2026-07-16 00:00 Japan Papers Foundations & Methods Translated from EN

Surprisal Adds Little to ECoG Language Encoding

Summary A bioRxiv preprint finds that adding a surprisal predictor on top of GPT-2 contextual embeddings contributes almost nothing extra to predicting high-gamma ECoG responses during speech comprehension.

Why it matters for speech/language BCI decoding pipelines, the finding suggests contextual embeddings may already implicitly capture what surprisal contributes, offering a methodological cue to simplify feature engineering in future neural decoding models rather than stacking redundant predictors.

BCIwiki (bciwiki.com) — The preprint was posted to bioRxiv on July 16, 2026, by author Sakuma, T., affiliated with Soka University. The study examines two features commonly used to predict neural responses during language comprehension — surprisal and contextual embeddings, both derived from large language models — asking whether surprisal contributes information beyond what embeddings already capture. Using public ECoG recordings of natural speech, the author fit word-aligned ridge encoding models comparing baseline stimulus features, GPT-2 XL contextual embeddings, and GPT-2 XL surprisal in predicting high-gamma ECoG responses.

According to the researcher, adding the surprisal predictor to a model that already included embeddings left held-out correlation essentially unchanged at the center lag, and the near-zero effect held within a prespecified equivalence margin across alternative lags and sensitivity analyses. The researcher reports that surprisal does not appear to act as an independent predictor, but is better understood as a compressed readout of the same broader predictive state that embeddings already capture. This is a preprint posted on bioRxiv and has not yet been peer reviewed.

Compiled by BCIwiki from public sources

Sources · 1

biorxiv.org 2026-07-16

Soka University ECoG Decoding Algorithms Speech Decoding Japan

Read original ↗

Suggest a correction Revisions · none 中 / EN