MetaAlign: Align Large Language Models with Diverse Preferences during Inference Time

TL;DR

MetaAlign enables large language models to dynamically adapt to diverse human preferences during inference, overcoming the static limitations of traditional alignment methods like RLHF and DPO.

Contribution

We introduce MetaAlign, a novel approach allowing LLMs to align with various preferences specified at inference time, enhancing flexibility over existing static alignment techniques.

Findings

LLMs trained with MetaAlign effectively adapt to different preferences during inference.

MetaAlign outperforms static alignment methods in aligning with diverse preferences.

Experimental results validate the feasibility of dynamic preference alignment in LLMs.

Abstract

Large Language Models (LLMs) acquire extensive knowledge and remarkable abilities from extensive text corpora, making them powerful tools for various applications. To make LLMs more usable, aligning them with human preferences is essential. Existing alignment techniques, such as Reinforcement Learning from Human Feedback (RLHF) and Direct Preference Optimization (DPO), typically embed predefined preferences directly within the model's parameters. These methods, however, often result in a static alignment that can not account for the diversity of human preferences in practical applications. In response to this challenge, we propose an effective method, \textbf{MetaAlign}, which aims to help LLMs dynamically align with various explicit or implicit preferences specified at inference time. Experimental results show that LLMs optimized on our meticulously constructed MetaAlign Dataset can effectively align with any preferences specified at the inference stage, validating the feasibility of MetaAlign. We hope that our work can provide some insights into the alignment of language models.