Autism as LLM-Style Cognitive Architecture - A Computational Framework for Understanding Neurodivergent Learning and Cognition

Current models of autism emphasise social deficits and restricted or repetitive behaviours, framing autistic cognition primarily through what it lacks compared to neurotypical processing.We propose a radical reframing: that a subset of autistic cognition may represent a fundamentally different but computationally valid learning architecture that shares core structural principles with Large Language Models (LLMs). Rather than viewing autistic traits as deficits or compensatory mechanisms, we argue they emerge naturally from an architecture optimised for statistical pattern recognition, threshold-based learning, and regenerative inference. This framework explains several previously puzzling features of autistic cognition: extended observation periods followed by sudden competence rather than gradual skill acquisition, exceptional pattern recognition coupled with difficulty generalising from limited examples, preference for explicit rules over implicit social learning, all-or-nothing learning curves, cognitive exhaustion from continuous environmental processing, and reported comfort with AI interaction overhuman socialisation. We demonstrate that capabilities often attributed to special interests or savant skills, particularly systems thinking and narrative fluency, may be emergent properties of this architectural difference rather than compensatory developments. The model generates testable predictions about learning trajectories, intervention efficacy, and optimal educationalapproaches, and suggests that many current autism interventions fail because they attempt to impose neurotypical incremental-learning paradigms onto threshold-based cognitive architectures.