Pith. sign in

REVIEW 4 major objections 4 minor 52 references

Predictive caching of last-block features doubles the throughput of pretrained flow-matching molecule generators at matched sample quality, and can reach roughly triple speed with minimal quality loss.

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 ↗

T0 review · deepseek-v4-flash

2026-08-04 11:24 UTC pith:3BZBFHUS

load-bearing objection Useful transfer of predictive caching to equivariant flow-matching molecule generation; speedups look real, but 'matched quality' is oversold and the smoothness evidence is thin. the 4 major comments →

arxiv 2510.04646 v2 pith:3BZBFHUS submitted 2025-10-06 cs.LG cs.AI

Predictive Feature Caching for Training-free Acceleration of Molecular Geometry Generation

Johanna Sommer , John Rachwan , Nils Fleischmann , Stephan G\"unnemann , Bertrand Charpentier This is my paper
classification cs.LG cs.AI
keywords predictive feature cachingflow matchingmolecular geometry generationSE(3) equivariancetraining-free accelerationTaylor series forecastAdams-Bashforth forecastinference speedup
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.

Flow-matching models that generate 3D molecular geometries pay a large inference cost because every solver step runs a full equivariant network. This paper argues that the network's last-block features evolve smoothly along the sampling trajectory, so those features can be cached at selected checkpoints and forecast in between with a Taylor-series or Adams-Bashforth extrapolation. Applied to a pretrained equivariant flow-matching generator on GEOM-Drugs, the method cuts the effective number of network evaluations from 100 to 51 and doubles throughput while matching or improving the base model's energy and strain metrics; at more aggressive caching it reaches about 3x speed with only a small validity drop. Because the forecast is a linear combination of cached features, it commutes with the SE(3) action, so equivariance is preserved without retraining. The authors also show the cache composes with graph compilation and TF32 to reach about 7x speedup overall.

Core claim

On its own terms, the paper establishes that predictive feature caching lets a pretrained SE(3)-equivariant flow-matching molecule generator run at about half the network evaluations with no measured loss of sample quality. On GEOM-Drugs, the base model samples 11.4 molecules/s at 100 Euler steps while cached variants at an effective 51 steps sample 21.8-22.1 molecules/s, with equal or better energy and strain and validity within noise. At effective 34 and 26 steps the quality difference stays small, whereas the uncached 51-step baseline already degrades clearly. Combining the cache with graph compilation and TF32 reaches about 7x end-to-end speedup, taking 10,000 molecules from over 14 minu

What carries the argument

The load-bearing mechanism is a last-block predictive cache. The model F(x_t) is the final transformer block of the equivariant backbone; its input features x_t move smoothly along the ODE trajectory because both the velocity field and the network are continuous in time. Every D solver steps the cache stores F(x_t) together with finite differences (Taylor variant) or the previous few block outputs (Adams-Bashforth variant). For the D-1 intermediate steps the full backbone is skipped and the output is a scalar linear combination of cached terms. That linearity is what preserves E(3)×S_N equivariance: the cached features transform under the same group action, and the forecast inherits it. The

Load-bearing premise

The claim rests on the assumption that the hidden features of the last network block evolve smoothly enough between solver steps for a short Taylor or Adams-Bashforth extrapolation to stay accurate; the authors illustrate this on one trajectory and note that errors are largest in early steps.

What would settle it

Sample a large, size-diverse set of molecules from GEOM-Drugs and compute the per-step forecast error of last-block features for the cached run. If the error spikes around bond-forming steps or grows with molecule size such that at 51 effective steps the cached model's validity or molecule stability falls below the 100-step baseline, the iso-quality claim is falsified. A simpler check: if increasing the cache interval D from 2 to 4 makes the energy and strain advantage disappear and validity drop by more than the reported noise, the smoothness assumption is breached.

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

If this is right

  • Pretrained flow-matching molecule generators can be accelerated without any training, data, or fine-tuning.
  • At a 2x operating point the model matches or improves energy and strain relative to the 100-step baseline, so high-throughput screening can use the cache without sacrificing conformer quality.
  • Caching is orthogonal to lossless system-level optimizations; combined with graph compilation and TF32 it yields about 7x and reduces 10,000-molecule generation from over 14 minutes to about 2 minutes.
  • Higher-order forecasts (second-order Taylor, third-order Adams-Bashforth) are the best operating points, and Adams-Bashforth consistently outperforms Taylor in the reported settings.
  • Simply reducing solver steps is not a substitute: the 51-step base model degrades, while the cached 51-effective-step model does not.

