Pith. sign in

REVIEW 2 major objections 8 minor 33 references

Swapping softmax for entmax in CLIP's final layers sharpens dense prediction

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · glm-5.2

2026-07-09 19:13 UTC pith:JRB2HAAG

load-bearing objection Solid empirical study with a useful finding, but the central proportionality claim is under-quantified. the 2 major comments →

arxiv 2607.07135 v1 pith:JRB2HAAG submitted 2026-07-08 cs.CV

Sparse Attention for Dense Open-Vocabulary Prediction in CLIP

Fatimah Zohra , Chen Zhao , Shuming Liu , Bernard Ghanem This is my paper
classification cs.CV
keywords CLIPsparse attentionalpha-entmaxopen-vocabulary segmentationdense predictionself-correlation attentionsoftmax sparsificationvision-language models
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

CLIP's contrastive training objective optimizes for image-level semantics, leaving the spatially localized features that dense prediction tasks need buried under coarse signals. This paper argues that the softmax normalizer in CLIP's final attention layers compounds the problem: because softmax assigns strictly positive weight to every token, each patch becomes a global mixture diluted by a long tail of irrelevant background and non-class tokens. The authors replace softmax with the alpha-entmax transform—a generalization that applies a data-dependent threshold mapping low-relevance attention scores exactly to zero while redistributing mass onto the most relevant tokens. This is done at inference time on a frozen CLIP encoder, adding no parameters and leaving pretrained weights untouched. The central finding is that the gain from this sparsification is proportional to how diffuse the underlying attention distribution is: where softmax spreads mass across many non-class tokens, entmax yields large improvements; where attention is already concentrated, it offers no benefit or even degrades performance. This proportionality holds across standard query-key attention, self-correlation variants (query-query, key-key, value-value), two model scales (ViT-B/16 and ViT-L/14), and both pixel-level segmentation and region-level fine-grained retrieval. The benefit widens at higher resolutions and on larger backbones, because the noisy tail grows with token count.

Core claim

The gain from attention sparsification in frozen CLIP is governed by a single quantitative property of the baseline attention distribution: how much probability mass each patch spreads over non-class tokens. When that tail is large, replacing softmax with alpha-entmax in the final layers denoises the distribution and sharpens dense predictions substantially; when the mass is already concentrated on same-class tokens, sparsification has nothing to prune and can hurt. This means the failure mode of CLIP's dense features can be partially diagnosed and predicted by measuring attention diffuseness, and the fix is a parameter-free, inference-time substitution that isolates attention density as the

What carries the argument

The alpha-entmax transform, which replaces the row-wise softmax in the final self-attention layers. Entmax solves a Tsallis-entropy-regularized optimization that yields a thresholded closed form: a single scalar threshold tau separates active coordinates from those mapped exactly to zero, with alpha controlling the entropy regularization (alpha approaching 1 recovers softmax, alpha equals 2 gives sparsemax). The threshold is data-dependent—computed per attention row from the sorted scores—so it adapts to each query's distribution and prunes only the low-relevance tail rather than applying a fixed sparsity pattern.

Load-bearing premise

The paper assumes that the noise obscuring dense predictions in CLIP originates primarily in the softmax normalizer of the final attention layers, manifesting as a low-relevance tail that can be cleanly separated and zeroed by a data-dependent threshold. If the degradation in dense features is instead caused by the representational geometry of the value vectors or earlier-layer mixing, sparsifying the final attention distribution alone would not produce the observed gains.

What would settle it

If one could show that the same dense-prediction gains are achievable by simply sharpening softmax with a fixed temperature (no zeroing), or by applying a random mask of matched sparsity, then the data-dependent thresholding mechanism of entmax would not be the operative cause and the paper's central mechanistic claim would fail. The paper addresses this with control experiments showing entmax outperforms both alternatives.

