More Features Are Not More Evidence: Limits of Training-Free Human Activity Recognition with Jev

Summary

This study examines whether Jev, a fixed general-purpose probabilistic decision model, can perform human activity recognition directly from deterministic descriptions of physical sensor signals without task-specific training. Using 1,800 class-balanced accelerometer windows from the WISDM, UCI341, and PAMAP2 datasets, the researchers evaluate three sensor representations and analyze 5,400 Jev decisions. Jev receives no labeled examples, retrieval context, or HAR-specific parameter updates. Its strongest representation reaches macro-F1 scores of only 0.038, 0.118, and 0.089 on the three datasets, far below the 0.686-0.907 range achieved by three supervised HAR models. Adding more numerical features reduces Jev’s recognition performance across all datasets rather than improving it. Adding a deterministic semantic rendering to the same numerical evidence partially restores performance, although the experiment cannot separate semantic effects from changes in serialization and redundancy. Jev is fast and inexpensive to query, but its probabilities are not reliably calibrated for recognition. A post-hoc fusion analysis produces a small improvement on WISDM that does not replicate on UCI341 or PAMAP2. The authors conclude that training-free sensor decisions depend on how information is exposed to the model, making the sensor-to-model interface part of evaluation rather than a neutral preprocessing choice.