Pith. sign in

REVIEW 3 major objections 5 minor 39 references

inversedMixup: Data Augmentation via Inverting Mixed Embeddings

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read This paper claims that interpolating two BERT sentence embeddings and decoding the result with a large language model yields readable, label-consistent training sentences, and that this decoding exposes a widespread 'manifold intrusion' pro

desk verdict Promising Mixup-inversion augmentation with an untested controllability claim and circular intrusion evidence. read the letter →

arxiv 2601.21543 v3 pith:HBGGWGIQ submitted 2026-01-29 cs.CL

Fanshuang Kong , Richong Zhang , Qiyu Sun , Zhijie Nie , Ting Deng , Chunming Hu This is my paper
classification cs.CL
keywords MixupdataaugmentationLLMinversiontextclassificationmanifoldintrusionembeddingalignmentfew-shotlearningBERT
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

The paper tries to establish that the two rival approaches to text data augmentation—embedding-level Mixup and token-level LLM generation—can be unified: interpolate sentence embeddings from a task model, then decode the mixed vector with an LLM into a human-readable sentence. To make that work, it introduces a three-stage training procedure that aligns BERT's output embedding space with LLaMA's input space through a lightweight adaptor. The paper also claims that the decoded sentences provide the first direct empirical evidence of 'manifold intrusion' in text Mixup, where the interpolated input does not semantically match the interpolated label. A sympathetic reader would care because this makes latent-space Mixup inspectable for the first time, and offers a practical fix: let the LLM assign hard labels to reconstructed sentences, which the paper shows improves downstream accuracy.

What carries the argument

The load-bearing component is the adaptor, a single-layer MLP that maps the task model's 768-dimensional output embeddings into the LLM's 4096-dimensional input embedding space. It is trained with a generation loss: the LLM must reconstruct the original sentence from the adapted embedding. After alignment, the Mixup operation happens in BERT space, and the mixed vector is pushed through the adaptor and decoded by the frozen LLM via a prompt that treats the adapted embedding as a soft token. This machinery converts an uninterpretable latent interpolation into an observable sentence, which is what enables the paper's manifold-intrusion analysis.

What would settle it

Decode a mixed embedding with λ=0.5 from two clearly unrelated sentences from different classes and have human judges decide whether the output meaningfully combines both; if a large fraction of outputs are nonsense or favor one class overwhelmingly, the inversion claim collapses. Additionally, swap the trained adaptor for a simple linear regression fitted on paired BERT and LLaMA embeddings from the same unlabeled corpus; if the manifold-intrusion rates change dramatically, the phenomenon is an artifact of the specific adaptor rather than a property of text Mixup.

Watch

Extended reading notes

Core claim

The central claim is that a mixed embedding, formed as a convex combination of two task-model sentence embeddings, can be inverted into a fluent sentence that combines the semantics of both sources, under a controllable mixing ratio. The inversion is achieved by an adaptor that maps the task model's output embeddings into the LLM's input space, trained first on unlabeled open-domain text and then refined on labeled task data. The paper further claims that inverting many such mixed embeddings reveals that text Mixup routinely produces sentences whose true label lies outside the interpolated label distribution—'inside' one of the original classes or 'outside' both—and that replacing soft inter

Load-bearing premise

The pipeline assumes that a single-layer adaptor trained on open-domain text can map BERT's output embeddings into LLaMA's input space faithfully enough that linear interpolations between BERT embeddings decode into sentences that actually blend the original meanings; if that mapping fails, the decoded sentences are arbitrary and the claimed manifold-intrusion measurements are artifacts of the inverter rather than of Mixup itself.

Editorial extensions

If this is right

  • Text Mixup becomes auditable: practitioners can decode interpolated embeddings into sentences and see exactly what the mix produced, instead of trusting soft labels blindly.
  • Manifold intrusion in text augmentation is confirmed as a real, widespread phenomenon that varies with mixing ratio: intrusion is more frequent near λ=0.5 and decreases as one endpoint dominates.
  • A simple mitigation—replacing interpolated soft labels with LLM-assigned hard labels for each reconstructed sentence—consistently improves accuracy across all tested interpolation ratios.
  • inversedMixup outperforms both embedding-level Mixup and prompt-based LLM augmentation in most few-shot settings across Yahoo, TREC, and AG News, with the largest gains on the dataset used during adaptor alignment.

