Mitigating the curse of detail: Scaling arguments for feature learning and sample complexityNoa RubinOrit Davidovichet al.2026ICLR 2026Conference paper