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..
Asymptotically exact variational flows via involutive MCMC kernels
Authors: Zuheng (David) Xu, Trevor Campbell
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments This section presents an empirical evaluation of the four proposed flows three IRF variants and homogeneous Mix Flows (collectively referred to as IRF flows since homogeneous Mix Flows can be viewed as a special case). We compare them against two normalizing flows, Real NVP [19] and Neural Spline Flow (NSF) [71], and against the No-U-Turn Sampler (NUTS) [72]. |
| Researcher Affiliation | Academia | Zuheng Xu Trevor Campbell Department of Statistics University of British Columbia [zuheng.xu | trevor]@stat.ubc.ca |
| Pseudocode | Yes | The detailed transition procedure of involutive MCMC is described in Algorithm 1 of Appendix A.2. Consider an auxiliary variable v defined on a space V, with conditional density ρ(v | x) given x X with respect to a base measure mv on V, and the augmented target density π(x, v) := π(x)ρ(v|x). Let m := m mv be the joint base measure on X V. For an involution g:X V X V, each transition from state x proceeds in three steps: |
| Open Source Code | Yes | Code for reproducing the main experimental results is available at: https://github.com/zuhengxu/Mix Flow.jl.git. |
| Open Datasets | Yes | Our synthetic experiments consist of four 2-dimensional targets used by Xu et al. [39]: the Banana [76], Neal s funnel [77], a cross-shaped Gaussian mixture, and a warped Gaussian distribution. ... and a latent Brownian motion model (Brownian; 32-dimensional) and the Log-Gaussian Cox process model (LGCP; 1600-dimensional) from the Inference Gym library [79]. |
| Dataset Splits | No | The paper describes training procedures and evaluation against ground truth or benchmarks, but does not explicitly provide training/test/validation split percentages or sample counts for the datasets used. |
| Hardware Specification | Yes | Experiments are conducted on the following platforms: a local machine equipped with an AMD Ryzen 9 5900X CPU and 64 GB of RAM, the ARC Sockeye computing platform at the University of British Columbia, and the high-performance compute cluster provided by the Digital Research Alliance of Canada. |
| Software Dependencies | No | For NUTS benchmarks, we use the Julia package Advanced HMC.jl [84] with default settings throughout. The paper mentions the use of Adam optimizer for training but does not specify its version or the versions of other software libraries used for its own implementation. |
| Experiment Setup | Yes | All IRF flows start from the same reference distribution q0: a mean-field Gaussian trained for 10K Adam steps with batch size 10 and learning rate 10 3. All IRF flows are evaluated with 64 i.i.d. draws, while normalizing flows use 1024. ... For Neural Spline Flows (NSF), we set the spline bandwidth to B = 30, and used K = 11 knots. ... Each normalizing flow is trained via 50,000 Adam steps of batch size 32; we grid-search both the learning rates {10 4, 10 3, 10 2} and flow layers {6, 10} ... All IRF variants use RWMH kernel, with the step size tuned to achieve a 0.8 acceptance rate using bisection search between 0.001 and 10. ... We set T = 5000 for the backward IRF and homogeneous Mix Flow and ensemble IRF Mix Flow, and set T = 4000 for the IRF Mix Flow. |