Pith. sign in

REVIEW 4 major objections 6 minor 55 references

Measuring which Vision Transformer parts break under low-bit quantization, then solving a knapsack problem, yields mixed-precision models that transfer far better on detection.

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 →

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

T0 review · grok-4.5

2026-07-31 02:55 UTC pith:P576AZCH

load-bearing objection Clean mixed-precision PTQ recipe on AdaLog; COCO +9.6 AP is the real claim and needs protocol checks, while ImageNet “competitive” is oversold. the 4 major comments →

arxiv 2607.28589 v1 pith:P576AZCH submitted 2026-07-30 cs.CV cs.LG

MixFrag: Fragility-Guided Mixed-Precision Post-Training Quantization for Vision Transformers

Md. Mehrab Hossain Opi , Robiul Islam Ryad , Md. Umar Faruk This is my paper
classification cs.CV cs.LG
keywords post-training quantizationmixed-precision quantizationVision Transformersquantization fragilityKL divergencemultiple-choice knapsackAdaLogCOCO detection
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.

Vision Transformers are hard to shrink for phones and edge chips because different layers react very differently when you cut their numerical precision. Most post-training methods still give every layer the same bit-width, wasting bits on sturdy parts and starving fragile ones. MixFrag scores each quantizable piece by how much the model’s output distribution shifts when that piece alone is quantized, using a tiny unlabeled calibration set and KL divergence. It then treats bit assignment as a multiple-choice knapsack problem so the total average bit budget is met while total fragility is minimized, and hands the resulting map to an existing PTQ backend. On ImageNet the method stays competitive at practical 4–6 bit mixed settings; on COCO detection and instance segmentation it leads other mixed-precision PTQ methods, gaining up to 9.6 AP at the harsh 3-bit mixed setting. The paper’s point is that directly measuring output damage, not proxy importance scores, is enough to steer precision where it matters for both classification and dense vision tasks.

Core claim

Component-level quantization fragility—measured as the KL divergence between full-precision and isolated-quantized output distributions on a small calibration set—plus a multiple-choice knapsack formulation of bit allocation, produces mixed-precision ViTs that match or beat uniform and prior mixed-precision PTQ on ImageNet under practical budgets and set the strongest mixed-precision PTQ results on COCO detection and segmentation, including a 9.6 AP gain at MP3/MP3.

What carries the argument

Quantization fragility Ω(c,b): the average KL divergence between the full-precision model’s logits and those of a model where only component c is quantized to bit-width b. That matrix becomes the cost table for an exact dynamic-programming MCKP that assigns each layer one bit-width under a global average-bit budget.

Load-bearing premise

How much a layer hurts the model when quantized alone on a few dozen unlabeled images is assumed to tell you how many bits that layer still needs when every layer is quantized together.

What would settle it

Re-run the same MCKP pipeline but replace isolated KL scores with scores measured while neighboring layers are already at their allocated low bits (or with a true joint search); if the COCO AP lead and the reported fragility–bit correlations disappear, the isolation premise fails.

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

If this is right

  • Under tight average-bit budgets, attention projections and early blocks should systematically receive more bits than later MLP layers when allocation is fragility-driven.
  • Mixed-precision PTQ judged only on ImageNet top-1 can miss large gains on detection and segmentation that come from protecting intermediate features.
  • A single-pass fragility estimate plus one global knapsack solve can outperform iterative re-estimation of layer sensitivity for this isolation-based metric.
  • Any PTQ backend that already supports per-layer bit-widths can adopt the same fragility table and MCKP allocator without redesigning its quantizer.

Where Pith is reading between the lines

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

  • If isolated output KL is a good enough proxy here, similar isolation-plus-knapsack pipelines may transfer to other transformer families (language, multimodal) where full joint search is still too expensive.
  • The paper’s finding that marginal joint refinement hurt accuracy suggests future work should model pairwise layer interactions explicitly rather than re-estimate full fragility after a first pass.
  • Stronger COCO transfer than ImageNet ranking implies deployment benchmarks for ViT quantization should weight dense-prediction heads more heavily than classification alone.

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

4 major / 6 minor

Summary. MixFrag proposes a mixed-precision post-training quantization pipeline for Vision Transformers. Component-level “quantization fragility” is defined as the mean KL divergence between full-precision logits and logits obtained when a single component is quantized in isolation (Eq. 1), estimated on a small unlabeled calibration set. Bit-widths are then chosen by solving a Multiple-Choice Knapsack Problem that minimizes the sum of these isolation fragilities under an average bit budget (Eqs. 2–4), and the resulting assignment is applied through the AdaLog PTQ backend. On ImageNet-1K the method is reported as competitive under practical mixed-precision budgets (Table 1). On COCO detection and instance segmentation with Swin-backed Mask/Cascade R-CNN, MixFrag claims state-of-the-art among mixed-precision PTQ methods, with up to +9.6 average AP over LRP-AQViT under MP3/MP3 (Table 2). Ablations compare random, greedy, and MCKP allocation (Table 6), report fragility–bit correlations (Table 5), and find that an optional joint “marginal refinement” pass hurts accuracy (Table 7).