Where Pith is reading between the lines

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

  • If the feature-smoothness assumption holds broadly, the same cache should transfer to other equivariant flow or diffusion generators, and possibly to latent-space molecule generators where the per-step backbone is even cheaper.
  • The largest forecast errors appear in early steps, which suggests an adaptive schedule that refreshes the cache more often at the start and less often later could push the speed-quality frontier beyond uniform caching intervals.
  • Smoothness may break precisely where chemistry changes fast, such as bond formation or steric clashes; testing cache behavior on large or flexible molecules would reveal whether the iso-quality claim survives across the full GEOM-Drugs distribution rather than only at aggregate metrics.
  • Because the method does not alter the trained vector field, it should compose with training-based accelerations such as distillation, potentially multiplying their gains rather than merely adding to them.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes a training-free predictive feature caching scheme for flow-matching molecular geometry generators. At selected checkpoint timesteps, the method stores the last-block output of the SE(3)-equivariant backbone and forecasts intermediate outputs using Taylor-series expansion (m=1,2) or Adams-Bashforth (j=2,3) predictors, thereby reducing the number of full network evaluations. Experiments on GEOM-Drugs and QM9 with the pretrained SemlaFlow model report roughly a 2× throughput increase at 51 effective steps and up to ~3.6× at 26 steps, with further composition with graph compilation and TF32 yielding up to ~7×. The authors argue that caching preserves equivariance and that energy/strain metrics are improved, while validity and Optimized RMSD are slightly worse.

Significance. If the claims hold, this is a practical, training-free acceleration for pretrained equivariant flow-matching molecule generators. It is orthogonal to training-based acceleration and system-level optimizations, and the equivariance preservation argument is elegant. The paper is transparent in sweeping cache intervals and orders, uses standard benchmarks (GEOM-Drugs, QM9), and reports multiple quality metrics. However, the headline claim of 'matched sample quality' is not fully supported by the data: Table 1 shows a systematic validity drop at all cached operating points, and the smoothness evidence for the iso-quality claim rests on a single molecule. These issues need to be addressed before the central claim can be accepted.

major comments (4)
  1. [Abstract / §5.2, Table 1] The abstract claims a 'twofold reduction in wall-clock inference time at matched sample quality.' Table 1 contradicts this: at 51 effective steps, validity drops from 0.88 (base 100) to 0.85–0.87 for all cached variants, and Optimized RMSD worsens from 0.86 to 0.87–0.88. Energy/strain improve, but validity is a primary quality metric. Please weaken the claim to 'comparable quality with a small validity drop' or provide a formal equivalence test with a pre-specified threshold.
  2. [§5.1, Fig. 2] The iso-quality claim rests on the assumption that last-block features evolve smoothly enough for low-order extrapolation. The only direct evidence is a single molecule's trajectory and an aggregate linear-prediction-error curve, with the authors noting errors are largest in the early steps. Please add a dataset-wide forecast-error analysis: distribution of prediction errors across the GEOM-Drugs test set, broken down by molecule size and time step, and relate forecast error to downstream quality metrics (e.g., validity). This is needed to rule out failure modes concentrated in a subset of molecules, which aggregate Table 1 metrics could mask.
  3. [§4, Eq. (7)] The Adams-Bashforth formula as written uses F(x_{t+k+i}) for i=1..j, i.e., future cached outputs, which cannot be causal. Standard AB predicts from past outputs. Please correct the indexing/signs or clarify the notation. As written, the AB method is not reproducible and the reported AB results are questionable.
  4. [§5.2 / Table 1 / Fig. 3] Experimental details are incomplete. (a) Report the cache interval D explicitly for each row; 'effective steps' alone does not uniquely determine D unless K is specified. (b) In Fig. 3, the combined 7× speedup configuration reports no quality metrics; please provide validity, energy, and strain for that setting to justify 'no significant loss in sampling quality.'
minor comments (4)
  1. [§5] Typo: 'eniqueness' should be 'uniqueness'. Also 'GEOM Drugs' vs 'GEOM-Drugs' is inconsistent.
  2. [Table 2] The caption says 'QM9 Drugs'; should be 'QM9'.
  3. [§5] Grammar: 'we employ GEOM Drugs a more meaningful benchmark' is missing 'as'.
  4. [Fig. 2] Specify whether the linear prediction error is normalized per molecule or per coordinate; the log-scale y-axis is otherwise ambiguous.

