Kaori 2.5: Zero-Shot Document Layout Synthesis and Multimodal Spatial OCR
Abstract
Extracting structured hierarchies from arbitrary multi-column documents, complex mathematical formulas, and historical manuscript scans remains a notorious challenge for classical OCR engines and generic vision-language models. In this technical report, we introduce Kaori 2.5, a compact 120M parameter spatial transformer engine engineered for zero-shot layout recovery and structured JSON extraction.
Kaori 2.5 employs multi-scale spatial cross-attention to construct dense 2D bounding graphs directly in latent space, eliminating bounding-box drift and multi-stage alignment pipelines. Experimental results demonstrate a 99.4% structural accuracy on complex multi-column scientific papers and historical manuscripts.
1. Spatial Transformer Architecture
Unlike classical OCR pipelines that decouple bounding-box detection from text recognition, Kaori treats document understanding as a unified spatial graph generation task. High-resolution document images are encoded into continuous multi-scale patch tokens that preserve microscopic typographical geometry.
Citation
@article{yadav2026kaori,
title={Kaori 2.5: Zero-Shot Document Layout Synthesis and Multimodal Spatial OCR via Multi-Scale Transformers},
author={Yadav, Darsh},
journal={Sorika Perception Lab Technical Report},
year={2026},
month={August},
url={https://sorikalabs.com/research/kaori}
}