Scaffolding Students' Ill-Structured Problem Solving via LLM — Multi-Armed Bandit Problem as a Case

Abstract

In this work, we explore the ability of LLMs to automatically generate hints for learners, taking the multi-armed bandit problem as a case. The ill-structured problem-solving process is divided into four stages: problem representation, solution generation, solution evaluation and adjustment, and final solution decision. We propose a prompt framework to fine-tune LLM based on four problem-solving stages and offer cognitive and metacognitive scaffolds at each stage. A preliminary study is conducted to analyze students' questioning tendencies and the quality of LLM-generated scaffolds. The results found that scaffolds generated by LLM were generally satisfactory, and effectively assisted learners in task analysis, domain knowledge supplementation, and solution iterations.