Research map / Large language models / Alignment and preferences
Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| use reference-guided judging (ablation) | GPT-4 judge · Math grading | failure rate: 14/20 -> 3/20 | Helped |
| use chain-of-thought judge prompt (ablation) | GPT-4 judge · Math grading | failure rate: 14/20 -> 6/20 | Helped |
| add few-shot judgment examples (ablation) | GPT-4 judge · Position bias test | consistency: 65.0 -> 77.5 | Helped |
| add few-shot judgment examples (ablation) | GPT-4 judge · MT-bench | agreement with humans | No clear effect |
| use 'score' prompt (absolute scores) instead of default (ablation) | LLM judges · Position bias test | consistency: GPT-4 65.0 -> 51.2; GPT-3.5 46.2 -> 55.0 | Mixed |
| use GPT-4 instead of weaker LLM as judge | LLM-as-a-judge · Repetitive list attack | failure rate: 91.3 -> 8.7 | Helped |
| fine-tune Vicuna-13B judge on arena votes as 3-way classification | Vicuna-13B judge · Position bias test | consistency: 16.2% to 65.0% | Helped |
| fine-tune on small high-quality conversations (4.8M tokens) | LLaMA-7B · MT-bench | MT-Bench score: 2.74 -> 5.95 | Helped |
| scale dialog fine-tuning data from 4.8M to 370M tokens | Vicuna-7B · MMLU 5-shot | accuracy: 37.3 -> 47.1 | Helped |
| present full conversation in single prompt for multi-turn judging | GPT-4 judge · MT-bench | judgment accuracy | Helped |
Closest papers by meaning
- Chatbot Arena: An Open Platform for Evaluating LLMs by Human Preference (ICML 2024 · similarity 0.94)
- A Statistical Framework for Ranking LLM-based Chatbots (ICLR 2025 · similarity 0.91)
- JudgeLM: Fine-tuned Large Language Models are Scalable Judges (ICLR 2025 · similarity 0.90)
- WildBench: Benchmarking LLMs with Challenging Tasks from Real Users in the Wild (ICLR 2025 · similarity 0.90)
- Bridging Human and LLM Judgments: Understanding and Narrowing the Gap (NeurIPS 2025 · similarity 0.90)
- JudgeBench: A Benchmark for Evaluating LLM-Based Judges (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.