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

[zef]

nice, i tried this experiment a few years ago and it failed, nice to see that it works now! spooky stuff