Pith. sign in

REVIEW 3 major objections 5 minor 41 references

Don't Reinvent the Wheel: Efficient Instruction-Following Text Embedding based on Guided Space Transformation

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read GSTransform claims instruction-following embeddings need no re-encoding: a single linear transform on precomputed vectors matches state-of-the-art quality at 6-300x lower real-time latency.

desk verdict The transformation-on-stored-vectors idea is genuinely new and the paper is well-built, but the headline quality and efficiency claims rest on a potentially contaminated evaluation and an incomplete time budget. read the letter →

arxiv 2505.24754 v1 pith:JZXKUHJ7 submitted 2025-05-30 cs.CL cs.AIcs.IR

Yingchaojie Feng , Yiqun Sun , Yandong Sun , Minfeng Zhu , Qiang Huang , Anthony K. H. Tung , Wei Chen This is my paper
classification cs.CLcs.AIcs.IR
keywords instruction-followingtextembeddingguidedspacetransformationcontrastivelearninglabeltaxonomycomputationalefficiencylargelanguagemodelsvectordatabases
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

Instruction-following text embeddings usually require pushing the entire corpus through the model again for each new instruction; GSTransform claims that is wasteful. The paper's thesis is that generic text embeddings already contain the information an instruction asks for, in latent form, and that a lightweight linear transformation can surface and reorganize it. To build that transformation cheaply, GSTransform samples a few thousand texts, uses an LLM to summarize and label them along the instruction's axis, and trains a single-layer encoder-decoder on those labels with a contrastive-plus-reconstruction loss. The authors report that this beats the state-of-the-art baselines on average across nine datasets while cutting real-time latency by 6-300x, making instruction-aware embeddings practical for large pre-indexed collections.

What carries the argument

The load-bearing object is the Label-guided Embedding Transformation: a single-layer linear encoder-decoder that maps a generic embedding into an instruction-aligned space. Its supervision comes from an Instruction-based Label Construction pipeline that summarizes sampled texts with an LLM under the user's instruction, embeds the summaries with a generic model, clusters them with k-means++, and uses contrastive LLM prompts to name each cluster; those labels annotate the sample texts, and the model is trained with a Euclidean contrastive loss that pulls same-label pairs together and pushes different-label pairs apart, plus a reconstruction loss that preserves the original space's structure.

What would settle it

Take a corpus where the instruction-specified attribute is absent from the text surface (for example, asking for author identity in anonymous news text) and generic embeddings provably carry no signal for it; if GSTransform's transformed embeddings still separate the classes, the claim that the linear transform surfaces latent instruction information cannot be right.

Watch

Extended reading notes

Core claim

In its own terms, the paper's discovery is that instruction-relevant information is already present in generic text embeddings and can be recovered by a learned linear map, so instruction following can be decoupled from embedding generation. The evidence is a transformation trained on 3,000 LLM-annotated in-domain texts that lifts the mean score across three instruction-awareness tasks from 55.31 (InBedder-Llama2) to 66.01, including a jump on Amazon Counterfactual clustering from 1.49 to 34.68. The transformation is model-agnostic: the same recipe improves UAE, Mxbai, and BGE backbones by large margins.

Load-bearing premise

The paper assumes that whatever an instruction asks about is already encoded in the generic embedding and can be extracted with a single linear layer.

Editorial extensions

If this is right

  • Corpora need to be encoded only once; subsequent instructions require a matrix multiply per embedding, so real-time latency stays near constant as the corpus grows.
  • The same generic index can serve many instructions, since each instruction defines its own transformation on the stored vectors.
  • Any future generic embedding model can be plugged in and made instruction-aware with the same 3,000-sample annotation budget.
  • Quality gains are largest where the baselines have headroom; datasets well covered by pretraining (AG-News, NYTClust) still improve but by less.
  • On corpora smaller than the annotation sample, direct re-encoding methods remain competitive, so the efficiency advantage emerges at scale.

