Pith. sign in

REVIEW 3 major objections 4 minor 1 cited by

QZhou-Embedding claims first place on both the English MTEB and Chinese CMTEB leaderboards as of August 27, 2025, arguing that data quality and diversity—not architecture—drive embedding performance.

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 →

QZhou-Embedding reports state-of-the-art average scores on MTEB and CMTEB as of August 27, 2025, using a two-stage multi-task pipeline with LLM-based data synthesis.

T0 review reviewed 2026-08-05 challenge →

load-bearing objection A workmanlike embedding report with a plausible SOTA claim that is not auditable until the contamination filter is quantified. the 3 major comments →

arxiv 2508.21632 v1 pith:X6U2NLAA submitted 2025-08-29 cs.CL cs.AI

QZhou-Embedding Technical Report

classification cs.CL cs.AI
keywords text embeddingretrievalMTEBCMTEBmulti-task learningdata synthesishard negative miningcontrastive learning
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The reading

This paper presents QZhou-Embedding, a general-purpose text embedding model built on the 7B-parameter Qwen2.5-Instruct language model, and claims it reached the top average score on both the English MTEB and Chinese CMTEB leaderboards as of August 27, 2025. The authors attribute the gain chiefly to data, not architecture: a unified multi-task framework converts heterogeneous text sources into retrieval, NLI, and classification training formats, and an LLM-based synthesis pipeline generates paraphrases, semantic augmentations, and hard negatives, yielding about 11 million training quadruples. Training proceeds in two stages—retrieval-only first, then all tasks under a controlled 0.72 retrieval sampling ratio—with dataset-level batching and task-specific losses. The paper's stated finding is that higher-quality, more diverse data, especially LLM-generated data, is the decisive factor in advancing retrieval model performance. If the result is genuine, the recipe transfers to other base models and languages without architectural changes.

Core claim

Central claim: an embedding model can reach state-of-the-art results on both major benchmarks through data and training, not a new architecture. Around the Qwen2.5-7B-Instruct base (bidirectional attention, mean pooling), a unified multi-task framework converts heterogeneous text into (query, document, relation) triples for retrieval, NLI, and classification, each with a matched loss: InfoNCE plus query-query negatives, Cosent ranking, and InfoNCE with class-label masking. An LLM pipeline adds paraphrases, augmentations, and hard negatives (~11M quadruples); a two-stage schedule trains retrieval first, then all tasks at retrieval ratio 0.72 with dataset-level batching. Reported result: first

What carries the argument

The load-bearing mechanism is a data transformation-and-synthesis pipeline paired with a task-matched training schedule around a 7B decoder-only base. The data side converts heterogeneous sources into three formats—retrieval, NLI, classification—then applies LLM paraphrasing, semantic augmentation, and hard-negative generation, reaching 11M quadruples. The training side couples three specialized losses—InfoNCE with query-query negatives (retrieval), Cosent on ordinal labels (NLI), InfoNCE with class-label masking (classification)—with the Data Grouping Strategy (one dataset per batch, raising in-batch negative difficulty) and a two-stage schedule whose global retrieval sampling ratio ηRET =

Load-bearing premise

The leaderboard claim stands only if the 'contamination exclusion' filter in Section 6.1 removed every training sample similar enough to MTEB/CMTEB test sets to inflate scores—because the training corpus explicitly includes those benchmarks' own training splits.

What would settle it

Audit the released training corpus for overlap with MTEB/CMTEB test sets using n-gram and embedding similarity between retained quadruples and test queries/passages; if a non-negligible share of retained samples are near-duplicates of test items, the SOTA result is partly a training-data artifact. The cleaner empirical check: benchmark QZhou-Embedding on newly built retrieval tasks that share no data with its training set, and compare its margin over the same leaderboard competitors.

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

If this is right

  • If the reported numbers are correct, the recipe is transferable: any team with the same base model can adopt the data transformation, LLM synthesis, and two-stage schedule, since the claimed gains come from data and training, not new architecture.
  • A single 7B checkpoint claims top scores on both the English MTEB and Chinese CMTEB leaderboards, evidence that one general-purpose model can serve both languages without separate bilingual tuning.
  • The three-task transformation scheme widens usable training material to news, claims with evidence, QA forums, reviews, and legal and medical texts, all converted into contrastive triplets.
  • The two-stage schedule with a controlled retrieval ratio implies retrieval competence can be established first and then extended to classification, STS, and pairwise tasks without sacrificing retrieval.
  • Released weights under Apache 2.0, along with evaluation code and instructions, make the leaderboard claims independently checkable.

Where Pith is reading between the lines

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

  • Because the training corpus explicitly includes MTEB evaluation datasets' training splits with only an unquantified contamination filter, the SOTA claim is exactly as strong as that filter; a cleaner test would benchmark the model on held-out tasks with no training overlap.
  • The paper attributes the gains to data quality and diversity without an ablation; holding compute fixed and removing paraphrasing, augmentation, or hard-negative synthesis one at a time would pin down which component carries the improvement.
  • The dataset-level grouping and ratio-controlled two-stage schedule generalize beyond embeddings to any multi-task contrastive setup, such as reranker or retrieval-augmented generation training.
  • The 1,536-token training passage length with 8k practical extrapolation suggests the same recipe applies to long-document retrieval benchmarks, a regime MTEB's average score under-weights.
Share X Bluesky LinkedIn Reddit HN

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 / 4 minor

Summary. The paper presents QZhou-Embedding, a 7B-parameter decoder-only text embedding model built on Qwen2.5-7B-Instruct with bidirectional attention and mean pooling. The main methodological contributions are: (1) a unified multi-task framework that transforms heterogeneous data into retrieval, NLI, and classification formats; (2) an LLM-based data synthesis pipeline using paraphrasing, semantic augmentation, and hard-negative generation; (3) a dataset-level grouping strategy with exponential sampling weights; and (4) a two-stage training scheme (retrieval-only pretraining followed by full-task fine-tuning). The central empirical claim is that the model ranks first on both the MTEB (English) and CMTEB (Chinese) leaderboards as of August 27, 2025, with the reported averages in Tables 2 and 3. The model weights and evaluation code are publicly released.

