Pith. sign in

REVIEW 4 major objections 5 minor 15 references

Context-dependent phrase vectors give token-level retrieval quality while storing fewer document vectors, by learning which spans should match jointly.

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

2026-08-04 01:23 UTC pith:2HTETY63

load-bearing objection Genuinely novel retrieval-unit learning, but the headline claims are reranking gains, not full-corpus retrieval gains — worth reading, worth revising, worth sending to referees. the 4 major comments →

arxiv 2608.00065 v1 pith:2HTETY63 submitted 2026-07-29 cs.AI cs.LG

H+ Embedding: Harmonizing Global and Token-Level Retrieval with Context-Dependent Phrases

Shusen Zhang , Junyi Hu , Ye Feng , Ziteng Wang , Zhaoyuan Pan , Guosheng Dong , Xiaojun Yuan , Jiangshou Hong
show 1 more author
Xiangzhi Wang
This is my paper
classification cs.AI cs.LG
keywords retrieval granularitycontext-dependent phraseslate interactionmulti-vector retrievalbudgeted unit selectionconditional random fieldmedical information retrievalweighted MaxSim
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.

The paper tries to establish that context-dependent phrases are a useful retrieval unit between a single global vector and token-level late interaction. It introduces H+ Embedding, a shared-encoder retriever that predicts variable-length phrase boundaries with a conditional random field, keeps uncovered tokens as singletons, and selects units by learned importance under a fixed vector budget. Across 16 scientific, medical, and bilingual tasks, the phrase branch improves over the global branch by 6.91 macro nDCG@10; in matched-candidate comparisons it nearly matches token-level interaction (62.94 vs. 62.95) while using 13.7% fewer document vectors. A sympathetic reader would care because this suggests practical retrieval systems can approach token-level accuracy at lower storage and scoring cost by making the retrieval unit itself learnable.

Core claim

The central claim is that retrieval-unit design is a joint problem of context-dependent partitioning and budgeted unit selection, and that solving it yields an intermediate quality-cost operating point. H+ Embedding uses a shared bidirectional encoder to produce global, phrase, and lexical views; a CRF labels each token as beginning, inside, or outside a phrase; units are mean-pooled contextual vectors; an importance head assigns softplus scores summed per unit; and at inference the top-B units are kept and scored by weighted MaxSim. The paper demonstrates that on terminology-intensive benchmarks, these learned phrase units consistently outperform the global branch, nearly match token-level

What carries the argument

The central object is the retrieval phrase: a contiguous, context-dependent span whose tokens act jointly as one unit of local relevance. Boundaries are predicted by a linear-chain Conditional Random Field (CRF) trained on teacher-provided BIO labels, with uncovered tokens kept as singleton units; each unit is a mean-pooled ℓ2-normalized vector. A shared importance head computes per-token scores, summed within each unit, and the top-B units by importance are retained under a per-text vector budget. Scoring uses weighted MaxSim, where query-side importance weights each unit's maximum similarity to a document unit. This machinery makes retrieval granularity itself learnable rather than fixed b

Load-bearing premise

The headline gains of Phrase over Global are measured when Phrase reranks only the top-1,000 documents that Global already retrieved from the full corpus, so the improvement could come from better reranking rather than from phrase units finding documents Global would miss.

What would settle it

Run Phrase as an independent full-corpus retriever on the same 16 tasks, indexing all document phrase vectors and searching without a Global candidate gate, under a comparable vector budget. If its nDCG@10 drops to or below Global's, the claimed gain is a reranking artifact rather than a retrieval-unit effect.

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

If this is right

  • Phrase retrieval improves over global single-vector retrieval by 6.91 macro nDCG@10 across 16 scientific and medical tasks, and it is better on 15 of 16 tasks.
  • Under a shared candidate pool, the phrase branch nearly matches token-level late interaction (62.94 vs. 62.95 macro nDCG@10) while storing 13.7% fewer document vectors.
  • When query and document unit counts are matched exactly, phrase units outperform token, bigram, and random-span units at every evaluated budget from B=4 to B=128, with the advantage over content-independent grouping becoming statistically significant at B=32.
  • Ablations show that contextual boundary placement and query-side importance weighting are complementary mechanisms; replacing summed phrase weights with uniform weights reduces nDCG@10 by 1.66 points.
  • On a 10-million-document corpus, phrase retrieval beats global retrieval in nDCG@100 (0.8777 vs. 0.8617), suggesting the benefit persists at scale.

