Pith. sign in

REVIEW 5 major objections 5 minor 33 references

Semantic Certainty Assessment in Vector Retrieval Systems: A Novel Framework for Embedding Quality Evaluation

T0 review · 5 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper proposes a per-query reliability score that predicts retrieval success and lifts Recall@10 by 9.4% over strong baselines.

desk verdict The 9.4% claim is contradicted by the paper's own tables, and the method never says how the score changes retrieval; the framework is a plausible incremental QPP idea that isn't currently evaluable. read the letter →

arxiv 2507.05933 v1 pith:3N25SPIA submitted 2025-07-08 cs.IR cs.CL

Y. Du This is my paper · ORCID
classification cs.IRcs.CL
keywords semanticreliabilityembeddingqualityqueryperformancepredictionquantizationrobustnessneighborhooddensitydenseretrievalvectorRecall@10
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper tries to establish that a retrieval system can predict how well it will answer a query using only the query embedding, without knowing the true relevant documents. The proposed score combines two cheap signals: how stable the embedding is under quantization and how densely its nearest neighbors cluster around it. If the claim is right, production retrieval systems gain a per-query confidence signal at under 5% extra retrieval time, enabling adaptive strategies such as reranking only uncertain queries. The paper reports consistent Recall@10 improvements across four datasets, averaging 9.4±1.2% over the strongest baseline.

What carries the argument

The load-bearing object is the semantic reliability score $R_q = 2 G_q I_q/(G_q + I_q)$, the harmonic mean of a quantization-stability proxy $G_q = \exp(-\|e_q - Q(e_q)\|^2 / 2\hat{\sigma}_q^2)$ and a neighborhood-density proxy $I_q = K / (\sum_{i=1}^K \|e_q - r_i\|^2 + \epsilon)$. The paper frames these proxies through the Semantic Gravity Well Model, in which each concept is a Gaussian well and deeper wells concentrate probability mass, making their embeddings more quantization-robust and more densely surrounded by semantically related vectors. The score is what carries the argument: it turns the abstract claim that reliable embeddings are geometrically stable and informationally dense into a per-query number used for prediction and adaptive retrieval.

What would settle it

Build or find a corpus where an irrelevant query's embedding is surrounded by a dense cluster of near-duplicate generic passages, compute the paper's neighborhood density and combined score, and check whether that query is scored high while its labeled Recall@10 is low; a robust pattern of such inversions would falsify the claim that density measures semantic certainty.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that semantic reliability of a query decomposes into an information-density factor and a geometric-stability factor, and that their harmonic mean predicts query-level retrieval performance. High-quality embeddings are said to sit in stable, dense regions of the embedding space; the score estimates this by measuring reconstruction error after product quantization (the stability proxy) and the inverse average distance to the $K$ nearest neighbors (the density proxy). The combined reliability score is claimed to correlate with Recall@10 (0.79 in the MS MARCO ablation) and to improve Recall@10 consistently, from +2.1% on MS MARCO to +3.2% on TREC-DL 2019, with an average gain of 9.4±1.2% over the strongest baseline.

Load-bearing premise

The load-bearing assumption is that the $K$ nearest neighbors of a query embedding are semantically meaningful, so that many close neighbors indicate high semantic reliability rather than corpus structure such as near-duplicate generic passages.

Editorial extensions

If this is right

  • A retrieval system can compute a certainty score for every incoming query and route only low-certainty queries to expensive reranking, expansion, or human review.
  • Because the score adds less than 5% to retrieval time, real-time quality monitoring of a query stream becomes practical without separate labeled judgments.
  • Certainty scores vary systematically by query type, with factual queries highest and ambiguous queries lowest, pointing toward augmenting training data with more ambiguous queries.
  • Embedding models can be compared or selected using their certainty statistics, since models differ in how reliably they represent factual, conceptual, and ambiguous queries.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • An untested extension: the same two signals could be computed on document-side embeddings to flag low-quality index regions, not only hard queries.
  • If the gravity-well interpretation is correct, the density proxy should respond monotonically when random neighbors are removed; a cheap validation is to sample subsets of neighbors and check that score ordering is stable.
  • In corpora with many near-duplicate generic passages, the density proxy may measure corpus artifacts rather than semantic certainty, so renormalizing density by collection-wide statistics is a natural correction not explored in the paper.
  • The stated bound $E[\text{Recall@K}] \geq 1 - \exp(-R_q K / 2D)$ suggests the score could be used to choose retrieval depth $K$ per query, an adaptive strategy the paper does not test.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 5 minor

