Pith. sign in

REVIEW 4 major objections 4 minor 45 references

DOGR: Leveraging Document-Oriented Contrastive Learning in Generative Retrieval

T0 review · 4 major / 4 minor · reviewed 2026-08-08 · deepseek-v4-flash

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

Pith's one-line read Contrastive learning lets generative retrieval rank documents by meaning, not just identifiers.

desk verdict Solid, well-ablated method paper; the SOTA claim is weakened by test-set selection of identifier length, but the core approach deserves review. read the letter →

arxiv 2502.07219 v2 pith:NNRNWVOW submitted 2025-02-11 cs.IR

Penghao Lu , Xin Dong , Yuansheng Zhou , Lei Cheng , Chuan Yuan , Linjian Mo This is my paper · ORCID
classification cs.IR
keywords generativeretrievalcontrastivelearningdocumentrankingnegativesamplingquery-documentrelevanceidentifierssequence-to-sequenceencoder-decoder
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

Generative retrieval replaces the external document index with a language model that writes document identifiers for a query, but earlier systems mainly learn query-to-identifier mappings and never compare the query with the full document. DOGR claims that retrieval quality improves when the model also learns a document-level semantic score through contrastive learning, and that the two signals should be multiplied at inference time. The paper builds this in two stages: first a generation stage teaches the model to produce an identifier from the query or from a document segment, then a ranking stage adds two contrastive losses over deliberately chosen negative documents while keeping the generation loss alive. On NQ320k and MS MARCO, DOGR reports the best generative-retrieval results among the compared methods, and its ablations show that both contrastive losses, the semantic score, and the fusion step each contribute. The significance is practical: if the claim holds, generative retrieval can disambiguate shared identifiers and operate on large corpora rather than memorising small identifier sets.

What carries the argument

The load-bearing mechanism is the two-stage learning loop with a fused relevance score. Stage one trains a T5-base encoder-decoder to generate a keyword-based identifier $z_d$ from either the query or a randomly sampled document segment, using teacher-forced cross-entropy. Stage two fine-tunes that model with two contrastive objectives: the prefix-oriented loss $L_{\mathrm{ci}}$ pulls a query near its positive document and pushes it away from documents that share the same identifier prefix, while the retrieval-augmented loss $L_{\mathrm{cq}}$ pushes it away from documents whose identifiers the stage-one model generated with high probability for that query, excluding the positive. An auxiliary generation loss $L_g$ is added so the contrastive terms do not destroy the model's ability to generate identifiers. At inference, constrained beam search produces candidate identifiers and documents are re-ranked by $\mathrm{rel}(q,d) = P(z_d \mid q) \cdot s(q,d)$, so the semantic score both breaks identifier collisions and corrects the order of the generated list.

What would settle it

Take a test collection with full relevance judgments, run the stage-one model on each query, and inspect the documents behind its top-k generated identifiers after removing the labelled positive. If a substantial share of those documents are judged relevant, DOGR's retrieval-augmented negatives include false negatives, and a version of the method that filters them out should outperform the published model; if filtering changes nothing, the retrieval-augmented loss is not doing the work the paper claims.

Watch

Extended reading notes

Core claim

DOGR's central claim is that query-document relevance in generative retrieval should be modelled as the product of the identifier generation probability and a semantic similarity between the query text and the full document, rather than by identifier probability alone. The paper defines the semantic score as $s(q,d) = h_q \cdot h_d^{\mathsf{T}}$, where $h_x$ is the mean-pooled encoder output for input $x$, and ranks candidate documents by $\mathrm{rel}(q,d) = P(z_d \mid q) \cdot s(q,d)$. To make that score meaningful, it introduces a two-stage training strategy: the identifier generation stage learns both $q \to z_d$ and document-segment inputs $d \to z_d$ with cross-entropy, and the document ranking stage adds a prefix-oriented contrastive loss and a retrieval-augmented contrastive loss, with an auxiliary generation loss to preserve identifier generation. The reported results are Recall@1 of 70.2 on NQ320k and MRR@10 of 22.5 on MS MARCO, above the generative retrieval baselines compared, and the framework also improves keyword, n-gram, first-token, and numerical identifier types. On the paper's own reading, this shows that direct query-document interaction is what previous generative retrievers were missing.

