REVISOR：超越文本反思，面向长视频理解的多模态内省推理
REVISOR: Beyond Textual Reflection, Towards Multimodal Introspective Reasoning in Long-Form Video Understanding
把纯文本的自我反思扩展为跨文本与视觉片段的反思过程，并用 Dual Attribution Decoupled Reward 在 GRPO 中约束推理与所选视频证据的因果对齐；做长视频 VQA 或多模态推理的工作可以直接借用这套「视觉片段回溯 + 解耦奖励」的强化学习框架，无需额外监督微调或外部模型。
原文依据 we propose REVISOR (REflective VIsual Segment Oriented Reasoning), a novel framework for tool-augmented multimodal reflection
摘要译文
由模型翻译 · 术语统一 · 请对照原文核对关键结论依赖纯文本重思过程的自反思机制在大多数多模态任务中表现良好。然而，当直接应用于长视频理解场景时，它们表现出明显的局限。其根本原因有两点：（1）长视频理解涉及更丰富、更动态的视觉输入，这意味着仅对文本信息进行重思是不够的，还需要一个专门针对视觉信息的进一步重思过程；（2）纯文本的反思机制缺乏跨模态交互能力，使其在反思过程中无法充分整合视觉信息。受这些洞见启发，我们提出 REVISOR（REflective VIsual Segment Oriented Reasoning，反思性视觉片段导向推理），一个面向工具增强多模态反思的新框架。REVISOR 使多模态大语言模型（MLLM）能够跨文本与视觉模态协同构建内省式反思过程，显著增强其长视频理解推理能力。为确保 REVISOR 在强化学习过程中学会准确回顾与问题高度相关的视频片段，我们设计了双归因解耦奖励（Dual Attribution Decoupled Reward, DADR）机制。该机制集成到 GRPO 训练策略中，强制模型推理与所选视频证据之间形成因果对齐。值得注意的是，REVISOR 框架无需额外的监督微调或外部模型，即可显著提升 MLLM 的长视频理解能力，并在 VideoMME、LongVideoBench、MLVU 和 LVBench 四个基准上取得令人印象深刻的结果。
英文原文摘要 1667 字符
Self-reflection mechanisms that rely on purely text-based rethinking processes perform well in most multimodal tasks. However, when directly applied to long-form video understanding scenarios, they exhibit clear limitations. The fundamental reasons for this lie in two points: (1) long-form video understanding involves richer and more dynamic visual input, meaning rethinking only the text information is insufficient and necessitates a further rethinking process specifically targeting visual information; (2) purely text-based reflection mechanisms lack cross-modal interaction capabilities, preventing them from fully integrating visual information during reflection. Motivated by these insights, we propose REVISOR (REflective VIsual Segment Oriented Reasoning), a novel framework for tool-augmented multimodal reflection. REVISOR enables MLLMs to collaboratively construct introspective reflection processes across textual and visual modalities, significantly enhancing their reasoning capability for long-form video understanding. To ensure that REVISOR can learn to accurately review video segments highly relevant to the question during reinforcement learning, we designed the Dual Attribution Decoupled Reward (DADR) mechanism. Integrated into the GRPO training strategy, this mechanism enforces causal alignment between the model's reasoning and the selected video evidence. Notably, the REVISOR framework significantly enhances long-form video understanding capability of MLLMs without requiring supplementary supervised fine-tuning or external models, achieving impressive results on four benchmarks including VideoMME, LongVideoBench, MLVU, and LVBench.