Reading between the lines

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

  • If the central claim holds, vector database systems could offer instruction-dependent similarity search from a static index, turning instructions into cheap query-time projections instead of a re-indexing operation.
  • A testable extension is whether non-linear label boundaries can be recovered by widening the single linear layer; the paper's ablations show FDA underperforms, suggesting the linear assumption is doing real work rather than being vacuous.
  • The fixed 3,000-sample budget sets an amortization threshold: for small corpora the method's overhead may exceed direct re-encoding (the paper's own latency tables show parity near 3,000 samples), so the efficiency claim is inherently about scale.
  • The label taxonomy ties each transformation to the specific instruction's semantics; generalization to instructions outside the constructed taxonomy is not tested and is the natural next question.
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

3 major / 5 minor

Summary. The paper proposes GSTransform, a framework for instruction-following text embeddings that avoids re-encoding the full corpus for each instruction. Instead, it precomputes generic embeddings once and applies a lightweight linear encoder-decoder transformation, whose parameters are learned from a small set of LLM-annotated texts. Labels are constructed by summarizing a randomly sampled subset according to the instruction, embedding those summaries, clustering them, and generating category labels with an LLM. Experiments on clustering, STS, and triplet alignment across nine datasets claim that GSTransform outperforms InstructOR and InBedder while being 6–300x faster in real-time processing. The paper also reports ablations, parameter sensitivity studies, a case study, and a public code repository.

Significance. If the empirical claims survive closer scrutiny, the core idea is a practically attractive efficiency contribution: a linear transformation on precomputed embeddings gives near-constant per-instruction latency, reuses existing vector stores, and requires only a few thousand in-domain annotated texts. The paper is clearly written, includes a broad benchmark, ablates each component, and releases code. However, the headline quality claim is currently undermined by a possible train/evaluation overlap and by a supervision-imbalanced comparison, so the significance of the reported 66.01 vs. 55.31 margin cannot be assessed from Table 1 as presented.

major comments (3)
  1. [§3.2 and §4.2] The manuscript never states that the 3,000 texts used for label construction and transformation training are excluded from the downstream evaluation sets. Section 3.2 trains on 'the instruction-annotated subset (e.g., 3,000 samples)', while Section 4.2 evaluates clustering on the full datasets, STS on 50,000 pairs 'sampled from annotated datasets', and triplet alignment on 50,000 triplets 'randomly sampled' from the same datasets. Because the contrastive loss in Section 3.2 directly optimizes the embeddings of those 3,000 texts to be separable by the LLM-generated labels, any overlap between the training subset and the evaluation instances makes Table 1 a measure of fitting to training labels rather than of instruction-following generalization. Please state explicitly whether any sampled texts appear in the evaluation sets and, if so, re-run all metrics with the training subset excluded or provide disjoint splits.
  2. [§4.1 and Table 1] The comparison to InstructOR and InBedder is supervision-imbalanced: the baselines are evaluated zero-shot, while GSTransform receives 3,000 in-domain LLM-labeled texts per dataset (Section 3.1). Even with a clean held-out evaluation, the reported gains may reflect the additional labeled supervision rather than the space-transformation mechanism. Please add a supervision-matched baseline—for example, a linear probe or a linear transformation trained on the same 3,000 labels, or an instruction-aware model fine-tuned on the same data—so that the quality comparison isolates the contribution of the proposed method.
  3. [§4.4, Table 2, and Abstract] The claimed '6–300× speedups in real-time processing' are computed from the per-instruction transformation latency only (77–87 seconds in Table 2) and exclude the one-time pre-computing time (626–1,689 seconds). This metric is defensible if embeddings are precomputed and reused across many instructions, but the paper should state this amortization assumption explicitly and qualify the claim as marginal per-instruction latency for an already-indexed corpus, rather than end-to-end time for a single instruction.
minor comments (5)
  1. [§3.2] The sentence 'The encoder transforms the each input embedding vector' contains a typo; it should read 'The encoder transforms each input embedding vector'.
  2. [§4.4] The sentence 'Big Patent comprises 67.1K text with character lengths ranging from 2.76K to 3.11M' appears twice in the same paragraph; please delete the duplicate.
  3. [§4.3] The explanation for varying gains across datasets ('degree of alignment between the datasets and pretraining data of (large) language models') is speculative and not directly evidenced; please either support it with an analysis or soften the wording.
  4. [Limitations] The Limitations section discusses sampling sensitivity and encoder-decoder design, but it does not mention the risk of transductive evaluation from using the same dataset for label construction, training, and evaluation. Please add a discussion of this risk and of how the released code handles train/test separation.
  5. [Appendix C, Table 7] The instruction for IntEmo Intent differs across methods: InBedder asks 'What does the customer need?' while GSTransform uses 'What does the user care about? Just tell me the name of the thing.' If the two phrasings are not behaviorally equivalent, this introduces uncontrolled variance; please either align the instructions or report evidence that the semantic intent is unchanged.

