🛰️ Daily AI Frontier
‹ back to 2026-08-12

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

Research Efficiency & Systems

Ranking

Overall 75
Content 80
Popularity 62

Observed public metrics from 1 member.

Merged summary

TL;DR - DistilVDR is a 524M-parameter end-to-end visual document retriever distilled from an 8B vision-language teacher, reaching ~87% of teacher quality while indexing an order of magnitude faster with a far smaller index. It matters because VDR deployment cost is currently dominated by multi-billion-parameter multi-vector models.

  • Bilateral (dual-student) distillation uses only a pointwise cosine alignment loss against the frozen teacher's embedding space — no relevance labels, negative sampling, or contrastive term.
  • Asymmetric encoder-only student mirrors VDR's text-query/image-document asymmetry: visual capacity concentrated on the document encoder, query side kept at 70M parameters.
  • DistilVDR-HiRes hits 61.74 average NDCG@5 on ViDoRe v1+v2+v3 (86.9% of the 8B teacher) and leads reproduced sub-1B baselines on the high-resolution-sensitive v3; DistilVDR-Fast reaches 59.98 with a 3x smaller visual-token budget.
  • Single-vector output yields a 15.6x smaller index for one million documents than the strongest sub-1B multi-vector baseline; code released as NanoVDR.

Sources (1)

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

arXiv cs.IR Zhuchenyang Liu, Ziyi Wang, Yao Zhang, Yu Xiao 2026-08-11 arXiv:2608.10636
Public signals Hugging Face upvotes 11
Providers: Hugging Face · Upvotes 11 OpenAlex · N/A Publisher · N/A Semantic Scholar · N/A X · N/A Fetched 2026-09-10 14:31:05.413857 UTC

TL;DR - DistilVDR is a 524M-parameter end-to-end visual document retriever distilled from an 8B vision-language teacher, reaching ~87% of teacher quality while indexing an order of magnitude faster with a far smaller index. It matters because VDR deployment cost is currently dominated by multi-billion-parameter multi-vector models.

  • Bilateral (dual-student) distillation uses only a pointwise cosine alignment loss against the frozen teacher's embedding space — no relevance labels, negative sampling, or contrastive term.
  • Asymmetric encoder-only student mirrors VDR's text-query/image-document asymmetry: visual capacity concentrated on the document encoder, query side kept at 70M parameters.
  • DistilVDR-HiRes hits 61.74 average NDCG@5 on ViDoRe v1+v2+v3 (86.9% of the 8B teacher) and leads reproduced sub-1B baselines on the high-resolution-sensitive v3; DistilVDR-Fast reaches 59.98 with a 3x smaller visual-token budget.
  • Single-vector output yields a 15.6x smaller index for one million documents than the strongest sub-1B multi-vector baseline; code released as NanoVDR.
item →