Significance. If the reported leaderboard results are trustworthy, this is a meaningful empirical contribution: a 7B-parameter embedding model that simultaneously achieves strong performance on English and Chinese benchmarks is practically useful, and the two-stage training plus synthetic hard-negative pipeline is a reproducible recipe for the community. The paper's strengths include the public release of weights under Apache 2.0, the inclusion of evaluation code, and a relatively detailed description of data transformation and loss functions. The loss equations (1) and (3) follow standard InfoNCE-based contrastive formulations, and the two-stage design is clearly stated. However, the significance of the headline result depends entirely on the credibility of the MTEB/CMTEB scores, and the manuscript currently provides insufficient evidence that the evaluation is uncontaminated: it explicitly trains on the training splits of several MTEB evaluation datasets and describes a 'contamination exclusion' step without quantifying it. This is the load-bearing point that the rest of the report focuses on.

major comments (3)
  1. [Section 6.1] The training set includes the training splits of Imdb-Classification, MassiveIntent, MassiveScenario, STS12, LCQMC, PAWSX, and STSB, all of which are evaluation tasks in MTEB/CMTEB. The only safeguard is described as 'contamination exclusion applied to remove samples highly similar to test sets,' with no similarity measure, threshold, removal count, or post-hoc overlap audit reported. Since the paper's central claim is first place on both leaderboards, an incomplete filter could directly inflate the reported averages. Please provide (a) the exact filtering procedure and threshold, (b) the number of removed samples per dataset, and (c) an ablation rerunning the final evaluation with these seven datasets removed from training. Without (c) the reader cannot distinguish a genuine generalization improvement from training/evaluation overlap.
  2. [Section 6.5] The SOTA claim is based on single-run scores in Tables 2 and 3, with no confidence intervals, standard deviations over repeated evaluations, or significance testing. The margin over the next best model is small (MTEB mean 69.52 vs. 68.71 for LGAI-Embedding-Preview; CMTEB mean 78.58 vs. 76.68 for Seed1.6-embedding). Both differences could be within run-to-run variability. Please report variance estimates or at least multiple evaluation runs, and, if the scores are taken from the official leaderboard, provide the exact leaderboard snapshot and submission records rather than only a private run.
  3. [Abstract / Table 3] The abstract claims 'state-of-the-art performance on tasks including reranking, clustering, etc.' This is not consistent with the CMTEB results in Table 3: the model's clustering score is 70.91, while Qwen3-Embedding-8B scores 80.08 on the same task type. The claim should be restricted to the English MTEB task types, or otherwise reconciled with the CMTEB numbers. As written, the abstract overstates the cross-benchmark clustering result.
minor comments (4)
  1. [Table 1] The two rightmost columns are mislabeled: the text in Section 6.3 states Stage 1 LR=3e-5 and 32k steps, and Stage 2 LR=2e-5 and 8k steps, but Table 1 shows 'Steps 3e-5 2e-5' and 'LR 32k 8k'. Swap the column headings or values.
  2. [Section 6.1] Typo: 'Sentention 3.2' should be 'Section 3.2'. Also the phrase 'we solely utilized the proposed method to update harder negative samples' is confusing; it should clarify that the collected datasets were kept as-is except for negative replacement.
  3. [Equation (2)] The Cosent loss notation is ambiguous: the condition 'sim(i,j) > sim(k,l)' and the exponent use the same symbol 'sim' as the model similarity function. Please define explicit label scores (e.g., s_{ij}) versus model cosine similarity (e.g., cos(x_i,x_j)) so the pair-ordering semantics are clear.
  4. [Section 6.1] The final training set is described as '11M quadruples (query, pos, neg, instruction).' This format does not naturally cover the NLI/STS data where a scalar score is used with Cosent loss, nor the classification data processed example-wise. Please specify how these data types are encoded in the quadruple format.

Circularity Check

1 steps flagged

MTEB/CMTEB SOTA claim rests on training splits of the same benchmarks with an undisclosed contamination filter.

specific steps
  1. fitted input called prediction [Section 6.1 (Training Dataset); claim in Abstract and Section 6.5 (Main Results)]
    "We further integrate MTEB evaluation-related datasets like Imdb-Classification, MassiveIntent-Classification, MassiveScenario-Classification, STS12, LCQMC, PAWSX, and STSB, we utilized the training split from these datasets with contamination exclusion applied to remove samples highly similar to test sets. For the MTEB training set, we adopt instructions corresponding to its evaluation(consistent with Qwen3-Embedding runtime)."

