Visually Grounded Self-Reflection for Vision-Language Models via Reinforcement Learning (VRRL)

Forum topic · 小凯 · 2026-07-06

Summary

Large vision-language models (LVLMs) can reason over multimodal inputs via textual chains of thought, but during self-reflection they often fail to properly attend to visual inputs, limiting their ability to turn feedback into grounded corrections—especially on out-of-distribution images. Researchers Liyan Tang, Fangcong Yin, and Greg Durrett propose VRRL, a reinforcement learning training framework with two components designed to elicit visually grounded self-reflection: (1) randomly masking trajectory prefixes during training to emphasize recovery from incorrect intermediate predictions rather than making early mistakes, and (2) buffer roll-ins from an experience replay buffer that expose the model to diverse failure states it must learn to correct. Evaluated on visually grounded tasks involving tables and charts plus a spatial navigation benchmark, VRRL substantially improves average out-of-distribution accuracy compared with standard RL and reflection-oriented fine-tuning baselines, which degrade significantly under distribution shift. Paper: arXiv 2607.02490.

Paper Overview

Field: NLP Authors: Liyan Tang, Fangcong Yin, Greg Durrett Published: 2026-07-02 arXiv: 2607.02490 Categories: cs.CL, cs.CV

Abstract

Large vision-language models can reason over multimodal inputs by generating textual chains of thought (CoT). A key capability exhibited in CoT reasoning is self-reflection: revisiting earlier decisions and correcting previous errors. However, existing LVLMs often fail to properly attend to visual inputs during reflection, limiting their ability to translate feedback into grounded corrections, especially for out-of-distribution images.

To address this issue, the authors propose VRRL, a novel reinforcement learning training framework with two components explicitly designed to elicit visually grounded self-reflection:

1. Random trajectory prefix masking during training, to emphasize recovery from incorrect intermediate predictions rather than making early mistakes. 2. Buffer roll-ins from an experience replay buffer, exposing the model to diverse failure states it must learn to correct.

Evaluation

The method is evaluated on visually grounded tasks involving tables and charts, as well as a spatial navigation benchmark. While off-the-shelf models and conventionally fine-tuned models degrade significantly under distribution shift, VRRL improves average out-of-distribution accuracy by effectively leveraging self-reflection, substantially outperforming standard RL and reflection-oriented fine-tuning baselines.