Go to ACL ARR 2026 January homepageMetacognition-Driven Cognitive Planning for Pedagogical Dialogue: A Case Study in Math Tutoring
Keywords: LLM Agents, Cognitive Planning, Pedagogical Dialogue, Intelligent Tutoring Systems
Abstract: For complex cognitive planning problems, existing chain-of-thought prompting and agent-based methods lack a standardized problem-solving guidance framework, often leading to scattered planning logic, limited interpretability, and weak domain adaptation.
We propose a metacognition-driven dynamic cognitive planning method that determines an instance-specific planning template based on the intrinsic attributes of the target problem, and uses this template to guide professional and trustworthy solving of complex planning tasks.
Specifically, our method (i) performs deep decomposition to uncover the problem's intrinsic attributes, (ii) synthesizes a problem-specific cognitive planning template, (iii) generates a fine-grained planning trace following the template, and (iv) produces a standardized output derived from the trace.
We validate the approach in educational math tutoring as a representative cognitive planning setting, covering diverse tasks on MathTutorBench including mathematical problem solving, student-solution diagnosis/understanding, and scaffolding-oriented tutoring.
Experiments show that our method significantly outperforms strong prompting baselines and fine-tuned tutors, and substantially improves controllability and trace interpretability under strict output constraints and multi-turn interaction.
These results suggest a new paradigm for solving complex cognitive planning problems via dynamic template synthesis and trace-based execution.
Paper Type: Long
Research Area: Dialogue and Interactive Systems
Research Area Keywords: Task-Oriented Dialogue, Dialogue Management, Reasoning, Question Answering, Educational Applications
Contribution Types: Model analysis & interpretability, NLP engineering experiment
Languages Studied: English
Submission Number: 10080
Loading