arxiv: 2604.20723 · v1 · submitted 2026-04-22 · 💻 cs.LG · cs.AI

Recognition: unknown

Tokenised Flow Matching for Hierarchical Simulation Based Inference

Giovanni Charles , Cosmo Santoni , Seth Flaxman , Elizaveta Semenova

Authors on Pith no claims yet

Pith reviewed 2026-05-10 01:00 UTC · model grok-4.3

classification 💻 cs.LG cs.AI

keywords simulation based inferencehierarchical modelsflow matchinglikelihood factorisationposterior estimationneural surrogatesinfectious disease modeling

0 comments

The pith

Likelihood factorisation allows tokenised flow matching to train hierarchical posterior estimators from single-site simulations.

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

In hierarchical simulation-based inference, global parameters are shared across exchangeable sites that each have their own parameters and observations. Existing methods still require multi-site simulations per training sample even after factorising the posterior. This paper instead factorises the likelihood, trains a neural surrogate on single-site simulations, and assembles synthetic multi-site observations from it. These synthetic data then train a tokenised flow matching model that estimates the full hierarchical posterior. The method is evaluated on a new benchmark plus infectious disease and fluid dynamics examples, yielding well-calibrated posteriors at lower simulator cost.

Core claim

By factorising the likelihood rather than the posterior and using a per-site neural surrogate to generate synthetic multi-site observations, tokenised flow matching amortises inference over the full hierarchical posterior from single-site simulations alone, producing well-calibrated posteriors on hierarchical benchmarks, infectious disease models, and computational fluid dynamics problems while reducing the number of simulator evaluations.

What carries the argument

Likelihood factorisation, in which a learned per-site neural surrogate assembles synthetic multi-site observations for training a tokenised flow matching posterior estimator.

If this is right

Posterior estimation for hierarchical models requires fewer full simulator runs during training.
Function-valued observations are handled directly inside the flow matching setup.
Well-calibrated posteriors are obtained on both synthetic hierarchical benchmarks and realistic models.
The method applies to any setting with exchangeable site-level parameters and shared globals.

Where Pith is reading between the lines

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

The same likelihood factorisation could be paired with other amortised inference techniques beyond flow matching.
The introduced benchmark offers a standard testbed for comparing future hierarchical SBI algorithms.
If the surrogate generalises across sites, the approach might support sequential addition of new sites without retraining from scratch.

Load-bearing premise

A learned per-site neural surrogate of the simulator can be used to assemble synthetic multi-site observations that preserve sufficient information to amortise inference for the full hierarchical posterior.

What would settle it

Generate synthetic multi-site data from the trained surrogate, estimate the posterior with the flow matching model, then compare posterior calibration and predictive coverage on held-out real multi-site observations from the same hierarchical simulator; systematic miscalibration relative to a multi-site trained baseline would falsify the claim.

Figures

Figures reproduced from arXiv: 2604.20723 by Cosmo Santoni, Elizaveta Semenova, Giovanni Charles, Seth Flaxman.

**Figure 2.** Figure 2: Method overview: Encoder-only transformer architecture for tokenised flow matching. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗

**Figure 3.** Figure 3: Posterior consistency measured by ℓ-C2ST (lower is better) across the hierarchical SBI benchmark. LF: Likelihood Factorisation sampling (Section 2.2), implemented with TFMPE. PF: Posterior Factorisation as proposed by Heinrich et al. (2024). NPE: Neural Posterior Estimation (Papamakarios & Murray, 2016; Boelts et al., 2024). SNPE: Sequential Neural Posterior Estimation (Papamakarios et al., 2019). Simform… view at source ↗

**Figure 4.** Figure 4: SEIR posterior comparison: a Pairplot of posterior estimates for global parameter [PITH_FULL_IMAGE:figures/full_fig_p010_4.png] view at source ↗

**Figure 5.** Figure 5: TARP calibration diagnostic for TFMPE on the seasonal SEIR model with 100 sites. [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗

**Figure 6.** Figure 6: Detailed appendix figures for the method overview: a Example token layout under the [PITH_FULL_IMAGE:figures/full_fig_p021_6.png] view at source ↗

