Abstract
Autism spectrum disorder involves differences in social communication, interaction and behavioral flexibility, yet conventional single-response Theory-of-Mind tasks rarely capture the trial-by-trial dynamics of social adaptation. We examined how autistic and non-autistic young people organized recent choice history, how they adjusted it across opponents, and how these patterns related to task performance. Twenty-seven autistic children and adolescents and 24 healthy controls completed the same two-choice hide-and-seek task under social and non-social framing against artificial opponents of graded sophistication. We compared aggregate performance, Bayesian model-based estimates of latent strategy use, and model-free history kernels measuring how current choices depended on recent self-actions and opponent actions. History kernels showed that autistic participants switched away from their previous choice more strongly against every opponent, whereas controls added lag-2 and opponent-coupled structure when the opponent adapted to them and reconfigured their kernels across opponents more. Within the same dataset, replaying participants’ kernels against empirically fitted opponents reconstructed accuracy (r = .86) and the direction of the group difference against the first-order opponent, against which switching provided less benefit than against the zero-order opponent. Against the first-order opponent, controls were more often described by first-order opponent modelling (influence learning) and autistic participants by heuristic policies. These findings suggest that autism-related differences in repeated interaction are expressed in how choice history is organized and reconfigured across opponents, beyond categorical differences in inferred strategies.