Reasoning tokens are a way to escape autoregressive woes. The model can generate a draft, then ponder on it, and use this to generate a final version