Pith. sign in

REVIEW 5 major objections 6 minor 1 cited by

QwT-v2: Practical, Effective and Efficient Post-Training Quantization

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Pith's one-line read A per-channel affine correction folded into quantization scales recovers most accuracy loss of low-bit post-training quantization with zero extra inference cost.

desk verdict The per-channel post-compensation idea is good, but Table 8—the only evidence for lossless folding—is corrupted, leaving the zero-overhead claim unverified. read the letter →

arxiv 2505.20932 v1 pith:2Z6JOGKK submitted 2025-05-27 cs.CV

classification cs.CV
keywords post-trainingquantizationchannel-wiseaffinecompensationinteger-onlyinferencemodelcompressionvisiontransformerslargelanguagemodelsdiffusion
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

QwT-v2 sets out to fix the main practical obstacles of QwT, a quantization method that adds external compensation matrices to a post-training quantized network: roughly 30% extra parameters, extra floating-point compute, and poor compatibility with integer-only hardware. The paper proposes replacing the full-matrix, block-level pre-compensation with a channel-wise affine compensation (CWAC) applied after each quantized layer, whose scale and shift are solved in closed form from a small calibration set. It then shows that this diagonal correction can be absorbed into the layer's quantization scale and offset, so runtime cost vanishes and the model runs on standard integer inference engines. The claimed result is accuracy equal to or better than QwT across image classification, detection, CLIP, diffusion, and LLM tasks, with only 1-2% extra parameters before folding.

What carries the argument

The carrying object is the channel-wise affine compensation (CWAC) vector pair $(\alpha, \beta)$, one scalar per output channel, obtained from the one-dimensional regression in Equations (3)–(5), with $\alpha_c$ the ratio of covariance to variance and $\beta_c$ the intercept. The identity that does the heavy lifting is Equation (10): the compensated output equals the standard integer-quantized linear layer with the multiplier replaced by $M' = \alpha S_x S_w / S_r$ and an extra folded offset $\lfloor \beta / (\alpha S_x S_w) \rceil$, so the correction lives entirely inside the existing scale and offset parameters. This converts a diagonal linear compensation into a reparameterization of the quantized layer itself, which is why the method adds zero floating-point operations and no extra tensors at inference time.

What would settle it

Measure the per-channel coefficient of determination $R^2$ of the regressions in Equation (3) on a held-out set: if the residuals are strongly nonlinear in $Y_{\mathrm{quant}}$ or the fit quality drops sharply off the calibration set, the affine correction cannot hold. A direct end-to-end test is a 4-bit fully quantized language model where applying the folded CWAC fails to lower perplexity versus the base PTQ model, or where the rounded-integer $\beta$ fold changes results materially.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that a per-channel affine map $Y_{\mathrm{comp}} = \alpha_c \odot Y_{\mathrm{quant},:c} + \beta_c$ fit by ordinary least squares on the calibration set aligns each quantized output channel with its floating-point counterpart, and that the fitted parameters can be folded into the inference engine's multiplier $M' = \alpha S_x S_w / S_r$ and an integer offset, eliminating the compensation as a separate operation. The shift from QwT's pre-compensation (regressing a block's output on its input through the block's nonlinearity) to post-compensation (regressing the output of a single quantized linear or convolutional layer directly on that layer's own output) is what makes the correction nearly diagonal and therefore cheap to fold. The empirical claim is that this simple correction recovers most of the accuracy QwT recovers, and in several settings improves on it, while adding 1-2% parameters and zero inference overhead.

Load-bearing premise

The per-channel quantization error is well approximated by an affine function of the quantized output, meaning one scale and one shift per channel fitted on 512 calibration samples carries over to the rest of the data distribution.

