Skip to content

State of the Art: Document Parsing & Layout Analysis (Mid-2026)

Type: Technical Survey Report
Author: Michael Pisman
Links: Full Report

Overview

As of mid-2026, the document parsing landscape has shifted from heuristic text extractors toward end-to-end Vision-Language Models (VLMs) and Data Extraction with Language Models (DELMs). This report surveys the top open-source frameworks and benchmarks.

Key Findings

  • DeepSeek-OCR 2 uses a 3B-parameter "Visual Causal Flow" architecture that processes document structure contextually, outperforming MinerU on dense scientific documents.
  • Datalab's Lift 9B VLM introduces schema-driven extraction — providing a JSON schema and Lift visually parses the PDF to populate it directly — operating at 25+ pages per second.
  • Frontier VLMs (Qwen3-VL, GLM-4.6V/GLM-OCR) process page images natively as "visual agents," eliminating the traditional pipeline separation of layout analysis, OCR, and text correction.

For building a production document parser: use DELM as the orchestration framework, Marker with Lift for general-purpose Markdown extraction, and DeepSeek-OCR 2 or GOT-OCR 2.0 for dense visual layouts (tables, math).

Relevance

This research directly informs the ingestion pipeline for Scriptorium and the lab's broader Knowledge Base architecture.