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

If this is right

  • If the diffuseness-to-gain proportionality generalizes, one could build a diagnostic that predicts whether sparsification will help a given model or layer by measuring how much attention mass falls on non-class tokens, without running the full downstream task.
  • The finding that the residual connection in the final block impedes localization more than the FFN suggests that architectural denoising (removing the residual) and attention sparsification target the same noise source from different angles, and could be combined or made complementary.
  • Because the benefit grows with token count, the method becomes more valuable as vision-language models scale to higher resolutions and larger patch grids, where softmax attention becomes increasingly diffuse.
  • The proportionality principle could extend to other transformer-based vision models whose dense features degrade due to global objectives, provided their final-layer attention exhibits the same diffuse-tail signature.

Where Pith is reading between the lines

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

  • If the value vectors themselves encode spatially degraded features (independent of the attention weights), then sparsifying attention would not help—but the paper's observed gains on self-correlated distributions suggest the value representations retain usable structure that is merely being diluted by the attention mixing, not destroyed.
  • A natural next step would be to learn alpha per-head rather than using a fixed value, since different heads may exhibit different diffuseness profiles and thus benefit from different sparsity strengths.
  • The proportionality finding implies a potential failure mode for entmax on models or layers where attention is already sharp: it could zero out relevant mass. This suggests an adaptive scheme that applies entmax only when measured diffuseness exceeds a threshold.
  • The interaction between sparsification and resolution scaling raises the question of whether entmax could enable efficient high-resolution inference by allowing aggressive sparsity without the quadratic cost penalty of dense softmax attention.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

2 major / 8 minor

Summary. This paper studies a training-free, inference-time modification to frozen CLIP visual encoders: replacing the row-wise softmax in the final self-attention layers with the α-entmax transform. The motivation is that softmax assigns strictly positive weight to every token pair, spreading attention across low-relevance background patches and obscuring the fine-grained, spatially localized features needed for dense prediction. Because entmax applies a data-dependent threshold that maps low scores exactly to zero, it acts as an implicit denoiser. The authors evaluate this substitution across multiple attention score variants (standard qkv and self-correlation variants such as qqv, kkv, vvv), two backbones (ViT-B/16, ViT-L/14), and two task families (open-vocabulary semantic segmentation on VOC, Context, ADE20K; fine-grained region-text retrieval on FG-OVD). The central empirical finding is that the gain from sparsification is proportional to how diffuse the baseline attention distribution is: distributions that spread mass off the target class benefit more, while already-concentrated distributions (e.g., vvv on ViT-B/16) gain little or degrade. The paper also provides ablations against random masking and temperature sharpening, depth sweeps, resolution sweeps, and architectural ablations of the final transformer block.

Significance. The paper addresses a well-recognized problem in the CLIP dense-prediction literature: the final attention layers spread probability mass globally, degrading per-patch discrimination. While prior training-free methods (MaskCLIP, SCLIP, GEM, ClearCLIP, NACLIP) modify the score source or value path, they all retain the softmax normalizer. This paper isolates the normalizer as a variable and shows that a principled, parameter-free sparsification (α-entmax) yields consistent gains on self-correlated distributions. The contribution is primarily analytical rather than a new state-of-the-art method, but the systematic study is valuable: the entmax substitution is parameter-free, leaves pretrained weights untouched, and the ablations (Table 3: random mask vs. temperature vs. entmax) cleanly isolate the data-dependent threshold as the source of benefit. The finding that gains scale with resolution and backbone size (Tables 5–6) is a falsifiable, practically useful observation. The work is a well-scoped addition to the training-free CLIP dense-prediction literature.

