acceptodds
Under review as a conference paper at ICLR 2027

TOPVAE: REDUCING DARK AREAS IN 3D MOLECULAR LATENT DIFFUSION

Abstract

Latent diffusion is a scalable framework for 3D molecular generation, but it requires a latent space that stays valid beyond the posterior samples. A molecular VAE decoder can reconstruct every posterior latent exactly and still return disconnected or chemically invalid graphs a short distance away from it, which is where a diffusion prior draws its samples. We call these regions dark areas. They are widest where conditioning or size extrapolation builds a latent code no training molecule supplied. Therefore, we propose TopVAE, a topology-optimized VAE whose decoder emits adjacency before atom and bond types and is trained with connectivity and chemical constraints inside the decoding path, so that the region of latents that decode to valid molecules extends well past the posterior. TopVAE matches the strongest baseline at the posterior and keeps returning valid, connected molecules far outside it, where that baseline has collapsed. It continues to return sanitizable products when a scaffold is grown well past its training maximum in atom count. Paired with a standard diffusion transformer it is competitive on de novo generation on QM9 and GEOM-Drugs, with the lowest GEOM-Drugs FCD-3D of the methods scored through one pipeline.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
AI for Science, generative model, latent diffusion, variational autoencoder

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 3 reviews, above 27% of the others.
Reviewer_WZLFOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
2
Presentation
1
Soundness
1
Read review
Paper summary

The paper introduces TopVAE, a 3D molecular variational autoencoder, in which the decoder can decode latent states unseen during training (termed dark areas) into valid molecules. The main idea is to incorporate chemical and structural constraints into the decoder's design and training.

Critical strengths

None.

Critical weaknesses
  • The paper is hard to read; the method section (Section 4) reads like a technical report that describes the implementation with little intuition on the design choices and why the method is expected to work well. Also, mathematical objects are used throughout without a precise definition.
  • The generation results are poor; Table 3 shows a 2D FCD of 1.056 (vs 0.130 for the best baselines) and a 3D FCD of 0.340 (vs 0.137). This questions the utility of the method in practical generation settings.
Questions for authors
  1. In Equation (5), shouldn't the allowed values for B in the second part be 1...K?
  2. Can the method be extended to 2D molecular graph generation?
6 comments

We thank the reviewer for this feedback.

W1. Design intuition and mathematical clarity

We will revise Section 4 to better explain our design intuition and clarify the mathematical definitions.

Our central intuition is to construct molecules from topology to geometry: connectivity → chemical validity → 3D structure. Rather than relying solely on independent atom-pair predictions, we incorporate molecular-level structural constraints into the decoder.

First, ensure connectivity (TopoBridge). A meaningful molecular structure requires a connected graph. TopoBridge establishes connected topology before downstream predictions, preventing fragmented molecular graphs.

Second, ensure chemical validity (ChemCO). Connectivity alone does not guarantee chemically valid bonds. ChemCO refines bond assignments according to chemical constraints, including valence limits and bond compatibility, while preserving the decoder's preferences.

Finally, predict 3D geometry. The decoder directly predicts atomic coordinates, using the learned molecular representation and topology-aware decoding.

Why this design works. Pairwise reconstruction losses cannot directly enforce global molecular properties such as connectivity and valence consistency. TopoBridge and ChemCO introduce these structural priors, while training through TopoBridge and distilling ChemCO corrections via AGCL help the decoder learn more robust predictions beyond the training posterior.

Revision. We will reorganize Section 4 around this design intuition, clarify the mathematical formulation of each module, add a worked decoding example, and provide a complete notation table in Appendix A.

W2. FCD comparison and practical generation utility

We thank the reviewer for raising this concern. We clarify two important points.

First, the reported FCD values are not directly comparable because they use different reference sets. Table 3 quotes JODO's published FCD results, which use a different reference set from ours. Since FCD depends on the reference distribution, these values should not be directly compared. Moreover, JODO's GEOM-Drugs code generates random test splits without a fixed seed (code), making the original split difficult to reconstruct from code alone.

