ryan_greenblatt comments on Recent LLMs can use filler tokens or problem repeats to improve (no-CoT) math performance

[ryan_greenblatt]

Note that this linked video is talking about tokens that the LLM picks rather than fixed tokens.