acceptodds
Under review as a conference paper at ICLR 2027

AtomSteer: Test-Time Constraint Steering for Atomistic Generative Models

19 views

Abstract

Diffusion and flow matching models offer scalable approaches to generating molecular conformers, protein-ligand complexes, and crystalline materials. However, satisfying structural geometry, accurate stereochemistry, and crystallographic symmetry remains challenging. Satisfying these requirements calls for corrections that also preserve the geometry already captured by the pretrained model. In this work, we propose AtomSteer, a test-time inference framework for steering pretrained atomistic generative models with structural constraints. During sampling, AtomSteer corrects intermediate structure predictions with bounded local projections, graph-rigid stereochemical transformations, and projections over periodic symmetries, accepting updates that improve constraint satisfaction while limiting changes to the rest of the structure. We demonstrate AtomSteer across three pretrained backbones, ET-Flow, Boltz-2, and FlowMM, spanning molecular conformer generation, protein-ligand complex structure prediction, and de novo crystal generation and crystal structure prediction. AtomSteer improves molecular validity and E/Z stereochemical accuracy while retaining conformer ensemble fidelity. For protein-ligand complexes, AtomSteer reduces ligand E/Z errors relative to experimental reference structures while achieving competitive docking success and PoseBusters validity. AtomSteer improves space-group distribution coverage in de novo crystal generation and improves target space-group recovery in crystal structure prediction. Our method provides a unified framework for adding chemical and crystallographic control to existing atomistic generative models without retraining or architectural changes.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Atomistic generative models, test-time steering, constrained generation, diffusion models, flow matching, molecular generation, protein–ligand structure prediction, materials generation

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_zmpQOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces AtomSteer, a test-time framework for enforcing structural constraints in pretrained atomistic generative models without retraining. It corrects intermediate endpoint predictions through structure-dependent projections and incorporates accepted corrections into the original sampling process. Experiments on molecular conformers, protein–ligand complexes, and crystals demonstrate improvements in geometric validity, stereochemical accuracy, and crystallographic constraint satisfaction.

Critical strengths
  1. The framework enables structural constraint steering without retraining or architectural changes, making it applicable to existing pretrained diffusion and flow models.

  2. The structure-dependent projections are well motivated: local coordinate corrections, rigid-fragment rotations, and rigid-body transformations address different violations while protecting surrounding geometry.

  3. Experiments across molecular conformers, protein–ligand complexes, and crystals demonstrate broad applicability, with improvements in geometric validity, stereochemical accuracy, and crystallographic constraint satisfaction.

Critical weaknesses
  1. The experiments do not sufficiently isolate the contribution of intermediate endpoint projection. A controlled comparison using the same projection operators only after generation is needed, together with ablations of the projection schedule, rigid-fragment corrections, and acceptance checks.

  2. The computational cost is insufficiently evaluated. Repeated geometric optimization introduces additional overhead, while the baselines use different particle counts and optimization procedures. Wall-clock measurements and quality–cost comparisons under matched computational budgets are needed to establish practical advantages.

  3. The method relies on numerous task-specific choices, including residual definitions, projection schedules, displacement bounds, and acceptance thresholds. Their sensitivity is not systematically examined, leaving unclear how much tuning is required when transferring the framework to new models or constraints.

Questions for authors

none

Reviewer_qKwLOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
1
Read review
Paper summary

This paper introduces the AtomSteer framework for test time enforcement of structural constraints for pretrained all atom generative models. During sampling, AtomSteer corrects endpoint with several domain-specific geometric correctors and use the corrected endpoint to continue sampling. Empirically, the method improves molecular stereochemistry, physical validity, and crystal symmetry recovery.

Critical strengths
  1. AtomSteer is a practical and general test time framework that can be integrated into different pretrained generative models. The projection designs are intuitive and preserve existing geometry while correcting violations.

  2. AtomSteer has strong empirical improvements on EZBench and on composition-only MP-20, and the improvements are consistent across different backbone models.

Critical weaknesses

The paper could include additional ablations. The paper does not compare the benefit of performing projections during sampling against a simple baseline that applies the same projections only to the final generated structures. Such a comparison would help determine the benefit of incorporating corrections into the generative trajectory over simple post hoc correction.

In addition, the current geometric correctors require substantial task-specific engineering. Could the authors ablate the contributions of individual components and report their computational efficiency? Would the proposed framework be more computationally efficient than unconstrained sampling followed by post hoc correction or filtering? It would also be helpful if the authors could provide a more general recipe for designing these correctors, or explain the reasoning behind the domain-specific design choices.

Questions for authors

See Critical Weaknesses.

Reviewer_3nxwOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper presents AtomSteer, a training-free framework that steers atomistic generative models through bounded, structure-dependent endpoint projections during sampling. Experiments on ET-Flow, Boltz-2, and FlowMM demonstrate improvements in molecular validity, double-bond geometry, and crystal symmetry recovery while largely preserving conformer coverage and docking accuracy. Overall, the approach is well motivated and has clear practical value.