Circularity Check

1 steps flagged · score 6.0 of 10

Table 1 quality gains are not fully predictive because the 3,000 fitted texts are not shown to be excluded from the clustering evaluation sets.

  1. fitted input called prediction [Section 3.1/3.2 vs Section 4.2, Table 1]
    "we randomly sample a representative subset (e.g., 3,000 texts in our experiments) for label construction; To train the model, we use the instruction-annotated subset (e.g., 3,000 samples), splitting it into 80% training and 20% validation; We assess clustering performance on three datasets: NYTClustering (NYTClust), Amazon CounterFactual (AmzCF), and MasakhaNews (MNews)."

    The 3,000-text subset is the training input to the encoder-decoder: the contrastive loss pulls same-label embeddings together and pushes different-label embeddings apart by the margin. The clustering evaluation in Table 1 is then scored with V-measure on the same listed datasets, and the paper never states that the 3,000 fitted texts are excluded from the clustering corpus. Thus the NYTClust/AmzCF/MNews V-measures include direct optimization targets of the contrastive loss, so part of the reported gain over zero-shot baselines is a fit to training labels rather than a prediction on unseen texts. Because the headline mean (66.01 vs 55.31) includes these clustering columns, the central quality claim is partially circular as reported, unless a held-out split is supplied.

full rationale

The only load-bearing circularity I can exhibit is the evaluation-protocol overlap: the same 3,000 texts that fit the linear transformation are not explicitly removed from the clustering sets whose V-measures populate Table 1. Otherwise the derivation is self-contained: the transformation is an ordinary supervised encoder-decoder trained with a stated dual objective, the efficiency claim follows from fixed-cost transformation of precomputed embeddings and is benchmarked independently, and there is no self-citation chain, uniqueness theorem, or ansatz smuggled in via citation. The paper should be read as reporting a promising supervised adaptation pipeline, but the headline quality advantage over zero-shot baselines is confounded by training/evaluation overlap and by supervision imbalance; a clean held-out comparison would resolve the partial circularity.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

All supervision comes from LLM annotations on in-domain samples; the method introduces no new physical entities. The listed hyperparameters and assumptions are the main external commitments.

free parameters (4)
  • number of clusters k = 50
    Set by hand as a default; the sensitivity study in Table 5 shows stability across 10 to 90, but k controls label taxonomy granularity and is not derived.
  • sampled subset size = 3,000
    Chosen by hand; Table 4 shows gains plateau around 3,000, but the value is a design choice made before evaluation.
  • loss weights beta1, beta2 = 1.0, 1.0
    Set to 1.0 by default with no tuning rationale; balances contrastive and reconstruction objectives.
  • contrastive margin m = not specified
    Introduced in the contrastive loss formula in Section 3.2 but no value is reported, making the exact training objective under-specified.
assumptions (4)
  • domain assumption Generic embeddings contain instruction-relevant information recoverable by a linear map
    Section 1 states the observation; Section 3.2 uses a linear encoder-decoder. If false, the transformation cannot produce instruction-aligned embeddings.
  • domain assumption LLM-generated summaries and labels are accurate, instruction-aligned, and mutually exclusive
    Section 3.1 relies on GPT-4o-mini for summarization, label generation, and classification; no verification of label quality is provided.
  • domain assumption A random 3,000-text sample is representative of each full corpus
    Section 3.1 samples randomly; the Limitations section notes sensitivity to data imbalance.
  • ad hoc to paper Equal-weight contrastive plus reconstruction loss preserves general semantics while enforcing instruction structure
    Section 3.2 sets beta1=beta2=1.0; no analysis shows this balance preserves the original embedding structure.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Don't Reinvent the Wheel: Efficient Instruction-Following Text Embedding based on Guided Space Transformation." pith.science (2026). https://pith.science/paper/JZXKUHJ7

