Researchers propose a method called 'grafting' to apply pre-training interventions to AI models without requiring a full post-training run, reducing iteration time and improving model stability.