VERDICT: A Framework for Evolving Legal Judgment Prediction

Legal Judgment Prediction (LJP) aims to identify relevant laws, charges, and penalties based on case details. Traditional methods often lack adaptability and interpretability, functioning merely as static predictors. In response, we present VERDICT, an innovative framework that mimics a collaborative panel of legal experts. This system employs specialized agents for various tasks such as fact organization, legal research, opinion formulation, and oversight verification. These agents work together in a systematic process of drafting, verifying, and revising, allowing for clear documentation of reasoning and revision justifications. To ensure the framework evolves with legal practices, we introduce the Hybrid Jurisprudential Memory (HJM), which is based on the Micro-Directive Paradigm. This memory system retains precedent standards and refines multi-agent verification paths into updated directives, promoting continuous learning. We tested VERDICT using the CAIL2018 dataset and a newly developed CJO2025 dataset, achieving top performance on CAIL2018 and demonstrating effective generalization on CJO2025. Our commitment to transparency and further research is underscored by the release of our code and datasets for public access.