MASLegalBench
收藏资源简介:
MASLegalBench是一个专门为多智能体系统(MAS)设计的法律推理基准,它利用了MAS在任务分解、智能体专业化和灵活训练方面的独特优势。该基准以GDPR为应用场景,包含了丰富的背景知识和复杂的推理过程,真实地反映了现实世界法律情况的复杂性。数据集由950个法律案例组成,每个案例都附有详细的背景信息,包括事实、相关法律规则、推理过程和结论。该数据集的构建过程是通过人工设计角色化的MAS,并使用不同的LLMs进行实验来完成的。MASLegalBench旨在解决现有法律推理模型在复杂推理、信息对接和领域知识方面的挑战,为MAS在法律领域的应用提供了评估基础。
MASLegalBench is a legal reasoning benchmark specifically tailored for multi-agent systems (MAS). It leverages the unique strengths of MAS in task decomposition, agent specialization, and flexible training. Taking GDPR as its targeted application scenario, the benchmark encompasses rich background knowledge and complex reasoning workflows, which authentically mirrors the complexity of real-world legal scenarios. The dataset consists of 950 legal cases, each accompanied by detailed background information including case facts, relevant legal provisions, reasoning processes, and final conclusions. The construction of MASLegalBench is completed by manually designing role-based multi-agent systems and conducting experiments with diverse large language models (LLMs). This benchmark aims to address the challenges faced by existing legal reasoning models in complex reasoning, information integration, and domain knowledge, providing a standardized evaluation basis for the application of multi-agent systems in the legal domain.

- 1通过香港科技大学 清华大学 · 2025年