Editorial extensions

If this is right

  • A 4-bit or 6-bit quantized model can be deployed on integer-only hardware with the same computation graph and latency as the base post-training quantization model, because the compensation lives only in the scale and offset.
  • The method requires only a small calibration set (512 samples) and a closed-form regression, so it can be applied in minutes without hyperparameter tuning on top of any existing PTQ baseline.
  • Because compensation is applied per quantized layer, it stacks consistently onto RepQ-ViT, Percentile, GPTQ, and Q-DiT, and improves accuracy on detection and CLIP vision-plus-text tasks beyond what QwT achieves.
  • The extra parameter cost drops from roughly 30% of the model size in QwT to about 1-2% in QwT-v2, and after folding into the inference engine even those parameters are removed.

Reading between the lines

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

  • The near-diagonal structure of the fitted post-compensation weights suggests the dominant quantization error is a per-channel gain-and-bias distortion rather than cross-channel mixing; a direct test would ask whether a full per-layer matrix yields meaningful further gains once the diagonal part is folded.
  • CWAC could in principle be re-estimated continuously during deployment, since it only needs per-channel means, variances, and covariances; the paper does not explore adapting the compensation under distribution shift.
  • The fine-grained placement after each quantized layer leaves open compositions with residual connections or normalization layers that the paper does not examine, such as applying a second affine correction after the residual add.
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

5 major / 6 minor

Summary. The paper proposes QwT-v2, a post-training quantization compensation method built on the authors' earlier QwT. Instead of QwT's full-matrix, block-level pre-compensation, QwT-v2 inserts per-channel affine compensation (CWAC) modules after each quantized linear or convolutional layer, fits the per-channel scale alpha_c and shift beta_c by closed-form least squares on 512 calibration samples, and claims that the compensation can be folded into the quantization scale and offset (Eqs. 9-11) so that inference remains integer-only with zero additional overhead. Experiments cover ImageNet classification, COCO detection, CLIP zero-shot recognition, DiT image generation, and LLaMA-3 language modeling, reporting large gains over the PTQ baselines and an extra-parameter cost of only about 1-2% of the model size, compared with roughly 30% for QwT.

Significance. If the claims hold, QwT-v2 would be a practically valuable extension: it is simple, closed-form, training-free at initialization, parameter-efficient, and potentially hardware-friendly after folding. The paper has several strengths: the closed-form regression in Eqs. (4)-(5) is straightforward; the evaluation is broad, spanning vision, multimodal, generation, and language tasks; the measured parameter overheads are consistently much smaller than QwT's; and the evaluation is not circular, since the compensation parameters are fit on calibration data and evaluated on held-out validation data. The main weakness is that two central claims are not fully supported by the reported evidence: the abstract's claim that QwT-v2 'matches or even outperforms QwT' is contradicted by several non-negligible accuracy gaps, and the claim that folding into the inference engine is lossless is supported only by a corrupted ablation table.

