MetaAligner: Towards Generalizable Multi-Objective Alignment of Language Models
Recent advancements in large language models (LLMs) aim to tackle heterogeneous human expectations and values via multi-objective preference alignment. However, existing methods are parameter-adherent to the policy model, leading to two key limitations: (1) the high-cost repetition of their alignment algorithms for each new target model; (2) they cannot expand to unseen objectives due to their static alignment objectives. In this work, we propose Meta-Objective Aligner (MetaAligner), a model that performs conditional weak-to-strong correction for weak responses to approach strong responses. MetaAligner is the first policy-agnostic and generalizable method for multi-objective preference alignment, which enables plug-and-play alignment by decoupling parameter updates from the policy models and facilitates zero-shot preference alignment for unseen objectives via in-context learning. Experimental results show that MetaAligner achieves significant and balanced improvements in multi-objective alignments on 10 state-of-the-art policy models, and outperforms previous alignment methods with down to 15.71x less GPU training hours. The model also effectively aligns unseen objectives, marking the first step towards generalizable multi-objective preference alignment.
PDF Abstract
