arXivDaily arXiv每日学术速递 周一至周五更新
arXiv周末暂无论文更新,休息一下吧,周末愉快~~
arXiv 2610.18299cs.CL
收藏

LegalRewardBench:用于法律偏好评估的奖励模型基准测试

LegalRewardBench: Benchmarking Reward Models for Legal Preference Evaluation

Harbin Institute of Technology(哈尔滨工业大学);ModelBest Inc.(模贝斯特公司);The University of Osaka(大阪大学)

Hui Huang, Jie Zhou, Hongli Zhou, Qundong Shi, Shuyuan Zheng, Yini Su, Shaochun Wang, Biyuan Lin, Muyun Yang, Tiejun Zhao

AI总结:

本研究推出首个法律领域奖励模型基准LegalRewardBench,评估三类模型后发现通用生成式模型是最稳健的法律奖励模型,微调模型迁移可靠性较低,还对比了输入组织与训练策略。

摘要:

奖励模型是法律任务偏好对齐的核心,但现有奖励模型基准仍主要聚焦于通用领域偏好。本研究推出LegalRewardBench,首个经人工验证的法律领域奖励模型评估基准。该基准将法律偏好评估划分为知识、理解、分析、决策、写作五个能力层级,偏好示例通过基于评分规则的自动评判结合人工验证构建。我们评估了三类评判模型,涵盖通用模型、微调模型以及生成式和判别式架构。实验结果表明,当前强大的通用生成式模型是最稳健的法律奖励模型,而现有微调奖励模型向法律偏好评估的迁移可靠性较低。我们还对输入组织方式和训练策略进行了详细对比。

英文摘要:

Reward models are central to preference alignment for legal tasks, yet existing reward model benchmarks are still primarily centered on general-domain preferences. In this work, we introduce LegalRewardBench, the first human-verified benchmark for evaluating reward models in the legal domain. The benchmark organizes legal preference evaluation into five capability levels: knowledge, understanding, analysis, decision, and writing, and preference examples are constructed through rubric-guided automatic judging combined with human verification. We evaluate three families of judges, covering both general and fine-tuned models, and generative and discriminative architectures. Experimental results show that strong general-purpose generative models are currently the most robust legal reward models, while existing fine-tuned reward models transfer less reliably to legal preference evaluation. We also conduct detailed comparisons of input organizations and training strategies.

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。

↑