Load-bearing premise

The load-bearing premise is that the documents reached through the stage-one model's top-k generated identifiers, excluding the positive document, are actually non-relevant to the query; if the first-stage model is imperfect, some of these assumed negatives are relevant in truth, and that would contaminate the contrastive signal.

Editorial extensions

If this is right

  • If the central claim is correct, identifier collisions cease to be a ranking dead end: documents sharing one identifier are ordered by their semantic match to the query instead of arbitrarily.
  • The two-stage recipe transfers across identifier schemes, so future identifier designs can inherit the contrastive ranking stage rather than redesigning the training objective.
  • The large relative gain on MS MARCO suggests semantic scoring is what lets generative retrieval leave small-corpus memorisation behind and work at passage-corpus scale.
  • The ablation pattern implies contrastive learning should be an addition to, not a replacement for, identifier generation; dropping the auxiliary generation loss costs 6.3 points of Recall@1.
  • A fused score that uses both signals outperforms either one alone, giving later systems a simple template: let the generative model propose, let the semantic model dispose.

Reading between the lines

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

  • A testable extension the authors do not run: replace retrieval-augmented negatives with random same-prefix negatives of equal count; if the gain vanishes, the benefit comes from hard-negative mining rather than from document-oriented contrastive learning per se.
  • The multiplication in $\mathrm{rel}(q,d)$ couples a probability with an unbounded dot product, so a learned temperature or additive weight between the two terms is a natural follow-up that the paper's own ablations do not explore.
  • Because prefix-oriented negatives are sampled inside the batch, some can be false negatives (relevant to the query without being the labelled positive); measuring that contamination on a fully judged subset would show how robust the contrastive signal is.
  • The same two-stage pattern, generate a compact code then contrastively rank full inputs, could transfer to other sequence-to-sequence settings, such as structured prediction or entity linking, whenever the output code is shorter than the input.
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

4 major / 4 minor

Summary. The paper proposes DOGR, a two-stage generative retrieval framework built on T5-base. In the first stage, an encoder-decoder model is trained to generate keyword-based document identifiers, using both query-to-identifier and document-to-identifier generation losses. In the second stage, the model is fine-tuned with two contrastive objectives: a prefix-oriented loss that uses in-batch documents sharing the identifier prefix as negatives, and a retrieval-augmented loss that uses documents under the model's top-k generated identifiers as negatives, together with an auxiliary generation loss. At inference, documents are ranked by the product of the identifier generation probability and a semantic dot-product score between the query and document encoder representations. Experiments on NQ320k and MS MARCO report state-of-the-art results among generative retrieval methods, with ablations attributing gains to each component and additional experiments showing the framework generalizes across lexical and numerical identifier types.

Significance. If the reported results are reliable, DOGR is a useful contribution to generative retrieval: it is one of the few methods that explicitly combines identifier generation with a learned query-document semantic similarity, and the ablations suggest the two contrastive objectives and the fusion ranking each contribute. The generalization experiments across identifier types are a strength, as they show the framework is not tied to the specific keyword identifier used in the main experiments. The paper also clearly describes its training objectives and inference procedure. However, the empirical SOTA claim is currently undermined by evaluation-protocol issues: a hyperparameter (identifier length) is selected directly on the test set, no variance or significance information is given, and the provenance of baseline numbers is inconsistent. The central idea is defensible, but the evidence as presented does not yet establish the claimed advantage over prior work.

