Project Description

Children on the autistic spectrum exhibit significant heterogeneity in communication, executive function, and learning styles, yet most educational systems and AI tools assume uniform learners. Consequently, they fail to align with individual cognitive profiles, undermining comprehension and engagement, while human-led personalized tutoring remains costly and non-scalable. Neurodevelopmental disorders (NDDs) affect an estimated 15-20% of the global student population, with autism spectrum disorder (ASD) affecting approximately 1 in 31 individuals in the UK, yet fewer than 2% of learners with ASD can access individualized tutoring. While generative AI, particularly large language models (LLMs), offers potential for scalable personalized support, over 80% of students currently use such tools in an unguided manner, despite limited evidence of effectiveness for neurodivergent populations, underscoring the need for adaptive, pedagogy-grounded solutions. This Autipilot Ph.D. project aims to explore the following core research questions: RQ1: What are the needs, constraints, expectations, gaps, and best practices across diverse stakeholders, and how do state-of-the-art GenAI models systematically benchmark against these criteria? RQ2: How to robustly quantify learner heterogeneity, grounded in RQ1, and integrate it into adaptive, pedagogy-grounded tutoring systems to address identified gaps? RQ3: How can this adaptive system be effectively integrated into human-in-the-loop settings to support tutors working with children with ASD while ensuring safety, reliability, and trust? The feasibility of this project is supported by prior work along similar lines, including Google's LearnLM initiative and Stanford's Tutor CoPilot. The expected outcomes include the development of public evaluation benchmarks and adaptive tutoring models. Additionally, the project aims to quantify heterogeneity through concordance with expert assessments; measure learning gains via pre-and post-assessments; and evaluate engagement through metrics such as session duration, completion rates, and dropout reduction. Finally, it seeks to deliver a human-centered tool that augments tutors' efforts, as evidenced by metrics such as the number of response edits, reduced time per session, and overall workload reduction.