Pith. sign in

REVIEW 3 major objections 5 minor 27 references

A single Matryoshka Hypencoder generates one set of Q-Net parameters that can be truncated to form valid smaller Q-Nets, letting one deployed model score documents at several speed–quality points without retraining.

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

2026-08-01 17:52 UTC pith:FLUGFIWB

load-bearing objection Genuinely new application of Matryoshka to Q-Net parameters, with decent empirical support, but the central truncation construction is undefined and one latency claim overshoots. the 3 major comments →

arxiv 2607.17457 v1 pith:FLUGFIWB submitted 2026-07-20 cs.IR

The Matryoshka Hypencoder

Majd Alkawaas , Sean MacAvaney This is my paper
classification cs.IR
keywords HypencoderMatryoshka Representation LearningQ-NetHyperNetworkretrieval efficiencyzero-shot retrievalparameter truncationBEIR
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.

Hypencoders encode each query as a small neural network—a Q-Net—that scores precomputed document embeddings, giving more expressive relevance estimation than a vector dot product but fixing the model at one size. This paper asks whether the Matryoshka idea—train representations so that prefixes of the vector are themselves useful, smaller representations—can be applied to the Q-Net's parameter vector instead of the embedding vector. The authors claim it can: by fine-tuning only the hyperhead with a Matryoshka-style margin loss, a single generated parameter set can be truncated to widths 128, 256, 512, or 768, each forming a valid Q-Net. Their experiments show the width-256 prefix (roughly 7x fewer parameters) is statistically indistinguishable in-domain from the full width-768 model on TREC DL'19 and DL'20, and the width-512 prefix (half the parameters) keeps zero-shot out-of-domain effectiveness intact on five BEIR tasks. The payoff they argue for is practical: one deployed model can move between quality and cost, and smaller Q-Nets score documents 1.6–3.4x faster in the tested settings.

Core claim

The paper's central claim is that the Matryoshka principle transfers from representation vectors to the parameter space of a hypernetwork-generated scoring function. A Matryoshka Hypencoder is a Hypencoder whose hyperhead is fine-tuned (with document and query transformers frozen) under a multi-objective MarginMSE loss computed at four widths: 128, 256, 512, and 768. At each width, the loss truncates the generated full parameter vector to a prefix, assembles a temporary Q-Net of that hidden width, and compares its query-document margin to a cross-encoder teacher's margin. The authors report that, in-domain, the width-256 Q-Net (591,616 parameters, about 7x fewer than the full 4,134,912) achi

What carries the argument

The central object is the truncatable Q-Net parameter vector. The hyperhead outputs a single large parameter set Θ_full for a Q-Net with hidden width 768; the Matryoshka loss (Eq. 4) evaluates a margin-MSE term at each width m ∈ M by taking the prefix Θ^{1:m} and assembling a temporary Q-Net of that width. The mechanism that carries the argument is the unstated construction that a prefix of the parameter vector maps to a valid narrower Q-Net—i.e., that the hyperhead's output is ordered so that each prefix corresponds to a complete set of layer weights and biases for a smaller MLP. This is what makes one trained model serve multiple deployment points: at scoring time you simply truncate to th

Load-bearing premise

The paper assumes that truncating the flat parameter vector to a prefix yields a valid, narrower Q-Net (correct layer shapes for the stated width), but the construction is never specified—no per-layer ordering or hidden-unit ordering is given—so a naive prefix could cut through a weight matrix and break the network.

What would settle it

Load the released checkpoint, generate Θ_full for a query, truncate to the first 197,632 parameters (the width-128 size), and run the Q-Net forward pass on a document embedding. If the truncated parameters do not assemble into a valid 6-layer MLP with hidden width 128—or if the reported MS MARCO dev MRR of 0.380 at width 768 and 0.375 at width 128 do not reproduce—the central claim that prefixes are valid Q-Nets fails.

