LegalWorld: A Life-Cycle Interactive Environment for Legal Agents

LegalWorld is an interactive environment designed to model the entire life cycle of Chinese civil litigation. It encompasses five interconnected stages—Legal Consultation, Complaint/Defense Drafting, First-Instance Trial, Appeal Drafting/Response, and Second-Instance Trial—each comprising specific sub-scenarios. This comprehensive approach ensures that the complexities and causal dependencies inherent in the litigation process are accurately represented, providing a realistic simulation for legal agents.

The environment is grounded in a substantial dataset of 75,309 paired Chinese civil judgments, which serves as the foundation for its simulations. To maintain consistency across the litigation process, LegalWorld incorporates reusable infrastructure elements such as local memory, global case memory, and a Skill/Tool library. These components ensure that each dispute remains consistent throughout its full life cycle, facilitating a more accurate evaluation of legal agents' capabilities.

In addition to its simulation capabilities, LegalWorld introduces LongJud-Bench, a framework designed to assess agent performance across all five connected stages. This evaluation is supported by 18,992 ratings from 217 legal-background evaluators, confirming that the environment's trajectories are procedurally faithful and role-consistent. The cross-model evaluation reveals significant divergences that aggregate scores alone cannot expose, highlighting the need for specialized models in areas like consultation, drafting, and courtroom advocacy.

LegalWorld is particularly beneficial for researchers and practitioners in the field of legal technology, offering a robust platform for testing and refining legal agents. Its detailed and interconnected simulation of the litigation process makes it an invaluable tool for those seeking to enhance the performance and reliability of legal AI systems.