Student or educator? Sign up with your school email and get 6 months of Pro free.
CoChat

The DOI matches its DataCite record; 59 of the 59 references it cites were found in a registry or an index; 4 of the 29 sources its claims cite were checked for support.

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

Open access

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.

Overview

CoChat Summary

TL;DR

This paper examines whether strong LLMs can serve as judges for chatbot responses by comparing their judgments with human preferences on two new benchmarks, MT-bench and Chatbot Arena. It finds that GPT-4 can reach over 80% agreement with human preferences, comparable to human-human agreement, while also documenting several judge biases and mitigation strategies.

Key findings

  • GPT-4’s judgments aligned closely with human experts on MT-bench: 85% agreement without ties, versus 81% agreement among humans; on Chatbot Arena, GPT-4 also showed high agreement with crowd votes, above 80%.
  • Position bias was substantial: with the default prompt, consistency was 65.0% for GPT-4, 46.2% for GPT-3.5, and 23.8% for Claude-v1; swapping positions and few-shot examples raised GPT-4 consistency to 77.5%.
  • Verbosity bias appeared in a “repetitive list” attack: Claude-v1 and GPT-3.5 failed 91.3% of cases, while GPT-4 failed 8.7%.
  • Reference-guided grading reduced math-judging failures from 14/20 under the default prompt to 3/20, with chain-of-thought at 6/20; self-enhancement was suggested by GPT-4 favoring itself with a 10% higher win rate and Claude-v1 with a 25% higher win rate, but the study could not confirm the bias.

Methods

  • The authors created MT-bench, 80 multi-turn questions across 8 categories with 10 questions each, and Chatbot Arena, a crowdsourced anonymous battle platform; they collected about 3K expert votes from 58 labelers on MT-bench and about 30K Arena votes, sampling 3K for analysis.
  • They compared LLM judges—mainly GPT-4, GPT-3.5, and Claude-v1—using pairwise comparison, single-answer grading, and reference-guided grading against human expert and crowd votes, measuring agreement and win rates.
  • They also evaluated LLaMA and Vicuna variants on MMLU, TruthfulQA, and MT-bench with GPT-4 single-answer grading to test how preference-based and capability-based benchmarks complement each other.

Limitations

  • The paper states that it emphasizes helpfulness and largely neglects safety, honesty, and harmlessness; it also notes that helpfulness dimensions such as accuracy, relevance, and creativity are combined into one metric, and its bias mitigations are preliminary.
  • The paper reports that self-enhancement bias could not be determined because of limited data and small differences, and that its position-bias test is difficult because the compared answers are very similar and sometimes indistinguishable even to humans.
  • The text makes evident that few-shot judging may introduce new biases and makes API calls about 4× more expensive, while zero-shot Vicuna-13B as a judge had low consistency (11.2–16.2%) and high error rates (22.5–78.8%), limiting its use as a scalable open-source judge.

Model-written summary. Check it against the paper before relying on it.

Integrity signals

Registry

DataCite · 10.48550/arxiv.2306.05685

Cites

59 of 59 references verified0 citation errors · 0 not checkable

Statistics

Nothing to recompute.

Verification

Claims the paper makes

3 Valid · 2 Not valid · 8 Not checked · 98 Own results and interpretations

60 more claims

Create a free account to see all 111 claims and every source verdict.

See all claims — free account
  • Probably not a substantial claim. 22
  • It may state two separate claims as one. 29
  • The quote only partly states this claim. 2

Create a free account to see these claims.

Claims other papers make about this one

Each is checked against this paper’s own text: does it say what the citing paper says it does?

106 more claims about this paper

Create a free account to see every claim other papers make about this one.

See all claims — free account

Citations

Cites

Show all 2 references

Research works printed but not linked: 2

  1. Chin-Yew Lin. Rouge: A package for automatic evaluation of summaries. In Text summarization branches out, pages 74-81, 2004.
  2. Zongheng Yang, Zhanghao Wu, Michael Luo, Wei-Lin Chiang, Romil Bhardwaj, Woosuk Kwon, Siyuan Zhuang, Frank Sifei Luan, Gautam Mittal, Scott Shenker, and Ion Stoica. SkyPilot: An intercloud broker for sky computing. In 20th USENIX Symposium on Networked Systems Design and Implementation (NSDI 23), pages 437-455, Boston, MA, April 2023. USENIX Association.

Cited by

See all 2,091

Metadata

Authors
Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E. P., Zhang, H., Gonzalez, J. E., & Stoica, I.
Venue
arXiv
Publisher
arXiv
Published
2023
Pages
46595-46623
ISSN
1049-5258 (ISSN-L)
OpenAlex
W7133239970
Cited by
1,888
References
2
Citations by year
2026: 1115 · 2025: 413 · 2024: 308 · 2023: 51
FWCI
2548.93
Type
conference paper
Language
en
Open access
Locations
  • neural information processing systems (publishedVersion)
  • arXiv (Cornell University) (submittedVersion)
  • arXiv (Cornell University)
Funders
  • VMware
  • Samsung
Topics
  • Social Sciences › Law › Artificial Intelligence in Law
  • Physical Sciences › Computer Science › Artificial Intelligence › Topic Modeling
  • Physical Sciences › Computer Science › Artificial Intelligence › Natural Language Processing Techniques
Keywords
large language models, LLM-as-a-judge, automated assessment, MT-Bench, Chatbot Arena, benchmarking, chatbots

About these checks

Registry, retraction, Cites and statistics results are computed, not judged, with one exception: a printed reference that no index lists is read once by a model before it counts as a citation error. Claims are extracted by a model and kept only when their quote appears verbatim in the text; each cited source is then checked, with a model’s help, for whether it says what the claim relies on. Those verdicts can be wrong — read the claim’s quote and the cited source before relying on one.

The claim verdicts this report counts are experimental: their accuracy is not yet measured.

A language model drew the claims from the paper on 2026-10-29, under version 3 of the claim definition (model deepseek-v4.1-flash, prompt extract-v3).

A second model labelled each claim on 2026-10-29 (model jev-1.13-20260917, definition revision gold-guideline-v0.2).

Cited sources are checked with a language model’s help (gpt-6-sol).

Statistics and arithmetic are recomputed, not judged.

Confidence bands are shown only for fields calibrated on our annotated test set. None is calibrated yet, so no band is shown.

What counts as a claim ›How we measure confidence ›How we check this ›

Are you an author, or do you see a problem with this page? Request a review

How to cite

Cite CoChat for the checks on this page, and the paper for its own findings.

This page
CoChat. (2026). Citation check: Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. https://app.cochat.ai/papers/10.48550/arxiv.2306.05685
The paper
Zheng, L., Chiang, W.-L., Sheng, Y., Zhuang, S., Wu, Z., Zhuang, Y., Lin, Z., Li, Z., Li, D., Xing, E. P., Zhang, H., Gonzalez, J. E., & Stoica, I. (2023). Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena. arXiv. https://doi.org/10.48550/arxiv.2306.05685

Data sources

Bibliographic data: Semantic Scholar, OpenAlex, Crossref.

Contains information from the Semantic Scholar Academic Graph Datasets, made available under the ODC Attribution License (ODC-BY 1.0).

Semantic Scholar’s data is described in: Kinney, R., et al. (2023). The Semantic Scholar Open Data Platform. arXiv:2301.10140.