Guiding Diffusion Models with Reinforcement Learning for Stable Molecule Generation

Chao Qu; Fenglei Cao; Junyi An; Xuan Zhang; Yuan Qi; Yunfei Shi; Zhijian Zhou; Zongkai Liu

REVIEW 3 major objections 7 minor 1 cited by

Guiding Diffusion Models with Reinforcement Learning for Stable Molecule Generation

T0 review · 3 major / 7 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Fine-tuning diffusion models with force-field rewards substantially improves molecular stability, validity, and equilibrium quality.

desk verdict RLPF is a credible, useful extension of DDPO to 3D molecular diffusion with physical rewards, but the size-invariant log-likelihood in §4.5 changes the objective in a way the paper doesn't justify, and several ablation tables are missing numbers. read the letter →

arxiv 2508.16521 v1 pith:E5GOHJN2 submitted 2025-08-22 cs.LG cs.AI

Zhijian Zhou , Junyi An , Zongkai Liu , Yunfei Shi , Xuan Zhang , Fenglei Cao , Chao Qu , Yuan Qi This is my paper · ORCID

classification cs.LGcs.AI

keywords 3Dmoleculargenerationequivariantdiffusionmodelsreinforcementlearningforce-fieldrewardstabilityPPOfine-tuningQM9GEOM-drug

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

This paper claims that a pretrained 3D diffusion model for molecules can be fine-tuned with reinforcement learning whose reward is a physical force-field evaluation, and that this post-training step substantially improves structural stability. The proposed method, RLPF, treats each denoising step as an action in a Markov decision process and scores the final molecule by the root-mean-square deviation of its atomic forces under DFT or GFN2-xTB. On QM9, fine-tuning EDM with DFT force rewards raises molecule stability from 82.0% to 93.37% and validity from 91.9% to 98.22%; on GEOM-drug, atom stability rises from 81.3% to 87.52%. The same recipe improves GeoLDM and UniGEM, and a control experiment shows continued supervised training on valid samples does not produce comparable gains. If correct, this makes physics-based RL fine-tuning a general post-training strategy for molecular diffusion models.

What carries the argument

The load-bearing object is the force-RMSD reward, the root-mean-square deviation of atomic forces computed on the final generated molecule. It is attached to a DDPO-style Markov decision process over denoising steps and optimized with PPO-style clipping and standardized advantages. A size-invariant log-likelihood, implemented as a masked average over valid atoms, keeps gradient contributions balanced across variable-size molecules. The force-field reward is what converts chemically valid generation into physically relaxed generation.

What would settle it

Before fine-tuning, draw a large sample from the pretrained model and compute the variance of force-RMSD rewards. If that variance is near zero, RLPF predicts negligible stability gain; if variance is large, molecule stability should jump by roughly the reported margin. A second check is to re-optimize generated post-RLPF molecules with the same force field: if the low residual forces are genuine equilibrium, energy minimization should change them little, and a different force field should also report low forces.

Watch

Extended reading notes

Core claim

RLPF's central claim is that direct physical feedback, in the form of atomic force residuals, is a usable reward for aligning diffusion-generated molecules with equilibrium structures. The paper formulates the reverse diffusion process as an MDP in which the state is the noised molecular latent, the action is the denoising step, and a scalar reward is assigned only at the terminal step: the force RMSD from DFT at the B3LYP/6-31G(2df,p) level or the semi-empirical GFN2-xTB method, or a valency-based stability signal. Policy updates use a PPO-style clipped surrogate objective with advantage normalization and a size-invariant masked log-likelihood so molecules with different atom counts contrib

Load-bearing premise

RLPF's gains depend on the base model having enough spread between stable and unstable outputs; if the pretrained model generates molecules of similar quality, all advantages shrink toward zero and the fine-tuning signal disappears.

