Vision-language models have made OCR-free visual document retrieval possible by representing documents as multiple vectors. But those representations come at a cost: models generate a large number of tokens to preserve fine visual detail, and that token bloat severely limits efficiency. A new arXiv paper, Unbalanced Optimal Transport for Efficient Visual Document Retrieval, takes on this bottleneck.
Rather than relying on the usual balanced matching between query and document tokens, the paper proposes an unbalanced optimal transport formulation. Based on the title and abstract, this is intended to make retrieval more efficient while retaining the advantages of multi-vector, OCR-free representations.
The abstract is brief and does not include quantitative results or implementation details. The contribution, as described, is a response to a known practical problem: dense visual representations are powerful but expensive.