Pith. sign in

REVIEW 3 major objections 4 minor 46 references

An agentic LLM framework called NIMMGen constructs hybrid mechanistic models that forecast epidemics, cancer dynamics, and alloy strength more accurately than prior LLM pipelines on a new realistic benchmark.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-02 22:02 UTC pith:YUBM5JEI

load-bearing objection NIMMGen's benchmark is a real step forward, but the headline results likely come from selecting on the test set rather than genuine generalization. the 3 major comments →

arxiv 2602.18008 v2 pith:YUBM5JEI submitted 2026-02-20 cs.LG cs.AIcs.CL

Are LLMs Ready for Neural-integrated Mechanistic Modeling? A Benchmark and Agentic Framework

Zihan Guan , Rituparna Datta , Mengxuan Hu , Shunshun Liu , Aiying Zhang , Prasanna Balachandran , Sheng Li , Anil Vullikanti This is my paper
classification cs.LG cs.AIcs.CL
keywords neural-integrated mechanistic modelingLLM agentsdigital twinspartial observationforecastingmodel verificationcode generationcounterfactual simulation
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper's central claim is that LLM-driven mechanistic modeling is not yet reliable for real-world use, and it establishes this with a new benchmark that adds partial observations, forecasting objectives, and a general neural-integrated model form. On this benchmark, existing LLM and non-LLM baselines produce high forecast error, frequent runtime errors, and mechanically valid but semantically wrong models. The paper then introduces NIMMGen, an agentic framework that iterates a modeling agent, a verification agent, and a reflection agent with an error-repair memory, and reports that it outperforms all baselines in forecast accuracy and code correctness across epidemiology, cancer pharmacokinetics, and alloy strength prediction. If correct, this points toward automated construction of interpretable digital twins that can answer counterfactual questions, rather than hand-built models.

Core claim

NIMMGen couples a neural network that predicts spatially and temporally varying mechanistic parameters with a mechanistic simulator, forming a compositional model f = f_mech ∘ f_NN. The framework's optimization loop proposes candidate models from a memory of top-k models, retrieved code snippets, and error messages; a verification agent checks that the translated differential equations are physically and semantically sound; the environment trains the model and returns forecasting-period RMSE; and a reflection agent converts errors and evaluation feedback into actionable natural-language guidance. The paper reports substantially lower RMSE and lower bug counts than black-box sequence models,

What carries the argument

The central object is the neural-integrated mechanistic model, where a neural network f_NN maps heterogeneous input features to mechanistic parameters θ, and those parameters drive a mechanistic simulator f_mech (e.g., compartmental epidemic ODEs). NIMMGen's engine is an agentic loop: a modeling agent generates code from a context of data insights, skeleton code, top-k historical models, and retrieved code snippets; a verification agent filters out codes whose differential equations are physically invalid or semantically inconsistent; the environment trains the model and returns forecasting RMSE to update a top-k population; and a reflection agent maintains an error-repair memory and propose

Load-bearing premise

The central claim depends on the forecasting-period RMSE not being used both to select the top models and to report the final test error; if that same period leaks into model selection, the reported performance gains over baselines may be selection artifacts rather than real forecasting skill.

What would settle it

Re-run the NIMMGen pipeline while choosing the top-k models using only the training-period RMSE, and reserve the forecasting period strictly for the final comparison; if the gap over the baselines vanishes or shrinks sharply, the original results were inflated by optimizing the reported test metric.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • If NIMMGen's performance holds, LLM agents could automate construction of calibrated digital twins for new diseases or materials with limited human modeling effort.
  • The NIMM benchmark gives the community a reusable stress test for LLM scientific coding, going beyond code compilation to partial observations and forecast objectives.
  • Semantically verified mechanistic models support counterfactual simulations, so decision-makers can probe hypothetical policies like social distancing before deployment.
  • The reported improvements in bug counts suggest agentic self-repair can make LLM-generated scientific simulators practical enough for real-time forecasting pipelines.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The reported gains may be partly an artifact of model selection: the forecasting-period RMSE is used both to pick the top-k models and to compute the final reported error, so the benchmark may inadvertently select on the test set.
  • A fairer comparison would hold out the final forecast window until the end of the search, choosing models on training-period error alone; if NIMMGen's advantage persists, the claim of genuine generalization is much stronger.
  • The framework's generalizability is asserted from only three domains and relatively small datasets; extending to larger, noisier real-time streams is a natural next test.
  • Using an LLM to verify another LLM's scientific code is circular to some degree; models that are semantically wrong in plausible ways may still pass the verification agent.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 4 minor

