Pith. sign in

REVIEW 3 major objections 5 minor 39 references

Measurement Plasticity: Sensor-Level Adaptation for Vision-Language Models

T0 review · 3 major / 5 minor · reviewed 2026-08-03 · deepseek-v4-flash

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

Pith's one-line read This paper argues that treating a camera's exposure triangle (ISO, shutter speed, aperture) as test-time 'physical prompts'—selecting and combining real captures that best match source features—improves VLM accuracy by up to 25.6 percentage

desk verdict The 'physical prompt' idea is real and worth testing, but the headline gains conflate capture budget with algorithmic gain, and the source-affinity selection is only validated on ImageNet-like classes. read the letter →

arxiv 2512.12571 v3 pith:TGPT5GDL submitted 2025-12-14 cs.CV

Boyeong Im , Wooseok Lee , Yoojin Kwon , Hyung-Sin Kim This is my paper
classification cs.CV
keywords test-timeadaptationvision-languagemodelsphysicalpromptsexposuretrianglesource-affinityscorehardvotingsensor-levelcovariateshiftmulti-viewcapture
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 test-time adaptation for vision-language models can and should happen at the sensor level, not only at the token or prompt level. It proposes MVP, which captures a scene under several exposure settings, scores each physical view by how closely its feature statistics match ImageNet-derived source statistics, keeps the most source-aligned views, filters their digitally augmented variants by predictive entropy, and hard-votes the survivors. On two Tiny-ImageNet-based sensor-shift benchmarks, MVP is reported to beat digital-only TTA/TPT on single Auto-Exposure captures by up to 25.6 percentage points and to add up to 3.4 points beyond pipelines that combine conventional sensor control with TTA. A sympathetic reader would care because if true, it means measurement-time choices—how light is integrated before the image exists—are a powerful, gradient-free lever for robustness that digital post-processing cannot recover.

What carries the argument

Source-affinity score (Eq. 1): for each candidate physical capture, the model's visual encoder produces per-layer mean and variance of image-token embeddings over a set of confident digital augmentations; the score is the negative summed squared distance from those statistics to precomputed ImageNet source statistics. This score is the selection criterion for which camera settings to keep. The second mechanism is entropy-gated hard voting: among the retained views' augmented predictions, the bottom gamma percent by Shannon entropy form a confident subset, and the final prediction is the class receiving the most votes (zero-temperature softmax). Together they replace gradient-based prompt ada

What would settle it

A decisive test would run MVP on a sensor-shift benchmark whose classes are entirely disjoint from ImageNet (e.g., a Places-derived or medical-imaging exposure dataset) and compare it against a control that selects physical views at random (or by confidence only) with identical entropy filtering and hard voting. If MVP's margin over the random-view control vanishes—or if replacing ImageNet source statistics with statistics from an unrelated dataset (say, Places) leaves performance unchanged—then the large reported gains likely stem from ImageNet/Tiny-ImageNet overlap rather than from a general

Watch

Extended reading notes

Core claim

The central claim is that choosing and combining real physical views at test time—control exercised before photons become an image—substantially improves VLM robustness under sensor-induced covariate shift. Concretely, MVP, built on a frozen MaPLe/CLIP encoder, captures M candidate views per scene, computes a source-affinity score per setting as the negative mean squared distance between per-layer feature mean/variance of the view's confident augmented versions and precomputed ImageNet source statistics, selects the top-5 settings, keeps the lowest-entropy 3% of their augmented views, and aggregates via zero-temperature softmax (hard voting). On ImageNet-ES and ImageNet-ES-Diverse, this is r

Load-bearing premise

The whole selection mechanism relies on ImageNet feature statistics being a faithful stand-in for the VLM's true source distribution, and since the evaluation benchmarks are built from Tiny-ImageNet—sharing ImageNet's classes—the affinity score operates in a regime where its reference distribution matches the test distribution, so transfer to non-ImageNet domains remains untested.

