Loading...
Blog
Plain-language breakdowns of the research shaping OCR and document AI.
14 min readarXiv:2610.11148
SP-DocReader: Cutting Vary-600K OCR Error Rates in Half with Difference-Aware Self-Play
On Qwen3-VL-4B, SP-DocReader reduces Vary-600K character error rate by roughly 54% versus two-epoch supervised fine-tuning, training an OCR module while keeping the backbone frozen.
19 min readarXiv:2610.11818
Small Models, Structured Data: Can Lightweight VLMs Replace Cloud APIs for Complex Document Extraction?
Can 4B–7B vision models extract structured JSON from messy archives on local GPUs? A new FAU Erlangen study benchmarks 8 open-weight VLMs across university collections.
10 min read
We Sent a Blank White Page to 184 OCR Engines. Some of Them Read It Anyway.
When you show modern vision models an empty 800-by-1000 white image, classical OCR engines return silence. Newer models compose UFO letters, Victorian court cases, and gut health ads.
15 min readarXiv:2610.07091
Stop Sending Every Page to a Vision Model: Lessons from Citigroup, EY, and NVIDIA Researchers
An ingestion preprint reports that 92% of pages in an internal corpus skipped page-level OCR, explains structure-aware chunking, and measures where extraction time goes.
6 min read
OpenOCR now catches empty results from Gemini 3.1 Flash-Lite, and from any engine that follows
Google's Gemini 3.1 Flash-Lite sometimes returns no text for a perfectly readable page while reporting success. OpenOCR now detects that on every engine, tells you why, keeps the rest of your PDF, and never bills you more than the provider charged.
14 min readarXiv:2610.02880
Found But Not Read: Why Vision-Language Models Still Benefit From OCR
Adding CPU OCR text raised document QA accuracy by 13 to 16 percentage points on FoveDoc-Bench, revealing a gap between retrieving evidence and using it.
15 min readarXiv:2610.01191
Segmenting Handwritten Bangla Without Deep Learning: An Image Processing Deep Dive
How KUET researchers used adaptive image processing and iterative contour counting to segment handwritten Bangla words across colored paper and shadows.
12 min read
Introducing OpenOCR Stratify
We’re designing Stratify to break complex documents into readable parts, choose the right tools for each part, and keep every result connected to its source.
17 min readarXiv:2609.40356
Video Scene Text Editing Still Breaks: How ViTeX-Bench Exposes the Trade-Offs
ViTeX-Bench tests precise text replacement in real videos. Its results expose trade-offs between correct characters, stable motion, and scene preservation.
14 min readarXiv:2609.28757
Teaching a 500M Vision Model for Navigation More Than Doubled Its OCR-Bench Score
Researchers post-trained a 500M vision-language model for blind and low-vision assistance, improving its OCR-Bench ANLS score by 101.5% and suggesting that spatial grounding can transfer to visual-text reasoning.
11 min readarXiv:2609.29240
Breaking the Hallucination Loop: How TOLA Cuts Diffusion Text Super-Resolution to One Step
Diffusion can sharpen blurry text, but an incorrect text prior can reinforce the wrong character during iterative restoration. TOLA uses one image prediction, cutting measured inference time from 9.95s to 132ms.
12 min readarXiv:2609.03445
The Compiler in the Loop: How Tencent's DocEDR Uses Rendering to Check and Repair OCR
Tencent's DocEDR uses rendered predictions to diagnose and repair OCR, raising formula diagnostic Case-F1 by 30.99 points over DOCR-Inspector-7B.
15 min readarXiv:2609.03595
How Far Can Synthetic Data Take Thai OCR? Inside the Recipe That Reached 1.24% Median CER
Controlled synthetic document reconstruction helped a 0.9B Thai OCR model outperform Typhoon OCR 7B on five evaluation sets, with important limits on handwriting and severe errors.
12 min readarXiv:2609.02486
Stop Feeding Ten Pages to Vision LLMs: Inside ViSAR’s Training-Free Adaptive Retrieval
Fixed top-$k$ visual retrieval can slow vision LLMs and degrade accuracy. ViSAR exploits multi-vector geometry to reduce context size and cut latency by up to 58.7%.
11 min readarXiv:2608.31107
The Hidden Capture-Device Bias in an OCR Benchmark: What VeriCam Measures Across Devices
VeriCam studies capture-device bias in LPLCv2 traffic images, using pairwise verification and graph clustering to group unknown capture-device classes.
12 min readarXiv:2608.28248
Synthetic Layouts, Better Vertical Japanese OCR: Inside Synth-JDoc
Synth-JDoc combines programmatic HTML/CSS layouts, generated images, and document degradation to improve vertical Japanese OCR in four of five evaluated vision-language models.
8 min readarXiv:2608.26921
Arabic Manuscripts Don't Read Top to Bottom, and Now There's a Dataset That Admits It
AraMS-28k is a 28,600-line dataset of real historical Arabic manuscripts that finally tells OCR models where marginal notes actually belong in the text.