Critical strengths
  1. The paper addresses an important problem in atomistic generation. The motivation for structure-dependent projections is clear: different structural constraints require different coordinated movements to correct violations while preserving the surrounding geometry.
  2. The method achieves substantial improvements on several key metrics, particularly those related to double-bond geometry, physical validity, and crystal symmetry.
  3. Evaluation across three pretrained models and multiple tasks demonstrates the practical value and broad applicability of the framework.
Critical weaknesses
  1. The current experiments primarily evaluate the full method. Key controlled comparisons are missing for intermediate projections, terminal refinement, and structure-dependent projection operators, leaving the individual benefits of these design choices unclear.
  2. The paper does not report computational costs, making it difficult to assess the additional overhead required to obtain the reported improvements.
  3. The novelty of the proposed method is kind of lukewarm. The author is suggested to further clarify the contribution compared with PCFM.
Questions for authors
  1. Within AtomSteer, how much additional benefit do rigid-fragment and rigid-body projections provide beyond Cartesian ChemGN? Component ablations would help isolate their individual contributions.
  2. Could the authors report inference times on representative inputs for all three base models, or at least one representative model, together with the additional overhead relative to the original sampling procedure under matched settings?
  3. Among targets successfully docked by each baseline (Vanilla and Boltz-steering), what fraction fail with AtomSteer, and how many of these cases also show improved constraint satisfaction?

Comments on the paper

1 comment
AuthorsOpenReview ↗

Final Remarks by the Authors

Dear Area Chair, Program Chairs, and Reviewers,

As the discussion period ends, we thank the reviewers for constructive engagement. Below, we summarize AtomSteer's contributions and the discussion period's key outcomes.

AtomSteer is a test-time framework that steers pretrained atomistic diffusion and flow matching models with structural constraints, without retraining. Each constraint is a residual on the atomistic graph, paired with a structure-dependent projection that moves only the atoms needed to reduce the violation. For example, AtomSteer corrects E/Z stereochemistry by rotating the whole substituent, resolves ligand-protein clashes, and enforces space group symmetry in periodic crystals.

The same framework improves four pretrained models:

  • Protein-ligand complexes (Boltz-2). Complexes with top-ranked pose docked within 2 Å RMSD, physically valid, and stereochemically correct (Success@1+stereo) rises from 36.5% with Boltz-steering to 45.2% with AtomSteer on 1,103 targets.
  • Molecular conformers (ET-Flow). The share of physically valid conformers (PB-valid) rises to 96.93%, and the error in E/Z geometry (E/Z W1) falls from 8.94° to 2.15°, while retaining distribution coverage.
  • Crystals (FlowMM). In crystal structure prediction, the share of first predictions that match both the structure and the space group (Joint@1) rises from 45.18% to 52.54%. Separately, in lattice motif generation, more crystals contain the requested motif and pass stability prescreening (joint rate) than with SCIGEN on all three motifs.
  • Molecular crystals (CLARI, added in the rebuttal). Applied without retuning to a new flow matching backbone and domain, AtomSteer's ET-Flow settings transfer and raise the share of physically valid structures (joint physical validity) from 67.8% to 85.3%.

All three reviewers recognized AtomSteer's broad applicability, and both reviewers who engaged with our rebuttal raised their ratings to 3.

Reviewer 3nxw found that AtomSteer addresses "an important problem in atomistic generation" and "achieves substantial improvements on several key metrics." Their requests concerned controlled comparisons, computational costs, and the contribution compared with PCFM. Projecting during sampling outperforms applying the same projections only to the final structures (45.2% vs 39.4% Success@1+stereo on Boltz-2), and each projection improves the constraint it targets (Appendices R1 and R2). PCFM projects with a Gauss-Newton step on the full state. Applying the Gauss-Newton control experiment lowers PB-valid on ET-Flow to 66.98%, against 96.93% for AtomSteer (Appendices R2 and R7). Following the reviewer's suggestion, Appendix R3 now describes AtomSteer's efficiency and its comparison with other approaches. The reviewer credited "the detailed comparison between PCFM and AtomSteer" and raised their rating to 3.

Reviewer qKwL described AtomSteer as a "practical and general test time framework" with "strong empirical improvements" that are "consistent across different backbone models." Their requests concerned computational efficiency against unconstrained sampling followed by post hoc correction or filtering, and a more general recipe. Post hoc correction reaches a joint rate of 4.2% on the triangular motif against 30.7% for AtomSteer, and filtering would need 109x and 307x as many samples to match it on the Kagome and triangular motifs (Appendices R1 and R3). Every constraint follows one recipe: a structural residual, structure-dependent projections, bounded corrections, and an acceptance check (Appendix R5). The reviewer has not followed up to our responses but maintains their rating of 3.

Reviewer zmpQ noted that AtomSteer "enables structural constraint steering without retraining or architectural changes," that its "structure-dependent projections are well motivated," and that the experiments "demonstrate broad applicability." Their requests concerned ablations of the projection schedule, wall-clock comparisons, and the tuning required for new models. Correcting every 8th step on ET-Flow retains AtomSteer's gains at 2.0x the time of vanilla sampling (Appendix R3). The transfer to a new model on molecular crystals, CLARI, required no retuning (Appendix R5). Following the discussion, the reviewer raised their rating to 3.

The revised PDF marks changes in blue, where Appendices R1-R7 hold the new experiments, a comparison with related methods (R6), and the comparison to PCFM (R7).

We appreciate the reviewers' thoughtful engagement, which has helped clarify and strengthen the paper's central contribution: AtomSteer is a general test-time framework that pairs each structural constraint with a structure-dependent projection, across molecules, protein-ligand complexes, inorganic crystals, and now molecular crystals. We will incorporate all promised revisions into the final version.

Thank you very much!

Warm regards,

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.