acceptodds
Under review as a conference paper at ICLR 2027

Evolutionary One-Step Generators: Fast and Diverse Sampling for Discrete Design

Abstract

Several discrete design tasks, such as molecular discovery, require diverse collections of useful candidates at low computational cost. High validity alone does not guarantee a useful candidate library: repeatedly generating the same valid structures leaves few distinct alternatives. Training for both feasibility and diversity is challenging because many relevant criteria can only be evaluated after hard decoding. To address this challenge, we propose EGO (Evolutionary Generators with One-step inference), a framework for training compact generators directly on discrete outputs. The method combines distribution matching with structural constraints and optional diversity or history-dependent rewards, using antithetic low-rank evolution strategies without requiring criterion-specific differentiable surrogates. Once trained, the generator produces the entire graph in a single neural-network evaluation. On molecular generation benchmarks, our compact generator achieves over the valid-and-unique yield per estimated dense operation compared to recent one-step flow-map baselines while retaining high chemical validity. In scaffold completion, EGO achieves an observed speedup over MoLeR in generation to SMILES and produces approximately as many filter-passing proposals within matched time budgets for generation and screening Beyond chemistry, EGO produces as many distinct held-out elite architectures as relaxed gradient training on NAS-Bench-101. The low generation cost may enable real-time candidate generation across discrete design tasks, supporting interactive exploration of constrained design spaces and rapid construction of candidate sets for downstream evaluation.

Primary area:
generative models
Keywords:
Discrete generative modeling, Evolution strategies, One-step generation, Graph generation

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_MxGHOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

A framework for training compact generators on discrete outputs. Samples are scored using various constraints and the resulting is optimized with antithetic evolution strategies. The generator allows for an entire graph to be generated in a single evaluation. The method is tested on standard datasets QM9 and ZINC250K.

Critical strengths

The one-step inference efficiency is shown through the VU yields, and also demonstrates overall training efficiency.

Critical weaknesses

Related works section not included in the main text, which would be useful for more context regarding the methodology and contributions. VU yield performance is mediocre. The strength in training budget and comparison does not compare to the stronger performers for VU yield. In particular, other one-step methods like MoFlow and the autoregressive methods are not mentioned. Further testing in various molecular objectives would be useful for more context regarding the quality of the generated molecules. Figure of methodology is not helpful and provides minimal information regarding the contributions, and overall presentation of the information is extremely dense and difficult to parse through.

Questions for authors

What exactly is fitness shaping for the experiments? What/Why do the molecular experiments use centered ranks? Additional information regarding the why and how to select the fitness would be great. Is there a specific reason related works was pushed to the appendix? It provides a lot of context and hurts the clarity of the methodology.

2 comments

Thank you for highlighting the presentation issues and for the constructive feedback. We agree that the context and supporting evidence needed to be easier to find. We have revised the manuscript (see the revised PDF) accordingly and welcome any further suggestions for improving it.

Our evaluation covers progressively larger molecular graphs: up to 9 heavy atoms on QM9, 38 on ZINC250K, and 88 on GuacaMol in our experiments. Larger graphs require coordinating more atom and bond decisions while satisfying chemical constraints. GuacaMol therefore extends the evaluation to larger molecular graphs. We also evaluate scaffold completion as a constrained molecular application (Appendix E). Beyond chemistry, we test neural architecture generation and examine both the benefits and limitations of the training framework.

1. Related work

We have added a Related Work section immediately after the introduction, while retaining the more detailed background and related work in Appendix B.

2. VU yield performance

Figure 2 and Appendix D compare EGO with autoregressive, iterative, and one-step generators. EGO achieves VU yields of 78.21% on QM9, 90.66% on ZINC250K, and 77.61% on GuacaMol. Its advantage concerns the balance between yield and inference cost. Compared with one-step methods such as CFM-ECLD, EGO achieves approximately 53× and 208× higher VU per estimated dense inference operation on QM9 and ZINC250K, respectively, despite lower absolute VU.