@misc{pith2026250524754,
  author       = {Pith},
  title        = {Pith review of: Don't Reinvent the Wheel: Efficient Instruction-Following Text Embedding based on Guided Space Transformation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JZXKUHJ7}},
  note         = {Machine review of arXiv:2505.24754}
}
read the original abstract

In this work, we investigate an important task named instruction-following text embedding, which generates dynamic text embeddings that adapt to user instructions, highlighting specific attributes of text. Despite recent advancements, existing approaches suffer from significant computational overhead, as they require re-encoding the entire corpus for each new instruction. To address this challenge, we propose GSTransform, a novel instruction-following text embedding framework based on Guided Space Transformation. Our key observation is that instruction-relevant information is inherently encoded in generic embeddings but remains underutilized. Instead of repeatedly encoding the corpus for each instruction, GSTransform is a lightweight transformation mechanism that adapts pre-computed embeddings in real time to align with user instructions, guided by a small amount of text data with instruction-focused label annotation. We conduct extensive experiments on three instruction-awareness downstream tasks across nine real-world datasets, demonstrating that GSTransform improves instruction-following text embedding quality over state-of-the-art methods while achieving dramatic speedups of 6~300x in real-time processing on large-scale datasets. The source code is available at https://github.com/YingchaojieFeng/GSTransform.

Figures

Figures reproduced from arXiv: 2505.24754 by the authors.

Figure 1
Figure 1. An illustration of instruction-following text embeddings. Depending on different instructions, the [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. GSTransform consists of two core components: (1) Instruction-based Label Construction, which summa [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 4
Figure 4. The cosine distance relations between a triplet [PITH_FULL_IMAGE:figures/full_fig_p009_4.png] view at source ↗
Figures from the paper (1 more)
Figure 3
Figure 3. Figure 3: The embedding visualization of UAE and GSTransform (UAE). We use t-SNE for projection and encode the country labels in different colors. We also examine the cosine distance relation for a triplet text under different instructions in [PITH_FULL_IMAGE:figures/full_fig_p…

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

41 extracted references · 15 canonical work pages

  1. [1]

    online" 'onlinestring :=

    ENTRY address archivePrefix author booktitle chapter edition editor eid eprint eprinttype howpublished institution journal key month note number organization pages publisher school series title type volume year doi pubmed url lastchecked label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block STRING...

  2. [2]

    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 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    David Ifeoluwa Adelani, Marek Masiak, Israel Abebe Azime, Jesujoba Alabi, Atnafu Lambebo Tonja, Christine Mwase, Odunayo Ogundepo, Bonaventure FP Dossou, Akintunde Oladipo, Doreen Nixdorf, et al. 2023. https://doi.org/10.18653/V1/2023.IJCNLP-MAIN.10 Masakhanews: News topic classification for african languages . In Proceedings of the 13th International Joi...

    doi:10.18653/v1/2023.ijcnlp-main.10 2023

  4. [4]

    Aggarwal and ChengXiang Zhai

    Charu C. Aggarwal and ChengXiang Zhai. 2012. https://doi.org/10.1007/978-1-4614-3223-4\_4 A survey of text clustering algorithms . Mining Text Data, pages 77--128

    doi:10.1007/978-1-4614-3223-4 2012

  5. [5]

    Cer, Mona T

    Eneko Agirre, Daniel M. Cer, Mona T. Diab, and Aitor Gonzalez - Agirre. 2012. https://aclanthology.org/S12-1051/ Semeval-2012 task 6: A pilot on semantic textual similarity . In Proceedings of the 6th International Workshop on Semantic Evaluation, SemEval@NAACL-HLT 2012, Montr \' e al, Canada, June 7-8, 2012 , pages 385--393

    work page 2012

  6. [6]