Where Pith is reading between the lines

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

  • The main evaluation compares Phrase reranking Global's top-1,000 candidates, so the practical system still needs a cheap first-stage gate; this paper does not yet demonstrate full-corpus phrase indexing with Phrase generating its own candidates. A natural extension is to test Phrase with its own approximate-nearest-neighbor index on the full corpus.
  • The near-token result at 13.7% lower vector count uses a fixed per-text budget; adaptive budgets that allocate more vectors to terminology-dense documents could widen the storage advantage while preserving quality.
  • Because the CRF is trained on LLM teacher labels and human boundary agreement is low, retrieval-optimal units may diverge from linguistic segmentation; a retrieval-driven phrase induction without teacher labels is a plausible next step and could reveal how much of the gain depends on the supervision source.
  • The gains are largest on medical and terminology-heavy benchmarks; the method's benefit on general-domain retrieval with less compositional vocabulary remains an open question that could be tested on broader retrieval suites.

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 / 5 minor

Summary. The paper proposes H+ Embedding, a multi-granularity bi-encoder built on a shared Qwen3 backbone, with three views: a single global vector, a CRF-predicted variable-length phrase partition with importance-guided budgeted selection and weighted MaxSim interaction, and an auxiliary lexical view. It reports three headline results: (i) the phrase view improves over the global view by 6.91 macro nDCG@10 across 16 tasks; (ii) phrase interaction nearly matches token-level interaction while using 13.7% fewer document vectors; and (iii) exact-budget comparisons show phrase units outperform content-independent grouping rules at moderate budgets. The paper is transparent that all public Phrase, Hybrid, and Lexical results are candidate-constrained reranks of the global top-1,000 (Section 4.2), and it includes controlled exact-count comparisons, bootstrap confidence intervals, Holm corrections, and explicit protocol labels.

Significance. The central idea—learning context-dependent variable-length retrieval units as an intermediate granularity between single-vector and token-level interaction—is timely and plausible. If the claims were fully supported, the paper would make a useful contribution to efficient multi-vector retrieval. The controlled experimental design is an unusual strength: candidate IDs are matched within pools, unit counts are exactly controlled, paired bootstrap CIs and Holm corrections are reported, and protocol limitations are honestly labeled. However, several headline claims go beyond what the protocols establish, especially the branch-level retrieval claim and the content-independent grouping claim. The paper's own tables (Table 14 vs. Table 15) and appendices (Appendix H) undercut these statements, so the present framing overstates the evidence.

