To work effectively with new human or AI partners, estimate their hidden capabilities from a few tasks, then use those estimates to plan collaborative actions—this works better than assuming partners are optimal or pre-training on populations.
This paper tackles ad-hoc teamwork—where an AI agent must collaborate with unknown partners on multiple tasks. The key innovation is inferring hidden partner capabilities (what actions they can reliably execute) without pre-training, then using these estimates to plan better joint actions. The method handles human unpredictability by considering multiple valid strategies, not just optimal ones.