MOSLIM:Align with diverse preferences in prompts through reward classification

TL;DR

MOSLIM introduces a flexible multi-objective alignment method for LLMs using a single reward model and policy, enabling diverse preference handling without preference-specific training, and outperforms existing methods efficiently.

Contribution

This work presents MOSLIM, a novel approach that uses a multi-head reward model and a single policy to align LLMs with multiple preferences without preference-specific training.

Findings

Outperforms current multi-objective methods in most benchmarks.

Requires fewer GPU resources compared to existing policy optimization techniques.

Effective across various reward model sizes and optimization methods.

Abstract

The multi-objective alignment of Large Language Models (LLMs) is essential for ensuring foundational models conform to diverse human preferences. Current research in this field typically involves either multiple policies or multiple reward models customized for various preferences, or the need to train a preference-specific supervised fine-tuning (SFT) model. In this work, we introduce a novel multi-objective alignment method, MOSLIM, which utilizes a single reward model and policy model to address diverse objectives. MOSLIM provides a flexible way to control these objectives through prompting and does not require preference training during SFT phase, allowing thousands of off-the-shelf models to be directly utilized within this training framework. MOSLIM leverages a multi-head reward model that classifies question-answer pairs instead of scoring them and then optimize policy model with a scalar reward derived from a mapping function that converts classification results from reward model into reward scores. We demonstrate the efficacy of our proposed method across several multi-objective benchmarks and conduct ablation studies on various reward model sizes and policy optimization methods. The MOSLIM method outperforms current multi-objective approaches in most results while requiring significantly fewer GPU computing resources compared with existing policy optimization methods.