major comments (4)
  1. [§4.2, Table 1, Abstract] The claim that 'its phrase retrieval branch exceeds the global retrieval branch by 6.91 macro nDCG@10' is not supported by the protocol. Section 4.2 states that all public Phrase, Hybrid, and Lexical results rerank the same full-corpus Global top-1,000 candidates, while Global performs full-corpus search. The observed gain is therefore a reranking gain over Global's candidates, not an independent branch comparison. Section 3.6 says branches may search independently and Eq. (2) defines a candidate union, but no experiment evaluates Phrase as a full-corpus first-stage retriever; Appendix G uses 'the same evaluation pipeline' and is also candidate-constrained. Please either report full-corpus phrase retrieval or reframe the paper's central claim as a two-stage global-gate + phrase-rerank system and adjust the abstract and contributions accordingly.
  2. [Abstract, §5.4, Table 2] The '13.7% fewer document vectors' saving is computed on the Vec./doc column, i.e., local document vectors only, in the shared-pool protocol. The deployed Phrase system also maintains a global index (Section 3.6), and the Token row also relies on an external candidate pool in this protocol. The quality-cost claim should be stated over the full system (global + local) with the same end-to-end retrieval protocol; as written, it overstates the savings. The relative saving remains roughly 13.7% when a single global vector per document is included, so this is primarily a precision/consistency issue in the abstract and RQ4 discussion.
  3. [Abstract, §5.2, Tables 14/15] The abstract's claim that phrase retrieval 'outperforms content-independent grouping rules under moderate vector budgets' is contradicted by the paper's own all-task exact-count results. In Table 14 (19-task macro), Whole word and Equal chunks are stronger than Phrase at every budget; at B=128 the values are 43.25 and 42.79 vs. 38.71. The statement is only supported for Bigram and Random spans in the non-CJK 15-task slice (Table 15) from B=32 onward. Please qualify the claim to name the comparison set (Bigram/Random, non-CJK tasks, B≥32) or remove it from the abstract.
  4. [§3.2, Appendix H] The paper says an LLM teacher provides explicit BIO supervision for the CRF, but the archived teacher arrays are identical to the CRF's outputs (Table 18 footnote: 'Archived teacher arrays match CRF (not independent)'), and the teacher checkpoint identifier was not retained (Appendix A). The human boundary audit therefore cannot independently validate the 'teacher-provided' boundary signal, and the Teacher row in Table 18 is circular. Please provide the original teacher labels or clearly state that the boundary supervision is not externally verifiable; if the labels were produced by the CRF itself, the learning mechanism described in §3.2 must be restated.
minor comments (5)
  1. [Table 1 caption] The caption does not repeat the candidate-constrained protocol; readers should be told in the main table that Phrase reranks Global top-1,000 while Global uses full-corpus search. A short parenthetical in the caption would prevent misinterpretation.
  2. [Appendix H, Table 18] Inter-annotator agreement is very low (Cohen's κ=0.166, symmetric Boundary F1=0.436), and the final reference is Annotator A. The Boundary F1 values are therefore reference-sensitive. This limitation is acknowledged in the appendix, but it should be surfaced in Section 5.3 when the 'linguistic plausibility' of the CRF units is discussed.
  3. [Appendix G] The term 'strict-level recall' is used without definition. Please define it in the appendix.
  4. [Section 2] The budgets (B_q, B_d) are used in the problem formulation before being defined. A one-sentence definition before Eq. (1) would improve readability.
  5. [Section 5.2, Table 6] The exploratory B=4 point is reported with a +2.22 Phrase-Token gain, but the main grid pre-specified B∈{8,...,128}. The text should clearly mark B=4 as exploratory and not part of the planned Holm family, as it currently does in Appendix D; consider adding the same caveat in Section 5.2.

Circularity Check

1 steps flagged

Boundary supervision reduces to the CRF's own outputs; retrieval comparisons are otherwise transparent and externally benchmarked.

specific steps
  1. self definitional [Appendix H, Table 18 footnote; Section 3.2, Eq. (18)]
    "The Teacher row is not an independent teacher-student comparison: the archived teacher boundary arrays are identical to CRF. ... † Archived teacher arrays match CRF (not independent)."

    Section 3.2 says 'An LLM teacher provides explicit BIO supervision for the CRF boundary predictor,' and Eq. (18) trains the CRF with Lseg = −log p_φ(y_BIO_q | q) − log p_φ(y_BIO_d+ | d+). Appendix H reveals that the archived teacher arrays are identical to the CRF's own outputs. Therefore the 'teacher' signal that supervises the boundary predictor is, by the paper's own admission, the model's own prediction. The claimed external, LLM-provided context-dependent boundary supervision is thus self-referential: the CRF is effectively trained to reproduce its own partitions, which are then used as the retrieval units in Eqs. (5) and (9). This does not make the retrieval scores equal to their inputs by construction, but it removes the independent grounding for the 'context-dependent phrase' inter

full rationale

The central retrieval comparisons are not circular in the equation-level sense: Phrase, Hybrid, and Lexical scores are measured on held-out MTEB tasks, and the paper explicitly discloses that these scores rerank the full-corpus Global top-1,000 candidates (Section 4.2), so the headline 6.91-point gain is a transparent two-stage evaluation rather than a prediction forced by construction. The exact-count budget sweeps, matched-candidate shared-pool comparisons, and bootstrap/Holm analyses provide external, falsifiable evidence. No load-bearing self-citation chain, uniqueness theorem, or fitted-parameter-as-prediction pattern is present. The one genuine circular step is the boundary-supervision chain: the asserted 'LLM teacher' labels are identical to the CRF's own outputs (Appendix H), making the teacher supervision self-referential. This is load-bearing for the 'context-dependent phrases' interpretation, but it does not force any of the retrieval score comparisons, so the paper is only partially self-referential rather than circular by construction.

