RLAIF: Scaling Reinforcement Learning from Human Feedback with AI Feedback

21 Sept 2023 (modified: 11 Feb 2024)Submitted to ICLR 2024EveryoneRevisionsBibTeX
Primary Area: generative models
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: reinforcement learning, rlhf, rlaif, nlp, large language models, llm, nlp, machine learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2024/AuthorGuide.
TL;DR: We show that RL from AI feedback (RLAIF) can achieves results to RL from human feedback (RLHF)
Abstract: Reinforcement learning from human feedback (RLHF) is an effective technique for aligning large language models (LLMs) to human preferences, but gathering high-quality human preference labels is a critical bottleneck. RL from AI Feedback (RLAIF) is an alternative solution that generates preferences labels using an off-the-shelf LLM in lieu of human annotators. We compare RLAIF and RLHF, and we find that RLAIF achieves improvements on par with RLHF, with both RL policies outperforming the baseline supervised fine-tuning policy by approximately 70\% for summarization and 60\% for helpful dialogue generation, as rated by human evaluators. Furthermore, when asked to rate RLAIF against RLHF in a head-to-head comparison, both are equally preferred. These results suggest that RLAIF can achieve human-level performance, offering a potential solution to the scalability limitations of RLHF.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Submission Number: 3762