Editorial extensions

If this is right

  • If MVP's results hold, test-time adaptation for VLMs can be done without gradients or model modification, purely by choosing which physical measurements to feed in—compatible with 'gray-box' APIs that expose only features or logits.
  • Physical multi-view capture provides a robustness axis that digital augmentations (geometric or photometric) cannot emulate, as shown by LDA of embeddings; sensor diversity is a distinct resource for handling covariate shift.
  • Selecting physical views by source affinity can complement existing sensor-control policies and digital TTA: the reported up-to-3.4 pp gains over the combination suggest both are additively usable.
  • MVP remains effective with reduced candidate sets (M as low as 6), giving a favorable latency–accuracy trade-off and supporting practicality when capture cost is constrained.
  • The source-affinity criterion reportedly generalizes across ViT-B/16 (with and without MaPLe initialization) and ResNet50, so it is not tied to one prompt-initialized architecture.

Reading between the lines

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

  • If the central claim is right, sensor-side selection should be treated as a new axis of test-time adaptation, orthogonal to prompt tuning; future pipelines could combine physical prompting with learned prompts rather than treating them as competing baselines.
  • A testable extension: white balance, focus distance, and sensor gain are other controllable physical parameters; the same source-affinity+entropy+voting recipe could be applied to them, and success there would strongly confirm the 'photons-not-tokens' thesis.
  • The dependence on ImageNet as a proxy for CLIP's true source distribution is a genuine limitation; on deployment domains whose classes and visual statistics differ from ImageNet, the affinity criterion may need per-domain source statistics or a learned reference distribution to keep its edge.
  • The reported gains (25+ pp) are large enough to invite a robustness check: comparing against a random physical-view selection with the same voting protocol would isolate how much of the gain comes from the affinity criterion versus multi-view voting alone—the paper's controlled comparison varies the selection criterion but does not include a random baseline.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 5 minor

Summary. The paper proposes Multi-View Physical-prompt (MVP), a forward-only, training-free test-time adaptation method for vision-language models. Instead of adapting after capture, MVP acquires multiple physical views of a scene by varying camera exposure parameters (ISO, shutter speed, aperture), selects the top-k views using a source-affinity score based on feature statistics (mean and variance per layer) computed against ImageNet, expands each selected view with digital augmentations, filters by predictive entropy, and aggregates with hard voting. Experiments on ImageNet-ES and ImageNet-ES-Diverse report large gains over digital-only TTA baselines under Auto-Exposure (up to 25.6 pp) and smaller gains over pipelines combining Lens sensor control with TTA (up to 3.4 pp). The paper also evaluates reduced capture budgets (CSA), runtime latency, ablations on k, γ, layer depth, and a controlled comparison against confidence-based selection.

Significance. If the claims hold, the paper opens a useful direction: moving test-time adaptation from post-capture prompt tuning to pre-capture sensor control. The method is simple, forward-only, and compatible with gray-box APIs. The controlled experiment in Table 6 (Appendix F) is a genuine strength: it fixes the number of physical views and the aggregation strategy, and shows source-affinity selection outperforms confidence-based selection. The paper also ships ablations on key hyperparameters and latency, which is more than many TTA papers. However, the headline 25–30 pp gains are driven largely by using 27 physical captures versus a single AE capture, and the source-affinity criterion is validated only on benchmarks whose classes come from ImageNet, the same distribution used to precompute the reference statistics. The layer ablation further shows the score can catastrophically mis-select views when the layer choice is changed. These issues limit the strength of the central claim as currently stated, though the core idea is defensible.

