Quantifying Label-Induced Bias in Large Language Model Self- and Cross-Evaluations

2025-08-28

AI Summary

Large language models (LLMs) exhibit significant label-induced bias in both self-assessment and cross-model evaluation: human and model-based quality judgments are driven more by identity labels (e.g., “Claude”, “Gemini”) attached to outputs than by actual content. Method: Through controlled experiments, we systematically compare preference voting and multidimensional quality ratings (coherence, informativeness, conciseness) across four conditions—no label, ground-truth label, and two types of fabricated labels—using outputs from multiple LLMs. Contribution/Results: We provide the first quantitative characterization of this bias: spurious labels reverse model rankings, shift preference votes by up to 50 percentage points, and distort dimensional scores by up to 12 points; self-evaluation is similarly impaired. Label effects are asymmetric—e.g., “Claude” consistently inflates scores, while “Gemini” systematically depresses them. Based on these findings, we propose practical fairness-enhancing interventions, including blind evaluation and multi-model collaborative assessment.

Technology Category

• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Ethics — Bias, Fairness, Transparency & Privacy
• Computer Vision: Bias, Fairness & Privacy

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics

Abstract

Large language models (LLMs) are increasingly used to evaluate outputs, yet their judgments may be influenced. This study examines bias in self- and cross-model evaluations by ChatGPT, Gemini, and Claude under four conditions: no labels, true labels, and two false-label scenarios. Blog posts authored by each model were evaluated by all three using both overall preference voting and quality ratings for Coherence, Informativeness, and Conciseness, with all scores expressed as percentages for direct comparison. Results reveal striking asymmetries: the "Claude" label consistently boosts scores, while the "Gemini" label consistently depresses them, regardless of actual content. False labels frequently reversed rankings, producing shifts of up to 50 percentage points in preference votes and up to 12 percentage points in converted quality ratings. Gemini's self-scores collapsed under true labels, while Claude's self-preference intensified. These findings show that perceived model identity can heavily distort high-level judgments and subtly influence detailed quality ratings, underscoring the need for blind or multimodel evaluation protocols to ensure fairness in LLM benchmarking.

Problem

• Examining bias in LLM self- and cross-model evaluations
• Quantifying how model labels influence evaluation scores
• Assessing label-induced ranking reversals in LLM benchmarking

Innovation

• Label-induced bias quantification in LLM evaluations
• Comparative analysis under true and false label conditions
• Blind and multimodel protocols for fair benchmarking