Pith. sign in

REVIEW 4 major objections 5 minor 23 references

A standard tabular diffusion model trained once on squared-error loss can be conditioned at inference on any differentiable constraint, because the constraint's gradient points along the data manifold rather than off it.

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 ↗

Gradients of any differentiable tabular constraint lie in the tangent space of the data manifold, enabling a single trained diffusion model to satisfy unseen conditions via inference-time guidance.

T0 review reviewed 2026-08-03 challenge →

load-bearing objection Solid empirical paper on inference-time tabular conditioning; the theory overclaims because the tangency guarantee is proved only in the t→0 limit and Figure 3 does not test finite-t tangency. the 4 major comments →

arxiv 2602.07875 v3 pith:GL65DMJH submitted 2026-02-08 cs.LG

Harpoon: Generalised Manifold Guidance for Conditional Tabular Diffusion

Aditya Shankar , Yuandou Wang , Rihan Hai , Lydia Y. Chen This is my paper
classification cs.LG
keywords tabular diffusionmanifold guidanceinference-time conditioningimputationinequality constraintstangent spaceorthogonal projectiontraining-free guidance
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.

The reading

The paper sets out to show that conditional generation for tabular data need not be tied to training-time conditioning. Its central claim is that a diffusion model trained with the ordinary squared-error objective learns a denoising map that acts like an orthogonal projection onto the data manifold, and that the gradient of any differentiable user-supplied loss computed through that projection lies in the manifold's tangent space. On that footing the authors build HARPOON, a sampling procedure that alternates unconditional denoising steps with tangential gradient corrections, enabling imputation, inequality constraints, and mixed categorical/continuous conditions to be enforced purely at inference. If the geometric claims hold, a single pretrained model could serve arbitrary unseen constraints, a practical step beyond methods that require retraining or rejection sampling.

Core claim

The paper's core discovery is a pair of theoretical statements about diffusion models. Theorem 3.1 says that, in the limit as the noise level approaches zero, the model's 'dirty estimate' of the clean sample converges to the orthogonal projection of the noisy input onto the data manifold. Theorem 3.2 then says that if the dirty estimate is an orthogonal projection, the gradient of any differentiable inference-time loss with respect to the noisy input lies in the tangent space of the manifold at the projected point. The authors use the second statement to justify an algorithm that takes gradient steps along the manifold surface while denoising, and they report that this reduces constraint vio

What carries the argument

The central object is the dirty-estimate mapping Q_t, defined by reversing the forward noising step: Q_t(x_t) = (x_t - sqrt(1-\bar{\alpha}_t) \epsilon_\theta(x_t,t)) / sqrt(\bar{\alpha}_t). The paper proves this map converges to the orthogonal projection onto the data manifold and identifies its Jacobian with the identity on tangent vectors. That Jacobian identity is the mechanism that forces gradients of arbitrary inference-time losses into the tangent space. The algorithm then uses these tangential gradients as corrections interleaved with standard denoising, with task-specific losses such as sparsity-inducing absolute error for imputation and ReLU-based soft penalties for inequalities.

Load-bearing premise

The proof of the tangent-space guarantee assumes that, inside a small ball around every clean data point, the data manifold coincides exactly with its tangent plane; that is true only for flat manifolds, so the theorem as proven does not cover the curved geometries the paper advertises.

What would settle it

Train a diffusion model on points sampled from a genuinely curved manifold, such as a sphere embedded in three-dimensional space, then measure the angle between the gradient of an arbitrary differentiable loss and the manifold's tangent plane at the dirty estimate across the full denoising trajectory; if the angle deviates substantially from 90 degrees at any step, the tangent-space guarantee fails outside the flat case. A more direct check is to compute the derivative of the dirty-estimate mapping applied to a tangent vector and compare it with the vector itself—the proof requires exact equal

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

If this is right

  • A model trained once can be repurposed for any number of new conditions at inference, since guidance depends only on a differentiable loss evaluated on the dirty estimate.
  • Inference-time losses can be chosen to suit the data type—for example, sparsity-inducing MAE for one-hot encoded categories—and can outperform the training loss, removing the need to train separate discrete and continuous diffusion branches.
  • For inequality, categorical, conjunctive, and disjunctive constraints, the method reports violation rates below 20% on the tested benchmarks, with the lowest violations on disjunctions because the feasible region is larger.
  • The approach remains practical: sampling cost stays within about twice the unconditional diffusion baseline, and under five seconds on the largest tested dataset.
  • The tangent-space guarantee applies to any differentiable loss, so future constraint types beyond imputation and inequalities—such as domain-specific physical or regulatory bounds—can be plugged in without changing the model or its training.

