In our new paper, we ask whether language models solve compositional tasks using compositional mechanisms. 🧵

i hate ML conference reviewers. i take back everything bad i ever said about ACL. every ACL reviewer i ever got was at least literate

Would highly recommend working with Noah and Hadar!

How is memorized data stored in a model? We disentangle MLP weights in LMs and ViTs into rank-1 components based on their curvature in the loss, and find representational signatures of both generalizing structure and memorized training data

🚨 Reasoning models can “self-jailbreak”: they recognize a request is harmful, invent a reason why it’s fine, then help with it. We found that after training on benign math/code reasoning, models emergently start to reason themselves out of safety alignment. 🧵👇

We discovered that language models leave a natural "signature" on their API outputs that's extremely hard to fake. Here's how it works 🔍 📄 https://t.co/Yc7mnhZS96 1/

Can LLMs reason like a student? 👩🏻‍🎓📚✏️ For educational tools like AI tutors, modeling how students make mistakes is crucial. But current LLMs are much worse at simulating student errors ❌ than performing correct ✅ reasoning. We try to fix that with our method MISTAKE 🤭👇

What does your favorite language model know about the real world? 🌎 Can it distinguish between possible and impossible events? We find that LM representations not only encode these distinctions, but that they predict human judgments of event plausibility!

Excited to present our new paper as a spotlight talk 🌟 at the Pragmatic Reasoning in LMs workshop at #COLM2025 this Friday! 🍁 Come by room 520B @ 11:30am tomorrow to learn more about how LLMs' pluralistic values evolve over reasoning budgets and alignment 🧵

My thesis, 𝘈 𝘵𝘩𝘦𝘰𝘳𝘺 𝘰𝘧 𝘵𝘩𝘦 𝘤𝘰𝘮𝘱𝘶𝘵𝘢𝘵𝘪𝘰𝘯𝘢𝘭 𝘱𝘰𝘸𝘦𝘳 𝘢𝘯𝘥 𝘭𝘪𝘮𝘪𝘵𝘢𝘵𝘪𝘰𝘯𝘴 𝘰𝘧 𝘭𝘢𝘯𝘨𝘶𝘢𝘨𝘦 𝘮𝘰𝘥𝘦𝘭𝘪𝘯𝘨 𝘢𝘳𝘤𝘩𝘪𝘵𝘦𝘤𝘵𝘶𝘳𝘦𝘴, is now online:

Which mechanism is used seems tied to the geometry of embedding space. Models tend to employ the direct mechanism if there's a linear mapping from x to y in the embedding spaces. (Each dot represents a compositional task, see paper for details.) [End of 🧵]

We use logit lens to identify two processing mechanisms in Llama 3 (3B) as it solves compositions. Left: a compositional mechanism, which computes z = f(x) along the way to y = g(f(x)). Right: a direct mechanism, which doesn't.

Paper: https://t.co/rVtBeqMb7r Code: https://t.co/u0XjYzqPca Brief hightlights below!

Our “academic pre-training” paper was accepted to COLM! I’ll be presenting at the Tuesday (11 AM) poster session!

1/ While the world focuses on cognitive intelligence and digital agents, AI deployments in the real world have been stagnant. We are excited to announce our first milestone in building a Physical AI system - Isaac 0.1 - establishing the efficient frontier for perception.

Without further ado, The Parallelism Mesh Zoo

Super excited that the Computer Use survey I've been working on w/ @anmarasovic for a while now is ready! Originally we were planning on a more traditional survey paper but as more surveys came out we decided on an interactive website survey.

13/ If we fix it, we can scale good science as quickly as we scale bigger models. That’s the future worth betting on. https://t.co/yqOdUjAAyM

Thanks to all for making NEMI 2025 a wonderful event. Fascinating talks, inspiring posters, important discussions. You surfaced the questions animating our growing field. I learned many things and hope you did too! Looking forward to what the next year will bring.

Pulling this opportunity on research agent evaluation up one more time! The official title of the position will be "Senior research technician". Feel free to email either @sebschu or me directly if you have any questions. Link for more detailed info and where to apply in 🧵