Pith. sign in

REVIEW 5 major objections 5 minor 42 references

Task-Specific Adaptation with Restricted Model Access

T0 review · 5 major / 5 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read Gray-box fine-tuning with only input/output adapters and gradient access matches full-access fine-tuning on retrieval benchmarks, despite hidden weights and architecture.

desk verdict A genuinely useful gray-box adaptation setting with a solid empirical core, but the abstract overclaims on distant domains and the paper needs error bars and a VPT comparison. read the letter →

arxiv 2502.00796 v1 pith:BEHGA6L6 submitted 2025-02-02 cs.CV

classification cs.CV
keywords gray-boxfine-tuninginput-outputadaptersprompttuninggradientaccessfrozenfoundationmodelsretrievalparameter-efficientadaptationtask
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper tries to establish that a frozen foundation model whose weights and architecture stay hidden can be adapted to new tasks almost as well as full-access fine-tuning, by training only lightweight adapters on the model's input and output and using the back-propagated gradient as the sole window into the model. The authors introduce two settings: DarkGray-box (DGA), which only plugs in an input adapter and an output adapter, and LightGray-box (LGA), which also injects learned tokens into intermediate transformer layers. Across text-image, text-video, and sketch-image retrieval plus classification, DGA lands within a few points of LoRA on near-distribution benchmarks while beating zero-shot by large margins. The paper itself notes that the approach is weaker when the target domain is far from the backbone's training domain, as on Sketchy. If true, this would let model providers offer task adaptation as a service without exposing weights or creating per-task model copies.

What carries the argument

The central object is the DarkGray-Box Input/Output Adapter (DGA) composition $B \circ F \circ A$, where $A$ is an identity-initialized affine map on image pixels or text embeddings and $B$ is a linear output layer. Because only $A$ and $B$ are trained and the backbone $F$ is frozen, the only information that must leave the model is the gradient of the loss with respect to $A$'s output; all layer sizes, weights, and intermediate activations stay hidden. The LightGray-box variant (LGA) relaxes this by adding one learned proxy token per transformer layer, which changes attention scores without modifying weights. The load-bearing piece is that a frozen foundation model's features can be redirected by input-space perturbations alone, without any internal modification.

What would settle it

Train DGA and LoRA on a set of target domains with measured distribution shift from the backbone's pretraining data, and check whether DGA's recall gap to LoRA grows past the near-distribution margin (roughly 0.3 to 1.7 R@1) as the shift increases; if the gap becomes as large as the 27.5-point Sketchy gap for any moderately shifted domain, the broad claim of competitiveness with full-access methods is refuted. Alternatively, if an existing gradient-inversion method recovers the backbone weights from DGA's exposed gradients, the gray-box privacy premise is false.

Watch

Extended reading notes

Core claim

The central claim is that task-specific adaptation can be done with gray-box access, where the provider exposes only gradient flow through entry points and keeps all weights and layer details hidden. The mechanism is the composition $B \circ F \circ A$, with the adapters $A$ and $B$ initialized as identity, so the frozen backbone $F$ initially behaves exactly as before. DGA trains a convolutional visual adapter, two textual tokens (an extra token and a shift token), and linear output adapters, totaling about 0.4% of backbone parameters. On COCO and Flickr30K retrieval with BLIP, DGA reaches R@1 of 53.18 and 83.7 versus LoRA's 53.48 and 85.4, improving over zero-shot values of 47.04 and 78.5; on MSR-VTT and VATEX video retrieval, DGA reaches R@1 of 37.24 and 41.03 versus LoRA's 37.72 and 41.63. The competitive claim is explicitly qualified: on the distant-domain Sketchy benchmark, DGA drops to 31.2 R@1 while LoRA achieves 58.7, so the paper's own evidence says gray-box adaptation is most effective when the new task is not far from the backbone's original domain.

Load-bearing premise

The load-bearing premise is that the model's input gradients are accessible and that changing only the input representation can push a frozen backbone's features far enough toward a new task; the paper also assumes, without testing, that exposing gradients does not undercut the privacy and intellectual-property motivation.

