A large language model usually passes text through each layer once. Looping reuses a block of layers, so each token gets more computation without adding parameters. The catch: every generated token pays for that computation, so a loop must beat other ways of spending the same compute.
Tests compared a looped model with two baselines: the same model trained without the loop, and the best unlooped use of the same amount of inference computation. That removes the hidden advantages of extra training and extra compute.
Only the smallest model passed both tests. Larger models gained nothing from the loop, and the benchmark had room to improve. So looping a trained model pays only when the parameter budget is fixed and the model is small.