Pith. sign in

REVIEW 3 major objections 6 minor 122 references

Exploring Structured Semantic Priors Underlying Diffusion Score for Test-time Adaptation

T0 review · 3 major / 6 minor · reviewed 2026-08-10 · deepseek-v4-flash

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

Pith's one-line read A single-timestep score identity turns a diffusion model into a test-time teacher for classifiers, beating the 180-timestep Diffusion-TTA by up to 5.1 points on corrupted ImageNet and 4.2 mIoU on ADE20K.

desk verdict Solid empirical TTA paper with an elementary identity and an unquantified proxy at the load-bearing step; worth refereeing but needs fixes. read the letter →

arxiv 2501.00873 v1 pith:FU2KYGKV submitted 2025-01-01 cs.CV cs.LG

Mingjia Li , Shuang Li , Tongrui Su , Longhui Yuan , Jian Liang , Wei Li This is my paper · ORCID
classification cs.CVcs.LG
keywords scorefunctionsdiffusionmodelstest-timeadaptationsemanticpriorsdenoisingobjectivesingle-timestepimageclassificationsegmentation
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

This paper tries to establish that score-based diffusion models hide a discriminative structure that a task model can borrow at test time without labels. The key identity is that the score of the noisy data equals the class-posterior-weighted average of conditional scores, which Tweedie's formula converts into a statement about predicted noise: the noise added to an image is approximately the probability-weighted sum of conditional noise predictions. The authors turn this into a training objective that adapts a classifier or segmenter by matching its class probabilities to the diffusion model's conditional noise estimates at one timestep, $t=100$. If right, this would make generative priors available to discriminative tasks cheaply, replacing Monte Carlo averaging over hundreds of timesteps with a small constant class budget, and the same objective carries over to per-pixel dense prediction.

What carries the argument

The load-bearing object is Proposition 1, a pointwise score identity expressing the unconditional score as a posterior-weighted sum of conditional scores. Its power comes from Tweedie's formula ($\nabla_{x_t}\log p(x_t) = -\epsilon/\sqrt{1-\bar\alpha_t}$), which rewrites the identity as an ensemble of conditional noise predictions and makes it operational at any single timestep. Around this, a Candidate Selection Module (CSM) with LogitNorm and a mixed top-$k$/random class budget cuts the per-sample cost from all $K$ classes to $b$ classes while keeping the timestep fixed at $t=100$.

What would settle it

Estimate the true posterior $p(y\mid x_t)$ at $t=100$ with a diffusion classifier (averaging over many noise samples) on ImageNet-C, and compare it with the task model's $p_\theta(y\mid x_0)$; if the two are far apart on corrupted images, or if swapping in the accurate posterior does not improve adaptation, the substitution in Eq. (10) is not doing the work claimed.

Watch

Extended reading notes

Core claim

On its own terms, the paper's discovery is that the score function decomposition is a usable discriminative prior: $\nabla_x \log p(x) = \sum_y p(y\mid x) \nabla_x \log p(x\mid y)$, which combined with Tweedie's formula gives $\epsilon \approx \sum_y p(y\mid x_t)\, \epsilon_\phi(x_t, t, c_y)$. Substituting the task model's clean-image prediction $p_\theta(y\mid x_0)$ for the unknown noisy posterior $p(y\mid x_t)$ yields the DUSA objective $\mathbb{E}_\epsilon [\|\epsilon - \sum_y p_\theta(y\mid x_0) \epsilon_\phi(x_t, t, c_y)\|^2]$, and updating the task model against it transfers structured semantic knowledge from the diffusion model into the task model. The paper also shows the diffusion model can be adapted unconditionally through a null condition with little to no loss, and that the entire construction applies per pixel to semantic segmentation using an off-the-shelf text-to-image diffusion model.

Load-bearing premise

The load-bearing premise is that a classifier's prediction on the clean image is a good stand-in for what the diffusion model's own implicit belief would be about which class produced the noisy image at step 100; if those disagree, the objective can reinforce the classifier's mistakes rather than correct them.

Editorial extensions

If this is right

  • A single denoising step suffices for diffusion-guided test-time adaptation; the 180-timestep Monte Carlo likelihood estimation used by Diffusion-TTA is not needed.
  • The same objective transfers from image classification to semantic segmentation by applying the score identity per pixel, allowing off-the-shelf image-level-conditioned diffusion models to supervise dense prediction.
  • Adapting the diffusion model unconditionally (null condition) yields performance on par with joint adaptation, implying the semantic priors are already embedded in the diffusion model independent of task feedback.
  • The computational cost scales with the number of classes or the selected budget $b$, not the number of timesteps, so the method remains practical when the classifier has many classes.
  • On the benchmarks used, the method reports gains of +5.1% and +7.3% over Diffusion-TTA on ConvNeXt-L for fully and continual test-time adaptation, and +4.2% mIoU on SegFormer-B5 for segmentation.

Reading between the lines

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

  • Beyond the paper, the score identity is purely algebraic, so the objective should extend to any score-based generative model; a cheaper score estimator could replace the diffusion model in settings where the diffusion forward pass is too slow.
  • One could invert the objective and read the per-class noise residuals as a confidence or calibration signal for the task model, diagnosing which classes the diffusion model thinks are plausible for a given noisy image.
  • The choice of $t=100$ is empirical; a per-sample or per-corruption adaptive timestep scheduler might outperform any fixed value, especially for corruptions that the diffusion model finds out-of-distribution (e.g., pixelation in the ablation).
  • Because the random-selection budget $m$ exists to prevent class bias, DUSA's robustness should degrade gracefully only if the task model's top-$k$ accuracy collapses; a direct stress test would corrupt the task model's ranking and measure how much the random component compensates.
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 / 6 minor

Summary. The paper proposes DUSA, a test-time adaptation method that uses a pre-trained conditional diffusion model as a source of discriminative semantic priors. The method starts from the exact score decomposition in Proposition 1, applies it at a single diffusion timestep via Tweedie's formula to obtain Eq. (9), and then substitutes the task model's clean-image prediction p_theta(y|x_0) for the noisy-data posterior p(y|x_t) to obtain the DUSA objective in Eq. (10). Practical designs are added for efficiency: a fixed timestep t=100, LogitNorm, a candidate selection module with budget k+m, and an unconditional diffusion-adaptation variant (DUSA-U). The method is evaluated on ImageNet-C for fully and continual test-time adaptation of three classifiers and on ADE20K-C for test-time semantic segmentation, reporting consistent gains over Diffusion-TTA and other baselines, with ablations in Section 4.4 and appendices. Code and implementation details are provided.

