Pith. sign in

REVIEW 4 major objections 5 minor 37 references

Test-Time Registers as Global Priors for Tokenized Image Generation

T0 review · 4 major / 5 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Vision transformers' outlier 'register' tokens can serve, training-free, as global priors that improve tokenized image generation.

desk verdict Solid diagnostic core and an honest but modest generation experiment; the main gap is an underspecified cross-model codebook alignment that leaves the training-free claim unverifiable. read the letter →

arxiv 2607.16824 v2 pith:EIXXVEEC submitted 2026-07-18 cs.CV

Cheng-Yao Hong , Yifan Wang , Yuewei Lin , Chenyu You This is my paper · ORCID
classification cs.CV
keywords attentionsinksregistersvisiontransformerstokenizedimagegenerationtraining-freeglobalpriorslow-frequencystructuretest-timeoptimization
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

RegToken asks whether the high-norm outlier tokens that vision transformers accumulate—often dismissed as attention sinks—actually carry usable global image structure. The paper argues they do: register-associated features concentrate energy in low frequencies more than either [CLS] readouts or patch means, and correlate with pixel-space low-frequency content. On that basis, RegToken extracts a compact register subspace from a frozen backbone and maps it into a frozen compact-token generation pipeline, improving ImageNet FID-5k from 20.5 to 20.1, SigLIP from 3.6 to 3.9, and cutting optimization steps to threshold from 74 to 52—all with no weight updates. The claim is that what look like artifacts are in fact lightweight global priors that a frozen decoder can exploit.

What carries the argument

The load-bearing object is the test-time register token: a single auxiliary token constructed from a frozen vision transformer's register subspace. It is produced by (i) NFN-based layer selection, which locates layers whose outputs are dominated by a few feature directions; (ii) TokenRank-guided subspace extraction, which ranks channels by their activity on outlier tokens and by their stationary attention centrality; and (iii) a projection-and-conservation update that moves activation energy from outlier patches into the register token while preserving the mean projection magnitude. The resulting vector is mapped to the decoder's discrete codebook by cosine-similarity nearest neighbor or a l

What would settle it

Run the identical RegToken pipeline but replace the extracted register vector with one whose feature dimensions are randomly permuted while keeping its norm; if FID-5k and SigLIP stay at 20.1/3.9 instead of degrading toward the no-prior baseline (21.2/3.5), the claim that the structured content matters would be refuted. A second check: substitute a patch-mean vector through the same codebook projection; if it matches RegToken's gains, registers are not special.

Watch

Extended reading notes

Core claim

The central discovery is that test-time register structure—feature directions associated with high-norm outlier tokens in a frozen ViT—encodes smooth, scene-level statistics such as illumination, color tone, and coarse layout, and that this structure can be captured by a three-stage procedure: NFN-based layer localization, TokenRank-guided subspace extraction, and a projection-and-conservation update that transfers register-subspace energy from outlier patches to a dedicated token. Mapped into a discrete codebook via nearest-neighbor or Procrustes alignment and inserted as a global token into a frozen 1D-token decoder, this training-free prior improves generation quality and text-image align

Load-bearing premise

The method assumes that a register feature vector extracted from a frozen vision backbone, after being projected into the discrete codebook of an unrelated frozen tokenizer, still carries the global structure that produced the generation gains.

Editorial extensions

If this is right

  • If registers really carry global low-frequency structure, the same extraction can serve as a cheap global prior for other frozen generative pipelines without retraining.
  • Inserting register priors improves convergence of test-time token optimization, suggesting they give the optimizer a better initialization.
  • Because the prior is extracted from a frozen backbone and mapped by codebook lookup, it can be applied to new decoders as long as a light alignment is available.
  • The low-frequency concentration diagnostic offers a principled way to choose where and what to inject, replacing hand-picked heuristics.
  • The gains transfer across backbone scales and token budgets, indicating the phenomenon is not peculiar to one architecture.

