acceptodds
Under review as a conference paper at ICLR 2027

Learning Molecules as Conformer Sets for Novel-Scaffold Selection at Billion Scale

Abstract

Molecular representations based on 1D, 2D, or a single 3D structure do not capture the conformational flexibility that governs binding, limiting the discovery of active compounds with novel scaffolds. We propose ZAO, a foundation model that represents each molecule as a conformer set and encodes 10 conformers per molecule as a single input. It has 238 million parameters and is pretrained on a corpus of 11 billion conformers, 12.4 times the size of the largest published conformer corpus. ZAO's frozen representation is stable under conformer regeneration, remains effective in regions structurally distant from known active compounds, and is complementary to 2D fingerprints. We evaluate this representation prospectively for the discovery of novel active compounds. From a 1.1-billion-molecule virtual library, we select 9,220 compounds across 31 targets, score them with absolute binding free-energy calculations, and compare the outcome with that of a Boltz-2 plus ESPSim pipeline. To the best of our knowledge, no prior evaluation of a molecular foundation model has measured the ability to select novel active compounds at this scale. Median per-target hit rates (ΔG ≤ -10 kcal/mol) are 26.4% versus 16.0%, with ZAO's rate higher on 22 of 30 targets. ZAO's advantage is greatest on targets regarded as difficult, including protein-protein interactions. At ΔG ≤ -12 kcal/mol, the gap widens: 14.6% versus 5.2%. ZAO and Boltz-2 agree only weakly, and the hit rate for compound-target pairs that both methods judge active is 53.0%, indicating that conformer-set representations can complement structure-based scoring. We also release ZAO's weights.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
foundation model, molecular foundation model, molecular representation learning, virtual screening, drug discovery

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_JDbQOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
1
Soundness
2
Read review
Paper summary

The paper introduces a QSAR modeling approach that uses a set of 3D conformers as the molecular representation rather than a single conformer. The reported results suggest state-of-the-art performance on some tasks, although the experimental setup and evaluation tasks are difficult to understand from the text, tables, and figure captions.

The summary is based on checking figures and tables and careful reading of the first 3 pages. I stopped reading the paper after that due to the poor presentation.

Critical strengths

S1. An interesting and relevant problem in molecular representation learning.

The paper addresses an important problem in molecular representation learning: representing molecules using multiple 3D conformers rather than a single structure. Since molecules exist as conformational ensembles, incorporating information from multiple conformers is a potentially valuable approach for improving molecular representations and downstream QSAR modeling.

Critical weaknesses

W1. The presentation is substantially below the standard expected for an ICLR submission, preventing a reliable scientific evaluation.

The manuscript is exceptionally difficult to follow due to major problems with its writing, organization, and formatting. Essential information needed to understand the methodology and experiments is frequently omitted from the main text, while numerous concepts and implementation details are introduced without adequate definition, motivation, or context.

Almost every paragraph refers to other sections or appendices, forcing the reader to navigate repeatedly through the manuscript merely to reconstruct the basic experimental procedure. This is made considerably worse by the lack of hyperlinks in cross-references, despite their support in the ICLR template. The unusual citation formatting further reduces readability.

For example, the first paragraph of Section 3.2 introduces two regression models, multiple feature families, the "union" of the models, a novelty requirement, and a candidate-selection procedure without explaining the prediction task, how the labels are defined, what "union" means operationally, or how the models are used to select candidates. Instead, the reader is directed to multiple appendices for information essential to understanding the central experiment.

I stopped my detailed reading at approximately page 3 because I could not reliably reconstruct the methodology and experimental setup from the presentation. Consequently, I cannot provide a meaningful assessment of the technical soundness or claimed contributions.

Clear, coherent, and auditable presentation is an essential component of scientific work, particularly at a top-tier conference. In my assessment, the paper requires a drastic rewrite and reorganization before it can be properly evaluated. Such a revision is far beyond the scope of a rebuttal.

My recommendation is therefore rejection on presentation and evaluability grounds, independently of the potential scientific merit of the work.

Questions for authors

None. The presentation issues are too extensive to be addressed through individual clarification questions during the rebuttal.

4 comments
AuthorsOpenReview ↗

Working cross-references and a self-contained opening to Section 3.2

We thank the reviewer for identifying plainly the passages where the presentation became hard to follow.

