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..
The Quotient Bayesian Learning Rule
Authors: Mykola Lukashchuk, Raphaël Trésor, Wouter Nuijten, Ismail Senoz, Bert Vries
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results on the Student-t distribution confirm that our method converges more rapidly and attains higher-quality solutions than previous variants of the Bayesian Learning Rule. These findings position quotient geometry as a unifying tool for efficient and principled inference across a broad class of latent-variable models. |
| Researcher Affiliation | Collaboration | 1Department of Electrical Engineering, Technical University of Eindhoven, the Netherlands 2Lazy Dynamics, Utrecht, the Netherlands 3GN Hearing, Eindhoven, the Netherlands EMAIL EMAIL |
| Pseudocode | Yes | Algorithm 1 The Quotient Bayesian Learning Rule Input: lifted prior parameters λ0, canonical projection π : Λ Ξ, data set D = {xi}N i=1, ELBO defined in the lifted space L(λ) (15), step size schedule {βt}t 0, tolerance ϵ 1: λ λ0 initialize in the lifted (joint) space 2: repeat 3: gθ θL(λ) compute natural gradient through the dual coordinates Eq. (6) 4: g λ Projker π gθ) project onto the horizontal space, defined in Eq. (16) 5: λ λ + βt g λ natural-gradient ascent step 6: until g λ 2 < ϵ 7: ξ π(λ) 8: return marginal variational posterior qξ ( ) |
| Open Source Code | Yes | The full, version-pinned codebase that recreates every number in Table 2 is archived at https:// anonymous.4open.science/r/MIRWB-C735. |
| Open Datasets | Yes | We run the methods for four different datasets that are taken from the UCI/Open ML repository: Breast Cancer Wisconsin (Diagnostic) 569 samples, 30 features [Wolberg et al., 1993]. Pima Indians Diabetes 442 samples, 10 features [Smith et al., 1988]. Sonar (Mines vs. Rocks) 208 samples, 60 features [Gorman and Sejnowski, 1988]. Spambase 4 601 samples, 57 features [Hopkins et al., 1999]. |
| Dataset Splits | Yes | Each dataset is split 80:20 (stratified) and feature-standardized using training statistics only. |
| Hardware Specification | Yes | All experiments were conducted on a Mac Book Pro (2021) equipped with an Apple M1 Pro chip and 32 GB of memory. |
| Software Dependencies | Yes | Python 3.11 (CPU-only); jax 0.6.0, numpy 2.2.4, scikit-learn 1.6.1, torch 2.6.0, pandas 2.2.3. A version-pinned pyproject.toml is included in the repository. |
| Experiment Setup | Yes | Epochs = 8,000; mini-batch size = 32. Step sizes: no hand-tuned learning rate. Do G/RDo G schedules determine ηt with ϵ = 10 3; curvature correction disabled by default (κ = 0). Monte-Carlo samples: 10 per update for BBVI ; 1 for NG variants (gradients), plus 64 fixed draws for the NG-LIN symmetric-KL distance used by RDo G. |