Do We Need Domain-Specific Time-Series Models? Insights from EEG Classification Benchmarks

Time-series foundation models (TS FMs) are becoming increasingly popular, alongside growing efforts to develop domain-specific foundation models such as EEG foundation models (EEG FMs). However, whether generic TS FMs reliably transfer to EEG remains unclear. This paper presents a broader investigation into the transferability of two state-of-the-art TS FMs—Mantis and MOMENT—to EEG classification. We evaluate both models under linear probing and fine-tuning across four public datasets: TUEV, FACED, BCI-IV-2A, and Error. We adopt cross-subject evaluation for TUEV and FACED, and compare population-level decoding with leave-one-out fine-tuning on BCI-IV-2A and Error. Our results show that fine-tuned TS FMs can outperform some EEG-specific baselines, but linear probing remains weak, and performance varies across downstream tasks. These findings extend prior benchmarks and highlight both the promise and limits of generic TS FMs for EEG decoding.