Parameter Efficient Reinforcement Learning from Human Feedback

TL;DR

This paper demonstrates that parameter-efficient RLHF using LoRA achieves similar alignment performance to traditional RLHF but with significantly reduced training time and memory usage, promoting wider adoption.

Contribution

It empirically evaluates PE-RLHF with LoRA, showing substantial efficiency gains while maintaining effectiveness across diverse tasks.

Findings

PE-RLHF achieves comparable performance to RLHF.

Training time reduced by up to 90%.

Memory footprint reduced by up to 50%.

Abstract

While Reinforcement Learning from Human Feedback (RLHF) effectively aligns pretrained Large Language and Vision-Language Models (LLMs, and VLMs) with human preferences, its computational cost and complexity hamper its wider adoption. To alleviate some of the computational burden of fine-tuning, parameter efficient methods, like LoRA were introduced. In this work, we empirically evaluate the setup of Parameter Efficient Reinforcement Learning from Human Feedback (PE-RLHF) that leverages LoRA fine-tuning for Reward Modeling, and Reinforcement Learning. We benchmark the PE-RLHF setup on six diverse datasets spanning summarization, harmless/helpful response generation, UI automation, and visual question answering in terms of effectiveness of the trained models, and the training resources required. Our findings show, for the first time, that PE-RLHF achieves comparable performance to RLHF, while significantly reducing training time (up to 90% faster for reward models, and 30% faster for RL), and memory footprint (up to 50% reduction for reward models, and 27% for RL). We provide comprehensive ablations across LoRA ranks, and model sizes for both reward modeling and reinforcement learning. By mitigating the computational burden associated with RLHF, we push for a broader adoption of PE-RLHF as an alignment technique for LLMs and VLMs.