**Figure 7.** Figure 7: Posterior consistency measured by ℓ-C2ST (lower is better) across the hierarchical SBI benchmark for TFMPE ablations. Each experiment is described in Section A.5. All results are averaged over 10 independently drawn observations per task. Posterior consistency was measured as the simulation budget N increased, with sites fixed at ns = 50. 24 [PITH_FULL_IMAGE:figures/full_fig_p024_7.png] view at source ↗

**Figure 8.** Figure 8: Posterior predictive check for haemodynamics calibration showing observed outlet flow [PITH_FULL_IMAGE:figures/full_fig_p025_8.png] view at source ↗

**Figure 9.** Figure 9: Global-parameter posterior for the 16-patient haemodynamics calibration experiment. [PITH_FULL_IMAGE:figures/full_fig_p026_9.png] view at source ↗

**Figure 10.** Figure 10: Outlet-specific local-parameter posterior for the 16-patient haemodynamics calibration [PITH_FULL_IMAGE:figures/full_fig_p027_10.png] view at source ↗

**Figure 11.** Figure 11: Posterior predictive check for the scaled [PITH_FULL_IMAGE:figures/full_fig_p028_11.png] view at source ↗

read the original abstract

The cost of simulator evaluations is a key practical bottleneck for Simulation Based Inference (SBI). In hierarchical settings with shared global parameters and exchangeable site-level parameters and observations, this structure can be exploited to improve simulation efficiency. Existing hierarchical SBI approaches factorise the posterior yet still simulate across multiple sites per training sample; We instead explore likelihood factorisation (LF) to train from single-site simulations. In LF sampling we learn a per-site neural surrogate of the simulator and then assemble synthetic multi-site observations to amortise inference for the full hierarchical posterior. Building on this, we propose Tokenised Flow Matching for Posterior Estimation (TFMPE), a tokenised flow matching approach that supports function-valued observations through likelihood factorisation. To enable systematic evaluation, we introduce a benchmark for hierarchical SBI. We validate TFMPE on this benchmark and on realistic infectious disease and computational fluid dynamics models, finding well-calibrated posteriors while reducing computational cost.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Desk Editor's Note private letter to a colleague

Single-site training for hierarchical SBI via surrogate assembly and tokenised flow matching, with a new benchmark but open questions on joint calibration.

read the letter

The one thing to know is that this work lets you amortise hierarchical SBI posteriors using only single-site simulator runs. They learn a neural surrogate per site from those runs, assemble synthetic multi-site observations, and then train a tokenised flow matching model on the full posterior. What is new is the use of tokenised flow matching to support function-valued observations under this likelihood factorisation. The paper also introduces a benchmark specifically for hierarchical SBI, which fills a gap. They test the method on that benchmark plus real models in infectious disease and computational fluid dynamics, and report that the posteriors are well-calibrated at lower simulation cost. The main soft spot is in the surrogate assembly. The global parameters create dependencies between sites, but the surrogate is fit only to single-site data. When you combine the synthetic observations, any missing correlations will distort the target for the flow matcher. This could produce posteriors that look calibrated on marginals but are off on the joint. The abstract mentions validation on benchmarks and real models, but without ablations that check the surrogate's reproduction of cross-site structure, it's difficult to judge how general the calibration is. Readers working on simulation-based inference for multi-level models will get value from the efficiency gain and the new benchmark. The paper is aimed at practitioners who face high simulation costs in hierarchical settings. The work engages honestly with the literature on SBI and hierarchical models. The central argument holds up in outline, though the details on surrogate fidelity would need checking in review. I recommend sending this to peer review.

Referee Report

2 major / 2 minor

Summary. The manuscript proposes Tokenised Flow Matching for Posterior Estimation (TFMPE) for hierarchical simulation-based inference. It introduces likelihood factorisation (LF) to train a per-site neural surrogate of the simulator from single-site simulations only, then assembles synthetic multi-site observations to amortise the full hierarchical posterior. TFMPE combines this with a tokenised flow-matching posterior estimator that handles function-valued observations. A new benchmark for hierarchical SBI is presented, and the method is evaluated on this benchmark plus infectious-disease and computational-fluid-dynamics models, with the claim that it produces well-calibrated posteriors at reduced computational cost.

