Pith. sign in

REVIEW 5 major objections 5 minor 48 references

LARGO: Low-Rank Regulated Gradient Projection for Robust Parameter Efficient Fine-Tuning

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

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

Pith's one-line read This paper claims that adding trainable per-matrix projection radii and SVD-based initialization to LoRA yields OOD accuracy above full fine-tuning on DomainNet and ImageNet with under a million trainable parameters.

desk verdict The empirical hunch is plausible but the paper does not specify a correct or reproducible method: the key update rule is algebraically inconsistent and the gradient formulas are malformed. read the letter →

arxiv 2506.12394 v1 pith:RLGEL57G submitted 2025-06-14 cs.CV cs.AI

Haotian Zhang , Liu Liu , Baosheng Yu , Jiayan Qiu , Yanwei Ren , Xianglong Liu This is my paper
classification cs.CVcs.AI
keywords parameter-efficientfine-tuninglow-rankadaptationLoRAout-of-distributionrobustnessgradientprojectiondomainshiftsingularvaluedecompositionvisualquestionanswering
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

LARGO is a parameter-efficient fine-tuning method built on LoRA. Its proposal is that robustness to distribution shift can be regulated locally, at the level of each low-rank matrix, by two trainable scalars per layer that scale the LoRA update, rather than by global or layer-wise weight projections. The paper reports that on a ViT-S backbone this reaches out-of-distribution averages of 34.26 on DomainNet and 34.97 on ImageNet with 0.42M and 0.68M trainable parameters, above full fine-tuning (33.37 and 33.70), LoRA (30.31 and 31.12), and TPGM (32.76 and 30.50), with competitive in-domain accuracy; a VQA experiment on a 3B vision-language model gives an OOD average of 42.55 versus 40.84 for LoRA. The authors also note that their hyperparameter search was non-exhaustive and that only LoRA-family adaptation, not adapters or prompt tuning, was tested.

What carries the argument

The load-bearing object is a pair of per-matrix trainable projection radii, $\gamma^a_t$ and $\gamma^b_t$, attached to the two LoRA factors $A$ and $B$ of each layer. During training, LARGO first computes the unconstrained low-rank updates $\tilde A_t$ and $\tilde B_t$, then rescales their product by $\gamma^a_t \gamma^b_t /(\|\tilde A_t\|_1 \|\tilde B_t\|_1)$ before adding it to $W_0$. The two scalars receive gradient flow through this rescaling, so each layer learns the size of its own update; this is the mechanism claimed to preserve inter-layer independence and pretrained-model robustness. The supporting object is the SVD initialization, which sets $A_0$ and $B_0$ from the leading singular vectors of $W_0$ scaled by a factor derived from $s_r = r/\|S_r\|_2^2$, so the starting update is near zero and the pretrained representation is not perturbed at initialization.

What would settle it

Compute, for any trained LARGO layer, $\|\tilde A_t \tilde B_t\|_1$ and $\|\tilde A_t\|_1 \|\tilde B_t\|_1$ and compare them with the update actually applied; because matrix norms are submultiplicative rather than multiplicative, they will generically differ, which would show that the implemented normalization is not the one described by Eq. (4.1.4). A complementary experiment is to retrain with $\gamma^a=\gamma^b=1$ and Kaiming initialization; if the OOD gains persist, the scalars and SVD initialization are not necessary for the reported robustness.