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

The model is a learned system; its claims rest on a pretrained encoder, teacher-labeled segmentation, a specific scoring rule, and a candidate-constrained evaluation protocol. The most fragile entries are the unverifiable teacher labels and the candidate gate.

free parameters (4)
  • loss weights λ_G, λ_T, λ_L, λ_seg = not reported
    Stage-2 objective L_stage2 = λ_G L_G + λ_T L_T + λ_L L_L + λ_seg L_seg; values are not specified and affect the learned trade-off among global, token, lexical, and segmentation objectives.
  • importance head parameters (w_imp, b_imp) = learned from data
    Softplus importance scores u_{x,i} are used for phrase selection and query weighting; their parameters are trained, so the reported budgeted-selection results depend on these learned values.
  • representation width r (128 vs 1024) = 128 for main results (MRL prefix)
    All main results use MRL-128; the dimensional sweep shows global benefits from 1024 while phrase is unchanged, so the width choice interacts with the claims.
  • query/document budgets B_q, B_d = swept {4,8,16,32,64,128}
    The quality-cost trade-off is defined by these budgets; the 'moderate budget' claim depends on B ≥ 32.
axioms (5)
  • domain assumption Qwen3-0.6B-Base, with causal mask replaced by bidirectional mask, provides contextual token states suitable for global, phrase, and lexical retrieval heads.
    The entire architecture is initialized from this pretrained encoder (Section 3.1); if the base model's representations are unsuitable, the method fails.
  • domain assumption LLM teacher BIO labels are a valid supervision signal for semantic phrase boundaries.
    Segmentation supervision (Eq. 18) comes from an LLM teacher; Appendix H shows the archived teacher arrays match the CRF, so this supervision is not independently verifiable.
  • ad hoc to paper Weighted MaxSim with sum-of-softplus unit importance is an appropriate scoring rule for budgeted phrase interaction.
    Eq. 9 defines this rule; the choice of sum vs uniform weighting is a design decision, and Table 4b shows the phrase-vs-token advantage reverses under uniform query weighting.
  • ad hoc to paper The candidate-constrained evaluation (reranking Global top-1,000) is a valid way to measure phrase-branch effectiveness for the main claim.
    Section 4.2 restricts all public phrase results to Global top-1,000 candidates; the 6.91-point gain is therefore not an end-to-end phrase retrieval result.
  • domain assumption MTEB/C-MTEB tasks and nDCG@10 are reliable proxies for retrieval quality.
    All conclusions rest on these benchmarks; no theoretical guarantee links benchmark nDCG to production retrieval performance.
invented entities (1)
  • Context-dependent retrieval phrase (variable-length span unit) independent evidence
    purpose: Intermediate retrieval unit between global vector and token subword; tokens in a phrase interact jointly via mean-pooled vector and weighted MaxSim.
    The paper provides ablations (Table 4a/b) showing boundary placement and importance weighting change retrieval scores, giving a falsifiable handle. It is a computational construct, not a physical entity.

pith-pipeline@v1.3.0-alltime-deepseek · 21511 in / 15232 out tokens · 157421 ms · 2026-08-04T01:23:44.252586+00:00 · methodology

0 comments
read the original abstract

Terminology-intensive retrieval, especially in medical settings, depends on preserving multi-word entities, abbreviations, numerical constraints, and compositional concepts. However, existing representations lie at two extremes: single-vector retrievers often over-compress local relevance signals, while token-level late interaction retains every tokenizer subword at substantial indexing, storage, and scoring cost. This mismatch raises a natural question: can context-dependent phrases provide a useful retrieval unit between global vectors and tokens? We introduce H+ Embedding, a unified multi-granularity retriever that predicts variable-length phrase partitions, preserves uncovered tokens as singletons, and applies importance-guided unit selection with weighted MaxSim interaction. Across 16 scientific, medical, and bilingual tasks, its phrase retrieval branch exceeds the global retrieval branch by 6.91 macro nDCG@10. It also nearly matches Token while using 13.7% fewer document vectors and outperforms content-independent grouping rules under moderate vector budgets. Context-dependent phrase interaction therefore provides an intermediate quality-cost point between global compression and token-level interaction for practical retrieval systems.

