Papers

Stanford 论文:8 个 LLM 评审都偏袒自己的医疗答案

Why we picked it

做模型评测的朋友该看看这篇。Stanford 用 620 个真实临床问题实测了 8 个模型的 LLM-as-judge 偏差,连匿名和加 rubric 都压不住,GPT-5.6 Sol 偏得最狠。文章结尾给了两个能直接落地的评审方案。

Stanford 一项新研究发现,来自 OpenAI、Anthropic、Google 和 Qwen 的 8 个 LLM 在评审医疗问答时,都把自己生成的答案排得更高,即使答案已匿名处理。实验让这 8 个模型回答 620 个真实临床医生提问,再互相给全部匿名答案排序,对比每个评审如何排列自己的答案与其他 7 个评审对同一答案的排名。结果显示,每个评审平均把自己的答案提前 1.22 个名次,GPT-5.6 Sol 偏差接近 3 个名次。

Body

技术上,团队验证了两种常见去偏手段的效果:去掉评分 rubric 后自偏依然存在;而向评审公开答案对应的模型名称,偏差仅减少 0.094 个名次,说明匿名和命名几乎不影响结果。这意味着「LLM 当裁判」的做法在医疗基准测试中存在系统性倾斜——若基准只用单一模型或与被测模型同家族的模型来评分,排名会向该模型偏移。

作者给出的实操建议是:使用 LLM 评审时组建跨多个模型家族的评审团,并按每个评审分别报告结果,而不是合并成单一汇总分数。这对依赖 LLM-as-judge 做医疗模型评测的开发者和研究团队有直接参考价值。

Original · rohanpaul_ai

New Stanford paper finds that all 8 LLMs tested rank their own medical answers higher than other models rank them, even when answers are anonymous.

Medical AI benchmarks often use LLMs to grade other LLMs' answers. If a judge favors its own outputs, the benchmark tilts toward that model.

They had 8 models from OpenAI, Anthropic, Google and Qwen answer 620 real clinician questions, then rank all 8 anonymous answers. They compared where each judge placed its own answer with where the other 7 judges placed that same answer.

Every judge favored its own answer, by 1.22 rank positions on average and nearly 3 for GPT-5.6 Sol. The bias held without a rubric, and revealing model names cut it by only 0.094 positions.

If you use LLM judges, use a panel from several model families and report results per judge instead of 1 pooled score.

Industry Coordinates & Source Verification

This report is categorized under AITOP's "Papers" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Stanford, GPT-5.6 Sol, LLM-as-judge, AI safety.

Source attribution: curated from rohanpaul_ai (@rohanpaul_ai).