Surprisal contributes little beyond contextual embeddings in high-gamma ECoG encoding

Adding GPT-2 XL word surprisal to ridge encoding models that already include GPT-2 XL contextual embeddings leaves held-out high-gamma ECoG prediction essentially unchanged in natural-speech ECoG.

• Why it matters: It is unclear whether surprisal provides independent predictive information for neural responses beyond what contextual embeddings capture.
• What they did: Using public ECoG recordings during natural speech, the study fit word-aligned ridge encoding models with baseline stimulus features plus GPT-2 XL embeddings and then added GPT-2 XL surprisal, evaluating held-out correlation.
• The result: The surprisal increment stayed within an equivalence margin across alternative lags, residualized surprisal, and frequency-band targets, but did not consistently meet the same criterion under banded ridge regularization or within labeled temporal subsets.

AI summary of the abstract · may contain errors · not medical advice · read the paper.