Can Large Language Models Grasp Legal Theories? Enhance Legal Reasoning with Insights from Multi-Agent Collaboration

TL;DR

A challenging task (confusing charge prediction) is introduced to better evaluate LLMs' understanding of legal theories and reasoning capabilities and a novel framework is proposed: Multi-Agent framework for improving complex Legal Reasoning capability (MALR).

Abstract

Large Language Models (LLMs) could struggle to fully understand legal theories and perform complex legal reasoning tasks. In this study, we introduce a challenging task (confusing charge prediction) to better evaluate LLMs' understanding of legal theories and reasoning capabilities. We also propose a novel framework: Multi-Agent framework for improving complex Legal Reasoning capability (MALR). MALR employs non-parametric learning, encouraging LLMs to automatically decompose complex legal tasks and mimic human learning process to extract insights from legal rules, helping LLMs better understand legal theories and enhance their legal reasoning abilities. Extensive experiments on multiple real-world datasets demonstrate that the proposed framework effectively addresses complex reasoning issues in practical scenarios, paving the way for more reliable applications in the legal domain.

Figures (12)

• Figure 1: The performance of LLMs on predicting the golden (Misappropriation of Public Fund) or confusing charge (Fund Misappropriation) for the cases from CAIL-2018 datasets. The horizontal axis represents 5 advanced promt methods to solve legal reasoning problems (detailed information is described in Section \ref{['section:Experiments']}). In each method, GPT-3.5 and GPT-4 both exhibit a significant performance gap.
• Figure 2: An example to demonstrate how a judge and an LLM apply legal rules to conclude whether a case satisfies a specific charge. This example outlines two confusing charges under Chinese criminal law: the Crime of Fund Misappropriation and the Crime of Misappropriation of public fund. The most significant difference between the two charges is whether the defendant is a state functionary. In the case description, the defendant is "the Deputy Director of the Bureau of Finance in X district", a position that qualifies as a state functionary. Therefore, the judge can easily infer that the case falls under the Crime of Misappropriation of public funds, rather than Fund Misappropriation. However, the LLM fails to predict the confusing charge.
• Figure 3: Our research framework in (A) and Adaptive Rule-Insights training process in (B).
• Figure 4: Case study for a given case. The green parts mean are the most critical information for distinguish the confusing charges, the red parts are contents that do not match the facts of the case.
• Figure 5: Prompt Template for Auto-Planer, Sub-task Agent and Self-Reflector