Researchers develop ConventionPlay, an AI training method that lets agents learn to lead or follow teammates depending on what works best

Researchers at an academic institution published ConventionPlay, a new reinforcement learning (a type of AI training where the system learns by trial and error) approach that teaches AI agents to work effectively with teammates of different skill levels. Unlike previous methods that only adapt to a partner's strategy, ConventionPlay agents actively figure out which teammate capability they're facing and choose whether to lead or follow to maximize the team's performance.

The method trains agents against partners with different capability limits—some can follow many strategies, others only a few. This teaches the agent to 'probe' or test what its partner can do, then steer the team toward the best shared strategy. In coordination tasks (scenarios where two agents must align their actions to succeed), ConventionPlay achieved higher efficiency than baseline methods, especially when different strategies had different payoff values.

This matters for any application where AI systems must collaborate with humans or other AI systems without pre-agreed rules: autonomous vehicle handoffs between drivers and autopilot, team robotics in factories, or multiplayer human-AI gameplay. Teams can now perform better even when teammates have unknown or mixed capabilities, rather than requiring everyone to learn the same fixed convention beforehand.