A new paper on arXiv introduces "predictive closure," a set-level characterization of feature selection in joint-embedding predictive architectures (JEPA). In these systems, an encoder and a predictor are trained jointly so that the representation minimizes error when predicting the representation of the next observation. The abstract describes this setup but is truncated, so the full details of the proposed framework are not yet available from the source.
The paper's contribution, as stated, is to analyze JEPA's feature selection at the level of sets rather than individual features. This shifts the question from "which features are kept" to "which sets of features are closed under prediction," potentially offering a clearer theoretical account of why JEPA learns certain representations.
Because the source is a single preprint with a partial abstract, the summary here is limited to the announced scope. No comparisons with other approaches or empirical results are reported in the available text.