acceptodds
Under review as a conference paper at ICLR 2027

Multi-Objective Generative Optimization via Expected Hypervolume Fine-Tuning

Abstract

Scientific discovery requires plausible designs spanning useful trade-offs among competing properties: a molecule with strong binding affinity, for example, may be difficult to synthesize. Pretrained generative models, e.g., flow and diffusion generators, provide a foundation for proposing such designs, but steering them toward high-quality, complementary trade-offs remains challenging. Inference-time search incurs high per-sample costs, while fine-tuning typically relies on fixed per-candidate rewards that do not capture how designs complement one another. We introduce HVFT, which adapts a single pretrained generator to maximize the expected hypervolume of its samples while controlling departure from the pretrained distribution. Expected hypervolume jointly rewards candidate quality and collective trade-off coverage. Nonetheless, we show that it cannot be expressed as the expectation of a fixed per-candidate reward and therefore cannot be optimized directly with standard techniques. A key observation of this work is that its first variation corresponds to the expected marginal hypervolume contribution of a candidate, which yields a tractable adaptive reward via an unbiased estimator based on model samples. HVFT uses this reward to decompose the objective into a sequence of KL-proximal scalar-reward subproblems. We establish concavity, relative smoothness, and convergence guarantees for the resulting distributional procedure. Across therapeutic-peptide and molecular design tasks spanning discrete diffusion and continuous flow-matching models, HVFT achieves state-of-the-art expected hypervolume and trade-off coverage.

Primary area:
generative models
Keywords:
multi-objective optimization, generative models, finetuning, hypervolume

Reviews

Among:
Shape:
1234this paper, Top 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.50 over 4 reviews, above 99% of the others.
Reviewer_RaPDOpenReview ↗
Rating
3(+1)
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes Hypervolume Fine-Tuning (HVFT), which fine-tunes a pretrained generator to improve the expected hypervolume of the set of independently generated candidates. The paper's central observation is that the first variation of this hypervolume objective can be expressed as an expected marginal hypervolume contribution. The authors estimate this quantity via Monte Carlo, using samples from the current generator. Then they apply Flow Density Control to construct successive scalar-reward fine-tuning subproblems. The paper also derives theoretical properties (e.g., convergence) of the distributional objective and evaluates the method on peptide and small-molecule design tasks using discrete diffusion and continuous flow-matching models.

Critical strengths
  1. The distributional formulation is a simple and intuitive way to value complementary candidates. The marginal-hypervolume reward bridges the full distributional objective to existing scalar-reward fine-tuning methods which is useful and tractable.
  2. The empirical hypervolume improvements are consistent across tested tasks and model families. HVFT also retains more diversity than several fine-tuning baselines.
Critical weaknesses
  1. A weakness regarding evaluation metrics is that the evaluation does not establish candidate-set quality under a fixed computational budget (e.g., fixed number of generation attempts). Hypervolume and design-space metrics are computed from pools of valid designs obtained through repeated generation (L324-327), while validity is reported separately. HVFT's peptide validity is only approximately 24-34%, and HVFT’s molecular results involve mixed trade-offs in validity and coverage. These conditional metrics are informative, but they leave uncertain whether HVFT provides better candidate sets when failed generation attempts consume the same total budget.
  2. Computational advantage is unclear. The motivation emphasizes amortizing the cost of inference-time search, but the end-to-end computational advantage remains unclear. Appendix B.2 reports execution times against fine-tuning counterparts, without quantifying the generation volume needed to recover the training cost relative to inference-time steering. The molecular experiments also lack a training-free search or guidance comparator; DiffusionBlend instead combines separately aligned models at sampling time.
Questions for authors
  1. How do the methods compare when hypervolume is evaluated under a fixed number of generation attempts, counting invalid designs as contributing no dominated volume? Would the principal rankings remain unchanged?
  2. Including training, objective evaluation, and generation costs, at approximately what number of valid outputs does HVFT become cheaper than inference-time steering?
  3. Which changes in the individual task properties account for the hypervolume improvements? Property distributions or success rates under jointly specified thresholds would help establish the practical significance of the gains.
3 comments
AuthorsOpenReview ↗

Response to Reviewer RaPD

Thank you for the thoughtful review and constructive suggestions. We appreciate the recognition that our marginal-hypervolume reward is “useful and tractable” and that the empirical improvements are “consistent across tested tasks and model families.” In response to the reviewer’s requests, we added evaluations accounting for invalid generation attempts, as well as matched time- and sampling-budget comparisons on both evaluated peptide tasks. HVFT retains its advantage in these evaluations. We also provide joint-property plots clarifying the gains.

