acceptodds
Under review as a conference paper at ICLR 2027

Expanding Generable Sets with One-Step Generators

Abstract

Generative scientific discovery aims to produce valid designs beyond the support of a static pretrained model. The goal is to expand its generable set over the valid design space using only black-box verifier feedback. Existing methods steer multi-step samplers toward unexplored regions, but their sampler implicitly determines how far queries travel and how strongly they remain near the initial model scope. We introduce Expanse, which grows the generable set with a 1-step generator by searching directly in design space with explicit control over exploration and travel. Rejected designs push the generator away from invalid regions, while inference settings trade coverage against validity. We also derive a reachability bound giving a necessary condition for a disconnected valid region to be found, in terms of generator support and query travel. With the same verifier budget and one function evaluation (NFE) per sample, Expanse covers of the connected valid region compared to for the multi-step baseline, while achieving higher coverage and validity overall.

Primary area:
generative models
Keywords:
Generable Set Expansion, Out-of-Distribution Discovery, Active Exploration

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 4 reviews, above 27% of the others.
Reviewer_daCHOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper addresses the problem of expanding the generable set of pretrained models under black-box verifier feedback. It proposes EXPANSE, which separates exploration from generator updates using occupancy-guided probing and rejection-aware training. Experiments on synthetic benchmarks and QM9 show improved coverage of valid regions and promising results in molecular validity and novelty compared with the evaluated baselines.

Critical strengths
  1. The paper addresses an important problem in generative scientific discovery by separating exploration from generator updates. The motivation is clear and the proposed framework is well structured.

  2. EXPANSE achieves substantial improvements over ActFlow in the 2D benchmark, particularly in discovering disconnected valid regions. The ablation studies provide useful evidence for the roles of seed perturbation, directed exploration, and replay separation.

  3. The reachability analysis provides a useful geometric interpretation of the exploration process. The paper is generally well written and discusses several limitations of the proposed approach.

Critical weaknesses
  1. The paper presents EXPANSE as a general framework for generable set expansion in generative scientific discovery. However, the proposed one-step formulation is mainly evaluated on a 2D synthetic benchmark, while the QM9 experiments use a few-step sampler. Evaluation of the one-step approach on more complex design tasks would strengthen the empirical support for the main claim.

  2. The exploration mechanism relies on occupancy-guided probing to discover unexplored regions. While the paper provides ablations on individual components, it lacks comparisons with exploration baselines such as random perturbation or distance-based novelty search. Such comparisons would help clarify the specific benefit of the proposed occupancy guidance.

Questions for authors

Please refer to the above weaknesses.

3 comments
AuthorsOpenReview ↗

Rebuttal response for Reviewer daCH

Thank you for the detailed review. We have addressed each point with new experiments, and uploaded the revised version of the paper with new text marked in blue.

Beyond the 2D benchmark. The one-step results are on ActFlow's released benchmark, where coverage is exact, now swept over gaps 0 to 6 (Table 6). The same two stages carry to molecules. On QM9 a four-step EXPANSE sampler finds 6.3× as many valid new clusters as ActFlow at four steps, 13.4× as many as ActFlow at 100 steps and 2.2× as many as the frozen teacher under equal compute (Table 7). Its count keeps growing while ActFlow's four-step count falls.

Table 6. Larger gaps and the coverage force. Coverage (%) at α=1 with ρ=Δ+0.5, which uses the gap and is not available in practice. EXPANSE has five seeds at Δ=4 and 6 and three at Δ=3 and 5. ActFlow has three seeds.

ΔEXPANSE component, no forceEXPANSE component, forceForce, τ=0.002EXPANSE total, no / with forceActFlow component (w=100 / w=13)ActFlow total (w=100 / w=13)
370.1 ± 11.776.6 ± 3.999.687.2 / 87.7not runnot run
40.066.8 ± 7.999.578.9 / 85.171.4 ± 2.5 / 52.6 ± 21.369.5 ± 1.5 / 75.1 ± 6.4
50.060.5 ± 12.599.978.6 / 84.1not runnot run
60.023.9 ± 24.298.177.3 / 79.561.5 ± 7.2 / 31.1 ± 33.068.0 ± 4.5 / 71.0 ± 8.6

Table 7. QM9 at 4 NFE under 16k NFE per cell, α=1. Entries are centers at τ=0.2 / new centers at cutoff 0.4 / validity (%), mean ± sample standard deviation over four seeds for EXPANSE and Rec-F and three for ActFlow. The frozen teacher at 4 NFE has 174 / 645 / 27.9. The two 100-NFE rows draw 160 molecules per cell.

MethodStartIteration 100Last
EXPANSE180±5 / 1235±12 / 86.8±0.7164±22 / 1363±274 / 73.5±6.2188±14 / 1438±81 / 69.0±3.7
Rec-F181±6 / 1231±16 / 86.9±0.4188±27 / 1485±317 / 77.4±1.9180±21 / 1564±227 / 78.0±3.5
ActFlow (w=10)178±3 / 620±15 / 28.6±1.1125±37 / 234±41 / 6.2±1.0156±22 / 230±23 / 5.9±0.6
ActFlow (no uncertainty)172±6 / 620±15 / 28.3±0.6110±18 / 212±42 / 5.7±1.0156±20 / 253±31 / 6.5±0.8
ActFlow (w=10, 100 NFE)42±2 / 47±5 / 96.2±1.749±2 / 86±12 / 85.2±1.067±8 / 107±10 / 87.3±2.2
ActFlow (no unc., 100 NFE)40±2 / 44±3 / 97.1±1.649±5 / 87±3 / 84.0±2.556±6 / 99±14 / 81.0±4.7

Exploration baselines. We added both baselines you named at the same seed scale and travel budget. At gap 2 occupancy ascent covers 80.5% of the component against 59.8% for random perturbation and 10.4% for distance-based novelty, with less than half the seed spread. At gaps 0 and 1 the three agree, so occupancy guidance pays exactly where reach is hard (Sec. 4.3, Table 4).

Table 4. Probe direction at Δ=2. The base method has no coverage force. It uses five seeds, α=1, τ=0.05, and the same σ0 and ρ in every row.

Probe directionComponent coverage (%)Queries in componentMedian first hit
Occupancy ascent80.5 ± 6.375186
Random direction59.8 ± 15.443922
kNN novelty10.4 ± 16.715982

Theory. The geometric analysis you found useful now proves upper and lower probability bounds on reaching a region, and it predicts the measured queries in the component (747 against 755 at gap 2, Table C).

Table C. Capture mass against logged queries, Δ=2, coverage force, five seeds (Appendix B.1, Fig. 9). These are Monte Carlo consistency checks, not certified bounds.

QuantityValue
Capture mass M_C (seed mean)747 (per seed 626 to 840)
Logged queries in the component755 (per seed 714 to 772)
Upper-bound total (seed mass within ρ of C)7542
Per-probe capture probability ν_t(C)0.6% to 4.9%
Per-probe upper bound20% to 29%
40 further runs, capture mass before first hit roundmean 0.53, max 3.3; share of runs above λ never exceeds e^{−λ}

Verifying these numbers. Each table here regenerates from the supplementary zip with one command: python -m drift.dump_numbers controls for Table 4, force for Table 6, qm9 for Table 7 and python -m drift.capture --plot for Table C. The occupancy probe also helps outside our loop: on ActFlow's own sampler it raises total coverage by 8 to 28 points, while random seed perturbation alone on the same sampler gives 62.9 and 58.3 total and 29.4 and 0.0 component coverage at gaps 1 and 2.

