Active Preference Optimization via Maximizing Learning Capacity

Seola Choi, Dongmin Park, Jaehyun Park, Jae-Gil Lee

Keywords: active learning, preference learning, preference optimization

TL;DR: Propose a novel active learning method for fine-tuning LLMs with preference feedback

Abstract: The success of deep learning in various complex tasks relies heavily on large amounts of annotated data, which can be prohibitively expensive to acquire. Techniques such as reinforcement learning with human feedback (RLHF) and direct preference optimization (DPO) have emerged as methods for fine-tuning models by leveraging human preferences, but they come with significant costs, especially when applied to large-scale language models (LLMs). Recent efforts to reduce these costs have focused on active preference optimization, which uses certainty-based selection to minimize the annotation burden. However, the two-step process of selecting uncertain input prompts and then acquiring completions can lead to sub-optimal pairings, potentially limiting model learning capacity. This paper suggests that divAPO eliminates suboptimal pairings that are typical of two-step methods and enhances learning capacity by selecting the most informative preference pairs in a single phase, taking into account both data distribution probabilities and preference model certainty. Through experiments on complicated Language tasks, we demonstrate that our method achieves significant performance improvements over existing approaches.

Primary Area: foundation or frontier models, including LLMs

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: 1795