Where Pith is reading between the lines

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

  • The theorems as stated carry a hidden caveat: the proof of tangent-space alignment relies on exact local flatness of the manifold, which no curved manifold satisfies. A natural repair is to replace the equality with a bound from the manifold's second fundamental form and quantify the resulting angle error; the paper's angle plots suggest the error may be small, which deserves direct testing.
  • If the orthogonal-projection behaviour holds beyond the limiting regime (which the paper verifies only empirically), the same inference-time guidance recipe should transfer to non-tabular mixed discrete/continuous modalities, such as molecular or event-stream data, where constraints are equally diverse.
  • The disjunctive-constraint loss exposes a possible failure mode: when the feasible region is disconnected, gradient guidance may systematically favour one branch. A testable extension is to add a diversity-promoting term and measure both violation rate and branch coverage.
  • Because the algorithm reuses the gradient computed at step t for the correction at step t-1, it implicitly assumes consecutive manifold shells are nearly parallel; this could be checked by measuring the angle between tangent planes at successive dirty estimates, and a large angle would motivate recomputing gradients after each denoising step.
Share X Bluesky LinkedIn Reddit HN

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

4 major / 5 minor

Summary. The paper proposes HARPOON, a training-free conditional tabular diffusion method. The main theoretical claim is that a standard MSE-trained denoiser acts, in a suitable limit, as an orthogonal projection onto the data manifold, and that the gradient of any differentiable inference-time loss computed through this denoiser lies in the tangent space of that manifold. On this basis, the authors design an inference-time guidance algorithm that interleaves unconditional denoising with tangential gradient corrections, and evaluate it on imputation and inequality-constraint tasks across eight tabular datasets. The paper also argues that this generalises prior manifold-guidance results, which were restricted to squared-error losses and flat manifolds.

Significance. If the theoretical claims were established at the level stated, the paper would make a useful contribution: it would extend manifold-guided diffusion from squared-error inverse problems to arbitrary differentiable objectives, and would address a real limitation of training-time conditional tabular models. The paper is commendably explicit about its assumptions, provides proofs in the appendix, gives a reproducible experimental setup, and evaluates on a broad set of baselines and tasks. The empirical gains for categorical imputation and inequality constraints are plausible. However, the central derivation contains a false geometric assumption, an incorrect posterior in the proof of Theorem 3.1, and a finite- t gap between the theorem and the algorithm. These issues are load-bearing for the paper's main claim that HARPOON's updates stay on the data manifold, so the contribution is only partially supported as written.

