Who Brings the Frisbee: Probing Hidden Hallucination Factors in Large Vision-Language Model via Causality Analysis

TL;DR

The paper tackles hallucinations in large vision-language models by proposing a causal hallucination probing framework that models the interaction between visual input, textual prompts, and latent representations. It introduces image, text, and embedding interventions grounded in backdoor adjustments to block confounding context factors and reduce erroneous outputs, evaluated on AMBER and COCO with InstructBLIP and mPLUG-Owl. Across image pasting, object removal, Foreground-Background text prompting, and embedding editing, the approach yields significant reductions in hallucination metrics (e.g., CHAIR, HAL, Cog) while preserving content coverage, without updating model parameters. The results highlight actionable pathways for mitigating multimodal hallucinations and open avenues for editing internal representations to curb spurious associations in LVLMs.

Abstract

Recent advancements in large vision-language models (LVLM) have significantly enhanced their ability to comprehend visual inputs alongside natural language. However, a major challenge in their real-world application is hallucination, where LVLMs generate non-existent visual elements, eroding user trust. The underlying mechanism driving this multimodal hallucination is poorly understood. Minimal research has illuminated whether contexts such as sky, tree, or grass field involve the LVLM in hallucinating a frisbee. We hypothesize that hidden factors, such as objects, contexts, and semantic foreground-background structures, induce hallucination. This study proposes a novel causal approach: a hallucination probing system to identify these hidden factors. By analyzing the causality between images, text prompts, and network saliency, we systematically explore interventions to block these factors. Our experimental findings show that a straightforward technique based on our analysis can significantly reduce hallucinations. Additionally, our analyses indicate the potential to edit network internals to minimize hallucinated outputs.

Figures (5)

• Figure 1: The InstructBLIP dai2024instructblip LVLM hallucinates a frisbee when describing a boy in the green field. There might be a spurious correlation between a boy and a frisbee. Meanwhile, the green field is another non-hallucinatory subject that might induce hallucinations. This underlying relation remains underexplored in the hallucination reduction research.
• Figure 2: (a) Ideal LVLM generation. (b) The causal graphical model for LVLM generation. (c-e) Deconfounded by (c) image intervention, (d) text intervention, and (e) embedding intervention.
• Figure 3: Our proposed image, text, and embedding intervention approaches correspond to Figure \ref{['fig:causality_graphs']} (c), (d), and (e).
• Figure 4: Embedding Saliency with timestamps in rows and dimensions in columns for (a) Instructblip and (b) mPLUG-Owl2. (c) and (d) show retrieval safe scores given either a non-hallucinatory word ($O_n$) or a commonly hallucinatory word ($O_h$) described in $\S$\ref{['ssec:exp3_embedding']} using InstructBLIP on AMBER dataset.
• Figure 5: A case with the FGBG approach continuing to hallucinate while image-pasting successfully reduces the hallucination, indicating the potential to explore cross-modality casual relations.