Reading between the lines

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

  • If register features are truly global low-frequency summaries, a testable extension is to manipulate the register token directly for style transfer or illumination editing, rather than using it only as a fidelity prior.
  • The same-image prior in the main protocol is a privileged setting; the paper's own cross-image results suggest a class-level prior, so a follow-up could measure how much of the gain survives under purely text-conditional generation with no image-derived prior at all.
  • Because the projection relies on cosine similarity to a codebook trained for reconstruction, the effect's strength likely depends on how well the backbone's global statistics align with the tokenizer's codebook; a systematic study varying tokenizer training data could bound the method's portability.
  • The method can be interpreted as a soft, training-free form of conditioning: even 'artifactual' internal tokens carry a usable bottleneck of global information worth reusing in generation.
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 / 5 minor

Summary. The paper studies test-time register tokens in vision transformers and asks whether the global, low-frequency structure they encode can be operationalized for tokenized image generation without retraining. The authors first present diagnostic evidence (Table 1) that register-token features have stronger low-frequency spectral concentration than [CLS] or patch-mean features across DINOv2 and OpenCLIP. They then propose RegToken, a training-free method that localizes register structure via NFN-based layer selection, identifies register-relevant channels with TokenRank, and constructs a register token via a projection-and-conservation update. This token is mapped to the codebook of a frozen HCT-style decoder and used as a global prior, with optional optimization of the inserted token only. The paper reports improvements on ImageNet generation metrics (e.g., FID-5k 20.5 to 20.1, SigLIP 3.6 to 3.9) and faster test-time optimization, while keeping the backbone and decoder frozen.

Significance. If the claims hold, the paper makes a useful contribution by turning an often-dismissed attention-sink phenomenon into a plug-and-play generative prior. The diagnostic section is careful: the low-frequency analysis is statistically grounded, uses paired tests, and is complemented by a pixel-space DCT correlation in the supplement. The method is also extensively ablated, and the supplement includes causal controls (low- vs. high-frequency decomposition, shuffled priors, matched-norm random priors) that strengthen the interpretation. The main limitation is that the headline generation gain is obtained in a same-image prior setting, which is acknowledged in the supplement but not in the abstract, and the cross-model codebook alignment in Eq. (10) is underspecified. The value of the paper therefore depends on whether the transferable-prior interpretation can be established outside the same-image diagnostic and whether the alignment step is exactly defined.

major comments (4)
  1. [§4.4, Eq. (10) and §5.2] The cross-model codebook alignment is underspecified. The text mentions 'whitening followed by nearest-neighbor search' and 'linear Procrustes mapping,' but gives no equations, no dimensionalities of r^(ℓ) vs. c_j, no data used to estimate any whitening covariance or Procrustes matrix, and no train/test split. If the mapping is fitted on ImageNet data, the method is not training-free; if it is fixed or random, the gains may be coincidental. Because Eq. (10) is the interface between the backbone and the decoder, the entire generation pipeline depends on this step. The authors should specify the exact operator, state whether any statistics are estimated and on which data, report sensitivity to the alignment choice, and provide reproducible code for this step.
  2. [§5, Table 4 and Table 5] The headline improvements (FID-5k 20.5→20.1, SigLIP 3.6→3.9) are obtained with a same-image prior: the register is extracted from the same real image whose class is used for generation. Table 5 shows a clear ordering (same-image 20.1 < same-class 20.6 < random 21.1), confirming that the prior source matters. The supplement calls this 'a source-conditioned diagnostic setting' rather than a zero-shot protocol, but the abstract and introduction present the gain without this caveat. Moreover, the Table 4 comparison against [CLS] prior is also same-image, so the claimed advantage over [CLS] may be specific to privileged target-image information. The authors should report cross-image same-class and random-source comparisons against the [CLS] prior under the identical protocol, and clearly label the same-image setting as an upper-bound diagnostic in the abstract and main text.
  3. [Table 4 and general evaluation] All FID-5k numbers are reported as point estimates with no error bars or repeated evaluations. Given that the reported gains are small (0.4 FID, 0.3 SigLIP) and FID-5k is known to have high variance, these differences may not be statistically meaningful. The supplement reports variance across seeds for CLIPScore trajectories but not for FID, IS, CLIP, or SigLIP. The authors should report mean ± std over at least 3–5 independent decoding runs for the main comparisons, or use a larger FID evaluation set, so the reader can judge whether the improvements exceed noise.
  4. [§4.3, Eq. (9)] The projection-and-conservation update is stated qualitatively: α is set to 'approximately preserve the mean projection magnitude,' but the exact formula for α is not given. Since α is one of many hyperparameters (κ, n, k, w, Δ, s, γ, ∼α, λ, η, h), the method's 'training-free' status and reproducibility would benefit from an explicit definition of α in terms of the computed statistics. The robustness ablations are reassuring, but the conservation step should be specified exactly rather than left as an approximation.
minor comments (5)
  1. [Abstract] The abstract states 'FID-5k 20.5 to 20.1' without noting that 20.5 is the [CLS]-prior baseline, not the no-prior baseline. This should be clarified.
  2. [Figure 2] The caption says 'PCA-projected token features and their 1-D FFT spectra along the feature dimension.' It would be clearer to state explicitly that the FFT is applied after PCA projection and to define the number of PCA components.
  3. [Table 2] The column header 'norm gap (Attention)' is ambiguous; it seems to list attention values but the units and definition should be stated in the caption.
  4. [§5.2] The sentence 'It reaches the same CLIPScore threshold in fewer steps (Steps@τ: 74→52)' refers to Table S2 but the τ definition appears only in Appendix A.3. A brief definition in the main text would help.
  5. [Appendix A.4] The text says 'we report FID-5k, IS, CLIP, SigLIP, and trajectory-based metrics under identical settings for all compared priors,' but the main paper does not show error bars for these metrics; the supplement should state explicitly which metrics are computed once versus averaged over seeds.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: RegToken's register diagnostics, token construction, and frozen-decoder evaluation are independently specified; the same-image prior is disclosed as a source-conditioned diagnostic, and remaining alignment under-specification is a reproducibility concern, not a circular reduction.

