Pith. sign in

REVIEW 3 major objections 6 minor 45 references

SynBridge: Bridging Reaction States via Discrete Flow for Bidirectional Reaction Prediction

T0 review · 3 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read SynBridge treats a chemical reaction as a discrete flow of atom and bond states between two molecular graphs, and reports state-of-the-art forward and retrosynthesis accuracy on USPTO-50K, USPTO-MIT, and Pistachio.

desk verdict A solid graph-based bidirectional reaction predictor with real gains over same-prior baselines, but the retro SOTA claim leans on an atom-count prior that SMILES baselines don't get, and the paper has several reporting sloppiness issues. read the letter →

arxiv 2507.08475 v1 pith:3SJGHI3X submitted 2025-07-11 cs.LG

Haitao Lin , Junjie Wang , Zhifeng Gao , Xiaohong Ji , Rong Zhu , Linfeng Zhang , Guolin Ke , Weinan E This is my paper · ORCID
classification cs.LG
keywords chemicalreactionpredictionretrosynthesisdiscreteflowmatchinggraph-to-graphtransformermoleculargraphsdiffusionbridgesmultitasklearning
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that a chemical reaction can be modelled directly as a discrete, bidirectional flow between the reactant molecular graph and the product molecular graph, rather than as a sequence of SMILES tokens. SynBridge, a graph-to-graph transformer trained with a discrete flow-matching objective, predicts atom types, aromaticity, formal charges, and bond orders as it walks the graph through intermediate states; the same model runs the flow forward for product prediction and backward for retrosynthesis, guided by a task token. On three benchmarks the authors report state-of-the-art Top-1 accuracy in both directions, and the largest gains occur in retrosynthesis, where the model is given the total reactant atom count and an atom mapping. If these results hold, a single graph-based generative model can cover both halves of reaction planning without autoregressive SMILES decoding, while remaining applicable to reactions with ionic species and charge changes.

What carries the argument

The load-bearing object is the discrete flow bridge, a flow-matching conditional probability path over a finite state space in which the conditional distribution at time $t$ is a convex combination of the source state, the target state, and the uniform distribution: $p_t(x|x_0,x_1)=\alpha_t\delta_{x_0}(x)+\beta_t\delta_{x_1}(x)+\sigma_t U_{1/K}(x)$, with $\alpha_0=\beta_1=1$ and $\alpha_1=\beta_0=\sigma_0=\sigma_1=0$. The coefficient $\gamma_t=\min\{\dot\alpha_t/\alpha_t,\dot\beta_t/\beta_t,\dot\sigma_t/\sigma_t\}$ defines a velocity field that transports probability mass along the path, and the network learns the target-endpoint distribution by cross-entropy. The same bridge runs forward (reactants to products) or backward (products to reactants) by swapping which endpoint the network is asked to recover, and a task token selects the direction. A graph-to-graph transformer with separate encoders for the current and source graphs, a cross-attention merge of the task token, and four decoding heads (atom type, aromaticity, charge, and pairwise bond type) carries the discrete distributions.

What would settle it

Run SynBridge and a sequence baseline such as T5Chem on a retrosynthesis test set where the product SMILES is given but the total reactant atom count and atom mapping are withheld, and compare Top-1 accuracy of the recovered reactant sets; if SynBridge's advantage largely disappears, the atom-count and mapping prior, rather than the discrete flow, is the source of the reported retro gains.

Watch

Extended reading notes

Core claim

SynBridge's central claim is that reaction prediction and retrosynthesis are the two directions of one discrete transportation problem. The paper defines the reactant graph $G_0$ and the product graph $G_1$ on the same $N$ atoms, with missing reactant atoms represented as dummy atoms, and builds a conditional probability path $p_t(x|x_0,x_1)=\alpha_t\delta_{x_0}(x)+\beta_t\delta_{x_1}(x)+\sigma_t U_{1/K}(x)$ for each discrete variable (atom type, aromatic flag, formal charge, bond order). A network $\varphi_\theta(G_t, G_\mp, x_{\mathrm{tsk}})$ is trained with cross-entropy to predict the opposite endpoint given the intermediate noisy graph and the source graph, and Euler-style discrete sampling moves the graph from one endpoint to the other. In experiments the model obtains the best Top-1 scores on USPTO-50K (95.9% forward, 79.4% retro), USPTO-MIT (88.4% forward, 37.8% retro), and Pistachio (94.8% forward, 66.1% retro) in single-task settings, with the retro gains attributed to the atom-count and atom-mapping priors.