major comments (3)
  1. [Section 4.2, Table 1; Appendix F, Table 6] The headline 'up to 25.6 pp' improvement over AE-only TTA compares MVP using 27 physical captures (or CSA-reduced sets) against baselines using a single Auto-Exposure physical view (5 AE captures averaged). This conflates the benefit of additional physical views with the benefit of the proposed selection criterion. The controlled comparison in Table 6, which fixes k and aggregation, shows the source-affinity score adds only about 1.2–1.2 pp over confidence-based selection on ImageNet-ES and 0.3–0.5 pp on ImageNet-ES-Diverse (k=5). Please present the matched-budget comparison as the primary evidence for the selection criterion, and clearly separate the 'multi-physical-view' effect from the 'source-affinity selection' effect in the abstract and conclusions.
  2. [Section 3.2, Eq. (1); Appendix I] The source-affinity score is validated only in a regime where test classes are drawn from the same distribution as the reference statistics: both ImageNet-ES and ImageNet-ES-Diverse are built from Tiny-ImageNet, and the source statistics are precomputed on ImageNet. This favorable overlap means the test content shares class and domain structure with the proxy source. No out-of-domain experiment is reported, so it is unknown whether the affinity score selects useful views when the deployment classes or capture conditions differ from ImageNet. This is load-bearing because the central claim is about general robustness improvement for VLMs. Please add at least one out-of-domain benchmark (e.g., an ImageNet-variant with different classes, or a non-ImageNet dataset with sensor shifts) or explicitly restrict the claims to ImageNet-like deployment domains.
  3. [Table 4, Section 5] The layer ablation demonstrates that the source-affinity score is highly sensitive to the hyperparameter L. Using layers 1–12 or 10–12 collapses ImageNet-ES-Diverse accuracy to 18–20%, far below the Auto-Exposure zero-shot baseline of 37.65, while layers 1–3 achieve 67.28. This indicates that the criterion is not stable across layer choices, and the optimal L=3 appears to have been selected using the benchmark itself. If a deployment domain differs, an uninformed choice of L could actively hurt performance. Please provide a principled method for setting L (e.g., validation on a held-out set, or a domain-agnostic heuristic) and analyze why later-layer statistics fail so dramatically. As it stands, the method's practical robustness is in question.
minor comments (5)
  1. [Section 2.3] Typo: 'adpting' should be 'adapting'.
  2. [Table 1] The MVP row in Table 1 lists only two numbers ('87.85 67.28') under a six-column dataset/condition header. Please clarify which columns these correspond to and report the full set of results across all evaluation protocols.
  3. [Figure 2] The two panels are not fully self-explanatory. Please define how 'feature-space distance' is computed for the density plot in (a) and the attention similarity metric in (b).
  4. [Section 3.2] The dual filtering step (α for source-affinity aggregation, γ for final entropy selection) is introduced quickly. A one-line summary in the main text would help readers understand the relationship between the two thresholds.
  5. [Appendix C, Table 5] Table 5 lacks column headers in the provided manuscript text; please fix the table formatting.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: source-affinity selection, entropy filtering, and hard voting are fixed, parameter-free operations; no prediction reduces to a fitted input or to a load-bearing self-citation.

full rationale

The derivation chain is self-contained. Eq. (1) computes a fixed negative distance between per-layer test-view feature statistics and precomputed ImageNet statistics; no parameter is fit to test labels or to reported accuracy. Eq. (2)-(3) are standard entropy filtering and zero-temperature hard voting, with no learned components. The self-citations to ImageNet-ES [2], ImageNet-ES-Diverse [3], and Lens [3] are used as benchmarks and as a sensor-control baseline, not as evidence for the core claim; there is no uniqueness theorem or ansatz imported through a self-citation. The Tiny-ImageNet/ImageNet overlap noted by the reader is a real domain-overlap confound that limits the transferability of the source-affinity criterion, and Table 4 shows the layer choice is sensitive; however, this is an external-validity/correctness concern, not a case where the prediction is equivalent to its input by construction. The headline gains (25.58 pp over AE, 3.4 pp over Lens+TTA) compare physical multi-view acquisition with single-view and Lens baselines under the same protocol; they do not follow algebraically from the definition of the score. Hence no circular step is present.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The method introduces no new physical or mathematical entities. Its empirical support relies on the assumption that ImageNet is a reliable proxy for the source domain, that physical views are a distinct augmentation axis, and that the tuned selection hyperparameters generalize beyond the chosen test benchmarks.

