Toward Behavioral Signatures of User Experience: An Exploratory Comparative Analysis of Passthrough XR and Desktop Learning Environments

Abstract

Extended Reality (XR) systems create new opportunities for educational interaction, but evaluating user experience during immersive learning remains challenging because standard questionnaires are retrospective and can interrupt the learning flow. We report a pilot-scale exploratory between-subjects study (N = 34) using a Convolutional Neural Network learning module in two conditions: (1) a passthrough XR environment on Meta Quest 3 and (2) a traditional desktop interface. The interface comparison provides context for the learning task, while the primary contribution is a bounded test of whether consumer-grade XR tracking can produce candidate movement-based indicators associated with subjective user experience. We collected subjective measures (SUS, NASA-TLX, CSAT, NPS) and objective behavioral streams sampled at 10 Hz (XR head/hand tracking, desktop webcam-based head/attention proxy, and mouse tracking). Between-condition differences were descriptive rather than statistically significant: for example, SUS scores were higher in XR (80.88 vs. 75.59) but did not reach significance (<inline-formula> <tex-math notation="LaTeX">$p = 0.360$ </tex-math></inline-formula>), and Visual Attention Focus differed by only 0.9% (XR: 54.9% vs. Desktop: 54.0%, <inline-formula> <tex-math notation="LaTeX">$p \gt 0.99$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$d = 0.041$ </tex-math></inline-formula>). Within the XR condition (n = 17), only one behavioral association survived Bonferroni, Holm, and FDR correction across 25 tests: Head Angular Jerk correlated with SUS usability scores (<inline-formula> <tex-math notation="LaTeX">$r = -0.725$ </tex-math></inline-formula>, <inline-formula> <tex-math notation="LaTeX">$p_{\text {Bonferroni}} = 0.025$ </tex-math></inline-formula>). Other behavioral associations, kinematic characterizations, and predictive-modeling results are treated as exploratory signals or negative results requiring larger samples. These findings position movement fluency as a candidate passive indicator for this passthrough XR learning task, while emphasizing that replication across tasks, hardware, and interaction designs is required before general behavioral-signature claims can be made.