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..

BrainMoE: Cognition Joint Embedding via Mixture-of-Expert Towards Robust Brain Foundation Model

Authors: Ziquan Wei, Tingting Dan, Tianlong Chen, Guorong Wu

NeurIPS 2025 | Venue PDF | LLM Run Details

Reproducibility Variable Result LLM Response
Research Type Experimental We evaluate the proposed Brain Mo E on 3 pre-training datasets, including UK Biobank (UKB), HCP Aging (HCPA), and HCP Young Adult (HCPYA), and 7 downstream datasets, including ADNI, ABIDE, PPMI, Taowu, SZ, HCPA, and HCPYA.
Researcher Affiliation Academia Departments of Computer Science and Psychiatry University of North Carolina at Chapel Hill Chapel Hill, NC 27599 EMAIL;EMAIL
Pseudocode No The paper includes diagrams illustrating the framework (Figure 3) and descriptive text for processes but does not present any structured pseudocode or algorithm blocks.
Open Source Code Yes The codes and data split settings can be acquired via this code repository7. 7https://github.com/Chrisa142857/brain_moe
Open Datasets Yes Public data is accessible via internet (UKB2, HCPA3, HCPYA4, ADNI5. PPMI, ABIDE, and Taowu can be found here6). The licenses to obtain those data can also be accessed on the websites. 2https://www.ukbiobank.ac.uk/ 3https://www.humanconnectome.org/ 4https://www.humanconnectome.org/study/hcp-young-adult/overview 5https://adni.loni.usc.edu/ 6https://auckland.figshare.com/articles/dataset/Neur IPS_2022_Datasets/21397377
Dataset Splits Yes Following previous works, our experiments are conducted with subject-level cross-validation (CV). The average score and the standard deviation are both listed. To make our results comparable with previous papers, HCPA, HCPYA, and ADNI use a 5-fold CV as same as [9, 31], while others use 10-fold as same as [33].
Hardware Specification Yes The experiments are done on a Linux system with one NVIDIA RTX 6000 Ada.
Software Dependencies No The paper mentions "FSL software [16]" but does not specify a version number for it or any other software dependency.
Experiment Setup Yes Batch size and learning rate are set as 128 and 1e-4, respectively. The maximum epoch is set as 200 and Chid 2048. Training will be early stopped if accuracy keeps dropping in 50 epochs.