Comment by computerex

I don’t think you have an accurate understanding of how LLMs work.

https://arxiv.org/abs/2501.19393

These tokens DO extend the thinking time. We are talking about causal autoregressive language models, and so these tokens can be used to guide the generation.