Editorial extensions

If this is right

RLPF can be dropped onto any pretrained 3D diffusion generator without collecting new data or retraining the backbone from scratch.
Force-field rewards can be chosen by cost: DFT for small molecules, GFN2-xTB for drug-sized molecules, with similar stability gains at much lower compute.
RLPF fine-tuning roughly halves both the number of samples and the wall-clock time required for rejection sampling of stable molecules.
Adding a property-prediction term to the force reward improves conditional generation accuracy while preserving stability gains.
Supervised continued training on valid samples does not reproduce RLPF's stability gains, indicating that the reinforcement signal, not extra data, drives the improvement.

Reading between the lines

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

A natural next step is to reuse the same pipeline with rewards beyond force residuals, such as binding energy, strain energy, or solvation free energy, to steer generation toward function rather than equilibrium.
Because RLPF's advantage signal depends on reward variance, monitoring that variance during training could serve as a diagnostic: when it collapses, stability gains should plateau. This follows directly from the paper's stated limitation.
The reduced novelty of the DFT-fine-tuned model may not be a defect, since QM9 is a nearly complete enumeration of small stable molecules; lower novelty may partly reflect stronger confinement to valid chemical space.
A stress test of the reward would be to re-minimize post-RLPF molecules with a different force field; if the forces are genuinely near equilibrium, the alternative force field should also report low residuals, and a mismatch would suggest overfitting to one energy surrogate.

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.

Desk Editor's Note

read the letter

The core claim here is believable: fine-tuning 3D diffusion models with force-field-based rewards improves stability and validity across multiple backbones, and the continued-training control in Table 6 is the right check. The method is a fairly direct transfer of DDPO/PPO to equivariant diffusion with a domain-specific reward, and the multi-backbone results (EDM, GeoLDM, UniGEM) plus the code release make it easy to take seriously. I credit the authors for stating their own limitation in Appendix D about dependence on base-model diversity; that's an honest and real weakness, not a hidden one.

The main soft spot is Eq. (13) in §4.5. The paper defines the reverse-transition log-probability as an average over valid atoms, dividing by |M_i|. That makes the importance ratio the N-th root of the actual trajectory density ratio, so the PPO surrogate is not estimating the J_RLPF in Eq. (6). This is not a constant scaling across molecules—N varies—so the effective reward is R/N, which reweights the objective toward small molecules. The paper calls this 'size-invariant' and says it's critical, but gives no derivation and no ablation comparing against the unnormalized ratio. If force rewards correlate with molecular size, some of the reported gains could come from this reweighting rather than from the physical feedback per se. This is testable and should be resolved before the central claim is clean.

Second, several tables are missing numbers. Tables 4, 5, and 9 have blank entries in the version I read. That's a rendering or compilation failure, but it makes key ablations unverifiable. Also, the paper doesn't compare against existing RL-based molecular diffusion methods like GDPO; since the novelty rests on the specific combination, at least one such baseline would help isolate the effect of the physics reward.

None of this is fatal. The paper is systematic, the fairness control is good, and the honest limitation in Appendix D suggests the authors understand their method's boundary. The missing derivations and numbers are fixable in revision. I'd send it to peer review, but I'd ask for a derivation of Eq. (13), an ablation with and without the per-atom normalization, and completed tables. A reader working on RL for generative models or molecule generation would get value from this.

Referee Report

3 major / 7 minor

Summary. The paper proposes RLPF (Reinforcement Learning with Physical Feedback), a post-training procedure that fine-tunes pretrained 3D equivariant diffusion models with PPO-style policy gradient updates. The reward is derived from atomic force residuals computed by DFT (B3LYP/6-31G(2df,p)) or GFN2-xTB, with a valency-based stability reward as an additional variant. The method is applied to EDM, GeoLDM, and UniGEM on the QM9 and GEOM-drug datasets. The main reported result is a large improvement in molecule stability (e.g., EDM on QM9: 82.0% to 93.37%) and validity, with smaller gains on GEOM-drug. The appendix contains a continued-training control, reward-function ablations, sampling-step ablations, a clipping-threshold ablation, conditional generation experiments, and an explicit statement of a limitation regarding base-model sampling diversity.