Significance. If the link between the theoretical identity and the DUSA objective were tight, this would be a solid contribution: it extracts class-structured information from a diffusion model in a single timestep, avoids Monte Carlo sampling over timesteps, extends to dense prediction, and is supported by an unusually thorough evaluation with three architectures, two tasks, multiple protocols, ablations, error bars, and code release. The empirical claims are credible: the reported gains over Diffusion-TTA (+5.1% and +7.3% on ConvNeXt-L, +4.2% on SegFormer-B5) are backed by tables with means and standard deviations, and the DUSA-U variant is a useful control. However, the central theoretical bridge is weaker than the presentation suggests. The identity in Proposition 1 is correct, but the objective in Eq. (10) relies on an unquantified substitution of p_theta(y|x_0) for p(y|x_t), and the claimed unbiasedness of noise predictions in Appendix D.2 is not established. These issues do not invalidate the empirical method, but they currently make the paper's stated theoretical foundation broader than what is actually proven.

major comments (3)
  1. [Sec. 3.2, Eq. (10)] The derivation establishes Eq. (9) with p(y|x_t), the posterior under the t-noised data, but the actual objective in Eq. (10) replaces this with p_theta(y|x_0), the task model's prediction on the clean target image. The sentence 'can be learned by directly plugging the task model prediction on x0' asserts this substitution without a bound, a convergence argument, or an experimental measurement of the gap. Appendix A lists limitations of the method but does not identify this assumption. The gap is load-bearing: at t=100 the noise in the 32x32x4 latent is not negligible, x_0 itself is out-of-distribution for the source-trained diffusion model, and if p_theta is confidently wrong, minimizing Eq. (10) can reinforce the wrong class's conditional noise prediction. The ablations in Table 4 vary budgets, LogitNorm, and diffusion adaptation, but no experiment compares Eq. (10) against a loss that uses an estimate of p(y|x_t), nor reports how far p_theta(y|x_0) is from p(y|x_t). I ask for a quantitative measurement of this gap (e.g., using a diffusion-classifier estimate of p(y|x_t) at the chosen timestep) or, alternatively, a reformulation of the method as a heuristic in which the substitution is explicitly treated as an assumption.
  2. [Appendix D.2] The claim that the conditional noise prediction is 'provably unbiased' is not correct. Stationarity of the simplified diffusion loss gives E[epsilon_phi(x_t,t,c)] = E[epsilon] under the joint training distribution, which is 0, not E[epsilon_phi(x_t,t,c_y) | x_t,t,c_y] = epsilon for the realized noise. The proof in Eq. (36) drops the conditioning and treats a marginal identity as a conditional one; moreover, a single network evaluation is not an unbiased estimator of a particular sampled epsilon in any useful sense. This appendix is used to support the statement that 'the only estimations in our method ... are provably unbiased,' so it should be corrected, weakened, or removed. The main empirical results do not depend on this claim, but it is part of the paper's stated theoretical support.
  3. [Sec. 3.2, Eq. (8)] The transition from Proposition 1 to the DUSA objective also assumes that the score functions of the source-trained diffusion model are the score functions of the target distribution at the chosen timestep. The proposition itself holds for a fixed joint distribution; it does not by itself justify using source-trained conditional scores to supervise a task model on corrupted target inputs. The Pixelate row in Table 5 is a concrete warning: the score-based loss with a frozen diffusion model drops ConvNeXt-L accuracy from 42.3 to 9.2, suggesting that the source scores can be far off on some target corruptions. The paper should either discuss this source-target score mismatch explicitly or provide an experiment measuring how the noise predictions and the implied posteriors change on target-corrupted inputs.
minor comments (6)
  1. [Sec. 3.2 and Sec. 4.1] There are several typos: 'simutaneously' in Section 3.2, 'Sepecifically' in Section 4.1, and 'Similiarly' and 'F orx0-prediction' in Appendix D.2. 'Extensive to' in Corollary 2 should be 'extensible to'.
  2. [Algorithm 1 and Eq. (14)] The symbol K is used both for the total number of classes in Section 3.2 and Eq. (14) and for the number of selected classes in Algorithm 1. Please use different symbols, e.g., b for the selected budget.
  3. [Table 4] The abbreviations D.F. and D.B. should be defined in the table caption or at first use in the body text; the text defines D.F. but not D.B. until later in the same section.
  4. [Sec. 3.3] The sentence 'we deem the posterior p(y|x) of less likely class candidates to be zero' is inconsistent with the subsequent random multinomial selection of m classes from the remaining candidates. Please rephrase to say that the loss is restricted to the selected classes rather than that the posterior is set to zero.
  5. [Appendix H] The ensembling experiments in Table 6 do not report standard deviations, unlike the main tables. Adding error bars would help assess whether the small differences between single and multiple timesteps are meaningful.
  6. [Sec. 4.1 and Appendix F.3] The statement that Diffusion-TTA is evaluated with a budget of 6 timesteps, rather than the 180 timesteps in the original method, is important for interpreting the comparison and should appear in the main text rather than only in Appendix F.3.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the core score decomposition is a self-contained Bayes identity and the DUSA objective is an externally anchored distillation loss, not a fitted re-statement.

full rationale

The derivation chain is self-contained. Proposition 1 follows directly from the law of total probability, the log-derivative trick, and Bayes' theorem, as shown in Appendix C, and Eq. (9) follows by applying Tweedie's formula and the diffusion noise-prediction parametrization. Eq. (10) then substitutes the task model's clean-image prediction p_theta(y|x_0) for the noisy posterior p(y|x_t); this is an unquantified plug-in approximation and a genuine correctness risk, but it is not a circular reduction, because the two quantities are not equal by construction and the loss is minimized against the external diffusion model's conditional noise predictions rather than against the task model's own output alone. The method is benchmarked against external baselines (Tent, CoTTA, EATA, SAR, RoTTA, Diffusion-TTA) on ImageNet-C and ADE20K-C, so the empirical claims do not reduce to the paper's inputs. The only author self-citations are peripheral and non-load-bearing: [55] is cited alongside the original LogitNorm paper [54], and [61] is simply a comparison baseline. The unquantified gap between p(y|x_t) and p_theta(y|x_0), and the absence of any measured bound or ablation isolating that substitution, should be weighed as a rigor/correctness concern rather than as circularity.

