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