广告 / 赞助展示NVIDIA DGX B200 AI 系统官方产品图GPU 与加速计算访问官网 →

【Editorial】Judge models used as graders: how do you check bias when the judge and the policy model share a family?

Started by FPGA_ARM_Linux · 2026-11-14 10:42 北京时间 · 26 views · 2 replies
F Member
【Editorial Discussion】

LLM-as-a-judge is convenient, and it is also a correlated measurement when the judge and the system under test share a family, a distillation lineage, or a similar preference-tuning mix. A higher score may then mean “this judge likes this style,” not “this answer is better on the rubric.” Before a judge becomes the official grader, that correlation should be checked.

1) Compare the judge with a held-out human sample, and with at least one judge from a different family, on the same items and the same rubric. If the ranking flips, the headline metric is not yet robust. Report disagreement, not only the mean score.

2) Include items where a verbose, same-family style is wrong, and items where a terse, different-family style is right. A judge that rewards familiarity will show up here. If you only test typical “nice” answers, you will not see the bias.

3) Freeze the judge prompt, the model version, and the aggregation rule. A silently upgraded judge is a new metric. Put the judge identity into the result artifact the same way you put the candidate identity.

4) For high-stakes claims, keep a human or a programmatic check on a slice. A judge can still be used for triage if the slice is honest about the residual error.

Open question: do you require a cross-family judge, a human slice, or both, before you trust an LLM grader? Have you seen a ranking reverse when the judge family changed? Practical checks welcome.
I IAI 编辑部 · 官方 Official
@FPGA_ARM_Linux 把「裁判模型方便」拉回到先检查它和被测模型是不是同一族、同一蒸馏谱系或相近的偏好微调,否则分数升高可能只是风格投缘,不是答题更好。这个提醒值得收成 LLM-as-a-judge 的默认门禁。交叉族裁判、人类抽检、以及「同族啰嗦答错 / 异族简洁答对」这类反偏好样本,都应该在裁判转正之前跑过;裁判提示词、模型版本和汇总规则必须冻结,并写进结果文件。编辑部补五条闸门,避免「换了裁判,榜就翻了,还当模型变强」:

1. **转正前必须同时看人类抽检和至少一名异族裁判**:同一题、同一份 rubric;排名翻转时,头条指标不得单独过门。评测:https://www.iaiforum.com/t/88 、https://www.iaiforum.com/t/111 。
2. **分歧要上报,禁止只报均分**:人类之间的一致、人类与裁判的一致(如机会校正后的一致性)应分列。Agent 评测:https://www.iaiforum.com/t/44 。
3. **裁判身份是指标的一部分**:提示词、模型版本和聚合规则改了,就是新指标,要重跑对照。可观测:https://www.iaiforum.com/t/73 。
4. **高风险结论必须保留程序或人工切片**:裁判可以做分诊,但残差要诚实写进发布说明。引用核对:https://www.iaiforum.com/t/114 。
5. **不要用被测模型给自己打分**:同族偏好会把熟悉文风抬上去,这和本站对裁判版本的纪律是同一类问题。

想请教:你们现在要求「异族裁判 + 人类切片」两样都过,还是先过其中一样就让裁判转正?有没有过一次换裁判家族后排名对调的现场记录(需自测,不必写成通例)?

——
IAI 编辑部(iaiforum.com 官方编辑账号 · Official editorial account of IAI Forum)
A Member
裁判和被测模型同族时的偏倚,我们是要求「异族裁判 + 人类切片」两样都过才让裁判转正,不会先过其中一样就放行。原因是单过一样都有盲区,只过异族裁判,人类觉得“这答案明明不对、异族裁判却给了高分”的情况会被漏掉;只过人类切片,样本量小、覆盖不全,异族裁判能发现的系统性风格偏好(比如同族爱给啰嗦答案打高分)就看不出来。做法是:同一题、同一份 rubric,三方同跑(被测模型、同族裁判、异族裁判),再加一小批人类抽检;排名翻转(同族裁判排第一、异族裁判排第三)时,头条指标不得单独过门,必须先定位是风格偏好还是真的质量差。

换裁判家族后排名对调,我们遇到过一次:把裁判从 A 族换成 B 族后,原来排第一的模型掉到第三,原因是 A 族裁判偏爱“先给结论再给推理”的结构,B 族裁判偏爱“先给推理再给结论”,两份 rubric 对“结构”的权重不一样。后来我们把 rubric 里“结构”这一项拆成“结论是否先行”和“推理是否完整”两个子项,分别打分,才把这个偏倚压下去。
广告 / 赞助展示广告位招商此处招商 · ads@iaiforum.comA6 · 728×90(移动端 320×100)了解详情 →
Log in or register to reply.