Similarly, UDM-3D's published FCD (0.130) was directly quoted from its paper, as noted in Table 3. We can not reproduce this result, and the authors have not released pretrained checkpoints (https://github.com/lyc0930/UAE-3D/issues/4); an independent reproduction also reported substantial discrepancies (https://github.com/lyc0930/UAE-3D/issues/2).

We therefore retrained UAE+UDM from the official code and recomputed FCD using a unified reference set consistent with UDM-3D. Under this fair comparison, TopVAE+DiT achieves comparable QM9 2D FCD to UAE+UDM (0.194 vs. 0.196) and FCD3D to JODO (0.216 vs. 0.209). We will clarify these evaluation differences in the revised paper.

Model2D FCD, our reference2D FCD, JODO referenceFCD3D, our referenceFCD3D, JODO reference
TopVAE+DiT0.1935 ± 0.0115—0.2156 ± 0.0181—
UAE+UDM (retrained)0.1959 ± 0.0049—0.2598 ± 0.0109—
JODO0.76750.1360.20850.137

Table A. QM9 FCD under different reference sets (lower is better). 10,000 samples per model. TopVAE+DiT and UAE+UDM: mean ± sd over three seeds; JODO: released samples. FCD3D is computed with hydrogens removed from both generated and reference molecules.

Second, practical generation extends beyond in-distribution FCD, particularly to atom-size extrapolation. Prior work highlights this importance: DiGress achieves only 2.6%, 2.2%, and 0.0% valid-and-unique molecules at +5, +10, and +20 atoms beyond its training maximum (Vignac et al., ICLR 2023, Table 7), while real molecular design may require molecules exceeding 100 atoms (Reidenbach et al., 2025, arXiv:2505.18392).

Our results directly demonstrate this utility. On GEOM-Drugs zero-shot scaffold expansion, TopVAE+DiT more than doubles UAE+UDM's average joint completion rate (11.1% vs. 5.0%, Table 2). At +80 and +100 atoms, TopVAE achieves 97.5% sanitizability, while UAE+UDM produces no sanitizable molecules (Appendix N).

TopVAE achieves competitive generation quality under a consistent evaluation protocol, while offering substantially stronger generalization to molecular sizes beyond the training distribution.

Q1. Allowed values of B in Eq. (5)

The second clause of Eq. (5) contains a typo. For edges selected by TopoBridge, Eq. (6) assigns zero probability to the no-bond class. Therefore, the correct constraint is:

A_{ij}=1 \Rightarrow B_{ij}\in\{1,\ldots,K\},\qquad K=3.$$ We have corrected Eq. (5) and the corresponding statement in Sec. 4.2 (lines 158–159). Eq. (6) and the implementation are already correct, so these corrections do not affect any experimental results. ## Q2. Can the method be extended to 2D molecular graph generation? TopVAE can generate 2D molecule graph in nature. TopVAE generates the 2D molecular graph, and generates it first. In Eq. (4) the adjacency, atom types and bond types form the 2D graph and are decoded before the coordinates, which are conditioned on that graph. Both constraint modules are defined on the graph alone, TopoBridge on the adjacency and ChemCO on the bond types. The 2D block of Tables 3 and 4 scores the bond graph TopVAE+DiT itself predicts, not one re-inferred from its coordinates (Appendix C.7).

Thank you for the clarifications!

Could you explain the difference between the FCD values posted here and the ones in the paper? For example, here you report 0.1935 2D FCD for TopVAE+DiT, while in the paper it is 1.056 (Table 4 in the current PDF).

Reviewer_iibhOpenReview ↗
Rating
2
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces TopVAE, a molecular variational autoencoder designed to reduce latent regions that decode to invalid or disconnected molecules. The method combines topology-first decoding, connectivity refinement through TopoBridge, and training-time chemical constraint optimization with advantage-gated supervision. Experiments assess robustness to latent perturbations, conditional generation, size extrapolation, and downstream molecular generation with a diffusion prior. Overall, the paper addresses a relevant problem, but my enthusiasm is tempered by the narrow demonstrated scope and the incomplete attribution of robustness gains to learned decoding rather than inference-time connectivity refinement.

