Research
Teaching machines to read documents — and to remember them.
Documents carry meaning in their text, their layout and their images at once, and they change over time. My work goes from learning joint vision-language representations of a page, to answering questions only from evidence, to the state a system should keep across documents and sessions.
§ 01 Current agenda
Persistent, evidence-grounded document state.
Survey · arXiv:2609.32041 · September 2026
Amnesia by Design, Memory By Necessity: Persistent State for Document Intelligence
Systems extract fields, reason over tables and ground answers to page regions, yet process each document as if it were the first. An amendment read the next day starts from scratch: no schema retained, no contradiction detected. It formalizes the state, its lifecycle and invariants, bounds the risk of remembering wrong things, audits ten benchmarks, and proposes a longitudinal harness with five counterfactual metrics.
Companion · arXiv 2026
Fresh at Decision, Stale at Action: Routing New Information to LLM Decisions Already Made
Decisions made by LLMs are carried out later; a decision ledger keeps them valid.
Grounding · arXiv 2025
MTRACE: Multilingual Retrieval-Augmented Generation for Temporally Diverse Text Corpora
Multilingual RAG that cites its evidence and abstains when it has none.
§ 02 Themes
Seven threads, 29 publications.
Each publication carries one or more themes. Follow a thread, or filter the full list.
Document understanding10
Cross-modal vision-language representations for classifying, retrieving and summarizing visually rich documents.
-
arXiv 2026Preprint
-
WACV Workshops 2025Workshop
-
WACV 2025Conference
-
ICDAR 2024Conference
-
Pattern Recognition 2023Journal
-
SSRN 2023Preprint
-
PhD thesis 2022Thesis
-
IJDAR 2021Journal
-
ICIP 2020Conference
-
CVPR Workshops 2020Workshop
Persistent state & decisions2
What document systems should retain across documents and sessions: provenance-linked state, supersession, governed forgetting, and decisions that stay valid after they are made.
-
arXiv 2026Preprint
-
arXiv 2026Preprint
Grounded retrieval & QA2
Retrieval-augmented generation that answers only from evidence, cites it, and abstains when the evidence is insufficient.
-
arXiv 2025Preprint
-
WACV 2025Conference
Scene text & low-resource scripts8
Detection, recognition and generation of text in complex, low-resource scripts such as Khmer, including handwriting.
-
ICPR Workshops 2026Workshop
-
WACV Workshops 2026Workshop
-
ICDAR 2025Conference
-
ICDAR Workshops 2025Workshop
-
ICDAR Workshops 2025Workshop
-
arXiv 2025Preprint
-
ACCV 2024Conference
-
ACET 2024Conference
Trust, identity & security3
Identity documents, document quality, and reputation systems for evaluating language models.
-
WACV Workshops 2025Workshop
-
COMPSAC 2024Conference
-
ICDAR Workshops 2019Workshop
Language & translation5
Neural machine translation for specialized domains, knowledge distillation, compression, and knowledge-graph-to-text generation.
-
arXiv 2026Preprint
-
EAMT 2026Conference
-
AICS 2025Conference
-
arXiv 2025Preprint
-
Applied Sciences 2025Journal
Graph learning2
Graph representations and benchmarks for structured document data.
-
IJDAR 2026Journal
-
GbRPR 2025Workshop
§ 03 Lineage
Where my models sit among document representations.
Table I of the survey traces representation learning for documents from BERT to instruction-tuned layout models. EAML, VLCDoC and GlobalDoc are mine. Every row shares one empty column: native state.
| Method | Year | Modality | Pretraining | Layout | Visual | OCR-free | Native state |
|---|---|---|---|---|---|---|---|
| BERT | 2019 | Text | MLM | – | – | – | – |
| LayoutLMv1–v3 | 2020–22 | Text+Layout+Vision | MLM/MIM | ✓ | ✓ | – | – |
| EAML My work: EAML: Ensemble Self-Attention-Based Mutual Learning Network for Document Image Classification | 2021 | Image+Text | Mutual learning | – | ✓ | – | – |
| SelfDoc | 2021 | Cross-modal | Self-supervised | ✓ | ✓ | – | – |
| UDoc | 2021 | Cross-modal | Unified alignment | ✓ | ✓ | – | – |
| DiT | 2022 | Vision | MIM | Implicit | ✓ | ✓ | – |
| Bi-VLDoc | 2022 | Cross-modal | Bidirectional VL | ✓ | ✓ | – | – |
| Donut | 2022 | Vision→Text | Seq2seq | Implicit | ✓ | ✓ | – |
| VLCDoC My work: VLCDoC: Vision-Language Contrastive Pre-Training Model for Cross-Modal Document Classification | 2023 | Cross-modal | VL contrastive | – | ✓ | – | – |
| Pix2Struct | 2023 | Vision→Text | Seq2seq | Implicit | ✓ | ✓ | – |
| UDOP | 2023 | Vision+Text+Layout | Generative | ✓ | ✓ | Partial | – |
| DocLLM / LayoutLLM | 2024 | Vision+Text+Layout | Instruction tuning | ✓ | ✓ | Partial | – |
| GlobalDoc My work: GlobalDoc: A Cross-Modal Vision-Language Framework for Real-World Document Image Retrieval and Classification | 2025 | Cross-modal | Unified VL | ✓ | ✓ | – | – |
13 representation models for documents, BERT (2019) to GlobalDoc (2025). Rows marked My work are my models and link to their publication pages. Source: Table I of the paper.
§ 04 People & places
Done with others.
I co-supervise PhD students on low-resource scene-text recognition and explainable multimodal document understanding, and work with groups across Europe and Asia.
Collaborations
- Computer Vision Center (CVC), Universitat Autònoma de Barcelona · Spain
- KU Leuven · Belgium
- L2TI · France
- South East Technological University (SETU) · Ireland
- Jožef Stefan Institute · Slovenia
- University of Science and Technology of Hanoi (USTH) · Vietnam
- Cambodia Academy of Digital Technology · Cambodia
- MNS University · Pakistan
- Allread ML Technologies · Spain, industry
- Siren · France, industry
Earlier work: PhD research statement (2019–2022).