Watch

Extended reading notes

Core claim

The paper's central claim is that the right place to impose robustness constraints in PEFT is inside the LoRA update itself. Concretely, after the unconstrained low-rank factors $\tilde A_t$ and $\tilde B_t$ are computed, LARGO produces the layer update $W_t = W_0 + (\tilde A_t \tilde B_t)\frac{\gamma^a_t \gamma^b_t}{\|\tilde A_t\|_1\|\tilde B_t\|_1}$, with per-layer scalars $\gamma^a_t$ and $\gamma^b_t$ learned by gradient descent. The design intent is to keep the fine-tuned weights near the pretrained ones while allowing each layer to choose its own update budget independently, avoiding the inter-layer coupling that makes projection baselines such as TPGM costly. The second component is an SVD-based initialization that starts the LoRA factors from scaled leading singular vectors of $W_0$, so that the initial update is close to zero. On this machinery the paper attributes its reported OOD gains, including DomainNet OOD average 34.26, ImageNet OOD average 34.97, and VQA OOD average 42.55, all with parameter counts at or below 0.68M.

Load-bearing premise

The load-bearing premise is that $\|\tilde A_t \tilde B_t\|_1 = \|\tilde A_t\|_1 \|\tilde B_t\|_1$, so the normalized update in Eq. (4.1.5) matches the one defined in Eq. (4.1.4); this equality is not an identity for the L1 norm, and if it fails the implemented update rule differs from the rule the paper derives.

Editorial extensions

If this is right

  • If the reported numbers hold, LARGO delivers OOD averages of 34.26 on DomainNet and 34.97 on ImageNet with 0.42M and 0.68M trainable parameters, beating full fine-tuning, LoRA, and TPGM on those benchmarks.
  • The method transfers to multimodal fine-tuning: on a 3B vision-language model trained on 10% of VQAv2, it reports 65.29 in-domain accuracy and 42.55 OOD average, above the LoRA baseline's 64.87 and 40.84.
  • The ablation results imply that the initial value of $\gamma$ matters: smaller values such as $10^{-8}$ give the best OOD averages on the 10% and 20% DomainNet splits, while the optimal value for the 50% split is $10^{-6}$.
  • The SVD scalar ablation suggests that an intermediate scaling factor of 0.5 balances OOD and in-domain accuracy across all data splits, so the initialization scheme contributes to the gains rather than being neutral.

Reading between the lines

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

  • Editorial inference: because $\|\tilde A_t \tilde B_t\|_1$ is generally smaller than $\|\tilde A_t\|_1 \|\tilde B_t\|_1$, the update implemented by Eq. (4.1.5) is not a projection onto the stated L1 ball; its practical effect may be closer to learned per-factor learning-rate scaling, an explanation the paper does not test.
  • Editorial inference: a direct comparison of LARGO against per-layer or per-matrix learning-rate schedules, with the same SVD initialization, would isolate whether the gains come from the constrained-update story or simply from more flexible optimization.
  • Editorial inference: applying the same scalar-regulated normalization to adapters or prompt tuning would test whether the mechanism generalizes beyond the LoRA parameterization; the paper's limitation note reports that this was not done.
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 / 5 minor

Summary. The paper proposes LARGO, a parameter-efficient fine-tuning method that augments LoRA with trainable scalar projection radii (gamma_a, gamma_b) per low-rank matrix and an SVD-based initialization of the LoRA factors. The authors claim state-of-the-art out-of-distribution (OOD) accuracy on DomainNet and ImageNet-family benchmarks using ViT-S with roughly 0.42M--0.68M trainable parameters, and additional gains on a PaliGemma-3B VQA setup. The central contribution is presented as a dynamical constraint mechanism that regulates layer-wise updates independently while preserving pretrained knowledge. The manuscript contains the proposed update equations, gradient formulas for the gamma scalars, an SVD initialization scheme, and extensive tables of results across image and vision-language benchmarks, but no code or error bars are provided.