Reading between the lines

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

  • If the alignment is genuinely general, the same invert-and-inspect pipeline could be used to audit other embedding-space manipulations, such as adversarial perturbations, style-transfer interpolation, or latent-space editing, turning any continuous operation on BERT vectors into readable text for human review.
  • The method's ceiling is set by the decoder: as LLMs grow more capable, the fidelity of inverted sentences should improve, potentially making the adaptor a smaller bottleneck and shifting the skill to generation-loss tuning.
  • The 'Inside/Outside' typology could be formalized into a quantitative diagnostic for any interpolation-based augmentation scheme, independent of the specific inverter used, if a reliable automatic judge is available.
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. inversedMixup proposes a three-stage text augmentation pipeline: align a BERT-base encoder to LLaMA-3-8B-Instruct's input space via a single-layer MLP adaptor trained on open-domain unlabeled data; refine the alignment on labeled task data; then interpolate task-model embeddings with λ ~ Beta(α,α), decode the mixed embedding through the adaptor and LLM into a human-readable sentence, and train the task model on the reconstructed sentences with either interpolated soft labels or LLM-assigned hard labels. The paper claims that this combines Mixup's controllability with LLM interpretability, reports accuracy gains in few-shot and fully supervised settings on Yahoo, TREC, and AG News, and uses the decoded sentences to provide what it calls the first empirical evidence of manifold intrusion in text Mixup, together with a hard-label mitigation strategy.

Significance. If the controllability claim holds, inversedMixup is a useful bridge between embedding-level Mixup and token-level LLM augmentation, and the decoded outputs offer a genuinely new way to study manifold intrusion in text. The paper has real experimental strengths: 10 random seeds, Wilcoxon signed-rank tests, two LLM backends, several datasets, and both few-shot and fully supervised regimes. These are meaningful strengths. However, the headline property of controllable mixing is never directly measured, the statistical evidence on unseen datasets is weaker than the text suggests, and the intrusion findings rest entirely on LLM-as-a-judge judgments without human validation. The core idea is defensible, but the load-bearing evidence is incomplete.

major comments (3)
  1. [§3.4, Eqs. (8)-(10)] The central claim of 'a controllable mixing ratio' is asserted but never measured. No experiment varies λ and measures the semantic content of the decoded sentence. Section 4.7 gives only qualitative examples at λ=0.5 and 0.9; Figure 7 aggregates intrusion rates but does not report the probability that the decoded label tracks the λ-majority source; Section 4.9 fixes λ only to compare soft vs. hard label training. Since the adaptor is a single-layer MLP and the decoder is a frozen LLM, it is not self-evident that interpolation structure survives decoding. Add a λ-response evaluation (e.g., human/independent-judge label distribution or embedding similarity to endpoints as a function of λ) to support the headline claim.
  2. [Table 2, §4.3] The statistical support for improvement over Mixup is strong only on Yahoo, which was used in the alignment and refinement stages. On the unseen datasets, the Wilcoxon p-values vs. Mixup are 0.0625 (TREC K=5), 0.246 (TREC K=10), 0.161 (AG News K=1), and 0.278 (AG News K=10); only TREC K=1 and AG News K=5 are below 0.05. The fully supervised AG News result (89.7 vs. 90.0 for Base) also does not support the claimed generalizability. Either provide more evidence (more seeds/datasets) or temper the conclusion to 'competitive on unseen datasets.'
  3. [§4.8, A.4, Figure 7] The 'first empirical evidence' of manifold intrusion is based entirely on an LLM-as-a-judge evaluation of 100 TREC sentences, with no human agreement, no confidence intervals, and no statement of which model generated the judged sentences. If the judged sentences come from the Table 3 ChatGPT-4o variant, the generator and judge are the same model, making the assessment circular; even for LLaMA-3-generated sentences, LLM judge bias is a known risk. Report human agreement on a subset and specify the generator/judge split explicitly.
minor comments (5)
  1. [§4.1.2] The Beta(α,α) concentration α, the number of alignment/refinement epochs, and the number of generated augmented samples per setting are not reported. These are necessary for reproducibility, especially since α controls the interpolation distribution.
  2. [§3.5] Cross-reference errors: §3.5 says 'Figure 7 illustrates the Inside/Outside intrusion issue' but Figure 7 is the quantitative result; the definition figure is Figure 3. §1 also references 'Figure 6' for examples, but Figure 6 appears only in §4.7.
  3. [Table 2] The table header '10−2' is easy to misread; state explicitly in the caption that reported values are p-values multiplied by 10², or show actual p-values. Also, 18 comparisons at α=0.05 are made without multiple-comparison correction; this should be acknowledged.
  4. [§4.5, Table 3] The ChatGPT-4o variant shows substantial drops on TREC K=1 (41.6→38.6) and AG News K=1 (77.7→69.4) relative to LLaMA-3. The text says inversedMixup 'consistently outperforms'; this sensitivity should be acknowledged.
  5. [A.2, §4.1] Minor typos: 'interplates' in A.2, 'Exprimental' in §4.1, 'Corresbonding' in the footnote, and 'representes' in the Figure 5 caption.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: the core augmentation pipeline is evaluated on held-out test sets; self-citations are background, and the LLM-as-judge concern is a validation caveat rather than a formal circular reduction.

