GLIE regenerates full page embeddings from 4 vectors, slashing visual document retrieval storage
_reachsumit · x · 2026-09-11
A new arXiv paper presents Generative Late-Interaction Embeddings (GLIE) to fix the storage cost of late-interaction visual document retrieval:
- Geometric findings: across three encoders, page embedding vectors lie exactly on the unit sphere and concentrate near a manifold of intrinsic dimension 5-6, enabling regeneration from few vectors.
- Free correction: k-means centroids fall inside the sphere, systematically underestimating MaxSim; normalizing them yields up to +0.093 nDCG@5.
- Method: learn k<<N vectors per page as a lightweight index; search runs on these k vectors, and a decoder expands only top candidates back to all N vectors for exact rescoring — no encoder retraining needed.
- Results: with just 4 vectors per page on ViDoRe v1, GLIE retains nearly 80% of uncompressed performance, beating prior compression methods.
Related event: GLIE Paper Compresses Late-Interaction Retrieval Storage 200x(4 posts)→
More from Research
- DeepMind researcher: all of mathematics will soon be formalizable, with agents hunting proofs — ChrSzegedy · 2026-10-28
- A fine-tuned 'cat obsession' control fools every AI welfare criterion — MirrorEthic_Anchor · 2026-10-28
- Sweeping every Qwen3-8B neuron, rated for weirdness by Haiku — Sauers_ · 2026-10-28
- Kevin Buzzard: Lean formalization revealed errors in theorems by 4 Fields medalists — Machine Learning Street Talk · 2026-10-28
- Follow-up repo machine-verifies Hamiltonian cycle results on top of OpenAI/Anthropic ETH refutation claims — xennygrimmato_ · 2026-10-28
- DiscoTrace wins EMNLP 2026 Best Paper: LLM answers lack humans' community-specific rhetorical diversity — TuhinChakr · 2026-10-28