YC 创业公司 Unsiloed AI 在 olmOCR-Bench 上超过 GPT-5.5
做文档自动化或者 RAG 的人可以看看 Unsiloed AI。它在 Allen AI 的 olmOCR-Bench 上拿了 88.0 分,比 GPT-5.5 的 84.6 分还高,专啃跨页表格、贷款件这类难文档,而且每个字段带置信度分数,金融场景更敢用。
YC 孵化的 Unsiloed AI 宣布其文档解析器在 Allen AI 的公开基准 olmOCR-Bench 上拿到 88.0 分,超过 GPT-5.5 的 84.6 分。olmOCR-Bench 包含 1,403 份 PDF 和 8,413 个单元测试,专门考察真实世界中的复杂文档。
这些文档包括贷款文件、保险理赔单,以及表格跨三页以上的年报。解析环节每读错一个单元格,下游智能体就会带进一个错误数字,所以 4 分的差距在实际应用中影响不小。
Unsiloed 把 PDF、扫描件和表格转成 Markdown 和 JSON 供智能体使用,并给每个提取出的字段附上置信度分数。这个设计对金融场景尤其关键,因为用户可以在置信度低时触发人工复核,而不是全盘相信解析结果。
A YC startup is out-reading GPT-5.5 on messy PDFs.
On olmOCR-Bench, Allen AI's public test with 1,403 PDFs and 8,413 unit tests, Unsiloed AI says its parser scores 88.0. GPT-5.5 lands at 84.6.
Four points sounds small until you look at what these documents are. Loan files, insurance claims, annual reports with tables that run over three pages. Every misread cell becomes a wrong number in whatever the agent does next.
Unsiloed turns PDFs, scans and spreadsheets into Markdown and JSON for agents, with a confidence score on every extracted field. That last part is what I'd want before trusting any of this in finance.
This report is categorized under AITOP's "Products" desk, focusing on US frontier AI labs and Silicon Valley technology leaders. Within our cross-regional analytical framework, it directly maps to AI developer toolchains, coding agents, and engineering workflows. Key entities and technical topics under continuous tracking include Unsiloed AI, GPT-5.5, olmOCR-Bench, OCR.
Source attribution: curated from kimmonismus (@kimmonismus).