From ColBERT to CITADEL: how late interaction learned to live inside an inverted index

Asked:“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 citation graph of 14 papers, 2020–2025, tracing how ColBERT-style contextual token retrieval was made indexable: lexical inverted lists (COIL, CSurF), learned lexical routing (CITADEL), sparsified late interaction (SLIM, SPLATE), centroid inverted lists (ColBERTv2), and the adjacent engine branch (PLAID, WARP). Solid arrows are explicit citation edges verified in the source pages; dashed arrows are conceptual lineage.

Recommended reading path

Start with ColBERT (2020) for late interaction, then COIL (2021) for contextual token vectors in lexical inverted lists, then CITADEL (2023) for learned dynamic routing, then SLIM and SPLATE for fully sparse late interaction. For the systems branch, follow ColBERTv2 → PLAID → WARP — efficient multi-vector engines built on centroid inverted lists rather than lexical ones.

COIL stores a contextualized vector per token in the term's inverted list, so each query token scores only documents that contain it — exact match with deep representations — aclanthology.org
CITADEL replaces the exact-match constraint with learned dynamic lexical routing of token vectors, and CSurF calls it concurrent work on controlled soft match — aclanthology.org, cs.cmu.edu
ColBERTv2 inverts by k-means centroids with quantized residuals — an inverted index over vector space, not vocabulary — and explicitly cites COIL and SPLADE — ar5iv.labs.arxiv.org
SLIM's verified reference list draws on CITADEL, PLAID, SPLADE v2 and the DeepImpact/COIL conceptual notes to sparsify late interaction onto standard inverted indexes — dl.acm.org
WARP (SIGIR 2025) cites ColBERT, ColBERTv2, PLAID, COIL, CITADEL and SPLATE, but is an efficiency engine for the centroid branch, not a lexical inverted-index method — dl.acm.org

The papers

PaperYearVenueIndex mechanismSource

Curated important-paper graph, not a complete bibliometric census. Built from 63 scholarly source rows across four result sets; 14 papers kept, 22 solid arrows drawn only where a source page explicitly lists the cited title, 5 dashed conceptual links. Years are formal venue years where available; arXiv preprint years can differ (CITADEL preprint 2022, ACL 2023). Unrelated 2025–2026 downstream papers were cut for scope.

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 · 6,068 pages in 11m 38s · Ask your own questionShare:XLinkedInReddit
Made with Keenable SELECT