Specifically regarding MoFlow, we thank the reviewer for prompting a closer examination of its evaluation protocol. Its published results include validity correction after graph generation, an aspect that we did not make sufficiently clear in the text. We evaluated public MoFlow checkpoints without this post-hoc validity correction, using five sampling runs of 10,000 proposals per dataset. On QM9, MoFlow achieves 90.19% VU, compared with 78.21% for EGO. On ZINC250K, MoFlow achieves 81.59% VU, compared with 90.66% for EGO. The same pattern holds for validity: 96.04% for MoFlow versus 93.46% for EGO on QM9, and 81.87% versus 94.02% on ZINC250K.

Furthermore, the evaluated MoFlow models contain 2.65 million and 130.35 million parameters on QM9 and ZINC250K, respectively, compared with 1.46 million and 6.32 million for EGO. MoFlow therefore has approximately 1.82× and 20.62× as many parameters.

Accordingly, we have updated the comparison tables (Appendix D.2 and D.3), Figure 2, and the main text to distinguish the correction protocols and make these quality and cost trade-offs more explicit. We welcome further suggestions to improve the clarity and completeness of the comparison.

3. Molecular quality beyond validity

The scaffold completion experiments in Section 5.2 and Appendix E jointly require QED ≥ 0.7, SA ≤ 4, and no additional PAINS alert types relative to the scaffold.

QED, the Quantitative Estimate of Drug-likeness, summarizes physicochemical properties associated with typical oral drugs. SA, Synthetic Accessibility, estimates synthesis difficulty. PAINS (Pan-Assay Interference Compounds) filters flag structural motifs associated with potential assay interference.

At 100,000 proposals, EGO produces 20,245 distinct qualifying completions, compared with 16,684 for MoLeR. These three screening criteria are neither explicitly rewarded during training nor used for checkpoint selection. We consider the ability to satisfy them jointly additional evidence of generation quality beyond the explicit training criteria. These scores could also be incorporated directly into the fitness in future applications.

4. Methodology figure and overall presentation

We have also redrawn Figure 1 from scratch to make the method easier to follow. It now separates training from inference and illustrates the four stages: perturb, generate, score, and update. A revised caption and a short explanation in the main text guide the reader through the method.

5. Fitness shaping

Fitness shaping converts scores into ES update utilities. For the main molecular runs, centered ranks preserve candidate ordering while reducing sensitivity to score magnitudes and outliers. Rarity is ranked separately to control its influence.

A key advantage of our method is its support for non-differentiable criteria on hard outputs. This avoids designing a differentiable surrogate for each chemical constraint or diversity reward. We select components according to the task: distribution matching for reference similarity, chemical penalties for feasibility, and optional rarity for exploration. Appendix C.1 specifies the fixed weights and ranking procedure.

Scaffold completion ranks several components separately (Appendix C.2), whereas NAS uses mean-centered raw fitness without rank shaping (Appendix C.3).

Thank you for your response. I have no further questions.

Reviewer_jBsUOpenReview ↗
Rating
4
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper investigates the problem of one-step generators for discrete structures, which is challenging since many relevant criteria are naturally defined only on the final discrete object. The authors propose EGO, a framework for training one-step discrete generators. Complete hard samples are scored using distribution-matching terms, structural constraints, and optional set- or history-dependent rewards, without requiring these criteria to be differentiable. The resulting objective is optimized using an evolutionary approach, i.e. an antithetic evolution strategy (ES). Comprehensive experiments were performed, evaluating EGO for molecular generation, scaffold completion, and NAS-Bench-101 on three benchmarks, using three generic gradient-based generators and a broad range of generative molecule models as baselines.

Critical strengths
  • Practical training approach for the generation of discrete structures
 As mentioned by the authors, direct generation, distribution matching, and evolution strategies are established ideas, but their contribution is to integrate them in a practical training approach for discrete structures.

  • Comprehensive experimental evaluation
 EGO is evaluated for molecular generation, scaffold completion, and NAS-Bench-101 on three benchmarks (QM9, ZINC250K, and GuacaMol). EGO is compared to three generic gradient-based generators and a broad range of generative molecule models. In summary, Overall, the experiments demonstrate that EGO achieves higher candidate yield under limited training time.

  • Theoretical analyses
 The authors prove theoretical results on finite-budget distinct yield and on the stability under shared Gumbel noise.

