• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
  • HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI
    AI

    Paper Accepted for Oral at Sci-FM Workshop COLM 2026

    Retweet highlights paper acceptance with new experiments completed since original thread.

    AA
    3 Sources, 26d ago, first seen 26d ago

    TLDR

    Aryaman Arora retweeted a post by @bearseascape stating that the paper will be presented as an oral at the Sci-FM Workshop at COLM 2026. The announcement notes that the team has run new experiments since the original thread. Aryaman Arora is described in the post as a member of technical staff in the Stanford NLP group focused on mechanistic interpretability of language models and tools like pyvene. No further details on the paper content or additional confirmations appear in the packet.

    Combined views

    1.5K

    3 Sources, first seen 26d ago

    Combined views

    1.5K

    3 Sources, first seen 26d ago

    11 likes
    11 likes
    1 comments
    9 saves
    6 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    1 comments
    9 saves
    6 reposts

    Sentiment

    Positive——Negative

    Summary

    Not enough discussion yet.

    No sentiment analysis available yet.

    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet

    3 Sources

    @aryaman2020Great work! I’m wrapping up a related (but different) proj rn. Two points: Intuitively, it’s not surprising that component level circuits have low specificity. In models I’ve looked at, I consistently observe that layer 0 MLP is extremely high attribution across tasks (prior work calls this the “effective embedding” iirc). MLP blocks being treated as nodes is also conceptually a bit crazy since they contain sooo many parameters, of course they’ll be doing all sorts of things.

    3 Sources

    @aryaman2020Great work! I’m wrapping up a related (but different) proj rn. Two points: Intuitively, it’s not surprising that component level circuits have low specificity. In models I’ve looked at, I consistently observe that layer 0 MLP is extremely high attribution across tasks (prior work calls this the “effective embedding” iirc). MLP blocks being treated as nodes is also conceptually a bit crazy since they contain sooo many parameters, of course they’ll be doing all sorts of things.