Skip to content
Search

Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena

Authors: Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li, Dacheng Li, Eric P. Xing, Hao Zhang, Joseph E. Gonzalez, Ion Stoica

Published June 9, 2023

Read the paper: arXiv 2306.05685 · PDF

Why it matters

Many teams use LLM-as-a-judge somewhere in their evals, and this is the paper that measured whether it can be trusted. The authors built two benchmarks to check it, MT-bench (multi-turn questions) and Chatbot Arena (crowdsourced head-to-head votes), and found that a strong judge matched both expert and crowd preferences at over 80% agreement.

The limits are the part to remember. The paper documents position bias (preferring the answer shown first or second), verbosity bias (preferring longer answers), self-enhancement bias (preferring its own answers) and weak reasoning when grading math and logic. It proposes ways to reduce some of these, which is why good judge setups look the way they do.

How to use it

When you compare two answers with a model, run the comparison twice with the order swapped and count a win only if it holds both ways, which is the rule the authors used. Tell the judge that length is not quality, and check whether it rewards padding. For math and reasoning questions, give the judge a reference answer; in the paper this worked better than asking the judge to solve the problem itself first. Use a different model from the one being graded where you can, and spot-check a sample of its grades against your own. The RAG eval harness playbook shows a judge in a working eval.

Abstract

Evaluating large language model (LLM) based chat assistants is challenging due to their broad capabilities and the inadequacy of existing benchmarks in measuring human preferences. To address this, we explore using strong LLMs as judges to evaluate these models on more open-ended questions. We examine the usage and limitations of LLM-as-a-judge, including position, verbosity, and self-enhancement biases, as well as limited reasoning ability, and propose solutions to mitigate some of them. We then verify the agreement between LLM judges and human preferences by introducing two benchmarks: MT-bench, a multi-turn question set; and Chatbot Arena, a crowdsourced battle platform. Our results reveal that strong LLM judges like GPT-4 can match both controlled and crowdsourced human preferences well, achieving over 80% agreement, the same level of agreement between humans. Hence, LLM-as-a-judge is a scalable and explainable way to approximate human preferences, which are otherwise very expensive to obtain. Additionally, we show our benchmark and traditional benchmarks complement each other by evaluating several variants of LLaMA and Vicuna. The MT-bench questions, 3K expert votes, and 30K conversations with human preferences are publicly available at https://github.com/lm-sys/FastChat/tree/main/fastchat/llm_judge.