Summary. The paper proposes a lightweight query-level embedding-quality score that combines quantization robustness and neighborhood density, and claims that this score predicts retrieval performance and can be used to improve Recall@10. The framework is motivated by a 'semantic gravity well' model, formalized through Definitions 1-5 and Theorems 1-3, and evaluated on MS MARCO, Natural Questions, TREC-DL 2019, and BEIR. The paper also reports a runtime overhead analysis, an ablation study, and an analysis of certainty scores across query types.

Significance. If the claimed effect were real, a cheap per-query confidence signal would be practically valuable for adaptive dense retrieval. The paper contains a useful survey of related work on query performance prediction and uncertainty quantification, and the runtime measurement in Table 5 is concrete. However, the central empirical claim is not supported by the paper's own tables, the retrieval mechanism is never specified, and the theoretical guarantees are asserted rather than derived. The paper therefore does not currently establish its main contribution.

major comments (5)
  1. [Abstract; Tables 1 and 3] The headline claim of a 9.4±1.2% average Recall@10 improvement is contradicted by the paper's own results. Table 3 reports per-dataset improvements of +2.1%, +2.2%, +3.2%, and +2.5%, whose average is about 2.5%; Table 1's absolute differences imply the same order of magnitude. No relative or absolute transformation of the reported numbers yields 9.4%, so the central empirical claim is internally inconsistent. Table 4 additionally reports a +7.6% gain for the combined score without stating the baseline used for that percentage, which is another unexplained discrepancy.
  2. [Algorithm 1; Algorithm 2] The paper never specifies how the computed reliability score changes the retrieval procedure. Algorithm 1 returns Rq, and Algorithm 2 only triggers a quality alert and states 'Apply adaptive retrieval strategy' in line 5, with the strategy left undefined. Consequently the reported Recall@10 improvements cannot be reproduced or even interpreted, because the evaluated system is not described.
  3. [Eq. (20); Algorithm 2; Table 4] The combination rule is inconsistent across the paper. Equation (20) defines Cq as an unweighted harmonic mean, Algorithm 2 uses Cq = α·Sq + β·Nq, and Table 4 reports 'Combined (α = 0.6)' without defining β or explaining how α was chosen. Since Table 4 is the only component-ablation evidence, this makes the ablation uninterpretable.
  4. [Theorem 1; Theorem 3] The theoretical support is not derived. Theorem 1 restates Definition 1 by plugging in specific formulas for I and G, so it is tautological rather than an independent decomposition result. Theorem 3 asserts P(miss) ≤ exp(−Rq K/2D) in Eq. (15) with no derivation; the phrase 'concentration of measure' does not justify an exponential bound in the reliability score, and the bound treats the estimated Rq from Eq. (13) as a known quantity without any error propagation.
  5. [Definition 4, Eq. (12)] The information density proxy assumes that the K nearest neighbors of the query embedding in the raw embedding space are semantically meaningful. In anisotropic or artifact-dominated embedding spaces, Iq will measure corpus geometry rather than semantic certainty, and the paper provides no experiment separating these effects, such as perturbing embeddings or removing near-duplicate passages. This concern directly affects the validity of the proposed metric.
minor comments (5)
  1. [Eq. (10)] The step from Eq. (9) to Eq. (10) appears algebraically incorrect: for a Gaussian well with ∇²φ = σ_c⁻²I, the expression Dσ_c²/λ_min·ε_q² evaluates to Dσ_c⁴ε_q², not Dσ_c²ε_q², so the displayed bound is not justified as written.
  2. [Main Results; Table 2] The text states that improvements are 'statistically significant across all test sets,' but Table 2 reports significance testing only for MS MARCO; p-values for the other datasets are not provided.
  3. [Related Work] Reference [18] is cited for 'cross-lingual dense retrieval,' but it is a paper on entity retrieval; this citation does not match the claim.
  4. [Experimental Setup; Algorithm 2] The method depends on several free parameters (K, σ², ε, α, τ), and the paper does not report how they were set or how sensitive the results are to their values. The limitation section acknowledges domain dependence, but no sensitivity analysis is given.
  5. [Table 6] Table 6 reports 'semantic certainty scores' for TCT-ColBERT, but this model is not otherwise described or included in the retrieval experiments, so the table's connection to the main evaluation is unclear.

