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

Research Question and Approach

This paper investigates the internal mechanics of LLM-as-judge bias. While existing research treats judges as black boxes—measuring how input perturbations change output scores—the authors argue that these biases are represented as specific geometric structures within the model's hidden states. By constructing a dataset of seven bias types (e.g., verbosity, prestige, bandwagon) across nine benchmarks, the researchers map how these biases manifest in the activation space of seven different LLM judges.

Geometric and Causal Findings

The researchers discovered that baseline inputs occupy a tight activation manifold, while biased inputs are displaced along a low-dimensional, type-specific subspace that becomes more pronounced at deeper layers. This geometry is not just a passive correlate; it acts as a causal handle. By steering hidden states along this identified subspace, the authors could reliably manipulate scores: forward shifts induced bias in clean inputs, while reverse shifts restored fair scoring in biased inputs. These interventions were significantly more effective than random directional shifts of the same magnitude.

Operational Utility

Beyond theoretical understanding, the authors demonstrate that this geometric account is operationally useful. They developed a linear projection method that uses these bias-direction features to predict when a judge will produce a degraded score. This approach significantly outperformed text-based baselines when tested on entirely unseen benchmarks, suggesting that the internal representation of bias is a robust, transferable feature of the model's decision-making process.

By shifting the focus from input-output noise to internal activation geometry, this work provides a unified framework for understanding, controlling, and predicting judge bias. This mechanistic approach offers a more precise way to audit and debias LLMs used in sensitive evaluation pipelines, moving beyond simple prompt engineering toward direct, representation-level interventions.