judge Logo
Jiawei Gu1*, Xuhui Jiang1*, Zhichao Shi1,2*, Hexiang Tan2, Xuehao Zhai3, Chengjin Xu1, Wei Li2, Yinghan Shen2, Shengjie Ma1,4, Honghao Liu1, Kun Zhang2, Saizhuo Wang1,6, Yuanzhuo Wang2, Wen Gao5, Lionel Ni6,7, Jian Guo1†
1IDEA Research, 2Chinese Academy of Sciences, 3Imperial College London, 4Renmin University of China, 5Peking University, 6HKUST, 7HKUST(GZ)

*Equal contribution

†Corresponding author

Paper arXiv Paper List Meta-eval Talk
judge pipeline

LLM-as-a-Judge Evaluation Pipeline.

🔔News

🔥 [2025-01-28]: We added analysis on LLM-as-a-Judge and o1-like Reasoning Enhancement, as well as meta-evaluation results on o1-mini, Gemini-2.0-Flash-Thinking-1219, and DeepSeek-R1!

🌟 [2025-01-16]: We shared and discussed the methodologies, applications (Finance, RAG, and Synthetic Data), and future research directions of LLM-as-a-Judge at BAAI Talk! 🤗
[Replay] [Methodology] [RAG & Synthetic Data]

🚀 [2024-11-23]: WWe released A Survey on LLM-as-a-Judge, exploring LLMs as reliable, scalable evaluators and outlining key challenges and future directions!

Abstract

Accurate and consistent evaluation is crucial for decision-making across numerous fields, yet it remains a challenging task due to inherent subjectivity, variability, and scale. Large Language Models (LLMs) have achieved remarkable success across diverse domains, leading to the emergence of "LLM-as-a-Judge," where LLMs are employed as evaluators for complex tasks. With their ability to process diverse data types and provide scalable and flexible assessments, LLMs present a compelling alternative to traditional expert-driven evaluations.

However, ensuring the reliability of LLM-as-a-Judge systems remains a significant challenge that requires careful design and standardization. This paper provides a comprehensive survey on LLM-as-a-Judge, offering a formal definition and a detailed classification, while focusing on addressing the core question: How to build reliable LLM-as-a-Judge systems? We explore strategies to enhance reliability, including improving consistency, mitigating biases, and adapting to diverse assessment scenarios. Additionally, we propose methodologies for evaluating the reliability of LLM-as-a-Judge systems, supported by a novel benchmark designed for this purpose. To advance the development and real-world deployment of LLM-as-a-Judge systems, we also discussed practical applications, challenges, and future directions. This survey serves as a foundational reference for researchers and practitioners in this rapidly evolving field.

Motivation

Meta-evaluation Pipeline and Results

We also conducted a meta-evaluation of improvement strategies for LLM-as-a-Judge systems using benchmarks such as LLMEval2 and EVALBIASBENCH to assess their effectiveness in optimizing evaluation performance and mitigating biases. Notably, we analyzed the performance of reasoning-enhanced LLMs, demonstrating their advancements over baseline models but highlighting their inconsistent improvements in alignment-related tasks.

eval_pipeline eval_benchmark

BibTeX

@article{gu2024survey,
  title={A Survey on LLM-as-a-Judge},
  author={Gu, Jiawei and Jiang, Xuhui and Shi, Zhichao and Tan, Hexiang and Zhai, Xuehao and Xu, Chengjin and Li, Wei and Shen, Yinghan and Ma, Shengjie and Liu, Honghao and others},
  journal={arXiv preprint arXiv:2411.15594},
  year={2024}
}