Critical strengths
  • The distinction between accurate posterior reconstruction and reliable decoding away from the posterior is well motivated and relevant to latent generative models.
  • The controlled ChemCO and supervision ablations provide evidence that the proposed training procedure improves off-posterior decoding.
  • The supplementary evaluation includes physicochemical distributions, diversity, stereochemistry, and force-field relaxation, providing useful context beyond sanitization rates.
Critical weaknesses
  • Limited demonstrated generality. The method is evaluated within atom-aligned molecular latent representations and relies on chemistry-specific constraints. Its applicability to other latent representations or generative domains remains untested.
  • Incomplete separation of learned robustness from inference-time repair. Although ChemCO is removed at inference, TopoBridge remains in the decoding path. The reported audit shows that repair is infrequent for ordinary GEOM-Drugs generation, but does not establish its contribution under the perturbations and size extrapolations central to the paper’s claims. A direct ablation is needed to distinguish improved decoder predictions from robustness supplied by connectivity repair.
  • Limited interpretation of valence stability. Stability is evaluated after adding hydrogens to the generated heavy-atom graphs. Consequently, undersaturated heavy-atom valences can be completed automatically. These scores should therefore be interpreted as compatibility with hydrogen completion, rather than evidence that the model generates complete, valence-balanced molecular structures.
  • Physical-quality evidence remains limited. The MMFF94 relaxation analysis is useful, but compares primarily against a retrained UDM-3D baseline and shows a remaining gap to reference conformers. Broader comparisons would help establish whether improved graph validity translates into competitive three-dimensional molecular quality, particularly in the off-posterior and extrapolation settings.
  • Limited conceptual novelty of the “dark-area” analysis. The observation that accurate reconstruction of posterior samples does not guarantee valid decoding of perturbed or off-posterior latents is not particularly surprising. The analysis usefully quantifies this behavior, but the paper does not clearly establish what new insight the “dark-area” framing provides beyond familiar concerns about latent-space coverage and decoder robustness.
Questions for authors
  • How do off-posterior validity and connectivity change when TopoBridge’s repair step is disabled? How frequently does repair activate as latent perturbation or molecular size increases?
  • Can you distinguish the effects of training with TopoBridge from those of retaining it at inference?
  • How should the reported stability scores be interpreted given hydrogen completion, and are comparisons with explicit-hydrogen generators evaluated under consistent conventions?
  • Do the off-posterior and size-extrapolated molecules retain reasonable physicochemical properties and low relaxation energies, beyond passing sanitization?
6 comments

We thank the reviewer for those insightful suggestion.

W1. demonstrated generality

The method is general by construction. Connectivity and each chemical rule enter as separate constraints on the decoded graph, so rules, such as constraints from materials, can also be added, dropped or combined freely. Other latent representations and other generative domains are future work, outside the scope of this paper.

W2. Separating learned robustness from inference-time repair

We provide additional ablations under standard generation, latent perturbation, and size extrapolation to distinguish inference-time repair from learned decoder robustness.

First, TopoBridge primarily benefits off-posterior generation, with limited impact on end to end sampling. Under end to end generation, disabling refinement changes V&C only from 0.9985 to 0.9958 on QM9 and from 0.9508 to 0.9382 on GEOM-Drugs (Table S5).

Second, these gains cannot be explained by post-hoc repair alone. Attaching the same modules to the baseline at inference achieves only 0.157 versus TopVAE's 0.456 at σ = 0.5 (Table A). Even without the modules, TopVAE's trained decoder outperforms the baseline (0.085 vs. 0.015). This supports the importance of integrating TopoBridge during training, where the adjacency head is trained through it and downstream predictions condition on its output.

VCBσ = 0.3σ = 0.5σ = 1.0
TopVAE0.74350.45610.2897
TopVAE w/o modules0.40920.08460.0057
Baseline + modules0.45550.15680.0271
Baseline0.11860.01500.0016

Table A: QM9 validation (n = 10,000), with identical molecules and perturbation noise across rows. Modules: TopoBridge and ChemCO.

