LegalMALR:Multi-Agent Query Understanding and LLM-Based Reranking for Chinese Statute Retrieval

TL;DR

LegalMALR addresses the challenge of retrieving applicable Chinese statutes from implicit, multi-issue queries by combining a Multi-Agent Query Understanding System (MAS) with a zero-shot LLM-based reranker. The MAS generates diverse reformulations to explore multiple doctrinal interpretations and uses GRPO to stabilize and optimize retrieval trajectories, followed by a final ranking by a large language model (Qwen-Max) that performs substantive legal reasoning. The framework is evaluated on STARD and CSAID, where it substantially outperforms strong RAG baselines in both in-domain and cross-domain settings, demonstrating improved recall and ranking through multi-perspective query interpretation and principled large-model reranking. The CSAID dataset and the GRPO-enabled MAS contribute new resources and techniques for robust statute retrieval, with practical implications for legal decision support and administrative compliance in Chinese law. Limitations include computation cost, reliance on embedding-based retrieval, and domain specificity, suggesting future work on distillation, faster one-step approximations, and multilingual/extensions to broader jurisdictions.

Abstract

Statute retrieval is essential for legal assistance and judicial decision support, yet real-world legal queries are often implicit, multi-issue, and expressed in colloquial or underspecified forms. These characteristics make it difficult for conventional retrieval-augmented generation pipelines to recover the statutory elements required for accurate retrieval. Dense retrievers focus primarily on the literal surface form of the query, whereas lightweight rerankers lack the legal-reasoning capacity needed to assess statutory applicability. We present LegalMALR, a retrieval framework that integrates a Multi-Agent Query Understanding System (MAS) with a zero-shot large-language-model-based reranking module (LLM Reranker). MAS generates diverse, legally grounded reformulations and conducts iterative dense retrieval to broaden candidate coverage. To stabilise the stochastic behaviour of LLM-generated rewrites, we optimise a unified MAS policy using Generalized Reinforcement Policy Optimization(GRPO). The accumulated candidate set is subsequently evaluated by the LLM Reranker, which performs natural-language legal reasoning to produce the final ranking. We further construct CSAID, a dataset of 118 difficult Chinese legal queries annotated with multiple statutory labels, and evaluate LegalMALR on both CSAID and the public STARD benchmark. Experiments show that LegalMALR substantially outperforms strong Retrieval-augmented generation(RAG) baselines in both in-distribution and out-of-distribution settings, demonstrating the effectiveness of combining multi-perspective query interpretation, reinforcement-based policy optimisation, and large-model reranking for statute retrieval.

Figures (5)

• Figure 1: Overview of LegalMALR. MAS expands retrieval coverage via diverse legal-perspective reformulations and iterative search, while the LLM eranker conducts statutory reasoning to produce accurate final rankings for implicit and colloquial legal queries.
• Figure 2: Overall architecture of LegalMALR. The upper row illustrates the conventional dense-retrieval pipeline in which a single embedding retriever issues one-shot retrieval over the user query. The middle row presents the proposed Multi-Agent Query Understanding System (MAS), which replaces this single retrieval decision with multi-perspective, multi-round reformulation and retrieval. The lower row depicts the complete LegalMALR framework, where MAS is further stabilised through GRPO and the merged candidate set is reranked by a large language model.
• Figure 3: End-to-end processing flow of LegalMALR. MAS iteratively analyses and reformulates the query, retrieves candidate statutes, and monitors coverage to decide termination. The final merged candidate pool is evaluated by a zero-shot LLM-based reranker to produce the ranked statute list.
• Figure 4: Ablation of iterative search depth in MAS. The One-step configuration executes only a single rewrite and retrieves a fixed top--10 list, whereas multi-step MAS and MAS+GRPO operate with dynamic candidate pool sizes (approximately 14 on average). Multi-step search substantially improves coverage, and GRPO further enhances retrieval effectiveness.
• Figure 5: Combined case study of MAS+GRPO behaviour. The left example illustrates a multi step retrieval trajectory in which the second reformulation exposes complementary legal elements and achieves full statutory coverage. The right example illustrates early stopping behaviour where the initial reformulation already captures the necessary statutory elements and the system terminates the iterative process. Recall values appear only as diagnostic indicators.