Luming

CTO @SoMarkAI | Seeking the optimal path from carbon to silicon intelligence.

[Luming @soul_code98]

If OXR works for you, give it a ⭐ on GitHub. https://t.co/q5KydRIUtI

[Luming @soul_code98]

Different workflows can need different layouts from the same page. That’s why OXR lets you replace layout analysis while keeping recognition and reading order.

[Binyang Qiu @BinyangQiu]

Document parsing has a long-overlooked performance tax: reading order.

Where you put that task matters. Run it as a separate stage, and parsing takes longer: reading order adds another sequential step. Combine it with layout analysis, and preparing training data gets harder: you need to label both the layout and the reading order.

In OXR, our open-source document parsing model, we run recognition and reading order in parallel within the same model, keeping layout analysis separate. This saves time while also making reading order content-aware. The separated layout model can be lighter and focus on detecting regions and their types, preserving layout accuracy while further speeding up parsing.

Keeping layout analysis separate also matters when adapting parsing to a business. The same page might need different structures: one workflow needs each image as a separate region; another needs the whole group kept together. Both are valid.

We made OXR’s layout analysis pluggable so you can swap in a model that fits your business while keeping OXR’s recognition and reading order capabilities.

I’ve written more about the design and trade-offs in “The Long-Overlooked Performance Cost in Document Parsing: Reading Order.”

Full article in the reply below.

[Luming @soul_code98]

@BinyangQiu I think of reading order as more than sorting boxes. Knowing what’s inside them helps you decide what comes next.

[Luming @soul_code98]

If you find OXR helpful, leave a ⭐ on GitHub. https://t.co/q5KydRImEa

[Luming @soul_code98]

With OXR’s RSI loop, parsing failures become a source of targeted training data—without manually labeling new samples. Binyang explains how we put that into practice.

[Binyang Qiu @BinyangQiu]

Errors as data: RSI helped us build a small document parsing model with SOTA accuracy and speed.

We built OXR, our open-source document parsing model, around the idea that a small model can go a long way if you systematically find its weaknesses and build the right training data.

Real documents keep exposing new weaknesses. But turning those failures into training data usually means finding bad examples, diagnosing them, and manually labeling new samples. We wanted model iteration to move without waiting for another round of manual data preparation.

So we built a recursive self-improvement (RSI) loop into post-training. To our knowledge, OXR is the first document parsing model to use this approach.

Each round, OXR parses unlabeled documents. A general VLM evaluates the results and analyzes errors and uncertainty; disagreement between OXR’s predictions for the same document also flags uncertain samples. A clustering model groups issues by cause, helping the VLM plan the next round of OXR training and generate targeted data.

This loop requires no manual labeling of new training data, speeding up model iteration.

I think this idea is worth exploring beyond document parsing, wherever clear, testable outcomes let failures guide what a model should learn next.

OXR is one example of how we've put this idea into practice: a 1.3B model with SOTA accuracy and speed. And we think there's still room to make it smaller.

Full write-up in the next post 👇

[Luming @soul_code98]

@BinyangQiu What I like about this loop: errors help us figure out what OXR needs to learn next. Tried OXR? We’d love your feedback ��️ https://t.co/seMzW7D0wj

[Luming @soul_code98]

@cudanexus @somarkai Thanks for asking! We went with the top models on olmOCR-Bench and OmniDocBench when we ran our tests.

[Luming @soul_code98]

@LoganMarkewich @somarkai Thanks! We focused on olmOCR-Bench and OmniDocBench at launch. If you feel like trying OXR on PareBench, I’d be curious to see how it does too!

[Luming @soul_code98]

@whispersofninja @somarkai Thanks for sharing! We haven’t run OXR on this benchmark yet. If you feel like giving it a go, we’d love to see your results!

