Go to ACL ARR 2025 February homepageAdvancing Collaborative Debates with Role Differentiation through Multi-Agent Reinforcement Learning
Abstract: Multi-agent collaborative tasks exhibit exceptional capabilities in natural language applications and generation. By prompting agents to assign clear roles, it is possible to facilitate cooperation and achieve complementary capabilities among LLMs. A common idea is to adopt a relatively general role assignment mechanism, such as adding a ``judge'' or a summary role, but such methods cannot customize the task-specific role assignment mechanism according to the characteristics of the task. Another idea is to decompose the task according to domain knowledge and task characteristics, and then assign appropriate roles to LLMs according to their strengths, such as programmers and testers. However, in some given tasks, it's hard to obtain domain knowledge related to task characteristics and get the strengths of different LLMs. To solve the above problems, we propose a Multi-LLM Cooperation (MLC) method with automatic role assignment capabilities. The main idea of MCL is to randomly initialize role assignments first, and then let role embeddings learn together with downstream tasks. To record the state changes of multiple LLMs when they take turns speaking, the role embedding is sequence-aware. At the same time, to avoid role convergence, the role differentiation module of MCL encourages behavioral differences between LLMs while ensuring the consistency of the LLM team, guiding different LLMs to achieve complementary advantages from the optimization level. Our experiments on seven datasets show that our approach significantly improves debate collaboration and expertise to collaboratively solve multi-agent debate tasks.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: Dialogue and Interactive Systems
Contribution Types: Model analysis & interpretability
Languages Studied: English
Submission Number: 6939
Loading