To keep this response concise while making the additional evidence easy to find, we have collected the supporting tables and plots in Appendix E. We will carefully incorporate the feedback, additional results, and clarifications into the revised manuscript. If the responses address your concerns, we would sincerely appreciate an increase in your scores to reflect this updated assessment.

1. Fixed generation budget

“How do the methods compare when hypervolume is evaluated under a fixed number of generation attempts, counting invalid designs as contributing no dominated volume? Would the principal rankings remain unchanged?”

Response.

Thank you for this helpful suggestion. Our original evaluation compares equal-sized sets of valid designs, reflecting downstream use after validity filtering. Following your suggestion, we additionally evaluated all methods under a fixed generation budget, accounting for unsuccessful attempts. Encouragingly, HVFT retains the highest 100-HV on all four evaluated tasks.

We evaluated each method using 1,000 generation attempts, without replacing invalid outputs. To compute 100-HV, we sampled subsets of 100 attempts and measured the hypervolume contributed by their valid members; invalid outputs contributed no volume.

HVFT achieves 100-HV of 11.04 versus 7.46 for the strongest baseline on GLAST, 10.34 versus 8.11 on GLP-1R, 3.39 versus 3.30 on Topology, and 4.65 versus 4.09 on Geometry. Results are averaged over two seeds. Full results, including standard errors, validity rates, and the remaining metrics, appear in Appendix E.1, Tables 9 and 10.

2. Computational cost and candidate quality

“Including training, objective evaluation, and generation costs, at approximately what number of valid outputs does HVFT become cheaper than inference-time steering?”

Response.

Thank you for this helpful suggestion. In some applications, the goal of fine-tuning is the adapted model itself, rather than a particular set of designs. Nonetheless, we agree that comparing the cost of producing candidate sets provides useful practical evidence. Because a break-even count based only on valid outputs does not account for quality differences, we compare candidate-set quality at equal total computational cost.

In Appendix E.2, we compare HVFT, PepTune, and pretrained-model sampling with greedy hypervolume selection (Base+Greedy) on GLAST and GLP-1R under matched wall-clock budgets (3,000–30,000 seconds) and sampling budgets (10,000–90,000 samples). The wall-clock budgets include training, objective evaluation, and generation until the final candidate set is obtained. HVFT achieves higher 100-HV than both baselines at every evaluated budget on both tasks. Results are averaged over two seeds, with full 100-HV results in Tables 11 and 12.

3. Property improvements and practical significance

“Which changes in the individual task properties account for the hypervolume improvements? Property distributions or success rates under jointly specified thresholds would help establish the practical significance of the gains.”

Response.

Thank you for this helpful suggestion. We added joint-property visualizations in Appendix E.3, Figures 4–8, to clarify which properties and trade-offs drive the hypervolume gains. These reveal notable improvements in non-fouling for peptides and QED for small molecules. We use joint plots rather than separate marginal distributions to show whether desirable property values occur in the same candidates.

Joint-threshold success rates would also be useful given application-specific acceptance criteria. Our benchmarks do not specify established joint acceptance thresholds. Without such criteria, choosing thresholds would introduce an additional preference over trade-offs that could affect the comparison. The joint-property plots provide complementary evidence without imposing a particular acceptance criterion.

I thank the authors for their detailed response and new experiments. Tables 9 and 10 sufficiently address my concern for fixed generation budget, although HVFT shows lower validity and coverage than Base and PepTune on GLP-1R. Tables 11 and 12 also persuasively show that HVFT is effective at matched budget. I also appreciate Figures 4-8, where the joint plots help visualize the practicality of HVFT. I have no further questions, and my concerns are sufficiently addressed in the revision. Hence, I raise my score.