[Luming @soul_code98]

@ngendlio @somarkai Thanks for asking! indra reports 47% table TEDS on OmniDocBench, while OXR reports 96.62%.——but Indra’s CPU efficiency is impressive! You can try both in your own documents

[Luming @soul_code98]

@somarkai Luming here, CTO at SoMark 👋 Give OXR a try on your own documents. Happy to answer questions about getting started!

[Luming @soul_code98]

@An_yhl @somarkai 你是否想了解：https://t.co/Ibbdng3Kfo，商业版支持这项功能🎁

[Luming @soul_code98]

If OXR works for you, give it a ⭐ on GitHub.
https://t.co/q5KydRIUtI

[Luming @soul_code98]

There are always documents we haven’t seen. By open-sourcing OXR, we hope more people can help us find those gaps and improve the model together.

[Binyang Qiu @BinyangQiu]

The next batch of documents may come from an unfamiliar industry, in another language, with a different layout. Can the model still parse it reliably?

It’s a question that keeps coming up in our work.

To tackle it, we turned to AI4AI—using AI to help improve AI—and built a recursive self-improvement (RSI) loop for post-training 🔁.With help from a general vision-language model, parsing errors and uncertain results inform the next training plan and targeted data generation.

This process helped us build SoMark, our production-grade document parsing engine. But as we served more customers, we realized there were still many kinds of documents we hadn't seen.

🎯 That's why we decided to open-source OXR. We hope the community will bring new use cases, reproducible problems, and domain knowledge, helping us improve the model and bring those improvements back to developers.

Developers also need accuracy, speed, and flexibility💡. We redesigned OXR to run recognition and content-aware reading order in parallel within one model, with independent, pluggable layout analysis. This speeds up parsing while preserving layout accuracy and lets developers customize layout analysis for their business needs.

Every one of the decisions serves one goal：keep improving how the model handles unfamiliar documents, and make those improvements easier to use.

Full story below 👇

[Luming @soul_code98]

@BinyangQiu Happy to share what we’ve been working on. Curious to see what people build with OXR!

[Luming @soul_code98]

Here’s what we’ve been working on — meet OXR! If you find OXR helpful, leave a ⭐ on GitHub.

[SoMark @somarkai]

👇 See OXR in action.

OXR recognizes 13 element types, including formulas, tables, chemical equations, chemical structures and code blocks, with bounding box coordinates.

It also handles newspapers, research papers, financial reports, resumes, exams, handwriting and more.

[Luming @soul_code98]

@somarkai Luming here, CTO at SoMark 👋 I’d start with the online demo. Try a document you work with regularly and see how it parses.

[Luming @soul_code98]

@somarkai Tiny dev PSA: let it pick an answer, not take the wheel. Your code still decides what happens next.

[Luming @soul_code98]

The MCP bit is my kind of rabbit hole. Hook it up to Codex or Claude Code, throw a document question at it, see what you can wire up. Have at it 🎁

[SoMark @somarkai]

Give your agent a document-focused version of JEV.

We’ve open-sourced Docision!!! 🚀

Inspired by JEV, Docision adds multimodal capabilities and is fine-tuned for document tasks. With just 4B parameters, it lets your agent classify receipts and check forms on your laptop.

Give it a try and let us know what you’d use it for. Already using JEV? We’d love to hear how Docision could fit into your workflow.

If you find it helpful, give it a star⭐
https://t.co/djYmP2B0zX

[Luming @soul_code98]

哈哈哈哈哈哈哈哈哈哈哈

[FomskyWei @Fomsky_Wei]

好了睡前我想到了一个绝妙的点子
然后我觉得我貌似成全网第一乐子人了..我现在真的很迷茫
我觉得这个事件会和某mac电鸡事件一样陪伴我很久了

你想办法自己去赚80美元重置一下,谢谢

卧槽codex你tmd给我退钱.今晚又睡不着了.
去你大爷的退钱!!!我让你自己去赚钱啊草!!!
我现在很迷茫.谁能救救我.