PaperlayerView exact revision on arXiv

Source-linked AI summary

Zatom-2: Multitask Pretraining on Atomistic Data for Generative Modeling across Domains

Miruna Cretu, Alex Abrudan, Antonia Panescu, Tynan Perez, Rishabh Anand, N. Benjamin Erichson, Michael W. Mahoney, Samuel Blau, Joseph Jacobson, Rafael Gómez-Bombarelli, Rex Ying, Tuomas Knowles, Pietro Liò, Alex Morehead

arXiv:2610.11454v1cs.LGcs.AI

TL;DR

Atomistic generative models have often been specialized to separate scientific domains, leaving open how large-scale data can support transfer across tasks and domains. Zatom-2 combines molecular and materials pretraining with generative and predictive objectives. It reports stronger molecular distribution learning than Zatom-1, strong molecule and material benchmark performance, force-regime control, and transfer to protein generation.

  • Problem

    Existing atomistic generative approaches are specialized by scientific domain, and the paper asks how large-scale atomistic data can support representations that transfer across tasks and domains.

    Source: §1 · p. 2

  • Method

    Zatom-2 jointly pretrains a unified atomistic model on OMol25 and OMat24 using generation, structure prediction, and auxiliary energy and force prediction objectives.

    Source: §1 · p. 2

  • Results

    Zatom-2 learns molecular distributions better than Zatom-1, performs strongly on molecule and material generation benchmarks, controls generated force regimes, and transfers to protein generation.

    Source: §1 · p. 2

  • Takeaways & Limitations

    The approach treats recurring local atomic environments in diverse chemical data as a potential basis for transfer to larger, data-scarce biomolecular structures.

    Source: §1 · p. 2

    Training uses only the 4M OMol25 and 1M OMat24 subsets, while within-regime force conditioning is not well captured on all data subsets.

    Source: §G.2 · p. 49

Abstract

from arXiv · show

Unified atomistic modeling has the potential to accelerate discovery in chemistry, materials science, and biology by bridging data-rich chemical domains and data-scarce biological contexts. However, existing generative approaches to atomistic modeling remain highly specialized to scientific disciplines (chemistry vs. biology) or do not leverage both high-volume organic (molecule) and inorganic (material) data for general-purpose pretraining. To this end, we introduce Zatom-2, an atomistic generative model pretrained on approximately five million structures from the OMol25 and OMat24 electronic structure datasets. Zatom-2 features a multiscale Transformer architecture coupled with conditional flow matching that supports force conditioning and foundational pretraining tasks such as generation, structure prediction, and prediction of molecular and material energies and forces. Empirically, Zatom-2 achieves better molecular distribution fidelity than Zatom-1 and achieves strong performance on existing molecule and material generation benchmarks. Zatom-2 demonstrates the ability to control sample generation across low- and high-force regimes, and enhances protein generation in a low-data setting through joint generative-predictive pretraining and transfer learning, increasing protein backbone designability in a length extrapolation setting from 67.8% without pretraining to 74.8% after finetuning on 2,000 protein domains.

1 INTRODUCTION

Atomistic models have largely remained separated by scientific domain, despite shared underlying structure in collections of interacting atoms. Zatom-2 combines molecular and materials pretraining with tasks intended to support generation and transfer across domains.

1 INTRODUCTION

  • Existing approaches do not train and finetune a single model across molecules, periodic materials, and biomolecules, although these systems share recurring local atomic environments.The paper frames cross-domain transfer as a central question for using large-scale atomistic data.Source: §1 · p. 2
  • OMol25 and OMat24 provide diverse molecular and periodic-system structures with energy and force labels, motivating atomistic pretraining for data-scarce settings.The paper emphasizes OMol25's chemical, conformational, and energy diversity as relevant to transfer.Source: §1 · p. 2
  • Zatom-2 combines OMol25 (4M) and OMat24 (1M) pretraining with generation, structure prediction, and auxiliary energy and force prediction across atomistic domains.Its architecture uses a unified atom1 and atom14 tokenization scheme for molecule, material, and all-atom protein tasks.Source: §1 · p. 2, Figure 1 · p. 3
  • Zatom-2 achieves better molecular distribution learning than Zatom-1, performs strongly on molecule and material generation benchmarks, and supports force-regime control and protein-generation transfer.The paper also highlights extrapolation to sequence lengths beyond the training distribution.Source: §1 · p. 2

2 PRELIMINARIES

The preliminaries introduce a shared coordinate-based representation for chemical and biological systems, then describe the flow-matching formulation used to generate structures.

