olmOCR 2: Unit Test Rewards for Document OCR

2025-10-22

AI Summary

This paper addresses core challenges in OCR conversion of digital printed documents (e.g., PDFs), including mathematical formula recognition, table parsing, and multi-column layout reconstruction. Methodologically, it proposes a reinforcement learning framework grounded in verifiable rewards: (1) a fine-grained reward mechanism guided by binary unit tests to enforce structural and semantic correctness; (2) a controllable synthetic document generation pipeline that produces large-scale, diverse training and evaluation samples covering complex layouts; and (3) a dedicated 7B-parameter vision-language model, olmOCR-2-7B-1025, trained end-to-end. Evaluated on the olmOCR-Bench English benchmark, the method achieves state-of-the-art performance, significantly outperforming prior approaches on three critical tasks—mathematical formula conversion, table parsing, and multi-column ordering—thereby validating the effectiveness of the “verifiable-reward + synthetic-test-driven” paradigm.

Technology Category

• Natural Language Processing: Code Generation / Program Synthesis from Natural Language
• Machine Learning: Large Multimodal Models (LMMs)
• Search and Optimization: Learning to Search

Application Category

• Economics, Online Markets and Human Computation: LLM based quality controls for crowd work
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

We present olmOCR 2, the latest in our family of powerful OCR systems for converting digitized print documents, like PDFs, into clean, naturally ordered plain text. olmOCR 2 is powered by olmOCR-2-7B-1025, a specialized, 7B vision language model (VLM) trained using reinforcement learning with verifiable rewards (RLVR), where our rewards are a diverse set of binary unit tests. To scale unit test creation, we develop a pipeline for generating synthetic documents with diverse and challenging layouts, known ground-truth HTML source code, and extracted test cases. We show that RL training on these test cases results in state-of-the-art performance on olmOCR-Bench, our English-language OCR benchmark, with the largest improvements in math formula conversion, table parsing, and multi-column layouts compared to previous versions. We release our model, data and code under permissive open licenses.

Problem

• Convert digitized print documents into clean text
• Improve OCR accuracy for complex layouts
• Enhance math formula and table parsing

Innovation

• Reinforcement learning with verifiable unit test rewards
• Synthetic document pipeline for diverse layout generation
• Specialized vision language model for OCR conversion