Self-Preference Bias in Rubric-Based Evaluation of Large Language Models
Jos\'e Pombal, Ricardo Rei, Andr\'e F. T. Martins
Why It Matters
What makes this one worth your time
Understanding and mitigating self-preference bias is crucial for fair and accurate evaluation of language models, which is essential for their development and deployment.
The study reveals self-preference bias in rubric-based evaluations of language models and explores mitigation strategies.
Summary
The paper investigates self-preference bias (SPB) in rubric-based evaluations of large language models, demonstrating that judges tend to favor outputs from their own model family, even with objective criteria. It shows that ensembling multiple judges can mitigate but not eliminate SPB, and highlights the impact of SPB on subjective rubrics in medical chat benchmarks.
Key contributions
- First study of self-preference bias in rubric-based evaluation of language models.
- Demonstration of self-preference bias in both objective and subjective evaluation settings.
- Analysis of factors contributing to self-preference bias in subjective rubrics.
Notable insights
- Self-preference bias persists even with objective evaluation criteria.
- Ensembling multiple judges reduces but does not fully eliminate self-preference bias.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2604.06996v2 Announce Type: replace-cross Abstract: LLM-as-a-judge has become the de facto approach for evaluating LLM outputs. However, judges are known to exhibit self-preference bias (SPB): they tend to favor outputs produced by themselves or by models from their own family. This skews evaluations and, thus, hinders model development, especially in settings of recursive self-improvement. We present the first study of SPB in rubric-based evaluation, an increasingly popular benchmarking paradigm where judges issue binary verdicts on individual evaluation criteria, instead of assigning holistic scores or rankings. Using IFEval and LiveCodeBench, benchmarks with programmatically verifiable rubrics, we show that SPB persists even when evaluation criteria are entirely objective: among rubrics where generators fail, judges can be more than 50% more likely to incorrectly mark them as satisfied when the output is their own. We also find that, similarly to other evaluation paradigms, ensembling multiple judges helps mitigate SPB, but without fully eliminating it. On HealthBench, a medical chat benchmark with subjective rubrics, we observe that SPB skews model scores by up to 10 points, a potentially decisive margin when ranking frontier models. We analyze the factors that drive SPB in this setting, finding that negative rubrics and subjective topics like communication and emergency referrals are particularly susceptible.