Do Large Language Models Mirror Human Brain Language Processing? Google’s Groundbreaking Findings

Overview

Google Research recorded cortical electrocorticography (ECoG) from participants engaged in natural, open‑ended conversations while they spoke and listened. Over 100 hours of neural data were collected. The speech‑to‑text large language model Whisper was run on the same audio, and three levels of internal representations were extracted for each word: low‑level acoustic embeddings, mid‑level phonetic embeddings, and high‑level lexical (word‑level) embeddings.

Methodology

For each word the researchers aligned the model embeddings with the ECoG signal in a time window from –2 s to +2 s relative to word onset. A linear encoding model (ridge regression) was trained on a subset of the data to map each embedding dimension to neural activity across electrodes. Model performance was evaluated on held‑out dialogue segments by computing the correlation between predicted and observed signals.

Key Findings

During speech perception, acoustic embeddings best predicted activity in the superior temporal gyrus (STG) shortly after the word was heard, whereas lexical embeddings peaked several hundred milliseconds later in Broca’s area (inferior frontal gyrus, IFG), reflecting semantic processing. During speech production the temporal order reversed: lexical embeddings predicted IFG activity ~500 ms before articulation, followed by acoustic embeddings predicting motor‑cortex activity as the phonetic plan was executed. The full temporal profile showed a sequence of planning in language areas, execution in motor areas, and auditory monitoring after speech.

Quantitatively, the encoding model achieved peak Pearson correlations of ~0.3–0.4 for the best electrodes, and the latency of the peak prediction matched known neurophysiological timings (STG ≈ 100 ms, IFG ≈ 300–500 ms).

Implications

The results demonstrate that embeddings from a speech‑to‑text LLM trained solely for transcription capture statistical regularities that align with human cortical dynamics during natural language processing. This provides a concrete computational framework for linking artificial neural network representations to brain activity, and suggests that large models can serve as proxies for studying the neural basis of language.

Related Work

Nature Neuroscience (2022) showed that human listeners predict upcoming words and that prediction confidence modulates post‑onset surprise.

Nature Communications (2024) reported geometric similarity between LLM embedding spaces and brain‑derived representational spaces.

Future Directions

The team plans to design biologically‑inspired neural architectures that exploit the observed brain‑model parallels to improve real‑world information processing and to test whether such models can better predict neural responses in more complex linguistic contexts.

References

https://research.google/blog/deciphering-language-processing-in-the-human-brain-through-llm-representations/

https://www.nature.com/articles/s41562-025-02105-9

https://x.com/GoogleAI/status/1903149951166902316

https://x.com/rohanpaul_ai/status/1903373048260284868

Code example

收
藏
，
分
享
、
在
看
，
给
个
三
连
击呗！