Circularity Check

0 steps flagged

No significant circularity: the predictive cache is an extrapolation of the model's own prior outputs and is benchmarked against external quality and throughput metrics.

full rationale

The paper's central mechanism (Eqs. 5-7) is a Taylor or Adams-Bashforth forecast of the last-block feature F(x_t) from previously cached network outputs. This is a genuine numerical extrapolation, not a quantity defined in terms of the evaluation target. The cache parameters (m, j, D) are swept and reported transparently across Table 1 rather than fit to a single headline quality number. The equivariance argument (Eq. 8) follows from the linearity of the forecast operators and the G-equivariance of the cached features; it is a mathematical derivation, not an imported self-citation. The claimed speedups are measured wall-clock throughput on an external benchmark (GEOM-Drugs) with a pretrained base model, and quality is assessed with independent metrics (validity, energy, strain, RMSD). The only self-citations (Ayadi et al., Ketata et al.) appear in related-work context and are not load-bearing. The skeptic's concern about feature smoothness across the full molecule-size distribution is an empirical robustness risk, not circular reasoning.

Axiom & Free-Parameter Ledger

3 free parameters · 5 axioms · 0 invented entities

The method introduces no new parameters to fit (its forecast orders and cache interval are swept, not fitted to the target metrics) and no new physical entities. Its load-bearing premise is temporal smoothness of the last-block features, which is only weakly verified.

free parameters (3)
  • Cache interval D = D=2,3,4 used for 51/34/26 effective steps
    Manually swept, not learned. Determines speed/quality trade-off.
  • Taylor order m = m=1,2 evaluated; m=0 naive caching
    Forecast order swept; higher orders generally better on GEOM-Drugs.
  • Adams-Bashforth step j = j=2,3
    Forecast order swept; j=3 best on GEOM-Drugs.
axioms (5)
  • domain assumption Last-block output features evolve smoothly in t over the cache interval D
    Section 4: 'x_t^l will vary smoothly with t'; supported by a single-molecule PCA (Fig. 1) and linear predictability error (Fig. 2), not by a dataset-wide analysis.
  • standard math The backbone is a continuous composition g_L∘...∘g_1 and vθ is continuous in (x,t)
    Used in Section 4 to justify feature smoothness; standard continuity assumption.
  • domain assumption Scalar linear combinations and finite differences of features commute with the E(3)×S_N action
    Section 4 Equivariance; requires features to lie in linear representations so that scalar weights preserve the group action.
  • domain assumption Pretrained SemlaFlow weights are a valid base model and default hyperparameters are appropriate
    Section 5: 'use pretrained weights provided by the authors'.
  • domain assumption Graph compilation and TF32 do not affect evaluation metrics
    Section 5.3: they state compilation and TF32 'without affecting evaluation metrics', but no supporting table is given.

pith-pipeline@v1.3.0-alltime-deepseek · 12083 in / 14913 out tokens · 111596 ms · 2026-08-04T11:24:10.305906+00:00 · methodology

0 comments
read the original abstract

Flow matching models generate high-fidelity molecular geometries but incur significant computational costs during inference, requiring hundreds of network evaluations. This inference overhead becomes the primary bottleneck when such models are employed in practice to sample large numbers of molecular candidates. This work discusses a training-free caching strategy that accelerates molecular geometry generation by predicting intermediate hidden states across solver steps. The proposed method operates directly on the SE(3)-equivariant backbone, is compatible with pretrained models, and is orthogonal to existing training-based accelerations and system-level optimizations. Experiments on the GEOM-Drugs dataset demonstrate that caching achieves a twofold reduction in wall-clock inference time at matched sample quality and a speedup of up to 3x compared to the base model with minimal sample quality degradation. Because these gains compound with other optimizations, applying caching alongside other general, lossless optimizations yield as much as a 7x speedup.

Figures

Figures reproduced from arXiv: 2510.04646 by Bertrand Charpentier, Johanna Sommer, John Rachwan, Nils Fleischmann, Stephan G\"unnemann.

