AI still can't read messy tables — benchmark shows why spatial attention keeps breaking

Researchers built a dataset to measure why AI systems fail at reading complex tables with hierarchical layouts. The problem is perceptual overload: as tables get more complex, the AI's attention mechanism collapses across too many regions at once, and the system loses track of where it is spatially.

This is an honest documentation of where current multimodal AI systems actually break down in a real-world domain. Tables are everywhere in finance, healthcare, and scientific research — not a toy problem. The researchers show the failure isn't conceptual reasoning, it's spatial attention collapsing under density. This means the next generation of improvements won't come from smarter logic, but from forcing the system to ground itself in pixel coordinates as it reasons through a table step by step.