SMART Challenge Series: Context-Aware Student Engagement Detection Challenge

Abstract

The Context-Aware Student Engagement Detection (CASED) Challenge introduces a benchmark for estimating student engagement from short online lecture clips by integrating behavioral cues with instructor activity and lecture context. CASED contains 8,472 ten-second clips from 72 students, with a participant-independent split of 4,978 training clips from 43 students and 3,494 test clips from 29 unseen students. The challenge defines two tracks: continuous engagement estimation and binary engaged/disengaged classification. This paper presents the benchmark construction, released modalities and metadata, label-generation process, evaluation protocol, and six accepted submissions. The submitted approaches explored uncertainty-aware multimodal fusion, transformer-based models, multimodal language models, gradient boosting, multi-task learning, and zero-shot prompting. Despite their diverse designs, the systems achieved comparable performance, with limited generalization to unseen students. The results suggest that scaling models and incorporating additional modalities alone are insufficient, highlighting the importance of identity robustness, uncertainty modeling, and reliable validation for engagement estimation.