文章背景与核心概要

长视频问答（QA）系统在处理复杂时间跨度和连贯叙事时，常常受困于刻板的时间分块机制，并且往往过度关注局部的“语义相关性”而非真正的“证据充分性”。当关键的时间或因果线索缺失时，这种局限性会导致推理不完整。

为了解决这些痛点，本文介绍了 REVEAL 这一新型基于智能体的框架。它通过自适应内存构建、评分标准引导的验证以及针对性的重新检索机制，在无需额外训练的前提下，显著超越了现有的先进模型。这证明了对证据充分性进行显式验证是实现更可靠、更准确的长视频推理的关键所在。

REVEAL: A Rubric-Guided Agent for Explicit Evidence Sufficiency Verification in Long-Video Question Answering

Authors: Caijun Yan, Yang Zhou, Meixing Shi, Haoran Sun, Yichen Li, Yuxiang Cai, Yankai Jiang
Date: August 9, 2026
Primary Subject: Computer Vision and Pattern Recognition (cs.CV)
arXiv ID: 2608.08612

Summary

现有长视频问答（QA）系统通常难以应对刚性的时间分块，并且侧重于单纯的“语义相关性”而非“证据充分性”。当缺少关键的时间或因果线索时，这会导致推理不完整。

Existing long-video question answering (QA) systems often struggle with rigid temporal chunking and a focus on mere "semantic relevance" rather than "evidence sufficiency." This leads to incomplete reasoning when critical temporal or causal clues are missing.

REVEAL 是一个创新的基于智能体的框架，通过以下方式克服了这些局限性： * 自适应内存构建： 利用基于视觉相似度的预处理将帧分组为自然、连贯的事件单元，从而创建离线全局上下文和动态的、由问题条件触发的在线内存。 * 评分标准引导的验证： 采用自动生成的“评分标准库”来显式验证检索到的证据是否足以回答问题。 * 定向重新检索： 如果验证失败，智能体会识别缺失的信息并执行定向重新检索以填补空白。

无需额外训练，REVEAL 持续优于最先进的模型，证明了对证据充分性进行显式验证是实现更可靠、更准确的长视频推理的关键。

REVEAL is a novel agent-based framework that addresses these limitations by: * Adaptive Memory Construction: Utilizing visual-similarity-based preprocessing to group frames into natural, coherent event units, creating both offline global context and dynamic, question-conditioned online memory. * Rubric-Guided Verification: Employing an automatically generated "rubric library" to explicitly verify if retrieved evidence is sufficient to answer a question. * Targeted Re-retrieval: If the verification fails, the agent identifies missing information and performs targeted re-retrieval to fill the gaps.

Without requiring additional training, REVEAL consistently outperforms state-of-the-art models, demonstrating that explicit verification of evidence sufficiency is key to more reliable and accurate long-video reasoning.

Key Features

1. 自适应视频内存

与使用固定长度分块（例如 10 秒间隔）的传统方法不同，REVEAL 根据视觉连贯性构建内存。这确保了连续事件不会被碎片化，从而能够更准确地表示视频的叙事流程。

1. Adaptive Video Memory

Unlike traditional methods that use fixed-length chunks (e.g., 10s intervals), REVEAL constructs memory based on visual coherence. This ensures that continuous events are not fragmented, allowing for a more accurate representation of the video's narrative flow.

2. 证据充分性验证

REVEAL 的核心创新在于将其从“基于相关性”的检索转变为“基于充分性”的验证。通过使用评分标准库，智能体充当批评者的角色，确定当前证据是否足以支持正确的答案。

2. Evidence Sufficiency Verification

The core innovation of REVEAL is its shift from "relevance-based" retrieval to "sufficiency-based" verification. By using a rubric library, the agent acts as a critic, determining whether the current evidence is enough to support a correct answer.

3. 迭代推理

当智能体确定证据不足时，它不会简单地进行猜测。它能准确指出缺少的内容——无论是特定的动作、因果关系还是时间细节——并指导定向搜索以检索必要的补充信息。

3. Iterative Reasoning

When the agent determines that evidence is insufficient, it does not simply guess. It pinpoints exactly what is missing—whether it be a specific action, a causal link, or a temporal detail—and directs a targeted search to retrieve the necessary complementary information.

Access & Resources

• View PDF
• HTML (Experimental)
• TeX Source
• DOI: 10.48550/arXiv.2608.08612

Metadata