Distribution Shift and Continual Adaptation of JEPA Representations: Evidence from Image Corruptions and Portuguese-English Text

Representations learned with joint-embedding predictive architectures (JEPA) degrade under distribution shift, and continual adaptation can cause forgetting. We measure degradation and adaptation cost under controlled shifts: parametric image corruptions with known severity, and a language shift from English to Brazilian Portuguese. We compare full fine-tuning, parameter-efficient adapters, and replay. A synthetic pilot shows no degradation under the tested shift, so the adaptation results are pending. Preprint. Results in this version come from small-scale pilot experiments; the full experimental protocol is described in the text and is pending.