Training only a subset of model parameters while keeping others frozen, often targeting parameters deemed important.