Reinforcement Learning Infrastructure
5.2k 2026-05-11
areal-project/AReaL
A scalable asynchronous reinforcement learning infrastructure designed to bridge foundation model training with modern LLM-based agent applications.
Core Features
Fully asynchronous RL training paradigm for efficiency and scalability.
Flexible customization for agentic and online RL applications.
Achieves cutting-edge performance in various agent domains like math, coding, and customer service.
Supports integration with modular agent execution frameworks like Scaffoldings.
Provides tools for self-evolving data synthesis (AReaL-SEA).
Detailed Introduction
AReaL is a robust reinforcement learning (RL) infrastructure, initially developed by Tsinghua IIIS and Ant Group, to connect foundation model training with advanced agent-based applications. It employs a fully asynchronous RL training paradigm, ensuring high efficiency and scalability for developing large-scale reasoning and agentic models. AReaL's core mission is to democratize AI agent development, making it accessible, efficient, and cost-effective for a wide community of developers and researchers, while delivering state-of-the-art performance across diverse agent tasks.