full rationale

The paper's claimed derivation chain is not circular in any load-bearing step. The motivating diagnostic (Section 3.1, Table 1, Fig. 2, Table S3) measures low-frequency concentration of register features and correlates it with pixel-space DCT energy; these measurements are independent of the downstream generation metrics (FID, IS, CLIP, SigLIP). RegToken's construction is specified as a forward procedure on frozen backbone activations: NFN-based layer scoring (Eqs. 3-5), token-level outlier detection, neuron-level channel ranking (Eq. 6), and projection-and-conservation updates (Eqs. 8-9). No equation defines the output as the input, and the method does not fit a parameter to the generation metric and then 'predict' that same metric. The generation evaluation uses an external frozen HCT/TiTok-style decoder and codebook, and compares RegToken against no-prior, random-prior, [CLS]-prior, test-time-register, and trained-register baselines under an identical protocol (Table 4). The same-image prior is explicitly and repeatedly disclosed as a 'source-conditioned diagnostic setting' and 'upper-bound analysis' (Section 5.2, Appendix A.4), with cross-image same-class and random-image controls (Table 5) showing transferability rather than a by-construction advantage. The cited prior work on registers (Darcet et al. 2024; Jiang et al. 2025) is external and is used as motivation, not as an author-imported uniqueness theorem; the paper's many author-overlapping citations appear only as related work and are not load-bearing. The manuscript even states its own limitations (Appendix H.3), including scope limits and cases where the prior helps less, which is inconsistent with a circularly forced result. The main remaining concern is under-specification of the cross-model alignment in Section 5.2 ('whitening followed by nearest-neighbor search or a linear Procrustes mapping' is named but not defined, and Eq. 10 omits any projection). This is a reproducibility and potential-fitting risk, but the paper does not provide enough detail to show that the claimed prediction reduces by construction to a fitted input, so it does not constitute demonstrated circularity under the stated rules.