On both formatting points the reviewer is right: in the revision, all section, appendix, figure, table, and citation references are clickable, and the citations are author-year throughout.

Section 3.2 now opens self-contained in this revision, not in a later camera-ready version. Specifically, it states the prediction target and the positive label in the main text, says what the union of the models means operationally, and lists the six screening stages.

On the navigation cost, everything needed to follow the experiment is now in the main text. The pointers that remain lead to supporting evidence, and removing them would cost a reader who wants to verify the results more than it saves one who does not. Each remaining pointer is now a hyperlink. Feature blocks, the common configuration, early enrichment (BEDROC20), the strong-hit threshold, and the definition of a novel scaffold are all given where they are first used.

In the reviewer's judgment, the rewrite needed here is beyond the scope of a rebuttal. We would ask the reviewer to weigh that judgment against what the revised manuscript now contains: the review names three specific defects, all three are repaired, and none of them required rewriting the manuscript as a whole. If the revised PDF is now adequate on those counts, we would be glad to hear it; if it is not, naming one further passage that is as hard to follow as the former opening of Section 3.2 would let us rewrite that passage within this discussion period.

Reviewer_JDbQOpenReview ↗

Presentation Concerns Remain Unresolved Despite Revisions

Thank you for your response. Fixing the broken cross-references and citation formatting is appreciated, but these are basic presentation requirements that should have been met in the original submission.

I also acknowledge the effort to clarify Section 3.2. However, even the revised version remains confusing. For example, the opening sentence introduces a continuous activity prediction task (suggesting regression) immediately followed by a binary positive-label definition (suggesting classification), without explaining the relationship between the two. The next sentence begins "As shown in Figure 2a, we fit two regression models [...]", yet Figure 2a provides no meaningful explanation beyond the label "Two regressors on ZAO and 2D features, ligand only."

More importantly, Section 3.2 was merely one example of a broader, systemic problem with the manuscript's clarity and organization, not an exhaustive list of issues to correct. While the revisions improve some passages, I do not see a comparable effort to systematically improve the clarity and organization of the manuscript. I disagree that "everything needed to follow the experiment is now in the main text." Important methodological choices still require consulting the appendices.

I therefore maintain my original assessment. Despite the revisions, I still consider the presentation substantially below the standard expected at a top-tier conference.

AuthorsOpenReview ↗

The revision beyond Section 3.2, and the two passages named

The body as a whole. The second revision works through the body of the paper. What a reader needs in order to follow the experiment is stated where it is first used, and what the appendices hold is the evidence behind a number rather than a definition the body depends on: the free-energy protocol, the probe design, the per-target tables, and the two control experiments added during this discussion. The paragraphs are shorter and each carries less at once.

The opening of Section 3.2. The reviewer is right that it set a continuous prediction task beside a binary label without relating the two. The paragraph now opens with one task, regression, and two cutoffs on the predicted scale: 6.0 defines the training positives that serve as the reference set for novelty, and stage (4) applies 7.0 to predictions. No model in the paper is trained as a classifier.

Figure 2a. The reviewer is right that the panel carried less than the sentence citing it implied. That citation is gone, and the panel is cited once, at the point where the funnel counts it shows are given.

If a further passage is as hard to follow as the former opening of Section 3.2, naming it would let us rewrite that passage within this period.

AuthorsOpenReview ↗

The third revision

A third revision has been uploaded. The wording has been made plainer and more consistent throughout, every section of the body now carries run-in headings, and one heading governs one block of text. The cross-references between sections, which the reviewer raised at the outset, have also been substantially reworked.

We thank the reviewer for the time spent on this paper and for the frankness of the assessment. It is what let us concentrate on the quality of the presentation.

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

This paper proposes ZAO, which is a molecular foundation model that encodes multiple molecular conformers into a representation. The authors evaluate ZAO using binding affinity-based virtual screening for 1.1 billion molecules across 31 targets, and ZAO achieves higher computational hit rates compared to the existing Boltz-ESPSim pipeline. The authors also show the benefits of multiple conformer-based pretraining and learned aggregation using geometric property predictions.

Critical strengths

To the best of my knowledge, this is the first study to incorporate multiple conformers to learn a molecular representation, rather than aggregating molecular representations for multiple conformers at the downstream stage. The virtual screening experiment for 1.1 billion molecules across 31 targets is a valuable large-scale evaluation of molecular representations.

