Research map / Large language models / Language, safety and interpretability
Silencer: From Discovery to Mitigation of Self-Bias in LLM-as-Benchmark-Generator
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Identifies self-bias in LLM-as-Benchmark-Generator (generators score higher on their own benchmarks), decomposes it into domain, style and wrong-label sub-biases, and proposes Silencer, which mitigates each at sample level and reweights benchmark ensembles to suppress self-bias.
Problem
LLMs evaluated on benchmarks they generate themselves achieve inflated performance (self-bias), which undermines the reliability of LLM-as-Benchmark-Generator evaluation.
Approach
Decompose self-bias into question domain, language style, and wrong label sub-biases; mitigate them per sample (attribute integration, cross paraphrase, label calibration) using heterogeneous generators, then ensemble benchmarks from multiple generators with weights set by evaluation consistency with the ensembled benchmark, iterated to convergence.
Key results
- Silencer reduces average self-bias from 0.113 to 0.009 across settings.
- Average Pearson correlation with human-annotated benchmarks improves from 0.655 to 0.833.
- On MATH math reasoning with BenchMaker, self-bias drops from 0.1205 to 0.0140 and r_p rises from 0.7863 to 0.9386.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Apply Silencer to BenchMaker benchmarks | BenchMaker · Math Reasoning, Language Understanding, and Commonsense Reasoning | evaluation effectiveness (Pearson correlation): 0.655 to 0.833 on average | Helped |
| Apply Silencer to BenchMaker benchmarks | BenchMaker · Math Reasoning, Language Understanding, and Commonsense Reasoning | self-bias: from 0.113 to 0.009 on average | Helped |
| Add Attribute Integration (ablation) | BenchMaker · Math Reasoning | evaluation effectiveness (Pearson correlation): 0.7863 -> 0.7892 | Helped |
| Add Cross Paraphrase (ablation) | BenchMaker · Language Understanding | evaluation effectiveness (Pearson correlation): 0.6816 -> 0.6840 | No clear effect |
| Add Label Calibration (ablation) | BenchMaker · Math Reasoning | evaluation effectiveness (Pearson correlation): 0.7863 -> 0.8302 | Helped |
| Add Attribute Integration, Cross Paraphrase, and Label Calibration (ablation) | BenchMaker · Math Reasoning | evaluation effectiveness (Pearson correlation): 0.7863 -> 0.8371 | Helped |
| Add naive ensemble to sample-level modules (ablation) | BenchMaker · Math Reasoning | evaluation effectiveness (Pearson correlation): 0.8371 -> 0.9216 | Helped |
| Replace naive ensemble with Bias-Neutralizing Ensemble (ablation) | BenchMaker with sample-level modules · Math Reasoning | evaluation effectiveness (Pearson correlation): 0.9216 -> 0.9386 | Helped |
| Apply Silencer | AttrPrompt · Math Reasoning, Language Understanding, and Commonsense Reasoning | evaluation effectiveness (Pearson correlation): 0.6179 -> 0.8641; 0.5843 -> 0.7725; 0.6106 -> 0.7762 | Helped |
| Increase generator count from T=3 to T=7 (ablation) | Naive Ensemble · Across settings | self-bias: 0.031 -> 0.008 | Helped |
| Increase generator count from T=3 to T=7 (ablation) | Reweighted Ensemble · Across settings | evaluation effectiveness (Pearson correlation): 0.912 -> 0.955 | Helped |
| Increase benchmark size from 50 to 200 (ablation) | Reweighted Ensemble · Across settings | self-bias: 0.014 -> 0.011 | Helped |
Surprising findings
- Some models show negative evaluation bias on model-generated benchmarks, attributed to data contamination on the human-annotated benchmark (verified with ConStat).
- Wrong label bias is the largest contributor to self-bias (about 67-80% relative contribution).
Limitations
- Incurs about 30% additional token cost.
- Relies on o3-mini as proxy for human annotators in disentangling sub-biases.
- Benchmark sizes evaluated are small (50 to 200 per sub-task).
Experimental setup
Datasets: MATH MMLU-Pro HellaSwag
Benchmarks: MATH MMLU-Pro HellaSwag
Metrics: self-bias Pearson correlation with human benchmark MMD Wasserstein distance word-frequency entropy 2-gram count label accuracy
Baselines: BenchMaker AttrPrompt Naive ensemble
Builds on: BenchMaker AttrPrompt ConStat
Closest papers by meaning
- Any Large Language Model Can Be a Reliable Judge: Debiasing with a Reasoning-based Bias Detector (NeurIPS 2025 · similarity 0.88)
- Benchmarking LLMs' Judgments with No Gold Standard (ICLR 2025 · similarity 0.87)
- From Crowdsourced Data to High-quality Benchmarks: Arena-Hard and Benchbuilder Pipeline (ICML 2025 · similarity 0.87)
- Self-Play Fine-Tuning Converts Weak Language Models to Strong Language Models (ICML 2024 · similarity 0.87)
- Self-contradictory Hallucinations of Large Language Models: Evaluation, Detection and Mitigation (ICLR 2024 · similarity 0.87)
- LLM Evaluators Recognize and Favor Their Own Generations (NeurIPS 2024 · similarity 0.87)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.