Legal Chain Resolver With Mixture of Experts and Multi-Agent System for Legal Assistance

Team

Supervisors

Table of content

  1. Abstract
  2. Related works
  3. Methodology
  4. Experiment Setup and Implementation
  5. Results and Analysis
  6. Conclusion
  7. Publications
  8. Links

Abstract

The Sri Lankan legal system faces significant barriers in accessibility, affordability, and efficiency, particularly within commercial law. High legal costs, language barriers, and limited digital legal resources make timely and accurate legal assistance challenging for many, especially SMEs. Existing AI-driven legal solutions in Sri Lanka predominantly rely on Natural Language Processing (NLP), which lacks structured legal reasoning and is prone to hallucinations. This project proposes an AI-driven legal assistance framework focused on Sri Lankan commercial law, integrating a Mixture-of-Experts (MoE) model, a multi-agent system, and knowledge graphs. The system is designed to enhance legal accuracy, interpretability, and efficiency by dynamically routing legal queries to domain-specific AI experts, structuring legal knowledge, and simulating real-world legal workflows. This approach aims to provide accessible, reliable, and context-aware legal support for businesses, legal professionals, and policymakers in Sri Lanka, ultimately improving legal decision-making and compliance

AI-driven legal assistance is an evolving field, with global advancements seen in models such as ChatLaw, LawGPT, LawNeo, and Sri Lanka’s own LawKey. Early legal AI relied on rule-based and statistical NLP systems, which struggled with legal language complexity. Modern solutions leverage deep learning and LLMs for tasks like legal text summarization and contract analysis but still face issues like hallucinations and lack of structured reasoning. ChatLaw stands out by integrating a Mixture-of-Experts (MoE) architecture, multi-agent collaboration, and knowledge graphs, significantly reducing hallucinations and improving legal reasoning. LawNeo uses composite model integration for cost-efficient domain adaptation, while LawKey focuses on accessibility within the Sri Lankan context using reinforcement learning and NLP. However, most solutions are limited by jurisdictional scope, lack of multilingual support, and insufficient explainability. These gaps motivate the development of a domain-specific, interpretable AI legal assistant tailored for Sri Lanka’s commercial law sector

Methodology

The proposed system’s methodology is structured as follows:

Experiment Setup and Implementation

The implementation plan includes:

Results and Analysis

As of the current stage, full-scale evaluation and testing are pending. The system is designed to be assessed via benchmarking against existing legal AI assistants, focusing on improvements in accuracy, reliability, and contextual legal reasoning. Planned evaluation phases include comparative benchmarking, user-based testing with legal professionals, legal validation against real-world case law, and scalability assessments. The expected outcomes are enhanced accuracy, reduced misinformation, and improved usability compared to traditional NLP-based legal assistants

Conclusion

This project addresses critical gaps in Sri Lankan legal assistance by developing an AI-powered framework that integrates Mixture-of-Experts, multi-agent systems, and knowledge graphs. The proposed solution aims to deliver reliable, interpretable, and context-aware legal support, specifically for commercial law. By leveraging structured legal knowledge and domain-specific expertise, the system seeks to improve accessibility, minimize misinformation, and support informed decision-making for businesses and legal professionals in Sri Lanka. Future work will focus on optimizing expert selection, enhancing real-time knowledge updates, and ensuring ethical compliance and data security

Publications