Significance. If the central claim holds, RLPF provides a practical and model-agnostic way to align 3D molecular diffusion models with physical force-field feedback. The method's rewards are external (DFT/xTB), the evaluation metrics are standard, and the paper includes a fairness control against continued training. The code is released, and the paper honestly states a key limitation in Appendix D. However, the policy-gradient objective is modified in a way that is not derived and is not equivalent to the stated RL objective; this affects the attribution of the reported gains. The experimental results are also presented with inconsistencies in key tables. The idea is promising, but the current manuscript requires substantial revision before the claims can be accepted.

major comments (3)

[§4.5, Eq. (13)] The masked per-atom average of the reverse-transition log-probability changes the importance-sampling ratio in Eq. (9). For a molecule with N atoms, log r becomes (1/N) times the true trajectory log-ratio. Consequently, the PPO surrogate in Eq. (11) does not estimate J_RLPF in Eq. (6); it rescales the effective reward by 1/N per trajectory. Since QM9 molecules range from a few to 29 atoms, this size normalization can reweight the objective toward small molecules if force rewards correlate with size. The paper's statement that the Gaussian normalization constant cancels is plausible, but the factor 1/N does not cancel and is not constant across molecules. No derivation or ablation with the unnormalized trajectory ratio is provided. This is load-bearing because the averaged likelihood is described as 'critical' to RLPF. Please provide a derivation, and either use the true trajectory log-pr
[§5.1 vs. §5.3, Tables 1 and 3] The EDM-RLPF results on QM9 are reported inconsistently across the two tables. Table 1 lists validity 98.22% and V·U 92.87%, while Table 3 lists validity 97.02% and a garbled V·U value for the same EDM-RLPF configuration. These are not the same numbers, and the paper does not explain the discrepancy. This makes the headline result difficult to reproduce. Please reconcile the tables or state explicitly which configuration/run each number corresponds to.
[Appendix A.2, Table 6] The fairness control is a useful experiment, but it does not fully isolate the effect of RL versus continued training. The control trains only on rejection-sampled valid molecules, whereas RLPF trains on all sampled trajectories and assigns a penalty to invalid structures. Thus the training distributions differ both in reward signal and in data composition. Please discuss this asymmetry and, if possible, add a control that uses the same accepted/penalized trajectory split without policy-gradient updates, or clarify why the current control is sufficient.

minor comments (7)

[§4.5, Eq. (13)] The notation is under-specified: the meaning of M_i, the index ranges for s and j, and the exact average (over atoms, features, or both) are not fully defined. Please clarify the formula and its relationship to the per-timestep transition distribution.
[Table 1] The baseline 'END [20]' should be 'EDN [20]' to match the reference and the rest of the text. The same typo appears in Table 2.
[Table 2] The 'Data (Ground Truth)' row is ambiguous: the column alignment suggests only one of Atom Stability or Validity is reported, but the entry '86.5' is not labeled. Please clarify which metric it refers to.
[Appendix A.1, Tables 4 and 5] Several numerical values in these tables are missing or corrupted (e.g., polarizability MAE values, all rows in Table 5, and parts of Table 4). Ablation results cannot be assessed from the current text. Please ensure all reported numbers are visible and precise.
[Appendix A.5, Table 9] The Stability and Force (xTB) rows report identical molecule stability (96.45%) and identical atom stability/validity values; please verify whether this is a formatting error or a genuine result, and explain why two different reward functions produce identical values on all metrics.
[Algorithm 1] The pseudocode mentions the Advantage A but does not define it; the definition appears only in Eq. (10). Please add a reference to the equation in the algorithm.
[§5.1] The paper uses N both for the number of denoising steps and later for the number of sampled trajectories, and Eq. (13) uses N for the number of atoms. Please use distinct symbols to avoid confusion.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: force-field rewards are external and the main stability metrics are independent of the training objective.