free parameters (5)
  • k (number of selected physical views) = 5
    Chosen via ablation on the test benchmarks (Fig 8a); controls the number of physical captures retained for voting.
  • γ (entropy percentile filter) = 3%
    Chosen via ablation (Fig 8b); retains the lowest-entropy (most confident) augmented views for voting.
  • α (confidence fraction for source-affinity aggregation) = 0.3
    Set in implementation details; fraction of N augmentations used to compute per-view feature statistics.
  • L (number of layers for source-affinity score) = 3
    Chosen via ablation (Table 4); early layers work best, later layers degrade performance.
  • CSA2 grid adjustment = 2^3 grids (vs Lens's 1^3)
    Hand-adjusted in App. E for 'fairer exposure variability'; changes the candidate set for Lens and MVP, potentially altering baseline comparisons.
assumptions (4)
  • domain assumption ImageNet approximates the source distribution of CLIP/MaPLe for feature alignment.
    Used to precompute source statistics in Eq. 1; original CLIP training data is unavailable, and ImageNet is a standard proxy (following PromptAlign [1]).
  • domain assumption Physical camera parameters provide an augmentation axis orthogonal to digital augmentations.
    Supported by LDA projections (Fig 5), but computed on a small sample; assumes the separation generalizes to the full datasets.
  • standard math Hard voting (zero-temperature softmax) is a robust aggregation for multi-view predictions.
    Taken from ZERO [7] and supported by the ablation in Table 5.
  • ad hoc to paper Early-layer feature statistics are sufficient to capture source-affinity.
    Chosen because later layers degrade performance (Table 4); this is an empirical design choice specific to this method and not derived from first principles.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Measurement Plasticity: Sensor-Level Adaptation for Vision-Language Models." pith.science (2026). https://pith.science/paper/TGPT5GDL

@misc{pith2026251212571,
  author       = {Pith},
  title        = {Pith review of: Measurement Plasticity: Sensor-Level Adaptation for Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/TGPT5GDL}},
  note         = {Machine review of arXiv:2512.12571}
}
read the original abstract

We propose Multi-View Physical-prompt (MVP) for Test-Time Adaptation (TTA), a forward-only framework that moves TTA from tokens to photons by treating the camera exposure triangle (i.e., ISO, shutter speed, and aperture) as physical prompts. At inference, MVP acquires selected multiple physical views using a source-affinity score, evaluates digitally augmented variants of each retained view and filters the lowest-entropy predictions, and aggregates predictions with hard voting. This selection-then-vote design is simple, calibration-friendly, and requires no gradients or model modifications. On ImageNet-ES and ImageNet-ES-Diverse, MVP outperforms digital-only TTA on both Auto-Exposure and a combination with conventional sensor control. MVP remains effective under reduced parameter candidates that lower capture latency, demonstrating its practicality.

Figures

Figures reproduced from arXiv: 2512.12571 by the authors.

Figure 1
Figure 1. Conceptual Figure of the Physical Prompts In the causal chain of scene → measurement → representation, physi￾cal prompts are controlled by sensor parameters (e.g., ISO, shutter speed, aperture) to minimize irrecoverable information loss during measurement, a loss that cannot be compensated by prompt tuning on post-capture images. speed, aperture) and lighting directly shape the photons that reach the sensor and thus… view at source ↗
Figure 2
Figure 2. Motivating findings. (a) Incorrect samples among Lens top-1 predictions form a bi-modal distribution, with a larger mode skewed toward the higher feature-space distances compared to the correct samples. (b) A strong correlation between attention score similarity and feature-space distance shows that samples closer in feature space exhibit more similar attention patterns. scene → measurement → representation. We ther… view at source ↗
Figure 3
Figure 3. Visualization of attention maps from different camera parameters. From the identical scene, we compare the attention maps of the source image (ImageNet) and varying parameter samples (ImageNet-ES-Diverse). terns most similar to the source, clearly distinguishing them from the other views, especially those chosen by the confidence-only strategy (Lens [3]). 3.2. Multi-View Physical-prompt for Test-Time Adaptation Sele… view at source ↗
Figures from the paper (9 more)
Figure 4
Figure 4. Figure 4: Overall Framework. Given a single scene, multiple physical captures are obtained by varying camera parameters along the exposure triangle—ISO, shutter speed, and aperture—forming a set of physical multi-views, as controllable physical prompts. 50 25 0 25 50 40 20 0 20 …
Figure 5
Figure 5. Figure 5: LDA projection of visual embeddings from five phys￾ical views. (Left) With digital transformation such as Random Crop or Horizontal Flip. (Right) With additional photometric aug￾mentations such as perturbing hue, saturation, and brightness. Vi￾sualizations for more sam…
Figure 7
Figure 7. Figure 7: Qualitative comparisons with the conventional sensor-control method. per test sample, showing that MVP delivers a more effective trade-off between performance and latency than the base￾lines. Specifically, AE attains the lowest latency but the weakest accuracy, consist…
Figure 6
Figure 6. Figure 6: Latency (ms) of adaptation and inference. Experi￾ments were conducted on NVIDIA RTX 4090 GPU and Intel(R) Xeon(R) Gold 6326 CPU. Color denotes the sensor control strat￾egy, and markers distinguish TTA variant under the same strategy. Evaluation with limited parameter s…
Figure 8
Figure 8. Figure 8: Ablation on parameter-view selection in MVP. (a) shows the selection of top-k physical parameter settings based on the source-affinity score, while (b) illustrates the confidence per￾centile thresholding used to filter augmented views for final pre￾diction. We fix k = …
Figure 9
Figure 9. Figure 9: Total latency (ms) of adaptation and inference [PITH_FULL_IMAGE:figures/full_fig_p012_9.png]
Figure 10
Figure 10. Figure 10: Additional LDA Projection Visualizations by Light Condition [PITH_FULL_IMAGE:figures/full_fig_p018_10.png]
Figure 11
Figure 11. Figure 11: Additional qualitative results 19 [PITH_FULL_IMAGE:figures/full_fig_p019_11.png]
Figure 12
Figure 12. Figure 12: Additional qualitative results 20 [PITH_FULL_IMAGE:figures/full_fig_p020_12.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

