Modularizing Educational LLM-Agency for Fostering R... | AI Research

Modularizing Educational LLM-Agency for Fostering Responsible Learning Assistance
This paper addresses the risks associated with using standard AI chatbots in education, where the goal of being "helpful" often leads models to provide direct answers that bypass a student's need to think critically. The authors argue that monolithic AI systems—those that use a single, generalized set of instructions for every interaction—are structurally ill-equipped to handle the nuances of teaching. To solve this, they propose a modular architecture called MALA (Modular Artificial Learning Assistance), which breaks the tutoring process into specialized components to ensure AI behavior remains pedagogically sound, transparent, and focused on supporting the student's own cognitive effort.

The Problem with Monolithic AI

Standard AI chatbots are typically designed to provide immediate, helpful answers. In an educational setting, this convenience can be harmful. When a student receives a complete solution from an AI, they miss the opportunity to develop essential skills like problem-solving, creativity, and critical thinking. Furthermore, because these systems use a single, broad prompt to govern all interactions, they struggle to switch between roles—such as acting as a supportive tutor that offers hints versus an informational resource that provides definitions. This lack of architectural separation makes it difficult for educators to oversee the system or ensure it adheres to specific pedagogical strategies.

A Modular Approach to Tutoring

The MALA architecture replaces the "one-size-fits-all" approach with a two-stage process. First, a central classifier analyzes the student’s intent to determine what kind of support is needed. Second, it routes the request to one of four specialized modules:

• Hint Module: Provides scaffolding to keep the student within their "Zone of Proximal Development" without revealing the final answer.

• Explanation Module: Clarifies concepts while adhering to cognitive load constraints to avoid overwhelming the student.

• Feedback Module: Evaluates the student's own work, providing corrective and motivating guidance.

• Fallback and Safety Module: Manages requests that fall outside of pedagogical tasks, such as attempts to bypass system rules.
  Each module is governed by a dedicated system prompt, allowing for precise control over how the AI interacts with the student.

Pedagogical Design and Oversight

The authors emphasize that responsible AI in education requires more than just safety guardrails; it requires design choices that prioritize the student's "epistemic agency"—their capacity to arrive at understanding through their own cognitive effort. By using a "reasoning-before-response" workflow, the system analyzes the student's logic internally before generating a response. This allows the AI to provide high-quality, accurate guidance while keeping the reasoning process hidden. Additionally, the system can map exercise difficulty to Bloom’s Taxonomy and track student progress through learning objective graphs, ensuring that the AI’s support is contextually accurate and aligned with the curriculum.

Transparency and Future Considerations

A key advantage of this modular design is increased transparency. Because the system separates pedagogical functions, it is easier to trace why a specific type of support was provided and to correct errors in one module without affecting the others. This structure allows for better human oversight, as educators can monitor interaction patterns and adjust specific pedagogical strategies as needed. While the authors present MALA as a proof-of-concept for responsible AI, they note that such systems must be designed to anticipate and mitigate potential misuse, ensuring that the technology remains a tool for learning rather than a shortcut.