This study presents a representation-level account of scoring bias in LLM-as-Judge models, analyzing the judge's hidden states rather than just input-output perturbations. The authors report findings across seven judges, seven bias types, and nine benchmarks.
- Biased inputs are displaced along a low-dimensional, type-specific subspace that sharpens with depth, while baseline inputs occupy a tight activation manifold.
- Steering hidden states along this subspace drives scoring in both directions, whereas random directions produce shifts an order of magnitude smaller.
- A linear projection onto bias-direction features anticipates judge failures on three unseen benchmarks, substantially outperforming text-based alternatives.
Reading bias as activation geometry unifies geometric structure, causal control, and operational prediction within a single framework.