Editorial extensions

If this is right

  • DGA can be offered as an API: a model provider keeps the backbone sealed, exposes only gradient flow, and still delivers task adaptation within a few points of LoRA on near-distribution benchmarks.
  • A single frozen backbone with multiple adapter pairs can handle many tasks at once, avoiding per-task backbone copies; the paper measures this setup at roughly nine times fewer FLOPs and one tenth the GPU memory in its multi-task experiment.
  • Gradient access extends lightweight adaptation to visual inputs, enabling video and sketch retrieval tasks that text-only black-box prompt optimization cannot address.
  • For distant target domains, gray-box adaptation should not be expected to match white-box methods; the paper shows LoRA and full fine-tuning remain necessary there, so the practical scope is limited to modest distribution shifts.

Reading between the lines

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

  • My inference: gray-box adapters and black-box text-prompt optimization operate on different entry points, so a combined pipeline may outperform either alone, though the paper does not test this.
  • My inference: because the paper does not test whether exposing input gradients allows an attacker to reconstruct the hidden weights, 'gray-box' should be read as reduced exposure rather than guaranteed protection; a gradient-inversion attack study would settle this concern.
  • My inference: the identity-initialized affine visual adapter produces images that look corrupted yet improve retrieval, suggesting the mechanism is closer to input-space adversarial optimization than to ordinary feature-space adaptation; this could be tested by measuring how much of the gain comes from the single shift token.
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. This paper introduces 'Gray-box' fine-tuning, a setting in which a pre-trained backbone is frozen and its architecture and weights remain hidden, with only gradient propagation exposed. The authors propose two variants: DarkGray-box (DGA), which learns a visual input adapter, a two-token textual input adapter, and linear output adapters, and LightGray-box (LGA), which additionally injects learnable proxy tokens into each transformer layer. The methods are evaluated on text-image retrieval (COCO, Flickr30K, Stanford-Cars), text-video retrieval (MSR-VTT, VATEX), image classification (ImageNet-1K, ImageNet-Sketch), and sketch-to-image retrieval (Sketchy), with additional experiments on diffusion, captioning, LLM, and CNN backbones in the appendix. The central claim is that the Gray-box approaches are competitive with full-access fine-tuning methods despite limited model access.

Significance. If the near-distribution results hold, DGA offers a practical and parameter-efficient alternative in settings where model weights and architecture must remain proprietary: it uses only 0.10-1.57% of backbone parameters, preserves a single inference flow, and on COCO retrieval achieves R@1 within 0.30 points of LoRA. The paper is also transparent about the method's limitation on distant domains, and the ablation study in Section 5 cleanly isolates the contribution of each adapter. However, the absence of error bars and the unqualified 'competitive' wording in the abstract currently overstate the certainty and scope of the empirical findings.

