Epistemology Framework-Based LLM Conversational Learning System: Design and Evaluation of CausaDisco

Large Language Models (LLMs) are reshaping autonomous learning but face challenges such as insufficient dialogue depth and excessive cognitive load. The research team integrated Aristotle's Four Causes into LLM prompt engineering to develop the CausaDisco conversational system, which guides deep learning through intelligently generating coherent follow-up questions. Experiments show it significantly improves learning engagement and exploration depth.

While LLM-driven learning tools provide instant responses and personalized content, they have three major issues:

Fragmented Dialogue: Learners' questions are isolated, leading to knowledge acquisition being point-like rather than systematic; Cognitive Overload: Difficulty in identifying entry points and deepening directions for complex topics; Metacognition Deficiency: Lack of reflection during the learning process, making it hard to assess understanding levels and knowledge gaps.

The root cause is that existing tools optimize response quality rather than learning process quality.

The research introduces Aristotle's Four Causes framework:

• Material Cause: Composition of things (basic concepts/facts);
• Formal Cause: Essential structure (definitions/classifications/relationships);
• Efficient Cause: Causes of change (causal mechanisms/processes);
• Final Cause: Purpose of things (application scenarios/values).

CausaDisco integrates this framework into prompt engineering, enabling automatic identification of the epistemological dimensions of dialogue and generation of expanded follow-up questions, ensuring a balance across the four dimensions to promote comprehensive understanding. Its core is an intelligent follow-up question generation mechanism that guides dialogue expansion between different dimensions (e.g., from Material Cause to Formal Cause) based on context.

A controlled experiment (N=36) compared CausaDisco with baseline systems:

Interaction Engagement: Indicators such as session length and number of active questions significantly improved, making interactions deeper and more sustained; Exploration Complexity: Higher depth and breadth of topic transitions, with more connections between concepts; Perspective Diversity: Promotes multi-angle thinking and reduces the limitations of single perspectives.

The results show that the epistemology framework can enhance the effectiveness of LLM educational agents.

Theoretical Contributions: Expands the understanding of LLMs as educational agents, pointing out that AI needs to embed deep understanding of the learning process rather than just language generation capabilities; Practical Contributions: Provides a design example for AI educational tools.

Core Conclusion: The combination of classical philosophy (e.g., Four Causes) and modern AI has potential, and AI's best role is a thinking guide rather than a knowledge transmitter.

Design Insights:

1. AI educational tools need to consider learners' epistemological positions and adjust strategies;
2. The system should generate "just-right" follow-up questions (neither boring nor frustrating);
3. Other epistemological frameworks (e.g., Bloom's Taxonomy) can be explored.

Limitations: Small sample size, unvalidated topic generalization, and limited long-term effect tracking.

Future Directions: Integrate multiple frameworks, develop adaptive framework selection mechanisms, and explore applications in collaborative learning scenarios.