NVIDIA Releases Polar — Framework for Training Code Agents

NVIDIA introduced Polar — a new framework for training language agents using reinforcement learning. The key feature is that it works without modifying existing agent harnesses, making it a universal solution for different environments and infrastructures.

How Polar Works

One of the main challenges in training AI agents is the incompatibility between the training pipeline and production harnesses. Often you have to choose: either modify the harness to meet training requirements, or lose access to optimal RL methods. Polar solves this problem elegantly.

The framework acts as an API proxy between the harness and inference server. It captures all token-level interactions and reconstructs from them trajectories that are fully ready for training via GRPO. This allows using advanced training methods directly with existing environments like Codex, Claude Code, and Pi, without changing a single line of their code.

Results on SWE-Bench Verified

NVIDIA researchers tested Polar based on Qwen3.5-4B — a compact model with 4 billion parameters. This is intentionally a small model to demonstrate that the improvement works not only for massive LLMs but also for resource-efficient solutions.

The results are impressive:

• Under Codex harness: +22.6 points on SWE-Bench Verified pass@1
• Under Claude Code harness: +4.8 points
• Under Pi harness: +6.2 points

For context: SWE-Bench Verified is a benchmark that measures how well an agent solves real coding tasks from open pull requests. This is not a synthetic test, but real code. A jump of 22.6 points under the Codex harness is a significant improvement, especially for a compact model.

Integration with NVIDIA Ecosystem

The framework is registered as a NeMo Gym environment, which allows using it within the standard NVIDIA ecosystem. This is an important step because it makes Polar not a one-off tool, but part of a larger platform.

The code is available in the ProRL Agent Server repository under an open license. This means that any developer can take Polar, install it, and train their own model on their own data using their own hardware.

"This demonstrates that efficient agent training does not require

changes to production infrastructure."

What This Means

For developers and companies, this opens a practical way to quickly improve their AI agents without rebuilding their entire infrastructure. NVIDIA demonstrates that even small models can improve significantly with the right training method. This is critical for deployment on edge devices and for overall computational resource savings.