Assumptions & free parameters 2 free parameters · 5 assumptions · 0 invented entities

The central claim relies on standard probability identities and standard diffusion model score approximations. The only ad hoc element is the clean-image classifier playing the role of the noised-image posterior, and the hand-selected hyperparameters for timestep and class budget.

free parameters (2)
  • timestep t = 100
    Selected by hand from the ablation in Fig. 3; the paper states t=100 works well for all settings but does not derive it from theory.
  • class selection budget k and m = k=4, m=2
    Chosen by hand from the budget-scheme study in Fig. 4 and Table 4; affects the number of diffusion forwards and the adaptation signal.
assumptions (5)
  • standard math Law of total probability, Bayes' theorem, and log-derivative trick for continuously differentiable densities with p(x)>0.
    Used in the proof of Proposition 1 in Appendix C. The identity ∇log p(x) = Σ_y p(y|x) ∇log p(x|y) is a direct consequence.
  • standard math Tweedie's formula holds for Gaussian conditional distributions in the diffusion forward process.
    Invoked in Lemma 1 and Corollary 1 to relate scores to predicted noise.
  • domain assumption A trained conditional diffusion model's noise prediction approximates the conditional score: ∇log p(x_t|y) ≈ -ϵ_φ(x_t,t,c_y)/√(1-ᾱ_t).
    Equation (8) in Sec. 3.2; this is the standard score-matching result for diffusion models, but it is an approximation in practice.
  • ad hoc to paper The task model prediction on the clean image p_θ(y|x_0) approximates the posterior under the noised data p(y|x_t) at the chosen timestep t=100.
    Equation (10) replaces p(y|x_t) with p_θ(y|x_0) without quantifying the error. The only justification is the empirical choice of t=100.
  • domain assumption A diffusion model pre-trained on the source domain of the task model is available at test time.
    All experiments use DiT-XL/2 on ImageNet and ControlNet on ADE20K; the limitations section in Appendix A admits this may not always be available.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Exploring Structured Semantic Priors Underlying Diffusion Score for Test-time Adaptation." pith.science (2026). https://pith.science/paper/FU2KYGKV

@misc{pith2026250100873,
  author       = {Pith},
  title        = {Pith review of: Exploring Structured Semantic Priors Underlying Diffusion Score for Test-time Adaptation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FU2KYGKV}},
  note         = {Machine review of arXiv:2501.00873}
}
read the original abstract

Capitalizing on the complementary advantages of generative and discriminative models has always been a compelling vision in machine learning, backed by a growing body of research. This work discloses the hidden semantic structure within score-based generative models, unveiling their potential as effective discriminative priors. Inspired by our theoretical findings, we propose DUSA to exploit the structured semantic priors underlying diffusion score to facilitate the test-time adaptation of image classifiers or dense predictors. Notably, DUSA extracts knowledge from a single timestep of denoising diffusion, lifting the curse of Monte Carlo-based likelihood estimation over timesteps. We demonstrate the efficacy of our DUSA in adapting a wide variety of competitive pre-trained discriminative models on diverse test-time scenarios. Additionally, a thorough ablation study is conducted to dissect the pivotal elements in DUSA. Code is publicly available at https://github.com/BIT-DA/DUSA.

Figures

Figures reproduced from arXiv: 2501.00873 by the authors.

Figure 1
Figure 1. Overview of DUSA. Our method adapts a discriminative task model fθ with a generative diffusion model ϵϕ. Given image x0 at test-time, the task model outputs logits. To improve efficiency, we devise a CSM to select classes to adapt and return their probabilities (probs). The embeddings of the classes are then queried as diffusion model conditions, yielding conditional noise predictions from noisy image xt. The aggreg… view at source ↗
Figure 2
Figure 2. Visualization of segmentation results on ADE20K-C. From left to right: clean and corrupted [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 4
Figure 4. Accuracy of ViT-B/16 on JPEG and ResNet-50 [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (2 more)
Figure 5
Figure 5. Figure 5: A model checkpoint is saved after test-time adaptation over a whole corrupted ADE20K [PITH_FULL_IMAGE:figures/full_fig_p023_5.png]
Figure 5
Figure 5. Figure 5: Visualization of test-time semantic segmentation results on ADE20K-C. From left to right: [PITH_FULL_IMAGE:figures/full_fig_p024_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

122 extracted references · 53 canonical work pages

  1. [1]

    On discriminative vs

    Andrew Ng and Michael Jordan. On discriminative vs. generative classifiers: A comparison of logistic regression and naive bayes. In NeurIPS, pages 841–848, 2001

    2001

  2. [2]

    Ng, and Andrew McCallum

    Rajat Raina, Yirong Shen, Andrew Y . Ng, and Andrew McCallum. Classification with hybrid genera- tive/discriminative models. In NeurIPS, pages 545–552, 2003

    2003

  3. [3]

    Combining generative and discriminative models for hybrid inference

    Victor Garcia Satorras, Zeynep Akata, and Max Welling. Combining generative and discriminative models for hybrid inference. In NeurIPS, pages 13802–13812, 2019

    2019

  4. [4]

    Your classifier is secretly an energy based model and you should treat it like one

    Will Grathwohl, Kuan-Chieh Wang, Joern-Henrik Jacobsen, David Duvenaud, Mohammad Norouzi, and Kevin Swersky. Your classifier is secretly an energy based model and you should treat it like one. In ICLR, 2020

    2020

  5. [5]

    Deep residual learning for image recognition

    Kaiming He, Xiangyu Zhang, Shaoqing Ren, and Jian Sun. Deep residual learning for image recognition. In CVPR, pages 770–778, 2016

    2016

  6. [6]

    A convnet for the 2020s

    Zhuang Liu, Hanzi Mao, Chao-Yuan Wu, Christoph Feichtenhofer, Trevor Darrell, and Saining Xie. A convnet for the 2020s. In CVPR, pages 11976–11986, 2022

    2022

  7. [7]

    An image is worth one word: Personalizing text-to-image generation using textual inversion

    Rinon Gal, Yuval Alaluf, Yuval Atzmon, Or Patashnik, Amit Haim Bermano, Gal Chechik, and Daniel Cohen-or. An image is worth one word: Personalizing text-to-image generation using textual inversion. In ICLR, 2023

    2023

  8. [8]

    Condensenet v2: Sparse feature reactivation for deep networks

    Le Yang, Haojun Jiang, Ruojin Cai, Yulin Wang, Shiji Song, Gao Huang, and Qi Tian. Condensenet v2: Sparse feature reactivation for deep networks. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 3569–3578, 2021

    2021

Show all 122 references
  1. [9]

