BCIwiki (bciwiki.com) — A contextual prior can lift the accuracy of neural-to-language decoding while making some of the errors it leaves behind more confident, according to a preprint uploaded to arXiv on October 6, 2026. The work, which has not been peer reviewed, examines how a prior reshapes candidate scores in speech retrieval on the MEG-MASC and MOUS datasets, using local decoding scores, a prior combined through additive shallow fusion, and the fused top-two margin as the confidence measure. Among initially incorrect predictions, the authors report a confidence-ordering reversal: a larger margin makes a repair more likely when the correct candidate starts near the top of the local ranking, but less likely when it starts lower.
On MEG-MASC, pooled correctness AUROC is 0.87, but the AUROC separating repairs from residual errors falls from 0.70 at initial ranks 2-3 to 0.39 at ranks 21-50. Errors starting beyond rank 20 account for 46.6% of all post-fusion errors. The authors trace this to the score arithmetic: a repair must first close the correct candidate's initial deficit, which limits its final margin, while a residual error can build a large margin between two incorrect candidates. A causal intervention that changes only the fusion weight moves the reversal to deeper ranks as predicted; under a word-level language model prior, it keeps moving after the accuracy gain peaks.
Reading local and prior scores separately improves selective decoding, with the decoder answering on 74.5% of windows instead of 56.7% while 92% of output sets still contain the correct candidate, the authors say. They argue that confidence after contextual fusion should retain the local and contextual evidence behind each prediction rather than the fused scores alone. The preprint runs 28 pages, with 4 figures and 18 tables.