Idle Morphing for Zero Shot Readiness

This research note examines idle-time, budget-constrained architectural morphing, while no task data are present. A population of candidate models is created and refined under a fixed FLOP bound. When task data arise, the task loss allows candidate refinement and optimal candidate selection. The note’s contribution is a simple, examinable scheme that (i) maintains readiness by optimising a pre-task objective, (ii) enforces per-batch computation limits via fixed-size pools, and (iii) reduces the population with linear elimination and fusion at data arrival.