Stale but Stable: Staleness-Adaptive Trust Regions for Stabilizing Asynchronous Reinforcement Learning

2026-07-21

AI Summary

This work addresses the challenge of stale samples in asynchronous reinforcement learning, where policy lag, engine latency, and expert routing heterogeneity lead to high policy mismatch that undermines existing trust region methods. The paper proposes Staleness-Adaptive Trust Region (SAT), which uniquely integrates trust region mechanisms with staleness awareness by employing a detached log-probability ratio as a staleness proxy to identify high-mismatch tail samples within a batch. SAT selectively contracts only the sign-chosen endpoint of the PPO clipping interval, imposing conservative constraints on highly stale samples while preserving standard updates for others. Combined with kernel scaling and sign-based selection, SAT achieves local interval containment and pointwise pessimism, reshaping update geometry under heterogeneous staleness. Evaluated on Qwen3-30B with SGLang and Megatron, SAT-GSPO w/ R3 attains state-of-the-art performance on AIME24 avg@8 (35.83 at lag 1, 34.79 at lag 8), significantly enhancing asynchronous training stability.

Abstract

Asynchronous reinforcement learning improves throughput by decoupling rollout generation from optimization, but staleness is an inevitable byproduct compounded by policy lag, engine delays, and mixture-of-experts routing. From a trust-region perspective, this mismatch is critical: training-inference divergence governs approximation error in finite-horizon bounds, whereas PPO clipping only gates sampled outward updates, acting as a sampled surrogate rather than a full-policy constraint. As a result, high-staleness updates remain weakly controlled in the asynchronous regime where stale rollouts matter most. We introduce the Staleness-Adaptive Trust Region (SAT), which uses the detached sampled log-ratio as a practical staleness proxy, identifies high-mismatch tails within each batch via staleness-based kernel scaling, and contracts only the sign-selected endpoint of the nominal PPO interval. This preserves baseline behavior on ordinary tokens while enforcing more conservative updates on newly intercepted outward bands. We prove local interval containment and pointwise pessimism relative to PPO, showing how the adaptive rule reshapes update geometry under heterogeneous staleness. We evaluate SAT in a decoupled asynchronous RL setup built on Qwen3-30B-A3B-Base, using SGLang as the inference engine and Megatron for training. In this setting, SAT-GSPO w/ R3 achieves the best observed AIME24 avg@8, reaching 35.83 at lag 1 and 34.79 at lag 8, while SAT-GSPO reaches 34.17 at lag 1. Adaptive clipping and routing replay act as complementary stabilizers targeting mismatch tails and routing inconsistency, respectively. Overall, aligning clip intervals with staleness heterogeneity effectively stabilizes asynchronous RL.

Problem

• asynchronous reinforcement learning
• staleness
• trust region
• policy lag
• training-inference divergence

Innovation

• Staleness-Adaptive Trust Region
• Asynchronous Reinforcement Learning
• Policy Staleness
• Adaptive Clipping
• Trust Region Optimization