arXiv new; ECCV comment; long-video evidence grounding · P2 · 2026-06-30

Reflect-R1：它和通用视觉自监督的关系在于：Reflect-R1 通过检索客观视觉证据做自校正，和视频 VLM 的 grounding/post-training 相关

中高相关；详见方法、贡献和实验边界。

Reflect-R1: Evidence-Driven Reflection for Self-Correction in Long Video Understanding　arXiv new; ECCV comment; long-video evidence grounding　原文链接

编号2606.27922 优先级P2 类别arXiv new; ECCV comment; long-video evidence grounding 会议arXiv new; ECCV comment; long-video evidence grounding 方法Reflect-R1 通过检索客观视觉证据做自校正，和视频 VLM 的 grounding/post-training 相关来源arXiv / OpenReview

先说结论。它和通用视觉自监督的关系在于：Reflect-R1 通过检索客观视觉证据做自校正，和视频 VLM 的 grounding/post-training 相关。中高相关；详见方法、贡献和实验边界。

**Figureure 1 · : Comparison between Internal Closed-Loop Reflection and Evidence-Driv**Fig. 1: Comparison between Internal Closed-Loop Reflection and Evidence-Driven Reflection. (a) Traditional closed-loop reflection relies solely on internal parametric knowledge, easily falling into the trap of blind confidence and failing to correct errors. (b) Reflect-R1 completely breaks the hallucination loop by formalizing an “intuition-verification-arbitration” pipeline, executing active searches to achieve genuine self-correction strictly grounded in objective retrieved evidence.这张图概括 Reflect-R1 的整体方法流程。阅读时先看模块之间传递的训练信号，再看作者如何把目标拆成可优化的子问题。

**Figureure 3 · : Decoupled training prevents policy coupling**Fig. 3: Decoupled training prevents policy coupling. (a) Joint end-to-end training collapses reflection into a trivial identity mapping of the initial intuition. (b) Our decoupled strategy stabilizes the preceding distribution, enabling the model to learn robust error-correction logic and achieve a widening performance gap.这张图/表用于判断 Reflect-R1 的实验收益来自哪里。重点看替换、消融或跨模型设置下趋势是否一致，而不是只看单个最高分。

核心问题

它和通用视觉自监督的关系在于：Reflect-R1 通过检索客观视觉证据做自校正，和视频 VLM 的 grounding/post-training 相关。

方法拆解

Reflect-R1 通过检索客观视觉证据做自校正，和视频 VLM 的 grounding/post-training 相关

主要贡献

中高相关；详见方法、贡献和实验边界。

实验看点

实验部分建议重点看两类证据：一是作者是否把方法收益和更强数据、更长训练、更大模型区分开；二是跨模型、跨数据或跨任务迁移是否还能保留同样趋势。

局限与读法

这篇论文的结论需要结合任务设置、训练数据规模和消融实验一起看；不要只凭单个指标判断它对通用视觉表征的价值。