From Synthesis to Clinical Assistance: A Strategy-Aware Agent Framework for Autism Intervention based on Real Clinical Dataset

A team of researchers from China (including Junhong Lai, Shuzhong Lai, Yanhao Yu, and others) has developed ASDAgent, a novel AI framework designed to address two critical bottlenecks in AI-assisted autism therapy: the scarcity of high-quality clinical dialogue data and the inability of general-purpose large language models (LLMs) to consistently follow Applied Behavior Analysis (ABA) protocols. While ABA is the gold standard for Early Intensive Behavioral Intervention (EIBI) for Autism Spectrum Disorder (ASD), standard LLMs often produce fluent but strategically flawed interactions. ASDAgent tackles this by splitting the problem into two specialized components: DoctorAgent uses an Observe-Think-Act-Correct (O-T-A-C) reasoning loop to make ABA execution explicit and controllable, preventing "strategy collapse"; ChildAgent employs probabilistic behavior modeling to simulate diverse and non-deterministic ASD response patterns, mitigating the data homogeneity that plagues synthetic datasets.

Experimental results show the framework's effectiveness: dialogues generated by ASDAgent closely match the strategy distribution of human therapists, with a KL divergence of just 0.083. In real clinical settings, ASDAgent achieves nearly 80% strategic consistency with human experts. Moreover, the synthetic data produced by ASDAgent successfully distills professional clinical knowledge into smaller language models (SLMs), significantly enhancing their therapeutic capabilities—meaning the framework can serve as both a data generator and a direct clinical assistant. The paper has been accepted to the 2026 ACL Main Conference, highlighting its significance in both machine learning and clinical AI applications.