2.1 ALL-ATOM REPRESENTATION AND MODELING

  • Atom14 represents each protein residue with 14 coordinate slots, while atom1 represents each molecular atom as a token with a 3D coordinate, enabling a shared all-atom framework.Residue and atom-level tokenization allow protein, molecule, and material data to use the same generation framework.Source: §2.1 · p. 3

2.2 CONDITIONAL FLOW MATCHING AND SAMPLING

  • Conditional flow matching learns an ODE transport from noise to data, and the sampler converts noisy coordinates into the flow parameterization before using predicted clean coordinates.This change of variables enables EDM-style sampling while retaining velocity-based flow-matching training.Source: §2.2 · pp. 3–4

3 ZATOM-2

Zatom-2 uses a multiscale atom–token Transformer and task-specific heads for atomistic prediction and generation. Force conditioning supplies a global descriptor of structural relaxation for generation and structure prediction.

3.1 MULTISCALE TRANSFORMER BACKBONE

  • The multiscale backbone passes atom features into a global token Transformer trunk and back to an atom decoder, with atom and token representations supporting local geometry and global information exchange.Task-specific heads use these representations for coordinate, identity, lattice, energy, and force predictions.Source: §3.1 · p. 4
  • The model uses a weighted multitask objective with task gates; force and energy supervision can be restricted to eligible, uncropped examples with valid annotations.The stated gating prevents target-force leakage in force-conditioned sample generation.Source: §3.1 · p. 5

3.2 FORCE CONDITIONING

  • Force conditioning supplies generation and structure prediction with the system's mean force norm as a global descriptor of structural relaxation.The motivation is to distinguish configurations with the same composition and inferred topology but different geometries.Source: §3.2 · p. 5

4 EXPERIMENTS

Experiments evaluate Zatom-2 pretrained on OMol25 and OMat24 across atomistic generation, structure prediction, scaling, and protein transfer. Results include force-regime control, molecule and material benchmark performance, and improved protein designability after pretraining.

4.1 ATOMISTIC PRETRAINING ON OMOL25 & OMAT24

  • The final 270M-parameter configuration, which adds force and energy prediction objectives, has the strongest overall performance across evaluated pretraining metrics.Earlier additions of structure prediction and OMat24 supervision preserve or partially restore generation quality and force consistency.Source: §4.1 · pp. 6–7
  • Force conditioning yields 0.69–0.93 eV ˚A−1 MAE and 0.76–1.00 AUROC across three settings, separating low- and high-force regimes more reliably than it matches force precisely.Force consistency is used as the primary metric for OMol25 structure prediction because the dataset lacks canonical reference conformers.Source: §4.1 · p. 7, Figure 2 · p. 7

4.2 GENERATION ON GEOM-DRUGS AND MP20

  • Zatom-2 reaches 95.72% validity on QM9, 96.90% validity on GEOM-Drugs, and 4.92% MetaSUN yield on MP20, improving or approaching state of the art.The MP20 yield is for unrelaxed, metastable, unique, and novel materials.Source: §4.2 · p. 8, Table 2 · p. 8

4.3 SCALING MODEL CAPACITY AND DATA COVERAGE

  • Full-dataset training outperforms 500k-example-per-dataset training across four validation metrics, while increasing model parameters provides additional gains, especially for OMat24 structure prediction.The study covers OMol25 molecular generation and force prediction, and OMat24 structure prediction and force prediction.Source: §4.3 · p. 8, Figure 3 · p. 8

4.4 TRANSFER TO LOW-DATA PROTEIN GENERATION

  • Designability rises from 67.81% to 74.80% with Zatom-2 (V) versus random initialization for protein lengths beyond the training range.The evaluation uses 129–256-residue sequences; pretraining benefits designability only when force conditioning is enabled.Source: §4.4 · p. 9, Table 3 · p. 9

4.5 LATENT SPACE ANALYSIS OF UNSEEN PROTEINS

  • Joint generative-predictive pretraining improves protein latent-space organization, with 88.1% versus 84.8% amino-acid 10-nearest-neighbor purity and a Calinski-Harabasz index of 374.9 versus 228.1.Perturbing a single final-token representation produces strongly localized decoder effects, with joint pretraining inducing greater perturbation sensitivity.Source: §4.5 · p. 10, Figure 4 · p. 10

5 CONCLUSION

Zatom-2 combines generative and predictive pretraining on OMol25 and OMat24, with reported strength across molecular, material, and protein-generation tasks. The authors also report scaling advantages and describe the work’s reproducibility and AI-use practices.

