Tag
Datalab released 'lift', an open-weights 9B vision model that extracts schema-constrained JSON directly from PDFs and images in a single step, matching Gemini Flash 3.5 accuracy at 3x the speed. It handles multi-page documents, is self-hostable via vLLM or Hugging Face, and is available under Apache-2.0 code / OpenRAIL-M model licenses.
Vik Paruchuri explains five ways PDFs misreport or omit the positions of words and characters, highlighting a key challenge in document parsing and OCR pipelines.
OCR researcher Vik Paruchuri highlights the dramatic progress in handwriting recognition models over the past two years, noting that outputs have gone from barely readable to nearly perfect transcription.
papero-pdf-text-extractor 是一个轻量级 PDF 结构解析工具，无需 ML 模型即可在 CPU 上提取阅读顺序、表格、公式和边界框，支持导出 Markdown、JSON、Word 和 Excel，适用于 RAG、搜索和 LLM 场景。
A DocInsights 2026 shared-task system paper describing EviCalc, a PDF document QA pipeline that pairs evidence passage selection with LLM-generated arithmetic expressions evaluated in local code, reporting 8.61% joint accuracy on DocSem's test set along with descriptive failure analysis, including OCR and block-grouping errors that merged relevant passages.
The paper introduces GAD-RL, a method that adaptively gates and attenuates on-policy distillation supervision during joint post-training of vision-language models to improve OCR transcription faithfulness, achieving 59.92% Micro Recall on CHAOS-Bench on Qwen3.5-2B, surpassing GRPO by 8.45 points.
研究人员提出了一种结合深度学习与图像预处理技术的OCR模型，用于自动提取韩文发票中的关键信息，在自建数据集上达到了87%的F1分数，且处理耗时极低。
The article compares AI classifiers Jev, Quinn, Leia, and Jeff for document tasks such as language detection and classification, showing Jev and Quinn as top performers across accuracy, cost, and latency.
Formalini 是一款新产品，可将任意文档、照片、扫描件或手写笔记快速转换为 JSON、Excel、表格等结构化数据，并支持个人文档管理、企业工作流集成（UiPath、SAP、Salesforce）以及供 AI 代理调用的 MCP 服务器。
Xiaomi-OCR-0 is a unified 0.8B vision-language model for document parsing and OCR-centric understanding, built from Qwen3.5-0.8B-Base, achieving #1 on Real5-OmniDocBench and outperforming larger models like Qwen3.5-2B on OCR VQA benchmarks.
PolyOCR-Venus 提出了一系列统一的 OCR 基础模型，结合共享指令框架、大规模数据引擎与Competence-Guided Policy Optimization，并发布修订版 OCRBench v2.1 基准，在多个文档理解基准上达到最先进的性能。
提出了基于表意文字描述序列(IDS)的IDSpect方法，通过细粒度的部件级与空间关系奖励信号改进中文文本渲染，GRPO后训练Qwen-Image在LongText和GenTextEval上取得领先表现。
小米发布 Xiaomi-OCR-0 技术报告：一个基于 Qwen3.5-0.8B 的紧凑型视觉-语言模型，通过自动化数据引擎构建约 1.7 亿样本语料，并结合文本锚定与混合任务强化学习，在多个文档解析与 OCR 基准上取得 SOTA 成绩。
The author developed a deterministic image routing system that classifies images to either send as raw bytes or extract text via OCR, achieving an 88% reduction in token consumption for LLM context integration.
Invofox launches Self Serve, a document parsing API that turns any document into structured JSON with 99%+ accuracy and SLA-backed guarantees, learning from user feedback to improve over time.
Opus 5.5 scores 93.9% on ParseBench for table parsing in PDFs, outperforming Opus 5 and other models, but is expensive compared to LlamaParse.
DocJev is a free, open-source Python tool for classifying and splitting complex document packets, supporting PDF, DOCX, and PPTX with optional OCR.
DocJev is a lightning-fast open-source library for document classification and splitting using Jev, with OCR backends like liteparse and LlamaParse, claiming 6x speed improvements over GPT-5.6-luna with equivalent accuracy.
An evaluation of over 100 models for document parsing, covering frontier VLMs, open-weight VLMs, OCR tools, and open-source parsers, with results shared via parsebench.
Baidu's open-source Unlimited-OCR model processes multiple PDF pages simultaneously, outperforming baselines like DeepSeek-OCR and supporting local execution with community integrations.