major comments (5)
  1. [Abstract, §1, §4.4, Table 7] The abstract's claim that the Gray-box approaches are 'competitive with full-access fine-tuning methods' is contradicted by the paper's own Sketchy results in Table 7: DGA R@1 is 31.20 versus LoRA 58.72 and Full FT 69.20, and even LGA reaches only 53.36. Section 6 concedes that this form of adaptation is 'less effective for more distant domains,' yet the abstract and Section 1 present the competitiveness claim without this qualification. The claim should be narrowed to near-distribution retrieval tasks, or the paper should define a quantitative margin of competitiveness and verify it on every benchmark.
  2. [§4 (Tables 2, 5, 6, 7)] No error bars, confidence intervals, or repeated-seed results are reported for any experiment. The headline comparisons on near-distribution tasks, such as the COCO R@1 gap of 0.30 points between DGA and LoRA in Table 2 and the MSR-VTT gap of 0.48 points in Table 5, are within the typical run-to-run variance of retrieval fine-tuning, so the reader cannot determine whether the claimed competitiveness is statistically meaningful. The authors should report mean and standard deviation over at least three seeds for the main tables, or otherwise justify the single-run protocol.
  3. [§3.2] The statement that 'adding non-linear activations does not provide additional benefits' for the visual input adapter is not supported by any reported experiment. No ablation with non-linear activations appears in Section 5 or in the appendices, so this claim should be either backed by an experiment or removed.
  4. [§4.4 and Table 6] The discussion states that LGA 'outperforms DGA and LP by a large margin' in cross-domain settings, but Table 6 shows the opposite on ImageNet-Sketch: DGA achieves 67.48 top-1 accuracy versus 60.06 for LGA. The narrative around LGA's advantage in cross-domain tasks needs to be reconciled with this result, or the claim restricted to the Sketchy experiment.
  5. [§2 and §6] The paper's motivation rests on privacy, safety, and IP protection, but Section 2 explicitly states that the authors 'do not assess the immunity of the Dark or LightGray-box concepts' to gradient-based recovery. Given that prior work cited in the paper (Milli et al., 2019; Horwitz et al., 2024) demonstrates that gradients or fine-tuned adapters can enable weight reconstruction, the abstract and introduction should either temper the privacy/IP claims or include a threat-model discussion and experiments addressing gradient-based extraction.
minor comments (5)
  1. [§4.1, Table 2] The text says the 'LLFT baseline dominates in both datasets,' but in Table 2 Full FT achieves higher R@1 than LLFT on Flickr30K (87.3 versus 86.5).
  2. [Table 1] The notation in Table 1 is confusing: the ✓ symbol is said to indicate 'partial requirements or information,' but the columns mix hidden information with requirements, and for LGA the LayersSizes entry uses ✓ while DGA uses ✔ for the same column. The table should clearly distinguish hidden information from exposed information.
  3. [Appendix B, Table 14] The text states that DGA 'consistently outperforms the Zero-Shot (ZS) and Linear Probing (LP) baselines' for the CLIP backbone on the COCO subsets, but Table 14 shows two counterexamples: on Furniture, DGA R@1 is 52.73 versus LP 52.83, and on Sea, DGA R@1 is 40.75 versus LP 42.81.
  4. [Appendix E] The implementation details do not specify the number of GPUs per node, the total number of seeds used for hyperparameter selection, or whether LoRA rank matching was based on the DGA parameter count per backbone as described; please clarify for reproducibility.
  5. [Throughout] The paper does not mention code release; providing the implementation would substantially strengthen the reproducibility of the reported comparisons.

Circularity Check

0 steps flagged · score 0.0 of 10

No circularity found: the competitiveness claim is an empirical result against external benchmarks, and no fitted quantity is renamed as a prediction.

full rationale

This paper reports an empirical evaluation of two adapter-based fine-tuning schemes (DGA and LGA) against standard white-box baselines on external benchmarks (COCO, Flickr30K, MSR-VTT, VATEX, ImageNet, Sketchy). The core claim that gray-box methods are competitive is a measured outcome, not a quantity derived from its own inputs. The method section (Section 3.2) defines the adapted model as B∘F∘A(x) with A and B initialized as identity, so the identity initialization explicitly preserves the zero-shot function and no fitted parameter is renamed as a prediction. The only tuning choice involving the baselines is the statement in Appendix E that LoRA rank is chosen to match DGA's parameter count; this is a controlled comparison, not a circular step. The paper's own Section 6 concedes that the method is less effective for distant domains like sketch-to-image, and Table 7 shows a large gap against LoRA on Sketchy; that is an evidentiary limitation on the breadth of the competitiveness claim, not a circularity. No load-bearing premise is justified by a self-citation chain, and no uniqueness theorem or ansatz is imported from prior work by the same authors. Accordingly, no circular step can be exhibited, and the appropriate score is 0.