5 CONCLUSION

  • Zatom-2 is pretrained on OMol25 (4M) and OMat24 (1M), and performs strongly on molecule and material benchmarks and in-distribution and out-of-distribution protein generation.The authors report that force and energy supervision offers better separation of atomic embeddings in PCA space, boosts generative performance, and enables stronger transfer to protein generation.Source: §5 · p. 10
  • The architecture shows performance advantages with data and model scaling, motivating the authors to propose larger-scale pretraining as future work.Source: §5 · p. 10
  • The authors report using generative AI tools for research and writing support, and state that they reviewed AI-assisted work and take responsibility for the final content.They also state that source code, documentation, data-loading infrastructure, and training and inference materials are freely available to support reproducibility.Source: AI USE STATEMENT · p. 11, REPRODUCIBILITY STATEMENT · p. 11

Zatom-2 Appendices

The appendices collect related work on molecular and periodic-material generation alongside dataset, model, evaluation, and additional-results materials. The related-work sections position Zatom-2 as a shared Transformer approach spanning these domains.

Zatom-2 Appendices

  • The appendices include dataset descriptions, model details, evaluation metrics, additional results, visualizations, and a limitations section.Source: Zatom-2 Appendices · pp. 18–19

A.1 GENERATIVE MODELING OF 3D MOLECULES

  • For 3D molecule generation, Zatom-2 retains a Transformer backbone without exact rotation equivariance, adds structure prediction and force-conditioned generation, and reports competitiveness on QM9 and GEOM-Drugs.The OMol25 experiments test broader chemical and conformational coverage.Source: §A.1 · p. 20

A.2 GENERATIVE MODELING OF PERIODIC MATERIALS

  • For periodic materials, Zatom-2 uses a common Transformer trunk with domain-specific fractional-coordinate and lattice-geometry prediction, and its MP20 results demonstrate competitive generation performance.The account attributes this capability to the new atom1 tokenizer, without dedicated materials-specific generation tracks for lattice parameters.Source: §A.2 · p. 20

A.3 REPRESENTATION LEARNING AND MULTITASK ATOMISTIC PREDICTION

This material frames Zatom-2’s pretraining as a test of shared generative, structural, and property-prediction representations, then describes the datasets and splits used across atomistic and protein settings.

A.3 REPRESENTATION LEARNING AND MULTITASK ATOMISTIC PREDICTION

  • Zatom-2 combines generative flow matching, structure prediction, and energy/force prediction in one pretraining mixture to test shared features and their usefulness during MLIP-heavy finetuning.This differs from UMA, whose sole objective is atomistic property prediction.Source: §A.3 · p. 20

B.1 OMOL25 AND OMAT24

  • OMol25 spans 83 atom types and diverse chemical and reactive systems; OMat24 includes 89 elements and both near-equilibrium and distorted structures.The work uses the OMol25 4M and OMat24 1M training splits, with force-norm distributions reported for each dataset.Source: §B.1 · p. 21

B.2 QM9

  • QM9 contains about 134,000 small organic molecules with at most nine heavy atoms, and the work uses Zatom-1’s 100,000-example training subset for joint molecule and material generation.Source: §B.2 · p. 23

B.3 MP20

  • MP20 contains 45,231 inorganic structures with at most 20 atoms per unit cell; the work uses Zatom-1’s corresponding 27,138 training structures.A separate Crystalite partition is prepared for structure prediction experiments.Source: §B.3 · p. 23

B.4 GEOM-DRUGS

  • GEOM-Drugs supplies energy-annotated conformer ensembles of larger, more flexible drug-like molecules for unconditional generation and conformer-generation evaluation.The conformer task uses a 200-molecule test split containing 14,324 reference conformers.Source: §B.4 · p. 23

B.5 SCOPE

  • SCOPe organizes experimentally determined protein domains by structural and evolutionary relationships; the work builds SCOPe-1k, -2k, and -4k by cycling through its 488 fold groups to preserve structural diversity.The SCOPe ASTRAL-40 representatives are clustered so no two domains share more than 40% sequence identity.Source: §B.5 · p. 23

B.6 DATASET LICENSES

The appendices specify dataset licensing, describe Zatom-2’s forward pass and readouts, and summarize pretraining and evaluation settings.

