A new study shows LLMs represent different data types based on their underlying meaning and reason about data in their dominant language.
A study by MIT researchers reveals that large language models (LLMs) process diverse data, including text, images, audio, and code, using a mechanism similar to the human brain's semantic hub. These models integrate information by relating inputs to their underlying meaning, often using their dominant language as a central medium. The research demonstrates that interventions in this central processing mechanism, using the dominant language, can influence model outputs. This finding has implications for improving LLM efficiency, enhancing multilingual models, and preventing language interference, where learning new languages can reduce accuracy in the primary language. The study highlights potential advancements in creating more versatile and effective language models.
