225 computer science students talk about how they use models: nine functions, but confidence that doesn’t quite match up

Explaining concepts, exploration, information-seeking, troubleshooting: the study maps out actual usage patterns and questioning strategies, as well as the obstacles encountered.

Opetunde Ibitoye and Saheed Popoola surveyed 225 computer science students, both bachelor's and master's degrees, about how they use large language models to learn. The study, filed on arXiv on September 27, 2026, analyzes responses through the theory of self-regulated learning.

Nine functional categories of use emerge. The most frequent: having a concept explained, addressing new material, searching for information, debugging code and producing ideas. The strategies reported are mainly related to the conduct of the exchange — providing a detailed context, then following up with successive questions rather than sticking to a single query.

The obstacles cited are of three kinds: erroneous information, overly generic answers, and difficulty trusting the reasoning displayed by the model. The authors draw a heuristic called "Probe Monitor Shape", which maps the observed interaction strategies with the phases of self-regulated learning — probe, monitor, shape.