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

If this is right

  • The frozen-encoder fine-tuning recipe is viable: a Hypencoder fine-tuned only in the hyperhead matches the original end-to-end model on 7 of 8 in-domain and zero-shot datasets, with no statistically significant difference.
  • In-domain, the width-256 prefix of the Matryoshka Q-Net uses about 7x fewer active parameters than the full width-768 Q-Net while remaining statistically indistinguishable on TREC DL'19 and DL'20.
  • Out-of-domain, the width-512 prefix (half the active parameters) maintains comparable zero-shot effectiveness across TREC-COVID, FiQA, DBPedia, NFCorpus, and Webis Touché; significant degradation appears at widths 256 and 128 on FiQA.
  • Smaller Q-Nets are faster: measured document-scoring throughput rises 1.6x at width 512, 3.4x at width 256, and 6.3x at width 128 on large collections, with the 128-size speedup falling short of the 20.9x theoretical parameter reduction.
  • Training smaller widths 64 and 32 did not converge, so the practical nesting range demonstrated is 128–768.

Where Pith is reading between the lines

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

  • If the prefix-truncation construction is made explicit (e.g., ordering parameters layer-by-layer and hidden-unit-by-hidden-unit), the same Matryoshka loss should transfer to other hypernetwork-generated outputs—such as LoRA adapters, classifier heads, or per-query attention parameters—giving any such module a built-in speed–quality dial.
  • The reported throughput numbers omit query encoding and approximate search; a full-system measurement might show the end-to-end speedup is smaller in practice, but on CPU, where scoring is compute-bound rather than memory-bandwidth-bound, the 128-width Q-Net could approach closer to the theoretical 20.9x parameter reduction.
  • A testable extension is to measure how effectiveness degrades with truncation at even finer granularities (e.g., width 192) to map the full rate of the trade-off curve; the paper only reports four points.
  • The failure of widths 32 and 64 to converge suggests a lower bound on the Matryoshka nesting range; whether that bound is a property of the Q-Net architecture or the margin-MSE teaching signal is an open question worth probing.

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

3 major / 5 minor

Summary. The paper proposes the Matryoshka Hypencoder, which extends the recently introduced Hypencoder architecture to generate query-specific scoring networks (Q-Nets) of multiple widths from a single model. The idea is to train the hyperhead so that prefixes of the full generated Q-Net parameter vector form valid, narrower Q-Nets, enabling efficiency-effectiveness trade-offs at deployment time. The authors fine-tune an existing Hypencoder with frozen transformer encoders and a multi-objective MarginMSE loss over several Q-Net widths (M = [128, 256, 512, 768]). They report in-domain and zero-shot BEIR results showing that smaller Q-Nets largely preserve effectiveness while increasing document scoring throughput, and that a lightweight frozen-encoder fine-tuning procedure is competitive with full end-to-end training.

Significance. If the central truncation construction is made precise and reproducible, the contribution is valuable: it extends Matryoshka-style nested representation learning to the parameter space of a query-specific neural scorer, potentially allowing a single retrieval model to be deployed at multiple cost points without retraining. The paper also ships code and evaluates on standard external benchmarks (MS MARCO, DL'19, DL'20, five BEIR sets), and the frozen-encoder fine-tuning procedure is a practical strength. However, the paper's main empirical claim of 'comparable in-domain effectiveness' is too strong as stated, and the core mechanism—the definition of the truncated Q-Net—is absent from the text, which currently makes the results irreproducible in principle.

major comments (3)
  1. [§3.1, Eq. (4)] The truncation operation Θ^{1:m}_q is never defined. The Q-Net is a 6-layer MLP with LayerNorm and residual connections (Eq. 2), so a flat prefix of the flattened parameter vector would cut mid-matrix and break layer shapes. A per-layer, per-hidden-unit ordering and a rule for truncating LayerNorm parameters are needed. Without this, s_m(p; Θ^{1:m}_q) in Eq. (4) is undefined, the training objective is ill-posed, and the central nesting claim is unverifiable. Please specify the exact parameter layout, the ordering used for prefixes, and how each prefix yields a valid narrower Q-Net. Also clarify how residual connections work when the hidden width m differs from the document embedding dimension: Eq. (2) adds x_i to a transformed vector, which requires square weight matrices unless the residual is omitted or the input is also truncated.
  2. [Abstract and Table 2] The abstract claims 'comparable in-domain effectiveness' for the 7× smaller Q-Net, but the paper's own Table 2 shows that every Matryoshka size is statistically significantly worse than the Original Hypencoder on MS MARCO Dev MRR@10 (.379–.381 vs .386, all marked †). The later qualification in §5.2 (that DL'19/DL'20 have near-complete assessments whereas Dev does not) is reasonable, but the abstract and conclusion do not reflect this caveat. The headline claim should be qualified to avoid overstating the result.
  3. [§5.4, Table 4] The reported throughput gains (1.6–3.4×) are only for Q-Net document scoring, explicitly excluding query encoding time. Since the paper never states that the hyperhead can generate only a prefix of the Q-Net parameters, it remains unclear whether smaller Q-Nets also reduce query-encoder cost. If the full parameter vector is always generated and then truncated, the end-to-end speedup for a single query is smaller than the scoring-only numbers suggest, particularly for short queries. Please clarify the deployment mechanism and report end-to-end latency (or justify why query encoding is negligible for the intended use case).
minor comments (5)
  1. [§3.1, footnote 1] The pilot study that selected the loss formulation is mentioned but not described. For reproducibility, list the variants considered or state the selection protocol.
  2. [§5.2] The sentence 'no significant degradations compared to the full-size Matryoshka Hypencoder (d=768) for d=512 and d=256' is easy to misread as a comparison with the Original Hypencoder. Rephrase to avoid ambiguity.
  3. [Abstract] Typo: 'approximately7×' should be 'approximately 7×'.
  4. [References] Reference [16] has 'doi:0.1145/...'—the leading '1' appears to be missing (should be '10.1145/...').
  5. [Table 2] The parameter counts for each Q-Net width should be derived from the architectural definition clarified in response to major comment 1. Currently the numbers cannot be verified without the missing truncation specification.

