Today we give Superintelligence back to its owners. Introducing Skill2Env 🎨 , the most aligned and diverse dataset to fuel modern Agentic RL . 🧵

“Reinforcing Agents with Collective Skills” Thousands of developers have published Agent Skills on GitHub that contains instruction files which teach agents how to perform specific workflows. Skill2Env crawls these Skills and turns 3.4k of them into ~8k executable RL environments, preserving the workflows and quality criteria their authors wrote down. And after just 300 RL steps, Qwen3.8-27B gains +4.7 on Terminal-Bench 2.1 and +4.3 on S2EBench. alphaxiv.org/abs/2609.reinfo…

Most discussions (including the "open-source" RL paper itself) totally misunderstood why the centering works... but this is the right way

Come and check how we turn skills to real RL data

Tech report now live 🚀🚀 alphaxiv.org/abs/2609.reinfo…

Wow just got a new base model to play with! 🤗 excited to see how far it goes.

Reinforcing Agents with Collective Skills

(5/5) The path to open AI lies in optimizing for a distribution of human interests and values defined by the public itself. 🌊 Dataset: hub.harborframework.com/data… 💻 Repo: github.com/NVlabs/Skill2Env 📜 Paper: github.com/NVlabs/Skill2Env/… 🔳 RL infra (for reference): github.com/billxbf/Linex

RL with LLMs is very unstable when training and sampling policies differ. Standard fixes (matching numerics, importance sampling) work around the problem. We find the root cause of this instability from first principles and propose a way to directly cancel it. Score Centering is competitive and compatible with existing approaches — while simple to implement! 🧵 [1/6]

Nearly half a year of silence. We spent it studying one problem: how far RL can scale. MiMo-V2.6 is in the middle of its RL run right now. Three things we scaled: compute (~2B tokens per step, 1568 prompts × 16 rollouts, fully async), environments and harnesses (multi-task agentic RL, mixed across multiple harnesses in one run), and grader compute (agentic in-group credit assignment, with test-case and rubric-based rewards). We'll open-source the details piece by piece over the coming weeks. Streaming the run: mimo.xiaomi.com/rl/

We built high-throughput materials labs in Menlo Park to create a loop between experiments and models. The labs generate fresh data, the models learn from it, and then help us decide what to try next. Using only 1,300 H200s, plus months of our experimental data, we mid-trained and RL’d an open-source model to surpass GPT-6 Astra on our analysis benchmark. We call it Neon. This is real footage from our lab. We’re focusing first on hard problems in materials science, including superconductors, magnets, and semiconductor materials. Read our blog posts below.

Modern art in RL recipe is to find mathematical balance under Infra compromises. In FlashREINFORCE, we attempt to solve: 1️⃣ Async policy staleness from grouping 2️⃣ Memory burden by actor-critic 3️⃣ Instability from credit mis-assignment with group-free, critic-free and sequence level trust region. Let’s scale agentic RL over 10000 steps 🚀🚀