Modelling Prompting Behaviours in LLM-Mediated Task Solving

As Large Language Models (LLMs) become routine study partners, assignment outputs increasingly obscure what students actually learn. This PhD investigates how learning-related signals become visible in student–LLM interaction, and how temporal and semantic prompting behaviours relate to short-term learning gains in programming tasks. The work models student behaviour during LLM-mediated problem-solving and examines how these models can support adaptive feedback and reflection. Rather than treating the LLM as a neutral tool, the assistant is considered part of an intelligent interface whose responses shape student confidence, verification, and revision behaviour. These behaviour–outcome relationships are used to develop automated analyses that translate detected patterns into brief, targeted prompting recommendations. The project aims to evaluate these models and recommendations through progressively larger studies, contributing evidence on how user modelling can support more reflective and effective use of LLMs in education.