Inside the Unfair Judge: A Mechanistic Interpretability Account of LLM-as-Judge Bias

2026-07-13

AI Summary

This work addresses the systematic biases exhibited by large language models (LLMs) when acting as evaluators, a phenomenon inadequately understood at the level of internal representations. Leveraging mechanistic interpretability, the study reveals for the first time that LLM evaluator bias manifests as low-dimensional geometric structures within hidden activation spaces, enabling the identification of bias subspaces. Through activation geometry analysis, subspace estimation, and hidden state interventions, the proposed approach is validated across seven evaluator models, seven bias types, and nine evaluation benchmarks. The introduced linear projection method not only causally intervenes to reverse biased scoring behaviors but also significantly outperforms text-based baselines on three unseen benchmarks, accurately predicting evaluator failure. This framework unifies the representation, control, and cross-benchmark prediction of evaluator bias within a single geometric paradigm.

Abstract

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/

Problem

• LLM-as-judge
• bias
• representation
• activation geometry
• interpretability

Innovation

• mechanistic interpretability
• representation geometry
• causal steering
• bias subspace
• LLM-as-judge