B.6 DATASET LICENSES

  • Dataset licensing terms are recorded for the appendix datasets and LeMat-Bulk, while software and pretrained-model licenses remain separate from data licenses.Source: §B.6 · p. 23, Table 4 · p. 24

C.1 MODEL FORWARD PASS

  • The forward pass recycles coordinate estimates across passes and returns atomistic, lattice, and optional energy and force readouts.The model uses multiresolution atom and token representations, with attention stacks and gated cross-attention between resolutions.Source: §C.1 · p. 25, Table 5 · p. 26

C.2 HYPERPARAMETERS

  • Joint pretraining samples OMol25 and OMat24 equally, and force conditioning is supplied for 80% of eligible generation and structure-prediction examples.With MLIP supervision, generation, structure prediction, and MLIP task probabilities are 0.34, 0.33, and 0.33; the OMol25/OMat24 sample evaluation uses 200 steps, while QM9/MP20 uses 100.Source: §C.2 · pp. 26–27

C.3 COMPUTE RESOURCES

Training used NVIDIA A100 GPUs with 80 GB of memory, across four- or sixteen-node jobs. Pretraining accumulated an approximate 458,752 tokens per optimizer update before two augmented diffusion views.

C.3 COMPUTE RESOURCES

  • Pretraining accumulated approximately 64 × 7,168 = 458,752 tokens per optimizer update before two augmented diffusion views.The per-GPU token budget was halved for Zatom-2-L experiments while preserving the approximate maximum.Source: §C.3 · p. 27
  • Jobs used four GPUs per node on NVIDIA A100 80 GB GPUs, running on either four nodes (16 GPUs) or sixteen nodes (64 GPUs).Source: §C.3 · p. 27
  • Training-cost estimates multiply elapsed run time by the number of GPUs used for each experiment.Source: §C.3 · p. 27

D EVALUATION METRICS

The appendix specifies evaluation metrics for force conditioning, generation, structure prediction, interatomic-potential prediction, and protein design. It also describes limits on interpreting distributional scores and notes that OMol25 structure-prediction training excludes examples with disconnected fragments after bond inference.

D.1 FORCE SELF-CONSISTENCY METRICS

  • Force self-consistency compares generated samples’ mean atomic force norms with conditioning targets, while UMA target MAE and AUROC assess force-norm accuracy and low/high-force discrimination.AUROC uses a 1 eV Å−1 cutoff, except OMol25 ANI2x, which uses 2.5 eV Å−1.Source: §D.1 · p. 28

D.2 MOLECULE GENERATION METRICS

  • Molecule metrics include RDKit validity and uniqueness, fingerprint-based diversity and novelty, and PoseBusters pass rates conditional on RDKit-valid samples.Source: §D.2 · pp. 28–29

D.3 MATERIAL GENERATION METRICS

  • Material metrics assess compositional and structural validity, uniqueness, fingerprint diversity and novelty, with force diagnostics evaluated using UMA’s OMat head.Source: §D.3 · p. 29

D.4 QM9 AND MP20 SAMPLE GENERATION METRICS

  • QM9 reports each PoseBusters pass rate as the fraction of generated molecules passing that test among RDKit-valid molecules; MP20's uniqueness and novelty rates divide qualifying counts by all attempted materials.For MP20, novelty means a unique valid material has no structural match in the 5M LeMat-Bulk reference dataset, rather than meeting the default OMat24 fingerprint similarity criterion.Source: §D.4 · pp. 29–30

D.5 GEOM-DRUGS SAMPLE GENERATION METRICS

  • GEOM-Drugs joint validity metrics count distinct valid molecules or PoseBusters-valid molecules over all attempted samples, rather than conditioning on RDKit validity.Source: §D.5 · p. 30

D.6 ATOMISTIC FRÉCHET DISTANCE

  • AFD compares Gaussian fits to reference and generated atomistic embeddings, but requires a domain-matched encoder and reference set, with scores comparable only at equal sample sizes.The embedding encoder defines which structural changes the score can detect; AFD does not certify individual-structure validity.Source: §D.6 · p. 30

