This study audits whether medical chain-of-thought (CoT) rationales in large language models actually reflect the reasoning that determines the answer. The authors applied a battery of 30 clinically motivated perturbations (such as reversing severity, flipping negation, swapping demographics, or ablating evidence) to questions from four medical QA benchmarks. They analyzed whether the CoT text and the model's answer updated appropriately after the edit. Across 14 models, they found that the chain often fails to register the edit and the answer does not flip, a pattern they call the Chain-Decoupling Rate (CDR). The average CDR was 72.9% on destructive edits. Further, corrupting the chain did not change accuracy, and removing CoT prompting altogether did not reduce accuracy. Two board-certified clinicians confirmed that 98.5% of 197 perturbed questions still had a defensible gold answer. The results hold across medical and reasoning fine-tuning, model scale, and closed-source models where only answer-side signals are observable.
Why this matters: CoT is often presented as transparent reasoning, especially in medical contexts where clinicians might rely on it for interpretability. This work shows that, for current models, the visible CoT does not faithfully represent the model's internal reasoning. It highlights a serious reliability gap for medical AI, where trust in explanations must be questioned. The framework and CDR metric offer a reusable way to audit CoT faithfulness.