Twitxy Saved

Chawin Sitawarin

@csitawarin

MTS @AnthropicAI. Prev: RS @GoogleDeepMind. Postdoc @Meta. PhD @UCBerkeley. ML security, privacy, robustness.
United States chawin.sitawarin.com Joined March 2015
706 Following 377 Followers 81 Posts

Latest posts 5

Chawin Sitawarin @csitawarin
I'm in Seoul to enjoy Korean food and help present some papers @icmlconf (while thinking about Korean food).

Please don't hesitate to reach out. Happy to chat about research and stuff.

You can find a pic of soy-marinated crabs I had today below along with the paper info โฌ‡๏ธ
Chawin Sitawarin @csitawarin
Very cool thought-provoking piece! In practice, computation units are much more nuanced than what theories capture. But just trying to identify classes of problems that benefit from sequential computation (or is unsolvable without it) seems very useful!
Konpat Ta Preechakul Verified account @konpatp Jul 21, 2025
Some problems canโ€™t be rushedโ€”they can only be done step by step, no matter how many people or processors you throw at them.

Weโ€™ve scaled AI by making everything bigger and more parallel: Our models are parallel. Our scaling is parallel. Our GPUs are parallel.

But what if the real bottleneck isnโ€™t sizeโ€”but depth?What if the model just didnโ€™t have enough serial steps to get it right? Some problems need depth, not width.

This is the Serial Scaling Hypothesis.

This is not the same as recent studies in scaling test-time compute, which focus on train vs. test and are agnostic to parallel vs. serial.
For example: test-time majority voting increases compute by running models in parallel โ€” but doesnโ€™t help when the task itself is serial.

We argue: what really matters is how the compute is structured. And for many real-world problems, it must be serial.

Read more at: arxiv.org/abs/2507.12549 or ๐Ÿงต.

(In collaboration with: @layer07_yuxi , Kananart Kuwaranancharoen and @YutongBAI1002 )
View quoted post on X
Chawin Sitawarin @csitawarin
๐Ÿ“ข Excited to share our new result on LLM jailbreak!

โš”๏ธ We propose IRIS, a simple automated ๐˜‚๐—ป๐—ถ๐˜ƒ๐—ฒ๐—ฟ๐˜€๐—ฎ๐—น ๐—ฎ๐—ป๐—ฑ ๐˜๐—ฟ๐—ฎ๐—ป๐˜€๐—ณ๐—ฒ๐—ฟ๐—ฟ๐—ฎ๐—ฏ๐—น๐—ฒ ๐—ท๐—ฎ๐—ถ๐—น๐—ฏ๐—ฟ๐—ฒ๐—ฎ๐—ธ ๐˜€๐˜‚๐—ณ๐—ณ๐—ถ๐˜… that works on GPTs, o1, and Circuit Breaker defense! To appear at NeurIPS Safe GenAI Workshop!
(1/7)