LegalWorld: A Life-Cycle Interactive Environment for Legal Agents
Songhan Zuo, Shengbin Yue, Tao Chiang, Guanying Li, Yun Song, Xuanjing Huang, Zhongyu Wei
Why It Matters
What makes this one worth your time
Understanding and evaluating legal agents in a life-cycle context can lead to more effective AI systems in legal domains, potentially improving legal processes and outcomes.
LegalWorld models the full life-cycle of civil litigation to evaluate legal agents' performance across interconnected stages.
Summary
The paper introduces LegalWorld, an interactive environment simulating the life-cycle of Chinese civil litigation, capturing causal dependencies across five stages. It includes a reusable infrastructure for maintaining consistency and a benchmark, LongJud-Bench, to evaluate legal agents' capabilities across these stages.
Key contributions
- Development of LegalWorld, an interactive environment for modeling civil litigation.
- Creation of LongJud-Bench, a benchmark for evaluating legal agents across multiple stages of litigation.
- Grounding the environment in a large dataset of 75,309 paired Chinese civil judgments.
Notable insights
- The use of a life-cycle model to capture cross-stage causal dependencies in legal processes.
- The integration of local and global memory systems to maintain consistency across the litigation life cycle.
Possible limitations
- Not stated in the abstract
Abstract
arXiv:2606.18728v1 Announce Type: new Abstract: Civil litigation is inherently a life-cycle process: what a lawyer drafts on day one constrains what unfolds at trial months later. Yet existing legal benchmarks evaluate isolated subtasks, and prior legal-agent simulators reinitialize each scenario from shared ground truth, leaving cross-stage causal dependencies unmodeled. We present LegalWorld, a life-cycle interactive environment that models Chinese civil litigation as a causally connected state chain of five stages (seven sub-scenarios), grounded in 75,309 paired Chinese civil judgments. We pair it with reusable infrastructure (local memory, global case memory, a Skill/Tool library) that keeps each dispute consistent across its full life cycle. Building on this environment, we construct LongJud-Bench to evaluate agent capability across all five connected stages. 18,992 ratings from 217 legal-background evaluators confirm that LegalWorld trajectories are procedurally faithful and role-consistent; and a capability-level cross-model evaluation reveals sharp divergences that aggregate scores cannot expose, with no single backbone leading across consultation, drafting, and courtroom advocacy. Detailed resources will be released publicly.