LLM Agent Collaboration for Educational Strategy Design

Abstract

Large language models (LLMs) can be organized into multi-agent systems (MAS) that simulate expert panels, but coordination remains a challenge. In education, effective support often requires integrating insights from data analytics, curriculum design, and pedagogy. We present a moderated, embedding-driven MAS framework for educational decision-making. Role-specialized LLM agents generate structured recommendations, which are embedded into a shared semantic space. A central moderator measures alignment and agreement, computes a consensus embedding, and translates it back into natural language. This consensus is returned to the agents in subsequent rounds, guiding convergence while preserving distinct perspectives. A simulated case study illustrates the process in the context of secondary education. The results highlight the potential of embedding-based moderation to produce coherent, interpretable recommendations.