Teaching Spatial Prepositions to Children With Autism Using Three‐Dimensional Matrix Training: A Demonstration of LLM‐Assisted Material Preparation

Matrix training is an evidence‐based approach for promoting recombinative generalization in children with autism spectrum disorder (ASD). However, the time and effort required to generate instructional materials for multidimensional matrices can limit its use in practice. In this study, we evaluated the effects of a three‐dimensional matrix training protocolusing a minimal set of 10 exemplars. To address practical barriers to material design, a large language model (LLM) was utilized strictly as an instructional planning support tool to assist teachers with exemplar generation. Ten children with ASD participated in a study employing a concurrent multiple‐probe design across participants, with target exemplars introduced sequentially across progressive sentence sets. The 10 baseline exemplars, selected by teachers with computational support from the LLM, were directly taught using a structured prompting sequence. Following instruction, 9 participants demonstrated immediate, high levels of recombinative generalization to 80 untrained combinations within the training matrix and to 54 novel combinations involving untrained stimuli in the generalization matrix. Maintenance data showed sustained performance at 2‐ and 4‐wk follow‐ups. These findings demonstrate that three‐dimensional matrix training with a minimal exemplar set efficiently produces generative language outcomes, and highlight the feasibility of integrating LLMs to support behavioral programming in applied settings.