Load-bearing premise

The retrosynthesis results assume that, at test time, the model is given the total number of reactant atoms $N$ and a one-to-one atom mapping between product and reactants (with missing reactant atoms supplied as dummy atoms), so if those priors are absent the reported retro Top-1 gains over sequence models like T5Chem may not transfer.

Editorial extensions

If this is right

  • A single trained SynBridge model can be used for both product prediction and retrosynthesis by reversing the direction of the discrete flow, so forward and retro capabilities come from one set of parameters rather than separate task-specific models.
  • Because the state space is graphs with explicit atom and bond variables rather than SMILES strings, the model can represent reactions involving ionic species and formal-charge changes without losing molecular topology.
  • In retrosynthesis, the model's strong Top-1 results (79.4% on USPTO-50K, 37.8% on USPTO-MIT, 66.1% on Pistachio) are achieved with the total reactant atom count and atom mapping supplied in advance, and the paper attributes the gap over sequence models to this prior.
  • Multi-step sampling improves accuracy up to a saturation point, indicating that the discrete flow gains come from iterative self-correction rather than a single feed-forward pass.
  • On the USPTO-MIT ablation, the uniform-noise term matters for self-correction: Top-1 forward accuracy is 77.5% at $\sigma=0.0$ and rises to 85.7% at $\sigma=2.0$, with the paper's reference setting at 88.4%.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The retrosynthesis evaluation always supplies the total reactant atom count and a one-to-one atom mapping; a reasonable extension is to measure SynBridge without that prior, e.g., by masking reactant atoms and letting the model propose the number of extra atoms, to see how much of the 79.4% Top-1 comes from the flow bridge rather than the prior.
  • The same discrete flow bridge could be applied to other paired graph-translation problems with known node correspondences, such as reagent suggestion, protecting-group selection, or molecular editing, since the machinery does not depend on the specific chemistry of bonds and charges.
  • The paper's comparison with sequence baselines is under a setting where those baselines do not receive the atom count; an editor-level test would check whether giving sequence models the same reactant-size hint restores parity, which would isolate the discrete-flow contribution.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

3 major / 6 minor

Summary. SynBridge is a discrete flow bridge model for chemical reaction prediction. It treats a reaction as a graph-to-graph translation between a reactant graph G0 and a product graph G1, using four discrete node/bond variables (atom type, aromaticity, formal charge, bond order) and a bidirectional training objective with a task token. The method is built on Discrete Flow Matching and is evaluated on USPTO-50K, USPTO-MIT, and Pistachio for both forward product prediction and retrosynthesis, reporting state-of-the-art or near-state-of-the-art Top-k accuracy in most settings. The paper also includes ablations on the input graph state, noise level σ, and number of sampling steps.

Significance. If the reported results are reproducible, SynBridge is a useful application of discrete flow matching to a new domain, and the unified bidirectional formulation is a practical step toward multi-task reaction prediction. The authors provide public code and trained models, which is a strength. However, the headline claim of state-of-the-art retrosynthesis performance is conditioned on giving the model the total reactant atom count and a one-to-one atom mapping, which is not available to the SMILES-based baselines; the retrosynthesis comparison is therefore not an equal-task comparison. The Pistachio dataset description is internally inconsistent, and the sampling noise schedule is described ambiguously. The mathematical core is sound but largely inherited from the cited discrete flow matching framework.

