Hof – Today, artificial intelligence can describe images, recognize objects, and explain complex relationships. The pace of development is remarkable: so-called vision-language models (VLMs) combine text and image understanding in impressive ways. Yet, of all things, they struggle with a seemingly simple task—counting. Researchers at the Institute for Information Systems (iisys) at Hof University of Applied Sciences are now working to address this issue. “Many common models are very good at recognizing what can be seen in an image—but not reliably how many objects there are,” explains Prof. Dr. René Peinl from the Institute for Information Systems (iisys) at Hof University of Applied Sciences. Errors become more frequent when there are more than four or five objects of the same type. Why Counting Is So Difficult for AI The problem runs deeper than it may appear at first glance. While humans can intuitively grasp small quantities, larger numbers must be actively counted. This crucial step is missing in many AI models. In addition, existing training data is often unsuitable. “Some datasets are too simple and only encourage pattern recognition—others are too complex or flawed, for example due to occluded objects or ambiguous questions,” says institute director Prof. Peinl. As a result, models tend to “guess” or rely on learned expectations—sometimes producing surprisingly incorrect results. The Solution from Hof: An Artificial Dataset To tackle this problem in a targeted way, iisys has developed the SITUATE dataset. Instead of using real photographs, the researchers generate artificial 3D scenes with clearly defined properties. “We wanted to create an environment in which we can precisely control what happens in the image—and what does not,” says Prof. Dr. René Peinl. These scenes contain geometric objects such as cubes, spheres, or cylinders, with clearly defined positions (e.g., “to the left of the table”), allowing for targeted