Human hippocampus uses similar semantic coding as large language models

Summary

Researchers recorded hippocampal neurons during narrative speech. They found that activity in the neuronal population encodes word meanings in a distributed and context-sensitive manner. The geometric structure of this activity, specifically the distance relationships, largely overlaps with the embedding spaces of large language models (LLMs). The strongest alignment is achieved with GPT-2 vectors.

Why it matters

This finding represents one of the most direct empirical bridges drawn between the coding scheme of a specific brain region and the representational geometry of transformer-based language models. It adds biological plausibility to the idea that LLMs capture something real about how humans process meaning. It repositions the hippocampus not merely as a memory store, but as an active semantic encoder during real-time language comprehension. This has significant implications for both cognitive theory and the interpretability of LLMs.

Highlights

• Neuronal population distance tracks semantic distance. This relationship remains parallel to LLM embedding geometry even after controlling for phoneme and grammar variables.
• Alignment peaks with GPT-2 rather than newer, larger models. Response variability parallels polysemy scores derived from the LLM. This indicates context-dependent coding.
• For semantically similar words in contextual embedding models, the neuronal-semantic distance correlation moves in the *inverse* direction. The authors attribute this to the noise-reduction benefits of contrastive coding.

Critical take

While the alignment achieved with GPT-2 is striking, it carries a caveat. GPT-2 is a 2019 model with 1.5 billion parameters. The 'closest match' expression may stem from the simplicity of the task, i.e., passive listening to narratives. This does not necessarily imply a deep computational isomorphism with the latest generation of LLMs. The observed inverse correlation for similar words is explained in a single sentence as the 'noise-reduction benefit of contrastive coding.' This explanation lacks a formal model or control to rule out simpler confounding factors, such as response saturation or recording noise in the hippocampal population.