Significance. If the method were well defined and the results reproducible, the idea of per-matrix trainable projection radii for LoRA would be a useful and low-overhead addition to the PEFT toolbox, and the reported parameter counts are attractive. The experimental coverage is broad, spanning DomainNet, ImageNet distribution-shift benchmarks, and multiple VQA OOD sets. However, the central mathematical specification is internally inconsistent, the exact update rule is ambiguous, and the empirical claims are not accompanied by code, error bars, or a consistent hyperparameter configuration. The strengths of the paper cannot compensate for the absence of a precisely specified, reproducible algorithm.

major comments (5)
  1. [Section 4.1.1, Eqs. (4.1.4) and (4.1.5)] The two displayed update rules are not equivalent. Eq. (4.1.4) has denominator ||bW_t - W0||_1 = ||bA_t bB_t||_1, while Eq. (4.1.5) uses ||bA_t||_1 ||bB_t||_1. The required identity ||AB||_1 = ||A||_1||B||_1 is false in general; for example, A = I_2 and B = [[1,1],[1,1]] gives ||AB||_1 = 4 but ||A||_1||B||_1 = 8. Since every experimental result depends on this update, the manuscript must state exactly which denominator was implemented, or the method is not well defined.
  2. [Section 4.1.2, Eqs. (4.1.6)-(4.1.8)] The gamma-gradient formulas contain an undefined scalar factor t and are not derivable from Eq. (4.1.5). A derivative of W_t = (bA_t bB_t) gamma_a gamma_b / (||bA_t||_1 ||bB_t||_1) + W0 with respect to gamma_a would involve the full denominator and both low-rank factors, not the displayed expression. Moreover, the right-hand side of Eq. (4.1.6) appears to be matrix-valued while the left-hand side is a scalar. The update rule for gamma is therefore not specified in an implementable or verifiable form.
  3. [Section 4.2, Eqs. (4.2.1)-(4.2.3)] The SVD initialization is not mathematically coherent. Eq. (4.2.2) introduces an undefined factor t and does not contain the scalar s, while Eq. (4.2.3) introduces s and a different expression, A0 <- s t^2 U_r sqrt(r)/||S_r||_2. Combining Eqs. (4.2.1) and (4.2.2) gives A0 B0 = t^2 (r/||S_r||_2^2) U_r V_r^T, which is neither equal to Eq. (4.2.3) nor clearly a low-rank approximation of W0. The initialization used in the experiments needs a single precise definition.
  4. [Section 5.4 and Appendix A] The experimental reporting is internally inconsistent. Table 4 lists LARGO hyperparameters as gamma=1e-4 and svd_scalar=0.5, while Section 5.4 states that the best gamma is 1e-8. Table 7 shows that for the 50% split, gamma=1e-4 with svd_scalar=0.5 gives OOD average 35.11 and gamma=1e-8 gives 34.85, whereas the main Table 1 reports LARGO OOD average 34.26. In Appendix Table 5, for the 10% split with gamma=1e-8, Kaiming initialization gives OOD average 30.85 while SVD with scalar 0.5 gives 30.35, contradicting the text's claim that SVD initialization outperforms Kaiming. These discrepancies make it impossible to determine which configuration produced the headline numbers.
  5. [Tables 1, 2, and 8] No error bars, no number of seeds, and no code are provided, and the statement that the source code 'will be released soon' is not a substitute for reproducibility. Combined with the definitional issues in Section 4, the central empirical claim that LARGO achieves state-of-the-art OOD accuracy cannot be checked from the manuscript as submitted.
minor comments (5)
  1. [Section 3, Notation] The paper states that ||X||_1 denotes the L1 norm but does not specify whether it is the entrywise L1 norm, the matrix-induced L1 norm, or a row-wise norm. This matters because Eqs. (4.1.3)-(4.1.5) rely on different possible interpretations, and Eq. (3.2.4) is explicitly row-wise.
  2. [Section 4.2, Eq. (4.2.2)] The symbol t is used both as an epoch index and as an undefined scalar in Eq. (4.2.2). Please use a distinct name for the initialization scale, and define it before it appears.
  3. [References] Reference [20] and reference [21] appear to be the same paper, and references [41] and [42] also appear to be the same paper. Please unify the citations.
  4. [Throughout] There are numerous typos and awkward phrasings, including 'initialze', 'Morevoer', 'We proposedLARGO', and 'Initiazation'. A careful proofreading pass is needed.
  5. [Section 6, Limitations] The limitations paragraph candidly states that the hyperparameter exploration is non-exhaustive and that adapters and prompt tuning were not evaluated; this is appropriate, but it qualifies the abstract's claim of 'minimal hyperparameter tuning' and 'state-of-the-art performance' across PEFT variants.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: LARGO's update rule is an empirical construction, not a derivation from its target results; the Eq. (4.1.4)/(4.1.5) discrepancy is a mathematical consistency defect, not circular reasoning.