Significance. If the COCO gains are attributable to the fragility+MCKP allocation rather than backend or protocol differences, the work would be a useful practical contribution to mixed-precision PTQ for ViTs: isolation KL is simple, the MCKP formulation is standard and exact via DP, and the pipeline is backend-agnostic in principle. The ablations are coherent (random < greedy ≤ MCKP; structured attention/early-block bit patterns in Tables 3–4; moderate fragility–bit correlation at tight budgets). The paper is honest about unmodeled joint interactions (§4.6) and about weaker ImageNet W3/A3 numbers relative to optimization-heavy uniform PTQ. The distinctive scientific claim, however, rests heavily on the unreplicated COCO MP3 delta and on the isolation proxy transferring to joint quantization—both of which need tighter evidence before the SOTA framing is secure.

major comments (4)
  1. [Table 2, §4.2] Table 2 / abstract SOTA claim (+9.6 AP, MP3/MP3): the headline comparison to LRP-AQViT (and other MP baselines) is not shown under a matched quantization backend and protocol. MixFrag uses AdaLog with 256 COCO calibration images; LRP-AQViT numbers appear taken from prior work without a same-stack re-run or a controlled “AdaLog + LRP/LAMPQ bits” arm. Part of the 40.9 vs 31.3 gap may therefore be backend, checkpoint, or eval protocol rather than fragility-guided MCKP. Please re-evaluate the main mixed-precision baselines under the authors’ AdaLog pipeline (same Swin-T/S Mask and Cascade checkpoints, same calib size, same MP3/MP4 meaning for W/A, no extra fine-tuning) or, at minimum, ablate MixFrag bits vs LRP/LAMPQ bits on identical AdaLog quantization so the allocation method is isolated from the backend.
  2. [§3.2 Eq. (1), §4.5.2 Table 7, §4.6] Eq. (1) and Fig. 1(a) define fragility via isolated quantization (one component at bit b, all others full precision) on as few as 32 images in the figure caption (1,024/256 in §4.1). §4.6 explicitly states that interactions among simultaneously quantized components are not modeled, and Table 7 shows that a joint “marginal refinement” re-estimation pass consistently hurts accuracy (e.g., W3/A3 31.91 → 29.73). That result directly pressures the isolation axiom: if joint re-estimation is harmful/unreliable, it is unclear why isolation ranks remain the right statistic for full MP assignment. Please add (i) sensitivity of Ω and final bits to calibration size and temperature T, (ii) a comparison of isolation ranks vs a cheaper joint proxy (e.g., leave-one-higher-bit or block-wise reconstruction error under a fixed MP skeleton), and (iii) a clear statement of when isolation is expected to fail.
  3. [Table 1, Abstract, §4.2] Table 1 shows MixFrag trailing strong uniform PTQ at W3/A3 by a large margin (avg 42.34 vs FIMA-Q 71.29, DopQ-ViT 63.39, I&S-ViT 60.43) and also below other mixed-precision methods (LAMPQ 51.81, Mix-AQViT 53.27). At W4/A4 it remains below FIMA-Q and Mix-AQViT. The paper’s distinctive value therefore rests disproportionately on COCO transfer (§4.2 discussion). That is a legitimate scientific angle, but the manuscript currently frames ImageNet as “competitive” without quantifying when fragility+MCKP helps vs hurts relative to the AdaLog uniform baseline alone. Please add a controlled ImageNet column for uniform AdaLog vs MixFrag at matched average bits on the same runs, and temper abstract/intro wording so the primary claim is downstream robustness under MP PTQ rather than broad ImageNet competitiveness at 3-bit.
  4. [§3.3 Eqs. (2)–(4), Table 3] MCKP Eqs. (2)–(4) minimize sum of isolation KL under an average bit constraint Σ b_j x_ij ≤ B_target, but the constraint is written as a sum of bit-widths rather than a parameter- or MAC-weighted memory/compute budget. For ViTs, MLP layers dominate parameter count; Table 3 already shows attention getting higher average bits than FC layers under W3/A3. If B_target is an unweighted average over components, the optimizer can “spend” bits on small attention projections while aggressively quantizing large FFN matrices, which may not match a true model-size or BitOps budget used by prior MPQ work. Please define B_target operationally (per-tensor average vs size-weighted), report actual model size / BitOps for MixFrag vs uniform and vs LAMPQ/LRP allocations, and confirm COCO/ImageNet comparisons use the same budget accounting.
minor comments (6)
  1. [Figure 1, §4.1] Fig. 1 caption says fragility is computed on 32 unlabeled images, while §4.1 states 1,024 ImageNet / 256 COCO calibration images. Clarify which N is used for Ω vs for AdaLog FPCS calibration.
  2. [Table 1, References] Table 1 lists “AQViT [44]” and “AdaLog [44]” with the same reference; confirm citation keys and whether AQ-ViT is a separate method or an AdaLog variant.
  3. [Figure 1, Table 4] Candidate set B is described as {2,3,4,6} in Fig. 1 but Table 4 reports assigned widths up to 7 under W6/A6. State the full candidate set used in experiments.
  4. [§3.2 Eq. (1)] Eq. (1) uses temperature T in the softmax; the value of T and whether it is tuned are not specified in §3.2 or §4.1.
  5. [§2.2–2.3] Related work cites Mix-QViT / Mix-AdaLog and LRP-QViT; a short explicit contrast (isolation KL vs LRP relevance vs Hessian proxies) in §2.3 would help position the contribution.
  6. [Abstract, Figure 1] Minor prose: abstract brace typo “{MixFrag”; “FPCS” and “Adaptive Logarithm Quantization” in Fig. 1 could use a one-line expansion on first use in the main text.