Assumptions & free parameters 6 free parameters · 4 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. The free parameters are standard training hyperparameters, and the axioms are the practical assumptions needed for the gray-box setting to work.

free parameters (6)
  • Learning rate = 1e-4 for CLIP, 5e-5 for BLIP/DINOv2, exponential decay 0.93
    Hand-chosen in Appendix E; standard optimizer setting.
  • Batch size = 256 (retrieval), 64 (Stanford-Cars, ImageNet-Sketch), 1024 (ImageNet-1K)
    Hand-chosen per dataset; affects optimization trajectory.
  • Training epochs = 25 (Stanford-Cars, ImageNet-1K), 30 (Sketchy, ImageNet-Sketch), 50 (COCO), 2 (Flickr30K), 20 (MSR-VTT), 40 (VATEX)
    Hand-chosen in Appendix E.
  • Text input adapter tokens = 2 (1 extra + 1 shift)
    Ablation in Table 17 shows both contribute; authors select this configuration for DGA.
  • LGA proxy tokens per layer = 1 per layer up to 12 layers
    Chosen to balance performance and exposure; Table 16 shows more tokens improve R@1 but increase gradient dimensionality.
  • LoRA rank matched to DGA parameters = r=8 (CLIP), r=2 (BLIP), r=25 (DINOv2)
    Set so that LoRA baselines have the same number of trainable parameters as DGA, per Appendix E.
assumptions (4)
  • domain assumption The pretrained backbone model is differentiable and the provider exposes gradients with respect to entry point inputs.
    This is the defining Gray-box assumption (Section 3.1), and all gradient-based training relies on it.
  • domain assumption The backbone's features are expressive enough that input-space perturbations can shift them into the target domain.
    DGA's success depends on this; the paper's own sketch-to-image results (Table 7) show it fails when the domain gap is large.
  • domain assumption The extra textual token is position-invariant because no positional encoding is applied to it.
    Section 3.2 justifies inserting the token at any position based on CLIP's positional invariance, which is architecture-specific.
  • domain assumption Adaptation can be achieved with only a shallow output linear layer on top of the frozen features.
    Output adapters are linear probes (Section 3.2), which is known to be a strong baseline but not a guarantee for all tasks.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Task-Specific Adaptation with Restricted Model Access." pith.science (2026). https://pith.science/paper/BEHGA6L6

@misc{pith2026250200796,
  author       = {Pith},
  title        = {Pith review of: Task-Specific Adaptation with Restricted Model Access},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BEHGA6L6}},
  note         = {Machine review of arXiv:2502.00796}
}
read the original abstract

The emergence of foundational models has greatly improved performance across various downstream tasks, with fine-tuning often yielding even better results. However, existing fine-tuning approaches typically require access to model weights and layers, leading to challenges such as managing multiple model copies or inference pipelines, inefficiencies in edge device optimization, and concerns over proprietary rights, privacy, and exposure to unsafe model variants. In this paper, we address these challenges by exploring "Gray-box" fine-tuning approaches, where the model's architecture and weights remain hidden, allowing only gradient propagation. We introduce a novel yet simple and effective framework that adapts to new tasks using two lightweight learnable modules at the model's input and output. Additionally, we present a less restrictive variant that offers more entry points into the model, balancing performance with model exposure. We evaluate our approaches across several backbones on benchmarks such as text-image alignment, text-video alignment, and sketch-image alignment. Results show that our Gray-box approaches are competitive with full-access fine-tuning methods, despite having limited access to the model.

Figures

Figures reproduced from arXiv: 2502.00796 by the authors.

Figure 1
Figure 1. An overview of our gray-box frameworks. Left: DarkGray-Box Input/Output Adapters (DGA) permits modifications only at the input and output levels while keeping the backbone model hidden and frozen. The only information available is the gradient flow (indicated by the orange-dotted arrow), which matches the shape of the last layer of the input adapter. Right: In contrast, LighGray-box (LGA) allows additional entry poi… view at source ↗
Figure 2
Figure 2. An overview of our Input Adapters. The visual input adapter (left) consists of 2D task-specific convolutional layers that preserve the image’s original size. The textual input adapter (right) includes two task-specific tokens: a “shift” token added to the original sequence tokens and an “extra” token appended to the original sequence as a contextual token. Both adapters transform the original input into a new repres… view at source ↗
Figure 3
Figure 3. Generated images by three different model versions, of Original (zero-shot), LoRA and LGA. 16 [PITH_FULL_IMAGE:figures/full_fig_p016_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Visualization of the input adapter’s influence on images. 17 [PITH_FULL_IMAGE:figures/full_fig_p017_4.png]
Figure 5
Figure 5. Figure 5: General schemes for handling N different tasks or domains. Top: A single optimized model designed for multiple tasks or domains. Bottom: A naive approach with N different models, one for each task. 18 [PITH_FULL_IMAGE:figures/full_fig_p018_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