major comments (5)
  1. [Section 4.6, Table 8] The only empirical evidence for the claim that rounding beta is negligible is Table 8, but this table is internally inconsistent: the Swin-T 6/6 row reports an unrounded top-1 accuracy of 90.0 against the FP32 baseline of 81.4, and the final block labeled ViT-B repeats the ResNet-50 values 75.3/75.1 and 76.7/76.7 from Table 1. These entries cannot be correct. Since Section 3.4 relies on this table to justify lossless folding, the zero-overhead integration claim is currently unverified and needs a corrected experiment.
  2. [Section 3.4, Eq. (10)] The folding derivation in Eq. (10) treats arithmetic as exact and does not account for output clipping. In Eq. (1), quantized values are restricted to [0, 2^b-1], and standard integer inference engines clip the requantized accumulator output. If the affine transform alpha*y + beta pushes the folded integer result outside this range, the folded computation will not reproduce the explicit floating-point compensation. The manuscript neither analyzes this case nor provides a valid experiment (Table 8 being unusable), so the paper should either add a clipping analysis or compare folded versus explicit compensation empirically.
  3. [Abstract, Table 9, Table 3] The abstract claims QwT-v2 'matches or even outperforms QwT in accuracy', but the reported results show several non-negligible losses: DeiT-T 4-bit (59.9 vs 61.4 in Table 9), ViT-B 4-bit (75.6 vs 76.3 in Table 9), and CLIP vision-only 6/6-bit (59.8 vs 60.3 in Table 3). These gaps are up to 1.5 points. Please either soften the claim to 'comparable accuracy' with enumerated exceptions or provide a more careful statistical comparison.
  4. [Section 3.3] The per-layer compensation parameters are fit independently on the outputs of the uncompensated quantized model, but at inference all compensations are inserted simultaneously, so later layers receive corrected activations. This distribution shift means the closed-form solutions are not jointly optimal for the final model, and the shift could also affect clipping after the folding in Section 3.4. The paper does not discuss or evaluate this effect; an ablation comparing one-pass fitting with an iterative or blockwise refit would make the method's claims more robust.
  5. [Table 6] In Table 6, Swin-T 4/4 accuracy jumps from 69.9 with 32 calibration images to 77.0 with 128 images, a 7.1-point increase, while the text reports only a 0.2-point gap over this range. This is either a typographical error or an indication that 32 images is insufficient in a way that deserves discussion; please correct or explain.
minor comments (6)
  1. [Section 4.4] In the discussion of Table 4, the sentence 'FID of QwT is 0.03 higher than QwT' contains a naming error and should compare QwT-v2 with QwT while stating the direction correctly.
  2. [Section 3.2, Eq. (1)] The notation 'clip(x, u, v)' is followed by 'clips x to the [a, b] range', which mixes variable names; use consistent names for the lower and upper bounds.
  3. [Section 1] 'memory footage' should be 'memory footprint'.
  4. [Section 4.6] The heading 'we conducted' should be capitalized, and 'Resnet-50' should be 'ResNet-50' for consistency.
  5. [Section 6] 'we are interesting in finding' should be 'we are interested in finding'.
  6. [Appendix A] The sentence 'the initialization of QwT-v2 is done on 4 RTX 3090 GPUs.' is separated from its context by a stray 'the computation.'; please clean up the sentence.

Circularity Check

0 steps flagged · score 2.0 of 10

No significant circularity: compensation is fit on calibration and evaluated on held-out data; minor QwT self-citations are not load-bearing.

full rationale

No significant circularity in the claimed derivation. The compensation parameters alpha_c and beta_c are obtained as the closed-form least-squares solution of Eq. (3) on a 512-sample calibration set, and every reported accuracy is measured on held-out validation data (ImageNet validation, COCO val, WikiText2/C4, etc.), so the main accuracy results are genuine out-of-sample evaluations rather than the fit being renamed as a prediction. The integration algebra in Eqs. (9)-(11) follows from the standard integer-only quantization definitions in Eqs. (6)-(8) and the affine compensation form; it is a rewriting of the same computation and does not import the target accuracy as an input. The paper does rely on the authors' own QwT paper (Fu et al., 2025) for experimental settings and for the diffusion assumption that quantization error mainly depends on the token x rather than the timestep t; this is a self-citation, but it is not load-bearing for the paper's central claim, and QwT-v2's effectiveness is also supported by independent comparisons across classification, detection, CLIP, diffusion, and LLaMA-3 tasks. The most serious problem is a correctness/validation issue rather than circularity: Table 8, the sole direct evidence for the claim that rounding beta is negligible, contains an impossible 90.0 top-1 for Swin-T 6-bit (above the 81.4 FP32 baseline) and a final block repeating the ResNet-50 values from Table 1, so the zero-overhead folded-inference claim is not backed by a valid experiment. That gap weakens the practical claim but does not make the derivation self-referential.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities; CWAC is a parameterization of existing tensors. The free parameters are the per-channel alpha and beta fits plus the user-chosen calibration set size. The main domain assumption is that a per-channel affine transform captures quantization error.

