Pith. sign in

REVIEW 4 minor 91 references

In flow matching, the noise–data coupling itself can embed property control into the learned flow, so a single noise scalar steers molecular rewards at inference with no oracle or guidance.

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 ↗

Property-aligned optimal-transport coupling in flow matching embeds a noise-space scalar that steers logP and QED distributions at inference with no oracle, reward model, or guidance.

T0 review reviewed 2026-07-13 challenge →

load-bearing objection Clean, scoped idea: property-sorted OT coupling turns a noise scalar into a free distributional control knob for flow-matching molecules, with honest residuals and opposite-size evidence that rules out size bias.

arxiv 2607.08781 v1 pith:U7PPIGLV submitted 2026-06-13 cs.LG cs.AIq-bio.QM

Reward Transport: Property Control in Flow Matching via Noise-Space Alignment

Kehan Guo , Yili Shen , Yujun Zhou , Yue Huang , Chujie Gao , Shiyi Du , Xiangliang Zhang This is my paper
classification cs.LG cs.AIq-bio.QM
keywords flow matchingoptimal transport couplingproperty-controlled generationmolecular generationCross-Entropy Methodnoise-space alignmentSELFIESreward transport
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

Flow matching usually treats the rule that pairs noise vectors with data points as a training detail. This paper argues that the pairing can instead be an alignment interface: sort noise by a scalar coordinate and molecules by a target property, then pair them by rank. That monotone coupling writes property structure into the learned flow field. At generation time, choosing the scalar selects a truncated property distribution—recovering one Cross-Entropy Method selection step when rank order is preserved—without any reward model, gradient guidance, or extra compute. On ZINC-250K and GuacaMol the same knob monotonically steers logP and QED, and it grows molecules for logP while shrinking them for QED, showing the structure is property-specific rather than a generic size bias. The result is a distribution-level control channel that is complementary to ordinary conditioning.

Core claim

Property-aligned monotone optimal-transport coupling embeds controllable structure into a flow-matching field so that, at inference, varying a single noise-space scalar steers the generated property distribution with no oracle, reward model, gradient guidance, or added computation. In the coupling-preserving limit, thresholding that scalar recovers the Cross-Entropy Method’s truncated reward distribution.

What carries the argument

Reward Transport: the 1-D monotone rearrangement that sorts noise by a scalar key s(z) and molecules by property y, then pairs them rank-by-rank; Proposition 1 shows that, when the learned flow preserves ranks, conditioning on s ≥ τ yields the data distribution truncated to the corresponding property quantile.

Load-bearing premise

The trained flow must keep enough of the training-time rank order between the noise scalar and the property; if that order collapses, the scalar knob loses its control.

What would settle it

Retrain under the same recipe and measure per-molecule rank correlation between the noise scalar and generated property: if that correlation stays near zero while group means still appear monotone, or if the same scalar produces identical structural responses for chemically opposite targets, the claimed coupling-to-flow transfer fails.

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

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

0 major / 4 minor

Summary. The paper reframes the noise–data coupling in flow matching as an alignment interface rather than a training heuristic. Reward Transport builds a property-aligned monotone OT coupling by sorting a scalar noise key s(z)=∥z̄∥₂ against a molecular property y and pairing rank-by-rank; a Direction MLP injects s into a Pre-LayerNorm Transformer. At inference, sweeping or thresholding s steers the generated property distribution with no oracle, reward model, or guidance. Proposition 1 shows that, in the coupling-preserving limit, this recovers one Cross-Entropy Method selection step. On ZINC-250K and GuacaMol the method yields group-mean ρ=1.000 for logP and consistent QED control, with opposite atom-count responses that rule out generic size bias; ablations, a sorting-key study, a same-backbone Conditional FM comparison, and an ε-prediction negative result with SNR analysis delimit the interface.