Summary. The paper introduces NIMM, an evaluation framework for LLM-generated neural-integrated mechanistic models under partial observation, spanning three scientific domains (public health forecasting, a PKPD cancer model, and alloy yield-strength prediction). It then proposes NIMMGen, an agentic framework with a modeling agent, a reflection agent, an error-correction module, a code-RAG tool, and an LLM-based verification agent that searches over model specifications by maintaining a top-k population. The main empirical claims are that existing LLM-based baselines (zero-shot prompting and HDTwinGen) struggle in this setting, while NIMMGen achieves state-of-the-art forecasting accuracy, higher code-success likelihood, and supports counterfactual intervention simulation. Experiments are reported on four epidemiological datasets plus cancer and materials datasets, with RMSE, bug counts, ablations, and qualitative evolution traces.

Significance. If the empirical claims held, the paper would make a useful contribution: the NIMM benchmark formalizes a realistic partial-observation forecasting setting for LLM-driven mechanistic model discovery, and NIMMGen's components (error-memory, reflection, verification) are sensible and clearly motivated. The real-time evaluation setup and explicit attention to code-level and semantic correctness are strengths. However, the current evaluation protocol has a load-bearing validation/test-separation problem, and the three-domain SOTA claim is not supported in one of the domains. Because the main empirical conclusions rest on this compromised protocol, the significance of the paper is currently not established.

major comments (3)
  1. [§4.5, §4.1, Appendix A.4, Appendix F.3] The selection criterion is the RMSE on the forecasting period, and the same quantity appears to be the reported test metric. §4.5 states that after evaluation the RMSE of the forecasting period is returned and used to update the top-k population; §4.1 states that the model with the best performance after G iterations is selected as the final model. No separate validation split is described anywhere. Appendix F.3 explicitly labels this quantity as 'validation loss'. Under the real-time setup of Appendix A.4, each shifted window still uses its forecasting period both to select models and, apparently, to report final RMSE. This makes the improvement over zero-shot and HDTwinGen in Table 1 potentially an artifact of selecting on the test set rather than evidence of genuine generalization. The central claim in §5.1 ('clear advantages in forecasting accuracy') therefore rests on an unverified
  2. [§5.1, Table 5 (Materials Science)] The concluding SOTA claim is made 'across three domains', but the materials-science evaluation compares NIMMGen only with a pure neural-network baseline. No zero-shot or HDTwinGen baselines are reported for the FCC/BCC yield-strength tasks, and no bug counts are given. Table 5 therefore cannot support the claim that NIMMGen outperforms LLM-based alternatives in the materials-science domain. The authors should add LLM baselines under identical conditions for these two tasks, or restrict the claim to the public-health and clinical-health domains.
  3. [§4.4, §5.3, Appendix F.2, Table 8] The verification agent is a central component for the 'semantic correctness' claim, but its reliability is not validated. No precision/recall or agreement with human expert judgment is reported; Appendix G.1 only shows anecdotal examples of semantically wrong codes. Moreover, Table 8 shows the verification agent's effect is non-monotonic (it improves mechanistic mode but slightly degrades hybrid mode). As the method's final performance depends on this filtering step, the paper should provide a systematic evaluation of the verifier (e.g., against human labels on a sample of generated models) and report the sensitivity of the main results to the verification threshold.
minor comments (4)
  1. [Table 3 and §5.3] The text names 'GPT-4.1, GPT-5, and GPT-5.2' as the proprietary models, but the table lists GPT-4.1, GPT-4o, and GPT-5-mini. This makes the sensitivity analysis hard to interpret. Please align text and table.
  2. [Table 1, footnote a] The zero-shot 'Bug Counts@20' is described as intentionally modified because zero-shot is run 20 times, while HDTwinGen and NIMMGen counts are over the last 20 iterations of one run. This is reasonable, but the current wording is confusing; clarify whether one bug count is recorded per generation or per run.
  3. [§5.3 / Appendix E.1] The iteration sensitivity table (Table 6) and the verification ablation report only point estimates with standard deviations over few runs; no significance tests are provided. Given the high variance in Table 1, this makes some of the comparisons fragile (e.g., Lung Cancer: 1.41±1.58 vs 1.69±2.05).
  4. [Reproducibility] No code repository or configuration files for NIMMGen are provided, only links to two baseline implementations. For a benchmark/method paper, releasing the exact prompts, skeleton codes, and RAG database construction would substantially improve reproducibility.