Please let us know if any further experiments or clarification are required during the discussion period or before the camera ready version. We will be happy to discuss and perform further experiments.

AuthorsOpenReview ↗

Follow-up for Reviewer daCH: new results since our rebuttal

Dear Reviewer daCH, thank you for your review. Our rebuttal added the two exploration baselines you named (random perturbation and kNN novelty), where occupancy ascent covers 80.5% of the disconnected component at Δ=2 against 59.8% and 10.4%. Since then we have posted three further results that bear on your points.

Generalisation without retuning. At held-out gaps Δ=1.5 and 2.5, with every setting frozen, EXPANSE leads ActFlow by 15 to 21 points of total coverage (Welch p=0.016 to 0.037) and reaches the disconnected component 3 to 10 times sooner.

Beyond ActFlow. Against Variational Search Distributions and Generative Bayesian Optimization as evaluated, both stay at 3.0% coverage when started from the pretrained model, against 90 to 93% for EXPANSE. With a uniform prior, VSD keeps pace only while the design space is small enough to sample, and EXPANSE holds 88.5 to 90.2% as the space grows 64-fold in area.

QM9. A factorial over the probe and the rejection mixture shows the two working as a pair: together they give the highest discovery per valid molecule of the three arms and a 5.8× smaller seed spread than consolidation alone. EXPANSE finds 6.3× and 13.4× as many valid new clusters as ActFlow at four and 100 steps.

Full tables are in our replies to mA4z and QiLn. If any concern remains, we would be grateful to hear it during the discussion period so we can respond with evidence.

Thank you for the authors’ response and the additional experiments. The newly added random perturbation and kNN novelty-search baselines have alleviated my earlier concern about the insufficient comparison of exploration mechanisms. However, with respect to the effectiveness of the single-step method on complex tasks, the response still does not provide sufficient evidence. The validation of single-step generation remains limited to 2D generation experiments. although the additional QM9 experiments strengthen the empirical support for the framework on complex tasks, since they employ a four-step sampler, the applicability of the single-step generation capability emphasized in the paper to complex tasks remains to be verified. Therefore, I maintain my original score.

Reviewer_HC9MOpenReview ↗
Rating
1
Confidence
4
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper tackles the problem of generable set expansion in the case of 1-step generators. They provide a new method, Expanse, and a theoretical reachability analysis via the introduced algorithm, as well as experimental results covering both 2D toy examples and a small molecules dataset (QM9).

Critical strengths
  • Substantially improves over ActFlow machinery to learn from negative samples. This is an important problem in discovery tasks where valid signal can be substantially sparse.
  • The new type of presented guarantees is interesting as it appears to be less demanding/weaker than the ones in prior work, but also requiring significantly less assumptions and providing a more geometric, rather than probabilistic viewpoint. This viewpoint is very original and a promising one for future analysis.
  • The 2D Island Gap visual experiments are particularly insightful and results are strong/convincing. This hints at the fact that the proposed mechanisms indeed result in plausibly stronger results in the presence of disconnected components.
Critical weaknesses
  • (Writing) General writing clarity is very weak. I strongly suggest the authors to manually re-write every paragraph making sure that the key messages are communicated, and removing excessive details when not needed. For instance, the introduction is too dense in writing, appears heavily LLM processed. I would suggest to keep it simpler, and high-level also by dropping details/ideas from there. Similarly, I would remove the explicit N_c bound from the list of contributions as it is not introduced yet. All terms used (e.g., push distribution) should be better introduced and defined. The experiments section is a single wall of text: make sure to divide it into subsections/paragraphs focused on specific experiments. Currently this work is overall very hard to understand.

  • (Experiments) The comparison with ActFlow on QM9 is unconvincing. In particular, since ActFlow can retrieve Rec-F, just by setting to zero the term weighting the variance in the acquisition function, its performance should be approximately lower bounded by that of Rec-F. Also, these results on QM9 do not seem to reproduce (even closely) the positive results from the ActFlow paper as from Fig. 8 it seems that all relevant metrics (e.g., coverage, novelty) decrease with ActFlow, which is likely due to improper tuning of the baseline.

I am open to change my assessment significantly if a significant rewriting increases clarity by a large margin, and if the point regarding QM9 comparison is properly addressed.

Questions for authors

I have no further questions for the authors.

4 comments
AuthorsOpenReview ↗

Rebuttal response for Reviewer HC9M

Thank you for the detailed review, and for offering to revisit your assessment if the writing and the QM9 comparison were addressed. We have addressed each point with new experiments, and uploaded the revised version of the paper with new text marked in blue.

Writing. The paper is significantly rewritten. The introduction is cut to the core idea, the N_C bound is out of the contribution list, every term including the push distribution is defined where it first appears, and the experiments are split into six subsections. The reachability result you found promising is also stronger, since Theorem 3.4 now proves upper and lower probability bounds in one equation.

QM9 comparison. We ran exactly the comparison you describe. ActFlow with its variance term switched off is its own recursive baseline, and it now sits next to ActFlow in Table 7. The two agree within seed spread at four steps and at 100 steps, so ActFlow does not fall below its own recursive baseline. The row called Rec-F is a different method of ours that keeps EXPANSE's drifting update, and the revision labels the two apart.

Reproducing ActFlow. At its native 100 steps our ActFlow run stays valid and more than doubles its new clusters over the loop. The decline you saw is a few-step effect. A flow-matching fine-tune of the distilled sampler drops four-step validity from 87% to under 1% in its first update, which identifies few-step fine-tuning and not tuning as the cause. Under equal compute EXPANSE finds 13.4× as many new clusters as ActFlow at 100 steps and 6.3× as many as at four steps. Nothing in the comparison points to a mistuned port, and Appendix C.2 lists the port settings and the remaining differences (the README guidance weight of 10 and 200 rather than 1000 rounds).

Table 7. QM9 at 4 NFE under 16k NFE per cell, α=1. Entries are centers at τ=0.2 / new centers at cutoff 0.4 / validity (%), mean ± sample standard deviation over four seeds for EXPANSE and Rec-F and three for ActFlow. The frozen teacher at 4 NFE has 174 / 645 / 27.9. The two 100-NFE rows draw 160 molecules per cell.

MethodStartIteration 100Last
EXPANSE180±5 / 1235±12 / 86.8±0.7164±22 / 1363±274 / 73.5±6.2188±14 / 1438±81 / 69.0±3.7
Rec-F181±6 / 1231±16 / 86.9±0.4188±27 / 1485±317 / 77.4±1.9180±21 / 1564±227 / 78.0±3.5
ActFlow (w=10)178±3 / 620±15 / 28.6±1.1125±37 / 234±41 / 6.2±1.0156±22 / 230±23 / 5.9±0.6
ActFlow (no uncertainty)172±6 / 620±15 / 28.3±0.6110±18 / 212±42 / 5.7±1.0156±20 / 253±31 / 6.5±0.8
ActFlow (w=10, 100 NFE)42±2 / 47±5 / 96.2±1.749±2 / 86±12 / 85.2±1.067±8 / 107±10 / 87.3±2.2
ActFlow (no unc., 100 NFE)40±2 / 44±3 / 97.1±1.649±5 / 87±3 / 84.0±2.556±6 / 99±14 / 81.0±4.7

Port fidelity. Our ActFlow port follows its released code and README: fine-tuning starts after 4096 valid samples at batch 64 (round ≈66, matching the 66 initial iterations of its paper), guidance weight 10 as its README suggests, and features from flow timestep 0.9. The remaining differences are 200 rather than 1000 rounds and scoring at four steps beside its native 100. At 100 steps it grows over the loop as its paper reports: centers rise from 42 to 67 and new clusters from 47 to 107 at 87% validity.