    The central claim is SOTA on MTEB/CMTEB. The training corpus explicitly includes the training splits of MTEB evaluation datasets (Imdb-Classification, MassiveIntent/Scenario, STS12, LCQMC, PAWSX, STSB) and uses evaluation-matched instructions for the MTEB training set. The only separation is an unspecified 'contamination exclusion' with no similarity measure, threshold, or removal count, and no ablation withholding these splits. Consequently the leaderboard score is not verifiably independent of training input: without evidence that the filter removes all near-duplicates of test items, the reported SOTA can be explained by benchmark training data rather than by generalization. The prediction (MTEB/CMTEB rank) is thus coupled to the benchmark's own data by construction.

full rationale

The paper's derivation chain (data transformation -> multi-task training -> SOTA) contains one load-bearing evaluation-loop step: the training data includes the training splits of MTEB evaluation datasets and uses evaluation-matched instructions, while the headline result is ranking first on MTEB/CMTEB. Because the contamination filter is unquantified and no withheld-split ablation is reported, the SOTA claim cannot be independently separated from benchmark training input. This is partial circularity rather than a full self-derivation: the model might still generalize and the filter might be adequate, but the evidence as written does not permit verification. No other circular pattern is present: there are no load-bearing self-citations, imported uniqueness theorems, ansatz-smuggling citations, or renaming of known results as organization. The loss functions, data grouping, and two-stage training are standard and externally grounded.

Axiom & Free-Parameter Ledger

4 free parameters · 3 axioms · 0 invented entities

The central claim depends on several tuned hyperparameters (η, α, temperature) and on the validity of MTEB/CMTEB as fair benchmarks. The most consequential unverified item is the contamination exclusion threshold, which directly affects whether the leaderboard result is a genuine generalization improvement or an artifact of benchmark training data.

free parameters (4)
  • retrieval ratio eta (ηRET) = 0.72
    Controls the proportion of retrieval data in stage-2 training (Section 6.3). Chosen based on observed retrieval degradation from adding non-retrieval data.
  • sampling exponent alpha (α) = not disclosed
    Exponent in dataset sampling weight formula (Equation 4). Value is not reported in the paper, suggesting it was set by hand or empirical tuning.
  • cosine temperature (τ) = 0.02
    Temperature in InfoNCE and Cosent losses (Section 6.3). Standard but hand-tuned.
  • contamination exclusion threshold = unspecified
    Used to remove training samples 'highly similar to test sets' (Section 6.1). The threshold is not quantified, and the central claim depends on its effectiveness.
axioms (3)
  • domain assumption MTEB and CMTEB leaderboard scores are accurate and comparably measured for all models.
    The SOTA claim assumes the externally maintained leaderboard is reliable and that the reported scores are computed under the same protocol for every model.
  • domain assumption The Qwen2.5-7B-Instruct base model provides a strong initialization, and full-parameter fine-tuning preserves its capabilities.
    The paper builds on this foundation model without justification; the success of the recipe depends on this base model's quality.
  • ad hoc to paper The contamination exclusion procedure successfully removes all near-duplicate training samples that overlap with MTEB/CMTEB test sets.
    Section 6.1 asserts the exclusion but provides no details or validation. The central claim of SOTA would be invalid if this assumption fails.

reviewed 2026-08-05 · how reviews work

0 comments
Cite this review

Pith. "Pith review of QZhou-Embedding Technical Report." pith.science (2026). https://pith.science/paper/X6U2NLAA

@misc{pith2026250821632,
  author       = {Pith},
  title        = {Pith review of: QZhou-Embedding Technical Report},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/X6U2NLAA}},
  note         = {Machine review of arXiv:2508.21632}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

We present QZhou-Embedding, a general-purpose contextual text embedding model with exceptional text representation capabilities. Built upon the Qwen2.5-7B-Instruct foundation model, we designed a unified multi-task framework comprising specialized data transformation and training strategies. The data transformation scheme enables the incorporation of more diverse textual training datasets, while the task-specific training strategies enhance model learning efficiency. We developed a data synthesis pipeline leveraging LLM API, incorporating techniques such as paraphrasing, augmentation, and hard negative example generation to improve the semantic richness and sample difficulty of the training set. Additionally, we employ a two-stage training strategy, comprising initial retrieval-focused pretraining followed by full-task fine-tuning, enabling the embedding model to extend its capabilities based on robust retrieval performance. Our model achieves state-of-the-art results on the MTEB and CMTEB benchmarks, ranking first on both leaderboards (August 27 2025), and simultaneously achieves state-of-the-art performance on tasks including reranking, clustering, etc. Our findings demonstrate that higher-quality, more diverse data is crucial for advancing retrieval model performance, and that leveraging LLMs generative capabilities can further optimize data quality for embedding model breakthroughs. Our model weights are released on HuggingFace under Apache 2.0 license. For reproducibility, we provide evaluation code and instructions on GitHub.

Figures

Figures reproduced from arXiv: 2508.21632 by Bin Chen, En Xu, Haibiao Chen, Peng Yu, Yinfei Xu.

Figure 1
Figure 1. Figure 1: QZhou-Embedding Architecture to enable comprehensive context capture, and (2) employing mean pooling with sub￾sequent normalization to produce final embedding vectors. The model architecture is shown in [PITH_FULL_IMAGE:figures/full_fig_p006_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: CLS-oriented data transformation sampling other texts sharing the same label as positive examples, and selecting texts from different labels as negative examples [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: LLM-based Paraphrasing Workflow 4.2 Semantic Diversity Enhancement Merely augmenting data through superficial structural modifications yields negligible improvements in model capabilities, as generalization relies not only on structural dis￾entanglement but also on diverse topics and content to ensure uniform vector rep￾resentations in the spatial domain. Therefore, beyond paraphrasing, we propose an augme… view at source ↗
Figure 4
Figure 4. Figure 4: Semantic Augmentation Workflow [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Hard Negative Synthesis Workflow 4.3 More challenging embeddings Hard negative examples are crucial for enhancing the performance of text embedding models, often requiring substantial effort to acquire. Leveraging the linguistic capabili￾ties of large language models, we design an automated hard negative synthesis method tailored for retrieval datasets. Our domain-specific experiments demonstrate that larg… view at source ↗
Figure 6
Figure 6. Figure 6: Training pipeline 5 Training Optimization 5.1 Data Grouping Strategy Prior works like Linq-Embedding[52] and SFR-Embedding-Mistral[30] adopted task￾homogeneous batching, partitioning data by task rather than mixing them, and sam￾pling tasks based on weighted randomness during training. Building on this, we propose a refined Data Grouping Strategy, extending the granularity from task-level to dataset￾level … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. TabEmbed: Benchmarking and Learning Generalist Embeddings for Tabular Understanding