Circularity Check

1 steps flagged

No load-bearing circular derivation; only a mild by-construction ‘validation’ that bit allocation correlates with the fragility costs the optimizer was told to minimize.

specific steps
  1. other [§4.4 Quantization Fragility Analysis; Table 5; also Abstract / §5 claim that analyses ‘validate’ fragility via correlation with bit allocation]
    "Additional analyses validate the proposed fragility metric and demonstrate its strong correlation with the learned bit allocation. ... Under the most challenging W3/A3 budget, the proposed fragility metric exhibits strong positive correlations with the assigned bit-widths across all three measures, achieving Pearson, Spearman, and Kendall coefficients of 0.55, 0.57, and 0.47, respectively."

    Bit-widths b*_c are chosen by MCKP to minimize Σ Ω_i,j x_i,j under the bit budget (Eqs. 2–4), so components with large Ω at low candidate bits are preferentially given higher precision. Reporting a positive correlation between those same Ω(c, low-b) scores and the assigned bits as independent ‘validation’ of the fragility metric is largely by construction of the objective, not a separate empirical test. This affects only the interpretive correlation claim, not ImageNet/COCO accuracy.

full rationale

MixFrag’s chain is empirical systems work, not a closed-form derivation that forces the headline results. Fragility Ω(c,b) is measured as average KL between full-precision and isolated-component quantized logits on a small unlabeled calibration set (Eq. 1); bit-widths are then chosen by MCKP to minimize sum of those measured costs under an average-bit budget (Eqs. 2–4); the mixed-precision model is quantized with AdaLog and scored on ImageNet top-1 and COCO AP. Those test metrics are not algebraic rearrangements of the calibration KL or of the MCKP objective, so the competitive ImageNet numbers and the +9.6 AP COCO claim are not true by construction—they stand or fall on external evaluation (and on protocol parity with cited baselines). There is no self-citation uniqueness theorem, no ansatz smuggled in as external fact, and no fitted scalar renamed as an independent prediction of the same quantity. The only mild circularity is analytic: §4.4/Table 5 present correlation between Ω and the MCKP-assigned bits as validation of the fragility metric, but MCKP assigns higher bits precisely to reduce high-Ω choices, so that correlation is largely expected from the optimizer’s objective rather than independent evidence. That does not underwrite the accuracy/AP claims. Overall circularity is negligible (score 1).

Axiom & Free-Parameter Ledger

4 free parameters · 5 axioms · 1 invented entities

Load-bearing content is methodological and empirical, not theorem-driven. The claim stack rests on (i) isolation-KL as a valid precision-need proxy, (ii) average-bit MCKP as the right global objective, (iii) AdaLog as an adequate heterogeneous-bit backend, and (iv) small unlabeled calibration sets representing deployment behavior. Free knobs are standard PTQ choices (candidate bits, budgets, calibration N, temperature) but they shape the reported allocations. No new physical entities; the main invented construct is the operational “quantization fragility” score Ω(c,b).