major comments (4)
  1. [Implementation Details and Table 4] The identifier token length is selected on the same NQ320k test set used for the headline results. The text states, 'After experimenting with different identifier token length, we set the token length of document identifiers to 8 and 12 for NQ320k and MS MARCO respectively,' and Table 4 reports Recall@1 on NQ320K(7830) for l = 3, 4, 6, 8, 10, 12. The chosen l = 8 corresponds to the maximum value in that table (70.2), while neighboring lengths give 68.9 and 69.8. Because the margin over the best generative baseline (GLEN, 69.1) is only 1.1 points, this test-set selection can explain the reported SOTA advantage. Please re-select the identifier length on a held-out validation split (or report that the choice is insensitive across a validation set) and then report the test-set numbers for the chosen configuration. The same concern applies to the other hyperparameters (tau, lambda_g, the number of retrieval-augmented negatives, and beam size), which are not described as being chosen on a separate validation split.
  2. [Experimental Results, Tables 1 and 2] No error bars, multiple-seed variance, or significance tests are reported. Several reported differences are very small (e.g., +0.3% on R@10 in Table 1, and +0.4% for 'w/o fusion rank' in Table 5), and for a method with multiple stochastic training stages these differences could easily be within run-to-run noise. Please provide at least three independent runs (or comparable variance estimates) for the main results and ablations, and indicate whether the observed differences are statistically significant. Without this, the claimed improvements, especially the small margins in Table 1, are not established.
  3. [Baselines and Table 1 / Table 3] The provenance of the baseline numbers is inconsistent and needs clarification. The note under Table 1 says 'We refer to the results of baselines reported by (Lee et al. 2023b),' but the Baselines section states that LTRGR was re-trained with a T5-base backbone for fairness, implying that at least some baselines were re-run. Moreover, the LTRGR and GLEN numbers in Table 3 (keyword row: 69.2 and 68.7) differ from the corresponding Table 1 values (67.5 and 69.1), even though both tables report results on NQ320K under the same keyword identifier. Please state explicitly which baselines were re-run, which were quoted, and why the same method has different scores in Tables 1 and 3. If the Table 3 baselines were produced under a different setup, describe that setup. This information is essential for assessing the claimed SOTA margin.
  4. [Document Ranking Stage, Retrieval-augmented Negative Sampling] The retrieval-augmented negative sampling assumes that all documents whose identifiers appear in the first-stage model's top-k output, excluding the positive document's identifier, are non-relevant to the query. Because the first-stage model is imperfect, some of these documents may actually be relevant, and treating them as negatives would corrupt the contrastive signal. The paper does not analyze the precision of these negatives or discuss this risk. Since the ablation shows that removing this component costs 1.3% R@1 (Table 5), the validity of this assumption is load-bearing for the method's design. Please measure the proportion of sampled negatives that are actually relevant (e.g., using the training relevance labels) or provide an ablation with a cleaner negative source to show that the observed gain is not an artifact of noisy labels.
minor comments (4)
  1. [Table 1] The DSI-QG row appears to contain copied values: the 'Unseen test' columns show exactly the BM25 unseen-test values (32.3, 61.9, 42.7), and the 'Seen test' MRR@100 is 39.5, which is far below that method's full-test MRR@100 of 69.5. Please verify and correct these entries.
  2. [Implementation Details, Training and Inference] The sentence 'batch size is set to 256 and 32' is ambiguous: it is unclear whether these values correspond to the two training stages or to the two datasets. Please specify the batch size for each stage and each dataset.
  3. [References] The reference list contains many entries that are never cited in the text and appear unrelated to the paper's topic, including Clancey (1979, 1983, 1984, 2021), Engelmore and Morgan (1986), Hasling et al. (1983, 1984), NASA (2015), Rice (1986), and Robinson (1980a, 1980b). There are also duplicate entries for Wang et al. (2022a and 2022b), which appear to be the same paper. Please clean the reference list to include only cited works and resolve the duplicates.
  4. [Throughout] There are several typos and unclear notations: 'augumented' should be 'augmented' in the Query Generation paragraph; 'MMR@100' in the Table 3 caption should be 'MRR@100'; and the description of h_x in Eq. (2) should specify how long documents are truncated or chunked before average pooling, since document length can affect the semantic score.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DOGR's contrastive losses and fusion scoring are defined independently of the reported results.

