Engineering Multi-Agent Large Language Model Systems for Mobile Adaptive Learning Environments

Source raw items (1)

• Semantic Scholar2026-07-23 07:15:18
  Engineering Multi-Agent Large Language Model Systems for Mobile Adaptive Learning Environments

  Introduction: Large Language Models (LLMs) have catalyzed rapid adoption in educational technology, with 37% of secondary teachers globally using generative AI tools. However, deployment predominantly occurs through monolithic architectures where single models handle diverse functions without specialization for adaptive learning constraints. This work presents the design and implementation of a mobile adaptive learning platform for primary mathematics education, featuring a specialized multi-agent LLM architecture. Method: The proposed system implements four functionally distinct agents—Pattern Analyzer for cognitive profiling, Curriculum Architect for personalized sequencing, Exercise Selector for dynamic content generation, and Feedback Provider for constructivist support—coordinated by a central orchestrator following the Mediator design pattern. Each agent is independently optimized with specialized prompts and intelligent selection between GPT-4o and GPT-4o-mini, incorporating hybrid routing between template responses and dynamic LLM generation for feedback.  Results: Development testing over three months consumed approximately 1 million tokens across 192 API requests, incurring $3.72 total cost and projecting per-student costs of $0.25-0.30, with the hybrid feedback strategy achieving approximately 85% cost reduction. A pilot test with a single student demonstrated technical feasibility through successful cognitive profile generation, personalized curriculum sequencing, and system integration. Conclusions: The multi-agent approach shows that specialized architectures enable targeted optimization unavailable in monolithic systems, establishing an engineering framework for adaptive learning platforms, while pedagogical effectiveness still requires future experimental validation with learners in controlled studies.