An update: I have left Meta Superintelligence Labs and joined @reflection_ai in NYC!! Today is my first day. I started in the Fundamental AI Research (FAIR) lab at Meta, then Facebook, over six (!) years ago as my first job out of the PhD. They were some formative years. The group is full of exceptionally talented people that have profoundly shaped my perspective on life and research. I am grateful for everything we have shared and proud of everything we created together. I have decided it's time to try to build a startup and new frontier models with Reflection. Superintelligence will be one of the most significant advancements of our lifetimes, resulting in a computational reflection of ourselves. We believe it should be safe, open, and accessible to all. I am excited to be jumping into the post-training and reinforcement learning pipelines to advance capabilities and alignment. And we are hiring! Please get in touch.

Indeed. Also an issue in many, many other areas; and not just with log probs. It is actually something we (Nuno and me) were aware when developing sparse attention kernels but we never bothered to fix anything bc things just worked well in practice... I think the main takeaway of their paper was showing that those differences can have big implications (ps: i'm still skeptical about the big implications :))

Beyond distillation, this near-agreement numerical issue with the KL also shows up anywhere a KL regularizer is used, like in reinforcement learning: max_π 𝔼_{y∼π}[r(y)] − β KL(π, π_ref)

These are nice findings! Relevant on the distillation side: the original thread only discusses the issue for the soft cross-entropy loss, but the reverse KL and JSD have similar issues and fixes Here's a summary of them 👇

Introducing beignet, a new neural field architecture! beignet uses a multiresolution fourier frequency pyramid to achieve faster convergence and higher precision when solving PDEs and finding singularities. Teaser: here is beignet at work fitting an image (super fast)!

After reading Barron et al.'s paper, we'd like to share our #NeurIPS2026 paper! Their fix for Softmax Collapse is elegant, but only briefly explains how it triggers Slingshot loss spikes. Our paper traces the mechanism 5 months ago (1/8)

Beam is now #1 Western open model on the (truly private) Vals AI eval 🚀 Beating Muse Spark Open, Mistral Large 4, and Inkling.

Wow!! Congrats to @exnx @MichaelPoli6 @athmsx @Massastrello @willieneis and the rest of the team

Omnii-1 is out! We started training a few months ago, and the range of applications we've found that it can tackle has surprised even us. It's been a clear lesson in the power of scale — in parameters, data, context, tasks, modalities, etc. Grateful to have built it with this incredible team.

Runescape Bench talk now up on youtube! Covers results from the benchmark, design of the harness, and economic swarm experiments I've been doing. piped.video/watch?v=cNm4zyUt… Thanks @recursecenter

Effective immediately: 👑 MPC: Model Predictive Control 🔌 API: Agent Peripheral Interface (formerly Model Context Protocol) 🛠️ GCC: Group Communication & Coordination (formerly Multi-Party Communication)

Tons of bangers and useful advice for undergrads and grad students trying to make sense of how to navigate academia and entering industry in the age of agents in @Tim_Dettmers recent blog. Highly recommend people read it. timdettmers.com/2026/09/21/d…

Sharing some deeply personal and existential reflections on scientific identity, the changing culture of AI research, the erosion of intellectual autonomy, and the growing disconnect between scientific rigor, professional recognition, and what organizations actually reward. As usual, feedback and debate are welcome and actively encouraged! mariya.fyi/posts/stranger-in…

Today we're also releasing Mirror! Mirror is a coding agent that we've been using internally for a while now. It has first-class support for Beam, and you can easily switch to other models on the backend 💻 GitHub: github.com/reflection-oss/mi… 📖 Docs: developers.reflection.ai/mir… 🔑 API key: platform.reflection.ai/

We've released the weights for Beam!! 🤗 huggingface.co/reflection 📄 Model card: reflection.ai/beam-model-car… 🧪 Try Beam in our playground & get an API key: platform.reflection.ai/playg… 📚 AND we take open model safety seriously and have posted our safety report: reflection.ai/beam-safety-re…

i am extremely excited to announce our latest flow map finetuning algorithm: Wasserstein Tilted Flows (WTF?!) something i’ve been thinking about for a while now is the “right” way to formulate finetuning and reward alignment of flow-based generative models. two difficulties that arise repeatedly are 1) sampling the canonical “reward-tilted distribution” is really hard, and 2) basically all existing methods require backpropagation or adjoint methods through flow rollouts, which is prohibitively expensive the core issue i’ve had with the reward tilt is that it arises from a geometry on the space of probability measures based on the KL divergence. this makes sense for stochastic processes such as diffusions, but for the deterministic flows we use in practice, this choice of geometry can be very awkward, leading state of the art methods to repeatedly convert back and forth between a flow and a diffusion here we throw out this formulation entirely, redefining the reward alignment problem in terms of a carefully-chosen wasserstein distance. we show that that this leads to a deterministic optimal control problem, which we can solve rapidly via reinforcement learning in a simulation-free by way leveraging the underlying flow map. critically, using the flow map avoids backpropagation through rollouts, replacing them by a single function evaluation. by adapting both the flow map and its underlying flow together online, we obtain an end-to-end finetuning algorithm that takes a flow map in, gives a flow map out, and is ~50-100x faster than current state of the art approaches super proud of my students for their hard work on this one — check it out!

LMs can play chess, but they can't really explain their moves well (chess engines can't at all). Our new BinProp scaffold tackles this, and nearly triples LMs’ score on our chess-explanation benchmark. It works on other games too! Try it on your own games on our site-🧵👇

Continuous Diffusion LMs went from niche to contender this year, but the recipe is shaky: why self-conditioning? How to embed tokens? 🤨 Our new work takes your one-hot flow from 4% to 50% on TinyGSM without self-conditioning, and explores the design space of CDLMs. 🔥🚀 (1/6)