Circularity Check

0 steps flagged

No significant circularity: the central claims are empirically benchmarked; the missing truncation specification is a reproducibility gap, not a circular derivation.

full rationale

The paper's central claims are empirical, not derived by construction. The Matryoshka loss (Eq. 4) is a training objective that exposes truncated parameter prefixes; it assumes, rather than proves, that a prefix of Θ_q^full can be assembled into a smaller Q-Net. That assumption is an architectural design choice, and the paper's contributions—comparable effectiveness at reduced width, OOD generalization, and throughput gains—are measured against external benchmarks (DL'19, DL'20, MS MARCO Dev, BEIR), not read off from the loss. There is no load-bearing self-citation: the base Hypencoder [11] and MRL [13] are independent prior works with no author overlap; the only self-citation [15] is for the ir_datasets library and is not used to justify any claim. The pilot-study footnote ('this version was the most reliable') indicates model selection over loss variants but does not state that selection was made on the reported test splits, so it cannot be shown to make the in-domain numbers circular. The omitted specification of how Θ_q^{1:m} maps onto the 6-layer MLP of Eq. (2) is a genuine reproducibility/correctness gap, but an underspecified construction is not an equivalence between input and output; it does not by itself make the empirical evaluation circular.

Axiom & Free-Parameter Ledger

3 free parameters · 6 axioms · 0 invented entities

No invented physical or conceptual entities are introduced. The paper's contribution rests on training-objective assumptions (nesting works in function-parameter space), an unspecified truncation-order assumption, and standard distillation/evaluation assumptions. The main free parameters are the Matryoshka dimension set and the unstated loss-weighting choice.

free parameters (3)
  • Matryoshka dimension set M = {128, 256, 512, 768}
    Chosen via pilot; smaller sizes 64 and 32 failed to converge (Section 4.1 footnote). Reported effectiveness and speedup numbers depend on this specific set.
  • Matryoshka loss weighting coefficients c_m = Not reported; Eq. 4 appears to use uniform weighting 1/|M|
    Footnote 1 states several loss variations were explored and 'this version was the most reliable'; the exact weighting and any schedule are not given, yet they shape the training objective.
  • Fine-tuning hyperparameters = Not reported
    Learning rate, number of epochs, batch size, and distillation margin details are absent; they are required to reproduce the frozen-encoder fine-tuning results.
axioms (6)
  • domain assumption Truncating the full generated parameter set to a prefix yields a valid Q-Net of smaller width.
    Eq. (4) computes losses over truncated prefixes Θ^{1:m}_q, but the Q-Net is a 6-layer residual MLP (Eq. 2); a flat prefix is only well-defined with a specific per-layer ordering that the paper never specifies.
  • domain assumption The Matryoshka nested-prefix principle transfers from representation vectors to the parameter space of a generated scoring function.
    This is the core hypothesis motivating Sections 2.2 and 3.1. It is not proven; the paper offers only empirical support.
  • domain assumption Frozen BERT encoders retain enough capacity that fine-tuning only the hyperhead gives performance comparable to end-to-end training.
    Section 3.2 introduces the lightweight procedure; RQ1 provides partial support, but Table 2 shows significant degradation on MS MARCO Dev and Table 3 on NFCorpus.
  • domain assumption Cross-encoder teacher margins Δ_s^T(q) are reliable relevance targets at every truncated width.
    The MarginMSE objective (Eq. 4) is trained against precomputed teacher scores from the original Hypencoder's MS MARCO preprocessing; teacher quality is assumed.
  • standard math Paired t-tests at p<0.05 are meaningful on the small query sets used.
    DL'19 has 43 queries and DL'20 has 54; low sample sizes reduce statistical power, yet the paper relies on these tests for non-significance claims.
  • domain assumption Zero-shot BEIR nDCG@10 measures out-of-domain generalization.
    The paper uses five BEIR datasets as its OOD benchmark (Section 4); transferability of these results to other domains is assumed.