    Cer, Mona T

    Eneko Agirre, Daniel M. Cer, Mona T. Diab, Aitor Gonzalez - Agirre, and Weiwei Guo. 2013. https://aclanthology.org/S13-1004/ *sem 2013 shared task: Semantic textual similarity . In Proceedings of the Second Joint Conference on Lexical and Computational Semantics, *SEM 2013, June 13-14, 2013, Atlanta, Georgia, USA , pages 32--43

    work page 2013

  7. [7]

    David Arthur and Sergei Vassilvitskii. 2007. http://dl.acm.org/citation.cfm?id=1283383.1283494 k-means++: the advantages of careful seeding . In Proceedings of the eighteenth annual ACM-SIAM Symposium on Discrete Algorithms (SODA), pages 1027--1035

    arXiv 2007

  8. [8]

    John, Noah Constant, Mario Guajardo - Cespedes, Steve Yuan, Chris Tar, Brian Strope, and Ray Kurzweil

    Daniel Cer, Yinfei Yang, Sheng - yi Kong, Nan Hua, Nicole Limtiaco, Rhomni St. John, Noah Constant, Mario Guajardo - Cespedes, Steve Yuan, Chris Tar, Brian Strope, and Ray Kurzweil. 2018. https://doi.org/10.18653/V1/D18-2029 Universal sentence encoder for english . In Proceedings of the 2018 Conference on Empirical Methods in Natural Language Processing: ...

    doi:10.18653/v1/d18-2029 2018

Show all 41 references
  1. [9]

    Chengliang Chai, Jiayi Wang, Nan Tang, Ye Yuan, Jiabin Liu, Yuhao Deng, and Guoren Wang. 2023. https://doi.org/10.1145/3580305.3599326 Efficient coreset selection with cluster-based methods . In Proceedings of the 29th ACM SIGKDD Conference on Knowledge Discovery and Data Mini...

    2023

  2. [10]

    Jacob Devlin, Ming - Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. https://doi.org/10.18653/V1/N19-1423 BERT: pre-training of deep bidirectional transformers for language understanding . In Proceedings of the 2019 Conference of the North A merican Chapter of the Associa...

    2019 doi

  3. [11]

    Quan Do. 2019. https://www.theseus.fi/bitstream/handle/10024/226938/Quan_Do.pdf Jigsaw unintended bias in toxicity classification

    2019

  4. [12]

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. https://doi.org/10.18653/v1/2021.emnlp-main.552 Simcse: Simple contrastive learning of sentence embeddings . In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 6894--6910

    2021 doi

  5. [13]

    Maarten Grootendorst. 2022. https://doi.org/10.48550/ARXIV.2203.05794 Bertopic: Neural topic modeling with a class-based TF-IDF procedure . arXiv

    2022 doi arXiv

  6. [14]

    Vladimir Karpukhin, Barlas Oguz, Sewon Min, Patrick Lewis, Ledell Wu, Sergey Edunov, Danqi Chen, and Wen - tau Yih. 2020. https://doi.org/10.18653/V1/2020.EMNLP-MAIN.550 Dense passage retrieval for open-domain question answering . In Proceedings of the 2020 Conference on Empir...

    2020 doi

  7. [15]

    Sean Lee, Aamir Shakir, Darius Koenig, and Julius Lipp. 2024. https://www.mixedbread.ai/blog/mxbai-embed-large-v1 Open source strikes bread - new fluffy embeddings model

    2024

  8. [16]

    Yibin Lei, Di Wu, Tianyi Zhou, Tao Shen, Yu Cao, Chongyang Tao, and Andrew Yates. 2024. https://doi.org/10.18653/V1/2024.ACL-LONG.546 Meta-task prompting elicits embeddings from large language models . In Proceedings of the 62nd Annual Meeting of the Association for Computatio...

    2024 doi

  9. [17]

    Xianming Li and Jing Li. 2024. https://doi.org/10.18653/V1/2024.ACL-LONG.101 Aoe: Angle-optimized embeddings for semantic textual similarity . In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (ACL), pages 1825--1839

    2024 doi

  10. [18]

    Sebastian Mika, Gunnar Ratsch, Jason Weston, Bernhard Scholkopf, and Klaus-Robert Mullers. 1999. https://doi.org/10.1109/NNSP.1999.788121 Fisher discriminant analysis with kernels . In Neural networks for signal processing IX: Proceedings of the 1999 IEEE signal processing soc...

