Yes — a quality-focused literature exists for ColBERT as a pre-ranker, and the OOD late-interaction reranker paper is its anchor

Asked:

“ok, and now lets check if there're important papers that focus on colbert quality instead of simplicity of retrieval -- like if we want to use our colbert as a preranker after real retrieval”

This follow-up scans the quality-focused slice of the expanded citation graph — 8 papers selected from 291 focused evidence records and a 257-row candidate appendix — and places each one on the pipeline stage it improves. What the IR literature calls a second-stage reranker is exactly your “pre-ranker”: a ColBERT-style MaxSim scorer over externally retrieved candidates, optionally before a final cross-encoder. Once candidates arrive from outside, ColBERT’s own ANN and index machinery is unnecessary.

The pipeline, with each paper on the stage it improves

1 · Candidate retriever

Your “real retrieval” — BM25, dense, or hybrid. Guard recall; ColBERT never sees what this stage drops.

Papers that train this stage

→

2 · ColBERT MaxSim pre-ranker

Multi-vector late interaction scores only the candidate set. No ANN, no inverted index — just token-level MaxSim. This is where the quality literature lives.

Papers that improve this stage

→

optional

3 · Final reranker

A cross-encoder or listwise model over the top of the pre-ranked list, at the smallest depth.

Adjacent modern rerankers

What the visual cannot say alone

Essential reading. Zhang, Li & Lin (2023) is the most directly aligned early paper: contextual late interaction as a reranker over candidates from first-stage retrievers of different kinds, with about 5% average improvement on out-of-distribution datasets, little added latency, and no in-domain degradation. openreview.net · arxiv.org

Modern anchor. Martinico, Nardini, Rulli & Venturini match your deployment pattern name-for-name — strong first stage, then multi-vector reranking — but note it is an ECIR 2026 publication and the extracted result fields are blank, so no candidate depths or numbers are claimed here. link.springer.com · arxiv.org

Feedback, not just reranking. ColBERT-PRF (journal version 2023) mines feedback embeddings from the initially retrieved documents and reports gains up to 26% MAP on TREC 2019 passage queries and 10% on TREC 2020; the paper says it can enhance reranking of initial results as well as a second retrieval. dl.acm.org · eprints.gla.ac.uk

Training beats indexing. ColBERTv2’s denoised supervision is the piece to reuse for a pre-ranker; TCT-ColBERT distills ColBERT into a single-vector model, so treat it as teacher evidence, not a reranker recommendation. arxiv.org · arxiv.org

Adjacent, licensed carefully. jina-reranker-v3 (2025) is listwise late interaction rather than classic MaxSim; a 0.6B model is on Hugging Face under non-commercial CC BY-NC 4.0. Extracted benchmark fields are blank, so no numbers are claimed. arxiv.org · huggingface.co

Practical starting point for the build

Retrieve generously enough to protect recall — the pre-ranker cannot recover what stage 1 drops
Rerank the top 100–1000 depending on latency budget and document length
Batch and cache document encodings when documents repeat across queries
Train on hard negatives sampled from your actual production first stage
Evaluate oracle recall@K of stage 1 and post-rerank nDCG/MRR separately
Test under distribution shift, not only in-domain
Compare against a cross-encoder at the same latency before committing

The top-100–1000 range is an engineering starting point, not a finding from a cited paper.

Decision matrix

Paper	Year	Pipeline role	Quality lever	First-stage independence	Best use	Evidence strength	Source
OOD generalization with contextualized late interaction	2023	Stage 2 · multi-vector reranker	Late interaction on top of a reranker	High — tested over retrievers of diverse natures	Essential reading; your exact pattern	Strong: ~5% avg OOD gain reported	arxiv.org
Multivector reranking, strong first-stage era	2026	Stage 2 · multi-vector reranker	Reranking framing for modern first stages	High by design	Modern anchor	Bibliographic only; result fields blank	arxiv.org
ColBERT-PRF	2023 (TWEB)	Stage 2 · feedback/expansion	Feedback embeddings from initial results	Medium — needs an initial retrieved set	Recall and expansion on top of candidates	Strong: up to 26% MAP TREC 2019, 10% TREC 2020	dl.acm.org
ColBERTv2	2022	Stage 2 · training recipe	Denoised supervision, residual compression	High — recipe transfers to reranking	Train your scorer this way	Strong background	arxiv.org
ColBERTer	2022	Stage 2 · compact scorer	Enhanced reduction, whole-word bags	High	Quality per byte under memory limits	Moderate	arxiv.org
TCT-ColBERT	2020	Stage 1 · distillation	ColBERT as tightly-coupled teacher	n/a — produces a single-vector retriever	Teacher/distillation lineage only	Moderate; not a multi-vector reranker	arxiv.org
Cross-architecture knowledge distillation	2021	Stage 1/2 · training	Margin-MSE across architectures	High	Loss function for student training	Moderate lineage evidence	arxiv.org
jina-reranker-v3	2025	Stage 3 · listwise reranker	Listwise “last” interaction, not MaxSim	High	Adjacent modern option; check CC BY-NC 4.0	Model available (0.6B); benchmarks blank	huggingface.co

Evidence: six extraction result sets over the expanded ColBERT citation graph — 19 rows on the 2023 OOD reranker, 7 on ColBERT-PRF, 21 on multivector reranking (ECIR 2026), 38 on jina-reranker-v3, 206 focused evidence rows, and a 257-row quality-focused candidate appendix covering 69 distinct titles. Numbers are quoted only where an extraction row states them; blank fields stay blank. The appendix was filtered to papers whose own contribution improves ColBERT effectiveness or scores external candidates; generic mentions were cut. This page complements, not replaces, the earlier index-focused graph.

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.