pith-pipeline@v1.3.0-alltime-deepseek · 10256 in / 13734 out tokens · 127261 ms · 2026-08-01T17:52:19.310751+00:00 · methodology

0 comments
read the original abstract

The Hypencoder is a recently-proposed retrieval approach that encodes queries as shallow neural networks ("Q-Nets") that estimate relevance over pre-computed document embeddings. Inspired by Matryoshka Representation Learning, we show that the Hypencoder can be extended to support multiple sizes of Q-Nets, allowing trade-offs between effectiveness and efficiency when deployed. We find that this "Matryoshka Hypencoder" achieves comparable in-domain effectiveness with approximately 7x fewer active parameters in-domain and half as many active parameters out-of-domain, which corresponds to a 1.6-3.4x increase in scoring throughput. This work paves the way for practical deployment of Hypencoders.

Figures

Figures reproduced from arXiv: 2607.17457 by Majd Alkawaas, Sean MacAvaney.

Figure 1
Figure 1. Figure 1: Overview of the Matryoshka Hypencoder architec [PITH_FULL_IMAGE:figures/full_fig_p001_1.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

27 extracted references · 4 linked inside Pith

  1. [1]

    Alexander Bondarenko, Maik Fröbe, Meriem Beloucif, Lukas Gienapp, Yamen Ajjour, Alexander Panchenko, Chris Biemann, Benno Stein, Henning Wachsmuth, Martin Potthast, and Matthias Hagen. 2020. Overview of Touché 2020: Argument Retrieval. InWorking Notes of CLEF 2020 - Conference and Labs of the Evaluation Forum, Thessaloniki, Greece, September 22-25, 2020 (...

    2020

  2. [2]

    Vera Boteva, Demian Gholipour Ghalandari, Artem Sokolov, and Stefan Riezler

  3. [3]

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, and Daniel Campos. 2021. Overview of the TREC 2020 deep learning track.CoRRabs/2102.07662 (2021). arXiv:2102.07662 https://arxiv.org/abs/2102.07662

    Pith/arXiv arXiv 2021

  4. [4]

    Voorhees

    Nick Craswell, Bhaskar Mitra, Emine Yilmaz, Daniel Campos, and Ellen M. Voorhees. 2020. Overview of the TREC 2019 deep learning track.CoRR abs/2003.07820 (2020). arXiv:2003.07820 https://arxiv.org/abs/2003.07820

    Pith/arXiv arXiv 2020

  5. [5]

    Ruiqi Guo, Philip Sun, Erik Lindgren, Quan Geng, David Simcha, Felix Chern, and Sanjiv Kumar. 2020. Accelerating Large-Scale Inference with Anisotropic Vector Quantization. InProceedings of the 37th International Conference on Machine Learning, ICML 2020, 13-18 July 2020, Virtual Event (Proceedings of Machine Learning Research, Vol. 119). PMLR, 3887–3896....

