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.
More from Research
- Survey proposes PAPAV: a unified framework for multimodal embodied agents — MikeShou1 · 2026-09-11
- Terence Tao warns AI math capability strongly correlates with cyberattack ability — sytelus · 2026-09-11
- European Math Society hails OpenAI's Navier–Stokes solution, flags closed-model access concern — i_dg23 · 2026-09-11
- Debate over "slop math": academia must adapt verification norms for AI output — _sathvikr · 2026-09-11
- Alibaba's RecGPT Brings Three LLMs into Taobao's Recommender Pipeline — _reachsumit · 2026-09-11
- Meta's Auto-RecSys Deploys Autonomous Research Agents for Multi-Day Recommender Experiments — _reachsumit · 2026-09-11