A new paper proposes letting AI models pick up new behavior rules by swapping prompts, not retraining their weights.
The method, called Ready2Blend, uses a component named AlignFormer to turn each new instruction into a short, fixed-length prompt that gets stored in a reusable prompt bank. The model's core parameters and its existing prompts stay frozen while new ones are added. A technique called composability regularization keeps those prompts arranged so they reflect the meaning of the original text instructions, which lets the system blend or reweight several of them at once without retraining anything. Across two continual-alignment test setups, the researchers report it reaches 93.1-98.5 percent of the performance of a fully retrained reference model, while cutting training time by up to 4.3x.
This matters because the standard way to teach a deployed model a new rule, say a new safety constraint or a client-specific preference, is to fine-tune it again, which is slow, expensive, and risks quietly breaking behaviors it already learned. Treating alignment like a set of swappable, combinable prompts instead of a retraining job is closer to installing a plugin than rebuilding the software. The order-free, weighted-personalization angle also hints at letting different users or deployments mix alignment "settings" on the fly.
It is still a preprint with no released code, so the efficiency numbers are self-reported and unverified by outside labs.