Critical weaknesses
  • Unclear benefits of using multiple conformers. To evaluate the benefits of incorporating multiple conformers, current manuscript only considers improvements in geometric property prediction but does not consider whether multiple conformers improve useful downstream predictions (e.g., binding affinity, stability). Although the authors consider binding affinity screening, this does show relative benefits with respect to established Boltz2 pipelines rather than showing the benefits of incorporating multiple conformers.

  • Lack of controlled comparison in virtual screening. The considered binding affinity screening comparison (Table 1) differs in search space and computational requirements, without a comparison under comparable budgets. This makes it difficult to assess whether ZAO provides a practical advantage in large-scale virtual screening.

Questions for authors
  • Can the authors provide the benefits of using multiple conformers with respect to a single conformer for useful downstream activity prediction, e.g., stability, binding affinity?

  • Can the authors comment on whether simple atom-wise conformational statistics, e.g., RMSF, provide similar improvements, or whether the structural flexibility considered in multiple conformers is significant?

  • Can the authors compare the computational budgets with respect to the Boltz-ESPSim pipeline? How does ZAO provide benefits to the screening results under controlled setups with respect to Boltz-ESPSim given the same budgets?

5 comments
AuthorsOpenReview ↗

Downstream value of the conformer set, the atom-wise fluctuation summary, and compute

We thank the reviewer for pressing on downstream value and on atom-wise fluctuation, and ran a new experiment for each; throughout, table numbers are those of the revision.

Q1. Pretraining on 10 conformers outperforms pretraining on one, measured on Section 3's activity-ranking task, not on a geometric probe. On the 31-target task of Section 3, per-target early enrichment (BEDROC20) is higher by +0.044 with gradient-boosted trees and +0.053 with kernel ridge regression (26 and 25 of 31 targets, p = 1.4e-05 and 1.3e-04), and fold enrichment in the top 1% is 20.8× against 16.2× and 18.8× against 14.2×, respectively. Apart from the pretraining input, 10 distinct conformers per molecule against the lowest-energy conformer copied into all 10 slots, the checkpoints are identical in molecules, order, seeds, compute, and token count; both receive the same 10 conformers at inference, and the downstream splits, feature blocks, and tuning are fixed (Table 20, Appendix G.3). Similarly, on conformer energy a linear probe recovers the target at 0.241 from the 1-conformer control, against 0.415 from the 10-conformer control (Table 19).

Q2. We built the atom-wise fluctuation summary the reviewer named, measured it on the same 36 metrics as every other feature set, and added it to Table 2. For each molecule we take each heavy atom's root-mean-square fluctuation about its mean position over the superposed 10 conformers and reduce that vector to 22 features, scored on the 19,974 molecules of that table with the probe and folds used there; Appendix E.1 describes the construction. On the six set-specific metrics it scores 0.619, closer to ZAO's 0.805 than any other hand-built or public feature set and well above the 2D descriptors (0.498) and 2D fingerprints (0.485), as a feature built from conformer spread should. However, on the other four categories it is the lowest-scoring feature set in the table: 0.112 against 0.422 on conformer energy, 0.246 against 0.858 on shape, 0.275 against 0.856 on pharmacophore distances, and 0.323 against 0.939 on 2D properties; over all 36 metrics it averages 0.326 against ZAO's 0.830. Substituting it for the ZAO feature block would discard most of the information encoded in that block, and we do not expect it to reproduce the improvement of Q1.

As expected of a representation of flexibility, the gap between the controls on set-specific metrics widens monotonically with molecular size, from 0.196 at 20 heavy atoms or fewer to 0.425 at 36 or more (Appendix G.2), and with the number of rotatable bonds (Figure 3b).

Q3. For ZAO the 10-conformer to 1-conformer cost ratio is 1.65× end to end, and the full run over 1.1 billion molecules cost about 1,360 GPU-hours of conformer generation and about 800 of embedding extraction, i.e. about 7 GPU-ms per molecule amortized over the library; the 20.5 ms in Table 18 is a single-machine, end-to-end figure on 1,962 molecules, where batching cannot help.

