🔭 Daily Papers (2026-09-22)

Document Understanding

Chart-RVR — A reinforcement learning framework that trains chart reasoning agents with verifiable process rewards, decomposing reasoning into auditable structure, evidence-table and derivation blocks instead of answer-only outputs.

Visual Text Generation

DuetGen — An autonomous visual text generator that jointly learns autoregressive layout planning and diffusion rendering so that rendering supervision shapes the planner's representations instead of optimizing the two stages separately.

📖 Contents

• Overview
• 🎉 News
• 📖 Contents
• 🔭 Daily Papers
• 🔍 Emerging Trends
• 📄 Document Parsing
• 📄 Document Understanding
• 📄 Visual Text Generation
• 📄 Specialized Model
• 📄 Benchmarks and Evaluation

Overview

A curated, continuously updated reading list of OCR in the era of large language models, covering document parsing and understanding, visual text generation, benchmarks, challenges, and future perspectives, with a focus on research around the past five years (2021–now).

Scope. This list tracks OCR in the LLM era: work that applies large vision-language or multimodal models to text-rich images and documents (parsing, understanding, benchmarks, and specialized text tasks). It is not a general document-AI list, a generic MLLM list, or a classical OCR-1.0 list; such work appears only when it directly bears on text-rich visual understanding.

A note on evaluation. Most recent systems are released as technical reports with self-reported numbers, private test sets, and inconsistent protocols, so cross-paper scores are rarely comparable in a rigorous sense. We list results as reported and, where known, indicate the evaluation basis. The field still lacks a unified, contamination-resistant, reproducible benchmark, and we see building one as a prerequisite for trustworthy leaderboard claims.

🎉 News

• [2026-2-11] 🔥 We release an open-source resource to help the community easily track recent OCR research!

Contributing. PRs welcome. One row per model, newest first; please include venue/date, affiliation, and a code or model link.

🔍 Emerging Trends

Emerging Trends (2023–2026)

• End-to-end VLM-based parsing replaces modular OCR pipelines.
• Reinforcement learning for layout and reading order modeling.
• OCR-free document understanding models.
• Scaling down: compact document VLMs under 1B parameters.
• Long-doc OCR is a scaling problem of tokens and consistency, not just context length.
• Structure (layout + logic) is the new accuracy.
• Generative OCR shifts the core risk from “misrecognition” to “hallucination”.
• Benchmarks are moving toward executable evaluation.
• Document agents and autonomous reasoning over PDFs.
• Document Agents need recoverability, not one-shot perfection.

📄 Document Parsing

Document parsing focuses on converting visually complex documents into structured, machine-readable representations. In the LLM era, parsing is no longer a pipeline of isolated modules, but increasingly unified within end-to-end VLM architectures.

📄 See full list at Document-Parsing.md

📄 Document Understanding

Document understanding extends beyond structural parsing to semantic comprehension and reasoning over visually rich documents.

📄 See full list at Document-Understanding.md

📄 Visual Text Generation

Visual Text Generation focuses on generating or editing legible, visually harmonious, and semantically consistent text within images, serving as the creative inverse of OCR. In the LLM era, it is no longer a task reliant on specialized modules, but is emerging as a foundational skill for general-purpose generative models.

📄 See full list at Visual-Text-Generation.md

📄 Specialized Model

Beyond general document parsing and understanding, traditional OCR research remains essential for specialized visual-text structures, restoration, temporal analysis, and forensic reliability.

📄 Document Dewarping

Document dewarping restores photographed or scanned pages to a geometrically rectified and readable form.

📄 Physical Structure Analysis

Physical structure analysis identifies document regions, layouts, and spatial relationships before semantic reasoning.

📄 Reading Order Prediction

Reading order prediction recovers the intended sequence of text and visual elements in complex page layouts.

📄 Mathematical Expression Recognition

Mathematical expression recognition converts printed or handwritten formulas into structured symbolic representations.

📄 Table & Chart Understanding

Table and chart understanding covers structure recognition, data extraction, grounding, retrieval, and visual reasoning over structured graphics.

📄 Scene Text Understanding

Scene text understanding unifies text detection, recognition, and spotting in natural images and related real-world settings.

📄 Text Removal & Editing

Text removal and editing modifies textual content in images while preserving visual consistency and surrounding appearance.

📄 Text Image Super-Resolution

Text image super-resolution restores low-resolution text regions while preserving character identity and legibility.

📄 Handwritten Document Analysis

Handwritten document analysis covers handwriting recognition, writer-related analysis, signatures, and document-level handwritten content.

📄 Video Text Analysis

Video text analysis studies temporal text detection, recognition, spotting, tracking, and reasoning across video frames.

📄 Historical Document Analysis

Historical document analysis addresses OCR, restoration, structure recovery, and interpretation for manuscripts and archival materials.

📄 Tampered Text Detection & Forensics

Tampered text detection and forensics studies document authenticity, manipulation localization, forgery detection, and evidence-grounded verification.

| Venue | Name | Primary affiliation | Title | GitHub