FLAME: Improving Legal Case Retrieval through Factor-aware Graph Modeling and Mixture-of-Experts

Abstract

Legal Case Retrieval (LCR) is a foundational task in intelligent legal systems that assists legal professionals in efficiently retrieving precedent cases from large-scale legal corpora. Unlike generic document retrieval, LCR demands deeper semantic modeling due to the complexity and domain-specific logic embedded in legal texts, where similarity depends crucially on the alignment of legal factors such as legal facts, dispute focus, and reasoning. While recent large language models (LLMs) have advanced legal document understanding, existing approaches generally treat legal documents as monolithic texts and overlook the rich internal structure of legal components and their complex juridical reasoning patterns. Moreover, they lack the flexibility to dynamically adjust the importance of legal factors based on specific case contexts. To address these challenges, we propose Factor-aware Legal cAse Matching with Expert Networks (FLAME), a novel framework that centers on legal factors as fundamental modeling units and dynamically identifies the most relevant ones between case pairs. FLAME comprises three core modules: (1) the Legal-Factor Extraction Module employs Multi-Agent Reinforcement Learning for instance-wise prompt selection and leverages LLMs to extract structured legal factors from raw case documents; (2) the Legal-Factor Interaction Module constructs a heterogeneous legal-factor graph using a Factor Graph Convolutional Network to model semantic relationships among factors; (3) the Legal-Factor Integration Module employs a Legal-Factor Aware Mixture-of-Experts mechanism to dynamically highlight salient legal factors through parallel expert assessment and gating-based weight determination. The framework is optimized through a principled two-stage process that first uses contrastive learning to train the core retrieval architecture, and then employs reinforcement learning for adaptive prompt refinement. Our approach enables interpretable and context-aware legal case retrieval by explicitly modeling factor-level semantics and interactions. Extensive experiments on two benchmark datasets demonstrate that FLAME significantly outperforms state-of-the-art baselines.

Keywords

Legal Case Retrieval / Reinforcement Learning / Graph Neural Networks / Mixture-of-Experts / Prompt Optimization

RIGHTS & PERMISSIONS

Higher Education Press 2026

Just Accepted

This article has successfully passed peer review and final editorial review, and will soon enter typesetting, proofreading and other publishing processes. The currently displayed version is the accepted final manuscript. The officially published version will be updated with format, DOI and citation information upon launch. We recommend that you pay attention to subsequent journal notifications and preferentially cite the officially published version. Thank you for your support and cooperation.