LawLuo: A Multi-Agent Collaborative Framework for Multi-Round Chinese Legal Consultation

TL;DR

LawLuo tackles the limitations of single-agent Chinese legal consultations by introducing a four-agent framework (receptionist, lawyer, secretary, boss) that emulates real law-firm workflows for multi-turn dialogue. It leverages three domain-fine-tuned datasets (receptionist Inquire–Lawyer Description, MURLED, LCRG) and a case-graph–based RAG to address ambiguity, with LoRA fine-tuning and PPO reinforcement learning guiding inter-agent optimization. Empirical results show LawLuo outperforms baselines in personalization, professionalism, ambiguity handling, and instruction-following over multiple rounds, while ablations confirm the importance of each component and the learned legal knowledge. The work demonstrates the practicality and impact of multi-agent collaboration for Chinese legal consultations and provides an open-source platform for further development in legal AI services.

Abstract

Legal Large Language Models (LLMs) have shown promise in providing legal consultations to non-experts. However, most existing Chinese legal consultation models are based on single-agent systems, which differ from real-world legal consultations, where multiple professionals collaborate to offer more tailored responses. To better simulate real consultations, we propose LawLuo, a multi-agent framework for multi-turn Chinese legal consultations. LawLuo includes four agents: the receptionist agent, which assesses user intent and selects a lawyer agent; the lawyer agent, which interacts with the user; the secretary agent, which organizes conversation records and generates consultation reports; and the boss agent, which evaluates the performance of the lawyer and secretary agents to ensure optimal results. These agents' interactions mimic the operations of real law firms. To train them to follow different legal instructions, we developed distinct fine-tuning datasets. We also introduce a case graph-based RAG to help the lawyer agent address vague user inputs. Experimental results show that LawLuo outperforms baselines in generating more personalized and professional responses, handling ambiguous queries, and following legal instructions in multi-turn conversations. Our full code and constructed datasets will be open-sourced upon paper acceptance.

Figures (10)

• Figure 1: The left side shows the single-agent architecture used by most legal consultation systems, producing superficial, generalized responses without understanding user intent and case details. The right side presents our proposed multi-agent framework, offering more personalized and professional answers.
• Figure 2: The multi-agent collaboration framework we propose for multi-round Chinese legal consultation. In this framework, the receptionist agent first assesses the user's consultation intent based on the initial input $u_0$ and selects the most suitable lawyer from the lawyer pool. Subsequently, the selected lawyer agent is responsible for engaging in multi-round dialogues with the user. During this process, the lawyer agent actively queries the user for case details via case graph-based RAG. Finally, the secretary agent organizes the dialogue records between the user and the lawyer, producing a comprehensive consultation report. The boss agent monitors the performance of the lawyer and secretary agents to ensure optimal outcomes.
• Figure 3: Distribution of the MURLED dataset across 16 different consultation domains.
• Figure 4: An example from the MURLED dataset. It can be seen that this dataset emphasizes the active guidance ability of training large legal models in multi-turn dialogues.
• Figure 5: Win rate of LawLuo compared to the baselines