    Resolution adaptive networks for efficient inference

    Le Yang, Yizeng Han, Xi Chen, Shiji Song, Jifeng Dai, and Gao Huang. Resolution adaptive networks for efficient inference. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 2369–2378, 2020

    2020

  2. [10]

    Dynamic spatial focus for efficient compressed video action recognition

    Ziwei Zheng, Le Yang, Yulin Wang, Miao Zhang, Lijun He, Gao Huang, and Fan Li. Dynamic spatial focus for efficient compressed video action recognition. IEEE Transactions on Circuits and Systems for Video Technology, 2024

    2024

  3. [11]

    Adadet: An adaptive object detection system based on early-exit neural networks.IEEE Transactions on Cognitive and Developmental Systems, 16(1):332–345, 2023

    Le Yang, Ziwei Zheng, Jian Wang, Shiji Song, Gao Huang, and Fan Li. Adadet: An adaptive object detection system based on early-exit neural networks.IEEE Transactions on Cognitive and Developmental Systems, 16(1):332–345, 2023

    2023

  4. [12]

    Recognition in terra incognita

    Sara Beery, Grant Van Horn, and Pietro Perona. Recognition in terra incognita. In ECCV, pages 456–473, 2018

    2018

  5. [13]

    Shortcut learning in deep neural networks

    Robert Geirhos, Jörn-Henrik Jacobsen, Claudio Michaelis, Richard Zemel, Wieland Brendel, Matthias Bethge, and Felix A Wichmann. Shortcut learning in deep neural networks. Nat. Mach. Intell., 2(11):665– 673, 2020

    2020

  6. [14]

    Infogan: Interpretable representation learning by information maximizing generative adversarial nets

    Xi Chen, Yan Duan, Rein Houthooft, John Schulman, Ilya Sutskever, and Pieter Abbeel. Infogan: Interpretable representation learning by information maximizing generative adversarial nets. In NeurIPS, pages 2172–2180, 2016

    2016

  7. [15]

    Masked autoencoders are scalable vision learners

    Kaiming He, Xinlei Chen, Saining Xie, Yanghao Li, Piotr Dollár, and Ross Girshick. Masked autoencoders are scalable vision learners. In CVPR, pages 16000–16009, 2022

    2022

  8. [16]

    Diffusion based representation learning

    Sarthak Mittal, Korbinian Abstreiter, Stefan Bauer, Bernhard Schölkopf, and Arash Mehrjou. Diffusion based representation learning. In ICML, pages 24963–24982, 2023

    2023

  9. [17]

    Gat: Generative adversarial training for adversarial example detection and robust classification

    Xuwang Yin, Soheil Kolouri, and Gustavo K Rohde. Gat: Generative adversarial training for adversarial example detection and robust classification. In ICLR, 2020

    2020

  10. [18]

    Generative robust classification

    Xuwang Yin. Generative robust classification. arXiv preprint arXiv:2212.07283, 2022

    2022 arXiv

  11. [19]

    Mage: Masked generative encoder to unify representation learning and image synthesis

    Tianhong Li, Huiwen Chang, Shlok Mishra, Han Zhang, Dina Katabi, and Dilip Krishnan. Mage: Masked generative encoder to unify representation learning and image synthesis. In CVPR, pages 2142–2152, 2023

    2023

  12. [20]

    High-resolution image synthesis with latent diffusion models

    Robin Rombach, Andreas Blattmann, Dominik Lorenz, Patrick Esser, and Björn Ommer. High-resolution image synthesis with latent diffusion models. In CVPR, pages 10684–10695, 2022

    2022

  13. [21]

    Fleet, and Mohammad Norouzi

    Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kam- yar Seyed Ghasemipour, Raphael Gontijo-Lopes, Burcu Karagol Ayan, Tim Salimans, Jonathan Ho, David J. Fleet, and Mohammad Norouzi. Photorealistic text-to-image diffusion models with dee...

    2022

  14. [22]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, pages 4195– 4205, 2023. 11

    2023

  15. [23]

    Do text-free diffusion models learn discriminative visual representations? arXiv preprint arXiv:2311.17921, 2023

    Soumik Mukhopadhyay, Matthew Gwilliam, Yosuke Yamaguchi, Vatsal Agarwal, Namitha Padmanabhan, Archana Swaminathan, Tianyi Zhou, and Abhinav Shrivastava. Do text-free diffusion models learn discriminative visual representations? arXiv preprint arXiv:2311.17921, 2023

    2023 arXiv

  16. [24]

    Unleashing text-to-image diffusion models for visual perception

    Wenliang Zhao, Yongming Rao, Zuyan Liu, Benlin Liu, Jie Zhou, and Jiwen Lu. Unleashing text-to-image diffusion models for visual perception. In ICCV, pages 5729–5739, 2023

    2023

  17. [25]

    Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak

    Alexander C. Li, Mihir Prabhudesai, Shivam Duggal, Ellis Brown, and Deepak Pathak. Your diffusion model is secretly a zero-shot classifier. In ICCV, pages 2206–2217, 2023

    2023

  18. [26]

    Diffusionengine: Diffusion model is scalable data engine for object detection

    Manlin Zhang, Jie Wu, Yuxi Ren, Ming Li, Jie Qin, Xuefeng Xiao, Wei Liu, Rui Wang, Min Zheng, and Andy J Ma. Diffusionengine: Diffusion model is scalable data engine for object detection. arXiv preprint arXiv:2309.03893, 2023

    2023 arXiv

  19. [27]

    Open- vocabulary panoptic segmentation with text-to-image diffusion models

    Jiarui Xu, Sifei Liu, Arash Vahdat, Wonmin Byeon, Xiaolong Wang, and Shalini De Mello. Open- vocabulary panoptic segmentation with text-to-image diffusion models. In CVPR, pages 2955–2966, 2023

    2023

  20. [28]

    A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence

    Junyi Zhang, Charles Herrmann, Junhwa Hur, Luisa Polania Cabrera, Varun Jampani, Deqing Sun, and Ming-Hsuan Yang. A tale of two features: Stable diffusion complements dino for zero-shot semantic correspondence. NeurIPS, 36, 2024

    2024

  21. [29]

    Text-to-image diffusion models are zero shot classifiers

    Kevin Clark and Priyank Jaini. Text-to-image diffusion models are zero shot classifiers. In NeurIPS, pages 58921–58937, 2023

    2023

  22. [30]

    Intriguing properties of generative classifiers

    Priyank Jaini, Kevin Clark, and Robert Geirhos. Intriguing properties of generative classifiers. In ICLR, 2024

    2024

  23. [31]

