[CVPR 2026] Official project page for
M3Grounder: Mask-Based Multi-Span and Multi-Granular Grounding for Document QA
BharatGen & IIIT Hyderabad
🔗 Paper · HF Space · Code · Poster
M3Grounder grounds document QA answers at pixel level — not just bounding boxes. It autoregressively generates answer text interleaved with [GROUND] tokens that link each answer span to its evidence region on the page, at phrase, line, and block granularity.
Key ideas:
- Mask-based grounding — fine-grained segmentation instead of coarse boxes
- Multi-span — one answer can cite multiple evidence regions
- Multi-granular — phrase ⊂ line ⊂ block, enforced by an enclosure loss
- Bleed-suppression loss — prevents masks from spilling into irrelevant regions
Static project page built with plain HTML/CSS/JS. Includes an interactive demo showcasing multi-span, multi-granular grounding across Finance, Legal, and Transcript documents.
index.html # entire site (single file)
static/
images/ # logos, teaser, document images
To run locally:
python -m http.server 8000
# open http://localhost:8000@inproceedings{venna2026m3grounder,
title = {M3Grounder: Mask-Based Multi-Span and Multi-Granular Grounding for Document QA},
author = {Venna, Venkata Kesav and Gunda, Sai Madhusudan and Jinka, Jyothi Swaroopa
and Rachakonda, Hrithik Sagar and Srinivasan, Anirudh and Sarvadevabhatla, Ravi Kiran},
booktitle = {Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
pages = {23685--23695},
year = {2026}
}