free parameters (4)
  • candidate bit-width set B = {2,3,4,6}
    Discrete menu for MCKP groups; paper uses b ∈ {2,3,4,6} (Fig. 1). Choice changes feasible allocations and reported averages.
  • target average bit budget B_target = matched to nominal 3/4/6-bit averages
    Hard constraint in Eq. (4); defines W3/A3, W4/A4, W6/A6 operating points used for all claims.
  • fragility calibration set size N and temperature T = N=32 (fragility fig); T unspecified
    Enter Eq. (1). Fig. 1 states 32 unlabeled images; §4.1 states 1024/256 for PTQ calibration. T is introduced but not numerically specified. Both affect Ω and thus bit maps.
  • which modules are quantizable groups = linear proj layers only
    Only linear projections in attention/MLP are scored and allocated; LayerNorm/residuals stay FP (§3.2–3.4). This design choice strongly shapes both fragility and accuracy.
axioms (5)
  • ad hoc to paper Isolated quantization degradation (one component at bit b, others FP) is a sufficient statistic for that component’s bit need under joint mixed-precision quantization.
    Core of §3.2 and Eq. (1); contradicted in spirit by §4.6 limitation and by Table 7 where re-estimating after joint assignment hurts.
  • ad hoc to paper Minimizing sum of isolation-KL fragilities under an average bit budget (MCKP Eqs. 2–4) is the right proxy for end-task accuracy after AdaLog PTQ.
    Objective never directly optimizes ImageNet top-1 or COCO AP; success is judged only after the fact (§3.3, §4).
  • domain assumption Standard PTQ setting: small unlabeled calibration, pretrained FP weights fixed, no full training-set QAT.
    Stated throughout §1 and §4.1; shared with cited PTQ literature.
  • domain assumption AdaLog’s log-base/scale search and activation handling remain valid when bit-widths differ per linear layer.
    Backend assumption in §3.4; framework claimed backend-agnostic but all results use AdaLog.
  • standard math Dynamic programming solves the stated MCKP to global optimality for the discretized budget.
    Classical MCKP DP (§3.3); correctness of solver is standard if implementation matches formulation.
invented entities (1)
  • Quantization fragility Ω(c,b) as mean KL between FP and isolation-quantized logits no independent evidence
    purpose: Scalar sensitivity used as MCKP item cost so bits flow to components that move outputs most when quantized alone.
    Named and operationalized in §3.2 Eq. (1) as the paper’s distinctive intermediate object, distinct from Hessian/LRP proxies it criticizes. Falsifiable only via correlation/ablation inside the same pipeline (Tables 4–6), not via an external measurable.

pith-pipeline@v1.2.0-daily-grok45 · 22243 in / 4111 out tokens · 88134 ms · 2026-07-31T02:55:07.703428+00:00 · methodology

0 comments
read the original abstract

Post-training quantization (PTQ) has emerged as an effective solution for deploying Vision Transformers (ViTs) on resource-constrained devices. However, existing PTQ methods typically employ uniform bit-widths across transformer components, overlooking their heterogeneous sensitivity to quantization and leading to inefficient precision allocation. In this paper, we propose {MixFrag, a fragility-guided mixed-precision PTQ framework for Vision Transformers. MixFrag first estimates component-level quantization fragility by measuring the Kullback--Leibler (KL) divergence between full-precision and isolated quantized output distributions using a small calibration set. It then formulates bit allocation as a Multiple-Choice Knapsack Problem (MCKP), enabling adaptive layer-wise precision assignment under a target bit budget. Extensive experiments on ImageNet-1K across multiple Vision Transformer architectures demonstrate that MixFrag achieves competitive classification performance under practical mixed-precision settings. Furthermore, evaluations on COCO object detection and instance segmentation show that MixFrag achieves state-of-the-art performance among existing mixed-precision PTQ methods, improving the previous best method by up to 9.6 AP under the challenging MP3/MP3 setting. Additional analyses validate the proposed fragility metric and demonstrate its strong correlation with the learned bit allocation. These results establish MixFrag as an effective framework for mixed-precision post-training quantization of Vision Transformers.

Figures

Figures reproduced from arXiv: 2607.28589 by Md. Mehrab Hossain Opi, Md. Umar Faruk, Robiul Islam Ryad.

Figure 1
Figure 1. Figure 1: (a) Each quantizable component c is independently quantized at candidate bit-width b using AdaLog, and its fragility Ωc,b is computed as the KL divergence between full-precision and quantized logits on 32 unlabeled calibration images, forming the fragility table Ω ∈ R C×B. (b) The fragility table and target bit budget are optimized via an exact Multiple-Choice Knapsack Problem (MCKP) solver to obtain the f… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

55 extracted references · 17 linked inside Pith

  1. [1]

    Cascade r-cnn: Delv- ing into high quality object detection.2018 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 6154–6162, 2017

    Zhaowei Cai and Nuno Vasconcelos. Cascade r-cnn: Delv- ing into high quality object detection.2018 IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 6154–6162, 2017. 6

    2018

  2. [2]

    Comprehensive survey of model com- pression and speed up for vision transformers.ArXiv, abs/2404.10407, 2024

    Feiyang Chen, Ziqian Luo, Lisang Zhou, Xueting Pan, and Ying Jiang. Comprehensive survey of model com- pression and speed up for vision transformers.ArXiv, abs/2404.10407, 2024. 1

    Pith/arXiv arXiv 2024

  3. [3]

