NVIDIA releases Polar, a framework for training coding agents

NVIDIA has introduced Polar, a new framework for training language agents with reinforcement learning. Its key feature is that it works without modifying existing agent harnesses, which makes it a universal solution for different environments and infrastructures.

How Polar works

One of the main difficulties in training AI agents is the incompatibility between the training pipeline and production harnesses. Teams often have to choose: either change the harness to fit training requirements, or lose access to the best RL methods. Polar solves this problem elegantly.

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

Results on SWE-Bench Verified

NVIDIA researchers tested Polar on Qwen3.5-4B, a compact model with 4 billion parameters. The model is deliberately small, to show that the improvement works not only for giant LLMs but also for resource-efficient solutions. The results are impressive:

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

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

Integration with the NVIDIA ecosystem

The framework is registered as a NeMo Gym environment, so it can be used within NVIDIA's standard ecosystem. This matters because it makes Polar part of a larger platform rather than a one-off tool.

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

"This shows that effective agent training does not require changing

production infrastructure."

What this means

For developers and companies, this opens a practical path to quickly improving their AI agents without rebuilding the whole infrastructure. NVIDIA shows that even small models can improve substantially with the right training method. That is critical for deployment on edge devices and for saving compute resources overall.