End-to-end full-page document parsing is a difficult task because it must preserve both the content and the structure of heterogeneous elements on a page. The authors of a new arXiv paper note that even when a model produces largely accurate output, it can still contain local errors and omissions. These small mistakes are easy to miss when evaluating only at the page level.
The paper introduces P2TPO, a method for structure-aligned page-to-token credit assignment. Rather than relying on a single page-level advantage signal, which does not explicitly capture where individual tokens go wrong, P2TPO aims to assign credit more precisely to tokens while respecting the document's structure.
The abstract is truncated in the source, so details of the method and experimental results are not available here. The motivation, however, is clear: improving full-page parsing requires not just overall accuracy, but also attention to local structure and token-level errors. P2TPO appears to be a step in that direction.