full rationale

This paper does not exhibit circular reasoning. LARGO's update rule (Eqs. 4.1.4-4.1.9) is proposed as an algorithmic construction, not as a quantity derived from the performance numbers it later reports. The reported OOD accuracies (Tables 1, 2, and 8) are measurements on external held-out benchmarks (DomainNet, the ImageNet series, and OOD VQA sets), and the scalars gamma_a, gamma_b and the SVD scaling parameter are trainable parameters or ordinary hyperparameters, not constants fitted to the OOD results. The ablation study (Tables 3, 5-7, 9-14) varies gamma and svd_scalar and reports all configurations, which is hyperparameter exploration rather than post hoc fitting disguised as prediction. The paper cites prior work (LoRA, TPGM, FTP, MARS), but none of these citations are self-citations by the present authors, and the load-bearing comparisons are external baselines rather than a self-citation chain. One nontrivial mathematical issue exists: the claimed equivalence between Eq. (4.1.4) and Eq. (4.1.5) requires ||A_t B_t||_1 = ||A_t||_1 ||B_t||_1, which is not an identity for the L1 norm; likewise, Eqs. (4.1.6)-(4.1.8) appear to contain an undefined factor t and omitted normalization factors. However, these are correctness and reproducibility defects in the algorithm specification, not circularity: the update rule would still be an input to the experiments, not an output of them. There is no step in which a reported prediction is equivalent by construction to an input, and no uniqueness theorem or self-citation is invoked to make the method's choice forced. The inherent limitations note explicitly states that hyperparameter exploration is non-exhaustive, which further supports that gamma values are tuned choices rather than fitted constants derived from the target results. Overall, the central claim is empirical and self-contained against external benchmarks, so the circularity score is 0.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

No new unobserved entities are introduced. The load is carried by two hand-tuned or learned scalar families and two unproved structural assumptions about norms and SVD scaling.

free parameters (3)
  • Initial projection radii gamma_a, gamma_b = 1e-4 for DomainNet main results, 1e-8 in ablations; final trained values not reported
    Per-matrix learnable controls of update magnitude; this is the core mechanism of LARGO.
  • SVD scalar s (svd_scalar) = 0.5 for main results; 0.1 and 1.0 ablated
    Manual scaling chosen by validation performance; directly controls the claimed deviation from pretrained weights.
  • LoRA rank r = 16
    Standard hyperparameter used for all methods; determines the number of trainable parameters but is not the central innovation.
assumptions (4)
  • ad hoc to paper L1 norm of a product equals the product of L1 norms, ||bA_t bB_t||_1 = ||bA_t||_1 ||bB_t||_1
    Needed for Eq. (4.1.4) to be equivalent to Eq. (4.1.5); generally false, only submultiplicativity holds.
  • ad hoc to paper Scalar-scaled SVD matrices A0 = U_r sqrt(sr), B0 = sqrt(sr) V_r^T approximate the top singular structure of W0
    Used in Eqs. (4.2.2) to (4.2.3); when sr is a scalar, A0B0 equals sr U_r V_r^T, which does not reproduce U_r S_r V_r^T unless all top singular values are equal.
  • domain assumption Staying close to pretrained weights in row-wise L1 distance preserves OOD robustness
    Inherited from MARS and TPGM and used to justify the constraints in Eq. (4.1.3); no new proof is supplied.
  • domain assumption Independent per-layer constraints improve OOD robustness without inter-layer gradient dependencies
    Central motivation repeated in Sections 4 and 6; no theoretical or controlled experimental support isolates this effect.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LARGO: Low-Rank Regulated Gradient Projection for Robust Parameter Efficient Fine-Tuning." pith.science (2026). https://pith.science/paper/RLGEL57G

