Pith. sign in

REVIEW 4 major objections 4 minor 48 references

SemPT: Semantic Prompt Tuning for Vision-Language Models

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read SemPT claims that shared attribute-level descriptions generated by a two-step LLM prompt build transferable semantic bridges that let vision-language prompt tuning generalize to unseen categories, and reports state-of-the-art results across

desk verdict Useful plug-in method with consistent small gains, but the 'unseen' generalization claim is compromised by transductive leakage of novel class names into the LLM prompting and training losses. read the letter →

arxiv 2508.10645 v1 pith:RCDWLNWJ submitted 2025-08-14 cs.CV

Xiao Shi , Yangjun Ou , Zhenzhong Chen This is my paper · ORCID
classification cs.CV
keywords prompttuningvision-languagemodelstransferlearningattribute-leveldescriptionslargelanguagezero-shotgeneralizationbase-to-novelCLIP
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

SemPT claims that prompt tuning for vision-language models fails to generalize to unseen classes because text embeddings are fragmented: category labels are too sparse and LLM-generated descriptions are too disparate to form connections across categories. The paper's remedy is to build a shared vocabulary of visual attributes with a two-step LLM prompt and to express every category, seen and unseen, as combinations of those attributes, creating explicit semantic bridges from the known to the unknown. It then weights attribute descriptions by visual relevance, fuses them into label embeddings through a residual projection, and trains label and attribute-enhanced objectives jointly. At inference, seen classes use label embeddings while unseen classes use attribute-enhanced embeddings. Across 15 benchmarks, the approach reports the best harmonic-mean results in base-to-novel, cross-dataset, and cross-domain settings when plugged into two existing prompt-tuning methods.

What carries the argument

The central object is the shared attribute vocabulary $\mathcal{A} = \{a_m\}_{m=1}^{M_a}$, produced by the first LLM prompt over both seen and unseen category names, together with the attribute-level description set $\mathcal{T} = \{t_i^j\}$ generated in a second prompt conditioned on $\mathcal{A}$. The shared vocabulary is what makes the space connected: unseen categories can be decomposed into attribute combinations already present in seen categories. The mechanism that makes the vocabulary useful is the visually guided weighting (top-K selection with temperature-scaled softmax over image–description cosine similarities) and the residual fusion $T_{\text{enh}}^{(i)} = (1-\$\alpha$) T_{\text{l

What would settle it

Run the exact SemPT setup on one of the 11 standard benchmarks but replace the shared attribute vocabulary with a matched set of per-category attributes generated independently (no shared vocabulary), keeping description count, fusion, and training identical; if base-to-novel harmonic mean does not drop meaningfully, the shared-attribute claim is unsupported. A second check is to read the two-step LLM outputs: if unseen categories share few or no attributes with seen categories, the semantic-bridge mechanism cannot be carrying the reported gains.

Watch

Extended reading notes

Core claim

The central claim is that shared attribute-level knowledge, rather than category names or free-form LLM descriptions, is the right carrier of transferable semantics for vision-language prompt tuning. The paper argues that embeddings built from labels alone or from unrelated descriptions form isolated clusters with no semantic path to unseen categories. SemPT's two-step prompting first asks an LLM for a compact vocabulary of visual attributes that applies across all categories, then asks it to express each category as short descriptions grounded in that vocabulary. The resulting attribute-aligned text embeddings are selected per image by top-K cosine similarity with temperature-scaled softmax

Load-bearing premise

The load-bearing premise is that an LLM, given the names of both seen and unseen categories and told to identify shared visual attributes, will produce a vocabulary that genuinely transfers across the seen–unseen boundary; the paper provides no LLM outputs, prompts, or manual inspection to verify that the shared attributes are real, distinct, and useful.

Editorial extensions

If this is right

  • SemPT can be layered onto existing prompt-tuning methods (PromptKD, MMRL) and reports consistent gains over those baselines across all four evaluation settings.
  • Improvements are strongest on fine-grained datasets like StanfordCars and OxfordPets, where attribute-level descriptions help separate visually similar classes.
  • The ablation shows the mixed inference rule (labels for seen, attribute-enhanced for unseen) beats using either embedding type for all categories, so category exposure should guide text representation choice.
  • Two-step prompting with a shared vocabulary outperforms both one-step prompting and category names alone, indicating the structure of LLM knowledge construction matters beyond description content.
  • Because only prompts and a small MLP are trained, SemPT retains few-shot compatibility, where it reports the best 16-shot average among compared methods.

