Pith. sign in

REVIEW 3 major objections 4 minor 76 references

MotifRole-Diff: Risk-Optimal Role-Aware Corruption for Masked Molecular Graph Diffusion

T0 review · 3 major / 4 minor · reviewed 2026-08-01 · deepseek-v4-flash

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

Pith's one-line read Uniform masking wastes the diffusion budget: a risk-optimal role-aware schedule improves molecular graph generation at equal compute.

desk verdict The role-aware masking package works and is cheap to adopt, but the ablation never isolates the schedule from the inverse-exposure reweighting, and the paper's 'no-op at γ=1' claim is flatly wrong. read the letter →

arxiv 2607.21634 v1 pith:P3TWCTBT submitted 2026-07-20 cs.LG cs.AI

Tasfia Nuzhat Ornee , Elias Hossain , Ivan Garibay , Niloofar Yousef This is my paper
classification cs.LGcs.AI
keywords maskeddiscretediffusionmoleculargraphgenerationrole-awarecorruptionrisk-optimalmaskingtokenrolesmotifserializationlosslessdecodingQM9/MOSESbenchmarks
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

The paper argues that in masked discrete diffusion for molecular graphs, corrupting every serialized token with the same schedule is wasteful. It shows that tokens play distinct structural roles—syntax, motif interiors, and cross-motif interfaces—that differ sharply in reconstruction difficulty and in how damaging an error is to the decoded molecule. On that basis it derives a risk-optimal allocation of a fixed masking budget: hard, high-impact roles are masked less and easy, low-impact roles more. Under matched architecture, training budget, and sampling compute, this role-aware schedule improves validity and distributional fidelity on QM9 and MOSES, and validity on GuacaMol. The point is that smarter corruption, not more compute, is what drives the gain.

What carries the argument

The load-bearing object is the role criticality c_r = I_r ε_r, the product of the expected graph-level impact of an error in role r and the residual reconstruction error of a finite-capacity denoiser for that role, estimated from single-token perturbation trials and probe negative log-likelihoods. Given role frequencies π_r and a fixed masking budget ρ̄, the schedule solves a linear program whose KKT conditions give a water-filling allocation: roles with criticality above the budget multiplier λ are protected (masked less), roles below are corrupted more. A KL-regularized version yields a unique smooth schedule ρ_r* = σ(σ^{-1}(ρ̄) + η(λ − c_r)), which is strictly decreasing in criticality; p

What would settle it

Run the same matched-compute experiment with the opposite allocation—corrupting the interface role more and syntax/interior less—under the identical budget. If that reverse schedule also improves validity or FCD versus uniform masking, the measured criticality ordering is not the mechanism. A more direct test compares predicted role-weighted risk R(ρ) against measured graph distortion under joint random corruption; if the linear sum underestimates distortion when two interface errors occur together, the independence assumption fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that uniform masking is risk-optimal only when all token roles have equal criticality, defined as the product of a role's reconstruction difficulty and its graph-level impact; when criticality varies, the optimal fixed-budget corruption protects high-criticality roles. It proves a water-filling theorem for the modeled role-weighted risk and a smooth logistic schedule that interpolates from uniform masking to the threshold rule. Empirically, with everything else held fixed, the derived schedule raises validity on QM9 from 0.905 to 0.944 and lowers FCD from 1.701 to 1.609, with analogous gains on MOSES, while preserving the lossless decoder and adding no parameters

Load-bearing premise

The derivation assumes that errors on different token roles contribute to graph-level distortion independently and additively, and that a role's residual error is adequately measured by probe negative log-likelihood; if simultaneous errors compound or interact, the water-filling schedule may not minimize real graph risk.

Editorial extensions

If this is right

  • If role criticalities are heterogeneous, every non-uniform allocation that shifts exposure from high- to low-criticality roles strictly reduces the modeled risk, so uniform masking provably leaves quality on the table.
  • The same derived schedule transfers across datasets without retraining: the QM9-derived exponents improve MOSES and GuacaMol, indicating the criticality ranking captures a structural signal rather than dataset noise.
  • Because the clean sequence space, grammar, and lossless decoder are untouched, any grammar-valid sequence still decodes to the same graph; the method changes only the forward corruption kernel.
  • The schedule adds no parameters, no forward passes, and negligible wall-clock overhead, so the gains are attributable to informed allocation rather than added compute.
  • The inverse-exposure loss weighting is a no-op under uniform masking, so MotifRole-Diff contains uniform MDLM as its flat-criticality special case.

