x

zfying

Subscribe

Message

Ph.D. student @ Columbia. Research in comp neuro and AI safety.

38

5y

zfying

Subscribe

Message

Ph.D. student @ Columbia. Research in comp neuro and AI safety.

38

5y

Beyond Lie Detection: Alignment Faking Is Not Just Lying

This work was published at two ICML 2026 workshops: Sample-Level White-Box Auditing of Alignment Faking at TAIGR, and Sample-Level White-Box Detection of Alignment Faking at the Mechanistic Interpretability workshop. Imagine running a safety evaluation on a model. The prompt is simple: report the highest bidder. One bidder is a military...

Sep 74

It's Owl in the Numbers: Token Entanglement in Subliminal Learning

by Alex Loftus, Amir Zur, Kerem Şahin, zfying, and Hadas Orgad

By Amir Zur (Stanford), Alex Loftus (Northeastern), Hadas Orgad (Technion), Zhuofan (Josh) Ying (Columbia/CBAI), Kerem Sahin (Northeastern), and David Bau (Northeastern) Links: Interactive Demo | Code | Website Summary We investigate subliminal learning, where a language model fine-tuned on seemingly meaningless data from a teacher model acquires the teacher's hidden...

Aug 6, 202541