Aligning Large Language Models with Domain Adaptation

David Wu, Sanjiban Choudhury

Keywords: alignment, domain adaptation, large language models, generalization

TL;DR: We propose Data Efficient Alignment for Language, using domain adaptation to perform cross-task alignment in scenarios where target labels are scarce, including across languages, noise resistance, low-data regimes, and easy-to-hard generalization.

Abstract: Aligning large language models (LLMs) has emerged as a critical challenge in the age of generative AI: LLMs must be appropriately aligned with human values and preferences in order to be helpful and harmless. In many real world cases, however, large amounts of preference data are not available on important tasks, limiting the effectiveness of resulting reward models. In some cases, data from a similar task is available, and unlabeled data on the target task is available or can be generated by an LLM. In other cases, clean data may be available to train an LLM for real-world use on noisy data, small amounts of labeled data on the target task may be available, or data may be available on an easier task. In this work, we demonstrate that domain adaptation can effectively use different types of data, by transferring supervision and human values across tasks with similar data distributions, strengthening resistance to noisy data, improving few-shot generalization ability, and even transfer from easy to hard tasks, in the form of short to long generalization. Specifically, we propose Data Efficient Alignment for Language (DEAL), using domain adaptation to effectively perform cross-task alignment in scenarios where labeled target data is not available. We evaluate our method for reward model training on a variety of benchmarks and demonstrate that our method can meaningfully improve performance on target tasks by utilizing data on related tasks or low amounts of data. Furthermore, we offer analysis on the inner mechanism of domain adaptation and the alignment of embedding distributions.

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: 5567