Reading between the lines

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

  • A direct consequence the paper leaves implicit: if shared attributes drive transfer, then unseen categories expressible as combinations of known attributes should be recognizable with zero labeled images, which can be tested by probing with synthetic attribute text.
  • The dependence on the specific LLM and prompt wording is unexplored; a natural extension is to vary the LLM and measure whether vocabulary agreement across LLMs predicts transfer performance.
  • The binary seen/unseen inference switch could be softened into a confidence-based blend of label and attribute embeddings, a variation the paper does not test but its dual-embedding view suggests.
  • The dual-embedding design could be adapted to class-incremental learning, where the seen set grows over time and attribute-enhanced embeddings could preserve old-class discrimination while extending to new classes.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 4 minor

Summary. SemPT is a prompt-tuning framework for CLIP-based vision-language models. The paper proposes to build a shared attribute vocabulary by prompting an LLM in two steps: first extract cross-category visual attributes from both seen and unseen class names, then generate attribute-level descriptions for every class. These descriptions are encoded, visually weighted by top-K selection with softmax attention, fused with label text embeddings via an MLP and residual connection, and trained with two softmax losses. At inference, label embeddings are used for seen classes and attribute-enhanced embeddings for unseen classes. The method is evaluated on 15 datasets under base-to-novel, cross-dataset, cross-domain, and few-shot settings, integrated with PromptKD and MMRL baselines. The authors report consistent but small improvements over these baselines, e.g., PromptKD w/SemPT reaches 84.04% harmonic mean in base-to-novel, 72.11% in cross-dataset, and 72.05% in cross-domain.

Significance. If validated, SemPT would offer a simple plug-in enhancement for existing prompt-tuning methods by explicitly modeling shared attribute-level semantics, and the adaptive seen/unseen inference rule is a sensible design. The paper is reasonably positioned relative to LLM-augmented prompt tuning (HPT, ArGue, CoCoLe, ProText) and tests a broad set of benchmarks with two strong baselines. However, the reported gains are modest (0.3–0.8 points), and the evidence is weakened by the fact that unseen class names and descriptions are used during training, by a concrete hyperparameter inconsistency, by the absence of error bars and significance tests, and by the lack of prompt templates or LLM outputs needed for reproducibility. The central claim of transfer via shared attributes is therefore not yet established.

