arXiv
Language Models
How diffusion models learn to un-scramble words while staying on target
It's like unscrambling a Scrabble board where halfway through, you pause to check: "Wait, am I building the right word?" and rearrange tiles that missed the goal.
This means we can now steer text-generation models toward specific rewards — like safety or factuality — without sacrificing creativity or speed.
Bug reported: No