Research

Paper

AI LLM February 19, 2026

Stable Asynchrony: Variance-Controlled Off-Policy RL for LLMs

Authors

Luke Huang, Zhuoyang Zhang, Qinghao Hu, Shang Yang, Song Han

Abstract

Reinforcement learning (RL) is widely used to improve large language models on reasoning tasks, and asynchronous RL training is attractive because it increases end-to-end throughput. However, for widely adopted critic-free policy-gradient methods such as REINFORCE and GRPO, high asynchrony makes the policy-gradient estimator markedly $\textbf{higher variance}$: training on stale rollouts creates heavy-tailed importance ratios, causing a small fraction of samples to dominate updates. This amplification makes gradients noisy and learning unstable relative to matched on-policy training. Across math and general reasoning benchmarks, we find collapse is reliably predicted by effective sample size (ESS) and unstable gradient norms. Motivated by this diagnosis, we propose $\textbf{V}$ariance $\textbf{C}$ontrolled $\textbf{P}$olicy $\textbf{O}$ptimization ($\textbf{VCPO}$), a general stabilization method for REINFORCE/GRPO-style algorithms that (i) scales learning rate based on effective sample size to dampen unreliable updates, and (ii) applies a closed-form minimum-variance baseline for the off-policy setting, avoiding an auxiliary value model and adding minimal overhead. Empirically, VCPO substantially improves robustness for asynchronous training across math, general reasoning, and tool-use tasks, outperforming a broad suite of baselines spanning masking/clipping stabilizers and algorithmic variants. This reduces long-context, multi-turn training time by 2.5$\times$ while matching synchronous performance, demonstrating that explicit control of policy-gradient variance is key for reliable asynchronous RL at scale.

Metadata

arXiv ID: 2602.17616
Provider: ARXIV
Primary Category: cs.LG
Published: 2026-02-19
Fetched: 2026-02-21 18:51

Related papers

Raw Data (Debug)
{
  "raw_xml": "<entry>\n    <id>http://arxiv.org/abs/2602.17616v1</id>\n    <title>Stable Asynchrony: Variance-Controlled Off-Policy RL for LLMs</title>\n    <updated>2026-02-19T18:40:51Z</updated>\n    <link href='https://arxiv.org/abs/2602.17616v1' rel='alternate' type='text/html'/>\n    <link href='https://arxiv.org/pdf/2602.17616v1' rel='related' title='pdf' type='application/pdf'/>\n    <summary>Reinforcement learning (RL) is widely used to improve large language models on reasoning tasks, and asynchronous RL training is attractive because it increases end-to-end throughput. However, for widely adopted critic-free policy-gradient methods such as REINFORCE and GRPO, high asynchrony makes the policy-gradient estimator markedly $\\textbf{higher variance}$: training on stale rollouts creates heavy-tailed importance ratios, causing a small fraction of samples to dominate updates. This amplification makes gradients noisy and learning unstable relative to matched on-policy training. Across math and general reasoning benchmarks, we find collapse is reliably predicted by effective sample size (ESS) and unstable gradient norms. Motivated by this diagnosis, we propose $\\textbf{V}$ariance $\\textbf{C}$ontrolled $\\textbf{P}$olicy $\\textbf{O}$ptimization ($\\textbf{VCPO}$), a general stabilization method for REINFORCE/GRPO-style algorithms that (i) scales learning rate based on effective sample size to dampen unreliable updates, and (ii) applies a closed-form minimum-variance baseline for the off-policy setting, avoiding an auxiliary value model and adding minimal overhead. Empirically, VCPO substantially improves robustness for asynchronous training across math, general reasoning, and tool-use tasks, outperforming a broad suite of baselines spanning masking/clipping stabilizers and algorithmic variants. This reduces long-context, multi-turn training time by 2.5$\\times$ while matching synchronous performance, demonstrating that explicit control of policy-gradient variance is key for reliable asynchronous RL at scale.</summary>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.LG'/>\n    <category scheme='http://arxiv.org/schemas/atom' term='cs.AI'/>\n    <published>2026-02-19T18:40:51Z</published>\n    <arxiv:primary_category term='cs.LG'/>\n    <author>\n      <name>Luke Huang</name>\n    </author>\n    <author>\n      <name>Zhuoyang Zhang</name>\n    </author>\n    <author>\n      <name>Qinghao Hu</name>\n    </author>\n    <author>\n      <name>Shang Yang</name>\n    </author>\n    <author>\n      <name>Song Han</name>\n    </author>\n  </entry>"
}