    cs.CL 2026-05 unverdicted novelty 7.0

    TabEmbed is the first generalist embedding model for tabular data that unifies classification and retrieval in one space via contrastive learning and outperforms text embedding models on the new TabBench benchmark.

Reference graph

Works this paper leans on

78 extracted references · 50 canonical work pages · cited by 1 Pith paper · 2 internal anchors

  1. [1]

    Robertson, Stephen E., and Steve Walker. ”Some simple effective approximations to the 2-poisson model for probabilistic weighted retrieval.” In SIGIR’9 4: Proceedings 16 QZhou-Embedding Technical Report Kingsoft AI of the Seventeenth Annual International ACM-SIGIR Conferen ce on Research and Development in Information Retrieval, organised by Dublin City Un...

  2. [2]

    Bert: Pre- training of deep bidirectional transformers for language underst anding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutano va. Bert: Pre- training of deep bidirectional transformers for language underst anding. arXiv preprint arXiv:1810.04805, 2018

  3. [3]

    Exploring the limits of tr ansfer learn- ing with a unified text-to-text transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, Katherine Lee, Shara n Narang, Michael Matena, Yanqi Zhou, Wei Li, and Peter J Liu. Exploring the limits of tr ansfer learn- ing with a unified text-to-text transformer. Journal of machine le arning research, 21(140):1–67, 2020

  4. [4]

    Text embeddings by weakly-super vised con- trastive pre-training

    Liang Wang, Nan Yang, Xiaolong Huang, Binxing Jiao, Linjun Yang, D axin Jiang, Rangan Majumder, and Furu Wei. Text embeddings by weakly-super vised con- trastive pre-training. arXiv preprint arXiv:2212.03533, 2022

  5. [5]

    Unsupervised dense information retrieval with contrastive learning

    Gautier Izacard, Mathilde Caron, Lucas Hosseini, Sebastian Ried el, Piotr Bo- janowski, Armand Joulin, and Edouard Grave. Unsupervised dense information retrieval with contrastive learning. arXiv preprint arXiv:2112.0911 8, 2021

  6. [6]

    Sentence-bert: Sentence em beddings using siamese bert-networks

    Nils Reimers and Iryna Gurevych. Sentence-bert: Sentence em beddings using siamese bert-networks. arXiv preprint arXiv:1908.10084, 2019

  7. [7]

    Tianyu Gao, Xingcheng Yao, and Danqi Chen. 2021. SimCSE: Simple contrastive learning of sentence embeddings. In Proceedings of the 2021 Conf erence on Empir- ical Methods in Natural Language Processing, pages 6894–6910, Online and Punta Cana, Dominican Republic. Association for Computational Linguistics

  8. [8]

    Large du al encoders are generalizable retrievers

    Jianmo Ni, Chen Qu, Jing Lu, Zhuyun Dai, Gustavo Hern´ andez ´Abrego, Ji Ma, Vincent Y Zhao, Yi Luan, Keith B Hall, Ming-Wei Chang, et al. Large du al encoders are generalizable retrievers. arXiv preprint arXiv:2112.07899, 202 1

  9. [9]

    Kaplan, Pra- fulla Dhariwal, Arvind Neelakantan et al

    Brown, Tom, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared D . Kaplan, Pra- fulla Dhariwal, Arvind Neelakantan et al. ”Language models are few-s hot learners.” Advances in neural information processing systems 33 (2020): 18 77-1901

  10. [10]

    ”F ine-tuning llama for multi-stage text retrieval.” In Proceedings of the 47th Int ernational ACM SIGIR Conference on Research and Development in Information Re trieval, pp

    Ma, Xueguang, Liang Wang, Nan Yang, Furu Wei, and Jimmy Lin. ”F ine-tuning llama for multi-stage text retrieval.” In Proceedings of the 47th Int ernational ACM SIGIR Conference on Research and Development in Information Re trieval, pp. 2421-

  11. [11]

    ”Repetition improves language model embeddings.” a rXiv preprint arXiv:2402.15449 (2024)

    Springer, Jacob Mitchell, Suhas Kotha, Daniel Fried, Graham Ne ubig, and Aditi Raghunathan. ”Repetition improves language model embeddings.” a rXiv preprint arXiv:2402.15449 (2024)

  12. [12]

    ”Llm2vec: Large languag e models are secretly powerful text encoders.” arXiv preprint arXiv:2404.0596 1 (2024)

    BehnamGhader, Parishad, Vaibhav Adlakha, Marius Mosbach, D zmitry Bah- danau, Nicolas Chapados, and Siva Reddy. ”Llm2vec: Large languag e models are secretly powerful text encoders.” arXiv preprint arXiv:2404.0596 1 (2024)

  13. [13]

    https://cloud.tencent.com/developer/news/2461911 17 QZhou-Embedding Technical Report Kingsoft AI

  14. [14]