D.7 STRUCTURE PREDICTION METRICS

  • Structure prediction uses symmetry-aware molecular RMSD and periodic-material match rates and normalized displacement; bond inference filtering removes 60.32% of OMol25 training examples.The excluded examples contain disconnected fragments after OpenBabel bond inference; material displacement is summarized only over matched targets.Source: §D.7 · p. 31, §E.1 · p. 32, Table 9 · p. 32

D.8 INTERATOMIC POTENTIAL METRICS; D.9 PROTEIN GENERATION METRICS

  • The appendix defines energy and force prediction errors and coverage, plus protein designability, sequence success, novelty, and structure-clustering measures.Protein designability requires at least one of eight refolds below 1.5 Å Cα scRMSD; novelty compares designable backbones against the canonical SCOPe-2k training set.Source: §D.8 · p. 31, §D.9 · pp. 31–32

E.2 MLIP PERFORMANCE BEFORE AND AFTER FINETUNING

Finetuning Zatom-2 for energy and force prediction reduces prediction errors, especially for molecules, while mixed-task finetuning largely retains generative capabilities. Protein checkpoint trajectories show faster convergence for pretrained models, and joint QM9+MP20 training gives Zatom-2 a 95.72 QM9 validity score.

E.2 MLIP PERFORMANCE BEFORE AND AFTER FINETUNING

  • After 40 finetuning epochs, Zatom-2 (V)’s force MAE falls 49% on OMol25 and 8% on OMat24, with lower aggregate energy errors in both domains.OMol25 force MAE changes from 0.253 to 0.130 eV Å−1 and OMat24 from 0.637 to 0.587 eV Å−1; extended finetuning reduces them to 0.097 and 0.564 eV Å−1, respectively.Source: §E.2 · p. 33, Table 10 · p. 34
  • Mixed-task finetuning largely retains generation and structure prediction, but materials-generation and structure-prediction metrics degrade before partly recovering with extended finetuning.For Zatom-2 (V), OMol25 generation AFD improves from 0.019 to 0.016 and structure-prediction RMSD changes from 1.762 to 1.729 Å; molecular force-conditioning MAE worsens with extended finetuning.Source: §E.2 · p. 33, Table 10 · p. 34

E.3 SCOPE-2K CHECKPOINT TRAJECTORIES

  • Generation-and-structure-pretrained Zatom-2 exceeds 85% designability at 10k local steps, while scratch-trained Zatom-2 and RFdiffusion3 first exceed 85% at 15k.The scratch-trained models deteriorate more sharply after peaking; pretrained models generally preserve high designability and novelty later in training.Source: §E.3 · p. 35, Figure 8 · p. 35

E.4.1 GENERATION ON QM9, GEOM-DRUGS, AND MP20

E.5 GEOM-DRUGS CONFORMER GENERATION BENCHMARK

On GEOM-Drugs conformer generation, Zatom-2 has higher precision but lower recall than GO-Flow. On MP20 material structure prediction, it approaches Crystalite’s match rate but has higher displacement error; the remaining figures visualize generated molecules, materials, and protein backbones.

E.5 GEOM-DRUGS CONFORMER GENERATION BENCHMARK

  • Zatom-2 achieves higher GEOM-Drugs conformer precision than GO-Flow, but lower recall, trading broader reference coverage for a larger fraction of close generated candidates.COV-R and COV-P measure recall and precision at a 1.25 Å RMSD threshold; MAT-R and MAT-P report directional minimum RMSDs.Source: §E.5 · p. 36, Table 14 · p. 37

E.6 MP20 MATERIAL STRUCTURE PREDICTION

  • 64.00% vs. 66.09% is Zatom-2’s MP20 match rate versus Crystalite’s; its RMSE is higher at 0.1156 vs. 0.0337.Zatom-2 is conditioned on composition to infer lattice geometry and fractional coordinates; the evaluation uses no hyperparameter tuning.Source: §E.6 · p. 37, Table 15 · p. 37

F VISUALIZATION

  • The visualization section presents generated molecular geometries, periodic structures, and protein backbones, alongside latent representations of OMol25 and OMat24 validation examples.Source: F VISUALIZATION · p. 38

