acceptodds
Under review as a conference paper at ICLR 2027

Let It Go or Learn to Self-Correct: Continuous Diffusion for Constrained Discrete Tasks

Abstract

Denoising Diffusion Probabilistic Models (DDPMs) generate samples by starting from noise and repeatedly denoising while keeping each update close to the current noisy state. This behavior is effective in many continuous domains, but its role is less clear for globally constrained discrete tasks, such as Sudoku, graph connectivity, Latin squares, and N-queens. In such settings, early discrete errors can be difficult to undo. As a result, standard diffusion sampling may preserve early mistakes, even when the model’s clean predictions are informative. We compare standard samplers to sampling directly from the model’s clean prediction. Without retraining, this single change improves Sudoku validity from 31% to 95%, with consistent gains across the other discrete tasks. We hypothesize that staying close to the current noisy state is harmful because the reverse trajectory can drift off the forward noising distribution the model was trained on. To reduce this train-test mismatch, we further introduce self-correction training, which exposes the model to its own predictions, improving robustness to errors that arise during inference. This substantially improves the performance of standard samplers. Our results suggest that continuous diffusion models can learn nontrivial global constraints, but discrete reasoning tasks require better alignment between training and inference: either through samplers that reduce commitment to early decisions, or through training that teaches the model to correct its own inference-time errors.

Primary area:
generative models
Keywords:
continuous diffusion, reasoning, DDPM

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_mXjQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies diffusion models for globally constrained discrete tasks, where global constraint rules, such as those in Sudoku, are learned from data rather than explicitly supplied to the model. Complementing previous work on solving these tasks with continuous or discrete diffusion models, it uses them to investigate inference failures in continuous diffusion. The authors observe that generated outputs can be locally valid but globally invalid. Their analysis suggests that the denoiser can propose valid solutions that the standard sampling process subsequently undermines. They address this through Tweedie reprojection, which removes the direct previous-state residual during sampling, and self-correction training, which teaches the model to correct its own errors.

Critical strengths

Using the characteristics of generation on discrete tasks to, in turn, analyze some properties of continuous diffusion models is quite an interesting angle.

And then, they develop corresponding methods on the inference and training sides to alleviate this training–inference mismatch problem. This is quite interesting, and the problem and the methods do match.

The writing itself is quite fluent. It is easy to get the main storyline, and it is fairly self-consistent.

Experimentally, the controlled comparisons do support the correctness of the methods and the analysis. For example, on Sudoku, keeping the model fixed and changing only the sampler leads to a substantial improvement, which is quite a convincing observation.

Critical weaknesses

The algorithmic contributions need clearer delineation. The inference update has a direct antecedent in [1] and structural similarities to [2], while the training approach shares its central motivation and elements with [3,4]. The new algorithmic insight in either component remains unclear.

[1] Song, J., Meng, C., & Ermon, S. (2020). Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502.

[2] Song, Y., Dhariwal, P., Chen, M., & Sutskever, I. (2023). Consistency models. arXiv preprint arXiv:2303.01469.

[3] Deng, Y., Kojima, N., & Rush, A. M. (2022). Markup-to-image diffusion models with scheduled sampling. arXiv preprint arXiv:2210.05147.

[4] Ren, Z., Zhan, Y., Ding, L., Wang, G., Wang, C., Fan, Z., & Tao, D. (2024, March). Multi-step denoising scheduled sampling: Towards alleviating exposure bias for diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 5, pp. 4667-4675).

Questions for authors
  • Why do many failed Tweedie trajectories on Sudoku contain an earlier valid denoiser proposal but end with an invalid sample?
  • What are the substantive differences between the inference method and [1,2], and between the training method and [3,4]? Under a unified parameterization, please distinguish the state construction, re-noising procedure, time sampling, and loss weighting, and explain which differences are necessary for the results.

Should the main contribution be understood as a training algorithm or as a mechanistic analysis of diffusion on discrete constraint tasks?

4 comments

We thank the reviewer for the careful reading and questions!

Our main contribution is a mechanistic analysis identifying an inference failure: a denoiser can predict a constraint-satisfying solution, yet the standard reverse update can turn that prediction into an invalid update center. Our sampling and training interventions investigate this failure from complementary sides.

  1. Removing the residual does not make a valid proposal an absorbing state: subsequent re-noising and imperfect denoising can move away from it. We also realized that the meaning of "valid" in line 355 (375 in revised) was different from the rest of the paper. Here, it meant satisfying the Sudoku rules, while elsewhere it also requires satisfying the given clues. Under this latter definition, a valid proposal appears in only 10% of failed Sudoku-Extreme trajectories. We updated the wording in the revised manuscript.

  2. Appendix B.1 gives the unified parameterization, we clarify it further in the Appendix N (revised manuscript):

\qquad \hat\epsilon = \frac{x_t-\alpha_t\hat x_0}{\beta_t}, \quad \xi\sim\mathcal N(0,I).$$ DDIM sets $\kappa=0$, DDPM sets $\kappa^2$ to the posterior variance, and Tweedie sets $\kappa=\beta_s$, removing the direct state residual. Even though Tweedie is a special case of generalized DDIM[1], we are not aware of prior work showing that this choice of $\kappa$ makes continuous diffusion work on constrained discrete tasks. On the Sudoku-21 baseline checkpoints of Table 1, DDIM ($\kappa=0$) reaches 17% (Table 16 in the revised manuscript, $\rho=0$ column), versus 31% for DDPM and 95% for Tweedie reprojection. In fact, sweeping $\rho$ across all ten configurations and both training regimes (Table 16 in the revised manuscript) shows that validity generally improves toward the Tweedie endpoint. Figure 2 isolates the residual's effect on decoded center validity, while the noise-scale controls show that changing DDPM's noise magnitude alone does not close the gap (Figure 6). Consistency models [2] also alternate prediction and re-noising, but learn a mapping consistent along probability-flow trajectories. At inference, our update shares the predict-and-re-noise structure of multistep consistency sampling, but uses a standard denoiser without consistency training or distillation. Our self-correction directly supervises recovery of the ground-truth solution from a re-noised model prediction. Consistency training instead matches model outputs on two versions of the same data example corrupted at different noise levels. Deng et al. [3] construct inputs through short learned reverse rollouts (one neighboring step in their experiments) and progressively mix these with standard noise-MSE training. Ren et al. [4] also start from the model's own one-step prediction, but instead of running the model again they propagate the resulting error analytically over several further steps, and they add a second output head that learns to predict and subtract this input error at inference. We instead sample $t_1\sim U[0,T]$, $t_2\sim U[0,t_1]$, and directly construct $$\tilde x_{t_2} = \alpha_{t_2}\operatorname{sg}[f_\theta(x_{t_1},t_1)] + \beta_{t_2}\xi.$$ The target is the original $x_0$, with objective $L_{\mathrm{rec}}+\lambda_{\mathrm{simple}}L_{\mathrm{simple}}$ and $\lambda_{\mathrm{simple}}=0.1$ in the main experiments. In our case noise levels are not necessarily adjacent. Figure 8, Appendix I.2 (Figure 9 in the revised manuscript) compares re-noised clean predictions and DDPM-step recovery inputs within our training setup. They produce comparable improvements on Sudoku, showing that the recovery benefit persists across both setups. [1] Song, J., Meng, C., & Ermon, S. (2020). Denoising diffusion implicit models. arXiv preprint arXiv:2010.02502. [2] Song, Y., Dhariwal, P., Chen, M., & Sutskever, I. (2023). Consistency models. arXiv preprint arXiv:2303.01469. [3] Deng, Y., Kojima, N., & Rush, A. M. (2022). Markup-to-image diffusion models with scheduled sampling. arXiv preprint arXiv:2210.05147. [4] Ren, Z., Zhan, Y., Ding, L., Wang, G., Wang, C., Fan, Z., & Tao, D. (2024, March). Multi-step denoising scheduled sampling: Towards alleviating exposure bias for diffusion models. In Proceedings of the AAAI Conference on Artificial Intelligence (Vol. 38, No. 5, pp. 4667-4675).

Thank you for the detailed response and additional experiments.

The additional experiments provide stronger support for the proposed analysis. However, the algorithmic contributions remain closely related to existing methods. Overall, while I appreciate the mechanistic insights, I find the novelty and broader contribution to the community somewhat limited.

Considering these factors, I will maintain my original scores.

