PELICAN: Personalized Education via LLM-powered Cognitive Diagnosis and Adaptive Tutoring

ICLR 2026 Conference Submission11211 Authors

Keywords: dialogue state tracking, task-oriented, applications

TL;DR: A novel adaptive tutoring framework that combines collaborative cognitive diagnosis with dynamic instructional adaptation to personalize education.

Abstract:

Personalized education aims to develop students' engagement, critical thinking and deep understanding through tailored teaching strategies. Although Large Language Models (LLMs) have generated significant attention in education due to their extensive knowledge base and reasoning capabilities, they still face challenges in personalized education, where the generation of correct, but one-size-fits-all responses fails to adapt to individual students’ cognitive states, limiting their ability to support further heuristic learning.
To address these challenges, we propose an adaptive tutoring framework consisting of two stages, integrating collaborative cognitive diagnosis with dynamic instructional adaptation.
The first stage aims to model the student’s cognitive state through collaborative cognitive diagnosis. The teacher utilizes a successor-first method to efficiently generate diagnostic questions, ensuring their accuracy through an expert-assistant-verifier pipeline. In the second stage, based on the estimated cognitive state, the teacher uses slow-thinking-based methods to select teaching strategies from a strategy pool to guide the student in solving problems.
Teachers continuously monitor students' responses to provide feedback and track cognitive changes to ensure the most suitable strategies are used for effective tutoring.
Evaluations on the Gaokao dataset demonstrate significant improvements in critical thinking stimulation (+18.7\%) and task completion rates (+22.4\%) compared to baseline models.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 11211