Delve into Visual Contrastive Decoding for Hallucination Mitigation of Large Vision-Language Models

TL;DR

Large vision-language models suffer from hallucinations when grounding outputs in visual content. The paper studies training-free visual contrastive decoding (CD) by applying multiple visually distorted samples (downsampling, diffusion noise, no image, and image editing) and analyzes their effects using entropy and distribution-distance metrics. It then proposes an entropy-weighted fusion to combine these CDs, achieving robust hallucination mitigation across diverse LVLMs and benchmarks (POPE and MME) without additional training. The results demonstrate that leveraging complementary CDs through entropy-guided fusion yields practical, cross-task improvements, offering a scalable strategy for reducing hallucinations in real-world visual-language systems.

Abstract

While large vision-language models (LVLMs) have shown impressive capabilities in generating plausible responses correlated with input visual contents, they still suffer from hallucinations, where the generated text inaccurately reflects visual contents. To address this, recent approaches apply contrastive decoding to calibrate the model's response via contrasting output distributions with original and visually distorted samples, demonstrating promising hallucination mitigation in a training-free manner. However, the potential of changing information in visual inputs is not well-explored, so a deeper investigation into the behaviors of visual contrastive decoding is of great interest. In this paper, we first explore various methods for contrastive decoding to change visual contents, including image downsampling and editing. Downsampling images reduces the detailed textual information while editing yields new contents in images, providing new aspects as visual contrastive samples. To further study benefits by using different contrastive samples, we analyze probability-level metrics, including entropy and distribution distance. Interestingly, the effect of these samples in mitigating hallucinations varies a lot across LVLMs and benchmarks. Based on our analysis, we propose a simple yet effective method to combine contrastive samples, offering a practical solution for applying contrastive decoding across various scenarios. Extensive experiments are conducted to validate the proposed fusion method among different benchmarks.

Figures (19)

• Figure 1: Performance gain brought by visually changed samples across different LVLMs and benchmarks (POPE li2023pope and MME fu2023mme). We find that contrastive decoding with different samples reaches varying performance gains across benchmarks and LVLMs, motivating us to develop a solution to leverage the advantage of each CD method in a combined manner.
• Figure 2: Illustration of various visually changed samples for different contrastive decoding strategies.
• Figure 3: The entropy analysis on the POPE and MME benchmarks.
• Figure 4: The analysis of probability distribution distance on the POPE and MME benchmarks.
• Figure 5: The revision behaviors and answering tendency of InstructBLIP with different visually changed samples.