39 extracted references · 6 linked inside Pith

  1. [1]

    Align your prompts: Test-time prompting with distribution align- ment for zero-shot generalization.Advances in Neural Infor- mation Processing Systems, 36:80396–80413, 2023

    Jameel Abdul Samadh, Mohammad Hanan Gani, Noor Hus- sein, Muhammad Uzair Khattak, Muhammad Muzammal Naseer, Fahad Shahbaz Khan, and Salman H Khan. Align your prompts: Test-time prompting with distribution align- ment for zero-shot generalization.Advances in Neural Infor- mation Processing Systems, 36:80396–80413, 2023. 2, 3, 4, 5, 6, 11

    2023

  2. [2]

    Unexplored faces of robustness and out-of-distribution: Co- variate shifts in environment and sensor domains

    Eunsu Baek, Keondo Park, Jiyoon Kim, and Hyung-Sin Kim. Unexplored faces of robustness and out-of-distribution: Co- variate shifts in environment and sensor domains. InPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 22294–22303, 2024. 2, 3, 5, 6, 11, 13

    2024

  3. [3]

    Adaptive camera sensor for vision models.arXiv preprint arXiv:2503.02170, 2025

    Eunsu Baek, Sunghwan Han, Taesik Gong, and Hyung-Sin Kim. Adaptive camera sensor for vision models.arXiv preprint arXiv:2503.02170, 2025. 2, 3, 4, 5, 6, 7, 11, 12, 13

    arXiv 2025

  4. [4]

    Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models.Ad- vances in neural information processing systems, 32, 2019

    Andrei Barbu, David Mayo, Julian Alverio, William Luo, Christopher Wang, Dan Gutfreund, Josh Tenenbaum, and Boris Katz. Objectnet: A large-scale bias-controlled dataset for pushing the limits of object recognition models.Ad- vances in neural information processing systems, 32, 2019. 1

    2019

  5. [5]

    Albumentations: fast and flexible image augmenta- tions.Information, 11(2):125, 2020

    Alexander Buslaev, Vladimir I Iglovikov, Eugene Khved- chenya, Alex Parinov, Mikhail Druzhinin, and Alexandr A Kalinin. Albumentations: fast and flexible image augmenta- tions.Information, 11(2):125, 2020. 5, 6

    2020

  6. [6]

    Conceptual 12m: Pushing web-scale image-text pre- training to recognize long-tail visual concepts

