VOPE: Revisiting Hallucination of Vision-Language Models in Voluntary Imagination Task

Xingming Long, Jie Zhang, Shiguang Shan, Xilin ChenarXiv 2025 · DOI · arXiv

Introduction

In the development of Vision-Language Models (LVLMs), "hallucination" has traditionally been defined as the generation of any content not explicitly present in the input image. Traditional metrics, such as CHAIR (Caption Hallucination Assessment with Image Relevance), indiscriminately penalize models for mentioning absent objects. However, as we move toward "voluntary imagination" tasks—such as creative storytelling or complex reasoning—we often want models to go beyond the pixels. The challenge lies in distinguishing between "bad" factual hallucinations and "good" creative imagination.

The VOPE Framework

To address this, researchers introduced VOPE (Voluntary-imagined Object Presence Evaluation), a framework that evaluates whether a model mistakenly believes an absent object is real or correctly identifies it as a narrative inference. The core of VOPE is a "recheck-based" evaluation pipeline.

The VOPE framework distinguishes between factual hallucinations and creative imagination by re-questioning the model about mentioned objects.

As illustrated above, the VOPE process involves three steps:

  1. Generative Task: The model is prompted with an image and a task requiring imagination (e.g., "What event might have just happened?").
  2. Object Extraction: All objects mentioned in the response are extracted.
  3. Presence Evaluation (The Recheck): For every mentioned object, the model is asked a follow-up question: "Is there a [object] in the image?"

Based on the model's self-evaluation, objects are categorized into two types of hallucinations:

  • Factual Hallucination ($Hal-D$): The model claims an object is present when it is not.
  • Imagination-based Hallucination ($Hal-I$): The model claims an object is merely an inference when it is actually present in the image, indicating a failure to track reality during creative modes.

Key Insights and Results

The study evaluated eight mainstream models, including GPT-4o, Qwen2.5-VL, and Gemini 2.0, using the MSCOCO dataset across varying levels of imagination tasks.

The Creativity Paradox

A primary finding is that factual accuracy does not guarantee imaginative consistency. For instance, Qwen2.5-VL achieved the lowest factual hallucination rates ($Hal-D$) but struggled significantly with $Hal-I$, frequently failing to recognize real objects when in "creative mode." Conversely, models like Gemma3 and Gemini 2.0 showed better consistency in tracking what they had imagined versus what was real, despite Gemini tending to generate more irrelevant imaginations.

Attention Drift and Mitigation

Using attention maps, the authors discovered that factual descriptions correlate with focused attention on specific image regions. In contrast, "imagination" results in scattered, low-intensity attention. Notably, standard hallucination mitigation techniques (like VCD or OPERA) designed to tether models to the image were found to have little to no positive effect on hallucinations during voluntary imagination tasks, and in some cases, even worsened them.

Conclusion

VOPE highlights that current LVLMs lack a robust internal "world model" capable of separating perceived reality from narrative fiction. The authors call for new mitigation strategies that focus on helping models maintain this internal distinction rather than simply forcing them to focus more intensely on the input pixels. Future work will involve refining these "recheck" mechanisms to account for potential self-correction biases in models.

Summary updated

Abstract

Most research on hallucinations in Large Vision-Language Models (LVLMs) focuses on factual description tasks that prohibit any output absent from the image. However, little attention has been paid to hallucinations in voluntary imagination tasks, such as story writing, despite this human-like cognitive ability being essential for real-world generative applications. To address this limitation, we introduce Voluntary-imagined Object Presence Evaluation (VOPE) -- a recheck-based evaluation benchmark for assessing LVLMs' grounding behavior in voluntary imagination tasks. Specifically, VOPE poses recheck-based questions to evaluate how an LVLM interprets the presence of the imagined objects in its own response. Rather than penalizing the imagined content itself, VOPE identifies hallucinations based on the correctness of the model's presence judgments for the generated objects. Built on this idea, we construct a dataset covering captioning, reasoning, and writing tasks with different levels of voluntary imagination. We apply VOPE to several mainstream LVLMs and hallucination mitigation methods, revealing two key findings: (1) most LVLMs hallucinate heavily during voluntary imagination, and their performance in presence evaluation is notably poor on imagined objects; (2) existing hallucination mitigation methods show limited effect in voluntary imagination tasks, making this an important direction for future research.
BibTeX
@misc{long2025vope,
  author = {Xingming Long and Jie Zhang and Shiguang Shan and Xilin Chen},
  title = {{VOPE: Revisiting Hallucination of Vision-Language Models in Voluntary Imagination Task}},
  year = {2025},
  eprint = {2511.13420},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2511.13420},
}
Figure