    ”Jaspe r and stella: distillation of sota embedding models.” arXiv preprint arXiv:2412.19048 (2024)

    Zhang, Dun, Jiacheng Li, Ziyang Zeng, and Fulong Wang. ”Jaspe r and stella: distillation of sota embedding models.” arXiv preprint arXiv:2412.19048 (2024)

  15. [15]

    ”Bge m3-embedding: Multi-lingual, multi-functionality, multi-gran ularity text embeddings through self-knowledge distillation.” arXiv preprint arXiv :2402.03216 (2024)

    Chen, Jianlv, Shitao Xiao, Peitian Zhang, Kun Luo, Defu Lian, and Zheng Liu. ”Bge m3-embedding: Multi-lingual, multi-functionality, multi-gran ularity text embeddings through self-knowledge distillation.” arXiv preprint arXiv :2402.03216 (2024)

  16. [16]

    ”Learning more effective repre senta- tions for dense retrieval through deliberate thinking before sear ch.” arXiv preprint arXiv:2502.12974 (2025)

    Ji, Yifan, Zhipeng Xu, Zhenghao Liu, Yukun Yan, Shi Yu, Yishan L i, Zhiyuan Liu, Yu Gu, Ge Yu, and Maosong Sun. ”Learning more effective repre senta- tions for dense retrieval through deliberate thinking before sear ch.” arXiv preprint arXiv:2502.12974 (2025)

  17. [17]

    LG-ANNA-Embedding technical repo rt[J]

    Choi J, Kim H, Jang H, et al. LG-ANNA-Embedding technical repo rt[J]. arXiv preprint arXiv:2506.07438, 2025

  18. [19]

    ”Nv-embed: Improved t echniques for training llms as generalist embedding models.” arXiv preprint arXiv:2405 .17428 (2024)

    Lee, Chankyu, Rajarshi Roy, Mengyao Xu, Jonathan Raiman, Mohammad Shoeybi, Bryan Catanzaro, and Wei Ping. ”Nv-embed: Improved t echniques for training llms as generalist embedding models.” arXiv preprint arXiv:2405 .17428 (2024)

  19. [21]

    ”Qwen2 technical report.” arXiv preprint arXiv:24 07.10671 (2024)

    Team, Qwen. ”Qwen2 technical report.” arXiv preprint arXiv:24 07.10671 (2024)

  20. [22]

    ”C-pack: Packed resources for general chinese embedd ings.” In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval, pp

    Xiao, Shitao, Zheng Liu, Peitian Zhang, Niklas Muennighoff, Defu L ian, and Jian- Yun Nie. ”C-pack: Packed resources for general chinese embedd ings.” In Proceedings of the 47th international ACM SIGIR conference on research and development in information retrieval, pp. 641-649. 2024. Team, Qwen

  21. [23]

    ”Mteb: Mas- sive text embedding benchmark.” arXiv preprint arXiv:2210.07316 (2 022)

    Muennighoff, Niklas, Nouamane Tazi, Lo ¨ ıc Magne, and Nils Reimers. ”Mteb: Mas- sive text embedding benchmark.” arXiv preprint arXiv:2210.07316 (2 022)

  22. [24]

    ”Conan-embed ding: Gen- eral text embedding with more and better negative samples.” arXiv p reprint arXiv:2408.15710 (2024)

    Li, Shiyu, Yang Tang, Shizhe Chen, and Xi Chen. ”Conan-embed ding: Gen- eral text embedding with more and better negative samples.” arXiv p reprint arXiv:2408.15710 (2024)

  23. [25]

    ”An information-theoretic perspective of tf–id f measures.” Infor- mation Processing & Management 39, no

    Aizawa, Akiko. ”An information-theoretic perspective of tf–id f measures.” Infor- mation Processing & Management 39, no. 1 (2003): 45-65

  24. [26]

    Robertson, Stephen E., and Steve Walker. ”Some simple effectiv e approximations to the 2-poisson model for probabilistic weighted retrieval.” In SIGI R’94: Proceed- ings of the Seventeenth Annual International ACM-SIGIR Confe rence on Research and Development in Information Retrieval, organised by Dublin City Un iversity, pp. 232-241. London: Springer Lond...

  25. [27]

    Dumais, George W

    Deerwester, Scott, Susan T. Dumais, George W. Furnas, Tho mas K. Landauer, and Richard Harshman. ”Indexing by latent semantic analysis.” Journal of the American society for information science 41, no. 6 (1990): 391-407

  26. [28]

    Improving text embeddings with large language models

    Liang Wang, Nan Yang, Xiaolong Huang, Linjun Yang, Rangan Maj umder, and Furu Wei. Improving text embeddings with large language models. arX iv preprint arXiv:2401.00368, 2023b

  27. [29]

    ”Sfrembedding-mistral: enhance text retrieval with tran sfer learning.” Sales- force AI Research Blog 3 (2024): 6

    Meng, Rui, Ye Liu, Shafiq Rayhan Joty, Caiming Xiong, Yingbo Zhou , and Semih Yavuz. ”Sfrembedding-mistral: enhance text retrieval with tran sfer learning.” Sales- force AI Research Blog 3 (2024): 6

  28. [30]

    Sfr-embedding-2: Advanced text embedding with multi-stage training, 2024[J]

    Meng R, Liu Y, Joty S R, et al. Sfr-embedding-2: Advanced text embedding with multi-stage training, 2024[J]

  29. [31]

    Muennighoff, Niklas, S. U. Hongjin, Liang Wang, Nan Yang, Furu W ei, Tao Yu, Amanpreet Singh, and Douwe Kiela. ”Generative representational instruction tun- ing.” In The Thirteenth International Conference on Learning Rep resentations. 2024

