Screenshot-to-Code Models Fill In the Pattern They Expect, Not the Pixels They See — 69.8% Bias on Card Widths, 80.2% on Font Sizes

arXiv 2608.03691·high signalXBlueskyLinkedInCopy link

The first benchmark for visual pattern-completion bias perturbs one localized element inside a repeated UI pattern and asks five frontier MLLMs to recover the masked width or font-size from the screenshot. Mean accuracy is 21.17% on card-width and just 7.89% on text font-size, with Codex-5.3 best at 68.61% on cards but collapsing to 13.89% on text and Flash-3.0 hitting 96.11% bias on text. Reasoning analysis shows models frequently identify the anomalous element correctly and then override it with the pattern-consistent answer — a concrete failure mode for anyone shipping design-to-code pipelines.