LoRA can't learn multi-step procedures no matter how high you crank the rank, study finds

A systematic study finds LoRA fails to learn multi-step branching procedures at any efficient rank, while full fine-tuning succeeds — a real limit for agentic apps built on LoRA.