

How can we determine whether an AI system actually draws on existing knowledge or bases its decisions primarily on learned patterns? This question is at the heart of a joint research paper by the German Research Center for Artificial Intelligence (DFKI) and Osnabrück University, which received the Best Student Paper Award at the IEEE International Conference on Tools with Artificial Intelligence (ICTAI) 2026 in Boca Raton, Florida.
The paper, “Explaining Knowledge Utilization in Heterogeneous GNNs for 3D Semantic Scene Graphs,” was co-authored by Tim Stelzner, Marian Renz, Felix Igelbrink, and Prof. Dr. Martin Atzmueller. It focuses on how artificial intelligence interprets spatial environments and incorporates additional knowledge about objects and their relationships.
To understand their surroundings, AI systems must not only recognize individual objects but also identify the spatial and semantic relationships between them. This is where 3D semantic scene graphs come into play. They represent, for example, which objects are present in a room and how they relate to one another.
Specialized Graph Neural Networks (GNNs) can infer such relationships from sensor data. Additional knowledge graphs provide further information, such as typical properties of objects and the relationships between them. However, it remains difficult to determine whether and how AI systems actually use this additional knowledge when making predictions.
This is precisely where the award-winning research comes in. The researchers adapted existing Explainable Artificial Intelligence (XAI) methods and introduced new evaluation metrics. These enable a systematic analysis of which information influences an AI model's decisions and what role the integrated knowledge plays in the process.
The findings show that the AI models examined rely primarily on learned statistical correlations rather than making targeted use of the semantic knowledge structures provided to them. When changes to the model architecture enforce greater use of this knowledge, its influence on decision-making increases, but the accuracy of certain predictions declines significantly.
The results demonstrate that simply providing additional information is not sufficient to ensure that AI systems use it effectively. At the same time, the study shows how explainable AI methods can help identify such weaknesses in model architectures and provide starting points for further improvements.