Pith. sign in

REVIEW 4 major objections 4 minor 18 references

Col-Bandit shows that late-interaction reranking can certify the exact top-K set while computing only a fraction of the token-level MaxSim scores, reducing query-time FLOPs by up to 5×.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-03 05:13 UTC pith:R3X6BOBD

load-bearing objection Genuinely new idea — adaptive MaxSim cell pruning — with consistent empirical wins, but the abstract overclaims and no configuration has a proven δ-valid guarantee. the 4 major comments →

arxiv 2602.02827 v2 pith:R3X6BOBD submitted 2026-02-02 cs.IR

Col-Bandit: Query-Time Top-K Estimation for Late-Interaction Retrieval

Roi Pony , Adi Raz Goldfarb , Oshri Naparstek , Idan Friedman , Udi Barzelay , Eli Schwartz This is my paper
classification cs.IR
keywords late-interaction retrievalColBERTMaxSim pruningtop-K identificationfinite-population banditsBernstein–Serfling boundsquery-time rerankingcompute-fidelity trade-off
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Col-Bandit targets the query-time bottleneck of multi-vector late-interaction retrievers (e.g., ColBERT), which normally compute every token-level MaxSim score for every candidate document. The paper's central claim is that most of those MaxSim entries are redundant: a document can be ruled in or out of the top-K using only the cells needed to separate the top-K from the rest under statistically grounded confidence bounds. Formulating reranking as a finite-population top-K identification problem, Col-Bandit reveals matrix entries adaptively and stops once the weakest winner's lower bound exceeds the strongest loser's upper bound. On text and multimodal benchmarks, the paper reports preserving high overlap with exhaustive top-5 reranking while reducing MaxSim FLOPs by up to 5×, with up to ~13× single-thread CPU speedups, and as a drop-in layer it needs no retraining or index changes. If this holds, late-interaction quality can be retained at a fraction of its usual cost.

Core claim

The core discovery is that the N×T matrix of token-level MaxSim scores contains exploitable redundancy of a specific kind: the top-K set can be certified after observing a small, adaptively chosen subset of the cells. For each document, Col-Bandit maintains hybrid lower/upper bounds that combine deterministic hard bounds from known token ranges with a variance-adaptive empirical Bernstein–Serfling radius, which shrinks as tokens are revealed and collapses when a row is fully scored. The algorithm repeatedly reveals cells for the weakest current winner and strongest current loser until their bounds separate, at which point the top-K is returned. With uniform-within-row reveals and αef=1 the r

What carries the argument

The load-bearing object is the finite-population empirical Bernstein–Serfling radius (Eq. 12): a confidence interval for each document's total score that uses the empirical variance of observed tokens, a finite-population correction that drives the radius to zero as the row is fully revealed, and a calibration factor αef that tunes the compute–fidelity trade-off. It is wrapped in a LUCB-style refinement loop that identifies the weakest winner and strongest loser, reveals the highest-uncertainty unrevealed token (with ϵ-greedy exploration), and stops on separation of their bounds.

Load-bearing premise

The load-bearing assumption is that the empirically tightened variance-adaptive radius (αef≈0.2 with ϵ-greedy token selection) remains a valid characterization of uncertainty when deciding to stop; if it is over-optimistic, the algorithm can halt early with a confidently wrong top-K, and only empirical calibration stands between that and the reported results.

What would settle it

Run the deployed configuration on a corpus with exhaustively computed top-5 sets, and record, for each query, the coverage at which Col-Bandit stops and whether its returned top-5 matches the exhaustive one. A single query where the algorithm stops at low coverage while returning a set that excludes a true top-5 document — or where the empirical stop-time distribution is systematically earlier than the uniform-within-row δ-valid mode's — would falsify the claim that the aggressive radius is calibrated.

Watch this falsifier. Get emailed when new claim-graph text bears on it.

