https://dylanchiang-dev.github.io/en/tags/legal-nlp/Legal NLPHugo Blox Builder (https://hugoblox.com)en-USWed, 17 Jun 2026 00:00:00 +0000https://dylanchiang-dev.github.io/media/icon_hu_982c5d63a71b2961.pnghttps://dylanchiang-dev.github.io/en/tags/legal-nlp/https://dylanchiang-dev.github.io/en/publication/legalworld/Wed, 17 Jun 2026 00:00:00 +0000https://dylanchiang-dev.github.io/en/publication/legalworld/<p>This paper is included in the <strong>EMNLP 2026 Main Conference</strong>.</p> <h2 id="abstract">Abstract</h2> <p>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.</p> <p>We present LegalWorld, a life-cycle interactive environment that models Chinese civil litigation as a causally connected state chain of five stages and seven sub-scenarios, grounded in 75,309 paired Chinese civil judgments. We also construct LongJud-Bench to evaluate agent capability across all five connected stages.</p> <p>Ratings from 217 legal-background evaluators, totaling 18,992 evaluations, confirm that LegalWorld trajectories are procedurally faithful and role-consistent. Capability-level cross-model evaluation further reveals divergences that aggregate scores cannot expose.</p>