major comments (4)
  1. [§3.1–3.2 and Eqs. (12)–(14)] The framework uses unseen-class names and attribute descriptions during training. Sec. 3.1 defines N = N_s + N_u, and Sec. 3.2 prompts the LLM with category examples from both C_s and C_u, producing attribute vocabulary A and descriptions T for all categories. The training losses in Eqs. (12)–(14) normalize over all N categories, so image embeddings are pushed away from unseen-class label and attribute-enhanced text embeddings. Thus the 'unseen' categories are not truly unseen at training time. The 'category names only' ablation in Table 4(b) also uses all N names, so it does not control for this. The reported gains in Table 1 could stem from transductive exposure to unseen class names rather than from the proposed shared-attribute transfer. Please add a control experiment that removes C_u from the LLM knowledge construction and from the loss denominators, and report base/novel HM. If th
  2. [§4.3 vs Table 4(d)] There is a direct contradiction in the loss-balancing hyperparameter. Implementation details (Sec. 4.3) state that β in Eq. (14) is set to 0.4. The ablation in Table 4(d) reports that the optimal β is 0.6, achieving 84.04% HM, which equals the main result reported in Table 1. It is therefore unclear which β was used for the final results. If β=0.6 was used, the implementation statement is wrong; if β=0.4 was used, the ablation's conclusion is unsupported. This must be corrected and the value of β for each experimental setting must be specified explicitly.
  3. [Tables 1–3 and Fig. 4] All results are reported as single numbers without variance, multiple seeds, or significance tests. The headline improvements are small (0.31–0.78 percentage points) and some per-dataset values decrease (e.g., ImageNet base drops from 80.83 to 80.71 in Table 1). Without error bars, it is not possible to tell whether the average gains are meaningful. Please report mean±std over at least three seeds and, where feasible, paired significance tests across the datasets.
  4. [§3.2 Knowledge Construction] The two-step prompting strategy is the core of the method, but the exact prompt templates and LLM outputs are not provided. No examples of the shared attribute vocabulary A or of the generated descriptions T are shown, and there is no manual or automatic verification that the attributes are shared or visually meaningful. This makes the method non-reproducible and leaves the central assumption of transferable semantic bridges unverified. Please include the full prompts in an appendix, provide sample outputs per dataset, and ideally add a human or automated consistency check.
minor comments (4)
  1. [Table 1, KIM row] In the FGVCAircraft column, KIM reports Base=42.80, Novel=37.70, but HM=30.10. The harmonic mean of those two numbers is about 40.1, suggesting a typo. Please check.
  2. [§3.1] There is a typo in 'Tlbl∈ R N N×d'; it should be 'T_lbl ∈ R^{N×d}'.
  3. [Table 4(b)] The 'one-step prompting strategy' is not defined. Please specify how it differs from the two-step strategy (e.g., no explicit attribute vocabulary) so the ablation is interpretable.
  4. [§4.3] The statement that all baselines were 'faithfully reproduced' would be easier to verify if code/checkpoints or a detailed reproduction protocol were provided. Consider adding a footnote with repository links or a reproducibility statement.

Circularity Check

2 steps flagged · score 6.0 of 10

Novel-class names leak into LLM knowledge construction and training losses, making the 'unseen' generalization claim partially transductive.

  1. self definitional [Section 3.2, Knowledge Construction, Steps 1-2; cf. Section 4.2.1]
    ""To enhance the quality of extracted attributes, we incorporate category examples from both Cs and Cu in the prompt, enabling the LLM to identify attributes that are generalizable across the category spectrum." ... "Based on A, we then prompt the LLM to generate multiple concise descriptions for each category ci ∈ Cs∪Cu.""

    Section 4.2.1 defines base-to-novel evaluation as training 'solely on the base categories' and evaluating on novel categories, i.e., the novel classes are supposed to be unseen. But Step 1 and Step 2 explicitly supply the LLM with novel class names (Cu) and request attribute-level descriptions for every ci ∈ Cs∪Cu. The attribute vocabulary A and descriptions T that later produce attribute-enhanced embeddings for novel categories are thus constructed directly from the names of the categories being predicted. The 'unseen' split is therefore definitional only: novel category identity is an input to knowledge construction, so novel-class accuracy is a transductive result rather than evidence of transfer to genuinely unseen categories.

  2. fitted input called prediction [Section 3.1 (N = Ns+Nu); Section 3.5.1, Eqs. (12)-(14); Section 3.5.2, Eq. (15)]
    ""where N = Ns + Nu is the total number of categories" ... "Llbl =− log exp(⟨v, T(y) lbl⟩/τ) / PN j=1 exp(⟨v, T( j) lbl⟩/τ)" ... "Lenh =− log exp(⟨v, T(y) enh⟩/τ) / PN j=1 exp(⟨v, T( j) enh⟩/τ)" ... "L = (1−β)·L lbl +β·L enh""

    Training images have base-class labels, but the softmax in both Llbl and Lenh sums over all N = Ns + Nu classes. Consequently the learnable textual prompt PT, visual prompt PV, and projection MLP are optimized using cosine similarities to novel-class label and attribute-enhanced embeddings. The novel-class embeddings later selected at inference (Eq. 15) are therefore not independent pre-trained anchors; they have been shaped by a training objective that included the novel class names as negatives. Reporting the resulting novel accuracy as 'generalization to unseen categories' is thus a fitted-input-called-prediction: the predicted novel-class text embeddings were effectively fitted, through shared prompt parameters, with the test class identities in the loop.