    Soravit Changpinyo, Piyush Sharma, Nan Ding, and Radu Soricut. Conceptual 12m: Pushing web-scale image-text pre- training to recognize long-tail visual concepts. InProceed- ings of the IEEE/CVF conference on computer vision and pattern recognition, pages 3558–3568, 2021. 1

    2021

  7. [7]

    Frustratingly easy test-time adaptation of vision-language models.Advances in Neural Information Processing Systems, 37:129062–129093, 2024

    Matteo Farina, Gianni Franchi, Giovanni Iacca, Massimil- iano Mancini, and Elisa Ricci. Frustratingly easy test-time adaptation of vision-language models.Advances in Neural Information Processing Systems, 37:129062–129093, 2024. 2, 3, 4, 5, 6

    2024

  8. [8]

    Diverse data augmentation with diffusions for effective test-time prompt tuning

    Chun-Mei Feng, Kai Yu, Yong Liu, Salman Khan, and Wangmeng Zuo. Diverse data augmentation with diffusions for effective test-time prompt tuning. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 2704–2714, 2023. 2

    2023

Show all 39 references
  1. [9]

    Dat- acomp: In search of the next generation of multimodal datasets.Advances in Neural Information Processing Sys- tems, 36:27092–27112, 2023

    Samir Yitzhak Gadre, Gabriel Ilharco, Alex Fang, Jonathan Hayase, Georgios Smyrnis, Thao Nguyen, Ryan Marten, Mitchell Wortsman, Dhruba Ghosh, Jieyu Zhang, et al. Dat- acomp: In search of the next generation of multimodal datasets.Advances in Neural Information Processing Sys-...

    2023

  2. [10]

    Benchmarking neu- ral network robustness to common corruptions and perturba- tions.arXiv preprint arXiv:1903.12261, 2019

    Dan Hendrycks and Thomas Dietterich. Benchmarking neu- ral network robustness to common corruptions and perturba- tions.arXiv preprint arXiv:1903.12261, 2019. 2

    1903 arXiv

  3. [11]

    The many faces of robust- ness: A critical analysis of out-of-distribution generalization

    Dan Hendrycks, Steven Basart, Norman Mu, Saurav Kada- vath, Frank Wang, Evan Dorundo, Rahul Desai, Tyler Zhu, Samyak Parajuli, Mike Guo, et al. The many faces of robust- ness: A critical analysis of out-of-distribution generalization. InProceedings of the IEEE/CVF internationa...

    2021

  4. [12]

    Natural adversarial examples

    Dan Hendrycks, Kevin Zhao, Steven Basart, Jacob Stein- hardt, and Dawn Song. Natural adversarial examples. In Proceedings of the IEEE/CVF conference on computer vi- sion and pattern recognition, pages 15262–15271, 2021. 2

    2021

  5. [13]

    Scaling up visual and vision-language representa- tion learning with noisy text supervision

    Chao Jia, Yinfei Yang, Ye Xia, Yi-Ting Chen, Zarana Parekh, Hieu Pham, Quoc Le, Yun-Hsuan Sung, Zhen Li, and Tom Duerig. Scaling up visual and vision-language representa- tion learning with noisy text supervision. InInternational conference on machine learning, pages 4904–4916. PMLR,

  6. [14]

    Clip-rt: Learning language- conditioned robotic policies from natural language supervi- sion.arXiv preprint arXiv:2411.00508, 2024

    Gi-Cheon Kang, Junghyun Kim, Kyuhwan Shim, Jun Ki Lee, and Byoung-Tak Zhang. Clip-rt: Learning language- conditioned robotic policies from natural language supervi- sion.arXiv preprint arXiv:2411.00508, 2024. 1

    2024 arXiv

  7. [15]

    Efficient test-time adaptation of vision-language models

    Adilbek Karmanov, Dayan Guan, Shijian Lu, Abdulmotaleb El Saddik, and Eric Xing. Efficient test-time adaptation of vision-language models. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14162–14171, 2024. 2, 3, 5, 6

    2024

  8. [16]

    Promptsync: Bridging domain gaps in vision-language models through class-aware prototype alignment and discrimination