If this is right

  • If Col-Bandit's results are correct, late-interaction pipelines can shrink query-time cost several-fold without changing the model or index, making state-of-the-art multi-vector retrieval practical at higher throughput.
  • The efficiency gains are largest for small K (top-1/top-5), which is the regime of precision-oriented retrieval tasks such as question answering, and the paper's experiments show savings persist, though reduced, at K=10.
  • Col-Bandit's pruning happens inside the exact scoring loop, so it is orthogonal to index-time compression and candidate-set reduction; those methods can be composed underneath to compound the speedups.
  • Because the stopping decision compares only the weakest winner against the strongest loser, the method automatically spends more compute on borderline candidates and almost none on clear winners or losers — the behavior the paper demonstrates by dominating static row-reveal baselines at matched coverage.
  • The αef knob yields a family of operating points rather than a single fixed speed–accuracy point, giving deployment teams a tunable frontier instead of a binary trade-off.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The formal δ-PAC guarantee is proven only for the uniform-within-row reveal mode with αef=1; the deployed aggressive mode with adaptive token selection is a calibrated heuristic, so production use that needs a hard certification should either stay in the valid mode or re-derive the radius.
  • The same reveal-and-certify framing could transfer to other scoring functions that decompose into bounded per-feature contributions, such as hybrid sparse-dense or feature-based rerankers, where components can be revealed on demand — the paper does not explore this.
  • The reported reductions are in FLOPs; realizing wall-clock GPU speedups likely requires the batched block-reveal implementation the paper lists as future work, so the CPU speedup numbers may not directly transfer to GPU deployments.
  • A natural untested combination is Col-Bandit on top of a compressed index (e.g., centroid or quantized token representations); since it prunes MaxSim operations inside scoring, the savings could multiply, though the interaction remains unmeasured.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. Col-Bandit proposes to identify the top-K documents in multi-vector late-interaction retrieval by adaptively revealing only the (document, query token) MaxSim entries needed to separate candidate scores. It maintains per-document deterministic hard bounds and a variance-adaptive empirical Bernstein–Serfling radius, then uses a LUCB-style rule that repeatedly compares the weakest current winner against the strongest current loser and stops when the lower bound of one exceeds the upper bound of the other. The paper reports substantial coverage/FLOP reductions on BEIR text retrieval (ColBERTv2, Jina-ColBERT-v2) and REAL-MM-RAG multimodal retrieval (Granite Vision Embedding), against two static baselines. A formal δ-PAC identification claim is made for a uniform-within-row mode with α_ef=1, while the deployed default uses ϵ-greedy token selection with α_ef<1 and is described as a calibrated heuristic.

Significance. If the reported reductions are reproducible, Col-Bandit is a practically useful drop-in query-time layer: it requires no retraining or index changes, and it is evaluated across multiple backbones and modalities. The paper's strengths include transparent ablations of exploration strategy and ANN-derived bounds, per-dataset breakdowns, and a candid Limitations section that concedes the strongest mode is a heuristic. However, the formal guarantee stated in §3.3 is not currently established, and the submitted abstract goes beyond the full text. As presented, the contribution is best viewed as an empirically calibrated pruning heuristic with promising cost-accuracy frontiers, not as a certified top-K method.