full rationale

The paper contains no mathematical derivation that reduces to a fitted quantity, and there is no load-bearing self-citation or ansatz-smuggled-via-citation pattern; the method is an empirical recipe evaluated against external benchmarks. The core circularity is protocol-level: in base-to-novel generalization, the categories labeled 'unseen' are not actually excluded from the method. Their names are fed to the LLM during knowledge construction (Section 3.2), and their label/attribute-enhanced text embeddings appear in the softmax denominators of the training losses (Eqs. 12-14), which update the shared learnable prompts and MLP. The same embeddings are then used at inference for novel classes (Eq. 15). This makes the novel-class prediction partially transductive: the model has been optimized with the test class identities as negatives, so the 'transfer to unseen categories' claim is not established as a genuinely held-out test. The ablation labeled 'category names only' (83.83 HM) also uses all N names and therefore does not control for this leakage. Cross-dataset and cross-domain results may be less affected if target class names are not in the training losses, but the paper does not report such a control. The empirical numbers are still measurements of a real, possibly useful transductive/LLM-assisted prompt-tuning method, so the circularity is partial rather than total: score 6.

Assumptions & free parameters 5 free parameters · 3 assumptions · 0 invented entities

No new physical or conceptual entities are introduced; the framework combines existing CLIP, LLM, and prompt tuning components.

free parameters (5)
  • alpha = 0.2
    Balance between label and enriched embeddings in Eq. 10; tuned by validation.
  • beta = 0.4 (implementation) / 0.6 (ablation optimum)
    Loss weight in Eq. 14; paper reports 0.4 in Section 4.3 but ablation Table 4(d) shows 0.6 is best and is used in final results.
  • K = 2
    Number of top attribute descriptions selected per category in Section 3.3; tuned via ablation.
  • tau = not stated; likely CLIP default
    Temperature in Eq. 3 and 4; no tuning described.
  • S = not stated
    Number of attribute descriptions generated per category; not reported.
assumptions (3)
  • domain assumption LLM-generated attribute descriptions grounded in shared attributes improve CLIP text embeddings for transfer.
    Core assumption of the paper; not proven, only empirically tested on benchmarks.
  • standard math Pre-trained CLIP encoders provide a reliable cross-modal embedding space where cosine similarity aligns visual and textual relevance.
    Underlying all methods; accepted as a baseline assumption.
  • domain assumption The base/novel category split in each dataset is representative of unseen-class generalization.
    The evaluation assumes that one random split reflects real-world unseen categories; no multiple splits.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SemPT: Semantic Prompt Tuning for Vision-Language Models." pith.science (2026). https://pith.science/paper/RCDWLNWJ

@misc{pith2026250810645,
  author       = {Pith},
  title        = {Pith review of: SemPT: Semantic Prompt Tuning for Vision-Language Models},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/RCDWLNWJ}},
  note         = {Machine review of arXiv:2508.10645}
}
read the original abstract

Visual transfer learning for unseen categories presents an active research topic yet a challenging task, due to the inherent conflict between preserving category-specific representations and acquiring transferable knowledge. Vision-Language Models (VLMs) pre-trained on large amounts of image-text pairs offer a promising solution. However, existing prompt tuning methods rely on sparse category labels or disparate LLM-generated descriptions, which fragment knowledge representation and hinder transferability. To address this limitation, we introduce Semantic Prompt Tuning (SemPT), a novel framework that tackles the generalization challenge by leveraging shared attribute-level knowledge across categories. Specifically, SemPT adopts a two-step prompting strategy to guide LLM in extracting shared visual attributes and generating attribute-level descriptions, capturing transferable semantic cues beyond labels while ensuring coherent structure. Then, visually guided weighting is applied to the embeddings of attribute-level descriptions to reduce noise from irrelevant attributes and enhance the text embeddings. Additionally, image embeddings are jointly aligned with both label and attribute-enhanced text embeddings, balancing discrimination for seen categories and transferability to unseen ones. Considering the availability of category exposure, our inference dynamically selects between standard label embeddings for seen categories and attribute-enhanced embeddings for unseen ones to ensure effective adaptation. Extensive experiments on 15 benchmark datasets demonstrate that SemPT achieves state-of-the-art performance across various settings, including base-to-novel generalization, cross-dataset transfer, cross-domain transfer, and few-shot learning.