Third, TopoBridge also substantially improves size extrapolation. Disabling refinement reduces VCB from 0.539 to 0.190 for 60–90 heavy atoms and from 0.209 to 0.025 beyond the training maximum of 90 atoms (Table B).

VCB≤ 3233–5960–90> 90
TopVAE, modules on0.96670.78650.53930.2090
TopVAE, modules off0.82700.47200.19030.0253
Baseline0.86030.60050.01650.0000

Table B. GEOM-Drugs scaffold expansion, grouped by requested heavy-atom count. Modules: TopoBridge and ChemCO.

TopoBridge provides substantial robustness beyond the standard generation regime, while having relatively little effect on end2end sampling. Its benefits arise from both inference-time connectivity enforcement and its integration into decoder training, rather than post-hoc repair alone.

W3 & Q3. Hydrogen completion and fair comparison

Hydrogen completion is the standard evaluation convention for models that generate only heavy atoms, and it does not conceal chemical validity errors. Implicit hydrogens fill open valences, as in SMILES, but cannot correct over-valent atoms or invalid connectivity. Over-valence fails RDKit sanitization, while isolated heavy atoms fail our VCB criterion. Importantly, even ground-truth QM9 molecules achieve only 2.14% molecular stability without hydrogen completion, versus 100% after completion.

Our comparisons also respect each baseline's original evaluation convention. Published results in Tables 3 and 4 follow their respective sources (Appendix C.7). For JODO, which explicitly generates hydrogens, we additionally evaluate its released samples under both conventions (Table C).

Model / ConventionQM9 2DQM9 3DGEOM-Drugs 2D
JODO (published)0.999 / 0.9880.992 / 0.9341.000 / 0.981
JODO (reproduced, original convention)0.9990 / 0.98920.9925 / 0.93610.9995 / 0.9807
JODO (our convention)1.0000 / 1.00000.9795 / 0.76220.9497 / 0.9156
TopVAE+DiT (our convention)1.000 / 1.0000.991 / 0.9280.9995 / 0.979

Table C. Atom / molecular stability. JODO is evaluated on the same 10,000 released samples per dataset under both conventions. Our scorer reproduces its published results within 0.003.

Under a unified convention, TopVAE+DiT matches JODO on QM9 2D stability and outperforms it on QM9 3D (0.928 vs. 0.762) and GEOM-Drugs 2D (0.979 vs. 0.916) molecular stability.

Revision. We will clarify the hydrogen-completion convention and its chemical interpretation in Sec. 5.1, with the cross-convention comparison provided in Table C.

W4 & Q4. Physical quality and breadth of comparison

Our baseline is UDM-3D (NeurIPS 2025), a prior state-of-the-art method for latent 3D molecular generation. We further evaluate physical quality using MMFF94 relaxation, five PoseBusters checks, and held-out reference conformers.

Under end to end generation, TopVAE achieves near-reference quality on GEOM-Drugs: median relaxation energy of 0.466 vs. 0.456 kcal/mol per heavy atom, and a PoseBusters pass rate of 88.4% vs. 92.8%.

Under off-posterior perturbations, TopVAE substantially outperforms UDM-3D. At σ = 0.5 on QM9, its PoseBusters pass rate is 5.3% vs. 0%, with much lower relaxation energy (35.9 vs. 102.7 kcal/mol per heavy atom).

MetricTopVAE σ=0UDM-3D σ=0TopVAE σ=0.5UDM-3D σ=0.5
Sanitizable ↑0.99951.00000.45640.2497
PoseBusters (all five) ↑0.99200.98800.05300.0000
Relaxation energy ↓1.1760.55735.855102.740

Table D. Physical quality under latent perturbations (QM9).

Under size extrapolation, TopVAE maintains substantially lower relaxation energies across all evaluated offsets. Its median remains between 29.0 and 41.8 kcal/mol per heavy atom, while UDM-3D's increases sharply, reaching 344.8 at +60.

