“The autoregressive prediction model used by LLMs, which generates one token at a time through a fixed number of layers, is not a good model for reasoning.”