Thank you sincerely for taking the time to review the additional results and for raising your rating. We are glad that the new evidence addresses your questions, and we will incorporate these results and clarifications into the revised manuscript. Thank you again for your contribution to the review process and for helping us strengthen our work.

Reviewer_uemyOpenReview ↗
Rating
4(+1)
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduces Hypervolume Fine-Tuning (HVFT), a fine-tuning approach that uses expected hypervolume as a distributional training objective for multi-objective generation. The goal is a generator whose samples jointly achieve high objective quality and broad trade-off coverage. Because expected hypervolume cannot be written as the expectation of a fixed per-candidate reward, the authors use Flow Density Control (De Santi et al., 2025). They show that the first variation of the objective is the expected marginal hypervolume contribution of a candidate, and use it as the scalar reward in a sequence of KL-proximal fine-tuning subproblems. HVFT is evaluated on therapeutic peptide design and small-molecule design, against state-of-the-art diffusion and flow-matching models. HVFT achieves the highest hypervolume on all tasks and, among the fine-tuned methods, the highest diversity and coverage on most of them.

Critical strengths
  1. The authors propose the first method to directly maximize the expected hypervolume in a fine-tuning scheme for a given base generator (e.g., a diffusion or flow model). Although they apply it to drug design, the approach is applicable beyond this setting and beyond the generative models considered.

  2. Empirically, the authors show that their method performs strongly compared with existing diffusion-based and flow-based MOO methods in two different drug design tasks.

Critical weaknesses
  1. The paper reuses the core ideas of De Santi et al. (2025), which the authors cite, and adapts them to the multi-objective optimization setting by choosing a specific scalar reward function. Thus, the paper's main contribution can be viewed as partly incremental or as a special case of that work, since that work already lays most of the groundwork.

  2. The inference-time guidance methods are compared with the fine-tuning approaches under different computational budgets, and the authors compare against only a few diffusion- and flow-based methods.

Questions for authors
  1. On 100-HV, PepTune (an inference-time guidance approach) is statistically indistinguishable from the base model while the finetuned methods show strong performance in the current setup. However, finetuning a model and then sampling is likely to have a higher total compute budget than inference-time methods. Can you add results with comparable compute budget between the methods and report the metrics?

  2. Can you show how the choice of η\eta (and α\alpha) affects the results?

  3. Can you add non-diffusion multi-objective baselines (e.g., MO-GFlowNets, which are cited, or evolutionary methods, e.g., NSGA-II)? How does HVFT compare to best-of-N from the base model, with Pareto filtering or greedy HV subset selection, at matched total compute?

  4. Can you show objective-space coverage for the non-dominated set for all the methods?

4 comments
AuthorsOpenReview ↗

Response to Reviewer uemy (Part 1 of 2)

Thank you for the thoughtful review and constructive suggestions. We appreciate the recognition that HVFT is the “first method to directly maximize the expected hypervolume in a fine-tuning scheme” and is applicable beyond the tasks and generative models considered. In response to the reviewer’s requests, we added matched-budget comparisons and a regularization sensitivity analysis, and clarify below our contributions beyond Flow Density Control.

To keep this response concise while making the additional evidence easy to find, we have collected the supporting tables and plots in new sections of Appendix E. We will carefully incorporate the feedback and additional results into the revised manuscript. If the responses address your concerns, we would sincerely appreciate an increase in your scores to reflect this updated assessment.

1. Contribution beyond Flow Density Control

“The paper reuses the core ideas of De Santi et al. (2025) … and adapts them to the multi-objective optimization setting by choosing a specific scalar reward function.”

Response.

Thank you for raising this point. Flow Density Control provides the general framework for optimizing nonlinear objectives over generative distributions. Our contribution is to develop expected hypervolume as a tractable training objective that rewards both candidate quality and complementary trade-offs among generated samples.