Assumptions & free parameters 12 free parameters · 5 assumptions · 1 invented entities

The central claim rests on a set of validation-tuned hyperparameters (kappa, n, k, w, Delta, s, gamma, beta, alpha, lambda, eta, h) and on domain assumptions about NFN-localized register structure and cross-model codebook transfer. The only invented entity is the auxiliary RegToken itself, which is a method artifact rather than a natural-kind entity, so its lack of independent evidence is less damaging than the 'graviton' case but still worth noting.

free parameters (12)
  • kappa (number of NFN-selected layers) = 5 (default; 1/3/5 swept)
    Selected on ImageNet-val FID; Table S6 shows sensitivity.
  • n (number of outlier patch tokens) = 3-8
    Used to define outlier token set; Section 4.1 defaults.
  • k (number of register neurons/channels) = 32 (default; 16/64/128/256/512 swept)
    Defines register subspace; Table S5 and S6 show validation-based selection.
  • w (window size for outlier detection) = 3
    Window around candidate layer; Section 4.1 default.
  • Delta (preceding layers for channel ranking) = 2
    Number of earlier layers averaged in Eq. (6).
  • s (register scaling factor) = 2.0 (default; 0.5-16 swept)
    Controls reallocation strength in Eq. (9); tuned on validation.
  • gamma (initialization strength) = validation choice from {0.25, 0.5, 0.75, 1.0}
    Strength of prior mixing in Eq. (11); selected on ImageNet-val.
  • beta (soft-bias strength) = 1.0 (default; 0.5/1/2 swept)
    Temperature for cosine prior in Eq. (12); validation-selected.
  • alpha (conservation coefficient) = not specified
    Eq. (9) says 'we set alpha to approximately preserve' the mean projection; no closed form is given.
  • lambda (regularization in Eq. 15) = 0.1-1
    Regularization strength during token optimization; validation-swept.
  • eta (optimization step size) = 1e-3 to 5e-3
    Step size for token optimization; validation-swept.
  • h (number of gated heads) = 1-3
    Top TokenRank heads used for head gating; Section E.5.
assumptions (5)
  • standard math Softmax attention and the attention matrix as a Markov chain with a stationary distribution.
    Used in TokenRank and spectral-gap analysis (Section 4.2).
  • standard math FFT/DCT spectral analysis is a valid way to measure 'smoothness' of token embeddings.
    Used throughout the diagnostic in Section 3.1 and Appendix C.
  • domain assumption High-NFN modules indicate where sink/register behavior becomes prominent.
    The paper inverts PLoP's use of NFN and states 'in practice these peaks coincide with layers where sink/register behavior becomes prominent' (Section 4.1).
  • domain assumption Register features encode global low-frequency scene statistics that are transferable to another model's codebook.
    Relies on prior work [5,14] plus correlations in Table S3; the cross-model projection in Section 4.4 assumes transferability.
  • ad hoc to paper The nearest-codebook mapping in Eq. (10) preserves prior usefulness after cross-model transfer.
    A design choice with no independent justification; the paper does not prove the projected token remains semantically aligned.
invented entities (1)
  • RegToken / test-time register token t_reg
    purpose: An auxiliary token inserted into a frozen decoder that carries global low-frequency register structure.
    Constructed from existing activations; its utility is shown only through the paper's internal generation experiments, with no external falsifiable handle such as a predicted measurable quantity.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Test-Time Registers as Global Priors for Tokenized Image Generation." pith.science (2026). https://pith.science/paper/EIXXVEEC

@misc{pith2026260716824,
  author       = {Pith},
  title        = {Pith review of: Test-Time Registers as Global Priors for Tokenized Image Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/EIXXVEEC}},
  note         = {Machine review of arXiv:2607.16824}
}
abstract

Attention-based models often develop attention sinks, where a small number of tokens repeatedly attract attention and accumulate unusually large activations. In vision transformers, these outliers are closely related to registers, which have been diagnostically linked to global, low-frequency image structure. Existing work has largely studied registers through interpretability analyses and linear probes, leaving open whether they can be operationalized as plug-and-play signals for generation without retraining. We revisit this question in tokenized image generation. Using OpenCLIP and DINOv2 on ImageNet, we find that test-time register features exhibit stronger low-frequency concentration than both [CLS] readouts and patch-mean features, and show a consistent (albeit moderate) correlation with pixel-space DCT low-frequency energy. Motivated by these diagnostics, we introduce RegToken, a training-free procedure that converts register structure into a small set of global prior tokens by (i) NFN-based layer localization, (ii) TokenRank-guided subspace extraction, and (iii) a projection-and-conservation update on the register subspace. Inserted into a frozen compact 1D token generation pipeline, RegToken improves ImageNet generation and alignment metrics (e.g., FID-5k 20.5 to 20.1, SigLIP 3.6 to 3.9) without modifying pretrained weights, and accelerates test-time optimization (Steps@$\tau$ 74 to 52). Overall, our results suggest that structures often viewed as attention artifacts can be repurposed as lightweight global priors for tokenized generation.

