Measuring and Improving Chain-of-Thought Reasoning in Vision-Language Models

Vision-language models (VLMs) can effectively act as visual assistants, interpreting questions about images and producing human-like responses. This work explores their abilities to demonstrate human-like reasoning. To address concerns about the consistency of VLMs’ reasoning, we introduce a chain-of-thought (CoT) consistency measure. We tackle the challenge of extensive human annotations by proposing an LLM-Human-in-the-Loop pipeline. Based on this pipeline, we build the CURE benchmark to measure both the zero-shot reasoning performance and consistency of VLMs. We evaluate state-of-the-art VLMs and find that even the best-performing model is unable to demonstrate strong visual reasoning capabilities and consistency, indicating that substantial efforts are required to enable VLMs to perform visual reasoning as systematically and consistently as humans. As an early step, we propose a two-stage training framework aimed at improving both the reasoning performance and consistency of VLMs without human annotations. The framework consists of two primary stages: supervised fine-tuning and learning from feedback, to guide VLMs in generating reasoning chains that exhibit both consistency and groundedness. Our framework exhibits a 4% relative improvement in reasoning performance and consistency. We release the dataset at https://github.com/Yangyi-Chen/CoTConsistency.