LegalWorld: A Life-Cycle Interactive Environment for Legal Agents

A team of researchers has introduced LegalWorld, a life-cycle interactive environment that models Chinese civil litigation as a causally connected state chain of five stages, aiming to address a gap in how legal AI agents are evaluated [1]. The environment, detailed in a paper submitted to arXiv on June 17, 2026, is grounded in 75,309 paired Chinese civil judgments [1][2]. Unlike existing legal benchmarks that evaluate isolated subtasks, LegalWorld simulates the entire litigation process, where actions at one stage constrain what is possible at later stages [2]. The system is paired with reusable infrastructure, including local memory, global case memory, and a Skill/Tool library, to keep each dispute consistent across its full life cycle [1][2]. Building on this environment, the researchers constructed LongJud-Bench, a benchmark designed to evaluate agent capability across all five connected stages [1][2]. To validate the procedural fidelity of the simulations, the team collected 18,992 ratings from 217 evaluators with legal backgrounds [1][2]. These ratings confirmed that the trajectories generated within LegalWorld are procedurally faithful and role-consistent [1][2]. The concept of modeling a legal life cycle mirrors the structure of corporate law, which governs the formation, funding, governance, and dissolution of a corporation [3]. The researchers' approach of creating a causally connected state chain is a departure from prior legal-agent simulators, which typically reinitialize each scenario from a shared ground truth, thereby leaving cross-stage causal dependencies unmodeled [2]. A cross-model evaluation using LongJud-Bench revealed sharp divergences in agent capabilities that aggregate scores could not expose, with no single model backbone leading across the distinct tasks of consultation, drafting, and courtroom advocacy [1][2]. The researchers stated that detailed resources for the project will be released publicly [1][2].