42 extracted references · 32 canonical work pages

  1. [1]

    Z., Tople, S., Paverd, A., and K \" o pf, B

    B \' e guelin, S. Z., Tople, S., Paverd, A., and K \" o pf, B. Grey-box Extraction of Natural Language Models . In ICML , volume 139 of Proceedings of Machine Learning Research, pp.\ 12278--12286. PMLR , 2021

  2. [2]

    Bharati, S., Mondal, M. R. H., Podder, P., and Prasath, V. B. S. Federated learning: Applications, challenges and future directions . Int. J. Hybrid Intell. Syst. , 18 0 (1-2): 0 19--35, 2022

  3. [3]

    A., Adeli, E., Altman, R

    Bommasani, R., Hudson, D. A., Adeli, E., Altman, R. B., Arora, S., von Arx, S., Bernstein, M. S., Bohg, J., Bosselut, A., Brunskill, E., Brynjolfsson, E., Buch, S., Card, D., Castellon, R., Chatterji, N. S., Chen, A. S., Creel, K., Davis, J. Q., Demszky, D., Donahue, C., Doumbouya, M., Durmus, E., Ermon, S., Etchemendy, J., Ethayarajh, K., Fei - Fei, L., ...

  4. [4]

    T., Luo, P., Lu, H., and Li, Z

    Chen, J., Yu, J., Ge, C., Yao, L., Xie, E., Wang, Z., Kwok, J. T., Luo, P., Lu, H., and Li, Z. PixArt- \( \) : Fast Training of Diffusion Transformer for Photorealistic Text-to-Image Synthesis . In ICLR , 2024

  5. [5]

    BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding

    Devlin, J., Chang, M., Lee, K., and Toutanova, K. BERT: Pre-training of Deep Bidirectional Transformers for Language Understanding . In NAACL-HLT , pp.\ 4171--4186. Association for Computational Linguistics, 2019

  6. [6]

    An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., and Houlsby, N. An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale . In ICLR , 2021

  7. [7]

    B., Donahue, J., Darrell, T., and Malik, J

    Girshick, R. B., Donahue, J., Darrell, T., and Malik, J. Rich Feature Hierarchies for Accurate Object Detection and Semantic Segmentation . In CVPR , pp.\ 580--587. IEEE Computer Society, 2014

  8. [8]

    Reconstructing Training Data From Trained Neural Networks

    Haim, N., Vardi, G., Yehudai, G., Shamir, O., and Irani, M. Reconstructing Training Data From Trained Neural Networks . In NeurIPS , 2022

Show all 42 references
  1. [9]

    Masked autoencoders are scalable vision learners

    He, K., Chen, X., Xie, S., Li, Y., Doll \'a r, P., and Girshick, R. Masked autoencoders are scalable vision learners. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pp.\ 16000--16009, 2022

  2. [10]

    Recovering the pre-fine-tuning weights of generative models

    Horwitz, E., Kahana, J., and Hoshen, Y. Recovering the pre-fine-tuning weights of generative models. arXiv preprint arXiv:2402.10208, 2024

  3. [11]

    J., Shen, Y., Wallis, P., Allen - Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W

    Hu, E. J., Shen, Y., Wallis, P., Allen - Zhu, Z., Li, Y., Wang, S., Wang, L., and Chen, W. LoRA: Low-Rank Adaptation of Large Language Models . In ICLR , 2022

  4. [12]

    U., Rasheed, H

    Khattak, M. U., Rasheed, H. A., Maaz, M., Khan, S. H., and Khan, F. S. MaPLe: Multi-modal Prompt Learning . In CVPR , pp.\ 19113--19122, 2023

  5. [13]

