A multi-agent framework with legal event logic graph for multi-defendant legal judgment prediction

Weikang Yuan, Kaisong Song, Zhuoren Jiang, Junjie Cao, Yujie Zhang, Chengyuan Liu, Jun Lin, Ji Zhang, Kun Kuang, Xiaozhong Liu

Abstract: Multi-defendant Legal Judgment Prediction (LJP) is a complex and challenging task in real-world legal scenarios. Existing approaches often struggle with analyzing intricate relationships among defendants and incorporating domain-specific legal expertise, particularly in penalty prediction. To address these challenges, we propose MAGLJP, a novel Multi-Agent framework with Legal Event Logic Graph for multi-defendant LJP. Our framework systematically decomposes the task into a Standard Operating Procedure, employing three specialized LLM-based agents: the Conviction Agent for law article and charge prediction, the Legal Knowledge Assistant Agent for legal knowledge integration, and the Sentencing Agent for penalty prediction. To support legal reasoning and effectively integrate domain knowledge, we introduce the Legal Event Logic Graph (LELG), a directed acyclic graph structure designed to represent and infer the complex relationships among criminal facts, legal knowledge, and sentencing outcomes. Additionally, we construct a comprehensive legal knowledge base that incorporates multiple levels of judicial interpretation. Extensive experiments on two benchmark multi-defendant LJP datasets show that MAGLJP significantly outperforms strong baselines, achieving state-of-the-art performance across all evaluation metrics. Tests conducted across diverse scenarios and case studies further demonstrate the robustness, generalization ability, and interpretability of MAGLJP, highlighting its strong applicability in real-world judicial settings.