major comments (3)
  1. [Abstract; Sec. 3.3.1; Sec. 3.3.2] The unqualified state-of-the-art claim for retrosynthesis is not supported by the reported evaluation. Section 3.3.1 states that 'during retrosynthesis, the total number of atoms is given, and the model directly predicts the atomic types,' and Sec. 3.3.2 states that T5Chem 'performs autoregressive SMILES generation without access to the true atom count.' Because SynBridge receives dummy atoms that encode N and a one-to-one atom mapping, while the SMILES baselines in Tables 2 and 5 must infer the number and identity of reactant atoms from the product SMILES alone, the abstract's claim of 'state-of-the-art performance in both forward and retrosynthesis tasks' overreaches. Please either qualify the claim as retrosynthesis under a known atom-count and atom-mapping prior, or add baselines that receive the same prior.
  2. [Table 1; Sec. 3.2] The Pistachio dataset size is internally inconsistent. Table 1 reports 2,508,278 training reactions and 2,810,278 total, but Sec. 3.2 states that after preprocessing the final dataset contains 478,192 reactions, split into 408,192 training, 30,000 validation, and 40,000 test. If the reported experiments actually use the 408k split, the 'large-scale' and 'scaling-up' discussion in Sec. 3.3.1 is not supported by the 2.5M figure, and the benchmark should be described as a filtered Pistachio subset rather than Pistachio proper. Please clarify which numbers were actually used.
  3. [Sec. 3.1; Table 6; Appendix B] There is a contradiction about the noise schedule used at sampling. Section 3.1 says 'in the sampling, σt is set to 0', but Table 6 reports the setting for Table 3 as 'with G+, σ=1, n=20', and Tables B1–B3 list σ=1 for all experiments. Please clarify whether σ=1 is used only during training while the sampler sets σ=0, and specify which setting produced each reported result. This is needed for reproducibility and for interpreting the ablation study.
minor comments (6)
  1. [Sec. 2.2.1] The sentence 'source samples x0 ∼ p0 to target samples x0 ∼ p1' should be 'target samples x1 ∼ p1'.
  2. [Sec. 2.2.2] The text says the neural network approximates 'p1(x|x0) = δx0(x)', but Eq. (4) correctly targets δx1(x); this appears to be a typo.
  3. [Sec. 2.3.1] The text before Eq. (13) says the source sample for retrosynthesis should be G−, but Eq. (13) and Algorithm 3 use G+ as the source and G− as the target; please correct the text.
  4. [Eq. (12); Eq. (14)] Eq. (12) uses vθ_{t+h}(Gt, G−) whereas the sampling algorithm in Eq. (6) evaluates the velocity at time t; similarly, Eq. (14) likely should use vθ_t rather than vθ_{t−h}.
  5. [Table 6] The rows under 'noise levels' do not state whether G+ is included as input; since the preceding row is labeled 'without input G+', please label these rows explicitly.
  6. [Table 1] The caption contains the typo 'Discreption'; it should be 'Description'.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SynBridge's predictions are not equivalent to its inputs; the atom-count prior is a stated task assumption that affects benchmark comparability but does not make the derivation circular.

full rationale

The paper's core derivation is self-contained. The discrete flow bridge is constructed from the external Discrete Flow Matching theorem (Ref. 24) and a defined conditional path (Eq. 2); the learning objective (Eq. 4) trains a network to approximate the target endpoint δx1, so the predicted graph states are not recovered from fitted parameters that already encode the answer. The retrosynthesis task is formulated with a known total atom count N and a one-to-one atom mapping, with reactant-only atoms represented as dummy nodes (Sec. 2.1). This is an explicit modeling assumption, not a circular derivation: the model still must predict atom types, charges, aromaticity, and bond orders for those nodes, and the paper candidly notes that T5Chem performs autoregressive SMILES generation without access to the true atom count (Sec. 3.3.1). That caveat concerns fairness of the comparison and external validity of the unqualified SOTA claim, not circularity of the method. Hyperparameters such as sigma and the number of sampling steps are tuned on validation data, which is standard practice and does not constitute a fitted input being called a prediction. No load-bearing self-citations or imported uniqueness theorems are present. Overall circularity score 0.