full rationale

The derivation chain for the central stability claim is self-contained. The reward in Eq. (12) is an external force-field RMSD (DFT or GFN2-xTB), and the main QM9/GEOM-drug metrics (atom stability, molecule stability, validity, V·U) are standard RDKit/valency metrics that are not defined in terms of the reward. No parameter is fitted to the reported stability numbers, and the PPO/advantage machinery is standard DDPO. The size-invariant log-likelihood in §4.5 is an explicit design choice; although it changes the objective (the importance ratio becomes a per-atom root and the effective reward is scaled by 1/|M_i|), this is a correctness/objective-mismatch concern, not a circular reduction of the result to an input. Appendix D candidly states the load-bearing diversity assumption, which is a limitation, not circularity. The conditional-generation experiment in A.1 uses the same pretrained predictor f in the reward and in the MAE evaluation, so the property-alignment improvement is partly reward-alignment; but this is an evaluation-overlap concern in an auxiliary appendix, not the paper's central derivation, and it does not make the stability claims circular. No load-bearing self-citation or imported-uniqueness argument appears.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim relies on standard DDPM/PPO machinery and on two domain-specific postulates: (1) force-field RMSD rewards are a valid proxy for molecular stability, and (2) the size-invariant log-likelihood (averaging Gaussian log-probabilities over atoms/features) preserves a useful policy gradient. The latter is asserted without derivation. No new physical entities are introduced.

free parameters (4)

PPO clipping threshold ε = 0.2
Chosen by hand; ablation (Fig 3, Appendix A.6) shows 0.2 balances stability and reward improvement.
Convergence threshold for force-based rewards = reward > -0.25
Empirically determined stopping criterion (Appendix B.2).
Convergence threshold for valency-based rewards = reward > 0.95
Empirically determined stopping criterion (Appendix B.2).
Conditional reward weighting λ = 0.5 (claimed best)
Ablation Table 5, but values are missing; text claims intermediate λ best.

assumptions (5)

standard math The reverse transition p(x_s | x_t) is Gaussian (standard DDPM assumption)
Invoked in Section 4.5, Eq (13), for computing log-probabilities.
standard math The denoising network is E(3)-equivariant, preserving rotational/translational/reflection symmetries
Required for the EDM framework (Section 3.1, Eq (2)).
domain assumption DFT (B3LYP/6-31G(2df,p)) and GFN2-xTB force calculations provide reliable measures of physical stability
Used to define the reward (Section 4.4); if the force fields are inaccurate, the reward misguides the model.
ad hoc to paper Averaging the Gaussian log-probability over atoms and features yields a valid importance sampling ratio for policy gradient optimization
Asserted in Section 4.5 without derivation; this is the size-invariant log-likelihood contribution.
domain assumption The pretrained base model samples diverse trajectories with sufficient reward variance
Stated as a limitation in Appendix D; if the base model is too narrow, advantages are uniformly small and RLPF fails.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Guiding Diffusion Models with Reinforcement Learning for Stable Molecule Generation." pith.science (2026). https://pith.science/paper/E5GOHJN2

@misc{pith2026250816521,
  author       = {Pith},
  title        = {Pith review of: Guiding Diffusion Models with Reinforcement Learning for Stable Molecule Generation},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/E5GOHJN2}},
  note         = {Machine review of arXiv:2508.16521}
}

read the original abstract

Generating physically realistic 3D molecular structures remains a core challenge in molecular generative modeling. While diffusion models equipped with equivariant neural networks have made progress in capturing molecular geometries, they often struggle to produce equilibrium structures that adhere to physical principles such as force field consistency. To bridge this gap, we propose Reinforcement Learning with Physical Feedback (RLPF), a novel framework that extends Denoising Diffusion Policy Optimization to 3D molecular generation. RLPF formulates the task as a Markov decision process and applies proximal policy optimization to fine-tune equivariant diffusion models. Crucially, RLPF introduces reward functions derived from force-field evaluations, providing direct physical feedback to guide the generation toward energetically stable and physically meaningful structures. Experiments on the QM9 and GEOM-drug datasets demonstrate that RLPF significantly improves molecular stability compared to existing methods. These results highlight the value of incorporating physics-based feedback into generative modeling. The code is available at: https://github.com/ZhijianZhou/RLPF/tree/verl_diffusion.