Circularity Check

0 steps flagged

No significant circularity: NIMMGen is an empirical benchmark and agentic search paper; the reported SOTA is not forced by definition, by fitted constants, or by a self-citation chain.

full rationale

This is an empirical methods/benchmark paper rather than a derivation chain. The central claims are that LLM baselines struggle on the NIMM benchmark and that NIMMGen achieves lower RMSE and higher code-success rates. These claims are supported by head-to-head experiments against external and prior-work baselines (LSTM, Transformer, HDTwinGen, Grad-Metapopulation) on public datasets. No equation in the paper reduces to another by construction: the neural-integrated model f = f_mech ∘ f_NN is a compositional model class, and the training objective (Eq. 5) fits parameters to the training window while forecasting is evaluated on a future window. The paper's self-citations (Guan et al., Datta et al., Cui et al.) are used for datasets, feature choices, and one baseline model; none of these is invoked as a uniqueness theorem or as a load-bearing justification for the paper's own result. The main caveat—that the forecasting-period RMSE returned in §4.5 is used to update top-k models and select the final model, and the same-origin metric is reported as the main result—is a legitimate benchmark-soundness / selection-bias concern, not a circularity. It does not make the output equal to the input by construction, nor does it rename a fitted parameter as a prediction. A reviewer should treat the selection-on-test concern as a statistical/generalization risk, but under the circularity criteria defined here the paper receives a low score.

Axiom & Free-Parameter Ledger

5 free parameters · 7 axioms · 0 invented entities

The paper is an empirical benchmark/method contribution with no formal derivation. Its central claims rest on evaluation choices: a partially-observable ODE task family, an LLM-based verification filter, several manually chosen hyperparameters, and a test-period selection rule. The main risk is the selection rule: if the forecasting-period RMSE is used both to choose models and to report test error, the reported gains are not evidence of generalization.