We do not have the Boltz-ESPSim pipeline's own total, since it was built outside this work, but we timed the operation it rests on, now recorded in Table 18: co-folding one complex with Boltz-2 and its affinity head on one H200 takes 30 to 55 s of GPU time over proteins of 125 to 766 residues, excluding the per-target MSA its compounds share. Per compound, that is three to four orders of magnitude above the ZAO pass.

As a per-compound measurement, it supports only the per-compound conclusion: the pipelines could not have searched equally large populations. We claim no advantage in total campaign compute, which we cannot establish without that group's own accounting. Due to the cost of free-energy scoring (about 72 GPU-hours per complex), a budget-matched comparison is out of reach here. Appendix B.3 is the nearest substitute: applying the Boltz-2 cutoff to both groups widens the gap from +5.7 to +16.5 points, though we do not present this as the controlled experiment the reviewer asked for.

For downstream value the revision adds a new measurement; for the comparison it quantifies the one asymmetry that could be addressed retrospectively. We hope this changes the reviewer's assessment, and we welcome any further comment.

The authors have partially addressed my concerns, and I would raise my rating to weak reject.

The additional activity-ranking experiments are helpful and support the downstream value of ZAO. However, they do not directly show how much ZAO contributes to the free-energy screening results in Table 1. These results are still difficult to interpret because the ZAO and Boltz-ESPSim pipelines differ in candidate pools, filtering and novelty constraints, and computational budgets. In addition, the ZAO pipeline itself includes several hand-designed filtering steps.

I understand that a full budget-matched comparison would be computationally expensive. However, even on a smaller candidate set or a subset of targets, it would be useful to compare candidate selection with ZAO, without ZAO, and with other molecular representations using the same initial candidate pool, representation-independent filters, learner, and selection budget.

The manuscript would also benefit from further editing. Some sentences are awkward or unclear. For example, the phrase “what stays frozen holds the information of a single conformer” is difficult to follow. The description of the screening pipeline in Section 3.2 is also very dense, with the model configuration, filtering steps, applicability-domain definition, and thresholds introduced in a single paragraph. Sections 4.1 and 4.2 discuss many representations, controls, metrics, and numerical results in a compressed manner.

AuthorsOpenReview ↗

The filtering steps, a comparison at smaller scale, and the writing

The filtering steps. The reviewer is right that Table 1 compares pipelines rather than representations, and that the ZAO group alone passes a hand-built filter chain. The stage in that chain which decides what chemistry each group can reach runs against the ZAO group. Only it is held to a 2D similarity cap against the training positives, which leaves its compounds further from the known active compounds than the comparison group's and far more often carrying ring systems the training data does not contain (Table 5). Proximity to known active compounds is what raises hit rates, as the near-analog control at 64.8% shows. The margin is therefore not something that stage produces. Appendix D.2 checks this directly, comparing the two groups at equal proximity to the known active compounds, and the ZAO group is ahead in all seven bands. The remaining stages do not work in that direction either. The physicochemical and size limits bound lipophilicity, aromatic ring count and molecular complexity, and computed binding free energy grows more favourable with all three. Those limits remove the compounds most likely to clear a fixed cutoff. The applicability domain is defined in ZAO's embedding space, and crediting the margin to it is crediting the margin to the representation.

A comparison at smaller scale. Three targets, three conditions and ten compounds each would take about 6,500 GPU-hours at roughly 72 GPU-hours per complex, and a hit rate measured over ten compounds would be dominated by noise. Appendices D.3 and H already measure what that comparison is for, how much of the selection model's ability to find active compounds comes from ZAO, on the same 31 targets with everything but the representation held fixed. Both end on the ranking of held-out active compounds rather than on free energy, and neither replaces Table 1.

D.3 takes the model as deployed and removes ZAO, holding the data, the splits, the learner and the hyperparameters fixed. Early enrichment is higher with ZAO on 28 of the 31 targets with gradient-boosted trees and on 29 with kernel ridge regression, and the active compounds recovered in the top 1% fall to about a third once the columns are gone. H then gives the condition without ZAO a per-target search of its own at the same budget, free to choose among five conventional representations, and the gap does not close, standing at +0.202 and +0.102 in early enrichment. Since the ranking decides which candidates clear the activity cutoff, a loss of that size changes which compounds a campaign would select.

