s³: Multi-Agent RL Lets LLM Swarms Self-Organize

Researchers introduced s³, a multi-agent reinforcement learning approach that enables large language model swarms to self-organize without predefined leaders, roles, or turn orders. Operating under a shared reward where the team succeeds if any agent solves the task, the system successfully coordinates information aggregation and decentralized exploration. Testing showed that swarms scaled past their training size and outperformed individual rewards and baseline methods. Open article →