Researchers have been studying 'reasoning shortcuts' in AI systems, which are rules that lead to correct predictions through unintended concepts. A recent framework analyzed these shortcuts using value relabelings and automorphism groups. However, the researchers found that this framework does not apply as stated to several benchmarks, including those from the CLE4EVR dataset. They also discovered that a simple method of padding can produce confident but false pathology in some cases. The study provides six theorems that give sufficient conditions for transitivity and its failure, and classifies Boolean transitivity exactly using automorphisms. Weakly supervised models were found to place all observed shortcuts at one level flagged by the theory, despite being trained on various domains. --- Why it matters: This research is important because it sheds light on the behavior of AI systems and their ability to generalize knowledge. The study's findings can help improve the design of AI models and prevent them from producing incorrect or misleading results. Source: https://arxiv.org/abs/2608.10420