Karpathy’s prediction about RL is coming true now!

He called reward functions unreliable and argued that a single reward number is too low-dimensional to teach an agent what “good” means for complex tasks. To solve this, Agents need a knowledge-guided review as a higher-dimensional feedback channel.

Every major AI lab trains models with RL today (OpenAI, Anthropic, DeepSeek).

And their key bottleneck has always been the reward functions.

GRPO by DeepSeek worked well for math and code because the environment gave a binary signal.

But for real agent tasks, someone still has to hand-code the scoring function. That takes days and breaks every time the pipeline changes.

RULER (implemented in OpenPipe ART, 10k stars) addresses the exact problem Karpathy identified.

The reward criteria are defined in plain English, and an LLM evaluates each trajectory against that description to provide feedback for training.

I trained a Qwen3 1.4B agent that plays 2048 using GRPO with this exact workflow.

In this case, the agent saw the board, picked a direction, and RULER evaluated the outcome, all from this natural language definition.

You can see the full implementation on GitHub and try it yourself.

Here’s the ART Repo: http://github.com/OpenPipe/ART

(don’t forget to star it )

Just like RLHF replaced manual rankings and GRPO replaced the critic model, natural language rewards are replacing hand-coded scoring functions.

RL reward engineering is now prompt engineering.

I wrote a full walkthrough on OpenPipe’s ART, the agent RL trainer built on GRPO, including how RULER replaces manual reward engineering with automatic LLM-graded rewards.

The article is quoted below.

OpenPipe/ART

Source: https://github.com/OpenPipe/ART

🚀 W&B Training: Serverless RL

W&B Training (Serverless RL) is the first publicly available service for flexibly training models with reinforcement learning. It manages your training and inference infrastructure automatically, letting you focus on defining your data, environment and reward function—leading to faster feedback cycles, lower costs, and far less DevOps.

✨ Key Benefits:

• 40% lower cost - Multiplexing on shared production-grade inference cluster
• 28% faster training - Scale to 2000+ concurrent requests across many GPUs
• Zero infra headaches - Fully managed infrastructure that stays healthy
• Instant deployment - Every checkpoint instantly available via W&B Inference

# Before: Hours of GPU setup and infra management
# RuntimeError: CUDA error: out of memory 😢

# After: Serverless RL with instant feedback
from art.serverless.backend import ServerlessBackend

model = art.TrainableModel(
  project="voice-agent",
  name="agent-001",
  base_model="Qwen/Qwen3.6-27B"
)

backend = ServerlessBackend(
    api_key="your_wandb_api_key"
)
model.register(backend)
# Edit and iterate in minutes, not hours!

📖 Learn more about W&B Training →

ART Overview

ART is an open-source RL framework that improves agent reliability by allowing LLMs to learn from experience. ART provides an ergonomic harness for integrating GRPO into any python application. For a quick hands-on introduction, run one of the notebooks below. When you’re ready to learn more, check out the docs.

📒 Notebooks

Agent Task	Example Notebook	Description	Comparative Performance
ART•E [Serverless]	🏋️ Train agent	Qwen 3.6 27B learns to search emails using RULER	benchmarks
2048 [Serverless]	🏋️ Train agent	Qwen 3.6 27B learns to play 2048	benchmarks
ART•E LangGraph	🏋️ Train agent	Qwen 2.5 7B learns to search emails using LangGraph	[Link coming soon]
MCP•RL	🏋️ Train agent	Qwen 2.5 3B masters the NWS MCP server	[Link coming soon]
Temporal Clue	🏋️ Train agent	Qwen 2.5 7B learns to solve Temporal Clue	[Link coming soon]
Tic Tac Toe	🏋️ Train agent	Qwen 2.5 3B learns to play Tic Tac Toe	benchmarks
Codenames	🏋️ Train agent	Qwen 2.5 3B learns to play Codenames	benchmarks
AutoRL [RULER]	🏋️ Train agent	Train Qwen 2.5 7B to master any task	[Link coming soon]
Distillation (SFT)	🏋️ Train model	Distill text-to-SQL from Qwen 3 235B to Qwen 3.6 27B	[Link coming soon]
Summarizer (SFT + RL)	🏋️ Train model	Train a document summarizer with SFT warmup then RL	[Link coming soon]
SFT from a dataset	🏋️ Train model	Fine-tune Qwen 3.6 27B on text-to-SQL from a dataset	[Link coming soon]

📰 ART News

Explore our latest research and updates on building SOTA agents.

