Full‑Lifecycle Legal Simulation World: One‑Click Run or Play the Case Yourself?

1. Introduction

Legal disputes involve repeated communication, document drafting, trial, appeal, and evidence exchange, forming a complex lifecycle that existing simulators only capture as isolated points. LEGALWORLD aims to change this by providing a full‑lifecycle, LLM‑agent‑driven legal environment.

2. Construction Method

LEGALWORLD builds case seeds from 75,309 paired first‑instance and appellate civil judgments collected from China Judgments Online, covering more than 500 civil causes. Each seed contains structured fields such as parties, claims, facts, evidence, first‑instance ruling, appeal request, and appellate ruling. The environment releases information gradually according to the current stage and role.

The system is a multi‑agent platform where three heterogeneous agents—Client, Lawyer, and Judge—interact through a continuous state chain spanning five stages and seven sub‑scenes: legal consultation, first‑instance document drafting, first‑instance trial, appeal document drafting, and appellate trial.

2.1 Agents and Roles

Client Agent : Represents the litigant, provides facts, expresses demands, and responds to inquiries; its behavior follows the Legal Client Persona Framework (LCPF) covering legal literacy, disclosure willingness, emotional stability, and narrative ability.

Lawyer Agent : Handles consultation, document drafting, and trial representation; serves as the primary target for evaluation.

Judge Agent : Controls trial flow, conducts investigations and deliberations, and issues judgments for both first‑instance and appellate courts.

Each role sees only the information it is allowed to access, mirroring real‑world confidentiality constraints.

2.2 Dual‑Level Memory

Local (scene) memory : Stores dialogue and actions within the current scene, ensuring continuity.

Global case memory : Persists facts, evidence, claims, procedural progress, and party positions across stages, enabling long‑term reasoning.

After each stage, new dialogues, documents, and judgments are written back to the appropriate memory, preventing the agents from restarting from scratch.

2.3 Skills and Tools

Skill : Prescribed procedural steps for each stage (e.g., client interview, claim drafting, appellate brief preparation, courtroom argument).

Tool : Executable utilities for memory read/write, document export, legal article retrieval, citation checking, document comparison, and benchmark execution.

The environment enforces which skills and tools each role may invoke at each stage, ensuring realistic constraints.

3. Experiment Setup

The authors introduce LongJud‑Bench , a benchmark that evaluates legal agents across eight capabilities: issue spotting, identification, claim construction, fact marshalling, evidence marshalling, position consistency, evidentiary advocacy, and legal reasoning. Human judges (217 legal experts) provided 18,992 ratings, and an LLM‑as‑Judge (Claude‑Sonnet‑4.6) supplied automatic scores for large‑scale testing.

4. Results

4.1 Environment Reliability

Human evaluation gave LEGALWORLD average scores of 8.96/10 for stage realism and 8.98/10 for role consistency, with 73% of ratings ≥9 and only 4.5% ≤6, indicating high fidelity of the simulated litigation trajectory.

4.2 Model Comparison

Six LLM backbones were placed as target lawyer agents in LEGALWORLD and assessed on LongJud‑Bench. No single model excelled in all abilities; some were stronger in document drafting and claim identification, while others performed better in courtroom evidence handling and reasoning. All models struggled most with the trial stage, which requires dynamic, multi‑turn integration of case memory, opposing arguments, and judge prompts.

4.3 Skill Accumulation Exploration

The authors experimented with Reflective Legal Skill (RLS) : after a case concludes, the lawyer agent reviews the full trajectory, extracts reusable legal practice rules, deduplicates them, and adds them as new skills for future cases. On three high‑frequency causes (divorce property, private lending, labor disputes), RLS raised the average LongJud‑Bench score from 61.56 to 65.29 (≈3.7 points), and also yielded positive transfer to held‑out cases of the same causes.

5. Conclusion

LEGALWORLD demonstrates that evaluating legal agents solely on isolated tasks overlooks the challenges of maintaining consistency across a case’s entire lifecycle. The environment provides a realistic testbed for long‑term, memory‑aware legal AI and suggests that future progress should focus on improving courtroom interaction and leveraging full‑process trajectories as training signals.