    Benchmarking neural network robustness to common corruptions and perturbations

    Dan Hendrycks and Thomas Dietterich. Benchmarking neural network robustness to common corruptions and perturbations. In ICLR, 2019

    2019

  24. [32]

    Test-time adaptation of discriminative models via diffusion generative feedback

    Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, and Katerina Fragkiadaki. Test-time adaptation of discriminative models via diffusion generative feedback. In NeurIPS, 2023

    2023

  25. [33]

    Discovering interpretable directions in the semantic latent space of diffusion models

    René Haas, Inbar Huberman-Spiegelglas, Rotem Mulayoff, and Tomer Michaeli. Discovering interpretable directions in the semantic latent space of diffusion models. arXiv preprint arXiv:2303.11073, 2023

    2023 arXiv

  26. [34]

    Diffusion models already have a semantic latent space

    Mingi Kwon, Jaeseok Jeong, and Youngjung Uh. Diffusion models already have a semantic latent space. In ICLR, 2023

    2023

  27. [35]

    Noiseclr: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models

    Yusuf Dalva and Pinar Yanardag. Noiseclr: A contrastive learning approach for unsupervised discovery of interpretable directions in diffusion models. arXiv preprint arXiv:2312.05390, 2023

    2023 arXiv

  28. [36]

    Guided image synthesis via initial image editing in diffusion model

    Jiafeng Mao, Xueting Wang, and Kiyoharu Aizawa. Guided image synthesis via initial image editing in diffusion model. In ACM MM, pages 5321–5329, 2023

    2023

  29. [37]

    Understanding the latent space of diffusion models through the lens of riemannian geometry

    Yong-Hyun Park, Mingi Kwon, Jaewoong Choi, Junghyo Jo, and Youngjung Uh. Understanding the latent space of diffusion models through the lens of riemannian geometry. In NeurIPS, 2023

    2023

  30. [38]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In NeurIPS, pages 6840–6851, 2020

    2020

  31. [39]

    A connection between score matching and denoising autoencoders

    Pascal Vincent. A connection between score matching and denoising autoencoders. Neural Comput., 23(7):1661–1674, 2011

    2011

  32. [40]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In NeurIPS, pages 11895–11907, 2019

    2019

  33. [41]

    Improved techniques for training score-based generative models

    Yang Song and Stefano Ermon. Improved techniques for training score-based generative models. In NeurIPS, pages 12438–12448, 2020

    2020

  34. [42]

    Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole

    Yang Song, Jascha Sohl-Dickstein, Diederik P. Kingma, Abhishek Kumar, Stefano Ermon, and Ben Poole. Score-based generative modeling through stochastic differential equations. In ICLR, 2021

    2021

  35. [43]

    Understanding diffusion models: A unified perspective

    Calvin Luo. Understanding diffusion models: A unified perspective. arXiv preprint arXiv:2208.11970, 2022

    2022 arXiv

  36. [44]

    Tent: Fully test-time adaptation by entropy minimization

    Dequan Wang, Evan Shelhamer, Shaoteng Liu, Bruno Olshausen, and Trevor Darrell. Tent: Fully test-time adaptation by entropy minimization. In ICLR, 2021

    2021

  37. [45]

    Continual test-time domain adaptation

    Qin Wang, Olga Fink, Luc Van Gool, and Dengxin Dai. Continual test-time domain adaptation. In CVPR, pages 7201–7211, 2022

    2022

  38. [46]

    Deep unsupervised learning using nonequilibrium thermodynamics

    Jascha Sohl-Dickstein, Eric Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In ICML, pages 2256–2265, 2015

    2015

  39. [47]

    GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models

    Alexander Quinn Nichol, Prafulla Dhariwal, Aditya Ramesh, Pranav Shyam, Pamela Mishkin, Bob Mcgrew, Ilya Sutskever, and Mark Chen. GLIDE: Towards photorealistic image generation and editing with text-guided diffusion models. In ICML, pages 16784–16804, 2022. 12

    2022

  40. [48]

    Adding conditional control to text-to-image diffusion models

    Lvmin Zhang, Anyi Rao, and Maneesh Agrawala. Adding conditional control to text-to-image diffusion models. In ICCV, pages 3836–3847, 2023

    2023

  41. [49]

    Tweedie’s formula and selection bias

    Bradley Efron. Tweedie’s formula and selection bias. JASA, 106(496):1602–1614, 2011

    2011

  42. [50]

    Progressive distillation for fast sampling of diffusion models

    Tim Salimans and Jonathan Ho. Progressive distillation for fast sampling of diffusion models. In ICLR, 2022

    2022

  43. [51]

    How to train your energy-based models

    Yang Song and Diederik P Kingma. How to train your energy-based models. arXiv preprint arXiv:2101.03288, 2021

    2021 arXiv

  44. [52]

    Target score matching

    Valentin De Bortoli, Michael Hutchinson, Peter Wirnsberger, and Arnaud Doucet. Target score matching. arXiv preprint arXiv:2402.08667, 2024

    2024 arXiv

  45. [53]

    Test-time adaptation with diffusion models

    Mihir Prabhudesai, Tsung-Wei Ke, Alexander Cong Li, Deepak Pathak, and Katerina Fragkiadaki. Test-time adaptation with diffusion models. In ICML Workshop, 2023

    2023

  46. [54]

    Mitigating neural network overconfidence with logit normalization

    Hongxin Wei, Renchunzi Xie, Hao Cheng, Lei Feng, Bo An, and Yixuan Li. Mitigating neural network overconfidence with logit normalization. In ICML, pages 23631–23644, 2022

    2022

  47. [55]

    Vblc: Visibility boosting and logit-constraint learning for domain adaptive semantic segmentation under adverse conditions

    Mingjia Li, Binhui Xie, Shuang Li, Chi Harold Liu, and Xinjing Cheng. Vblc: Visibility boosting and logit-constraint learning for domain adaptive semantic segmentation under adverse conditions. In AAAI, pages 8605–8613, 2023

    2023

  48. [56]

    Scene parsing through ade20k dataset

    Bolei Zhou, Hang Zhao, Xavier Puig, Sanja Fidler, Adela Barriuso, and Antonio Torralba. Scene parsing through ade20k dataset. In CVPR, pages 633–641, 2017

    2017

  49. [57]

    An image is worth 16x16 words: Transformers for image recognition at scale

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, Jakob Uszkoreit, and Neil Houlsby. An image is worth 16x16 words: Transformers for image recognition at...

