Right Diagnoses, Decorative Reasoning: A Perturbation Audit of Medical Chain-of-Thought investigates whether the reasoning chains produced by Large Language Models (LLMs) in medical contexts actually drive their final answers or if they are merely decorative text. The authors, Mengzhu Xu, Jifan Gao, Xia Jiang, Yaoxin Wu, and Xi Long, find that in most cases, these chains do not reflect the clinical evidence provided in the questions.
The Problem with Medical Reasoning Chains
Clinicians often rely on Chain-of-Thought (CoT) rationales to verify the accuracy of medical AI. However, current evaluations treat these chains as "black boxes," assuming that if the final answer is correct, the reasoning must be sound. The authors argue that if a model’s rationale does not track the underlying clinical evidence, the model is not truly reasoning; it is simply narrating an answer after the fact.
How the Audit Works
The researchers developed a medical perturbation audit to test if LLM reasoning is "load-bearing." They used a battery of 30 operators to modify either the reasoning chain or the medical question itself. These edits included clinically significant changes such as:
Severity Reversal: Changing "acute" to "chronic." * Negation Flip: Negating clinical hinge phrases.
Demographic Swap: Changing age or sex markers.
Evidence Ablation: Removing specific factual sentences.
The team then performed a joint analysis, tracking whether the model updated its reasoning chain and whether it changed its final answer in response to these edits.
Key Findings
The study applied this framework to 14 LLMs across four medical benchmarks. The results suggest that medical CoT is largely decorative:
Chain-Decoupling Rate (CDR): The researchers found a 72.9% panel-wide decoupling rate. This means that in nearly 73% of cases involving destructive clinical edits, the model neither updated its reasoning chain to reflect the new information nor changed its final answer.
Accuracy Stability: Corrupting the reasoning chain or removing CoT prompting entirely did not significantly reduce the models' accuracy.
Clinician Validation: Two board-certified clinicians reviewed 197 perturbed questions and confirmed that 98.5% of the edits left the original gold-standard answer defensible, meaning the models' failure to react to these changes was not a result of the edits making the questions unanswerable.
Franklin Analysis
The evidence provided by the authors suggests a significant gap between the appearance of medical reasoning and the actual computational process of these models. By introducing the Chain-Decoupling Rate (CDR) as a metric, the paper provides a specific, measurable way to distinguish between models that are "faithful and robust" and those that are "decoupled and narrative." The consistency of these results across open-weight, reasoning-tuned, and closed-source models indicates that this decoupling is a widespread characteristic of current medical LLMs rather than an issue isolated to a specific architecture.
Comments (0)
to join the discussion
No comments yet
Be the first to share your thoughts!