CooT: Learning to Coordinate In-Context with Coordination Transformers

In three linesCooT is a multi-agent framework using in-context learning for real-time adaptation to unfamiliar partners. Evaluated on Overcooked and Google Research Football, it outperforms population-based methods, gradient-based fine-tuning, and Meta-RL baselines without parameter updates.

Summary generated by Claude — human-verified