Well, you can take the output of a first pass and pass it back through the model like AR “reasoning” models do at inference time.