Inside the Unfair Judge: Mechanistic Interpretability of LLM-as-Judge Bias
arXiv paper · source date 2026-07-13 · added 2026-08-07 15:52:43 · updated 2026-08-07 15:52:43 · Open original blog
1
Problems / challenges / motivations
- LLM-judge bias is well documented at the input-output level (position bias, verbosity bias, self-preference) but not explained mechanistically.
- Debiasing is therefore trial-and-error prompt engineering.
- There is no way to predict in advance whether a judge will fail on a new benchmark.
2
Key ideas
- Judge biases live on a low-dimensional, type-specific activation subspace that sharpens with model depth.
- Steering along those directions causally induces or removes biased scoring — evidence the subspace is the mechanism, not a correlate.
- A linear projection onto the bias directions predicts judge failures on unseen benchmarks.
- Evaluated across 7 judges, 7 bias types, and 9 benchmarks, with a public project page.
3
Why it matters for evals
- It turns "validate your judge" from advice into a diagnostic you can run before deploying the judge on new data.
- It exemplifies the month's shift toward treating the judge as an instrument with characterizable failure modes rather than an oracle.
- Caveat: a preprint. The linear-subspace account and the steering results need independent replication, and predictive power on unseen benchmarks is demonstrated only on the studied judge and bias set.
Comments
No comments yet.