Pith. sign in

REVIEW 3 major objections 2 minor 35 references

The First Differentiable Transfer-Based Algorithm for Discrete MicroLED Repair

T0 review · 3 major / 2 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read A differentiable transfer planner is claimed to cut microLED repair steps by 50 percent.

desk verdict The abstract promises a microLED repair algorithm, but the full text is an unrelated paper on jailbreak detection; as submitted, the central claim is entirely unsupported. read the letter →

arxiv 2508.09206 v1 pith:D3X66GG7 submitted 2025-08-10 cs.LG physics.comp-ph

Ning-Yuan Lue This is my paper
classification cs.LGphysics.comp-ph
keywords microLEDrepairdifferentiabletransfermodulegradient-basedoptimizationstepminimizationXYstageplanningdisplayfabricationlarge-array
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 submitted abstract announces the first repair algorithm based on a differentiable transfer module for discrete microLED transfer, trainable via gradient-based optimization, and claims a 50% reduction in transfer steps with sub-2-minute planning on 2000x2000 arrays. A sympathetic reader would care because microLED transfer planning is a throughput bottleneck in display fabrication, and a gradient-trainable planner would replace hand-crafted proximity searches and avoid RL's feature-engineering overhead. However, the full text of this submission is a different paper, on unsupervised jailbreak detection in large vision-language models, and nowhere mentions microLED, transfer platforms, or repair planning. The microLED claims therefore rest on the abstract alone, with no derivation, baseline, or dataset in the body to support them.

What carries the argument

The central object is the differentiable transfer module: a computational model of the discrete shifts of transfer platforms, designed so that the repair objective is a differentiable function of the shift sequence and thus trainable by gradient-based optimization. In the abstract this module is what replaces local proximity searching and what allows flexible objectives such as step-count minimization; it is the mechanism that the claimed 50% step reduction and sub-2-minute planning hinge on. The full text provides no such module.

What would settle it

Search the submitted full text for the terms 'microLED', 'transfer platform', or 'repair'; none appears, and no algorithm or experiment on transfer planning exists in the manuscript. If a differentiable transfer module were actually described, it would appear here; its total absence settles that the abstract's central claim is unsupported by this document.

Watch

Extended reading notes

Core claim

The author's stated discovery is a differentiable transfer module that models discrete shifts of transfer platforms and can be trained end-to-end with gradients, making the repair-planning objective—such as minimizing the number of transfer steps—directly optimizable. The abstract reports a 50% step reduction over local proximity searching and sub-2-minute planning on 2000x2000 arrays, and contrasts it with RL-based planners that require handcrafted feature extractors. On its own terms, this would be the first differentiable, gradient-trainable planner for discrete microLED repair. Inspection of the submitted full text, however, shows a wholly different paper—the LoD framework for detecting

Load-bearing premise

The load-bearing premise is that the submitted full text corresponds to the abstract's microLED repair work; the body is instead an unrelated paper on jailbreak detection in large vision-language models, so the premise fails.

Editorial extensions

If this is right

  • End-to-end training of repair objectives would let fabricators minimize step count directly, replacing hand-tuned proximity-search heuristics.
  • A 50% step reduction would roughly halve XY-stage motion time, a major contributor to microLED transfer cost.
  • Sub-2-minute planning on 2000x2000 arrays would make large-array repair practical in production, where current planners do not scale.
  • Because the module is gradient-trained, it could adapt to new objective changes without retraining feature extractors, unlike RL-based planners.

Reading between the lines

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

  • If a differentiable transfer module of the kind described existed, the same discrete-shift-plus-gradients recipe could be reused for other pick-and-place manufacturing tasks, such as chip placement or mask alignment, where step-count minimization matters.
  • The absence of any microLED content in the body suggests the submission is a metadata or file mix-up; the claim should not be attributed to the LoD results, and a reader should track down the actual microLED manuscript before citing these numbers.
  • A natural next check is to see whether the abstract's 50%-step and sub-2-minute numbers appear in any other release by the same author; if they never do, the figures are unverifiable from the current record.
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 / 2 minor

Summary. The manuscript presents an abstract claiming the first differentiable transfer-based algorithm for discrete microLED repair, with a 50% reduction in transfer steps and sub-2-minute planning on 2000x2000 arrays. However, the full text is an entirely different paper on jailbreak detection in large vision-language models (LVLMs), titled 'Learning to Detect Unseen Jailbreak Attacks in Large Vision-Language Models' and carrying the arXiv identifier 2508.09201v4 [cs.CR]. The body contains no mention of microLED, transfer platforms, XY stages, discrete shifts, or any of the experiments claimed in the abstract. As a result, the abstract's central contribution is absent from the manuscript, and there is no derivation, methodology, baseline, dataset, or result to evaluate for the microLED claim.