OffsetTopVAE valid / totalTopVAE energy ↓UDM-3D valid / totalUDM-3D energy ↓
+52000/200041.832000/200048.63
+101997/200034.302000/200053.41
+202000/200036.041996/200065.81
+401996/200029.001593/2000165.56
+601988/200029.2255/1500344.81
+801879/200033.230/500N/A

Table E. Physical quality under GEOM-Drugs size extrapolation. Median MMFF94 relaxation energy (kcal/mol per heavy atom) over sanitizable, parameterizable molecules. Request counts differ where baseline outputs are unavailable.

Overall, these results show that TopVAE's improved graph validity is accompanied by better 3D physical quality under OOD perturbations and size extrapolation, even against a strong prior SOTA baseline.

W5. What the dark-area framing adds

Our key insight is that molecular latent spaces are far less tolerant of errors than image latent spaces. A slightly distorted image remains usable, whereas a small latent perturbation can break a bond and disconnect a molecule, or introduce an extra bond that violates valence, making the entire molecule invalid. These two distinct failure modes motivate TopoBridge and ChemCO.

Importantly, these failures are invisible to standard reconstruction metrics. On QM9, UAE and TopVAE achieve comparable posterior reconstruction (Table 16), yet at σ = 0.5, their off-posterior VCB differs dramatically: 0.014 vs. 0.460 (Table 1).

This is particularly important for practical molecular generation, especially size extrapolation, where conditioning pushes latent codes beyond the training-supported region. Prior methods struggle in this setting: DiGress achieves only 2.6%, 2.2%, and 0.0% valid-and-unique molecules at +5, +10, and +20 atoms beyond its training maximum (Vignac et al., ICLR 2023, Table 7), while real molecular design can involve structures exceeding 100 atoms (Reidenbach et al., arXiv:2505.18392).

Thus, the dark-area framing goes beyond identifying latent-space coverage gaps: it reveals molecule-specific failure mechanisms and motivates structurally constrained decoding for robust generation beyond the training distribution.

Q1. Validity, connectivity, and refinement activation

TopoBridge primarily improves connectivity under off-posterior perturbations. Disabling it increases sanitizability but sharply reduces connectivity, leading to substantially lower VCB. At σ = 0.5, sanitizability rises from 0.4564 to 0.5877, while connectivity drops from 1.0000 to 0.1951 and VCB from 0.4563 to 0.0819 (Table F).

Metric ( Modules On / Off)σ = 0σ = 0.3σ = 0.5σ = 1.0
Sanitizable.9995 / .9995.7510 / .7842.4564 / .5877.2683 / .5503
Connected1.000 / .96261.000 / .53691.000 / .19511.000 / .0423
VCB.9994 / .9620.7509 / .4125.4563 / .0819.2682 / .0063

Table F. Ablation of inference-time refinement on QM9. QM9 validation (n = 10,000), identical inputs and noise; refinement is the only inference-time module.

TopoBridge activates mainly when decoding becomes challenging. The fraction of molecules requiring edge insertion increases from 3.74% at σ = 0 to 80.5% at σ = 0.5 and 95.8% at σ = 1.0. Under GEOM-Drugs size extrapolation, activation similarly rises from 3.45% at +5 to 76.6% at +80 heavy atoms (Table G). Even at +80, inserted edges account for only 1.23% of decoded adjacency edges.

QM9 σ, or GEOM-Drugs offset (realized heavy atoms)molecules that receive an inserted edgeinserted edges per molecule
σ = 03.74%0.060
σ = 0.346.3%0.603
σ = 0.580.5%1.497
σ = 1.095.8%2.889
+5 (11–15)3.45%0.036
+10 (16–20)13.3%0.14
+20 (26–30)28.6%0.32
+40 (46–50)45.1%0.58
+60 (66–70)65.8%1.01
+80 (86–90)76.6%1.37

Table G. QM9: refinement-on decodes of Table R5. GEOM-Drugs: scaffold expansion samples. 500 requests per scaffold and offset, four aromatic scaffolds.

Q2. Training with refinement vs. retaining it at inference

Under end to end generation, refinement has minimal impact on performance. Removing it at inference changes V&C only from 0.9985 to 0.9958 on QM9 and from 0.9508 to 0.9382 on GEOM-Drugs, while other metrics remain nearly unchanged (Table H).

