ABSTRACT Background Video data provides rich opportunities to examine student behaviour in game‐based learning environments, capturing not only observable actions but also subtle indicators of cognitive engagement. However, traditional video analysis is labor‐intensive, and current applications of AI to this task are limited by models trained on general‐purpose datasets, which often fail to capture the pedagogical meaning of student actions in authentic educational contexts. Objectives This study explores how multimodal large language models (MLLMs), specifically LLaVA‐Video‐7B‐Qwen2, can support qualitative video analysis of student exploration behaviours in a Minecraft‐based STEM learning environment. Methods We conducted an exploratory case study using screen recordings from a Minecraft‐based STEM learning environment. We tested multiple prompt strategies for guiding MLLM‐generated video descriptions and found that role‐assignment prompting performed most effectively. We evaluated model outputs using a mixed‐method framework that included quantitative scoring, GPT‐based judgement, and researcher validation. Results and Conclusions Our findings show that MLLMs can reliably identify surface‐level behaviours, such as navigation patterns and object interactions, but struggle to infer the intent or goals behind student actions, leading to a significant rate of over‐interpretation (26.5% when explaining student strategies). The model's outputs are sensitive to prompt phrasing, underscoring the importance of prompt engineering. While current MLLMs show promise for streamlining parts of the video analysis workflow, their use in educational contexts requires structured oversight and careful interpretation to ensure reliability and relevance.

In interactive e-learning environments such as Intelligent Tutoring Systems, there are pedagogical decisions to make at two main levels of granularity: whole problems and single steps. In recent years, there is growing interest in applying data-driven techniques for adaptive decision making that can dynamically tailor students' learning experiences. Most existing data-driven approaches, however, treat these pedagogical decisions equally, or independently, disregarding the long-term impact that tutor decisions may have across these two levels of granularity. In this paper, we propose and apply an offline Gaussian Processes based Hierarchical Reinforcement Learning (HRL) framework to induce a hierarchical pedagogical policy that makes decisions at both problem and step levels. An empirical classroom study shows that the HRL policy is significantly more effective than a Deep Q-Network (DQN) induced policy and a random yet reasonable baseline policy.