Language models learn to block distracting medical sentences
EXCISE: Locating and Excising Distracting Sentences in Medical Question Answering
Computation and Language
Summary
Medical question-answering models can get confused by sentences that don’t relate to the question, leading to wrong answers. The authors show that if the model is told to ignore these distracting sentences, it can recover the right answer without harming correct cases. They created a method called EXCISE that automatically detects and blocks these irrelevant sentences during answering, even without knowing in advance which sentence distracts. This approach improved accuracy significantly on their new test set and other medical benchmarks compared to previous methods.
What this means in practice
- •For clinical ai developers: Improve medical QA systems by automatically identifying and blocking misleading sentences to increase answer accuracy.
- •For healthcare chatbot teams: Enhance robustness of chatbots answering medical questions by reducing the impact of irrelevant information without prior manual labeling.
Authors
Yerim Hong, Sohyun Son, Heeseung Chung, Mujeen Sung
Abstract
Language models are increasingly considered for clinical decision support, where their answers should be grounded in clinical evidence and remain robust to irrelevant information. Prior work shows that a single irrelevant sentence can lower the accuracy of medical models, but not where inside the model its effect can be removed, or how to do so without knowing which sentence it is. We find that blocking attention to a known distractor restores the answer, whereas head masking can suppress correct answers. We also show that the model's hidden states mark the distractor as the sentence that least fits the case even when the model follows it. To remove a distractor's effect without knowing which sentence it is, we propose EXCISE, an inference-time method that blocks the sentence most influencing the answer, but only if the model's hidden states mark it as inconsistent with the case. On MedQA-TD (MedQA with Targeted Distractors), a benchmark we construct in which each distractor targets a specific wrong option, EXCISE gains 16.7-22.8 accuracy points at a 1.5-2.1-point cost on the same questions without distractors, outperforming prompting, rewriting, and steering baselines for all three models. Its gains extend to MedDistractQA and BiasMedQA, as well as chain-of-thought generation.