Figures

Figures reproduced from arXiv: 2508.10645 by the authors.

Figure 1
Figure 1. Illustration of existing methods and the proposed SemPT framework. (a) Label-based methods rely on sparse category [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Overview of the SemPT framework. Training Stage: In the KC module, GPT-4o is used in a two-step prompting strategy to extract shared visual attributes and generate attribute-level descriptions T across all categories. The SA module selects the top-K most visually relevant descriptions per category based on image-text similarity, followed by softmax-weighted averaging to obtain attribute-aligned text embeddings Tˆ (i… view at source ↗
Figure 3
Figure 3. Two-step prompting strategy for knowledge construction using LLM. Step 1 identifies diverse and complementary [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Comparison between our method and other existing methods under the Few-Shot Learning setting. [PITH_FULL_IMAGE:figures/full_fig_p010_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 43 canonical work pages

  1. [1]

    Radford, J

    A. Radford, J. W. Kim, C. Hallacy, A. Ramesh, G. Goh, S. Agarwal, G. Sastry, A. Askell, P. Mishkin, J. Clark, and G. Krueger. Learning transferable visual models from natural language supervision. In ICML, pages 8748–8763, 2021

    work page 2021

  2. [2]

    Rasheed, M

    H. Rasheed, M. U. Khattak, M. Maaz, S. Khan, and F. S. Khan. Fine-tuned CLIP models are efficient video learners. In CVPR, pages 6545–6554, 2023

    work page 2023

  3. [3]

    Y . Wei, H. Hu, Z. Xie, Z. Liu, Z. Zhang, Y . Cao, J. Bao, D. Chen, and B. Guo. Improving CLIP fine-tuning perfor- mance. In ICCV, pages 5439–5449, 2023

    work page 2023

  4. [4]

    Z. Duan, H. Cheng, D. Xu, X. Wu, X. Zhang, X. Ye, and Z. Xie. Cityllava: E fficient fine-tuning for VLMs in city scenario. In CVPR, pages 7180–7189, 2024

    work page 2024

  5. [5]

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu. Learning to prompt for vision-language models. Int. J. Comput. Vis., 130(9):2337–2348, 2022

    work page 2022

  6. [6]

    K. Zhou, J. Yang, C. C. Loy, and Z. Liu. Conditional prompt learning for vision-language models. In CVPR, pages 16816–16825, 2022

    work page 2022

  7. [7]

    M. U. Khattak, H. Rasheed, M. Maaz, S. Khan, and F. S. Khan. MaPLe: Multi-modal prompt learning. In CVPR, pages 19113–19122, 2023

    work page 2023

  8. [8]

    M. U. Khattak, S. T. Wasim, M. Naseer, S. Khan, M. H. Yang, and F. S. Khan. Self-regulating prompts: Founda- tional model adaptation without forgetting. In ICCV, pages 15190–15200, 2023

    work page 2023

Show all 48 references
  1. [9]

    C. Zhao, Y . Wang, X. Jiang, Y . Shen, K. Song, D. Li, and D. Miao. Learning domain invariant prompt for vision-12 language models. IEEE Trans. Image Process., 33:1348– 1360, 2024

    2024

  2. [10]

    Yang, R.-Y

    L. Yang, R.-Y . Zhang, Y . Wang, and X. Xie. MMA: Multi- modal adapter for vision-language models. InCVPR, pages 23826–23837, 2024

    2024

  3. [11]

    S. Xuan, M. Yang, and S. Zhang. Adapting vision-language models via learning to inject knowledge. IEEE Trans. Image Process., 33:5798–5809, 2024

    2024

  4. [12]

    Z. Li, X. Li, X. Fu, X. Zhang, W. Wang, S. Chen, and J. Yang. PromptKD: Unsupervised prompt distillation for vision-language models. In CVPR, pages 26617–26626, 2024

    2024

  5. [13]

    Y . Wang, X. Jiang, D. Cheng, D. Li, and C. Zhao. Learning hierarchical prompt with structured linguistic knowledge for vision-language models. In AAAI, pages 5749–5757, 2024