What drives the disconnected-component result. You found the island-gap experiments convincing, and a new control isolates their mechanism. On ActFlow's own sampler, seed noise alone covers 29.4% of the component at gap 1 and 0.0% at gap 2; adding our occupancy ascent raises this to 86.9% and 56.3%, and total coverage from 62.9 and 58.3 to 90.5 and 86.1 (three seeds each, paper Table 5).

Verifying these numbers. The supplementary zip holds the scored QM9 results for every checkpoint and seed of every row in Table 7, with the ActFlow port and its launch arguments. python -m drift.dump_numbers qm9 and qm9nfe100 reproduce Table 7 from these scored results without retraining, and drift.evaluate_qm9 is the scorer that produced them.

Please let us know if any further experiments or clarification are required during the discussion period or before the camera ready version. We will be happy to discuss and perform further experiments.

AuthorsOpenReview ↗

Follow-up for Reviewer HC9M: new results since our rebuttal

Dear Reviewer HC9M, thank you again for offering to revisit your assessment once the writing and the QM9 comparison were addressed. Since our rebuttal we have added three sets of results in the discussion.

QM9: what each component adds (seeds 0–2, last checkpoint, equal compute)

ArmNew clustersPer 100 validValidity (%)
Rec-F (EXPANSE's consolidation only)1552 ± 27753.3 ± 5.179.6
+ probe1285 ± 42353.3 ± 12.561.7
+ rejection mixture1110 ± 1348.3 ± 2.371.1
+ probe + rejection mixture (EXPANSE)1474 ± 4857.7 ± 4.570.1

Rec-F is EXPANSE without acquisition and mixture, and ActFlow's own recursive baseline sits beside ActFlow in Table 7. With the rejection mixture added, validity recovers, discovery per valid molecule is the highest of the three arms, and the seed spread is 5.8× smaller than Rec-F's. EXPANSE still finds 6.3× as many valid new clusters as ActFlow at four steps and 13.4× as many at ActFlow's native 100 steps.

Held-out gaps on the island benchmark you found convincing. All settings were frozen, and Δ=1.5 and 2.5 were never used for tuning. EXPANSE leads ActFlow by 15 to 21 points of total coverage (Welch p=0.016 to 0.037) and reaches the disconnected component 3 to 10 times sooner (after 64 and 86 queries, against 198 to 836).

Broader baselines. We compared against Variational Search Distributions and Generative Bayesian Optimization as evaluated (126 runs). Started from the pretrained model, both stay at 3.0% coverage, against 90 to 93% for EXPANSE. Given a uniform prior over the design space, VSD matches EXPANSE only while that space is small enough to sample, and falls to 59% as it grows. EXPANSE stays at 88.5 to 90.2%. With its published regularizer added, GenBO still does not expand (at most 4.2% coverage, 0.0% of the component in all 27 runs).

Full tables are in our replies to mA4z and QiLn. If any concern remains, we would value hearing it during the discussion period so we can address it directly.

Reviewer_HC9MOpenReview ↗

The paper is still unreadable even after the suggested rewriting

The writing and exposition quality of the paper is extremely low for the standards of this community. Although, as I mentioned in my initial review, the work contains promising ideas and results, the "rewriting" resulted still in an unreadable paper. In the following, I aim to support my point my discussing the first 4 sentences of the current (rewritten) abstract, namely:

Pretrained generative models rarely produce valid designs outside their training data, yet scientific discovery needs exactly those designs. Multi-step samplers fix how far queries travel from the old model. We introduce EXPANSE, which splits each round into covering and consolidation. A probe searches design space directly, with explicit control over how widely and how far queries travel.

The first sentence is wrong stated as is: generative models nearly always generate new data beyond the training set. I know what the authors meant to express, but it is not written here. The second sentence is extremely poorly connected with the first, "travel" is undefined, "old model" is undefined. Moreover, the sentence does not make logical sense. The level of clarity is already extremely low. The next sentence (3rd) starts discussing algorithmic aspects saying "[EXPANSE] splits each round" without neither stating the goal of the algorithm nor saying rounds of what process. The next sentence starts with "A probe searches" without minimally contextualize what this "probe" is relative to what has been introduced earlier.

I then analyzed several other parts of the updated paper, including method (Sec. 3), and Experiments (Sec. 4), and similar problems persist with high frequency. This work is a clear rejection as it is currently unreadable. I strongly suggest the authors to invest sufficient time into rewriting the paper from scratch, manually (or at least critically), and well.

AuthorsOpenReview ↗

We respectfully disagree

Thank you for reading the revision closely. We respectfully disagree with the conclusion that the paper is unreadable. The abstract is written to the length limit, and each term you raise is defined where it first appears in the body.

Travel is the probe's budget ρ = Kδ (Sec. 3.2, Lemma 3.1), the old model is the pretrained generator being expanded (Sec. 3.1), and a round is one iteration of Algorithm 1.

On the first sentence, we believe the distinction is already in the text. It concerns valid designs outside the training data, not new samples in general, and a generative model producing novel samples near its training distribution is precisely the behaviour that motivates expanding its generable set; as the reviewer notes, the intended meaning was clear. For other sections please point out where the problems persist in high frequency and we will fix it. The vague comments help no one.

We note that this comment raises no concern about the method or the results, and that the one technical point in the original review, the QM9 comparison with ActFlow's recursive baseline, was answered with new experiments and has not been followed up. We would still welcome any technical feedback that could improve the method.

Reviewers mA4z and QiLn engaged with the method and results in technical detail, and QiLn now plans to raise their score, which suggests the exposition supports careful reading. You described the ideas as promising, the island-gap results as strong and convincing, and the reachability viewpoint as original.

Sentence-level polish such as defining "old model" and "travel" in the abstract itself is quick to do, and we will make those edits once the discussion period ends, when there is time to give the writing the attention it deserves. We have currently improved the abstract and are focusing on polishing the entire paper from scratch.

We hope the final assessment weighs those contributions alongside the presentation, since phrasing at this level can be fixed in a camera-ready revision. We will take your sentence-level comments into a further revision, and we thank you for your time.

We were glad to see that the substance of the paper came through clearly enough for you to call the ideas promising and the results strong, and we trust the same care will carry over to the final assessment.

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

This paper proposes EXPANSE, a method intended to expand the set of valid designs that a pretrained generator can produce, using feedback from a black-box verifier. The method builds on ideas from ActFlow, but replaces trajectory-based sampling with direct probing in design space and iteratively updates the generator from accepted and rejected queried designs. Experiments are reported on a synthetic disconnected-region benchmark and QM9.

Critical strengths
  1. The problem is relevant and potentially impactful. Expanding a generative model beyond the support of its initial training distribution using only black-box validity feedback is an interesting and practically relevant problem, particularly in scientific design settings where valid candidates may be expensive to obtain and validity may only be assessable through a simulator, experiment, or other non-differentiable verifier. The paper’s focus on discovering valid regions that are not represented by the initial generator is therefore well motivated.

  2. The paper provides theoretical results that appear sensible. The initial lemmas seem correct to me, though I could not verify the main results.

  3. The experimental section includes useful ablations. Figure 7 provides ablations of several method components, while Figures 3–4 study sensitivity to key query parameters. This is useful given that the proposed method has multiple interacting design choices.

Critical weaknesses
  1. The paper is difficult to follow and not sufficiently self-contained. Many details crucial for understanding the method are either only informally introduced, scattered across the paper and appendix, or assumed from ActFlow/Drifting. For example, I had difficulty identifying precise definitions and roles for α\alpha, rmin⁡r_{\min}, thinning, ΠX\Pi_{\mathcal X}, d^\hat d, σt\sigma_t, and the occupancy field. Naming a quantity is not a substitute for defining it mathematically or operationally. As written, I could not develop enough confidence in my understanding of the methodology to thoroughly assess the main theoretical results. The paper would benefit from a concise overview of one complete iteration, a notation table, clearer pseudocode, and a direct explanation of which components are inherited from ActFlow versus newly introduced.

  2. The scope of the “one-step” claim and model class is unclear. The method repeatedly updates the generator during the expansion loop, so its efficiency claim appears to concern sampling-time NFE rather than avoiding model retraining. Moreover, the core formulation assumes a deterministic map, x=fθ(z,α)x = f_\theta(z, \alpha). This seems natural for the one-step implicit generator in the synthetic experiment and perhaps deterministic flow/ODE-style samplers. It does not directly describe a standard stochastic DDPM reverse process, which involves a sequence of random transitions. The QM9 experiment also uses a 22–44-NFE fine-tuned flow model rather than a one-step generator. The paper should clarify the exact intended scope: which model classes does EXPANSE apply to without modification, and which parts are specific to deterministic implicit or flow-style generators?

  3. The empirical comparison is limited. The main baseline is ActFlow, with , with Rec-F added for QM9. I believe the paper should more carefully position itself against generative black-box search methods that also learn an evolving proposal distribution, e.g., Variational Search Distributions (Steinberg et al., ICLR 2025) and Generative Bayesian Optimization (Oliveira et al., ICLR 2026). These methods iteratively learn generative proposal distributions from batch black-box feedback and support binary utilities, such as what a black-box verifier can provide. Their usual formulations may require explicit model densities and hence may not be drop-in replacements for the flow-matching setup used here. However, this is not necessarily a fundamental obstacle to an end-to-end comparison, since they could be instantiated with a density-based generator and compared under matched query and compute budgets. At minimum, the paper should clarify whether the intended contribution claim is restricted to sample-only flow-style generators, or whether it concerns generable-set expansion more broadly.

Questions for authors
  1. Can the authors give a self-contained definition of a complete EXPANSE iteration, including the exact meanings of α\alpha, γ\gamma, rmin⁡r_{\min}, thinning, and d^\hat d?

  2. Which method components require a deterministic generator x=fθ(z,α)x=f_\theta(z,\alpha)? Would the method apply directly to a standard stochastic DDPM, or only after replacing its sampler with a deterministic or distilled alternative?

  3. Can the authors add stronger baselines, more seeds, and controls that isolate the effects of initialization and replay from the proposed probing mechanism?

7 comments
AuthorsOpenReview ↗

Rebuttal response for Reviewer QiLn

Thank you for the detailed review. We have addressed each point, and uploaded the revised version of the paper with new text marked in blue.

Self-contained method. Appendix A.1 is a notation table for every symbol you listed, Algorithm 1 states one complete round, and thinning is defined where it first appears. Sec. 3.1 states what comes from ActFlow (the coverage objective), what comes from Drifting (the update field) and what is new (the probe, occupancy depletion and the rejection mixture).

Scope. The efficiency claim is about sampling and acquisition cost per round (Sec. 3.1). The probe needs only terminal samples, so it applies to any sampler, and we demonstrate that by running it on ActFlow's multi-step sampler. Only the drifting update needs a deterministic differentiable map from noise to design. A stochastic sampler such as DDPM qualifies once distilled into such a map, and on QM9 we distill the pretrained flow into a four-step map.

Positioning. Variational Search Distributions and Generative Bayesian Optimization are now discussed in Sec. 2. Both fit a proposal through its explicit density and target optimization or rare-class search, while EXPANSE targets sample-only generators and coverage of a valid set. ActFlow poses exactly our objective and is the baseline on both benchmarks: across guidance weights, samplers and step counts on the 2D benchmark, and at four and 100 steps on QM9.

Seeds and controls. EXPANSE now uses five seeds on the 2D benchmark (three at gaps 3 and 5) and four on QM9. Each control isolates one factor at gap 2. Removing the free positives moves component coverage by under two points (Table B), so initialization does not drive the result. Removing replay separation lowers total coverage from 89.5 to 76.1 and component coverage from 83.9±2.7 to 43.8±21.2, so replay matters and is reported. Replacing occupancy ascent with random or distance-based directions costs 21 and 70 points (Table 4). Our probe alone lifts ActFlow's total coverage by 8 to 28 points (Table 5).

Table 4. Probe direction at Δ=2. The base method has no coverage force. It uses five seeds, α=1, τ=0.05, and the same σ0 and ρ in every row.

Probe directionComponent coverage (%)Queries in componentMedian first hit
Occupancy ascent80.5 ± 6.375186
Random direction59.8 ± 15.443922
kNN novelty10.4 ± 16.715982

Table 5. Our probe on ActFlow's sampler. Entries are total / component coverage, then validity (%), at α=1. ActFlow with our probe changes only the query points: its guidance is off for them, and its fine-tuning, replay and verifier are the released ones. It has three seeds, ActFlow three (five at Δ=2), EXPANSE five.

MethodΔ=1 Cov. / comp.Δ=1 Val.Δ=2 Cov. / comp.Δ=2 Val.
ActFlow (w=100)62.4 / 69.596.769.8 / 74.896.5
ActFlow (w=13)71.2 / 74.696.378.1 / 72.396.5
ActFlow with our probe90.5 / 86.996.086.1 / 56.397.1
EXPANSE89.8 / 85.993.490.4 / 83.593.3

Table B. Free positives, Δ=2, α=1, base method, five seeds.

SettingTotal coverage (%)Component coverage (%)Validity (%)
With free positives89.980.5 ± 6.393.7
Without (fine-tuning gated until 64 accepted positives)89.778.8 ± 7.794.7

Theory. Theorem 3.4 now proves upper and lower probability bounds on reaching a region, through one label-free quantity that is defined in the theorem and listed in the notation table.

One round, self-contained. (1) Seeds: draw B samples from the generator and add Gaussian noise of scale σ0. (2) Probe: each seed climbs the occupancy field for K steps of length δ while the bandwidth h shrinks, so it travels at most ρ=Kδ. The occupancy field is the posterior variance of a random-feature Bayesian linear model over the locations of all past queries and uses no labels. (3) Verify: the B endpoints are labelled. An accepted point enters the positive buffer only if it lies at least r_min from every stored positive (thinning), and a rejected point enters the negative buffer. (4) Consolidate: S drifting steps move generated samples toward the positives and away from a push set of generated samples with negatives mixed in at rate γ. The generator takes α=1/(1−γ) as input, so α sets coverage against validity at inference.

Verifying these numbers. Each control here regenerates from the supplementary zip with one command: python -m drift.dump_numbers controls, afprobe and free for Tables 4, 5 and B.

Please let us know if any further experiments or clarification are required during the discussion period or before the camera ready version. We will be happy to discuss and perform further experiments.

Reviewer_QiLnOpenReview ↗

Remaining concerns on evaluation and impact

Thank you for the revision and for offering further experiments and clarification. My main remaining concern is the breadth of the experimental evaluation and what it establishes about the significance of the contribution beyond an improvement over ActFlow.

Could you clarify which contribution claims are specifically relative to ActFlow and which concern the broader state of the art? Given the setting of adapting a pretrained generator using black-box verifier feedback, I would like to see comparisons with other methods reasonably applicable to this setting. Your related-work discussion identifies potentially relevant alternatives, but I am not yet convinced that the stated distinctions establish their incompatibility with the evaluated setting. In particular, redistributing mass toward initially low-density regions may expand the density-level generable set without changing mathematical support. Likewise, differences in objective or density requirements do not, by themselves, rule out a meaningful experimental comparison, with reasonable adaptations where necessary.

These comparisons would help establish the practical value of the approach beyond ActFlow and help me assess the depth and potential impact of the contribution for the ICLR community. I recognize that the remaining two to three days may not permit a meaningful additional evaluation, and I would not expect a rushed comparison. If addressing this concern requires more time, a strengthened evaluation in a future resubmission may be the more appropriate route.

AuthorsOpenReview ↗

Response to Reviewer QiLn: VSD and GenBO comparisons, and scope of claims

Thank you for the follow-up and for pointing out the comparisons we need to make. We ran the experiments in full today on 4 GPUs in parallel. VSD and GenBO on the benchmark, 126 runs, three seeds per cell, with the same verifier, the same 32k query budget and the same scorer as every other row in the paper. The experiment is small in two dimensions, so it could be done carefully and completely. The results help sharpen the evaluation section, and we are grateful for the suggestion.

Scope of claims. Every quantitative comparison in the paper is relative to ActFlow, which introduced generable set expansion and is its only prior method. Three contributions do not depend on a baseline: the probe is portable (it lifts ActFlow's own coverage by 8 to 28 points), the rejection mixture gives a coverage-validity control at inference, and Theorem 3.4 bounds reach from both sides. We do not claim state of the art for black-box optimization or active generation, and the revision states this.

Setup. VSD (Eq. 10, class probability estimator) and GenBO (PI utility, balanced forward KL), each with a 64-component Gaussian mixture as the explicit density. Both papers are discrete, so two adaptations were needed and we state them: a reparameterized gradient for VSD in place of the score-function estimator, and GenBO without importance weights (as in its experiments) and without its parameter regularizer. Both methods require a prior p0. We sweep it from the pretraining law, which is the setting of generable set expansion, to a uniform distribution over the design-space box.

Table A. Total / component coverage / validity (%), tau = 0.05, 32k queries. EXPANSE: 5 seeds (Table 1). VSD, GenBO: 3 seeds.

MethodPrior p0Gap 0Gap 1Gap 2
EXPANSE (alpha=1)pretrained generator92.6 / 91.4 / 94.689.8 / 85.9 / 93.490.4 / 83.5 / 93.3
VSDpretraining law3.0 / 0.0 / 1003.0 / 0.0 / 1003.0 / 0.0 / 100
GenBOpretraining law3.0 / 0.0 / 1003.0 / 0.0 / 1003.0 / 0.0 / 100
VSDGaussian, std 0.527.2 / 41.5 / 99.821.1 / 0.0 / 99.721.3 / 0.0 / 99.7
VSDGaussian, std 1.048.6 / 92.8 / 99.439.3 / 28.5 / 99.135.1 / 0.0 / 99.3
VSDGaussian, std 2.065.1 / 97.9 / 99.265.9 / 89.5 / 98.961.9 / 47.3 / 98.7
VSDuniform over box94.1 / 89.6 / 99.091.8 / 84.3 / 98.788.0 / 51.4 / 98.7
GenBOGaussian, std 0.527.4 / 41.4 / 83.421.3 / 0.0 / 70.321.4 / 0.0 / 70.2
GenBOGaussian, std 1.048.6 / 93.2 / 55.339.2 / 28.2 / 43.334.9 / 0.0 / 40.5
GenBOGaussian, std 2.062.4 / 94.0 / 25.857.7 / 60.2 / 22.755.4 / 21.5 / 20.6
GenBOuniform over box47.7 / 41.0 / 99.645.0 / 20.2 / 99.443.4 / 26.6 / 99.2

1. In the setting of the paper, neither method expands. Starting from the pretrained model, VSD and GenBO remain at 3.0% coverage at every gap and place no query in the component in 288k queries each. This follows directly from their objectives: both converge to p0 times the utility, and the pretrained model already lies inside the valid set, so the target is the pretrained model itself. EXPANSE starts from the same model and reaches 90 to 93%.

2. Your point about density is correct, and the sweep quantifies it. A density model can raise mass in low-density regions and so enlarge the generable set. How far it does so is set by the prior: coverage grows monotonically with the share of the space that p0 already covers. For these methods the expansion is supplied through p0. In EXPANSE it is produced by the probe.

3. With a uniform prior, VSD is competitive when the box is tight and degrades as it grows. EXPANSE does not. A uniform prior presupposes a bounded space that can be sampled directly. We therefore swept the size of that box at gap 2, keeping the valid set, budget and scoring fixed. EXPANSE uses the box only as a clamp.

Table B. Total / component coverage / validity (%), gap 2, tau = 0.05, 3 seeds per cell.

Box half-widthValid share of boxEXPANSEVSD, uniformGenBO, uniform
5.5 (benchmark)9.9%88.5 / 81.0 / 93.688.0 / 51.4 / 98.743.4 / 26.6 / 99.2
112.5%90.2 / 78.8 / 92.890.0 / 89.3 / 98.741.2 / 9.8 / 98.0
220.62%89.2 / 77.6 / 94.376.2 / 52.9 / 98.634.5 / 0.1 / 96.0
440.15%88.6 / 75.8 / 94.259.1 / 20.5 / 99.028.7 / 0.0 / 93.4
Reviewer_mA4zOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3(+1)
Read review
Paper summary

The paper follows ActFlow and studies expanding the "generable set" of a generator under a black-box binary verifier. Each round is split into covering and consolidation. Covering is a probe seeded at scale σ0\sigma_0 that ascends a label-free RFF occupancy field within a travel budget ρ\rho. Consolidation is a drifting update with rejections mixed into the push side at rate γ\gamma, and conditioning on α=1/(1−γ)\alpha=1/(1-\gamma) gives a coverage-validity knob at inference. A necessary reachability bound is derived from probe geometry. Experiments are on the 2D ActFlow benchmark with an added disconnected component, and on QM9.

Critical strengths
  • Splitting covering from consolidation makes each knob measurable on its own. I like this.
  • Thorough ablations (Fig. 3, 4, 7).
  • Honest about what fails (αα on QM9, the ActFlow port, the 350× claim).
Critical weaknesses
  • The evidence is essentially one 2D toy. This is my main concern. On QM9 the model is a 2–4 NFE fine-tune, α\alpha does not transfer, and Theorem 3.4 does not apply (D.4). ActFlow gets 6.1% validity and the authors suspect their own port, so Rec-F is the only real baseline. At 4 NFE EXPANSE finds 168 centers. The frozen teacher alone finds 174 at the same NFE and budget (Sec. 1). So on the discovery count the whole loop does not beat the untouched teacher, only on validity.
  • The headline gains measure density spreading, not reach. At ActFlow's own τ=0.01\tau=0.01, every method except ActFlow (SDE) is above 93% at every gap (C.2). So ActFlow (ODE) already reaches the component, which contradicts Fig. 1 (middle). At τ=0.05\tau=0.05, seed noise alone (ρ=0\rho=0) gives 96.1% connected coverage, above the full method's 90.7% (Fig. 7). So the abstract's 90.7 vs. 78.8 is not evidence for the probe. A plausible source is rminr_{min} thinning: matching a uniformly thinned buffer spreads density, which is exactly what τ=0.05\tau=0.05 rewards.
  • Operating points are mixed. Fig. 2 is at α=1\alpha=1, where EXPANSE validity is not reported (it appears below ActFlow (ODE) in Fig. 5i). The "both metrics" claim uses α=1.25\alpha=1.25, with total coverage 80.8 to 83.3. Against ActFlow's 77.8 at Δ=2\Delta=2, the margin may be 3 points.
  • The attribution argument does not hold. C.2 says a component swap is impossible because the probe needs terminal coordinates. But at β=0\beta=0 the probe only uses a sample as a seed, and D.1 already runs the probe with a multi-step sampler. So "ActFlow sampler + probe" is feasible and missing. Without it no margin in Fig. 2 is attributable. The 57 free positives and ρ=2.0\rho=2.0 (equal to the largest Δ\Delta) make this worse.
  • Theory is shallow. Theorem 3.4 is one containment plus Markov. Corollary 3.5 is vacuous at the main setting: with ρ=Δ=2\rho=\Delta=2, the second term of Eq. 12 equals 1 for every rr. B.1 admits N^C\hat{N}_C restates the ρ+σ0d>Δ\rho+\sigma_0\sqrt{d}>\Delta threshold. Prop. 3.3 is two lines of algebra and, per the authors, does not describe the implemented update.
  • Much of Sec. 3.2 is unused. The tilted law and mean-shift prior (Eq. 1–3) live in the β\beta term, and every reported result uses β=0\beta=0. The probe is then plain occupancy ascent, and the one-step generator only supplies seeds. The claim that a one-step generator forces the separation (Sec. 1) is not what drives the results (D.1 says as much).

Minor issues:

  • Contribution (3) has the direction backwards. "Every run that missed had N^C<1\hat{N}_C<1" supports sufficiency, not necessity.
  • QM9 numbers are at iteration 100 of a 200-round loop, and Fig. 8 stops at 150.
  • Which generator is the "Pre-trained" row in Fig. 2? EXPANSE's own init at τ=0.05\tau=0.05 is not reported (A.1).
  • Consolidation is batch-limited: at N=128N=128 the component falls to 4.6 at Δ=2\Delta=2.
  • "Eq. equation 5", "island" vs. "component", "transfering".
Questions for authors
  • On QM9, new centers (1468) exceed total centers (168) at the same cutoff and from the same valid pool. Fig. 9a is labelled "clusters reached". Which definition is used?
  • Can you report all methods at τ=0.01\tau=0.01 and τ=0.1\tau=0.1, with component coverage separately?
  • What does ActFlow get when starting from the same 57 thinned positives?
  • How were σ0=1\sigma_0=1 and ρ=2\rho=2 chosen? Was any Δ\Delta held out?
  • Is ActFlow's σt\sigma_t actually label-dependent? If it is a GP regression variance, it is label-free too, and contribution (1) is not a difference.
4 comments
AuthorsOpenReview ↗

Rebuttal response for Reviewer mA4z

Thank you for the detailed review. We have addressed each point with new experiments, and uploaded the revised version of the paper with new text marked in blue.

Evidence beyond one benchmark. The gap is now swept from 0 to 6 on ActFlow's own benchmark, and QM9 is a full second benchmark with ActFlow at four and 100 steps and every checkpoint scored (Sec. 4.6, Table 7). The 168 and 174 centers you compared come from a coarse packing that cannot distinguish methods, since every four-step method lands between 156 and 188. The count that separates them is new valid clusters, and there EXPANSE finds 2.2× as many as the frozen teacher (1438 against 645, p<0.001) and keeps growing over the loop. Appendix D.3 defines both counts. At gap 6 ActFlow covers more of the component at τ=0.05 (61.5 against 23.9, p=0.023), while at τ=0.002 both cover nearly all of it (98.6% and 98.1%), and EXPANSE stays ahead in total coverage (Table 6). Rec-F reaches a higher raw count than EXPANSE (1564 against 1438) because of its higher validity, and at a matched pool of 100 valid molecules the two tie (57.8±3.7 against 55.0±5.4).

Table 6. Larger gaps, α=1, ρ=Δ+0.5 (uses the gap, not available in practice). ActFlow: w=100, 3 seeds.

ΔComp. no forceComp. forceComp. force, τ=0.002Total no/with forceActFlow comp./total
40.066.8±7.999.578.9/85.171.4/69.5
60.023.9±24.298.177.3/79.561.5/68.0

Table 7. QM9, 4 NFE unless noted, equal compute, α=1.

MethodNew clusters, start → lastCenters, lastValid %
EXPANSE1235 → 1438±8118869.0
Rec-F1231 → 1564±22718078.0
ActFlow620 → 230±231565.9
ActFlow, no uncertainty620 → 253±311566.5
ActFlow, 100 NFE47 → 107±106787.3
ActFlow, no unc., 100 NFE44 → 99±145681.0
Frozen teacher64517427.9

Reach versus density spreading. The disconnected component isolates reach, and we report it separately everywhere. Seed noise without the probe covers 96% of the connected region and none of the component, while the full method covers 84% of the component (three seeds; 80.5% over five seeds, Table 4) with the same replay, update and thinning. Thinning therefore cannot explain it. ActFlow's ODE sampler also reaches the component, which Fig. 1 and Sec. 4.2 now state, and EXPANSE is still ahead on the component at gaps 0 to 2, significantly so at gap 1 (Table 3).

Operating points. Tables 1 to 3 report α=1 and α=1.25 at every gap with total coverage, component coverage, validity and tests (validity is the third entry in Table 1 here and is Table 2 in the paper). At α=1 ActFlow is 1.5 to 4.3 points more valid. At α=1.25 EXPANSE is more valid than ActFlow at every gap and covers 5 to 11 points more, so the lead does not depend on the operating point.

