EEG-FM-Bench: A Comprehensive Benchmark for the Systematic Evaluation and Diagnostic Analyses of EEG Foundation Models

The article introduces EEG-FM-Bench, a standardized benchmark for evaluating electroencephalography foundation models (EEG-FMs) across 14 datasets and 10 paradigms, addressing the inconsistencies in current evaluation protocols. Key findings highlight the role of multi-task learning as a regularizer, the limitations of pre-training due to gradient conflicts, and the importance of objective alignment and EEG-specific design in transfer performance. This benchmark facilitates fair comparisons and reproducible analyses, enhancing the interpretability of EEG-FMs for researchers and practitioners in the field.