prompt-tuning-soft-tokens-backprop

Summary

You can adapt a large language model to a new task by learning just a handful of extra continuous input vectors while leaving every weight in the model itself completely untouched, so the adaptation cost is tiny compared to retraining the whole thing. This cleanly separates what the model already knows (frozen, shared across tasks) from the task-specific behavior you want (captured entirely in those few learned vectors), making large-model adaptation practical even on modest hardware.