Figures

Figures reproduced from arXiv: 2607.16824 by the authors.

Figure 1
Figure 1. Overview of our proposed RegToken. (a) Prior test-time register insertion methods [14] rely on heuristic layer and channel selection, which limits their transfer￾ability across architectures. We instead localize register structure using NFN-based layer scores and TokenRank-guided channel importance, and construct register to￾kens through a projection-and-conservation interpolation rule. (b) The resulting test￾time r… view at source ↗
Figure 2
Figure 2. Spectral structure of token embeddings. We visualize PCA-projected to￾ken features and their 1-D FFT spectra along the feature dimension for three token types: register, patch-mean, and [CLS]. Across both OpenCLIP and DINOv2, [REG] features exhibit stronger low-frequency concentration than patch-mean features and differ systematically from [CLS] representations. This pattern suggests that register tokens encode smoo… view at source ↗
Figure 3
Figure 3. Effect of register insertion on DINOv2 features. [PITH_FULL_IMAGE:figures/full_fig_p010_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Correlation between λ2 and outlier tokens. Each dot corresponds to one ImageNet-val image. The x-axis shows the fraction of patch tokens above the pre-95th norm threshold, and the y-axis shows λ2 of the attention Markov chain (median over heads). Vanilla test-time regi…
Figure 5
Figure 5. Figure 5: Qualitative results. Following the HCT protocol, register-based priors gener￾ate diverse images from one input image using the prompt “a photo of the [class]”. Every two rows form a group: the upper row shows input images and the lower row shows the corresponding gener…

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

37 extracted references

  1. [1]

    In: ICML (2025)

    Beyer, L.L., Li, T., Chen, X., Karaman, S., He, K.: Highly compressed tokenizer can generate without training. In: ICML (2025)

    2025

  2. [2]

    In: Larochelle, H., Ranzato, M., Hadsell, R., Balcan, M., Lin, H

    Brown, T.B., Mann, B., Ryder, N., Subbiah, M., Kaplan, J., Dhariwal, P., Nee- lakantan, A., Shyam, P., Sastry, G., Askell, A., Agarwal, S., Herbert-Voss, A., Krueger, G., Henighan, T., Child, R., Ramesh, A., Ziegler, D.M., Wu, J., Win- ter, C., Hesse, C., Chen, M., Sigler, E., Litwin, M., Gray, S., Chess, B., Clark, J., Berner, C., McCandlish, S., Radford...

    2020

  3. [3]

    In: CVPR (2022)

    Chang, H., Zhang, H., Jiang, L., Liu, C., Freeman, W.T.: Maskgit: Masked gener- ative image transformer. In: CVPR (2022)

    2022

  4. [4]

    In: CVPR (2023)

    Cherti, M., Beaumont, R., Wightman, R., Wortsman, M., Ilharco, G., Gordon, C., Schuhmann, C., Schmidt, L., Jitsev, J.: Reproducible scaling laws for contrastive language-image learning. In: CVPR (2023)

    2023

  5. [5]

    In: ICLR (2024)

    Darcet, T., Oquab, M., Mairal, J., Bojanowski, P.: Vision transformers need reg- isters. In: ICLR (2024)

    2024

  6. [6]

    In: ICLR (2021)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Transformers for image recognition at scale. In: ICLR (2021)

    2021

  7. [7]

    CoRR (2025)

    Erel, Y., Dünkel, O., Dabral, R., Golyanik, V., Theobalt, C., Bermano, A.H.: At- tention (as discrete-time markov) chains. CoRR (2025)

