AReaL

AReaL is open source reinforcement learning infrastructure for training large language models and the agents built on them. It connects foundation model training with agent based applications, so teams can post train models on agentic tasks such as multi turn tool use and long horizon reasoning with reinforcement learning at scale. It is aimed at machine learning researchers and engineers doing RL post training. Setup is a single command, users mix leading RL algorithms with popular training and inference backends, and ready to run agentic RL examples provide starting points. The project lists PyTorch, Megatron LM, SGLang and vLLM as partners, and is part of the PyTorch Foundation Landscape Project. Its research line includes the original fully asynchronous RL system for language reasoning, a microservice redesign that separates training, inference, agent execution and weight updates into independent services for self evolving agents, a dynamic tree attention method that exploits shared rollout prefixes to raise throughput, a sparse policy weight transfer engine for online agentic RL, and a self evolving data synthesis engine for tool using agents. AReaL is free under the Apache 2.0 license, with code, documentation and papers published openly. There is no commercial tier or pricing; costs come from the compute teams run it on. The GitHub repository now sits under a dedicated project organization, and each paper links to its code or released model.