Humans squint and read it; multimodal models say no text: original-only training lifts reading from 26.50 to 88.61

Multimodal LLMs can report that an image contains no text even when a string is embedded in its coarse structure and a human recovers it by squinting or viewing the image at reduced scale. This failure had not been named or studied as its own class; it appears in production moderation traffic, where domains, contact strings and slogans are drawn into coarse image structure so that human viewers read them while the reviewing model reports no text. Four commercial and open-weight MLLMs reach only 8.3 to 24.4 ANLS on the 1076 original production moderation images, so the failure is not confined to one checkpoint.

With frozen weights, hidden text leaves a strong coarse signal in the vision encoder while the identity of the individual glyph is only weakly represented. Downscaling or low-pass filtering was already known to make such text legible, but which property of those transforms is responsible was not established. On 500 matched pairs that differ only in whether a string was embedded, text presence reaches 0.956 to 0.981 AUC, while glyph identity reaches 0.718 for Chinese and 0.612 for Latin; restricting the probe to glyph-bearing locations raises Chinese to 0.799, identifying spatial selection as a major bottleneck.

During training a frozen teacher sees a downscaled view and the correct transcription, and its per-step token distribution supervises a student that sees only the original image. Earlier options either appended a downscaled view at inference, which changes the deployed input, or supervised with hard one-hot labels, which discards the relative probabilities among visually plausible glyphs. On the 1076 production moderation images the method reaches 88.61 ANLS, against 26.50 for the base model, 58.96 when the corrective view is appended only at inference, 72.47 for single-view SFT and 73.55 for DAPO.

After training, glyph representations become more discriminative and reading depends causally on the visual regions that carry the glyphs. Before training, glyph identity was a weak signal in the encoder and reading was not concentrated on glyph locations. Chinese glyph-identity AUC rises from 0.718 to 0.975 and Latin from 0.612 to 0.779; replacing glyph-bearing visual-token embeddings with the mean of non-glyph tokens drops accuracy from 90.5% to 71.5%, while replacing an equal number of random background tokens moves it only from 90.5% to 91.0%.

The method targets visual moderation settings that must read text carried in coarse image structure while receiving only the original image at inference, and it is evaluated on 1076 production-derived moderation images plus a controlled synthetic diagnostic suite. The diagnosis lists spatial selection and high-frequency interference as major contributors, with visual token count a possible additional factor that is not fully disentangled from frequency. Training can reorganize partially recoverable evidence; at conditioning strength 0.5, where strokes are largely destroyed, the base and trained models are statistically indistinguishable.

High-frequency interference and visual token count are not fully disentangled, and the token-count effect is only a possible contributor. After training, general benchmarks degrade in a task-dependent way: AI2D moves from 86.66 to 85.33, DocVQA from 95.93 to 92.98, and OCRBench from 879 to 833. The method currently relies on visual evidence that remains recoverable from the input and on a fixed privileged supervision setup.