Significance. If the claims hold, the work opens a distinct, distribution-level control channel for flow matching that is complementary to classifier-free guidance and conditional generation and that incurs zero inference overhead. The opposite structural programs for logP versus QED, the component ablations (OT, DirEmb, unmasked MSE), the sorting-key robustness, the GuacaMol replication, and the analytic attenuation of coupling gradients under ε-prediction are concrete, falsifiable contributions. Public code further strengthens reproducibility. The scoped claim—distributional rather than pointwise control under x̂₁/velocity prediction—is carefully stated and of clear interest to molecular generation and generative modeling more broadly.

minor comments (4)
  1. In §4.1 and Proposition 1 the realized per-molecule ρ_per (0.57 logP, 0.22 QED) is correctly treated as residual slack, but a short explicit sentence in the main text quantifying how far the distributional approximation sits from the ideal CEM truncation would help readers who only skim the proof appendix.
  2. Table 1 and the GuacaMol tables report validity/uniqueness at 100%/≥99%; the extended FCD/SA audit in Appendix M is valuable but could be cross-referenced more prominently in §5.2 so that the quality–steerability trade-off is visible without leaving the main narrative.
  3. Notation for the sorting key alternates between s(z), ∥z̄∥₂ and the normalized ŝ; a single consistent definition early in §4 would reduce minor ambiguity when reading Algorithm 1 and the inference paragraph together.
  4. Appendix J’s SNR derivation is clear; a one-line pointer in §6 to the explicit attenuation factor E_t[α_t] would make the architectural boundary easier to cite without opening the appendix.

Circularity Check

0 steps flagged

No significant circularity: Proposition 1 is classical monotone rearrangement plus CEM truncation under an explicit coupling-preserving limit; empirical control is measured against external RDKit properties.

full rationale

The central theoretical claim (Proposition 1 / Appendix A) states that, under the monotone OT coupling of a 1-D noise key s(z) with property y and in the coupling-preserving limit rho_per -> 1, conditioning on s >= tau recovers the data distribution truncated to the corresponding upper property quantile—the functional form of one CEM selection step. This follows directly from the classical monotone-rearrangement theorem (Villani 2003) and the definition of CEM truncation (Rubinstein 1999); neither is fitted to the paper’s data, and the paper itself reports imperfect realized rho_per (0.57 logP, 0.22 QED) so the claim is distributional rather than pointwise. Empirical deltas, group-mean rho = 1.000, opposite size programs, GuacaMol replication, and ablations (OT / Dir / unmask) are measured on held-out generations scored by external RDKit oracles, not by quantities that were fitted into the coupling. The Direction-MLP and unmasked-MSE fixes are architectural necessities discovered by ablation, not redefinitions of the target. The epsilon-prediction negative result and SNR derivation further delimit scope rather than close a self-referential loop. No self-definitional step, fitted-input-called-prediction, load-bearing self-citation uniqueness theorem, or renaming of a known result is present. Score 1 reflects only the mild, non-load-bearing fact that the paper cites standard OT/CEM literature that any such construction must cite.

Axiom & Free-Parameter Ledger

5 free parameters · 4 axioms · 1 invented entities

The central claim rests on standard optimal-transport and flow-matching machinery plus a small set of architectural and training choices needed to keep the 1-D coupling signal alive inside a Pre-LN Transformer that generates variable-length discrete sequences. No new physical entities are postulated; the free parameters are ordinary ML hyper-parameters whose values are reported.

free parameters (5)
  • CE loss weight lambda = 1
    Fixed at 1 in the combined MSE+CE objective; balances continuous flow signal against discrete token decoding.
  • Direction-MLP size and learning-rate multiplier = ~591k params, 10x LR on strong base
    Adds ~591k parameters (~1%); strong-base runs use 10x LR on the MLP. Chosen to inject s after Pre-LN erasure.
  • Two-stage fine-tune length (5-10 epochs) = 5-10 epochs
    Warm-start from 120-epoch base then short OT fine-tune; from-scratch training shows a transient signal window that collapses.
  • s normalization (mu_s, sigma_s) = dataset-dependent
    Training-set statistics used to standardize the direction signal before the MLP.
  • Euler steps at inference = 50
    Fixed at 50 for all reported sweeps; affects residual rank preservation.
axioms (4)
  • standard math 1-D monotone rearrangement is the unique OT plan under any convex cost (Villani).
    Invoked to justify rank-by-rank pairing of s(z) with y(x) as optimal transport (Eq. 4, Prop. 1).
  • domain assumption Flow matching with linear interpolant and x1- or velocity prediction yields a Bayes-optimal target that depends on the coupled data point.
    Required for the coupling-induced gradient to reach the Direction MLP; contrasted analytically with epsilon-prediction (Appendix J).
  • domain assumption SELFIES tokenization plus RDKit property oracles correctly measure validity and the target scalars logP/QED.
    All validity and property numbers rest on these external tools.
  • ad hoc to paper Unmasked MSE on zero PAD embeddings anchors global noise magnitude and blocks the length shortcut.
    Discovered empirically; necessary for the scalar s to control property rather than sequence length (Sec. 5.3).
invented entities (1)
  • Reward Transport coupling (property-sorted 1-D OT + DirEmb) independent evidence
    purpose: Turns the training coupling into a distributional control interface so that a noise scalar steers generated property distributions.
    The procedure is new; independent evidence is the reported rho/delta sweeps and opposite structural responses, which are falsifiable by re-running the public code.

reviewed 2026-07-13 · how reviews work

0 comments
Cite this review

Pith. "Pith review of Reward Transport: Property Control in Flow Matching via Noise-Space Alignment." pith.science (2026). https://pith.science/paper/U7PPIGLV

@misc{pith2026260708781,
  author       = {Pith},
  title        = {Pith review of: Reward Transport: Property Control in Flow Matching via Noise-Space Alignment},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/U7PPIGLV}},
  note         = {Machine review of arXiv:2607.08781}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

The coupling in flow matching -- the rule pairing noise vectors with data points -- is typically treated as a computational choice. We show that this coupling can instead serve as an alignment interface: by matching noise and data according to a target molecular property, it embeds controllable structure directly into the learned flow field. Building on this view, we introduce Reward Transport, which uses optimal transport coupling at training time to align a scalar noise-space coordinate with molecular rewards; at inference, varying this coordinate steers the generated distribution without requiring an oracle, reward model, gradient guidance, or additional computation. In the coupling-preserving limit, thresholding this coordinate recovers the Cross-Entropy Method's truncated reward distribution, providing a principled, continuously adjustable distribution-level control knob. Empirically, on ZINC-250K and GuacaMol, sweeping the scalar induces monotone control of logP and consistent QED control over its operating range; most tellingly, the same knob produces opposite structural responses for different targets, growing molecules for logP but shrinking them for QED, which rules out a generic size bias. The interface is complementary to classifier-free guidance and conditional flow matching, while a negative result under epsilon-prediction diffusion clarifies where coupling-level alignment is structurally absent. Code: https://github.com/KehanGuo2/reward-transport

Figures

Figures reproduced from arXiv: 2607.08781 by Chujie Gao, Kehan Guo, Shiyi Du, Xiangliang Zhang, Yili Shen, Yue Huang, Yujun Zhou.

Figure 1
Figure 1. Figure 1: Reward Transport overview. (A) Standard flow matching pairs noise and data indepen￾dently, producing no property structure in noise space. (B) Property-aligned OT coupling sorts noise vectors by norm and molecules by target property, then pairs them rank-by-rank, creating a monotone mapping between noise coordinates and molecular properties. (C) At inference, sampling from the upper tail of the noise distr… view at source ↗
Figure 2
Figure 2. Figure 2: Rank preservation emerges through the flow. UMAP of xt for the logP model (n=1,400); top row coloured by s, bottom by actual logP. Structure is absent at t=0 and aligned by t=1 (per￾molecule ρ=0.57, p<10−120); a no-OT base model shows no alignment (Appendix K). 5 [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Property control and distribution shift. Left: Mean property vs. s for logP and QED (n=1000 per s). Gray dashed: atom count. Both achieve ρ=1.000 with opposite size responses. Right: Full logP distributions across seven s values, shifting continuously rightward. Dashed line: ZINC mean (logP=2.46). QED distributions in Appendix B. Does the coupling transfer to the flow field? [PITH_FULL_IMAGE:figures/full_… view at source ↗
Figure 4
Figure 4. Figure 4: Rank preservation emerges within the first OT fine-tuning epoch. Per￾molecule ρ(s, y) across training checkpoints; ρ≈0 at base, lifted within one epoch, plateau <1 is the distributional gap Prop. 1 de￾scribes. Rank preservation as a training dynamic. Propo￾sition 1 describes the coupling-preserving limit ρper → 1; [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: QED distribution shift. Generated QED distributions across all seven s values (n=1000 per s). Distributions shift rightward with increasing s, though the effect is smaller than for logP. Dashed line: ZINC training mean (QED=0.728). Standard configuration [PITH_FULL_IMAGE:figures/full_fig_p015_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Rank preservation: QED model. UMAP projections of xt at t=0 (top) and t=1 (bottom), colored by s (left) and by actual QED (right). At t=0, no structure is present. At t=1, both gradients align (per-molecule ρ(s, QED)=0.22), though with more overlap than the logP model ( [PITH_FULL_IMAGE:figures/full_fig_p016_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Negative control. Top: logP model. Bottom: QED model. Each group shows base model (top row) vs. Reward Transport (bottom row), colored by s (left), actual property (center), and shuffled s (right). The base model shows no s-structure (ρ=0.047 for logP, 0.032 for QED). The Reward Transport model shows aligned gradients (ρ=0.570 / 0.220), which disappear under label permutation. L Same-Backbone Comparison ag… view at source ↗
Figure 8
Figure 8. Figure 8: Failure modes visualised. See text above. 28 [PITH_FULL_IMAGE:figures/full_fig_p028_8.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

91 extracted references · 4 linked inside Pith

  1. [2]

    Improving and generalizing flow-based generative models with minibatch optimal transport , author=. Trans. Mach. Learn. Res. , year=

  2. [3]

    Methodology and Computing in Applied Probability , volume=

    The Cross-Entropy Method for Combinatorial and Continuous Optimization , author=. Methodology and Computing in Applied Probability , volume=

  3. [4]

    2003 , publisher=

    Topics in Optimal Transport , author=. 2003 , publisher=

    2003

  4. [5]

    Irwin, John J and Sterling, Teague and Mysinger, Michael M and Bolstad, Erin S and Coleman, Ryan G , journal=

  5. [6]

    Machine Learning: Science and Technology , volume=

    Krenn, Mario and H. Machine Learning: Science and Technology , volume=

  6. [7]

    Journal of Cheminformatics , volume=

    Estimation of Synthetic Accessibility Score of Drug-like Molecules Based on Molecular Complexity and Fragment Contributions , author=. Journal of Cheminformatics , volume=

  7. [8]

    Journal of Chemical Information and Modeling , volume=

    Blaschke, Thomas and Ar. Journal of Chemical Information and Modeling , volume=

  8. [9]

    Advances in Neural Information Processing Systems , year=

    Graph Convolutional Policy Network for Goal-Directed Molecular Graph Generation , author=. Advances in Neural Information Processing Systems , year=

  9. [11]

    Zeng, Cheng and others , journal=

  10. [12]

    International Conference on Machine Learning , year=

    Exploring Chemical Space with Score-based Out-of-distribution Generation , author=. International Conference on Machine Learning , year=

  11. [15]

    International Conference on Learning Representations , year=

    Flow Straight and Fast: Learning to Generate and Transfer Data with Rectified Flow , author=. International Conference on Learning Representations , year=

  12. [16]

    International Conference on Learning Representations , year=

    Building Normalizing Flows with Stochastic Interpolants , author=. International Conference on Learning Representations , year=

  13. [17]

    International Conference on Learning Representations , year=

    Score-Based Generative Modeling through Stochastic Differential Equations , author=. International Conference on Learning Representations , year=

  14. [18]

    Advances in Neural Information Processing Systems , volume=

    Denoising Diffusion Probabilistic Models , author=. Advances in Neural Information Processing Systems , volume=

  15. [19]

    Equivariant Diffusion for Molecule Generation in

    Hoogeboom, Emiel and Satorras, V. Equivariant Diffusion for Molecule Generation in. International Conference on Machine Learning , volume=

  16. [20]

    Scientific Reports , volume=

    Optimization of Molecules via Deep Reinforcement Learning , author=. Scientific Reports , volume=

  17. [21]

    and Tiwari, Mo and Bengio, Emmanuel , journal=

    Bengio, Yoshua and Lahlou, Salem and Deleu, Tristan and Hu, Edward J. and Tiwari, Mo and Bengio, Emmanuel , journal=

  18. [22]

    Advances in Neural Information Processing Systems , volume=

    Torsional Diffusion for Molecular Conformer Generation , author=. Advances in Neural Information Processing Systems , volume=

  19. [23]

    Brown, Nathan and Fiscato, Marco and Segler, Marwin H. S. and Vaucher, Alain C. , journal=

  20. [24]

    Foundations and Trends in Machine Learning , volume=

    Computational Optimal Transport , author=. Foundations and Trends in Machine Learning , volume=

  21. [25]

    Advances in Neural Information Processing Systems , year=

    Elucidating the Design Space of Diffusion-Based Generative Models , author=. Advances in Neural Information Processing Systems , year=

  22. [26]