@misc{pith2026250612394,
  author       = {Pith},
  title        = {Pith review of: LARGO: Low-Rank Regulated Gradient Projection for Robust Parameter Efficient Fine-Tuning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RLGEL57G}},
  note         = {Machine review of arXiv:2506.12394}
}
read the original abstract

The advent of parameter-efficient fine-tuning methods has significantly reduced the computational burden of adapting large-scale pretrained models to diverse downstream tasks. However, existing approaches often struggle to achieve robust performance under domain shifts while maintaining computational efficiency. To address this challenge, we propose Low-rAnk Regulated Gradient Projection (LARGO) algorithm that integrates dynamic constraints into low-rank adaptation methods. Specifically, LARGO incorporates parallel trainable gradient projections to dynamically regulate layer-wise updates, retaining the Out-Of-Distribution robustness of pretrained model while preserving inter-layer independence. Additionally, it ensures computational efficiency by mitigating the influence of gradient dependencies across layers during weight updates. Besides, through leveraging singular value decomposition of pretrained weights for structured initialization, we incorporate an SVD-based initialization strategy that minimizing deviation from pretrained knowledge. Through extensive experiments on diverse benchmarks, LARGO achieves state-of-the-art performance across in-domain and out-of-distribution scenarios, demonstrating improved robustness under domain shifts with significantly lower computational overhead compared to existing PEFT methods. The source code will be released soon.

Figures

Figures reproduced from arXiv: 2506.12394 by the authors.

Figure 1
Figure 1. Overview of the LARGO method. The process computes the gradient for whole LARGOlayers, followed by parallel updates of the low-rank matrices A and B and their respective trainable scalars γ a and γ b . This constraint update mechanism dynamically modulates weight adjustments, ensuring efficient and balanced adaptation while preserving pretrained robustness. Trainable Projected Gradient Method: TPGM framework introdu… view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

48 extracted references · 34 canonical work pages

  1. [1]

    Bałazy, M

    K. Bałazy, M. Banaei, K. Aberer, and J. Tabor. Lora-xs: Low-rank adaptation with extremely small number of parameters.arXiv preprint arXiv:2405.17604, 2024

    arXiv 2024

  2. [2]

    Beyer, A

    L. Beyer, A. Steiner, A. S. Pinto, A. Kolesnikov, X. Wang, D. Salz, M. Neumann, I. Alabdulmohsin, M. Tschan- nen, E. Bugliarello, et al. Paligemma: A versatile 3b vlm for transfer.arXiv preprint arXiv:2407.07726, 2024

    arXiv 2024

  3. [3]

    Biderman, J

    D. Biderman, J. Portes, J. J. G. Ortiz, M. Paul, P. Greengard, C. Jennings, D. King, S. Havens, V . Chiley, J. Frankle, et al. Lora learns less and forgets less.arXiv preprint arXiv:2405.09673, 2024

    arXiv 2024

  4. [4]

    J. P. Bigham, C. Jayant, H. Ji, G. Little, A. Miller, R. C. Miller, R. Miller, A. Tatarowicz, B. White, S. White, et al. Vizwiz: nearly real-time answers to visual questions. InProceedings of the 23nd annual ACM symposium on User interface software and technology, pages 333–342, 2010

    work page 2010

  5. [5]

    C.-F. R. Chen, Q. Fan, and R. Panda. Crossvit: Cross-attention multi-scale vision transformer for image classification. InProceedings of the IEEE/CVF international conference on computer vision, pages 357–366, 2021

    work page 2021

  6. [6]