Significance. If the central claims hold, the work offers a practical route to lower the simulation burden in hierarchical SBI by exploiting exchangeable site structure. The introduction of a dedicated benchmark is a constructive contribution that could facilitate future comparisons. The technical choice of tokenised flow matching for function-valued data is novel within the SBI literature and, if shown to be robust, could influence amortised inference methods more broadly.

major comments (2)

[Abstract and §4] Abstract and §4 (validation experiments): the claim that TFMPE 'yields well-calibrated posteriors' is not accompanied by any description of the calibration diagnostics employed (coverage, PIT histograms, or posterior predictive checks), surrogate accuracy metrics, or controls for bias introduced by assembling synthetic multi-site data from per-site surrogates. Because the LF procedure relies on the surrogate reproducing not only marginals but also the dependence structure induced by shared global parameters, the absence of an ablation isolating this effect makes the calibration claim load-bearing and currently unsupported.
[§3] §3 (likelihood factorisation and TFMPE): the statement that synthetic multi-site observations assembled from the per-site surrogate are 'distributionally sufficient' for amortised posterior inference is presented without a formal argument or empirical test showing that cross-site correlations are preserved. Any systematic under-dispersion or missing global-site dependence in the surrogate would directly bias the flow-matching target and produce miscalibrated hierarchical posteriors, yet no such diagnostic is reported.

minor comments (2)

[§3] The notation distinguishing the per-site surrogate parameters from the tokenised flow-matching parameters is introduced without an explicit table or equation reference, making it difficult to track which quantities are learned in each stage.
[§4] Figure captions for the benchmark results should explicitly state the number of independent runs and the precise definition of 'computational cost' (wall-clock time, number of simulator calls, or both) to allow direct comparison with baselines.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive and detailed review. The comments highlight important gaps in the presentation of calibration evidence and the justification for likelihood factorisation. We address each point below and will revise the manuscript accordingly to strengthen the supporting evidence.

read point-by-point responses

Referee: [Abstract and §4] Abstract and §4 (validation experiments): the claim that TFMPE 'yields well-calibrated posteriors' is not accompanied by any description of the calibration diagnostics employed (coverage, PIT histograms, or posterior predictive checks), surrogate accuracy metrics, or controls for bias introduced by assembling synthetic multi-site data from per-site surrogates. Because the LF procedure relies on the surrogate reproducing not only marginals but also the dependence structure induced by shared global parameters, the absence of an ablation isolating this effect makes the calibration claim load-bearing and currently unsupported.

Authors: We agree that the original manuscript provided insufficient detail on calibration diagnostics and lacked explicit controls for potential bias from synthetic data assembly. In the revision we will add a new subsection to §4 that reports: (i) coverage probabilities at 50%, 90% and 95% credible levels, (ii) PIT histograms for both global and site-level parameters, and (iii) posterior predictive checks on held-out multi-site observations. We will also report surrogate accuracy via MSE and log-likelihood on single-site test simulations. Finally, we will include an ablation that compares posteriors obtained from true multi-site simulations against those obtained from LF-assembled synthetic observations, thereby isolating any effect on dependence structure. These additions will make the calibration claims fully supported. revision: yes
Referee: [§3] §3 (likelihood factorisation and TFMPE): the statement that synthetic multi-site observations assembled from the per-site surrogate are 'distributionally sufficient' for amortised posterior inference is presented without a formal argument or empirical test showing that cross-site correlations are preserved. Any systematic under-dispersion or missing global-site dependence in the surrogate would directly bias the flow-matching target and produce miscalibrated hierarchical posteriors, yet no such diagnostic is reported.

Authors: We accept that the manuscript did not supply a formal argument or explicit empirical test for preservation of cross-site dependence under likelihood factorisation. The LF construction relies on the conditional independence of sites given the global parameters, which in principle induces the correct joint distribution once the surrogate is conditioned on the shared globals; however, we will revise §3 to state this assumption explicitly and to note its limitations. In addition, we will add empirical diagnostics in the revised §4 (and in the new benchmark section) that compare the empirical covariance and dispersion of synthetic versus real multi-site observations, including a quantitative check on cross-site correlation recovery. These tests will directly address the concern about under-dispersion or missing dependence. revision: yes

Circularity Check

0 steps flagged

No significant circularity; forward methodological proposal

full rationale

The paper introduces likelihood factorisation (LF) to train per-site neural surrogates from single-site simulations, then assembles synthetic multi-site data for amortised hierarchical posterior inference via tokenised flow matching (TFMPE). No equations or claims reduce the reported posterior calibration or efficiency gains to quantities defined by the fitted parameters themselves, nor do they rely on self-citation chains for uniqueness theorems, ansatzes, or renamings of known results. Validation uses an introduced benchmark plus external infectious-disease and CFD models, keeping the derivation chain self-contained and externally falsifiable.

