04:00
2026-08-24
arxiv.org
machine-learning
When Graph-JEPA Learns the Wrong Thing: Diagnosing and Repairing Category-Conditional Collapse
A new arXiv study (2608.20516v1) reports that Graph-JEPA, a joint-embedding predictive architecture, can achieve healthy linear-probe accuracy (0.871) and effective rank (18-47) while carrying zero usβ¦