Circularity Check

2 steps flagged · score 6.0 of 10

Theorem 1 restates Definition 1, and Theorem 3 assumes the bound it claims to prove; the empirical 9.4% headline is also not supported by the paper's own tables.

  1. self definitional [Definition 1 (Eq. 1) and Theorem 1 (Eq. 2) with proof Eqs. (3)-(6), Section 'Theoretical Framework']
    "Definition 1 (Semantic Reliability): ... R(q) = I(E(q)) · G(E(q)) (1) ... Theorem 1 (Semantic Reliability Decomposition): For any embedding eq = E(q), the semantic reliability decomposes into: R(q) = E[log p(eq|N(eq))] · exp(−E[∥eq − T(eq)∥2]) (2) ... I(eq) = E[log p(eq|N(eq))] (3) ... G(eq) = exp(−E[∥eq − T(eq)∥2]) (5)."

    Definition 1 defines semantic reliability as the product of information density I(E(q)) and geometric stability G(E(q)). Theorem 1 then states that reliability decomposes into an information-density factor and a geometric-stability factor. In the proof, the information-density term is set to E[log p(eq|N(eq))] and the geometric-stability term is set to exp(−E[∥eq − T(eq)∥2]), which are exactly the two factors in the defining equation. No independent derivation is supplied; Eq. (2) is Eq. (1) with concrete formulas substituted for I and G. The 'decomposition' is therefore true by construction and adds no new content beyond the definition.

  2. other [Theorem 3 (Reliability-Performance Bound), Eqs. (14)-(17), Section 'Performance Guarantee']
    "Proof of Theorem 3 : The bound follows from the concentration of measure in high-dimensional spaces. For embeddings with high semantic reliability, the probability of finding relevant documents within the top-K decreases exponentially with the reliability score: P (miss relevant doc) ≤ exp(−Rq·K/2D) (15)."

    The theorem asserts E[Recall@K] ≥ 1 − exp(−Rq·K/2D), which is algebraically equivalent to P(miss relevant doc) ≤ exp(−Rq·K/2D). The proof simply restates that inequality as its key step and invokes 'concentration of measure' without deriving the exponential dependence from any distributional or geometric condition. The performance bound is thus assumed rather than proven: the theorem's conclusion is fed back into the proof as the premise. This is circular as a derivation and means the claimed theoretical guarantee does not independently establish a relationship between Rq and recall.

full rationale

The framework's derivation chain contains two places where a stated result is equivalent to its own premises. Theorem 1 ('Semantic Reliability Decomposition') is Definition 1 restated: after writing R(q) = I(E(q)) · G(E(q)) in Eq. (1), the proof defines I(eq) and G(eq) to be exactly the two factors appearing in Eq. (2), so nothing is decomposed or derived. Theorem 3 is an assumed performance guarantee: the proof asserts the exponential miss-probability bound that the theorem claims to establish, with no intermediate concentration-of-measure argument, so the bound is an input rather than a consequence. These two items matter because they are the paper's theoretical justification for equating the hand-built score Rq with retrieval recall. The empirical evaluation is an independent check of the proxy, and that part is not circular in itself; however, the abstract's 9.4±1.2% improvement is not consistent with Table 3's reported +2.1%, +2.2%, +3.2%, +2.5% (average about 2.5%), and Table 4's 'Combined (α = 0.6)' value is presented without a tuning or validation protocol. Those are consistency and validity problems rather than circularity, but they reinforce that the central claims are not as robust as stated. Because the core theoretical statements reduce to definitions or assumptions while the constituent metrics are still evaluated against held-out test data, a partial-circularity score of 6 is appropriate.

