VLM Reasoning Chains Reveal Epistemic Signals for Uncertainty

New research investigates how Visual Language Models (VLMs) express uncertainty, particularly when operating in a "thinking mode" where they generate reasoning chains. The study empirically characterized the behavior of answer entropy in these models, identifying three distinct patterns across different VLM architectures like Qwen3-VL-8B-Thinking, GLM-4.1V-9B-Thinking, and InternVL3-8B. A key finding is that the entropy within these reasoning chains provides a more robust signal for uncertainty quantification compared to the entropy of the final answer token distribution. This holds true across various models and is particularly pronounced for free-form answers and harder reasoning tasks. The research also documents structured abstention behaviors in VLMs, suggesting a practical abstention gate can significantly improve accuracy by allowing models to decline uncertain queries.