Mitigating Hallucination in Multimodal Large Language Model via Hallucination-targeted Direct Preference Optimization

TL;DR

Hallucination-targeted Direct Preference Optimization is introduced to reduce hallucinations in MLLMs and achieves superior performance across multiple hallucination evaluation datasets, surpassing most state-of-the-art (SOTA) methods and highlighting the potential of the approach.

Abstract

Multimodal Large Language Models (MLLMs) are known to hallucinate, which limits their practical applications. Recent works have attempted to apply Direct Preference Optimization (DPO) to enhance the performance of MLLMs, but have shown inconsistent improvements in mitigating hallucinations. To address this issue more effectively, we introduce Hallucination-targeted Direct Preference Optimization (HDPO) to reduce hallucinations in MLLMs. Unlike previous approaches, our method tackles hallucinations from their diverse forms and causes. Specifically, we develop three types of preference pair data targeting the following causes of MLLM hallucinations: (1) insufficient visual capabilities, (2) long context generation, and (3) multimodal conflicts. Experimental results demonstrate that our method achieves superior performance across multiple hallucination evaluation datasets, surpassing most state-of-the-art (SOTA) methods and highlighting the potential of our approach. Ablation studies and in-depth analyses further confirm the effectiveness of our method and suggest the potential for further improvements through scaling up.

Figures (9)

• Figure 1: Overview of our three kinds of Hallucinated-targeted Preference data. Better view on the digital screen.
• Figure 2: CHAIR scores under different max new tokens
• Figure 3: Performance of LLaVA-v1.5-7B w/ and w/o conflicts on AMBER, details in \ref{['cf-lab']}.
• Figure 4: Scaling law in HDPO with different data sizes. Lower CHAIR$_s$ and CHAIR$_i$ are better, while higher AMBER Score is better.
• Figure 5: System Prompt used in LCH