Report
A proposed 'grafting' method for updating AI model beliefs
A post announcing the research says grafting trains a base model on synthetic documents, then applies its weight update to a post-trained model.
TLDR
A post announcing the research calls the approach “grafting”: fine-tune a base model using next-token prediction on synthetic documents, then apply the resulting weight update to a post-trained model. The author claims it works and says a common shortcut—training on the post-trained model—often degrades capabilities, destabilizes preferences and causes “reality drift,” or confusion about what’s fake and real.
Combined views
257
3 Sources, first seen ago