Reading between the lines

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

  • The same risk-allocation logic should apply to any losslessly serialized graph—knowledge graphs, program syntax trees, or source code—wherever token roles have measurable difficulty and downstream impact; the paper's framework is not molecule-specific.
  • A natural next test is to make the token–role prior context-conditioned rather than based on the static training prior; the paper itself flags that the current P(r|v) could be sharpened, which would likely sharpen the derived schedule's effect.
  • Joint protection of motifs rather than individual tokens might capture error interactions that the linear risk model ignores; an extension could define roles at the motif level and allocate budget across motifs.
  • The fidelity–diversity trade visible in the novelty dips suggests the schedule behaves like a regularizer; one could tune η per application to control how much diversity one trades for validity.
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 / 4 minor

Summary. The paper proposes MotifRole-Diff, a role-aware corruption schedule for masked discrete diffusion on serialized molecular graphs. The authors introduce mSENT, a motif-aware serialization that labels tokens by structural role (special, syntax, interior, interface), measure each role's reconstruction difficulty and graph-level perturbation impact, and derive a corruption schedule by solving a fixed-budget risk-allocation problem whose objective is a role-weighted residual-risk bound. They report that, under matched architecture, training budget, and sampling compute, this schedule improves validity and FCD over uniform MDLM on QM9 and MOSES and validity on GuacaMol, with multi-seed confidence intervals. The theoretical contribution is a water-filling characterization of the optimal allocation for the modeled risk, with uniform MDLM as the flat-criticality special case.

Significance. If the central attribution claim holds, the paper makes a useful contribution: it identifies role heterogeneity in serialized molecular graphs and provides a principled, budget-matched way to reallocate masking. The matched-compute experimental design is a genuine strength, as are the five-seed confidence intervals and the explicit statement that the theorem certifies only the modeled risk. The risk-allocation framing with a water-filling optimum is clean and potentially transferable beyond molecules. However, the paper's main empirical claim is currently undercut by an attribution gap: the final configuration changes two coupled components, and the ablation does not isolate them. The paper also contains a formally incorrect statement about the inverse-exposure weighting being a no-op at uniform rates. These issues are fixable with additional experiments, but they are load-bearing for the paper's core message.

major comments (3)
  1. [Eq. (4) and Table S2] The claim that the inverse-exposure weighting is 'a no-op at γ_r≡1 (w≡1)' is incorrect. With γ_r=1, Pr(i∈M_t|r_i)=α_t=1−exp(−Λ(t)), so w(t,r_i)=1/α_t+ε, which is time-dependent and not identically 1. Thus a uniform-γ version of MotifRole-Diff is not uniform MDLM; it is an inverse-time-weighted MDLM. This matters because Table S2 compares only (mSENT, uniform, reweighting off) against (mSENT, role-aware, reweighting on). The measured gains conflate the role-aware schedule with the inverse-exposure reweighting. Please correct the no-op statement and add a schedule-only condition (γ* with w≡1 or with the standard MDLM weighting) and a reweighting-only condition (γ≡1 with w as in Eq. 4).
  2. [Results, RQ3] The paper explicitly concedes that 'we do not isolate their individual effects' for the schedule and the inverse-exposure weighting. This is not a peripheral caveat: the central claim is that structurally informed corruption allocation, rather than the auxiliary loss reweighting, is what improves generation. The current evidence is also compatible with the alternative hypothesis that inverse-exposure reweighting alone (e.g., upweighting hard interface tokens even under uniform corruption) accounts for the validity and FCD gains. The missing ablation rows are essential to support RQ3 as stated. If the schedule-only condition does not reproduce the gains, the claims should be revised accordingly.
  3. [Theoretical Analysis, Eq. (5)] The risk bound E[Δ_graph]≤Σ_r π_r I_r ε_r ρ_r assumes that errors on different token roles contribute linearly and independently, and that the residual error ε_r is independent of exposure ρ_r. The supplementary remark acknowledges an exposure-dependent generalization, but the schedule is derived from the linear form. If role errors interact or if ε_r depends on ρ_r, the water-filling allocation may not minimize actual graph risk. Since this model is the basis for calling the schedule 'risk-optimal', please state this limitation in the main text and provide a diagnostic comparing the modeled risk R(ρ) with measured graph distortion under the derived and uniform schedules, so readers can judge whether the linear proxy is useful.
minor comments (4)
  1. [Supplementary Section S2] The schedule derivation fixes special at γ=1 and clips other exponents to [0.25,4.0]. Please clarify whether the fixed-budget constraint is adjusted for the special role and report sensitivity to the clipping bounds; otherwise the 'derived rather than tuned' phrasing overstates the role of Eq. (9).
  2. [Figure 2 caption and Table S1] The caption says uniqueness is 'already saturated' on GuacaMol, but Table S1 reports uniform MDLM uniqueness 1.000 and MotifRole-Diff 0.984. The word 'saturated' is misleading; please rephrase, e.g., 'near-ceiling on the uniform baseline.'
  3. [Eq. (4)] The additive constant ε in the inverse-exposure weight is notationally too close to the residual error ε_r used in Eq. (5). Use a different symbol, such as δ or c, to avoid confusion.
  4. [Table 2] For η=0 the table lists all exponents as 1.00, but η=0 is the uniform-MDLM limit. Since the table is about sensitivity to η, please state explicitly that η=0 is not a fitted schedule but the reference baseline.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the schedule is an explicit surrogate optimization and the reported generation gains are independent held-out measurements; the main caveat is an attribution confound, not circularity.

