Reproduce “Judging LLM-as-a-Judge with MT-Bench and Chatbot Arena” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Lianmin Zheng, Wei-Lin Chiang, Ying Sheng, Siyuan Zhuang, Zhanghao Wu, Yonghao Zhuang, Zi Lin, Zhuohan Li and 5 more

The authors evaluate whether GPT-4’s judgments of chatbot answers align with human preferences in MT-bench and Chatbot Arena. They compare answers and test answer-order bias. GPT-4 agrees with humans on 85% of non-tied MT-bench votes and with Arena crowds on 87% of non-tied votes; all-vote rates are 66% and 64%. These findings concern helpfulness evaluations, not a general replacement for human review.

Key insights

  • Powerful LLMs, particularly GPT-4, can serve as effective automated judges for evaluating other LLMs, showing high agreement with human preferences.
  • LLM judges are susceptible to biases like position and verbosity bias, which can be mitigated through methods such as response swapping and reference-guided prompting.
  • A hybrid evaluation framework combining traditional capability benchmarks with human-preference-focused benchmarks (like MT-bench and Chatbot Arena) offers a more comprehensive assessment of LLMs.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.