BCIwiki (bciwiki.com) — Language models can raise P300 speller typing speed by up to ~45% with within-subject training and ~75% with across-subject training over conventional approaches, according to simulations on EEG data from 78 subjects reported in PLOS ONE on October 1, 2026 (PMID 42821610; DOI 10.1371/journal.pone.0349281).
The P300 speller brain-computer interface (BCI) helps people with amyotrophic lateral sclerosis (ALS) and similar conditions communicate by translating EEG responses into text, but slow typing and the need for subject-specific calibration limit its use. Earlier work showed that large language models such as GPT-2 can improve performance by predicting words; whether the gains were model-specific, and where the ceiling lies, had not been established.
The researchers built a multi-model theoretical analysis framework, evaluated a wide range of language models within a unified decoding framework, and introduced an idealized language model to set an upper bound on achievable performance. They also added cross-subject classifier training to reduce calibration demands and test generalization.
Several high-performing models, despite architectural differences, came within 5% of the theoretical bound, which the authors say indicates diminishing returns from further model scaling. The gains held for both within-subject and across-subject classifiers. The authors conclude that the main bottleneck is shifting from language modeling to neural signal decoding. The work is based on offline simulations and was not tested in live use by patients.
Authors are affiliated with the University of Illinois Urbana-Champaign, the University of Washington and the University of California, Los Angeles, among others.