Research map / Large language models / Language, safety and interpretability
Tuning LLM Judge Design Decisions for 1/1000 of the Cost
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.
Systematically tunes LLM-judge hyperparameters (open-weight model, prompt format, temperature, order averaging) using multi-objective multi-fidelity search with successive halving and non-dominated sorting, finding judges that match or beat prior judges at a fraction of the cost.
Problem
LLM-as-a-judge configurations vary confoundingly across papers (model, prompt, hyperparameters), and evaluating a judge via Spearman correlation on a grid of models is prohibitively expensive, so judge design choices are rarely tuned systematically.
Approach
Uses human agreement (cheaper, lower-variance metric) as the optimization target and searches a 4,480-configuration judge space with multi-fidelity successive halving plus non-dominated sorting over accuracy and cost, yielding open-weight judges at roughly 1/1000 the cost of Alpaca-Eval/Arena-Hard style evaluation.
Key results
- Tuned open-weight judges reach 0.49 human agreement on LMSys test (Ours-large) vs 0.50 for Arena-Hard judge at lower cost (0.48 vs 1.2 USD per 1K annotations), and outperform PandaLM-7B (0.38) and JudgeLM-7B (0.42).
- On PandaLM test set, Ours-medium reaches 0.78 human agreement, above GPT-4 (0.67) and PandaLM-70B (0.67).
- The full search over 4,480 configurations costs about 2.1K USD versus an estimated ~2M USD for Alpaca-Eval/Arena-Hard-style evaluation of the same configurations.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Increase the judge model size | Qwen2.5 judge · LMSys | human agreement: 0.36 $\pm$ 0.006 -> 0.50 $\pm$ 0.006 | Helped |
| Increase the judge model size | Qwen2.5 judge · Alpaca-Eval and Arena-Hard | Spearman correlation: 0.09 $\pm$ 0.189; 0.90 $\pm$ 0.052; 0.88 $\pm$ 0.074 | Mixed |
| Use human agreement instead of Spearman correlation to rank judge configurations | Judge evaluation · Judge scaling | ability to distinguish 32B and 72B judges | Helped |
| Use the tuned judge configuration | LMSys judge evaluation · LMSys test instructions | human agreement: 0.42 +/- 0.01 -> 0.45 +/- 0.01 | Helped |
| Use the tuned judge configuration | LMSys judge evaluation · LMSys test instructions | cost per 1K annotations: 1.2 -> 0.21 | Helped |
| Use the tuned judge configuration | Arena-Hard judge evaluation · Arena-Hard | Spearman correlation: 0.90 +/- 0.06; 0.93 +/- 0.05 | Mixed |
| Use the tuned judge configuration | PandaLM judge evaluation · PandaLM test set | human agreement: 0.67 -> 0.78 | Helped |
| Use pair output format (ablation) | LLM judge configurations · Hyperparameter analysis | human agreement | Helped |
| Lower the judge temperature (ablation) | LLM judge configurations · Hyperparameter analysis | human agreement | Helped |
| Average judgments across both answer orders (ablation) | LLM judge configurations · Hyperparameter analysis | human agreement | Helped |
| Provide an example in the prompt (ablation) | Large LLM judges · Hyperparameter analysis | human agreement | Helped |
| Provide an example in the prompt (ablation) | Small LLM judges · Hyperparameter analysis | human agreement | Hurt |
Surprising findings
- Asking the judge to provide an explanation or its own answer hurts performance, contrary to chain-of-thought expectations.
- Scaling judge model size alone is insufficient; prompt and other hyperparameters have large influence even for a fixed model.
Limitations
- Selection optimizes only accuracy and cost, not position, verbosity or stylistic biases; flip rate correlates negatively with human agreement (r = -0.789), so biases are not explicitly controlled.
- Human agreement on LMSys is low (around 0.5) because the chosen instructions are harder, and results are tied to the current generation of open models and to the LMSys distribution.
Experimental setup
Datasets: LMSys Chatbot Arena battles Alpaca-Eval Arena-Hard PandaLM test set Chatbot Arena ELO ratings
Benchmarks: LMSys PandaLM test set Arena-Hard
Metrics: human agreement Spearman correlation cost per 1K annotations
Baselines: Arena-Hard judge (GPT-4 / Claude) PandaLM-7B JudgeLM-7B Length baseline GPT-4 GPT-3.5 PandaLM-70B
Builds on: successive halving non-dominated sorting Arena-Hard Alpaca-Eval
Closest papers by meaning
- JudgeLM: Fine-tuned Large Language Models are Scalable Judges (ICLR 2025 · similarity 0.92)
- RocketEval: Efficient automated LLM evaluation via grading checklist (ICLR 2025 · similarity 0.91)
- tinyBenchmarks: evaluating LLMs with fewer examples (ICML 2024 · similarity 0.90)
- MLLM-as-a-Judge: Assessing Multimodal LLM-as-a-Judge with Vision-Language Benchmark (ICML 2024 · similarity 0.90)
- JudgeBench: A Benchmark for Evaluating LLM-Based Judges (ICLR 2025 · similarity 0.89)
- An Empirical Analysis of Uncertainty in Large Language Model Evaluations (ICLR 2025 · similarity 0.89)
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.