Assumptions & free parameters 4 free parameters · 5 assumptions · 0 invented entities

The central claim rests on borrowed flow-matching theory, a custom scheduler, and a task formulation that includes atom-mapping and atom-count priors. The free parameters are the uniform noise level, the number of sampling steps, and the data-filtering choices; the axioms are the DFM theorem and the domain assumptions about atom mapping and state representation. No new physical entities are introduced.

free parameters (4)
  • Uniform noise level sigma (training scheduler) = 1 (training); 0 (sampling, per Sec. 3.1)
    Appears in Eq. 8; ablation in Sec. 3.4 shows sigma=2.0 gives Top-1 85.7 vs sigma=0.0 77.5 on USPTO-MIT forward, yet main tables report sigma=1 for training and sigma=0 for sampling, which conflicts with the ablation and is not resolved.
  • Number of sampling steps n = 20
    Ablation shows n=1 gives 86.2, n=100 gives 89.5, n=20 gives 87.9 on USPTO-MIT forward; n=20 is used for all main results without justification for dataset-specific tuning.
  • Element vocabulary size = 72 types + dummy (M=73)
    The choice of 72 elements (Appendix A) excludes reactions containing other elements, which is a data-filtering decision that affects both datasets and the model's applicability.
  • RXNMapper confidence threshold = 0.8
    Applied to Pistachio preprocessing; discards a large fraction of reactions, affecting the reported dataset size and results.
assumptions (5)
  • standard math Theorem 3 of Discrete Flow Matching (Gat et al., 2024) gives the conditional velocity in Eq. 3 and the CE training objective.
    The velocity field and the training objective are taken from the cited DFM theorem; the paper does not re-derive it.
  • domain assumption One-to-one atom mapping between reactant and product atoms always exists and is available at test time.
    Task formulation in Sec. 2.1 assumes each product atom corresponds to a reactant atom, with extra reactant atoms marked as dummy type M. This requires atom-mapped reactions, which are not available in general retrosynthesis settings.
  • domain assumption The four discrete variables (atom type, aromaticity, formal charge, bond order) are a sufficient state description for reaction prediction.
    Sec. 2.1 featurizes graphs with these variables; stereochemistry, isotopes, and reaction conditions are ignored.
  • ad hoc to paper The scheduler of Eq. 8 (alpha_t, beta_t, sigma_t) is a valid probability path for the discrete flow bridge.
    The paper chooses sigma*sqrt(t(1-t)) with sigma in [0,2] without a derivation; it is a design choice that the ablation shows strongly affects performance.
  • ad hoc to paper Post-hoc data filtering (72-element vocabulary, <=80 atoms, RXNMapper confidence >0.8) does not bias benchmark comparisons.
    Appendix A and Sec. 3.2 describe exclusions that alter the evaluation set; the effect on the reported SOTA numbers is not quantified.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SynBridge: Bridging Reaction States via Discrete Flow for Bidirectional Reaction Prediction." pith.science (2026). https://pith.science/paper/3SJGHI3X

@misc{pith2026250708475,
  author       = {Pith},
  title        = {Pith review of: SynBridge: Bridging Reaction States via Discrete Flow for Bidirectional Reaction Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/3SJGHI3X}},
  note         = {Machine review of arXiv:2507.08475}
}
read the original abstract

The essence of a chemical reaction lies in the redistribution and reorganization of electrons, which is often manifested through electron transfer or the migration of electron pairs. These changes are inherently discrete and abrupt in the physical world, such as alterations in the charge states of atoms or the formation and breaking of chemical bonds. To model the transition of states, we propose SynBridge, a bidirectional flow-based generative model to achieve multi-task reaction prediction. By leveraging a graph-to-graph transformer network architecture and discrete flow bridges between any two discrete distributions, SynBridge captures bidirectional chemical transformations between graphs of reactants and products through the bonds' and atoms' discrete states. We further demonstrate the effectiveness of our method through extensive experiments on three benchmark datasets (USPTO-50K, USPTO-MIT, Pistachio), achieving state-of-the-art performance in both forward and retrosynthesis tasks. Our ablation studies and noise scheduling analysis reveal the benefits of structured diffusion over discrete spaces for reaction prediction.