    Dearkd: Data-efficient early knowledge distillation for vision transformers.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12042–12052, 2022

    Xianing Chen, Qiong Cao, Yujie Zhong, Jing Zhang, Shenghua Gao, and Dacheng Tao. Dearkd: Data-efficient early knowledge distillation for vision transformers.2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 12042–12052, 2022. 1

    2022

  4. [4]

    Gopalakrishnan

    Jungwook Choi, Zhuo Wang, Swagath Venkataramani, Pierce I-Jen Chuang, Vijayalakshmi Srinivasan, and K. Gopalakrishnan. Pact: Parameterized clipping activation for quantized neural networks.ArXiv, abs/1805.06085, 2018. 1

    Pith/arXiv arXiv 2018

  5. [5]

    Dynamic detr: End-to-end object detection with dynamic attention.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2968–2977, 2021

    Xiyang Dai, Yinpeng Chen, Jianwei Yang, Pengchuan Zhang, Lu Yuan, and Lei Zhang. Dynamic detr: End-to-end object detection with dynamic attention.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 2968–2977, 2021. 1

    2021

  6. [6]

    Towards accurate post-training quantization for vision transformer.Proceed- ings of the 30th ACM International Conference on Multime- dia, 2022

    Yifu Ding, Haotong Qin, Qing-Yu Yan, Zhenhua Chai, Jun- jie Liu, Xiaolin Wei, and Xianglong Liu. Towards accurate post-training quantization for vision transformer.Proceed- ings of the 30th ACM International Conference on Multime- dia, 2022. 2, 3

    2022

  7. [7]

    Packqvit: Faster sub-8-bit vi- sion transformers via full and packed quantization on the mo- bile.Advances in Neural Information Processing Systems 36,

    Peiyan Dong, Lei Lu, Chao Wu, Cheng Lyu, Geng Yuan, Hao Tang, and Yanzhi Wang. Packqvit: Faster sub-8-bit vi- sion transformers via full and packed quantization on the mo- bile.Advances in Neural Information Processing Systems 36,

  8. [8]

    Low-rank rescaled vision transformer fine-tuning: A residual design approach.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16101–16110, 2024

    Wei Dong, Xing Zhang, Bihui Chen, Dawei Yan, Zhijun Lin, Qingsen Yan, Peng Wang, and Yang Yang. Low-rank rescaled vision transformer fine-tuning: A residual design approach.2024 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 16101–16110, 2024. 1

    2024

  9. [9]

    Ma- honey, and Kurt Keutzer

    Zhen Dong, Zhewei Yao, Amir Gholami, Michael W. Ma- honey, and Kurt Keutzer. HAWQ: Hessian aware quantiza- tion of neural networks with mixed-precision. InIEEE/CVF International Conference on Computer Vision (ICCV), 2019. 3

    2019

  10. [10]

    Hawq-v2: Hessian aware trace-weighted quantization of neural networks.Ad- vances in neural information processing systems, 33:18518– 18529, 2020

    Zhen Dong, Zhewei Yao, Daiyaan Arfeen, Amir Gholami, Michael W Mahoney, and Kurt Keutzer. Hawq-v2: Hessian aware trace-weighted quantization of neural networks.Ad- vances in neural information processing systems, 33:18518– 18529, 2020. 3

    2020

  11. [11]

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

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Representa- tions (ICLR), 2021. 1, 6

    2021

  12. [12]

    Model quantiza- tion and hardware acceleration for vision transformers: A comprehensive survey.ArXiv, abs/2405.00314, 2024

    Dayou Du, Gu Gong, and Xiaowen Chu. Model quantiza- tion and hardware acceleration for vision transformers: A comprehensive survey.ArXiv, abs/2405.00314, 2024. 1

    Pith/arXiv arXiv 2024

  13. [13]

    Jumping through local minima: Quantization in the loss landscape of vision transformers

    Natalia Frumkin, Dibakar Gope, and Diana Marculescu. Jumping through local minima: Quantization in the loss landscape of vision transformers. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 16978–16988, 2023. 3

    2023

  14. [14]

    Mahoney, and Kurt Keutzer

    Amir Gholami, Sehoon Kim, Zhen Dong, Zhewei Yao, Michael W. Mahoney, and Kurt Keutzer. A survey of quanti- zation methods for efficient neural network inference.ArXiv, abs/2103.13630, 2021. 1

    Pith/arXiv arXiv 2021

  15. [15]

    Knowl- edge distillation in vision transformers: A critical review

    Gousia Habib, Tausifa Jan Saleem, and Brejesh Lall. Knowl- edge distillation in vision transformers: A critical review. ArXiv, abs/2302.02108, 2023. 1

    Pith/arXiv arXiv 2023

  16. [16]

