mDPO: Conditional Preference Optimization for Multimodal Large Language Models

TL;DR

This paper introduces mDPO, a novel preference optimization method for multimodal large language models that effectively addresses the unconditional preference problem, leading to improved performance and reduced hallucination.

Contribution

mDPO is the first multimodal preference optimization method that explicitly prevents language-only preference bias and incorporates a reward anchor to enhance model alignment.

Findings

mDPO outperforms existing methods on multiple benchmarks.

It significantly reduces hallucination in multimodal LLMs.

The approach is effective across different model sizes.

Abstract

Direct preference optimization (DPO) has shown to be an effective method for large language model (LLM) alignment. Recent works have attempted to apply DPO to multimodal scenarios but have found it challenging to achieve consistent improvement. Through a comparative experiment, we identify the unconditional preference problem in multimodal preference optimization, where the model overlooks the image condition. To address this problem, we propose mDPO, a multimodal DPO objective that prevents the over-prioritization of language-only preferences by also optimizing image preference. Moreover, we introduce a reward anchor that forces the reward to be positive for chosen responses, thereby avoiding the decrease in their likelihood -- an intrinsic problem of relative preference optimization. Experiments on two multimodal LLMs of different sizes and three widely used benchmarks demonstrate that mDPO effectively addresses the unconditional preference problem in multimodal preference optimization and significantly improves model performance, particularly in reducing hallucination.