    Dancette, R

    C. Dancette, R. Cadene, D. Teney, and M. Cord. Beyond question-based biases: Assessing multimodal shortcut learning in visual question answering. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 1574–1583, 2021

    work page 2021

  7. [7]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. Imagenet: A large-scale hierarchical image database. In2009 IEEE conference on computer vision and pattern recognition, pages 248–255. Ieee, 2009

    2009

  8. [8]

    N. Ding, Y . Qin, G. Yang, F. Wei, Z. Yang, Y . Su, S. Hu, Y . Chen, C.-M. Chan, W. Chen, et al. Parameter- efficient fine-tuning of large-scale pre-trained language models.Nature Machine Intelligence, 5(3):220–235, 2023

    work page 2023

Show all 48 references
  1. [9]

    Dosovitskiy, L

    A. Dosovitskiy, L. Beyer, A. Kolesnikov, D. Weissenborn, X. Zhai, T. Unterthiner, M. Dehghani, M. Minderer, G. Heigold, S. Gelly, J. Uszkoreit, and N. Houlsby. An image is worth 16x16 words: Transformers for image recognition at scale. InInternational Conference on Learning Re...

    2021

  2. [10]

    El-Nouby, M

    A. El-Nouby, M. Klein, S. Zhai, M. Á. Bautista, V . Shankar, A. T. Toshev, J. M. Susskind, and A. Joulin. Scalable pre-training of large autoregressive image models. InForty-first International Conference on Machine Learning, 2024

    2024

  3. [11]

    Ganin, E

    Y . Ganin, E. Ustinova, H. Ajakan, P. Germain, H. Larochelle, F. Laviolette, M. March, and V . Lempitsky. Domain-adversarial training of neural networks.Journal of machine learning research, 17(59):1–35, 2016

    2016

  4. [12]

    H. Gouk, T. Hospedales, and massimiliano pontil. Distance-based regularisation of deep networks for fine-tuning. InInternational Conference on Learning Representations, 2021

    2021

  5. [13]

    Goyal, T

    Y . Goyal, T. Khot, D. Summers-Stay, D. Batra, and D. Parikh. Making the v in vqa matter: Elevating the role of image understanding in visual question answering. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 6904–6913, 2017. 9

    2017

  6. [14]

    K. He, X. Zhang, S. Ren, and J. Sun. Delving deep into rectifiers: Surpassing human-level performance on imagenet classification. InProceedings of the IEEE international conference on computer vision, pages 1026–1034, 2015

    2015

  7. [15]

    Hendrycks, S

    D. Hendrycks, S. Basart, N. Mu, S. Kadavath, F. Wang, E. Dorundo, R. Desai, T. Zhu, S. Parajuli, M. Guo, D. Song, J. Steinhardt, and J. Gilmer. The many faces of robustness: A critical analysis of out-of-distribution generalization. InProceedings of the IEEE/CVF International ...

    2021

  8. [16]

    Hendrycks, K

    D. Hendrycks, K. Zhao, S. Basart, J. Steinhardt, and D. Song. Natural adversarial examples.CVPR, 2021

    2021

  9. [17]

    E. J. Hu, Y . Shen, P. Wallis, Z. Allen-Zhu, Y . Li, S. Wang, L. Wang, and W. Chen. Lora: Low-rank adaptation of large language models.arXiv preprint arXiv:2106.09685, 2021

    2021 arXiv

  10. [18]

    Huang, J

    C. Huang, J. Tian, B. Maneechotesuwan, S. Chopra, and Z. Kira. Directional gradient projection for robust fine-tuning of foundation models.arXiv preprint arXiv:2502.15895, 2025

    2025 arXiv

  11. [19]

    D.-H. Jang, S. Yun, and D. Han. Model stock: All we need is just a few fine-tuned models. InEuropean Conference on Computer Vision, pages 207–223. Springer, 2024