full rationale

The central claim—that inversedMixup reconstructs mixed embeddings into controllable, human-interpretable sentences and improves task-model accuracy—is implemented through a concrete pipeline (Eqs. 1–12) and evaluated on external held-out test sets (Table 1, Yahoo/TREC/AG News). No parameter is fitted directly to the test metric; the adaptor is trained with generation and classification losses, and the final accuracy is measured against unseen data. The self-citations in the reference list ([3], [11]) are used only as baseline descriptions and background, not as load-bearing justifications of the method's validity; the LLM-inversion premise relies on external prior work ([4,15,24,34]). The manifold-intrusion analysis does rely on LLM-as-a-judge, so the 'first empirical evidence' is partly self-confirmatory: the LLaMA-3 generator and the ChatGPT-4o judge are different models, and the downstream augmentation results are objective, so this is a potential confound rather than a reduction-by-construction. The skeptical criticism about controllability (no λ-response measurement) is a missing-support gap, not circularity. Overall, the derivation is self-contained and externally benchmarked, with no step where a claimed prediction is equivalent to its input by definition.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The method rests on three empirical domain assumptions (adaptor alignment transferability, interpolation-decodability of mixed embeddings, and LLM-as-judge validity) and several unreported hand-chosen hyperparameters; no new physical or conceptual entities are introduced, and no formal proof is supplied.

free parameters (4)
  • Mixup Beta(alpha, alpha) concentration = not reported
    Controls the mixing ratio distribution in Eq (9); chosen by hand, not stated in the paper, and directly determines the amount of interpolation in generated samples.
  • Number of refinement/alignment epochs and steps = not reported
    The three-stage training requires iterative fine-tuning (Eq 6-7) but the exact steps/epochs, batch sizes, and amount of unlabeled data are not specified; these choices affect final alignment quality.
  • Number of generated augmented samples per setting = not reported (varied in Figure 5)
    The method can generate 'potentially unlimited' samples; the count used in Table 1 is not given, and sample count strongly affects downstream accuracy.
  • Fixed interpolation ratios in Section 4.9 = lambda values in Figure 8 (not enumerated)
    The intrusion-alleviation study fixes lambda rather than sampling from Beta, and the exact values are left to the figure; this hand-choice influences the conclusion that hard labels help.
assumptions (5)
  • domain assumption A single-layer MLP adaptor can map BERT's 768-d sentence embeddings into LLaMA-3-8B's 4096-d input embedding space well enough that the LLM reconstructs the original sentence (Eq 2-3).
    Entire pipeline depends on this geometric alignment; no guarantee, purely empirical.
  • domain assumption LLM inversion of an interpolated embedding yields a sentence whose semantics are a meaningful blend of the two source sentences (Section 3.4, Eq 10).
    Load-bearing for the augmentation claim; prior work shows reconstruction of original sentences, not of arbitrary convex combinations.
  • domain assumption ChatGPT-4o-as-judge correctly classifies generated sentences and identifies 'Inside'/'Outside' intrusion without human validation (Section 4.8, A.4).
    The manifold-intrusion evidence and the hard-label mitigation both rest on this unvalidated judge.
  • standard math Standard statistics: Wilcoxon signed-rank test and accuracy evaluation (Section 4.3).
    Assumed background; no issue.
  • domain assumption Mixup interpolation in linear embedding space preserves label semantics (vanilla Mixup assumption; Eq 9-11).
    The paper itself challenges this with the intrusion evidence; the method initially relies on soft labels.

