Inference of Latent Task Models in Human–AI Collaboration Under Uncertainty

Abstract

Mixed human-AI teams often operate with limited explicit communication under task uncertainty, so effective coordination requires quickly identifying which latent task or environment model best explains the observed interaction. This paper develops implicit task-hypothesis inference, enabling an AI agent to maintain a posterior over a finite set of candidate task models using its own experience and passively observed human state trajectories. We model collaboration as a task-uncertain cooperative Markov decision process (MDP) and derive a fully recursive Bayesian update that combines the agent’s experience with human data. Unobserved human actions are modeled and integrated using a bounded-rational, cooperative policy model. The resulting posterior can be used for a wide range of downstream tasks, including online decision rules that either act under the most likely hypothesis (maximum a posteriori) or account for uncertainty by weighting decisions across hypotheses according to the posterior. We analyze when human trajectories are informative by studying their sensitivity to human decision reliability and identifying conditions under which the human likelihood provides limited additional discrimination among hypotheses. Numerical results on grid-world rescue in maze environments with hidden environmental structures demonstrate faster hypothesis identification and higher cooperative returns than competing inference methods.