    1999

  11. [19]

    Corrado, and Jeffrey Dean

    Tom \' a s Mikolov, Ilya Sutskever, Kai Chen, Gregory S. Corrado, and Jeffrey Dean. 2013. https://proceedings.neurips.cc/paper/2013/hash/9aa42b31882ec039965f3c4923ce901b-Abstract.html Distributed representations of words and phrases and their compositionality . In Proceedings ...

    2013

  12. [20]

    Dehai Min, Zhiyang Xu, Guilin Qi, Lifu Huang, and Chenyu You. 2025. https://aclanthology.org/2025.naacl-long.234/ Unihgkr: Unified instruction-aware heterogeneous knowledge retrievers . In Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Associa...

    2025

  13. [21]

    Niklas Muennighoff, Hongjin Su, Liang Wang, Nan Yang, Furu Wei, Tao Yu, Amanpreet Singh, and Douwe Kiela. 2025. https://openreview.net/forum?id=BC4lIvfSzv Generative representational instruction tuning . In The Thirteenth International Conference on Learning Representations (ICLR)

    2025

  14. [22]

    Niklas Muennighoff, Nouamane Tazi, Loic Magne, and Nils Reimers. 2023. https://aclanthology.org/2023.eacl-main.148/ MTEB: Massive Text Embedding Benchmark . In Proceedings of the 17th Conference of the European Chapter of the Association for Computational Linguistics (EACL), p...

    2023

  15. [23]

    Hanseok Oh, Hyunji Lee, Seonghyeon Ye, Haebin Shin, Hansol Jang, Changwook Jun, and Minjoon Seo. 2024. https://doi.org/10.48550/ARXIV.2402.14334 INSTRUCTIR: A benchmark for instruction following of information retrieval models . arXiv

    2024 doi arXiv

  16. [24]

    James O'Neill, Polina Rozenshtein, Ryuichi Kiryo, Motoko Kubota, and Danushka Bollegala. 2021. https://doi.org/10.18653/V1/2021.EMNLP-MAIN.568 I wish I would have loved this one, but I didn't - A multilingual dataset for counterfactual detection in product review . In Proceedi...

    2021 doi

  17. [25]

    Malte Ostendorff, Till Blume, Terry Ruas, Bela Gipp, and Georg Rehm. 2022. https://doi.org/10.1145/3529372.3530912 Specialized document embeddings for aspect-based similarity of research papers . In Proceedings of the 22nd ACM/IEEE Joint Conference on Digital Libraries (JCDL), page 7

    2022

  18. [26]

    Letian Peng, Yuwei Zhang, Zilong Wang, Jayanth Srinivasa, Gaowen Liu, Zihan Wang, and Jingbo Shang. 2024. https://doi.org/10.18653/V1/2024.ACL-LONG.27 Answer is all you need: Instruction-following text embedding via answering the question . In Proceedings of the 62nd Annual Me...

    2024 doi

  19. [27]

    Jeffrey Pennington, Richard Socher, and Christopher D. Manning. 2014. https://doi.org/10.3115/V1/D14-1162 Glove: Global vectors for word representation . In Proceedings of the 2014 Conference on Empirical Methods in Natural Language Processing (EMNLP), pages 1532--1543

    2014 doi

  20. [28]

    Nils Reimers and Iryna Gurevych. 2019. https://doi.org/10.18653/V1/D19-1410 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 ...

    2019 doi

  21. [29]

    Andrew Rosenberg and Julia Hirschberg. 2007. https://aclanthology.org/D07-1043/ V-measure: A conditional entropy-based external cluster evaluation measure . In Proceedings of the 2007 joint conference on empirical methods in natural language processing and computational natura...

    2007

  22. [30]

    Paul R \"o ttger, Haitham Seelawi, Debora Nozza, Zeerak Talat, and Bertie Vidgen. 2022. https://doi.org/10.18653/v1/2022.woah-1.15 Multilingual H ate C heck: Functional tests for multilingual hate speech detection models . In Proceedings of the Sixth Workshop on Online Abuse a...

