Search Abstracts | Symposia | Slide Sessions | Poster Sessions
Are predictable words processed earlier? Faster? Evidence from naturalistic MEG encoding
Poster Session A, Wednesday, September 30, 11:00 am - 1:00 pm, Wangari Maathai
Theo Desbordes1, Nicolas Piron1, Sophie Schwartz1,2, Nina Kazanina1; 1University of Geneva, 2Swiss Center for Affective Sciences
Language comprehension unfolds as a rapid cascade of computational operations: from acoustic and phonological analysis, through lexical access, syntactic structure building, and semantic interpretation, to discourse-level integration. Prediction is increasingly viewed as a central part of language processing (Kuperberg & Jaeger, 2016, Lang. Cogn. Neurosci.; Heilbron et al., 2022, PNAS), but direct evidence that word predictability modulates the timing and/or speed of neural processes remains scarce. To address this gap, three subjects from Armeni et al. (2022, Sci. Data) listened to ~10 h of The Adventures of Sherlock Holmes during magnetoencephalography (MEG) recording. Per-word hidden states were extracted from multiple Large Language Models (LLMs) of the LLaMA family, under two regimes: a predictive regime in which the current word is excluded from the context window, so that the embedding reflects only what preceding text allows the model to anticipate, and a standard regime in which the current word participates in its own context, so that the model integrates the current word into its context representation. We examined the embeddings from each layer of the model, as layer depth in causal transformers is known to index a hierarchy of representations from surface to abstract. Time-resolved Ridge regression mapped embeddings to MEG sensor activity layer by layer and for each variant (standard or predictive) separately. First, we replicated recent results (Goldstein et al., 2025, Nat. Commun.; Raugel et al., 2025, NeurIPS) showing, under both regimes, an ordered relationship between layer depth and peak encoding latency: deeper layers align with later neural responses, as if successive computational steps in the model mapped onto successive stages of neural processing. Going further, we found that under the predictive regime the encoding response separates into two distinct latency windows, each traversing the full layer hierarchy but at different speeds. One peaks at or near the word onset; the other peaks ~500 ms post-onset and is markedly slower. Crucially, the temporal alignment between peak latency and layer depth in the first latency window was stronger for more predictable words, potentially reflecting pre-activated neural representations of upcoming semantic content. The late window falls within the classical lexical access and post-lexical integration window, and the strength of the temporal alignment there was stronger for less predictable words. This suggests that the same cascade of computational steps unfolds for both surprising and unsurprising words, but predictable words undergo the full LLM layer hierarchy earlier and faster. Consistent with recent results (Dou et al., 2025, PLOS Comput. Biol.), these findings challenge a common assumption of epoch-averaging designs in neurolinguistics: that neural responses to words are time-locked to stimulus onset. Ongoing work extends these results with i) perturbation-based variance partitioning across acoustic, lexical, syntactic, and semantic content, ii) source-space localization of the two latency windows, iii) transposition of these methods to an intracranial EEG naturalistic speech production dataset, where we expect a different temporal ordering of representations.
Topic Areas: Computational Approaches, Speech Perception