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..
NeurIPT: Foundation Model for Neural Interfaces
Authors: Zitao Fang, Chenxuan Li, Hongting Zhou, Shuyang Yu, Guodong DU, Ashwaq Qasem, Yang Lu, Jing Li, Junsong Zhang, Sim Kuan Goh
NeurIPS 2025 | Venue PDF | LLM Run Details
| Reproducibility Variable | Result | LLM Response |
|---|---|---|
| Research Type | Experimental | Empirical evaluations across eight downstream BCI datasets, via finetuning, demonstrated Neur IPT consistently achieved state-of-the-art performance, highlighting its broad applicability and robust generalization. |
| Researcher Affiliation | Academia | Zitao Fang1 Chenxuan Li1 Hongting Zhou1 Shuyang Yu2 Guodong Du3 Ashwaq Qasem1 Yang Lu4 Jing Li5 Junsong Zhang4 Sim Kuan Goh1 1Xiamen University Malaysia 2Columbia University 3The Hong Kong Polytechnic University 4Xiamen University 5Harbin Institute of Technology (Shenzhen) EMAIL EMAIL EMAIL |
| Pseudocode | No | The paper describes methods textually and mathematically in sections 2.1 and 2.2, but does not include any explicitly labeled pseudocode or algorithm blocks. |
| Open Source Code | Yes | Our project is available at this https URL. |
| Open Datasets | Yes | NEURIPT is pre-trained using more than 2,000 hours of data collected from public datasets, with the eight downstream datasets explicitly excluded. For more details on pre-training datasets, please refer to Appendix E.1. |
| Dataset Splits | Yes | Experimental Configuration Following CBra Mod [21], subject 1 to 28 are set to training set, subject 29 to 32 are set to validation set and subject 33 to 36 are set to test set. ... The training subjects were further divided into training and validation subsets in an 80% / 20% ratio. |
| Hardware Specification | Yes | The pre-training process was conducted for approximately 400K steps, employing an effective batch size of 480 and bfloat16 mixed-precision training on eight NVIDIA Ge Force RTX 4090 GPUs. |
| Software Dependencies | Yes | We implemented NEURIPT using Python 3.9.19 and Py Torch 2.3.0 with CUDA 12.1 and cu DNN 8902. |
| Experiment Setup | Yes | Pre-training stage was trained using the Adam W optimizer combined with the One Cycle learning rate strategy [29] (upper learning rate 3e-4, divided factor 25, final divided factor 1e4, and cosine annealing strategy). The pre-training process was conducted for approximately 400K steps, employing an effective batch size of 480 and bfloat16 mixed-precision training on eight NVIDIA Ge Force RTX 4090 GPUs. For more details on implementation settings and hyperparameters, please refer to Appendix D and Table 13. |