Landing AI tops DocVQA with a record 99.16% accuracy

Anyone who has built a document parser or a question-answering tool for contracts knows the pain: every single time a user asks something new, you have to push the heavy scanned image back through a vision-language model. It is slow, and the token bills pile up fast.

Landing AI took a fundamentally different route and quietly broke a major benchmark. The team hit 99.16% accuracy on the DocVQA benchmark, passing Qwen2-VL's previous record of 97.25% and topping human performance, which typically lands between 96% and 98%.

Parse once, query forever

Standard DocVQA setups are rigid. Every time a question arrives, the model ingests the original document image alongside the prompt. That works in benchmark leaderboards, but in real production environments, running multi-modal inferences repeatedly is prohibitively expensive.

Landing AI decoupled seeing the document from answering questions about it, adopting a parse-once, query-unlimited workflow.

First, a specialized model looks at the visual layout. Landing AI used its in-house DPT-2 (Document Pre-trained Transformer 2) to convert the scanned page into structured outputs in a single pass. This includes Markdown text, JSON schemas, and precise visual bounding boxes with layout coordinates and confidence scores.

Second, question answering becomes purely textual. Downstream queries bypass visual models entirely. The experiment fed the structured text and bounding box context directly into Claude Sonnet 4.5 using a hybrid prompting strategy.

The benefits are immediate. You process the heavy image exactly once. Whether you ask one question or a hundred, downstream costs drop to cheap text-only inference. Even better, because every field carries spatial coordinates, every single answer can be traced straight back to its location on the original scan.

Accounting for 45 errors

The evaluation ran across the DocVQA validation set of 1,286 real-world scanned documents from the UCSF industry library. The test covers fact retrieval, complex tables, and spatial reasoning under strict case-insensitive exact matching.

The performance jumps were clear. Using raw Markdown alone yielded 95.36%. Once JSON-based visual grounding and guided prompts were added, accuracy climbed to 99.16%.

The team was unusually candid about its failures. Out of the entire set, they found only 45 true errors: 13 came from upstream parsing flaws like OCR confusion or occlusions, 18 stemmed from LLM reasoning mistakes, and 9 were tricky spatial questions outside the core parsing scope. Another 18 cases were excluded due to ambiguous ground-truth labels in the original dataset.

To prove it, Landing AI open-sourced the pre-parsed files, hybrid prompt templates, evaluation scripts, predictions, and an interactive gallery displaying 63 successful cases and 45 error cases.

Why it matters

In our view, the most impressive part of this result is not just the 99.16% number, but the clear reminder that massive end-to-end multimodal models are not the only viable way forward for enterprise documents.

There is a prevailing belief that feeding raw pixels directly into giant frontier models is the future. But in daily business operations, brute-force multimodal calls run into ugly realities: massive GPU memory footprints, spiraling inference bills, and zero auditability when an answer hallucinates.

Decoupling the parser from the reasoning engine looks like classic engineering pragmatism, but paired with modern LLMs, it offers a dramatic leap in efficiency. For teams dealing with millions of contracts, invoices, and compliance reports, this means documents can be parsed once into structured offline assets. Everything downstream, from RAG search to privacy auditing, becomes faster and vastly cheaper.

The risk, of course, is error propagation. If the front-end layout model misreads a crucial table cell or drops a footnote, the downstream LLM has no way to recover. Even so, by proving that modular architecture can outperform brute-force vision models while cutting operational overhead, Landing AI may force enterprise teams to rethink how they build document workflows.