9 min readarXiv:2608.27169
An AI Read a 2,000-Year-Old Seal and Just Made Up a Poem
A new benchmark tests OCR on 3,000 years of Chinese artifacts. Even the best model gets under 58% right, and it hallucinates with total confidence.
8 min readarXiv:2608.26091
Reading Highway Plans Without Losing the Drawing: A Look at PlanSightRAG
A new RAG system skips OCR entirely and reasons over plan images directly, hitting 91% retrieval accuracy on DOT drawings.
8 min readarXiv:2608.22346
The Exam Dataset That Grades on Rubrics, Not Vibes
A new dataset pairs 485 scanned exam answers with the actual rubrics used to grade them, criterion by criterion, not just a final score.
8 min readarXiv:2608.22366
Feeding a Bad OCR Guess to an AI Corrector: When It Helps and When It Wrecks the Page
A new study on Arabic manuscripts finds that "let the AI fix the OCR" only works when the OCR's guess is close enough to be worth trusting.
8 min readarXiv:2608.18696
An old-school OCR pipeline just beat Gemini, GPT and Claude at reading Sanskrit
A fine-tuned CRAFT+GNN+CTC pipeline reads historical Sanskrit manuscripts more accurately than Gemini, GPT and Claude, after correcting just three pages.
10 min readarXiv:2608.18957
Inside the 22.6-Million-Image Pipeline Harvard Built to Find the Pictures OCR Ignores
Harvard's new pipeline pulled 22.6 million illustrations out of a million scanned books, using tiny models instead of expensive VLMs.
8 min readarXiv:2608.20011
When Aligning an Image Model Makes It Worse at Spelling
A new paper shows why preference-tuning flow models can quietly wreck generated text, and fixes it with a method that lifts OCR accuracy 47.5%.
9 min readarXiv:2608.20122
The OCR Benchmark Built to Make GPT-5 Fail
GPT-4o reads warped, watermarked, or AI-blended text with 23% accuracy. A new training method nearly triples that, without breaking normal OCR.
8 min readarXiv:2608.12898
Why Photographing a Document Still Breaks Most OCR Models
NaviDC-OCR handles scanned and phone-photographed documents in one model, tackling warped pages, tables, and formulas that trip up existing VLM-based parsers.
8 min readarXiv:2608.15698
Your Chatbot Can't Find the Chart That Has the Answer. This Paper Tries to Fix That.
ConceptFormer teaches retrieval models to find evidence in charts and layouts, not just text, beating both OCR and visual baselines by double digits.
9 min readarXiv:2608.12911
The OCR Model That Guesses Your Birthday When It Can't Read It
A new benchmark shows document-AI models don't just misread blurry ID fields, they fill them in from memorized training data, leaking real people's info.
9 min readarXiv:2608.13783
ImageNet Has Been Grading Models on a Broken Answer Key for 14 Years
A new reannotation of ImageNet-1k found 12% of its labels wrong and a third of images multilabel, and fixing it moves accuracy by up to 6 points.
8 min readarXiv:2607.27558
The AI That Reads Blueprints Like an Engineer, Not Like a Scanner
Drawing-Recode links dimension text on scanned CAD drawings to the exact geometry it measures, then writes editable 3D CAD code from the result.
9 min readarXiv:2607.27902
One Patch Is Enough: How a 2B-Parameter Model Out-Read GPT-5.4 on Street Signs
A 2B-parameter model beat GPT-5.4, Claude, and Gemini at reading street signs by pointing at exactly one image patch per word.
8 min readarXiv:2607.25390
The One-Step Trick That Might Make Diffusion Models Actually Useful for OCR
A new paper shows one deterministic diffusion pass, tuned the right way, beats slower multi-step restoration on classification, detection, and OCR.
8 min readarXiv:2607.24651
Stop Asking AI to Draw Boxes: A Better Way to Prove Where an Answer Came From
A new study finds that asking vision-language models to quote their evidence instead of drawing bounding boxes roughly doubles how often they cite the right part of a document.
8 min readarXiv:2607.22200
The OCR Model Is Reading Your Blank Margins, and That's Why It's Slow
A new paper cuts VLM-based OCR's visual tokens by 50% by teaching the model to ignore blank margins, with almost no accuracy loss.
8 min readarXiv:2607.15509
Three AI Models Walked In and Designed Their Own OCR Networks From Scratch
GPT-5, GPT-4o, and Claude Sonnet 4 ran 270 experiments designing their own handwriting-recognition networks, no human architect involved.
8 min readarXiv:2607.17052
Skipping Vision Layers Works, Until You Ask It to Read Text
An evolutionary search finds cheaper vision routes for VLMs that beat naive pruning, but task-tailored shortcuts can quietly wreck OCR accuracy.
