---
title: "《不公判官内部：LLM作为评判者偏见的机制可解释性分析》"
source: "arXiv AI/CL/LG（论文）"
source_tier: "t2"
url: "https://arxiv.org/abs/2607.11871v1"
permalink: "https://feed.astralor.com/items/itm_1bc5cce454463fdaa0f42583"
published: "2026-07-13T17:55:19.000Z"
score: 79
channel: "tech"
tags: ["LLM-as-Judge","偏见","因果干预","机制可解释性","激活几何"]
---

# 《不公判官内部：LLM作为评判者偏见的机制可解释性分析》

## AI 点评

> 这篇论文真正让我眼前一亮的是它将偏见从‘噪声’重新定义为‘几何结构’，从而同时获得理解和控制的能力。对想提升LLM评估可靠性的工程师来说，其中的线性投影预测方法几乎可以直接拿来用；对机制可解释性研究者，它展示了激活子空间如何兼具描述性、因果性与预测性。值得一读。

## 速览

本文提出LLM-as-judge的评分偏见不仅存在于输入输出层面，还可以在隐藏状态的表征层面进行解释。通过跨7个法官模型、7种偏见类型和9个基准的实验，发现偏见输入在激活空间中沿低维子空间偏移，沿该子空间进行因果控制可双向调节评分，且简单线性投影能预测未知基准上的失败，性能远超基于文本的方法。

## 关键点

- 偏见输入在激活空间中沿类型特定的低维子空间偏移，该子空间随深度加深而锐化。
- 沿偏见子空间进行因果控制（前向/逆向）可双向调节评分，且效应远超随机方向。
- 通过线性投影到偏见方向特征，可预测三个完全未见基准上的评判失败，优于基于文本的基线。
- 将偏见视为激活几何结构而非输入输出噪声，统一了结构、因果控制与操作预测。

## 全文

Existing studies of LLM-as-judge scoring bias work predominantly at the input-output level: they perturb inputs, measure score deltas, and propose prompt-level mitigations. We argue that the same biases admit a representation-level account in the judge's hidden state, complementary to the input-output view and operationally useful in ways it does not afford. We report three findings, across seven judges, seven bias types, and nine benchmarks. Geometry: baseline judging inputs occupy a tight activation manifold while biased inputs are displaced along a low-dimensional, type-specific subspace that sharpens with depth and is recovered consistently by three families of estimators. Causal control: steering hidden states along this subspace drives scoring in both directions, forward shifts reproducing biased scoring on clean inputs and reverse shifts restoring baseline scoring on biased ones, while matched-norm random directions produce shifts an order of magnitude smaller. Operational: a simple linear projection onto the same bias-direction features anticipates judge failures on three entirely unseen benchmarks, substantially outperforming text-based alternatives. Reading bias as activation geometry, rather than as input-output noise, unifies geometric structure, causal control, and operational prediction within a single framework. The project page is available at https://xzx34.github.io/unfair-judge/

---

引用前请核对原文;本文由星际流动整理:https://feed.astralor.com/items/itm_1bc5cce454463fdaa0f42583