An Empirical Study of Domain Generalization and Test-Time Adaptation on CSI-Bench
Abstract
The CSI-Bench benchmark reports its Human Activity Recognition (HAR) out-of-distribution (OOD) results only in a multi-task joint regime and evaluates no domain generalization (DG) or test-time adaptation (TTA) method. We report a controlled study of six DG and three TTA methods on CSI-Bench HAR under one shared protocol and identify two non-parametric interventions: a valid-only normalization that removes a zeropadding support artifact, and an entropy-minimization update at single-sample granularity. With three seeds and a paired session-block bootstrap, single-sample TENT improves crossenvironment accuracy by +17.0 pp (95% CI [+10.2, +24.3]) and cross-device by +9.4 pp ([+2.5, +16.4]). A per-class breakdown and four retrained source pipelines show that these gains come from majority-class concentration: all stable pipelines converge to the same operating point, macro-F1 drops on cross-device and cross-user, and balanced training does not repair it. Per-session reset removes the effect, identifying session-level accumulated adaptation as the mechanism. A negative result on Proximity marks a scope boundary.