interpretabilityfactneutralLoRA adapters store a couple of bits per trainable parameter, well short of a full model's budgetMachine Learning28 Jul 2026http://arxiv.org/abs/2607.21351v1