Brief

Researchers have introduced LegalWorld, an interactive environment designed to simulate the entire lifecycle of legal agents in Chinese civil litigation. This system models the process across five distinct stages, maintaining consistency through local and global memory, and a skill/tool library. To evaluate agent capabilities within this framework, they developed LongJud-Bench, which uses over 18,000 ratings from legal professionals to assess procedural faithfulness and role consistency. Initial evaluations using LongJud-Bench revealed significant performance differences among various AI models across different legal tasks, indicating that aggregate scores do not fully capture an agent's overall competence. AI