    2024

  6. [14]

    Roy and A

    S. Roy and A. Etemad. Consistency-guided prompt learn- ing for vision-language models. In ICLR, 2024

    2024

  7. [15]

    X. Tian, S. Zou, Z. Yang, and J. Zhang. ARGUE: Attribute- guided prompt tuning for vision-language models. In CVPR, pages 28578–28587, 2024

    2024

  8. [16]

    Zhang, K

    Y . Zhang, K. Yu, S. Wu, and Z. He. Conceptual codebook learning for vision-language models. In ECCV, pages 235–251, 2024

    2024

  9. [17]

    C. Jia, Y . Yang, Y . Xia, Y .-T. Chen, Z. Parekh, H. Pham, Q. Le, Y .-H. Sung, Z. Li, and T. Duerig. Scaling up visual and vision-language representation learning with noisy text supervision. In ICML, pages 4904–4916, 2021

    2021

  10. [18]

    J. Li, D. Li, C. Xiong, and S. Hoi. Blip: Bootstrapping language-image pre-training for unified vision-language understanding and generation. In ICML, pages 12888– 12900, 2022

    2022

  11. [19]

    J. Li, D. Li, C. Xiong, and S. Hoi. Blip-2: Bootstrapping language-image pre-training with frozen image encoders and large language models. In ICML, 2023

    2023

  12. [20]

    Alayrac, J

    J.-B. Alayrac, J. Donahue, P. Luc, A. Miech, I. Barr, Y . Hasson, K. Lenc, A. Mensch, K. Millican, M. Reynolds, R. Ring, et al. Flamingo: A visual language model for few-shot learning. In NeurIPS, pages 23716–23736, 2022

    2022

  13. [21]

    J. Lin, H. Yin, W. Ping, P. Molchanov, M. Shoeybi, and S. Han. Vila: On pre-training for visual language models. In CVPR, pages 26689–26699, 2024

    2024

  14. [22]

    K. Su, X. Zhang, S. Zhang, J. Zhu, and B. Zhang. To boost zero-shot generalization for embodied reasoning with vision-language pre-training. IEEE Trans. Image Process., 33:5370–5381, 2024

    2024

  15. [23]

    Lester, R

    B. Lester, R. Al-Rfou, and N. Constant. The power of scale for parameter-efficient prompt tuning. In EMNLP, pages 3045–3059, 2021

    2021

  16. [24]

    Guo and X

    Y . Guo and X. Gu. MMRL: Multi-modal representation learning for vision-language models. arXiv:2503.08497, 2025

    2025 arXiv

  17. [25]

    Z. Lu, J. Bai, X. Li, Z. Xiao, and X. Wang. Task-to-instance prompt learning for vision-language models at test time. IEEE Trans. Image Process., 34:1908–1920, 2025

    1908

  18. [26]

    C. Cui, Z. Liu, S. Gong, L. Zhu, C. Zhang, and H. Liu. When adversarial training meets prompt tuning: Adversar- ial dual prompt tuning for unsupervised domain adaptation. IEEE Trans. Image Process., 34:1427–1440, 2025

    2025

  19. [27]

    OpenAI. ChatGPT. https://www.openai.com/ chatgpt, 2023. [Online; accessed 2023]

    2023

  20. [28]

    Touvron, T

    H. Touvron, T. Lavril, G. Izacard, X. Martinet, M.-A. Lachaux, T. Lacroix, B. Rozière, N. Goyal, E. Hambro, F. Azhar, and A. Rodriguez. LLaMA: Open and efficient foundation language models. arXiv:2302.13971, 2023

    2023 arXiv

  21. [29]

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Milli- can, and D. Silver. Gemini: A family of highly capable multimodal models. arXiv:2312.11805, 2023

    2023 arXiv

  22. [30]

    Anthropic. Claude. https://www.anthropic.com/ product, 2024. [Online; accessed 2024]

    2024

  23. [31]

    Y . Pu, X. Wu, L. Yang, and S. Wang. Learning prompt- enhanced context features for weakly-supervised video anomaly detection. IEEE Trans. Image Process., 33:4923– 4936, 2024