free parameters (2)
  • Per-channel affine compensation parameters (alpha_c, beta_c) = Not reported per channel; estimated by closed-form covariance and mean over the calibration set
    For every output channel of every quantized linear or convolutional layer, alpha and beta are fit on calibration data via Equations (4)-(5); the central improvement claim depends on these fitted values.
  • Calibration set size = 512 samples for QwT-v2 initialization, 32 samples for baseline PTQ
    Chosen by hand; the ablation in Table 6 shows only small sensitivity (0.2-0.5 accuracy points), so it is not a strong free parameter, but it is a user-chosen input.
assumptions (4)
  • standard math Uniform quantization formulation of Equations (1)-(2)
    Standard definition of uniform affine quantization, used throughout the paper.
  • domain assumption Per-channel affine correction sufficiency
    Section 3.3 posits that minimizing ||Y_full - (alpha ⊙ Y_quant + beta)|| per channel recovers quantization error; this is empirically validated but not derived.
  • domain assumption Diffusion quantization error depends mainly on input token, not timestep
    Adopted from Fu et al. (2025) to justify setting t=0 for compensation initialization in Section 4.4; the current paper does not test this assumption independently.
  • domain assumption Rounding of folded beta is negligible
    Section 3.4 and Table 8 claim the error from rounding beta/(alpha Sx Sw) to an integer is negligible; only one ablative table supports this and that table contains inconsistencies.

how reviews work

0 comments
Cite this review

Pith. "Pith review of QwT-v2: Practical, Effective and Efficient Post-Training Quantization." pith.science (2026). https://pith.science/paper/2Z6JOGKK

@misc{pith2026250520932,
  author       = {Pith},
  title        = {Pith review of: QwT-v2: Practical, Effective and Efficient Post-Training Quantization},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/2Z6JOGKK}},
  note         = {Machine review of arXiv:2505.20932}
}
read the original abstract

Network quantization is arguably one of the most practical network compression approaches for reducing the enormous resource consumption of modern deep neural networks. They usually require diverse and subtle design choices for specific architecture and tasks. Instead, the QwT method is a simple and general approach which introduces lightweight additional structures to improve quantization. But QwT incurs extra parameters and latency. More importantly, QwT is not compatible with many hardware platforms. In this paper, we propose QwT-v2, which not only enjoys all advantages of but also resolves major defects of QwT. By adopting a very lightweight channel-wise affine compensation (CWAC) module, QwT-v2 introduces significantly less extra parameters and computations compared to QwT, and at the same time matches or even outperforms QwT in accuracy. The compensation module of QwT-v2 can be integrated into quantization inference engines with little effort, which not only effectively removes the extra costs but also makes it compatible with most existing hardware platforms.

Figures

Figures reproduced from arXiv: 2505.20932 by the authors.

Figure 1
Figure 1. The left figure (a) shows pre-compensation in QwT (top) and our post-compensation in [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. The architecture of QwT-v2’s compensation in a vision Transformer. Compensations are [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

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

  1. YOLOv8-SMOT: An Efficient and Robust Framework for Real-Time Small Object Tracking via Slice-Assisted Training and Adaptive Association

    cs.CV 2025-07 conditional novelty 4.0 of 10

    A YOLOv8 detector trained on overlapping slices plus an OC-SORT tracker with EMA motion direction and expanded IoU distance penalty achieves 55.205 SO-HOTA on the SMOT4SB public test set.

Reference graph

Works this paper leans on

45 extracted references · 40 canonical work pages · cited by 1 Pith paper

  1. [1]

    write newline

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

  2. [2]

    BERT : Pre-training of deep bidirectional transformers for language understanding

    Jacob Devlin, Ming-Wei Chang, Kenton Lee, and Kristina Toutanova. BERT : Pre-training of deep bidirectional transformers for language understanding. In North A merican Chapter of the Association for Computational Linguistics , pages 4171--4186, 2019

  3. [3]

    Tom B. Brown, Benjamin Mann, Nick Ryder, Melanie Subbiah, Jared Kaplan, Prafulla Dhariwal, Arvind Neelakantan, Pranav Shyam, Girish Sastry, Amanda Askell, Sandhini Agarwal, Ariel Herbert-Voss, Gretchen Krueger, Tom Henighan, Rewon Child, Aditya Ramesh, Daniel M. Ziegler, Jeffrey Wu, Clemens Winter, Christopher Hesse, Mark Chen, Eric Sigler, Mateusz Litwin...

