Multimodal Multi-Agent Empowered Legal Judgment Prediction

A research team led by Zhaolu Kang and Junhao Gong has introduced JurisMMA, a novel multimodal multi-agent framework designed to predict outcomes in legal cases. Published on arXiv and accepted to ICASSP 2026, this system addresses limitations in traditional legal prediction methods that struggle with complex cases involving multiple allegations and diverse evidence types. Unlike previous approaches relying on statistical analysis or role-based simulations, JurisMMA organizes legal reasoning into distinct stages using specialized AI agents that can handle both textual case descriptions and multimodal video-text evidence.

The researchers simultaneously released JurisMM, a substantial new dataset containing over 100,000 recent Chinese judicial records that include both text and video-text data, enabling comprehensive evaluation of multimodal legal AI systems. Experiments demonstrate the framework's effectiveness on both this new dataset and the established LawBench benchmark. This work represents a significant advancement in legal AI by providing a structured, agent-based approach to case analysis that could eventually assist legal professionals in predicting outcomes, analyzing evidence patterns, and standardizing judicial processes across complex cases.