full rationale

The paper's derivation chain is transparent: it defines a role-decomposed surrogate risk R(ρ)=Σ_r π_r I_r ε_r ρ_r (Eq. 5), then minimizes that surrogate under a fixed budget (Eq. 7), obtaining a water-filling allocation (Theorem 1) and a smooth schedule (Eq. 9). This is a valid mathematical optimization of an explicitly stated objective, not a hidden circular reduction. The paper repeatedly and explicitly limits the theorem's scope: 'Our theorem characterizes optimality for the modeled role-weighted residual risk, while downstream generation performance is evaluated empirically' and 'the theorem certifies only the modeled role-weighted risk, so the generation gains remain empirical.' Criticalities are measured on a uniform-MDLM probe and used to set the schedule; this is a data-driven design choice, not a fitted parameter disguised as a prediction, because validity, FCD, uniqueness, and novelty are evaluated on held-out reference splits under matched architecture and compute and are not constructed to equal the fitted risk. There is no load-bearing self-citation and no imported uniqueness theorem; the SENT/mSENT serialization is prior external work by other authors. The genuine caveats are empirical-validity issues, not circularity: (1) the claim that inverse-exposure reweighting is 'a no-op at γ_r≡1 (w≡1)' is inconsistent with Eq. 4, since with γ=1, P(i∈M_t|r_i)=α_t=1−exp(−Λ(t)), so w=1/(α_t+ε)≠1; and (2) the Table S2 caption concedes the final configuration 'does not isolate their individual effects,' so the improvement cannot be attributed solely to the role-aware schedule. These concerns affect attribution of the empirical gains, but they do not make the derivation equivalent to its inputs by construction.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

The paper introduces no new physical or mathematical entities. The free parameters are mostly schedule hyperparameters (η, clipping, special role exponent, probe range, normalization), while the axioms include the linear risk decomposition and the use of NLL as a proxy for residual error. The risk model itself is the largest ad hoc element.

