A Multi-Agent Consensus Framework for Explainable Automated Assessment of Language Ability in Children with Autism Spectrum Disorder

Language ability reflects the developmental level and social communication in children with autism spectrum disorder (ASD), making scalable and explainable assessment essential. Existing large language model (LLM)-based methods mainly rely on single-agent reasoning or aggregate independent outputs, offering limited insight into disagreement identification, evidence exchange, and consensus formation. To address this limitation, this paper proposes a Multi-Agent Consensus framework that formulates language assessment as a structured consensus process among LLM-based agents. Using parent–child free-play transcripts, three Expert Agents independently assess language comprehension, vocabulary competence, logical expression, and communication skills under shared criteria. A Judge Agent identifies dimension-level disagreements, marks dimensions with consistent judgments as resolved, coordinates evidence-based negotiation, and adjudicates remaining conflicts. On a real-world dataset of 79 children, the framework achieves a Pearson correlation of 0.843 with an external standardized language development measure, outperforming Single Agent (0.802) and Initial Expert Aggregation (0.814), while maintaining strong cross-run stability. Ablation and process analyses show that consensus negotiation integrates initial judgments, resolves disagreements, and improves assessment stability. Case analysis and human expert evaluations support the plausibility, evidential grounding, and transparency of the generated reports. These findings suggest that multi-agent consensus can enhance automated language assessment for children with ASD while providing structured, traceable explanations. The correlations support criterion validity, but clinical validity and effectiveness in practice require further evaluation given the limited sample size and data sources.