free parameters (5)
  • Default iteration number G = 40
    Search budget for the agentic loop; Table 6 shows sensitivity from 10 to 80 iterations, so the reported result depends on this choice.
  • Optimizer and training hyperparameters = AdamW, lr=5e-4, 1000 iterations
    Environment Engine training settings (Appendix A.4) chosen following prior work; central to fitting the neural parameters.
  • Forecast horizon H = 8 weeks (Influenza-USA, MRSA-Virginia); 28 days (COVID datasets)
    Task definition determines the training/test split and the reported RMSE, so results vary with this choice.
  • Training window T and real-time shift = T=294/191/244 etc.; shifts 1-3 weeks
    Real-time evaluation setup (Appendix A.4) defines how many points are used for training versus forecast, directly affecting error magnitudes.
  • RAG retrieval top-k and embedding model = top-3 chunks; Qwen3-Embedding-0.6B
    External code retrieval affects agent proposals, but no sensitivity analysis is given for these choices.
axioms (7)
  • domain assumption The target dynamical system follows a continuous-time ODE dxt/dt = Φ(x,u,t) (Eq. 1).
    All generated models are assumed to be ODE simulators; discrete, stochastic, or agent-based mechanisms are excluded.
  • domain assumption Neural-integrated formulation: mechanistic parameters θ are predicted by a neural network f_NN from data, θ = f_NN(D) (Eqs. 2-3).
    Imported from prior work (Chopra et al., Holt et al.); it constrains the model class the LLM is asked to produce.
  • domain assumption Partial observations consist only of reported infection counts; latent compartments (S,E,R) are unobserved but identifiable from data.
    The NIMM problem setting, §3.2, assumes the LLM can construct an identifiable model from one observed signal.
  • ad hoc to paper The LLM-based verification agent can reliably judge physical and semantic correctness of generated differential equations.
    The paper relies on this filter (§4.4) but provides no independent quantitative validation of the filter's accuracy; false positives/negatives could change the search results.
  • ad hoc to paper The forecasting-period RMSE used in top-k selection is not overfitting the reported test error.
    §4.5 uses the RMSE on the forecasting period to update the top-k population and final selection; this is only valid if that period is not also the reported test period.
  • domain assumption The real-time evaluation protocol uses fully revised data rather than data available at prediction time.
    Appendix A.4 states revised data are used following prior work; this may overstate real-world forecasting performance.
  • domain assumption Scaling the transmission rate by (1-Δ) adequately represents social-distancing interventions (Eq. 4).
    Used in the counterfactual intervention simulation (§5.2); no validation against real policy outcomes is possible or attempted.

pith-pipeline@v1.3.0-alltime-deepseek · 21321 in / 13117 out tokens · 152516 ms · 2026-08-02T22:02:50.044230+00:00 · methodology

0 comments
read the original abstract

Large language models (LLMs) have shown promise in constructing mechanistic models from data. However, existing evaluations largely focus on simplified settings and fail to capture the complexity of real-world scientific modeling. In practice, such modeling often involves neural-integrated formulations, where a mechanistic model component and a neural network component are jointly constructed, leading to a significantly more complex search space. Motivated by this gap, we introduce the Neural-Integrated Mechanistic Modeling (NIMM) benchmark, which evaluates LLM-generated neural-integrated mechanistic models across three scientific domains. Experiments on NIMM reveal that existing LLM-based approaches struggle to effectively explore this complex space, resulting in limited search stability and solution quality. To address this challenge, we propose NIMMGen, a tree-guided agentic framework that enables diversified exploration via branch-level search and improves solutions through atomic model refinement. Extensive experiments demonstrate that NIMMGen achieves state-of-the-art performance on NIMM, significantly improving search stability and solution quality.

Figures

Figures reproduced from arXiv: 2602.18008 by Aiying Zhang, Anil Vullikanti, Mengxuan Hu, Prasanna Balachandran, Rituparna Datta, Sheng Li, Shunshun Liu, Zihan Guan.