Critical weaknesses
  • Old gradient-based baselines
 The gradient-based baselines are from 1992, 2013, and from 2017, resp. Are there no more recent SOTA methods?
Questions for authors
  1. To reduce the memory required by the population, you use rank-one perturbations. What performance would you gain with higher-rank perturbations?

  2. The gradient-based baselines are fairly old (from 1992/2013/2017). Are there not more recent ones?

1 comment

Thank you for your constructive feedback! We have added experiments addressing both questions (see the revised PDF) and would welcome recommendations for other suitable estimators or comparisons.

1. More recent gradient estimators

Following your suggestion, we also tested ReinMax from “Bridging Discrete and Backpropagation: Straight-Through and Beyond” (Liu et al., NeurIPS 2023) using the architecture and 30-minute active training budget of Table 21. Our adaptation retained the baseline optimizer settings and temperature, with a numerical correction to the connectivity surrogate.

All three seeds reached 0% validity and strict VUN at 20 minutes. The two available evaluations at 30 minutes remained at zero. Large but finite gradient spikes were followed by almost deterministic, invalid outputs.

Molecular validity requires coordinated atom and bond choices. ReinMax adjusts the gradient through sampling, but that gradient still comes from our continuous surrogate. Improving the surrogate does not necessarily improve discrete chemical validity. The observed saturation suggests that updates concentrated on invalid structures. This is a possible explanation under our settings, not a general limitation of ReinMax.

2. Higher-rank perturbations

Following your suggestion, we added a dedicated comparison of ranks 1, 2, and 4 using the ZINC250K protocol of Table 21. Architecture and hyperparameters were fixed to those of the Table 21 experiments, with three seeds and 30 minutes of active training. Perturbations were normalized to preserve their variance.

Your question helped us identify a trade-off that depends on the training budget. Strict VUN at 30 minutes was 77.91 ± 7.92%, 77.91 ± 3.36%, and 78.27 ± 4.40% for ranks 1, 2, and 4, respectively, reported as mean ± sample SD. Higher ranks improved early mean yield, but added little at 30 minutes while increasing time per update by approximately 7%.

These results support rank one at this budget, while suggesting potential benefits from higher ranks under shorter budgets. The ablation is now included in the revised appendix (Appendix D.3.5) and referenced in the method section.

Reviewer_f51KOpenReview ↗
Rating
3
Confidence
4
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents EGO, a framework for training one-step generators directly using criteria evaluated on hard discrete outputs. It combines distribution matching, structural constraints, and optional diversity rewards through antithetic low-rank evolution strategies. The resulting generator produces complete candidates in a single network evaluation. Experiments on molecular generation, scaffold completion, and neural architecture generation demonstrate its usefulness for producing distinct qualifying candidates under limited computational budgets.

Critical strengths

Cross-domain validation supports general applicability. Beyond molecular generation and scaffold completion, the paper evaluates neural architecture generation on NAS-Bench-101. This provides meaningful evidence that the framework can accommodate discrete design tasks with different structural constraints and success criteria.

The contribution boundaries are clearly articulated. The authors explicitly acknowledge established ES methods as their optimization foundation. They position their contributions around the hard-output training framework, finite-budget candidate-generation formulation, and reusable one-step generator implementation, rather than presenting existing ES mechanisms as a new optimization algorithm. This distinction is clear and appropriately scoped.

The evaluation is practically motivated and transparent about trade-offs. Measuring distinct qualifying candidates addresses the needs of candidate-library construction more directly than validity alone. The authors also acknowledge trade-offs involving distributional fidelity, average quality, and diversity, which strengthens the interpretation of their results.

Critical weaknesses

The training–inference trade-off needs more complete quantification. ES requires evaluating multiple parameter perturbations. Although matched training-time comparisons are provided, reporting total training costs and the sampling volume needed to offset additional training overhead would clarify practical benefits.

The sources of efficiency gains could be better isolated. Compact MLPs and one-step decoding already offer inexpensive inference. Further comparisons among ES, continuous relaxations, and score-function estimators under matched architectures, model sizes, and training budgets would help establish the training framework’s independent contribution.

Higher distinct-candidate yield does not necessarily imply better overall quality. Relaxed training achieves higher mean accuracy and better distribution matching in the NAS experiments. Yield curves across quality thresholds and larger sampling budgets would better characterize when EGO is preferable.

