AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

Rethinking the Role of Token Retrieval in Multi-Vector Retrieval

Jinhyuk Lee, Zhuyun Dai, Sai Meher Karthik Duddu, Tao Lei, Iftekhar Naim, Ming-Wei Chang, Vincent Y Zhao

NeurIPS 2023 · Architectures and efficiency › Attention and transformers · approach: Language model pretraining and induction · 7 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
train token retrieval with in-batch top-k token alignment rather than document-level sum-of-max alignmentT5-based multi-vector retrieval · BEIR (13 datasets)nDCG@10 average: 50.8 -> 52.7Improved
score candidates from retrieved tokens and impute missing similarities rather than gathering and scoring all document tokensXTR · MS MARCO-derived scoring estimateFLOPs/query: 0.36 × 10^9 -> 0.09 × 10^6Improved
reduce token retrieval depth from k′=40,000 to k′=1,000 (ablation)XTR-base · BEIR (13 datasets)nDCG@10 average: 49.1 -> 47.9Hurt
reduce token retrieval depth from k′=40,000 to k′=1,000 (ablation)XTR-base · MS MARCOnDCG@10: 45.0 -> 43.2Hurt
train multilingual retrieval on English MS MARCO plus MIRACL rather than English MS MARCO alonemXTR-xxl · MIRACL (18 languages)nDCG@10 average: 63.5 -> 65.9Improved

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy