acceptodds
Withdrawn from ICLR 2027

Benchmarking Flow-Based Models for Inverse Design Problems

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Inverse design problems are pervasive across science and engineering, including inverse kinematics, materials design, and drug design. Additionally, ill-posed inverse design problems often admit multiple design solutions that produce similar target responses, causing traditional deep learning methods to fail and motivating the development of specialized deep inverse models (DIMs). Recent flow-based models, particularly Flow Matching (FM), have achieved state-of-the-art performance across diverse generative modeling tasks and have recently been applied to specific inverse-design applications, yet their behavior relative to existing DIMs remains poorly characterized. In this work, we systematically benchmark several recent flow-based inverse models, including single-step variants, against competitive DIMs across a diverse collection of publicly available inverse-design benchmarks. We characterize each method by both inference speed and the accuracy of the inverse designs it produces. Our results reveal a pronounced accuracy-speed tradeoff among existing DIMs, with flow-based models frequently achieving competitive performance and occupying portions of the Pareto-optimal frontier. We further find that the relative advantages of different flow-based approaches vary substantially across tasks, highlighting the importance of considering both solution quality and inference cost when selecting inverse models.

Primary area:
generative models
Keywords:
inverse design, flow matching, generative models, deep inverse models, probabilistic inverse modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_Gv6QOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

The paper under review presents a benchmark for inverse design tasks and elaborate on the methodology of evaluation. The inverse design algorithms are invertible neural networks that have recently been published. The evaluation protocol unifies performance assessment for data driven inverse design over six data sets retrieved from recent publications. Variations in the amount of data and generated designs are presented. The multi-criteria evaluation contains inference speed along with the precision of predivtions.

Critical strengths

The inverse design problem is well picked and of broad interest. Working towards a unified evaluation protocol for such tasks is interesting.

Critical weaknesses

My main objection against the paper is its very limited contribution and zero novelty. Essentially, the paper only collects data sets and models from the literature and runs some evaluation. There are no deep thoughts how to expand the evaluation protocol beyond this. There are no real mistakes of flaws in this paper and the intention is ok, but just running some evaluation protocol with a few models and a few data set is very clearly miles away from what should be expected from an ICLR publication.

Also, one of the two metrics employed - the inference speed - is largely irrelevant as all these models are so small that one can use all of them in interactive mode. There is therefore not much value in this metric.

Looking into the contribution sections as written by the author, only the first point is a very limited real contribution. Already the second bullet is not of interest and the argument that the iMF is very recent and therefore has not been tested before can't really count. The third novelty is plotting a Pareto plot with - as I discussed above - with one largely irrelevant metric.

To be honest, I don't understand why this paper has been submitted here. This leaves me with high perplexity.

Questions for authors

I see no real way how my opinion on this paper can be changed, therefore I don't have any question.

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

This paper aims to benchmark flow-based generative models (e.g., Flow Matching and improved MeanFlow) for amortized inverse design problems through conditional generation. The evaluation uses six existing benchmarks with low-dimensional design parameters to compare best-of-(T) accuracy and generation speed. It reports competitive accuracy for Flow Matching on several tasks and the fastest generation for improved MeanFlow, with relative performance depending on the task and proposal budget.

Critical strengths

None

Critical weaknesses

W1. The first major issue is the limited contribution and narrow scope. The authors choose “generative models” as the primary area for this work, but provide no methodological contribution. Even from the perspective of "datasets and benchmarks", the paper provides no new dataset, and the comparison is very weak. The experimental evaluation is much weaker than that of many methodological papers.

W2. The second major issue is the insufficiently reliable comparison. Only FM, iMF, and cDDPM receive fresh hyperparameter searches, while cINN and MDN use existing configurations without retuning. Also, no independent multi-seed results or uncertainty estimates are reported. I find that the main conclusions provide little new understanding e.g. one-step models are faster, and additional integration steps often improve accuracy.

Questions for authors

Q1 What is the “NA” model in Figure 3, and why is it absent from the other tables and plots?

Q2. I am skeptical that whether best-of-(T) selection can reliably compensate for inaccurate generation, particularly in high-dimensional settings (curse of dimensionality) . Given that the study focuses on small-scale inverse design, the authors should clarify its applicability and limitations in larger parameter spaces.

Q3 As a benchmark paper, I am surprised of the absence of independent multi-seed results and uncertainty quantification.

Q4 Presentation suggestion: The overall readability could be improved. Several parts could be significantly compressed or moved to the appendix, such as the detailed description of the well-known RK4 method in Equation 4. Figure 2 occupies almost an entire page despite its relatively low information density; a more compact layout would improve the presentation.

Reviewer_DPy3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper benchmarks flow-based generative models for inverse design: finding design parameters that produce a desired response. It compares these models with established inverse-design methods on six benchmarks, examining accuracy, inference time, and the effect of generating multiple candidate designs.

Critical strengths
  1. Relevant problem. Inverse design is an important engineering task, where finding parameters that achieve a desired response can require expensive search.
  2. Multiple benchmark tasks. The evaluation covers six problems spanning analytical functions, kinematics, and optical design.
  3. Use of established benchmarks. Reusing public tasks and existing benchmark infrastructure facilitates comparison with prior work.
Critical weaknesses
  1. The novelty of the evaluation perspective needs clearer justification. Earlier inverse-design benchmarks already examined accuracy across proposal budgets and computation time. The paper should distinguish its contribution more precisely from these studies.
  2. Coverage of relevant competitors is incomplete. Comparisons with methods such as Mixture Manifold Networks and recent inverse-specific flow approaches would help establish whether the findings extend beyond the selected baselines.
  3. The design spaces are relatively low-dimensional. All six tasks involve only 2–14 design variables. Although these problems can still be challenging, the experiments provide limited evidence about scaling to substantially larger or different design representations.
Questions for authors
  1. What specifically distinguishes the proposed evaluation from earlier benchmarks that studied inverse-design accuracy over time and across proposal budgets?
  2. What criteria guided the selection of competing methods? Could the authors discuss how including Mixture Manifold Networks or recent inverse-specific flow models might affect the conclusions?
  3. Which conclusions should be restricted to the evaluated low-dimensional settings, and what evidence, if any, supports their generalization to larger or different design representations?

Comments on the paper

1 comment

Dear Reviewers,

We want to thank you for the time and care you put into reviewing our work and for giving us your feedback. We have decided to withdraw this submission and focus on strengthening our paper. Your comments and suggestions will help guide our revisions in order to prepare a future submission.

We appreciate your thoughtful engagement with our work.

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

Related papers

Discussion (0)

Sign in to comment.