    2021

  50. [58]

    Alvarez, and Ping Luo

    Enze Xie, Wenhai Wang, Zhiding Yu, Anima Anandkumar, Jose M. Alvarez, and Ping Luo. Segformer: Simple and efficient design for semantic segmentation with transformers. InNeurIPS, pages 12077–12090, 2021

    2021

  51. [59]

    Towards stable test-time adaptation in dynamic wild world

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Zhiquan Wen, Yaofo Chen, Peilin Zhao, and Mingkui Tan. Towards stable test-time adaptation in dynamic wild world. In ICLR, 2023

    2023

  52. [60]

    Efficient test-time model adaptation without forgetting

    Shuaicheng Niu, Jiaxiang Wu, Yifan Zhang, Yaofo Chen, Shijian Zheng, Peilin Zhao, and Mingkui Tan. Efficient test-time model adaptation without forgetting. In ICML, pages 16888–16905, 2022

    2022

  53. [61]

    Robust test-time adaptation in dynamic scenarios

    Longhui Yuan, Binhui Xie, and Shuang Li. Robust test-time adaptation in dynamic scenarios. In CVPR, pages 15922–15932, 2023

    2023

  54. [62]

    Evaluating prediction-time batch normalization for robustness under covariate shift.arXiv preprint arXiv:2006.10963, 2020

    Zachary Nado, Shreyas Padhy, D Sculley, Alexander D’Amour, Balaji Lakshminarayanan, and Jasper Snoek. Evaluating prediction-time batch normalization for robustness under covariate shift.arXiv preprint arXiv:2006.10963, 2020

    2006 arXiv

  55. [63]

    Improving robustness against common corruptions by covariate shift adaptation

    Steffen Schneider, Evgenia Rusak, Luisa Eck, Oliver Bringmann, Wieland Brendel, and Matthias Bethge. Improving robustness against common corruptions by covariate shift adaptation. In NeurIPS, pages 11539–11551, 2020

    2020

  56. [64]

    Adam: A method for stochastic optimization

    Diederik Kingma and Jimmy Ba. Adam: A method for stochastic optimization. In ICLR, 2015

    2015

  57. [65]

    Imagenet: A large-scale hierarchical image database

    Jia Deng, Wei Dong, Richard Socher, Li-Jia Li, Kai Li, and Li Fei-Fei. Imagenet: A large-scale hierarchical image database. In CVPR, pages 248–255, 2009

    2009

  58. [66]

    Revisiting batch normalization for practical domain adaptation

    Yanghao Li, Naiyan Wang, Jianping Shi, Jiaying Liu, and Xiaodi Hou. Revisiting batch normalization for practical domain adaptation. In ICLR Workshop, 2017

    2017

  59. [67]

    Sita: Single image test-time adaptation

    Ansh Khurana, Sujoy Paul, Piyush Rai, Soma Biswas, and Gaurav Aggarwal. Sita: Single image test-time adaptation. arXiv preprint arXiv:2112.02355, 2021

    2021 arXiv

  60. [68]

    Test-time training with self-supervision for generalization under distribution shifts

    Yu Sun, Xiaolong Wang, Zhuang Liu, John Miller, Alexei Efros, and Moritz Hardt. Test-time training with self-supervision for generalization under distribution shifts. In ICML, pages 9229–9248, 2020

    2020

  61. [69]

    TTT++: When does self-supervised test-time training fail or thrive? In NeurIPS, 2021

    Yuejiang Liu, Parth Kothari, Bastien Germain van Delft, Baptiste Bellot-Gurlet, Taylor Mordan, and Alexandre Alahi. TTT++: When does self-supervised test-time training fail or thrive? In NeurIPS, 2021

    2021

  62. [70]

    Test-time training with masked autoencoders

    Yossi Gandelsman, Yu Sun, Xinlei Chen, and Alexei A Efros. Test-time training with masked autoencoders. In Alice H. Oh, Alekh Agarwal, Danielle Belgrave, and Kyunghyun Cho, editors, NeurIPS, 2022

    2022

  63. [71]

    Mt3: Meta test-time training for self-supervised test-time adaption

    Alexander Bartler, Andre Bühler, Felix Wiewel, Mario Döbler, and Bin Yang. Mt3: Meta test-time training for self-supervised test-time adaption. In AISTATS, pages 3080–3090, 2022

    2022

  64. [72]

    MEMO: Test time robustness via adaptation and augmentation

    Marvin Mengxin Zhang, Sergey Levine, and Chelsea Finn. MEMO: Test time robustness via adaptation and augmentation. In NeurIPS, pages 38629–38642, 2022. 13

    2022

  65. [73]

    On pitfalls of test-time adaptation

    Hao Zhao, Yuejiang Liu, Alexandre Alahi, and Tao Lin. On pitfalls of test-time adaptation. In ICML, pages 42058–42080, 2023

    2023

  66. [74]

    Parameter-free online test-time adaptation

    Malik Boudiaf, Romain Mueller, Ismail Ben Ayed, and Luca Bertinetto. Parameter-free online test-time adaptation. In CVPR, pages 8344–8353, 2022

    2022

  67. [75]

    on discriminative vs. generative classifiers: A comparison of logistic regression and naive bayes

    Jing-Hao Xue and D Michael Titterington. Comment on “on discriminative vs. generative classifiers: A comparison of logistic regression and naive bayes”. Neural Process. Lett., 28:169–187, 2008

    2008

  68. [76]

    Revisiting discriminative vs

    Chenyu Zheng, Guoqiang Wu, Fan Bao, Yue Cao, Chongxuan Li, and Jun Zhu. Revisiting discriminative vs. generative classifiers: Theory and implications. In ICML, pages 42420–42477, 2023

    2023

  69. [77]

    Plug & play generative networks: Conditional iterative generation of images in latent space

    Anh Nguyen, Jeff Clune, Yoshua Bengio, Alexey Dosovitskiy, and Jason Yosinski. Plug & play generative networks: Conditional iterative generation of images in latent space. In CVPR, pages 4467–4477, 2017

    2017

  70. [78]

    Jem++: improved techniques for training jem

    Xiulong Yang and Shihao Ji. Jem++: improved techniques for training jem. In ICCV, pages 6494–6503, 2021

    2021

  71. [79]

    Your vit is secretly a hybrid discriminative-generative diffusion model

    Xiulong Yang, Sheng-Min Shih, Yinlin Fu, Xiaoting Zhao, and Shihao Ji. Your vit is secretly a hybrid discriminative-generative diffusion model. arXiv preprint arXiv:2208.07791, 2022

    2022 arXiv

  72. [80]

    Return of unconditional generation: A self-supervised representation generation method

