Play Favorites: A Statistical Method to Measure Self-Bias in LLM-as-a-Judge

2025/08/08 by Evangelia Spiliopoulou, Spiliopoulou, Evangelia, Riccardo Fogliato +11 · 1 voice · 7 citations
Computer Science · Social Sciences · #Artificial Intelligence (cs.AI) #Computation and Language (cs.CL) #Computational and Text Analysis Methods #Ethics and Social Impacts of AI #Explainable Artificial Intelligence (XAI) #FOS: Computer and information sciences #cs.AI #cs.CL

paper · pdf · doi:10.48550/arxiv.2508.06709

openalex publication_date 2025/08/08 · arxiv published 2025/08/08 · arxiv updated 2025/08/08 · openalex created_date 2025/10/10 · openalex updated_date 2026/07/28

Abstract

Large language models (LLMs) can serve as judges that offer rapid and reliable assessments of other LLM outputs. However, models may systematically assign overly favorable ratings to their own outputs, a phenomenon known as self-bias, which can distort evaluations of true model performance. Previous studies often conflate genuine differences in model quality with bias or incorrectly assume that evaluations from LLMs and humans follow the same rating distributions. In this work, we present a statistical framework that explicitly formalizes assumptions under which self-bias can be identified and estimated. Our method models the difference in the scoring distribution that LLM-as-a-judge assigns to its own completions compared to other models, while accounting for the underlying quality of the completions provided by an independent, third-party judge (e.g., humans). Our method reliably isolates and quantifies self-bias, even when models vary in ability, ensuring that genuine performance differences are not mistaken for self-bias. We conduct an empirical analysis of self-bias on a large dataset (>5000 prompt-completion pairs) consisting of expert human annotations and judgments from nine different LLM judges. We find that some models, such as GPT-4o and Claude 3.5 Sonnet, systematically assign higher scores to their own outputs. These models also display family-bias; systematically assigning higher ratings to outputs produced by other models of the same family. Our findings highlight potential pitfalls of using LLM judges and offer practical guidance to mitigate biases when interpreting automated evaluations.

Citations

• From Generation to Judgment: Opportunities and Challenges of LLM-as-a-judge
• Self-Preference Bias in LLM-as-a-Judge
• GPT-4o System Card
• Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge
• The Llama 3 Herd of Models
• OffsetBias: Leveraging Debiased Data for Tuning Evaluators
• A Framework for Efficient Model Evaluation through Stratification, Sampling, and Estimation
• Large Language Models are Inconsistent and Biased Evaluators
• Prometheus 2: An Open Source Language Model Specialized in Evaluating Other Language Models
• Replacing Judges with Juries: Evaluating LLM Generations with a Panel of Diverse Models
• LLM Evaluators Recognize and Favor Their Own Generations
• Pride and Prejudice: LLM Amplifies Self-Bias in Self-Refinement
• Humans or LLMs as the Judge? A Study on Judgement Biases
• LLMs as Narcissistic Evaluators: When Ego Inflates Evaluation Scores
• JudgeLM: Fine-tuned Large Language Models are Scalable Judges
• Verbosity Bias in Preference Labeling by Large Language Models
• Generative Judge for Evaluating Alignment
• Benchmarking Cognitive Biases in Large Language Models as Evaluators
• FLASK: Fine-grained Language Model Evaluation based on Alignment Skill Sets
• PRD: Peer Rank and Discussion Improve Large Language Model based Evaluations
• Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
• PandaLM: An Automatic Evaluation Benchmark for LLM Instruction Tuning Optimization
• Large Language Models are not Fair Evaluators
• Can Large Language Models Be an Alternative to Human Evaluations?
• OpenAssistant Conversations -- Democratizing Large Language Model Alignment
• Human-like Summarization Evaluation with ChatGPT
• G-Eval: NLG Evaluation using GPT-4 with Better Human Alignment
• GPT-4 Technical Report
• Large Language Models Are State-of-the-Art Evaluators of Translation Quality
• Self-Instruct: Aligning Language Models with Self-Generated Instructions
• Affective Coherence Monitoring for Transformer-Based Language Models
• On the Limitations of Reference-Free Evaluations of Generated Text
• Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedback
• Red Teaming Language Models with Language Models
• Understanding Dataset Difficulty with V-Usable Information
• maars: Tidy Inference under the 'Models as Approximations' Framework in R
• On Faithfulness and Factuality in Abstractive Summarization
• Model-free Study of Ordinary Least Squares Linear Regression
• Don't Give Me the Details, Just the Summary! Topic-Aware Convolutional\n Neural Networks for Extreme Summarization
• A Practitioner’s Guide to Cluster-Robust Inference
• Reference-Guided Verdict: LLMs-as-Judges in Automatic Evaluation of Free-Form QA
• BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

Cited by

• One Battle After Another: Probing LLMs' Limits on Multi-Turn Instruction Following with a Benchmark Evolving Framework
• Orchestrating Human-AI Teams: The Manager Agent as a Unifying Research Challenge
• Measuring and Mitigating Identity Bias in Multi-Agent Debate via Anonymization
• Towards Verified Code Reasoning by LLMs
• Extreme Self-Preference in Language Models
• RedNote-Vibe: A Dataset for Capturing Temporal Dynamics of AI-Generated Text in Social Media
• Self-Exploring Language Models for Explainable Link Forecasting on Temporal Graphs via Reinforcement Learning

Discussions

• Totally! And they rate outputs from their own model family higher. arxiv.org/abs/2508.06709 At this stage, automated text review should probably average across multiple models. [bsky, 3 points, 0 comments]

Related

• LLM Evaluators Recognize and Favor Their Own Generations
• Beyond the Surface: Measuring Self-Preference in LLM Judgments
• Self-Preference Bias in LLM-as-a-Judge
• Do LLM Evaluators Prefer Themselves for a Reason?
• Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge