This study aims to develop a system that utilizes large language models (LLMs) to generate specific instructional advice optimized for each teacher's unique classroom characteristics.
Based on actual classroom dialogue data, we trained an LLM to replicate a “teacher dialogue model.” To enhance the accuracy of this replication, we iteratively optimized a meta-prompt defining the LLM's behavior, generating a “replication prompt” that reflects each teacher's dialogue tendencies. By adopting the analysis of differences between multiple teachers' prompts as a methodology, we confirmed that these differences clearly suggest individual teaching styles and challenges.
This approach enables the generation of individually optimized advice that goes beyond conventional, generic feedback. Through reproduced dialogues leveraging expert insights, it provides an effective means to support teachers in efficiently improving their instructional capabilities.