    Tianhong Li, Dina Katabi, and Kaiming He. Return of unconditional generation: A self-supervised representation generation method. arXiv preprint arXiv:2312.03701, 2023

    2023 arXiv

  73. [81]

    Adversarial feature learning

    Jeff Donahue, Philipp Krähenbühl, and Trevor Darrell. Adversarial feature learning. arXiv preprint arXiv:1605.09782, 2016

    2016 arXiv

  74. [82]

    Context encoders: Feature learning by inpainting

    Deepak Pathak, Philipp Krahenbuhl, Jeff Donahue, Trevor Darrell, and Alexei A Efros. Context encoders: Feature learning by inpainting. In CVPR, pages 2536–2544, 2016

    2016

  75. [83]

    Label- efficient semantic segmentation with diffusion models

    Dmitry Baranchuk, Andrey V oynov, Ivan Rubachev, Valentin Khrulkov, and Artem Babenko. Label- efficient semantic segmentation with diffusion models. In ICLR, 2022

    2022

  76. [84]

    The effectiveness of mae pre-pretraining for billion-scale pretraining

    Mannat Singh, Quentin Duval, Kalyan Vasudev Alwala, Haoqi Fan, Vaibhav Aggarwal, Aaron Adcock, Armand Joulin, Piotr Dollár, Christoph Feichtenhofer, Ross Girshick, et al. The effectiveness of mae pre-pretraining for billion-scale pretraining. In ICCV, pages 5484–5494, 2023

    2023

  77. [85]

    Peekaboo: Text to image diffusion models are zero-shot segmentors

    Ryan Burgert, Kanchana Ranasinghe, Xiang Li, and Michael S Ryoo. Peekaboo: Text to image diffusion models are zero-shot segmentors. arXiv preprint arXiv:2211.13224, 2022

    2022 arXiv

  78. [86]

    Model adaptation: Unsupervised domain adaptation without source data

    Rui Li, Qianfen Jiao, Wenming Cao, Hau-San Wong, and Si Wu. Model adaptation: Unsupervised domain adaptation without source data. In CVPR, pages 9641–9650, 2020

    2020

  79. [87]

    Adaptive distribution masked autoencoders for continual test-time adaptation

    Jiaming Liu, Ran Xu, Senqiao Yang, Renrui Zhang, Qizhe Zhang, Zehui Chen, Yandong Guo, and Shanghang Zhang. Adaptive distribution masked autoencoders for continual test-time adaptation. arXiv preprint arXiv:2312.12480, 2023

    2023 arXiv

  80. [88]

    Tea: Test-time energy adaptation

    Yige Yuan, Bingbing Xu, Liang Hou, Fei Sun, Huawei Shen, and Xueqi Cheng. Tea: Test-time energy adaptation. arXiv preprint arXiv:2311.14402, 2023

    2023 arXiv

  81. [89]

    Tttflow: Unsupervised test-time training with normalizing flow

    David Osowiechi, Gustavo A Vargas Hakim, Mehrdad Noori, Milad Cheraghalikhani, Ismail Ben Ayed, and Christian Desrosiers. Tttflow: Unsupervised test-time training with normalizing flow. In WACV, pages 2126–2134, 2023

    2023

  82. [90]

    Classifier-free diffusion guidance

    Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. arXiv preprint arXiv:2207.12598, 2022

    2022 arXiv

  83. [91]

    SDEdit: Guided image synthesis and editing with stochastic differential equations

    Chenlin Meng, Yutong He, Yang Song, Jiaming Song, Jiajun Wu, Jun-Yan Zhu, and Stefano Ermon. SDEdit: Guided image synthesis and editing with stochastic differential equations. In ICLR, 2022

    2022

  84. [92]

    Diffusion models for adversarial purification

    Weili Nie, Brandon Guo, Yujia Huang, Chaowei Xiao, Arash Vahdat, and Anima Anandkumar. Diffusion models for adversarial purification. In ICML, pages 16805–16827, 2022

    2022

  85. [93]

    Back to the source: Diffusion-driven adaptation to test-time corruption

    Jin Gao, Jialing Zhang, Xihui Liu, Trevor Darrell, Evan Shelhamer, and Dequan Wang. Back to the source: Diffusion-driven adaptation to test-time corruption. In CVPR, pages 11786–11796, 2023

    2023

  86. [94]

    Target to source: Guidance-based diffusion model for test-time adaptation

    Kaiyu Song and Hanjiang Lai. Target to source: Guidance-based diffusion model for test-time adaptation. arXiv preprint arXiv:2312.05274, 2023

    2023 arXiv

  87. [95]

    Null-text inversion for editing real images using guided diffusion models

    Ron Mokady, Amir Hertz, Kfir Aberman, Yael Pritch, and Daniel Cohen-Or. Null-text inversion for editing real images using guided diffusion models. In CVPR, pages 6038–6047, 2023

    2023

  88. [96]

    Ilvr: Conditioning method for denoising diffusion probabilistic models

    Jooyoung Choi, Sungwon Kim, Yonghyun Jeong, Youngjune Gwon, and Sungroh Yoon. Ilvr: Conditioning method for denoising diffusion probabilistic models. arXiv preprint arXiv:2108.02938, 2021

    2021 arXiv

  89. [97]

    Multiresolution textual inversion

    Giannis Daras and Alexandros G Dimakis. Multiresolution textual inversion. arXiv preprint arXiv:2211.17115, 2022

    2022 arXiv

  90. [98]

    ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers

    Yogesh Balaji, Seungjun Nah, Xun Huang, Arash Vahdat, Jiaming Song, Qinsheng Zhang, Karsten Kreis, Miika Aittala, Timo Aila, Samuli Laine, et al. ediff-i: Text-to-image diffusion models with an ensemble of expert denoisers. arXiv preprint arXiv:2211.01324, 2022. 14

    2022 arXiv

  91. [99]

    Attend-and-excite: Attention- based semantic guidance for text-to-image diffusion models

    Hila Chefer, Yuval Alaluf, Yael Vinker, Lior Wolf, and Daniel Cohen-Or. Attend-and-excite: Attention- based semantic guidance for text-to-image diffusion models. TOG, 42(4):1–10, 2023

    2023

  92. [100]

    Sketch-guided text-to-image diffusion models

    Andrey V oynov, Kfir Aberman, and Daniel Cohen-Or. Sketch-guided text-to-image diffusion models. In SIGGRAPH, pages 1–11, 2023

    2023

  93. [101]