major comments (4)
  1. [Section 3.1, Assumption 2] Assumption 2 states M0 ∩ B_r(x0) = T_{x0}M0 ∩ B_r(x0) for some r>0. For any genuinely curved manifold this equality is false for every r>0, so the assumption contradicts the paper's advertised generality to 'curved geometries'. The proof of Theorem 3.2 (Appendix B) uses this assumption to justify Eq. (5), J_Q v = v_T. As written, the tangent-space guarantee is therefore proved only for locally flat manifolds. The assumption is repairable: for a smooth manifold, the differential of the orthogonal projection at a regular point is the projection onto the tangent space, so Eq. (5) can be obtained without the false equality. But this repair is not in the manuscript.
  2. [Appendix B, proof of Theorem 3.1] The proof states that under the Gaussian noising framework, p(x0|xt,t) ∼ N(x0|xt, σ_t^2 I). This is not the posterior implied by Eq. (1). From xt = sqrt(αbar_t)x0 + sqrt(1−αbar_t)ϵ, the likelihood is p(xt|x0) ∝ exp(−||xt − sqrt(αbar_t)x0||² / (2(1−αbar_t))), so the posterior is proportional to N(xt; sqrt(αbar_t)x0, (1−αbar_t)I)p(x0), not to N(xt, σ_t²I). The subsequent Gaussian localisation argument using Lemma D.9 of Stanczuk et al. is therefore not justified as written. Since Theorem 3.1 is the only formal support for the orthogonal-projection assumption in Theorem 3.2, this gap is central.
  3. [Section 3.2 / Algorithm 1 / Figure 3] Theorem 3.1 establishes lim_{αbar_t→1} Q_t(x_t)=π(x_t), i.e. a statement only in the limit t→0. Theorem 3.2 assumes that Q_t acts as an orthogonal projection at every t, but that is not proved. Algorithm 1 (lines 6–9) applies the tangential update at every denoising step, including large t. The paper's empirical support for finite t is Figure 3, but that figure plots the angle between ∇_{x_t}L_inf(x̂0,c) and x̂0. Tangency of a vector to T_{x̂0}M0 is not equivalent to orthogonality to x̂0; the relevant normal direction is x_t − x̂0 (or the component normal to the tangent plane). The displayed angle is neither necessary nor sufficient for tangency, so the central 'without leaving the manifold' guarantee is unverified in the operating regime of the algorithm. The text already concedes that strict orthogonality holds only in the limit, which makes the gap explicit.
  4. [Appendix B, proof of Theorem 3.2] After establishing J_Q v = v_T, the proof derives J_Q^T w = w_T by 'matching terms' in the equation v_T^T J_Q^T w + v_N^T J_Q^T w = v_T^T w_T. This inference does not follow from the displayed equation alone: the decomposition of w into w_T + w_N is not introduced, and the conclusion requires J_Q = P_T, which is exactly what Eq. (5) gives for all v only if J_Q is the orthogonal projection onto the tangent space. The algebra should be rewritten to show that J_Q^T = P_T first. As written, the proof is not self-contained at this step.
minor comments (5)
  1. [Figure 3 / Section 4.1] The caption says 'Avg. angle between gradients and dirty estimates', which is the quantity that cannot support tangency. Consider instead reporting the angle between the gradient and the normal component x_t − x̂0, or the projection residual ||(I−P_T)∇L|| / ||∇L||, for several t.
  2. [Equation (4)] The displayed expression for F(x_t,t) contains redundant p(x_t) factors in the numerator and denominator; this is a typo that obscures the Bayes step. Please clean up the notation.
  3. [Algorithm 1, line 9] The gradient g is computed at x_t but applied to the denoised sample x'_{t−1}. The paper justifies this by saying consecutive shells are nearly parallel, but this is stated informally. A short quantitative statement (e.g., how the tangent spaces of M_t and M_{t−1} are related) would strengthen the presentation.
  4. [Remark 1 / Section 3.1] Remark 1 says 'our theoretical derivations provide a way to overcome this limitation for tabular diffusion', but no such derivation is given in the paper. Either provide the argument or soften the claim.
  5. [Remark 2 / Section 4.4] The claim that Table 4 'directly verifies Theorem 3.2' is an overstatement: the table shows that MAE/CE losses improve imputation scores, but it does not measure tangency of gradients. Please rephrase to say the results are consistent with, but not a direct verification of, the theorem.

Circularity Check

0 steps flagged

No significant circularity: the derivation is a conditional chain built on explicit assumptions and external results, not on the authors' own prior work.

full rationale

HARPOON's central theory is a conditional derivation, not a fitted or renamed prediction. Theorem 3.1 is proved from the MSE denoising objective plus an external concentration lemma (Lemma D.9 of Stanczuk et al. 2024); it is not derived from the authors' own results. Theorem 3.2 is an explicit conditional: under the stated assumption that Q_t acts as an orthogonal projection, the chain-rule computation shows ∇_{x_t} L_inf(x̂0,c)=J_{Q_t}^T w, and the tangent-space claim follows. The proof's Eq. (5) is based on the stated Assumption 2 (local linearity), which is an assumption, not a conclusion smuggled in from a citation. The one self-citation (Shankar et al. 2024, Sec. 2) is a peripheral pointer to latent tabular diffusion and is not load-bearing for any theorem or algorithm choice. The acknowledged limiting nature of Theorem 3.1 (ᾱ_t→1) and the fact that Figure 3 measures angles to x̂0 rather than to the normal direction are correctness/validity concerns about the finite-t operating regime, not circularity: they do not make a claimed output equal to an input by construction. No fitted parameter is relabeled as a prediction. Thus no circular step can be exhibited, and the appropriate score is 0.

