English

Asymmetric REINFORCE for off-Policy Reinforcement Learning: Balancing positive and negative rewards

Machine Learning 2025-12-01 v2 Computation and Language

Abstract

Reinforcement learning (RL) is increasingly used to align large language models (LLMs). Off-policy methods offer greater implementation simplicity and data efficiency than on-policy techniques, but often result in suboptimal performance. In this work, we study the intermediate range of algorithms between off-policy RL and supervised fine-tuning by analyzing a simple off-policy REINFORCE algorithm, where the advantage is defined as A=rVA=r-V, with rr a reward and VV some tunable baseline. Intuitively, lowering VV emphasizes high-reward samples, while raising it penalizes low-reward ones more heavily. We first provide a theoretical analysis of this off-policy REINFORCE algorithm, showing that when the baseline VV lower-bounds the expected reward, the algorithm enjoys a policy improvement guarantee. Our analysis reveals that while on-policy updates can safely leverage both positive and negative signals, off-policy updates benefit from focusing more on positive rewards than on negative ones. We validate our findings experimentally in a controlled stochastic bandit setting and through fine-tuning state-of-the-art LLMs on reasoning tasks.

Keywords

Cite

@article{arxiv.2506.20520,
  title  = {Asymmetric REINFORCE for off-Policy Reinforcement Learning: Balancing positive and negative rewards},
  author = {Charles Arnal and Gaëtan Narozniak and Vivien Cabannes and Yunhao Tang and Julia Kempe and Remi Munos},
  journal= {arXiv preprint arXiv:2506.20520},
  year   = {2025}
}