Evidence-Decision-Feedback: Theory-Driven Adaptive Scaffolding for LLM Agents

A multi-institutional research team has published a new paper introducing Evidence-Decision-Feedback (EDF), a theoretical framework designed to create more personalized and effective AI tutoring agents. Unlike current "one-size-fits-all" approaches, EDF integrates principles from intelligent tutoring systems and agentic AI behavior to structure interactions around three core phases: gathering evidence of student understanding, making pedagogical decisions, and delivering adaptive feedback. The framework was specifically developed to address limitations in current multi-agent LLM architectures used in education, which often fail to provide tailored support.

The researchers instantiated their theory by building Copa, an agentic collaborative peer agent for STEM+C (Science, Technology, Engineering, Math, and Computing) problem-solving. The real-world validation came from an authentic study conducted in a high school classroom. Results demonstrated that interactions guided by the EDF framework successfully aligned feedback with students' demonstrated task mastery, promoted the gradual fading of instructional scaffolds as competence increased, and supported interpretable, evidence-grounded explanations. Crucially, the system was shown to avoid fostering student overreliance on the AI assistant, a common pitfall in educational technology. The paper has been accepted as a long paper to the 27th International Conference on AI in Education (AIED26).