Answers are grounded in this paper and cite their sources.
Classifying scanned pages by document type is a routing step in enterprisedocument processing. Layout-aware models reach high accuracy but rely onlarge-scale document pretraining and full fine-tuning, and seldom expose why adecision was made. We study a deliberately lightweight alternative, GLaDViT,which keeps CLIP and BERT encoders frozen, derives a layout descriptor fromTesseract hOCR word boxes, and trains only a gated fusion head. Because theclassifier is linear in the gated sum of modality embeddings, every class logitdecomposes exactly into per-modality terms, which yields an attribution that isfaithful to the computation by construction, unlike the gate weights alone. Ona label-corrected release of Tobacco-3482 (2737 pages, fixed 2186/272/279split), GLaDViT reaches \((93.7±0.4%)\) accuracy over three seeds (95% Wilsoninterval 90.2--96.0%), against 87.7% for the best single modality; addinglayout to vision and text contributes 2.8 accuracy points and gating a further1.1. The layout stream nevertheless receives at most about 10% of the gatemass, indicating that gate weights are mixing coefficients rather thanimportance scores. A triple InfoNCE alignment of the three streams, used as awarm-up on frozen features, lowers accuracy to 90.3%; we analyse this negativeresult. Code and hOCR tooling will be made publicly available.
No takes yet. Share an insight, caveat, or question.
Murat Şimşek (2026) studied this question.
Synapse has enriched 8 closely related papers on similar clinical questions. Consider them for comparative context: