A one-step AI generator reported valid-and-unique yields on three molecular benchmarks and more distinct held-out elite architectures than two comparison methods in a separate test. It reported yields of 78.21% on QM9, 90.66% on ZINC250K and 77.61% on GuacaMol, with one network evaluation for each proposal.
The paper asks how a one-step proposal distribution can be shaped directly by criteria evaluated on final hard outputs. That matters when only a finite number of candidates can be proposed: distinct yield depends both on how often proposals succeed and on whether probability is spread across different successful identities.
The method, called EGO, trains one-step discrete generators directly on hard samples, meaning the finished discrete candidates used for evaluation. It combines distribution matching and structural constraints with optional rewards tied to a set of outputs or their history, and uses antithetic low-rank evolution strategies.
Promising yields, with caveats
After 30 minutes of active ZINC250K training, EGO reached 77.91% on the strict valid-and-unique-novel measure. EGO without its rarity reward reached 53.73%, while STE reached 29.50% and Soft 5.78%. The corresponding sample standard deviations across three seeds were 7.92, 3.70, 1.76 and 9.86 percentage points, respectively. STE and Soft completed 5.64 and 6.06 times as many updates as EGO.
Another QM9 comparison, using matched rewards, put EGO at 38.39% on the strict valid-and-unique-novel measure versus 12.05% for RLOO. The paired difference was 26.34 percentage points, with an unadjusted 95% interval from 22.81 to 29.86 points. The recipes still differed in learning rate, perturbation settings and reference reuse.
Against CFM-ECLD, EGO's valid-and-unique yield per estimated dense-projection operation was about 50 times higher on QM9 and 200 times higher on ZINC250K. Its valid-and-unique yield itself was lower by 7.24 percentage points on QM9 and 2.37 points on ZINC250K. These are estimated-operation comparisons under differing source protocols, not matched wall-clock speed tests.
A practical test of scaffold completion
A separate scaffold completion test used 100 held-out queries and 1,000 attempts per query. For a result to count as usable, it had to be chemically valid, connected and contain the supplied scaffold.
At matched generation-and-screening time, EGO screened 1,892.7 proposals per query on average, with a sample standard deviation of 63.7, versus 256.0 for MoLeR, with a standard deviation of 0.0. EGO produced 547.3 filter-passing proposals per query, with a sample standard deviation of 15.9, versus 54.5 for MoLeR, with a standard deviation of 0.4. That was 10.05 times as many. The figures summarize three replicate evaluations of the same 100-query panel and are descriptive rather than a significance test.
At a separate fixed budget of 100,000 proposals, EGO returned 20,245 distinct filter-passing completions, with a sample standard deviation of 755, versus 16,684 for MoLeR, with a standard deviation of 79. The observed generation-to-SMILES time showed a 44.3-fold reduction for EGO. That timing excluded screening, and the methods differed in their training and checkpoint sources.
More variety did not win every test
Outside molecular generation, NAS-Bench-101 offered another test of candidate variety. From 10,000 proposals, EGO returned 221.7 distinct held-out elite architectures, with a sample standard deviation of 7.5. STE returned 138.0, with a standard deviation of 28.9, and Soft 143.7, with a standard deviation of 18.6. That was 1.61 times as many alternatives as STE and 1.54 times as many as Soft.
That advantage did not hold in every test. On binarized MNIST, Soft improved on EGO in every paired seed on the reported distribution and predicted-digit coverage measures. EGO nevertheless recorded 99.98% exact image uniqueness, with a sample standard deviation of 0.03%. The paper cautions that exact image uniqueness does not establish semantic diversity.
The authors say the benefits are task-dependent. A higher distinct-candidate yield can trade off against distributional fidelity, average proposal quality or coverage. Population evaluation raises training cost, and scalability to substantially larger generators has not been established. In molecular tests, passing the filters is not evidence of experimental usefulness.
The paper is a preprint identified as arXiv:2610.08367v1 and dated 6 October 2026. It says a documented code repository will be provided as supplementary material and made public when a public version is released. The supplied declarations do not report a funding source or conflict-of-interest disclosure.
Paper data and sources
Original title: Evolutionary One-Step Generators: Fast and Diverse Sampling for Discrete Design
Authors: Marcus Vukojevic, Erik Nielsen, Veronica Lachi et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-10-06
DOI: Not available
Original paper · Full text