PulseAugur
EN
LIVE 08:02:17

DistilVDR: Compact Visual Document Retriever Achieves 87% of Teacher Model Performance

By PulseAugur Editorial · [3 sources] ·

Researchers have developed DistilVDR, a compact end-to-end visual document retriever that uses dual-student distillation. This new system, with 524 million parameters, is distilled from a larger 8 billion parameter teacher model, eliminating the need for relevance labels or negative sampling. DistilVDR offers two variants, DistilVDR-HiRes and DistilVDR-Fast, which achieve performance close to the teacher model while significantly reducing index size and improving indexing speed compared to existing sub-1 billion parameter models. AI

IMPACT This research offers a more efficient approach to visual document retrieval, potentially enabling wider adoption of large-scale document analysis systems.

RANK_REASON The cluster describes a new research paper detailing a novel model for visual document retrieval.

Read on arXiv cs.IR (Information Retrieval) →

AI-generated summary · Google Gemini · from 3 sources. How we write summaries →

DistilVDR: Compact Visual Document Retriever Achieves 87% of Teacher Model Performance

COVERAGE [3]

  1. arXiv cs.CL TIER_1 English(EN) · Zhuchenyang Liu, Ziyi Wang, Yao Zhang, Yu Xiao ·

    DistilVDR: A Compact End-to-End Visual Document Retriever via Dual-Student Distillation

    arXiv:2608.10636v1 Announce Type: cross Abstract: Visual document retrieval (VDR) is dominated by multi-billion-parameter models that are slow to index at full corpus scale and expensive to serve. Prior compression routes either train a smaller multi-vector encoder from scratch o…

  2. arXiv cs.IR (Information Retrieval) TIER_1 English(EN) · Yu Xiao ·

    DistilVDR: A Compact End-to-End Visual Document Retriever via Dual-Student Distillation

    Visual document retrieval (VDR) is dominated by multi-billion-parameter models that are slow to index at full corpus scale and expensive to serve. Prior compression routes either train a smaller multi-vector encoder from scratch or distil only the query side; neither yields a com…

  3. Hugging Face Daily Papers TIER_1 English(EN) ·

    DistilVDR: A Compact End-to-End Visual Document Retriever via Dual-Student Distillation

    DistilVDR is a compact 524M vision-document retriever distilled from an 8B teacher using cosine alignment without relevance labels, achieving near-teacher accuracy with far smaller indexes and faster indexing.

RELATED ENTITIES

RELATED TOPICS