Assumptions & free parameters 5 free parameters · 4 assumptions · 2 invented entities

The framework rests on a Gaussian 'gravity well' model of semantics, an unproved exponential bound relating the reliability score to recall, and the assumption that embedding neighborhoods are semantically meaningful. All of these are assumed rather than derived. The empirical evaluation additionally depends on hyperparameters K, σ², ε, α, and τ, none of which are specified with a validation protocol.

free parameters (5)
  • K (neighbor count)
    Number of nearest neighbors in Eq. 12 and Algorithm 1; no value or selection procedure is reported.
  • σ² (variance for quantization stability)
    Used in Eq. 18 and Definition 3; said to be estimated from the embedding distribution, but the estimation method is unspecified.
  • ε (regularization term)
    Appears in Eq. 12 and Definition 4; no value is given, and it directly affects the density score.
  • α (combination weight) = 0.6
    Table 4 uses a combined score with α=0.6; the value appears to be selected in the ablation to maximize correlation on the same data, with no validation protocol.
  • τ (alert threshold)
    Threshold in Algorithm 2 for triggering quality alerts; no value or tuning method is provided.
assumptions (4)
  • ad hoc to paper Semantic concepts induce Gaussian gravity wells with centroid μc and variance σ_c², Eq. 7.
    Introduced to justify the relationship between curvature and quantization stability; no independent evidence that real embedding neighborhoods are Gaussian.
  • ad hoc to paper The probability of missing a relevant document satisfies P(miss) ≤ exp(-Rq K / 2D), Eq. 15.
    This exponential concentration bound is asserted without derivation and is the entire content of Theorem 3.
  • domain assumption The K nearest vectors in the embedding index are semantically coherent neighbors for the query.
    The neighborhood density score in Definition 4 assumes nearest neighbors carry semantic information rather than corpus artifacts.
  • ad hoc to paper Quantization noise is small enough that quantization error is bounded by the inverse Hessian of the well, Eqs. 9-10.
    No derivation links PQ truncation error to the local curvature of a Gaussian density.
invented entities (2)
  • Semantic Reliability Score R(q)
    purpose: Combines quantization stability and neighborhood density into a query-level quality estimate.
    The metric is validated only on the same retrieval datasets used for tuning α; it has no external falsifiable prediction.
  • Semantic Gravity Well model
    purpose: Conceptual justification for why stability and density should correlate with retrieval performance.
    The model is introduced post hoc, and no separate test distinguishes it from alternative explanations.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Semantic Certainty Assessment in Vector Retrieval Systems: A Novel Framework for Embedding Quality Evaluation." pith.science (2026). https://pith.science/paper/3N25SPIA

@misc{pith2026250705933,
  author       = {Pith},
  title        = {Pith review of: Semantic Certainty Assessment in Vector Retrieval Systems: A Novel Framework for Embedding Quality Evaluation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3N25SPIA}},
  note         = {Machine review of arXiv:2507.05933}
}
abstract

Vector retrieval systems exhibit significant performance variance across queries due to heterogeneous embedding quality. We propose a lightweight framework for predicting retrieval performance at the query level by combining quantization robustness and neighborhood density metrics. Our approach is motivated by the observation that high-quality embeddings occupy geometrically stable regions in the embedding space and exhibit consistent neighborhood structures. We evaluate our method on 4 standard retrieval datasets, showing consistent improvements of 9.4$\pm$1.2\% in Recall@10 over competitive baselines. The framework requires minimal computational overhead (less than 5\% of retrieval time) and enables adaptive retrieval strategies. Our analysis reveals systematic patterns in embedding quality across different query types, providing insights for targeted training data augmentation.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

33 extracted references · 30 canonical work pages

  1. [1]

