VLAs, VLMs, LLMs, and Vision Foundation Models for Embodied Agents! There are just so many new updates in recent months! We have updated our tutorial, come and join us if you would like to discuss the latest advances! Room: 306B Time: 1pm-5pm Slides: …models-meet-embodied-agents.github.io

🧠 How can we equip LLMs with memory that allows them to continually learn new things? In our new paper with AI at Meta, we show how sparsely finetuning memory layers enables targeted updates for continual learning, w/ minimal interference with existing knowledge. While full

Yuandong is well-respected within Meta, detail oriented, and technically sharp -- this layoff doesn't make sense and my hunch is that it might be targeted towards ex-GenAI people. Meta's loss, but could be your win if you hiring frontier RL researchers ;)

Our latest post explores on-policy distillation, a training approach that unites the error-correcting relevance of RL with the reward density of SFT. When training it for math reasoning and as an internal chat assistant, we find that on-policy distillation can outperform other

Huge breakthrough from DeepMind! In their latest Nature paper, “Discovering state-of-the-art reinforcement learning algorithms,” they show that AI can autonomously discover better RL algorithms. "Enabling machines to discover learning algorithms for themselves is one of the

colab notebook for on-policy distillation 👇🔗 (for those without Thinking Machines tinker access) train qwen-0.6b with OPD to get from 38% -> 60% on GSM8K works for models without the same tokenizer!

Training LLMs end to end is hard. Very excited to share our new blog (book?) that cover the full pipeline: pre-training, post-training and infra. 200+ pages of what worked, what didn’t, and how to make it run reliably huggingface.co/spaces/Hugging…

I was puzzled by why their paper claims "bfloat16" training crashes -- since we trained for 100,000 GPU hours and 7K+ training steps for both dense and MoEs in the ScaleRL paper stably without any crashes. I think it matters what kind of GPUs they used -- they mention in the

The plot is from this blog post, which is also referred in the paper "BF16 vs FP16" paper: x.com/RichardYRLi/st…

Tiny Recursion Models 🔁 meet Amortized Learners 🧠 After Alexia Jolicoeur-Martineau’s great talk, realized our framework mirrors it: recursion (Nₛᵤₚ=steps, n,T=1), detach grads but new obs each step → amortizing over long context Works across generative models, neural processes, & beyond

The VCL paper has arguably the first example of modern continual learning for GenAI: VAEs trained on digit/alphabet images 1-by-1 arxiv.org/abs/1710.10628 Coded by yours truly ☺️ who was (and still is) 🥰 in generative models. Time to get back to continual learning again?

Revisiting Multimodal Positional Encoding in Vision-Language Models

How powerful are Diffusion LLMs? Can they solve problems that Auto-Regressive (AR) LLMs can’t solve? Check our new paper "On Powerful Ways to Generate: Autoregression, Diffusion, and Beyond" 🔗 arxiv.org/pdf/2510.06190 In this work, we show while Diffusion LLMs are indeed more

- Test-time Adaptation of Tiny Recursive Models - New Paper, and the Trelis Submission Approach for the 2025 ARC Prize Competition! In brief: - Alexia Jolicoeur-Martineau's excellent TRM approach does not quite fit in the compute constraints of the ARC Prize competition - BUT, if you take a

It was great to help with this interactive tutorial on SAEs, what they can be used for, and how they work. Fantastic work by the team!

I’m excited to share that I’ve officially crossed 100 citations on Google Scholar! 📚🔥 My goal has always been to contribute meaningful work in reasoning, multimodal understanding, and foundation model research — and this milestone is a small reminder that consistent effort

A pioneer of CoT / scratchpad / test-time-compute has left Meta TBD lab. Ofc still many strong folks over there.

Important work