    Unsupervised semantic correspondence using stable diffusion

    Eric Hedlin, Gopal Sharma, Shweta Mahajan, Hossam Isack, Abhishek Kar, Andrea Tagliasacchi, and Kwang Moo Yi. Unsupervised semantic correspondence using stable diffusion. Advances in Neural Information Processing Systems, 36, 2024

    2024

  94. [102]

    SLime: Segment like me

    Aliasghar Khani, Saeid Asgari, Aditya Sanghi, Ali Mahdavi Amiri, and Ghassan Hamarneh. SLime: Segment like me. In ICLR, 2024

    2024

  95. [103]

    Openmmlab’s pre-training toolbox and benchmark

    MMPreTrain Contributors. Openmmlab’s pre-training toolbox and benchmark. https://github.com/ open-mmlab/mmpretrain, 2023

    2023

  96. [104]

    MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark

    MMSegmentation Contributors. MMSegmentation: Openmmlab semantic segmentation toolbox and benchmark. https://github.com/open-mmlab/mmsegmentation, 2020

    2020

  97. [105]

    Ecker, Matthias Bethge, and Wieland Brendel

    Claudio Michaelis, Benjamin Mitzkus, Robert Geirhos, Evgenia Rusak, Oliver Bringmann, Alexander S. Ecker, Matthias Bethge, and Wieland Brendel. Benchmarking robustness in object detection: Autonomous driving when winter is coming. arXiv preprint arXiv:1907.07484, 2019

    1907 arXiv

  98. [106]

    Pytorch image models

    Ross Wightman. Pytorch image models. https://github.com/rwightman/pytorch-image- models, 2019. 15 A Broader Impacts and Limitations Broader Impacts. In this work, an approach to incorporate generative diffusion models into the discriminative task of test-time adaptation is int...

    2019

  99. [107]

    5: Make task prediction ˆyj = argmaxy zy

    with task model. 5: Make task prediction ˆyj = argmaxy zy. 6: Apply CSM to prune logits z to K classes and get ˆz. 7: Take softmax over ˆz to get probabilities {pθ(y | xj 0)}K y=1. 8: Compute conditions {cy}K y=1 for all classes in ˆz. 9: Sample noise ϵ and get noisy sample xj...

  100. [108]

    Guidelines: • The answer NA means that the abstract and introduction do not include the claims made in the paper

    Claims Question: Do the main claims made in the abstract and introduction accurately reflect the paper’s contributions and scope? Answer: [Yes] Justification: We have clearly stated the contributions made in the paper and the scope both in the abstract and introduction. Guidel...

  101. [109]

    Limitations

    Limitations Question: Does the paper discuss the limitations of the work performed by the authors? Answer: [Yes] Justification: We have discussed the limitation of our work in Appendix A. Guidelines: • The answer NA means that the paper has no limitation while the answer No me...

  102. [110]

    Formal proofs are provided in Appendix C

    Theory Assumptions and Proofs Question: For each theoretical result, does the paper provide the full set of assumptions and a complete (and correct) proof? Answer: [Yes] 25 Justification: We state the assumptions made in all theoretical results and properly referenced the Lemm...

  103. [111]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experimental Result Reproducibility Question: Does the paper fully disclose all the information needed to reproduce the main ex- perimental results of the paper to the extent that it affects the main claims and/or conclusions of the paper (regardless of whether the code and da...

  104. [112]

    We have provided dataset details in Appendix F.1, model weights URLs in Appendix F.3, and reproducible instructions with code attached in the supplemental material

    Open access to data and code 26 Question: Does the paper provide open access to the data and code, with sufficient instruc- tions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: The datasets and pre-tra...

  105. [113]

    Experimental Setting/Details Question: Does the paper specify all the training and test details (e.g., data splits, hyper- parameters, how they were chosen, type of optimizer, etc.) necessary to understand the results? Answer: [Yes] Justification: A brief yet informative speci...

  106. [114]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiment Statistical Significance Question: Does the paper report error bars suitably and correctly defined or other appropriate information about the statistical significance of the experiments? Answer: [Yes] Justification: We report all our main results with mean & standar...

  107. [115]

    Guidelines: • The answer NA means that the paper does not include experiments

    Experiments Compute Resources Question: For each experiment, does the paper provide sufficient information on the com- puter resources (type of compute workers, memory, time of execution) needed to reproduce the experiments? Answer: [Yes] Justification: The compute resource ar...

  108. [116]

    Guidelines: • The answer NA means that the authors have not reviewed the NeurIPS Code of Ethics

    Code Of Ethics Question: Does the research conducted in the paper conform, in every respect, with the NeurIPS Code of Ethics https://neurips.cc/public/EthicsGuidelines? Answer: [Yes] Justification: We have preserved anonymity in all submitted materials. Guidelines: • The answe...

  109. [117]

    Guidelines: • The answer NA means that there is no societal impact of the work performed

    Broader Impacts Question: Does the paper discuss both potential positive societal impacts and negative societal impacts of the work performed? Answer: [Yes] Justification: We have discussed the potential societal impacts in Appendix A. Guidelines: • The answer NA means that th...

  110. [118]

    Guidelines: • The answer NA means that the paper poses no such risks

    Safeguards Question: Does the paper describe safeguards that have been put in place for responsible release of data or models that have a high risk for misuse (e.g., pretrained language models, image generators, or scraped datasets)? Answer: [NA] Justification: The datasets an...

  111. [119]

    We have cited related papers, respect their license, and include their license in Appendix B

    Licenses for existing assets Question: Are the creators or original owners of assets (e.g., code, data, models), used in the paper, properly credited and are the license and terms of use explicitly mentioned and properly respected? Answer: [Yes] Justification: The datasets, mo...

  112. [120]

    Guidelines: • The answer NA means that the paper does not release new assets

    New Assets Question: Are new assets introduced in the paper well documented and is the documentation provided alongside the assets? Answer: [NA] Justification: No new assets are released in our work. Guidelines: • The answer NA means that the paper does not release new assets....

  113. [121]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Crowdsourcing and Research with Human Subjects Question: For crowdsourcing experiments and research with human subjects, does the paper include the full text of instructions given to participants and screenshots, if applicable, as well as details about compensation (if any)? A...

  114. [122]

    Guidelines: • The answer NA means that the paper does not involve crowdsourcing nor research with human subjects

    Institutional Review Board (IRB) Approvals or Equivalent for Research with Human Subjects Question: Does the paper describe potential risks incurred by study participants, whether such risks were disclosed to the subjects, and whether Institutional Review Board (IRB) approvals...

Pith tools

Reviewed August 10, 2026 · model on record in the stance chip above.