Bradley Segal

Hongjian Zhou, Xinyu Zou, Jinge Wu et al.

Large language models (LLMs) now reach expert-level scores on medical licensing exams, encouraging the assumption that high scores imply safe medical judgment while patients increasingly use them for health advice. We show this assumption is fragile: when misleading context is injected into questions that LLMs originally answer correctly, they abandon the correct answer. We call the ability to maintain correct judgment under adversarial context epistemic resilience, and introduce MedMisBench to measure it. MedMisBench contains 10,932 medical question items and 48,889 misleading context-option pairs spanning medical reasoning, agentic capability, and patient-journey evaluation. Across 11 model configurations, mean accuracy falls from 71.1% on original questions to 38.0% under focused misleading context, with 51.5% attack success. The most damaging injections are formal, rule-like fabrications: authority-framed falsehoods reach 69.5% attack success and exception-poisoning claims reach 64.1%. A 14-member clinical panel from 7 countries identified serious potential harm in 38.2% of reviewed cases. MedMisBench exposes a structural blind spot in LLM evaluation in medical settings: existing benchmarks measure what models know, but not whether they preserve correct medical judgment under misleading context.

Bradley Segal, Joshua Fieggen, David Clifton et al. · oxford

Ensuring the generalisability of clinical machine learning (ML) models across diverse healthcare settings remains a significant challenge due to variability in patient demographics, disease prevalence, and institutional practices. Existing model evaluation approaches often rely on real-world datasets, which are limited in availability, embed confounding biases, and lack the flexibility needed for systematic experimentation. Furthermore, while generative models aim for statistical realism, they often lack transparency and explicit control over factors driving distributional shifts. In this work, we propose a novel structured synthetic data framework designed for the controlled benchmarking of model robustness, fairness, and generalisability. Unlike approaches focused solely on mimicking observed data, our framework provides explicit control over the data generating process, including site-specific prevalence variations, hierarchical subgroup effects, and structured feature interactions. This enables targeted investigation into how models respond to specific distributional shifts and potential biases. Through controlled experiments, we demonstrate the framework's ability to isolate the impact of site variations, support fairness-aware audits, and reveal generalisation failures, particularly highlighting how model complexity interacts with site-specific effects. This work contributes a reproducible, interpretable, and configurable tool designed to advance the reliable deployment of ML in clinical settings.