Figure 1
Figure 1. Figure 1: Projection onto the first two principal components of a single molecule’s generation tra￾jectory. Both coordinates and atom types evolve smoothly over solver steps. Recall from section 3 that we sample by in￾tegrating x˙ t = vθ(xt, t) with xt = (c, a, b), where vθ is implemented by a shared backbone. Let the backbone be a composition of functions gL ◦· · · ◦ g1. At solver time t, we denote the input to blo… view at source ↗
Figure 2
Figure 2. Figure 2: Trajectories of an atom’s coordi￾nates and evaluation of linear predictability error over sampling steps. As a base model, we utilize SemlaFlow [Irwin et al., 2025] and use pretrained weights provided by the authors for both the QM9 and GEOM Drugs dataset. Unless explicitly stated otherwise, we use the default hyperparameters of the model reported for each of the datasets. All metrics for SemlaFlow present… view at source ↗
Figure 3
Figure 3. Figure 3: Inference time and memory overhead to sample 10,000 molecular geometries of various acceleration method combinations [PITH_FULL_IMAGE:figures/full_fig_p006_3.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

52 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    Diffusion Models for Molecules : A Survey of Methods and Tasks , February 2025

    Liang Wang, Chao Song, Zhiyuan Liu, Yu Rong, Qiang Liu, Shu Wu, and Liang Wang. Diffusion Models for Molecules : A Survey of Methods and Tasks , February 2025. URL http://arxiv.org/abs/2502.09511. arXiv:2502.09511 [cs]

    arXiv 2025

  2. [2]

    Diffusion Models in \ textit\ De Novo \ \ Drug Design

    Amira Alakhdar, Barnabas Poczos, and Newell Washburn. Diffusion Models in \ textit\ De Novo \ \ Drug Design . Journal of Chemical Information and Modeling, 64 0 (19): 0 7238--7256, October 2024

    2024

  3. [3]

    A survey of generative AI for de novo drug design: new frontiers in molecule and protein generation

    Xiangru Tang, Howard Dai, Elizabeth Knight, Fang Wu, Yunyang Li, Tianxiao Li, and Mark Gerstein. A survey of generative AI for de novo drug design: new frontiers in molecule and protein generation. Briefings in Bioinformatics, 25 0 (4), May 2024. ISSN 1467-5463, 1477-4054

    2024

  4. [4]

    De novo generated combinatorial library design

    Simon Viet Johansson, Morteza Haghir Chehreghani, Ola Engkvist, and Alexander Schliep. De novo generated combinatorial library design. Digital Discovery, 3 0 (1): 0 122--135, 2024. doi:10.1039/D3DD00095H. URL https://pubs.rsc.org/en/content/articlelanding/2024/dd/d3dd00095h. Publisher: Royal Society of Chemistry

    work page doi:10.1039/d3dd00095h 2024

  5. [5]

    Jenkins, Peter S

    Lingling Shen, Jian Fang, Lulu Liu, Fei Yang, Jeremy L. Jenkins, Peter S. Kutchukian, and He Wang. Pocket Crafter : a 3D generative modeling based workflow for the rapid generation of hit molecules in drug discovery. Journal of Cheminformatics, 16 0 (1): 0 33, March 2024. ISSN 1758-2946. doi:10.1186/s13321-024-00829-w

    work page doi:10.1186/s13321-024-00829-w 2024

  6. [6]

    Towards DNA - Encoded Library Generation with GFlowNets

    Michał Koziarski, Mohammed Abukalam, Vedant Shah, Louis Vaillancourt, Doris Alexandra Schuetz, Moksh Jain, Almer van der Sloot, Mathieu Bourgey, Anne Marinier, and Yoshua Bengio. Towards DNA - Encoded Library Generation with GFlowNets . GEM workshop ICLR 2024, April 2024

    2024

  7. [7]

    Straight- Line Diffusion Model for Efficient 3D Molecular Generation , June 2025

    Yuyan Ni, Shikun Feng, Haohan Chi, Bowen Zheng, Huan-ang Gao, Wei-Ying Ma, Zhi-Ming Ma, and Yanyan Lan. Straight- Line Diffusion Model for Efficient 3D Molecular Generation , June 2025. URL http://arxiv.org/abs/2503.02918. arXiv:2503.02918 [cs]

    Pith/arXiv arXiv 2025

  8. [8]

    Accelerating the Generation of Molecular Conformations with Progressive Distillation of Equivariant Latent Diffusion Models

    Romain Lacombe and Neal Vaidya. Accelerating the Generation of Molecular Conformations with Progressive Distillation of Equivariant Latent Diffusion Models . Generative and Experimental Perspectives for Biomolecular Design Workshop at the 12th International Conference on Learning Representations, April 2024

    2024

  9. [9]

    Geometric Latent Diffusion Models for 3D Molecule Generation

    Minkai Xu, Alexander Powers, Ron Dror, Stefano Ermon, and Jure Leskovec. Geometric Latent Diffusion Models for 3D Molecule Generation . The Fortieth International Conference on Machine Learning, May 2023

    2023

  10. [10]

