RCPS at EMNLP 2026

RCPS was accepted to the EMNLP 2026 Industry Track in August and presented on 25 October in Virtual Presentation Session 1. The paper is in the ACL Anthology at pp. 1531 to 1544, by Sang-Woo Son, Hyeong-seob Kim, Hyeonsang Kim, Hyun-woo Cho and Jinmo Kim.

It ranks document parsers by how well retrieval works on their output rather than by how clean the output looks. It started with WigtnOCR, the parser we run in production, which a better-looking parser beat on an intrinsic score and then lost to on retrieval.

The 663-question probe and a reference implementation are public. The measurements, and what the evaluation does not cover, are in the tech report.