BCIwiki (bciwiki.com) — Adjusting how much a P300 speller trusts its language model on a trial-by-trial basis did not beat a tuned fixed weight in a reanalysis of 3,373 selections from 47 people with amyotrophic lateral sclerosis (ALS), according to a preprint posted to medRxiv on September 3, 2026. The study, listed under BIDMC, used the archived BigP3BCI dataset and has not been peer reviewed.
P300 spellers let users select characters by detecting the P300 response in EEG. Published fusion schemes fix the trust placed in a language-model prior regardless of how reliable each trial is. The researchers tested whether a reliability estimate could improve on that, tuning both a fixed weight and an adaptive policy out-of-fold in a matched search space across 22 evaluable language-model priors of up to 46.7B parameters. Two representative priors, GPT-2 and a classical 5-gram, received additional naive and mechanistic analyses.
No prior's 95% confidence interval favored adaptive fusion under the fair comparison, even though oracle headroom went unexploited at every scale. Under GPT-2, the naive comparison was significantly worse for adaptive fusion, and both representative priors converged to a degenerate or near-degenerate solution in the fair comparison. 3 further controllers also failed to turn the headroom into benefit. The fixed-fused posterior's output probability outperformed the best controller at flagging errors, with 2.8- to 3.8-fold enrichment.
The authors concluded that a tuned fixed weight is a difficult-to-beat default across the tested scale range and that reliability estimation gave no deployable adaptive advantage. They recommend validating adaptive weighting against a fairly tuned baseline across model families and scales.