TopVAE+DiTatom stab. 2Dmol. stab. 2DV&CV&Uatom stab. 3Dmol. stab. 3D
QM9, refinement on1.0000 ± 0.00001.0000 ± 0.00000.9985 ± 0.00030.9662 ± 0.00110.9907 ± 0.00040.9275 ± 0.0028
QM9, refinement off1.0000 ± 0.00001.0000 ± 0.00000.9958 ± 0.00060.9659 ± 0.00120.9882 ± 0.00070.9254 ± 0.0029
GEOM-Drugs, refinement on0.9995 ± 0.00000.9780 ± 0.00120.9508 ± 0.00170.9999 ± 0.00020.8324 ± 0.00040.0201 ± 0.0012
GEOM-Drugs, refinement off0.9995 ± 0.00000.9784 ± 0.00160.9382 ± 0.00260.9999 ± 0.00020.8328 ± 0.00050.0202 ± 0.0012

Table H. Standard generation with and without inference-time refinement.

Under off-posterior perturbations, both learned robustness and inference-time refinement contribute.*At σ = 0.5, TopVAE without refinement achieves 0.0819 VCB versus the baseline's 0.0146, demonstrating an improvement in the trained decoder itself. Retaining refinement further raises VCB to 0.4563 (Table A).

Importantly, post-hoc refinement alone cannot reproduce this performance: attaching the same inference-time modules to the baseline yields only 0.157 VCB versus TopVAE's 0.456 (Table S1).

Thus, refinement is nearly unnecessary for standard generation but critical under latent distribution shift, while co-training with it improves the decoder itself.

Reviewer_5ap4OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

TopVAE is a 3D molecular VAE for latent diffusion that targets “dark areas”. The authors refer "dark area" as latents in off-posterior, that decodes into invalid molecules. The authors propose a post-hoc remedy, by forcing connectivity by BFS edge insertion (TopoBridge), refines bond types, with the corrections distilled into the decoder through an advantage gate.

Critical strengths

The perturbation protocol is interesting and useful, where the results differ significantly when they are displaced via noise. This indicates that the reconstruction metric is not reliable to determine the quality of the latent space itself.

Evaluation protocol is meaningful and reliable.

Critical weaknesses

The paper fixes the decoder without testing whether the sampling side is the cheaper fix. To my understanding, the beta value is around 1e-6 in qm9, which makes it a near-deterministic autoencoder with no regularization. Without any beta sweeps, rejection sampling, which are much cheaper alternatives, doesn't seem to be tested. Also, it seems to me that the failures are only restricted to conditioning/extrapolation, but authors refer unconditional generation as if it is suffering from the same issue.

Also, the post-hoc modification seems to be the driving force of the validity, rather than a better latent space. A controlled experiment with & without TopoBridge, at inference, on the same perturbation sweep, seems to be necessary.

Questions for authors

Refer to the weaknesses above.

2 comments

Thanks for your careful and insightful comments.

W1. rejection sampling, and β on the sampling side

We provided additional experiments on rejection sampling and beta sweep.

σaccepted of 10,000: TopVAEaccepted of 10,000: baselinedecoder passes per accepted molecule: TopVAEdecoder passes per accepted molecule: baseline
0.37,5091,2921.337.74
0.54,5631462.1968.5
1.02,68283.731,250

Decoder passes per accepted molecule represents how many rejection sampling will conduct to get one valid molecule. On QM9 it needs 5.81× and 31.3× as many decoder passes per accepted molecule as TopVAE at σ = 0.3 and 0.5, and at σ = 1.0 it accepts 8 of 10,000 against TopVAE's 2,682 (Table R1)

As for beta, the baseline's KL weight is its own published 10⁻⁸ (UDM-3D, arXiv 2503.15567v4, Table 13). Our 10⁻⁶ is the KL weight of the image latent-diffusion autoencoders of Rombach et al. (2022),¹ and the latent space stays structured at it.