  4. [4]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, et al. An image is worth 16x16 words: Transformers for image recognition at scale. In International Conference on Learning Representations, pages 1--21, 2021

  5. [5]

    Mask R-CNN

    Kaiming He, Georgia Gkioxari, Piotr Dollár, and Ross Girshick. Mask R-CNN . In IEEE International Conference on Computer Vision, pages 2980--2988, 2017

  6. [6]

    S peech GPT : Empowering large language models with intrinsic cross-modal conversational abilities"

    Dong Zhang, Shimin Li, Xin Zhang, Jun Zhan, Pengyu Wang, Yaqian Zhou, and Xipeng Qiu. S peech GPT : Empowering large language models with intrinsic cross-modal conversational abilities". In Findings of the Association for Computational Linguistics: EMNLP 2023, pages 15757--15773, 2023

  7. [7]

    Learning transferable visual models from natural language supervision

    Alec Radford, Jong Wook Kim, Chris Hallacy, et al. Learning transferable visual models from natural language supervision. In International Conference on Machine Learning, pages 8748--8763, 2021

  8. [8]

    Visual instruction tuning

    Haotian Liu, Chunyuan Li, Qingyang Wu, and Yong Jae Lee. Visual instruction tuning. In Proceedings of the 37th International Conference on Neural Information Processing Systems (NeurIPS), 2023

Show all 45 references
  1. [9]

    Pruning and quantization for deep neural network acceleration: A survey

    Tailin Liang, John Glossner, Lei Wang, Shaobo Shi, and Xiaotong Zhang. Pruning and quantization for deep neural network acceleration: A survey. Neurocomputing, 461 0 (C): 0 370–403, 2021

  2. [10]

    BRECQ : Pushing the limit of post-training quantization by block reconstruction

    Yuhang Li, Ruihao Gong, Xu Tan, et al. BRECQ : Pushing the limit of post-training quantization by block reconstruction. In International Conference on Learning Representations, pages 1--12, 2021

  3. [11]

    Smoothquant: accurate and efficient post-training quantization for large language models

    Guangxuan Xiao, Ji Lin, Mickael Seznec, Hao Wu, Julien Demouth, and Song Han. Smoothquant: accurate and efficient post-training quantization for large language models. In Proceedings of the 40th International Conference on Machine Learning, 2023

  4. [12]

    Learned step size quantization

    Steven K Esser, Jeffrey L McKinstry, Deepika Bablani, Rathinakumar Appuswamy, and Dharmendra S Modha. Learned step size quantization. In International Conference on Learning Representations, pages 1--10, 2020

  5. [13]

    Quantized feature distillation for network quantization

    Ke Zhu, Yin-Yin He, and Jianxin Wu. Quantized feature distillation for network quantization. In AAAI Conference on Artificial Intelligence, pages 11452--11460, 2023

  6. [14]

    Q-ViT : Accurate and fully quantized low-bit Vision Transformer

    Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Peng Gao, and Guodong Guo. Q-ViT : Accurate and fully quantized low-bit Vision Transformer . In Advances in Neural Information Processing Systems, pages 34451 -- 34463, 2024

  7. [15]

    Quantization without tears

    Minghao Fu, Hao Yu, Jie Shao, Junjie Zhou, Ke Zhu, and Jianxin Wu. Quantization without tears. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), 2025

  8. [16]

    Gonzalez, Hao Zhang, and Ion Stoica

    Woosuk Kwon, Zhuohan Li, Siyuan Zhuang, Ying Sheng, Lianmin Zheng, Cody Hao Yu, Joseph E. Gonzalez, Hao Zhang, and Ion Stoica. Efficient memory management for large language model serving with pagedattention. In Proceedings of the ACM SIGOPS 29th Symposium on Operating Systems...