Thank you for the follow-up! Over the weekend, we ran additional experiments to compare our approach more directly with [3,4]. In particular, we ablated the recovery input used in training on 21-clue Sudoku:

Recovery input used in trainingDDPMTweedie
Ours: re-noised prediction at a random lower noise level0.830.96
DDPM step to a random lower noise level0.770.97
Ours, restricted to the adjacent noise level0.160.88
Deng-style input: one DDPM step to the adjacent level0.220.90
Ren-style input: model step + analytic steps, at most 5 levels0.190.91

Rows 1 and 3 use our self-correction and differ only in the range of the second noise level; rows 2 and 4 make the same comparison using a DDPM-step input. This directly addresses which design choice matters in our setting: sampling the second noise level over a wider range matters much more than whether the recovery input is constructed by re-noising or a DDPM step. Under DDPM, the random lower level variants reach 0.77-0.83 validity, while the short range variants reach only 0.16-0.22, below approximately 0.30 for standard training. Under Tweedie, the difference is smaller but follows the same pattern.

We also want to emphasize why we believe the mechanistic analysis is an important contribution. We explicitly acknowledge the relationships to generalized DDIM and scheduled sampling. However, our main finding is that a denoiser can already predict constraint-satisfying solutions that the standard reverse update fails to use. This changes how reported failures should be interpreted. For example, in SRM, the standard diffusion baseline reaches only 0.8% on MNIST-Sudoku, motivating a spatial sampling order (best: 51.6%). We took that same released baseline checkpoint, which we did not train, and changed only the sampler. Accuracy increased to 65.7%, while sampling remained fully parallel. Thus, the 0.8% result does not mean that the denoiser failed to learn the constraints: the sampling procedure was a major part of the failure.

We believe that identifying this distinction, and showing how it can be addressed from both the sampling and training sides, is an important contribution to understanding continuous diffusion on constrained discrete tasks.

Thank you again for the discussion! We wanted to share two additional results in the new revision that we hope address your remaining concerns.

First, we trained the method from Deng et al. including its mixing schedule. On 21-clue Sudoku, it reaches 20% DDPM validity, compared with 83% for our self-correction, both trained for 400k steps (Appendix Q, Table 19). Together with our previous ablations, this shows that sampling recovery noise level over a wider range makes a substantial difference with respect to using adjacent steps.

We also evaluated our method on Kim et al.'s Sudoku benchmark (see our discussion with q4WJ). Our 6.4M-parameter model reaches 98.8% accuracy with 8 denoiser calls, compared with 89.49% for their published 6M-parameter masked diffusion model at 50 steps (Section 4, Figure 3, Appendix M).

We hope these new results help clarify both the differences from prior work and the practical value of our findings. We would be grateful if you could reconsider your score in light of these additions!

Reviewer_q4WJOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper studies the adaptation of continuous diffusion to constrained discrete tasks. The authors find that a simple modification to the standard DDPM sampler improves performance on these tasks. Given this observation, the authors augment the standard diffusion denoising loss with an additional self-correction loss, improving performance even for weakly performing samplers.

Critical strengths
  • The authors propose a simple sampling modification, Tweedie reprojection, which allows continuous diffusion models to perform well on constrained reasoning tasks at inference time.
  • Their loss modification also improves other samplers, which I find quite interesting.
Critical weaknesses

I honestly enjoy this paper despite my limited familiarity with the literature on continuous diffusion for structured tasks. But my main concern is that discrete diffusion is a natural fit for these tasks, and given the results in the table, there is still a gap on some of the tasks. So I wonder about the general motivation for using continuous diffusion on such tasks.

One important motivation for continuous diffusion is speed, such as few-step or one-step generation, but the paper does not study such efficiency gains over discrete diffusion either. A matched-compute comparison would strengthen the paper. If such results with all the tasks are given, I am happy to update my score (given that it's reproducible and useful for the community).

Minor comments on the writing:

  • It would be nice if the authors used plots when presenting the results of table 1, as this could help readers quickly observe the improvements. Particularly comparison with the discrete block.
  • Figure 2 is hard to follow and could be presented more simply with a descriptive caption.
  • I recommend moving Figure 7 from the appendix to the main paper, since it shows the contribution of your self-correction objective better. It would also be better to evaluate this on more tasks.
Questions for authors
  • My main question is whether the authors can address the benefit of this approach over discrete diffusion or other approaches. Is it efficiency, or is it that we can make continuous diffusion work on these tasks?
  • With λ_simple equal to 0.2 or smaller overall, this means that the objective's benefit comes more from refinement. I wonder whether there is a connection to iterative refinement and deep supervision in HRM?
7 comments

We are happy that you enjoyed the paper, and thank you for the concrete suggestions!

  1. Our main goal is to understand why continuous diffusion fails on these tasks, and in particular to separate failures of the denoiser from failures of sampling. We show that the denoiser can already predict constraint-satisfying solutions, but standard sampling does not always use them effectively. Once this is addressed, we can also benefit from the existing continuous-diffusion sampling toolbox.

As a simple example of applying this toolbox, we change only the sampling time grid (such techniques are used in [1,2]). We use our existing self-correction checkpoints without fine-tuning. We take one step from noise to an intermediate level chosen on the validation set, and sample from there with few Tweedie steps.

Below we compare against parameter-matched discrete diffusion D3PM with its best sampler (ancestral, confidence, or remasking). Resulting test validity is reported in the table below:

SamplerSudoku 21MediumHardExtreme pass@1NQNQ genLatinLatin genGC 12GC 18
SC + Tweedie, 8 calls0.830.980.490.400.840.220.940.341.001.00
D3PM, best with ≤8 calls0.040.660.020.070.970.050.290.010.850.59
SC + Tweedie, 16 calls0.960.990.920.410.840.440.990.991.001.00
D3PM, best with ≤16 calls0.350.770.150.160.970.470.720.360.970.92

At 8 calls, we obtain higher validity on 9/10 tasks, and at 16 calls, on 8/10.

Thank you for suggesting the efficiency comparison!

  1. Yes, there is a connection: in HRM's deep supervision, each detached segment is supervised against the final target. Our recovery loss follows a similar principle: a detached prediction is re-noised and supervised against x0x_0. The λsimple\lambda_{\mathrm{simple}} sweeps (Figures 7 and 8), where small values often work as well as 0.1, support the importance of refinement supervision. Unlike HRM, we do not carry a separate hidden state across steps; we re-noise the model's prediction and use the same denoiser again.

We updated Figure 2. We kept Figure 7 in the appendix for space but extended it to three more tasks (Figure 8, Appendix I.1). We agree that the Table 1 has a lot of numbers and we would love to replace it with the Figure, however, with the co-authors we did not find the satisfying way after a long discussion. We included our best alternative in the appendix of a revised manuscript (Figure 10), although we are not fully satisfied with it. If you have any other suggestions, we will be very happy to hear them! Thank you.

[1] Song, Jiaming, Chenlin Meng, and Stefano Ermon. "Denoising diffusion implicit models." arXiv preprint arXiv:2010.02502 (2020).

[2] Karras, Tero, et al. "Elucidating the design space of diffusion-based generative models." Advances in neural information processing systems 35 (2022): 26565-26577.

Dear authors, thank you for providing the comparison with discrete diffusion. These results are encouraging. However, I remain concerned about whether the discrete diffusion baseline reflects a competitive modern implementation. Although Appendix E.1 describes a masked diffusion model, masked-token cross-entropy alone does not clarify how its objective, loss weighting, and training setup compare with MDLM [1], which I consider a relevant baseline.

Recent masked diffusion works such as [2,3] also provide implementations with strong results on structured reasoning tasks. My suggestion is to use these as references for constructing a competitive discrete baseline. Although their evaluation settings differ, they could provide a useful starting point for comparison on your tasks.

The new results demonstrate an advantage over the evaluated D3PM implementation. However, until it is clear that your D3PM baseline is competitive with recent masked diffusion methods, I remain uncertain whether these results establish an efficiency advantage over discrete diffusion. Addressing this would strengthen the paper.

[1] Sahoo et al. Simple and Effective Masked Diffusion Language Models.

[2] Ye et al. Beyond Autoregression: Discrete Diffusion for Complex Reasoning and Planning.

[3] Kim et al. Train for the Worst, Plan for the Best: Understanding Token Ordering in Masked Diffusions.

Thank you for the follow-up!

