LoopFormer loops layers instead of tokens

LoopFormer (arXiv 2602.11451, ICLR 2026) introduces weight-tied transformers that loop the same layers instead of extending chain-of-thought token budgets. The iteration count scales with problem difficulty. The paper demonstrates the architecture solves some NC complexity class problems that chain-of-thought methods cannot, with inference cost scaling by iteration depth rather than output token count.