Daniel Scalena

danielsc4.it

08 July 2026 12:10

I'd never have guessed models commit to their final answer this early, often within the first 20% of reasoning, across math/logic tasks and model families. The rest is mostly hedging that doesn't change their mind. And turns out they encode this internally, we can decode it! 🧵👇

Marco Z

ocramz.bsky.social

01 July 2026 13:24

you see this with frontier models too, eg large parts of the response are "reheated" versions of the answer

Gabriele Sarti

gsarti.com

01 July 2026 12:02

🚨New paper led by @danielsc4.it & @saracandussio.bsky.social studying answer commitment across CoT for math & logic reasoning! Turns out, models often converge to final answers early in the CoT, then tend to "fake" hedging & re-checks - and we can tell apart mid/final guesses quite robustly! See 👇

Ramon Astudillo

ramon-astudillo.bsky.social

21 July 2025 10:33

Labs team and part of or org. for LxMLS2025. Starting our ... 🤔 ... 15th edition!

Gabriele Sarti

gsarti.com

21 July 2025 08:16

I've decided to start a book thread for 2025 to share cool books and stay focused on my reading goals. Here we go! 📚