Rohan Paul
@rohanpaul_ai
A small layout model boxes each block, and the 1.3B OXR model reads every box and sets reading order, so layout can be swapped per domain. In the RSI loop, a general VLM (vision language model) flags OXR’s wrong or uncertain pages, groups the failure reasons, and generates new labeled pages for each group to train on, with no human labeling.
