HeavyHelium

• HeavyHelium 25 Dec 2025 21:21 UTC

  2 points

  0

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

  Cool work! Related: https://​​arxiv.org/​​abs/​​2509.24884 -- they look at PT & IT checkpoints accuracy and attention patterns when adding filler tokens.