We have no such evidence that LLMs do.
That's a pretty significant difference between the next-token predictor and the squishy lump of meat.