Tag
The paper proposes SciJEPA, a citation-free framework for scientific document representation using asymmetric within-document predictive learning, and demonstrates that regularization improves performance.