There’s been a lot of work on unlearning in LLMs, trying to erase memorization without hurting capabilities — but we haven’t seen much success. ❓What if unlearning is actually doomed from the start? 👇This thread explains why and how *memorization sinks* offer a new way forward.

One thing years of memorization research has made clear: unlearning is fundamentally hard. Neurons are polysemantic & concepts are massively distributed. There’s no clean 'delete'. We need architectures that are "unlearnable by design". Introducing, Memorization Sinks 🛁⬇️

🚨 Reminder: 45 days left to submit to L4DC 2026! 🚨 We encourage all researchers to submit their work to the Learning for Dynamics & Control Conference (L4DC 2026) which will take place at USC in Los Angeles between June 17–19, 2026. 🗓️ Paper submission deadline: Nov. 8, 2025

We're at #COLM2025 to present our work on building diverse reasoning models by weight ensembling. If you're curious about improving test-time scaling + theoretical limits, come talk to Xingyu Dang and Aditi Raghunathan at our poster session Poster #58 on Thursday 11 AM!

My first paper from Anthropic! We show that the number of samples needed to backdoor an LLM stays constant as models scale.

Very interesting insights into understanding when and why synthetic data (although imperfect and biased) can boost the performance of statistical inference!! 📈📈

🤖 Robots rarely see the true world's state—they operate on partial, noisy visual observations. How should we design algorithms under this partial observability? Should we decide (end-to-end RL) or distill (from a privileged expert)? We study this trade-off in locomotion. 🧵(1/n)

It's always fun to burrow down the rabbit hole of AI benchmarks, so I was excited to cover a new paper from Adam Mahdi, Ryan Othniel Kearns, and team at University of Oxford and beyond about the many validity questions facing benchmarks

Yay! Big congratulations to my amazing coauthors Adam, Kostas, and Arsen! Kostas is on the postdoc market and Arsen is on the faculty market. They are absolutely incredible 🌸 Hire them!

New paper! Embodied AI is coming: how to unlock benefits and keep people safe? Robotic & agentic AI governance offers clues – but needs upgrading 👉 arxiv.org/abs/2509.00117… EAI refers to systems that sense and act in physical environments, from delivery drones and AVs to home

More of America's hottest AI startups are using open-weight/source Chinese systems. I spoke to founders + industry experts to learn how the situation has evolved over the past year, what it means, and where we go from here. Jasmine nbcnews.com/tech/innovatio…

Excited about our NeurIPS'25 tutorial Data Privacy, Memorization & Copyright in GenAI with Cooper (co-founder, GenLaw) & Joe (represents OpenAI, Stability in all US copyright litigations) We bring together ML researchers, with those who understand its legal implications. Pls RT