F.1 GENERATED MOLECULES

  • The molecule galleries show selected QM9 samples and force-conditioned OMol25 samples spanning all ten OMol25 subsets in a 1,000-sample batch.QM9 examples were filtered for connected, sanitizable neutral RDKit graphs; OMol25 examples show original atom identities and coordinates without sanitization or post hoc optimization.Source: §F.1 · pp. 38–39, Figure 9 · p. 38, Figure 10 · p. 39

F.2 GENERATED MATERIALS

  • The material galleries show selected MP20 unit-cell compositions and nine force-conditioned OMat24 examples, with the latter not filtered by LeMat-GenBench validity or MetaSUN criteria.The OMat24 examples are limited to cells with at most 40 atoms; neither relaxation nor symmetry refinement is applied.Source: §F.2 · pp. 40–41, Figure 11 · p. 40, Figure 12 · p. 41

F.3 GENERATED PROTEINS

  • The protein gallery selects designable helix-rich, mixed, and sheet-rich backbones near lengths 60, 90, and 120 residues after SCOPe-2k finetuning.Samples come from a 1,027-backbone evaluation with 13 samples at each length from 50 to 128 residues; designability uses a best-of-eight Cα scRMSD threshold of 1.5 Å.Source: §F.3 · p. 42, Figure 13 · p. 42

F.4 LATENT REPRESENTATIONS OF OMOL25 AND OMAT24

Energy/force prediction pretraining is associated with changes in how Zatom-2 represents molecules and materials, including greater concentration of variance in the first principal component. Atom-level embeddings show clearer OMol25 subset organization with this pretraining, while most molecular and material atoms remain distinguishable in the full representation space.

F.4 LATENT REPRESENTATIONS OF OMOL25 AND OMAT24

  • System embeddings average final atom representations across each example; the visualization samples 768 molecules and 704 materials under identical preprocessing.Examples contain 2–512 atoms, and each PCA panel is fitted independently.Source: §F.4 · p. 43
  • 95.4% versus 51.2% for OMol25 and 86.1% versus 44.3% for OMat24: energy/force prediction pretraining concentrates more variance in the first principal component.These values compare models with versus without energy/force prediction pretraining.Source: §F.4 · p. 43, Figure 14 · p. 44
  • With energy/force supervision, most molecular and material atoms remain largely distinguishable in the full representation space, despite partial overlap in projections.The degree of overlap in joint projections varies by element type and pretraining configuration.Source: §F.4 · p. 45, Figure 15 · p. 46
  • 12.6–14.9% higher fraction of ten nearest neighbors share an atom’s OMol25 subset label for H, C, N, and O with energy/force pretraining.The comparison uses same-element neighbors from other validation examples and gives each query example equal weight.Source: §F.4 · p. 45, Figure 16 · p. 47

G.1 BROADER IMPACTS

  • Shared atomistic representations could accelerate discovery and reduce separate-model training, but training costs, data biases, and dual-use risks require responsible assessment.The paper calls for application-specific safety assessment, screening of proposed designs, and experimental validation before deployment.Source: §G.1 · p. 49

G.2 LIMITATIONS

The work identifies limits in training-data scale, force conditioning, force-consistency evaluation, protein-generation data, and energy/force prediction performance.

G.2 LIMITATIONS

  • Training uses 4M OMol25 and 1M OMat24 examples, while each full dataset contains 100M examples.Training on the entire datasets is left for future work.Source: §G.2 · p. 49
  • Force conditioning distinguishes low- and high-force regimes, but does not capture within-regime conditioning well on all data subsets.The authors leave improved conditioning mechanisms, including per-atom force-vector conditioning, to future work.Source: §G.2 · p. 49
  • The UMA-based force-consistency metric is error-prone, although its mean-force-norm MAEs against dataset labels are 0.00587 eV Å−1 for OMol25 and 0.02462 eV Å−1 for OMat24.The reported comparisons are in Appendix D.1.Source: §G.2 · p. 49
  • Protein-generation experiments use a small, though diverse, protein dataset; transfer to large-scale protein datasets via finetuning remains future work.Source: §G.2 · p. 49
  • Energy and force prediction are auxiliary generative-training tasks, not intended to match specialized MLIPs, and results leave room for improvement.The authors leave these improvements for future work.Source: §G.2 · p. 49

Related papers

Loading 2610.11454v1…