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 1 Pith paper

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

Atomic Design Transformer: Scaffold-Conditioned 3D Molecule Generation via xTB-Reward Reinforcement Learning
physics.comp-ph 2026-07 conditional novelty 6.0 of 10

A plain causal transformer that tokenizes atom positions in local frames generates 3D molecules directly; RL against an xTB relaxation reward lifts topology-preserving valid yield from ~50% to ~95%.

Reference graph

Works this paper leans on

35 extracted references · 34 canonical work pages · cited by 1 Pith paper

[1]

Denoising diffusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. �� , 33:6840–6851, 2020

2020
[2]

Generative adversarial networks

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. �� , 63(11): 139–144, 2020

work page 2020
[3]

An introduction to variational autoencoders

Diederik P Kingma, Max Welling, et al. An introduction to variational autoencoders. �� , 12(4):307–392, 2019

work page 2019
[4]

E (n) equivariant graph neural networks

Vıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In �� , pages 9323–9332. PMLR, 2021

work page 2021
[5]

Equiformer: Equivariant graph attention transformer for 3d atomistic graphs

Yi-Lun Liao and Tess Smidt. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. �� , 2022

work page 2022
[6]

Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds

Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. �� , 2018

work page 2018
[7]

Geodiff: A geometric diffusion model for molecular conformation generation

Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. �� , 2022

work page 2022
[8]

Torsional diffusion for molecular conformer generation

Bowen Jing, Gabriele Corso, Jeffrey Chang, Regina Barzilay, and Tommi Jaakkola. Torsional diffusion for molecular conformer generation. �� , 35:24240–24253, 2022

work page 2022

Show all 35 references

[9]

Equivariant diffusion for molecule generation in 3D

Emiel Hoogeboom, Víctor Garcia Satorras, Clément Vignac, and Max Welling. Equivariant diffusion for molecule generation in 3D. In Kamalika Chaudhuri, Stefanie Jegelka, Le Song, Csaba Szepesvari, Gang Niu, and Sivan Sabato, editors, �� ...

2022
[10]

Learning to summarize with human feedback

Nisan Stiennon, Long Ouyang, Jeffrey Wu, Daniel Ziegler, Ryan Lowe, Chelsea V oss, Alec Radford, Dario Amodei, and Paul F Christiano. Learning to summarize with human feedback. �� , 33:3008–3021, 2020

2020
[11]

Markov decision processes

Martin L Puterman. Markov decision processes. �� , 2:331–434, 1990

1990
[12]

Simple statistical gradient-following algorithms for connectionist reinforcement learning

Ronald J Williams. Simple statistical gradient-following algorithms for connectionist reinforcement learning. �� , 8:229–256, 1992

1992
[13]

Geometric latent diffusion models for 3d molecule generation

Minkai Xu, Alexander S Powers, Ron O Dror, Stefano Ermon, and Jure Leskovec. Geometric latent diffusion models for 3d molecule generation. In �� , pages 38592–38610. PMLR, 2023

2023
[14]

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. �� , 2024

2024
[15]

Quantum chemistry structures and properties of 134 kilo molecules

Raghunathan Ramakrishnan, Pavlo O Dral, Matthias Rupp, and O Anatole V on Lilienfeld. Quantum chemistry structures and properties of 134 kilo molecules. �� , 1(1):1–7, 2014

2014
[16]

Geom, energy-annotated molecular conformations for property prediction and molecular generation

