acceptodds
Under review as a conference paper at ICLR 2027

Packora: Systematic Design for Generative Molecular Crystal Structure Prediction

Abstract

Molecular crystal structure prediction (CSP) is important in pharmaceuticals, agrochemicals, and organic electronics, where subtle differences in molecular conformation and packing can strongly affect material properties. We present Packora, a generative model for molecular CSP that jointly predicts atomic coordinates and the lattice from molecular graphs. To accommodate diverse crystal compositions and available structural information, Packora offers greater conditioning flexibility than existing approaches, accepting any subset of molecular conformers, stereochemical labels, and space-group information while supporting multi-component and organometallic crystals. Beyond this conditioning flexibility, Packora integrates effective design choices for atomistic generative modeling: cacheable pairwise refinement, effective training objectives and numerical solvers, conditioning dropout, and balanced scaling of pairwise and single representations. We also report systematic studies of architecture, training, conditioning, inference, and scaling that we conducted to make the design choices. Finally, we introduce a new evaluation suite for generative molecular CSP that separates generation and ranking, following the CCDC CSP blind test protocol. Packora outperforms baselines in both settings, achieving the highest coverage across all six generation benchmarks, higher experimental-form recovery, lower experimental-form ranks, and faster convergence in ranking.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
generative models, molecular crystals, material discovery, crystal structure prediction

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_fdkWOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

This paper presents PACKORA, a flow-based generative model for molecular crystal structure prediction (CSP). The proposed framework integrates several architectural and training design choices and supports flexible conditioning on molecular conformers, stereochemistry, and space-group information. The authors systematically investigate model architecture, training objectives, conditioning strategies, inference algorithms, and scaling. Extensive experiments demonstrate that PACKORA consistently outperforms competitive baselines across multiple structure-generation benchmarks and achieves improved performance in downstream structure ranking.

Critical strengths

Important and relevant research topic. Molecular crystal structure prediction is an important and challenging problem with broad applications in materials science and drug discovery. Developing effective generative models for this task is valuable to the research community.

  1. Clear writing and presentation. The paper is well organized and clearly written. The motivation, methodology, and experimental setup are presented in a logical manner.

  2. Well-designed evaluation protocol. I appreciate the effort to establish a comprehensive evaluation framework that separates structure generation from downstream relaxation and ranking. This distinction is important for understanding the performance of generative CSP models and provides a useful benchmark protocol for future research.

  3. Comprehensive empirical study and strong experimental results. A major strength of this paper is its systematic investigation of important model design choices, including architecture, training objectives, conditioning, inference, and scaling. The proposed model consistently achieves strong performance against competitive baselines across multiple benchmarks. More importantly, the systematic findings provide useful empirical insights into effective modeling strategies for flow-based molecular CSP.

Critical weaknesses
  1. Limited methodological novelty. Most of the individual components employed in PACKORA. Consequently, the methodological novelty is relatively limited. However, I view this work primarily as a systematic design study rather than a proposal of a fundamentally new generative modeling approach. From this perspective, the comprehensive empirical investigation constitutes meaningful contributions.

  2. Suggestions for improving the evaluation. While the experimental evaluation is comprehensive overall, I have a few suggestions that could further strengthen the paper:

  • Statistical robustness of structure ranking. Although the results are encouraging, it would be helpful to report the variability across different random seeds or provide uncertainty estimates to better assess the robustness of the improvements.

  • Evaluation under matched computational budgets. The current experiments mainly compare different methods using the same number of generated candidates. Since the methods have different generation costs, it would be interesting to compare their performance under the same computational budget (e.g., GPU time).

  • Analysis of generated structure diversity. The improvements in crystal solve rate may result from better individual predictions, greater diversity among generated structures, or both.

Questions for authors

See weakness.

4 comments

Dear Reviewer fdkW,

Thank you for your positive review and for viewing Packora as a systematic design study. We incorporated all three of your suggestions:

  1. Structure ranking now uses three sampling seed sets;
  2. Cost is compared over the full CSP pipeline;
  3. A new analysis shows that the solve-rate gain comes mainly from more accurate individual predictions.

In the revised PDF, changes for your review are in green\color{#1B7F3B}{\text{green}}.

W1. Limited methodological novelty: most of the individual components employed in Packora are not new.

