GLIE retains 80% of ViDoRe nDCG@5 at four vectors per page

At four vectors per page, down from the roughly one thousand vectors that late-interaction visual retrieval typically stores, a new method called GLIE keeps nearly 80% of the uncompressed system's nDCG@5 on the ViDoRe v1 benchmark. The best prior post-hoc compression stalls at 70%.

The arxiv preprint describing it, "Generative Late-Interaction Embeddings For Visual Document Retrieval," attributes the gain to a geometry observation. Across three encoders the authors tested, the page vectors "lie exactly on the unit sphere and concentrate near a manifold of intrinsic dimension five to six." Because that manifold has so few degrees of freedom, the paper argues, "the full set of vectors can be regenerated from only a few."

GLIE trains a small network to do that regeneration. Search runs against k << N vectors per page, and a decoder then expands only the top candidates back to all N vectors for exact rescoring. The paper reports the whole system uses "a 415K-parameter network fitted in under three GPU-minutes on just a thousand training pages."

A smaller free-lunch finding sits inside the same paper. Standard k-means centroids fall inside the sphere, causing MaxSim scores to be under-estimated; normalizing them to the surface is worth "up to +0.093 nDCG@5 over raw centroids" with no retraining. At a matched training budget, the authors add, fine-tuning the encoder "does not reach even the training-free stage of GLIE."

The abstract publishes no query-time latency numbers for the decoder-expansion step, and the reported results run on ViDoRe v1 and v2 only.