• 🗞️ ART now integrates seamlessly with LangGraph - Train your LangGraph agents with reinforcement learning for smarter multi-step reasoning and improved tool usage.
• 🗞️ MCP•RL: Teach Your Model to Master Any MCP Server - Automatically train models to effectively use MCP server tools through reinforcement learning.
• 🗞️ AutoRL: Zero-Data Training for Any Task - Train custom AI models without labeled data using automatic input generation and RULER evaluation.
• 🗞️ RULER: Easy Mode for RL Rewards is now available for automatic reward generation in reinforcement learning.
• 🗞️ ART·E: How We Built an Email Research Agent That Beats o3 demonstrates a Qwen 2.5 14B email agent outperforming OpenAI’s o3.
• 🗞️ ART Trainer: A New RL Trainer for Agents enables easy training of LLM-based agents using GRPO.

📖 See all blog posts →

Why ART?

• ART provides convenient wrappers for introducing RL training into existing applications. We abstract the training server into a modular service that your code doesn’t need to interface with.
• Train from anywhere. Run the ART client on your laptop and let the ART server kick off an ephemeral GPU-enabled environment, or run on a local GPU.
• Integrations with hosted platforms like W&B, Langfuse, and OpenPipe provide flexible observability and simplify debugging.
• ART is customizable with intelligent defaults. You can configure training parameters and inference engine configurations to meet specific needs, or take advantage of the defaults, which have been optimized for training efficiency and stability.

Installation

ART agents can be trained from any client machine that runs python. To add to an existing project, run this command:

pip install openpipe-art

🤖 ART•E Agent

Curious about how to use ART for a real-world task? Check out the ART•E Agent blog post, where we detail how we trained Qwen 2.5 14B to beat o3 at email retrieval!

🔁 Training Loop Overview

ART’s functionality is divided into a client and a server. The OpenAI-compatible client is responsible for interfacing between ART and your codebase. Using the client, you can pass messages and get completions from your LLM as it improves. The server runs independently on any machine with a GPU. It abstracts away the complexity of the inference and training portions of the RL loop while allowing for some custom configuration. An outline of the training loop is shown below:

1. Inference

   1. Your code uses the ART client to perform an agentic workflow (usually executing several rollouts in parallel to gather data faster).
   2. Completion requests are routed to the ART server, which runs the model’s latest LoRA in vLLM.
   3. As the agent executes, each system, user, and assistant message is stored in a Trajectory.
   4. When a rollout finishes, your code assigns a reward to its Trajectory, indicating the performance of the LLM.

2. Training

   1. When each rollout has finished, Trajectories are grouped and sent to the server. Inference is blocked while training executes.
   2. The server trains your model using GRPO, initializing from the latest checkpoint (or an empty LoRA on the first iteration).
   3. The server saves the newly trained LoRA to a local directory and loads it into vLLM.
   4. Inference is unblocked and the loop resumes at step 1.

This training loop runs until a specified number of inference and training iterations have completed.

🧩 Supported Models

ART should work with most vLLM/HuggingFace-transformers compatible causal language models, or at least the ones supported by Unsloth. Gemma 3 does not appear to be supported for the time being. If any other model isn’t working for you, please let us know on Discord or open an issue on GitHub!

🤝 Contributing

ART is in active development, and contributions are most welcome! Please see the CONTRIBUTING.md file for more information.

📖 Citation

@misc{hilton2025art,
  author = {Brad Hilton and Kyle Corbitt and David Corbitt and Saumya Gandhi and Angky William and Bohdan Kovalevskyi and Andie Jones},
  title = {ART: Agent Reinforcement Trainer},
  year = {2025},
  publisher = {GitHub},
  journal = {GitHub repository},
  howpublished = {\url{https://github.com/openpipe/art}}
}

⚖️ License

This repository’s source code is available under the Apache-2.0 License.

🙏 Credits

ART stands on the shoulders of giants. While we owe many of the ideas and early experiments that led to ART’s development to the open source RL community at large, we’re especially grateful to the authors of the following projects:

• Unsloth
• vLLM
• trl
• torchtune

Finally, thank you to our partners who’ve helped us test ART in the wild! We’re excited to see what you all build with it.