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

45 extracted references · 33 canonical work pages

  1. [1]

    Nature 555(7698), 604–610 (2018)

    Segler, M.H., Preuss, M., Waller, M.P.: Planning chemical syntheses with deep 16 neural networks and symbolic ai. Nature 555(7698), 604–610 (2018)

    work page 2018

  2. [2]

    Angewandte Chemie International Edition in English 28(5), 455–473 (1989)

    Corey, E.J.: The logic of chemical synthesis: multistep synthesis of complex car- bogenic molecules. Angewandte Chemie International Edition in English 28(5), 455–473 (1989)

    work page 1989

  3. [3]

    Analyst 111(3), 221–231 (1986)

    Law, J.: Expert systems in chemistry research. Analyst 111(3), 221–231 (1986)

    work page 1986

  4. [4]

    : Molecular transformer: A model for uncertainty-calibrated chemical reaction prediction

    Schwaller, P., et al. : Molecular transformer: A model for uncertainty-calibrated chemical reaction prediction. ACS Central Science 5(9), 1572–1583 (2019)

    work page 2019

  5. [5]

    : Prediction of organic reaction outcomes using machine learning

    Coley, C.W., et al. : Prediction of organic reaction outcomes using machine learning. ACS Central Science 3(5), 434–443 (2017)

    work page 2017

  6. [6]

    In: Advances in Neural Information Processing Systems (NeurIPS), pp

    Jin, W., Coley, C.W., Barzilay, R., Jaakkola, T.: Predicting organic reaction outcomes with weisfeiler–lehman network. In: Advances in Neural Information Processing Systems (NeurIPS), pp. 2607–2616 (2017)

    work page 2017

  7. [7]

    arXiv preprint arXiv:1810.04805 (2018)

    Devlin, J., et al.: Bert: Pre-training of deep bidirectional transformers for language understanding. arXiv preprint arXiv:1810.04805 (2018)

    arXiv 2018

  8. [8]

    : Language models are few-shot learners

    Brown, T., et al. : Language models are few-shot learners. Advances in neural information processing systems 33, 1877–1901 (2020)

    work page 2020

Show all 45 references
  1. [9]

    : Mapping the space of chemical reactions using attention- based neural networks

    Schwaller, P., et al. : Mapping the space of chemical reactions using attention- based neural networks. Nature Machine Intelligence 3(2), 144–152 (2021)

    2021

  2. [10]

    Chemical Science 10(2), 370–377 (2019)

    Coley, C.W., et al.: A graph-convolutional neural network model for the prediction of chemical reactivity. Chemical Science 10(2), 370–377 (2019)

    2019

  3. [11]

    In: Meila, M., Zhang, T

    Bi, H., Wang, H., Shi, C., Coley, C.W., Tang, J., Guo, H.: Non-autoregressive electron redistribution modeling for reaction prediction. In: Meila, M., Zhang, T. (eds.) Proceedings of the 38th International Conference on Machine Learn- ing (ICML). Proceedings of Machine Learnin...

    2021

  4. [12]

    In: International Conference on Learning Representations (ICLR) (2024)

    Igashov, I., Schneuing, A., Segler, M., Bronstein, M.M., Correia, B.: Retrobridge: Modeling retrosynthesis with markov bridges. In: International Conference on Learning Representations (ICLR) (2024)

    2024

  5. [13]

    arXiv preprint arXiv:2010.02398 (2021)

    Somnath, V., et al.: Learning graph-based generative models for template-free retrosynthesis. arXiv preprint arXiv:2010.02398 (2021)

    2021 arXiv

  6. [14]

    arXiv preprint arXiv:2103.15334 (2021)

    Kim, J., et al.: T5chem: A self-supervised transformer model for language modeling of molecules and reactions. arXiv preprint arXiv:2103.15334 (2021)

    2021 arXiv

  7. [15]

    arXiv preprint arXiv:2006.11239 (2020) 17

    Ho, J., et al.: Denoising diffusion probabilistic models. arXiv preprint arXiv:2006.11239 (2020) 17

    2020 arXiv

  8. [16]

    In: ICLR (2021)

    Song, J., Meng, C., Ermon, S.: Denoising diffusion implicit models. In: ICLR (2021)

    2021

  9. [17]

    In: International Conference on Learning Representations (ICLR) (2021)

    Song, Y., Sohl-Dickstein, J., Kingma, D.P., Kumar, A., Ermon, S., Poole, B.: Score-based generative modeling through stochastic differential equations. In: International Conference on Learning Representations (ICLR) (2021)