Table 1. Total / component coverage / validity (%), τ=0.05. EXPANSE: 5 seeds, coverage force. ActFlow: released code, 100 steps, 3 seeds (5 at Δ=2). Pre-trained (EXPANSE's own initial generator): 2.4 / 0.0.

MethodΔ=0Δ=0.5Δ=1Δ=2
ActFlow w=10073.9/80.1/96.172.1/74.5/97.262.4/69.5/96.769.8/74.8/96.5
ActFlow w=1361.3/89.4/97.167.7/78.8/95.271.2/74.6/96.378.1/72.3/96.5
EXPANSE α=192.6/91.4/94.691.5/87.1/92.989.8/85.9/93.490.4/83.5/93.3
EXPANSE α=1.2582.2/82.1/99.182.8/80.0/97.981.5/80.1/97.583.4/74.7/97.8

Table 3. Welch tests, EXPANSE minus ActFlow (points).

Metric (α)ΔActFlowDiffp
Total (1)0w=100+18.70.018
Total (1)1w=100+27.40.017
Total (1)2w=100+20.70.0004
Total (1)2w=13+12.40.0014
Component (1)1w=100+16.40.03
Component (1)1w=13+11.30.035
Component (1)2w=100+8.70.07
Total (1.25)2w=13+5.30.029
Validity (1.25)2w=13+1.30.022
Component (1)6w=100−37.60.023

Attribution. We ran the swap you proposed. Our probe on ActFlow's sampler, with everything else in ActFlow unchanged, lifts its total coverage from 62 to 91 at gap 1 and from 70 to 86 at gap 2 (Table 5). The free positives do not drive the result, since removing them moves component coverage by under two points (80.5 with them, 78.8 without).

Table 5. Total / component coverage / validity (%), α=1. Only the query points change in ActFlow.

MethodΔ=1Δ=2
ActFlow w=10062.4/69.5/96.769.8/74.8/96.5
ActFlow w=1371.2/74.6/96.378.1/72.3/96.5
ActFlow + our probe90.5/86.9/96.086.1/56.3/97.1
EXPANSE89.8/85.9/93.490.4/83.5/93.3
AuthorsOpenReview ↗

Rebuttal response for Reviewer mA4z (continued)

Theory. Theorem 3.4 is now two-sided, which also settles the direction you raised in contribution (3). The capture mass M_C is the expected number of probes that land in a region given the history, and it needs no labels. The expected number of queries in the region equals the expected capture mass E[M_C], and the seed mass within travel distance bounds it from above for any acquisition rule. With probes independent given the history, as ours are, the chance of missing the region while M_C ≥ λ is at most e^{−λ}. Reach therefore needs an expected capture mass of order one, and a capture mass that is almost surely at least log(1/η) gives it with probability 1−η. The measurements agree. At gap 2 the capture mass predicts 747 queries in the component against 755 logged, and in 40 further runs the share of runs whose capture mass before the first hit exceeds λ never exceeds e^{−λ} (Appendix B.1, Fig. 9). The capture mass is the tight quantity and the geometric bound is its closed-form envelope, about ten times larger at the main setting. Corollary 3.5 is retained as the qualitative statement of what must bridge the gap, and Sec. 3.5 says plainly that it is vacuous at the main setting.

Prior term and Sec. 3.2. All main results use β=0, and the sweep supports that choice since raising β costs the component (Fig. 7c). Sec. 3.4 and Appendix B.4 state exactly what Proposition 3.3 describes, and the claim that a one-step generator forces the separation is removed.

Consistency and step count. Across seeds the standard deviation of EXPANSE's total coverage is 0.6 to 1.7 points against 3.6 to 32.0 for ActFlow, and 2.0 to 3.1 against 4.4 to 20.9 on the component. At one step ActFlow covers 0.3% of the valid set at gap 2, and its best total coverage is 70.7% at 10 steps. Against that 10-step operating point EXPANSE at α=1.25 leads on total coverage (83.4 vs 70.7), component coverage (74.7 vs 60.4) and validity (97.8 vs 97.2) at 35× its throughput, so the lead does not rest on a 3-point margin at one operating point. Both loops start at the same level: EXPANSE's initial generator gives 2.4/0.0 and ActFlow's base flow 2.6/0.0 at gap 2.

QM9 over the full loop. Every checkpoint is now scored, including the last (Table 7). EXPANSE grows from 1235 to 1438 new clusters (p=0.017), by 7% to 23% in every seed, while ActFlow at four steps falls by 63%.

Verifying these numbers. Each table here regenerates from the supplementary zip with one command: python -m drift.dump_numbers table and stats for Tables 1 to 3, afprobe for Table 5, free for the free-positive control, qm9 for Table 7, and python -m drift.capture --plot for the capture-mass check.

Specific questions. (1) Appendix D.3 defines both counts. Centers at τ=0.2 are the coarse packing, and new clusters at cutoff 0.4 count valid molecules outside the reference. (2) Tables 1 and 3 report every method at both operating points with the component separate. (3) We tested the free positives from the other side by removing them from EXPANSE, which moves component coverage by under two points. They contain no component point and never enter the occupancy posterior. (4) Seed scale and probe schedule were set at gap 2 and reused unchanged at gaps 0, 0.5 and 1. Travel at gaps 3 to 6 is set from the gap, and those rows test consolidation. The coverage force was set at gaps 4 and 6 and reused unchanged elsewhere. (5) ActFlow's two-dimensional uncertainty is label-free too, and Appendix B.2 now says so. What differs is the field the probe climbs, a persistent multi-scale occupancy used as a gradient with an explicit travel budget.

Seed noise versus ascent on ActFlow's sampler. To separate the seed perturbation from occupancy ascent in Table 5, we ran ActFlow with seed noise only (σ0=1, zero travel, everything else as in Table 5, three seeds). It gives 62.9/29.4/97.0 at gap 1 and 58.3/0.0/97.0 at gap 2 (total/component/validity), against 90.5/86.9/96.0 and 86.1/56.3/97.1 with ascent. The lift on ActFlow comes from occupancy ascent, not from seeding.

The typographical errors are fixed, the Pre-trained row is EXPANSE's own initial generator, and Appendix B.3 quantifies the batch-size effect.

Please let us know if any further experiments or clarification are required during the discussion period or before the camera ready version. We will be happy to discuss and perform further experiments.

Thank you for the extensive revision. The released-code ActFlow baseline, the probe-on-ActFlow swap, and the operating-point tables resolve several of my concerns, and I have raised my Soundness score. I keep my rating for four reasons:

  • The probe shows no benefit outside 2D. On QM9, Rec-F (EXPANSE without acquisition and mixture) matches EXPANSE per valid molecule (55.0 vs. 57.8) and exceeds it in raw count (1564 vs. 1438).
  • Component coverage versus ActFlow is inconsistent. EXPANSE is significantly ahead at Δ=0.5 and 1, not at Δ=2 (p=0.07), and behind at Δ=6 (p=0.023). The probe also lowers ActFlow's component coverage at Δ=2 (74.8 to 56.3).
  • Threshold. At τ=0.01 all methods except ActFlow (SDE) exceed 93% total coverage, and at τ=0.002 both methods cover ~98% of the component at Δ=4 and 6. Could you report component coverage at τ=0.01 for Δ≤2? If ActFlow also saturates there, the gains at τ=0.05 reflect density allocation rather than reach.
  • No held-out gap. Hyperparameters were set at Δ=2, 4 and 6, and ρ from the gap for Δ≥3

A non-toy result where the probe improves over Rec-F would change my assessment.

