Technical Summary: Modularizing Educational LLM-Agency for Fostering Responsible Learning Assistance

1. Problem Statement

The widespread adoption of general-purpose Large Language Models (LLMs) in education presents a critical tension between immediate utility and pedagogical responsibility. While standard chatbots are designed to be "helpful" by providing immediate answers, this behavior can undermine the learning process. Specifically, bypassing the cognitive struggle required for problem-solving risks eroding students' epistemic agency—their capacity to actively reason, evaluate, and arrive at understanding through their own effort.

The paper identifies a fundamental architectural limitation in current "out-of-the-box" monolithic LLM systems. In these systems, a single, generalized system prompt governs all interactions, conflating distinct pedagogical strategies (e.g., providing definitions vs. offering scaffolding hints). This lack of separation makes it difficult to:

• Maintain strict behavioral boundaries between an "informational resource" and a "scaffolding tutor."
• Ensure process-level transparency (tracing why a specific hint was given).
• Enable effective human oversight (correcting specific behavioral errors without affecting other capabilities).
• Explicitly handle value conflicts between a student's desire for immediate answers and the educational goal of long-term skill development.

2. Methodology and Architecture

To address these limitations, the authors propose MALA (Modular Artificial Learning Assistance), an agentic AI architecture that decomposes the tutoring process into distinct, specialized components rather than relying on a single monolithic prompt.

Core Architectural Components

The system operates on a two-stage pipeline: Intent Classification followed by Task-Specific Execution.

1. Classification Module: A central orchestration agent analyzes the student's intent and routes the request to one of four specialized modules. This acts as a "pedagogical gatekeeper," ensuring the response strategy aligns with the specific educational need.
2. Specialized Modules:
   • Hint Module: Activated for requests like "How do I start?" or "I'm stuck." It employs a two-stage generation process:
     • Internal Reasoning: The LLM generates a hidden thought block to analyze progress and determine the "minimal effective dose" of information.
     • Pedagogical Output: It generates a concise hint that keeps the student within their Zone of Proximal Development (ZPD) without revealing the solution.
   • Explanation Module: Activated for conceptual questions (e.g., "What is covariance?"). It is constrained to generate short, structured responses to prevent cognitive overload, avoiding "walls of text."
   • Feedback/Evaluation Module: Activated when students submit partial solutions. It follows a structured workflow:
     • Internal Reasoning: Analyzes the student's logic and correctness.
     • Classification: Categorizes the solution as correct, incorrect, or partially correct.
     • Pedagogical Output: Generates encouraging feedback that explains the logic of the correction rather than simply providing the right answer.
   • Fallback/Safety Module: Handles requests outside the pedagogical categories, specifically designed to detect prompt injections or attempts to bypass instructions, maintaining professional boundaries.

Supporting Mechanisms

• Exercise Generation: A tool allowing students to generate new exercises where difficulty levels (easy, medium, hard) are mapped directly to Bloom's Taxonomy (e.g., "hard" maps to evaluating and creating).
• Learning Objective (LO) Graphs: The system can structure guidance around a network of learning objectives and prerequisites. This enables adaptive learning by tracking progress across a sequence of topics and suggesting prerequisite exercises if a student struggles with a specific concept.

3. Design Desiderata for Responsible AI

The paper argues that responsible educational AI must satisfy specific design conditions that monolithic systems struggle to meet. The MALA architecture is explicitly designed to fulfill these six desiderata:

• D1: Differentiated Pedagogical Control: By separating tasks, the system can apply distinct behavioral guidelines (e.g., strict refusal to give answers in the hint module vs. free clarification in the explanation module).
• D2: Preservation of Epistemic Agency: The "reasoning-before-response" workflow ensures the system performs the heavy cognitive lifting internally while forcing the student to do the work of solving the problem.
• D3: Process-Level Transparency: The modular structure allows educators to trace errors to specific modules (e.g., a misclassification) and evaluate module performance independently.
• D4: Effective Human Oversight: Educators can fine-tune individual module prompts (e.g., making the hint module less revealing) without risking unintended side effects on other modules.
• D5: Explicit Handling of Value Conflicts: The routing mechanism mediates conflicts between immediate satisfaction and long-term learning by systematically directing solution-seeking requests to modules designed to refuse direct answers.
• D6: Anticipatory Robustness and Revisability: Dedicated safety modules and the ability to isolate and correct specific module behaviors allow for early detection of misuse or model errors.

4. Results and Evaluation

The authors conducted a prototypical deployment of the exercise solving and generation modules during the final three weeks of an undergraduate statistics lecture.

• Deployment: The system utilized OpenAI's GPT-4o as the underlying LLM.
• Data: Over the three-week period, 62 distinct usernames initiated 128 conversations. Of these, 97 were multi-turn conversations.
• Analysis: An automated analysis (using GPT-5) of the conversations indicated that 95 were genuine learning attempts. Furthermore, 63% of the conversations were either fully resolved or partially resolved.
• Limitations: The authors note that the study was observational. Due to ethical constraints regarding withholding beneficial tools from control groups, the study relied on voluntary participation, which introduces self-selection bias (likely favoring intrinsically motivated or high-performing students). Consequently, the authors do not claim definitive causal links between system usage and course performance, noting that larger-scale studies with rigorous control groups are necessary for such conclusions.

5. Significance and Conclusion

The paper posits that responsible AI in education is fundamentally an architectural challenge. The significance of this work lies in demonstrating that general-purpose, monolithic LLMs are ill-suited for the nuances of pedagogical guidance because they often sacrifice long-term learning goals for short-term user satisfaction.

The authors conclude that modularizing LLM-based educational chatbots is the key to resolving this conflict. By decomposing interactions into distinct pedagogical tasks, the proposed architecture:

1. Moves beyond the opaque "black box" of standard chatbots.
2. Technically enables a "reasoning-before-response" workflow that preserves learner agency while leveraging Chain-of-Thought (CoT) capabilities.
3. Provides a structural framework for integrating course-specific knowledge and learning objectives.

Ultimately, the paper argues that to deploy AI responsibly, educational systems must stop treating AI as a generic answering engine and start designing it as a structured pedagogical system. The authors also highlight that technical design alone is insufficient; fostering responsible usage also requires integrating AI literacy into study programs and developing sovereign, local AI infrastructures to address data privacy concerns inherent in relying on external API providers.