Medical vision-language models can describe what they see in a chest X-ray, but they struggle with why. Chain-of-thought reasoning helps — the model explains its steps — but without causal structure, the “reasoning” follows spurious correlations. A pleural effusion appears alongside cardiomegaly often enough that the model learns to predict one from the other, even when the causal relationship in a specific patient runs differently.
Yan et al. (arXiv:2603.23085) build MedCausalX, which explicitly models causal chains in medical reasoning. The architecture introduces causal tokens and verify tokens that let the model autonomously decide when to perform causal analysis and when to check its own reasoning. A contrastive dataset (CRMed) provides paired examples: the correct causal chain and a spurious one, with fine-grained anatomical annotations showing where the reasoning diverges.
The training uses error-attributed reinforcement learning — when the model makes a mistake, the loss function attributes the error to the specific step in the causal chain that went wrong, not to the final prediction. This allows targeted correction of reasoning failures rather than blunt retraining.
Results: +5.4 points in diagnostic consistency, -10+ points in hallucination rate, and the best spatial grounding (IoU) across benchmarks.
The through-claim: medical reasoning errors aren't perception errors — they're causal reasoning errors. The model sees the finding correctly but connects it to the wrong cause. Fixing this requires causal structure in the reasoning chain, not better vision. The hallucination lives in the arrow between observations, not in the observations themselves.