Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

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

Peiwen Yuan, Yiwei Li, Shaoxiong Feng, Xinglin Wang, Yueqi Zhang, Jiayi Shi, Chuyi Tan, Boyuan Pan, Yao Hu, Kan Li

TL;DR

This work formalizes self-bias in LLM-generated benchmarks and decomposes it into three sub-biases. It introduces Silencer, a framework that combines sample-level mitigations—Attribute Integration, Cross Paraphrase, and Label Calibration—with a bias-neutralizing ensemble at the benchmark level to suppress self-bias. Empirical results across multiple tasks and models show substantial reductions in self-bias and substantial gains in alignment with human-annotated benchmarks, with average Pearson correlation improving from 0.655 to 0.833. The approach demonstrates good generalizability and provides practical guidance on leveraging multiple generators to produce more reliable benchmarks for evaluating LLMs.

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.

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

TL;DR

This work formalizes self-bias in LLM-generated benchmarks and decomposes it into three sub-biases. It introduces Silencer, a framework that combines sample-level mitigations—Attribute Integration, Cross Paraphrase, and Label Calibration—with a bias-neutralizing ensemble at the benchmark level to suppress self-bias. Empirical results across multiple tasks and models show substantial reductions in self-bias and substantial gains in alignment with human-annotated benchmarks, with average Pearson correlation improving from 0.655 to 0.833. The approach demonstrates good generalizability and provides practical guidance on leveraging multiple generators to produce more reliable benchmarks for evaluating LLMs.

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.

Paper Structure

This paper contains 31 sections, 3 theorems, 25 equations, 3 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related Works
  3. LLM-as-Benchmark-Generator
  4. Self-Bias of LLMs
  5. Preliminary
  6. Formal Definition and Verification of Self-Bias
  7. Disentangling of Self-Bias
  8. Silencer Framework
  9. Motivations
  10. Sample-wise Sub-biases Mitigation
  11. Attribute Integration.
  12. Cross Paraphrase.
  13. Label Calibration.
  14. Benchmark-wise Self-bias Suppression
  15. Experiments
  16. ...and 16 more sections

Key Result

Lemma 1

There exists at least one $\alpha^\ast\in\Delta^{N-1}$ with $F^{\!*}(\alpha^\ast)=\alpha^\ast$.

Figures (3)

  • Figure 1: Evaluation bias of LLMs across tasks. The value at position $(\mathcal{M}_i,\mathcal{M}_j)$ represents $\mathcal{B}(\mathcal{M}_j|\mathcal{M}_i)$ (Eq. \ref{['equation: B']}), the evaluation bias of $\mathcal{M}_j$ on the benchmark generated by $\mathcal{M}_i$.
  • Figure 2: Overall illustration of our Silencer framework. Each generator LLM is represented by a unique color. Colored arrows indicate the process by which a certain LLM takes the content on the left as input and produces the content on the right as output (excluding the gray arrows).
  • Figure 3: T-SNE visualization results of the raw generated benchmarks (a), those with mitigated language style bias (b), and those with both language style and question domain biases suppressed (c) on the Language Understanding task. The presence of 13 clusters is probably due to the task comprising 13 sub-tasks. We use text-embedding-ada-002 text-embedding-ada-002 as the embedding model.

Theorems & Definitions (6)

  • Lemma 1: Existence
  • proof
  • Theorem 1: Uniqueness & Contraction
  • proof : Sketch
  • Theorem 2: Global convergence
  • proof