full rationale

The paper is an empirical method paper rather than a formal derivation, and I found no step in which a claimed prediction or first-principles result is equivalent by construction to its inputs. The central inference equation, rel(q,d) = P(z_d|q) * s(q,d), is an explicitly proposed fusion rule combining the identifier generation probability with a separately defined semantic score; it is not derived from the contrastive losses nor does it presuppose the reported Recall@1 or MRR values. The contrastive objectives L_ci and L_cq define query-document semantic similarity through averaged encoder hidden states and standard temperature-scaled softmax losses, with negative documents obtained either from same-prefix in-batch sampling or from the first-stage model's top-k generated identifiers. The retrieval-augmented negative sampling does use the model's own output to construct hard negatives, but this is ordinary hard-negative mining: the negatives are training signals, not a restatement of the final ranking metric, and their quality is empirically validated by the ablation study rather than assumed. No self-citations are load-bearing, and no uniqueness theorem or prior author result is invoked to force the method's choices. The only notable protocol concern is that the identifier length is selected after evaluating on the NQ320k test set in Table 4, which may inflate the reported SOTA margin; however, this is an evaluation-validity issue, not circularity in the derivation-chain sense, and it does not make the method's equations reduce to their own outputs. The empirical claims are benchmarked against external datasets and baselines, so the central content is self-contained.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claim depends on standard learning objectives, domain assumptions about identifier semantics, and two hand-chosen scoring components. No new physical or mathematical entities are introduced. The largest burden is the assumption that automatically mined negatives are true negatives.

free parameters (5)
  • identifier token length = 8 (NQ320k), 12 (MS MARCO)
    Selected after experiments (Table 4); the choice affects conflict rate and retrieval quality.
  • retrieval-augmented negative count k = 4
    Chosen by hand; no sensitivity analysis shown.
  • temperature tau = 0.5
    Chosen for the contrastive losses; no sensitivity analysis shown.
  • generation loss weight lambda_g = 0.1
    Chosen to balance generation and contrastive objectives; no sensitivity analysis shown.
  • beam size = 100
    Chosen for inference-time constrained beam search.
assumptions (5)
  • standard math Standard cross-entropy and InfoNCE losses provide valid training signals.
    Invoked throughout Equations 1 to 4.
  • domain assumption Keyword-based identifiers capture enough document semantics to serve as generation targets and contrastive anchors.
    Relies on KeyBERT keywords in the Identifier Generation Stage.
  • domain assumption Documents sharing identifier prefixes are semantically similar enough to act as hard negatives.
    Assumed in the Prefix-oriented Negative Sampling paragraph.
  • domain assumption First-stage top-k identifiers retrieve documents that are valid negatives for a query.
    Assumed in Retrieval-augmented Negative Sampling; depends on the imperfect first-stage model.
  • ad hoc to paper The multiplicative fusion of generation probability and semantic score ranks documents better than either alone.
    Inference section defines rel(q,d) = P(z_d|q) * s(q,d); supported only by ablation comparisons, not by a first-principles argument.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DOGR: Leveraging Document-Oriented Contrastive Learning in Generative Retrieval." pith.science (2026). https://pith.science/paper/NNRNWVOW

@misc{pith2026250207219,
  author       = {Pith},
  title        = {Pith review of: DOGR: Leveraging Document-Oriented Contrastive Learning in Generative Retrieval},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/NNRNWVOW}},
  note         = {Machine review of arXiv:2502.07219}
}
read the original abstract