Figures

Figures reproduced from arXiv: 2608.00065 by Guosheng Dong, Jiangshou Hong, Junyi Hu, Shusen Zhang, Xiangzhi Wang, Xiaojun Yuan, Ye Feng, Zhaoyuan Pan, Ziteng Wang.

Figure 1
Figure 1. Figure 1: Selected-task nDCG@10 under the CRF-only dimensional setting. Teal denotes H [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of H+ Embedding. A shared encoder produces global, phrase-level, and lexical representations, grouping medical concepts such as “type 2 diabetes” and “chronic kidney disease” into phrase-level retrieval units. • We identify a granularity mismatch in terminology￾intensive retrieval and formulate retrieval-unit design as the joint learning of context-dependent units and budgeted unit selection under… view at source ↗
Figure 3
Figure 3. Figure 3: Architecture of H+ Embedding. Token importance scores are max-pooled by vocabulary into lexical weights and aggregated within CRF-defined phrases into phrase weights for top-B selection and query-side weighted MaxSim. 3.2 Learning Context-Dependent Retrieval Units A linear-chain CRF (Lafferty, McCallum, and Pereira 2001) predicts a BIO label yi ∈ {O, B, I} for each contextual token state. Each B tag starts… view at source ↗
Figure 4
Figure 4. Figure 4: Task-level deployed Phrase-Global nDCG@10 dif [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] 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

15 extracted references · 5 linked inside Pith

  1. [3]

    Formal, T.; Piwowarski, B.; and Clinchant, S

    Reduc- ing the Footprint of Multi-Vector Retrieval with Minimal Performance Impact via Token Pooling.arXiv preprint arXiv:2409.14683. Formal, T.; Piwowarski, B.; and Clinchant, S

    Pith/arXiv arXiv

  2. [4]

    The two random controls are averaged per System Params Dim

    All methods use the same H+ Embed- ding MRL checkpoint, raw 128-dimensional states, CRF- only decoding for Learned Phrase, frozen Global top-1,000 candidates, and exactly matched realized query and docu- ment unit counts. The two random controls are averaged per System Params Dim. Protocol NFC. SciFact MedQA ChatDr. Feedback Bioinfo. Biology IIYi H+ Embed...

    2027

  3. [7]

    InProceedings of the 2021 Conference on Empirical Methods in Natural Language Pro- cessing, 3661–3672

    Phrase Retrieval Learns Passage Retrieval, Too. InProceedings of the 2021 Conference on Empirical Methods in Natural Language Pro- cessing, 3661–3672. Li, Y.; Li, Z.; Zhang, K.; Dan, R.; Jiang, S.; and Zhang, Y

    2021

  4. [8]

    InFindings of the Association for Com- putational Linguistics: ACL 2022, 926–937

    Using Interactive Feedback to Improve the Ac- curacy and Explainability of Question Answering Systems Post-Deployment. InFindings of the Association for Com- putational Linguistics: ACL 2022, 926–937. MacAvaney,S.;Mallia,A.;andTonellotto,N.2025.Efficient Constant-Space Multi-Vector Retrieval. InEuropean Con- ference on Information Retrieval, 237–245. Cham...

