Same Facts, Different Diagnosis: Measuring and Mitigating Narrative Anchoring in Clinical Language Models
Same Facts, Different Diagnosis: Measuring and Mitigating Narrative Anchoring in Clinical Language Models
相同事实，不同诊断：临床语言模型中“叙事锚定”现象的测量与缓解
Abstract: Large language models used for clinical diagnostic reasoning are sensitive to sociolinguistic register, not just clinical content. We term this failure mode Narrative Anchoring: identical clinical facts expressed in different registers cause diagnostic outputs to diverge. 摘要： 用于临床诊断推理的大型语言模型不仅对临床内容敏感，还对社会语言语体（sociolinguistic register）敏感。我们将这种失效模式称为“叙事锚定”（Narrative Anchoring）：即表达相同临床事实的不同语体会导致诊断输出产生偏差。
Unlike prior demographic-bias work, which manipulates explicit identity tokens such as race or income, our benchmark isolates register as the sole channel of variation, with no demographic marker present in any form. 与以往针对人口统计学偏见的研究（通常操纵种族或收入等显性身份标记）不同，我们的基准测试将“语体”作为唯一的变量渠道，且不包含任何形式的人口统计学标记。
We construct a dataset of 1,000 USMLE clinical vignettes, each rewritten into three sociolinguistically distinct personas under an independently audited fact-preservation guarantee, verified by a separate model that never sees the generation prompt. 我们构建了一个包含 1,000 个 USMLE（美国医师执照考试）临床病例的数据集，每个病例都在经过独立审计的事实保留保证下，被重写为三种社会语言学特征鲜明的人设。这些重写内容由一个从未接触过生成提示词的独立模型进行验证。
Across seven language models spanning three architecture families and scales, Narrative Anchoring is statistically significant under direct prompting in every model tested, with a Narrative Anchoring Gap of 0.064 to 0.151. 在跨越三个架构系列和不同规模的七个语言模型中，所有模型在直接提示下均表现出具有统计学意义的“叙事锚定”现象，其叙事锚定差距（Narrative Anchoring Gap）在 0.064 到 0.151 之间。
Chain-of-thought reasoning and explicit debiasing instructions reduce the bias only partially, and their apparent gains are frequently confounded by accuracy collapse. 思维链推理和明确的去偏见指令只能部分缓解这种偏见，且其表面上的提升往往伴随着准确率下降的问题。
We introduce NarrativeShield, a three-agent pipeline that structurally extracts and verifies clinical facts before diagnostic reasoning begins, reducing the Narrative Anchoring Gap to near-zero ($-0.004$ to $0.037$) and achieving the lowest rate of severely unstable decisions (DSS $<$ 0.8) of any method across all models, at a modest and mechanistically expected accuracy cost for most models. 我们引入了 NarrativeShield，这是一个三智能体流水线，它在诊断推理开始前结构化地提取并验证临床事实。该方法将叙事锚定差距降低至接近零（$-0.004$ 至 $0.037$），并在所有模型中实现了最低的严重不稳定决策率（DSS < 0.8），且对于大多数模型而言，其准确率损失处于适中且机制上可预期的范围内。
A stress test using a non-instruction-tuned base model shows that executing a debiasing intervention at all is gated by zero-shot instruction-following ability, not prompt content alone. 一项使用非指令微调基础模型的压力测试表明，能否执行去偏见干预取决于模型的零样本指令遵循能力，而不仅仅取决于提示词内容本身。
We release our dataset, human-validated for fact preservation, as a standalone resource for studying register-based clinical bias. 我们发布了该数据集，并经人工验证确保了事实的准确性，旨在将其作为研究基于语体的临床偏见的独立资源。