major comments (4)
  1. [Abstract; Tables 4–7] The submitted abstract claims '~13× single-thread CPU speedups' and 'cutting MaxSim FLOPs by up to ~8×' at ≥90% top-5 fidelity. The full text says 'up to 5×' and explicitly states that realizing wall-clock speedups requires batched implementations (Limitations). Tables 5–7 show a maximum top-5 saving of about 5.9× (ArguAna, 17% coverage), so even the full text's 'up to 5×' is at the edge of the data. The ~8× and ~13× numbers appear nowhere in the full text. This inconsistency must be corrected in revision.
  2. [§3.3, §4.3, Appendix A/C] The δ-PAC guarantee is not proven for the radius in Eq. 12. Appendix C claims δ-validity for uniform-within-row mode with α_ef=1, but Appendix A explicitly states that Eq. 12 'omits the O(1/n) bias term' of the empirical Bernstein–Serfling inequality. Since the Appendix C radius has the same form, the statement Pr(|S_i−Tμ̂_i| ≤ Tσ̂_i√(2log(c/δ)/n)√ρ_n) ≥ 1−δ is not a consequence of Bardenet & Maillard. Because Algorithm 1 stops when LCB_{i+} ≥ UCB_{i-}, an optimistic radius can lead to early stopping with a confidently wrong top-K. A correct bound (including the bias term and exact constants) or an explicit downgrade of the δ-PAC claim is required.
  3. [Appendix C; Alg. 1] The proof sketch in Appendix C unions over fixed sample sizes n, but the algorithm chooses documents and token positions adaptively and stops at a data-dependent time. Under adaptive selection, the set of revealed cells in a row is not a uniformly random subset of size n, so the fixed-n uniform-sampling concentration statement does not directly apply to the realized stopping rule. The remark in Appendix A that 'the procedure requires full separation of the top-K set' is informal and does not constitute an optional-stopping argument. A time-uniform or stopped-process concentration inequality is needed for any certified guarantee.
  4. [§5.1, §5.3, Abstract] The calibration of α_ef is load-bearing for the headline claim. The default α_ef=0.2 appears to be selected by sweeping α_ef on the same BEIR and REAL-MM-RAG corpora whose fidelity/coverage numbers are then reported; §5.1 says operating points are generated by sweeping α_ef over [10^-3,1]. Table 1 reports per-dataset coverage needed to hit 90%/95% overlap, so the savings are per-target operating points, not one fixed configuration. The abstract's 'preserves ≥90% fidelity ... on every corpus' with an implied single setting is not supported. The authors should either evaluate a fixed default with a validation split or clearly state that results are per-dataset tuned.
minor comments (4)
  1. [Eq. 12; Appendix A] The notation in Eq. 12 writes ρn_i while Appendix A defines ρn; please clarify that ρ is indexed by the observed count n_i. Also, the text says α_ef=1 is 'unshrunk', but Eq. 12 simply multiplies the whole radius by α_ef, so α_ef=1 is not a theorem-derived radius.
  2. [Eq. 13–14] For n_i≤1 the text sets r_i^eff=+∞ and relies only on hard bounds. Since LCB_i and UCB_i are clipped to the hard bounds, this is well-defined, but a brief worked example for n_i=1 would help.
  3. [Table 1 caption] The 'Savings (vs. Full)' column is defined as 100%/Mean coverage; the 'up to' language in the abstract should be tied to the best per-dataset point, not the mean. Please make this explicit in the table caption.
  4. [Appendix C] The constant c in the logarithmic term is not defined. If c comes from Bardenet & Maillard's theorem, give its value or explain that it is absorbed into α_ef.

Circularity Check

0 steps flagged

No circularity: reported fidelity/cost are measured on the algorithm's own adaptive frontier; alpha sweep and omitted-bias note are limitations, not definitional reductions.

full rationale

The derivation chain is not circular. Col-Bandit's stopping rule is an algorithmic construction; its overlap with full late-interaction scoring is measured empirically, not derived from its own confidence bounds. The variance-adaptive radius (Eq. 12) is explicitly a calibrated decision radius, and the Limitations section states the deployed mode 'should be viewed as a calibrated decision heuristic rather than a formal certificate,' so the empirical results do not masquerade as a theorem. The alpha_ef sweep used to draw the quality-coverage frontier is a standard in-sample characterization; the reported coverage-to-target-quality values are measured outcomes of the policy, not quantities reconstructed from alpha_ef. The Appendix A admission that Eq. (12) omits the O(1/n) bias term, and the resulting tension with Appendix C's delta-validity claim, is a proof gap / correctness risk, not circularity. Self-citations (REAL-MM-RAG benchmark, Granite-Vision model) are data/model citations and do not supply a load-bearing theorem or ansatz. No equation-level reduction of a prediction to a fitted input was found.

Axiom & Free-Parameter Ledger

4 free parameters · 4 axioms · 0 invented entities

The algorithm contributes an adaptive cell-reveal policy; the load-bearing extras are (i) a variance radius whose formal guarantee does not cover the deployed configuration, (ii) an in-sample calibrated α_ef, and (iii) exact-ANN-derived token bounds in the experiments. No new physical entities are proposed.