    2025

  8. [8]

    CoRR (2026)

    Gao, W., Wang, Y., Ma, Y., Yang, C., Li, W., You, C.: Neurosonic: Conditional flow matching for eeg-to-speech reconstruction. CoRR (2026)

    2026

Show all 37 references
  1. [9]

    In: ICLR (2025)

    Gu, X., Pang, T., Du, C., Liu, Q., Zhang, F., Du, C., Wang, Y., Lin, M.: When attention sink emerges in language models: An empirical view. In: ICLR (2025)

    2025

  2. [10]

    CoRR (2026)

    Guo, L., Wang, Y., Wen, T., Wang, Y., Feng, A., Chen, B., Jegelka, S., You, C.: Csrv2: Unlocking ultra-sparse embeddings. CoRR (2026)

    2026

  3. [11]

    In: WACV (2024)

    Han, K., Sun, S., Le, T.T., Yan, X., Ma, H., You, C., Xie, X.: Hybrid neural diffeomorphic flow for shape representation and generation via triplane. In: WACV (2024)

    2024

  4. [12]

    CoRR (2025)

    Hayou, S., Ghosh, N., Yu, B.: Plop: Precise lora placement for efficient finetuning of large models. CoRR (2025)

    2025

  5. [13]

    IEEE TMI (2024)

    Huang, X., Li, H., Cao, M., Chen, L., You, C., An, D.: Cross-modal conditioned reconstruction for language-guided medical image segmentation. IEEE TMI (2024)

    2024

  6. [14]

    NeurIPS (2025)

    Jiang, N., Dravid, A., Efros, A.A., Gandelsman, Y.: Vision transformers don’t need trained registers. NeurIPS (2025)

    2025

  7. [15]

    In: ICLR (2025)

    Kang, S., Kim, J., Kim, J., Hwang, S.J.: See what you are told: Visual attention sink in large multimodal models. In: ICLR (2025)

    2025

  8. [16]

    In: ICCV (2025)

    Kim, D., He, J., Yu, Q., Yang, C., Shen, X., Kwak, S., Chen, L.: Democratizing text-to-image masked generative models with compact text-aware one-dimensional tokens. In: ICCV (2025)

    2025

  9. [17]

    IEEE TPAMI (2021)

    Liu, F., Wu, X., You, C., Ge, S., Zou, Y., Sun, X.: Aligning source visual and target language domains for unpaired video captioning. IEEE TPAMI (2021)

    2021

  10. [18]

    NeurIPS (2022)

    Liu, F., Yang, B., You, C., Wu, X., Ge, S., Liu, Z., Sun, X., Yang, Y., Clifton, D.: Retrieve, reason, and refine: Generating accurate and faithful patient instructions. NeurIPS (2022)

    2022

  11. [19]

    NeurIPS (2021)

    Liu, F., You, C., Wu, X., Ge, S., Sun, X., et al.: Auto-encoding knowledge graph for unsupervised medical report generation. NeurIPS (2021)

    2021

  12. [20]

    CoRR (2025)

    Lu, J., Song, L., Xu, M., Ahn, B., Wang, Y., Chen, C., Dehghan, A., Yang, Y.: Atoken: A unified tokenizer for vision. CoRR (2025)

    2025

  13. [21]

    In: ICLR (2026)

    Luo, J., Fan, W.C., Wang, L., He, X., Rahman, T., Abolmaesumi, P., Sigal, L.: To sink or not to sink: Visual information pathways in large vision-language models. In: ICLR (2026)

    2026

  14. [22]

    TMLR (2024) RegToken: Registers as Global Priors for Image Generation 17

    Oquab, M., Darcet, T., Moutakanni, T., Vo, H.V., Szafraniec, M., Khalidov, V., Fernandez, P., Haziza, D., Massa, F., El-Nouby, A., Assran, M., Ballas, N., Galuba, W., Howes, R., Huang, P., Li, S., Misra, I., Rabbat, M., Sharma, V., Synnaeve, G., Xu, H., Jégou, H., Mairal, J., ...