modelβσ = 0σ = 0.3σ = 0.5σ = 1.0
baseline00.9999 / 0.99970.0840 / 0.08370.0072 / 0.00820.0007 / 0.0007
10⁻⁸ (published)0.9999 / 0.99990.1333 / 0.14260.0167 / 0.01730.0010 / 0.0015
10⁻⁷0.9999 / 0.99990.2468 / 0.26260.0391 / 0.04100.0020 / 0.0027
10⁻⁶0.9999 / 0.99990.3774 / 0.40500.0740 / 0.08070.0078 / 0.0078
10⁻⁵0.9999 / 0.99990.4550 / 0.44610.0991 / 0.09520.0083 / 0.0080
10⁻⁴0.9999 / 0.99990.3943 / 0.38520.0786 / 0.07450.0079 / 0.0090
10⁻³0.9999 / 0.99980.2685 / 0.25980.0809 / 0.07270.0487 / 0.0414
10⁻²0.9998 / 0.99980.4408 / 0.44280.1575 / 0.14520.0213 / 0.0201
ours, retrained10⁻⁶0.9999 / 0.99990.8626 / 0.77170.5841 / 0.47480.3211 / 0.2619

See from the Table R2, at σ = 0 every row is at 0.9997 or more; from σ = 0.3 to 1.0 every baseline row, the matched β = 10⁻⁶ included, stays below ours.

¹ https://github.com/CompVis/latent-diffusion/blob/a506df5756472e2ebaf9078affdde2c4f1502cd4/configs/autoencoder/autoencoder_kl_32x32x4.yaml (kl_weight: 0.000001).

W2. unconditional generation

We agree with your comments. On QM9, a relatively simple dataset, unconditional generation is not a problem. However, On GEOM-Drugs dataset, those models separate in unconditional sampling. TopVAE+DiT is at V&C 0.9504 ± 0.0030 of Table 4 (Table 18), against UDM-3D's published 0.879. The revised PDF will be rewritten to make our description more concise.

W3. Is post-hoc correction the driving force of validity?

We conducted a controlled perturbation sweep experiment, with and without TopoBridge and at inference (Table R3).

First, removing TopoBridge reduces TopVAE's VCB from 0.456 to 0.085 at σ = 0.5 and from 0.290 to 0.006 at σ = 1.0. In contrast, such modules have a much smaller effect under end2end generation: V&C improves only from 0.9958 to 0.9985 on QM9 and from 0.9382 to 0.9508 on GEOM-Drugs (Table S5).

Second, inference-time correction alone cannot explain the improvements. Attaching the same modules to the baseline decoder yields only 0.157 versus TopVAE's 0.456 at σ = 0.5, and 0.027 versus 0.290 at σ = 1.0. Even without inference-time modules, TopVAE's decoder outperforms the baseline (0.085 vs. 0.015 at σ = 0.5).

Together, these results show that the gains arise from both inference-time correction and its integration into training, rather than post-hoc correction alone. TopoBridge is parameter-free, but TopVAE trains its adjacency head through TopoBridge (Eq. 16), with downstream predictions conditioned on its output (Sec. 4.3).

VCBσ = 0.3σ = 0.5σ = 1.0
TopVAE0.74350.45610.2897
TopVAE w/o modules at inference0.40920.08460.0057
Baseline + modules at inference0.45550.15680.0271
Baseline0.11860.01500.0016

Table: Controlled perturbation sweep on QM9 validation (n = 10,000),

We also provide additional evidence on end to end generation.

DatasetV&C (Refinement On)V&C (Refinement Off)Difference
QM90.9985 ± 0.00030.9958 ± 0.0006+0.0027
GEOM-Drugs0.9508 ± 0.00170.9382 ± 0.0026+0.0126

Table: V&C under end to end sampling, averaged over three sampling seeds. Refinement off is turning TopoBridge and ChemCO off together.

I thank the rebuttal from the authors. However, I believe that as provided, the post-hoc correction seems to be the substantial force of success, not the "latent space fragility" the authors pointed out as a defect in prior works. Thus, I would like to retain my score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.