Source-linked AI summary
Multimodal Multi-Agent Empowered Legal Judgment Prediction
Zhaolu Kang, Junhao Gong, Qingxi Chen, Hao Zhang, Jiaxin Liu, Rong Fu, Zhiyuan Feng, Yuan Wang, Simon Fong, Kaiyue Zhou
TL;DR
Legal Judgment Prediction remains challenged by multiple charges, multimodal evidence, and complex trial interactions. The paper introduces JurisMMA, a structured multi-agent courtroom framework, and JurisMM, a large recent judicial dataset. Across key LJP tasks, experiments and ablations report that JurisMMA outperforms existing state-of-the-art methods.
Problem
Legal Judgment Prediction systems struggle with multiple charges or laws, multimodal evidence, and the procedural and interactive complexity of real-world trials.
Method
JurisMMA decomposes trial tasks into structured stages using a six-agent courtroom module, legal knowledge retrieval, and appeal procedures.
Results
JurisMMA outperforms existing state-of-the-art methods across Law Articles, Charges, and Terms of Penalty prediction.
Takeaways & Limitations
JurisMMA and JurisMM provide a multi-agent framework and large, recent text and video-text dataset for legal judgment prediction.
Abstract
from arXiv · showhide
Legal Judgment Prediction (LJP) aims to predict the outcomes of legal cases based on factual descriptions, serving as a fundamental task to advance the development of legal systems. Traditional methods often rely on statistical analyses or role-based simulations but face challenges with multiple allegations, diverse evidence, and lack adaptability. In this paper, we introduce JurisMMA, a novel framework for LJP that effectively decomposes trial tasks, standardizes processes, and organizes them into distinct stages. Furthermore, we build JurisMM, a large dataset with over 100,000 recent Chinese judicial records, including both text and multimodal video-text data, enabling comprehensive evaluation. Experiments on JurisMM and the benchmark LawBench validate our framework's effectiveness. These results indicate that our framework is effective not only for LJP but also for a broader range of legal applications, offering new perspectives for the development of future legal methods and datasets.
1. INTRODUCTION
Legal Judgment Prediction must handle specialized legal language, multiple charges and laws, multimodal evidence, and complex procedural interactions. JurisMMA addresses these challenges with a structured multi-agent courtroom framework and JurisMM, a large recent judicial dataset.
- Challenges: Specialized legal language and rising case volumes create significant challenges for Legal Judgment Prediction systems.Existing approaches also struggle to model the procedural and interactive aspects of real-world trials.
- Challenges: Existing frameworks often fail to predict multiple charges or laws and to use multimodal evidence such as trial recordings.These limitations can hinder outcome accuracy.
- Challenges: Simplistic role-based simulations overlook shared agent functions and procedural design, producing one-sided conclusions and illogical trial outcomes.
- Proposed framework: JurisMMA simulates court proceedings through a six-agent courtroom module and a knowledge retrieval module providing recent rulings and legal statutes.The courtroom module includes Junior Judges, Senior Judges, Assistants, the Chief Judge, Legal Counsel, and Defendants, with appeal handling for dissatisfied Defendants.
- Contributions: JurisMM combines single-text and multimodal video-text judicial data, comprising more than 100,000 recent cases.Experiments and ablations report that JurisMMA outperforms existing state-of-the-art methods across Law Articles, Charges, and Terms of Penalty prediction.
2. METHOD
JurisMMA organizes adjudication into structured stages that combine judicial agents, legal retrieval, first-instance decision-making, defendant objections, and appeal review. The workflow integrates case facts with legal knowledge and can revise the judgment through a second-instance process.
- JurisMMA assigns six agent roles to distinct stages that collaborate across the adjudication cycle.The roles are Junior Judge, Senior Judge, Chief Judge, Assistant, Defendant, and Legal Counsel.
- Stage I: Information Extraction: The Junior Judge extracts crime details and criminal motive, while the Senior Judge checks consistency with the full case text and produces a standardized case summary.
- Stage II: Legal Information Retrieval: The Assistant retrieves statutes, charge definitions, and precedents from multiple legal knowledge bases using dense, hybrid, and top-k retrieval procedures.The retrieved results are consolidated into a unified Case Information Document.
- Stage III: First-instance Decision: The Chief Judge combines extracted facts, the case summary, and retrieved legal information to select charges and issue the first-instance judgment.Charges are aligned with case facts and precedent information, while sentencing adjusts the statutory base sentence using precedent influence and mitigating factors.
- Stages IV–VI: Response and Appeal: Defendant objections must diverge from the judgment and be supported by the original case text; valid objections form the basis for an appeal brief integrating supporting laws and contextual reasoning.The appeal brief triggers the second-instance process, where the Chief Judge re-executes earlier stages and renders the final judgment.
3. EXPERIMENTS AND RESULTS
The experiments evaluate JurisMMA on large text and multimodal legal datasets, using diverse baselines, LawBench, and ablations. Results show strong performance across legal judgment tasks and benefits from multimodal evidence and the framework’s key components.
- Dataset and Resources: JurisMM includes 101,544 text cases, 83 multimodal video-text samples, and a domain-specific repository of legal articles, charges, and precedents.The text subset contains 94,494 training cases and 7,050 test cases; the repository includes 438 articles, 483 charges, and 46,365 representative cases.
- Experimental Settings: JurisMMA is compared with neural-network and large-language-model baselines, including GPT-4o and Qwen2.5-VL models.Neural baselines include TextCNN, TOPJUDGE, and MPBFN; evaluated LLMs include GLM-4V-9B, mPLUG-Owl-7B, Qwen variants, and GPT-4o.
- Overall Performance: JurisMMA consistently outperforms all baselines across Law Articles, Charges, and Terms of Penalty prediction on JurisMM-Text.Law Articles prediction is easiest, Charges prediction follows, and Terms of Penalty prediction is the most challenging.
- Overall Performance: JurisMMA outperforms five general-purpose LLMs on all selected LawBench tasks, supporting the adaptability of its legal-tailored multi-agent framework.LawBench evaluates accuracy, macro-precision, macro-recall, macro-F1, and, for sentence-length prediction, Normalized Log-Distance.
- Multimodal Evaluation: Visual information consistently improves accuracy and macro-F1 across tasks on 83 JurisMM-Video cases.The visual information enriches contextual understanding in multimodal legal judgment prediction.
- Ablation Study: Removing either the knowledge base or multi-agent collaboration causes significant performance drops across evaluated backbone models.The ablation results identify both components as important regardless of whether GPT-4o or Qwen2.5-VL-7B is used.
4. CONCLUSION
JurisMMA addresses legal judgment prediction complexity by structuring trial reasoning into stages and simulating courtroom procedures with coordinated agents. JurisMM supports this approach with over 100,000 text and video-text judicial cases.
- Framework: JurisMMA decomposes legal judgment prediction into structured stages and simulates realistic courtroom procedures.The framework uses coordinated LLM-based agents to address multivariable interactions and procedural nuances in real-world legal cases.
- Dataset: JurisMM contains over 100,000 judicial cases spanning unimodal textual data and multimodal video-text samples.The dataset is described as large-scale and up to date.
- Conclusion: Together, JurisMMA and JurisMM support legal judgment prediction involving complex procedural and multimodal case information.This conclusion follows the paper’s stated pairing of a structured multi-agent framework with a large multimodal dataset.