Large language models are known to predict neural responses when people process natural language, but exactly which linguistic properties produce this brain alignment has remained opaque. A new preprint on arXiv introduces SCRIBE, an interpretable semantic representation framework, to decompose that alignment into understandable components.
The abstract describes the motivation: LLMs predict neural responses during naturalistic language comprehension, yet the linguistic information underlying the alignment is difficult to interpret. SCRIBE is presented as the tool to address this, though the abstract is truncated and does not yet specify the full method or results.
Because only the abstract was available for this digest, the details of how SCRIBE works and what it found are not yet reported here. The paper's contribution, as stated, is to make brain–LLM alignment interpretable rather than treating it as a black-box correlation.