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..

CAMO: Convergence-Aware Multi-Fidelity Bayesian Optimization

Authors: WEI XING, Zhenjie Lu, Akeel Shah

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We assess CAMO on synthetic benchmarks, including continuous and discrete MFBO tasks, as well as real-world engineering design tasks. We compare the results to those of: (1) BOCA with a standard GP [3], (2) Fabolas [16], and (3) SMAC3 [27]. On discrete fidelity tasks, we compare the results to: (1) AR [28], (2) Res GP [29], and (3) a GP [6], and discrete MFBO baselines (1) MF-UCB [30], (2) MF-EI [10], and (3) cf KG [17].
Researcher Affiliation Academia Wei W. Xing University of Sheffield EMAIL SUSTeh & Shenzhen University EMAIL Akeel A. Shah Chongqing University EMAIL
Pseudocode No The paper describes methods and processes in narrative text and mathematical formulations but does not include any explicitly labeled pseudocode or algorithm blocks with structured steps.
Open Source Code Yes Code is available at https://github.com/Ice Lab-X/CAMO.
Open Datasets Yes We consider: (1) three canonical continuous-fidelity tasks [31], the Park, Currin and Branin functions; and (2) three further synthetic continuous-fidelity tasks [32], the nonlinear sin, Forrester, and Bohachevsky functions. In the latter 3 we set f(x, t) = (1 w(t))flow(x) + w(t)fhigh(x) with w(t) = ln(9t + 1). All functions are defined in Appendix I.
Dataset Splits Yes In each case, 10 low-fidelity and 4 high-fidelity designs were randomly selected to form the initial training set.
Hardware Specification Yes All experiments were performed on a workstation with an AMD 7800x CPU, Nvidia RTX4080 GPU, and 32GB RAM.
Software Dependencies No Except for DNN-MFBO, Fabolas, and SMAC3 (original implementations and default settings), all methods were implemented in Pytorch. While Pytorch is mentioned, no specific version number is provided for it or any other software.
Experiment Setup Yes Each model is updated for 200 steps using an Adam optimizer with a learning rate of 0.01 to ensure model convergence. In each case, 10 low-fidelity and 4 high-fidelity designs were randomly selected to form the initial training set. We repeated the experiments 20 times with random seeds and report the mean values. The query costs were set to c(t) = 10t.