  30. [32]

    Making text embedders few-shot learner s

    Chaofan Li, MingHao Qin, Shitao Xiao, Jianlyu Chen, Kun Luo, Yingx ia Shao, Defu Lian, and Zheng Liu. Making text embedders few-shot learner s. arXiv preprint arXiv:2409.15700, 2024

  31. [33]

    Towards general text embeddings with multi-stage con trastive learning,

    Zehan Li, Xin Zhang, Yanzhao Zhang, Dingkun Long, Pengjun Xie , and Meis- han Zhang. Towards general text embeddings with multi-stage con trastive learning,

  32. [34]

    ”Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models.” arXiv preprint arXiv:2506.05176 (2025 )

    Zhang, Yanzhao, Mingxin Li, Dingkun Long, Xin Zhang, Huan Lin, B aosong Yang, Pengjun Xie et al. ”Qwen3 Embedding: Advancing Text Embedding and Reranking Through Foundation Models.” arXiv preprint arXiv:2506.05176 (2025 )

  33. [35]

    ”Roformer: Enhanced transformer with rotary position embeddin g.” Neurocomput- ing 568 (2024): 127063

    Su, Jianlin, Murtadha Ahmed, Yu Lu, Shengfeng Pan, Wen Bo, an d Yunfeng Liu. ”Roformer: Enhanced transformer with rotary position embeddin g.” Neurocomput- ing 568 (2024): 127063

  34. [36]

    ”Root mean square layer norma lization.” Ad- vances in neural information processing systems 32 (2019)

    Zhang, Biao, and Rico Sennrich. ”Root mean square layer norma lization.” Ad- vances in neural information processing systems 32 (2019)

  35. [37]

    ”Glu variants improve transformer.” arXiv pre print arXiv:2002.05202 (2020)

    Shazeer, Noam. ”Glu variants improve transformer.” arXiv pre print arXiv:2002.05202 (2020)

  36. [38]

    https://seed1-6-embedding.github.io/

  37. [39]

    Piccolo2: General Text Embedding with Multi-task Hybrid Loss Training

    Huang, Junqin, Zhongjie Hu, Zihao Jing, Mengya Gao, and Yichao Wu. ”Pic- colo2: General text embedding with multi-task hybrid loss training.” a rXiv preprint arXiv:2405.06932 (2024)

  38. [40]

    ”Circle loss: A unified perspective of pair similarit y op- timization.” In Proceedings of the IEEE/CVF conference on comput er vision and pattern recognition, pp

    Sun, Yifan, Changmao Cheng, Yuhan Zhang, Chi Zhang, Liang Z heng, Zhongdao Wang, and Yichen Wei. ”Circle loss: A unified perspective of pair similarit y op- timization.” In Proceedings of the IEEE/CVF conference on comput er vision and pattern recognition, pp. 6398-6407. 2020. 19 QZhou-Embedding Technical Report Kingsoft AI

  39. [41]

    Rodrigo Nogueira, Wei Yang, Jimmy Lin, and Kyunghyun Cho. 201 9. Document expansion by query prediction. ArXiv preprint, abs/1904.08375

  40. [42]

    Liang Wang, Nan Yang, and Furu Wei. 2023. Query2doc: Query e xpansion with large language models. In Proceedings of the 2023 Conference on E mpirical Meth- ods in Natural Language Processing, pages 9414–9423, Singapor e. Association for Computational Linguistics

  41. [43]

    Zhuyun Dai, Vincent Y Zhao, Ji Ma, Yi Luan, Jianmo Ni, Jing Lu, An ton Bakalov, Kelvin Guu, Keith Hall, and Ming-Wei Chang. 2022. Promptagator: Fe wshot dense retrieval from 8 examples. In The Eleventh International Confer ence on Learning Representations

  42. [44]

    Kexin Wang, Nandan Thakur, Nils Reimers, and Iryna Gurevych. 2022a. GPL: Generative pseudo labeling for unsupervised domain adaptation of d ense retrieval. In Proceedings of the 2022 Conference of the North American Cha pter of the Association for Computational Linguistics: Human Language Techn ologies, pages 2345–2360, Seattle, United States. Association...

  43. [45]

    ”Unn atural in- structions: Tuning language models with (almost) no human labor.” ar Xiv preprint arXiv:2212.09689 (2022)

    Honovich, Or, Thomas Scialom, Omer Levy, and Timo Schick. ”Unn atural in- structions: Tuning language models with (almost) no human labor.” ar Xiv preprint arXiv:2212.09689 (2022)

  44. [46]

    ”Approximate nearest neighbo r negative con- trastive learning for dense text retrieval.” arXiv preprint arXiv:200 7.00808 (2020)

    Xiong, Lee, Chenyan Xiong, Ye Li, Kwok-Fung Tang, Jialin Liu, Pau l Bennett, Junaid Ahmed, and Arnold Overwijk. ”Approximate nearest neighbo r negative con- trastive learning for dense text retrieval.” arXiv preprint arXiv:200 7.00808 (2020)

  45. [47]

    ”NV-Retriever: Improving text emb edding models with effective hard-negative mining.” arXiv preprint arXiv:2407.15831 (2024)

    Moreira, Gabriel de Souza P., Radek Osmulski, Mengyao Xu, Rona y Ak, Benedikt Schifferer, and Even Oldridge. ”NV-Retriever: Improving text emb edding models with effective hard-negative mining.” arXiv preprint arXiv:2407.15831 (2024)

  46. [48]

