Published October 1, 2026, in PLOS ONE, this full-text study tested how multiple language models could support a P300 speller, a brain-computer interface that translates EEG responses into text. The analysis used existing EEG data from 78 healthy adults aged 20 to 35.
How performance changed With within-subject classifier training, GPT-2 plus Dijkstra-based word completion reached 75.3 bits per minute, about 45% above the random-flashing baseline and within 4% of the study's word-performance bound. Across-subject analyses increased from about 34 to about 60 bits per minute, with improvements appearing across several high-performing model families.
Where the bottleneck may be moving Once word prediction became sufficiently accurate, larger language models produced diminishing returns. The authors argue that neural decoding and signal acquisition now limit this framework more than language-model scale. Across-subject training may also reduce calibration demands.
The clinical boundary This was an offline simulation, not a live communication trial. Participants did not have ALS or another severe motor impairment. Real-time accuracy, fatigue, usability and communication value in intended users still require prospective testing.
Source and disclosure Full text checked. The authors reported no specific funding and declared no competing interests. Prepared with AI assistance and source-checked October 2, 2026.