Exploring Causes and Mitigation of Hallucinations in Large Vision Language Models

2025-02-24

AI Summary

Large Vision-Language Models (LVLMs) frequently generate hallucinated content—i.e., spurious objects or attributes not present in the input image—thereby compromising caption reliability. We observe that hallucinations often manifest as tokens whose generation exhibits low dependence on visual input, and thus propose an image-dependence signal as a principled hallucination indicator. To operationalize this insight, we introduce the first token-level hallucination classifier trained on parallel image-absent and image-augmented inference traces, enabling fine-grained hallucination detection without access to ground-truth annotations. Furthermore, we design a plug-and-play controllable decoding strategy that dynamically suppresses hallucinatory tokens during inference. Our approach requires no model fine-tuning or architectural modification, ensuring strong generalization across LVLMs and seamless deployment. Extensive evaluation across multiple benchmarks demonstrates significant hallucination reduction while preserving caption quality, fluency, and lexical diversity.

Technology Category

• Computer Vision: Large Vision Models
• Machine Learning: Large Multimodal Models (LMMs)
• Natural Language Processing: (Large) Language Models

Application Category

• Search and Retrieval-Augmented AI: Search Tool Learning with LLM: Teaching LLMs to invoke search and make use of retrieved information
• Economics, Online Markets and Human Computation: Cost models of using LLMs in production systems
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

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.

Problem

• Identify hallucination causes in LVLMs
• Develop automated hallucination detection
• Control hallucination rates in captioning

Innovation

• Automated pipeline detects hallucinated objects
• Token-level classifier uses hidden representations
• Decoding strategy controls hallucination rates