acceptodds
Under review as a conference paper at ICLR 2027

MolGAN: Few-Step 3D Molecule Generation with Hierarchical Reasoning

Abstract

Diffusion models generate high-quality three-dimensional molecules but typically require long iterative sampling trajectories. We introduce MolGAN, a few-step molecular generator that combines energy-based adversarial training with the multi-timescale computation of the Hierarchical Reasoning Model. This organization allows substantial latent computation without exposing a long sequence of molecular states. On QM9 and GEOM-DRUG, MolGAN achieves state-of-the-art few-step performance across molecular-quality and sampling-efficiency metrics. Its learned energy model additionally enables test-time scaling, improving molecular-property scores on GEOM-DRUG as the compute budget increases.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
3D molecule generation, few-step generation, hierarchical reasoning models, generative adversarial networks, test-time scaling

Reviews

Among:
Shape:
1234this paper, Bottom 9%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.50 over 4 reviews, above 9% of the others.
Reviewer_7LnaOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MolGAN, a teacher-free adversarial model for 3D molecular generation using a small number of external refinement passes. The generator adapts the Hierarchical Reasoning Model with recurrent high- and low-level latent states to iteratively refine atom types and coordinates. Training combines a relativistic energy-based adversarial objective, persistent carries, stochastic refinement with truncated gradients, VISReg, and a learned PAD atom for varying molecule size. On QM9 and GEOM-DRUG, MolGAN reports strong validity and stability with four external passes, and its discriminator is further used for candidate selection and trajectory-based early stopping. Lower-energy candidate selection improves several reported molecular-property scores. Overall, the results suggest that adversarial hierarchical refinement can generate valid molecules with short external trajectories. However, the claimed quality–efficiency advantage is not yet established once internal recurrent computation and the substantially lower QM9 diversity are considered.

Critical strengths
  1. A promising teacher-free approach to short-trajectory molecular generation. The paper offers an alternative to diffusion distillation by combining hierarchical recurrent refinement with adversarial training. Joint atom/coordinate prediction and the learned PAD category provide a flexible mechanism for molecule generation without requiring a pretrained diffusion teacher.

  2. Encouraging validity and stability across two datasets. At four external passes, MolGAN reports competitive stability and validity on QM9 and strong atom stability and validity on GEOM-DRUG. The discriminator-based candidate-selection experiment is also interesting, as larger candidate pools improve QED, normalized SA, and Lipinski scores.

Critical weaknesses
  1. The quality–efficiency claim lacks comparable compute accounting and is weakened by low diversity. A MolGAN “NFE” contains multiple internal HRM module evaluations, each involving Transformer computation. Therefore, external pass count alone is not a comparable measure of computational cost. The paper does not report the relevant recurrence counts, model size, latency, throughput, or memory, so Figure 2 demonstrates efficiency in terms of external trajectory length rather than actual computation. The QM9 diversity result is also concerning: four-pass MolGAN achieves only 34.71% valid-and-unique samples versus 72.70% for four-step FlashMol, despite similar validity. This substantially reduces its practical value for virtual screening and molecular exploration. A stronger efficiency comparison should therefore measure the cost of obtaining a fixed number of valid unique molecules.

  2. The proposed mechanisms and generated 3D distribution are insufficiently validated. MolGAN combines hierarchical recurrence, persistent carries, truncated gradients, coordinate rescaling, PAD atoms, and VISReg, but quantitative ablations isolating these components are largely absent. In particular, there is no comparison against a conventional recurrent Transformer under matched parameter and compute budgets. The molecular evaluation is also incomplete. GEOM-DRUG lacks uniqueness, novelty, and coverage metrics, while the 3D evaluation does not directly examine bond-angle or torsion distributions, steric clashes, or conformational energies. Candidate selection improves several scalar property metrics but also increases the PAINS rate, and its effects on diversity and 3D quality are not reported. Thus, it remains unclear whether selection broadly improves molecular quality or concentrates generation into a narrower region of chemical space.

Questions for authors
  1. Compute and efficiency. What are the values of the internal recurrence parameters, Transformer depth, and generator parameter count? Can the authors report latency and throughput against the main few-step baseline on identical hardware, including the cost of generating a fixed number of valid unique molecules?

  2. Diversity and evaluation protocol. How are uniqueness and (V\times U) computed, including sample count, molecular representation, and stereochemical conventions? Were all baselines evaluated under the same protocol? Please also report uniqueness and novelty on GEOM-DRUG and discuss the low QM9 diversity.

  3. Contribution of hierarchical computation. How does MolGAN compare with a conventional recurrent Transformer under matched parameters and computation? Please provide controlled ablations for persistent carries, truncated gradients, coordinate scaling, and VISReg, ideally with variation across independent runs.