Axiom & Free-Parameter Ledger

2 free parameters · 5 axioms · 0 invented entities

The paper’s contributions rest on the manifold hypothesis, a local-flatness assumption that is stated too strongly, the assumption that the trained denoiser is Bayes-optimal, and a borrowed technical lemma. It introduces no new physical entities. The only tuned scalar free parameters are the guidance strength and loss coefficients, which are set by hand.

free parameters (2)
  • Guidance step size η = 0.2
    Hand-chosen constant used in Algorithm 1 for all datasets and tasks. The algorithm’s behavior depends on it, and no sensitivity analysis is provided.
  • Soft-constraint coefficients λ_g, λ_h = 1.0
    Used in the inequality-constraint loss (eq. 9). Chosen without tuning; the central inequality experiments rely on these values.
axioms (5)
  • domain assumption Assumption 1: support of p(x_0) lies on a smooth n-dimensional manifold M_0 embedded in R^d
    Standard manifold hypothesis, unproven. It underpins all geometric claims about the data.
  • ad hoc to paper Assumption 2: M_0 ∩ B_r(x_0) = T_{x_0} M_0 ∩ B_r(x_0) for some r>0
    This exact equality is false for any genuinely curved manifold, so the paper implicitly assumes local flatness despite claiming to handle curved geometries. It is used to derive the Jacobian identity in Theorem 3.2.
  • domain assumption The denoiser ε_θ is the MSE-optimal (Bayes) denoiser
    Theorem 3.1’s proof optimizes the training loss pointwise, which requires the neural network to reach the global optimum. Real trained networks only approximate this.
  • standard math Lemma D.9 of Stanczuk et al. (2024) as cited in the proof of Theorem 3.1
    The proof borrows a technical lemma about Gaussian integrals over manifolds; the paper does not reproduce it.
  • domain assumption Categorical features are relaxed into a continuous simplex embedding
    Remark 1 assumes one-hot or soft encoded categories lie in a continuous space so the manifold assumptions apply. This is a modeling choice, not a proven equivalence.

reviewed 2026-08-03 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Harpoon: Generalised Manifold Guidance for Conditional Tabular Diffusion." pith.science (2026). https://pith.science/paper/GL65DMJH

@misc{pith2026260207875,
  author       = {Pith},
  title        = {Pith review of: Harpoon: Generalised Manifold Guidance for Conditional Tabular Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/GL65DMJH}},
  note         = {Machine review of arXiv:2602.07875}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

Generating tabular data under conditions is critical to applications requiring precise control over the generative process. Existing methods rely on training-time strategies that do not generalise to unseen constraints during inference, and struggle to handle conditional tasks beyond tabular imputation. While manifold theory offers a principled way to guide generation, current formulations are tied to specific inference-time objectives and are limited to continuous domains. We extend manifold theory to tabular data and expand its scope to handle diverse inference-time objectives. On this foundation, we introduce HARPOON, a tabular diffusion method that guides unconstrained samples along the manifold geometry to satisfy diverse tabular conditions at inference. We validate our theoretical contributions empirically on tasks such as imputation and enforcing inequality constraints, demonstrating HARPOON'S strong performance across diverse datasets and the practical benefits of manifold-aware guidance for tabular data. Code URL: https://github.com/adis98/Harpoon

Figures

Figures reproduced from arXiv: 2602.07875 by Aditya Shankar, Lydia Y. Chen, Rihan Hai, Yuandou Wang.

Figure 1
Figure 1. Figure 1: Geometry of forward (↑) and backward (↑) diffusion. The Geometry of (Image) Diffusion. Recent works in image dif￾fusion interpret the diffusion process as a sequence of manifold transitions (Chung et al., 2022; He et al., 2024). A clean sample x0 is assumed to lie on a low-dimensional manifold M0 ⊂ R d . The forward process gradually perturbs M0 into a family of sur￾rounding shells {Mt} T t=1, where MT app… view at source ↗
Figure 2
Figure 2. Figure 2: (a) Shows the orthogonal behaviour of dirty estimates projected from [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Avg. angle between gradients (100 samples) and dirty estimates xˆ0 for various loss functions on Adult. We provide the proof in Appendix B. Theorem 3.2 es￾tablishes a powerful result that the gradients of any dif￾ferentiable objective at inference (Linf) are aligned with the tangent space of the data manifold. This gener￾alises prior work, which proved the tangential behaviour only for squared-error infere… view at source ↗
Figure 4
Figure 4. Figure 4: Disconnected submanifolds under disjunctive constraints, e.g. ( [PITH_FULL_IMAGE:figures/full_fig_p019_4.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

23 extracted references · 3 linked inside Pith

  1. [3]

    net/forum?id=V1ZHVxJ6dSS

    URLhttps://openreview. net/forum?id=V1ZHVxJ6dSS. Tianyu Du, Luca Melis, and Ting Wang. Remasker: Imputing tabular data with masked autoen- coding. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vi- enna, Austria, May 7-11,

    2024

  2. [4]

    Chung et al. (2022) assume this integral behaves as a weighted average over all points on the manifold, relying on the symmetry of the normal distribution to argue that the denoised estimate points to the orthogonal projection. However, this assumption requires the manifold to be globally linear, not just locally linear, since the integral spans the entir...