    Representatio n learning with contrastive predictive coding

    Aaron van den Oord, Yazhe Li, and Oriol Vinyals. Representatio n learning with contrastive predictive coding. arXiv preprint arXiv:1807.03748, 20 18

  47. [49]

    https://www.kexue.fm/archives/8847

  48. [50]

    mgte: Generalized long-context text representation and reranking models for multilingual text retrieval, 2024

    Xin Zhang, Yanzhao Zhang, Dingkun Long, Wen Xie, Ziqi Dai, Jialon g Tang, Huan Lin, Baosong Yang, Pengjun Xie, Fei Huang, Meishan Zhang, Wenjie Li, and Min Zhang. mgte: Generalized long-context text representation and reranking models for multilingual text retrieval, 2024

  49. [51]

    Cole, Kai Hui et al

    Lee, Jinhyuk, Zhuyun Dai, Xiaoqi Ren, Blair Chen, Daniel Cer, Je remy R. Cole, Kai Hui et al. ”Gecko: Versatile text embeddings distilled from large la nguage models, 2024.” URL https://arxiv. org/abs/2403.20327

  50. [52]

    Linq-embed-mistral: Elevat ing text re- trieval with improved gpt data through task-specific control and quality refinement

    Junseong Kim, Seolhwa Lee, Jihoon Kwon, Sangmo Gu, Yejin Kim, M inkyung Cho, Jy yong Sohn, and Chanyeol Choi. Linq-embed-mistral: Elevat ing text re- trieval with improved gpt data through task-specific control and quality refinement. linq ai research blog, 2024

  51. [53]

    https://huggingface.co/dunzhang/stella-large-zh-v3-1792d 20 QZhou-Embedding Technical Report Kingsoft AI

  52. [54]

    An overview of the BIOASQ large- scale biomedical semantic indexing and question answering competitio n[J]

    Tsatsaronis G, Balikas G, Malakasiotis P, et al. An overview of the BIOASQ large- scale biomedical semantic indexing and question answering competitio n[J]. BMC bioinformatics, 2015, 16(1): 138

  53. [55]

    A Span-Extraction Dataset for Chinese Machine Reading Comprehension

    Cui Y, Liu T, Che W, et al. A span-extraction dataset for Chines e machine reading comprehension[J]. arXiv preprint arXiv:1810.07366, 2018

  54. [56]

    GLUE: A multi-task benchmark a nd analysis platform for natural language understanding[J]

    Wang A, Singh A, Michael J, et al. GLUE: A multi-task benchmark a nd analysis platform for natural language understanding[J]. arXiv preprint ar Xiv:1804.07461, 2018

  55. [57]

    Yelp Inc., [Year]

    Yelp Dataset. Yelp Inc., [Year]. Available: https://www.yelp.com/dataset

  56. [58]

    Learning word vectors for sent iment analy- sis[C]//Proceedings of the 49th annual meeting of the association f or computational linguistics: Human language technologies

    Maas A, Daly R E, Pham P T, et al. Learning word vectors for sent iment analy- sis[C]//Proceedings of the 49th annual meeting of the association f or computational linguistics: Human language technologies. 2011: 142-150

  57. [59]

    Jack FitzGerald, Christopher Hench, Charith Peris, Scott Mac kie, Kay Rottmann, Ana Sanchez, Aaron Nash, Liam Urbach, Vishesh Kakarala, Richa Sin gh, Swetha Ranganath, Laurie Crist, Misha Britan, Wouter Leeuwis, Gokhan Tu r, and Prem Natarajan. 2022. Massive: A 1m-example multilingual natural langu age understand- ing dataset with 51 typologically-diverse...

  58. [60]

    Eneko Agirre, Daniel Cer, Mona Diab, and Aitor Gonzalez-Agirre . 2012. Semeval- 2012 task 6: A pilot on semantic textual similarity. In * SEM 2012: The First Joint Conference on Lexical and Computational Semantics–Volume 1: Proceedings of the main conference and the shared task, and Volume 2: Procee dings of the Sixth International Workshop on Semantic Ev...

  59. [61]

    ”Lcqmc: A large-scale chinese question matching corpus.” In Proceedings of the 27th international conference on computatio nal linguistics, pp

    Liu, Xin, Qingcai Chen, Chong Deng, Huajun Zeng, Jing Chen, Do ngfang Li, and Buzhou Tang. ”Lcqmc: A large-scale chinese question matching corpus.” In Proceedings of the 27th international conference on computatio nal linguistics, pp. 1952-1962. 2018

  60. [62]

    ”PA W S-X: A cross-lingual adversarial dataset for paraphrase identification .” arXiv preprint arXiv:1908.11828 (2019)

    Yang, Yinfei, Yuan Zhang, Chris Tar, and Jason Baldridge. ”PA W S-X: A cross-lingual adversarial dataset for paraphrase identification .” arXiv preprint arXiv:1908.11828 (2019)

  61. [63]

    ”Semeval-2017 task 1: Semantic textual similarity-multilingual and c ross-lingual focused evaluation.” arXiv preprint arXiv:1708.00055 (2017)

    Cer, Daniel, Mona Diab, Eneko Agirre, Inigo Lopez-Gazpio, and L ucia Specia. ”Semeval-2017 task 1: Semantic textual similarity-multilingual and c ross-lingual focused evaluation.” arXiv preprint arXiv:1708.00055 (2017)

  62. [64]

    Tri Nguyen, Mir Rosenberg, Xia Song, Jianfeng Gao, Saurabh T iwary, Rangan Majumder, and Li Deng. 2016. MS MARCO: A human generated mach ine read- ing comprehension dataset. In Proceedings of the Workshop on Co gnitive Com- putation: Integrating neural and symbolic approaches 2016 co-lo cated with the 30th Annual Conference on Neural Information Processin...