free parameters (4)
  • α_ef (calibration knob) = 0.2 (default; swept over [1e-3,1])
    Eq. 12's radius is scaled by α_ef. The default is chosen on the evaluation corpora rather than from the theory, since the δ-validity proof only covers α_ef=1 with uniform-within-row sampling.
  • δ (error tolerance) = 0.01
    Used in the radius and PAC statement; fixed in experiments with no sensitivity analysis.
  • ϵ (exploration probability) = 0.1
    Hand-set in the dynamic ϵ-greedy policy. Section 4.1 promises a sensitivity study of ϵ, but Section 5.3 does not report one.
  • k' (ANN neighbors per query token) = 10
    Determines candidate set size and the token-level bounds in Eq. 15; a deployment choice rather than part of the core algorithm.
axioms (4)
  • domain assumption Every MaxSim entry H_{i,t} lies in a known bounded interval [a,b]
    Needed for the hard bounds in Eqs. 10-11 and for the Bernstein-Serfling support assumption; for cosine similarity on normalized vectors this is [-1,1] (Section 2.1).
  • ad hoc to paper The empirical Bernstein-Serfling radius remains valid in the deployed ϵ-greedy mode with α_ef<1
    Appendix C proves δ-validity only for uniform-within-row sampling with α_ef=1. Algorithm 1 uses ϵ-greedy token selection and α_ef=0.2, relying on calibration rather than a certificate, as the Limitations admit.
  • domain assumption The ANN stage provides valid per-token upper bounds for unretrieved documents
    Eq. 15 sets b_{i,t}=s_{k'}^{(t)} for documents not retrieved for token t; this is only valid if the ANN search is exact enough that any higher MaxSim would have been retrieved. Experiments use precomputed exact kNN per token, but real approximate ANN indexes may violate the bound.
  • standard math Union bounding over N documents and n=1..T sample sizes with δ/(NT) gives time-uniform control
    Appendix C uses a standard union bound. It relies on fixed N and fixed T, which hold in the evaluation setup.

pith-pipeline@v1.3.0-alltime-deepseek · 18252 in / 14704 out tokens · 150409 ms · 2026-08-03T05:13:31.265493+00:00 · methodology

0 comments
read the original abstract

Multi-vector late-interaction retrievers such as ColBERT achieve state-of-the-art quality, but their query-time cost is dominated by exhaustively computing token-level MaxSim interactions for every candidate document. The MaxSim scores of $N$ candidates against $T$ query tokens form an $N\times T$ matrix whose row-sums are the late-interaction scores, and identifying the top-$K$ rarely requires every entry. We introduce Col-Bandit, a query-time estimator of the exhaustive-MaxSim top-$K$: it reveals matrix entries in batches, maintains a finite-population Bernstein-Serfling confidence interval on each candidate's score, and permanently drops any document whose upper bound falls below the $K$-th largest lower bound, computing only the cells needed to separate the top-$K$. A single relaxation knob $\alpha_{\mathrm{ef}}\in(0,1]$ tunes the compute-fidelity trade-off. We deploy $\alpha_{\mathrm{ef}}{=}0.2$, while $\alpha_{\mathrm{ef}}{=}1$ admits a $\delta$-PAC guarantee under a simplified radius. On BEIR and REAL-MM-RAG, Col-Bandit preserves $\geq 90\%$ fidelity to the exhaustive top-$5$ on every corpus while cutting MaxSim FLOPs by up to ${\sim}8\times$, for up to ${\sim}13\times$ single-thread CPU speedups across x86 and ARM. A drop-in reranking layer, it needs no retraining or index changes.

Figures

Figures reproduced from arXiv: 2602.02827 by Adi Raz Goldfarb, Eli Schwartz, Idan Friedman, Oshri Naparstek, Roi pony, Udi Barzelay.

Figure 1
Figure 1. Figure 1: Schematic of Col-Bandit. Given a query (e.g., ”human mobility...”) and a set of candidate documents (e.g., Nature, Auto), the goal is to identify the Top-2 relevant documents. (A) Full ColBERT determines the exact score for every document by summing all interaction cells (MaxSims), requiring 100% of the compute budget. (B) Col-Bandit approximates these sums using partial cell observations. By adaptively re… view at source ↗
Figure 2
Figure 2. Figure 2: Cost-Accuracy trade-off for Col-Bandit compared to Random Reveal (Doc-Uniform) and Greedy Top-Margin (Doc-TopMargin) across three retrieval settings (text and multimodal). Each star marker denotes a Col-Bandit operating point obtained by sweeping the relaxation parameter αef. The top-right corner (Overlap@5 = 1.0, Cost = 100%) corresponds to full exhaustive scoring. these per-query-token maxima: S(d; Q) ≜ … view at source ↗
Figure 3
Figure 3. Figure 3: Taxonomy of efficient late-interaction retrieval. Meth￾ods are classified by when they prune (index-time vs. query-time) and what they prune. Col-Bandit is the first to dynamically prune the atomic interaction matrix H during query-time scoring. Identification in bandits (Kalyanakrishnan et al., 2012; Chen et al., 2014). The multi-armed bandit (MAB) framework has been extensively studied for resource-const… view at source ↗
Figure 4
Figure 4. Figure 4: Exploration Strategy Ablation. Trade-off on Jina ColBERTv2 / HotPotQA. The dynamic ϵ-greedy policy (pur￾ple) consistently dominates static warm-up schedules (green), avoiding wasteful reveals on easy queries [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Cost-Accuracy trade-off for Col-Bandit compared to Random Reveal (Doc-Uniform) and Greedy Top-Margin (Doc-TopMargin) across three retrieval settings (text and multimodal). Each star marker denotes a Col-Bandit operating point obtained by sweeping the relaxation parameter αef. The top-right corner (Overlap@K=1.0, Cost=100%) corresponds to full exhaustive scoring . 14 [PITH_FULL_IMAGE:figures/full_fig_p014_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Cost-Accuracy trade-off for Col-Bandit compared to Random Reveal (Doc-Uniform) and Greedy Top-Margin (Doc-TopMargin) across three retrieval settings (text and multimodal). Each star marker denotes a Col-Bandit operating point obtained by sweeping the relaxation parameter αef. The top-right corner (Overlap@K=1.0, Cost=100%) corresponds to full exhaustive scoring . B.2. Extended Retrieval Effectiveness (Top-… view at source ↗
Figure 8
Figure 8. Figure 8: Effect of ANN bounds and calibration. Quality–coverage trade-off for Col-Bandit with and without ANN-derived token bounds across four document retrieval settings. The without ANN bounds variant corresponds to uniform-within-row token reveals; in this setting, the unshrunk radius (αef = 1) matches the conditions of the empirical Bernstein–Serfling bound (Appendix C). Fix a document i and let Oi be the set o… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