    2021

  10. [18]

    arXiv preprint arXiv:2112.10752 (2021)

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. arXiv preprint arXiv:2112.10752 (2021)

    2021 arXiv

  11. [19]

    In: ICML (2022)

    Nichol, A.Q., Dhariwal, P., Ramesh, A., Shyam, P., Mishkin, P., McGrew, B., Sutskever, I., Chen, M.: Glide: Towards photorealistic image generation and editing with text-guided diffusion models. In: ICML (2022)

    2022

  12. [20]

    In: NeurIPS (2022)

    Saharia, C., Chan, W., Saxena, S., Li, L., Whang, J., Denton, E., Ghasemipour, S.K.S., Ayan, B.K., Steiner, S., Bras, R.L., Mahler, A., Garnier, R., Courville, A., Parmar, N., Salter, G., Ho, J., Salimans, T., Fleet, D., Norouzi, M.: Photorealistic text-to-image diffusion mode...

    2022

  13. [21]

    In: ECCV (2024)

    Ma, N., Goldstein, M., Albergo, M.S., Boffi, N.M., Vanden-Eijnden, E., Xie, S.: Sit: Exploring flow and diffusion-based generative models with scalable interpolant transformers. In: ECCV (2024)

    2024

  14. [22]

    In: ICLR (2023)

    Gong, S., Li, M., Feng, J., Wu, Z., Kong, L.: Diffuseq: Sequence-to-sequence text generation with diffusion models. In: ICLR (2023)

    2023

  15. [23]

    arXiv preprint arXiv:2206.08791 (2022)

    Liu, C., Tzen, B., Lee, J.D., Yang, S.: Flow matching for generative modeling. arXiv preprint arXiv:2206.08791 (2022)

    2022 arXiv

  16. [24]

    https://arxiv.org/abs/2407.15595

    Gat, I., Remez, T., Shaul, N., Kreuk, F., Chen, R.T.Q., Synnaeve, G., Adi, Y., Lipman, Y.: Discrete Flow Matching (2024). https://arxiv.org/abs/2407.15595

    2024 arXiv

  17. [25]

    In: arXiv Preprint arXiv:2303.08979 (2023)

    Albergo, M.S., Boffi, N.M., Vanden-Eijnden, E.: Stochastic interpolants: A uni- fying framework for flows and diffusions. In: arXiv Preprint arXiv:2303.08979 (2023)

    2023 arXiv

  18. [26]

    In: International Conference on Learning Representations (ICLR) (2024)

    Zhou, L., Lou, A., Khanna, S., Ermon, S.: Denoising diffusion bridge models. In: International Conference on Learning Representations (ICLR) (2024). Poster. https://openreview.net/forum?id=FKksTayvGo

    2024

  19. [27]

    In: Inter- national Conference on Learning Representations (ICLR) (2025)

    Kim, J.H., Kim, S., Moon, S., Kim, H., Woo, J., Kim, W.Y.: Discrete diffusion schr¨ odinger bridge matching for graph transformation. In: Inter- national Conference on Learning Representations (ICLR) (2025). Poster. https://openreview.net/forum?id=tQyh0gnfqW 18

    2025

  20. [28]

    In: ICLR (2023)

    Lipman, Y., Chen, R.T.Q., Ben-Hamu, H., Nickel, M., Le, M.: Flow matching for generative modeling. In: ICLR (2023)

    2023

  21. [29]

    arXiv preprint arXiv:2412.06264 (2024)

    Lipman, Y., Havasi, M., H¨ older-Rieth, P., Shaul, N., Le, M., Karrer, B., Chen, R.T.Q., L´ opez-Paz, D., Ben-Hamu, H., Gat, I.: Flow matching guide and code. arXiv preprint arXiv:2412.06264 (2024)

    2024 arXiv

  22. [30]

    Transactions on Machine Learning Research (TMLR) (2024)

    Tong, A., Fatras, K., Malkin, N., Huguet, G., Zhang, Y., Rector-Brooks, J., Wolf, G., Bengio, Y.: Improving and generalizing flow-based generative models with minibatch optimal transport. Transactions on Machine Learning Research (TMLR) (2024)

    2024

  23. [31]

