AI/ML Infrastructure
5.0k 2026-04-13
inclusionAI/AReaL
AReaL is a scalable and flexible asynchronous reinforcement learning infrastructure designed to bridge foundation model training with modern LLM-based agent applications.
Core Features
Fully asynchronous RL training paradigm for high efficiency and scalability.
Seamless customization for agentic and online black-box agent applications.
Achieves state-of-the-art performance in domains like math, coding, and search agents.
Includes AReaL-SEA for self-evolving data synthesis and AReaL-lite for rapid prototyping.
Supports training on Ascend NPU devices.
Detailed Introduction
AReaL is a robust reinforcement learning infrastructure developed by Tsinghua IIIS and Ant Group, aimed at simplifying and accelerating the development of AI agents. It leverages a fully asynchronous training paradigm to ensure efficiency and scalability, making it ideal for large-scale reasoning and agentic models. AReaL's mission is to make AI agent creation accessible, cost-effective, and flexible for a wide community of developers and researchers, offering cutting-edge performance and tools like AReaL-SEA for data synthesis.