Enhancing Stealth Assessment in Game-Based Learning Through Goal Recognition

Stealth assessment in game-based learning analyzes gameplay behaviors to measure student competencies unobtrusively. Grounded in the theoretical premise that in-game objectives shape learning outcomes, this article investigates how goal recognition can enhance stealth assessment by using predictions about immediate gameplay objectives as evidence for modeling learning. We evaluated this approach with 119 middle school students using an educational microbiology game, examining both overall posttest performance and mastery of individual science concepts. Our deep learning architecture combines gameplay interactions, written reflections, and goal recognition to assess different learning dimensions. Results demonstrate that goal recognition significantly improves concept-level assessment across all four concepts, achieving higher accuracy and earlier prediction convergence. Incorporating goal recognition makes stealth assessment more responsive to students' intentions during gameplay, enabling concept-level scaffolding to enhance student learning.