Mechanistic View Reveals How Bias Lives Inside LLM Judges
Study shows LLM-as-judge bias is encoded in activation geometry, enabling causal steering and better failure prediction than text-based methods.
A new study reframes LLM-as-judge scoring bias as a phenomenon rooted in the model's hidden activations rather than just input-output behavior. Across seven judge models, seven bias types, and nine benchmarks, the authors show that biased inputs shift along a low-dimensional, bias-type-specific subspace in activation space, with the effect sharpening at deeper layers.
Critically, the team demonstrates causal control: steering hidden states along this subspace can induce biased scoring on clean inputs or restore unbiased scoring on biased ones, while random directions of matched magnitude produce far weaker effects. This confirms the identified subspace genuinely encodes the bias mechanism, not just a correlation.
The findings also have practical value—a simple linear projection onto these bias-direction features predicts judge failures on three entirely unseen benchmarks, outperforming text-based detection methods. For engineers building LLM evaluation pipelines, this offers a mechanistic, activation-based tool for anticipating and mitigating judge bias beyond prompt-level fixes.