EUVIMEDEUVIMED entdecken
Crossref

Accelerating RLHF Training with Reward Variance Increase

Zonglin Yang, Zhexuan Gu, Houduo Qi, Yancheng Yuan

SIAM Journal on Mathematics of Data Science · 31.12.2026 · 8 (4) · 1393-1416

DOI: 10.1137/25m1764281 ↗

Kurz gefasst

Fachartikel in der Zeitschrift SIAM Journal on Mathematics of Data Science, online erschienen am 16.11.2026. Verfasst von Zonglin Yang und 3 weiteren Autor:innen. Die Arbeit nennt 49 Literaturangaben. Gefördert von: Hong Kong Research Grants Council, Hong Kong Polytechnic University.

AbstractBibliografische AngabenZitierenVerwandte Studien

Abstract (Originalsprache)

Reinforcement learning from human feedback (RLHF) is an essential technique for ensuring that large language models (LLMs) are aligned with human values and preferences during the posttraining phase. As an effective RLHF approach, group relative policy optimization (GRPO) has demonstrated success in many LLM-based applications. However, efficient GRPO-based RLHF training remains a challenge. Recent studies reveal that a higher reward variance of the initial policy model leads to faster RLHF training. Inspired by this finding, we propose a practical reward adjustment model to accelerate RLHF training by provably increasing the reward variance and preserving the relative preferences and reward expectation. Our reward adjustment method inherently poses a nonconvex optimization problem, which is NP-hard to solve in general. To overcome the computational challenges, we design a novel [Formula: see text] algorithm, where the complexity is dominated by sorting rewards, to find a global solution of the nonconvex reward adjustment model by explicitly characterizing the extreme points of the feasible set. As an important application, we naturally integrate this reward adjustment model into the GRPO algorithm, leading to a more efficient GRPO with reward variance increase (GRPOVI) algorithm for RLHF training. As an interesting byproduct, we provide an indirect explanation for the empirical effectiveness of GRPO with rule-based reward for RLHF training, as demonstrated in DeepSeek-R1. Experiment results demonstrate that the GRPOVI algorithm can significantly improve the RLHF training efficiency compared to the original GRPO algorithm.

Förderung

  • Hong Kong Research Grants Council
  • Hong Kong Polytechnic University

Zitieren

APA

Yang, Z., Gu, Z., Qi, H., & Yuan, Y. (2026). Accelerating RLHF Training with Reward Variance Increase. SIAM Journal on Mathematics of Data Science, 8(4), 1393-1416. https://doi.org/10.1137/25m1764281

Vancouver

Yang Z, Gu Z, Qi H, Yuan Y. Accelerating RLHF Training with Reward Variance Increase. SIAM Journal on Mathematics of Data Science. 2026;8(4):1393-1416. doi:10.1137/25m1764281

BibTeX
@article{Yang2026,
  author = {Yang, Zonglin and Gu, Zhexuan and Qi, Houduo and Yuan, Yancheng},
  title = {Accelerating RLHF Training with Reward Variance Increase},
  journal = {SIAM Journal on Mathematics of Data Science},
  year = {2026},
  volume = {8},
  number = {4},
  pages = {1393--1416},
  doi = {10.1137/25m1764281}
}
Kostenlos starten →