CVPR Findings2026
Hi3Doc: Hierarchical Tri-Level Representations for Multimodal Long-Document Understanding
Wanying Zhou, Zhuo Chen, Jianzhi Lu, Chenxi Ma, Weimin Tan, Bo Yan
Abstract
Real-world document question answering remains highly challenging for AI models, largely due to two core factors: document length and multimodal complexity. Long documents can span hundreds of pages with answers crossing page boundaries, while visually rich documents combine heterogeneous modalities such as images, charts, tables and layouts. Existing methods typically handle only one side of this problem, leaving a unified treatment of intra-page multimodal representations and inter-page relations missing. We propose Hi3Doc, the first hierarchical and unified framework for multimodal long-document understanding. Hi3Doc models documents at three semantic levels in a top-down manner-chunk level, page level, and element level-yielding comprehensive representations from global structure to local fine-grained content. During retrieval, Hi3Doc adopts a dual-branch strategy that combines chunk-level semantic matching with elementlevel feature matching. Extensive experiments on multiple benchmarks demonstrate that Hi3Doc achieves stateof-the-art retrieval performance and strong QA capability. Notably, Hi3Doc is particularly effective on cross-modal QA and challenging queries whose supporting evidence is distributed across multiple pages. The code is available at https://github.com/czzz79/Hi3Doc.
Ask Lune
Know this paper in minutes, not an afternoon.
Lune can read every page of it and the top-tier papers that cite it, and gives you its claims, its evidence and its limits, each one cited to the paper it comes from. No skimming, no AI slop.
Lineage
Builds on25
- Chain of Agents: Large Language Models Collaborating on Long-Context TasksYusen Zhang, Ruoxi Sun, Yanfei Chen, Tomas Pfister et al.NeurIPS 2024 · 297 citations
- SlideVQA: A Dataset for Document Visual Question Answering on Multiple ImagesRyota Tanaka, Kyosuke Nishida, Kosuke Nishida, Taku Hasegawa et al.AAAI 2023 · 178 citations
- Document Understanding Dataset and Evaluation (DUDE)Jordy Van Landeghem, Rafal Powalski, Rubèn Tito, Dawid Jurkiewicz et al.ICCV 2023 · 130 citations
- A Human-Inspired Reading Agent with Gist Memory of Very Long ContextsKuang-Huei Lee, Xinyun Chen, Hiroki Furuta, John F. Canny et al.ICML 2024 · 106 citations
- DocFormerv2: Local Features for Document UnderstandingSrikar Appalaraju, Peng Tang, Qi Dong, Nishant Sankaran et al.AAAI 2024 · 68 citations
Related papers
- DREAM: Integrating Hierarchical Multimodal Retrieval with Multi-page Multimodal Language Model for Documents VQAJinxu Zhang, Qiyuan Fan, Yongqi Yu, Yu ZhangACM MM 2025
- Attention as Selector: Unlocking VLM Attention for Long Document Page RetrievalMinfeng Zhu, Linxin Bao, Wei Chen, Linchao ZhuACL 2026
- SV-RAG: LoRA-Contextualizing Adaptation of MLLMs for Long Document UnderstandingJian Chen, Ruiyi Zhang, Yufan Zhou, Tong Yu et al.ICLR 2025
- SimpleDoc: Multi-Modal Document Understanding with Dual-Cue Page Retrieval and Iterative RefinementChelsi Jain, Yiran Wu, Yifan Zeng, Jiale Liu et al.EMNLP 2025 · 1 citation
- CREAM: Coarse-to-Fine Retrieval and Multi-modal Efficient Tuning for Document VQAJinxu Zhang, Yongqi Yu, Yu ZhangACM MM 2024 · 4 citations
