Stress-testing medical large language models reveals latent safety pathology beyond benchmark accuracy

• URL: http://arxiv.org/abs/2606.07929v1
• Date: Sat, 06 Jun 2026 01:39:14 GMT
• Title: Stress-testing medical large language models reveals latent safety pathology beyond benchmark accuracy
• Abstract summary: Large language models (LLMs) are entering clinical practice based on benchmark accuracy that may fail to detect safety-relevant failure modes.<n>Here we present AI-MASLD, a stress-audit framework that adapts the logic of metabolic stress testing from hepatology to the evaluation of clinical perturbations.
• Score: 25.403894453021817
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: Large language models (LLMs) are entering clinical practice based on benchmark accuracy that may fail to detect safety-relevant failure modes. Here we present AI-MASLD, a stress-audit framework that adapts the logic of metabolic stress testing from hepatology to the evaluation of clinical LLMs. Using 240 clinical cases across six narrative perturbation probes, we subjected seven models to double-stress testing and quantified performance through three indices: metabolic index (MI), perturbation flip rate (PFR), and counterfactual fairness index (CFI). Under clean baseline conditions, all models performed uniformly well. Under realistic narrative stress, performance diverged sharply, revealing two distinct stress-response phenotypes. Quantized models exhibited pseudonormalization, in which low flip rates hid functional collapse. Medical supervised fine-tuning systematically degraded logical stability, fairness, and information extraction. An open-weight model matched or exceeded proprietary alternatives on every safety dimension. These findings establish narrative stress auditing as a necessary complement to accuracy-based evaluation.

Related papers

• Beyond Scalar Scores: Exploring LLM-based Metrics for Clinical Significance Evaluation in Radiology Reports [49.5225801722164]
  Existing metrics obscure this requirement by reducing report quality to a medically ungrounded scalar.<n>We study this boundary using ReEvalMed benchmark as testbed and evaluate metric-level clinical significance.
  arXiv  Detail & Related papers  (2026-06-17T08:10:30Z)
• When Large Language Models Fail in Healthcare: Evaluating Sensitivity to Prompt Variations [1.0152838128195467]
  Large Language Models (LLMs) are increasingly used in healthcare for tasks such as clinical question answering, diagnosis support, and report summarization.<n>Despite their promise, these models remain highly sensitive to subtle prompt perturbations, both lexical and syntactic, posing serious risks in safety-critical clinical applications.<n>We categorize perturbations into natural and adversarial types and examine their effect on model consistency, accuracy, and reliability in clinical reasoning tasks.
  arXiv  Detail & Related papers  (2026-06-05T13:07:11Z)
• Counterfactual Stress Testing for Image Classification Models [12.520708194901735]
  Deep learning models in medical imaging often fail when deployed in new clinical environments.<n>A central challenge is underspecification, where models with similar validation performance exhibit divergent real-world failure modes.<n>We introduce a counterfactual stress testing framework based on causal generative models that create realistic "what if" images.
  arXiv  Detail & Related papers  (2026-05-11T17:36:16Z)
• Beyond Prompt: Fine-grained Simulation of Cognitively Impaired Standardized Patients via Stochastic Steering [58.85421622061983]
  We propose StsPatient for the fine-grained simulation of cognitively impaired patients.<n>We capture domain-specific features by extracting steering vectors from contrastive pairs of instructions and responses.<n>StsPatient significantly outperforms baselines in both clinical authenticity and severity controllability.
  arXiv  Detail & Related papers  (2026-04-14T02:37:46Z)
• Delving Aleatoric Uncertainty in Medical Image Segmentation via Vision Foundation Models [56.29123284262618]
  This work proposes leveraging the universal representation capabilities of visual foundation models to estimate inherent data uncertainty.<n>We analyze the feature diversity of the model's decoded representations and quantify their singular value energy to define the semantic perception scale for each class.<n>Based on this foundation, we design two uncertainty-driven application strategies: (1) the aleatoric uncertainty-aware data filtering mechanism to eliminate potentially noisy samples and enhance model learning quality; and (2) the dynamic uncertainty-aware optimization strategy that adaptively adjusts class-specific loss weights during training based on the semantic perception scale, combined with a label denoising mechanism to improve training stability
  arXiv  Detail & Related papers  (2026-04-13T03:59:54Z)
• The Stationarity Bias: Stratified Stress-Testing for Time-Series Imputation in Regulated Dynamical Systems [0.098314893665023]
  Time-series imputation benchmarks use random masking and shape-agnostic metrics.<n>We formalize this bias and propose a emphStratified Stress-Test that partitions evaluation into Stationary and Transient regimes.
  arXiv  Detail & Related papers  (2026-02-17T15:05:56Z)
• A systematic evaluation of uncertainty quantification techniques in deep learning: a case study in photoplethysmography signal analysis [1.6690512882610855]
  Deep learning models can be used to continuously monitor physiological parameters outside of clinical settings.<n>There is risk of poor performance when deployed in practical measurement scenarios leading to negative patient outcomes.<n>Here we implement eight uncertainty (UQ) techniques to models trained on two clinically relevant prediction tasks.
  arXiv  Detail & Related papers  (2025-10-31T22:54:13Z)
• EchoBench: Benchmarking Sycophancy in Medical Large Vision-Language Models [82.43729208063468]
  Recent benchmarks for medical Large Vision-Language Models (LVLMs) emphasize leaderboard accuracy, overlooking reliability and safety.<n>We study sycophancy -- models' tendency to uncritically echo user-provided information.<n>We introduce EchoBench, a benchmark to systematically evaluate sycophancy in medical LVLMs.
  arXiv  Detail & Related papers  (2025-09-24T14:09:55Z)
• Organ-Agents: Virtual Human Physiology Simulator via LLMs [66.40796430669158]
  Organ-Agents is a multi-agent framework that simulates human physiology via LLM-driven agents.<n>We curated data from 7,134 sepsis patients and 7,895 controls, generating high-resolution trajectories across 9 systems and 125 variables.<n>Organ-Agents achieved high simulation accuracy on 4,509 held-out patients, with per-system MSEs 0.16 and robustness across SOFA-based severity strata.
  arXiv  Detail & Related papers  (2025-08-20T01:58:45Z)
• Beyond Benchmarks: Dynamic, Automatic And Systematic Red-Teaming Agents For Trustworthy Medical Language Models [87.66870367661342]
  Large language models (LLMs) are used in AI applications in healthcare.<n>Red-teaming framework that continuously stress-test LLMs can reveal significant weaknesses in four safety-critical domains.<n>A suite of adversarial agents is applied to autonomously mutate test cases, identify/evolve unsafe-triggering strategies, and evaluate responses.<n>Our framework delivers an evolvable, scalable, and reliable safeguard for the next generation of medical AI.
  arXiv  Detail & Related papers  (2025-07-30T08:44:22Z)
• Robustness Stress Testing in Medical Image Classification [26.094688963784254]
  We employ stress testing to assess model robustness and subgroup performance disparities in disease detection models. We apply stress tests to measure the robustness of disease detection models for chest X-ray and skin lesion images. Our experiments indicate that some models may yield more robust and equitable performance than others.
  arXiv  Detail & Related papers  (2023-08-14T02:02:56Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.