    2024

  12. [20]

    Kumar, A

    A. Kumar, A. Raghunathan, R. Jones, T. Ma, and P. Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. 2022

    2022

  13. [21]

    Kumar, A

    A. Kumar, A. Raghunathan, R. M. Jones, T. Ma, and P. Liang. Fine-tuning can distort pretrained features and underperform out-of-distribution. InInternational Conference on Learning Representations, 2022

    2022

  14. [22]

    Y . Lee, A. S. Chen, F. Tajwar, A. Kumar, H. Yao, P. Liang, and C. Finn. Surgical fine-tuning improves adaptation to distribution shifts. InThe Eleventh International Conference on Learning Representations, 2023

    2023

  15. [23]

    Lester, R

    B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt tuning. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, pages 3045– 3059, 2021

    2021

  16. [24]

    D. Li, J. Li, H. Le, G. Wang, S. Savarese, and S. C. Hoi. Lavis: A one-stop library for language-vision intelligence. InProceedings of the 61st Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 31–41, 2023

    2023

  17. [25]

    Lialin, V

    V . Lialin, V . Deshpande, and A. Rumshisky. Scaling down to scale up: A guide to parameter-efficient fine-tuning.arXiv preprint arXiv:2303.15647, 2023

    2023 arXiv

  18. [26]

    Liu, C.-Y

    S.-Y . Liu, C.-Y . Wang, H. Yin, P. Molchanov, Y .-C. F. Wang, K.-T. Cheng, and M.-H. Chen. Dora: Weight- decomposed low-rank adaptation. InForty-first International Conference on Machine Learning, 2024

    2024

  19. [27]

    Z. Liu, H. Hu, Y . Lin, Z. Yao, Z. Xie, Y . Wei, J. Ning, Y . Cao, Z. Zhang, L. Dong, et al. Swin transformer v2: Scaling up capacity and resolution. In2022 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 11999–12009. IEEE, 2022

    2022

  20. [28]

    H. Lu, M. Ding, Y . Huo, G. Yang, Z. Lu, M. Tomizuka, and W. Zhan. Uniadapter: Unified parameter-efficient transfer learning for cross-modal modeling.arXiv preprint arXiv:2302.06605, 2023

    2023 arXiv

  21. [29]

    F. Meng, Z. Wang, and M. Zhang. Pissa: Principal singular values and singular vectors adaptation of large language models.Advances in Neural Information Processing Systems, 37:121038–121072, 2024

    2024

  22. [30]

    Y . Ni, S. Zhang, and P. Koniusz. Pace: Marrying generalization in parameter-efficient fine-tuning with consistency regularization.Advances in Neural Information Processing Systems, 37:61238–61266, 2024

    2024

  23. [31]

    R. Pan, X. Liu, S. Diao, R. Pi, J. Zhang, C. Han, and T. Zhang. Lisa: layerwise importance sampling for memory-efficient large language model fine-tuning.Advances in Neural Information Processing Systems, 37:57018–57049, 2024

    2024

  24. [32]

    X. Peng, Q. Bai, X. Xia, Z. Huang, K. Saenko, and B. Wang. Moment matching for multi-source domain adaptation. InProceedings of the IEEE/CVF international conference on computer vision, pages 1406–1415, 2019

    2019

  25. [33]

    Recht, R

    B. Recht, R. Roelofs, L. Schmidt, and V . Shankar. Do ImageNet classifiers generalize to ImageNet? In K. Chaudhuri and R. Salakhutdinov, editors,Proceedings of the 36th International Conference on Machine Learning, volume 97 ofProceedings of Machine Learning Research, pages 53...

    2019

  26. [34]

    B. Z. Reichman, A. Sundar, C. Richardson, T. Zubatiy, P. Chowdhury, A. Shah, J. Truxal, M. Grimes, D. Shah, W. J. Chee, et al. Outside knowledge visual question answering version 2.0. InICASSP 2023-2023 IEEE International Conference on Acoustics, Speech and Signal Processing (...