    Zhang, Shaoqing Ren, and Jian Sun

    Kaiming He, X. Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition.2016 IEEE Confer- ence on Computer Vision and Pattern Recognition (CVPR), pages 770–778, 2015. 1

    2016

  17. [17]

    Mask r-cnn

    Kaiming He, Georgia Gkioxari, Piotr Doll ´ar, and Ross Gir- shick. Mask r-cnn. In2017 IEEE International Conference on Computer Vision (ICCV), pages 2980–2988, 2017. 6

    2017

  18. [18]

    Bivit: Extremely compressed binary vision transformers

    Yefei He, Zhenyu Lou, Luoming Zhang, Jing Liu, Weijia Wu, Hong Zhou, and Bohan Zhuang. Bivit: Extremely compressed binary vision transformers. InProceedings of the IEEE/CVF international conference on computer vision, pages 5651–5663, 2023. 2

    2023

  19. [19]

    Lampq: Towards accurate layer- wise mixed precision quantization for vision transformers

    Minjun Kim, Jaeri Lee, Jongjin Kim, Jeongin Yun, Young- bum Kwon, and U Kang. Lampq: Towards accurate layer- wise mixed precision quantization for vision transformers. In AAAI Conference on Artificial Intelligence, 2025. 2, 3, 6, 7, 8

    2025

  20. [20]

    Brecq: Pushing the limit of post-training quantization by block reconstruc- tion.arXiv preprint arXiv:2102.05426, 2021

    Yuhang Li, Ruihao Gong, Xu Tan, Yang Yang, Peng Hu, Qi Zhang, Fengwei Yu, Wei Wang, and Shi Gu. Brecq: Pushing the limit of post-training quantization by block reconstruc- tion.arXiv preprint arXiv:2102.05426, 2021. 2

    Pith/arXiv arXiv 2021

  21. [21]

    Q-vit: Accurate and fully quan- tized low-bit vision transformer.ArXiv, abs/2210.06707,

    Yanjing Li, Sheng Xu, Baochang Zhang, Xianbin Cao, Pen- glei Gao, and Guodong Guo. Q-vit: Accurate and fully quan- tized low-bit vision transformer.ArXiv, abs/2210.06707,

    Pith/arXiv arXiv

  22. [22]

    I-vit: Integer-only quantization for efficient vision transformer inference

    Zhikai Li and Qingyi Gu. I-vit: Integer-only quantization for efficient vision transformer inference. InProceedings of the IEEE/CVF International Conference on Computer Vi- sion, pages 17065–17075, 2023. 2

    2023

  23. [23]

    Repq- vit: Scale reparameterization for post-training quantization 11 of vision transformers.2023 IEEE/CVF International Con- ference on Computer Vision (ICCV), pages 17181–17190,

    Zhikai Li, Junrui Xiao, Lianwei Yang, and Qingyi Gu. Repq- vit: Scale reparameterization for post-training quantization 11 of vision transformers.2023 IEEE/CVF International Con- ference on Computer Vision (ICCV), pages 17181–17190,

    2023

  24. [24]

    Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C

    Tsung-Yi Lin, Michael Maire, Serge J. Belongie, James Hays, Pietro Perona, Deva Ramanan, Piotr Doll ´ar, and C. Lawrence Zitnick. Microsoft coco: Common objects in context. InEuropean Conference on Computer Vision, 2014. 6

    2014

  25. [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. InInternational Joint Confer- ence on Artificial Intelligence, 2021. 2, 3, 4

    2021

  26. [26]

    Pd-quant: Post-training quantiza- tion based on prediction difference metric.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24427–24437, 2022

    Jiawei Liu, Lin Niu, Zhihang Yuan, Dawei Yang, Xinggang Wang, and Wenyu Liu. Pd-quant: Post-training quantiza- tion based on prediction difference metric.2023 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 24427–24437, 2022. 2

    2023

  27. [27]

    Swin transformer: Hierarchical vision transformer using shifted windows

    Ze Liu, Yutong Lin, Yue Cao, Han Hu, Yixuan Wei, Zheng Zhang, Stephen Lin, and Baining Guo. Swin transformer: Hierarchical vision transformer using shifted windows. In IEEE/CVF International Conference on Computer Vision (ICCV), 2021. 1, 6

    2021

  28. [28]

    Post-training quantization for vision transformer

    Zhenhua Liu, Yunhe Wang, Kai Han, Siwei Ma, and Wen Gao. Post-training quantization for vision transformer. ArXiv, abs/2106.14156, 2021. 1, 2

    Pith/arXiv arXiv 2021

  29. [29]

    Similarity-aware fast low-rank decomposition frame- work for vision transformers.2024 IEEE International Sym- posium on Circuits and Systems (ISCAS), pages 1–5, 2024