    Karpukhin, V ., Oguz, B., Min, S., Lewis, P., Wu, L., Edunov, S., Chen, D., and Yih, W. T. (2020). Dense passage retrieval for open-domain question answering. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP) , pages 6769–6781

    work page 2020

  2. [2]

    F., Liu, J., Ben- nett, P., Ahmed, J., and Overwijk, A

    Xiong, L., Xiong, C., Li, Y ., Tang, K. F., Liu, J., Ben- nett, P., Ahmed, J., and Overwijk, A. (2021). Approxi- mate nearest neighbor negative contrastive learning for dense text retrieval. InProceedings of the International Conference on Learning Representations (ICLR)

    work page 2021

  3. [3]

    and Zaharia, M

    Khattab, O. and Zaharia, M. (2020). ColBERT: Effi- cient and effective passage search via contextualized late interaction over BERT. InProceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 39–48

    work page 2020

  4. [4]

    and Gurevych, I

    Reimers, N. and Gurevych, I. (2019). Sentence-BERT: Sentence embeddings using Siamese BERT-networks. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), pages 3982–3992

    work page 2019

  5. [5]

    Thakur, N., Reimers, N., R ¨uckl´e, A., Srivastava, A., and Gurevych, I. (2021). BEIR: A heterogeneous benchmark for zero-shot evaluation of information re- trieval models. In Proceedings of the Neural Informa- tion Processing Systems Track on Datasets and Bench- marks

    work page 2021

  6. [6]

    Hofst ¨atter, S., Althammer, S., Schr ¨oder, M., Sertkan, M., and Hanbury, A. (2021). Improving efficient neu- ral ranking models with cross-architecture knowledge distillation. arXiv preprint arXiv:2010.02666

    arXiv 2021

  7. [7]

    Luan, Y ., Eisenstein, J., Toutanova, K., and Collins, M. (2021). Sparse, dense, and attentional representations for text retrieval. Transactions of the Association for Computational Linguistics, 9:329–345

    work page 2021

  8. [8]

    V oorhees, E. M. (2005). The TREC robust retrieval track. In ACM SIGIR Forum, volume 39, pages 11–20

    work page 2005

Show all 33 references
  1. [9]

    and Yom-Tov, E

    Carmel, D. and Yom-Tov, E. (2010). Estimating the Query Difficulty for Information Retrieval . Morgan & Claypool Publishers

    2010

  2. [10]

    Ethayarajh, K. (2019). How contextual are contextu- alized word representations? Comparing the geometry of BERT, ELMo, and GPT-2 embeddings. InProceed- ings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th Interna- tional Joint Conference ...

    2019

  3. [11]

    Qiu, X., Sun, T., Xu, Y ., Shao, Y ., Dai, N., and Huang, X. (2020). Pre-trained models for natural language pro- cessing: A survey. Science China Technological Sci- ences, 63(10):1872–1897

    2020

  4. [12]

    Rogers, A., Kovaleva, O., and Rumshisky, A. (2020). A primer on neural network models for natural language processing. Journal of Artificial Intelligence Research, 57:615–731

    2020

  5. [13]

    Zhao, Y ., Scholer, F., and Tsegay, Y . (2008). Effective pre-retrieval query performance prediction using sim- ilarity and variability evidence. In Proceedings of the 30th European Conference on IR Research, pages 52– 64

    2008

  6. [14]

    Shtok, A., Kurland, O., and Carmel, D. (2012). Us- ing statistical decision theory and relevance models for query-performance prediction. In Proceedings of the 35th international ACM SIGIR conference on Research and development in information retrieval , pages 259– 268

    2012

  7. [15]

    and Tanguy, L

    Mothe, J. and Tanguy, L. (2015). Linguistic features to predict query difficulty. In Proceedings of the 38th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 7–10

    2015

  8. [16]

    Zhao, J., Huang, J., and Ye, N. (2021). Learning to pre- dict query difficulty for neural information retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Informa- tion Retrieval, pages 1845–1849

    2021

  9. [17]

