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..
Rao-Blackwellised Reparameterisation Gradients
Authors: Kevin H. Lam, Thang Bui, George Deligiannidis, Yee Whye Teh
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | 5 Experiments We evaluate the benefits of initial training with the R2-G2 estimator for probabilistic models. We consider two standard tasks with variational Bayesian models that utilise a mean-field approximation of the posterior: image classification with BNNs and generative modelling with hierarchical VAEs. |
| Researcher Affiliation | Academia | Kevin H. Lam Department of Statistics University of Oxford Thang D. Bui School of Computing Australian National University George Deligiannidis Department of Statistics University of Oxford Yee Whye Teh Department of Statistics University of Oxford EMAIL, EMAIL |
| Pseudocode | Yes | Algorithm 1 Forward Pass with R2-G2 Gradients. Input: matrix A, noise vector ϵ. Compute z = Aϵ. Compute β = conjugate_gradient(A, z). Compute ϵ = A β . Compute z = Aϵ . Output: stop_gradient z z ) + z . |
| Open Source Code | Yes | Question: Does the paper provide open access to the data and code, with sufficient instructions to faithfully reproduce the main experimental results, as described in supplemental material? Answer: [Yes] Justification: We provide the code needed to run our experiments. |
| Open Datasets | Yes | Appendix J mentions: MNIST (Le Cun et al., 2010): Creative Commons Attribution-Share Alike 3.0 license CIFAR-10 (Krizhevsky and Hinton, 2009): MIT license Fashion-MNIST (Xiao et al., 2017): MIT license Omniglot (Lake et al., 2015): MIT license |
| Dataset Splits | Yes | We used the standard train and test splits of both datasets. |
| Hardware Specification | Yes | Compute resources All experiments were run on a single NVIDIA V100 GPU. |
| Software Dependencies | No | The paper mentions "Py Torch (Paszke et al., 2019)" and "Adam optimiser for all experiments Kingma and Ba (2015)", but it does not specify version numbers for these software components. Therefore, it does not provide specific ancillary software details with version numbers required for reproduction. |
| Experiment Setup | Yes | Optimisation We used a batch size of 80 and the Adam optimiser for all experiments Kingma and Ba (2015). We do not add regularisation such as weight decay, dropout or batch normalisation layers. Other optimisation parameters are listed in Table 4. |