The technical contribution lies in establishing the properties of this objective and deriving the quantities needed to optimize it within FDC. We show that its first variation equals the expected marginal hypervolume contribution and construct an unbiased sample-based estimator, making the adaptive reward computable from generated samples. We also derive a coverage-probability representation that establishes concavity and prove an explicit relative-smoothness bound supporting the convergence analysis. These HV-specific results connect the proposed objective to an implementable fine-tuning procedure.

The experiments demonstrate the practical value of this formulation. HVFT uses existing fine-tuning methods as inner solvers yet achieves higher hypervolume than their standalone versions across all five tasks. This supports the benefit of adapting the reward to account for complementarity among generated candidates, across both discrete diffusion and continuous flow-matching models.

2. Comparisons under matched computational budgets

“Can you add results with comparable compute budget between the methods and report the metrics?”

Response.

Thank you for this helpful suggestion. We added comparisons against PepTune and pretrained-model sampling with greedy hypervolume selection (Base+Greedy) on GLAST and GLP-1R under matched wall-clock budgets (3,000–30,000 seconds) and sampling budgets (10,000–90,000 samples). The budgets include all computation until the final candidate set is obtained, including training, objective evaluation, and generation.

Encouragingly, HVFT achieves higher 100-HV than both baselines at every evaluated budget on both tasks. Results are averaged over two seeds; full results and evaluation details appear in Appendix E.2, Tables 11 and 12.

3. Sensitivity to regularization

“Can you show how the choice of η (and α) affects the results?”

Response.

Thank you for this helpful suggestion. We added a sensitivity analysis of the local step-size parameter η, while keeping the pretrained-model penalty fixed. Since the local KL penalty is weighted by η⁻¹, larger η permits larger updates through weaker regularization, while smaller η imposes stronger regularization.

We tested five η settings spanning four orders of magnitude on each of GLAST, GLP-1R, Topology, and Geometry. Hypervolume remains competitive across a broad range of settings, indicating that the gains do not depend on a narrowly tuned parameter. Peptide tasks generally favor smaller η values, corresponding to stronger local regularization, whereas molecular tasks perform well with larger η values and weaker local regularization. Full results, including diversity, coverage, validity, and available standard errors, appear in Appendix E.4, Tables 13 and 14.

AuthorsOpenReview ↗

Response to Reviewer uemy (Part 2 of 2)

4. Additional baselines and pretrained-model sampling

“Can you add non-diffusion multi-objective baselines … e.g., NSGA-II? How does HVFT compare to best-of-N from the base model, with Pareto filtering or greedy HV subset selection, at matched total compute?”

Response.

Thank you for these helpful suggestions. We added NSGA-II under a sampling budget matched to HVFT, broadening the comparison beyond diffusion- and flow-based methods. We use the recombination operators of Tripp and Hernández-Lobato (2023) and restrict peptide lengths to those comparable to the base model’s outputs. NSGA-II achieves strong hypervolume performance, but HVFT achieves higher hypervolume and coverage. Full results appear in Appendix E.5, Table 15.

We also emphasize that beyond candidate-set quality, HVFT produces a reusable adapted generator, which is valuable when the application requires a model for subsequent generation. MO-GFlowNets offer a relevant broader comparison, although their standard formulation trains a generative policy for reward-directed sampling rather than adapting the pretrained generators considered here. A comparison would therefore require a separate model and training setup, whose adaptation and validation for our benchmarks were not feasible within the rebuttal period.

We also added pretrained-model sampling with Pareto filtering and hypervolume subset selection (Base+Greedy), addressing the reviewer’s suggested “best-of-N” comparison. This baseline uses the same sampling budget as HVFT (90,000 samples), maintaining a Pareto archive and greedily selecting 1,000 designs by hypervolume if the final archive exceeds that size. On GLAST and GLP-1R, HVFT achieves higher 100-HV than Base+Greedy under this matched sampling budget. Full results appear in Appendix E.5, Table 15.

5. Objective space coverage

“Can you show objective-space coverage for the non-dominated set for all the methods?”

Response.

Thank you for this suggestion. Figures 9–13 in Appendix E.6 show the objective-space coverage of the non-dominated sets for all compared methods, illustrating the property values and trade-offs represented by their generated designs.

I appreciate the authors detailed rebuttal and the new experimental results. The revised manuscript resolves the issues I raised, and I have no remaining questions. Thus, I have increased my score.

Thank you sincerely for reviewing the additional results and raising your rating. We are glad that our responses resolved your concerns, and we will incorporate the results and clarifications into the revised manuscript. Thank you again for helping us strengthen our work.

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

Objects have several numerical properties, each of which we want to be as high as possible. There is a tradeoff between properties' values: some objects are strictly dominated by other objects, and some are are not (these are on Pareto frontier). Fix the number of objects; the reasonable objective to optimize is the total volume of space dominated by these objects.

Suppose we have base generative model. We want to fine-tune it in such a way that the resulting model has two qualities 1) if we sample from it, the group of samples dominates large volume (on average) 2) the distribution of outputs are not too far from base model (KL divergence term)

Clearly, the added value of a single object to the objective depends on other objects already present in the group. A reasonable metric here is the volume which is dominated by this object and not dominated by other objects of the group (which is randomly sampled and averaged over, so this quantity is hard to measure directly). The paper constructs unbiased sample-based estimator of this quantity (together with changes to KL). Based on this estimator, the paper builds multi-step fine-tuning method to get from base generator model to the model with two qualities above.

Critical strengths

The method is novel and theoretically justified. The experiments show that the method outperforms previous methods on Hypervolume metrics.

Critical weaknesses

None

Questions for authors

Am I correct that that Diversity (Vendi Score) and Coverage are less important than Hypervolume for this setting, since they are affected by objects which are not on Pareto frontier?

3 comments
AuthorsOpenReview ↗

Response to Reviewer BoRE

Thank you for the encouraging review and support of our work. We appreciate the reviewer’s assessment that our method is “novel and theoretically justified” and the paper is “clear and easy to follow.” We address the reviewer’s question below and thank them again for their positive assessment.

1. Role of design-space diversity and coverage

“Am I correct that Diversity (Vendi Score) and Coverage are less important than Hypervolume … since they are affected by objects which are not on Pareto frontier?”

Response.

Yes, both diversity and coverage are computed using designs beyond those on the Pareto frontier. Hypervolume is our primary metric for objective quality and trade-off coverage, while diversity and coverage measure the structural variety of generated designs. They help assess whether hypervolume gains come at the expense of structural diversity and, because they are not directly optimized, provide complementary evidence beyond the training objective.

Thank you! I keep the rating "4: Clear acceptance"

Thank you for confirming your assessment and for your continued support of our work!

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

The paper fine-tunes a pretrained generator (discrete diffusion or continuous flow-matching) so that a set of n samples maximizes expected hypervolume (HV) in objective space, KL-regularized toward the pretrained prior. The central observation: expected HV is a non-additive (set-level) functional and cannot be written as E[fixed per-candidate reward]; but within the Flow Density Control framework, its first variation equals the expected marginal hypervolume contribution of a candidate against background sets drawn from the current model (Proposition 1). The authors prove concavity (via a coverage-probability representation), relative smoothness (Pinsker), and geometric convergence to a unique maximizer. Empirically, the proposed HVFT attains SOTA 10-/100-HV on three therapeutic-peptide targets and two molecular-design tasks, while also improving design-space diversity and coverage.

Critical strengths

●Genuinely elegant central identity. First-variation-of-expected-HV = expected marginal HV contribution is a clean, correct, and reusable result.

●Self-contained, rigorous theory. The coverage representation (Eq. 14) makes concavity transparent and the diminishing-returns/submodular interpretation precise. Relative-smoothness leading to geometric contraction (Lemmas 1–3, Thm 4) is a complete, correct chain.

●Strong, modular empirics across two modalities. Using existing fine-tuners as inner solvers and still beating them standalone is the right experiment and cleanly isolates the value of the sequential marginal-HV reward. Beating DiffusionBlend without any preference-grid specification is a meaningful result.

Critical weaknesses

●Training objective = evaluation metric. HV is both optimized and the primary metric, while baselines optimize sum-scalarization. This is not unfair, but it structurally favors HVFT; the diversity/coverage/validity metrics are what carry the comparison, and that should be foregrounded. An orthogonal downstream metric (e.g., docking/experimental proxy not in the reward set) would materially strengthen the claims.

●Novelty is bounded by Flow Density Control. The outer-loop machinery is directly De Santi et al. (2025); the contribution is the HV-specific first variation + estimator + smoothness constant Ln=n(n−1)CM/4L_n=n(n−1)C^{M/4}. Substantial and non-trivial, but the paper should be precise that it is instantiating an existing framework for a new, important objective rather than introducing the optimization framework.

Questions for authors

●The estimator (Eq. 8) has unbiasedness per fixed x, but you optimize over the model that generates x; does holding background sets fixed within an inner solve introduce bias in the gradient w.r.t. model parameters? A variance/bias discussion of the coupled estimator would help.

●How were r_ref chosen, and how sensitive are the rankings in Tables 1–2 to ±20% perturbations of r_ref and of the dipole scaling factor?

●Can you add a spread-aware scalarization baseline (random HV scalarizations) to separate "cardinal set-level reward" from "better-tuned scalarization"?

4 comments
AuthorsOpenReview ↗

Response to Reviewer uNPf (Part 1 of 2)

Thank you for the thoughtful review. We sincerely appreciate the description of our central identity as “genuinely elegant,” the theory as “self-contained” and “rigorous,” and the experiments as “strong” and “modular.” In response to the reviewer’s suggestions, we provide additional evaluation beyond hypervolume, alongside clarifications of the estimator and our contributions within Flow Density Control. We will provide additional results addressing the requests for a spread-aware baseline and reference-point sensitivity analysis shortly.

To keep this response concise while making the additional evidence easy to find, we have collected the supporting tables and plots in Appendix E. We will carefully incorporate the feedback and additional results into the revised manuscript. If the responses address your concerns, we would sincerely appreciate an increase in your scores to reflect this updated assessment.

1. Evaluation beyond hypervolume

“HV is both optimized and the primary metric … the diversity/coverage/validity metrics … should be foregrounded. An orthogonal downstream metric … would materially strengthen the claims.”

Response.

Thank you for this helpful observation. Hypervolume directly measures our stated goal of generating high-quality, complementary trade-offs. To assess whether the gains extend beyond this metric, we added inverted generational distance (IGD) and IGD+. IGD measures the average distance from a reference front to the nearest generated candidates; IGD+ measures only shortfalls relative to that front. Lower values are better for both metrics.

For each task, we construct a fixed empirical reference front from the non-dominated points pooled across all methods’ 1,000-candidate sets. We evaluate subsets of 100 candidates sampled without replacement, averaging over 1,000 subset draws within each seed, and report means and standard errors across three seeds. HVFT achieves the best scores on both IGD and IGD+, supporting the gains under additional objective-space metrics. Full results appear in Appendix E.7, Tables 16 and 17 for peptides and molecules, respectively.

2. Relationship to Flow Density Control

“The outer-loop machinery is directly De Santi et al. (2025); the contribution is the HV-specific first variation + estimator + smoothness constant. … The paper should be precise … rather than introducing the optimization framework.”

Response.

We appreciate this point. The manuscript identifies Flow Density Control as the underlying distributional optimization framework in Sections 2 and 3. Our contributions concern the expected-hypervolume objective: its first variation and estimator, concavity and relative-smoothness analysis, and practical use as an adaptive reward for generative fine-tuning. We will make this distinction more prominent to ensure that the scope of our contributions is clear.

3. Fixed backgrounds and gradient estimation

“Does holding background sets fixed within an inner solve introduce bias in the gradient w.r.t. model parameters? A variance/bias discussion of the coupled estimator would help.”

Response.

Thank you for this question. We see two possible interpretations and address both below.

Holding the background distribution fixed is precisely the update defined in Equation (7) and supported by our convergence analysis. At outer iteration k, the inner subproblem optimizes the first variation evaluated at pk. This reward remains fixed while the generator is updated and is recomputed at the next outer iteration. Consequently, no gradient term required by the inner subproblem is omitted.