The writing. The second revision goes through the body of the paper. Subjects built on abstract phrases now name what acts, dense paragraphs are split, and each number is tied to the condition it belongs to instead of being left to position.

We would be glad to know whether these change the reviewer's assessment.

Reviewer_k2CTOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
3(+1)
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes ZAO, a conformer-set encoding model that combines 3D geometry with a 2D bond graph, pretrained on roughly 1.1 billion molecules. The authors combine the frozen representation with fingerprints and physicochemical descriptors for virtual screening against 31 targets, selecting 9,220 compounds in total. On 30 paired targets, the median computed hit rates of ZAO and Boltz-ESPSim are 26.4% and 16.0%, respectively. The paper further analyzes the representation through linear probes, controlled pretraining, and 2D long-range retrieval.

Critical strengths

1. The main text reports screening results at scale, across many targets.

ZAO selects candidates from roughly 1.1 billion molecules. Under a unified evaluation protocol, its computed hit rate is higher in 22 of the 30 paired targets, with a median paired difference of +5.7 percentage points (Table 1, (p=0.0043)). This provides large-scale evidence for the practical effect of the proposed screening pipeline.

2. The screening results address hit rate and structural novelty together.

Under the Tanimoto < 0.6 constraint relative to training positives, the median per-target fraction of novel scaffolds among selected compounds, and among their hit subsets, is 100.0% in both cases (Section 3.3, Table 5). The results cover candidates structurally distinct from known actives, rather than only close analogues.

Critical weaknesses

1. The main comparison does not separate model gains from screening-setup effects.

ZAO searches roughly 1.1 billion molecules, while Boltz-ESPSim processes roughly 11,000 to 96,000 candidates per target. ZAO requires Tanimoto < 0.6 and keeps one molecule per chemotype cluster; Boltz-ESPSim has no equivalent 2D novelty constraint, keeps up to five molecules per scaffold, and selects a different number of candidates (C.1, C.2). The authors position this as a pipeline-level comparison, but the advantage in Table 1 cannot yet be attributed to the ZAO representation alone. A ranking comparison under a shared candidate pool, shared constraints, and a fixed number of selected compounds is missing.

2. The downstream gain of conformer-set training has not been directly tested.

Appendix G compares set-based and single-conformer pretraining under the same architecture, data, and matched compute, but the endpoint is a linear probe. D.3 already contains an activity-ranking ablation that removes ZAO features: top-1% enrichment drops from 41.0x to 14.4x for GBT and from 39.6x to 12.0x for KRR; that experiment compares the whole ZAO feature block, not the conformer set against single conformers. In addition, the ablation reuses a unified configuration after feature removal, which supplements dimension-related hyperparameter control but does not follow the per-target tuning pipeline of the main text. The existing results therefore support the incremental value of ZAO features, but do not yet show whether set-based training improves downstream activity ranking, or by how much relative to a conventionally tuned feature model.

Questions for authors

1. Comparison setup: Could you apply uniform novelty and de-replication rules on a shared candidate sub-library, fix the number of selected compounds per target, and report a ranking or enrichment comparison for both methods? If not, please state explicitly that Table 1 supports a pipeline-level comparison, not an independent advantage of the representation.

2. Conformer sets: Could you apply the two controlled checkpoints of Appendix G to the same downstream activity prediction task, with fixed data splits, other features, and tuning budget, comparing per-target BEDROC20 and top-1% enrichment? This control requires no new wet-lab or free-energy computations.

3. Conventional feature baselines: In D.3, if the models with and without ZAO features were each tuned through the main-text feature-combination search and hyperparameter procedure, with the same tuning budget, would the early-enrichment differences persist?

Minor Comments

  1. Please document how the applicability-domain thresholds per target in C.1 were chosen, and report candidate counts at each stage after the six-step filter, to help separate the effect of model ranking from that of the filtering rules.

  2. The public release in A.4 retains about 50.9% of label pairs and all 140 targets. Please clarify which screening and ablation results can be reproduced with the public data and which rely on commercial data.

3 comments
AuthorsOpenReview ↗

New experiments for Q2 and Q3, and the scope of Table 1

Throughout, table numbers are those of the revision, not of the submitted version.