    2020

  6. [6]

    Dai, and Quoc V

    David Ha, Andrew M. Dai, and Quoc V. Le. 2017. HyperNetworks. In5th Interna- tional Conference on Learning Representations, ICLR 2017, Toulon, France, April 24-26, 2017, Conference Track Proceedings. OpenReview.net. https://openreview. net/forum?id=rkpACe1lx

    2017

  7. [7]

    Faegheh Hasibi, Fedor Nikolaev, Chenyan Xiong, Krisztian Balog, Svein Erik Bratsberg, Alexander Kotov, and Jamie Callan. 2017. DBpedia-Entity v2: A Test Collection for Entity Search. InProceedings of the 40th International ACM SIGIR Conference on Research and Development in Information Retrieval, Shinjuku, Tokyo, Japan, August 7-11, 2017, Noriko Kando, Te...

    arXiv 2017

  8. [8]

    Kalervo Järvelin and Jaana Kekäläinen. 2002. Cumulated gain-based evaluation of IR techniques.ACM Trans. Inf. Syst.20, 4 (2002), 422–446. doi:10.1145/582415. 582418

    doi:10.1145/582415 2002

  9. [9]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen-tau Yih. 2020. Dense Passage Retrieval for Open- Domain Question Answering. InProceedings of the 2020 Conference on Empirical Methods in Natural Language Processing, EMNLP 2020, Online, November 16-20, 2020, Bonnie Webber, Trevor Cohn, Yulan He, and Ya...

    doi:10.18653/v1/2020.emnlp-main.550 2020

  10. [10]

    Omar Khattab and Matei Zaharia. 2020. ColBERT: Efficient 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, SIGIR 2020, Virtual Event, China, July 25-30, 2020, Jimmy X. Huang, Yi Chang, Xueqi Cheng, Jaap Kamps, Vaness...

    arXiv 2020

  11. [11]

    Julian Killingback, Hansi Zeng, and Hamed Zamani. 2025. Hypencoder: Hyper- networks for Information Retrieval. InProceedings of the 48th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2025, Padua, Italy, July 13-18, 2025, Nicola Ferro, Maria Maistro, Gabriella Pasi, Omar Alonso, Andrew Trotman, and Suzan Ver...

    arXiv 2025

  12. [12]

    Hrishikesh Kulkarni, Sean MacAvaney, Nazli Goharian, and Ophir Frieder. 2023. Lexically-Accelerated Dense Retrieval. InProceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2023, Taipei, Taiwan, July 23-27, 2023, Hsin-Hsi Chen, Wei-Jou (Edward) Duh, Hen-Hsen Huang, Makoto P. Kato, Josiane Mo...

    arXiv 2023

  13. [13]

    Kakade, Prateek Jain, and Ali Farhadi

    Aditya Kusupati, Gantavya Bhatt, Aniket Rege, Matthew Wallingford, Aditya Sinha, Vivek Ramanujan, William Howard-Snyder, Kaifeng Chen, Sham M. Kakade, Prateek Jain, and Ali Farhadi. 2022. Matryoshka Representation Learning. InAdvances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing Systems 2022, NeurIPS 2022...

    2022

  14. [14]

    Xianming Li, Zongxi Li, Jing Li, Haoran Xie, and Qing Li. 2025. ESE: Espresso Sentence Embeddings. InThe Thirteenth International Conference on Learning Representations, ICLR 2025, Singapore, April 24-28, 2025. OpenReview.net. https: //openreview.net/forum?id=plgLA2YBLH

    2025

  15. [15]

    Sean MacAvaney, Andrew Yates, Sergey Feldman, Doug Downey, Arman Cohan, and Nazli Goharian. 2021. Simplified Data Wrangling with ir_datasets. InSIGIR ’21: The 44th International ACM SIGIR Conference on Research and Development in Information Retrieval, Virtual Event, Canada, July 11-15, 2021, Fernando Diaz, Chirag Shah, Torsten Suel, Pablo Castells, Rosie...

    arXiv 2021

  16. [16]

    Craig Macdonald, Nicola Tonellotto, and Zhili Shen. 2026. A Replicability Study of Joint Product Quantisation for Effective Space-Efficient Dense Retrieval. InSIGIR ’26: The 49th International ACM SIGIR Conference on Research and Development in SIGIR ’26, July 20–24, 2026, Melbourne, VIC, Australia Majd Alkawaas and Sean MacAvaney Information Retrieval, M...

    arXiv 2026

  17. [17]

    Macedo Maia, Siegfried Handschuh, André Freitas, Brian Davis, Ross McDer- mott, Manel Zarrouk, and Alexandra Balahur. 2018. WWW’18 Open Challenge: Financial Opinion Mining and Question Answering. InCompanion of the The Web Conference 2018 on The Web Conference 2018, WWW 2018, Lyon , France, April 23-27, 2018, Pierre-Antoine Champin, Fabien Gandon, Mounia ...