  4. Reproducibility and discriminator state. Please specify the main training hyperparameters, coordinate scaling, PAD sampling, dataset splits, and molecule-reconstruction procedure. How are generator and energy-model carries initialized and reused at evaluation time? In particular, are VISReg views and candidate energies evaluated under identical carry states?

  5. Test-time selection and early stopping. Can the authors provide matched random-selection results for each candidate-pool size, with uncertainty, uniqueness, molecular size, and geometric-quality metrics? How should the increased PAINS rate be interpreted? For early stopping, what is the net computational saving after including discriminator evaluations and the extra pass required to detect an energy increase?

Reviewer_G7xcOpenReview ↗
Rating
2
Confidence
4
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes MolGAN, which combines HRM with adversarial training for 3D molecular generation. MolGAN uses multiple low- and high-level updates within each refinement pass to generate qualified molecules with a short sampling trajectory. The learned energy model also supports candidate selection at test time, improving QED, normalized SA, and Lipinski scores on GEOM-DRUG.

Critical strengths
  • Adapting HRM to 3D molecular generation is an interesting direction, given its promising performance on structured reasoning tasks in long CoT problems such as Sudoku and optimal path finding. This architecture appears well suited to exploring efficient generation without without relying on a long sampling trajectory.
  • The energy model enables candidate selection without additional parameter updates, demonstrating test-time improvements in several molecular properties (QED, normalized SA, and Lipinski scores).
Critical weaknesses
  • A single NFE includes multiple low and high-level HRM forward passes, thus it is difficult to claim that MolGAN enables few-shot generation. The authors should report the actual values of n and m used in their experiments. Further, reporting the wall-clock time analysis compared to baselines would strengthen their contribution.
  • The manuscript's organization and experimental description need improvement. The Sec. 1 provides a limited overview of proposed approach. Moving Sec. 2.2 to the last paragraph of Sec. 1 and elaborating more on the motivation for HRM would improve clarity. Moreover, the essential experimental setup details are missing, e.g., m and n, number of layers in generator and discriminator, training configurations, and hardware. This hinder reproducibility of the work.
Questions for authors
  • Could the authors explain why the performance of V ×\times U highly decreases compared to the baselines (72.70% for FlashMol vs. 34.71%)? Similarly, additional interpretations on why test-time scaling fails to improve PAINS would help understanding the tradeoff of MolGAN.
  • What motivated the choice of GAN over diffusion or flow-matching objectives? Teacher-free training can also be achieved using the latter frameworks.
Reviewer_R8zwOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper proposes MolGAN, a teacher-free 3D molecular generator combining adversarial training with a Hierarchical Reasoning Model (HRM). The generator uses recurrent high- and low-level Transformer states to refine atom types and coordinates, while a discriminator trains the generator and ranks samples at inference. Experiments on QM9 and GEOM-DRUG report competitive stability and validity with few refinement passes. On GEOM-DRUG, selecting low-energy molecules from larger candidate pools improves average QED, normalized synthetic accessibility and the number of Lipinski rules satisfied.

Critical strengths
  1. A teacher-free approach to few-pass generation. The paper combines HRM-style recurrent computation with adversarial training and produces molecules without distilling a pretrained diffusion teacher. Applying this architecture to 3D molecular generation is a reasonable idea.

  2. Promising stability and validity at a short external trajectory length. In the reported comparison with four-step FlashMol, MolGAN improves QM9 molecular stability and GEOM-DRUG atom stability while maintaining comparable validity. Reporting results across one to four passes also shows how these metrics change with the refinement budget.

Critical weaknesses

1. Applying HRM to molecular generation is not enough on its own; its contribution needs to be demonstrated. HRM is an existing architecture, and iterative refinement is already familiar in molecular generation (e.g., PocketXMol uses diffusion-based iterative refinement). The paper shows that the architecture works as a generator, but not what its hierarchical, multi-timescale structure adds: there is no comparison with a single recurrent Transformer, a non-hierarchical network of comparable compute, or a simpler recursive model such as TRM, and the effect of persistent carries is not isolated. In addition, an equivariant Transformer would arguably be more natural for 3D coordinates. HRM itself is not equivariant, so adopting it as is is understandable, but it leaves the design incomplete. At the very least, the paper should explain why an equivariant variant was not used (for example, if it was tried and did not work), and otherwise discuss the choice thoroughly.