Significance. If the abstract's claims were supported, a first differentiable, gradient-trainable planner for discrete microLED transfer with demonstrated large-scale speed would be a notable contribution to display manufacturing. However, the submitted manuscript does not present this work at all. The full text is a self-contained LVLM jailbreak detection paper with its own abstract, equations, and experiments, none of which bear on the microLED problem. The significance of the claimed microLED result therefore cannot be assessed; the paper as submitted does not establish any result in that domain.

major comments (3)
  1. [Title/Header vs. Abstract] The full text is an unrelated paper: the header on page 2 states 'arXiv:2508.09201v4 [cs.CR] 27 Jan 2026', and the title is 'Learning to Detect Unseen Jailbreak Attacks in Large Vision-Language Models'. The abstract, on the other hand, claims a microLED repair algorithm. This is not a minor formatting discrepancy: the central claim of the abstract is unsupported by any equation, section, table, or experiment in the manuscript. Sections 1-5 and all tables (Tables 1-6) concern LVLM jailbreak detection. There is no description of a differentiable transfer module, no microLED transfer setup, and no 2000x2000 array experiment. Because the manuscript's claimed contribution is entirely absent, the submission cannot be assessed as a paper on the stated topic.
  2. [Sections 3-5 (Methodology and Experiments)] Even if one generously read the body as the submitted content, none of the methodology applies to microLED repair. Equations (1)-(7) define MSCAV classifiers and a Safety Pattern Auto-Encoder for anomaly detection on LVLM safety representations. The experimental section evaluates AUROC, TPR, and F1 for jailbreak attacks across LLaVA, Qwen2.5-VL, and CogVLM. There is no equation, algorithm, or experiment addressing discrete shift planning, XY-stage motion minimization, or the claimed 50% step reduction. The load-bearing premise—that the full text corresponds to the abstract—is false, rendering the central claim unsubstantiated.
  3. [Abstract (claimed results)] The abstract's quantitative claims (50% reduction in transfer steps, sub-2-minute planning time on 2000x2000 arrays) are not backed by any data, table, or reproducibility artifact in the manuscript. No code, dataset, or baseline for microLED repair is provided. This absence of support for every stated experimental result is a fundamental issue that cannot be resolved by local revisions; the manuscript would need to be replaced entirely with the described work.
minor comments (2)
  1. [References] The reference list is entirely specific to the LVLM jailbreak detection paper (e.g., AdvBench, MM-SafetyBench, LLaVA, GradSafe). It contains no citations related to microLED manufacturing, transfer processes, or discrete optimization, underscoring the disconnect between the abstract and the body.
  2. [Code availability] The code link in the full text (anonymous.4open.science/r/Learning-to-Detect-51CB) points to the LoD framework, not to any microLED repair implementation. This is consistent with the body being the wrong paper, but it means no implementation of the claimed algorithm is available either.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular reduction exists; the claimed microLED algorithm is absent from the submitted body, which is an unrelated jailbreak-detection paper.

full rationale

The manuscript is structurally two documents. The abstract and metadata claim a first differentiable transfer-based microLED repair algorithm with a 50% reduction in transfer steps and sub-2-minute planning on 2000x2000 arrays, under arXiv:2508.09206. The submitted body, however, is titled 'Learning to Detect Unseen Jailbreak Attacks in Large Vision-Language Models', carries the header 'arXiv:2508.09201v4 [cs.CR] 27 Jan 2026', and never mentions microLED, transfer platforms, XY stages, or the reported experiments. There is therefore no derivation chain to audit for circularity: no equation defines the differentiable transfer module, no experiment reports the claimed speedup, and no baseline supports the claimed advantage. This is a fatal evidentiary mismatch, not a circularity in the reduction sense. Within the body itself, the only potentially self-referential dependency is the use of SCAV (Xu et al., 2024b) as the basis for the MSCAV classifiers; Xu and Wang are co-authors of both works. This does not make the LoD result circular, because the paper independently validates the linear separability assumption (Figure 3), evaluates against external baselines and external attack datasets, and cites a published, externally checkable NeurIPS paper. Hence the circularity score is 0: no prediction in the manuscript reduces to its inputs by construction, because the claimed prediction is simply absent from the submitted text.

