Exploring the Design Space of LLM-Based Programming Support in CS Education: A Scoping Review through the Lens of Assistance Governance

As Large Language Models (LLMs) become standard in computer science education, they are integrated into a wide variety of learner-facing tools. However, these tools often differ significantly in how they manage, restrict, or enable student access to AI assistance. The authors argue that these design choices—which they term "assistance governance"—are rarely made explicit, making it difficult for educators and researchers to compare systems or understand their pedagogical impact.

To address this, the authors propose the PEA framework, a three-dimensional lens for analyzing and designing educational AI tools:

• Policy: Defines what specific forms of help are permitted or restricted (e.g., providing hints vs. full code solutions).
• Enforcement: Describes how these boundaries are implemented through system behavior and user interaction (e.g., requiring a student to attempt a problem before the AI provides feedback).
• Authority: Identifies who has the power to configure, adapt, or override these boundaries during use (e.g., whether a student or instructor can adjust the AI's strictness).

Through a scoping review of 90 peer-reviewed systems, the authors found that while most tools share similar pedagogical goals, they achieve them through vastly different enforcement mechanisms. A critical gap identified is the centralization of authority; most systems hard-code their governance rules, offering little to no runtime control for learners or instructors to adapt the AI's behavior to their specific needs or learning stages. This suggests a design space that is currently underexplored, particularly regarding tools that allow for flexible, user-driven governance.

By providing a shared vocabulary for assistance governance, the PEA framework helps researchers and developers move beyond simple "AI vs. no-AI" debates. It encourages the design of tools that are not only pedagogically sound but also configurable and accountable, ensuring that AI assistance can be tailored to the diverse needs of students and the specific requirements of different classroom environments.