major comments (2)
  1. §1 and §5: The paper's central claim is that entmax gain is 'proportional to how diffuse the underlying attention is.' The evidence for this is indirect: cross-distribution comparisons (Table 1: qkv gains more than vvv) confound diffuseness with other properties that differ between score sources (grouping quality, value geometry, head specialization), and resolution/backbone scaling (Tables 5–6) confounds diffuseness with token count and model capacity. The one piece of direct per-patch evidence—Figure 1 (right), plotting per-patch entmax-minus-softmax gain against off-target attention fraction—is presented without a correlation coefficient, regression line, sample size, or variance estimate. It is therefore impossible to assess whether this is a strong linear relationship or a weak, high-variance trend. Quantifying this plot (Pearson/Spearman correlation, regression, N, confidence bands
  2. §5.1 and Table 1: The vvv-on-ViT-B/16 exception (where entmax degrades performance) is explained post-hoc as 'already concentrated, so there is no tail to prune.' This explanation is plausible but not isolated from alternatives: the value-path geometry could be insensitive to attention sparsification for reasons unrelated to diffuseness (e.g., value vectors already carry localized information regardless of attention weighting). The paper does not hold diffuseness fixed while varying other factors, so the proportionality claim is underdetermined. A controlled experiment—e.g., artificially diffusing the vvv distribution on B/16 and showing that entmax then helps—would strengthen the causal claim. At minimum, the authors should acknowledge that the cross-distribution argument conflates diffuseness with score-source identity.
minor comments (8)
  1. §5.1: 'yeilds' should be 'yields.'
  2. §5.1 and §5.2: 'effected' should be 'affected' (e.g., 'is not effected as much' → 'is not affected as much').
  3. §3.1: The choice of α=1.2 as the default is used throughout but the justification is only empirical (Figure 5 sweep). A brief note on why α=1.2 is a reasonable operating point (vs. α=1.5 used in some NLP work) would help reproducibility.
  4. Figure 1 (left): The caption mentions a green × marker for the query patch, but the marker is difficult to locate in the figure. A more prominent marker or arrow would help.
  5. Table 1: The 'v-only' reference row appears in Tables 5–6 but not in Table 1, making cross-table comparison harder. Adding it or noting its absence would keep tables self-contained.
  6. §4.1: The background threshold for VOC is swept and the best is reported, but the actual threshold value(s) used are not stated. Reporting the selected threshold would aid reproducibility.
  7. Reference 33 (Zohra et al. 2026) is a self-citation that appears in the evaluation protocol context. It is not load-bearing for the central claim, but the authors should ensure it is not presented as a prerequisite result.
  8. Figure 5 caption: 'Sweeps α' is slightly ambiguous about whether α is the entmax parameter or a learning rate. Consistent notation would help.

Circularity Check

0 steps flagged

No circularity: the paper applies an externally defined mathematical transform (α-entmax) to externally provided models (CLIP) and evaluates on external benchmarks.

full rationale

The paper's central claim—that entmax gain is proportional to baseline attention diffuseness—is an empirical hypothesis tested against external benchmarks (Pascal VOC, Pascal Context, ADE20K, FG-OVD) using externally defined mathematical tools (α-entmax from Blondel et al. 2019, Peters et al. 2019) and externally provided pretrained models (CLIP ViT-B/16, ViT-L/14). The derivation of entmax (§3.1) is standard and self-contained, recovering softmax as α→1 and sparsemax at α=2, with no re-derivation tailored to fit the results. The one self-citation (ref 33, Zohra et al. 2026, beta-CLIP) appears only in a list of evaluation-protocol references and is not load-bearing for the sparsification claim. No equation or definition reduces to its own inputs by construction. The proportionality claim is an empirical observation, not a derivation: the paper does not define diffuseness in terms of entmax gain or vice versa. While the skeptic correctly notes that the evidence for proportionality is indirect and confounded, that is a correctness/evaluative concern, not a circularity concern. The derivation chain is self-contained against external benchmarks with no self-citation chain forcing the result.

Axiom & Free-Parameter Ledger

3 free parameters · 3 axioms · 0 invented entities

The paper introduces no new entities or forces. The free parameters (α, depth, threshold) are standard hyperparameters tuned on validation data. The axioms are domain assumptions from prior CLIP literature or standard mathematical properties.

free parameters (3)
  • α (entmax sparsity parameter) = 1.2
    Chosen empirically; the paper sweeps α and finds 1.2 optimal. It is a hyperparameter selected based on performance.
  • Number of final layers sparsified (depth) = 1-2
    Selected empirically via the depth sweep in Figure 5. Not derived from first principles.
  • Background threshold (VOC) = 0.85 (mentioned in Fig 4 caption)
    Swept for best mIoU on Pascal VOC as stated in §4.1.
axioms (3)
  • domain assumption CLIP's dense-prediction noise originates primarily in the last layers' attention normalizer.
    Stated in §1 and §3. This motivates intervening only in the final layers. If noise were in earlier layers or value vectors, the method would not work.
  • domain assumption Self-correlation attention distributions (qq, kk, vv) localize better than qk for dense prediction.
    Adopted from prior work (GEM, SCLIP, etc.) and used as the basis for evaluating entmax across these variants.
  • standard math The α-entmax transform provides a data-dependent threshold that correctly identifies irrelevant tokens.
    Mathematical property of entmax from Blondel et al. 2019. Used as the denoising mechanism.