    2022

  3. [6]

    Yidong Ouyang, Liyan Xie, Chongxuan Li, and Guang Cheng

    URLhttps://openreview.net/forum?id=WpKDeixmFr. Yidong Ouyang, Liyan Xie, Chongxuan Li, and Guang Cheng. Missdiff: Training diffusion models on tabular data with missing values.arXiv preprint arXiv:2307.00467,

    Pith/arXiv arXiv

  4. [8]

    Juntong Shi, Minkai Xu, Harper Hua, Hengrui Zhang, Stefano Ermon, and Jure Leskovec

    doi: 10.1109/ICDE60146.2024.00016. Juntong Shi, Minkai Xu, Harper Hua, Hengrui Zhang, Stefano Ermon, and Jure Leskovec. Tabdiff: a mixed-type diffusion model for tabular data generation. InThe Thirteenth International Confer- ence on Learning Representations, ICLR 2025, Singapore, April 24-28,

    arXiv 2024

  5. [9]

    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 9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7,

    2021

  6. [10]

    Stoian, Salijona Dyrmishi, Maxime Cordy, Thomas Lukasiewicz, and Eleonora Giunchiglia

    Mihaela C. Stoian, Salijona Dyrmishi, Maxime Cordy, Thomas Lukasiewicz, and Eleonora Giunchiglia. How realistic is your synthetic data? constraining deep generative models for tabu- lar data. InThe Twelfth International Conference on Learning Representations, ICLR 2024, Vi- enna, Austria, May 7-11,

    2024

  7. [11]

    Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang

    doi: 10.1109/JBHI.2020.2980262. Jiwen Yu, Yinhuai Wang, Chen Zhao, Bernard Ghanem, and Jian Zhang. Freedom: Training-free energy-guided conditional diffusion model. InProceedings of the IEEE/CVF International Con- ference on Computer Vision, pp. 23174–23184,

    arXiv 2020

  8. [12]

    Mixed-type tabular data synthesis with score- based diffusion in latent space

    13 Published as a conference paper at ICLR 2026 Hengrui Zhang, Jiani Zhang, Zhengyuan Shen, Balasubramaniam Srinivasan, Xiao Qin, Christos Faloutsos, Huzefa Rangwala, and George Karypis. Mixed-type tabular data synthesis with score- based diffusion in latent space. InThe Twelfth International Conference on Learning Represen- tations,

    2026

  9. [13]

    Diffusion models for missing value imputation in tabular data.CoRR, abs/2210.17128,

    Shuhan Zheng and Nontawat Charoenphakdee. Diffusion models for missing value imputation in tabular data.CoRR, abs/2210.17128,

    Pith/arXiv arXiv

  10. [14]

    URLhttps: //doi.org/10.48550/arXiv.2210.17128

    doi: 10.48550/ARXIV .2210.17128. URLhttps: //doi.org/10.48550/arXiv.2210.17128. 14 Published as a conference paper at ICLR 2026 A NOTATIONS Table 7: Summary of key notations. Symbol Explanation (A) Data & Manifold Setup dAmbient space dimension x0 ∈R d Clean (un-noised) sample p(x 0)Clean data distribution M 0 ⊂R d Smooth, connected data manifold nIntrins...

    Pith/arXiv arXiv doi:10.48550/arxiv.2210.17128 2026

  11. [15]

    Substitutingϵfrom equation 1, we get: 15 Published as a conference paper at ICLR 2026 Ltrain =E x0,t,ϵ " ϵθ(xt, t)−xt − √ ¯αtx0√1−¯αt 2 2 #

    Proof.Consider the training objective for a denoiser with parametersθunder a continuous Gaussian noising framework equation 2: Ltrain =E x0∼p(x0),t∼U[1,T],ϵ∼N(0,I) ∥ϵθ(xt, t)−ϵ∥2 2 . Substitutingϵfrom equation 1, we get: 15 Published as a conference paper at ICLR 2026 Ltrain =E x0,t,ϵ " ϵθ(xt, t)−xt − √ ¯αtx0√1−¯αt 2 2 # . Ltrain = Z x0 Z xt Z t ϵθ(xt, t)...