    Pith/arXiv arXiv 2022

  5. [10]

    InProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies, 3715–3734

    ColBERTv2: Effective and Efficient Re- trieval via Lightweight Late Interaction. InProceedings of the 2022 Conference of the North American Chapter of the Association for Computational Linguistics: Human Lan- guage Technologies, 3715–3734. Wadden, D.; Lin, S.; Lo, K.; Wang, L. L.; van Zuylen, M.; Cohan, A.; and Hajishirzi, H

    2022

  6. [11]

    InProceedings of the 2020 Confer- ence on Empirical Methods in Natural Language Processing (EMNLP), 7534–7550

    Fact or Fiction: Verify- ing Scientific Claims. InProceedings of the 2020 Confer- ence on Empirical Methods in Natural Language Processing (EMNLP), 7534–7550. Wang, L.; Yang, N.; Huang, X.; Jiao, B.; Yang, L.; Jiang, D.; Majumder, R.; and Wei, F

    2020

  7. [12]

    Weller, O.; Boratko, M.; Naim, I.; and Lee, J

    Text Embeddings by Weakly-Supervised Contrastive Pre-training.arXiv preprint arXiv:2212.03533. Weller, O.; Boratko, M.; Naim, I.; and Lee, J

    Pith/arXiv arXiv

  8. [14]

    InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 641–649

    C-Pack: Packed Resources for General Chi- nese Embeddings. InProceedings of the 47th International ACM SIGIR Conference on Research and Development in Information Retrieval, 641–649. arXiv:2309.07597. Yang,A.;Li,A.;Yang,B.;Zhang,B.;Hui,B.;Zheng,B.;Yu, B.;Gao,C.;Huang,C.;Lv,C.;etal.2025. Qwen3Technical Report.arXiv preprint arXiv:2505.09388. Zhang, X.; Li,...

    Pith/arXiv arXiv 2025

  9. [511]

    A Full-Text Learning to Rank Dataset for Medical Infor- mation Retrieval

    Boteva,V.;Gholipour,D.;Sokolov,A.;andRiezler,S.2016. A Full-Text Learning to Rank Dataset for Medical Infor- mation Retrieval. InEuropean Conference on Information Retrieval, 716–722. Cham: Springer International Publish- ing. Chaffin, A

    2016

  10. [2019]

    Sentence-BERT: Sen- tenceEmbeddingsUsing SiameseBERT-Networks. InPro- ceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP- IJCNLP), 3982–3992. Santhanam, K.; Khattab, O.; Saad-Falcon, J.; Potts, C.; and Zaharia, M

    2019

  11. [2020]

    InProceedings of the 2020 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP), 6769–6781

    Dense Passage Retrieval for Open-Domain Question Answering. InProceedings of the 2020 Conference on Empirical Methods in Natural Lan- guage Processing (EMNLP), 6769–6781. Khattab, O.; and Zaharia, M

    2020

  12. [2021]

    InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2288–2292

    SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking. InProceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, 2288–2292. Hofstätter, S.; Khattab, O.; Althammer, S.; Sertkan, M.; and Hanbury,A.2022. IntroducingNeuralBagofWhole-Words withColBERTer:ContextualizedLateInteractionsusingEn-...

    2022

  13. [2022]

    arXiv:2112.09118

    Unsupervised Dense In- formationRetrievalwithContrastiveLearning.Transactions on Machine Learning Research. arXiv:2112.09118. Jha, R.; Wang, B.; Günther, M.; Mastrapas, G.; Sturua, S.; Mohr,I.;Koukounas,A.;Akram,M.K.;Wang,N.;andXiao, H.2024.Jina-ColBERT-v2:AGeneral-PurposeMultilingual Late Interaction Retriever. InProceedings of the Fourth Workshop on Mul...

    Pith/arXiv arXiv 2024

  14. [2024]

    InFindings of the Association for Computational Linguistics: ACL 2024, 2318–2335

    M3-Embedding: Multi-Linguality, Multi- Functionality, Multi-Granularity Text Embeddings through Self-Knowledge Distillation. InFindings of the Association for Computational Linguistics: ACL 2024, 2318–2335. Clavié, B.; Chaffin, A.; and Adams, G

    2024

  15. [2026]

    InProceedings of the International Conference on Learning Representations

    On the Theoretical Limitations of Embedding-Based Retrieval. InProceedings of the International Conference on Learning Representations. arXiv:2508.21038. Xiao, S.; Liu, Z.; Zhang, P.; Muennighoff, N.; Lian, D.; and Nie, J.-Y

    arXiv