    C., Lo, W., Doll \' a r, P., and Girshick, R

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A. C., Lo, W., Doll \' a r, P., and Girshick, R. B. Segment Anything . In ICCV , pp.\ 3992--4003. IEEE , 2023

  6. [14]

    3D Object Representations for Fine-Grained Categorization

    Krause, J., Stark, M., Deng, J., and Fei - Fei, L. 3D Object Representations for Fine-Grained Categorization . In ICCV Workshops 2013, Sydney, Australia, December 1-8, 2013 , pp.\ 554--561. IEEE Computer Society, 2013

  7. [15]

    S., Parikh, A

    Krishna, K., Tomar, G. S., Parikh, A. P., Papernot, N., and Iyyer, M. Thieves on Sesame Street! Model Extraction of BERT-based APIs . In ICLR , 2020

  8. [16]

    W., Hassoun, J., Keutzer, K., and Gholami, A

    Kwon, W., Kim, S., Mahoney, M. W., Hassoun, J., Keutzer, K., and Gholami, A. A Fast Post-Training Pruning Framework for Transformers . In NeurIPS , 2022

  9. [17]

    Post-training deep neural network pruning via layer-wise calibration

    Lazarevich, I., Kozlov, A., and Malinin, N. Post-training deep neural network pruning via layer-wise calibration . In ICCVW , pp.\ 798--805. IEEE , 2021

  10. [18]

    The Power of Scale for Parameter-Efficient Prompt Tuning

    Lester, B., Al - Rfou, R., and Constant, N. The Power of Scale for Parameter-Efficient Prompt Tuning . In EMNLP , pp.\ 3045--3059, 2021

  11. [19]

    Li, J., Li, D., Xiong, C., and Hoi, S. C. H. BLIP: Bootstrapping Language-Image Pre-training for Unified Vision-Language Understanding and Generation . In ICML , pp.\ 12888--12900, 2022

  12. [20]

    Li, J., Li, D., Savarese, S., and Hoi, S. C. H. BLIP-2: Bootstrapping Language-Image Pre-training with Frozen Image Encoders and Large Language Models . CoRR, abs/2301.12597, 2023. doi:10.48550/arXiv.2301.12597. URL https://doi.org/10.48550/arXiv.2301.12597

  13. [21]

    Li, X. L. and Liang, P. Prefix-Tuning: Optimizing Continuous Prompts for Generation . In ACL/IJCNLP , pp.\ 4582--4597, 2021

  14. [22]

    J., Hays, J., Perona, P., Ramanan, D., Doll \' a r, P., and Zitnick, C

    Lin, T., Maire, M., Belongie, S. J., Hays, J., Perona, P., Ramanan, D., Doll \' a r, P., and Zitnick, C. L. Microsoft COCO: Common Objects in Context . In ECCV , pp.\ 740--755, 2014

  15. [23]

    Language Models as Black-Box Optimizers for Vision-Language Models

    Liu, S., Yu, S., Lin, Z., Pathak, D., and Ramanan, D. Language Models as Black-Box Optimizers for Vision-Language Models . In CVPR , pp.\ 12687--12697. IEEE , 2024

  16. [24]

    P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks

    Liu, X., Ji, K., Fu, Y., Du, Z., Yang, Z., and Tang, J. P-Tuning v2: Prompt Tuning Can Be Comparable to Fine-tuning Universally Across Scales and Tasks . CoRR, abs/2110.07602, 2021

  17. [25]

    Exploring Models and Data for Remote Sensing Image Caption Generation

    Lu, X., Wang, B., Zheng, X., and Li, X. Exploring Models and Data for Remote Sensing Image Caption Generation . IEEE Trans. Geosci. Remote. Sens. , 2018

  18. [26]

    D., and Hardt, M

    Milli, S., Schmidt, L., Dragan, A. D., and Hardt, M. Model Reconstruction from Model Explanations . In FAT , pp.\ 1--9. ACM , 2019

  19. [27]

    GPT-4 Technical Report

    OpenAI. GPT-4 Technical Report . CoRR, abs/2303.08774, 2023

  20. [28]

    G., Sharma, V., Synnaeve, G., Xu, H., J \' e gou, H., Mairal, J., Labatut, P., Joulin, A., and Bojanowski, P

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El - Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P., Li, S., Misra, I., Rabbat, M. G., Sharma, V., Synnaeve, G., Xu, H., J \' e gou, H., Mair...

