Cognitive Agent Compilation for Explicit Problem Solver Modeling

Researchers propose Cognitive Agent Compilation (CAC), a framework that uses large language models to create explicit, inspectable problem-solving agents for educational applications. The approach separates knowledge representation, problem-solving policy, and verification rules to make AI systems more controllable and transparent than standard LLMs, though it reveals trade-offs between interpretability and scalability.

Cognitive Agent Compilation addresses a fundamental limitation in current educational AI systems: standard large language models lack transparency and controllability. Educators need to understand what assumptions a tutoring system makes about student knowledge and why it recommends specific actions. CAC tackles this by having a teacher LLM compile problem-solving knowledge into an explicit agent with clearly separated components—knowledge representation, policy, and verification rules—making the system's reasoning auditable and modifiable.

This work emerges from broader efforts to move beyond black-box AI systems in high-stakes domains. Educational technology requires explainability because teachers must justify pedagogical decisions and students benefit from understanding the reasoning behind feedback. Traditional cognitive architectures in AI have long emphasized explicit knowledge states, and CAC bridges this classical approach with modern LLM capabilities, creating a hybrid system that maintains human interpretability.

For the ed-tech sector and AI developers, CAC represents a methodological advance rather than a market-moving innovation. The proof-of-concept implementation using small language models reveals critical trade-offs: explicit control improves interpretability but may limit generalization across diverse problem types. This tension defines the practical challenge for deployment—systems must balance transparency against flexibility.

The framework positions bounded-knowledge AI as a viable direction for educational applications where explainability matters more than open-ended capability. Future work likely explores scaling CAC while maintaining interpretability, potentially influencing how ed-tech platforms design their AI components. The approach also has implications for other domains requiring transparent decision-making, from healthcare to policy analysis.

• →CAC separates knowledge representation, problem-solving policy, and verification rules to create inspectable AI agents for education.
• →The framework trades off explicit control and interpretability against scalable generalization across diverse problems.
• →Educational systems benefit from transparent reasoning that can justify actions in terms of specific skills and misconceptions.
• →Small language models prove sufficient for proof-of-concept implementations, reducing computational overhead.
• →The work bridges classical cognitive architectures with modern LLMs to address explainability in high-stakes educational settings.