    Advances in Neural Information Processing Systems , year=

    Generative Modeling by Estimating Gradients of the Data Distribution , author=. Advances in Neural Information Processing Systems , year=

  23. [27]

    Advances in Neural Information Processing Systems , year=

    Neural Ordinary Differential Equations , author=. Advances in Neural Information Processing Systems , year=

  24. [28]

    International Conference on Machine Learning , year=

    Deep Unsupervised Learning Using Nonequilibrium Thermodynamics , author=. International Conference on Machine Learning , year=

  25. [29]

    Advances in Neural Information Processing Systems , year=

    Sinkhorn Distances: Lightspeed Computation of Optimal Transport , author=. Advances in Neural Information Processing Systems , year=

  26. [30]

    Diffusion

    De Bortoli, Valentin and Thornton, James and Heng, Jeremy and Doucet, Arnaud , booktitle=. Diffusion

  27. [31]

    International Conference on Machine Learning , year=

    Multisample Flow Matching: Straightening Flows with Minibatch Couplings , author=. International Conference on Machine Learning , year=

  28. [32]

    International Conference on Machine Learning , year=

    Discrete Diffusion Modeling by Estimating the Ratios of the Data Distribution , author=. International Conference on Machine Learning , year=

  29. [33]

    Advances in Neural Information Processing Systems , year=

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

  30. [34]

    , booktitle=

    Li, Xiang Lisa and Thickstun, John and Gulrajani, Ishaan and Liang, Percy and Hashimoto, Tatsunori B. , booktitle=

  31. [35]

    International Conference on Machine Learning , year=

    Junction Tree Variational Autoencoder for Molecular Graph Generation , author=. International Conference on Machine Learning , year=

  32. [36]

    Shi, Chence and Xu, Minkai and Zhu, Zhaocheng and Zhang, Weinan and Zhang, Ming and Tang, Jian , booktitle=

  33. [37]

    Zang, Chengxi and Wang, Fei , journal=

  34. [38]

    ACS Central Science , volume=

    Generating Focused Molecule Libraries for Drug Discovery with Recurrent Neural Networks , author=. ACS Central Science , volume=

  35. [39]

    Vignac, Clement and Krawczuk, Igor and Siraudin, Antoine and Wang, Bohan and Cevher, Volkan and Frossard, Pascal , booktitle=

  36. [40]

    Advances in Neural Information Processing Systems , year=

    Training Language Models to Follow Instructions with Human Feedback , author=. Advances in Neural Information Processing Systems , year=

  37. [41]

    Advances in Neural Information Processing Systems , year=

    Direct Preference Optimization: Your Language Model Is Secretly a Reward Model , author=. Advances in Neural Information Processing Systems , year=

  38. [42]

    Weininger, David , journal=

  39. [43]

    Journal of Chemical Information and Computer Sciences , volume=

    Prediction of Physicochemical Parameters by Atomic Contributions , author=. Journal of Chemical Information and Computer Sciences , volume=

  40. [44]

    Advanced Drug Delivery Reviews , volume=

    Experimental and Computational Approaches to Estimate Solubility and Permeability in Drug Discovery and Development Settings , author=. Advanced Drug Delivery Reviews , volume=

  41. [46]

    ICML , year=

    Scaling Rectified Flow Transformers for High-Resolution Image Synthesis , author=. ICML , year=

  42. [47]

    and Boffi, Nicholas M

    Ma, Nanye and Goldstein, Mark and Albergo, Michael S. and Boffi, Nicholas M. and Vanden-Eijnden, Eric and Xie, Saining , booktitle=

  43. [48]

    ACL , year=

    Neural Machine Translation of Rare Words with Subword Units , author=. ACL , year=

  44. [49]

    Diffusion Models Beat

    Dhariwal, Prafulla and Nichol, Alex , booktitle=. Diffusion Models Beat

  45. [50]

    ICCV , year=

    Scalable Diffusion Models with Transformers , author=. ICCV , year=

  46. [51]

    ICML , year=

    On Layer Normalization in the Transformer Architecture , author=. ICML , year=

  47. [52]

    Nature chemistry , volume=

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

    2012

  48. [53]

    Albergo and Eric Vanden-Eijnden

    Michael S. Albergo and Eric Vanden-Eijnden. Building normalizing flows with stochastic interpolants. In International Conference on Learning Representations, 2023

    2023

  49. [54]

