我们是否在正确地评估LLM作为裁判？
Are We on the Right Way to Assessing LLM-as-a-Judge?
AI总结：
本文提出Sage评估套件，通过局部和全局一致性指标评估LLM作为裁判的可靠性，发现当前先进模型在复杂情况下存在显著偏好不一致问题，并揭示情境偏好现象。
摘要：
将LLM作为裁判（LLM-as-a-Judge）已广泛应用于评估方法，并在模型训练中作为监督奖励使用。然而，现有的LLM-as-a-Judge基准主要依赖于人工标注的地面真相，这引入了人为偏差，削弱了评估的可靠性，并施加了可扩展性限制。为克服这些限制，我们引入了Sage，一种新的评估套件，能够在不依赖任何人工标注的情况下评估LLM裁判的质量。受理性选择理论公理的启发，Sage引入了两个新的测量LLM-as-a-Judge的视角：局部自一致性（成对偏好稳定性）和全局逻辑一致性（在完整偏好集上的传递性）。我们通过结合结构化基准问题与真实用户查询，精心整理出包含650个问题的数据集。我们的实验表明，我们的指标具有稳定性，并且与监督基准如LLMBar和RewardBench2具有高度相关性，证实了Sage作为评估套件的可靠性，用于评估LLM-as-a-Judge的鲁棒性和准确性。基于Sage，我们发现当前最先进的LLM在作为裁判进行评分和成对设置时存在显著的可靠性问题；即使是最顶尖的模型，Gemini-2.5-Pro和GPT-5，在近四分之一的困难案例中也未能保持一致的偏好。我们将其归因于一种新的现象，称为情境偏好，解释了为什么显式的评分标准或准则可以帮助模型在答案对之间保持一致的判断。我们的进一步分析显示，微调的LLM-as-a-Judge是一种可行的方法来提升性能，而基于小组的裁判以及深度推理可以增强判断的一致性。我们还发现人类判断中存在显著的一致性问题，这表明人工标注可能不是一个可靠的黄金标准。
英文摘要：
LLM-as-a-Judge has been widely adopted as an evaluation method and served as supervised rewards in model training. However, existing benchmarks for LLM-as-a-Judge are mainly relying on human-annotated ground truth, which introduces human bias that undermines the assessment of reliability and imposes scalability constraints. To overcome these limitations, we introduce Sage, a novel evaluation suite that assesses the quality of LLM judges without necessitating any human annotation. Inspired by axioms of rational choice theory, Sage introduces two new lenses for measuring LLM-as-a-Judge: local self-consistency (pair-wise preference stability) and global logical consistency (transitivity across a full set of preferences). We curate a dataset of 650 questions by combining structured benchmark problems with real-world user queries. Our experiments demonstrate both the stability of our metrics and their high correlation with supervised benchmarks like LLMBar and RewardBench2, confirming Sage's reliability as an evaluation suite for the robustness and accuracy of LLM-as-a-Judge. Based on Sage, we reveal that current state-of-the-art LLMs exhibit significant reliability problems when acting as judges in both scoring and pairwise settings; even the top-performing models, Gemini-2.5-Pro and GPT-5, fail to maintain consistent preferences in nearly a quarter of difficult cases. We attribute this to a new phenomenon called situational preference, which explains why explicit rubrics or criteria can help the model judge consistently across answer pairs. Our further analysis shows that finetuned LLM-as-a-Judge is a feasible method to boost performance, and the panel-based judge as well as deep reasoning can enhance the judging consistency. We also find substantial inconsistency in human judgments, which indicates that human annotation may not be a reliable gold standard.
机构、总结、中文摘要均由AI分析而来，可能存在错误，请以论文正文为准。