We agree, and thank you for assessing Packora as a design study whose systematic investigation is a meaningful contribution. While the methods are not new, several of its findings are new for generative molecular CSP:

  • Loss: L1L^1 endpoint regression clearly outperforms L2L^2 (Sol_C 0.486 vs 0.407).
  • Backbone: Pairmixer, without triangle attention, outperforms the Pairformer used by OXtal (0.474 vs 0.462); Pairformer leads early in training but degrades later.
  • Conditioning: training with templates half the time helps even when none is given. Without optional inputs, p_tpl = 0.5 reaches 0.459, versus 0.442 for a model never trained with templates.
  • Scaling: widening only the single track raises per-sample success but lowers target coverage; widening both tracks recovers coverage and keeps the per-sample success.

W2a. Statistical robustness of structure ranking: please report the variability across different random seeds or provide uncertainty estimates.

We incorporated your comment by repeating the structure-ranking evaluation with three sampling seed sets. Packora's lead holds across seeds: on both benchmarks, Packora-M and Packora-L recover more experimental forms than CLARI-H, rank them higher, and relax about twice as fast. Specifically, for each model, we generated two more sets of 1,000 candidates per CSD entry with new sampling seeds and evaluated them with the same relaxation and ranking pipeline. The table reports the mean ± SD over the three sets (best in bold, second best in italics).

Single-polymorph (26 targets)

ModelTop-10 ↑Mean rank ↓Convergence rate† (%) ↑
CLARI-H14.0 ± 1.04.6 ± 0.714.7 ± 0.4%
Packora-M17.3 ± 0.62.1 ± 0.141.6 ± 0.2%
Packora-L18.0 ± 1.01.9 ± 0.351.8 ± 0.2%

Multi-polymorph (29 forms)

ModelTop-10 ↑Mean rank ↓Convergence rate† (%) ↑
CLARI-H11.3 ± 0.681.7 ± 20.713.2 ± 0.2%
Packora-M13.7 ± 0.635.1 ± 10.543.1 ± 1.5%
Packora-L13.0 ± 1.750.6 ± 4.858.1 ± 0.7%

†Convergence rate is the fraction of candidates converged within 100 BFGS steps.

We revised Section 4.3, Fig. 5 and Appendix N (Tables 19–22) to report the three-seed results. We appreciate your suggestion.

W2b. Evaluation under matched computational budgets: how do the methods compare under the same computational budget (e.g., GPU time)?

At same GPU time budget, Packora is more accurate than CLARI-H. We optimized our inference procedure (will be publicly released), and at equal throughput of about 35 candidates/s, Packora-M solves 36.4% of targets versus 27.3% for CLARI-H, where the solve rates (strict) are averaged over the test set with 30 candidates per target:

ModelStepsThroughput (candidates / s) ↑Solve rate (%) ↑
CLARI-H2035.027.3
CLARI-H2003.530.3
Packora-M20035.636.4
Packora-L20010.338.9

Also, please note that Packora's candidates also relax faster than CLARI-H's, because they converge in fewer optimization steps. Relaxation dominates the per-candidate cost of the pipeline, so Packora is also cheaper end to end.

We revised Appendix F.2 to add these results and will report our optimized inference code with our paper.

W2c. Analysis of generated structure diversity: do the solve-rate gains result from better individual predictions, greater diversity among generated structures, or both?

The gain comes mainly from better individual predictions: Packora candidates are correct about twice as often, and concentrate near low-energy minima. Three analyses support this.

  1. Per-candidate success (Table A). Packora doubles Sol_S, the chance that a single candidate is correct.
  2. Gain decomposition (Table B). We split Packora-L's Sol_C@30 gain over CLARI-H by which model solves each target. About 80% of the gain comes from targets both models solve, where Packora finds correct candidates more often, and the remaining 20% from targets only Packora-L solves. Packora-M and the standard criterion show the same pattern.
  3. Unique structures after relaxation. We relax all candidates and cluster those within 10 kJ/mol of the lowest energy with StructureMatcher. Packora places 1.3–1.6× more candidates in this window than CLARI-H, but they form the same number of unique structures or fewer: 123 and 138 vs 141 per target on single-polymorph, and 329 and 325 vs 435 per system on multi-polymorph. Its gain therefore does not come from greater diversity.

