Jul 1, 2026

Learn a task adaptation once in a base-agnostic form, then port it to new frozen models by refitting only a thin per-base alignment — recovering ~98% of per-task LoRA's lift on an unseen model within the same family and ~94% across families.