    2022 doi

  23. [31]

    Eva Sharma, Chen Li, and Lu Wang. 2019. https://doi.org/10.18653/V1/P19-1212 BIGPATENT: A large-scale dataset for abstractive and coherent summarization . In Proceedings of the 57th Conference of the Association for Computational Linguistics (ACL), pages 2204--2213

    2019 doi

  24. [32]

    Smith, Luke Zettlemoyer, and Tao Yu

    Hongjin Su, Weijia Shi, Jungo Kasai, Yizhong Wang, Yushi Hu, Mari Ostendorf, Wen-tau Yih, Noah A. Smith, Luke Zettlemoyer, and Tao Yu. 2023. https://aclanthology.org/2023.findings-acl.71/ One embedder, any task: Instruction-finetuned text embeddings . In Findings of the Associ...

    2023

  25. [33]

    Weiwei Sun, Zhengliang Shi, Wu Long, Lingyong Yan, Xinyu Ma, Yiding Liu, Min Cao, Dawei Yin, and Zhaochun Ren. 2024. https://aclanthology.org/2024.emnlp-main.778 MAIR: A massive benchmark for evaluating instructed retrieval . In Proceedings of the 2024 Conference on Empirical ...

    2024

  26. [34]

    Nandan Thakur, Nils Reimers, Andreas R \" u ckl \' e , Abhishek Srivastava, and Iryna Gurevych. 2021. https://datasets-benchmarks-proceedings.neurips.cc/paper/2021/hash/65b9eea6e1cc6bb9f0cd2a47751a186f-Abstract-round2.html BEIR: A heterogeneous benchmark for zero-shot evaluati...

    2021

  27. [35]

    Gomez, Lukasz Kaiser, and Illia Polosukhin

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones, Aidan N. Gomez, Lukasz Kaiser, and Illia Polosukhin. 2017. https://proceedings.neurips.cc/paper/2017/hash/3f5ee243547dee91fbd053c1c4a845aa-Abstract.html Attention is all you need . In Advances in Neural I...

    2017

  28. [36]

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Majumder, and Furu Wei. 2023. https://doi.org/10.48550/arXiv.2401.00368 Improving text embeddings with large language models . arXiv

    2023 doi arXiv

  29. [37]

    Lawrie, and Luca Soldaini

    Orion Weller, Benjamin Chang, Sean MacAvaney, Kyle Lo, Arman Cohan, Benjamin Van Durme, Dawn J. Lawrie, and Luca Soldaini. 2025 a . https://aclanthology.org/2025.naacl-long.597/ Followir: Evaluating and teaching information retrieval models to follow instructions . In Proceedi...

    2025

  30. [38]

    Lawrie, Ashwin Paranjape, Yuhao Zhang, and Jack Hessel

    Orion Weller, Benjamin Van Durme, Dawn J. Lawrie, Ashwin Paranjape, Yuhao Zhang, and Jack Hessel. 2025 b . https://openreview.net/forum?id=odvSjn416y Promptriever: Instruction-trained retrievers can be prompted like language models . In The Thirteenth International Conference ...

    2025

  31. [39]

    Shitao Xiao, Zheng Liu, Peitian Zhang, and Niklas Muennighoff. 2023. https://doi.org/10.48550/ARXIV.2309.07597 C-pack: Packaged resources to advance general chinese embedding . arXiv

    2023 doi arXiv

  32. [40]

    Xiang Zhang, Junbo Jake Zhao, and Yann LeCun. 2015. https://proceedings.neurips.cc/paper/2015/hash/250cf8b51c773f3f8dc8b4be867a9a02-Abstract.html Character-level convolutional networks for text classification . In Advances in Neural Information Processing Systems 28: Annual Co...

    2015

  33. [41]

    Wenjie Zhuo, Yifan Sun, Xiaohan Wang, Linchao Zhu, and Yi Yang. 2023. https://doi.org/10.18653/V1/2023.ACL-LONG.677 Whitenedcse: Whitening-based contrastive learning of sentence embeddings . In Proceedings of the 61st Annual Meeting of the Association for Computational Linguis...

    2023 doi

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.