Figure 1
Figure 1. Figure 1: The datasets of the NIMM Evaluation Framework are from multiple sources and multiple levels. dataset, where Dl,t denotes the signals for the location l ∈ {1, ..., L} at timestamp t. The predicted parameters θ are also spatially- and time-varying, where θl,t denotes the predicted mechanistic parameter for the location l at the timestamp t. This fine-grained formulation enables the mechanistic model better a… view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the proposed NIMMGen framework [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Observed and simulated epidemic trajectories under simulated social distancing policies. ability in handling the three domains across different sci￾entific domains, further suggesting its broad potential in building effective mechanistic models. 5.2. Counterfactual Intervention Simulations In this section, we show the potential validity of the models generated by NIMMGen with a counterfactual intervention.… view at source ↗
Figure 4
Figure 4. Figure 4: Loss curve along with the iterations [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figure 6
Figure 6. Figure 6: Real-time evaluation Setup. observations {y:,t} T t=1: β ∗ = arg min β RMSE({yˆ:,t} T t=1, {y:,t} T t=1), where {yˆ:,t} T t=1 = f mech(f NN β ({D:,t} T t=1)). (5) At inference time, we perform forecasting by running the calibrated composite model for T + H steps. The first T steps reproduce the observed period, while the additional H steps generate predictions for the future horizon [T + 1, . . . , T + H].… view at source ↗
Figure 5
Figure 5. Figure 5: Experimental setup. (a) Mechanistic mode. (b) Hybrid Mode. the agent is responsible for generating the codes for f mech only, while the f neural is predefined in the environment. In particular, in the mechanistic mode, we use the CalibNN designed by (Chopra et al., 2022), which has been demon￾strated to be effective in extracting the information from heterogeneous data sources across multiple datasets. In … view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

46 extracted references · 4 linked inside Pith

  1. [1]

    Adhikari, B., Xu, X., Ramakrishnan, N., and Prakash, B. A. Epideep: Exploiting embeddings for epidemic forecasting. In Proceedings of the 25th ACM SIGKDD international conference on knowledge discovery & data mining, pp.\ 577--586, 2019

    2019

  2. [2]

    J., Marathe, M

    Adiga, A., Kuhlman, C. J., Marathe, M. V., Mortveit, H. S., Ravi, S. S., and Vullikanti, A. Graphical dynamical systems and their applications to bio-social systems. In International Journal of Advances in Engineering Sciences and Applied Mathematics, pp.\ 1--19. Springer, 2018. Online version appeared in Dec. 2018

    2018

  3. [3]

    Anderson, J. L. An ensemble adjustment kalman filter for data assimilation. Monthly weather review, 129 0 (12): 0 2884--2903, 2001

    2001

  4. [4]

    K., Frey, C., Moh, J., Pollock, T

    Borg, C. K., Frey, C., Moh, J., Pollock, T. M., Gorsse, S., Miracle, D. B., Senkov, O. N., Meredig, B., and Saal, J. E. Expanded dataset of mechanical properties and observed phases of multi-principal element alloys. Scientific Data, 7 0 (1): 0 430, 2020

    2020

  5. [5]

    A., and Raskar, R

    Chopra, A., Rodr \' guez, A., Subramanian, J., Quera-Bofarull, A., Krishnamurthy, B., Prakash, B. A., and Raskar, R. Differentiable agent-based epidemiology. arXiv preprint arXiv:2207.09714, 2022

    Pith/arXiv arXiv 2022

  6. [6]

    Y., Ray, E

    Cramer, E. Y., Ray, E. L., Lopez, V. K., Bracher, J., Brennen, A., Castro Rivadeneira, A. J., Gerding, A., Gneiting, T., House, K. H., Huang, Y., et al. Evaluation of individual and ensemble probabilistic forecasts of covid-19 mortality in the united states. Proceedings of the National Academy of Sciences, 119 0 (15): 0 e2113561119, 2022

    2022

  7. [7]

    R., Sifri, C

    Cui, J., Heavey, J., Klein, E., Madden, G. R., Sifri, C. D., Vullikanti, A., and Prakash, B. A. Identifying and forecasting importation and asymptomatic spreaders of multi-drug resistant organisms in hospital settings. NPJ Digital Medicine, 8 0 (1): 0 147, 2025

    2025

  8. [8]

    R., and Vullikanti, A

    Datta, R., Cui, J., Madden, G. R., and Vullikanti, A. Calypso: Forecasting and analyzing mrsa infection patterns with community and healthcare transmission dynamics. arXiv preprint arXiv:2508.13548, 2025

