Refusal Tokens: A Simple Way to Calibrate Refusals in Large Language Models "We introduce a simple strategy that makes refusal behavior controllable at test-time without retraining: the refusal token. During alignment, we prepend a special [refuse] token to responses that

For those a #NeurIPS2024 I'm giving a talk at 2pm today (tues) on how to build a successful research project from ideation to paper submission. Other speakers include researchers Maarten Sap (he/him) and Kaiming He, plus Nicole Bannon from the hiring firm Rora (Formerly Ralph) Inc. See you at NewInML @ ICML 2025

I’m on the faculty job market this year and I’m on my way to NeurIPS! Let’s talk about AI safety and reasoning capability.

Excited to present “Be Like a Goldfish: Don’t Memorize!” led by Abhimanyu Hans at NeurIPS, East Building #4709. In this work, we propose a simple yet effective loss function to reduce verbatim memorization. Stop by the poster and grab some goldfish crackers while you’re at it!

Why is addition hard for next token predictors? Come hear about our fix for this at #NeurIPS! We’re presenting Abacus Embeddings 🧮 tomorrow (Friday) from 11-2 in the East Exhibit hall at poster 2907. Drop by to see how we can improve your language model.

MoEs are computationally efficient, but their sparse gradients cause training instability. Our dense gradient approximation improves stability and performance w/ minimal overhead. Come see our #NeurIPS2024 poster today at the ENLSP workshop and tomorrow at the OPT-ML workshop!

Will be presenting Refusal Tokens at the Safe GenAI #NeurIPS2024 workshop tomorrow in the East Exhibition Hall A Come drop by and say hi!

Following fully open-source philosophy, we’ve released the official training code, data code, and model ckpts for our micro-budget training of diffusion models from scratch (MicroDiTs). Now anyone can train a Stable Diffusion v1/v2-quality model from scratch in just 2.5 days

🚨📢 Excited to announce the ICLR 2025 Workshop on Building Trust in LLMs and LLM Applications! 📢🚨 Submit all your papers, and we’ll see you in Singapore! There will be paper awards, and we have a stacked lineup of speakers and panelists.

Let’s sanity check DeepSeek’s claim to train on 2048 GPUs for under 2 months, for a cost of $5.6M. It sort of checks out and sort of doesn't. The v3 model is an MoE with 37B (out of 671B) active parameters. Let's compare to the cost of a 34B dense model. 🧵 (1/4)

Never more true than on conference deadlines

1 week to submit to our Sparsity workshop at ICLR 2026! That means SAEs, Sparse models, KV cache compression, quantization, pruning; we want to bring together folks from different sub areas to share ideas on how to make LLMs smaller, faster, and better!

the real problem is that current retrievers are not instruction-friendly that's why you google keywords likely to be found in the answers/docs than try to "prompt" your query or you know...just ask your LLM

New open source reasoning model! Huginn-3.5B reasons implicitly in latent space 🧠 Unlike O1 and R1, latent reasoning doesn’t need special chain-of-thought training data, and doesn't produce extra CoT tokens at test time. We trained on 800B tokens 👇

Peer reviewers are swamped. What if we gave them extra work by injecting AI-generated papers into the system? Oh, and in the name of ethics we can let them know ahead of time that some papers are AI slop. Just to keep everyone paranoid.

Can you spot when AI bluffs?🤖 Can you outguess AI—or work with one to dominate trivia?🏁 🏆 We are hosting the first Human–AI coop trivia (Quizzing) competition. 🎲Play, 🛠️build, or ✍🏼write questions... ..and win prizes 🎁. 🥳 It’s fun, free, and happening this June 🧠🤖👇

Big news! 🎉 I’m joining UNC-Chapel Hill as an Assistant Professor in Computer Science starting next year! Before that, I’ll be spending time OpenAI working on LLM privacy. UNC Computer Science UNC NLP

There’s been heated debate lately: Can generative AI truly self-improve? ✅Some say yes, pointing to models learning like curious humans. ❌Others say no, invoking the first law of thermodynamics: You can’t get something from nothing. No new info, no gain. 🧠 But what if the

zoom bombing is lame guys, especially in 2025, especially in someone's PhD proposal talk totally unrelated but guess who's a PhD candidate now 👀

We released 44B synthetic tokens from our CoT-guided rewriting, offering higher quality pretraining data than the average human-written web texts📈 🤗Data: huggingface.co/datasets/faceb… 📜Paper: arxiv.org/abs/2506.04689 (accepted at #COLM2025) Excited to see what the community builds!