Simon Axelrod and Rafael Gomez-Bombarelli. Geom, energy-annotated molecular conformations for property prediction and molecular generation. �� , 9(1):185, 2022

2022
[17]

Automatic chemical design using a data-driven continuous representation of molecules

Rafael Gómez-Bombarelli, Jennifer N Wei, David Duvenaud, José Miguel Hernández-Lobato, Benjamín Sánchez-Lengeling, Dennis Sheberla, Jorge Aguilera-Iparraguirre, Timothy D Hirzel, Ryan P Adams, and Alán Aspuru-Guzik. Automatic chemical design using a data-driven continuous repr...

2018
[18]

Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules

Niklas Gebauer, Michael Gastegger, and Kristof Schütt. Symmetry-adapted generation of 3d point sets for the targeted discovery of molecules. �� , 32, 2019

2019
[19]

Symphony: Symmetry-equivariant point- centered spherical harmonics for molecule generation

Ameya Daigavane, Song Kim, Mario Geiger, and Tess Smidt. Symphony: Symmetry-equivariant point- centered spherical harmonics for molecule generation. �� , 2023. 10

2023
[20]

Equivariant neural diffusion for molecule generation

François Cornet, Grigory Bartosh, Mikkel N Schmidt, and Christian A Naesseth. Equivariant neural diffusion for molecule generation. In �� , 2024

2024
[21]

Equivariant 3d-conditional diffusion model for molecular linker design

Ilia Igashov, Hannes Stärk, Clément Vignac, Arne Schneuing, Victor Garcia Satorras, Pascal Frossard, Max Welling, Michael Bronstein, and Bruno Correia. Equivariant 3d-conditional diffusion model for molecular linker design. �� , pages 1–11, 2024

2024
[22]

A survey of large language models

Wayne Xin Zhao, Kun Zhou, Junyi Li, Tianyi Tang, Xiaolei Wang, Yupeng Hou, Yingqian Min, Be- ichen Zhang, Junjie Zhang, Zican Dong, et al. A survey of large language models. �� , 2023

2023
[23]

Training diffusion models with reinforcement learning

Kevin Black, Michael Janner, Yilun Du, Ilya Kostrikov, and Sergey Levine. Training diffusion models with reinforcement learning. �� , 2023

2023
[24]

Reinforcement learning for fine-tuning text-to-image diffusion models

Ying Fan, Olivia Watkins, Yuqing Du, Hao Liu, Moonkyung Ryu, Craig Boutilier, Pieter Abbeel, Moham- mad Ghavamzadeh, Kangwook Lee, and Kimin Lee. Reinforcement learning for fine-tuning text-to-image diffusion models. �� , 36, 2024

2024
[25]

A multi-composition reinforcement learning framework for isomer discovery in 3d

Bjarke Hastrup and François Cornet Tejs Vegge Arghya Bhowmik. A multi-composition reinforcement learning framework for isomer discovery in 3d. �� , 2024

2024
[26]

Graph diffusion policy optimization

Yijing Liu, Chao Du, Tianyu Pang, Chongxuan Li, Min Lin, and Wei Chen. Graph diffusion policy optimization. �� , 2024

2024
[27]

Gfn2-xtb—an accurate and broadly parametrized self-consistent tight-binding quantum chemical method with multipole electrostatics and density-dependent dispersion contributions

Christoph Bannwarth, Sebastian Ehlert, and Stefan Grimme. Gfn2-xtb—an accurate and broadly parametrized self-consistent tight-binding quantum chemical method with multipole electrostatics and density-dependent dispersion contributions. �� ...

