Research map / Large language models / Language, safety and interpretability
How Much Do Large Language Model Cheat on Evaluation? Benchmarking Overestimation Under the One-Time-Pad-Based Framework
Zi Liang, Liantong Yu, Zhang Shiyu, Qingqing Ye, Haibo Hu
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| regenerate SCP test cases with 32 different random seeds from the same articles (ablation) | ArxivRollBench-2024b-CS · ArxivRollBench-2024b-CS, Llama3-8B | accuracy std: std below 1 point (0.0074 to 0.0095) | No clear effect |
| use Sequencing vs Prediction vs Cloze generation strategy (ablation) | ArxivRollBench · Correlation with ChatbotArena (Spearman) | Spearman correlation: S 0.76, P 0.73, C 0.61 | Mixed |
| newer model versions within a series (e.g., Phi series) | LLM series · ArxivRollBench vs public benchmarks | accuracy and RSI | Mixed |
| measure public vs private benchmark gap | Qwen and Phi series · public vs ArxivRollBench | Absolute RSI: Qwen2.5-72B 1.41; Phi-3-mini 1.27 | Hurt |
| compare public vs private benchmark performance by domain | GPT-J-6B, Phi2, Llama3-8B, Yi1.5-34B, Llama3.1-Nemotron-70B · Econ, Q-Fin, Bio, Phy domains | Absolute RSI | Hurt |
| scale up to Llama3.1-70B or Nemotron-70B | Llama family · Private benchmarks | RS_II vs normalized RS_II: Nemotron-70B RS_II 1.30%, normalized 3.88% | Mixed |
Closest papers by meaning
- Training on the Benchmark Is Not All You Need (AAAI 2025 · similarity 0.90)
- Cheating Automatic LLM Benchmarks: Null Models Achieve High Win Rates (ICLR 2025 · similarity 0.89)
- OR-Bench: An Over-Refusal Benchmark for Large Language Models (ICML 2025 · similarity 0.89)
- DOMAINEVAL: An Auto-Constructed Benchmark for Multi-Domain Code Generation (AAAI 2025 · similarity 0.89)
- How Much Can We Forget about Data Contamination? (ICML 2025 · similarity 0.89)
- Overestimation in LLM Evaluation: A Controlled Large-Scale Study on Data Contamination’s Impact on Machine Translation (ICML 2025 · similarity 0.88)
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.
