A Deep Neural Network-Based Multi-Agent Mixture-of-Experts Framework for Generating AI Educational Content in Malagasy Language for Children (5–10)

This paper introduces a novel multi-agent framework for generating AI educational content in Malagasy, tailored for children aged 5 to 10. The system combines transformer-based large language models (LLMs) with deep neural networks (DNNs) and a Mixture-of-Experts (MoE) mechanism to dynamically adapt lesson structure and complexity based on the learner’s age. Our architecture integrates (1) an age-sensitive embedding model, (2) a content generator that selects specialized expert agents, (3) a retrieval-augmented generation (RAG) module, (4) a neural translator for Malagasy, and (5) an adaptive LoRA-enhanced DNN for linguistic correction with human-in-the-loop feedback. We also introduce a chain-of-thought (CoT) trace to ensure full explainability across agents. Unlike monolithic LLMs such as GPT-4o, GPT-4.1, GPT-4.5, or Gemini 2.x, our modular design enables localized language refinement, cultural adaptation, and progressive fine-tuning. Experiments demonstrate promising educational coherence and linguistic quality compared to five top-performing multilingual LLMs in low-resource language generation. This work presents a concrete step toward inclusive, AI-assisted pedagogy for underrepresented languages and learners in resource-constrained settings.