pith-pipeline@v1.1.0-glm · 18787 in / 2190 out tokens · 334146 ms · 2026-07-09T19:13:19.777955+00:00 · methodology

0 comments
read the original abstract

Contrastive Language-Image Pre-training (CLIP) relies on softmax-based self-attention, a strictly positive distribution that assigns probability mass to every pair of tokens-even semantically irrelevant ones. While these dense softmax weights are effective for gathering broad context during pre-training, they spread attention across many low-salience tokens, producing noise that obscures the fine-grained, spatially localized cues required for dense, open-vocabulary prediction. We study an inference-time substitution of the row-wise softmax in the final visual self-attention layers with the $\alpha$-entmax transform, applied across both the standard query-key attention and self-correlation variants. Because entmax applies a data-dependent threshold that maps low scores exactly to zero, it acts as an implicit denoiser, zeroing contextually irrelevant dependencies while redistributing mass onto the most relevant tokens. We evaluate on open-vocabulary tasks-dense semantic segmentation (Pascal VOC, Pascal Context, ADE20K) and fine-grained retrieval (FG-OVD)-and find the gain from attention sparsification is proportional to how much the baseline attention spreads off the target class.

Figures

Figures reproduced from arXiv: 2607.07135 by Bernard Ghanem, Chen Zhao, Fatimah Zohra, Shuming Liu.

Figure 1
Figure 1. Figure 1: Attention Mass in Softmax vs. Entmax Distributions. Left: [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-patch attention diagnostics. CLIP ViT-B/16, last block. For a central query patch shown as (green ×), we visualise the attention induced by each distribution— the standard qk and the self-correlated qq, kk, vv—under softmax (top row of each panel) and entmax (α=1.2, bottom row). Titles report the effective number of attended keys. The qk attention is globally diffuse throughout the image, whereas the s… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative Visualization of Segmentation Maps and Heatmaps. [PITH_FULL_IMAGE:figures/full_fig_p014_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Entmax sparsity vs. depth on Pascal VOC on ViT-L/14 [PITH_FULL_IMAGE:figures/full_fig_p019_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Additional per-patch attention diagnostics. [PITH_FULL_IMAGE:figures/full_fig_p020_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Additional per-patch attention diagnostics. [PITH_FULL_IMAGE:figures/full_fig_p021_7.png] view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

33 extracted references · 33 canonical work pages · 13 internal anchors

  1. [1]

    The devil is in the fine-grained details: Evaluating open-vocabulary object detectors for fine-grained understanding

    Bianchi, L., Carrara, F., Messina, N., Gennaro, C., Falchi, F.: The devil is in the fine-grained details: Evaluating open-vocabulary object detectors for fine-grained understanding. arXiv (2023).https://doi.org/10.48550/arxiv.2311.17518

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2311.17518 2023

  2. [2]

    In: Chaudhuri, K., Sugiyama, M

    Blondel, M., Martins, A.F.T., Niculae, V.: Learning classifiers with fenchel-young losses: Generalized entropies, margins, and algorithms. In: Chaudhuri, K., Sugiyama, M. (eds.) Proceedings of the 22nd International Conference on Artificial Intelligence and Statistics. Proceedings of Machine Learning Research, vol. 89, pp. 606–615. PMLR (April16–18 2019),...

    work page 2019

  3. [3]

    In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)

    Bousselham, W., Petersen, F., Ferrari, V., Kuehne, H.: Grounding everything: Emerging localization properties in vision-language transformers. In: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR). pp. 3828–3837 (June 2024)

    work page 2024

  4. [4]

