Reflect-R1: Evidence-Driven Reflection for Self-Correction in Long Video Understanding

2026-06-26

AI Summary

This work addresses critical limitations in existing long-form video understanding methods, where multimodal reflection mechanisms often lack grounding in external objective evidence, leading to hallucinations and poor self-correction capabilities. Additionally, multi-stage reinforcement learning suffers from policy coupling and insufficient training data. To overcome these challenges, the authors propose Reflect-R1, a three-stage intuition–verification–arbitration reasoning pipeline that dynamically retrieves visual evidence to validate initial predictions and resolves conflicts through iterative temporal search, enabling evidence-driven self-correction. The framework introduces SD-GRPO, a novel stage-decoupled reinforcement learning algorithm, alongside a large-scale training set of 120,000 samples to mitigate policy coupling and data scarcity. Experiments demonstrate state-of-the-art performance on benchmarks such as VideoMME and LongVideoBench, with significantly improved factual correction rates.

Abstract

Current multimodal reflection mechanisms for long video understanding predominantly rely on closed-loop self-reflection within internal parameters. Lacking objective external evidence, models are frequently trapped in blind confidence and often fail to correct errors. Furthermore, applying reinforcement learning to multi-stage reflection pipelines introduces severe policy coupling, which is exacerbated by a critical scarcity of dedicated training data. To address these limitations, this work proposes Reflect-R1, the first Evidence-Driven self-correction framework for long video understanding. The framework constructs a three-stage pipeline consisting of intuition, verification, and arbitration. By dynamically retrieving objective visual evidence to verify initial intuitions and autonomously executing multiple temporal searches to resolve conflicts, it completely breaks the hallucination loop. To overcome policy coupling, we design a stage-decoupled reinforcement learning algorithm named SD-GRPO that independently computes advantage functions across different reasoning stages. Concurrently, we construct a dataset of 120K samples to bridge the training data gap. Extensive experiments on benchmarks such as VideoMME and LongVideoBench demonstrate that Reflect-R1 achieves state-of-the-art performance. Our method significantly improves the genuine rectification rate and enables authentic self-correction strictly grounded in objective evidence.

Problem

• long video understanding
• self-correction
• evidence-driven reflection
• policy coupling
• hallucination

Innovation

• Evidence-Driven Reflection
• Self-Correction
• Stage-Decoupled Reinforcement Learning
• Long Video Understanding
• Hallucination Mitigation