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..

Variational Transdimensional Inference

Authors: Laurence Davies, Daniel MacKinlay, Rafael Oliveira, Scott SIsson

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental Numerical experiments show the performance of VTI on challenging problems that scale to high-cardinality model spaces. ... We present experiments involving synthetic and real data on two representative applications: robust variable selection and directed acyclic graphs.
Researcher Affiliation Academia 1University of New South Wales 2CSIRO Data61 EMAIL, $EMAIL ˆEMAIL
Pseudocode Yes Algorithm 1 Stochastic optimization with UCB sampling (Section C.3) Algorithm 2 Vectorized Lehmer decode via leftover mask (Section G.6) Algorithm 3 Original MADE (Final Layer Construction) (Section G.8) Algorithm 4 MADE+ (Final Layer Construction) (Section G.8)
Open Source Code Yes 1Py Torch CUDA code for all experiments is available at https://github.com/daviesl/avti.
Open Datasets Yes Real data example in flow cytometry: Sachs et al. [44] use Bayesian networks to analyze multiparameter single-cell data for deriving causal influences in cellular signaling networks of human immune cells.
Dataset Splits No The range of data sizes are n = 16, 32, 4, 128, 256, 512, 1024, where n < nmax simply takes the first n samples. ... The paper does not specify explicit train/test/validation splits for any of its experiments.
Hardware Specification Yes VTI inference was conducted on a cluster of GPU nodes with mixed Nvidia RTX3090 and H100 cards. On the former we used float32 precision for MLP architectures, the latter used float64.
Software Dependencies No Py Torch CUDA code for all experiments is available at https://github.com/daviesl/avti. ... implemented in Py Torch by [14]. While PyTorch and CUDA are mentioned, specific version numbers are not provided in the text.
Experiment Setup Yes Each of Figures 5 10 is a replicate of Figure 2 in the main text, showing a sweep of 10 randomly generated data sets... using three different variational families: diagonal Gaussian MLP... a composition of 5 affine masked autoregressive flows each with 5 hidden blocks, and a composition of 4 rational quadratic spline masked autoregressive flows each with 6 hidden blocks. (Section F.1). For DAG: We set the hidden layer width to H = 10. We set the number of nodes to Nd = 10. We omit the bias parameters b(1) j , b(2) j for all edges, i.e. set β = 0. The edge inclusion probability is set to ρEdge = 0.5. (Section G.4).