    Hu, Mo Tiwari, and Emmanuel Bengio

    Yoshua Bengio, Salem Lahlou, Tristan Deleu, Edward J. Hu, Mo Tiwari, and Emmanuel Bengio. GFlowNet foundations. Journal of Machine Learning Research, 24 0 (210): 0 1--55, 2023

    2023

  50. [55]

    Quantifying the chemical beauty of drugs

    G Richard Bickerton, Gaia V Paolini, J \'e r \'e my Besnard, Sorel Muresan, and Andrew L Hopkins. Quantifying the chemical beauty of drugs. Nature chemistry, 4 0 (2): 0 90--98, 2012

    2012

  51. [56]

    REINVENT 2.0: An AI tool for de novo drug design

    Thomas Blaschke, Josep Ar \'u s-Pous, Hongming Chen, Christian Marber, Thierry Kogej, and Ola Engkvist. REINVENT 2.0: An AI tool for de novo drug design. Journal of Chemical Information and Modeling, 60 0 (12): 0 5918--5922, 2020

    2020

  52. [57]

    Nathan Brown, Marco Fiscato, Marwin H. S. Segler, and Alain C. Vaucher. GuacaMol : Benchmarking models for de novo molecular design. Journal of Chemical Information and Modeling, 59 0 (3): 0 1096--1108, 2019

    2019

  53. [58]

    Ricky T. Q. Chen, Yulia Rubanova, Jesse Bettencourt, and David Duvenaud. Neural ordinary differential equations. In Advances in Neural Information Processing Systems, 2018

    2018

  54. [59]

    Sinkhorn distances: Lightspeed computation of optimal transport

    Marco Cuturi. Sinkhorn distances: Lightspeed computation of optimal transport. In Advances in Neural Information Processing Systems, 2013

    2013

  55. [60]

    Diffusion S chr \"o dinger bridge with applications to score-based generative modeling

    Valentin De Bortoli, James Thornton, Jeremy Heng, and Arnaud Doucet. Diffusion S chr \"o dinger bridge with applications to score-based generative modeling. In Advances in Neural Information Processing Systems, 2021

    2021

  56. [61]

    Generative modeling via drifting

    Mingyang Deng, He Li, Tianhong Li, Yilun Du, and Kaiming He. Generative modeling via drifting. arXiv preprint arXiv:2602.04770, 2026

    Pith/arXiv arXiv 2026

  57. [62]

    Diffusion models beat GAN s on image synthesis

    Prafulla Dhariwal and Alex Nichol. Diffusion models beat GAN s on image synthesis. In NeurIPS, 2021

    2021

  58. [63]

    Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions

    Peter Ertl and Ansgar Schuffenhauer. Estimation of synthetic accessibility score of drug-like molecules based on molecular complexity and fragment contributions. Journal of Cheminformatics, 1 0 (1): 0 1--11, 2009

    2009

  59. [64]

    Scaling rectified flow transformers for high-resolution image synthesis

    Patrick Esser, Sumith Kulal, Andreas Blattmann, et al. Scaling rectified flow transformers for high-resolution image synthesis. In ICML, 2024

    2024

  60. [65]

    Classifier-free diffusion guidance

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

    Pith/arXiv arXiv 2022

  61. [66]

    Denoising diffusion probabilistic models

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. In Advances in Neural Information Processing Systems, volume 33, pages 6840--6851, 2020

    2020

  62. [67]

    Equivariant diffusion for molecule generation in 3D

    Emiel Hoogeboom, V \' ctor Garcia Satorras, Cl \'e ment Vignac, and Max Welling. Equivariant diffusion for molecule generation in 3D . In International Conference on Machine Learning, volume 162, pages 8867--8887, 2022

    2022

  63. [68]

    Improving molecular graph generation with flow matching and optimal transport

    Xiaoyang Hou, Tian Zhu, Milong Ren, Dongbo Bu, Xin Gao, Chunming Zhang, and Shiwei Sun. Improving molecular graph generation with flow matching and optimal transport. arXiv preprint arXiv:2411.05676, 2024

    Pith/arXiv arXiv 2024

  64. [69]

    ZINC : A free tool to discover chemistry for biology

    John J Irwin, Teague Sterling, Michael M Mysinger, Erin S Bolstad, and Ryan G Coleman. ZINC : A free tool to discover chemistry for biology. Journal of Chemical Information and Modeling, 52 0 (7): 0 1757--1768, 2012

    2012

  65. [70]