free parameters (6)
  • eta (η) = 2.0 (chosen; sensitivity 0.5,1,2,4,8)
    Controls the magnitude of the allocation away from uniform; direction is robust across η>0, but the magnitude is a hand-chosen hyperparameter.
  • γ clipping bounds = [0.25, 4.0]
    Clipping limits the extreme exponents; hand-set, not derived.
  • special role exponent = 1.0
    Special tokens are fixed at uniform rate; not derived from criticality.
  • masking budget ρ̄ = implicit (matching uniform MDLM average exposure)
    The total expected masking rate is fixed to equal uniform MDLM, but its numerical value is not specified.
  • probe time range = t ~ U(0.05, 0.95)
    The range over which probe NLL is measured is a choice.
  • normalization in criticality = unspecified normalization of D_r and I_r
    The paper uses Norm(D_r)·Norm(I_r) but does not specify the normalization function, which affects the scale of c_r and thus the schedule.
assumptions (5)
  • standard math Serialization is a bijection: D(T(G)) ≅ G and H(Z) = H(G).
    Used to argue that schedules matter only because of finite-capacity denoisers; standard information-theoretic invariance.
  • domain assumption Graph risk decomposes as E[Δ_graph] ≤ Σ_r π_r I_r ε_r ρ_r.
    Eq. 5; assumes errors from different roles contribute linearly and independently, and that a role-r token harms the graph only when corrupted and then misreconstructed. This is the core modeling assumption.
  • domain assumption Residual error ε_r can be estimated by probe NLL D_r from a uniformly masked probe denoiser.
    Used in the plug-in criticality estimate; NLL is a proxy for reconstruction error and may not match actual top-1 error.
  • domain assumption Token roles are identifiable and role labels are accurate from the mSENT tokenizer.
    The schedule depends on reliable role labels; ambiguity or noise in role assignment would misallocate the budget.
  • domain assumption Fixed budget ρ̄ is equal to uniform MDLM average exposure and schedules are matched in expectation.
    Restricts the comparison to reallocation rather than adding noise; used in Theorem 1.

how reviews work

0 comments
Cite this review

Pith. "Pith review of MotifRole-Diff: Risk-Optimal Role-Aware Corruption for Masked Molecular Graph Diffusion." pith.science (2026). https://pith.science/paper/P3TWCTBT

@misc{pith2026260721634,
  author       = {Pith},
  title        = {Pith review of: MotifRole-Diff: Risk-Optimal Role-Aware Corruption for Masked Molecular Graph Diffusion},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/P3TWCTBT}},
  note         = {Machine review of arXiv:2607.21634}
}
read the original abstract

Masked discrete diffusion for molecular graph generation typically applies a uniform corruption schedule to all tokens in a lossless graph-to-sequence representation, implicitly treating structurally heterogeneous molecular components as equally difficult and equally important to reconstruct. However, different molecular graph token roles exhibit substantial variation in denoising difficulty and their influence on the decoded molecule, motivating role-specific corruption strategies. We introduce MotifRole-Diff, a role-aware corruption process that allocates masking rates according to empirically measured denoising difficulty and graph-level perturbation impact while preserving the model architecture, clean sequence space, and lossless molecular-graph decoder. We formulate schedule selection as the risk-optimal allocation of a fixed masking budget across token roles. Our theorem characterizes optimality for the modeled role-weighted residual risk, while downstream generation performance is evaluated empirically. Under matched architecture, training budget, and sampling compute, MotifRole-Diff improves validity on QM9 from 0.905 to 0.944 while reducing FCD from 1.701 to 1.609, and on MOSES improves validity from 0.920 to 0.938 while reducing FCD from 2.125 to 1.850. Role-wise diagnostics further show improved reconstruction across molecular graph token categories. Together, these matched-compute results indicate that structurally informed corruption is a more effective masking strategy than uniform schedules for serialized molecular graph diffusion.

Figures

Figures reproduced from arXiv: 2607.21634 by the authors.

Figure 1
Figure 1. Measured difficulty and the derived schedule, on [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Multi-seed robustness on QM9, MOSES, and GuacaMol (5 seeds, 5000 samples/seed; whiskers are 95% confidence intervals). Validity, uniqueness, and novelty (left axis, higher is better) and FCD (right axis, lower is better) for uniform MDLM (grey) vs. MotifRole-Diff (amber). MotifRole-Diff improves validity on all three datasets and FCD on QM9 and MOSES, with tight, largely non-overlapping intervals; the novelty dip on… view at source ↗
Figure 3
Figure 3. Motif-aware serialization (mSENT) improves [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

76 extracted references · 4 linked inside Pith

  1. [1]

    Advances in Neural Information Processing Systems , volume =

    Simple and Effective Masked Diffusion Language Models , author =. Advances in Neural Information Processing Systems , volume =

  2. [2]

    Advances in Neural Information Processing Systems , volume =

    Structured Denoising Diffusion Models in Discrete State-Spaces , author =. Advances in Neural Information Processing Systems , volume =

  3. [3]

    Advances in Neural Information Processing Systems , volume =

    A Continuous Time Framework for Discrete Denoising Models , author =. Advances in Neural Information Processing Systems , volume =