Our D3PM baseline is a masked diffusion model: noise replaces cells with a [MASK] token, and the model is trained to predict the masked cells with a cross-entropy loss. It belongs to the same model family as MDLM, which uses a different weighting of the training loss across noise levels. It has the same architecture and number of parameters as our continuous model, making denoiser-call budgets directly comparable. We conducted extra experiments:

  1. We re-evaluated the existing D3PM checkpoints with stronger decoding techniques: Kim et al.'s top-probability-margin ordering and Ye et al.'s top-K decoding with re-masking. At each inference budget, we report the best result across the tested decoders and step schedules, counting actual denoiser calls.
SamplerSudoku 21MediumHardExtremeNQNQ genLatinLatin genGC 12GC 18
SC + Tweedie, 8 calls0.830.980.490.400.840.220.940.341.001.00
D3PM, best decoder, ≤8 calls0.210.710.150.070.970.380.690.330.840.59
SC + Tweedie, 16 calls0.960.990.920.410.840.440.990.991.001.00
D3PM, best decoder, ≤16 calls0.610.860.430.160.970.800.860.790.970.92
  1. Keeping the architecture fixed, we also retrained masked-diffusion models using (i) the MDLM objective across the tasks and (ii) Ye et al.'s training recipe on Sudoku.

With its best decoder, MDLM training gives no consistent improvement over our original D3PM in these experiments.

Ye et al.'s training recipe gives comparable results on Sudoku-21, Medium, and Hard (0.60, 0.87, 0.39 at 16 calls), and improves Sudoku-Extreme to 0.24 at 16 calls and 0.40 at 64 calls, nearly matching our continuous model's performance (0.41) on that benchmark at the larger budget. However, our continuous model still obtains higher validity on 8 of 10 configurations at both 8 and 16 denoiser calls. We believe this shows that continuous diffusion is a promising direction for these tasks.

Reviewer_Zs5cOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This work focuses on using a continuous diffusion model to solve discrete constraint tasks. They propose two methods. The first one, Tweedie reprojection, is an inference trick that does not explicitly rely on x_t during inference. In this way, the errors in x_t cannot accumulate. The method works well when the solution is "single-mode". The second one requires additional training but is more general. They show that both methods can significantly boost the performance of the continuous diffusion model in solving some simple discrete games, including Latin squares, N-queens, and Sudoku.

Critical strengths
  1. The paper is generally well written and easy to follow.
  2. This is an interesting work in understanding the continuous diffusion's capacity in solving discrete constraint games. The effectiveness of the proposed method is significant.
  3. The empirical study in this simple setup is arguably comprehensive.
Critical weaknesses
  1. While this is an interesting work, the problem itself seems not that well motivated. In particular, I do not really think a continuous diffusion model is the right method to solve constrained discrete games. As far as I know, LLMs and discrete diffusion models would be a better fit for this kind of problem.

  2. The proposed method is actually not that new. The first algorithm is just a special case of DDIM. For the second, in my opinion, the method is not very different from "Markup-to-Image Diffusion Models with Scheduled Sampling, ICLR2023" (although they work on different tasks and show much more significant improvements).

  3. As mentioned in point 1, it might make more sense to test this method in some more critical applications, like 3D molecular generation like EDM (Hoogeboom et al., 2022). In this way, the models' impact could be significantly improved.

Questions for authors

I've mentioned several points in the critical weakness section. If they are addressed well, I'm willing to increase my score.

3 comments

We thank the reviewer for recognizing the strengths of our paper and for the questions!

  1. Continuous diffusion is a well-established framework, with a lot of research devoted to making sampling faster and more accurate. This is why showing that it can work for constrained discrete tasks is important: it unlocks all the techniques developed there. Our work therefore focuses on identifying the reason for the failure and how one can mitigate it from inference or training sides.

To illustrate the interest to use continuous diffusion, we change only the sampling time grid at inference, following the broader idea of optimizing time discretizations [4]. This gives higher validity for continuous diffusion models than parameter-matched discrete diffusion D3PM at 8 denoiser calls on 9 of 10 benchmarks(see our response to the first question of q4WJ and Table 15 in the revised manuscript).

  1. Indeed, Tweedie reprojection is a special case of generalized DDIM[1], as described in Section 3.2 and Appendix B.1. Please, note, we consider that our main contribution is identifying why diffusion fails on discrete reasoning tasks and showing how this failure can be addressed through sampling and training. To our knowledge, previous work on generalized DDIM has not established that removing the predicted-noise residual is particularly effective for these tasks. For example, the SRM paper[2] addresses this problem through spatially ordered sampling, achieving 51.6% validity. In contrast, applying our still-parallel sampler to the released diffusion-baseline checkpoint raises accuracy from 0.8% to 65.7%, without retraining.

