arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2403.02839cs.CL
收藏

关于用于 LLM 评估的 LLM-as-a-Judge 的实证研究:微调的 Judge Model 并非 GPT-4 的通用替代品

An Empirical Study of LLM-as-a-Judge for LLM Evaluation: Fine-tuned Judge Model is not a General Substitute for GPT-4

Faculty of Computing, Harbin Institute of Technology(哈尔滨工业大学计算机学院);School of Computer Science, Beijing Institute of Technology(北京理工大学计算机学院);Baidu Inc.(百度公司)

Hui Huang, Xingyuan Bu, Hongli Zhou, Yingqi Qu, Jing Liu, Muyun Yang, Bing Xu, Tiejun Zhao

AI总结:

本研究发现,尽管微调的 Judge Model 在域内测试集上表现优异甚至超越 GPT-4,但在泛化性、公平性和适应性上逊色于 GPT-4,其本质是任务特定分类器,无法作为通用替代品。

摘要:

近期,利用大型语言模型(LLM)评估其他 LLM 质量的趋势日益增长。许多研究基于开源 LLM 微调了 judge models 用于评估。尽管声称微调的 judge models 取得了与 GPT-4 相当的评估能力,但在本研究中,我们对 LLM-as-a-Judge 进行了实证研究。研究结果表明,尽管微调的 judge models 在域内测试集上表现优异甚至超越 GPT-4,但它们在泛化性、公平性和适应性等多个维度上均逊色于 GPT-4。我们还揭示,微调的 judge model 本质上作为任务特定分类器运行,从而施加了这些限制。

英文摘要:

Recently, there has been a growing trend of utilizing Large Language Model (LLM) to evaluate the quality of other LLMs. Many studies have fine-tuned judge models based on open-source LLMs for evaluation. While the fine-tuned judge models are claimed to achieve comparable evaluation capability with GPT-4, in this work, we conduct an empirical study of LLM-as-a-Judge. Our findings indicate that although the fine-tuned judge models achieve high performance on in-domain test sets, even surpassing GPT-4, they underperform GPT-4 across several dimensions, including generalizability, fairness and adaptability. We also reveal that the fine-tuned judge model inherently operates as a task-specific classifier, consequently imposing the limitations.

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。

↑