acceptodds
Under review as a conference paper at ICLR 2027

BFMT: Enhancing Search Capabilities of Tree Sampler via Bootstrap Flow-Map Tree

Abstract

The ability to efficiently navigate high-dimensional spaces to identify optimal candidates—whether designing targeted molecular structures or generating images with preferred attributes—stands as a central pillar of modern scientific progress. Recent tree-based samplers enable highly scalable inference-time search by eliminating the reward-gradient bottleneck inherent to particle-based methods. Despite their potential, existing tree-based samplers are fundamentally constrained by the massive number of function evaluations (NFEs) required for node valuation, crippling their utility in exploration-heavy tasks. Furthermore, their reliance on small, uniform transitions at each depth precludes dynamic, adaptive search capabilities. To address this, we introduce Bootstrap Flow-Map-Tree (a.k.a BFMT), a novel computationally efficient sampling framework that enables full tree-path construction from any tree depth using a single function evaluation, drastically reducing computational overhead while providing critical foresight for sequential sampling. By enabling dynamic transition time-step scheduling, BFMT efficiently allocates its sampling budget, smoothly transitioning from broad global exploration to fine-grained local refinement of high-utility modes discovered through exploration. Extensive experiments and ablation studies across diverse domains—from high-dimensional images to large-scale molecules—demonstrate BFMT's superiority over baseline approaches in both search and alignment tasks.

Primary area:
generative models
Keywords:
Flow Maps, Tree Search, Inference Time Search

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_m4ebOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper addresses the high cost of evaluating candidate solutions under limited search budgets in tasks such as drug discovery and image generation. It proposes BFMT, a training-free sampling framework that combines generative models with tree search. BFMT uses Flow Maps to reduce node evaluation costs and a bootstrapped sufficient-statistic mechanism to construct stochastic branches that preserve diversity. Guided by historical feedback, non-uniform generation steps, and the remaining budget, it adaptively shifts from broad exploration to fine-grained local refinement, aiming to find more high-quality candidates with limited computation and feedback.

Critical strengths
  1. theoretical results are strong, such as "the sampling error decays polynomially with respect to the total sampling budget"

  2. this paper is well written.

  3. the studied problem is important and related works are well presented.

Critical weaknesses
  1. do not use underlines in the title. do not use colors in the main paper. too many color using for texts and formulations! make this paper looks informal. The overall writing style is a bit strange, like a teaching note in top US university (text box using etc).

  2. for "Budget-Aware Node Selection", is there some theoretical guarentee?

  3. line 239, Starting -> starting

  4. too many descriptive words, it is recommended to use some psedo codes. can not follow very well.

  5. QED, SA attributes are too easy to attain in the python rdk toolkit, will them really represent the true merit of a mol. how far is the designed mol. from real world application like drug design.

  6. position of related works is a bit strange.

  7. too many contributions. it is better to use three to summarize them.

Questions for authors
  1. will the proposed UCT tree is proved to find the global optimal?

  2. mol. design papers (related works) are substantial, e.g., https://arxiv.org/html/2609.01608v1.

the compared baselines are too small. only one is used to do comparison.

  1. is it possible to include some Bayesian opt. methods as baselines?

  2. is there any numerical results for image tasks?

  3. how is the scaling law of the proposed method, like for very difficult math/code tasks, like finding a counter case for difficult math problem

4 comments
AuthorsOpenReview ↗

Response to Reviewer m4eb (Part1: W1 to W7, Q1)

We thank m4eb for the detailed feedback. We are pleased that Reviewer m4eb finds the problem important and the theoretical results strong. Below, we clarify all the raised concerns.

W1: do not use underlines in the title. do not use colors in the main paper. too many color using for texts and formulations! make this paper looks informal. The overall writing style is a bit strange, like a teaching note in top US university (text box using etc).

We thank the reviewer for the feedback regarding the visual presentation and style. Following your suggestions, we have revised the manuscript accordingly: (a) removed all non-standard colors from the text and mathematical formulations; (b) removed the background colors from text boxes; and (c) removed the underline from the title.

W2: for "Budget-Aware Node Selection", is there some theoretical guarantee?

Thanks for the question. Budget-aware node selection is designed for the finite-budget regime. UCT's exploration bonus is calibrated for an unbounded horizon and has no notion of a depleting budget. BASE addresses this directly by tying the exploitation strength to the remaining budget. It explores broadly when resources are plentiful and concentrates on high-value subtrees as the budget runs out. While it does not have a strict guarantee, Fig. 5 shows that this yields consistently higher reward than UCT at every feedback budget we test.

W3: line 239, Starting → starting

Thanks for catching this. We have fixed this in the updated version.

W4: too many descriptive words, it is recommended to use some pseudo codes. can not follow very well.

We appreciate your suggestion. We agree that including pseudo-code makes the approach much easier to follow. Accordingly, we have added pseudo-code to Appendix Q in the revised manuscript.

W5: QED, SA attributes are too easy to attain in the python rdk toolkit, will them really represent the true merit of a mol. how far is the designed mol. from real world application like drug design.

We thank the reviewer for raising this point. BFMT is a general-purpose, reward-agnostic sampler for feedback-efficient sampling; the molecular experiment is intended to show that it carries over from continuous (image) to discrete (molecular) generation without modification. QED, SA and their composition are standard and intuitive objectives for benchmarking molecular generation, in a discrete-data domain that complements our image experiments.

That said, we agree that a more application-specific objective strengthens the evaluation. Following the reviewer's suggestion, we evaluated BFMT on a Vina docking-score task, a considerably more demanding objective. For this task, we selected protein 4PXX (human carbonic anhydrase II), a well-characterized protein target with a structurally defined binding site and known small-molecule inhibitors, making it a suitable benchmark for docking-guided molecular generation. We compared CFM and BFMT on this task, steering both models towards generating molecules with more favorable (i.e., lower) docking scores, which may indicate stronger binding affinity to the target protein. BFMT improves over CFM under the same feedback budget, as reported in the table below and the density figure in the Appendix. We have added these results and the corresponding discussion to the revised manuscript (see Appendix T).

Table: Protein 4PXX docking experiments (lower is better).

Method200 NFE600 NFE1000 NFE
CFM-6.892-6.767-6.782
BFMT-7.2429-7.2411-7.2855

W6: position of related works is a bit strange.

Thank you for the helpful feedback regarding the structure. While we placed the related work section after the experiments to maintain a continuous narrative flowing directly from the introduction's motivation into the methodology, we completely agree that readability is paramount. We would be very happy to move the related work section before the method section in the final manuscript if the reviewer feels it improves the flow for the reader.

W7: too many contributions. it is better to use three to summarize them.

Thank you for the suggestion. Accordingly, we have restructured the contributions into three distinct bullet points in the updated manuscript to improve readability.

Q1: will the proposed UCT tree is proved to find the global optimal?

The proposed BFMT tree sampler guarantees convergence in total variation distance toward the target optimal policy rather than strict global optimality. We have stated this distinction explicitly in the revised manuscript (see Lines 270-271).

AuthorsOpenReview ↗

Response to Reviewer m4eb (Part2: Q2 to Q5)

Q2: mol. design papers (related works) are substantial, e.g., https://arxiv.org/html/2609.01608v1. The compared baselines are too small. only one is used to do comparison.

Thank you for this question. While our work focuses specifically on inference-time sampling, the suggested approach relies on fine-tuning, representing a distinct methodological direction. Nonetheless, we appreciate the pointer: we have added a paragraph on fine-tuning methods to the related work, including the suggested reference (lines 479-490).

And because BFMT functions as an inference-time sampler capable of building upon any state-of-the-art flow-based generative model, we adopt the recently introduced flow-based generative model (i.e., CFM) as the underlying generative prior. Comparing our approach directly against CFM therefore allows us to isolate and evaluate the specific effectiveness of the BFMT framework. The purpose of the molecular experiment is to demonstrate the generalizability of BFMT across different domains, and the comparison with CFM further highlights its effectiveness in a fundamentally different modality. In contrast, we provide a more extensive experimental comparison with existing state-of-the-art approaches across two different computer vision problem settings.

Q3: is it possible to include some Bayesian opt. methods as baselines?

Thank you for this question. Bayesian Optimization approaches do not fit in this setting. BFMT's objective is to sample from a reward-tilted distribution, so the goal is to produce diverse, high-reward samples, which is suitable in settings where the deliverable is a diverse batch and/or the reward is a proxy. However, Bayesian optimization directly optimizes for the single candidate that maximizes the reward, with no distributional guarantee.

Q4: is there any numerical results for image tasks?

Yes. We report numerical image generation results on ImageNet-1K search, compositional alignment (GenAI-Bench prompts) tasks, and quantity-aware alignment (T2I-CompBench++ prompts) tasks, against DTS, FKS, DAS, and MFM, as functions of both NFE and feedback budget (Figs. 3, 10–12). We report additional quantitative ablation study results, e.g., BASE vs UCT in Fig. 5, impact of BSS in Fig. 6, and impact of dynamic transition steps in Fig. 7.

While these results are illustrated graphically in the main text, in response to the reviewer's comment, we have also compiled the corresponding numerical values into tables for both NFEs and feedback budgets of 100, 300, and 500. Each entry reports the mean and standard error computed across 5 random seeds, and these new tables are included in Appendix U.

ImageReward results for search experiments (NFE).

Method100 NFE300 NFE500 NFE
DAS-2.403 ± 0.072-2.141 ± 0.087-1.663 ± 0.091
FKS-2.243 ± 0.127-2.185 ± 0.106-1.788 ± 0.087
DTS-1.544 ± 0.153-1.585 ± 0.121-1.342 ± 0.113
BFMT0.336 ± 0.0530.592 ± 0.0350.694 ± 0.020

ImageReward results for search experiments (feedback budget).

Method100 budgets300 budgets500 budgets
DAS-2.234 ± 0.083-1.874 ± 0.077-1.345 ± 0.079
FKS-2.183 ± 0.092-1.638 ± 0.087-1.540 ± 0.073
DTS-1.346 ± 0.095-0.332 ± 0.083-0.149 ± 0.069
BFMT0.336 ± 0.0530.592 ± 0.0350.694 ± 0.020

ImageReward results for alignment experiments (NFE).

Method100 NFE300 NFE500 NFE
DAS-1.037 ± 0.063-0.233 ± 0.0660.204 ± 0.058
FKS-0.974 ± 0.0520.279 ± 0.0770.330 ± 0.067
DTS0.408 ± 0.0840.587 ± 0.0570.705 ± 0.070
BFMT1.597 ± 0.0121.725 ± 0.0071.789 ± 0.005

ImageReward results for alignment experiments (feedback budget).

Method100 budgets300 budgets500 budgets
DAS0.217 ± 0.0360.567 ± 0.0410.980 ± 0.027
FKS0.343 ± 0.0290.375 ± 0.0331.037 ± 0.031
DTS0.704 ± 0.0330.836 ± 0.0401.240 ± 0.021
BFMT1.597 ± 0.0121.725 ± 0.0071.789 ± 0.005

Q5: how is the scaling law of the proposed method, like for very difficult math/code tasks or finding a counter case for difficult math problem.

Thank you for the question! We agree that exploring scaling behavior on complex reasoning tasks, such as finding counterexamples for challenging mathematical problems, is an exciting direction. The core motivation of this work is to establish a principled, computationally efficient foundation for budget-constrained tree sampling. Because BFMT is designed without domain-specific assumptions, we evaluated it across two diverse high-dimensional settings (continuous image generation and discrete molecule design), where it consistently demonstrated strong effectiveness. We look forward to investigating BFMT in these demanding math and language reasoning domains in future work.

Reviewer_m4ebOpenReview ↗

thx for rbt

thx for your rbt, this work do have some contribution, but the experiments are still a bit weak to me. tend to keep current rating.

Thanks for recognizing the contribution of the paper. However, we would greatly appreciate if you could point out which aspects of our experiments you feel remain insufficient, as we are eager to address any remaining reservations while the discussion window remains open.

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

For inference-time alignment of diffusion/flow models, search-based methods require a large number of function evaluations (NFEs) as search scales. The paper proposes Bootstrap Flow-MapTree, which utilizes a bootstrapped sufficient-statistics scheme and requires only a single function evaluation regardless of tree depth. The method is validated through experiments on both image and molecule generation.

Critical strengths

The proposed method addresses a critical problem in tree-search particle-based methods for inference-time scaling: the number of function evaluations (NFEs) also scales with the search. It utilizes Progressive BSS transitions, requiring only one NFE to estimate x_0 while generating different x_t-1 samples. The method is solid and comes with theoretical guarantees.

Critical weaknesses
  1. The paper does not report quality metrics beyond the reward itself, which makes it difficult to assess whether the method suffers from reward over-optimization.

  2. The molecular reward appears relatively simple. If possible, the authors could include a more challenging objective, such as docking score, to better evaluate the method under harder optimization settings.

Questions for authors

Please see weakness

2 comments
AuthorsOpenReview ↗

Response to Reviewer eUqq

We thank eUqq for the constructive feedback and are pleased that Reviewer eUqq found our method solid. We clarify all the raised concerns below.

W1 & Q1: The paper does not report quality metrics beyond the reward itself, which makes it difficult to assess whether the method suffers from reward over-optimization.

We thank the reviewer for this excellent question. To verify that BFMT remains robust to reward over-optimization, we evaluate its search and alignment performance using independent, held-out standard reward models, specifically CLIP-FlanT5 (i.e., VQA score) and PickScore, while BFMT is constructed using ImageReward. As summarized in the tables below, BFMT consistently outperforms the baseline methods across these held-out reward models under both NFE and feedback budget constraints (in both search and alignment settings). These generalization results indicate BFMT's ability to avoid reward over-optimization compared to the baselines. We have included these comprehensive results in Appendix V.

Search evaluation over 500 NFEs.

MethodCLIP-FlanT5PickScore
DAS0.171517.7835
FKS0.183217.5484
DTS0.220717.9022
BFMT0.880620.7699

Search evaluation over 500 feedback budgets.

MethodCLIP-FlanT5PickScore
DAS0.418918.5094
FKS0.393718.2460
DTS0.485018.6251
BFMT0.880620.7699

Alignment evaluation over 500 NFEs.

MethodCLIP-FlanT5PickScore
DAS0.530318.3559
FKS0.581019.1204
DTS0.900420.3603
BFMT0.974521.6085

Alignment evaluation over 500 feedback budgets.

MethodCLIP-FlanT5PickScore
DAS0.823820.1166
FKS0.848320.3047
DTS0.912620.7001
BFMT0.974521.6085

W2 & Q2: The molecular reward appears relatively simple. If possible, the authors could include a more challenging objective, such as docking score, to better evaluate the method under harder optimization settings.

We thank the reviewer for this suggestion. The purpose of the molecular experiment is to show that BFMT generalizes across fundamentally different problem settings, from continuous (image) to discrete (molecular) data, without any change to the method. For the same reason, we used QED, SA, and their composition, which are intuitive (we want drug-like molecules that are easy to synthesize), allowing us to illustrate the method's capabilities without introducing domain-specific complexity.

However, we agree that a more challenging application-specific objective strengthens the evaluation. Following the reviewer's suggestion, we evaluated BFMT on a Vina docking-score task, a considerably more demanding objective. For this task, we selected protein 4PXX (human carbonic anhydrase II), a well-characterized protein target with a structurally defined binding site and known small-molecule inhibitors, making it a suitable benchmark for docking-guided molecular generation. We compared CFM and BFMT on this task, steering both models towards generating molecules with more favorable (i.e., lower) docking scores, which may indicate stronger binding affinity to the target protein. BFMT improves over CFM under the same feedback budget, as reported in the table below and the density figure in the Appendix. We have added these results and the corresponding discussion to Appendix T of the revised manuscript.

Table: Protein 4PXX docking experiments (lower is better).

Method200 NFE600 NFE1000 NFE
CFM-6.892-6.767-6.782
BFMT-7.2429-7.2411-7.2855

Dear Reviewer eUqq,

As the discussion period is nearing its end, we want to follow up to ensure that all raised concerns have been addressed satisfactorily. If any residual questions or concerns remain, we would be happy to clarify them before the discussion period closes.

Thanks,

Authors

Reviewer_z5fTOpenReview ↗
Rating
3(+1)
Confidence
1
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper introduces Bootstrap Flow-Map Tree (BFMT), a training-free inference-time sampling framework for feedback-efficient search and alignment using pretrained generative models. BFMT combines conditional flow maps with bootstrap sufficient statistics, allowing a complete stochastic denoising trajectory to be constructed from a node with a single neural function evaluation. Therefore BFMT avoids computationally expensive denoising rollouts to estimate intermediate node values. It also introduces dynamic timestep scheduling to transition from broad exploration to local refinement, and a budget-aware node-selection strategy that incorporates feedback from previously evaluated samples. The authors provide theoretical justification and evaluate the approach on image generation and molecular optimization tasks. Experiments demonstrate improved reward scores and computational efficiency compared with relevant baselines.

Critical strengths
  1. BFMT significantly reduces the computational cost of tree search compared to baselines.

  2. The proposed methodology is supported by theoretical analysis, including derivations establishing the validity of the stochastic transitions and convergence guarantees for the sampling policy.

  3. The paper presents an extensive experimental evaluation, particularly for image generation, with comparisons against several competitive baselines and ablation studies. BFMT consistently demonstrates improved reward scores under constrained evaluation budgets.

Critical weaknesses

This section focuses primarily on molecular generation, and I acknowledge the paper's main emphasis on performance and computational efficiency in image synthesis. I am open to increasing the score if the concerns below are addressed or clarified.

  1. The molecular generation experiments appear more like a proof-of-concept case study rather than a comprehensive evaluation against previous molecular generation methods.

  2. Several details of the molecular generation experiments are missing:

    • While the authors consider QED-only and combined QED–SA optimization, the precise formulation of the combined reward is not specified.
    • Synthetic accessibility scores are not reported separately, making it difficult to assess the contribution of each property to the overall improvement.
    • Additional metrics such as structural diversity and novelty are missing.
    • Vanilla CFM sampling consistently achieves higher uniqueness than BFMT across the reported experiments, but this is not acknowledged or discussed.
  3. The claim that BFMT can explore diverse modes is primarily supported by qualitative examples in image generation. Quantitative diversity measurements would provide stronger evidence for this claim.

  4. Minor presentation concerns: The results section is quite crowded, and the transition between molecular generation and image alignment experiments could be clearer. Tables reporting molecular generation metrics have one-line captions.

Questions for authors
  1. How exactly is the combined QED–SA reward calculated?
  2. Is there a computational cost comparison (e.g. NFEs or inference time) between BFMT and vanilla CFM for molecular generation?
6 comments
AuthorsOpenReview ↗

Response to Reviewer z5fT (Part1: W1 to W2.3, Q1)

We thank z5fT for the detailed feedback. We have addressed all the raised concerns with additional experimental support and updated the manuscript to incorporate these new analyses.

W1: The molecular generation experiments appear more like a proof-of-concept case study.

Thanks for raising this point. Yes, the purpose of the molecular experiment is to show that BFMT generalizes across fundamentally different problem settings, from continuous (image) to discrete (molecular) data, without any change to the method. We used QED, SA, and their composition as the reward, which are intuitive (we want drug-like molecules that are easy to synthesize), allowing us to illustrate the method's capabilities without introducing domain-specific complexity. However, we agree that a more application-specific objective in the molecule generation domain strengthens the evaluation. Following the reviewer’s suggestion, we evaluated BFMT on a Vina docking-score task, a well-established domain-specific objective. For this task, we selected protein 4PXX (human carbonic anhydrase II), a well-characterized protein target with a structurally defined binding site and known small-molecule inhibitors, making it a suitable benchmark for docking-guided molecular generation. We steer CFM and BFMT towards generating molecules with more favorable docking scores. BFMT improves over CFM under the same feedback budget. We have discussed about it and reported the results in Appendix T.

W2.1 & Q1: How exactly is the combined QED–SA reward calculated?

Thanks for the detailed question. We define the combined objective as the geometric mean of QED and a normalized synthetic accessibility score rSAr_{\mathrm{SA}}, as defined below:

Score(x)=QED(x)⋅rSA(x),rSA(x)=clip(10−SA(x)9, 0, 1).\mathrm{Score}(x) = \sqrt{\mathrm{QED}(x)\cdot r_{\mathrm{SA}}(x)}, \qquad r_{\mathrm{SA}}(x) = \mathrm{clip}\left(\frac{10 - \mathrm{SA}(x)}{9},\, 0,\, 1\right).

We have included this detail in Lines 397-400. The geometric mean has been previously used in prior works (see e.g., REINVENT 4, Loeffler et al., 2024).

W2.2: Synthetic accessibility scores are not reported separately, making it difficult to assess the contribution of each property.

Thank you for this comment. We have now reported the individual QED and SA scores separately in the following tables, allowing the contribution of each property to the overall improvement to be assessed independently through another run. In particular, BFMT achieves a higher mean QED score and a lower (better) mean SA score than CFM, resulting in an improved combined Score. We have also clarified that the combined Score is calculated as the geometric mean of QED and SA, as defined in Equation 13 in the revised draft. Additionally, we report novelty (N), which indicates the fraction of generated compounds that are not present in the training set. The updated results are reported below.

Table: Impact of individual reward metrics and novelty (N) with 1000 feedback budgets.

Reward r = QED:

MethodN ↑Mean QED ↑
CFM0.98760.4796
BFMT0.99260.5620

Reward r = Score:

MethodN ↑Mean QED ↑Mean SA ↓Score ↑
CFM0.98980.48884.77890.5325
BFMT0.99480.54864.55590.5761

Additionally, we plot the distributions of QED and SA scores separately to show the overall distributional behavior and analyze the distributional shift. We have included these additional analyses in Appendix Section R. The plots further strengthen our conclusion regarding BFMT's ability over composition of different reward metrics.

W2.3: Additional metrics such as structural diversity and novelty are missing.

Great point. We have now added novelty results to Table 6 in the Appendix (see our previous answer). BFMT achieves higher novelty in both evaluation settings (0.9926 vs. 0.9876 and 0.9948 vs. 0.9898), indicating that both methods generate predominantly novel molecules, with BFMT achieving a slight improvement.

To further investigate the structural diversity of the generated compounds, we conduct two additional analyses based on ECFP4 fingerprints. First, for each generated compound, we computed its Tanimoto similarity to every compound in the training set and retained the maximum similarity. The average maximum similarity is 0.40 for BFMT and 0.39 for CFM, suggesting that the generated molecules are not simply reproducing compounds from the training set.

Second, we assessed internal diversity by computing the maximum pairwise Tanimoto similarity among the generated compounds. This value is 0.32 for BFMT and 0.25 for CFM, indicating that the most similar pair of generated compounds is somewhat more similar for BFMT than for CFM.

The full distributions of maximum fingerprint similarities to the training set are reported in Appendix S. These analyses provide complementary insights into the novelty and structural diversity of the generated compounds.

