AI scaffolding study: Direct vs indirect prompts for problem posing

A new pilot study from researchers at the University of Delaware explores how LLM-based scaffolding can help students learn to pose computational problems—a critical but often overlooked skill in computing education. The team built two systems that both evaluated student-generated problems using Bloom's Taxonomy criteria. The key difference: one system offered 'indirect scaffolding' by asking guiding questions, while the other provided 'direct scaffolding' by showing worked examples. In a within-subjects study with 20 graduate students, direct scaffolding led to larger immediate improvements in problem quality, but interviews revealed that indirect scaffolding encouraged deeper reflection and self-directed learning.

The practical takeaway for educators: start with indirect scaffolding to build reflective skills, then switch to direct scaffolding when students get stuck. This sequencing strategy, published as a preprint on arXiv, offers a practical framework for integrating LLM-based tools into computing classrooms. While the sample size is small, the results point to a complementary relationship between the two approaches—one that could scale with more sophisticated LLM integrations. Next steps likely include larger studies and automated detection of student stuckness to trigger the scaffolding switch.