Learning Loss Landscapes in Preference Optimization
Keywords: Preference optimization, mirror descent
Abstract: We present a framework to discover preference optimization algorithms specialized to particular scenarios, in a theoretically sound and computationally efficient setting. We start by designing a novel family of PO algorithms based on mirror descent, which we call Mirror Preference Optimization (MPO). MPO recovers existing methods like Direct Preference Optimization (DPO) and Odds-Ratio Preference Optimization (ORPO) for specific choices of the mirror map. Given specific properties of preference datasets, such as mixed-quality or noisy data, we show that we can efficiently search the MPO class to find specialized algorithms that outperform current baselines. Namely, we leverage evolutionary strategies and preference datasets generated on MuJoCo environments to systematically evaluate and optimize MPO algorithms on hand-crafted scenarios. We demonstrate the resulting PO algorithms successfully transfer to a Large Language Model (LLM) alignment task, where they demonstrate superior robustness in handling mixed-quality datasets.
Primary Area: reinforcement learning
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: 2402
Loading