    In: Proc

    Correia, G.M., Niculae, V., Martins, A.F.: Adaptively sparse transformers. In: Proc. EMNLP-IJCNLP (to appear) (2019)

    work page 2019

  5. [5]

    Dong, X., Bao, J., Zheng, Y., Zhang, T., Chen, D., Yang, H., Zeng, M., Zhang, W., Yuan, L., Chen, D., Wen, F., Yu, N.: Maskclip: Masked self-distillation advances contrastive language-image pretraining (2023)

    work page 2023

  6. [6]

    International Journal of Computer Vision111(1), 98–136 (Jan 2015)

    Everingham, M., Eslami, S.M.A., Van Gool, L., Williams, C.K.I., Winn, J., Zisser- man, A.: The pascal visual object classes challenge: A retrospective. International Journal of Computer Vision111(1), 98–136 (Jan 2015)

    work page 2015

  7. [7]

    Fang, Y., Wang, W., Xie, B., Sun, Q., Wu, L., Wang, X., Huang, T., Wang, X., Cao, Y.: Eva: Exploring the limits of masked visual representation learning at scale (2022)

    work page 2022

  8. [8]

    In: Proceedings of the Winter Conference on Applications of Computer Vision (WACV)

    Hajimiri, S., Ben Ayed, I., Dolz, J.: Pay attention to your neighbours: Training-free open-vocabulary semantic segmentation. In: Proceedings of the Winter Conference on Applications of Computer Vision (WACV). pp. 5061–5071 (February 2025)

    work page 2025

  9. [9]

    Jia, C., Yang, Y., Xia, Y., Chen, Y.T., Parekh, Z., Pham, H., Le, Q.V., Sung, Y., Li, Z., Duerig, T.: Scaling up visual and vision-language representation learning with noisy text supervision (2021)

    work page 2021

  10. [10]

    In: The Thirty-eighth Annual Conference on Neural Information Processing Systems (2024), https://openreview.net/forum?id=nExI4FuKWD

    Jing, D., He, X., Luo, Y., Fei, N., Yang, G., Wei, W., Zhao, H., Lu, Z.: FineCLIP: Self-distilled region-based CLIP for better fine-grained understanding. In: The Thirty-eighth Annual Conference on Neural Information Processing Systems (2024), https://openreview.net/forum?id=nExI4FuKWD

    work page 2024

  11. [11]

    In: European Conference on Computer Vision

    Lan, M., Chen, C., Ke, Y., Wang, X., Feng, L., Zhang, W.: Clearclip: Decomposing clip representations for dense vision-language inference. In: European Conference on Computer Vision. pp. 143–160. Springer (2024)

    work page 2024

  12. [12]

    In: CVPR (2022)

    Li, F., et al.: Grounded language-image pre-training. In: CVPR (2022)

    work page 2022

  13. [13]

    Li, Y., Wang, H., Duan, Y., Zhang, J., Li, X.: A closer look at the explainability of contrastive language-image pre-training (2024),https://arxiv.org/abs/2304. 05653

    work page 2024

  14. [14]

    Ma, Z., Hong, J., Gul, M.O., Gandhi, M., Gao, I., Krishna, R.: Crepe: Can vision- language foundation models reason compositionally? (2023),https://arxiv.org/ abs/2212.07796

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  15. [15]

    02068 16 F

    Martins, A.F.T., Astudillo, R.F.: From softmax to sparsemax: A sparse model of attention and multi-label classification (2016),https://arxiv.org/abs/1602. 02068 16 F. Zohra et al

    work page 2016

  16. [16]

    Martins, P.H., Marinho, Z., Martins, A.F.T.: Sparse text generation (2020),https: //arxiv.org/abs/2004.02644

    work page internal anchor Pith review Pith/arXiv arXiv 2020

  17. [17]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2014)

    Mottaghi, R., Chen, X., Liu, X., Cho, N.G., Lee, S.W., Fidler, S., Urtasun, R., Yuille, A.: The role of context for object detection and semantic segmentation in the wild. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR) (June 2014)

    work page 2014

  18. [18]

