Hippocampal Population Activity Tracks Semantic Distance in Narrative Speech, Aligning Closest with GPT-2 Embeddings

Researchers at Baylor, Rice, Berkeley and Texas Children's Hospital implanted microwires in the hippocampus of patients undergoing epilepsy monitoring and captured single-unit responses while participants heard continuous narratives. They fit encoding models that predicted firing rates from contextual word embeddings and found robust semantic tuning distributed across neurons rather than localized to single cells or fixed categories. Distance between neural population vectors scaled with semantic distance derived from the same stories processed by language models. The hippocampus is classically linked to episodic memory, yet these data indicate it also maintains dynamic semantic maps during comprehension. This extends prior work on concept cells by showing that meaning is represented through contrastive population codes that reduce noise for polysemous terms, a pattern the authors attribute to mechanisms also exploited in LLM training. The study therefore bridges memory and language literatures that have often been studied separately. Future experiments must test whether these hippocampal codes are necessary for comprehension by combining the same recording approach with targeted disruption or closed-loop stimulation. Larger cohorts across age and language backgrounds are required to determine if the GPT-2 alignment generalizes or reflects corpus-specific statistics. Longitudinal recordings could clarify whether semantic tuning strengthens with repeated exposure to narrative contexts.