  9. [17]

    ReActNet : Towards precise binary neural network with generalized activation functions

    Zechun Liu, Zhiqiang Shen, Marios Savvides, and Kwang-Ting Cheng. ReActNet : Towards precise binary neural network with generalized activation functions. In European Conference on Computer Vision, volume 12359 of LNIP, pages 143--159, 2020

  10. [18]

    Network quantization with element-wise gradient scaling

    Junghyup Lee, Dohyung Kim, and Bumsub Ham. Network quantization with element-wise gradient scaling. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6444--6453, 2021

  11. [19]

    Lsq+: Improving low-bit quantization through learnable offsets and better initialization

    Yash Bhalgat, Jinwon Lee, Markus Nagel, Tijmen Blankevoort, and Nojun Kwak. Lsq+: Improving low-bit quantization through learnable offsets and better initialization. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR) Workshops, pages 69...

  12. [20]

    PTQ4ViT : Post-training quantization for Vision Transformers with twin uniform quantization

    Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. PTQ4ViT : Post-training quantization for Vision Transformers with twin uniform quantization. In European Conference on Computer Vision, volume 13672 of LNCS, pages 191--207, 2022

  13. [21]

    GPTQ : Accurate post-training quantization for generative pre-trained Transformers

    Elias Frantar, Saleh Ashkboos, Torsten Hoefler, and Dan Alistarh. GPTQ : Accurate post-training quantization for generative pre-trained Transformers . In International Conference on Learning Representations, pages 1--12, 2023

  14. [22]

    Up or down? adaptive rounding for post-training quantization

    Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Christos Louizos, and Tijmen Blankevoort. Up or down? adaptive rounding for post-training quantization. In International Conference on Machine Learning, pages 7197--7206, 2020

  15. [23]

    QDROP : Randomly dropping quantization for extremely low-bit post-training quantization

    Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. QDROP : Randomly dropping quantization for extremely low-bit post-training quantization. In International Conference on Learning Representations, pages 1--12, 2022

  16. [24]

    RepQ-ViT : Scale reparameterization for post-training quantization of Vision Transformers

    Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. RepQ-ViT : Scale reparameterization for post-training quantization of Vision Transformers . In IEEE/CVF International Conference on Computer Vision, pages 17181--17190, 2023

  17. [25]

    FQ-ViT : Post-training quantization for fully quantized Vision Transformer

    Yang Lin, Tianyu Zhang, Peiqin Sun, Zheng Li, and Shuchang Zhou. FQ-ViT : Post-training quantization for fully quantized Vision Transformer . In International Joint Conference on Artificial Intelligence, pages 1173--1179, 2022

  18. [26]

    Quantization and training of neural networks for efficient integer-arithmetic-only inference

    Benoit Jacob, Skirmantas Kligys, Bo Chen, Menglong Zhu, Matthew Tang, Andrew Howard, Hartwig Adam, and Dmitry Kalenichenko. Quantization and training of neural networks for efficient integer-arithmetic-only inference. In Proceedings of the IEEE Conference on Computer Vision an...

  19. [27]

    TensorFlow Lite , 2024

    Google Inc. TensorFlow Lite , 2024. URL https://www.tensorflow.org/lite. Version 2.16.1

  20. [28]

    Fully quantized network for object detection

    Rundong Li, Yan Wang, Feng Liang, Hongwei Qin, Junjie Yan, and Rui Fan. Fully quantized network for object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 2805--2814, 2019

  21. [29]

    ImageNet : A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. ImageNet : A large-scale hierarchical image database. In IEEE Conference on Computer Vision and Pattern Recognition, pages 248--255, 2009

  22. [30]

    Swin Transformer : Hierarchical Vision Transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, et al. Swin Transformer : Hierarchical Vision Transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision, pages 9992--10002, 2021

  23. [31]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In IEEE Conference on Computer Vision and Pattern Recognition, pages 770--778, 2016

  24. [32]

