Reproduce “A Survey on LLM-as-a-Judge” on OpenResearch

About this paper

Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen and 8 more

The survey presents a practical framework for building and evaluating language-model judges, organizing reliability around agreement with humans, bias, and adversarial robustness. It reviews prompt design, model selection, and output processing, then compares strategies in a simple meta-evaluation. For GPT-3.5-turbo, adding explanations coincided with agreement falling from 54.72% to 52.47%, showing no universal reliability recipe.

Key insights

• Reliability is identified as the paramount challenge for LLM-as-a-Judge adoption, requiring deliberate design to ensure consistency, bias mitigation, and robustness.
• GPT-4-turbo consistently exhibits superior performance in human alignment and demonstrates fewer biases compared to other models, positioning it as a current leading model for evaluation tasks.
• The effectiveness of improvement strategies varies; multi-round majority voting significantly enhances reliability and mitigates biases, while 'providing evaluations with explanations' and 'self-validation' show limited or even negative impacts.

Summary by alphaXiv.