  21. [29]

    Efficiently Scaling Transformer Inference

    Pope, R., Douglas, S., Chowdhery, A., Devlin, J., Bradbury, J., Heek, J., Xiao, K., Agrawal, S., and Dean, J. Efficiently Scaling Transformer Inference . In Song, D., Carbin, M., and Chen, T. (eds.), MLSys , 2023

  22. [30]

    W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I

    Radford, A., Kim, J. W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., and Sutskever, I. Learning Transferable Visual Models From Natural Language Supervision . In Meila, M. and Zhang, T. (eds.), ICML , 2021

  23. [31]

    Learning multiple visual domains with residual adapters

    Rebuffi, S., Bilen, H., and Vedaldi, A. Learning multiple visual domains with residual adapters . In Advances in Neural Information Processing Systems, pp.\ 506--516, 2017

  24. [32]

    S., Berg, A

    Russakovsky, O., Deng, J., Su, H., Krause, J., Satheesh, S., Ma, S., Huang, Z., Karpathy, A., Khosla, A., Bernstein, M. S., Berg, A. C., and Fei - Fei, L. ImageNet Large Scale Visual Recognition Challenge . CoRR , abs/1409.0575, 2014

  25. [33]

    The sketchy database: learning to retrieve badly drawn bunnies

    Sangkloy, P., Burnell, N., Ham, C., and Hays, J. The sketchy database: learning to retrieve badly drawn bunnies . ACM Trans. Graph. , 35 0 (4): 0 119:1--119:12, 2016

  26. [34]

    Can't Steal? Cont-Steal! Contrastive Stealing Attacks Against Image Encoders

    Sha, Z., He, X., Yu, N., Backes, M., and Zhang, Y. Can't Steal? Cont-Steal! Contrastive Stealing Attacks Against Image Encoders . In CVPR , pp.\ 16373--16383. IEEE , 2023

  27. [35]

    K., and Ristenpart, T

    Tram \` e r, F., Zhang, F., Juels, A., Reiter, M. K., and Ristenpart, T. Stealing Machine Learning Models via Prediction APIs . In 25th USENIX Security Symposium, USENIX Security 16, Austin, TX, USA, August 10-12, 2016 , pp.\ 601--618. USENIX Association, 2016

  28. [36]

    N., Kaiser, L., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, L., and Polosukhin, I. Attention is All you Need . In NeurIPS , pp.\ 5998--6008, 2017

  29. [37]

    Wang, H., Ge, S., Lipton, Z., and Xing, E. P. Learning Robust Global Representations by Penalizing Local Predictive Power . In NeurIPS , pp.\ 10506--10518, 2019

  30. [38]

    Connecting the Dots: Collaborative Fine-tuning for Black-Box Vision-Language Models

    Wang, Z., Liang, J., He, R., Wang, Z., and Tan, T. Connecting the Dots: Collaborative Fine-tuning for Black-Box Vision-Language Models . In ICML , 2024

  31. [39]

    Zheng, L., Yang, Y., and Hauptmann, A. G. Person Re-identification: Past, Present and Future . CoRR, abs/1610.02984, 2016

  32. [40]

    C., and Liu, Z

    Zhou, K., Yang, J., Loy, C. C., and Liu, Z. Conditional Prompt Learning for Vision-Language Models . In CVPR , pp.\ 16795--16804, 2022

  33. [41]

    MeLo : Low-rank adaptation is better than fine-tuning for medical image diagnosis, 2024

    Zhu, Y., Shen, Z., Zhao, Z., Wang, S., Wang, X., Zhao, X., Shen, D., and Wang, Q. MeLo : Low-rank adaptation is better than fine-tuning for medical image diagnosis, 2024

  34. [42]

    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 gl...

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.