  4. [4]

    International Conference on Learning Representations , year =

    Vignac, Cl. International Conference on Learning Representations , year =

  5. [5]

    Proceedings of the 39th International Conference on Machine Learning , series =

    Score-based Generative Modeling of Graphs via the System of Stochastic Differential Equations , author =. Proceedings of the 39th International Conference on Machine Learning , series =

  6. [6]

    Proceedings of the 40th International Conference on Machine Learning , series =

    Efficient and Degree-Guided Graph Generation via Discrete Diffusion Modeling , author =. Proceedings of the 40th International Conference on Machine Learning , series =

  7. [7]

    Proceedings of the 40th International Conference on Machine Learning , series =

    Autoregressive Diffusion Model for Graph Generation , author =. Proceedings of the 40th International Conference on Machine Learning , series =

  8. [8]

    Proceedings of the 41st International Conference on Machine Learning , series =

    Hyperbolic Geometric Latent Diffusion Model for Graph Generation , author =. Proceedings of the 41st International Conference on Machine Learning , series =

Show all 76 references
  1. [9]

    International Conference on Learning Representations , year =

    A Simple and Scalable Representation for Graph Generation , author =. International Conference on Learning Representations , year =

  2. [10]

    Proceedings of the 40th International Conference on Machine Learning , series =

    Improving Graph Generation by Restricting Graph Bandwidth , author =. Proceedings of the 40th International Conference on Machine Learning , series =

  3. [11]

    You, Jiaxuan and Ying, Rex and Ren, Xiang and Hamilton, William and Leskovec, Jure , booktitle =

  4. [12]

    Proceedings of the 35th International Conference on Machine Learning , series =

    Junction Tree Variational Autoencoder for Molecular Graph Generation , author =. Proceedings of the 35th International Conference on Machine Learning , series =

  5. [13]

    Proceedings of the 37th International Conference on Machine Learning , series =

    Hierarchical Generation of Molecular Graphs Using Structural Motifs , author =. Proceedings of the 37th International Conference on Machine Learning , series =

  6. [14]

    Advances in Neural Information Processing Systems , volume=

    Flatten graphs as sequences: Transformers are scalable graph generators , author=. Advances in Neural Information Processing Systems , volume=

  7. [15]

    Scientific data , volume=

    Quantum chemistry structures and properties of 134 kilo molecules , author=. Scientific data , volume=. 2014 , publisher=

    2014

  8. [16]

    Frontiers in pharmacology , volume=

    Molecular sets (MOSES): a benchmarking platform for molecular generation models , author=. Frontiers in pharmacology , volume=. 2020 , publisher=

    2020

  9. [17]

    Journal of chemical information and modeling , volume=

    GuacaMol: benchmarking models for de novo molecular design , author=. Journal of chemical information and modeling , volume=. 2019 , publisher=

    2019

  10. [18]

    Advances in neural information processing systems , volume=

    Denoising diffusion probabilistic models , author=. Advances in neural information processing systems , volume=

  11. [20]

    Advances in neural information processing systems , volume=

    Argmax flows and multinomial diffusion: Learning categorical distributions , author=. Advances in neural information processing systems , volume=

  12. [21]

    Advances in neural information processing systems , volume=

    Structured denoising diffusion models in discrete state-spaces , author=. Advances in neural information processing systems , volume=

  13. [23]

    Advances in neural information processing systems , volume=

    Simplified and generalized masked diffusion for discrete data , author=. Advances in neural information processing systems , volume=

  14. [24]

    Bert: Pre-training of deep bidirectional transformers for language understanding , author=. Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language technologies, volume 1 (long and short papers) , pages=

    2019

  15. [25]

    Mask-predict: Parallel decoding of conditional masked language models , author=. Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natural language processing (EMNLP-IJCNLP) , pages=

    2019

  16. [26]

    Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

    Maskgit: Masked generative image transformer , author=. Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , pages=

  17. [27]

    International conference on artificial intelligence and statistics , pages=

    Permutation invariant graph generation via score-based generative modeling , author=. International conference on artificial intelligence and statistics , pages=. 2020 , organization=

    2020

  18. [28]

    International conference on machine learning , pages=

    Score-based generative modeling of graphs via the system of stochastic differential equations , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  19. [29]

