Exploring Causes and Mitigation of Hallucinations in Large Vision Language Models

TL;DR

The paper investigates object hallucinations in Large Vision-Language Models during image captioning, showing that hallucinations often emerge in later tokens and can be driven by language priors rather than visual features. It introduces a detection-and-decoding framework that compares token-level hidden representations from image-present and image-absent inferences, paired with an automated caption-annotation pipeline and a token-level classifier trained on automatically labeled data. A novel decoding strategy integrates the classifier into generation to reduce hallucinations without fine-tuning. Experiments on COCO with LLaVA-1.5-7B demonstrate reductions in CHAIR_s and CHAIR_i while preserving caption quality, suggesting a practical path toward more reliable multimodal generation.

Abstract

Large Vision-Language Models (LVLMs) integrate image encoders with Large Language Models (LLMs) to process multi-modal inputs and perform complex visual tasks. However, they often generate hallucinations by describing non-existent objects or attributes, compromising their reliability. This study analyzes hallucination patterns in image captioning, showing that not all tokens in the generation process are influenced by image input and that image dependency can serve as a useful signal for hallucination detection. To address this, we develop an automated pipeline to identify hallucinated objects and train a token-level classifier using hidden representations from parallel inference passes-with and without image input. Leveraging this classifier, we introduce a decoding strategy that effectively controls hallucination rates in image captioning at inference time.

Figures (6)

• Figure 1: The proportion of labels across token positions in the response
• Figure 2: (a) The workflow compares LLaVA's consistency in object existence judgments across two tasks; (b) The histogram displays the distribution of $p_{\text{exist}}$ values predicted by the Pipeline for Yes/No responses in discriminative tasks.
• Figure 3: Comparative visualization of token probabilities and distributions with and without image input. The caption annotations are manually labeled: gray indicates UNSURE, and orange highlights INACCURATE content.
• Figure 4: Training process for token-level hallucination classifier.
• Figure 5: Token-level classification at inference time.