Llm judges perform differently based on protocols and usage methods
From Judgment Quality to Downstream Utility: Rethinking LLM-as-a-Judge for Open-Ended Tasks
Artificial Intelligence
Summary
Judging outputs from AI language models is tricky because often there's no single correct answer. The authors studied how different ways of asking AI judges for opinions affect both how good their judgments are and how useful those judgments prove later. They found that better judgment quality doesn't always mean better results when using those judgments, and how you design the judging process really matters. Also, guiding AI systems with these judgments during testing can improve performance depending on how it's done.
What this means in practice
- •For ai development teams: Optimize AI judgment protocols to better train policies and improve language model outputs in complex tasks.
- •For customer support teams: Improve AI-assisted response selection during live interactions by using judge-guided methods tested for performance gains.
Authors
Zheng Zhang, Lufei Li, Xinyue Tan, Yuanhao Zeng, Ziwei Shan, Yexin Li, Kan Ren
Abstract
LLM-as-a-Judge is increasingly used to evaluate policy responses on open-ended tasks that lack ground-truth answers. Existing work often directly converts the resulting judgments into reward signals for policy training, paying limited attention to intrinsic judgment quality and largely restricting the use of Judges to training-time supervision. We systematically investigate judgment quality and downstream utility by examining both how judgments are elicited and how they are used. For judgment elicitation, we vary the Judge protocol along three dimensions: verdict granularity, critique usage, and evaluation batching. For judgment usage, beyond policy training, we extend Judge to test-time inference through Best-of-N selection, Judge-guided revision, and beam search. We find that, (i) Surprisingly, judgment quality and downstream utility do not always align. (ii) Judge protocol design substantially affects both intrinsic judgment quality and downstream utility. (iii) Judge guidance effectively converts test-time compute into performance gains, with benefits varying across inference strategies. Our results call for a multifaceted evaluation of LLM Judges on open-ended tasks, encompassing intrinsic judgment quality, and downstream utility.