Published paper
Marcelo Akira Inuzuka, Marcos Paulo C. M. Queiroz, Jhonata Fernandes Cordeiro, and Eduardo Fraga Pereira. 2026.
Acessilia at Dr.DocBench 2026: Agentic Document Parsing through Canonical Tree Adjudication.
In The Proceedings of Workshop on Document Intelligence and Understanding, pages 448–459, Budapest, Hungary. Association for Computational Linguistics.
Official results on the Dr.DocBench hidden test set (509 pages, EvalAI evaluation)
| System |
Text |
Reading order |
Tables (TEDS) |
Formulas (CDM) |
Overall |
| Baseline (MinerU only) |
67.65 |
53.44 |
47.36 |
0.00 |
58.09 |
| Acessília (final submission) |
74.67 |
72.30 |
58.31 |
0.22 |
70.37 |
Source: Table 1 in Inuzuka et al. (2026). A 12.28-point overall gain over the baseline. Formula recognition (CDM) remains a limitation acknowledged in the paper.
Open-source research software
Benchmarks and datasets