    Peters, B., Niculae, V., Martins, A.F.T.: Sparse sequence-to-sequence models (2019), https://arxiv.org/abs/1905.05702

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  19. [19]

    Qiu, C., Wu, Y., Ke, W., Bai, X., Zhang, T.: Refining clip’s spatial awareness: A visual-centric perspective (2025),https://arxiv.org/abs/2504.02328

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  20. [20]

    ICML (2021)

    Radford, A., Kim, J.W., Hallacy, C., et al.: Learning transferable visual models from natural language supervision. ICML (2021)

    work page 2021

  21. [21]

    Ranasinghe, K., McKinzie, B., Ravi, S., Yang, Y., Toshev, A., Shlens, J.: Perceptual grouping in contrastive vision-language models (2023),https://arxiv.org/abs/ 2210.09996

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  22. [22]

    Shao,T.,Tian,Z.,Zhao,H.,Su,J.:Explorethepotentialofclipfortraining-freeopen vocabulary semantic segmentation (2024),https://arxiv.org/abs/2407.08268

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  23. [23]

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A.N., Kaiser, L., Polosukhin, I.: Attention is all you need (2023)

    work page 2023

  24. [24]

    Voita, E., Talbot, D., Moiseev, F., Sennrich, R., Titov, I.: Analyzing multi-head self-attention: Specialized heads do the heavy lifting, the rest can be pruned (2019), https://arxiv.org/abs/1905.09418

    work page internal anchor Pith review Pith/arXiv arXiv 2019

  25. [25]

    SCLIP: Rethinking Self-Attention for Dense Vision-Language Inference

    Wang, F., Mei, J., Yuille, A.: Sclip: Rethinking self-attention for dense vision- language inference. arXiv preprint arXiv:2312.01597 (2023)

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  26. [26]

    Wu, S., Zhang, W., Xu, L., Jin, S., Li, X., Liu, W., Loy, C.C.: Clipself: Vision transformer distills itself for open-vocabulary dense prediction (2024),https:// arxiv.org/abs/2310.01403

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  27. [27]

    In: CVPR (2022)

    Wu, X., et al.: Regionclip: Region-based language Image pre-training. In: CVPR (2022)

    work page 2022

  28. [28]

    Xie, C., Wang, B., Kong, F., Li, J., Liang, D., Zhang, G., Leng, D., Yin, Y.: Fg-clip: Fine-grained visual and textual alignment (2025),https://arxiv.org/abs/2505. 05071

    work page 2025

  29. [29]

    Yuksekgonul, M., Bianchi, F., Kalluri, P., Jurafsky, D., Zou, J.: When and why vision-language models behave like bags-of-words, and what to do about it? (2023), https://arxiv.org/abs/2210.01936

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  30. [30]

    Zhai, X., Mustafa, B., Kolesnikov, A., Beyer, L.: Sigmoid loss for language image pre-training (2023),https://arxiv.org/abs/2303.15343

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  31. [31]

    Zhou, B., Zhao, H., Puig, X., Xiao, T., Fidler, S., Barriuso, A., Torralba, A.: Semanticunderstandingofscenesthroughtheade20kdataset.Int.J.Comput.Vision 127(3), 302–321 (Mar 2019).https://doi.org/10.1007/s11263- 018- 1140- 0 , https://doi.org/10.1007/s11263-018-1140-0

    work page doi:10.1007/s11263- 2019

  32. [32]

    Zhou, Y., Fritz, M., Keuper, M.: Multimax: Sparse and multi-modal attention learning (2025),https://arxiv.org/abs/2406.01189

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    arXiv (2026).https://doi

    Zohra, F., Zhao, C., Itani, H., Ghanem, B.:beta-CLIP: Text-conditioned contrastive learning for multi-granular vision-language alignment. arXiv (2026).https://doi. org/10.48550/arxiv.2512.12678 Sparse Attention in CLIP 17 A Extended Quantitative Analysis A.1 Resolution Scaling for Dense Segmentation on ViT-B/16 Table 7 extends the resolution sweep to dens...

    work page doi:10.48550/arxiv.2512.12678 2026