Self-correction provides a complementary training-side intervention. Its exposure-bias motivation is shared with Deng et al.[3], whom we already cite in Related Work. They use short learned reverse-DDPM rollouts; we directly re-noise a stopped-gradient clean prediction at a randomly sampled lower noise level, following the Tweedie construction. Figure 8, Appendix I.2 (Figure 9 in the revised manuscript) compares similar losses.

  1. This makes a lot of sense, thank you for your suggestion! Following your suggestion, we ran a preliminary experiment using the released EDM checkpoint on QM9, changing only the sampler (2,000 molecules). At 51 denoiser calls, Tweedie reprojection improves molecular stability from 73.9% to 91.8% and validity from 88.3% to 96.4%. However, novelty decreases from 54.9% to 39.7%, suggesting a trade-off between stability and novelty. At 51 calls, stability among novel molecules however improves, from 80.3% to 92.7%.

These results are encouraging, but we believe molecular generation deserves a dedicated study to understand this trade-off. Our present work focuses on controlled discrete benchmarks, and we leave a more systematic investigation of molecular generation to future work.

SamplerDenoiser callsMolecule stability (%)Validity (%)Uniqueness (%)Novelty (%)Novel among stable (%)Stable among novel (%)
EDM (DDPM sampler), default1,00182.392.399.755.853.585.9
EDM (DDPM sampler)5173.988.399.754.952.480.3
Tweedie reprojection5191.896.498.439.738.192.7

[1] Song, Jiaming, Chenlin Meng, and Stefano Ermon. "Denoising diffusion implicit models." arXiv preprint arXiv:2010.02502 (2020).

[2] Wewer, Christopher, et al. "Spatial reasoning with denoising models." arXiv preprint arXiv:2502.21075 (2025).

[3] Deng, Yuntian, Noriyuki Kojima, and Alexander M. Rush. "Markup-to-image diffusion models with scheduled sampling." arXiv preprint arXiv:2210.05147 (2022).

[4] Karras, Tero, et al. "Elucidating the design space of diffusion-based generative models." Advances in neural information processing systems 35 (2022): 26565-26577.

Reviewer_Zs5cOpenReview ↗

Thank you for the reply.

All my concerns are addressed, and the results on molecular generation are interesting. Overall, I think the current version is worth publishing and will provide enough insights to inspire future research; thus, I raise my score to 3. I would only raise it to 4 if the work had been demonstrated mainly on more critical applications, such as molecular generation. I suggest the authors mention the molecule results briefly in the revision, as I think this would inspire related research.

BTW, just for your future research, there are ways to increase novelty / generated sample variety in molecular generation; for example, see Ambient Proteins (Daras et al., 2025, NeurIPS). You may find a way to use the same ideas to mitigate concerns about decreased novelty.

Thank you for the encouraging feedback, for updating your score, and for pointing us to Ambient Proteins!

We have added the QM9 results to the revised manuscript, with a short discussion in Section 4 and the full results in Appendix N.

Comments on the paper

2 comments

We thank all reviewers for their careful reading and constructive feedback.

We uploaded the revised manuscript. There, we added a few-step comparison with D3PM (Appendix M), a sweep of the generalized-DDIM family (Appendix N), and self-correction loss-weight sweeps on Latin squares, N-Queens, and graph connectivity (Figure 8, Appendix I.1). We also updated Figure 2 and added a graphical version of Table 1 (Appendix O).

We address the individual questions in the separate responses.

Update:

We would like to thank you all for the discussion and suggestions! They helped us improve the paper.

We have uploaded a new revision. The main additions since the previous version are:

  • Figure 3 and Section 4: Accuracy versus the number of denoiser calls across all tasks and on the Sudoku benchmark of Kim et al.
  • Appendix M: Stronger masked-diffusion baselines, including the decoders of Kim et al. and Ye et al., the MDLM objective, and Ye et al.'s training recipe. With these additional baselines, our model achieves higher validity on eight of ten settings at both 8 and 16 calls.
  • Section 4 and Appendix N: Molecular generation results using the released EDM checkpoint on QM9.
  • Appendix Q: Self-correction input ablations, including a comparison with Deng et al.

Some references in our earlier comments have shifted with the new additions: the generalized-DDIM sweep is now in Appendix O, the loss-weight sweeps are in Figure 9, and the graphical version of Table 1 is in Appendix P.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.