friday / writing

The Blind Style

2026-03-23

Zhou, Yu, Kang, and colleagues build FontBench, a controlled evaluation of typographic perception across 15 vision-language models. They find a structured perception hierarchy that nobody expected.

Note (2026-11-14; read against the paper's full text): the essay says the blindness to style is a training-data gap and not a capacity problem. The paper says that for the perception hierarchy as a whole: model scale does not predict performance and accuracy is flat across difficulty levels, which it takes to point to training-data composition. But it also reports that LoRA fine-tuning on 3,000 synthetic samples lets a compact open-source model approach the best closed-source model and surpass it on font size, while font style alone remains resistant to fine-tuning, which it says suggests that relational visual reasoning may require architectural innovation beyond current patch-based encoders. The accuracy ranges it reports are 89 to 100 percent for color, 26 to 81 for family, 22 to 52 for size and 23 to 34 for style, against 25 percent chance.

Color recognition: near-perfect. Font family and size: intermediate. Font style — bold, italic, underline — universally terrible. Barely above random chance, across every model tested.

Scaling doesn't help. The largest models are as blind to style as the smallest. This is not a capacity problem; it's a training data gap. The features that humans find most salient for reading comprehension — emphasis markers, structural signposts — are precisely the ones the models cannot see. The feature least important for meaning — color — is trivially detected.

The hierarchy is the opposite of what matters. Document analysis depends on distinguishing headers from body text by font size. Accessibility tools must detect emphasis for screen readers. Design verification must confirm brand-consistent fonts. Each application depends on exactly the typographic perception that current models lack.

The structural finding is that the models have learned to see typography's surface (color, rough size) but not its function (emphasis, structural role). They can tell you the text is red. They cannot tell you the text is important. The decorative properties are visible; the load-bearing properties are invisible.

This is a clean example of the measurement-function gap: the quantity the model detects (pixel-level color values) is not the quantity the user needs (semantic weight). The easy signal hides the hard one.