    Cache Me if You Can : Accelerating Diffusion Models through Block Caching , 2023

    Felix Wimbauer, Bichen Wu, Edgar Schoenfeld, Xiaoliang Dai, Ji Hou, Zijian He, Artsiom Sanakoyeu, Peizhao Zhang, Sam Tsai, Jonas Kohler, Christian Rupprecht, Daniel Cremers, Peter Vajda, and Jialiang Wang. Cache Me if You Can : Accelerating Diffusion Models through Block Caching , 2023. URL https://arxiv.org/abs/2312.03209. Version Number: 2

    Pith/arXiv arXiv 2023

  11. [11]

    DeepCache : Accelerating Diffusion Models for Free , December 2023

    Xinyin Ma, Gongfan Fang, and Xinchao Wang. DeepCache : Accelerating Diffusion Models for Free , December 2023. URL http://arxiv.org/abs/2312.00858. arXiv:2312.00858 [cs]

    Pith/arXiv arXiv 2023

  12. [12]

    Faster Diffusion : Rethinking the Role of the Encoder for Diffusion Model Inference

    Senmao Li, Taihang Hu, Joost van de Weijer, Fahad Shahbaz Khan, Tao Liu, Linxuan Li, Shiqi Yang, Yaxing Wang, Ming-Ming Cheng, and Jian Yang. Faster Diffusion : Rethinking the Role of the Encoder for Diffusion Model Inference . 38th Conference on Neural Information Processing Systems, October 2024. URL http://arxiv.org/abs/2312.09608

    Pith/arXiv arXiv 2024

  13. [13]

    FORA : Fast - Forward Caching in Diffusion Transformer Acceleration , July 2024

    Pratheba Selvaraju, Tianyu Ding, Tianyi Chen, Ilya Zharkov, and Luming Liang. FORA : Fast - Forward Caching in Diffusion Transformer Acceleration , July 2024. URL http://arxiv.org/abs/2407.01425. arXiv:2407.01425 [cs]

    Pith/arXiv arXiv 2024

  14. [14]

    \ delta\ - DiT : A Training - Free Acceleration Method Tailored for Diffusion Transformers , June 2024

    Pengtao Chen, Mingzhu Shen, Peng Ye, Jianjian Cao, Chongjun Tu, Christos-Savvas Bouganis, Yiren Zhao, and Tao Chen. \ delta\ - DiT : A Training - Free Acceleration Method Tailored for Diffusion Transformers , June 2024. URL http://arxiv.org/abs/2406.01125. arXiv:2406.01125 [cs]

    Pith/arXiv arXiv 2024

  15. [15]

    From Reusing to Forecasting : Accelerating Diffusion Models with TaylorSeers

    Jiacheng Liu, Chang Zou, Yuanhuiyi Lyu, Junjie Chen, and Linfeng Zhang. From Reusing to Forecasting : Accelerating Diffusion Models with TaylorSeers . International Conference on Computer Vision, August 2025 a

    2025

  16. [16]

    AB - Cache : Training - Free Acceleration of Diffusion Models via Adams - Bashforth Cached Feature Reuse , April 2025

    Zichao Yu, Zhen Zou, Guojiang Shao, Chengwei Zhang, Shengze Xu, Jie Huang, Feng Zhao, Xiaodong Cun, and Wenyi Zhang. AB - Cache : Training - Free Acceleration of Diffusion Models via Adams - Bashforth Cached Feature Reuse , April 2025. URL http://arxiv.org/abs/2504.10540. arXiv:2504.10540 [stat]

    Pith/arXiv arXiv 2025

  17. [17]

    Timestep Embedding Tells : It 's Time to Cache for Video Diffusion Model

    Feng Liu, Shiwei Zhang, Xiaofeng Wang, Yujie Wei, Haonan Qiu, Yuzhong Zhao, Yingya Zhang, Qixiang Ye, and Fang Wan. Timestep Embedding Tells : It 's Time to Cache for Video Diffusion Model . Computer Vision and Pattern Recognition Conference, March 2025 b . URL http://arxiv.org/abs/2411.19108

    Pith/arXiv arXiv 2025

  18. [18]

    Region- Adaptive Sampling for Diffusion Transformers , February 2025 c

    Ziming Liu, Yifan Yang, Chengruidong Zhang, Yiqi Zhang, Lili Qiu, Yang You, and Yuqing Yang. Region- Adaptive Sampling for Diffusion Transformers , February 2025 c . URL http://arxiv.org/abs/2502.10389. arXiv:2502.10389 [cs]

    arXiv 2025

  19. [19]

