MIT researchers discovered that contemporary large language models (LLMs) can process varied data types comparably to the human brain. They observed that LLMs feature a central processing strategy similar to the brain's 'semantic hub,' which merges information from different modalities, like visual and tactile. By utilizing the primary language of these models, they could manipulate outputs even when processing inputs in different languages. This insight could lead to advancements in training LLMs for more effective handling of diverse data.
"LLMs are big black boxes. They have achieved very impressive performance, but we have very little knowledge about their internal working mechanisms. I hope this can be an early step to better understand how they work so we can improve upon them and better control them when needed."
The researchers demonstrate that they can intervene in a model's semantic hub by using text in the model's dominant language to change its outputs, even when the model is processing data in other languages.
Collection
[
|
...
]