Vision-Language Models Struggle with Negation, Study Reveals

A new study from MIT reveals a critical flaw in vision-language models (VLMs): they struggle to understand negation. This limitation can lead to significant errors in real-world applications, potentially impacting diagnoses in healthcare and quality control in manufacturing.

The research highlights that VLMs, which are trained on vast datasets of images and captions, often fail to recognize words like “no” or “doesn’t,” leading to misinterpretations of image content. Kumail Alhamoud, an MIT graduate student and lead author of the study, emphasizes the potential for “catastrophic consequences” if these models are used blindly without addressing this fundamental issue.

To demonstrate this vulnerability, the researchers tested VLMs’ ability to identify negation in image captions. The models often performed no better than random chance. Following this, they created a dataset of images paired with captions including negation words describing missing objects. This dataset was used to retrain a VLM, resulting in improved performance in tasks requiring the model to identify images that do not contain specific objects.

While retraining showed promising results, the researchers caution that this is not a complete solution. They emphasize the need for further investigation into the root causes of this problem. Marzyeh Ghassemi, an associate professor at MIT, stresses the importance of intensive evaluation before deploying VLMs in high-stakes settings.

The study identifies “affirmation bias” as a key reason for this failure, where VLMs tend to ignore negation words and focus solely on the objects present in the images. This issue persists across various VLMs, regardless of how negation is expressed.

To address this, the researchers developed datasets incorporating negation words. By finetuning VLMs with these datasets, they achieved performance gains in image retrieval and multiple-choice question answering tasks. Alhamoud hopes this work will encourage users to carefully test VLMs before deployment and spark further research into more robust solutions.

The team suggests future research directions, including training VLMs to process text and images separately and developing specialized datasets for specific applications like healthcare.

The research will be presented at the Conference on Computer Vision and Pattern Recognition.