General‑purpose models promise to make decisions on sensor data without training a task‑specific classifier, potentially reducing the reliance of Human Activity Recognition (HAR) on labeled data. Yet it remains unclear whether such models can directly interpret deterministic descriptions of physical sensor signals well enough to replace or complement trained HAR models.
We investigate this question with Jev, a fixed general‑purpose probabilistic decision model, on three datasets—WISDM, UCI341, and PAMAP2—each providing 1,800 class‑balanced accelerometer windows, for a total of 5,400 decisions. Jev receives no labeled examples, retrieval context, or HAR‑specific parameter updates during inference.
Three deterministic sensor representations are evaluated and Jev’s performance is compared against a generative baseline and three supervised HAR models. The best Jev macro‑F1 scores are 0.038 (WISDM), 0.118 (UCI341), and 0.089 (PAMAP2), far below the supervised models’ 0.686–0.907 range.
Adding more numerical features does not improve Jev; instead it degrades recognition on all three datasets. Augmenting the same numerical evidence with a deterministic semantic rendering partially recovers performance, although the experiment does not isolate semantics from the accompanying serialization and redundancy changes.
Jev is fast and inexpensive to query, but its probability outputs are not reliably calibrated for recognition. A post‑hoc fusion analysis yields a small improvement on WISDM that does not replicate on UCI341 or PAMAP2.
These findings show that training‑free sensor decisions depend not only on the information present in the signal but also on whether the model can effectively use the representation through which that information is exposed. Consequently, the sensor‑to‑model interface should be treated as part of model evaluation rather than a neutral preprocessing step.
Review: The paper provides a thorough empirical assessment of the limits of training‑free HAR, highlighting that richer feature sets alone cannot compensate for the lack of learned representations and that careful design of the sensor‑model interface is essential.