If the question instead concerns reusing a finite collection of sampled backgrounds, a separate consideration arises. The reward estimate is unbiased for each fixed candidate, but optimizing against the same sets can lead to overfitting to their sampling noise. Using more background sets improves estimation accuracy, and we draw fresh sets at each outer iteration. We will clarify this finite-sample consideration in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer uNPf (Part 2 of 2)

4. Reference-point selection and sensitivity

“How were rref chosen, and how sensitive are the rankings in Tables 1–2 to ±20% perturbations of rref and of the dipole scaling factor?”

Response.

For peptides, we set the reference coordinates for non-fouling, non-hemolysis, and solubility to zero, their lower bound. For binding and permeability, we chose reference values below the minima observed in 10,000 samples from the pretrained model. For molecules, the objectives have a lower bound of zero for valid designs; we used reference coordinates of −1 to match the scores assigned to invalid outputs.

We will provide the reference-point sensitivity results shortly.

We were unable to complete the dipole-scaling ablation within the rebuttal period, as assessing its effect on training would require rerunning all compared methods. Our experience with HVFT suggests that modest changes to objective scaling are unlikely to substantially affect its performance. We prioritized the reference-point sensitivity analysis and spread-aware baseline comparison to assess the robustness of the reported rankings and the benefit of the proposed reward.

5. Random hypervolume scalarization baseline

“Can you add a spread-aware scalarization baseline (random HV scalarizations) to separate ‘cardinal set-level reward’ from ‘better-tuned scalarization’?”

Response.

Thank you for this suggestion. We will provide results for a random hypervolume scalarization baseline shortly.

AuthorsOpenReview ↗

Reference-Point Robustness and Random Scalarization Results

Thank you again for the constructive suggestions. Complementing the clarifications and results provided in our earlier response, we have now completed the two outstanding experiments requested by the reviewer. We believe that, together, these responses address the reviewer’s main concerns and strengthen the paper. If the reviewer agrees, we would appreciate an upward revision of the score to reflect this updated assessment.

4. Reference-point sensitivity

“How sensitive are the rankings … to ±20% perturbations of rref …?”

Response.

Following the reviewer’s suggestion, we evaluated three alternative reference points on each of GLAST and Geometry. We generated 3,000 valid candidates from the base model and recorded the minimum observed value of each objective. For each coordinate, we set the perturbation size to 20% of the distance between this minimum and the original reference value. We then constructed three alternative reference points by independently adding or subtracting this amount from each original coordinate, with the signs chosen at random. The three points use different combinations of increases and decreases across coordinates, testing changes in multiple directions rather than a uniform shift.

Among the methods compared in this experiment, only HVFT uses the hypervolume reference point in its training objective. Changing the reference point therefore raises two distinct questions for HVFT: whether its advantage persists when only the evaluation reference point changes, and whether its performance is robust when the reference point used for training also changes.

To address the first question, we evaluated the model originally fine-tuned with HVFT under each alternative reference point, denoted HVFT (unmatched ref.). To address the second, we fine-tuned a separate model with HVFT for each alternative reference point and evaluated it using that same point, denoted HVFT (matched ref.). In both comparisons, we retained the existing baseline models and recomputed their hypervolume using the corresponding alternative reference point.

Both HVFT (unmatched ref.) and HVFT (matched ref.) achieve higher 100-Hypervolume and 10-Hypervolume than every compared baseline under all three alternative reference points on both tasks. The models from the two settings obtain similar hypervolume values, supporting robustness to changes in the reference point used for both evaluation and training. Full results appear in Appendix E.8, Tables 18–19.

5. Random hypervolume scalarization baseline

“Can you add a spread-aware scalarization baseline (random HV scalarizations) to separate ‘cardinal set-level reward’ from ‘better-tuned scalarization’?”

Response.

Following the reviewer’s suggestion, we evaluated a random hypervolume scalarization baseline (RHVS) on GLAST and Geometry using the same pretrained generators and inner solvers as HVFT.