    Anant Khandelwal. Promptsync: Bridging domain gaps in vision-language models through class-aware prototype alignment and discrimination. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 7819–7828, 2024. 2

    2024

  9. [17]

    Maple: Multi-modal prompt learning

    Muhammad Uzair Khattak, Hanoona Rasheed, Muham- mad Maaz, Salman Khan, and Fahad Shahbaz Khan. Maple: Multi-modal prompt learning. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19113–19122, 2023. 2, 5, 6, 11, 12

    2023

  10. [18]

    A decade’s battle on dataset bias: Are we there yet? InThe Thirteenth International Conference on Learning Representations

    Zhuang Liu and Kaiming He. A decade’s battle on dataset bias: Are we there yet? InThe Thirteenth International Conference on Learning Representations. 1

  11. [19]

    Swapprompt: Test-time prompt adaptation for vision- language models.Advances in Neural Information Process- ing Systems, 36:65252–65264, 2023

    Xiaosong Ma, Jie Zhang, Song Guo, and Wenchao Xu. Swapprompt: Test-time prompt adaptation for vision- language models.Advances in Neural Information Process- ing Systems, 36:65252–65264, 2023. 2

    2023

  12. [20]

    Learning transferable visual models from natural language supervi- sion

    Alec Radford, Jong Wook Kim, Chris Hallacy, Aditya Ramesh, Gabriel Goh, Sandhini Agarwal, Girish Sastry, Amanda Askell, Pamela Mishkin, Jack Clark, et al. Learning transferable visual models from natural language supervi- sion. InInternational conference on machine learning, p...

    2021

  13. [21]

    Do vision trans- formers see like convolutional neural networks?Advances in neural information processing systems, 34:12116–12128,

    Maithra Raghu, Thomas Unterthiner, Simon Kornblith, Chiyuan Zhang, and Alexey Dosovitskiy. Do vision trans- formers see like convolutional neural networks?Advances in neural information processing systems, 34:12116–12128,

  14. [22]

    Do imagenet classifiers generalize to im- agenet? InInternational conference on machine learning, pages 5389–5400

    Benjamin Recht, Rebecca Roelofs, Ludwig Schmidt, and Vaishaal Shankar. Do imagenet classifiers generalize to im- agenet? InInternational conference on machine learning, pages 5389–5400. PMLR, 2019. 1, 2

    2019

  15. [23]

    A mathematical theory of communi- cation.The Bell system technical journal, 27(3):379–423,

    Claude E Shannon. A mathematical theory of communi- cation.The Bell system technical journal, 27(3):379–423,

  16. [24]

    Large vlm-based vision- language-action models for robotic manipulation: A survey

    Rui Shao, Wei Li, Lingsen Zhang, Renshan Zhang, Zhiyang Liu, Ran Chen, and Liqiang Nie. Large vlm-based vision- language-action models for robotic manipulation: A survey. arXiv preprint arXiv:2508.13073, 2025. 1

    2025 arXiv

  17. [25]

    O-tpt: Orthogonality constraints for calibrating test-time prompt tuning in vision-language models

    Ashshak Sharifdeen, Muhammad Akhtar Munir, Sanoojan Baliah, Salman Khan, and Muhammad Haris Khan. O-tpt: Orthogonality constraints for calibrating test-time prompt tuning in vision-language models. InProceedings of the Computer Vision and Pattern Recognition Conference, pages ...

    2025

  18. [26]

    Test- time prompt tuning for zero-shot generalization in vision- language models.Advances in Neural Information Process- ing Systems, 35:14274–14289, 2022

    Manli Shu, Weili Nie, De-An Huang, Zhiding Yu, Tom Goldstein, Anima Anandkumar, and Chaowei Xiao. Test- time prompt tuning for zero-shot generalization in vision- language models.Advances in Neural Information Process- ing Systems, 35:14274–14289, 2022. 2, 5, 6

    2022

  19. [27]

    Just shift it: Test-time prototype shifting for zero-shot generaliza- tion with vision-language models

    Elaine Sui, Xiaohan Wang, and Serena Yeung-Levy. Just shift it: Test-time prototype shifting for zero-shot generaliza- tion with vision-language models. In2025 IEEE/CVF Win- ter Conference on Applications of Computer Vision (WACV), pages 825–835. IEEE, 2025. 2, 3