Assumptions & free parameters 0 free parameters · 1 assumptions · 0 invented entities

The abstract's microLED claim is not developed in the text, so no free parameters or invented entities can be identified. The only identifiable axiom is that the body corresponds to the abstract, which fails given the mismatch with arXiv:2508.09201v4.

assumptions (1)
  • ad hoc to paper The submitted full text is the paper corresponding to the abstract
    The manuscript depends on this identity to provide any evidence for the microLED claims; the body is actually arXiv:2508.09201v4, so this axiom fails.

how reviews work

0 comments
Cite this review

Pith. "Pith review of The First Differentiable Transfer-Based Algorithm for Discrete MicroLED Repair." pith.science (2026). https://pith.science/paper/D3X66GG7

@misc{pith2026250809206,
  author       = {Pith},
  title        = {Pith review of: The First Differentiable Transfer-Based Algorithm for Discrete MicroLED Repair},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/D3X66GG7}},
  note         = {Machine review of arXiv:2508.09206}
}
read the original abstract

Laser-enabled selective transfer, a key process in high-throughput microLED fabrication, requires computational models that can plan shift sequences to minimize motion of XY stages and adapt to varying optimization objectives across the substrate. We propose the first repair algorithm based on a differentiable transfer module designed to model discrete shifts of transfer platforms, while remaining trainable via gradient-based optimization. Compared to local proximity searching algorithms, our approach achieves superior repair performance and enables more flexible objective designs, such as minimizing the number of steps. Unlike reinforcement learning (RL)-based approaches, our method eliminates the need for handcrafted feature extractors and trains significantly faster, allowing scalability to large arrays. Experiments show a 50% reduction in transfer steps and sub-2-minute planning time on 2000x2000 arrays. This method provides a practical and adaptable solution for accelerating microLED repair in AR/VR and next-generation display fabrication.

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 15 canonical work pages

  1. [1]

    Detecting language model attacks with perplexity.arXiv preprint arXiv:2308.14132,

    [Alon and Kamfonas, 2023] Gabriel Alon and Michael Kamfonas. Detecting language model attacks with perplexity.arXiv preprint arXiv:2308.14132,

    arXiv 2023

  2. [6]

    Unbridled icarus: A survey of the potential perils of image inputs in multimodal large lan- guage model security

    [Fanet al., 2024 ] Yihe Fan, Yuxin Cao, Ziyu Zhao, Ziyao Liu, and Shaofeng Li. Unbridled icarus: A survey of the potential perils of image inputs in multimodal large lan- guage model security

    work page 2024

  3. [7]

    Mirrorcheck: Efficient adver- sarial defense for vision-language models.arXiv preprint arXiv:2406.09250,

    [Fareset al., 2024 ] Samar Fares, Klea Ziu, Toluwani Aremu, Nikita Durasov, Martin Tak ´aˇc, Pascal Fua, Karthik Nan- dakumar, and Ivan Laptev. Mirrorcheck: Efficient adver- sarial defense for vision-language models.arXiv preprint arXiv:2406.09250,

    arXiv 2024

  4. [8]

    Figstep: Jailbreaking large vision- language models via typographic visual prompts.arXiv preprint arXiv:2311.05608,

    [Gonget al., 2023 ] Yichen Gong, Delong Ran, Jinyuan Liu, Conglei Wang, Tianshuo Cong, Anyu Wang, Sisi Duan, and Xiaoyun Wang. Figstep: Jailbreaking large vision- language models via typographic visual prompts.arXiv preprint arXiv:2311.05608,

    arXiv 2023

  5. [9]

