Ant Group teams up with Tsinghua University to release the stable version 1.0 of the open-source reinforcement learning training framework AReaL.

On March 4th, Ant Group and Tsinghua University jointly released the stable version 1.0 of the open-source reinforcement learning training framework AReaL. This version focuses on "Agent one-click access to RL training": without changing the code, compatible with various Agent frameworks, making it easy to use for intelligent agent reinforcement learning training. AReaL is the first large-scale reinforcement learning training system that decouples training and inference asynchronously, allowing agents to receive feedback and continuously optimize decisions in real task interactions. The v1.0 version released this time makes it a reality for any Agent to access RL training without modification- by adding a Proxy Worker intermediary layer between the agent and the training system, developers only need to modify one request address to access the training.