    2025

  20. [28]

    Yfcc100m: The new data in multimedia research

    Bart Thomee, David A Shamma, Gerald Friedland, Ben- jamin Elizalde, Karl Ni, Douglas Poland, Damian Borth, and Li-Jia Li. Yfcc100m: The new data in multimedia research. Communications of the ACM, 59(2):64–73, 2016. 1

    2016

  21. [29]

    Unbiased look at dataset bias

    Antonio Torralba and Alexei A Efros. Unbiased look at dataset bias. InCVPR 2011, pages 1521–1528. IEEE, 2011. 1

    2011

  22. [30]

    Learning robust global representations by penalizing local predictive power.Advances in neural information pro- cessing systems, 32, 2019

    Haohan Wang, Songwei Ge, Zachary Lipton, and Eric P Xing. Learning robust global representations by penalizing local predictive power.Advances in neural information pro- cessing systems, 32, 2019. 2

    2019

  23. [31]

    Dynaprompt: Dynamic test-time prompt tuning.arXiv preprint arXiv:2501.16404, 2025

    Zehao Xiao, Shilin Yan, Jack Hong, Jiayin Cai, Xiaolong Jiang, Yao Hu, Jiayi Shen, Qi Wang, and Cees GM Snoek. Dynaprompt: Dynamic test-time prompt tuning.arXiv preprint arXiv:2501.16404, 2025. 2

    2025 arXiv

  24. [32]

    Tiny imagenet visual recognition challenge

    Xuan Yang. Tiny imagenet visual recognition challenge. 5

  25. [33]

    C-tpt: Calibrated test-time prompt tuning for vision- language models via text feature dispersion.arXiv preprint arXiv:2403.14119, 2024

    Hee Suk Yoon, Eunseop Yoon, Joshua Tian Jin Tee, Mark Hasegawa-Johnson, Yingzhen Li, and Chang D Yoo. C-tpt: Calibrated test-time prompt tuning for vision- language models via text feature dispersion.arXiv preprint arXiv:2403.14119, 2024. 2, 5, 6

    2024 arXiv

  26. [34]

    Maxime Zanella and Ismail Ben Ayed. On the test-time zero- shot generalization of vision-language models: Do we really need prompt learning? InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 23783–23793, 2024. 2, 3, 5, 6

    2024

  27. [35]

    Understanding bias in large-scale visual datasets.Advances in Neural Informa- tion Processing Systems, 37:61839–61871, 2024

    Boya Zeng, Yida Yin, and Zhuang Liu. Understanding bias in large-scale visual datasets.Advances in Neural Informa- tion Processing Systems, 37:61839–61871, 2024. 1

    2024

  28. [36]

    Historical test-time prompt tuning for vision foundation models.Advances in Neural Information Pro- cessing Systems, 37:12872–12896, 2024

    Jingyi Zhang, Jiaxing Huang, Xiaoqin Zhang, Ling Shao, and Shijian Lu. Historical test-time prompt tuning for vision foundation models.Advances in Neural Information Pro- cessing Systems, 37:12872–12896, 2024. 2

    2024

  29. [37]

    Conditional prompt learning for vision-language mod- els

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Conditional prompt learning for vision-language mod- els. InProceedings of the IEEE/CVF conference on com- puter vision and pattern recognition, pages 16816–16825,

  30. [38]

    Learning to prompt for vision-language models.In- ternational Journal of Computer Vision, 130(9):2337–2348,

    Kaiyang Zhou, Jingkang Yang, Chen Change Loy, and Ziwei Liu. Learning to prompt for vision-language models.In- ternational Journal of Computer Vision, 130(9):2337–2348,

  31. [2022]

    a photo of a class

    2, 5, 6 10 From Tokens to Photons: Test-Time Physical Prompting for Vision-Language Models Supplementary Material A. Evaluation on different architectures To verify that the effectiveness ofMVPdoes not de- pend on a specific prompting architecture, we further eval- uate our fr...

Pith tools

Reviewed August 3, 2026 · model on record in the stance chip above.