ACL 2026 paper: EEG-to-text benchmark COFETT achieves teacher-forcing-free decoding

A team led by Zihan Zhang (Harbin Institute of Technology, Shanghai Innovation Institute) has published a landmark paper at ACL 2026 tackling the long-standing debate over whether electroencephalography (EEG) can reliably decode text without artificial evaluation tricks. The researchers found that previous EEG-to-text (EEG2Text) models only appeared to work because they relied on teacher forcing — a method that feeds ground-truth tokens during inference, hiding the model's inability to generate meaningful sequences independently. Using a neuropsychology-inspired paradigm, they show that earlier benchmarks neglected the inherent instability of EEG signals, leading to inflated claims and sparking skepticism about whether EEG even contains decodable linguistic content.

To address this, the team assembled the Corpus Of EEG-To-Text (COFETT) using a 128-channel high-density EEG cap — significantly denser than typical consumer headsets. COFETT enables genuine teacher-forcing-free evaluation and distinguishes model performance far better than existing datasets. In comparisons, it achieves state-of-the-art discriminative power, providing a rigorous testbed for non-invasive brain-to-text research. The benchmark is fully open-sourced, giving researchers a standardized way to measure real-world feasibility. This work moves EEG2Text from academic trickery toward practical applications, potentially restoring communication for people with severe paralysis without requiring invasive brain surgery.