EEG-Captioner ties generated descriptions to measurable signals

An October 22 preprint introducing EEG-Captioner pairs EEG segments with questions about measurable signal properties and feature-grounded captions. The dataset contains 160,781 segments covering spectral, spatial, temporal and lateral characteristics.

The model combines waveform and frequency-related representations and trains with both question answering and caption supervision. The authors report improved caption factual consistency compared with caption-only adaptation.

The target is accurate description of a recording, not reading a person's private thoughts. Grounding the output in properties that can be checked gives researchers a more useful test than whether a generated sentence merely sounds like a plausible EEG report.