how reviews work

0 comments
Cite this review

Pith. "Pith review of inversedMixup: Data Augmentation via Inverting Mixed Embeddings." pith.science (2026). https://pith.science/paper/HBGGWGIQ

@misc{pith2026260121543,
  author       = {Pith},
  title        = {Pith review of: inversedMixup: Data Augmentation via Inverting Mixed Embeddings},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/HBGGWGIQ}},
  note         = {Machine review of arXiv:2601.21543}
}
read the original abstract

Mixup generates augmented samples by linearly interpolating inputs and labels with a controllable ratio. However, since it operates at the latent embedding level, the resulting samples are not human-interpretable. In contrast, LLM-based augmentation methods produce sentences via prompts at the token level, yielding readable outputs but offering limited control over the generation process. Inspired by recent advances in LLM inversion, which reconstructs natural language from embeddings and helps bridge the gap between latent embedding space and discrete token space, we propose inversedMixup, a unified framework that combines the controllability of Mixup with the interpretability of LLM-based generation. Specifically, inversedMixup aligns the output embedding space of a task-specific model with the input embedding space of an LLM, so that mixed embeddings can be reconstructed, under a controllable mixing ratio, into human-interpretable sentences. This interpretability provides the first empirical evidence of the manifold intrusion phenomenon in text Mixup. Building on this, we extend inversedMixup into a three-stage data augmentation method, and introduce a simple yet effective strategy to mitigate manifold intrusion during augmentation. Extensive experiments demonstrate the effectiveness and generalizability of our approach in both few-shot and fully supervised scenarios.

Figures

Figures reproduced from arXiv: 2601.21543 by the authors.

Figure 1
Figure 1. Illustration of inversedMixup. Mixed embedding [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed inversedMixup framework. (a) Adaptor alignment with unlabeled data aligns the task [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. ‘Inside’ refers to cases where the interpolated sam [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (3 more)
Figure 5
Figure 5. Figure 5: Comparison of different LLM-based augmentation [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Case study illustrating the manifold intrusion phenomenon under different interpolation ratios ( [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 8
Figure 8. Figure 8: Comparison between soft labels and hard labels for [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 1 canonical work pages

  1. [1]

    Nicholas Carlini, Florian Tramer, Eric Wallace, Matthew Jagielski, Ariel Herbert- Voss, Katherine Lee, Adam Roberts, Tom Brown, Dawn Song, Ulfar Erlingsson, et al. 2021. Extracting training data from large language models. In 30th USENIX security symposium (USENIX Security 21) . 2633–2650

    2021

  2. [2]

    Jiaao Chen, Zichao Yang, and Diyi Yang. 2020. MixText: Linguistically-Informed Interpolation of Hidden Space for Semi-Supervised Text Classification. InProceed- ings of the 58th Annual Meeting of the Association for Computational Linguistics . 2147–2157. doi:10.18653/v1/2020.acl-main.194

    doi:10.18653/v1/2020.acl-main.194 2020

  3. [3]

    Junfan Chen, Richong Zhang, Zheyan Luo, Chunming Hu, and Yongyi Mao. 2023. Adversarial word dilution as text data augmentation in low-resource regime. In Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 37. 12626–12634

    2023

  4. [4]

    Yiyi Chen, Heather Lent, and Johannes Bjerva. 2024. Text embedding inversion security for multilingual language models.arXiv preprint arXiv:2401.12192 (2024)

    arXiv 2024

  5. [5]

    Wanyun Cui and Qianle Wang. 2024. Ada-Instruct: Adapting Instruction Gener- ators for Complex Reasoning. In Findings of the Association for Computational Linguistics: EMNLP 2024, Yaser Al-Onaizan, Mohit Bansal, and Yun-Nung Chen (Eds.). Association for Computational Linguistics, Miami, Florida, USA, 6967–

    2024

  6. [6]