    UniCP : A Unified Caching and Pruning Framework for Efficient Video Generation , February 2025

    Wenzhang Sun, Qirui Hou, Donglin Di, Jiahui Yang, Yongjia Ma, and Jianxun Cui. UniCP : A Unified Caching and Pruning Framework for Efficient Video Generation , February 2025. URL http://arxiv.org/abs/2502.04393. arXiv:2502.04393 [cs]

    Pith/arXiv arXiv 2025

  20. [20]

    Zhengyao Lv, Chenyang Si, Junhao Song, Zhenyu Yang, Yu Qiao, Ziwei Liu, and Kwan-Yee K. Wong. FasterCache : Training - Free Video Diffusion Model Acceleration with High Quality , March 2025. URL http://arxiv.org/abs/2410.19355. arXiv:2410.19355 [cs]

    Pith/arXiv arXiv 2025

  21. [21]

    DiTFastAttn : Attention Compression for Diffusion Transformer Models

    Zhihang Yuan, Hanling Zhang, Pu Lu, Xuefei Ning, Linfeng Zhang, Tianchen Zhao, Shengen Yan, Guohao Dai, and Yu Wang. DiTFastAttn : Attention Compression for Diffusion Transformer Models . 38th Conference on Neural Information Processing Systems, October 2024

    2024

  22. [22]

    Learning a Continuous Representation of 3D Molecular Structures with Deep Generative Models

    Matthew Ragoza, Tomohide Masuda, and David Ryan Koes. Learning a Continuous Representation of 3D Molecular Structures with Deep Generative Models . MLSB workshop at NeurIPS, November 2020

    2020

  23. [23]

    Niklas W. A. Gebauer, Michael Gastegger, and Kristof T. Schütt. Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules, January 2020. URL http://arxiv.org/abs/1906.00957. arXiv:1906.00957 [stat]

    Pith/arXiv arXiv 2020

  24. [24]

    An Autoregressive Flow Model for 3D Molecular Geometry Generation from Scratch

    Youzhi Luo and Shuiwang Ji. An Autoregressive Flow Model for 3D Molecular Geometry Generation from Scratch . 42nd International Conference on Machine Learning, 2022

    2022

  25. [25]

    Fuchs, Ingmar Posner, and Max Welling

    Victor Garcia Satorras, Emiel Hoogeboom, Fabian B. Fuchs, Ingmar Posner, and Max Welling. E(n) Equivariant Normalizing Flows . 35th Conference on Neural Information Processing Systems, January 2022

    2022

  26. [26]

    Equivariant Diffusion for Molecule Generation in 3D

    Emiel Hoogeboom, Victor Garcia Satorras, Clément Vignac, and Max Welling. Equivariant Diffusion for Molecule Generation in 3D . International Conference on Machine Learning, June 2022

    2022

  27. [27]

    MDM : Molecular Diffusion Model for 3D Molecule Generation , September 2022

    Lei Huang, Hengtong Zhang, Tingyang Xu, and Ka-Chun Wong. MDM : Molecular Diffusion Model for 3D Molecule Generation , September 2022. URL http://arxiv.org/abs/2209.05710. arXiv:2209.05710 [cs]

    Pith/arXiv arXiv 2022

  28. [28]

    Learning Joint 2D & 3D Diffusion Models for Complete Molecule Generation , June 2023

    Han Huang, Leilei Sun, Bowen Du, and Weifeng Lv. Learning Joint 2D & 3D Diffusion Models for Complete Molecule Generation , June 2023. URL http://arxiv.org/abs/2305.12347. arXiv:2305.12347 [q-bio]

    Pith/arXiv arXiv 2023

  29. [29]

    MiDi : Mixed Graph and 3D Denoising Diffusion for Molecule Generation , June 2023

    Clement Vignac, Nagham Osman, Laura Toni, and Pascal Frossard. MiDi : Mixed Graph and 3D Denoising Diffusion for Molecule Generation , June 2023. URL http://arxiv.org/abs/2302.09048. arXiv:2302.09048 [cs]

    Pith/arXiv arXiv 2023

  30. [30]

    Coarse-to- Fine : a Hierarchical Diffusion Model for Molecule Generation in 3D

    Bo Qiang, Yuxuan Song, Minkai Xu, Jingjing Gong, Bowen Gao, Hao Zhou, Weiying Ma, and Yanyan Lan. Coarse-to- Fine : a Hierarchical Diffusion Model for Molecule Generation in 3D . The Fortieth International Conference on Machine Learning, May 2023

    2023

  31. [31]