10 min readarXiv:2607.19315
Why the Busiest-Looking Scans in Your Training Set Might Be Hurting Your OCR Model
A new label-free metric counts how many feature directions an image lights up, and pruning by it helps OCR models by cutting the busiest documents.
7 min readarXiv:2607.20385
Persian Has 110 Million Speakers and Almost No OCR Data. This Paper Fakes 343,000 Pages to Fix That
A new dataset fakes 343,000 Persian document images to fix OCR's worst blind spot: a script whose letters change shape depending on their neighbors.
9 min readarXiv:2607.18217
The Video AI That Has to Get Your Logo Right
A new video-generation paper feeds OCR output straight into a diffusion model so product logos and label text stop turning into mush.
9 min readarXiv:2607.16060
Can AI Draw You a Pretty Chart Without Lying About the Data?
A new benchmark shows image models routinely botch artistic charts, wrong bar heights, garbled labels, mismatched legends, and tests a fix that keeps both the art and the math honest.
10 min readarXiv:2607.09068
The Map Benchmark That Catches AI Models Faking Vision
OmniMapBench shows top AI models topping out at 75% on map questions, exposing how many "visual" benchmarks are secretly solvable by text alone.
8 min readarXiv:2607.10400
Why AI Document Readers Choke in the Middle of Long PDFs
A new synthetic benchmark shows top vision-language models get measurably worse at reading the middle of long documents, and chart understanding falls apart too.
8 min readarXiv:2607.13458
The Hidden Reason OCR Chokes on Curved and Tilted Text
A new positional encoding trick teaches OCR transformers to think in two dimensions, and the gains land exactly where you'd expect: curved, rotated, and warped text.
8 min readarXiv:2607.14041
The OCR Specialist That Never Trained for the Job
A new paper routes historical Manchu manuscripts to OCR specialists by handwriting style, and the best expert for each style often wasn't trained on it.
9 min readarXiv:2607.11192
The Benchmark Where the Best AI Model Still Fails 85% of the Time on a PDF
GDP.pdf tests AI on real professional PDFs, not toy OCR tasks. The best model passed 15% of items. Here's why that gap matters.
8 min readarXiv:2607.11562
Why a Model That's Great at Photos Is Bad at Reading Your Invoice
MonkeyOCRv2 trains its vision encoder on 113 million document images instead of borrowing one built for cats and cars, and a tiny 0.7B model beats a 3B rival.
8 min readarXiv:2607.09488
The Free Lunch Hiding in Your Vision Encoder: What SigLIP-HD Gets Right
A new training trick makes vision encoders read text better at the same resolution and same inference cost, no bigger images required.
6 min readarXiv:2607.08014
A Factory That Never Shares Its Photos, But Still Learns to Read Labels
FedTR trains OCR models across factories that can't share data, hitting 95.5% word accuracy without a single image leaving the building.
7 min readarXiv:2403.04473
TextMonkey: Teaching a Vision-Language Model to Point at the Line It Just Read
TextMonkey skips OCR entirely and reads documents as images, then backs up every answer with a bounding box showing where it looked.
7 min readarXiv:2309.11419
Microsoft's KOSMOS-2.5: One OCR Model, Two Very Different Jobs
Microsoft's KOSMOS-2.5 reads documents two ways at once: bounding-box OCR and full markdown reconstruction, from a single 1.3B-parameter model.
8 min readarXiv:2502.18443
Your PDFs Are Worth Trillions of Tokens. Most OCR Pipelines Are Throwing Them Away.
olmOCR shows how a fine-tuned 7B vision model can beat GPT-4o at reading PDFs, at 1/35th the cost per page.
7 min readarXiv:2109.10282
TrOCR: What Happens When You Throw Out OCR's Entire Playbook
Microsoft's TrOCR ditches CNNs and RNNs entirely, using two pre-trained Transformers to beat state-of-the-art OCR models with a much simpler pipeline.
7 min readarXiv:2111.15664
Donut Skips OCR Entirely, and It Still Reads Your Receipts
Donut reads documents without running OCR first. A look at how skipping text recognition altogether beat OCR pipelines on speed and accuracy.
7 min readarXiv:2409.01704
GOT-OCR2.0: A 580M-Parameter Model That Reads Text, Formulas, Sheet Music, and Charts
GOT is a 580M-parameter end-to-end model that reads text, math, tables, sheet music, and chemical diagrams with one prompt-driven system, no pipeline required.
8 min readarXiv:2308.13418
Why the Best PDF-to-Math Tool Before Nougat Got Equations Right Less Than 10% of the Time
Meta's Nougat reads a scanned paper and writes out LaTeX math directly, skipping OCR and bounding boxes entirely. Here's how, and where it still breaks.
8 min readarXiv:2412.02210
CC-OCR Put Nine Top AI Models Through 7,058 Real Documents, and Most Failed the Basics
CC-OCR tested nine leading multimodal AI models on messy real-world documents. Even GPT-4o and Gemini struggled with rotated text, foreign scripts, and repetition loops.