ExtractBench：14个文档提取系统各有感知盲区

Codex在扫描和手写场景准确率超93%，但处理旋转或纯图像页面时降至约80%。专用API表现相反，旋转和手写识别良好，扫描准确率为81%。

Gemini 3.5 Flash从干净PDF的88.

6%跌至扫描件的71.1%。

LlamaIndex的Agentic Plus是唯一无盲区系统，在旋转、扫描和手写三项上分别达到95.9%、93.

9%和93.8%。

Every document extraction system has a perception blind spot. We mapped them. For ExtractBench, we tested 14 systems on documents that weren't born digital: 1950s regulatory filings, hand-filled tax forms, and pages degraded with fax thresholding, photocopier tone curves, sensor noise, and phone-camera capture. The failures don't overlap. 💡 Codex reads scans and handwriting above 93%, then drops to ~80% on rotated or image-only pages. 💡 Specialized APIs are the exact inverse: fine on rotation and handwriting, 81% on scans 💡 Gemini 3.5 Flash falls from 88.6% to 71.1% the moment a page is scanned. You benchmark on clean PDFs. Production sends you a shadowed photocopy from 1953. Our new Extract Tier, Agentic Plus, was the only system with no blind spot: 95.9% / 93.9% / 93.8% across rotated, scanned, and handwritten, a 2-point spread where others swing 10+. Learn more about ExtractBench 👇️ Bl lnkd.in/gNm97fXp ceB6 Pap lnkd.in/euAfScWx kQpF

This report is categorized under AITOP's "Models" desk. Key entities and technical topics under continuous tracking include ExtractBench, Codex, Gemini 3.5 Flash, LlamaIndex.

Source attribution: curated from LlamaIndex (@llama_index). Cross-source corroboration: 2 independent sources have covered this development (including Jerry Liu), confirming industry-wide signal strength.