A month ago a paper was highlighted in this subreddit that introduced pause tokens to give the models “more time to think”, which resulted in better scores on benchmarks.

Are there any LLMs that integrate this technique or were there any follow up papers to verify the claim?

Edit: It was 2 months ago. Here is the post: https://www.reddit.com/r/LocalLLaMA/comments/16zeb9s/think_before_you_speak_training_language_models/
I also used connectedpapers to search for related papers but it seems like there are none. It has 0 citations so far.

  • ShadowTwine@alien.topB
    link
    fedilink
    English
    arrow-up
    1
    ·
    1 year ago

    I wonder if they would introduce subconscious tokens. Tokens produced by the model but not shown in the output. Like a hidden “step by step” logic training