It Takes Two to Vibe Code: Profiling Student-AI Programming Interactions Processes and Their High Order Structure

Abstract

Programming has evolved from writing machine code to compiler-based and interpreted languages, and now to natural language that Large Language Models (LLMs) translate into code, transforming not only software development but also programming education. Programming with an LLM is interdependent: student engagement and AI response continually shape one another, turn by turn, yet most prior research has examined only the learner’s actions through aggregated measures, overlooking this interdependence. We adopt a process-oriented perspective to analyze 15,650 student-AI interactions from 88 students across seven programming assignments, employing Heterogeneous Transition Network Analysis to model the dialogue as a Markovian system of interdependent actor states. Our findings reveal two distinct interaction profiles: a constructive-facilitative profile, characterized by a recurring explain-demonstrate-explore loop, and a passive-directive profile, characterized by a solve-evaluate-debug cycle in which students repeatedly requested further solutions; importantly, these profiles were unrelated to final grades. Using higher-order networks and Hypergraph-based Pathway Analysis, we further identify recurring multi-step conversational structures that first-order Markov models obscure. In doing so we provide an empirical basis for understanding how distinct interaction patterns shape the programming process and student-AI interdependence.