ColNanoVDR: Document-Free Query Distillation for Multi-Vector Visual Document Retrieval via Optimal Transport
A 149M text-only student can query existing multi-vector VDR indexes at up to 26x the speed while keeping about 95% of teacher NDCG@5.
ColNanoVDR distills multi-vector visual document retrievers without encoding or caching the training pages. Its OTW objective aligns student and teacher query tokens with entropic optimal transport, avoiding any need for matching tokenizations. The authors say the alignment cost bounds the MaxSim score difference on every page. In identical training, OTW matches score distillation while reading 12.6x less cached teacher data. HF Daily Papers' note
ColNanoVDR distills multi-vector visual document retrievers without encoding or caching the training pages. Its OTW objective aligns student and teacher query tokens with entropic optimal transport, avoiding any need for matching tokenizations. The authors say the alignment cost bounds the MaxSim score difference on every page. In identical training, OTW matches score distillation while reading 12.6x less cached teacher data. HF Daily Papers' note
score 4