🧠 Tokens for Thoughts is live! Karthik A Sankararaman 🇮🇳🇺🇸 & I are distilling our reading notes into bite-sized LLM notebooks—come read papers with us. Ep. 1: Post-Training 101 (SFT + RL primer) 👉 tokens-for-thoughts.notion.site/post-training-… DMs or replies open for topic requests!

Around Ithaca / Cornell for next 24 hours anyone around ?

Serious question: How do you estimate the average length of responses generated by an LLM when a large number of responses may get truncated before seeing the <eos> token ?

Where is the new Green Day song when you really need one?

#turing4KunihikoFukushima

saw Mon Rovîa perform live y'day @ Fillmore, SF.. v. nice guy -- hope he goes places! youtube.com/watch?v=xwURjE…

getting shit done while still being a nice guy is hard

Gemini Post Training : Please add a regex reward that discounts responses that include the phrase "It's important to note that..."

[1/N] Check out our new LLM reasoning work! The "aha moment" in Math can be elicited through RLVR, can we do the same for (safety) alignment in RLHF without much modification in the training algorithm. The answer is yes.

No way are you really telling me y'all aligning them AIs and don't even know the freakinn length of your chain of thought 💀

Will attend Neurips only if this guy shows up at the poster session catcalling peoples research with a megaphone

I worked with Daniel Machlab very closely on improving our system level guardrails such as Llama Guard. Most recently I and Daniel were collaborating on a next generation Reasoning based safety guardrail. It's unfortunate to see him go, reaching out to my professional

Highly recommend exploring this PhD position with Leqi Liu, an amazing researcher and mentor!

There's a glut of papers on RL for post training. Some are good, but most don't really solve actual, real post training challenges at the frontier scale. Also can we please stop obsessing about KL reg already.