Large Language Models Mimic Human Brains by Reasoning Across Diverse Data Types
Large language models (LLMs) are rapidly evolving, transcending their initial text-processing capabilities to perform diverse tasks across various data types. From understanding multiple languages to generating computer code, solving math problems, and interpreting images and audio, LLMs are becoming increasingly versatile. Now, MIT researchers have uncovered that the way these models process assorted data shares fascinating similarities with the human brain.
Researchers at MIT delved into the inner workings of LLMs to understand how they handle such a wide array of data. Their findings suggest that LLMs utilize mechanisms akin to the human brain’s “semantic hub,” located in the anterior temporal lobe, which integrates semantic information from visual, auditory, and tactile inputs. This hub is connected to modality-specific “spokes” that route information. Similarly, LLMs abstractly process data from diverse modalities in a centralized manner. For example, an LLM with English as its dominant language uses English as a central medium to process inputs in other languages or reason about arithmetic and computer code.
The researchers demonstrated their ability to influence a model’s semantic hub by using text in its dominant language to alter outputs, even when the model processes data in other languages. According to Zhaofeng Wu, an EECS graduate student and lead author of the research paper, these findings could pave the way for training future LLMs with enhanced abilities to handle 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,” says Wu.
The study builds upon prior research that hinted at English-centric LLMs using English to perform reasoning processes on various languages. Wu and his collaborators expanded this concept by conducting an in-depth study into the mechanisms LLMs use to process diverse data.
An LLM dissects input text into tokens (words or sub-words) and assigns a representation to each token, enabling it to explore relationships and generate subsequent words. For images or audio, tokens correspond to specific regions or sections. The researchers found that the model’s initial layers process data in its specific language or modality, akin to the brain’s modality-specific spokes. Then, the LLM converts tokens into modality-agnostic representations, similar to how the brain’s semantic hub integrates diverse information. This means that the model assigns similar representations to inputs with similar meanings, regardless of their data type.
To test their hypothesis, the researchers passed pairs of sentences with identical meanings in different languages through the model and measured the similarity of the model’s representations. They also fed an English-dominant model text in a different language, like Chinese, and measured the similarity of its internal representation to English versus Chinese. These experiments consistently showed that the model’s representations were similar for sentences with similar meanings and that the tokens processed in its internal layers were more like English-centric tokens than the input data type. “A lot of these input data types seem extremely different from language, so we were very surprised that we can probe out English-tokens when the model processes, for example, mathematic or coding expressions,” Wu noted.
The researchers believe that LLMs learn this semantic hub strategy during training because it is an efficient way to process varied data. This allows knowledge to be shared across languages, avoiding the need to duplicate information. They also found that intervening in the model’s internal layers with English text while processing other languages predictably changed the model outputs. This phenomenon could be leveraged to encourage information sharing across diverse data types, potentially boosting efficiency. However, it’s also important to consider concepts that are not translatable across languages, such as culturally specific knowledge. Future research could explore how to balance maximum sharing with language-specific processing mechanisms.
These insights could also improve multilingual models. Understanding an LLM’s semantic hub could help prevent language interference, where an English-dominant model loses accuracy in English after learning another language.
Mor Geva Pipek, an assistant professor in the School of Computer Science at Tel Aviv University, who was not involved with this work, notes, “Understanding how language models process inputs across languages and modalities is a key question in artificial intelligence. This paper makes an interesting connection to neuroscience and shows that the proposed ‘semantic hub hypothesis’ holds in modern language models, where semantically similar representations of different data types are created in the model’s intermediate layers… The hypothesis and experiments nicely tie and extend findings from previous works and could be influential for future research on creating better multimodal models and studying links between them and brain function and cognition in humans.”
