Image Tokens Matter: Mitigating Hallucination in Discrete Tokenizer-based Large Vision-Language Models via Latent Editing

TL;DR

This paper identifies that hallucinations in large vision-language models with discrete image tokens are caused by visual priors from token co-occurrences, and proposes a latent editing method to suppress absent tokens, reducing hallucinations.

Contribution

The paper introduces a co-occurrence graph and a GNN-based clustering approach to identify visually absent tokens responsible for hallucinations, and proposes a latent editing technique to mitigate this issue.

Findings

Hallucinations are linked to clusters of tokens with high absent token correlation.

Suppressing absent tokens during generation reduces hallucination frequency.

The proposed method preserves model expressivity while mitigating hallucinations.

Abstract

Large Vision-Language Models (LVLMs) with discrete image tokenizers unify multimodal representations by encoding visual inputs into a finite set of tokens. Despite their effectiveness, we find that these models still hallucinate non-existent objects. We hypothesize that this may be due to visual priors induced during training: When certain image tokens frequently co-occur in the same spatial regions and represent shared objects, they become strongly associated with the verbalizations of those objects. As a result, the model may hallucinate by evoking visually absent tokens that often co-occur with present ones. To test this assumption, we construct a co-occurrence graph of image tokens using a segmentation dataset and employ a Graph Neural Network (GNN) with contrastive learning followed by a clustering method to group tokens that frequently co-occur in similar visual contexts. We find that hallucinations predominantly correspond to clusters whose tokens dominate the input, and more specifically, that the visually absent tokens in those clusters show much higher correlation with hallucinated objects compared to tokens present in the image. Based on this observation, we propose a hallucination mitigation method that suppresses the influence of visually absent tokens by modifying latent image embeddings during generation. Experiments show our method reduces hallucinations while preserving expressivity. Code is available at https://github.com/weixingW/CGC-VTD/tree/main