MathAssist:A Memory-Efficient On-DeviceAI Mathematics Tutor for Low-Resource Environments

Mathematics is at the heart of Science, Engineering and Technology (SET) education [1]. Unfortunately, many learners experience difficulties in comprehending mathematical ideas because of a lack of personalised assistance, inadequate learning
resources and difficulties in grasping problem-solving steps. This problem becomes even more acute in developing
countries like Bangladesh, where about 16.4 per cent of learners encounter difficulties in learning mathematics and 53.1 per cent of them could not attain satisfactory academic
performance.
Thanks to recent progress in Artificial Intelligence (AI), modern intelligent tutoring systems allow adaptive learning and automate problem-solving [5]. Nonetheless, the existing solutions, including popular applications like Photomath and Khan
Academy, are based primarily on cloud computing and a continuous Internet connection [6]. This restricts the use of these technologies in the rural regions of Bangladesh, which have poor access to the Internet and other digital infrastructure.
To overcome the limitations mentioned above, the current research aims to propose MathAssist, an offline mobile AI-assisted learning framework combining Retrieval-Augmented Generation (RAG) and lightweight Large Language Models (LLMs) to provide context-aware and step-by-step mathematical assistance with educational data retrieved from a
local knowledge base. The developed memory-efficient Serial RAG allows deploying the proposed system on resource-limited mobile devices and improves the reliability of mathematical
reasoning.
Major contributions of the current work are listed below:
• Development of an offline AI-assisted mathematics
learning framework with multimodal interactions and
lightweight execution allowing running on mobile devices with less than 2 GB RAM.
• Proposal of the novel Serial RAG approach incorporating semantic retrieval and local knowledge indexing along with LLM reasoning, where benchmarking experiments reveal WizardMath as the best mathematical reasoning model among all LLMs considered in the research.