Chain-of-thought prompting has become a common way to improve reasoning in multimodal models, but it does not guarantee correct final answers. According to a new arXiv paper, the reasoning process can drift away from the image, and by the time the answer is generated, the trace is already fixed. This makes it hard to recover from earlier mistakes.

The paper, titled "The Image Still Has a Say: Final-Answer Repair for Multimodal Chain-of-Thought," proposes a repair step that revisits the final answer using the image. Rather than trying to correct the entire reasoning trace, the method focuses on the last stage of generation, where the image can still provide grounding.

The authors argue that this targeted repair is a useful complement to existing chain-of-thought approaches. The abstract does not include experimental results, so the effectiveness of the method is not yet detailed in the provided text.