Paper Overview
Field: NLP Authors: Liyan Tang, Fangcong Yin, Greg Durrett arXiv: 2507.03230
Summary
Large vision-language models (LVLMs) can reason over multimodal inputs by generating textual chains of thought (CoT). A key capability 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, the paper proposes a novel reinforcement learning training framework VRRL, with two components explicitly designed to elicit visually grounded self-reflection:
1. Random prefix masking: trajectory prefixes are randomly masked during training, emphasizing recovery from incorrect intermediate predictions rather than merely avoiding early mistakes. 2. Buffered roll-ins: roll-ins sampled from an experience replay buffer expose the model to diverse failure states that it must learn to correct.
The approach is evaluated on visual grounding tasks involving tables and charts, as well as spatial navigation benchmarks. While off-the-shelf and conventionally fine-tuned models degrade significantly under distribution shift, VRRL substantially improves average out-of-distribution accuracy over standard RL and reflection-oriented fine-tuning baselines by effectively leveraging self-reflection.
Original 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, we propose a novel reinforcement learning training framework VRRL, with two components explicitly designed to elicit visually grounded self-reflection. First, we randomly mask trajectory prefixes during training to emphasize recovery from incorrect intermediate predictions rather than making early mistakes. Second, we introduce buffered roll-ins from an experience replay buffer to expose the model to diverse failure states that it must learn to correct. We evaluate our approach on visual grounding tasks involving tables and charts, and spatial navigation benchmarks. Our method significantly improves average OOD accuracy over standard RL and reflection-oriented fine-tuning baselines.