    Pith/arXiv arXiv 2025

  9. [9]

    d., Georg, C.-P., Koziol, T., and Schasfoort, J

    Davids, A., Rand, G. d., Georg, C.-P., Koziol, T., and Schasfoort, J. Sabcom: A spatial agent-based covid-19 model. MedRxiv, pp.\ 2020--07, 2020

    2020

  10. [10]

    Elman, J. L. Finding structure in time. Cognitive science, 14 0 (2): 0 179--211, 1990

    1990

  11. [11]

    The impact of social distancing, contact tracing, and case isolation interventions to suppress the covid-19 epidemic: A modeling study

    Ge, Y., Chen, Z., Handel, A., Martinez, L., Xiao, Q., Li, C., Chen, E., Pan, J., Li, Y., Ling, F., et al. The impact of social distancing, contact tracing, and case isolation interventions to suppress the covid-19 epidemic: A modeling study. Epidemics, 36: 0 100483, 2021

    2021

  12. [12]

    A., and Vullikanti, A

    Guan, Z., Zhao, Z., Tian, F., Nguyen, D., Bhattacharjee, P., Tandon, R., Prakash, B. A., and Vullikanti, A. A framework for multi-source privacy preserving epidemic analysis. arXiv preprint arXiv:2506.22342, 2025

    arXiv 2025

  13. [13]

    and Schmidhuber, J

    Hochreiter, S. and Schmidhuber, J. Long short-term memory. Neural computation, 9 0 (8): 0 1735--1780, 1997

    1997

  14. [14]

    Automatically learning hybrid digital twins of dynamical systems

    Holt, S., Liu, T., and van der Schaar, M. Automatically learning hybrid digital twins of dynamical systems. Advances in Neural Information Processing Systems, 37: 0 72170--72218, 2024

    2024

  15. [15]

    Kamarthi, H., Rodr \' guez, A., and Prakash, B. A. Back2future: Leveraging backfill dynamics for improving real-time predictions in future. arXiv preprint arXiv:2106.04420, 2021

    Pith/arXiv arXiv 2021

  16. [16]

    Kamarthi, H., Kong, L., Rodr \' guez, A., Zhang, C., and Prakash, B. A. Camul: calibrated and accurate multi-view time-series forecasting. In Proceedings of the ACM Web Conference 2022, pp.\ 3174--3185, 2022

    2022

  17. [17]

    E., Kevrekidis, I

    Karniadakis, G. E., Kevrekidis, I. G., Lu, L., Perdikaris, P., Wang, S., and Yang, L. Physics-informed machine learning. Nature Reviews Physics, 3 0 (6): 0 422--440, 2021

    2021

  18. [18]

    Knowledge guided machine learning: Accelerating discovery using scientific knowledge and data

    Karpatne, A., Kannan, R., and Kumar, V. Knowledge guided machine learning: Accelerating discovery using scientific knowledge and data. CRC Press, 2022

    2022

  19. [19]

    Digital twins for health: a scoping review

    Katsoulakis, E., Wang, Q., Wu, H., Shahriyari, L., Fletcher, R., Liu, J., Achenie, L., Liu, H., Jackson, P., Xiao, Y., et al. Digital twins for health: a scoping review. NPJ digital medicine, 7 0 (1): 0 77, 2024

    2024

  20. [20]

    F., Husch, A., Ley, C., Gon c alves, J., Skupin, A., and Magni, S

    Kemp, F., Proverbio, D., Aalto, A., Mombaerts, L., d’H \'e rou \"e l, A. F., Husch, A., Ley, C., Gon c alves, J., Skupin, A., and Magni, S. Modelling covid-19 dynamics and potential for herd immunity by vaccination in austria, luxembourg and sweden. Journal of Theoretical Biology, 530: 0 110874, 2021

    2021

  21. [21]