    2024

  24. [32]

    M. U. Khattak, M. F. Naeem, M. Naseer, L. Van Gool, and F. Tombari. Learning to prompt with text only supervision for vision-language models. In AAAI, pages 4230–4238, 2025

    2025

  25. [33]

    J. Deng, W. Dong, R. Socher, L.-J. Li, K. Li, and L. Fei-Fei. ImageNet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009

    2009

  26. [34]

    Fei-Fei, R

    L. Fei-Fei, R. Fergus, and P. Perona. Learning generative visual models from few training examples: An incremental bayesian approach tested on 101 object categories. In CVPRW, pages 178–178, 2004

    2004

  27. [35]

    Krause, M

    J. Krause, M. Stark, J. Deng, and L. Fei-Fei. 3d object representations for fine-grained categorization. In ICCVW, pages 554–561, 2013

    2013

  28. [36]

    O. M. Parkhi, A. Vedaldi, A. Zisserman, and C. V . Jawahar. Cats and dogs. In CVPR, pages 3498–3505, 2012

    2012

  29. [37]

    S. Maji, E. Rahtu, J. Kannala, M. Blaschko, and A. Vedaldi. Fine-grained visual classification of aircraft. arXiv:1306.5151, 2013

    2013 arXiv

  30. [38]

    Nilsback and A

    M.-E. Nilsback and A. Zisserman. Automated flower clas- sification over a large number of classes. InICVGIP, pages 722–729, 2008

    2008

  31. [39]

    Bossard, M

    L. Bossard, M. Guillaumin, and L. Van Gool. Food-101: Mining discriminative components with random forests. In ECCV, pages 446–461, 2014

    2014

  32. [40]

    J. Xiao, J. Hays, K. A. Ehinger, A. Oliva, and A. Torralba. SUN database: Large-scale scene recognition from abbey to zoo. In CVPR, pages 3485–3492, 2010

    2010

  33. [41]

    Cimpoi, S

    M. Cimpoi, S. Maji, I. Kokkinos, S. Mohamed, and A. Vedaldi. Describing textures in the wild. In CVPR, pages 3606–3613, 2014

    2014

  34. [42]

    Helber, B

    P. Helber, B. Bischke, A. Dengel, and D. Borth. EuroSAT: A novel dataset and deep learning benchmark for land use and land cover classification. IEEE J. Sel. Topics Appl. Earth Observ. Remote Sens., 12(7):2217–2226, 2019.13

    2019

  35. [43]

    Soomro, A

    K. Soomro, A. R. Zamir, and M. Shah. UCF101: A dataset of 101 human actions classes from videos in the wild. arXiv:1212.0402, 2012

    2012 arXiv

  36. [44]

    Recht, R

    B. Recht, R. Roelofs, L. Schmidt, and V . Shankar. Do ImageNet classifiers generalize to ImageNet? In ICML, pages 5389–5400, 2019

    2019

  37. [45]

    H. Wang, S. Ge, Z. Lipton, and E. P. Xing. Learning robust global representations by penalizing local predictive power. In NeurIPS, 2019

    2019

  38. [46]

    Hendrycks, K

    D. Hendrycks, K. Zhao, S. Basart, J. Steinhardt, and D. Song. Natural adversarial examples. In CVPR, pages 15262–15271, 2021

    2021

  39. [47]

    Hendrycks, S

    D. Hendrycks, S. Basart, N. Mu, S. Kadavath, F. Wang, E. Dorundo, R. Desai, T. Zhu, S. Parajuli, M. Guo, and D. Song. The many faces of robustness: A critical analysis of out-of-distribution generalization. In ICCV, pages 8340– 8349, 2021

    2021

  40. [48]

    Paszke, S

    A. Paszke, S. Gross, F. Massa, A. Lerer, J. Bradbury, G. Chanan, T. Killeen, Z. Lin, N. Gimelshein, L. Antiga, A. Desmaison, A. Köpf, E. Yang, Z. DeVito, M. Raison, A. Tejani, S. Chilamkurthy, B. Steiner, L. Fang, J. Bai, and S. Chintala. PyTorch: An imperative style, high- pe...

    2019

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.