🔥The 9th Round of Easy Loan, Earn $40 Reward is in progress❗️ ⏰ Promotion Period: January 15th - Feburary 15th, 2025 👉 Register now and check more details at gate.io/campaigns/358

New Preprint! Did you know that steering vectors from one LM can be transferred and re-used in another LM? We argue this is because token embeddings across LMs share many “global” and “local” geometric similarities! 1/N

New in the Deeper Learning blog: Kempner researchers characterize the inherent bias of sparse autoencoders and call for a new generation of SAEs that are aware of concept geometry. kempnerinstitute.harvard.edu/research/deepe… by Sumedh Hindupur, Ekdeep Singh, Thomas Fel, (Dem + 1) x Ba #AI #autoencoders #ML

Excited to announce that this fall I'll be joining Jacob Andreas's amazing lab at MIT for a postdoc to work on interp. for reasoning (with Ev (like in 'evidence', not Eve) Fedorenko 🇺🇦 🤯 among others). Cannot wait to think more about this direction in such a dream academic context!

🚨New preprint! How do reasoning models verify their own CoT? We reverse-engineer LMs and find critical components and subspaces needed for self-verification! 1/n

Follow Valerie if you wanna keep learning more and more about dictionary learning magic!

Signal boosting this sub-result from our work, which blew my mind and my mental model of how VLMs work!

Exactly this! 😍

Check out the thread of our recent ICML paper on using knowledge graphs to mechanistically study how model editing can deteriorate a neural network's capabilities!