ZINA: Multimodal Fine-grained Hallucination Detection and Editing

TL;DR

ZINA introduces a fine-grained hallucination detection and editing framework for multimodal large language models, utilizing a new dataset and outperforming existing methods in accuracy.

Contribution

The paper presents ZINA, a novel approach for detecting and editing hallucinations in MLLMs at a detailed level, along with a new annotated dataset for training and evaluation.

Findings

ZINA outperforms GPT-4o and Llama-3.2 in detection accuracy.

Constructed VisionHall dataset with 6.9k real outputs and 20k synthetic samples.

Demonstrated effectiveness of fine-grained hallucination editing.

Abstract

Multimodal Large Language Models (MLLMs) often generate hallucinations, where the output deviates from the visual content. Given that these hallucinations can take diverse forms, detecting hallucinations at a fine-grained level is essential for comprehensive evaluation and analysis. To this end, we propose a novel task of multimodal fine-grained hallucination detection and editing for MLLMs. Moreover, we propose ZINA, a novel method that identifies hallucinated spans at a fine-grained level, classifies their error types into six categories, and suggests appropriate refinements. To train and evaluate models for this task, we construct VisionHall, a dataset comprising 6.9k outputs from twelve MLLMs manually annotated by 211 annotators, and 20k synthetic samples generated using a graph-based method that captures dependencies among error types. We demonstrated that ZINA outperformed existing methods, including GPT-4o and Llama-3.2, in both detection and editing tasks.