EAZY: Eliminating Hallucinations in LVLMs by Zeroing out Hallucinatory Image Tokens

2025-03-10

AI Summary

Large Vision-Language Models (LVLMs) suffer from object hallucination—generating objects not present in the input image. This work is the first to localize the root cause of hallucination at the image token level, revealing that only ~1.5% of high-attention image tokens dominate hallucinatory generation. To address this, we propose EAZY: a zero-shot, training-free, and architecture-agnostic intervention method. EAZY automatically identifies hallucination-relevant tokens via attention analysis and unsupervised importance estimation, then applies adaptive zero-masking to them. Evaluated across multiple LVLM architectures and benchmark datasets, EAZY consistently mitigates hallucination without compromising original task performance. It improves unsupervised hallucination detection accuracy by 15%, demonstrating precise, lossless, and generalizable hallucination suppression.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: Language Grounding & Multi-modal NLP

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work

Abstract

Despite their remarkable potential, Large Vision-Language Models (LVLMs) still face challenges with object hallucination, a problem where their generated outputs mistakenly incorporate objects that do not actually exist. Although most works focus on addressing this issue within the language-model backbone, our work shifts the focus to the image input source, investigating how specific image tokens contribute to hallucinations. Our analysis reveals a striking finding: a small subset of image tokens with high attention scores are the primary drivers of object hallucination. By removing these hallucinatory image tokens (only 1.5% of all image tokens), the issue can be effectively mitigated. This finding holds consistently across different models and datasets. Building on this insight, we introduce EAZY, a novel, training-free method that automatically identifies and Eliminates hAllucinations by Zeroing out hallucinatorY image tokens. We utilize EAZY for unsupervised object hallucination detection, achieving 15% improvement compared to previous methods. Additionally, EAZY demonstrates remarkable effectiveness in mitigating hallucinations while preserving model utility and seamlessly adapting to various LVLM architectures.

Problem

• Addresses object hallucination in Large Vision-Language Models (LVLMs).
• Identifies and removes hallucinatory image tokens to reduce errors.
• Introduces EAZY, a training-free method for hallucination mitigation.

Innovation

• Identifies hallucinatory image tokens automatically
• Zeroes out hallucinatory tokens to eliminate hallucinations
• Improves hallucination detection by 15% without training