Notice: The reproducibility variables underlying each score are classified using an automated LLM-based pipeline, validated against a manually labeled dataset. LLM-based classification introduces uncertainty and potential bias; scores should be interpreted as estimates. Full accuracy metrics and methodology are described in Coakley et alK. L. Coakley, T. Snelleman, H. Hoos, and O. E. Gundersen, "The embrace of open science: An analysis of a decade of AI research and 56 800 conference papers," Under Review, 2026..
Multilevel neural simulation-based inference
Authors: Yuga Hikida, Ayush Bharti, Niall Jeffrey, Francois-Xavier Briol
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | We demonstrate through both theoretical analysis and extensive experiments that our method can significantly enhance the accuracy of SBI methods given a fixed computational budget. Our extensive experiments on models from finance, synthetic biology, and cosmology also demonstrate the significant computational advantages provided by our method. |
| Researcher Affiliation | Academia | Yuga Hikida Aalto University EMAIL Ayush Bharti Aalto University EMAIL Niall Jeffrey University College London EMAIL François-Xavier Briol University College London EMAIL |
| Pseudocode | Yes | Algorithm 1 MLMC gradient adjustment |
| Open Source Code | Yes | The code to reproduce our experiments is available at https://github.com/yugahikida/multilevel-sbi. |
| Open Datasets | Yes | We now consider a cosmological simulator using the CAMELS suite [Villaescusa-Navarro et al., 2021, 2023] one of the most computationally intensive cosmological simulations to date which comprises both lowand high-fidelity data. |
| Dataset Splits | Yes | We use 20% of the data as validation set for early stopping. |
| Hardware Specification | Yes | For all the experiments, we use a Mac M4 CPU with 16 GB memory for the training of neural networks. |
| Software Dependencies | No | For optimisation, we use batch gradient descent with the Adam optimiser [Kingma and Ba, 2015] using Pytorch [Paszke et al., 2017]. For the construction of the conditional density estimator, we use the SBI library [Tejero-Cantero et al., 2020]. We also use the Bayes Flow package [Radev et al., 2023b] for some of the figures. For the high-fidelity simulator, we use an accurate approximation implemented in Scipy [Virtanen et al., 2020]. |
| Experiment Setup | Yes | We set the learning rate to 10 4 for all the experiments. ... NLE: We use the neural spline flow (NSF) [Durkan et al., 2019]. We pick 10 bins, span of [ 7, 7] and 1 coupling layer since we only have one dimensional input. The conditioner for the NSF is a multilayer perceptron neural network (MLP) with 3 hidden layers of 50 units, and 10% dropout, trained for 10, 000 epochs. NPE: We use NSF with 3 bins, span of [ 3, 3] and 3 coupling layers. The conditioner for the NSF is a MLP with 2 hidden layers of 50 units, and 10% dropout, trained for 800 epochs. |