Axiom & Free-Parameter Ledger

2 free parameters · 2 axioms · 1 invented entities

The central claim rests on standard SBI assumptions plus the paper-specific step of assembling synthetic observations from per-site surrogates. Free parameters are the weights of the neural surrogate and flow matching networks. No invented physical entities are introduced.

free parameters (2)

Neural network parameters for per-site simulator surrogate
Fitted to single-site simulation data to approximate the true simulator.
Parameters of the tokenised flow matching posterior estimator
Trained on assembled synthetic observations to estimate the hierarchical posterior.

axioms (2)

domain assumption The likelihood of hierarchical observations factorises across sites
Invoked to justify training from single-site simulations only.
ad hoc to paper Synthetic multi-site observations assembled from per-site surrogates are distributionally sufficient for amortised posterior inference
Core unproven step of the LF sampling procedure described in the abstract.

invented entities (1)

Tokenised Flow Matching for Posterior Estimation (TFMPE) no independent evidence
purpose: To perform posterior estimation on function-valued observations using likelihood factorisation
Newly introduced method whose properties are demonstrated only within this work.

pith-pipeline@v0.9.0 · 5461 in / 1535 out tokens · 80536 ms · 2026-05-10T01:00:04.533817+00:00 · methodology

discussion (0)

Reference graph

Works this paper leans on

29 extracted references · 24 canonical work pages · 3 internal anchors

[1]

doi: 10.1007/s10439-021-02841-9

ISSN 0090-6964. doi: 10.1007/s10439-021-02841-9. URLhttps://pmc.ncbi.nlm.nih.gov/articles/PMC8671284/. Arruda, J., Pandey, V., Sherry, C., Barroso, M., Intes, X., Hasenauer, J., and Radev, S. T. Compositional amortized inference for large-scale hierarchical bayesian models. 5

work page doi:10.1007/s10439-021-02841-9
[2]

Compositional amortized inference for large-scale hierarchical Bayesian models

doi: 10.48550/arXiv.2505.14429. URLhttp://arxiv.org/abs/2505.14429. Blanco, P. J. and Müller, L. O. One-dimensional blood flow modeling in the cardiovascular system. from the conventional physiological setting to real-life hemodynamics.International Journal for Numerical Methods in Biomedical Engineering, 41(3):e70020,

work page internal anchor Pith review Pith/arXiv arXiv doi:10.48550/arxiv.2505.14429
[3]

doi: 10.1002/cnm.70020

ISSN 2040-7939. doi: 10.1002/cnm.70020. URLhttps://europepmc.org/article/med/40077955. Boelts, J., Deistler, M., Gloeckler, M., Tejero-Cantero, Á., Lueckmann, J.-M., Moss, G., Steinbach, P., Moreau, T., Muratore, F., Linhart, J., Durkan, C., Vetter, J., Miller, B. K., Herold, M., Ziaeemehr, A., Pals, M., Gruner, T., Bischoff, S., Krouglova, N., Gao, R., L...

work page doi:10.1002/cnm.70020 2040
[4]

URL https://arxiv.org/abs/2411.17337v1

doi: 10.48550/arXiv.2411.17337. URL https://arxiv.org/abs/2411.17337v1. Chen, R. T. Q., Rubanova, Y., Bettencourt, J., and Duvenaud, D. Neural ordinary differential equations.arXiv:1806.07366 [cs, stat], 12

work page doi:10.48550/arxiv.2411.17337
[5]

Neural Ordinary Differential Equations

URLhttp://arxiv.org/abs/1806.07366. Dax, M., Wildberger, J., Buchholz, S., Green, S. R., Macke, J. H., and Schölkopf, B. Flow matching for scalable simulation-based inference. 10

work page internal anchor Pith review arXiv
[6]

ArXivabs/2305.17161 (2023)

URLhttp://arxiv.org/abs/2305.17161. Deistler, M., Goncalves, P. J., and Macke, J. H. Truncated proposals for scalable and hassle-free simulation-based inference.arXiv, October

work page arXiv
[7]

