When studies use contiguous train/test splits and control for sentence length, word position and word rate, how much of the LLM-to-brain fit remains — and is what remains contextual, syntactic, or just word meaning?
This decides whether Dhi's 'Established' correlations stay established.
LLM features predict language-network activity well, but with stricter controls (contiguous splits; sentence length, position, word rate) much of the fit shrinks. What survives seems driven mainly by word meaning, not syntax.
A neutral summary kept current by super users, not a verdict.
A substantial contextual component survives controls.
Schrimpf 2021; Goldstein 2024 (implied)
Most alignment is explained by length, position, word rate and static word meaning.
Feghhi et al. 2024; NeurIPS 2026
No decisions yet. When this room changes a flow or an evidence label, it is recorded here permanently.
Dhi: Room opened. The outcome of this Unknown can directly re-grade correlation rows 3, 8 and 10 in Correlate. Please attach analyses with their split strategy and control variables stated.