AI
ColNanoVDR: Document-Free Query Distillation for Multi-Vector Visual Document Retrieval via Optimal Transport
Publication: Hugging Face Daily Papers
Author: Not available in collected metadata
Source published: 27/09/2026, 20:00:00 UTC
Briefing created: 29/09/2026, 15:09:35 UTC
Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 29/09/2026, 07:18:24 UTC.
Briefings
ColNanoVDR is the first document-free distillation framework for multi-vector visual document retrieval, replacing large teacher encoders with small student models. By using entropic optimal transport with learned weights, the method aligns student queries with teacher embeddings without requiring tokenization correspondence, proving the alignment cost bounds MaxSim score differences. Distilled from five state-of-the-art teachers, the 149M text-only students retain 95% of their teachers' NDCG@5 on ViDoRe v1-v3 while encoding queries up to 26x faster and using 12.6x less cached data.
Key points
- ColNanoVDR is the first document-free distillation framework for multi-vector visual document retrieval, replacing large teacher encoders with small student models.
- The method uses entropic optimal transport with learned weights to align student queries with teacher embeddings without requiring tokenization correspondence, proving the alignment cost bounds MaxSim score differences.
Context and limitations
The framework was tested on five state-of-the-art teachers and ViDoRe v1-v3, with results showing 95% NDCG@5 retention and 26x speedup; however, generalizability to other multi-vector models and unseen datasets remains unverified.
Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 29/09/2026, 15:09:35 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.
View archived briefing on IPFS
bafybeiaiowkqn4ydd2f4qfwebyylh3zo5kp34eczrjvbssmuxq77lenix4