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..
Robust and Computation-Aware Gaussian Processes
Authors: Marshal Sinaga, Julien Martinelli, Samuel Kaski
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical results confirm that solving these challenges jointly leads to superior performance across both clean and outlier-contaminated settings, both on regression and high-throughput Bayesian optimization benchmarks. We evaluate the performance of our proposed RCa GP against concurrent baselines on a range of regression datasets, including a real-world dataset, followed by several high-throughput Bayesian Optimization tasks. |
| Researcher Affiliation | Academia | Marshal Sinaga ELLIS Institute Finland, Aalto University EMAIL Julien Martinelli ELLIS Institute Finland, Aalto University EMAIL Samuel Kaski ELLIS Institute Finland, Aalto University, University of Manchester EMAIL |
| Pseudocode | Yes | Algorithm 1 Expert-guided robust mean prior algorithm |
| Open Source Code | Yes | Our implementation is available at https://github.com/Marshal Arijona/RCa GP. |
| Open Datasets | Yes | We evaluate GP regression on four UCI datasets with asymmetric outliers (details in Section H.1) ... The dataset can be found at https://www.cs.toronto.edu/~delve/data/boston/ boston Detail.html. Hartmann 6D. The widely used Hartmann benchmark function [39]. |
| Dataset Splits | Yes | Average test set mean absolute error, negative log-likelihood, and clock-time (in seconds), with 1 std, for 20 train-test splits. (Table 1 Caption) We initialize the optimization process for all baselines with 250 data points sampled uniformly across the search space. (Section 7) Proportion of test set 0.2 (Table S4) |
| Hardware Specification | Yes | For the UCI regression experiments, all models including our proposed method and the baselines were executed on a compute cluster consisting of two machines: one equipped with dual AMD EPYC 7713 processors (64 cores each, 2.0 GHz), and another with dual Intel Xeon Gold 6148 processors (20 cores each, 2.4 GHz). Bayesian Optimization experiments were conducted on a cluster with four NVIDIA V100 GPUs, each with 32 GB of memory. |
| Software Dependencies | No | All baselines employ a Matérn-5/2 kernel and are implemented in GPy Torch [10]. |
| Experiment Setup | Yes | Table S4: Hyperparameter settings used for RCa GP. Most of them remain consistent across all the tasks. Optimizer ADAM Learning rate 0.01 Minibatch size n-data Number of iterations for optimizing ELBO 50 ϵ 0.2 n-inducing points (SVGP) 100 Projection-dim (RCa GP) 5 Mean-prior m (RCa GP and RCSVGP) 1 n Pn j yj Mean-prior m (SVGP) 0 β 1.0 σ2 corr. 1.0 |