Do LLMs Think Like Humans? New MIT Study Suggests They Might

February 20th 2025

A new MIT study reveals that large language models (LLMs) may process diverse data similarly to the human brain. Neuroscientists have long believed that the human brain has a “semantic hub” in the anterior temporal lobe, where information from different sensory modalities converges for abstract reasoning. Researchers found LLMs function similarly, using a central, generalized process to interpret varied inputs, such as languages, images, audio, computer code, and math.

MIT's study shows that LLMs convert diverse inputs into modality-agnostic representations, much like how humans integrate meaning across senses. For example, when an English-dominant LLM processes a Chinese sentence, it internally “thinks” in English before generating a Chinese response. This central reasoning hub allows LLMs to handle complex cross-modal tasks without duplicating knowledge for each modality.

The study also revealed that intervening in the model's internal layers using English text could predictably alter outputs in other languages. This could lead to more efficient AI models by encouraging shared knowledge while preserving language-specific nuances. While this strategy boosts efficiency, it raises questions about culturally specific knowledge that cannot always be translated across languages or data types.

The findings offer significant implications for future LLM development. Scientists could use the insights to create better multimodal models, improve multilingual AI, and enhance cross-modal reasoning. However, there are concerns about whether this uniform processing might overlook unique cultural contexts or lead to language interference. Researchers now aim to strike a balance between shared knowledge and language-specific processing, ensuring more adaptable and contextually aware AI systems.

This breakthrough not only sheds light on how LLMs think but also suggests AI might be mirroring cognitive structures found in the human brain. As AI continues advancing, understanding how it processes information could help build smarter, safer, and more culturally sensitive models.

Source: MIT News