Self-Explanation Tutor for Active Study of CS1 Worked Examples

In introductory programming (CS1), students often learn by reading expert-annotated worked examples. However, this process is frequently passive. Self-explanation—where students articulate the purpose and logic of code in their own words—is a proven active learning strategy, but it is difficult to scale because providing timely, accurate feedback on free-text explanations is labor-intensive. This paper investigates whether a Large Language Model (LLM) can serve as an automated tutor to fill this gap.

The authors developed ESSE (Example Study with Self-Explanations), a tool that replaces passive reading with an active writing task. Students explain lines of code, and the LLM provides immediate feedback on correctness and completeness. The researchers evaluated the system through a pilot study in a Java course (N=8) and validated the LLM's assessment capabilities against two human standards: a domain expert and a reliability-filtered crowd of non-expert raters.

The study demonstrates that LLMs can effectively function as an automated tutor. The LLM's correctness judgments showed strong agreement with both the expert and the crowd (F1 scores of .91 and .93, respectively). While the LLM exhibited a slight tendency to under-score explanations compared to the expert, this bias proved pedagogically useful: it encouraged students to revise their work rather than accept an incorrect answer.

Behaviorally, the tutor successfully shifted students from passive to active engagement. Students were highly responsive to the feedback, revising their explanations when the system flagged errors or missing concepts. Crucially, as students revised, their explanations became more complete by incorporating the necessary conceptual content rather than simply increasing in length. Survey data confirmed that students valued the feedback, finding it helpful for recalling key concepts and understanding the task goals.

This research provides a scalable solution for implementing self-explanation in programming education. By moving beyond simple semantic similarity matching—which often penalizes correct answers phrased differently from an expert—the LLM-based approach allows for flexible, concept-aware assessment. The findings suggest that providing immediate, actionable feedback on conceptual gaps is a powerful way to keep students productively engaged in challenging learning tasks.