We thank the reviewer for their constructive feedback. Our response and proposed revisions for the concerns raised by the reviewer are as follows

1

We will ensure that figure sizes remain consistent throughout the appendix, particularly improving the font readability of Figure 3. Thank you for pointing this out. We have separated the molecules into separate high-resolution images for better legibility: https://imgur.com/4pAWr4Q; https://imgur.com/a/jGxFnhY; https://imgur.com/a/uroAmCe; https://imgur.com/a/0biixYL. We’ll update the figure in the revised version of the paper. These molecules were sampled from the same pretrained model as before.

2

Our approach does not involve post-hoc RDKit filtering or penalization. Instead, A-GFN ensures atomic valency correctness at every generation step, preventing invalid states. This is crucial for atom-based GFlowNets, which operate in a vast combinatorial space, unlike fragment-based GFN, which is limited to predefined fragments. Without valency constraints, the model would frequently generate infeasible structures, making exploration intractable. Molecular validity is just one of many evaluation criteria. The key metric #Modes measures how many diverse molecules satisfy all four pretraining conditionals (TPSA, Num Rings, QED, SAS). A-GFN significantly outperforms fragment-based GFN in this regard, demonstrating superior exploration—not merely benefiting from valency constraints. Thus, valency enforcement is an inherent design choice, not a post-hoc filtering step, and does not confer an artificial advantage.

3

All baselines (Fragment-GFN, REINVENT, MolDQN, GraphMCTS) were fine-tuned under identical compute constraints: 24 hours on a single V100 GPU, ensuring a fair comparison. A-GFN pretraining lasted ~12 days on 4 A100 GPUs (250K steps), but a 100K-step checkpoint (~5 days) was sufficient for fine-tuning, as reward plateaued (see https://imgur.com/a/hLITsk5 ; to be included in the revised paper). While pretraining incurs a one-time cost, it benefits multiple downstream tasks. Once pretrained, A-GFN fine-tuning takes just 24 hours on a V100, making it comparable to molecular optimization baselines. This aligns with established pretraining and finetuning practices seen in ChemBERTa and MoLFormer.

4.1

We would like to clarify that our primary contribution is not that atomic-level design enhances pretraining, but rather the other way, i.e., pretraining is essential for A-GFN to function effectively. Without pretraining, A-GFN fails to generate viable molecules. Our experiments confirm that training A-GFN from scratch (Task train AGFN) consistently fails to satisfy task constraints.

4.2

Using ZINC for pretraining could introduce overlap with evaluation tasks like logP optimization, but it is common in pretraining large models to have some solutions for the downstream task to be contained in large scale pretraining data as a broadly representative pretraining data distribution is desired. Zero-shot OOD molecule generation is non-trivial, even for experts. To ensure benchmarking remains challenging, we chose property thresholds so that <25% of ZINC molecules meet the optimization criteria. Additionally, our Novelty metric measures unique molecules not present in ZINC, confirming that solutions stem from exploration rather than memorization.

4.3

Some baselines (Fragment-GFN, REINVENT) were pretrained using their respective methods, while others (MolDQN, GraphMCTS) were trained from scratch, per the PMO benchmark (https://github.com/wenhao-gao/mol_opt/tree/main) . We used publicly available pretrained models where applicable, ensuring fair comparisons. A-GFN’s superior performance is due to better exploration, not an unfair pretraining advantage.

5

We acknowledge that our paper does not include detailed descriptions of the baseline implementations, and we appreciate you bringing this to our attention. All baseline methods used in our work were directly adopted from the publicly available GitHub repository: https://github.com/wenhao-gao/mol_opt/tree/main. This repository is associated with the preprint https://arxiv.org/pdf/2206.12411, which includes some relevant implementation details, such as whether a method involves pretraining (check Table 7). However, specific aspects such as pretraining compute or further fine-tuning configurations are not comprehensively documented there either. We will add these details in the revised version of the paper for better transparency and reproducibility.

We sincerely appreciate your thoughtful and constructive review of our paper. We are pleased to see that you find our work well-motivated, comprehensive in evaluation, and a notable contribution to the application of GFlowNets in molecular design. Additionally, we are grateful for your recognition that our claims are well-supported by experimental evidence, with no significant concerns raised regarding our methodology, theoretical foundation, or experimental design. Based on the feedback from other reviewers, we have further improved our paper to enhance clarity, strengthen key arguments,

Given that no major weaknesses have been identified in your review and that our contributions are acknowledged as valuable to the field, we kindly ask you to reconsider and raise your current score. Your support would help ensure that this contribution reaches the broader community and advances research at the intersection of GFlowNets and molecular generation.

1. Why does TB sometimes outperform RTB in single-objective tasks (Table 3)? Is it due to over-regularization in RTB?

Thank you for raising this important question. The observed performance difference stems from fundamental differences in how TB and RTB balance optimization objectives: Yes, RTB's design introduces over-regularization in single-objective tasks due to its explicit anchoring to the pretrained prior. While this prior (e.g., a chemical validity model) helps maintain desirable properties in multi-objective settings, it creates conflicting constraints when optimizing for a single objective (which does not capture these desirable properties). We made initial efforts to address this trade-off in our current draft (lines 297-303), where we discuss how RTB’s prior anchoring leads to its weaker performance when compared to TB. We would like to point out that this effect would diminish when the prior already overlaps with high-reward solutions, and/or if tasks require balancing multiple constraints (e.g., both chemistry rules and potency). We appreciate this opportunity to clarify and will emphasize this trade-off more explicitly in our revision.

2. How sensitive is the method to the choice of inexpensive reward descriptors?

This is an excellent question, and we appreciate the reviewer's interest in understanding the sensitivity of our method to the choice of inexpensive reward descriptors. The selection of pretraining properties plays a crucial role in shaping the learned policy, and our choices were guided by well-established principles in drug discovery. Specifically, we selected QED (quantitative estimate of drug-likeness), SAS (synthetic accessibility score), TPSA (topological polar surface area), and the number of rings, as these properties are widely used in molecular generation and optimization tasks due to their strong correlation with drug-likeness and pharmacokinetic properties [1].

Empirically, we found that these four descriptors provided a well-balanced pretraining signal that improved sampling efficiency while preserving chemical diversity. Importantly, we observed that adding additional constraints led to premature convergence and mode collapse, as the optimization problem became overly restrictive, significantly reducing the number of valid and diverse molecules that could be generated. This aligns with findings in other large-scale generative modeling approaches, where excessive constraints during pretraining can limit the exploration capacity of the model and degrade its generalization ability in downstream fine-tuning.

3. How would a fragment-structured GFN compare to A-GFN?

We thank the reviewer for this question. The finetuned A-GFN generally outperforms fragment based GFN across all tasks considered in terms of diversity of molecules as well as the success percentage, which quantifies the ratio of molecules simultaneously satisfying all the enforced objectives. We have detailed these results in Table 12- 20 in Appendix F.

[1] Wellnitz, James, et al. "STOPLIGHT: a hit scoring calculator." Journal of Chemical Information and Modeling 64.11 (2024): 4387-4391.