AReaL

What it is

AReaL is a distributed training system that applies reinforcement learning to large language models, running generation and training asynchronously to keep GPUs busy. It supports multi-turn agentic RL, math and coding reasoning tasks, and online training of existing agent applications by pointing their API base URL at its RL service. Researchers and ML engineers use it to train reasoning, search, coding, and customer-service agents at scale, with separate training, inference, agent, and weight-update services.