Multimodal Structured Foundation Models for Noisy Documents: A 110M Encoder Where Structure Matches Scale on Clinical Key--Value Generation

Foundation models for structured data have so far assumed that the data arrives clean and tabular. In healthcare, however, a large share of structured records is locked inside scanned paper reports of heterogeneous layout (free-text notes, multi-column forms, and table-heavy lab and pathology reports) recovered through error-prone OCR. The standard pipeline turns these into rows of key--value pairs and treats the OCR step as upstream ground truth, even though its error rate varies systematically with capture quality and layout. We argue that this regime calls for a multimodal view of structured generation, in which OCR quality signals are treated as a first-class modality alongside text, and key--value structure is treated as a first-class pretraining target. We instantiate this view as a noise-aware encoder whose input layer fuses text with seven OCR-derived reliability signals, and whose pretraining replaces MLM/NSP with two structure-aware objectives over key--value pairs. On 3,582 OCR-derived clinical-report pages, an 110M-parameter model trained this way outperforms strong Chinese encoder baselines by up to +4.6 F1 points on end-to-end key--value pairing and beats a 0.6B medical decoder LLM by +31.1 F1 points --- a result that runs against parameter-count-as-scaling-axis intuition. Treating noise as signal and structure as supervision is, in our setting, a stronger lever than scale, and we sketch the implications for tabular and document-grounded structured foundation models more broadly.