Large Language Models (LLMs) have shown remarkable performance in Natural Language Processing (NLP) tasks, but they still face significant challenges in complex reasoning tasks due to their limitations in deeply analyzing implicit knowledge and logic. However, existing Multi-Agent System (MAS) may produce redundant or conflicting information during parallel reasoning, reducing response quality and masking key insights. To address these issues, we propose Role-of-Thought (RoT), an adaptive MAS that dynamically generates domain-specific expert agents to enhance reasoning capabilities. RoT divides complex tasks into multi-stage sub-tasks, assigns them to relevant expert agents, and integrates the reasoning processes into a coherent reasoning chain. By leveraging the strengths of multiple expert agents and multi-stage reasoning, RoT effectively extracts and applies implicit knowledge, improving the accuracy and adaptability of reasoning. Experimental results demonstrate that RoT significantly outperforms baseline reasoning methods, showing enhanced reasoning performance in various tasks.
Zeng et al. (2026) studied this question.