AuthorsOpenReview ↗

Response to Reviewer mA4z

Dear Reviewer mA4z, thank you for raising Soundness and for naming the tests that matter. We ran each one.

QM9 (point 1). Rec-F is EXPANSE's own consolidation stage: the same distilled sampler, drifting update and replay, without acquisition and the rejection mixture. The 6.3× and 13.4× margins over ActFlow are EXPANSE's, and Rec-F inherits them. To isolate the two remaining components we ran the full factorial (seeds 0–2, last checkpoint, equal compute):

ArmNew clustersPer 100 validValidity (%)
Rec-F1552 ± 27753.3 ± 5.179.6
+ probe1285 ± 42353.3 ± 12.561.7
+ mixture1110 ± 1348.3 ± 2.371.1
+ probe + mixture (EXPANSE)1474 ± 4857.7 ± 4.570.1

This is the non-toy probe result you asked for. With the mixture in place, adding the probe raises new clusters from 1110 to 1474 (Welch p=0.0035) and discovery per valid molecule from 48.3 to 57.7 (p=0.049). Neither component alone reaches the full method. Together they give the highest discovery per valid molecule of all four arms and a 5.8× smaller seed spread than Rec-F (48 against 277). The probe proposes beyond the generator's mass, and the mixture keeps the generator off what the verifier rejects. The inference-time α dial does not transfer to QM9, as the paper states, but the mixture's training-time role does. QM9 is ActFlow's molecular task and is used here for a direct comparison. Rec-F's undirected queries are already 76% valid there, so QM9 does not isolate reach. The disconnected component does, and the held-out gaps and the controlled swap below test it directly.

Held-out gaps (point 4). Every setting is frozen exactly as in Table 1: seed scale and probe schedule set at Δ=2, coverage force set at Δ=4 and 6, and travel ρ=2.0. Δ=1.5 and 2.5 were never used for any choice. At Δ=2.5 the travel budget is shorter than the gap.

τ=0.05, α=1Δ=1.5 total / comp. / val.First hitΔ=2.5 total / comp. / val.First hit
ActFlow w=10071.3±6.4 / 71.3±7.5 / 96.719867.9±6.3 / 69.3±6.7 / 96.3517
ActFlow w=1375.0±4.0 / 77.3±1.6 / 96.463573.4±5.8 / 69.3±3.0 / 96.0836
EXPANSE90.4±1.5 / 86.6±3.2 / 94.16488.8±2.0 / 74.6±5.8 / 93.986

EXPANSE (5 seeds; ActFlow 3) leads total coverage by 15 to 21 points against both variants (p=0.016 to 0.037). It leads the component at both gaps, by 9.3 points over the better variant at Δ=1.5 (p=0.0016), and reaches it 3 to 10 times sooner. Its held-out coverage is within 1.6 points of its tuned values, so the settings transfer without retuning.

Threshold (point 3). At τ=0.01 (about 8 of 200k samples per cell), component coverage at Δ≤2 is 98.5–98.8% for EXPANSE, 95.2–99.9% for w=100 and 98.7–99.2% for w=13, so every method touches the component. The generable set Ω^τ is defined by density, and the task is to place density across the valid set. That is where EXPANSE wins. At Δ=2 and τ=0.02 it covers 98.9% total and 96.6% of the component, against 91.6 / 94.7 (w=100) and 96.1 / 92.5 (w=13). At τ=0.05 it leads by 12 to 21 points. Reach is measured where it is the bottleneck: on ActFlow's own sampler, with everything but query placement fixed, seed noise reaches 0.0% of the component at Δ=2 and occupancy ascent reaches 56.3%.

Component coverage (point 2). With the held-out gaps added, EXPANSE is ahead of the better ActFlow variant on the component at all six gaps from 0 to 2.5. The lead is significant at Δ=1 (p=0.035) and Δ=1.5 (p=0.0016). At Δ=6, where travel is set from the gap, EXPANSE still leads total coverage by 11.5 points over w=100 (p=0.027). In the swap, ActFlow's 74.8 comes from its guided trajectory, which is switched off for probe queries. The matched reference, the same seeding without ascent, reaches 0.0% of the component. Ascent lifts it to 56.3%, and total coverage from 58.3 to 86.1.

All new tables will go into the final version, each with one-command reproduction in the supplementary material. We have also posted the GenBO regularizer results requested by Reviewer QiLn in that thread, and would be grateful if you considered them as well.

Thank you for the active discussion, and please tell us anything else you would like to see in the final version. We used extensive computation resources to run everything on demand in the last few days. We would really appreciate if the reviewer can consider raising the score.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Meta response to AC and reviewers

Dear Area Chair and reviewers,

Thank you for the careful reviews and constructive feedback. We have addressed each point with new experiments, and uploaded the revised version of the paper with new text marked in blue. Your feedback helped us make it a substantially stronger paper.

Revised version updates

  • Theory. Theorem 3.4 now proves upper and lower probability bounds on reaching a region, through one label-free quantity, and measurements match it (Sec. 3.5, Appendix B.1, Fig. 9).
  • Attribution. Random and distance-based exploration, our probe on ActFlow's own sampler, and removal of the free positives and of replay separation are added to the paper.
  • Baselines. ActFlow is rerun from its released code at both guidance weights, both samplers and one to 100 steps, with three seeds for ActFlow (five at gap 2), five for EXPANSE, and significance tests.
  • Larger gaps. The gap is swept to 6, and a new coverage force lets the generator absorb regions far from its mass.
  • QM9. Every method is scored under equal compute at every checkpoint, against ActFlow at four and 100 steps, its own recursive baseline, a matched baseline of ours and the frozen teacher.
  • Writing. The paper is substantially rewritten with a short introduction, a notation table, one algorithm for a full round and six experiment subsections.
  • Reproducibility. The supplementary zip regenerates every table and figure in the paper and in these responses from the logged runs, without retraining or checkpoints. It contains the code, the arguments and training log of all 362 runs, the scored QM9 results for every checkpoint and seed, and cached coverage histograms. Its README maps each table and figure to one command, and jobs/rerun.py rebuilds the exact training command of any run from its saved arguments, so any row can also be retrained.

What the results show

  • EXPANSE covers 12 to 19 points more of the valid set than the better ActFlow variant at gaps 0 to 2, in one step and with a several times smaller seed spread, and it stays ahead in total coverage out to gap 6.
  • The probe is portable. On ActFlow's own sampler it lifts ActFlow's total coverage by 8 to 28 points, and seed perturbation alone on the same sampler covers 0% of the component at gap 2, so the lift comes from occupancy ascent.
  • Occupancy ascent beats random and distance-based exploration on the disconnected component by 21 and 70 points.
  • The capture mass, computed from the probe alone without labels, matches the logged queries in the component, 747 against 755.
  • One model trades coverage for validity at inference, and at α=1.25 it is both more valid and broader than ActFlow at every gap from 0 to 2.
  • On QM9 EXPANSE finds 6.3× as many valid new clusters as ActFlow at four steps and 13.4× as many as ActFlow at its native 100 steps.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.