aiqbench analysis
This research introduces CRAFT, a framework that isolates specific attention heads responsible for two critical safety failures in medical vision-language models: arbitration failure (text overriding correct visual data) and brake failure (confident output despite insufficient evidence). By demonstrating that these failure modes are mediated by spatially disjoint head populations, the study enables targeted interventions that correct errors without retraining or degrading performance on clean inputs. This mechanistic insight provides a concrete pathway for improving the reliability of AI systems in high-stakes clinical environments.
CRAFT: Causal Responsibility and Failure Tracing in Medical Vision Language Models
arXiv:2609.38810v1 Announce Type: cross 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.
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.
Computer Vision and Pattern Recognition (cs.CV); Artificial Intelligence (cs.AI)
arXiv:2609.38810 [cs.CV]
(or arXiv:2609.38810v1 [cs.CV] for this version)
https://doi.org/10.48550/arXiv.2609.38810 Focus to learn more arXiv-issued DOI via DataCite (pending registration)
arXivLabs is a framework that allows collaborators to develop and share new arXiv features directly on our website.
Both individuals and organizations that work with arXivLabs have embraced and accepted our values of openness, community, excellence, and user data privacy. arXiv is committed to these values and only works with partners that adhere to them.
Have an idea for a project that will add value for arXiv's community? Learn more about arXivLabs.
Full text under the publisher's open licence · 343 words
Related coverage
Game Design Creativity evals
LessWrong · Oct 12, 2026
Building trust in AI agents: A production-ready evaluation framework with Amazon Bedrock AgentCore
AWS · Oct 16, 2026
AI in healthcare: applications and best practices
Databricks · Jul 24, 2026
Generate Trajectories, Reasoning Traces, and Auto-Labels with NVIDIA Alpamayo 2 Super
NVIDIA · Aug 4, 2026