    Geometry-Complete Diffusion for 3D Molecule Generation and Optimization

    Alex Morehead and Jianlin Cheng. Geometry- Complete Diffusion for 3D Molecule Generation and Optimization . MLDD workshop @ ICLR, May 2024. doi:10.48550/arXiv.2302.04313. URL http://arxiv.org/abs/2302.04313

    work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2302.04313 2024

  32. [32]

    Diffusion-based Molecule Generation with Informative Prior Bridges , September 2022

    Lemeng Wu, Chengyue Gong, Xingchao Liu, Mao Ye, and Qiang Liu. Diffusion-based Molecule Generation with Informative Prior Bridges , September 2022. URL http://arxiv.org/abs/2209.00865. arXiv:2209.00865 [cs]

    Pith/arXiv arXiv 2022

  33. [33]

    Applications of Modular Co - Design for De Novo 3D Molecule Generation , May 2025

    Danny Reidenbach, Filipp Nikitin, Olexandr Isayev, and Saee Paliwal. Applications of Modular Co - Design for De Novo 3D Molecule Generation , May 2025. URL http://arxiv.org/abs/2505.18392. arXiv:2505.18392 [cs]

    Pith/arXiv arXiv 2025

  34. [34]

    Accelerating 3D Molecule Generation via Jointly Geometric Optimal Transport

    Haokai Hong, Wanyu Lin, and Kay Chen Tan. Accelerating 3D Molecule Generation via Jointly Geometric Optimal Transport . The Thirteenth International Conference on Learning Representations, March 2025

    2025

  35. [35]

    UniGEM : A Unified Approach to Generation and Property Prediction for Molecules

    Shikun Feng, Yuyan Ni, Yan Lu, Zhi-Ming Ma, Wei-Ying Ma, and Yanyan Lan. UniGEM : A Unified Approach to Generation and Property Prediction for Molecules . arXiv, April 2025. URL http://arxiv.org/abs/2410.10516. arXiv:2410.10516 [cs]

    Pith/arXiv arXiv 2025

  36. [36]

    SemlaFlow -- Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching

    Ross Irwin, Alessandro Tibo, Jon Paul Janet, and Simon Olsson. SemlaFlow -- Efficient 3D Molecular Generation with Latent Attention and Equivariant Flow Matching . 28th International Conference on Artificial Intelligence and Statistics, February 2025

    2025

  37. [37]

    Equivariant Flow Matching with Hybrid Probability Transport

    Yuxuan Song, Jingjing Gong, Minkai Xu, Ziyao Cao, Yanyan Lan, Stefano Ermon, Hao Zhou, and Wei-Ying Ma. Equivariant Flow Matching with Hybrid Probability Transport . 37th Conference on Neural Information Processing Systems, December 2023

    2023

  38. [38]

    Mixed Continuous and Categorical Flow Matching for 3D De Novo Molecule Generation , April 2024

    Ian Dunn and David Ryan Koes. Mixed Continuous and Categorical Flow Matching for 3D De Novo Molecule Generation , April 2024. URL http://arxiv.org/abs/2404.19739. arXiv:2404.19739 [q-bio]

    Pith/arXiv arXiv 2024

  39. [39]

    Joshi, Xiang Fu, Yi-Lun Liao, Vahe Gharakhanyan, Benjamin Kurt Miller, Anuroop Sriram, and Zachary W

    Chaitanya K. Joshi, Xiang Fu, Yi-Lun Liao, Vahe Gharakhanyan, Benjamin Kurt Miller, Anuroop Sriram, and Zachary W. Ulissi. All-atom Diffusion Transformers : Unified generative modelling of molecules and materials, May 2025. URL http://arxiv.org/abs/2503.03965. arXiv:2503.03965 [cs]

    Pith/arXiv arXiv 2025

  40. [40]

    Unified Guidance for Geometry - Conditioned Molecular Generation

    Sirine Ayadi, Leon Hetzel, Johanna Sommer, Fabian Theis, and Stephan Günnemann. Unified Guidance for Geometry - Conditioned Molecular Generation . 38th Conference on Neural Information Processing Systems, January 2025

    2025

  41. [41]

    Lift Your Molecules : Molecular Graph Generation in Latent Euclidean Space

    Mohamed Amine Ketata, Nicholas Gao, Johanna Sommer, Tom Wollschläger, and Stephan Günnemann. Lift Your Molecules : Molecular Graph Generation in Latent Euclidean Space . The Thirteenth International Conference on Learning Representations, June 2024

    2024

  42. [42]

