NVIDIA 推出 Polar：專為 Codex、Claude Code 與 Qwen Code 設計的 GRPO 訓練保留 Token 推出框架

Reinforcement learning for language agents is growing more complex. Agents now manage multi-turn tool use, long-running contexts, and multi-agent orchestration. The main engineering challenge is connecting existing agent software to training pipelines without breaking how those tools work. NVIDIA’s research team introduced Polar, a rollout framework that lets researchers run reinforcement learning over any agent harness without modifying that harness. The Core Problem Polar Solves An ‘agent harness’ is a tool like Codex CLI, Claude Code, Qwen Code, or Pi. These harnesses manage system prompts, tool formatting, context engineering, and how the agent submits patches. These details directly affect agent behavior at evaluation time. Traditional RL infrastructure requires harness logic to be rewritten behind a framework-owned environment API — typically env.init(), env.step(), env.reset() in the OpenAI Gym style. Every new harness requires new integration code. That integration can also lose execution details specific to the native harness path. Polar’s key observation is that every LLM-based agent must call a model. That model API boundary is a common interface outside the agent itself. Instead of integrating inside the harness, Polar places a proxy at that boundary. How the Proxy Works For each incoming model request, the gateway proxy performs four steps: Detect the provider API — using the request path and headers, it distinguishes Anthropic Messages, OpenAI Chat Completions, OpenAI Responses, and Google generateContent-style calls. Normalize the request — converts roles, content parts, tool definitions, and generation parameters into the OpenAI Chat Completions shape used by the local inference server. Capture token-level data — stores request messages, response messages, prompt token IDs, sampled response token IDs, finish reason, and log probabilities. Return the provider shape — transforms the response back into the schema the harness expects. For streaming requests, Polar obtains a non-streaming upstream response and emits a synthetic provider-shaped stream. This preserves compatibility with harnesses that expect server-sent events while ensuring complete token capture. The only required change to an existing harness is pointing its model base URL at the gateway. https://arxiv.org/pdf/2605.24220 Architecture: Rollout Server and Gateway Nodes Polar has two core components: The rollout server accepts a TaskRequest and expands it into num_samples independent sessions. Each session carries a session ID, task ID, timeout budget, runtime specification, agent specification, trajectory builder, evaluator, and callback URL. The server dispatches sessions to gateway nodes and accepts callbacks when sessions complete. Gateway nodes own the lifecycle of each session — starting the runtime, running the harness, building trajectories, evaluating output, and teardown. The gateway also hosts the proxy endpoint for that session’s model calls, keeping completion capture tied to the session registry. Within each gateway, isolated worker pools handle INIT, RUNNING, and POSTRUN stages. A bounded READY buffer holds initialized runtimes until a run slot is available. CPU-heavy runtime preparation and evaluator prewarm proceed off the critical path, without blocking active GPU-bound agent execution. If a harness times out after model calls have been captured, the gateway still enters POSTRUN so partial traces can be recovered. Built-in evaluators include a session-completion reward, a configurable test-on-output evaluator, and a SWE-Bench/SWE-Gym harness evaluator. Custom evaluators can be added through a registry interface. Polar currently supports Docker and rootless Apptainer runtimes. Built-in harness shortcuts include codex, claude_code, gemini_cli, qwen_code, opencode, and pi. Trajectory Reconstruction: Per Request vs. Prefix Merging After a session completes, Polar reconstructs trainable trajectories from captured model calls. Two strategies are available: The per_request builder treats every model call as one independent trace. It is lossless per individual call but fragments multi-turn sessions. A single coding problem can produce hundreds of per-request traces, increasing the burden on downstream trainers. The prefix_merging builder reconstructs longer traces where the harness session preserves append-only conversation histories. It partitions completions into ordered chains by verifying a strict token-prefix relation between adjacent completions. Sub-agents, context compaction boundaries, and parallel agent branches naturally form separate chains. Within each merged trace, only sampled assistant tokens are marked trainable. Canonical interstitial tokens receive a loss mask of zero. Ablation Results The research team benchmarks both strategies on the same model, hardware, and topology over three training steps. Metricper_requestprefix_mergingTrainer updates1,185218Wall-clock time189.5 min35.2 minSpeedup—5.39×Avg. rollout GPU utilization20.4%87.7% SWE-Bench Verified Results Training uses standard GRPO on the Qwen3.5-4B base model. The dataset is SkyRL-v0-293-data SWE-Gym (293 tasks, 1 epoch, rollout batch size 4, 16 samples per prompt) with the Slime trainer. All experiments use prefix_merging for trajectory construction. Training Rollout Reward Progress (pass@1) HarnessFirst 10 StepsLast 10 StepsCodex9.5%54.5%Claude Code28.8%67.0%Qwen Code61.6%66.0%Pi61.6%76.2% SWE-Bench Verified Final Scores HarnessBasePolar RLGainCodex3.8%26.4%+22.6 ptsClaude Code29.8%34.6%+4.8 ptsQwen Code34.6%35.2%+0.6 ptsPi34.2%40.4%+6.2 pts The largest gain is under Codex. Codex presents an unfamiliar action protocol and patch-submission style to a Qwen model not originally trained on that harness. Polar attaches the reward signal to the actual sampled tokens flowing through the Codex execution path, so GRPO optimizes the behavior the model uses at evaluation time. Under the native Qwen Code harness, where the base model is already well-aligned, Polar still delivers a 0.6 point gain. Offline SFT Data Generation Polar can also serve as a distributed offline data generation service with no changes to the runtime. The research team demonstrates this using Qwen3.5-122B-A10B on an 8×H100 server (TP=8, max_model_len=32,768) with the pi harness against 1,638 instances from seven SWE-Gym repositories. A trajectory is accepted into the SFT corpus only if the SWE-Bench evaluation harness confirms the agent’s patch resolves every FAIL_TO_PASS test and leaves every PASS_TO_PASS test green. RepositoryAttemptsAcceptedRategetmoto/moto34318453.6%python/mypy25710139.3%conan-io/conan712738.0%pydantic/pydantic812429.6%iterative/dvc2194520.5%pandas-dev/pandas4779819.7%dask/dask1412517.7%Total1,63850430.8% The run cost roughly 64 GPU-hours. Accepted trajectories average 104 messages per session and 51 assistant turns. Framework Comparison SystemAsync RLAsync Rollout StagingRollout as ServiceHarness AgnosticPolar✓✓✓✓ProRL Agent✓✓✓✗SkyRL-Agent✓✓✗partialPRIME-RL✓✗✗✗Agent Lightningpartial✗partialpartialrLLMpartial✗✗✗OpenClaw-RL✓✗✗partial Polar is the only system in this comparison with first-class support across all four properties. Strengths and Limitations Strengths No harness code changes required — the proxy intercepts at the model API boundary Provider-agnostic: supports Anthropic, OpenAI Chat, OpenAI Responses, and Google API formats natively prefix_merging reduces trainer updates from 1,185 to 218 and cuts wall-clock time 5.39× Works for both online RL and offline SFT data generation with the same runtime Harness-native RL delivers large gains for unfamiliar execution paths — 22.6 pts on Codex Partial traces are recovered when a harness times out mid-session Released as open source under NeMo Gym Limitations Reward design, evaluator quality, and distribution shift remain the researcher’s responsibility Requires the harness to support a configurable model base URL Token-level capture depends on the servi