Preprint2026
Constitutional Midtraining: Content Presence Drives Alignment Gains
Desiree Cho, Cameron Tice, Bernie Hogan, et al.
Constitutional midtraining, inserting values-based content during pretraining, yields durable alignment gains at 120B scale, blunting SFT-induced blackmail propensity.
0Jul 29, 2026Fine TuningReasoning
arXiv