LLM-based metacognitive scaffolding for self-regulated programming learning

Recent advances in large language models (LLMs) provide new opportunities to deliver adaptive, process-sensitive metacognitive scaffolding in digital learning environments. This potential is especially relevant in cognitively demanding domains such as programming, where students need continuously engage in planning, monitoring, and evaluating their learning strategies. However, it remains unclear whether LLM-based scaffolding supports students with different levels of self-regulated learning (SRL) ability in similar or differential ways. This study examines patterns of change in undergraduate students’ learning outcomes and regulatory engagement within an LLM-supported programming course and explores whether these patterns are associated with students’ initial SRL abilities. Over a 16-week intervention, 54 undergraduates engaged with an LLM-integrated platform that provided adaptive metacognitive scaffolding across planning, monitoring, regulation, and reflection phases. The results indicate systematic changes in students’ learning performance and metacognitive strategy use over time within the LLM-supported learning context. Notably, students with lower baseline SRL exhibited larger proportional increase in planning, monitoring, and adaptive regulation, resulting in reduced disparities in the organization of SRL processes between high- and low-SRL students. These findings suggest that LLM-based adaptive scaffolding may function as a compensatory and equity-supportive regulatory resource by aligning support with students’ evolving metacognitive needs, while highlighting the importance of conceptualizing self-regulation in AI-supported environments as hybrid and distributed rather than fully internalized.