AI Paper Atlas

Research map / Probabilistic and scientific modeling / Variational and generative sampling

Tokenised Flow Matching for Hierarchical Simulation Based Inference

Giovanni Charles, Cosmo Nazzareno Santoni, Seth Flaxman, Elizaveta Semenova

ICML 2026 · Probabilistic and scientific modeling › Variational and generative sampling · approach: Flow matching and flow sampling · 0 citations (OpenAlex)

Read the paperCheck your idea against 137,217 papers

AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
use likelihood-factorised single-site training instead of fully flattened NPETFMPE · Hierarchical SIR (n_s=50, N=5000)l-C2ST: 1.91e-01 -> 1.21e-04Improved
use LF instead of posterior factorisation on Gaussian Mixture at n_s=50TFMPE · Hierarchical Gaussian Mixture (N=5000)l-C2ST: 2.01e-01 -> 1.62e-03Improved
use LF instead of posterior factorisation on Two MoonsTFMPE · Hierarchical Two Moons (N=10,000)l-C2ST: 5.69e-02 -> 1.16e-01Hurt
replace surrogate samples with true simulator samples (Direct) (ablation)TFMPE with LF · Hierarchical Two MoonscalibrationImproved
swap LF for PF factorisation under TFMPE backbone (ablation)TFMPE · Hierarchical Two MoonscalibrationNo clear effect
add group embedding for site membership (ablation)TFMPE · hierarchical benchmark tasksl-C2STImproved
use MLP vector field instead of transformer (ablation)TFMPE with LF · hierarchical benchmark tasksl-C2STNo clear effect
share one estimator for surrogate and posterior (Joint) (ablation)TFMPE with LF · hierarchical benchmark tasksstabilityHurt
use LF training on haemodynamics with 2 patientsTFMPE · Haemodynamics CFD (N=1000)training cost (h): 6.6 h -> 3.3 hImproved
draw observations from neural surrogate instead of CFDTFMPE · Haemodynamics per-patient samplingwall time: 11.88 s -> 6.41 msImproved
scale to n_s=100 sites with LF versus PF on Gaussian MixtureTFMPE · Hierarchical Gaussian Mixture (n_s=100)l-C2ST: 2.44e-01 -> 1.52e-02Improved
use LF instead of PF on SLCP at n_s=100TFMPE · Hierarchical SLCP (n_s=100)l-C2ST: 1.33e-02 -> 2.75e-02Hurt

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy