具按需取帧，替代对密集采样帧的穷举解析；做长视频 VQA 或视频推理时可直接复用这套按需取帧的智能体框架，在帧数大幅下降的同时保持甚至提升精度。
原文依据 actively seek answer-critical evidence instead of exhaustively parsing the full video
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论视频智能体模型推动了具有挑战性的视频-语言任务的发展。然而，大多数智能体方法仍然严重依赖对密集采样视频帧的贪心解析，导致计算成本高昂。我们提出 VideoSeek，一种长时程视频智能体，它利用视频逻辑流主动搜索对答案至关重要的证据，而不是穷举式地解析整个视频。这一思路使模型能够使用远少的帧数，同时保持甚至提升其视频理解能力。VideoSeek 在「思考-行动-观察」循环中运行，并配备精心设计的工具包以采集多粒度的视频观测。该设计支持在累积观测之上进行查询感知的探索，并支撑实用的视频理解与推理。在四个具有挑战性的视频理解与推理基准上的实验表明，VideoSeek 在取得较高准确率的同时，所用帧数远少于先前的视频智能体和独立 LMM。值得注意的是，VideoSeek 在其基座模型 GPT-5 的基础上于 LVBench 上取得 10.2 个绝对点的提升，同时使用的帧数减少了 93%。进一步的分析凸显了利用视频逻辑流、强大的推理能力以及工具包设计的互补作用的重要意义。
英文原文摘要 1256 字符
Video agentic models have advanced challenging video-language tasks. However, most agentic approaches still heavily rely on greedy parsing over densely sampled video frames, resulting in high computational cost. We present VideoSeek, a long-horizon video agent that leverages video logic flow to actively seek answer-critical evidence instead of exhaustively parsing the full video. This insight allows the model to use far fewer frames while maintaining, or even improving, its video understanding capability. VideoSeek operates in a think-act-observe loop with a well-designed toolkit for collecting multi-granular video observations. This design enables query-aware exploration over accumulated observations and supports practical video understanding and reasoning. Experiments on four challenging video understanding and reasoning benchmarks demonstrate that VideoSeek achieves strong accuracy while using far fewer frames than prior video agents and standalone LMMs. Notably, VideoSeek achieves a 10.2 absolute points improvement on LVBench over its base model, GPT-5, while using 93% fewer frames. Further analysis highlights the significance of leveraging video logic flow, strong reasoning capability, and the complementary roles of toolkit design.