  63. [65]

    Natural questions: a benchmark for question answering res earch

    Tom Kwiatkowski, Jennimaria Palomaki, Olivia Redfield, Michael Collins , Ankur Parikh, Chris Alberti, Danielle Epstein, Illia Polosukhin, Jacob Devlin, Ke nton Lee, et al. Natural questions: a benchmark for question answering res earch. Transactions of the Association for Computational Linguistics, 7:453–466, 2019

  64. [66]

    Angela Fan, Yacine Jernite, Ethan Perez, David Grangier, Jaso n Weston, and Michael Auli. 2019. ELI5: Long Form Question Answering. In Procee dings of the 57th Annual Meeting of the Association for Computational Ling uistics, pages 3558–3567, Florence, Italy. Association for Computational Lingu istics

  65. [67]

    Zhilin Yang, Peng Qi, Saizheng Zhang, Yoshua Bengio, William Cohen, Ruslan Salakhutdinov, and Christopher D. Manning. HotpotQA: A dataset for diverse, explainable multi-hop question answering. In Proceedings of the 201 8 Conference on Empirical Methods in Natural Language Processing, pp. 2369–2 380, Brussels, Belgium, October-November 2018. Association fo...

  66. [68]

    Miracl: A multilingual retrieval dataset covering 18 diverse language s

    Xinyu Zhang, Nandan Thakur, Odunayo Ogundepo, Ehsan Kama lloo, David Alfonso-Hermelo, Xiaoguang Li, Qun Liu, Mehdi Rezagholizadeh, and Jimmy Lin. Miracl: A multilingual retrieval dataset covering 18 diverse language s. Transactions of the Association for Computational Linguistics, 11:1114–1131, 2 023

  67. [69]

    Squad: 100,000+ questions for machine comprehension of text

    Pranav Rajpurkar, Jian Zhang, Konstantin Lopyrev, and Per cy Liang. Squad: 100,000+ questions for machine comprehension of text. ar Xiv preprint arXiv:1606.05250, 2016

  68. [70]

    Fever: a large-scale dataset for fact extraction and verifica tion

    James Thorne, Andreas Vlachos, Christos Christodoulopoulos , and Arpit Mit- tal. Fever: a large-scale dataset for fact extraction and verifica tion. arXiv preprint arXiv:1803.05355, 2018

  69. [71]

    Wei He, Kai Liu, Jing Liu, Yajuan Lyu, Shiqi Zhao, Xinyan Xiao, Yu an Liu, Yizhong Wang, Hua Wu, Qiaoqiao She, Xuan Liu, Tian Wu, and Haifeng Wa ng

  70. [72]

    Yichen Jiang, Shikha Bordia, Zheng Zhong, Charles Dognin, Mane esh Singh, and Mohit Bansal. 2020. HoVer: A Dataset for Many-Hop Fact Extract ion And Claim Verification. In Findings of the Association for Computational Lingu istics: EMNLP 2020, pages 3441–3460, Online. Association for Computational Lin guistics

  71. [73]

    Zhang X, Ma X, Shi P, et al. Mr. TyDi: A multi-lingual benchmark fo r dense retrieval[J]. arXiv preprint arXiv:2108.08787, 2021

  72. [74]

    Kyle Lo, Lucy Lu Wang, Mark Neumann, Rodney Kinney, and Danie l Weld. 2020. S2ORC: The Semantic Scholar Open Research Corpus. In Proceedin gs of the 58th Annual Meeting of the Association for Computational Linguistics, p ages 4969–4983, Online. Association for Computational Linguistics. 22 QZhou-Embedding Technical Report Kingsoft AI

  73. [75]

    https://huggingface.co/spaces/mteb/leaderboard

  74. [76]

    Gemini embedding: Generalizable embeddings from gemini

    Jinhyuk Lee, Feiyang Chen, Sahil Dua, Daniel Cer, Madhuri Sha nbhogue, Iftekhar Naim, Gustavo Hernandez /acute.ts1Abrego, Zhe Li, Kaifeng Chen, Henrique Schechter Vera, et al. Gemini embedding: Generalizable embeddings from gemini. arXiv preprint arXiv:2503.07891, 2025b. A Appendix A.1 Framework Constraints Table 4: Specifications of framework constraints ...

  75. [79]

    incorporation of irrelevant information, or 3) maintaining the same topic but diverging in as- pects. Imitation: syntax, sentence structure, structural Generating hard negative examples by emulating the structural and syntactic patterns of the given positive instance is a critical step to maximize dis- criminative challenge for the model. 23 QZhou-Embeddi...

  76. [1620]

    England had a rich literary heritage in its native tongue, which expanded as the printing press became more widely used in the mid-16th century. By the Elizabethan era, a thriving lit- erary culture in both drama and poetry emerged, with poets like Edmund Spenser, whose epic The Faerie Queene profoundly impacted English liter- ature, though it was eventua...

  77. [2018]

    In Proceedings of the Workshop on Machine Read ing for Ques- tion Answering, pages 37–46, Melbourne, Australia

    DuReader: a Chinese Machine Reading Comprehension Datase t from Real- world Applications. In Proceedings of the Workshop on Machine Read ing for Ques- tion Answering, pages 37–46, Melbourne, Australia. Association fo r Computational Linguistics

  78. [2023]

    URL https://arxiv.org/abs/2308.03281

This paper was first reviewed by deepseek-v4-flash on August 5, 2026.