Consensus Selection over Prompt-Diverse VLM Transcriptions for Expert-Level Document Parsing

Abstract

We describe a training-free system for the Dr.DocBench document-parsing challenge, which asks for Markdown transcriptions of difficult book pages and scores text, reading order, display formulas and tables. One general vision-language model transcribes every page many times under different prompts, image settings and window sizes. A page-level minimum-Bayes-risk (MBR) selector, whose utility is the challenge's own text and reading-order matching, picks one transcription per page. Deterministic rules derived from a close reading of the evaluator refine it, and a gated word-level consensus layer fixes recognition slips that independent OCR engines confirm. Because display formulas and tables are scored by best match, the submission also carries LaTeX and HTML candidate blocks. The final system scored 84.54 overall on the 509 test pages (Text_ED 0.1641, TEDS 71.86, CDM 81.84, reading order 85.65), rank 7. We document the full development path over 134 commits and more than 25 scored leaderboard submissions, ablate every component, quantify which development-set gains transferred to the hidden test set (selection changes: 68%; convention rules: none), locate the remaining error with diagnostic submissions, and report a large set of measured negative results, including five stronger commercial models that did not beat the ensemble. System description for the Dr.DocBench track of the DocInsights 2026 shared task (Workshop on Document Intelligence and Understanding, EMNLP 2026); rank 7 on the confirmed final leaderboard: https://drdocbench-challenge.abaka-pages.com/#leaderboard. Code (MIT): https://github.com/pranshurastogi29/drdocbench-challenge. Report and poster: https://pranshurastogi29.github.io/drdocbench/