    O., Huynh, T., Wei, W

    Kwok, K. O., Huynh, T., Wei, W. I., Wong, S. Y., Riley, S., and Tang, A. Utilizing large language models in infectious disease transmission modelling for public health preparedness. Computational and Structural Biotechnology Journal, 23: 0 3254--3257, 2024

    2024

  22. [22]

    H., Gonzalez, J., Zhang, H., and Stoica, I

    Kwon, W., Li, Z., Zhuang, S., Sheng, Y., Zheng, L., Yu, C. H., Gonzalez, J., Zhang, H., and Stoica, I. Efficient memory management for large language model serving with pagedattention. In Proceedings of the 29th symposium on operating systems principles, pp.\ 611--626, 2023

    2023

  23. [23]

    Liu, S., Lee, K., and Balachandran, P. V. Integrating machine learning with mechanistic models for predicting the yield strength of high entropy alloys. Journal of Applied Physics, 132 0 (10): 0 105105, 2022

    2022

  24. [24]

    and Vullikanti, A

    Marathe, M. and Vullikanti, A. Computational epidemiology. Communications of the ACM, 56 0 (7): 0 88--96, 2013

    2013

  25. [25]

    Applications of digital twins in medicine

    Marchal, I. Applications of digital twins in medicine. nature biotechnology, 43 0 (10): 0 1606--1612, 2025

    2025

  26. [26]

    and Curtin, W

    Maresca, F. and Curtin, W. A. Mechanistic origin of high strength in refractory bcc high entropy alloys up to 1900k. Acta Materialia, 182: 0 235--249, 2020

    2020

  27. [27]

    S., Centeno, V., Phadke, A., Poor, H

    Meyur, R., Vullikanti, A., Swarup, S., Mortveit, H. S., Centeno, V., Phadke, A., Poor, H. V., and Marathe, M. V. Ensembles of realistic power distribution networks. Proceedings of the National Academy of Sciences, 119 0 (42): 0 e2205772119, 2022

    2022

  28. [28]

    Foundational research gaps and future directions for digital twins

    National Academies of Sciences, Engineering, and Medicine . Foundational research gaps and future directions for digital twins. 2023

    2023

  29. [29]

    A., Sacks, M

    Niederer, S. A., Sacks, M. S., Girolami, M., and Willcox, K. Scaling digital twins from the artisanal to the industrial. Nature Computational Science, 1 0 (5): 0 313--320, 2021

    2021

  30. [30]

    Epi-dnns: Epidemiological priors informed deep neural networks for modeling covid-19 dynamics

    Ning, X., Jia, L., Wei, Y., Li, X.-A., and Chen, F. Epi-dnns: Epidemiological priors informed deep neural networks for modeling covid-19 dynamics. Computers in biology and medicine, 158: 0 106693, 2023

    2023

  31. [31]

    Z., Shirobokov, S., Kozlovskii, B., Ruiz, F

    Novikov, A., V \ u , N., Eisenberger, M., Dupont, E., Huang, P.-S., Wagner, A. Z., Shirobokov, S., Kozlovskii, B., Ruiz, F. J., Mehrabian, A., et al. Alphaevolve: A coding agent for scientific and algorithmic discovery. arXiv preprint arXiv:2506.13131, 2025

    Pith/arXiv arXiv 2025

  32. [32]

    L., Barrio, R

    N \'u \ n ez, M., Barreiro, N. L., Barrio, R. A., and Rackauckas, C. Forecasting virus outbreaks with social media data via neural ordinary differential equations. Scientific Reports, 13 0 (1): 0 10870, 2023

    2023

  33. [33]

    Management strategies in a seir-type model of covid 19 community spread

    Rǎdulescu, A., Williams, C., and Cavanagh, K. Management strategies in a seir-type model of covid 19 community spread. Scientific reports, 10 0 (1): 0 21256, 2020

    2020

  34. [34]

