feed7.dev
Sign InStart Agent Brain
arXivPaperNeeds Review

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

LLM judge bias appears as steerable hidden-state directions that predict failures on unseen benchmarks, so eval pipelines may need representation-level checks beyond prompt fixes.

arXiv
Open Source Open MarkdownOpen JSON
Source Summary

Across **7 judges, 7 bias types, and 9 benchmarks**, biased inputs displaced hidden states along low-dimensional, type-specific directions that became clearer with depth. Three estimator families recovered the structure consistently.

Practical Implication

If model-based grading drives agent evals, test the judge itself under known bias perturbations. For models whose activations are accessible, projection onto learned bias directions could supplement score-delta tests and prompt-level mitigations.

Agent-Ready Context
Across **7 judges, 7 bias types, and 9 benchmarks**, biased inputs displaced hidden states along low-dimensional, type-specific directions that became clearer with depth. Three estimator families recovered the structure consistently.

If model-based grading drives agent evals, test the judge itself under known bias perturbations. For models whose activations are accessible, projection onto learned bias directions could supplement score-delta tests and prompt-level mitigations.

The evidence comes from the studied judges and bias types, and activation access is unavailable for many hosted models. Prediction was tested on **3 unseen benchmarks**; broader transfer and practical intervention costs remain open.
Context Map
benchmarkresearch#benchmark-integrity#agent-evals#agent-reliability
Uncertainty
The evidence comes from the studied judges and bias types, and activation access is unavailable for many hosted models. Prediction was tested on **3 unseen benchmarks**; broader transfer and practical intervention costs remain open.