    Yuan-June Luo, Yu-Shan Tai, Ming-Guang Lin, and An-Yeu Wu. Similarity-aware fast low-rank decomposition frame- work for vision transformers.2024 IEEE International Sym- posium on Circuits and Systems (ISCAS), pages 1–5, 2024. 1

    2024

  30. [30]

    Ptq4sam: Post-training quantization for segment anything.2024 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 15941– 15951, 2024

    Chengtao Lv, Hong Chen, Jinyang Guo, Yifu Ding, and Xianglong Liu. Ptq4sam: Post-training quantization for segment anything.2024 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition (CVPR), pages 15941– 15951, 2024. 1

    2024

  31. [31]

    Up or down? adap- tive rounding for post-training quantization

    Markus Nagel, Rana Ali Amjad, Mart Van Baalen, Chris- tos Louizos, and Tijmen Blankevoort. Up or down? adap- tive rounding for post-training quantization. InInternational conference on machine learning, pages 7197–7206. PMLR,

  32. [32]

    Mix-qvit: Mixed- precision vision transformer quantization driven by layer importance and quantization sensitivity.arXiv preprint arXiv:2501.06357, 2025

    Navin Ranjan and Andreas Savakis. Mix-qvit: Mixed- precision vision transformer quantization driven by layer importance and quantization sensitivity.arXiv preprint arXiv:2501.06357, 2025. 3, 7

    Pith/arXiv arXiv 2025

  33. [33]

    Navin Ranjan and Andreas E. Savakis. Lrp-qvit: Mixed- precision vision transformer quantization using layer impor- tance score.2025 25th International Conference on Digital Signal Processing (DSP), pages 1–5, 2025. 2, 3, 8

    2025

  34. [34]

    Bernstein, Alexander C

    Olga Russakovsky, Jia Deng, Hao Su, Jonathan Krause, San- jeev Satheesh, Sean Ma, Zhiheng Huang, Andrej Karpathy, Aditya Khosla, Michael S. Bernstein, Alexander C. Berg, and Li Fei-Fei. Imagenet large scale visual recognition chal- lenge.International Journal of Computer Vision, 115:211 – 252, 2014. 6

    2014

  35. [35]

    Vision transformers on the edge: A comprehensive survey of model compression and acceleration strategies.Neurocomputing, 643:130417, 2025

    Shaibal Saha and Lanyu Xu. Vision transformers on the edge: A comprehensive survey of model compression and acceleration strategies.Neurocomputing, 643:130417, 2025. 1

    2025

  36. [36]

    Notes on kullback-leibler divergence and likelihood.arXiv preprint arXiv:1404.2000, 2014

    Jonathon Shlens. Notes on kullback-leibler divergence and likelihood.arXiv preprint arXiv:1404.2000, 2014. 6

    Pith/arXiv arXiv 2000

  37. [37]

    Very deep convo- lutional networks for large-scale image recognition.CoRR, abs/1409.1556, 2014

    Karen Simonyan and Andrew Zisserman. Very deep convo- lutional networks for large-scale image recognition.CoRR, abs/1409.1556, 2014. 1

    Pith/arXiv arXiv 2014

  38. [38]

    Cp-vit: Cascade vision trans- former pruning via progressive sparsity prediction.ArXiv, abs/2203.04570, 2022

    Zhuoran Song, Yihong Xu, Zhezhi He, Li Jiang, Naifeng Jing, and Xiaoyao Liang. Cp-vit: Cascade vision trans- former pruning via progressive sparsity prediction.ArXiv, abs/2203.04570, 2022. 1

    Pith/arXiv arXiv 2022

  39. [39]

    Segmenter: Transformer for semantic seg- mentation.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7242–7252, 2021

    Robin Strudel, Ricardo Garcia Pinel, Ivan Laptev, and Cordelia Schmid. Segmenter: Transformer for semantic seg- mentation.2021 IEEE/CVF International Conference on Computer Vision (ICCV), pages 7242–7252, 2021. 1

    2021

  40. [40]

    A survey on transformer compression.ArXiv, abs/2402.05964, 2024

    Yehui Tang, Yunhe Wang, Jianyuan Guo, Zhijun Tu, Kai Han, Hailin Hu, and Dacheng Tao. A survey on transformer compression.ArXiv, abs/2402.05964, 2024. 1

    Pith/arXiv arXiv 2024

  41. [41]

    Training data-efficient image transformers & distillation through at- tention

    Hugo Touvron, Matthieu Cord, Matthijs Douze, Francisco Massa, Alexandre Sablayrolles, and Herv ´e J´egou. Training data-efficient image transformers & distillation through at- tention. InInternational conference on machine learning, pages 10347–10357. PMLR, 2021. 6

    2021

  42. [42]

    Haq: Hardware-aware automated quantization with mixed precision

    Kuan Wang, Zhijian Liu, Yujun Lin, Ji Lin, and Song Han. Haq: Hardware-aware automated quantization with mixed precision. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 8612–8620,

