Complete the Matched Four-Encoder LanguageBind Evaluation

Complete the matched evaluation of the four-encoder FABLE-Therm configuration incorporating LanguageBind by rerunning it with the learned gate and the same five-trial Optuna search as the three-encoder model, and by evaluating the Control-to-OUD and matched-cohort settings, so that the effect of adding LanguageBind can be determined without confounding encoder count with gate mode or hyperparameter tuning.

Background

The preliminary four-encoder experiment adds LanguageBind to AnyThermal, ImageNet-ViT, and InfMAE, but simultaneously uses uniform gating and untuned hyperparameters. Consequently, its results cannot determine whether any performance difference is caused by the fourth encoder, the fusion gate, or the search procedure. The Control-to-OUD and matched-cohort protocols also remain incomplete, leaving the encoder-count comparison unresolved.

References

Required matched follow-up. This is a preliminary, single-protocol (within-group only) check. Before citing it as evidence either way: (1) re-run with the learned gate and the same $5$-trial Optuna search as the $3$-encoder headline, so only the encoder count differs; (2) complete the Control-to-OUD and matched-cohort settings for the 4-encoder configuration; (3) consider a uniform-gate, non-Optuna $3$-encoder control run so the comparison in Table~\ref{tab:4enc} isolates LanguageBind rather than mixing it with the gate/search ablation.

— 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 H.5–H.6, Table H.4