An Empirical Study on Preference Tuning Generalization and Diversity Under Domain Shift

TL;DR

This paper systematically investigates how preference tuning of language models generalizes across domains, comparing different objectives and adaptation strategies, and finds pseudo-labeling effectively mitigates domain shift effects.

Contribution

It provides a comprehensive analysis of alignment generalization under domain shift and evaluates adaptation strategies, highlighting pseudo-labeling's effectiveness.

Findings

Preference-tuning performance degrades under domain shift.

Different alignment objectives show systematic variation in generalization.

Pseudo-labeling strategies significantly reduce domain-shift degradation.

Abstract

Preference tuning aligns pretrained language models to human judgments of quality, helpfulness, or safety by optimizing over explicit preference signals rather than likelihood alone. Prior work has shown that preference-tuning degrades performance and reduces helpfulness when evaluated outside the training domain. However, the extent to which adaptation strategies mitigate this domain shift remains unexplored. We address this challenge by conducting a comprehensive and systematic study of alignment generalization under domain shift. We compare five popular alignment objectives and various adaptation strategies from source to target, including target-domain supervised fine-tuning and pseudo-labeling, across summarization and question-answering helpfulness tasks. Our findings reveal systematic differences in generalization across alignment objectives under domain shift. We show that adaptation strategies based on pseudo-labeling can substantially reduce domain-shift degradation