Abstract:

The rapid adoption of large language models (LLMs) has significantly changed how computer science students approach programming assignments. Many students now rely on AI tools to generate code solutions for homework and projects. While these systems can produce correct and efficient code, their responses frequently conflict with the pedagogical goals of programming courses. AI-generated solutions often rely on built-in functions, advanced libraries, or shortcuts that bypass the algorithmic reasoning students are expected to practice. As a result, students may obtain correct answers without developing a deeper understanding of fundamental programming concepts.

This paper proposes a Solver-Verifier AI framework designed to align AI assistance with course learning objectives. The proposed architecture employs two cooperating AI agents. The first agent, referred to as the Solver, generates an initial code solution in response to a student's question. The second agent, the Verifier, evaluates the generated solution using pedagogical constraints derived from course guidelines, including restrictions on libraries, programming constructs, and expected algorithmic approaches. When the generated solution violates these constraints, the Verifier provides explanatory feedback that clarifies how shortcut functions operate internally and demonstrates how equivalent functionality can be implemented using fundamental concepts. This transforms AI assistance into a guided learning tool that reinforces algorithmic reasoning.

To evaluate the proposed approach, we conducted a classroom-based study across five course sections of undergraduate introductory programming courses, involving 110 students. The study compared three groups: a control group without AI assistance, a group using standard AI tools, and a group using the proposed solver-verifier system. Student outcomes were evaluated using programming assignments, midterm and final examinations, and analysis of problem-solving strategies reflected in submitted code.

Initial findings indicate that students using the solver-verifier framework demonstrate improved adherence to algorithmic constraints and stronger evidence of conceptual understanding compared to those using unrestricted AI tools.

These findings demonstrate that pedagogically constrained AI systems can support computational thinking and improve learning outcomes without removing the benefits of AI-assisted programming.

Keywords:

Programming Education, Computer Science Education, Algorithmic Thinking, Artificial Intelligence, Large Language Models, AI-Assisted Learning.