Heterogeneous Neural Predictivity from Language Models During Naturalistic Comprehension

The paper's real contribution isn't that language models correlate with brain signals (known), but that it isolates linguistic structure encoding from confounds that typically inflate predictive claims. By controlling for temporal misalignment, acoustic artifacts, and model capacity, the authors establish a floor for what counts as genuine linguistic alignment versus statistical noise.

This connects directly to the mechanistic interpretability thread running through recent coverage. The emotion vectors work (Apertus and Gemma diverging in layer-wise patterns) and the framing-sensitivity instability paper both examine how internal representations encode structure. This study extends that lens to the brain itself, asking whether transformer geometry reflects biological necessity or architectural choice. Unlike the multilingual safety audits (RedVox, SamaVaani) which expose deployment gaps, this work operates at the level of representation alignment, similar to how GAVEL debugs multimodal hallucinations by localizing errors rather than just measuring accuracy.

If follow-up work shows that fine-tuned or instruction-aligned models predict neural activity differently than frozen base models, that would signal whether alignment procedures degrade or enhance biological correspondence. Conversely, if predictive power remains stable across model families and scales, it suggests linguistic structure encodes in transformers somewhat inevitably.