Medical vision language models show distinct causes of diagnosis errors
CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models
Computer Vision and Pattern RecognitionArtificial Intelligence
Summary
Medical AI models that look at pictures and text can sometimes make mistakes by trusting wrong text over the image or giving confident answers without enough visual proof. The authors identified specific parts inside these models that cause these errors, which act separately. They showed that removing or adjusting these parts can fix the errors without hurting the model’s normal performance. This helps improve the safety of AI used in medical diagnosis.
What this means in practice
- •For medical ai developers: Improve the safety of clinical diagnosis systems by targeting specific internal components that cause wrong reliance on text or overconfidence.
- •For healthcare technology teams: Maintain trust in AI-assisted medical tools by reducing diagnostic errors without requiring costly retraining of models.
Authors
Chunzheng Zhu, Jiaqi Zeng, Hongbo Zhao, Yihang Chen, Yijun Wang, Jianxin Lin
Abstract
As vision language models are increasingly deployed in clinical diagnosis, understanding how they internally resolve competing visual and textual signals becomes a safety imperative. Existing mechanistic analyses remain confined to unimodal text and offer no explanation for why a single misleading sentence can override a correct image based diagnosis, or why a model commits to a confident answer despite insufficient visual evidence. We find that these two safety risks, arbitration failure where textual context overrides visual grounding and brake failure where the model commits without adequate evidence, are mediated by spatially disjoint attention head populations: arbitration heads form a mid-to-deep wideband reflecting cross-layer evidence competition, while brake heads concentrate in a narrow middle-to-late layer band that regulates evidence sufficiency and abstention behavior. To ground these observations in causal circuitry, we introduce CRAFT, which localizes each failure mode to a minimal causal head set via dual criteria and verifies necessity and sufficiency through temporal probes and Tuned Lens trajectory analysis. Excising arbitration heads sharply reduces conflict following with negligible degradation on clean inputs, while excising brake heads restores appropriate abstention under degraded visual evidence. The two interventions target spatially disjoint head sets and produce distinct corrective effects, underscoring the mechanistic separability of the failure modes. Experiments across multiple medical VQA benchmarks and VLM architectures validate both the localization and interventions, demonstrating that the identified heads causally drive each failure mode and that targeted modulation generalises without retraining. The code is available at GitHub repository.