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers). 4171–4186

    2019

  7. [7]

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al-Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Amy Yang, Angela Fan, et al. 2024. The llama 3 herd of models. arXiv e-prints (2024), arXiv–2407

    2024

  8. [8]

    Jiahui Gao, Renjie Pi, LIN Yong, Hang Xu, Jiacheng Ye, Zhiyong Wu, WEIZHONG ZHANG, Xiaodan Liang, Zhenguo Li, and Lingpeng Kong. 2023. Self-Guided Noise-Free Data Generation for Efficient Zero-Shot Learning. In The Eleventh International Conference on Learning Representations . https://openreview.net/ forum?id=h5OpjGd_lo6

    2023

Show all 39 references
  1. [9]

    Jiawei Gu, Xuhui Jiang, Zhichao Shi, Hexiang Tan, Xuehao Zhai, Chengjin Xu, Wei Li, Yinghan Shen, Shengjie Ma, Honghao Liu, et al . 2024. A survey on llm-as-a-judge. arXiv preprint arXiv:2411.15594 (2024)

    2024 arXiv

  2. [10]

    Hongyu Guo. 2020. Nonlinear mixup: Out-of-manifold data augmentation for text classification. In Proceedings of the AAAI Conference on Artificial Intelligence , Vol. 34. 4044–4051

    2020

  3. [11]

    Hongyu Guo, Yongyi Mao, and Richong Zhang. 2019. Augmenting data with mixup for sentence classification: An empirical study. arXiv preprint arXiv:1905.08941 (2019)

    2019 arXiv

  4. [12]

    Sungnyun Kim, Gihun Lee, Sangmin Bae, and Se-Young Yun. 2020. Mixco: Mix-up contrastive learning for visual representation. arXiv preprint arXiv:2010.06300 (2020)

    2020 arXiv

  5. [13]

    Diederik P Kingma and Jimmy Ba. 2014. Adam: A method for stochastic opti- mization. arXiv preprint arXiv:1412.6980 (2014)

    2014 arXiv

  6. [14]

    Jens Lehmann, Robert Isele, and et al. 2015. DBpedia – A large-scale, multilingual knowledge base extracted from Wikipedia. Semantic Web (Jan 2015), 167–195. doi:10.3233/sw-140134

    2015 doi

  7. [15]

    Haoran Li, Mingshi Xu, and Yangqiu Song. 2023. Sentence embedding leaks more information than you expect: Generative embedding inversion attack to recover the whole sentence. arXiv preprint arXiv:2305.03010 (2023)

    2023 arXiv

  8. [16]

    Mengyu Li, Yonghao Liu, Fausto Giunchiglia, Xiaoyue Feng, and Renchu Guan

  9. [17]

    Xin Li and Dan Roth. 2002. Learning question classifiers. In COLING 2002: The 19th International Conference on Computational Linguistics

    2002

  10. [18]

    Zhuoyan Li, Hangxiao Zhu, Zhuoran Lu, and Ming Yin. 2023. Synthetic Data Generation with Large Language Models for Text Classification: Poten- tial and Limitations. In Proceedings of the 2023 Conference on Empirical Meth- ods in Natural Language Processing , Houda Bouamor, Jua...

    2023 doi

  11. [19]

    Julian McAuley and Jure Leskovec. 2013. Hidden factors and hidden topics: understanding rating dimensions with review text. InProceedings of the 7th ACM conference on Recommender systems . 165–172

    2013

  12. [20]

    John X Morris, Wenting Zhao, Justin T Chiu, Vitaly Shmatikov, and Alexander M Rush. 2023. Language model inversion. arXiv preprint arXiv:2311.13647 (2023)

    2023 arXiv

  13. [21]

    Tianyu Pang, Kun Xu, and Jun Zhu. 2019. Mixup inference: Better exploiting mixup to defend adversarial attacks. arXiv preprint arXiv:1909.11515 (2019)

    2019 arXiv

  14. [22]

    Rico Sennrich, Barry Haddow, and Alexandra Birch. 2015. Improving neural ma- chine translation models with monolingual data. arXiv preprint arXiv:1511.06709 (2015)

    2015 arXiv

  15. [23]

    Connor Shorten and Taghi M Khoshgoftaar. 2019. A survey on image data augmentation for deep learning. Journal of big data 6, 1 (2019), 1–48

