Quantifying and Mitigating Self-Preference Bias of LLM Judges

2026-04-24

AI Summary

This study addresses the self-preference bias (SPB) inherent in large language models (LLMs) when used as evaluators—a systematic tendency to favor their own generated responses, which undermines the reliability of automated evaluation. The work presents the first fully automated framework to quantify SPB by constructing equally high-quality response pairs without human annotation. Leveraging statistical methods, it disentangles an LLM’s discriminative capability from its biased preference, and introduces a multidimensional, structured evaluation paradigm grounded in cognitive load decomposition to mitigate SPB. Experiments across 20 mainstream LLMs demonstrate an average SPB reduction of 31.5%, while revealing that model capability exhibits no positive correlation—and may even be negatively correlated—with SPB severity, thereby offering a novel pathway toward trustworthy automatic evaluation.

Technology Category

• Natural Language Processing: Interpretability, Analysis, and Evaluation of NLP Models
• Machine Learning: Evaluation and Analysis
• Knowledge Representation and Reasoning: Preferences

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Large Language Models (LLM) for user modeling and recommendation
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems

Abstract

LLM-as-a-Judge has become a dominant approach in automated evaluation systems, playing critical roles in model alignment, leaderboard construction, quality control, and so on. However, the scalability and trustworthiness of this approach can be substantially distorted by Self-Preference Bias (SPB), which is a directional evaluative deviation in which LLMs systematically favor or disfavor their own generated outputs during evaluation. Existing measurements rely on costly human annotations and conflate generative capability with evaluative stance, and thus are impractical for large-scale deployment in real-world systems. To address this issue, we introduce a fully automated framework to quantifying and mitigating SPB, which constructs equal-quality pairs of responses with negligible quality differences, enabling statistical disentanglement of discriminability from bias propensity without human gold standards. Empirical analysis across 20 mainstream LLMs reveals that advanced capabilities are often uncorrelated, or even negatively correlated, with low SPB. To mitigate this bias, we propose a structured multi-dimensional evaluation strategy grounded in cognitive load decomposition, which reduces SPB by 31.5\% on average.

Problem

• Self-Preference Bias
• LLM-as-a-Judge
• automated evaluation
• evaluation bias
• model alignment

Innovation

• Self-Preference Bias
• LLM-as-a-Judge
• bias mitigation
• automated evaluation
• cognitive load decomposition