    Microsoft COCO : Common objects in context

    Tsung-Yi Lin, Michael Maire, Serge Belongie, et al. Microsoft COCO : Common objects in context. In European Conference on Computer Vision, volume 8693 of LNCS, pages 740--755, 2014

  25. [33]

    Cascade R-CNN : Delving into high quality object detection

    Zhaowei Cai and Nuno Vasconcelos. Cascade R-CNN : Delving into high quality object detection. In IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6154--6162, 2018

  26. [34]

    The llama 3 herd of models

    Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, and others. The llama 3 herd of models. arXiv:2407.21783, 2024

  27. [35]

    Pointer sentinel mixture models

    Merity Stephen, Xiong Caiming, Bradbury James, et al. Pointer sentinel mixture models. In International Conference on Learning Representations, pages 1--11, 2017

  28. [36]

    Exploring the limits of transfer learning with a unified text-to-text Transformer

    Colin Raffel, Noam Shazeer, Adam Roberts, et al. Exploring the limits of transfer learning with a unified text-to-text Transformer . Journal of Machine Learning Research, 21 0 (140): 0 1--67, 2020

  29. [37]

    Social IQ a: Commonsense reasoning about social interactions

    Maarten Sap, Hannah Rashkin, Derek Chen, Ronan Le Bras, and Yejin Choi. Social IQ a: Commonsense reasoning about social interactions. In Conference on Empirical Methods in Natural Language Processing and the International Joint Conference on Natural Language Processing, pages ...

  30. [38]

    HellaSwag : Can a machine really finish your sentence? In Annual Meeting of the Association for Computational Linguistics, page 4791–4800, 2019

    Rowan Zellers, Ari Holtzman, Yonatan Bisk, Ali Farhadi, and Yejin Choi. HellaSwag : Can a machine really finish your sentence? In Annual Meeting of the Association for Computational Linguistics, page 4791–4800, 2019

  31. [39]

    Piqa: Reasoning about physical commonsense in natural language

    Yonatan Bisk, Rowan Zellers, Ronan Le Bras, Jianfeng Gao, and Yejin Choi. Piqa: Reasoning about physical commonsense in natural language. In AAAI Conference on Artificial Intelligence, pages 7432--7439, 2020

  32. [40]

    WinoGrande : an adversarial winograd schema challenge at scale

    Keisuke Sakaguchi, Ronan Le Bras, Chandra Bhagavatula, and Yejin Choi. WinoGrande : an adversarial winograd schema challenge at scale. Communications of the ACM, 64 0 (9): 0 99–106, 2021

  33. [41]

    Think you have solved question answering? try arc, the ai2 reasoning challenge

    Peter Clark, Isaac Cowhey, Oren Etzioni, Tushar Khot, Ashish Sabharwal, Carissa Schoenick, and Oyvind Tafjord. Think you have solved question answering? try arc, the ai2 reasoning challenge. arXiv:1803.05457, 2018

  34. [42]

    Boolq: Exploring the surprising difficulty of natural yes/no questions

    Christopher Clark, Kenton Lee, Ming-Wei Chang, Tom Kwiatkowski, Michael Collins, and Kristina Toutanova. Boolq: Exploring the surprising difficulty of natural yes/no questions. In Annual Conference of the North American Chapter of the Association for Computational Linguistics,...

  35. [43]

    Can a suit of armor conduct electricity? a new dataset for open book question answering

    Todor Mihaylov, Peter Clark, Tushar Khot, and Ashish Sabharwal. Can a suit of armor conduct electricity? a new dataset for open book question answering. In Conference on Empirical Methods in Natural Language Processing, pages 2381--2391, 2018

  36. [44]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. In International Conference on Learning Representations, pages 1--18, 2019

  37. [45]

    Q-DiT : Accurate post-training quantization for diffusion Transformers

    Lei Chen, Yuan Meng, Chen Tang, et al. Q-DiT : Accurate post-training quantization for diffusion Transformers . arXiv:2406.17343, 2024

Pith tools

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