AuthorsOpenReview ↗

Response to Reviewer z5fT (Part2: W2.4 to W4, Q2)

W2.4: Vanilla CFM sampling consistently achieves higher uniqueness than BFMT across the reported experiments, but this is not acknowledged or discussed.

Thank you for pointing this out. We would also like to clarify that uniqueness is calculated among valid molecules only. Therefore, the uniqueness metric alone does not reflect the fraction of all generated molecules that are both valid and unique. For example, in the combined-reward setting, the fractions of generated molecules that are both valid and unique are 0.9365 for CFM and 0.9297 for BFMT. Thus, while CFM achieves slightly higher uniqueness, the difference in the fraction of valid and unique molecules is relatively small (0.0068 points). We consider this a reasonable trade-off given that BFMT achieves a higher overall score (0.5746 vs. 0.5173 for CFM), indicating better generated compound quality according to the combined QED and SA objective. We have added a discussion about this in the main text (see Lines 415-422).

W3: The claim that BFMT can explore diverse modes is primarily supported by qualitative examples in image generation. Quantitative diversity measurements would provide stronger evidence for this claim.

We thank the reviewer for raising this great point. We agree that reporting quantitative diversity measurements will further strengthen the conclusion. Following your suggestion, we compute the LPIPS diversity score for the search problem with feedback budgets of 300 and 500. As shown in the table below, BFMT outperforms the baselines by a significant margin across different feedback budgets. These quantitative findings highlight BFMT's capability to generate diverse high-quality images. We have included this analysis in Appendix P.

Table: Diversity analysis on the image generation search task.

Feedback budgetDiversity ↑FKSDASDTSMFM (BoN)Ours
300LPIPS0.23620.19293.07415.001838.4722
500LPIPS0.32830.21277.87977.057542.2586

W4: Minor presentation concerns: The results section is quite crowded, and the transition between molecular generation and image alignment experiments could be clearer. Tables reporting molecular generation metrics have one-line captions.

Thank you for the feedback. To improve the overall flow and clarity, we have added an introductory paragraph bridging the image alignment and molecular generation sections (see Lines 384-388). Additionally, we have expanded the caption for the table reporting molecule generation results to include explicit details regarding the feedback budget constraint.

Q2: Is there a computational cost comparison (e.g. NFEs or inference time) between BFMT and vanilla CFM for molecular generation?

Yes. To ensure a fair comparison, we keep the number of function evaluations the same for both CFM and BFMT across all the experiments. Concretely, we consider NFE = 1000 in Table 1, NFE = 200 in Table 2, and NFE = 600 in Table 4.

Reviewer_z5fTOpenReview ↗

Reply to Rebuttals by Reviewer z5fT

Thank you for comprehensively addressing the weakness points and the questions with additional experiments. As I'm reviewing the responses, I just have a few more comments, which are provided below:

we evaluated BFMT on a Vina docking-score task, a well-established domain-specific objective.

Thank you for including this experiment. BMFT appears significantly better in this case. Could you please clarify what is the reward function in this experiment? Is it just the docking score, or a combindation of docking score and generation score (QED and SA) similar to the initial experiments? Further, I also think including docked images from Vina would be great.

Computational cost comparison

Thank you for clarifying. From the analysis in the image and molecule generation, I understand that BMFT is much more feedback-efficient compared to the baselines. However, my comment is related to the computational cost: is BMFT faster than CFM for the same NFE (and likewise for image synthesis)? A comparison of wall-clock times, or other appropriate measures would be more useful.

Presentation concerns

Thanks for the modifications. I still think the results section can be neatly organized into subsections (for image, molecule and alignment). Right now, they are all "paragraph" styled sections, and the transition from image generation to molecule generation, and then to alignment experiments (images again) seems abrupt.

Comments on the paper

1 comment

Dear reviewers, we thank you again for your time and thoughtful comments on our work. We have provided detailed responses to the concerns, and with the discussion time winding down, we hope that you have a chance to consider these. If you have any additional questions, we would be happy to address them in the remaining time.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.