    Jaakkola

    Wengong Jin, Regina Barzilay, and Tommi S. Jaakkola. Junction tree variational autoencoder for molecular graph generation. In International Conference on Machine Learning, 2018

    2018

  66. [71]

    Jaakkola

    Bowen Jing, Gabriele Corso, Jeffrey Chang, Regina Barzilay, and Tommi S. Jaakkola. Torsional diffusion for molecular conformer generation. In Advances in Neural Information Processing Systems, volume 35, 2022

    2022

  67. [72]

    Elucidating the design space of diffusion-based generative models

    Tero Karras, Miika Aittala, Timo Aila, and Samuli Laine. Elucidating the design space of diffusion-based generative models. In Advances in Neural Information Processing Systems, 2022

    2022

  68. [73]

    SELFIES : A robust representation of semantically constrained graphs with an example application in chemistry

    Mario Krenn, Florian H \"a se, AkshatKumar Nigam, Pascal Friederich, and Al \'a n Aspuru-Guzik. SELFIES : A robust representation of semantically constrained graphs with an example application in chemistry. Machine Learning: Science and Technology, 1 0 (4): 0 045024, 2020

    2020

  69. [74]

    Exploring chemical space with score-based out-of-distribution generation

    Seul Lee, Jaehyeong Jo, and Sung Ju Hwang. Exploring chemical space with score-based out-of-distribution generation. In International Conference on Machine Learning, 2023

    2023

  70. [75]

    Flow matching for generative modeling

    Yaron Lipman, Ricky TQ Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow matching for generative modeling. arXiv preprint arXiv:2210.02747, 2022

    Pith/arXiv arXiv 2022

  71. [76]

    Flow straight and fast: Learning to generate and transfer data with rectified flow

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow straight and fast: Learning to generate and transfer data with rectified flow. In International Conference on Learning Representations, 2023

    2023

  72. [77]

    Albergo, Nicholas M

    Nanye Ma, Mark Goldstein, Michael S. Albergo, Nicholas M. Boffi, Eric Vanden-Eijnden, and Saining Xie. SiT : Exploring flow and diffusion-based generative models with scalable interpolant transformers. In ECCV, 2024

    2024

  73. [78]

    Scalable diffusion models with transformers

    William Peebles and Saining Xie. Scalable diffusion models with transformers. In ICCV, 2023

    2023

  74. [79]

    Computational optimal transport

    Gabriel Peyr \'e and Marco Cuturi. Computational optimal transport. Foundations and Trends in Machine Learning, 11 0 (5--6): 0 355--607, 2019

    2019

  75. [80]

    Aram-Alexandre Pooladian, Heli Ben-Hamu, Carles Domingo-Enrich, Brandon Amos, Yaron Lipman, and Ricky T. Q. Chen. Multisample flow matching: Straightening flows with minibatch couplings. In International Conference on Machine Learning, 2023

    2023

  76. [81]

    The cross-entropy method for combinatorial and continuous optimization

    Reuven Y Rubinstein. The cross-entropy method for combinatorial and continuous optimization. Methodology and Computing in Applied Probability, 1 0 (2): 0 127--190, 1999

    1999

  77. [82]

    Marwin H. S. Segler, Thierry Kogej, Christian Tyrchan, and Mark P. Waller. Generating focused molecule libraries for drug discovery with recurrent neural networks. ACS Central Science, 4 0 (1): 0 120--131, 2018

    2018

  78. [83]

    GraphAF : A flow-based autoregressive model for molecular graph generation

    Chence Shi, Minkai Xu, Zhaocheng Zhu, Weinan Zhang, Ming Zhang, and Jian Tang. GraphAF : A flow-based autoregressive model for molecular graph generation. In International Conference on Learning Representations, 2020

    2020

  79. [84]

    Weiss, Niru Maheswaranathan, and Surya Ganguli

    Jascha Sohl-Dickstein, Eric A. Weiss, Niru Maheswaranathan, and Surya Ganguli. Deep unsupervised learning using nonequilibrium thermodynamics. In International Conference on Machine Learning, 2015

    2015

  80. [85]

    Generative modeling by estimating gradients of the data distribution

    Yang Song and Stefano Ermon. Generative modeling by estimating gradients of the data distribution. In Advances in Neural Information Processing Systems, 2019

    2019

Showing first 80 references.

This paper was first reviewed by grok-4.5 on July 13, 2026.