    Yaron Lipman, Ricky T. Q. Chen, Heli Ben-Hamu, Maximilian Nickel, and Matt Le. Flow Matching for Generative Modeling , February 2023. URL http://arxiv.org/abs/2210.02747. arXiv:2210.02747 [cs]

    Pith/arXiv arXiv 2023

  43. [43]

    Improving and generalizing flow-based generative models with minibatch optimal transport

    Alexander Tong, Kilian Fatras, Nikolay Malkin, Guillaume Huguet, Yanlei Zhang, Jarrid Rector-Brooks, Guy Wolf, and Yoshua Bengio. Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research, March 2024

    2024

  44. [44]

    Flow Straight and Fast : Learning to Generate and Transfer Data with Rectified Flow , September 2022

    Xingchao Liu, Chengyue Gong, and Qiang Liu. Flow Straight and Fast : Learning to Generate and Transfer Data with Rectified Flow , September 2022. URL http://arxiv.org/abs/2209.03003. arXiv:2209.03003 [cs]

    Pith/arXiv arXiv 2022

  45. [45]

    Forecasting When to Forecast : Accelerating Diffusion Models with Confidence - Gated Taylor , August 2025

    Xiaoliu Guan, Lielin Jiang, Hanqi Chen, Xu Zhang, Jiaxing Yan, Guanzhong Wang, Yi Liu, Zetao Zhang, and Yu Wu. Forecasting When to Forecast : Accelerating Diffusion Models with Confidence - Gated Taylor , August 2025. URL http://arxiv.org/abs/2508.02240. arXiv:2508.02240 [cs]

    arXiv 2025

  46. [46]

    Dral, Matthias Rupp, and O

    Raghunathan Ramakrishnan, Pavlo O. Dral, Matthias Rupp, and O. Anatole von Lilienfeld. Quantum chemistry structures and properties of 134 kilo molecules. Scientific Data, 1 0 (1): 0 140022, August 2014. ISSN 2052-4463. doi:10.1038/sdata.2014.22. URL https://www.nature.com/articles/sdata201422. Publisher: Nature Publishing Group

    doi:10.1038/sdata.2014.22 2014

  47. [47]

    GEOM : Energy -annotated molecular conformations for property prediction and molecular generation, February 2022

    Simon Axelrod and Rafael Gomez-Bombarelli. GEOM : Energy -annotated molecular conformations for property prediction and molecular generation, February 2022. URL http://arxiv.org/abs/2006.05531. arXiv:2006.05531 [physics]

    Pith/arXiv arXiv 2022

  48. [48]

    Navigating the Design Space of Equivariant Diffusion - Based Generative Models for De Novo 3D Molecule Generation , November 2023

    Tuan Le, Julian Cremer, Frank Noé, Djork-Arné Clevert, and Kristof Schütt. Navigating the Design Space of Equivariant Diffusion - Based Generative Models for De Novo 3D Molecule Generation , November 2023. URL http://arxiv.org/abs/2309.17296. arXiv:2309.17296 [cs]

    Pith/arXiv arXiv 2023

  49. [49]

    Morris, and Charlotte M

    Martin Buttenschoen, Yael Ziv, Garrett M. Morris, and Charlotte M. Deane. An evaluation of unconditional 3D molecular generation methods. GEM workshop ICLR 2025, May 2025

    2025

  50. [50]

    Rdkit documentation, 2013

    Greg Landrum. Rdkit documentation, 2013

    2013

  51. [51]

    Morris, and Charlotte M

    Martin Buttenschoen, Garrett M. Morris, and Charlotte M. Deane. PoseBusters : AI -based docking methods fail to generate physically valid poses or generalise to novel sequences. Chemical Science, 15 0 (9): 0 3130--3139, 2024. ISSN 2041-6520, 2041-6539. doi:10.1039/D3SC04185A. URL http://arxiv.org/abs/2308.05777

    Pith/arXiv arXiv 2024

  52. [52]

    PyTorch : An Imperative Style , High - Performance Deep Learning Library , December 2019

    Adam Paszke, Sam Gross, Francisco Massa, Adam Lerer, James Bradbury, Gregory Chanan, Trevor Killeen, Zeming Lin, Natalia Gimelshein, Luca Antiga, Alban Desmaison, Andreas Köpf, Edward Yang, Zach DeVito, Martin Raison, Alykhan Tejani, Sasank Chilamkurthy, Benoit Steiner, Lu Fang, Junjie Bai, and Soumith Chintala. PyTorch : An Imperative Style , High - Perf...

    Pith/arXiv arXiv 2019