Deistler, P

doi: 10.48550/arXiv.2210.04815. Dillon, J. V., Langmore, I., Tran, D., Brevdo, E., Vasudevan, S., Moore, D., Patton, B., Alemi, A., Hoffman, M., and Saurous, R. A. TensorFlow Distributions.arXiv, November

work page doi:10.48550/arxiv.2210.04815
[8]

and Langmore, Ian and Tran, Dustin and Brevdo, Eugene and Vasudevan, Srinivas and Moore, Dave and Patton, Brian and Alemi, Alex and Hoffman, Matt and Saurous, Rif A

doi: 10.48550/arXiv.1711.10604. Dormand, J. R., El-Mikkawy, M. E. A., and Prince, P. J. High-Order Embedded Runge-Kutta- Nystrom Formulae.IMA J. Numer. Anal., 7(4):423–430, October

work page doi:10.48550/arxiv.1711.10604
[9]

doi: 10.1093/imanum/7.4.423

ISSN 0272-4979. doi: 10.1093/imanum/7.4.423. Durkan, C., Bekasov, A., Murray, I., and Papamakarios, G. Neural Spline Flows.arXiv, June

work page doi:10.1093/imanum/7.4.423
[10]

Durkan, A

doi: 10.48550/arXiv.1906.04032. Flaxman, S., Mishra, S., Gandy, A., Unwin, H. J. T., Mellan, T. A., Coupland, H., Whittaker, C., Zhu, H., Berah, T., Eaton, J. W., Monod, M., Ghani, A. C., Donnelly, C. A., Riley, S., Vollmer, M. A. C., Ferguson, N. M., Okell, L. C., and Bhatt, S. Estimating the effects of non-pharmaceutical interventions on COVID-19 in Eur...

work page doi:10.48550/arxiv.1906.04032 1906
[11]

Flaxman, et al

ISSN 1476-4687. doi: 10.1038/s41586-020-2405-7. Geffner, T., Papamakarios, G., and Mnih, A. Compositional score modeling for simulation-based inference

work page doi:10.1038/s41586-020-2405-7
[12]

B., Stern, H

15 Gelman, A., Carlin, J. B., Stern, H. S., Dunson, D. B., Vehtari, A., and Rubin, D. B. Bayesian data analysis third edition (with errors fixed as of 15 february 2021)

2021
[13]

Accessed: 2026-01-27

URL https://gadm.org/. Accessed: 2026-01-27. Gloeckler, M., Deistler, M., Weilbach, C., Wood, F., and Macke, J. H. All-in-one simulation-based inference. 7

2026
[14]

Gloeckler, M

doi: 10.48550/arXiv.2404.09636. URLhttp://arxiv.org/abs/2404.09636. Habermann, D., Bürkner, P.-C., Radev, S. T., Bulling, A., Kühmichel, L., and Schmitt, M. Amortized bayesian multilevel models. 8

work page doi:10.48550/arxiv.2404.09636
[15]

Heinrich, L., Mishra-Sharma, S., Pollard, C., and Windischhofer, P

URLhttp://arxiv.org/abs/2408.13230. Heinrich, L., Mishra-Sharma, S., Pollard, C., and Windischhofer, P. Hierarchical neural simulation- based inference over event ensembles. 2

work page arXiv
[16]

Hierarchical Neural Simulation-Based Inference Over Event Ensembles

doi: 10.48550/arXiv.2306.12584. URL http: //arxiv.org/abs/2306.12584. Hermans, J., Begy, V., and Louppe, G. Likelihood-free mcmc with amortized approximate ratio estimators

work page doi:10.48550/arxiv.2306.12584
[17]

The No-U-Turn Sampler: Adaptively Setting Path Lengths in Hamiltonian Monte Carlo

URLhttp://arxiv.org/abs/1111.4246v1. Kidger, P.On Neural Differential Equations. PhD thesis, University of Oxford,

work page Pith review arXiv
[18]

arXiv preprint , year =

doi: 10.48550/arXiv.2302.03026. Linhart, J., Gramfort, A., and Rodrigues, P. L. C. L-c2st: Local diagnostics for posterior approxi- mations in simulation-based inference. 6

work page doi:10.48550/arxiv.2302.03026
[19]

Linhart, A

URLhttp://arxiv.org/abs/2306.03580v2. Lipman, Y., Chen, R. T. Q., Ben-Hamu, H., Nickel, M., and Le, M. Flow matching for generative modeling. 10

work page arXiv
[20]