18 extracted references · 11 linked inside Pith

  1. [1]

    and Bubeck, S

    Audibert, J.-Y . and Bubeck, S. Best arm identification in multi-armed bandits. InCOLT-23th Conference on learning theory-2010, pp. 13–p,

    2010

  2. [2]

    This stage is the computational bottleneck

    by evaluating allN×TMaxSim operations, revealing all matrix cells. This stage is the computational bottleneck. A.2. Datasets and Models We evaluate Col-Bandit on five widely used text retrieval datasets from the BEIR benchmark (Thakur et al., 2021):Ar- guAna(Wachsmuth et al., 2018),Quora(Thakur et al., 2021),SciDocs(Cohan et al., 2020),NQ(Kwiatkowski et a...

    2021

  3. [6]

    URL https://arxiv.org/abs/2405.19504. Ding, S. and Suel, T. Faster top-k document retrieval using block-max indexes. InProceedings of the 34th interna- tional ACM SIGIR conference on Research and develop- ment in Information Retrieval, pp. 993–1002,

    Pith/arXiv arXiv

  4. [12]

    A study on token pruning for colbert.arXiv preprint arXiv:2112.06540,

    Lassance, C., Maachou, M., Park, J., and Clinchant, S. A study on token pruning for colbert.arXiv preprint arXiv:2112.06540,

    Pith/arXiv arXiv

  5. [13]

    PLAID: An efficient engine for late interaction retrieval

    Santhanam, K., Khattab, O., Potts, C., and Zaharia, M. PLAID: An efficient engine for late interaction retrieval. InProceedings of the 31st ACM International Conference on Information & Knowledge Management, CIKM ’22, pp. 1747–1756, New York, NY , USA, 2022a. Associa- tion for Computing Machinery. doi: 10.1145/3511808. 3557325. Santhanam, K., Khattab, O.,...

    doi:10.1145/3511808 2022

  6. [14]

    Best arm identifi- cation for prompt learning under a limited budget.arXiv preprint arXiv:2402.09723,

    Shi, C., Yang, K., Yang, J., and Shen, C. Best arm identifi- cation for prompt learning under a limited budget.arXiv preprint arXiv:2402.09723,

    Pith/arXiv arXiv

  7. [16]

    R., Schwartz, E., Barzelay, U., and Karlinsky, L

    Wasserman, N., Pony, R., Naparstek, O., Goldfarb, A. R., Schwartz, E., Barzelay, U., and Karlinsky, L. Real-mm- rag: A real-world multi-modal retrieval benchmark.arXiv preprint arXiv:2502.12342,

    Pith/arXiv arXiv

  8. [18]

    Yang, Z., Qi, P., Zhang, S., Bengio, Y ., Cohen, W., Salakhut- dinov, R., and Manning, C

    URL https: //arxiv.org/abs/2507.05513. Yang, Z., Qi, P., Zhang, S., Bengio, Y ., Cohen, W., Salakhut- dinov, R., and Manning, C. D. Hotpotqa: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 2018 conference on empirical methods in natural language processing, pp. 2369–2380,

    Pith/arXiv arXiv 2018

  9. [19]

    In our framework, the relaxation factor αef practically compensates for this approximation

    Theorem 4.3. In our framework, the relaxation factor αef practically compensates for this approximation. Furthermore, while the stopping time is adaptive, the procedure requires full separation of the top-K set, making it substantially less sensitive to optional stopping risks compared to classical sequential hypothesis tests. A.1. Two-Stage Retrieval Pip...

