On May 28th, NVIDIA Polar announced a 594% surge in training performance for Codex

On May 28, NVIDIA officially open-sourced Polar, a brand-new reinforcement learning training framework highlighted by its zero-modification integration capability. It enables mainstream code agents including Codex, Claude Code and Qwen Code to connect directly to GRPO (Generalized Relative Policy Optimization) reinforcement learning training without any alterations to their native source code.

Industry Pain Points: The High Barrier Blocking Agent Reinforcement Learning

Modern code agents are evolving from single-step tasks to long, complex workflows such as repository-wide code edits and operating system interactions, which heavily rely on mature execution frameworks. However, traditional reinforcement learning integration suffers from three major flaws:

1. Exorbitant integration costs: Developers are forced to rewrite core code logic to conform to standardized interfaces like env.init() and env.step(), creating extremely cumbersome workflows.
2. Loss of critical information: Refactoring often erases tool invocation details, multi-turn dialogue context and sub-agent collaboration logic, resulting in low-quality training signals.
3. Fragmented ecosystems: Disparate agent frameworks cannot interoperate, making training resources impossible to reuse across platforms.

Core Breakthrough: Turn API Boundaries Into Training Entry Points

Instead of revamping underlying execution frameworks, Polar implements a transparent proxy layer at the model API level. Its core architecture is designed as follows:

• Black-box compatibility: A Gateway is deployed between execution frameworks and inference servers to intercept and forward all requests, supporting any standard API from Anthropic, OpenAI, Google and other vendors.
• Automatic trajectory reconstruction: While forwarding traffic, the framework real-time records prompts, sampled tokens, log probabilities and other metrics to auto-generate trajectory data required for reinforcement learning.
• Highly asynchronous architecture: A Rollout Server handles task scheduling and data persistence, while Gateway Nodes manage request lifecycles. Preheating buffer pools and parallel processing completely eliminate GPU training bottlenecks.

Stunning Performance: Dramatic Gains in Both Training Efficiency & Model Effectiveness

Benchmark tests on the SWE-Bench Verified dataset using the Qwen3.5-4B model demonstrate transformative improvements when combining Polar with GRPO:

• Codex framework: pass@1 jumps from 3.8% to 26.4%, a 594.74% increase
• Claude Code framework: rises from 29.8% to 34.6%
• Pi framework: climbs from 34.2% to 40.4%

Training efficiency also sees massive upgrades. With the prefix_merging strategy enabled, total training duration is cut by a factor of 5.39, and GPU utilization surges from 20.4% up to 87.7%.

Industry Implications: AI Agent Training Enters the Industrial Engineering Era

NVIDIA’s open-source release of Polar is equivalent to building a dedicated high-speed highway for reinforcement learning pipelines for AI agents:

1. Researchers can reuse countless open-source code frameworks with zero extra overhead to launch GRPO training.
2. The barrier to entry for GPU computing power is drastically lowered, allowing small and mid-sized teams to train state-of-the-art code agents.
3. It standardizes and scales the evolution roadmap of AI coding agents, shifting development from manual lab tuning to industrialized mass production.

Moving forward, developers will no longer waste engineering hours adapting models to training frameworks, enabling faster, more efficient iteration cycles for all code-based AI agents.