    Arabzadeh, N., Zarrinkalam, F., Jovanovic, J., and Bagheri, E. (2021). Neural embedding-based metrics for pre-retrieval query performance prediction. In Pro- ceedings of the 44th International ACM SIGIR Con- ference on Research and Development in Information Retrieval, pages 1492–1496

    2021

  10. [18]

    Asai, A., Hashimoto, T., Hajishirzi, H., Socher, R., and Xiong, C. (2021). Learning dense representations for entity retrieval. In Proceedings of the 23rd Con- ference on Computational Natural Language Learning (CoNLL), pages 528–537

    2021

  11. [19]

    Santhanam, K., Khattab, O., Saad-Falcon, J., Clark, C., and Zaharia, M. (2021). ColBERTv2: Effective and ef- ficient retrieval via lightweight late interaction. arXiv preprint arXiv:2112.01488

    2021 arXiv

  12. [20]

    Faruqui, M., Tsvetkov, Y ., Rastogi, P., and Dyer, C. (2016). Problems with evaluation of word embeddings using word similarity tasks. In Proceedings of the 1st Workshop on Evaluating Vector-Space Representa- tions for NLP, pages 30–35

    2016

  13. [21]

    Neal, R. M. (2012). Bayesian learning for neural net- works. Springer Science & Business Media

    2012

  14. [22]

    and Ghahramani, Z

    Gal, Y . and Ghahramani, Z. (2016). Dropout as a Bayesian approximation: Representing model uncer- tainty in deep learning. In International Conference on Machine Learning, pages 1050–1059

    2016

  15. [23]

    Lakshminarayanan, B., Pritzel, A., and Blundell, C. (2017). Simple and scalable predictive uncertainty es- timation using deep ensembles. In Advances in Neural Information Processing Systems, pages 6402–6413

    2017

  16. [24]

    Guo, C., Pleiss, G., Sun, Y ., and Weinberger, K. Q. (2017). On calibration of modern neural networks. In International Conference on Machine Learning, pages 1321–1330

    2017

  17. [25]

    Ovadia, Y ., Fertig, E., Ren, J., Nado, Z., Sculley, D., Nowozin, S., Dillon, J., Lakshminarayanan, B., and Snoek, J. (2019). Can you trust your model’s uncer- tainty? Evaluating predictive uncertainty under dataset shift. In Advances in Neural Information Processing Systems, ...

    2019

  18. [26]

    Zamani, H., Mitra, B., Song, X., Craswell, N., and Tiwary, S. (2020). Neural ranking models with weak supervision. In Proceedings of the 43rd International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 2005–2008

    2020

  19. [27]

    and Ounis, I

    He, B. and Ounis, I. (2004). Inferring query perfor- mance using pre-retrieval predictors. In International Symposium on String Processing and Information Re- trieval, pages 43–54

    2004

  20. [28]

    Cronen-Townsend, S., Zhou, Y ., and Croft, W. B. (2002). Predicting query performance. In Proceedings of the 25th Annual International ACM SIGIR Confer- ence on Research and Development in Information Re- trieval, pages 299–306

    2002

  21. [29]

    J ´egou, H., Douze, M., and Schmid, C. (2011). Product quantization for nearest neighbor search. IEEE Trans- actions on Pattern Analysis and Machine Intelligence, 33(1):117–128

    2011

  22. [30]

    Martinez, J., Clement, J., Zouaq, A., and Suominen, H. (2021). Learned quantization for neural informa- tion retrieval. In Proceedings of the 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, pages 1508–1512

    2021

  23. [31]

    Ge, T., He, K., Ke, Q., and Sun, J. (2013). Optimized product quantization for approximate nearest neigh- bor search. In Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition , pages 2946–2953

    2013

  24. [32]

    Johnson, J., Douze, M., and J ´egou, H. (2017). Billion- scale similarity search with GPUs. arXiv preprint arXiv:1702.08734

    2017 arXiv

  25. [33]

    Guo, R., Sun, P., Lindgren, E., Geng, Q., Simcha, D., Chern, F., and Kumar, S. (2020). Accelerating large- scale inference with anisotropic vector quantization. In International Conference on Machine Learning, pages 3887–3896

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.