    2020

  10. [1998]

    Team, I. R. Granite-vision-3.3-2b-embedding, 2025a. URL https://huggingface.co/ibm-granite/ granite-vision-3.3-2b-embedding. Team, N. Nomic embed multimodal: Interleaved text, image, and screenshots for visual document retrieval, 2025b. URL https://nomic.ai/blog/posts/ nomic-embed-multimodal. Thakur, N., Reimers, N., R ¨uckl´e, A., Srivastava, A., and Gur...

    Pith/arXiv arXiv

  11. [2011]

    DESSERT: An efficient algorithm for vector set search with vector set queries

    Engels, J., Coleman, B., Lakshman, V ., and Shrivastava, A. DESSERT: An efficient algorithm for vector set search with vector set queries. InAdvances in Neu- ral Information Processing Systems 36 (NeurIPS 2023),

    2023

  12. [2014]

    Reducing the footprint of multi-vector retrieval with minimal per- formance impact via token pooling.arXiv preprint arXiv:2409.14683,

    Clavi´e, B., Chaffin, A., and Adams, G. Reducing the footprint of multi-vector retrieval with minimal per- formance impact via token pooling.arXiv preprint arXiv:2409.14683,

    Pith/arXiv arXiv

  13. [2015]

    URL https: //arxiv.org/abs/1309.4029

    doi: 10.3150/14-BEJ605. URL https: //arxiv.org/abs/1309.4029. Broder, A. Z., Carmel, D., Herscovici, M., Soffer, A., and Zien, J. Efficient query evaluation using a two-level re- trieval process. InProceedings of the twelfth international conference on Information and knowledge management, pp. 426–434,

    Pith/arXiv arXiv

  14. [2019]

    org/abs/1902.09465

    URL https://arxiv. org/abs/1902.09465. Jha, R., Wang, B., G ¨unther, M., Mastrapas, G., Sturua, S., Mohr, I., Koukounas, A., Akram, M. K., Wang, N., and Xiao, H. Jina-colbert-v2: A general-purpose multilingual late interaction retriever.arXiv preprint arXiv:2408.16672,

    Pith/arXiv arXiv 1902

  15. [2020]

    MUVERA: Multi-vector retrieval via fixed dimensional encodings

    Dhulipala, L., Hadian, M., Jayaram, R., Lee, J., and Mir- rokni, V . MUVERA: Multi-vector retrieval via fixed dimensional encodings. InAdvances in Neural Informa- tion Processing Systems 37 (NeurIPS 2024),

    2024

  16. [2021]

    K., Mohr, I., Ungureanu, A., Wang, B., Eslami, S., Martens, S., Werk, M., Wang, N., et al

    G¨unther, M., Sturua, S., Akram, M. K., Mohr, I., Ungureanu, A., Wang, B., Eslami, S., Martens, S., Werk, M., Wang, N., et al. jina-embeddings-v4: Universal embeddings for multimodal multilingual retrieval. InProceedings of the 5th Workshop on Multilingual Representation Learning (MRL 2025), pp. 531–550,

    2025

  17. [2023]

    Faysse, M., Sibille, H., Wu, T., Omrani, B., Viaud, G., Hudelot, C., and Colombo, P

    URL https://openreview.net/forum? id=kXfrlWXLwH. Faysse, M., Sibille, H., Wu, T., Omrani, B., Viaud, G., Hudelot, C., and Colombo, P. Colpali: Efficient document retrieval with vision language models.arXiv preprint arXiv:2407.01449,

    Pith/arXiv arXiv

  18. [2024]

    Cohan, A., Feldman, S., Beltagy, I., Downey, D., and Weld, D. S. Specter: Document-level representation learning using citation-informed transformers.arXiv preprint arXiv:2004.07180,

    Pith/arXiv arXiv 2004