    arXiv 2018

  18. [18]

    Malkov and Dmitry A

    Yury A. Malkov and Dmitry A. Yashunin. 2020. Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs.IEEE Trans. Pattern Anal. Mach. Intell.42, 4 (2020), 824–836. doi:10.1109/TPAMI.2018. 2889473

    doi:10.1109/tpami.2018 2020

  19. [19]

    Mackenzie, and Torsten Suel

    Antonio Mallia, Michal Siedlaczek, Joel M. Mackenzie, and Torsten Suel. 2019. PISA: Performant Indexes and Search for Academia. InProceedings of the Open- Source IR Replicability Challenge co-located with 42nd International ACM SIGIR Conference on Research and Development in Information Retrieval, OSIRRC@SIGIR 2019, Paris, France, July 25, 2019 (CEUR Work...

    2019

  20. [20]

    Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh Tiwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. InProceedings of the Workshop on Cognitive Computation: Integrating neural and symbolic approaches 2016 co-located with the 30th Annual Conference on Neural Information Processing Systems...

    2016

  21. [21]

    Rodrigo Nogueira and Kyunghyun Cho. 2019. Passage Re-ranking with BERT. CoRRabs/1901.04085 (2019). arXiv:1901.04085 http://arxiv.org/abs/1901.04085

    Pith/arXiv arXiv 2019

  22. [22]

    Robertson, Steve Walker, Susan Jones, Micheline Hancock-Beaulieu, and Mike Gatford

    Stephen E. Robertson, Steve Walker, Susan Jones, Micheline Hancock-Beaulieu, and Mike Gatford. 1994. Okapi at TREC-3. InProceedings of The Third Text REtrieval Conference, TREC 1994, Gaithersburg, Maryland, USA, November 2-4, 1994 (NIST Special Publication, Vol. 500-225), Donna K. Harman (Ed.). National Institute of Standards and Technology (NIST), 109–12...

    1994

  23. [23]

    Nandan Thakur, Nils Reimers, Andreas Rücklé, Abhishek Srivastava, and Iryna Gurevych. 2021. BEIR: A Heterogenous Benchmark for Zero-shot Evaluation of Information Retrieval Models.CoRRabs/2104.08663 (2021). arXiv:2104.08663 https://arxiv.org/abs/2104.08663

    Pith/arXiv arXiv 2021

  24. [24]

    Voorhees, Tasmeer Alam, Steven Bedrick, Dina Demner-Fushman, William R

    Ellen M. Voorhees, Tasmeer Alam, Steven Bedrick, Dina Demner-Fushman, William R. Hersh, Kyle Lo, Kirk Roberts, Ian Soboroff, and Lucy Lu Wang. 2020. TREC-COVID: constructing a pandemic information retrieval test collection. SIGIR Forum54, 1 (2020), 1:1–1:12. doi:10.1145/3451964.3451965

    arXiv 2020

  25. [25]

    Shuai Wang, Shengyao Zhuang, Bevan Koopman, and Guido Zuccon. 2025. 2D Matryoshka Training for Information Retrieval. InProceedings of the 48th In- ternational ACM SIGIR Conference on Research and Development in Information Retrieval, SIGIR 2025, Padua, Italy, July 13-18, 2025, Nicola Ferro, Maria Maistro, Gabriella Pasi, Omar Alonso, Andrew Trotman, and ...

    arXiv 2025

  26. [26]

    Orion Weller, Michael Boratko, Iftekhar Naim, and Jinhyuk Lee. 2025. On the Theoretical Limitations of Embedding-Based Retrieval.CoRRabs/2508.21038 (2025). arXiv:2508.21038 doi:10.48550/ARXIV.2508.21038

    doi:10.48550/arxiv.2508.21038 2025

  27. [2016]

    InAdvances in Information Retrieval - 38th European Conference on IR Research, ECIR 2016, Padua, Italy, March 20-23, 2016

    A Full-Text Learning to Rank Dataset for Medical Information Retrieval. InAdvances in Information Retrieval - 38th European Conference on IR Research, ECIR 2016, Padua, Italy, March 20-23, 2016. Proceedings (Lecture Notes in Computer Science, Vol. 9626), Nicola Ferro, Fabio Crestani, Marie-Francine Moens, Josiane Mothe, Fabrizio Silvestri, Giorgio Maria D...

    doi:10.1007/978-3-319-30671-1_58 2016