AI ENGVisual Encyclopedia

MODULE 7: EVALUATION & BENCHMARKING · SCENE 20

LLM-as-a-Judge: Bias Audit

Positional flips, length bias, and judge Elo deltas — measuring and mitigating the referee's own errors.

POSITIONAL BIASBIAS = 0.20 · SWAP-AUGMENT!WIN 64% → ΔELO 100
position0.20length0.12self-pref0.22MITIGATION: SWAP AND RE-JUDGE · LENGTH-CONTROLLED WIN RATES · CROSS-FAMILY JUDGES

This judge prefers the first answer it reads — close comparisons flip with answer order. Run every pair twice, swapped.
GO DEEPER: THE JUDGE BIAS AUDIT IN THE EVALUATION WORLD →

TECHNICAL BREAKDOWNModule 7: Evaluation & Benchmarking

LLM-as-a-judge: auditing the referee

LLM judges scale preference evaluation to thousands of comparisons, but they carry measurable biases: position (preferring the first answer), length (preferring the longer), and self-preference. An unaudited judge corrupts every win-rate downstream.

Positional Bias

Run every comparison twice with answers swapped. Win-rate flips reveal order dependence; report the bias-corrected average.

Length Bias

Verbose answers win more than their quality warrants. Length-controlled win rates (AlpacaEval-LC) strip the confound.

Judge Elo

Pairwise win rates convert to Elo via a Bradley-Terry fit — the same model as Module 4, now scoring models instead of responses.

MATHEMATICAL FORMULATION · POSITIONAL BIAS & ELO
bias = |p_first − 0.5|; ΔElo = 400·log₁₀( wr / (1 − wr) )

A judge that prefers the first answer 70% of the time carries bias 0.2 — enough to flip close comparisons. A 64% win rate converts to +128 Elo; a 55% rate, only +40.

REAL-WORLD PRODUCTION ENGINEERING
  • MT-Bench and Chatbot Arena popularized the pattern: judge rubric + swap-augmented pairs + length control.
  • Self-preference is real: a judge from the same model family rates its own outputs higher. Use a judge from a different family, or an ensemble.