Reproduce “CooT: Learning to Coordinate In-Context with Coordination Transformers” on OpenResearch

About this paper

Huai-Chih Wang, Hsiang-Chun Chuang, Hsi-Chun Cheng, Dai-Jie Wu, Shao-Hua Sun

CooT introduces a transformer-based framework that applies in-context learning to enable agents to adapt their coordination strategies to unseen partners in multi-agent settings. The system achieved superior performance over baselines in environments like Overcooked and Google Research Football, demonstrating rapid and robust adaptation to novel and non-stationary partner behaviors.

Key insights

• In-context learning, traditionally applied to task generalization, can be effectively adapted for dynamic partner generalization in multi-agent coordination scenarios.
• Conditioning on full, raw interaction histories through a transformer is more effective for robust partner inference and adaptation than relying on compressed latent representations.
• Efficient, in-context adaptive behavior by AI agents is not only quantitatively superior but also qualitatively perceived by human partners as more collaborative and adaptive.

Summary by alphaXiv.