Table A. Per-candidate success Sol_S and solve rate Sol_C@30 (%; best in bold, second best in italics)

ModelSol_S ↑Sol_C@30 ↑
CLARI-H4.727.3
Packora-M9.036.4
Packora-L9.238.9

Table B. Contribution to Packora-L's Sol_C@30 gain over CLARI-H, by target group (points)

Target groupTargetsContribution
Both models solve514+9.1
Only Packora-L solves138+2.9
Only CLARI-H solves46−0.4
Neither solves1940.0
Total892+11.6

We added this analysis to Appendix H.1.


Thank you again for suggestions that strengthened the paper; we are happy to answer any further questions.

Thanks for the response from the authors. I do not have further concerns. I will keep my positive rating.

Thank you for reviewing our response and for keeping your positive rating. We appreciate your time and helpful feedback throughout the review process.

Reviewer_Ha9XOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
2
Soundness
2
Read review
Paper summary

PACKORA is a flow-matching model that predicts atomic coordinates and the unit cell together, starting from a molecular graph. It accepts any subset of three optional conditions: 3D conformers, stereochemical labels and space group. It also supports multi-component and organometallic crystals. The architecture puts Pairmixer for pair refinement in front of a pair-biased DiT trunk. The paper also builds an evaluation suite that separates structure generation from structure ranking. PACKORA reports the highest solve rate on all six generation benchmarks, and better experimental-form recovery, and ranks on two ranking benchmarks.

Critical strengths

The design study is the contribution and it is done well. The caching result is a concrete win with a hard number, since 0.005 SolC buys a 20.1× reduction in 200-step sampling time. Negative results are reported, including that pre-entry Pairmixer beats the shipped configuration. The protocol separates generation from ranking under one relaxation pipeline, which fixes a real confound in prior work. Code is released and compute is itemized.

Critical weaknesses
  • My biggest concern is that close calls have no training-seed variance. Several decisions rest on small margins: 0.474 against 0.462 (Pairmixer against Pairformer), 0.479 against 0.474 (pre-entry against post-entry), 0.459 against 0.450 (template dropout). The bootstrap in the main results resamples 30 candidates from a pool of 1,000. This captures the sampling variance of one trained model, not the variance between training runs. At about 434 GPU-hours per run, full seed replication is not affordable, and I accept that. But two seeds on the three or four decisions with margins below about 0.015 would be proportionate. Please add them.

  • There is no limitations section. The paper goes from §5 to §6 Conclusion. Please add it.

  • The design choices are selected in a different regime from the final models. The ablations train 300 epochs on crystals of at most 300 atoms (§5.1). The shipped models train in two curriculum stages up to 512 atoms (§4.1). So the recipe is tuned on smaller molecules and shorter schedules, then used at a larger scale. §5.6 shows that conclusions do shift with capacity, and that is the same kind of regime dependence. No experiment confirms that the selected choices still hold at the final scale. Please verify a few key choices at final scale.

  • The parameter-count matching is uneven across the comparison. The models run from 88M to 354M: OXtal 90M, CLARI-M 88M, CLARI-L 173M, CLARI-H 354M, PACKORA-M 88M, PACKORA-L 187M. PACKORA-M against CLARI-M is a clean matched pair. PACKORA-L beating CLARI-H at about half the parameters would be a strong claim. Please state which comparisons are parameter-matched.

Questions for authors

Please see the weakness.

3 comments

Dear Reviewer Ha9X,

Thank you for your detailed review and for recognizing the design study, the caching trade-off and our reporting of negative results. In the revised PDF, changes for your review are in orange\color{#B35900}{\text{orange}}.

W1. Close calls have no training-seed variance: please add two seeds on the three or four decisions with margins below about 0.015.

We agree, and we are replicating all four decisions with margins below 0.015 on two new training seeds (124 and 125). Please note that this requires ten additional training runs, about 4,200 H200 GPU-hours (~175 GPU-days) and about 9,000 CPU-core-hours of COMPACK evaluation, so not all of them can finish within the discussion period. All runs except the seed-125 Pairformer, whose training alone takes about 1,200 GPU-hours, are complete; its result will be included in the camera-ready version.

Averaged over seeds, three decisions keep the paper's winner, and the fourth, the noisy-input entry, becomes a tie. Post-entry, which we chose for its 20.1× faster sampling, matches pre-entry (+0.000 ± 0.003), so the speedup costs no accuracy. The table shows the mean over seeds 123–125 (123–124 for the backbone so far) and the SE of the paired difference across seeds.

DecisionCompared (chosen first)Mean Sol_C (difference ± SE)Same winner as single seed?
BackbonePairmixer vs Pairformer0.474 vs 0.466 (+0.008 ± 0.004)✓
Template dropoutp_tpl = 0.5 vs 00.458 vs 0.445 (+0.013 ± 0.003)✓
Time samplerBeta-uniform vs Beta0.482 vs 0.475 (+0.007 ± 0.004)✓
Noisy-input entrypost vs pre0.475 vs 0.475 (+0.000 ± 0.003)Tie

The revised PDF reports these results per decision in a new appendix (Appendix Q), and Section 5.2 now states that post- and pre-entry perform on par across seeds.


W2. There is no limitations section: please add it.

We appreciate your suggestion and have incorporated your comment by adding the following paragraph to the conclusion:

Limitations. Our design studies vary one component at a time while holding the others fixed. They therefore characterize the effect of each choice rather than identify a jointly optimal configuration. Moreover, the preferred configuration may depend on the candidate budget; for instance, the model widths that maximize per-sample success differ from those that maximize target coverage.


W3. The design choices are selected in a different regime from the final models (300 epochs on at most 300 atoms, versus two curriculum stages up to 512 atoms): please verify a few key choices at final scale.

We will verify the two key decisions at final scale, template dropout (p_tpl = 0 versus the shipped 0.5) and pre- versus post-entry Pairmixer, and report them in the camera-ready version; these runs cannot finish within the discussion period. A final-scale Packora-M run takes 857 H200 GPU-hours, twice a 300-epoch ablation run, and evaluating it on the six benchmarks takes about 1,200 CPU-core-hours of COMPACK matching. These runs share our GPUs with the 4,200 GPU-hours experiments for addressing W1.


W4. The parameter-count matching is uneven across the comparison (88M to 354M): please state which comparisons are parameter-matched.

Two comparisons are parameter-matched, Packora-M vs CLARI-M (88M each) and Packora-L vs CLARI-L (187M vs 173M), and Packora wins both on all six benchmarks under both the standard criterion (at least 8 of 15 molecules matched) and the strict one (all 15). Packora-M is also matched to OXtal (88M vs 90M) and outperforms it on all five benchmarks OXtal reports.

CLARI-H (354M) is not parameter-matched, yet Packora-L, with about half its parameters, beats it on all six benchmarks under both criteria, and Packora-M, with a quarter, does so on all six under the standard criterion and on five of six under the strict one, tying on CSP7.

Solve rate (%) with 30 candidates per target, averaged over all 892 targets of the six benchmarks

ModelParametersStandardStrict
CLARI-M88M49.916.1
Packora-M88M63.436.4
CLARI-L173M54.222.1
Packora-L187M65.938.9
CLARI-H354M58.227.3

We revised section 4.2 to add: "The gain holds at matched size: Packora-M and Packora-L outperform the parameter-matched CLARI-M (88M each) and CLARI-L (187M vs. 173M) on every benchmark, and Packora-L leads CLARI-H with about half its parameters."


Thank you again for your comments, which helped us strengthen the paper; we are happy to answer any further questions.

I have read the responses. I thank the authors for theie efforts. Most of my concerns have been addressed.

The regime mismatch is still open. I accept that 857 GPU-hours per run cannot fit in the discussion period, and I am satisfied with the plan to verify template dropout and pre/post-entry at final scale in the camera-ready. I would like this held to.

I will keep my score at 3. My confidence is 2 because crystal structure prediction is next to my area and not in it, so I defer to the other reviewers on the benchmark comparisons.

Thank you for reading our responses and for your continued support. We understand the regime mismatch is still open, and we remain committed to including the final-scale template dropout and pre/post-entry in the camera-ready version. We appreciate your thoughtful engagement.

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

The paper proposes PACKORA, a generative model trained with a variational flow matching objective for molecular crystal structure prediction that jointly predicts the atomic coordinates and lattice parameters. The model takes in molecular graphs, stoichiometry, and a formula-unit count. It supports multicomponent and organometallic crystals, and optional conditioning inputs of molecular conformers, stereochemical labels, and space group information. The model combines a Pairmixer-based pair refinement with a diffusion transformer, and allows condition-dependent pair representations to be computed and reused during sampling. The authors conduct rigorous and systematic studies across changes in architecture, training, conditioning, sampling, and model scaling, which is quite commendable and transparent as these directly inform their design choices.

Critical strengths
  1. Systematic design studies with clear controlled comparisons. The authors conducted rigorous studies across architecture, training, conditioning, sampling/solver, and scaling. I appreciate the breadth of these controlled comparisons, and these give useful guidance for molecular CSP model design.

  2. Improved crystal recovery before and after relaxation. At 30 candidates per target sample, PACKORA-L can improve recovery over previous method, CLARI-H, from 27.5% to 39.4% and from 15.5% to 29.6% for Teaching and Flexible configs respectively. On the separate multi-polymorph benchmark, their method also recovers 21/29 experimental crystal forms compared to 18 for CLARI-H. These results provide evidence of improved candidate quality for molecular CSP.

Critical weaknesses
  1. Conditioning accuracy is not directly evaluated. PACKORA claims "flexible conditioning on molecular conformers, stereochemical labels, and space-group information" as one main contribution. However, the experiments only measure the end-results of crystal solve rates without actually reporting whether the generated structures satisfy the supplied R/S, E/Z, or space-group labels. The improved recovery rates do not establish whether the structures they generate consistently follow the conditions requested. It will be useful to evaluate on R/S, E/Z and space-group accuracy between the ground truth and their generations, with and without conditioning information, for example.

  2. Lack of clear attribution of the gains over previous methods. The controlled design studies are rigorous and useful, but the final comparison with CLARI-H combines changes across architecture, training, conditioning, and solver/sampling. PACKORA can use optional conditional inputs and the reported protocol uses the 200-step EDM-Heun solver as compared to 20-step Heun for CLARI-H.

Questions for authors
  1. Can the authors report agreement between PACKORA's generated structures and the supplied stereochemical (R/S and E/Z) and space-group information (i.e. accuracy of stereochemistry of generated structure, and space group detected under a tolerance)? Reporting these results alongside crystal recovery / solve rates and experimental form recovery will help further establish how reliable PACKORA satisfies the requested conditions. It will be important to see the improvements because this is one of your core contributions for conditional inputs.

  2. Contribution of optional inputs: How does the final PACKORA-L model compare with CLARI-H when conformer and stereochemical conditioning are turned off? This ablation will further clarify the improvements.

  3. Comparison at similar solver/generation cost: Can the authors compare their model with other methods such as CLARI-H under the same generation-time budgets, potentially the same sampling steps? This is more of a minor issue, but it can help understand generation quality-cost tradeoffs.

4 comments

Dear Reviewer PJos,

Thank you for your careful review and for recognizing our controlled design studies and improved crystal recovery. We answer your three points with new experiments: generated structures follow the supplied conditions, and Packora's lead over CLARI-H holds with all optional inputs turned off and at the same number of sampling steps. In the revised PDF, changes for your review are in blue\color{#1F5FBF}{\text{blue}}.

W1/Q1. Conditioning accuracy is not directly evaluated: do generated structures satisfy the supplied R/S, E/Z and space-group labels?

Generated structures follow the supplied conditions: each optional input raises agreement with its own label, and on the held-out benchmarks Packora-L matches the target space group more than three times as often as CLARI-H (63.7% vs 19.0% at 0.5 Å).

We checked every raw (unrelaxed) sample against its labels, perceiving stereochemistry with the CSD Python API and space groups with spglib. A sample counts as correct only if every stereochemistry labelled site agrees. We report the validation split (30 samples per target) and the six held-out benchmarks (final models, 1,000 samples per target), with CLARI-H, which takes no optional inputs, as a reference.

Stereochemistry: % of samples with all labelled sites correct

SplitModelR/S (%)E/Z (%)
ValidationPackora without stereo32.593.8
Packora with stereo58.995.9
BenchmarksCLARI-H32.586.6
Packora-L without stereo37.587.8
Packora-L with stereo74.894.0

Space group (SG): % of samples whose detected space group equals the target

SplitModelAccuracy (%)
ValidationPackora without SG66.6
Packora with SG85.4
BenchmarksCLARI-H19.0
Packora-L without SG42.4
Packora-L with SG63.7

We note that the space group accuracy usually improves with relaxation. We also note that existing inorganic CSP methods (e.g., SymmCD, DiffCSP++) that systematically enforce space group constraints through the asymmetric unit are not directly applicable to our case, e.g., the asymmetric unit can contain only part of a molecule. Developing a systematic way to condition on space groups for molecular CSP would be an interesting future work.

