📄 中文摘要
多智能体大型语言模型（LLM）和视觉语言模型（VLM）辩论系统在解决复杂问题时，通常会为每个智能体分配专门的角色。然而，现有系统并未充分利用模型的专业化能力来决定哪个模型应该扮演哪个角色。针对这一问题，提出了一种动态角色分配框架，该框架在实际辩论开始之前运行一个“元辩论”来选择合适的智能体。元辩论分为两个阶段：首先是提案阶段，在此阶段，候选智能体根据其潜在角色提供量身定制的论点。这些论点旨在展示智能体在该特定角色下的能力和专业知识。例如，如果一个角色是“事实核查员”，候选智能体可能会提供其在检索和验证信息方面的成功案例或方法。随后是评估阶段，一个独立的裁判或评估机制会根据这些提案对候选智能体进行评估，并最终决定哪个智能体最适合哪个角色。这种动态分配机制旨在优化团队组成，确保每个角色都由最擅长该任务的智能体担任，从而提升整个辩论系统的性能和问题解决效率。通过这种方式，系统能够更智能地利用不同模型的独特优势，避免了预设或随机角色分配可能带来的次优结果。该框架的引入，使得多智能体系统在面对复杂任务时，能够更灵活、高效地组织其内部资源，以期达成更优的协同效果。
📄 English Summary
Dynamic Role Assignment for Multi-Agent Debate
Multi-agent large language model (LLM) and vision-language model (VLM) debate systems commonly assign specialized roles for complex problem-solving. However, existing approaches do not leverage model specializations to determine which model should fill which role. Addressing this limitation, we propose a dynamic role assignment framework that conducts a "Meta-Debate" to select suitable agents prior to the actual debate. The meta-debate proceeds in two stages. The first stage is the proposal phase, where candidate agents present role-tailored arguments. These arguments are designed to showcase an agent's capabilities and expertise for a specific role. For instance, if a role is designated as a "fact-checker," a candidate agent might provide examples of its successful information retrieval and verification methods. Following this, the second stage is the evaluation phase, where an independent judge or evaluation mechanism assesses the candidate agents based on their proposals, ultimately deciding which agent is best suited for each role. This dynamic assignment mechanism aims to optimize team composition, ensuring that each role is filled by the agent most proficient in that task, thereby enhancing the overall performance and problem-solving efficiency of the debate system. By intelligently utilizing the unique strengths of different models, this approach circumvents suboptimal outcomes that might arise from static or random role assignments. The introduction of this framework enables multi-agent systems to organize their internal resources more flexibly and efficiently when tackling complex tasks, striving for superior collaborative effects.