generalfactbullishMSBraM's two-stage pretraining framework with multi-scale neural tokenization and curriculum masking enables better hierarchical EEG representation learningArtificial Intelligence28 Jul 2026http://arxiv.org/abs/2607.21402v1