To explore different directions in objective space, we fine-tuned five models, each using a different randomly sampled hypervolume scalarization, and combined their outputs. This allows each model to specialize in its assigned scalarization while retaining the pretrained architectures and existing inner solvers. Because training a separate model for each scalarization is computationally costly, we limited the comparison to five models. Another approach would condition a single generator on the scalarization weights. However, our pretrained models do not include this capability, so introducing it would require modifying their conditioning interface and developing and validating a corresponding fine-tuning procedure. This would constitute a separate methodological development.

We note that a key contribution of HVFT is precisely that a single fine-tuned generator can represent complementary trade-offs without conditioning on scalarization weights or modifying the base architecture.

We generated 200 valid designs from each scalarization model. When computing 100-Hypervolume and 10-Hypervolume, we sampled 20 and 2 designs per model, respectively, ensuring that all five scalarizations were represented. HVFT achieves higher 100-Hypervolume and 10-Hypervolume than RHVS on both tasks. Its 100-Hypervolume is 11.67 versus 8.26 for RHVS on GLAST and 4.988 versus 4.296 on Geometry (Appendix E.9, Tables 20–21).

As the rebuttal period nears its end, we would be grateful for your assessment of our responses and additional results. We believe these additions have strengthened the paper and would sincerely appreciate consideration of a higher score if they address your concerns. Thank you again for your valuable suggestions.

Comments on the paper

2 comments
AuthorsOpenReview ↗

Global Response to the Reviewers and Area Chair

We thank the reviewers for their careful evaluation and constructive feedback. We are grateful for their positive assessments of the method’s novelty, theoretical foundations, and empirical performance. In response to their requests, we have strengthened the paper with:

  • Budget-controlled comparisons: Evaluations that account for invalid generation attempts, matched end-to-end time and sampling budgets, and additional comparisons with NSGA-II and pretrained-model sampling with greedy hypervolume selection.

  • Evidence beyond hypervolume: IGD/IGD+ results and joint-property and non-dominated-set visualizations that help assess the quality and trade-offs of generated designs.

  • Robustness and methodological clarification: Regularization sensitivity analyses across four tasks, a clearer account of our contributions within Flow Density Control, and an explanation of the fixed-background estimator.

The supporting tables and figures are collected in Appendix E for convenient reference. Reference-point sensitivity and random hypervolume scalarization results will follow shortly. We will incorporate the feedback, additional evidence, and clarifications into the revised manuscript.

We hope these additions support a positive assessment of the strengthened paper. If our responses address the reviewers’ concerns, we would appreciate an upward revision of their scores to reflect their updated assessment.

AuthorsOpenReview ↗

Closing Summary and Thanks

As the rebuttal period draws to a close, we sincerely thank the reviewers for their thoughtful feedback and engagement during the discussion, as well as the Area Chair for overseeing the review process.

We appreciate the reviewers’ recognition of the central identity as “genuinely elegant,” the theory as “self-contained” and “rigorous,” and the experiments as “strong” and “modular.” The discussion helped us strengthen the evidence for our main contribution: a principled method for fine-tuning a pretrained generator to produce complementary trade-offs without modifying its architecture or conditioning on scalarization weights.

In response to the reviews, we added:

  • Broader comparisons: fixed-attempt and matched-budget evaluations, plus comparisons with NSGA-II, pretrained-model sampling with greedy selection, and random hypervolume scalarization.
  • Evidence beyond hypervolume: IGD/IGD+ results and joint-property and non-dominated-set visualizations.
  • Robustness and clarification: regularization and reference-point sensitivity analyses, alongside explanations of our contributions within Flow Density Control and the fixed-background estimator.

Together, the theoretical analysis and expanded empirical evidence strengthen the case for expected hypervolume as a practical fine-tuning objective. Supporting results are collected in Appendix E and will be incorporated into the revised manuscript alongside the reviewers’ feedback.

We appreciate that Reviewers RaPD and uemy increased their scores and that Reviewer BoRE reaffirmed their clear acceptance recommendation. We also provided detailed responses and the requested additional experiments for Reviewer uNPf, although we have not yet received their follow-up assessment.

Thank you again for your time and for helping us strengthen our work.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.