EMNLP 2026 · Budapest, Hungary · October 2026
Self-Improving Agentic Document Parsing System
Invited oral talk after finishing 2nd in the Dr.DocBench Challenge, with an overall score of 85.83. The system follows four steps: map every page into blocks, read each block with the cheapest tool that passes a check, assemble everything back in reading order, and learn from every page through an improvement loop with human approval.
Score progression
Overall score on the Dr.DocBench leaderboard, submission by submission.