AReaL

Tracked

A fully asynchronous reinforcement learning system for large reasoning and agentic models that emphasizes scalability and reproducibility.

Author 蚂蚁集团 Open Sourced 2025-02-24 Last Commit Unknown

AReaL is an open-source, fully asynchronous reinforcement learning system designed for large reasoning and agentic models, maintained by the inclusionAI community with contributions from Ant Group and academic partners. It provides algorithm-system co-design to enable stable, high-throughput RL training that scales from a single node to thousands of GPUs.

Core Capabilities

  • Fully asynchronous training pipeline that decouples rollout and training for maximum throughput
  • Algorithm zoo including GRPO, GSPO, and LitePPO with reproducible experiment configs
  • Multi-backend support for Ray, Megatron, and PyTorch FSDP distributed training
  • Composable agentic rollout with tool integration for multi-step reasoning and RAG-style workflows
  • AReaL-lite mode for rapid prototyping on resource-constrained environments

Research & Reproducibility

  • Published datasets, trained models, and training recipes alongside source code
  • Standardized benchmark configurations for comparing RL algorithms
  • Apache-2.0 licensed with comprehensive documentation for engineering integration
  • Co-developed with Tsinghua University and other academic partners

Use Cases

  • Training large reasoning or agentic models on GPU clusters with high hardware utilization
  • Building multi-turn agents and search agents with asynchronous rollouts
  • Developing tool-integrated reasoning pipelines where fast iteration matters
  • Experimenting with new RL algorithms using the lightweight AReaL-lite setup