After 2023 the ColBERT-with-inverted-index family branches into six distinct approaches

Asked:“you can include less popular papers that arose starting from 2023” — following “could you show me citation graph of important papers for information retrieval that use colbert with inverted index — you can start with COIL and CITADEL papers”

A curated lineage of 23 papers, 2020–2025, connected by 33 explicit citation edges taken from the papers themselves. Solid, saturated nodes use a lexical inverted or sparse index directly; lighter, dashed nodes are adjacent branches — token-retrieval optimization, pruning, centroid streaming, bit-vector and fixed-dimensional alternatives, and retrieval engines. No citation counts or popularity metrics were measured.

Solid outline and saturated colour = the paper builds directly on a lexical inverted or sparse index. Dashed outline and muted fill = adjacent branch that does not literally use one. Small nodes are less-popular or niche references surfaced by the 2025 pruning paper. Arrows point from cited paper to citing paper.

The 2025 lossless token-pruning paper is the graph’s widest join: it explicitly cites 17 ancestors, from ColBERT and COIL to niche pruning work such as ColBERTer, learned token pruning and static pruning — arxiv.org
WARP (SIGIR 2025) merges the two main 2023 branches, explicitly citing both the sparse-index line (CITADEL, SPLATE, COIL) and the token-retrieval line (XTR) on top of ColBERTv2 and PLAID — dl.acm.org
2023 is the branching year: CITADEL (ACL), SLIM (SIGIR), COILcr (ECIR), CSurF (ICTIR) and XTR (NeurIPS) each take the COIL/ColBERT idea in a different indexing direction — aclanthology.org
PLAID SHIRTTT (SIGIR 2024) extends the centroid line to streaming terabyte-scale collections via hierarchical sharding, citing ColBERT, ColBERTv2 and PLAID — user.eng.umd.edu

Curated papers

PaperYearVenueBranchHow it indexes / retrievesSource

Further candidates / adjacent work

A discovery appendix of plausible post-2022 indexed late-interaction and multi-vector papers found during the search. These are not evidence of direct citation edges in the graph above; scope labels are approximate and metadata may be partial.

PaperYearVenueSource

Curated technical lineage, not an exhaustive citation-index export. Built from a 14-paper curated node table, 33 deduplicated explicit citation edges, a 161-row evidence set and a 166-row candidate table gathered from ACL Anthology, ACM DL, NeurIPS, arXiv, author pages and Google Research. Year corrections applied: XTR is NeurIPS 2023, CITADEL is ACL 2023, ColBERTv2 and PLAID are 2022; SPLATE and XTR title variants deduplicated. Arrows encode only citations stated in the papers; no citation counts or popularity measures were taken. Appendix trimmed to distinct titles for space.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.

Keenable · made with SELECT · 2,871 pages in 9m 28s · Ask your own questionShare:XLinkedInReddit
Made with Keenable SELECT