2. The experimental evidence does not support the claimed few-step quality–efficiency advantage. First, the reported NFE comparison does not establish a quality–efficiency frontier. NFE is a reasonable proxy for samplers that share a backbone (e.g., those built on GeoLDM), but here one NFE is a full refinement pass containing n(m+1)n(m+1) evaluations of MolGAN's own Transformer stacks. The paper gives neither nn and mm nor stack sizes, parameter counts, latency or FLOPs, yet Figure 2 and the abstract present the result as a quality–efficiency frontier. Wall-clock time or throughput on the same hardware is needed, including candidate generation and discriminator costs for energy-based selection. Second, the low valid-and-unique rate calls the practical usefulness of the generator into question. At four passes, QM9 valid-and-unique is 34.71%, less than half of FlashMol's 72.70%, despite MolGAN's higher validity. Validity and stability alone do not show that the model covers a useful distribution. This is not an assertion that MolGAN copies its training set, but the low diversity makes a distribution-level evaluation necessary: FCD, together with uniqueness and novelty on both datasets. The paper also states that VISReg increases uniqueness and valid-and-unique by approximately 20%, but gives no ablation table or precise definition; the values before and after VISReg are needed. Third, the GEOM-DRUG and Table 2 evaluations are limited. For GEOM-DRUG, Table 1 reports only atom stability and validity, with no molecular stability, diversity or novelty. In Table 2, QED, SA and Lipinski scores improve with energy selection, but PAINS rises from 0.59% to 0.89% and mean LogP drops from 3.03 to 2.02. Without reference values from the real data or results from other generators, it is hard to tell how good these results actually are.

3. The choice to omit explicit bond modeling needs justification. Established methods such as MiDi and SemlaFlow jointly model connectivity and 3D structure and report strong results. These methods are not slow to train, so limited computational resources do not explain excluding this family. The paper should discuss the trade-off and compare against representative graph-and-geometry generators on quality and cost.

Overall assessment. I lean reject. The paper shows that an HRM-based GAN can generate molecules, with encouraging stability and validity at a short external trajectory length. However, it does not establish what HRM adds, the efficiency claim lacks a meaningful cost comparison, and the low valid-and-unique rate raises a serious question about the usefulness of the generated distribution. The omission of explicit bond-generating baselines and limited discussion of equivariance further weaken the contribution.

Minor presentation points:

  • Figure 3 is difficult to read: the molecular structures are too small to make out, so the figure gives little insight into what refinement changes.
  • Table 1 does not highlight the best values, although its caption claims state-of-the-art few-step performance; marking the best values under a clearly stated comparison would make it easier to read.
Questions for authors
  1. Cost. What are nn and mm, the depth and parameter count of each stack, the main training settings, and the sampling latency or throughput of MolGAN and FlashMol on the same hardware (including the energy model and additional candidates for selection)?
  2. Contribution of HRM. At comparable parameter and compute budgets, how does the full model compare with a single recurrent Transformer, a non-hierarchical generator, iterative refinement without persistent carries, or TRM? Did the authors try an equivariant adaptation?
  3. Diversity. How are uniqueness and V×U computed, and with how many samples? Please report FCD, uniqueness and novelty on both datasets, and the values with and without VISReg (does "approximately 20%" mean a relative or an absolute increase?). For Table 2, please add reference values from the real data or results from other generators.

Suggestion: As future work, compare with SemlaFlow or another explicit bond-generating model at matched wall-clock budgets, and report molecule-level stability and geometric quality on GEOM-DRUG.

Reviewer_ibGsOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes MolGAN, a GAN-based framework for few-step 3D molecule generation that incorporates hierarchical reasoning through recurrent high- and low-level latent updates. Experiments on QM9 and GEOM-DRUG demonstrate that MolGAN generates molecules in one to four external refinement steps while retaining reasonable chemical validity.

Critical strengths

The paper applies hierarchical reasoning to 3D molecule generation, enabling generation in a small number of external refinement steps while maintaining reasonable chemical validity.

Critical weaknesses

1.The presentation and allocation of content are unbalanced. The Introduction is too brief to adequately motivate the method and clearly articulate its core novelty. Section 2.2 also provides very limited coverage of hierarchical reasoning, citing only a single work.

2.The few-step inference claim requires stronger computational evidence. Each MolGAN step contains multiple low- and high-level updates, yet the actual values of the corresponding hyperparameters, (m) and (n), are not reported. Consequently, fewer external refinement steps do not necessarily imply fewer internal network evaluations or lower computational cost. Wall-clock inference time and computational complexity should be reported to substantiate the efficiency claim.

3.Comparing a GAN-based generator with diffusion models solely by the number of sampling steps does not provide a fair assessment of efficiency, particularly given the internal computation within each MolGAN step. The evaluation should include relevant GAN-based baselines and compare actual inference costs under matched conditions.

4.MolGAN exhibits substantially lower valid-and-unique rates ((V× U)) than competing methods on QM9 in Table 1. Even at NFE = 4, it achieves only 34.71%, compared with 72.70% for four-step FlashMol. This indicates a substantial diversity limitation despite its relatively high validity.

Questions for authors

Could the authors report MolGAN’s wall-clock inference time and compare it with the baselines under the same hardware, batch size, and number of generated molecules, accounting for all internal low- and high-level updates?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.