Medical fine-tuning reshapes rather than reduces patient-facing clinical risk in language models

Evaluating six 7-9B models across six categories of clinical harm reveals that risk-bearing language is statistically independent of factuality, with domain adaptation suppressing drug mentions while amplifying overconfidence.

Primary source: arXiv:2610.24807: https://arxiv.org/abs/2610.24807

What’s new: Benchmarking medical large language models has traditionally focused almost exclusively on factual correctness against medical examination benchmarks or curated question datasets. However, high factual accuracy does not guarantee patient safety: an AI model can articulate entirely accurate pathophysiological facts while simultaneously issuing reckless directives, misjudging triage urgency, or offering inappropriate dosage advice. In Medical Fine-Tuning Reshapes, Not Reduces, Patient-Facing AI Risk: A Rubric-Based, LLM-Judge Evaluation Framework (arXiv:2610.24807, 20 Oct 2026), researcher Savan Doshi introduces a clinical risk evaluation framework that systematically audits responses across six risk categories: treatment directives, contraindications, dosage expressions, urgency-triage cues, high-alert medication mentions, and overconfident assertions.

Using a three-judge LLM ensemble anchored on clinical harm taxonomies, the study evaluates six 7B to 9B instruction-tuned models (general-purpose base models including Mistral-7B, Phi-3.5-mini, and Llama-3.1-8B, alongside medically adapted variants BioMistral-7B, OpenBioLLM-Llama3-8B, and Med42-Llama3-8B) across 200 authentic patient inquiries from the K-QA benchmark. The findings show that risk-bearing language is statistically independent of factual hallucination rates (pooled Spearman rho = -0.065, p = 0.24). Furthermore, medical adaptation does not decrease aggregate clinical risk; instead, it shifts the risk distribution. Continued pre-training adaptations successfully curb dangerous high-alert medication mentions but significantly inflate overconfident assertions. Meanwhile, clinical instruction-tuning effectively suppresses unverified dosage expressions without increasing overconfidence.

Why it matters: Deploying LLMs in patient-facing clinical contexts requires evaluating communicative safety alongside informational accuracy. Because clinical risk is orthogonal to factual hallucination, optimizing models solely for leaderboard factuality benchmarks can create a false sense of clinical readiness. Understanding how different adaptation techniques redistribute risk profiles allows developers to combine targeted supervised instruction-tuning with strict guardrails against overconfident directive language.

Caveats: The evaluation relied on simulated LLM ensembles as primary judges calibrated against clinical rubrics rather than multi-specialist physician panels grading each query manually. In addition, the study examined 7B to 9B parameter models, leaving open whether 70B+ frontier models exhibit identical risk redistribution trade-offs under domain adaptation.