A Leap Forward in Understanding AI Language Processing

(EPFL, May 13, 2025)

Unlocking artificial intelligence's language skills hinges on less than 1% of its neurons. Badr AlKhamissi and Prof. Martin Schrimpf from EPFL led a team that found key neurons in Large Language Models (LLMs) crucial for language processing, similar to the human brain's system. Turning off these units greatly reduces the model's linguistic abilities, enhancing our understanding of AI and linking artificial and human cognition, and opening new paths for treating neurological disorders. The team studied 18 LLMs using a neuroscience-inspired technique to identify these vital neurons. They compared unit activity during coherent sentence processing versus random words, uncovering so-called "language-selective units." Deactivating these units led to a significant drop in language performance, mirroring neuroscience methods of assessing brain area importance. This discovery could allow for more focused AI development and could revolutionize disease diagnosis and treatment, highlighting EPFL's interdisciplinary approach that combines computer science, life sciences, and engineering.