AiWo at EMNLP 2026: Three papers on efficient visual document retrieval
We’re delighted to share that one of AiWo’s doctoral researchers, Zhuchenyang Liu, has three papers accepted to the 2026 Conference on Empirical Methods in Natural Language Processing (EMNLP 2026), a flagship global venue in natural language processing and artificial intelligence.
- Conference: https://2026.emnlp.org/
- Date: 24 Oct 2026 – 29 Oct 2026, Budapest, Hungary
About the research
Visual document retrieval (VDR) finds relevant pages in large collections of visually rich documents, such as financial reports with charts, scientific papers with figures, and industrial manuals with diagrams. Rather than relying on Optical Character Recognition (OCR) pipelines, VDR treats document pages directly as images, preserving layouts, tables, and figures. Despite its high retrieval accuracy, the paradigm faces adoption hurdles: state-of-the-art retrievers rely on multi-billion-parameter models that require GPUs, and multi-vector approaches produce prohibitively large search indices.
The three papers tackle this problem from complementary angles, compressing the query encoder, the document index, and finally the entire retrieval system.
Accepted papers and key contributions
- NanoVDR: Distilling a 2B Vision-Language Retriever into a 70M Text-Only Encoder for Visual Document Retrieval (Main Conference)
Observes that documents are visually complex while queries are just short text, and exploits this asymmetry: a large vision-language model indexes documents offline, while a distilled text-only encoder of only 69M parameters handles queries. The resulting model retains 95.1% of the teacher’s retrieval quality with 50× lower query latency on CPU. - Structural Anchor Pruning: Training-Free Multi-Vector Compression for Visual Document Retrieval (Main Conference)
Introduces a training-free, query-agnostic method that identifies the most informative image patches from the middle layers of the model. It reduces index storage by 90% while retaining 93–96% of retrieval quality, without any retraining or per-model tuning. - DistilVDR: A Compact End-to-End Visual Document Retriever via Dual-Student Distillation (Workshop on Document Intelligence and Understanding)
Compresses both sides of retrieval into a 524M end-to-end system distilled from an 8B vision-language teacher, requiring no relevance labels during training and producing a compact single-vector index.
Why it matters
Together, these contributions make high-accuracy visual document search feasible on edge and consumer-grade hardware. This directly addresses stringent privacy requirements in industrial settings, where proprietary technical documentation and enterprise documents cannot leave local infrastructure. For fieldworkers, this opens the way to fast, on-premise search over manuals, drawings, and maintenance records.
Authors
DistilVDR: Zhuchenyang Liu, Yao Zhang, Yu Xiao (Aalto University); Ziyi Wang (Independent Researcher, Netherlands)
NanoVDR: Zhuchenyang Liu, Yao Zhang, Yu Xiao (Aalto University)
Structural Anchor Pruning: Zhuchenyang Liu, Ziyu Hu, Yao Zhang, Yu Xiao (Aalto University)