SAIL: Self-improving Efficient Online Alignment of Large Language Models

ICLR 2025 Conference Submission12435 Authors

Keywords: RLHF, Alignment, Online Alignment, Self-Play

TL;DR: We introduce SAIL, an efficient online RLHF approach that addresses distribution shift and reduces reliance on preference oracles for improved LLM alignment.

Abstract: Reinforcement Learning from Human Feedback (RLHF) is a critical method for aligning large language models (LLMs) with human preferences. However, existing offline alignment approaches, such as DPO, IPO, and SLiC, rely heavily on static datasets of human preferences, often leading to suboptimal performance. Recent efforts in the literature have moved towards online RLHF methods, but they lack a unified framework and suffer from distribution shift issues. In this work, we formalize online LLM alignment as a bilevel optimization problem. By reducing this formulation to a more computationally efficient single-level first-order method, utilizing reward-policy equivalence, we propose SAIL (Self-improving Efficient Online Alignment).SAIL generates new samples and iteratively refines model alignment through online exploration and regulation of preference labels. This enables continuous, self-improving alignment and generalizes prior online RLHF methods as special cases. Compared to state-of-the-art RLHF methods, SAIL delivers significant performance gains, with up to 11.6\% improvement in win rate and a 3.6-point increase in evaluation rewards, while maintaining low computational overhead.

Primary Area: alignment, fairness, safety, privacy, and societal considerations

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.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2025/AuthorGuide.

Reciprocal Reviewing: I understand the reciprocal reviewing requirement as described on https://iclr.cc/Conferences/2025/CallForPapers. If none of the authors are registered as a reviewer, it may result in a desk rejection at the discretion of the program chairs. To request an exception, please complete this form at https://forms.gle/Huojr6VjkFxiQsUp6.

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: 12435