    2023

  27. [35]

    Roy and A

    S. Roy and A. Etemad. Consistency-guided prompt learning for vision-language models.arXiv preprint arXiv:2306.01195, 2023

    2023 arXiv

  28. [36]

    Russakovsky, J

    O. Russakovsky, J. Deng, H. Su, J. Krause, S. Satheesh, S. Ma, Z. Huang, A. Karpathy, A. Khosla, M. Bern- stein, et al. Imagenet large scale visual recognition challenge.International journal of computer vision, 115:211–252, 2015

    2015

  29. [37]

    Saito, S

    K. Saito, S. Yamamoto, Y . Ushiku, and T. Harada. Open set domain adaptation by backpropagation. In Proceedings of the European conference on computer vision (ECCV), pages 153–168, 2018

    2018

  30. [38]

    M. Shah, X. Chen, M. Rohrbach, and D. Parikh. Cycle-consistency for robust visual question answering. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 6649–6658, 2019

    2019

  31. [39]

    Sheng, A

    S. Sheng, A. Singh, V . Goswami, J. Magana, T. Thrush, W. Galuba, D. Parikh, and D. Kiela. Human- adversarial visual question answering.Advances in Neural Information Processing Systems, 34:20346–20359, 2021

    2021

  32. [40]

    J. Tian, Z. He, X. Dai, C.-Y . Ma, Y .-C. Liu, and Z. Kira. Trainable projected gradient method for robust fine-tuning. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7836–7845, 2023

    2023

  33. [42]

    Tian, Y .-C

    J. Tian, Y .-C. Liu, J. S. Smith, and Z. Kira. Fast trainable projection for robust fine-tuning.Advances in Neural Information Processing Systems, 36, 2024

    2024

  34. [43]

    Tzeng, J

    E. Tzeng, J. Hoffman, K. Saenko, and T. Darrell. Adversarial discriminative domain adaptation. InProceedings of the IEEE conference on computer vision and pattern recognition, pages 7167–7176, 2017

    2017

  35. [44]

    H. Wang, S. Ge, Z. Lipton, and E. P. Xing. Learning robust global representations by penalizing local predictive power. InAdvances in Neural Information Processing Systems, pages 10506–10518, 2019

    2019

  36. [45]

    H. Wang, X. Yang, J. Chang, D. Jin, J. Sun, S. Zhang, X. Luo, and Q. Tian. Parameter-efficient tuning of large-scale multimodal foundation model. InThirty-seventh Conference on Neural Information Processing Systems, 2023

    2023

  37. [46]

    Wortsman, G

    M. Wortsman, G. Ilharco, J. W. Kim, M. Li, S. Kornblith, R. Roelofs, R. G. Lopes, H. Hajishirzi, A. Farhadi, H. Namkoong, et al. Robust fine-tuning of zero-shot models. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 7959–7971, 2022

    2022

  38. [47]

    Xuhong, Y

    L. Xuhong, Y . Grandvalet, and F. Davoine. Explicit inductive bias for transfer learning with convolutional networks. InInternational Conference on Machine Learning, pages 2825–2834. PMLR, 2018

    2018

  39. [48]

    E. B. Zaken, S. Ravfogel, and Y . Goldberg. Bitfit: Simple parameter-efficient fine-tuning for transformer-based masked language-models.arXiv preprint arXiv:2106.10199, 2021. 11 Appendix A. Experiment Details In this section, we consolidate all implementation specifics and ext...

    2021

  40. [49]

    public repository to fine-tune all methods. All methods are trained on 3 NVIDIA RTX 4090 GPUs using a fixed configuration: learning rate of 1e-4, weight decay of 1e-4, linear warmup followed by cosine annealing scheduler, batch size of 4 per GPU, and gradient accumulation step...

    1938

Pith tools

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