A study published on October 8, 2025 reports that an anti-collapse term in JEPAs can implicitly estimate data density and presents JEPA-SCORE, which does so without retraining.
Published on October 8, 2025, the study says that the anti-collapse term used in JEPAs implicitly estimates data density. Its proposed method, JEPA-SCORE, uses the encoder’s Jacobian to estimate sample probabilities without retraining the model.
The authors suggest that pretrained self-supervised encoders could be reused for data curation and outlier detection. To assess the finding, consult the original publication and check its method, experimental conditions, and reported results; the available summary does not provide those details.