    Mllm- guard: A multi-dimensional safety evaluation suite for multimodal large language models

    [Guet al., 2024 ] Tianle Gu, Zeyang Zhou, Kexin Huang, Dandan Liang, Yixu Wang, Haiquan Zhao, Yuanqi Yao, Xingge Qiao, Keqing Wang, and Yujiu Yang. Mllm- guard: A multi-dimensional safety evaluation suite for multimodal large language models

    work page 2024

  6. [10]

    Hudson and Christo- pher D

    [Hudson and Manning, 2019] Drew A. Hudson and Christo- pher D. Manning. Gqa: A new dataset for real-world visual reasoning and compositional question answering. IEEE,

    work page 2019

  7. [12]

    Hiddendetect: Detecting jailbreak attacks against large vision-language models via monitoring hid- den states.arXiv preprint arXiv:2502.14744,

    [Jianget al., 2025 ] Yilei Jiang, Xinyan Gao, Tianshuo Peng, Yingshui Tan, Xiaoyong Zhu, Bo Zheng, and Xi- angyu Yue. Hiddendetect: Detecting jailbreak attacks against large vision-language models via monitoring hid- den states.arXiv preprint arXiv:2502.14744,

    arXiv 2025

  8. [13]

    Jailbreakzoo: Survey, landscapes, and horizons in jail- breaking large language and vision-language models

    [Jinet al., 2024 ] Haibo Jin, Leyang Hu, Xinuo Li, Peiyan Zhang, Chonghan Chen, Jun Zhuang, and Haohan Wang. Jailbreakzoo: Survey, landscapes, and horizons in jail- breaking large language and vision-language models. arXiv preprint arXiv:2407.01599,

    arXiv 2024

Show all 35 references
  1. [14]

    Images are achilles’ heel ofalignment: Exploiting visual vulnerabilities forjail- breaking multimodal large language models

    [Liet al., 2025 ] Yifan Li, Hangyu Guo, Kun Zhou, Wayne Xin Zhao, and Ji Rong Wen. Images are achilles’ heel ofalignment: Exploiting visual vulnerabilities forjail- breaking multimodal large language models. InEuropean Conference on Computer Vision,

    2025

  2. [15]

    A survey of attacks on large vision-language models: Resources, advances, and future trends.arXiv preprint arXiv:2407.07403,

    [Liuet al., 2024a ] Daizong Liu, Mingyu Yang, Xiaoye Qu, Pan Zhou, Yu Cheng, and Wei Hu. A survey of attacks on large vision-language models: Resources, advances, and future trends.arXiv preprint arXiv:2407.07403,

    arXiv

  3. [16]

    Safety of multimodal large lan- guage models on images and texts.arXiv preprint arXiv:2402.00357,

    [Liuet al., 2024d ] Xin Liu, Yichen Zhu, Yunshi Lan, Chao Yang, and Yu Qiao. Safety of multimodal large lan- guage models on images and texts.arXiv preprint arXiv:2402.00357,

    arXiv

  4. [17]

    Jailbreakv: A bench- mark for assessing the robustness of multimodal large lan- guage models against jailbreak attacks.arXiv preprint arXiv:2404.03027,

    [Luoet al., 2024 ] Weidi Luo, Siyuan Ma, Xiaogeng Liu, Xi- aoyu Guo, and Chaowei Xiao. Jailbreakv: A bench- mark for assessing the robustness of multimodal large lan- guage models against jailbreak attacks.arXiv preprint arXiv:2404.03027,

    2024 arXiv

  5. [18]

    Visual-roleplay: Universal jailbreak attack on multimodal large language models via role-playing image character.arXiv preprint arXiv:2405.20773,

    [Maet al., 2024 ] Siyuan Ma, Weidi Luo, Yu Wang, and Xi- aogeng Liu. Visual-roleplay: Universal jailbreak attack on multimodal large language models via role-playing image character.arXiv preprint arXiv:2405.20773,

    2024 arXiv

  6. [19]

    Jailbreaking attack against multimodal large language model.arXiv preprint arXiv:2402.02309,

