Reproduce “Improving LLM-as-a-Judge Inference with the Judgment Distribution” on OpenResearch

About this paper

Victor Wang, Michael J.Q. Zhang, Eunsol Choi

This work demonstrates that leveraging the full probability distribution of judgment tokens from large language models (LLMs) significantly improves their performance as evaluators. The study finds that using the mean of the judgment distribution consistently outperforms traditional greedy decoding, and that Chain-of-Thought prompting often reduces evaluation accuracy by sharpening the judgment distribution.

Key insights

• The mean of the judgment distribution consistently outperforms the mode (greedy decoding) as an inference method for LLM-as-a-judge across diverse tasks and models, providing more accurate and calibrated judgments.
• Chain-of-Thought prompting frequently harms LLM-as-a-judge performance by 'sharpening' the judgment distribution, which reduces its spread and can lead to a loss of valuable signal.
• Continuous methods for extracting judgments are generally more accurate than discrete methods due to lower tie rates, and increasing score granularity can mitigate the tie problem for discrete methods.

Summary by alphaXiv.