Questions for authors

Above

1 comment

Thank you for the constructive feedback. Your comments highlighted the need to make evidence easier to find. We have added direct appendix references in the revised main text (see the revised PDF) and clarified our comparisons. We welcome further suggestions to improve clarity.

1. Training costs and sampling volume

Appendix D.5 reports training durations and inference costs separately. Table 27 gives EGO optimizer times of approximately 0.30 hours on QM9 and 2.10 hours on ZINC250K. CFM reports approximately 48 and 144 hours of elapsed training. The table specifies the accounting scope of each measurement.

Table 26 relates valid and unique yield to estimated dense inference work. EGO achieves approximately 53× and 208× higher efficiency than CFM ECLD on QM9 and ZINC250K. These partial operation counts provide computational context, but do not establish a measured end-to-end break-even volume.

Within the controlled comparison of Table 21, population evaluations are included in the shared 30-minute active training budget. At 10,000 evaluation proposals per seed, EGO achieves higher yield without additional active training time to amortize.

2. Sources of efficiency gains

We agree that inexpensive inference comes from the compact architecture and one-step generation. Our controls hold the architecture fixed.

Table 21 also matches model size, GPU, and active training time. Strict valid, unique and novel yield is 77.9% for EGO, 53.7% without rarity, 29.5% for STE, and 5.8% for Soft.

Table 13 compares EGO and RLOO with identical architectures, hard rewards, and feature scales, excluding rarity and rank shaping. At 11.52 million scored graphs per method, strict VUN is 38.4% versus 12.1%. The revised main text directly references these complementary controls in Appendices D.3.4 and D.2.4.

3. Quality and diversity in NAS

We agree that higher distinct yield does not imply better overall quality. Table 2 shows better mean accuracy and distribution matching for Soft. EGO nevertheless returns more distinct held-out architectures meeting the same elite threshold, with 221.7 versus 143.7 per 10,000 proposals.

We have added curves across validation accuracy thresholds and budgets up to 100,000 proposals in Appendix F.1, Figure 9. At the 99th percentile threshold, EGO returns 31.0 distinct qualifying architectures per 10,000 proposals versus 19.3 for Soft. Soft can perform better at smaller budgets. The revised discussion explicitly describes this dependence on the quality threshold and sampling budget.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Summary of Rebuttal Updates and Manuscript Revisions

Dear Chairs and Reviewers,

We thank you all for the feedback on our manuscript. We have uploaded a revised version of the full paper, with changes highlighted in blue. During the rebuttal, we addressed the reviewers’ comments through additional experiments and revisions:

  • Clarified training and inference cost accounting, which was not sufficiently clear, and added direct appendix references to make the controlled comparisons easier to find. (f51K)
  • Added NAS results across quality thresholds and budgets up to 100,000 proposals to examine whether the distinct-yield advantage persists under stricter quality requirements and different sampling budgets. (f51K)
  • Tested an additional gradient estimator, ReinMax, proposed in “Bridging Discrete and Backpropagation: Straight-Through and Beyond” (Liu et al., NeurIPS 2023). We used the architecture and 30-minute active training budget of Table 21 and discussed the results under our settings. (jBsU)
  • Compared perturbation ranks 1, 2, and 4 under the same training-time budget to assess whether higher ranks improve performance, finding early gains but similar final yield. (jBsU)
  • Added a Related Work section in the main text, with an expanded discussion in Appendix B. We also redrew the method figure and clarified fitness shaping and evaluation protocols to make the methodology easier to follow. (MxGH)
  • Evaluated public MoFlow checkpoints without post-hoc validity correction and updated the comparisons to distinguish evaluation protocols and clarify quality and cost trade-offs. (MxGH)

Best, The authors

AuthorsOpenReview ↗

Follow-up on our rebuttal and manuscript revisions

Dear Reviewers,

Thank you again for your constructive feedback, which has helped us clarify several aspects of the paper. As the discussion period draws to a close, we would appreciate hearing whether our responses and manuscript revisions address your concerns, or whether there is anything else we could clarify in the remaining time.

Thank you for your time.

Best, The authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.