Flow Matching for Generative Modeling

URLhttp://arxiv.org/abs/2210.02747v2. Lueckmann, J.-M., Boelts, J., Greenberg, D. S., Gonçalves, P. J., and Macke, J. H. Benchmarking simulation-based inference. 4

work page internal anchor Pith review Pith/arXiv arXiv
[21]

Papamakarios, G., Sterratt, D

URLhttps://proceedings.neurips.cc/paper_files/paper/2016/file/ 6aca97005c68f1206823815f66102863-Paper.pdf. Papamakarios, G., Sterratt, D. C., and Murray, I. Sequential neural likelihood: Fast likelihood-free inference with autoregressive flows. 1

2016
[22]

Papamakarios, D

URLhttp://arxiv.org/abs/1805.07226. Pfaller, M. R., Pham, J., Verma, A., Pegolotti, L., Wilson, N. M., Parker, D. W., Yang, W., and Marsden, A. L. Automated generation of 0d and 1d reduced-order models of patient-specific blood flow.International Journal for Numerical Methods in Biomedical Engineering, 38(10): e3639,

work page arXiv
[23]

doi: 10.1002/cnm.3639

ISSN 2040-7939. doi: 10.1002/cnm.3639. URLhttps://pmc.ncbi.nlm.nih.gov/ articles/PMC9561079/. 16 Radev, S. T., Schmitt, M., Pratz, V., Picchini, U., Köthe, U., and Bürkner, P.-C. Jana: Jointly amortized neural approximation of complex bayesian models. 6

work page doi:10.1002/cnm.3639 2040
[24]

Rodrigues, P

URLhttp://arxiv.org/ abs/2302.09125. Rodrigues, P. L., Moreau, T., Louppe, G., and Gramfort, A. Hnpe: Leveraging global parameters for neural posterior estimation. 2

work page arXiv
[25]

Taylor-LaPole, A

URLhttp://arxiv.org/abs/2102.06477. Taylor-LaPole, A. M., Paun, L. M., Lior, D., Weigand, J. D., Puelz, C., and Olufsen, M. S. Parameter selection and optimization of a computational network model of blood flow in single-ventricle patients.Journal of the Royal Society Interface, 22(223):20240663,

work page arXiv
[26]

doi: 10.1098/rsif.2024.0663

ISSN 1742-5689. doi: 10.1098/rsif.2024.0663. URL https://royalsocietypublishing.org/rsif/article/22/223/ 20240663/90759/Parameter-selection-and-optimization-of-a. Zaheer, M., Kottur, S., Ravanbhakhsh, S., Poczós, B., Salakhutdinov, R. R., and Smola, A. J. Deep sets.Advances in Neural Information Processing Systems, 3

work page doi:10.1098/rsif.2024.0663 2024
[27]

Deep Sets

doi: 10.48550/arXiv.1703.06114. URLhttp://arxiv.org/abs/1703.06114. 17 A Appendix A.1 Derivation of the Compositional Posterior Factorisation We derive the factorisationp(θ|y) ∝p (θ)1−ns Qns s=1 p(θ|ys)for i.i.d. observations y = (y1, . . . , yns). Applying Bayes’ rule to the full posterior and using conditional independence: p(θ|y)∝p(θ) nsY s=1 p(ys|θ).(...

work page Pith review doi:10.48550/arxiv.1703.06114
[28]

Direct" experiment revealed that little inconsistency is due to surrogate approximation error, as its metrics closely track TFMPE’s. The

while keeping the TFMPE tokenised flow-matching backbone, group embeddings, optimiser, and schedule fixed. Two TFMPE estimators are trained sequentially: a global estimatorqϕg(θg |y )on simulations with variable site countsn∈ { 1, . . . , ns} (drawn via stick-breaking so the simulation budget is spent exactly), and a local estimatorqϕl(ηs |θ g, ys)on sing...

2024
[29]

Terminal outlets use RCR (Windkessel) boundary conditions

24 where A0 is the reference area andβ is a vessel stiffness parameter. Terminal outlets use RCR (Windkessel) boundary conditions. Hierarchical parameters.Global parameters are θg = (logβ scale,logµ,logQ in), where βscale rescales a baseline stiffness profile,µ is blood viscosity, andQin sets inflow amplitude. We treat each patientas a site s∈ { 1, . . . ...

1918