Paper Overview
Field: NLP Authors: Liyan Tang, Fangcong Yin, Greg Durrett arXiv: 2507.03230
Abstract
Large vision-language models (LVLMs) 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 a novel reinforcement learning training framework VRRL, 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 avoiding early mistakes. 2. Buffered roll-ins from an experience replay buffer, exposing the model to diverse failure states that it must learn to correct.
Evaluation
The approach is evaluated on visual grounding 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 achieves substantially higher average out-of-distribution accuracy by effectively leveraging self-reflection, outperforming both standard RL and reflection-oriented fine-tuning baselines.
Original abstract (abridged):
> 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... We evaluate our approach on visual grounding tasks involving tables and charts...