Generative Late-Interaction Embeddings For Visual Document Retrieval
GLIE stores only a few vectors per page, then regenerates full page embeddings only for top candidates before rescoring.
The paper says late-interaction visual document retrieval is accurate but storage-heavy, with roughly 1,000 vectors per page. Its authors argue that page vectors sit on the unit sphere and cluster near a low-dimensional manifold, which makes aggressive compression less destructive if centroids are normalized and used generatively. With four vectors per page on ViDoRe v1, GLIE keeps nearly 80% of the uncompressed system’s nDCG@5, compared with 70% for the best prior post-hoc method. The reported decoder is small, 415K parameters, and was trained in under three GPU-minutes on 1,000 pages.
HF Daily Papers' note
The paper says late-interaction visual document retrieval is accurate but storage-heavy, with roughly 1,000 vectors per page. Its authors argue that page vectors sit on the unit sphere and cluster near a low-dimensional manifold, which makes aggressive compression less destructive if centroids are normalized and used generatively. With four vectors per page on ViDoRe v1, GLIE keeps nearly 80% of the uncompressed system’s nDCG@5, compared with 70% for the best prior post-hoc method. The reported decoder is small, 415K parameters, and was trained in under three GPU-minutes on 1,000 pages.
HF Daily Papers' note
score 5