arXivDaily arXiv每日学术速递 周一至周五更新
arXiv 2402.14740cs.LG
收藏

回归基础:重新审视用于大语言模型人类反馈学习的 REINFORCE 风格优化

Back to Basics: Revisiting REINFORCE Style Optimization for Learning from Human Feedback in LLMs

Cohere For AI(Cohere 人工智能研究院);Cohere

Arash Ahmadian, Chris Cremer, Matthias Gallé, Marzieh Fadaee, Julia Kreutzer, Olivier Pietquin, Ahmet Üstün, Sara Hooker

AI总结:

本文重新审视 RLHF 中的优化方法,证明简单的 REINFORCE 风格变体可在低成本下超越 PPO 及 DPO、RAFT 等免强化学习方法。

摘要:

以基于人类反馈的强化学习(RLHF)形式进行的 AI 对齐,越来越被视为高性能大语言模型的关键要素。近期的文献将近端策略优化(PPO)定位为 RLHF 中强化学习部分的标准方法。然而,PPO 既涉及高昂的计算成本,又需要敏感的超参数调优。我们认为,导致 PPO 发展起来的大部分动机性原则在 RLHF 中实际关切较小,并主张采用一种计算开销更低、同时能保持甚至提升性能的方法。我们在强化学习的背景下重新审视了基于人类偏好的对齐问题表述。以简洁性作为指导原则,我们表明 PPO 的许多组件在 RLHF 场景中是不必要的,并且远更简单的 REINFORCE 风格优化变体在性能上优于 PPO 以及新提出的 DPO 和 RAFT 等“免强化学习”方法。我们的工作表明,针对大语言模型对齐特性的细致适配,能够以低成本从在线强化学习优化中获益。

英文摘要:

AI alignment in the shape of Reinforcement Learning from Human Feedback (RLHF) is increasingly treated as a crucial ingredient for high performance large language models. Proximal Policy Optimization (PPO) has been positioned by recent literature as the canonical method for the RL part of RLHF. However, it involves both high computational cost and sensitive hyperparameter tuning. We posit that most of the motivational principles that led to the development of PPO are less of a practical concern in RLHF and advocate for a less computationally expensive method that preserves and even increases performance. We revisit the formulation of alignment from human preferences in the context of RL. Keeping simplicity as a guiding principle, we show that many components of PPO are unnecessary in an RLHF context and that far simpler REINFORCE-style optimization variants outperform both PPO and newly proposed "RL-free" methods such as DPO and RAFT. Our work suggests that careful adaptation to LLMs alignment characteristics enables benefiting from online RL optimization at low cost.

机构、总结、中文摘要均由AI分析而来,可能存在错误,请以论文正文为准。

↑