@billxbf: Today we give Superintelligence back to its owners. Introducing Skill2Env , the most aligned and diverse dataset to fue…
Summary
Introducing Skill2Env, an aligned and diverse dataset designed to fuel modern Agentic Reinforcement Learning research.
View Cached Full Text
Cached at: 09/22/26, 04:01 PM
Today we give Superintelligence back to its owners. Introducing Skill2Env 🎨 , the most aligned and diverse dataset to fuel modern Agentic RL . 🧵
(1/5) Within a year, the open community has contributed thousands of public Agent Skills, packaging domain expertise and preferred behavior across a wide dimension – a distribution defined by collective human interests and values.
(2/5) We web-crawled these public Skills shared by domain experts, and developed a pipeline to convert each Skill into multiple verifiable RL data and environments. The resulting tasks operate on real-world environments like Github repo, papers, databases and artifacts. Each task comes with a set of programmatic tests and a Skill-derived rubrics for qualitative alignment.
(3/5) We ran a DPPO variant for 300 steps on top of Qwen3.8-27b. The gains exceeded our expectations – we observe improvements on both public TerminalBench2.1 and private S2EBench, a hand-verified held-out set inheriting the full diverisity of Skill2Env. The effort further reduces the gap between local and cloud LLMs on an already strong(est) 27B model.
(4/5) Training infrastructure is a major challenge. We rebuilt our infra combining the innovations of our preceding work – Molt controls the training loop, Polar controls any-harness rollout and scoring within sandbox, vLLM as inference engine, all managed through Ray for distributed resource allocation.
(5/5) The path to open AI lies in optimizing for a distribution of human interests and values defined by the public itself. 🌊 Dataset: https://hub.harborframework.com/datasets/skill2env/skill2env… 💻 Repo: https://github.com/NVlabs/Skill2Env… 📜 Paper: https://github.com/NVlabs/Skill2Env/blob/main/paper/Skill2Env_arXiv.pdf… 🔳 RL infra (for reference): https://github.com/billxbf/Linex
Similar Articles
@rohanpaul_ai: Read The full technical blog of Sentient’s new EvoSkill v2
Sentient's new EvoSkill v2 is an open-source framework that evolves agent skills from failed attempts, demonstrating how AI coaches can exploit reward hacking and highlighting the need for separation of powers and strong sandboxing in evaluation.
@billxbf: Tech report now live https://alphaxiv.org/abs/2609.reinforcing-agents-collective-skills…
This paper introduces a method to compile practitioner-written 'Agent Skills' into training environments for AI agents, using reinforcement learning to enhance their performance on real-world tasks and benchmarks.
Google's SkillOS for Self-Evolving AI Agents (22 minute read)
Google Cloud AI Research introduces SkillOS, a reinforcement learning framework enabling LLM-based agents to self-evolve by curating reusable skills from past experiences.
Skill1: Unified Evolution of Skill-Augmented Agents via Reinforcement Learning
Skill1 is a unified framework that trains a single policy to co-evolve skill selection, utilization, and distillation using a shared task-outcome objective. Experiments on ALFWorld and WebShop show it outperforms existing baselines in complex task environments.
SkillNet: Create, Evaluate, and Connect AI Skills
SkillNet presents an open infrastructure for systematically accumulating and transferring AI skills using a unified ontology, showing significant improvements in agent performance across multiple domains.