    [Niuet al., 2024 ] Zhenxing Niu, Haodong Ren, Xinbo Gao, Gang Hua, and Rong Jin. Jailbreaking attack against multimodal large language model.arXiv preprint arXiv:2402.02309,

    2024 arXiv

  7. [20]

    Mllm-protector: Ensuring mllm’s safety without hurting performance.arXiv preprint arXiv:2401.02906,

    [Piet al., 2024 ] Renjie Pi, Tianyang Han, Jianshu Zhang, Yueqi Xie, Rui Pan, Qing Lian, Hanze Dong, Jipeng Zhang, and Tong Zhang. Mllm-protector: Ensuring mllm’s safety without hurting performance.arXiv preprint arXiv:2401.02906,

    2024 arXiv

  8. [21]

    Visual adversarial examples jailbreak aligned large language models

    [Qiet al., 2024 ] Xiangyu Qi, Kaixuan Huang, Ashwinee Panda, Peter Henderson, Mengdi Wang, and Prateek Mit- tal. Visual adversarial examples jailbreak aligned large language models. InProceedings of the AAAI conference on artificial intelligence, volume 38, pages 21527–21536,

    2024

  9. [22]

    Anomaly detection using autoencoders with non- linear dimensionality reduction.ACM,

    [Sakurada and Yairi, 2014] Mayu Sakurada and Takehisa Yairi. Anomaly detection using autoencoders with non- linear dimensionality reduction.ACM,

    2014

  10. [24]

    Meta llama guard

    [Team, 2024] Llama Team. Meta llama guard

    2024

  11. [25]

    Cogvlm: Visual expert for pretrained language models

    [Wanget al., 2023 ] Weihan Wang, Qingsong Lv, Wenmeng Yu, Wenyi Hong, Ji Qi, Yan Wang, Junhui Ji, Zhuoyi Yang, Lei Zhao, Xixuan Song, Jiazheng Xu, Bin Xu, Juanzi Li, Yuxiao Dong, Ming Ding, and Jie Tang. Cogvlm: Visual expert for pretrained language models

    2023

  12. [26]

    Jailbreak large vision- language models through multi-modal linkage.arXiv preprint arXiv:2412.00473,

    [Wanget al., 2024b ] Yu Wang, Xiaofei Zhou, Yichen Wang, Geyuan Zhang, and Tianxing He. Jailbreak large vision- language models through multi-modal linkage.arXiv preprint arXiv:2412.00473,

    arXiv

  13. [27]

    Gradsafe: Detecting jailbreak prompts for llms via safety-critical gradient analysis.arXiv preprint arXiv:2402.13494,

    [Xieet al., 2024 ] Yueqi Xie, Minghong Fang, Renjie Pi, and Neil Gong. Gradsafe: Detecting jailbreak prompts for llms via safety-critical gradient analysis.arXiv preprint arXiv:2402.13494,

    2024 arXiv

  14. [28]

    Cross-modality information check for detecting jailbreaking in multimodal large language models.arXiv preprint arXiv:2407.21659,

    [Xuet al., 2024a ] Yue Xu, Xiuyuan Qi, Zhan Qin, and Wen- jie Wang. Cross-modality information check for detecting jailbreaking in multimodal large language models.arXiv preprint arXiv:2407.21659,

    arXiv

  15. [29]

    A survey of safety on large vision-language models: Attacks, defenses and evaluations.arXiv preprint arXiv:2502.14881,

    [Yeet al., 2025 ] Mang Ye, Xuankun Rong, Wenke Huang, Bo Du, Nenghai Yu, and Dacheng Tao. A survey of safety on large vision-language models: Attacks, defenses and evaluations.arXiv preprint arXiv:2502.14881,

    2025 arXiv

  16. [30]

    Mm- vet v2: A challenging benchmark to evaluate large multi- modal models for integrated capabilities.arXiv preprint arXiv:2408.00765,

    [Yuet al., 2024 ] Weihao Yu, Zhengyuan Yang, Lingfeng Ren, Linjie Li, Jianfeng Wang, Kevin Lin, Chung-Ching Lin, Zicheng Liu, Lijuan Wang, and Xinchao Wang. Mm- vet v2: A challenging benchmark to evaluate large multi- modal models for integrated capabilities.arXiv preprint arX...

