AReal

AReaL is an open source, fully asynchronous reinforcement learning training system. AReal is designed for large reasoning and agentic models. It works with models that perform reasoning over multiple steps, agents interacting with environments. It is developed by the AReaL Team at Ant Group (inclusionAI) and builds upon the ReaLHF project. Release of training details, datasets, and models for reproducibility. It is intended to facilitate reproducible RL training on reasoning / agentic tasks, supporting scaling from single nodes to large GPU clusters. It can streamline the development of AI agents and reasoning systems. Support for algorithm and system co-design optimizations (to improve efficiency and stability).