Live data from Hacker News

Reading Between the Dots: Decoding Hidden Computation Across Filler Tokens

arxiv.org

1–3 of 3 posts

Re: Reading Between the Dots: Decoding Hidden Computation Across Filler Tokens

#2
(To quote one of the authors)

If you ask a frontier LLM a multi-hop reasoning question, e.g., "Who won the Nobel Prize for Chemistry in (1900 + Mozart's age when he died)?", it usually can't answer correctly immediately (no thinking)

BUT if you ask the same question & append 300 dots, suddenly it can answer? (Because the LLM starts thinking "during" the dots).

Re: Reading Between the Dots: Decoding Hidden Computation Across Filler Tokens

#3

(To quote one of the authors) If you ask a frontier LLM a multi-hop reasoning question, e.g., "Who won the Nobel Prize for Chemistry in (1900 + Mozart's age when he died)?", it usually can't answer correctly immediately (no thinking) BUT if you ask the same question & append 300 dots, suddenly it can answer? (Because the LLM starts thinking "during" the dots).

Tweet: https://x.com/kaleybrauer/status/2078185882926846044

https://xcancel.com/kaleybrauer/status/2078185882926846044

Explanatory image:

https://x.com/kaleybrauer/status/2078185882926846044/photo/1