强化学习NeurIPS 20252025

REINFORCE++: Improved Policy Gradient Methods

Kasun Amarasinghe, Kit Rodolfa, Hemank Lamba, Rayid Ghani et al.DeepMind

摘要

We present REINFORCE++, which revisits and significantly improves the classic REINFORCE algorithm for modern deep learning settings. REINFORCE++ introduces variance reduction techniques, adaptive baselines, and efficient gradient estimation that make policy gradient methods competitive with actor-critic approaches.

policy gradientREINFORCEvariance reductionadaptive baselinesgradient estimation

技术细节

数据集
HH-RLHFAtariDMControl
模型骨架

LLaMA-2 7B + REINFORCE++ Variant

编码器

Transformer Encoder (LLaMA-2)

解码器

REINFORCE++ Policy Decoder (32 layers)

京ICP备2026064258号-1