Generative retrieval constitutes an innovative approach in information retrieval, leveraging generative language models (LM) to generate a ranked list of document identifiers (docid) for a given query. It simplifies the retrieval pipeline by replacing the large external index with model parameters. However, existing works merely learned the relationship between queries and document identifiers, which is unable to directly represent the relevance between queries and documents. To address the above problem, we propose a novel and general generative retrieval framework, namely Leveraging Document-Oriented Contrastive Learning in Generative Retrieval (DOGR), which leverages contrastive learning to improve generative retrieval tasks. It adopts a two-stage learning strategy that captures the relationship between queries and documents comprehensively through direct interactions. Furthermore, negative sampling methods and corresponding contrastive learning objectives are implemented to enhance the learning of semantic representations, thereby promoting a thorough comprehension of the relationship between queries and documents. Experimental results demonstrate that DOGR achieves state-of-the-art performance compared to existing generative retrieval methods on two public benchmark datasets. Further experiments have shown that our framework is generally effective for common identifier construction techniques.

Figures

Figures reproduced from arXiv: 2502.07219 by the authors.

Figure 1
Figure 1. An overview of the framework of DOGR. The training phase can be divided into two stages. The first is the identifier [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 26 canonical work pages

  1. [1]

    Bevilacqua, M.; Ottaviano, G.; Lewis, P. S. H.; Yih, S.; Riedel, S.; and Petroni, F. 2022. Autoregressive Search Engines: Generating Substrings as Document Identifiers. In Koyejo, S.; Mohamed, S.; Agarwal, A.; Belgrave, D.; Cho, K.; and Oh, A., eds., Advances in Neural Information Processing Systems 35: Annual Conference on Neural Information Processing S...

    work page 2022

  2. [2]

    D.; Izacard, G.; Riedel, S.; and Petroni, F

    Cao, N. D.; Izacard, G.; Riedel, S.; and Petroni, F. 2021. Autoregressive Entity Retrieval. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7, 2021 . OpenReview.net

    work page 2021

  3. [3]

    Chen, J.; Zhang, R.; Guo, J.; de Rijke, M.; Chen, W.; Fan, Y.; and Cheng, X. 2023. Continual Learning for Generative Retrieval over Dynamic Corpora. In Frommholz, I.; Hopfgartner, F.; Lee, M.; Oakes, M.; Lalmas, M.; Zhang, M.; and Santos, R. L. T., eds., Proceedings of the 32nd ACM International Conference on Information and Knowledge Management, CIKM 202...

    work page 2023

  4. [4]

    Clancey, W. J. 1979. Transfer of Rule-Based Expertise through a Tutorial Dialogue . Ph.D. diss., Dept.\ of Computer Science, Stanford Univ., Stanford, Calif

    1979

  5. [5]

    Clancey, W. J. 1983. Communication, Simulation, and Intelligent Agents: Implications of Personal Intelligent Machines for Medical Education . In Proceedings of the Eighth International Joint Conference on Artificial Intelligence (IJCAI-83) , 556--560. Menlo Park, Calif: IJCAI Organization

    1983

  6. [6]

    Clancey, W. J. 1984. Classification Problem Solving . In Proceedings of the Fourth National Conference on Artificial Intelligence, 45--54. Menlo Park, Calif.: AAAI Press

    1984

  7. [7]

    Clancey, W. J. 2021. The Engineering of Qualitative Models . Forthcoming

    2021

  8. [8]

    Engelmore, R.; and Morgan, A., eds. 1986. Blackboard Systems. Reading, Mass.: Addison-Wesley

    1986

Show all 45 references
  1. [9]

    Gao, T.; Yao, X.; and Chen, D. 2021. SimCSE: Simple Contrastive Learning of Sentence Embeddings. In Moens, M.; Huang, X.; Specia, L.; and Yih, S. W., eds., Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, EMNLP 2021, Virtual Event / Punta...