2019
[28]

Diffusion-based molecule generation with informative prior bridges

Lemeng Wu, Chengyue Gong, Xingchao Liu, Mao Ye, and Qiang Liu. Diffusion-based molecule generation with informative prior bridges. �� , 35:36533–36545, 2022

2022
[29]

Unified generative modeling of 3d molecules via bayesian flow networks.�� , 2024

Yuxuan Song, Jingjing Gong, Yanru Qu, Hao Zhou, Mingyue Zheng, Jingjing Liu, and Wei-Ying Ma. Unified generative modeling of 3d molecules via bayesian flow networks.�� , 2024

2024
[30]

Cormorant: Covariant molecular neural networks

Brandon Anderson, Truong Son Hy, and Risi Kondor. Cormorant: Covariant molecular neural networks. �� , 32, 2019

2019
[31]

E (n) equivariant normalizing flows

Victor Garcia Satorras, Emiel Hoogeboom, Fabian Fuchs, Ingmar Posner, and Max Welling. E (n) equivariant normalizing flows. �� , 34:4181–4192, 2021

2021
[32]

Top-n: Equivariant set and graph generation without exchangeability

Clément Vignac and Pascal Frossard. Top-n: Equivariant set and graph generation without exchangeability. In �� , 2022. URL �� . 11 A Extended experiments and analysis A.1 Conditional molecule generation on QM9 In this section, we inves...

2022
[33]

This captures both the intermediate states �t and the final molecular structure ��

Sample Trajectories: The pre-trained model �θold is used to generate � molecular trajec- tories by denoising latent variables over � timesteps. This captures both the intermediate states �t and the final molecular structure ��
[34]

These values serve as scalar rewards ��

Calculate Rewards : The generated molecules �� are evaluated using physically grounded reward functions, such as DFT- or xTB-based force deviation, or valency-based stability. These values serve as scalar rewards ��
[35]

The importance sampling ratio �k t �� is computed using log- likelihood scores from Section 4.5

Fine-tune with RL: For each trajectory �, the reward �� k� �k� is normalized to obtain an advantage estimate �� k t . The importance sampling ratio �k t �� is computed using log- likelihood scores from Section 4.5. A PPO-style clipped policy objective is optimized to update ...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.

[1] [1]

Denoising diffusion probabilistic models

Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffusion probabilistic models. �� , 33:6840–6851, 2020

2020

[2] [2]

Generative adversarial networks

Ian Goodfellow, Jean Pouget-Abadie, Mehdi Mirza, Bing Xu, David Warde-Farley, Sherjil Ozair, Aaron Courville, and Yoshua Bengio. Generative adversarial networks. �� , 63(11): 139–144, 2020

work page 2020

[3] [3]

An introduction to variational autoencoders

Diederik P Kingma, Max Welling, et al. An introduction to variational autoencoders. �� , 12(4):307–392, 2019

work page 2019

[4] [4]

E (n) equivariant graph neural networks

Vıctor Garcia Satorras, Emiel Hoogeboom, and Max Welling. E (n) equivariant graph neural networks. In �� , pages 9323–9332. PMLR, 2021

work page 2021

[5] [5]

Equiformer: Equivariant graph attention transformer for 3d atomistic graphs

Yi-Lun Liao and Tess Smidt. Equiformer: Equivariant graph attention transformer for 3d atomistic graphs. �� , 2022

work page 2022

[6] [6]

Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds

Nathaniel Thomas, Tess Smidt, Steven Kearnes, Lusann Yang, Li Li, Kai Kohlhoff, and Patrick Riley. Tensor field networks: Rotation-and translation-equivariant neural networks for 3d point clouds. �� , 2018

work page 2018

[7] [7]

Geodiff: A geometric diffusion model for molecular conformation generation

Minkai Xu, Lantao Yu, Yang Song, Chence Shi, Stefano Ermon, and Jian Tang. Geodiff: A geometric diffusion model for molecular conformation generation. �� , 2022

work page 2022

[8] [8]

Torsional diffusion for molecular conformer generation

Bowen Jing, Gabriele Corso, Jeffrey Chang, Regina Barzilay, and Tommi Jaakkola. Torsional diffusion for molecular conformer generation. �� , 35:24240–24253, 2022

work page 2022