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.
Experiments
Quantitative Results
The results on the Yes-No question tasks, including POPE and MME. P-R, P-P, and P-A denote the random, popular, adversarial task of POPE respectively. The best performance within each set is marked bold, and the secong highest performance is underlined.
Analysis
Entropy
Entropy represents the uncertainty of the prediction, serving as an indicator to determine the extent of inducing hallucinations given different visually changed samples.
Anaysis of entropy on POPE and MME benchmarks.
Probability Distribution Distance
We mainly apply the Hellinger distance to estimate the distance between probability distributions. Intuitively, when the distance increases, the prediction distribution of the original visual input and visually changed samples differs more, leading to a high risk of obtaining hallucination results from visually changed samples.
Anaysis of probability distribution distance on POPE and MME benchmarks.
Revision Behaviors
We investigate the behavior of each contrastive sample via three metrics: revised-correct samples, revised-wrong samples, and answering tendency on the Yes-No Questions from POPE and MME benchmarks.
The revision behaviors and answering tendency on POPE and MME benchmarks.
Pairwise Overlap of Rectified Answers
We investigate the complementary properties among different CD methods via calculating the pairwise overlap of rectified answers with Jaccard similarity coefficient (known as IoU).
The pairwise overlap of revise-correct samples.
Visualizaiton
We apply
InstructPix2Pix to edit the input images with the given editing textual instructions (e.g., the queried objects). The following examples ar the image editing results on POPE benchmark.
