AI Models Master Word Games But Fail Spatial Reasoning Tests

Puzzles and games have been central to the development of artificial intelligence since the very beginning. [1] In late 2024, scientists from Columbia University demonstrated that the most capable models could only solve 18 percent of New York Times Connections puzzles. [1] By early 2025, however, certain models were able to solve these puzzles almost perfectly every time. [1]

Despite these advances, today's models still struggle with subtle changes in classic riddles and demonstrate a particular weakness in visual puzzles. [1] Modern language models fail abysmally at mental rotation problems, which require determining if different images show the same object from varying angles. [1] These systems still appear unable to manipulate 3D objects in the way spatial thinkers like mechanical engineers can. [1]

The rapid progression from an 18% success rate to near-perfect performance on the NYT Connections puzzle highlights how quickly language-based AI reasoning can evolve. However, the persistent failure in spatial and visual tasks like mental rotation indicates a significant divergence between human cognition and machine learning architectures. While text-based reasoning is accelerating, physical world comprehension remains a major hurdle. What the sources don't address: How exactly AI developers plan to bridge the gap between two-dimensional visual processing and genuine three-dimensional spatial manipulation.