Depth Before Strategy: Temporal Choice-History Signatures Reveal Reduced Recursive Social Planning in Autism
Year: 2026
Type: preprint
Abstract: Autism spectrum disorder (ASD) is clinically defined through differences in social communication, interaction, and behavioral flexibility, yet the instruments used to detect it rarely capture the trial-by-trial dynamics through which social adaptation unfolds. Here we ask a deliberately methodological question: when autistic participants and healthy controls adapt to a repeated social interaction, does the diagnostically relevant signal live in the cognitive strategy they are inferred to use, or in the temporal structure of their choices? We tested this in a repeated hide-and-seek task in which autistic participants and healthy controls performed the same two-choice decision problem under social and non-social framing and against artificial opponents of graded sophistication. We compared three readings of the same behavior: aggregate performance summaries, Bayesian model-based estimates of latent strategy use, and model-free choice-sequence history kernels measuring how current choices depended on recent self-actions and opponent actions. Aggregate behavior showed context-specific, rather than global, ASD-HC differences. Bayesian model selection revealed condition-dependent shifts in inferred strategy distributions, especially in first-order opponent conditions, where healthy controls were more often assigned to adaptive opponent-modeling strategies and autistic participants to simpler heuristic policies. However, these strategy summaries carried limited individual-level group information. In contrast, choice-history kernels preserved a robust temporal signature. Autistic participants showed stronger immediate self-history dependence in simpler opponent conditions, whereas healthy controls showed stronger lag-2 self-history and opponent-history dependence when the opponent required adaptive tracking. Fixed lag 1:8 kernel features outperformed behavioral and model-based task-derived representations in ASD-HC classification, while clinical variables remained the strongest reference representation. These findings suggest that the clinically relevant variation in repeated social adaptation resides less in a categorical shift between latent strategies than in the temporal organization of choice history. Model-free history kernels thus provide an interpretable complement to generative modeling for computational phenotyping in ASD, situating the present findings within a broader computational framework linking recursive Theory-of-Mind reasoning and individual differences in the depth of social planning that have already proven useful as phenotyping markers in other populations.
Source: PsyArXiv (OSF Preprints)
Language: English
Subfield: Cognitive Neuroscience
Field: Neuroscience
Domain: Life Sciences
Open Access status: green
OpenAlex ID: W7168024287