    2019

  16. [24]

    Guy Tennenholtz, Yinlam Chow, Chih-Wei Hsu, Jihwan Jeong, Lior Shani, Azamat Tulepbergenov, Deepak Ramachandran, Martin Mladenov, and Craig Boutilier

  17. [25]

    Vikas Verma, Alex Lamb, Christopher Beckham, Amir Najafi, Ioannis Mitliagkas, David Lopez-Paz, and Yoshua Bengio. 2019. Manifold mixup: Better represen- tations by interpolating hidden states. In International conference on machine learning. PMLR, 6438–6447

    2019

  18. [26]

    Jason Wei and Kai Zou. 2019. EDA: Easy data augmentation techniques for boosting performance on text classification tasks. arXiv preprint arXiv:1901.11196 (2019)

    2019 arXiv

  19. [27]

    Robert F Woolson. 2007. Wilcoxon signed-rank test.Wiley encyclopedia of clinical trials (2007), 1–3

    2007

  20. [28]

    Xing Wu, Chaochen Gao, Meng Lin, Liangjun Zang, Zhongyuan Wang, and Songlin Hu. 2022. Text smoothing: Enhance various data augmentation methods on text classification tasks. arXiv preprint arXiv:2202.13840 (2022)

    2022 arXiv

  21. [29]

    Yuan Wu, Diana Inkpen, and Ahmed El-Roby. 2020. Dual mixup regularized learning for adversarial domain adaptation. In Computer Vision–ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXIX 16 . Springer, 540–555

    2020

  22. [30]

    Minghao Xu, Jian Zhang, Bingbing Ni, Teng Li, Chengjie Wang, Qi Tian, and Wenjun Zhang. 2020. Adversarial domain adaptation with domain mixup. In Proceedings of the AAAI conference on artificial intelligence , Vol. 34. 6502–6509

    2020

  23. [31]

    Rongkang Xu, Yongcheng Zhang, Kai Ren, Yu Huang, and Xiaomei Wei. 2024. Probabilistic interpolation with mixup data augmentation for text classification. In International Conference on Intelligent Computing . Springer, 410–421

    2024

  24. [32]

    Yue Yu, Yuchen Zhuang, Jieyu Zhang, Yu Meng, Alexander Ratner, Ranjay Kr- ishna, Jiaming Shen, and Chao Zhang. 2023. Large Language Model as Attributed Training Data Generator: A Tale of Diversity and Bias. In Thirty-seventh Confer- ence on Neural Information Processing System...

    2023

  25. [33]

    Hongyi Zhang, Moustapha Cisse, Yann N Dauphin, and David Lopez-Paz. 2017. mixup: Beyond empirical risk minimization. arXiv preprint arXiv:1710.09412 (2017)

    2017 arXiv

  26. [34]

    Xingjian Zhang, Ziyang Xiong, Shixuan Liu, Yutong Xie, Tolga Ergen, Dongsub Shim, Hua Xu, Honglak Lee, and Qiaozhu Me. 2024. Map2Text: New Content Generation from Low-Dimensional Visualizations. arXiv e-prints (2024), arXiv– 2412

    2024

  27. [35]

    Xiang Zhang, Junbo Zhao, and Yann LeCun. 2015. Character-level convolutional networks for text classification.Advances in neural information processing systems 28 (2015)

    2015

  28. [36]

    Description

    Shengyao Zhuang, Bevan Koopman, Xiaoran Chu, and Guido Zuccon. 2024. Understanding and mitigating the threat of vec2text to dense retrieval systems. In Proceedings of the 2024 Annual International ACM SIGIR Conference on Research and Development in Information Retrieval in the...

    2024

  29. [2023]

    arXiv preprint arXiv:2310.04475 (2023)

    Demystifying embedding spaces using large language models. arXiv preprint arXiv:2310.04475 (2023)

    2023 arXiv

  30. [2024]

    arXiv preprint arXiv:2405.11524 (2024)

    Simple-sampling and hard-mixup with prototypes to rebalance contrastive learning for text classification. arXiv preprint arXiv:2405.11524 (2024)

    2024

  31. [6984]

    doi:10.18653/v1/2024.findings-emnlp.409

    2024 doi

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.