Proceedings of the Annual Meeting of the Cognitive Science Society

In many pedagogical contexts, teachers guide learners through direct physical intervention—a modality we term state intervention. While common in human interaction, the computational mechanisms that allow agents to learn from being physically moved remain underexplored. We propose that the effectiveness of state intervention hinges on the interpretational alignment between teacher and learner: the learner must infer whether an intervention functions as a suggestion, a correction, or a non-pedagogical event. We investigated this problem using a 1D navigation task where human teachers (N=64) trained Q-learning agents under four different interpretation types. Our results show that agents learn most effectively when they interpret interventions as pedagogical signals—either as recommended actions (suggestion) or as discouraging actions (impede). In contrast, interpreting interventions as environment resets (reset) or mere interruptions (interrupt) led to significantly poorer performance and learning plateaus. Crucially, providing teachers with a visual representation of the agent's internal beliefs (Q-table) did not improve teaching effectiveness, suggesting that aligning the agent's learning rules with human intuition is more critical than information transparency. These findings highlight the importance of designing "socially-aware" learners that treat physical interaction as intentional communication.