Mitigating style-image hallucination in large vision language models

Abstract

LLMs are widely applied across various domains, yet a significant challenge remains—their performance deteriorates sharply in out-of-domain scenarios, often leading to increased hallucinations. Despite its importance, this phenomenon has received limited attention in academic research. To address this, we first construct a benchmark dataset using style transfer techniques and employ it to evaluate the out-of-domain performance of several popular large-scale models. Building upon these findings, we introduce CopeCap, a lightweight image captioning model that leverages collaborative prompting to achieve strong out-of-domain performance without requiring additional training.