    2021

  2. [10]

    Grootendorst, M. 2020. KeyBERT: Minimal keyword extraction with BERT

    2020

  3. [11]

    W.; Clancey, W

    Hasling, D. W.; Clancey, W. J.; and Rennels, G. 1984. Strategic explanations for a diagnostic consultation system. International Journal of Man-Machine Studies, 20(1): 3--19

    1984

  4. [12]

    W.; Clancey, W

    Hasling, D. W.; Clancey, W. J.; Rennels, G. R.; and Test, T. 1983. Strategic Explanations in Consultation---Duplicate . The International Journal of Man-Machine Studies, 20(1): 3--19

    1983

  5. [13]

    Karpukhin, V.; Oguz, B.; Min, S.; Lewis, P. S. H.; Wu, L.; Edunov, S.; Chen, D.; and Yih, W. 2020. Dense Passage Retrieval for Open-Domain Question Answering. In Webber, B.; Cohn, T.; He, Y.; and Liu, Y., eds., Proceedings of the 2020 Conference on Empirical Methods in Natural...

    2020

  6. [14]

    Khosla, P.; Teterwak, P.; Wang, C.; Sarna, A.; Tian, Y.; Isola, P.; Maschinot, A.; Liu, C.; and Krishnan, D. 2020. Supervised Contrastive Learning. In Larochelle, H.; Ranzato, M.; Hadsell, R.; Balcan, M.; and Lin, H., eds., Advances in Neural Information Processing Systems 33:...

    2020

  7. [15]

    P.; Alberti, C.; Epstein, D.; Polosukhin, I.; Devlin, J.; Lee, K.; Toutanova, K.; Jones, L.; Kelcey, M.; Chang, M.; Dai, A

    Kwiatkowski, T.; Palomaki, J.; Redfield, O.; Collins, M.; Parikh, A. P.; Alberti, C.; Epstein, D.; Polosukhin, I.; Devlin, J.; Lee, K.; Toutanova, K.; Jones, L.; Kelcey, M.; Chang, M.; Dai, A. M.; Uszkoreit, J.; Le, Q.; and Petrov, S. 2019. Natural Questions: a Benchmark for Q...

    2019

  8. [16]

    Lee, H.; Kim, J.; Chang, H.; Oh, H.; Yang, S.; Karpukhin, V.; Lu, Y.; and Seo, M. 2023 a . Nonparametric Decoding for Generative Retrieval. In Rogers, A.; Boyd - Graber, J. L.; and Okazaki, N., eds., Findings of the Association for Computational Linguistics: ACL 2023, Toronto,...

    2023

  9. [17]

    Lee, K.; Chang, M.; and Toutanova, K. 2019. Latent Retrieval for Weakly Supervised Open Domain Question Answering. In Korhonen, A.; Traum, D. R.; and M \` a rquez, L., eds., Proceedings of the 57th Conference of the Association for Computational Linguistics, ACL 2019, Florence...

    2019

  10. [18]

    Lee, S.; Choi, M.; Lee, J.; and X. 2023 b . GLEN: Generative Retrieval via Lexical Index Learning. In Bouamor, H.; Pino, J.; and Bali, K., eds., Proceedings of the 2023 Conference on Empirical Methods in Natural Language Processing, EMNLP 2023, Singapore, December 6-10, 2023 ,...

    2023

  11. [19]

    Lewis, M.; Liu, Y.; Goyal, N.; Ghazvininejad, M.; Mohamed, A.; Levy, O.; Stoyanov, V.; and Zettlemoyer, L. 2020. BART: Denoising Sequence-to-Sequence Pre-training for Natural Language Generation, Translation, and Comprehension. In Jurafsky, D.; Chai, J.; Schluter, N.; and Tetr...

    2020

  12. [20]

    Li, Y.; Yang, N.; Wang, L.; Wei, F.; and Li, W. 2023. Multiview Identifiers Enhanced Generative Retrieval. In Rogers, A.; Boyd - Graber, J. L.; and Okazaki, N., eds., Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers...

    2023

  13. [21]

    Li, Y.; Yang, N.; Wang, L.; Wei, F.; and Li, W. 2024. Learning to Rank in Generative Retrieval. In Wooldridge, M. J.; Dy, J. G.; and Natarajan, S., eds., Thirty-Eighth AAAI Conference on Artificial Intelligence, AAAI 2024, Thirty-Sixth Conference on Innovative Applications of ...