Q2. We ran the control the reviewer proposes, and the conformer set is ahead on the activity endpoint. Per-target early enrichment (BEDROC20) is higher by +0.044 with gradient-boosted trees and +0.053 with kernel ridge regression, on 26 and 25 of 31 targets (p = 1.4e-05 and 1.3e-04), and fold enrichment in the top 1% is 20.8× against 16.2× and 18.8× against 14.2×, respectively. As in Appendix D.3, the targets, splits, folds, hyperparameters, and common configuration are unchanged; only the embeddings filling the ZAO feature block differ. Each difference is the median of the per-target differences (Appendix G.3, Table 20).

Q3. Even when both conditions are tuned at an equal budget, the difference persists. Both now receive a per-target search of their own, run with the main text's procedure on the same 31 targets, at a 40-trial budget, in one hyperparameter space and one set of splits, differing only in whether the ZAO feature block may be selected. Early enrichment stays higher with the block, by +0.202 with gradient-boosted trees (25 of 31, p = 3.5e-06) and +0.102 with kernel ridge regression (26 of 31, p = 1.4e-06). Given a free choice, the search keeps the block on 29 and 27 of 31 targets. Where it did not, the conditions end up level (-0.011 median difference, against +0.165 where the block was kept), so the difference is not search noise. Compared to Appendix D.3, the gradient-boosted difference is essentially unchanged while the kernel ridge difference falls by more than half, +0.255 to +0.102, and stays decisively above zero (Table 21, Appendix H).

W1 / Q1. A shared-pool ranking comparison would need a fresh campaign at about 72 GPU-hours per complex, beyond this discussion period. Therefore we take the fallback the reviewer offers and state it plainly: Table 1 supports a pipeline-level comparison, not an independent advantage of the representation. Note that the caption of Table 1 in the revision now states this.

On the setup, the one asymmetry the paper quantifies favors the comparison group, not ours. Specifically, the Boltz-ESPSim group was filtered and ordered by Boltz-2 before free-energy scoring, and the ZAO group was not; applying that cutoff to both widens the gap from +5.7 to +16.5 points (Appendix B.3). Nor is +5.7 flattering: it is the smallest of the three pooling rules' differences, against +6.1 and +11.4 (Appendix B.3), and rises to +21.3 on the three targets where the scoring protocol's correlation with measured activity exceeds 0.7 (Appendix B.4). On breadth, searching 1.1 billion molecules is a handicap, not an advantage: a predictive model loses accuracy the further it is applied from its training data (the applicability-domain problem). However, candidate pools, per-target counts, and the selection constraints still differ.

In addition, the groups are not trading hit rate for novelty: the per-target median fraction carrying a ring system absent from the target's training positives is 81.9% for the ZAO group against 24.2% (Table 5), and none of its 9,220 compounds appears in the selection model's training labels or in the training split of the supervised pretraining set.

Minor comments. Section 3.2 now states the six screening stages with their thresholds, including the applicability-domain floor: a percentile of the nearest-neighbor similarity distribution inside each target's training set, between the 10th and the 40th according to how much labeled data that target has, with a fingerprint Tanimoto floor of 0.15. At the time, only the final counts were recorded, and the per-stage counts the reviewer asks for cannot now be reconstructed. Appendix D.3 separates ranking from filtering, holding all filters fixed and removing only the ZAO feature block from the ranking model; Appendix H does the same with each condition tuned separately. On reproducibility, the screening of Section 3 and the ablations of Appendices D.3, G.3, and H rest on the selection model's labels, part of which come from a commercial source; a public-data-only rebuild retains 50.9% of the pairs and all 140 targets, and the results can be reproduced at that coverage. Except for the control rows of Tables 2 and 13, Section 4.1 and Appendices E and G.2 need only the released weights and public molecules; Sections 4.2 and 4.3 share the label dependence.

We would be glad to know whether these experiments change the reviewer's assessment.

Thank you for the rebuttal. My main concerns have been addressed, so I'm glad to raise my score from 2 to 3. By the way, I think the presentation really needs improvement to meet the standard expected of an accepted paper.

AuthorsOpenReview ↗

The third revision

A third revision has been uploaded, addressed to the presentation the reviewer asked us to improve. The four contributions and their order, the section titles and their order, and the data behind every figure are as submitted.

We thank the reviewer for the engagement across the discussion period and for the specific questions, which led to the controls now in Appendices G.3 and H.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.