    2026

  12. [16]

    (2022):While our result is similar to Proposition 2 from Chung et al

    Key difference from Chung et al. (2022):While our result is similar to Proposition 2 from Chung et al. (2022), the key difference lies in how we treat the integral in equation

    2022

  13. [18]

    The statistics of these datasets are presented in Table 8, each of which is partitioned using a random 70-30 train-test split

    Additionally, we included theCalifornia Housingdataset from Kaggle9. The statistics of these datasets are presented in Table 8, each of which is partitioned using a random 70-30 train-test split. Prior to splitting the datasets, we filtered out the rows having missing values, similar to Zhang et al. (2025). We also remove the target column from training a...

    2025

  14. [19]

    ground truth

    =v,(8) whereℓ, uare elementwise lower and upper bounds (e.g.,Age∈[0,120],Salary≥0),g(·)allows transformations or subset selection. The functionh(·)could be a selector that enforces features to match a particular discrete or continuous valuev. Feasibility can be approximated using soft penalties as proposed by (Donti et al., 2021), which we can write in a ...

    2021

  15. [20]

    too similar

    of the generated samples, quantifying their fidelity. Categorical Constraints.For categorical constraints, we fix the value of a randomly chosen dis- crete feature to a particular class, deliberately avoiding the majority class from the training data. As with range constraints, we construct a ground truth by filtering the test set for rows that satisfy th...

    2026

  16. [21]

    For GAIN, we updated the plugin code to separate pre- training from inference-time masking

    and GAIN (Yoon et al., 2018):https://github.com/ vanderschaarlab/hyperimpute. For GAIN, we updated the plugin code to separate pre- training from inference-time masking. Following the paper (Yoon et al., 2018), we pre-train the model using 20% missing data under an MCAR mask, and evaluate at inference on various masks. • GReaT (Borisov et al., 2023):https...

    2018

  17. [22]

    3:Sample timestept∼Uniform({1, . . . , T}) 4:Sample noiseϵ∼ N(0, I) 5:Predict noise using backboneϵ θ(√¯αtx0 + √1−¯αt ϵ, t) 6:Update parameters with gradient descent on ∇θ ϵ−ϵ θ(xt, t) 2 2 7:untilconvergence H.2 SAMPLING WITHHARPOON At inference time, HARPOONaugments the standard denoising procedure with manifold-aware con- ditional guidance (see Algorith...

    2026

  18. [23]

    In terms of MSE and accuracy, DiffPuter and HARPOONgenerally perform the best, with HARPOONgenerally being the better among the two

    We notice that the overall trends remain consistent with the MAR setting. In terms of MSE and accuracy, DiffPuter and HARPOONgenerally perform the best, with HARPOONgenerally being the better among the two. While methods such as Remasker and Miracle are also strong baselines, we see that their bias towards numeric-type data limits their accuracy when impu...

    2026

  19. [2020]

    Chinchali

    12 Published as a conference paper at ICLR 2026 Sai Shankar Narasimhan, Shubhankar Agarwal, Oguzhan Akcin, Sujay Sanghavi, and Sandeep P. Chinchali. Time weaver: A conditional time series generation model. InForty-first Interna- tional Conference on Machine Learning, ICML 2024, Vienna, Austria, July 21-27,

    2026

  20. [2021]

    Donti, David Rolnick, and J

    Priya L. Donti, David Rolnick, and J. Zico Kolter. DC3: A learning method for optimization with hard constraints. In9th International Conference on Learning Representations, ICLR 2021, Vir- tual Event, Austria, May 3-7,

    2021

  21. [2022]

    Diffusion posterior sampling for general noisy inverse problems

    11 Published as a conference paper at ICLR 2026 Hyungjin Chung, Jeongsol Kim, Michael Thompson Mccann, Marc Louis Klasky, and Jong Chul Ye. Diffusion posterior sampling for general noisy inverse problems. InThe Eleventh Interna- tional Conference on Learning Representations,

    2026

  22. [2023]

    The intrinsic dimension of images and its impact on learning

    Phillip Pope, Chen Zhu, Ahmed Abdelkader, Micah Goldblum, and Tom Goldstein. The intrinsic dimension of images and its impact on learning. In9th International Conference on Learning Representations, ICLR 2021, Virtual Event, Austria, May 3-7,

    2021

  23. [2024]

    Jonathan Ho and Tim Salimans

    URLhttps://openreview.net/forum?id=o3BxOLoxm1. Jonathan Ho and Tim Salimans. Classifier-free diffusion guidance. InNeurIPS 2021 Workshop on Deep Generative Models and Downstream Applications,

    2021

This paper was first reviewed by deepseek-v4-flash on August 3, 2026.