Unlocking the Power of Multi-Agent LLM for Reasoning: From Lazy Agents to Deliberation

Recent research highlights significant advancements in large language models (LLMs) applied to complex reasoning tasks within multi-agent frameworks. In these settings, a meta-thinking agent is responsible for proposing plans, while a reasoning agent executes these plans through interactive conversations. This division of roles facilitates collaborative problem-solving and has demonstrated promising performance outcomes. However, a notable challenge has been identified concerning lazy agent behavior, where agents may underperform or rely excessively on others, potentially hindering overall system effectiveness. The interaction mode between agents, characterized by conversational exchanges, is central to this collaborative approach. Addressing the lazy agent issue remains a key focus for improving multi-agent LLM systems. These findings build on ongoing developments in reinforcement learning and large language models, underscoring the evolving landscape of AI reasoning capabilities.