Guiding solution insight with human-compatible virtual agents

Skill acquisition in social contexts depends on learners’ ability to attune to the behaviors of their partners, whose actions help structure environments that guide attention and shape the learning of adaptive behaviors. While greater coupling is often associated with learning and collaborative problem-solving, recent perspectives suggest that the dynamics of interpersonal coupling are more complex, arising from the interplay of two opposing forces – integration and individualization – which define social interaction as a metastable system. This study examined how these dynamics influence performance and solution insight in a two-person problem-solving game with a latent solution that, when discovered, enables superior performance. Using multidimensional cross-recurrence quantification analysis (MdCRQA), participants who discovered and exploited the latent solution exhibited greater perceptual-motor coupling during the pre-discovery period than non-discoverers. Coupling declined after discovery, with further reductions linked to successful task performance, consistent with the metastable view of social interaction. Finally, we demonstrate that AI agents embodying the decision policies of pre-discoverers can, when working with novice human participants, scaffold environments that increases the likelihood of participants discovering the latent solution. These findings are discussed in the context of designing human-compatible AI agents that facilitate the self-organization of adaptive behaviors in humans.