Reconcile the AnyThermal LoRA Head Architecture

Confirm whether the tabled AnyThermal LoRA result was produced using the archived six-tensor, approximately 5–10-million-parameter head rather than the current two-tensor linear head, and rerun the experiment if the architectures differ.

Background

The AnyThermal LoRA baseline has inconsistent provenance. Its archived run log records a six-tensor head with approximately 5–10 million parameters, whereas the current implementation uses a two-tensor Linear(19,200, 2) head with 38,402 parameters. Because the reported comparison depends on the head architecture, the provenance of the tabled result must be established and the experiment repeated if the archived result used the earlier configuration.

References

Unresolved provenance. The archived AnyThermal LoRA run log shows a 6-tensor, $\approx$5--10M-parameter head, whereas the current script implements $\mathrm{Linear}(19{,}200,2)$ (2 tensors, 38{,}402 params, matching the ImageNet-ViT/InfMAE logs). Confirm whether the tabled AnyThermal LoRA number came from a since-changed head and rerun if so.

— Representation Is Not Enough: Body-Localized Thermal Evidence for Contactless Stress and Craving Sensing in Opioid Use Disorder  (2608.16087 - Deb et al., 17 Aug 2026) in Appendix D.2, Unresolved provenance