Document QA
FinMMDocR
Multi-step numerical reasoning over financial page images where the evidence sits in the image, not a handed text layer.
As of 13 August 2026. Conditional pack. BUPT-Reasoning-Lab FinMMDocR, AAAI 2026. Recheck data terms. No Fable 5 or GPT-5.6 official table is cited.
FinMMDocR tests multi-step numerical reasoning over financial page images. Evidence in the image is a different treatment from TAT-QA's already-structured tables.
Until terms are approved, the pack cannot enter a commercial artifact. Paper system tables stay in the paper until replayed on a frozen revision.
- Scale
- AAAI 2026 multimodal financial-document benchmark. Images and text ship in separate folders.
- Provenance
- BUPT-Reasoning-Lab FinMMDocR, AAAI 2026. Recheck data terms before acquisition.
- Contamination
- Medium. Public since 2026. Recheck license before a commercial run.
- Cannot establish
- A 2026 VLM product score, or TAT-QA substitution. No official Fable 5 or GPT-5.6 table is cited here.
Metrics, kept separate
- Official task accuracy
- Scenario awareness
- Multi-step computation accuracy
No cited public figure is installed for this pack. Method, scale, and gates still apply. A missing chart is not a pass rate.
What the figures bound
The pack is listed as the multimodal financial-document diagnostic the market reaches for when the text layer is missing. That is why it is here. It is not TAT-DQA, and it is not a VLM product score.
No chart is installed without a cited official cell.
What this page does not conclude
- A commercial run before terms are approved
- GPT-5.6 or Fable 5 visual financial QA
- Hebbia or AlphaSense document-product performance
- An averaged multimodal winner
Cited as calibration on these memos — not a score of those products.
Not attached to a software, advisory, or automation memo yet.