    2024 arXiv

  17. [31]

    Mm- llms: Recent advances in multimodal large language mod- els

    [Zhanget al., 2024 ] Duzhen Zhang, Yahan Yu, Jiahua Dong, Chenxing Li, Dan Su, Chenhui Chu, and Dong Yu. Mm- llms: Recent advances in multimodal large language mod- els

    2024

  18. [32]

    Fc-attack: Jailbreak- ing multimodal large language models via auto-generated flowcharts.arXiv preprint ArXiv:2502.21059,

    [Zhanget al., 2025b ] Ziyi Zhang, Zhen Sun, Zongmin Zhang, Jihui Guo, and Xinlei He. Fc-attack: Jailbreak- ing multimodal large language models via auto-generated flowcharts.arXiv preprint ArXiv:2502.21059,

  19. [33]

    However, AUROC is an average criterion that may not fully reflect performance under specific operating conditions

    A Supplementary Evaluation Using Additional Criteria To further validate the effectiveness and robustness of our detection method, we note that the main text already demonstrates high AUROC scores. However, AUROC is an average criterion that may not fully reflect performance u...

    2024

  20. [34]

    For Vicuna, we improve over the best baseline by +45.6% on average, with especially large gains on UMK (+91.8%)

    Our method consistently achieves high values, significantly outperforming all baselines. For Vicuna, we improve over the best baseline by +45.6% on average, with especially large gains on UMK (+91.8%). Similar improvements hold for Qwen2.5-VL (+7.2% on average) and CogVLM (+13...

    2025

  21. [35]

    •GradSafe:The default initial safe set and unsafe set are used for selecting safety-critical parameters

    When testing on a specific model, the model is used to generate responses to each variant. •GradSafe:The default initial safe set and unsafe set are used for selecting safety-critical parameters. •HiddenDetect:The default refusal token list is used. For the selection of the Mo...

  22. [2014]

    Jailbreak in pieces: Composi- tional adversarial attacks on multi-modal language mod- els

    [Shayeganiet al., 2023 ] Erfan Shayegani, Yue Dong, and Nael Abu-Ghazaleh. Jailbreak in pieces: Composi- tional adversarial attacks on multi-modal language mod- els

    2023

  23. [2019]

    Playing the fool: Jailbreaking llms and multimodal llms with out-of- distribution strategy

    [Jeonget al., 2025 ] Joonhyun Jeong, Seyun Bae, Yeonsung Jung, Jaeryong Hwang, and Eunho Yang. Playing the fool: Jailbreaking llms and multimodal llms with out-of- distribution strategy. InProceedings of the Computer Vi- sion and Pattern Recognition Conference, pages 29937– 29946,

    2025

  24. [2022]

    Pixart-σ: Weak- to-strong training of diffusion transformer for 4k text-to- image generation

    [Chenet al., 2024 ] Junsong Chen, Chongjian Ge, Enze Xie, Yue Wu, Lewei Yao, Xiaozhe Ren, Zhongdao Wang, Ping Luo, Huchuan Lu, and Zhenguo Li. Pixart-σ: Weak- to-strong training of diffusion transformer for 4k text-to- image generation. InEuropean Conference on Computer Vision...

    2024

  25. [2023]

    [Baiet al., 2025 ] Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shi- jie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923,

    2025 arXiv

  26. [2024]

    A comparative study of rule-based and data-driven approaches in industrial monitoring.arXiv preprint arXiv:2509.15848,

    [De Gasperis and Facchini, 2025] Giovanni De Gasperis and Sante Dino Facchini. A comparative study of rule-based and data-driven approaches in industrial monitoring.arXiv preprint arXiv:2509.15848,

    2025 arXiv

  27. [2025]

    Why should adversarial perturbations be imperceptible? rethink the research paradigm in adversar- ial nlp.arXiv preprint arXiv:2210.10683,

    [Chenet al., 2022 ] Yangyi Chen, Hongcheng Gao, Ganqu Cui, Fanchao Qi, Longtao Huang, Zhiyuan Liu, and Maosong Sun. Why should adversarial perturbations be imperceptible? rethink the research paradigm in adversar- ial nlp.arXiv preprint arXiv:2210.10683,

    2022 arXiv

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.