AI still can't reason through physics diagrams the way physicists do

Researchers built the first test that measures whether AI can actually understand Feynman diagrams — the visual shorthand physicists use to predict particle behavior. State-of-the-art AI systems fail systematically: they can't enforce the basic rules physics requires (conservation laws, symmetry constraints) and they break the global logic that holds a diagram together, suggesting current AI lacks the structured reasoning needed for theoretical physics.

For years, AI researchers have claimed their systems can do scientific reasoning because they score well on generic physics benchmarks. But those benchmarks mostly test whether AI can extract information from diagrams, not whether it understands what the diagrams mean. This test reveals what that gap actually looks like: AI can look at a Feynman diagram and name its parts, but it can't tell you whether the diagram obeys the laws of physics. That matters because if AI can't handle the logical constraints that give scientific notation its power, it can't actually participate in the kind of reasoning physicists depend on for discovery.