We revised Section 5.4 and Appendix M to reflect these results.

W2/Q2. Lack of clear attribution of the gains: how does the final Packora-L model compare with CLARI-H when conformer and stereochemical conditioning are turned off?

With all optional inputs turned off, Packora-L still outperforms CLARI-H: it leads on all six benchmarks under both standard and strict criterion. We sampled Packora-L with every optional input (conformer template, stereochemical labels, space group) off, the same information CLARI-H receives, and kept everything else as in the main results: 1,000 candidates per target, 30-candidate solve rates by bootstrap, and COMPACK matching. A target is solved if a candidate matches at least 8 of its 15 molecules (standard) or all 15 (strict); best is in bold and second-best is in italics.

CriterionModelInputsRigidFlexibleCSP5CSP6CSP7TeachingAll 892
StandardCLARI-H—74.039.659.841.213.459.058.2
Packora-Lnone79.548.278.843.322.564.764.2
Packora-Lstereo + template (paper)79.553.079.343.323.866.365.9
StrictCLARI-H—42.415.523.08.81.127.527.3
Packora-Lnone46.724.346.320.61.137.336.8
Packora-Lstereo + template (paper)47.629.648.925.42.539.438.9

Over all 892 targets, the additional conditions improve Packora-L by 1.7 points (standard) and 2.1 points (strict), so about 80% of the gain over CLARI-H remains without them. Even Packora-M without inputs, at 88M parameters, exceeds the 354M CLARI-H on all six under the strict criterion.

We added Appendix K to report these results.

W2/Q3. Comparison at similar solver/generation cost: how does Packora compare with CLARI-H under the same generation-time budget, potentially the same sampling steps?

Packora outperforms CLARI-H both at the same number of sampling steps and at the same generation-time budget. As shown in table below, increasing CLARI-H inference steps to 200 (to match Packora) raises its solve rate by only 2–3 points. Over all 892 targets, Packora-L leads 65.9% vs 60.8% (standard) and 38.9% vs 30.3% (strict). In particular, Packora without conditioning and at matched sampling steps still leads CLARI-H (64.2% and 36.8%).

Solve rate (%) with 30 candidates per target. Bold: best; italics: second best.

CriterionModelStepsInputsRigidFlexibleCSP5CSP6CSP7TeachingAll 892
StandardCLARI-H (paper)20—74.039.659.841.213.459.058.2
CLARI-H200—77.343.268.047.318.361.460.8
Packora-L (paper)200stereo + template79.553.079.343.323.866.365.9
Packora-L200none79.548.278.843.322.564.764.2
StrictCLARI-H (paper)20—42.415.523.08.81.127.527.3
CLARI-H200—47.418.130.311.40.730.530.3
Packora-L (paper)200stereo + template47.629.648.925.42.539.438.9
Packora-L200none46.724.346.320.61.137.336.8

Please note that, even if we compare Packora and CLARI-H GPU-hour, Packora still leads, since our new speed-only optimization (which will be publicly released) allows Packora-M at 200 steps to generate as many candidates per second as CLARI-H at 20 steps (35.6 vs 35.0 on one H200).

All in all, this implies that Packora's gain over CLARI comes not only from inference time changes (e.g., more input and increased sampling budget), but also from our model architecture and training choices explored in our studies. We hope that reviewer's concern on lack of clear attribution of gains has been addressed.

These changes are reflected in Appendix K and Appendix F.2.


Thank you again for your comments, which helped us strengthen the paper; we are happy to answer any further questions.

Hi authors,

Thank you for the clearly written rebuttal and additional experiments. The new conditioning accuracy results address my concerns directly. They also show that the model can effectively use the supplied conditions and meet the stereochemistry and space group recovery better than other approaches or without conditioning them in PACKORA.

While these conditions are not perfectly satisfied, a possible future work can look more into constrained generation. Nevertheless, your results strengthened the paper and addressed my concerns. I will update my score to a 3, all the best!

Thank you for reading our rebuttal closely and for raising your score. We’re glad the conditioning accuracy results addressed your concerns. We agree that the conditions are not yet satisfied perfectly, and we see constrained generation as a promising next step. Thank you again for your constructive feedback.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.