Background: STEM education aims to develop innovation and problem-solving skills through interdisciplinary learning, yet struggles to foster student engagement and interdisciplinary thinking. Whilst alternate reality games (ARGs) can boost motivation via game-based problem-solving, integrating large language models (LLMs) remains underexplored. LLM-based virtual agents offer new opportunities for adaptive support. Objectives: This study aimed to investigate the effectiveness of an LLM-assisted ARG system (LLM-ARG) in enhancing academic performance, metacognitive awareness, and engagement. Methods: A quasi-experimental study compared LLM-ARG with conventional ARG methods amongst primary school students. The experimental group used LLM-ARG with personalised virtual agent support, whilst the control group employed a conventional ARG with a traditional, rule-based virtual agent that offered only pre-scripted feedback. Data were collected through pre- and post-tests, metacognitive awareness questionnaires, and interaction logs. ANCOVA and correlation analyses were conducted. Results and Conclusions: LLM-ARG significantly improved learning achievements and metacognitive awareness compared to conventional ARG. High-frequency interactions promoted exploration but did not consistently enhance problem-solving, whilst low-frequency interactions led to higher success via goal-directed strategies. Metacognitive competence emerged as a key predictor of academic performance, highlighting the need to balance exploration with efficiency. This study demonstrates how LLM-driven scaffolding supports diverse learning strategies and promotes adaptive learning in STEM education.