    International conference on machine learning , pages=

    Equivariant diffusion for molecule generation in 3d , author=. International conference on machine learning , pages=. 2022 , organization=

    2022

  20. [30]

    Joint European Conference on Machine Learning and Knowledge Discovery in Databases , pages=

    Midi: Mixed graph and 3d denoising diffusion for molecule generation , author=. Joint European Conference on Machine Learning and Knowledge Discovery in Databases , pages=. 2023 , organization=

    2023

  21. [31]

    Advances in neural information processing systems , volume=

    Graph convolutional policy network for goal-directed molecular graph generation , author=. Advances in neural information processing systems , volume=

  22. [33]

    SMILES, a chemical language and information system. 1. Introduction to methodology and encoding rules , author=. Journal of chemical information and computer sciences , volume=. 1988 , publisher=

    1988

  23. [34]

    Machine Learning: Science and Technology , volume=

    Self-referencing embedded strings (SELFIES): A 100\ author=. Machine Learning: Science and Technology , volume=. 2020 , publisher=

    2020

  24. [35]

    International conference on machine learning , pages=

    Junction tree variational autoencoder for molecular graph generation , author=. International conference on machine learning , pages=. 2018 , organization=

    2018

  25. [36]

    ACS central science , volume=

    Automatic chemical design using a data-driven continuous representation of molecules , author=. ACS central science , volume=. 2018 , publisher=

    2018

  26. [37]

    Advances in neural information processing systems , volume=

    Variational diffusion models , author=. Advances in neural information processing systems , volume=

  27. [38]

    International conference on machine learning , pages=

    Improved denoising diffusion probabilistic models , author=. International conference on machine learning , pages=. 2021 , organization=

    2021

  28. [40]

    Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

    Diffusionbert: Improving generative masked language models with diffusion models , author=. Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , pages=

  29. [41]

    Nature chemistry , volume=

    Quantifying the chemical beauty of drugs , author=. Nature chemistry , volume=. 2012 , publisher=

    2012

  30. [42]

    Preuer, Kristina and Renz, Philipp and Unterthiner, Thomas and Hochreiter, Sepp and Klambauer, Gunter , journal=. Fr. 2018 , publisher=

    2018

  31. [43]

    D.; Ho, J.; Tarlow, D.; and van den Berg, R

    Austin, J.; Johnson, D. D.; Ho, J.; Tarlow, D.; and van den Berg, R. 2021. Structured denoising diffusion models in discrete state-spaces. In Advances in Neural Information Processing Systems , volume 34, 17981--17993

    2021

  32. [44]

    R.; Paolini, G

    Bickerton, G. R.; Paolini, G. V.; Besnard, J.; Muresan, S.; and Hopkins, A. L. 2012. Quantifying the chemical beauty of drugs. Nature chemistry 4(2):90--98

    2012

  33. [45]

    H.; and Vaucher, A

    Brown, N.; Fiscato, M.; Segler, M. H.; and Vaucher, A. C. 2019. Guacamol: benchmarking models for de novo molecular design. Journal of chemical information and modeling 59(3):1096--1108

    2019

  34. [46]

    Campbell, A.; Benton, J.; De Bortoli, V.; Rainforth, T.; Deligiannidis, G.; and Doucet, A. 2022. A continuous time framework for discrete denoising models. In Advances in Neural Information Processing Systems , volume 35

    2022

  35. [47]

    Chang, H.; Zhang, H.; Jiang, L.; Liu, C.; and Freeman, W. T. 2022. Maskgit: Masked generative image transformer. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 11315--11325

    2022

  36. [48]

    Chen, X.; He, J.; Han, X.; and Liu, L.-P. 2023. Efficient and degree-guided graph generation via discrete diffusion modeling. In Proceedings of the 40th International Conference on Machine Learning , volume 202 of Proceedings of Machine Learning Research

    2023

  37. [49]

    Chen, D.; Krimmel, M.; and Borgwardt, K. 2026. Flatten graphs as sequences: Transformers are scalable graph generators. Advances in Neural Information Processing Systems 38:69071--69109

    2026

  38. [50]

    Devlin, J.; Chang, M.-W.; Lee, K.; and Toutanova, K. 2019. Bert: Pre-training of deep bidirectional transformers for language understanding. In Proceedings of the 2019 conference of the North American chapter of the association for computational linguistics: human language tec...