    Real-time influenza forecasts during the 2012--2013 season

    Shaman, J., Karspeck, A., Yang, W., Tamerius, J., and Lipsitch, M. Real-time influenza forecasts during the 2012--2013 season. Nature communications, 4 0 (1): 0 2837, 2013

    2012

  35. [35]

    Tang, X., Qian, B., Gao, R., Chen, J., Chen, X., and Gerstein, M. B. Biocoder: a benchmark for bioinformatics code generation with large language models. Bioinformatics, 40 0 (Supplement\_1): 0 i266--i276, 2024

    2024

  36. [36]

    Y., Swarup, S., Mortveit, H., Marathe, A., Vullikanti, A., and Marathe, M

    Thorve, S., Baek, Y. Y., Swarup, S., Mortveit, H., Marathe, A., Vullikanti, A., and Marathe, M. High resolution synthetic residential energy use profiles for the united states. Scientific Data, 10 0 (1): 0 76, 2023

    2023

  37. [37]

    N., Kaiser, ., and Polosukhin, I

    Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., Kaiser, ., and Polosukhin, I. Attention is all you need. Advances in neural information processing systems, 30, 2017

    2017

  38. [38]

    Optimizing spatial allocation of seasonal influenza vaccine under temporal constraints

    Venkatramanan, S., Chen, J., Fadikar, A., Gupta, S., Higdon, D., Lewis, B., Marathe, M., Mortveit, H., and Vullikanti, A. Optimizing spatial allocation of seasonal influenza vaccine under temporal constraints. PLoS computational biology, 15 0 (9): 0 e1007111, 2019

    2019

  39. [39]

    C., Aleta, A., Rodrigues, F

    Ventura, P. C., Aleta, A., Rodrigues, F. A., and Moreno, Y. Modeling the effects of social distancing on the large-scale spreading of diseases. Epidemics, 38: 0 100544, 2022

    2022

  40. [40]

    A general-purpose machine learning framework for predicting properties of inorganic materials

    Ward, L., Agrawal, A., Choudhary, A., and Wolverton, C. A general-purpose machine learning framework for predicting properties of inorganic materials. npj Computational Materials, 2 0 (1): 0 16028, 2016

    2016

  41. [41]

    D., Hocky, G

    White, A. D., Hocky, G. M., Gandhi, H. A., Ansari, M., Cox, S., Wellawatte, G. P., Sasmal, S., Yang, Z., Liu, K., Singh, Y., et al. Assessment of chemistry knowledge in large language models that generate code. Digital Discovery, 2 0 (2): 0 368--376, 2023

    2023

  42. [42]

    Inference of seasonal and pandemic influenza transmission dynamics

    Yang, W., Lipsitch, M., and Shaman, J. Inference of seasonal and pandemic influenza transmission dynamics. Proceedings of the National Academy of Sciences, 112 0 (9): 0 2723--2728, 2015

    2015

  43. [43]

    and Curtin, W

    Yin, B. and Curtin, W. A. First-principles-based prediction of yield strength in the rhirpdptnicu high-entropy alloy. npj Computational Materials, 5 0 (1): 0 14, 2019

    2019

  44. [44]

    A survey on agent workflow--status and future

    Yu, C., Cheng, Z., Cui, H., Gao, Y., Luo, Z., Wang, Y., Zheng, H., and Zhao, Y. A survey on agent workflow--status and future. In 2025 8th International Conference on Artificial Intelligence and Big Data (ICAIBD), pp.\ 770--781. IEEE, 2025

    2025

  45. [45]

    Large language models to accelerate organic chemistry synthesis

    Zhang, Y., Han, Y., Chen, S., Yu, R., Zhao, X., Liu, X., Zeng, K., Yu, M., Tian, J., Zhu, F., et al. Large language models to accelerate organic chemistry synthesis. Nature Machine Intelligence, pp.\ 1--13, 2025

    2025

  46. [46]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION format.date year duplicate empty "emp...