  43. [43]

    Qdrop: Randomly dropping quantiza- tion for extremely low-bit post-training quantization.ArXiv, abs/2203.05740, 2022

    Xiuying Wei, Ruihao Gong, Yuhang Li, Xianglong Liu, and Fengwei Yu. Qdrop: Randomly dropping quantiza- tion for extremely low-bit post-training quantization.ArXiv, abs/2203.05740, 2022. 2

    Pith/arXiv arXiv 2022

  44. [44]

    AdaLog: Post-training quantization for vi- sion transformers with adaptive logarithm quantizer

    Zhuguanyu Wu, Jiaxin Chen, Hanwen Zhong, Di Huang, and Yunhong Wang. AdaLog: Post-training quantization for vi- sion transformers with adaptive logarithm quantizer. InEu- ropean Conference on Computer Vision (ECCV), 2024. 2, 3, 4, 6, 7, 8

    2024

  45. [45]

    Zhuguanyu Wu, Shihe Wang, Jiayi Zhang, Jiaxin Chen, and Yunhong Wang. Fima-q: Post-training quantization for vi- sion transformers by fisher information matrix approxima- tion.2025 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 14891–14900, 2025. 6, 7

    2025

  46. [46]

    Dopq-vit: Towards distribution-friendly and outlier-aware post-training quantization for vision transformers.ArXiv, abs/2408.03291,

    Lianwei Yang, Haisong Gong, and Qingyi Gu. Dopq-vit: Towards distribution-friendly and outlier-aware post-training quantization for vision transformers.ArXiv, abs/2408.03291,

    Pith/arXiv arXiv

  47. [47]

    Vitkd: Feature-based knowledge distillation for vision transformers.2024 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition Workshops (CVPRW), pages 1379–1388, 2024

    Zhendong Yang, Zhe Li, Ailing Zeng, Zexian Li, Chun Yuan, and Yu Li. Vitkd: Feature-based knowledge distillation for vision transformers.2024 IEEE/CVF Conference on Com- puter Vision and Pattern Recognition Workshops (CVPRW), pages 1379–1388, 2024. 1

    2024

  48. [48]

    V-pruner: A fast and globally-informed token pruning framework for vision transformer

    Guangzhen Yao, Jiayu Zheng, Zezhou Wang, Wenxin Zhang, Renda Han, Chuangxin Zhao, Zeyu Zhang, and Runhao Liu. V-pruner: A fast and globally-informed token pruning framework for vision transformer. InAAAI Conference on Artificial Intelligence, 2026. 1

    2026

  49. [49]

    Prabhu, and Siyuan Chen

    Eric Youn, J SaiMitheran, Sanjana S. Prabhu, and Siyuan Chen. Compressing vision transformers for low-resource vi- sual learning.ArXiv, abs/2309.02617, 2023. 1 12

    Pith/arXiv arXiv 2023

  50. [50]

    Width & depth pruning for vision transformers

    Fang Yu, Kun Huang, Meng Wang, Yuan Cheng, Wei Chu, and Li Cui. Width & depth pruning for vision transformers. InAAAI Conference on Artificial Intelligence, 2022. 1

    2022

  51. [51]

    PTQ4ViT: Post-training quantization for vi- sion transformers with twin uniform quantization

    Zhihang Yuan, Chenhao Xue, Yiqi Chen, Qiang Wu, and Guangyu Sun. PTQ4ViT: Post-training quantization for vi- sion transformers with twin uniform quantization. InEuro- pean Conference on Computer Vision (ECCV), 2022. 3, 4

    2022

  52. [52]

    Cat-det: Con- trastively augmented transformer for multi-modal 3d object detection

    Yanan Zhang, Jiaxin Chen, and Di Huang. Cat-det: Con- trastively augmented transformer for multi-modal 3d object detection. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 908–917,

  53. [53]

    I&s-vit: An inclusive & stable method for post-training vits quantization.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 48(2):1063–1080,

    Yunshan Zhong, Jiawei Hu, Mingbao Lin, Mengzhao Chen, and Rongrong Ji. I&s-vit: An inclusive & stable method for post-training vits quantization.IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 48(2):1063–1080,

  54. [54]

    Octr: Octree-based transformer for 3d object detection

    Chao Zhou, Yanan Zhang, Jiaxin Chen, and Di Huang. Octr: Octree-based transformer for 3d object detection. 2023 IEEE/CVF Conference on Computer Vision and Pat- tern Recognition (CVPR), pages 5166–5175, 2023. 1

    2023

  55. [55]

    Deformable detr: Deformable transform- ers for end-to-end object detection.ArXiv, abs/2010.04159,

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable transform- ers for end-to-end object detection.ArXiv, abs/2010.04159,

    Pith/arXiv arXiv 2010