    2024

  14. [22]

    V.; Gupta, J.; Tay, Y.; Dehghani, M.; Tran, V

    Mehta, S. V.; Gupta, J.; Tay, Y.; Dehghani, M.; Tran, V. Q.; Rao, J.; Najork, M.; Strubell, E.; and Metzler, D. 2023. DSI++: Updating Transformer Memory with New Documents. In Bouamor, H.; Pino, J.; and Bali, K., eds., Proceedings of the 2023 Conference on Empirical Methods in...

    2023

  15. [23]

    NASA . 2015. Pluto: The 'Other' Red Planet. https://www.nasa.gov/nh/pluto-the-other-red-planet. Accessed: 2018-12-06

    2015

  16. [24]

    Nguyen, T.; Rosenberg, M.; Song, X.; Gao, J.; Tiwary, S.; Majumder, R.; and Deng, L. 2016. MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. CoRR, abs/1611.09268

    2016 arXiv

  17. [25]

    H.; Constant, N.; Ma, J.; Hall, K

    Ni, J.; \' A brego, G. H.; Constant, N.; Ma, J.; Hall, K. B.; Cer, D.; and Yang, Y. 2021. Sentence-T5: Scalable Sentence Encoders from Pre-trained Text-to-Text Models. CoRR, abs/2108.08877

    2021 arXiv

  18. [26]

    H.; Ma, J.; Zhao, V

    Ni, J.; Qu, C.; Lu, J.; Dai, Z.; \' A brego, G. H.; Ma, J.; Zhao, V. Y.; Luan, Y.; Hall, K. B.; Chang, M.; and Yang, Y. 2022. Large Dual Encoders Are Generalizable Retrievers. In Goldberg, Y.; Kozareva, Z.; and Zhang, Y., eds., Proceedings of the 2022 Conference on Empirical M...

    2022

  19. [27]

    Nogueira, R.; Lin, J.; and Epistemic, A. 2019. From doc2query to docTTTTTquery. Online preprint, 6(2)

    2019

  20. [28]

    Raffel, C.; Shazeer, N.; Roberts, A.; Lee, K.; Narang, S.; Matena, M.; Zhou, Y.; Li, W.; and Liu, P. J. 2020. Exploring the Limits of Transfer Learning with a Unified Text-to-Text Transformer. J. Mach. Learn. Res., 21: 140:1--140:67

    2020

  21. [29]

    X.; Liu, J.; Wu, H.; Wen, J.; and Wang, H

    Ren, R.; Zhao, W. X.; Liu, J.; Wu, H.; Wen, J.; and Wang, H. 2023. TOME: A Two-stage Approach for Model-based Retrieval. In Rogers, A.; Boyd - Graber, J. L.; and Okazaki, N., eds., Proceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume ...

    2023

  22. [30]

    Rice, J. 1986. Poligon: A System for Parallel Problem Solving . Technical Report KSL-86-19, Dept.\ of Computer Science, Stanford Univ

    1986

  23. [31]

    Robertson, S.; Zaragoza, H.; et al. 2009. The probabilistic relevance framework: BM25 and beyond. Foundations and Trends in Information Retrieval , 3(4): 333--389

    2009

  24. [32]

    E.; and Walker, S

    Robertson, S. E.; and Walker, S. 1997. On relevance weights with little relevance information. In Proceedings of the 20th annual international ACM SIGIR conference on Research and development in information retrieval, 16--24

    1997

  25. [33]

    Robinson, A. L. 1980 a . New Ways to Make Microcircuits Smaller. Science, 208(4447): 1019--1022

    1980

  26. [34]

    Robinson, A. L. 1980 b . New Ways to Make Microcircuits Smaller---Duplicate Entry . Science, 208: 1019--1026

    1980

  27. [35]

    Tang, Y.; Zhang, R.; Guo, J.; Chen, J.; Zhu, Z.; Wang, S.; Yin, D.; and Cheng, X. 2023. Semantic-Enhanced Differentiable Search Index Inspired by Learning Strategies. In Singh, A. K.; Sun, Y.; Akoglu, L.; Gunopulos, D.; Yan, X.; Kumar, R.; Ozcan, F.; and Ye, J., eds., Proceedi...

