AI-Powered Personalised Mobile Learning Using Large Language Models: Design, Implementation, and a Simulation-Based Feasibility Evaluation

Abstract

The significance of mobile learning (m-learning) has been greatly highlighted by the advent of technology due to its flexibility, accessibility and capability of being used to learn at any time and place. The development of Large Language Models (LLMs) offers great potential for creating adaptive, conversational and content-aware learning experiences via mobile apps. Most existing mobile-learning systems, however, still provide largely static content and shallow personalisation, which limits sustained learner engagement and adaptive support. This paper presents the design and implementation of an Android-based, LLM-powered personalised m-learning platform that integrates four components: a conversational tutoring agent, an Item-Response-Theory (IRT)-based adaptive quiz engine, a Retrieval-Augmented Generation (RAG) content-recommendation module, and an analytics-driven feedback loop. We describe the system architecture, the prompt-orchestration and knowledge-tracing pipelines, and a rigorous quasi-experimental evaluation protocol (pre-/post-test, System Usability Scale (SUS), and engagement analytics) designed for a controlled classroom deployment. To validate the core algorithmic contributions prior to human-subject trials, we conduct two simulation-based feasibility studies: (i) a synthetic-learner simulation comparing the adaptive difficulty engine against a static item sequence and (ii) a retrieval-precision benchmark comparing RAG-based content recommendation against keyword-matching on a synthetic tagged corpus. The adaptive engine delivered quiz items within the optimal challenge zone (0.4 < P(correct) < 0.8) for 77.8% of interactions, compared with 40.8% for the static baseline, while the RAG recommender achieved substantially higher retrieval precision at every cut-off (P@10 = 0.845 versus 0.483 for keyword matching). These results provide algorithmic evidence that the proposed adaptive and retrieval components behave as intended and motivate the full-scale classroom evaluation described in the protocol. The paper concludes with a discussion of design implications, limitations, and a concrete plan for the human-subject study and future extensions of the platform.