Silencer: From Discovery to Mitigation of Self-Bias in LLM-as-Benchmark-Generator

2025-05-27

AI Summary

This paper identifies and systematically validates “self-bias” in large language models (LLMs) acting as benchmark generators—i.e., inflated performance estimates on self-generated evaluation sets. This bias stems from three interrelated sub-biases: domain mismatch, stylistic inconsistency, and erroneous labeling. To address this, we formally define self-bias for the first time and propose Silencer, a general-purpose debiasing framework. Silencer employs multi-LLM collaborative generation, sample-level and benchmark-level heterogeneity modeling, bias溯源 (origin tracing), and adaptive filtering—requiring neither human annotation nor model fine-tuning. Experiments demonstrate near-zero residual self-bias, with evaluation validity—measured by Pearson correlation against human-constructed benchmarks—improving from 0.655 to 0.833. Moreover, Silencer exhibits strong cross-model and cross-task generalization.

Technology Category

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

Application Category

• Search and Retrieval-Augmented AI: Web evaluation methodologies and metrics
• User Modeling, Personalization and Recommendation: Fairness-aware retrieval and ranking
• Economics, Online Markets and Human Computation: Humans versus LLMs for data annotation and labeling

Abstract

LLM-as-Benchmark-Generator methods have been widely studied as a supplement to human annotators for scalable evaluation, while the potential biases within this paradigm remain underexplored. In this work, we systematically define and validate the phenomenon of inflated performance in models evaluated on their self-generated benchmarks, referred to as self-bias, and attribute it to sub-biases arising from question domain, language style, and wrong labels. On this basis, we propose Silencer, a general framework that leverages the heterogeneity between multiple generators at both the sample and benchmark levels to neutralize bias and generate high-quality, self-bias-silenced benchmark. Experimental results across various settings demonstrate that Silencer can suppress self-bias to near zero, significantly improve evaluation effectiveness of the generated benchmark (with an average improvement from 0.655 to 0.833 in Pearson correlation with high-quality human-annotated benchmark), while also exhibiting strong generalizability.

Problem

• Identifying self-bias in LLM-generated benchmarks
• Analyzing sub-biases from domain, style, and labels
• Mitigating bias via multi-generator heterogeneity framework

Innovation

• Leverages multiple generators' heterogeneity
• Neutralizes bias at sample and benchmark levels
• Suppresses self-bias to near zero