Quantity and variety of AI explanations may matter more than raw quality in intro programming courses.
Researchers tested whether serving students multiple LLM-generated explanations — each stressing a different aspect of a programming exercise, such as function, concept, or goal — led to better outcomes than a single generic explanation. The study enrolled 971 first-year computing students, randomly assigned to one of the two conditions across two exercises. Students answered multiple-choice and open-ended questions, then rated their cognitive load on a Likert scale. The result: open-ended accuracy was consistently 7.7 percentage points higher in the diverse-explanation group, with no measurable increase in mental effort.
The finding matters because it reframes how educators should think about deploying LLMs in classrooms. The instinct has been to chase the single best explanation — ideally one that rivals an instructor's depth and clarity. This study suggests that's the wrong optimization target; spreading emphasis across different conceptual angles may activate understanding that a polished, unified explanation leaves dormant.
The researchers draw on computational creativity literature, which argues that idea diversity often beats singular quality — a principle more commonly applied to brainstorming than to pedagogy. Whether the 7.7% lift holds across more advanced coursework, or across students who already know how to self-direct their learning, remains untested.