    2024

  15. [23]

    In: Meila, M., Zhang, T

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: Meila, M., Zhang, T. (eds.) ICML (2021)

    2021

  16. [24]

    CoRR (2025)

    Siméoni, O., Vo, H.V., Seitzer, M., Baldassarre, F., Oquab, M., Jose, C., Khalidov, V., Szafraniec, M., Yi, S., Ramamonjisoa, M., et al.: Dinov3. CoRR (2025)

    2025

  17. [25]

    CoRR (2024)

    Sun, M., Chen, X., Kolter, J.Z., Liu, Z.: Massive activations in large language models. CoRR (2024)

    2024

  18. [26]

    In: AAAI (2026)

    Sun, S., Xu, J., de Araujo, G., Zhou, S., Zhang, H., Huang, Z., You, C., Xie, X.: Coma: Compositional human motion generation with multi-modal agents. In: AAAI (2026)

    2026

  19. [27]

    CoRR (2025)

    Wang, H., Kai, J., Bai, H., Hou, L., Jiang, B., He, Z., Lin, Z.: Fourier-vlm: Com- pressing vision tokens in the frequency domain for large vision-language models. CoRR (2025)

    2025

  20. [28]

    CoRR (2026)

    Wang, Y., Ma, Y., Li, W., You, C.: Let eeg models learn eeg. CoRR (2026)

    2026

  21. [29]

    CoRR (2025)

    Wen,T.,Wang,Y.,Zeng,Z.,Peng,Z.,Su,Y.,Liu,X.,Chen,B.,Liu,H.,Jegelka,S., You, C.: Beyond matryoshka: Revisiting sparse coding for adaptive representation. CoRR (2025)

    2025

  22. [30]

    In: ICLR (2024)

    Xiao, G., Tian, Y., Chen, B., Han, S., Lewis, M.: Efficient streaming language models with attention sinks. In: ICLR (2024)

    2024

  23. [31]

    CoRR (2025)

    Yang, L., Zhang, Z., Jain, A., Cao, S., Yuan, B., Chen, Y., Jia, Z., Netravali, R.: Less is more: Training-free sparse attention with global locality for efficient reasoning. CoRR (2025)

    2025

  24. [32]

    Nature Communications (2025)

    You, C., Dai, H., Min, Y., Sekhon, J.S., Joshi, S., Duncan, J.S.: Uncovering mem- orization effect in the presence of spurious correlations. Nature Communications (2025)

    2025

  25. [33]

    In: CVPR (2024)

    You, C., Mint, Y., Dai, W., Sekhon, J.S., Staib, L., Duncan, J.S.: Calibrating multi-modal representations: A pursuit of group robustness without annotations. In: CVPR (2024)

    2024

  26. [34]

    In: NeurIPS (2024)

    Yu, Q., Weber, M., Deng, X., Shen, X., Cremers, D., Chen, L.: An image is worth 32 tokens for reconstruction and generation. In: NeurIPS (2024)

    2024

  27. [35]

    In: CVPR (2025)

    Zha, K., Yu, L., Fathi, A., Ross, D.A., Schmid, C., Katabi, D., Gu, X.: Language- guided image tokenization for generation. In: CVPR (2025)

    2025

  28. [36]

    inserted prior

    Zhao, J., Wang, T., Abid, W., Angus, G., Garg, A., Kinnison, J., Sherstinsky, A., Molino, P., Addair, T., Rishi, D.: Lora land: 310 fine-tuned llms that rival gpt-4, A technical report. CoRR (2024) 18 C. Hong et al. Supplementary Material A Generation Protocol and Evaluation D...

    2024

  29. [37]

    In particular, under Prior + opt, only the designated global slotz0 is updated, whilez1:T and all network weights remain frozen

    (15) A.2 Which Tokens Are Fixed and Which Are Optimized Table S1 summarizes the optimization variables for each setting. In particular, under Prior + opt, only the designated global slotz0 is updated, whilez1:T and all network weights remain frozen. A.3 Steps@τ, AUC, Threshold...

Pith tools

Reviewed August 1, 2026 · model on record in the stance chip above.