Griffin Farrow, Lily Sijia Li, Jack Johnson, Tingyan Wang, Philip Torr, William Bolton, Fabio J. Fehr
Hallucinations can undermine clinician trust in LLMs, making it important that evaluation methods capture clinically relevant errors. Rubric-based evaluation has become the leading approach for assessing LLMs in medicine, but it is unclear whether rubric scores reflect such errors. We first study this in a controlled setting using MedHallu, finding that more specific rubrics better distinguish correct from hallucinated responses. To test this systematically, we develop a taxonomy of medical hallucination types and a clinician-validated error-injection pipeline that creates matched correct and error-injected responses. Across HealthBench, HealthBench Professional, and LiveMedBench, our clinically relevant hallucinations are missed by rubrics, often leaving scores unchanged. We find that rubrics are most effective when explicitly checking facts, and are less effective for additional or unexpected errors they do not anticipate. A preliminary retrieval-based factuality check recovers some of the rubric-blind errors, suggesting a complementary approach. These findings reveal systematic blind spots in current medical evaluation of LLMs and suggest that rubric scores alone are insufficient to establish clinical reliability, potentially undermining clinician trust and confidence in clinical deployment.
Jack A. Johnson, Meghali Banerjee University of Oxford, Joseph Crawford University of Oxford, James Welch University of Oxford, Jim Davies NIHR Oxford Biomedical Research Centre, Oxford, Tingyan Wang University of Oxford
Patient health literacy is critical to health outcomes, yet medical discharge summaries remain inaccessible to many patients due to jargon and complex language. Large language models (LLMs) offer a promising means of bridging this gap, but their deployment in resource-constrained hospital environments demands lightweight, privacy-preserving solutions. We evaluate a range of open- and closed-source LLMs on the MeDiSumQA dataset, comprising real patient discharge summaries paired with lay questions and clinician-verified answers, and demonstrate that larger open-source models achieve accuracy and semantic similarity performance comparable to GPT-5. We then introduce LAMP-MedQA, a lightweight multi-agent framework for patient-oriented medical question answering. The framework decomposes the task into two sequential stages: question-relevant evidence extraction and patient-facing answer simplification. Each stage is governed by an automated, metric-driven feedback loop that enables iterative self-correction without human-in-the-loop supervision. Using Qwen2.5-7B-Instruct for generation agents and Phi-3.5-Mini-Instruct for reviewer/verifier agents, it achieves significantly lower FKGL than zero-shot GPT-5, indicating better readability, and obtains the highest simplification quality (SARI) among all evaluated models, while remaining broadly competitive on accuracy and semantic similarity. This competitiveness is further improved by an offline medical glossary, which narrows the gap in n-gram overlap and contextual-similarity metrics. These results suggest that collaborative lightweight agents represent a viable approach to improving health literacy in clinical settings. Our code is available at: https://github.com/JackJ3636/LAMP-MedQA
尚无 AI 解读，展示原文摘要
cs.CL · cs.MA
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 4: Student Research Workshop) · 2026 · 正式发表