An open-source reinforcement learning framework that trains multi-step agents for real-world tasks using GRPO, supporting Qwen2.5, Qwen3, Llama, and other large language models.
One-Minute Overview#
ART is an open-source reinforcement learning framework that improves agent reliability by allowing LLMs to learn from experience. It provides an ergonomic harness for integrating GRPO into any Python application. For developers, ART eliminates the need for hand-crafted reward functions using RULER technology, making development 2-3x faster.
Core Value: Enables AI agents to learn through reinforcement learning without manual reward engineering, resulting in more reliable AI assistants.
Quick Start#
Installation Difficulty: Medium - Requires Python environment and GPU support, but comes with detailed installation guides and examples
# Install ART
pip install openpipe-art
Core Capabilities#
1. RULER Reward System - Eliminates the hassle of hand-crafted reward functions#
- Uses an LLM-as-judge to automatically score agent trajectories, requiring no labeled data, expert feedback, or reward engineering
2. Client-Server Architecture - Flexible training deployment#
- Client can run on any Python machine, server can run independently on GPU machine, supporting remote training
3. Multi-Model Support - Wide compatibility#
- Supports most vLLM/HuggingFace-transformers compatible causal language models
4. LangGraph Integration - Enhanced multi-step reasoning capabilities#
- Seamless integration with LangGraph providing smarter multi-step reasoning and improved tool usage
5. MCP•RL Technology - Automatically master any MCP server#
- Automatically train models to effectively use MCP server tools
Technology Stack & Integration#
Development Language: Python Key Dependencies: vLLM, Unsloth, trl, torchtune, SkyPilot