GACET: A Graph-Aware Cross-Domain EEG Transformer with the CD-MWL Benchmark

ICLR 2026 Conference Submission12889 Authors

Keywords: Neuroscience, Computational Neuroscience, BCI, EEG

TL;DR: This paper introduces CD-MWL, the largest publicly available multi-day EEG mental-workload dataset, and GACET, a graph-aware transformer that achieves state-of-the-art cross-days brain-state decoding.

Abstract: Robust generalization of brain state decoding across days remains a grand challenge for brain-computer interfaces (BCIs), throttling the real-world deployment of applications like mental-workload (MWL) estimation. This long-standing problem has been difficult to address, largely due to the scarcity of public corpora suitable for rigorous long-term evaluation. To establish the first robust benchmark for this challenge, we introduce CD-MWL, the longest public MWL dataset to date: 42 hours of EEG from 14 participants over three days. Building on this benchmark, we argue that true generalization requires models that are not only high-performing but also neuroscientifically plausible. We therefore propose GACET, a Graph-Aware Cross-domain EEG Transformer that achieves superior generalization by dynamically learning the brain's underlying functional connectivity in a neuroscientifically plausible manner. Through topology-aware message passing on this learned graph, GACET not only delivers significant performance gains over all state-of-the-art methods but also provides a transparent window into its decision-making process. Crucially, we demonstrate that the model's learned connectivity patterns align with established neuroscience, establishing a powerful, evidence-based link between its interpretability and robust performance. All data, code, training logs and a one-command reproduction script are publicly available at https://anonymous.4open.science/r/GACET-B6F8.

Primary Area: applications to neuroscience & cognitive science

Submission Number: 12889