# Staleness-Learning Rate Scaling Laws for Asynchronous RLHF

arXiv: https://arxiv.org/abs/2607.01083 · HTML: https://mlsanity.com/paper/2607.01083/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Reward modeling and RLHF, one of 10 directions in LLM alignment / RLHF / preference optimisation.
- Builds on 45 of its 54 references that are in the map.
- First posted on arXiv 1 Jul 2026.

## Placed in

- LLM alignment / RLHF / preference optimisation / [Reward modeling and RLHF](https://mlsanity.com/alignment/reward-modeling-and-rlhf/)

## Builds on (most cited first, 30 of 45)

- [Training language models to follow instructions with human feedback](https://mlsanity.com/paper/2203.02155/) (2022)
- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [Qwen3 Technical Report](https://mlsanity.com/paper/2505.09388/) (2025)
- [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://mlsanity.com/paper/2501.12948/) (2025)
- [Efficient Memory Management for Large Language Model Serving with PagedAttention](https://mlsanity.com/paper/2309.06180/) (2023)
- [Measuring Mathematical Problem Solving With the MATH Dataset](https://mlsanity.com/paper/2103.03874/) (2021)
- [Let's Verify Step by Step](https://mlsanity.com/paper/2305.20050/) (2023)
- [DAPO: An Open-Source LLM Reinforcement Learning System at Scale](https://mlsanity.com/paper/2503.14476/) (2025)
- [Asynchronous Methods for Deep Reinforcement Learning](https://mlsanity.com/paper/1602.01783/) (2016)
- [HybridFlow: A Flexible and Efficient RLHF Framework](https://mlsanity.com/paper/2409.19256/) (2024)
- [Solving Quantitative Reasoning Problems with Language Models](https://mlsanity.com/paper/2206.14858/) (2022)
- [Kimi k1.5: Scaling Reinforcement Learning with LLMs](https://mlsanity.com/paper/2501.12599/) (2025)
- [Tulu 3: Pushing Frontiers in Open Language Model Post-Training](https://mlsanity.com/paper/2411.15124/) (2024)
- [IMPALA: Scalable Distributed Deep-RL with Importance Weighted Actor-Learner Architectures](https://mlsanity.com/paper/1802.01561/) (2018)
- [signSGD: Compressed Optimisation for Non-Convex Problems](https://mlsanity.com/paper/1802.04434/) (2018)
- [Gradient Descent on Neural Networks Typically Occurs at the Edge of Stability](https://mlsanity.com/paper/2103.00065/) (2021)
- [DD-PPO: Learning Near-Perfect PointGoal Navigators from 2.5 Billion Frames](https://mlsanity.com/paper/1911.00357/) (2019)
- [Safe and Efficient Off-Policy Reinforcement Learning](https://mlsanity.com/paper/1606.02647/) (2016)
- [AReaL: A Large-Scale Asynchronous Reinforcement Learning System for Language Reasoning](https://mlsanity.com/paper/2505.24298/) (2025)
- [Small-scale proxies for large-scale Transformer training instabilities](https://mlsanity.com/paper/2309.14322/) (2023)
- [Dissecting Adam: The Sign, Magnitude and Variance of Stochastic Gradients](https://mlsanity.com/paper/1705.07774/) (2017)
- [Why Transformers Need Adam: A Hessian Perspective](https://mlsanity.com/paper/2402.16788/) (2024)
- [Asynchronous Parallel Stochastic Gradient for Nonconvex Optimization](https://mlsanity.com/paper/1506.08272/) (2015)
- [Noise Is Not the Main Factor Behind the Gap Between SGD and Adam on Transformers, but Sign Descent Might Be](https://mlsanity.com/paper/2304.13960/) (2023)
- [On the SDEs and Scaling Rules for Adaptive Gradient Algorithms](https://mlsanity.com/paper/2205.10287/) (2022)
- [The Art of Scaling Reinforcement Learning Compute for LLMs](https://mlsanity.com/paper/2510.13786/) (2025)
- [Sharper Convergence Guarantees for Asynchronous SGD for Distributed and Federated Learning](https://mlsanity.com/paper/2206.08307/) (2022)
- [Slow and Stale Gradients Can Win the Race: Error-Runtime Trade-offs in Distributed SGD](https://mlsanity.com/paper/1803.01113/) (2018)
- [Asynchronous SGD Beats Minibatch SGD Under Arbitrary Delays](https://mlsanity.com/paper/2206.07638/) (2022)
- [StreamRL: Scalable, Heterogeneous, and Elastic RL for LLMs with Disaggregated Stream Generation](https://mlsanity.com/paper/2504.15930/) (2025)

## Built on by (most cited first, 2 of 2)

- [Probe the Harness: Setup Checks for Stale-Data RL Comparisons in Language Models](https://mlsanity.com/paper/2610.02911/) (2026)
- [Two Clocks: Why Larger Rollout Batches Learn Slower but Fail on Schedule in Reinforcement Learning with Verifiable Rewards](https://mlsanity.com/paper/2610.18853/) (2026)

## Abstract (arXiv)

Asynchronous reinforcement learning improves the throughput of large language model post-training by decoupling rollout generation from policy optimization, but introduces a mismatch between the behavior and learner policies. How the resulting policy staleness couples with the learning rate to govern training stability and collapse time remains poorly understood. We investigate this coupling in vanilla GRPO through controlled sweeps of the synchronization interval S and constant learning rate η on Llama-3.2-1B/3B, complemented by experiments on Qwen3-8B. We identify two empirical scaling laws: (i) Stability-boundary scaling: the largest stable learning rate scales approximately as S-1, yielding a stability boundary characterized by an approximately constant product Sη. (ii) Collapse-time scaling: among collapsing runs, estimated collapse times scale approximately as η-1, corresponding to a model- and setup-dependent cumulative learning-rate budget that aligns across synchronization intervals in the Llama sweeps. We interpret these laws through a local analysis of the behavior-dependent GRPO surrogate and a complementary mean-field model. Under local regularity conditions, the analysis yields an O(Sη) upper bound on the staleness-induced update bias that resets at synchronization. The mean-field model shows how sufficiently strong positive feedback can sustain directional drift when update directions persist across synchronization cycles. When drift speed saturates under optimizer normalization, this mechanism predicts exit from a local surrogate-validity region after an approximately fixed cumulative learning rate. Together, these findings motivate a practical calibration rule: estimate the stability threshold and collapse budget from a coarse sweep, then jointly select S and η for the intended training horizon.