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. J. Mach. Learn. Res. 21(140), 1–67 (2020)

    2020

  24. [32]

    In: Machine Learning: Science and Technology (2022)

    Irwin, R., Dimitriadis, S., He, J., Bjerrum, E.J.: Chemformer: A pre-trained transformer for computational chemistry. In: Machine Learning: Science and Technology (2022)

    2022

  25. [33]

    In: ECCV (2022)

    Jia, M., Tang, L., Chen, B.-C., Cardie, C., Belongie, S., Hariharan, B., Lim, S.-N.: Visual prompt tuning. In: ECCV (2022)

    2022

  26. [34]

    arXiv preprint arXiv:2109.01134 (2021)

    Zhou, K., Yang, J., Loy, C.C., Liu, Z.: Learning to prompt for vision-language models. arXiv preprint arXiv:2109.01134 (2021)

    2021 arXiv

  27. [35]

    arXiv preprint arXiv:2106.09553 (2021)

    Ross, J., Belgodere, B., Chenthamarakshan, V., Padhi, I., Mroueh, Y., Das, P.: Large-scale chemical language representations capture molecular structure and properties. arXiv preprint arXiv:2106.09553 (2021)

    2021 arXiv

  28. [36]

    arXiv preprint arXiv:2211.11720 (2022)

    Shen, S., Yang, S., Zhang, T., Zhai, B., Gonzalez, J.E., Keutzer, K., Darrell, T.: Multitask vision-language prompt tuning. arXiv preprint arXiv:2211.11720 (2022)

    2022 arXiv

  29. [37]

    In: ICML (2021)

    Ramesh, A., Pavlov, M., Goh, G., Gray, S., Voss, C., Radford, A., Chen, M., Sutskever, I.: Zero-shot text-to-image generation. In: ICML (2021)

    2021

  30. [38]

    In: ICML (2021)

    Radford, A., Kim, J.W., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision. In: ICML (2021)

    2021

  31. [39]

    arXiv preprint arXiv:2302.05543 (2023)

    Zhang, L., Agrawala, M.: Adding conditional control to text-to-image diffusion models. arXiv preprint arXiv:2302.05543 (2023)

    2023 arXiv

  32. [40]

    In: Journal of Machine Learning Research (2020) 19

    Raffel, C., Shazeer, N., Roberts, A., Lee, K., Narang, S., Matena, M., Zhou, Y., Li, W., Liu, P.J.: Exploring the limits of transfer learning with a unified text-to-text transformer. In: Journal of Machine Learning Research (2020) 19

    2020

  33. [41]

    In: ICML Workshop on Robustness and Reliability of Machine Learning Systems (2021)

    Tamkin, A., Monajatipoor, M., Naik, N., Jurafsky, D., Goodman, N.: Language models as multitask learners. In: ICML Workshop on Robustness and Reliability of Machine Learning Systems (2021)

    2021

  34. [42]

    In: ACL (2021)

    Li, X.L., Liang, P.: Prefix-tuning: Optimizing continuous prompts for generation. In: ACL (2021)

    2021

  35. [43]

    https://arxiv.org/abs/2006.15426

    Sacha, M., Blaz, M., Byrski, P., Dabrowski-Tumanski, P., Chrominski, M., Loska, R., Wlodarczyk-Pruszy´ nski, P., Jastrzebski, S.: Molecule Edit Graph Attention Network: Modeling Chemical Reactions as Sequences of Graph Edits (2021). https://arxiv.org/abs/2006.15426

    2021 arXiv

  36. [44]

    In: Advances in Neural Information Processing Systems, vol

    Dai, H., Li, C., Coley, C., Dai, B., Song, L.: Retrosynthesis prediction with conditional graph logic network. In: Advances in Neural Information Processing Systems, vol. 32 (2019)

    2019

  37. [45]

    Schwaller, P., Hoover, B., Reymond, J.-L., Strobelt, H., Laino, T.: Extraction of organic chemistry grammar from unsupervised learning of chemical reactions. Science Advances 7(15), 4166 (2021) https://doi.org/10.1126/sciadv.abe4166 20 Appendix A Data Preprocess We consider th...

    2021 doi

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.