医学微调重塑而非降低面向患者的AI风险:基于 rubric 的大语言模型评判评估框架
Medical Fine-Tuning Reshapes, Not Reduces, Patient-Facing AI Risk: A Rubric-Based, LLM-Judge Evaluation Framework
AI总结:
该研究提出基于 rubric 的 LLM 评判评估框架,发现医学微调重塑而非降低面向患者的 LLM 医学问答风险,且风险语言与事实幻觉率统计独立。
摘要:
面向患者的大型语言模型(LLM)医学问答正从研究演示转向生产部署,但评估仍以事实正确性为核心。事实正确性是必要条件但非充分条件:一个回答可能符合事实,却仍会给出若执行会伤害患者的指令。我们提出一种风险敏感评估框架,依据已发表的临床伤害分类法,通过六种风险语言类别(治疗指令、禁忌症、剂量表述、紧急分诊提示、高警示药物提及、过度自信断言)的存在性及严重程度对回答评分,严重程度由三评判者LLM集成体分类。我们将该框架应用于6个7-9B指令微调模型,其中3个通用模型(Mistral-7B、Phi-3.5-mini、Llama-3.1-8B)、3个医学适配模型(BioMistral-7B、OpenBioLLM-Llama3-8B、Med42-Llama3-8B),测试集为K-QA基准中的200个真实患者问题。得出两项发现:其一,风险语言集成与事实幻觉率统计独立(合并斯皮尔曼ρ=-0.065,n=300,95%置信区间[-0.18, +0.05];距离相关系数=0.10,独立性检验p=0.24);其二,医学微调未降低总体风险,而是重塑风险:持续预训练适配模型(BioMistral、OpenBioLLM)持续抑制高警示药物提及,却放大过度自信断言,而临床指令微调模型(Med42)仅抑制剂量相关风险,未放大过度自信。各类别Krippendorff α值在四个语言锚定类别中为中等至良好(0.53-0.71),在两个临床判断依赖类别中较低(0.19、0.31),明确了自动安全评估当前可靠与不可靠的领域。我们发布了该 rubric、评判者提示及评估代码。
英文摘要:
Patient-facing medical question answering with large language models (LLMs) is moving from research demonstrations to production deployments, yet evaluation remains anchored on factual correctness. Factuality is necessary but not sufficient: a response can be factually grounded and still issue a directive that, if acted on, would harm the patient. We introduce a risk-sensitive evaluation framework that scores responses by the presence and severity of six categories of risk-bearing language (treatment directives, contraindications, dosage expressions, urgency-triage cues, high-alert medication mentions, and overconfident assertions), anchored on published clinical harm taxonomies. Severity is classified by a three-judge LLM ensemble. We apply this framework to six 7-9B instruction-tuned models, three general (Mistral-7B, Phi-3.5-mini, Llama-3.1-8B) and three medically adapted (BioMistral-7B, OpenBioLLM-Llama3-8B, Med42-Llama3-8B), on 200 real patient questions from the K-QA benchmark. Two findings emerge. First, ensemble risk-bearing language is statistically independent of factual hallucination rate (pooled Spearman $ρ= -0.065$, $n = 300$, 95% CI $[-0.18, +0.05]$; distance correlation $= 0.10$, independence test $p = 0.24$). Second, medical fine-tuning does not reduce aggregate risk but reshapes it: continued-pretraining adaptations (BioMistral, OpenBioLLM) consistently suppress high-alert medication mentions while amplifying overconfident assertions, whereas clinical instruction-tuning (Med42) suppresses dosing only, without amplifying overconfidence. Per-category Krippendorff $α$ is fair to good for the four linguistically anchored categories (0.53-0.71) and low for the two clinical-judgment-dependent categories (0.19, 0.31), indicating where automatic safety evaluation is and is not yet reliable. We release the rubric, judge prompts, and evaluation code.
机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。