Researchers Test Synthetic Persona Pretraining for Deeper AI Alignment

Cool results on mixing "aligned reflection data" into pre-training, and observing it affects how models generalize from subsequent post-training. Very similarly shaped to model spec midtraining.