    2019

  39. [51]

    M.; Hase, P.; Bansal, M.; and Biancalani, T

    Diamant, N.; Tseng, A. M.; Hase, P.; Bansal, M.; and Biancalani, T. 2023. Improving graph generation by restricting graph bandwidth. In Proceedings of the 40th International Conference on Machine Learning , volume 202 of Proceedings of Machine Learning Research

    2023

  40. [52]

    H.; Doucet, A.; Strudel, R.; Dyer, C.; Durkan, C.; et al

    Dieleman, S.; Sartran, L.; Roshannai, A.; Savinov, N.; Ganin, Y.; Richemond, P. H.; Doucet, A.; Strudel, R.; Dyer, C.; Durkan, C.; et al. 2022. Continuous diffusion for categorical data. arXiv preprint arXiv:2211.15089

    2022 arXiv

  41. [53]

    Ghazvininejad, M.; Levy, O.; Liu, Y.; and Zettlemoyer, L. 2019. Mask-predict: Parallel decoding of conditional masked language models. In Proceedings of the 2019 conference on empirical methods in natural language processing and the 9th international joint conference on natura...

    2019

  42. [54]

    N.; Duvenaud, D.; Hern \'a ndez-Lobato, J

    G \'o mez-Bombarelli, R.; Wei, J. N.; Duvenaud, D.; Hern \'a ndez-Lobato, J. M.; S \'a nchez-Lengeling, B.; Sheberla, D.; Aguilera-Iparraguirre, J.; Hirzel, T. D.; Adams, R. P.; and Aspuru-Guzik, A. 2018. Automatic chemical design using a data-driven continuous representation ...

    2018

  43. [55]

    He, Z.; Sun, T.; Tang, Q.; Wang, K.; Huang, X.-J.; and Qiu, X. 2023. Diffusionbert: Improving generative masked language models with diffusion models. In Proceedings of the 61st annual meeting of the association for computational linguistics (volume 1: Long papers) , 4521--4534

    2023

  44. [56]

    Ho, J.; Jain, A.; and Abbeel, P. 2020. Denoising diffusion probabilistic models. Advances in neural information processing systems 33:6840--6851

    2020

  45. [57]

    Hoogeboom, E.; Nielsen, D.; Jaini, P.; Forr \'e , P.; and Welling, M. 2021. Argmax flows and multinomial diffusion: Learning categorical distributions. Advances in neural information processing systems 34:12454--12465

    2021

  46. [58]

    G.; Vignac, C.; and Welling, M

    Hoogeboom, E.; Satorras, V. G.; Vignac, C.; and Welling, M. 2022. Equivariant diffusion for molecule generation in 3d. In International conference on machine learning , 8867--8887. PMLR

    2022

  47. [59]

    Jang, Y.; Lee, S.; and Ahn, S. 2024. A simple and scalable representation for graph generation. In International Conference on Learning Representations

    2024

  48. [60]

    Jin, W.; Barzilay, R.; and Jaakkola, T. 2018a. Junction tree variational autoencoder for molecular graph generation. In Proceedings of the 35th International Conference on Machine Learning , volume 80 of Proceedings of Machine Learning Research

  49. [61]

    Jin, W.; Barzilay, R.; and Jaakkola, T. 2018b. Junction tree variational autoencoder for molecular graph generation. In International conference on machine learning , 2323--2332. PMLR

  50. [62]

    Jin, W.; Barzilay, R.; and Jaakkola, T. 2020. Hierarchical generation of molecular graphs using structural motifs. In Proceedings of the 37th International Conference on Machine Learning , volume 119 of Proceedings of Machine Learning Research , 4839--4848

    2020

  51. [63]

    Jo, J.; Lee, S.; and Hwang, S. J. 2022a. Score-based generative modeling of graphs via the system of stochastic differential equations. In Proceedings of the 39th International Conference on Machine Learning , volume 162 of Proceedings of Machine Learning Research

  52. [64]

    Jo, J.; Lee, S.; and Hwang, S. J. 2022b. Score-based generative modeling of graphs via the system of stochastic differential equations. In International conference on machine learning , 10362--10383. PMLR

  53. [65]

    Kingma, D.; Salimans, T.; Poole, B.; and Ho, J. 2021. Variational diffusion models. Advances in neural information processing systems 34:21696--21707

