Debiasing Online Preference Learning via Preference Feature Preservation
Keywords: large language model, alignment, bias, preference
Abstract: While various preferred features determine human preferences, current preference learning frameworks for large language models (LLMs) simplify them with binary pairwise comparisons and scalar rewards. This simplification could make LLMs' responses biased to mostly preferred features such as longer responses which would be exacerbated in online learning scenarios as the biases can be accumulate continuously throughout the iterations.
To address these challenges, we propose a novel framework called PFP (Preference Feature Preservation). The key idea of PFP is maintaining the distribution of human preference features throughout the online preference learning process. Specifically, PFP first trains a feature classifier using the existing offline pairwise human preference data.
Then, using this classifier and the distribution preserving optimization, PFP maps appropriate preference features for each input instruction during online learning.
Lastly, PFP trains LLM using the existing preference learning framework, by incorporating the preference feature of each data into system prompts and enabling LLM to explicitly handle various human preferences. Our experiments demonstrate that PFP successfully mitigates the bias in preference features that arise during online learning, and achieves superior performance compared to previous preference learning methods on general benchmarks including AlpacaEval 2.0 and MT-Bench. We also observe that PFP almost resolves a length bias issue, a long-standing problem of online preference learning, even though it was not specifically designed to tackle this.
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.
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: 12921
Loading