Researchers introduce CRAFT, a method for localizing and mitigating failures in medical vision language models. CRAFT identifies two failure modes: arbitration failure and brake failure, and provides a way to excise problematic attention heads to improve model performance.