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

TL;DR

This work addresses object hallucination in LVLMs built with discrete image tokenizers by identifying visual priors arising from token co-occurrence. It introduces Context-Guided Clustering (CGC) to learn co-occurrence-based token groups via a GNN and contrastive learning, followed by K-means clustering to form dominant visual clusters. It then mitigates hallucinations through Visual Token Decontamination (VTD), a latent-space editing technique that subtracts the influence of visually absent tokens from intermediate representations during generation, with CGC performed offline. Empirical results across AMBER, Object HalBench, and MME demonstrate that CGC+VTD reduces hallucinations while preserving or enhancing perceptual and cognitive capabilities and operates with favorable efficiency, making it a practical addition to discrete-token LVLMs.

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

Figures (9)

• Figure 1: Overview of our proposed method, which consists of two key components: Context-Guided Clustering (CGC) and Visual Token Decontamination (VTD). CGC organizes image tokens into clusters according to their co-occurrence patterns. VTD identifies potentially hallucinative tokens based on visual context and suppresses their influence during autoregressive generation. During inference, only VTD is applied while CGC is performed only once for every model offline.
• Figure 2: Illustration of the CGC pipeline. Codebook entries are first used to help construct a co-occurrence graph over image tokens, followed by a learning process to obtain graph-based embeddings, which are then used to cluster image tokens.
• Figure 3: Left: Hallucination HitRate for K from 1-10. Middle: Comparison between $C_2$ tokens identified by our CGC method versus naive K-means. Right: HitRate@5 of $C_2$ tokens from the ten most dominant clusters.
• Figure 4: The three heatmaps show CHAIR score difference between our method with the Nucleus Sampling baseline. Blue means improvement and Red means deterioration. It is shown that our method benefits from a wide range of parameter combinations.
• Figure 5: We show how editing weight $\gamma$, cluster size, number of GNN layers and editing layer affect our method. We report CHIAR score for generative tasks and F1 for discriminative tasks using Janus-Pro-7B. Overall, our method is relatively robust to hyperparameters.