Design and Implementation of a Mathematics Learning Support App Using Local LLMs

Source abstract

Students increasingly use artificial intelligence (AI) to support their learning. However, differences in the models and features available under free and paid plans may create disparities in learning opportunities when students cannot afford subscriptions. We developed a mathematics learning support app to provide an AI-assisted learning environment for students who cannot consistently pay for AI services. The app runs a large language model (LLM) on the user's device, avoiding per-question charges from external inference services and allowing us to offer the app free of charge. AI Sensei: High School Mathematics accepts questions as text or photographs and explains solutions in Japanese with reference to teaching materials. We separate question interpretation from explanation generation and supply the information needed at each stage. In a development comparison on an Apple Silicon Mac, removing instructions that were unnecessary for question interpretation reduced the wait before an answer appeared. When implementing the app on iOS, we identified differences in tokenhistory initialization and image input by comparing model invocation and execution conditions with the Mac implementation.