    2023

  28. [36]

    P.; Schuster, T.; Cohen, W

    Tay, Y.; Tran, V.; Dehghani, M.; Ni, J.; Bahri, D.; Mehta, H.; Qin, Z.; Hui, K.; Zhao, Z.; Gupta, J. P.; Schuster, T.; Cohen, W. W.; and Metzler, D. 2022. Transformer Memory as a Differentiable Search Index. In Koyejo, S.; Mohamed, S.; Agarwal, A.; Belgrave, D.; Cho, K.; and O...

    2022

  29. [37]

    N.; Kaiser, L.; and Polosukhin, I

    Vaswani, A.; Shazeer, N.; Parmar, N.; Uszkoreit, J.; Jones, L.; Gomez, A. N.; Kaiser, L.; and Polosukhin, I. 2017. Attention Is All You Need. arXiv:1706.03762

    2017 arXiv

  30. [38]

    Wang, Y.; Hou, Y.; Wang, H.; Miao, Z.; Wu, S.; Chen, Q.; Xia, Y.; Chi, C.; Zhao, G.; Liu, Z.; Xie, X.; Sun, H.; Deng, W.; Zhang, Q.; and Yang, M. 2022 a . A Neural Corpus Indexer for Document Retrieval. In Koyejo, S.; Mohamed, S.; Agarwal, A.; Belgrave, D.; Cho, K.; and Oh, A....

    2022

  31. [39]

    Wang, Y.; Hou, Y.; Wang, H.; Miao, Z.; Wu, S.; Chen, Q.; Xia, Y.; Chi, C.; Zhao, G.; Liu, Z.; Xie, X.; Sun, H.; Deng, W.; Zhang, Q.; and Yang, M. 2022 b . A Neural Corpus Indexer for Document Retrieval. In Koyejo, S.; Mohamed, S.; Agarwal, A.; Belgrave, D.; Cho, K.; and Oh, A....

    2022

  32. [40]

    Wang, Z.; Zhou, Y.; Tu, Y.; and Dou, Z. 2023. NOVO: Learnable and Interpretable Document Identifiers for Model-Based IR . In Frommholz, I.; Hopfgartner, F.; Lee, M.; Oakes, M.; Lalmas, M.; Zhang, M.; and Santos, R. L. T., eds., Proceedings of the 32nd ACM International Confere...

    2023

  33. [41]

    N.; Ahmed, J.; and Overwijk, A

    Xiong, L.; Xiong, C.; Li, Y.; Tang, K.; Liu, J.; Bennett, P. N.; Ahmed, J.; and Overwijk, A. 2021. Approximate Nearest Neighbor Negative Contrastive Learning for Dense Text Retrieval. In 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austri...

    2021

  34. [42]

    Yoon, S.; Kim, C.; Lee, H.; Jang, J.; and Seo, M. 2023. Continually Updating Generative Retrieval on Dynamic Corpora. CoRR, abs/2305.18952

    2023 arXiv

  35. [43]

    Zhuang, S.; Ren, H.; Shou, L.; Pei, J.; Gong, M.; Zuccon, G.; and Jiang, D. 2022. Bridging the Gap Between Indexing and Retrieval for Differentiable Search Index with Query Generation. CoRR, abs/2206.10128

    2022 arXiv

  36. [44]

    , " * write output.state after.block = add.period write newline

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint howpublished institution isbn journal key month note number organization pages publisher school series title type volume year label extra.label sort.label short.list INTEGERS output.state before.all...

  37. [45]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 gl...

Pith tools

Reviewed August 8, 2026 · model on record in the stance chip above.