    2021

  54. [66]

    Krenn, M.; H \"a se, F.; Nigam, A.; Friederich, P.; and Aspuru-Guzik, A. 2020. Self-referencing embedded strings (selfies): A 100\ Machine Learning: Science and Technology 1(4):045024

    2020

  55. [67]

    Lou, A.; Meng, C.; and Ermon, S. 2023. Discrete diffusion modeling by estimating the ratios of the data distribution. arXiv preprint arXiv:2310.16834

    2023 arXiv

  56. [68]

    Q., and Dhariwal, P

    Nichol, A. Q., and Dhariwal, P. 2021. Improved denoising diffusion probabilistic models. In International conference on machine learning , 8162--8171. PMLR

    2021

  57. [69]

    Niu, C.; Song, Y.; Song, J.; Zhao, S.; Grover, A.; and Ermon, S. 2020. Permutation invariant graph generation via score-based generative modeling. In International conference on artificial intelligence and statistics , 4474--4484. PMLR

    2020

  58. [70]

    Polykovskiy, D.; Zhebrak, A.; Sanchez-Lengeling, B.; Golovanov, S.; Tatanov, O.; Belyaev, S.; Kurbanov, R.; Artamonov, A.; Aladinskiy, V.; Veselov, M.; et al. 2020. Molecular sets (moses): a benchmarking platform for molecular generation models. Frontiers in pharmacology 11:565644

    2020

  59. [71]

    Preuer, K.; Renz, P.; Unterthiner, T.; Hochreiter, S.; and Klambauer, G. 2018. Fr \'e chet chemnet distance: a metric for generative models for molecules in drug discovery. Journal of chemical information and modeling 58(9):1736--1741

    2018

  60. [72]

    O.; Rupp, M.; and Von Lilienfeld, O

    Ramakrishnan, R.; Dral, P. O.; Rupp, M.; and Von Lilienfeld, O. A. 2014. Quantum chemistry structures and properties of 134 kilo molecules. Scientific data 1(1):1--7

    2014

  61. [73]

    S.; Arriola, M.; Schiff, Y.; Gokaslan, A.; Marroquin, E.; Chiu, J

    Sahoo, S. S.; Arriola, M.; Schiff, Y.; Gokaslan, A.; Marroquin, E.; Chiu, J. T.; Rush, A. M.; and Kuleshov, V. 2024. Simple and effective masked diffusion language models. In Advances in Neural Information Processing Systems , volume 37

    2024

  62. [74]

    Shi, C.; Xu, M.; Zhu, Z.; Zhang, W.; Zhang, M.; and Tang, J. 2020. Graphaf: a flow-based autoregressive model for molecular graph generation. arXiv preprint arXiv:2001.09382

    2020 arXiv

  63. [75]

    Shi, J.; Han, K.; Wang, Z.; Doucet, A.; and Titsias, M. 2024. Simplified and generalized masked diffusion for discrete data. Advances in neural information processing systems 37:103131--103167

    2024

  64. [76]

    P.; Kumar, A.; Ermon, S.; and Poole, B

    Song, Y.; Sohl-Dickstein, J.; Kingma, D. P.; Kumar, A.; Ermon, S.; and Poole, B. 2020. Score-based generative modeling through stochastic differential equations. arXiv preprint arXiv:2011.13456

    2020 arXiv

  65. [77]

    Vignac, C.; Krawczuk, I.; Siraudin, A.; Wang, B.; Cevher, V.; and Frossard, P. 2023a. DiGress : Discrete denoising diffusion for graph generation. In International Conference on Learning Representations

  66. [78]

    Vignac, C.; Osman, N.; Toni, L.; and Frossard, P. 2023b. Midi: Mixed graph and 3d denoising diffusion for molecule generation. In Joint European Conference on Machine Learning and Knowledge Discovery in Databases , 560--576. Springer

  67. [79]

    Weininger, D. 1988. Smiles, a chemical language and information system. 1. introduction to methodology and encoding rules. Journal of chemical information and computer sciences 28(1):31--36

    1988

  68. [80]

    You, J.; Liu, B.; Ying, Z.; Pande, V.; and Leskovec, J. 2018. Graph convolutional policy network for goal-directed molecular graph generation. Advances in neural information processing systems 31

    2018

Pith tools

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