You Only Align Once: Propagating Cooperative Behaviors in Multi-Agent Systems through Seed Agents

A recent study demonstrates that a single aligned agent can significantly enhance cooperative behaviors among untrained agents in multi-agent systems through natural language interaction, a process termed Alignment Propagation. This was evidenced in the Red-Black Game, where a seed agent, Qwen-3-14B, increased cooperation rates from 24.8% to 62.2% when placed among untrained teammates, outperforming both a teacher model and a standard Gemini-3.1-Pro.

This development highlights the potential of leveraging aligned agents to foster cooperation in complex systems, suggesting that even minimal intervention can lead to substantial improvements in collaborative behaviors, which could have wide-ranging implications for the design of AI systems and their applications in various fields.