Research map / Large language models / Language, safety and interpretability
How to Correctly Report LLM-as-a-Judge Evaluations
Chungpa Lee, Thomas Zeng, Jongwon Jeong, Jy-yong Sohn, Kangwook Lee
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 |
|---|---|---|---|
| apply Rogan-Gladen bias correction to naive judge score | LLM-as-a-judge accuracy · Chatbot Arena win rate (GPT-4.1-mini judge) | bias | Improved |
| use bias-corrected CI instead of naive p-hat CI | LLM-as-a-judge accuracy · Monte Carlo (q0,q1)=(0.7,0.9) | coverage probability: near-zero coverage -> close to nominal 95% | Improved |
| adaptive calibration allocation (m0 proportional to (1/p-1)sqrt(kappa) m1) vs equal split (ablation) | bias-corrected CI · Monte Carlo, m=200 | CI length | Improved |
| bias-correct the estimate before ranking | Gemini-3-Flash judge · Chatbot Arena ranking | exact ranking recovery: 0% -> 37% | Improved |
| bias-correct the estimate before ranking | Claude-Haiku-4.5 judge · Chatbot Arena ranking | exact ranking recovery: 7% -> 40% | Improved |
| bias-correct the estimate before ranking (10% calibration) | GPT-4.1 judge · Chatbot Arena ranking | Kendall's tau: 0.981 -> 0.915 | Hurt |
| increase calibration fraction from 10% to 40% (ablation) | bias-corrected estimator, GPT-4.1 judge · Chatbot Arena ranking | exact ranking recovery: 46% -> 75% | Improved |
| use difference/calibration-only/conditional calibration estimators under label distribution shift | existing bias-correction estimators · Monte Carlo distribution shift | bias | Hurt |
| use misclassification-adjusted estimator under label distribution shift | bias-corrected estimator · Monte Carlo distribution shift | bias | Improved |
| use calibrated LLM judge instead of human-only evaluation (current judges on Chatbot Arena) | bias-corrected estimator · Chatbot Arena | variance | Hurt |
| use GPT-4.1-mini judge on AlpacaEval instead of Chatbot Arena | bias-corrected estimator · AlpacaEval | variance vs human-only | Improved |
Closest papers by meaning
- An Empirical Analysis of Uncertainty in Large Language Model Evaluations (ICLR 2025 · similarity 0.93)
- Benchmarking LLMs via Uncertainty Quantification (NeurIPS 2024 · similarity 0.91)
- Large Language Models Must Be Taught to Know What They Don’t Know (NeurIPS 2024 · similarity 0.91)
- Position: Don't Use the CLT in LLM Evals With Fewer Than a Few Hundred Datapoints (ICML 2025 · similarity 0.90)
- Can LLMs Express Their Uncertainty? An Empirical Evaluation of Confidence Elicitation in LLMs (ICLR 2024 · similarity 0.90)
- Justice or Prejudice? Quantifying Biases in LLM-as-a-Judge (ICLR 2025 · similarity 0.90)
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.
