OCRBench Leaderboard

What this benchmark measures

• text recognition in images
• document-oriented visual question answering
• key information extraction

Relevant to:

• document processing model selection
• OCR pipeline screening

Failure modes it exercises:

• recognition failure on handwritten, multilingual, and non-semantic text

Method and its limits

A composite OCR evaluation spanning 29 datasets across five text-related visual task types: text recognition, scene text-centric VQA, document-oriented VQA, key information extraction, and handwritten mathematical expression recognition.

• A single composite score spans five distinct task types with different difficulty profiles.
• The authors identify multilingual text, handwritten text, non-semantic text, and mathematical expression recognition as persistent weak areas, which a headline score obscures.

Dataset

Dataset

OCRBench

Type

aggregated OCR and document understanding datasets

Freshness

aging

Benchmark release

2023-05-13

29 constituent datasets as registered in the primary paper. The date records first paper submission; the paper was revised through 2024.

How to read this score

The composite score is a useful screen, but per-task results matter more: the paper specifically flags handwritten, multilingual, and non-semantic text as weak areas for the models tested.

Similarity to real tasks: High — Established from the primary methodology: OCRBench aggregates 29 datasets covering text recognition, scene-text and document-oriented visual question answering, key information extraction, and handwritten mathematical expression recognition. These are the same operations a document-processing pipeline performs, so similarity to real work is high. Throughput, cost, and page-layout handling at scale remain out of scope.

Data contamination risk: Medium — Constituent datasets are public and widely used in multimodal training. No run-specific overlap audit is registered.

Benchmark gaming risk: Low — Scoring is based on recognition correctness against ground-truth text, which limits shortcut strategies; prompt format still affects extraction behaviour.

What you still need to test yourself

• Test on your own document types, scripts, and scan quality.
• Measure throughput and cost per page, which the benchmark does not cover.

This benchmark supports decisions about:

• Shortlist multimodal models for document and OCR workloads.
• Anticipate weakness on handwriting and non-Latin scripts.

Limitations

• A composite score hides large per-task variation.
• Not evidence of throughput, cost, or layout handling at production scale.

Primary source reviewed 2026-10-11. OCRBench is treated as an aging benchmark first published in 2023 and revised through 2024.

What this evidence does not cover

• Only one model is currently scored on this page, so no comparison is supported.

Sources

• OCRBench: On the Hidden Mystery of OCR in Large Multimodal Models Liu, Li, Huang, Yang, Yu, Li, Yin, Liu, Jin, Bai · accessed 2026-10-11Supports the 29-dataset composition, the five task types, and the stated weaknesses on multilingual, handwritten, non-semantic, and mathematical text; it does not support any model score shown here.
• OCRBench evaluation release Liu et al. (OCRBench authors) · accessed 2026-10-25The authors' OCRBench release, matching the paper's title. It establishes the task inventory and scoring assets; it does not establish per-model results.

How this benchmark is scored

Primary source: https://github.com/Yuliang-Liu/MultimodalOCR

Continue your research

Use these connected ModelRefs sections to compare alternatives, inspect implementation paths, and review the evidence and governance boundaries relevant to OCRBench Leaderboard.

• Browse models
• Review providers
• Inspect benchmarks
• Explore workflows
• Learn AI concepts
• Read decision guides
• Review trust and methodology
• Start a decision path