Chinese researchers have shown that models of artificial intelligence (AI) of large language and multimodal capacity (MLLM) can form mental representations o
The finding, carried out by the Institute of Automation of the Academy of Sciences of China in collaboration with the Center for Excellence in Intelligent Sciences and Technology, was published in the journal Nature Machine Intelligence.
The study combines behavior and neuroimaging analysis for Explore if these models really “understand” objects, beyond recognizing their forms.
Through a task known as “the intruder” – consistent in identifying which of three objects is the most different from the other two – the researchers gathered 4.7 million decisions of AI models when comparing trios of 1,854 objects of the natural world.
The analysis allowed to identify 66 stable and highly interpretable conceptual dimensions, that showed a semantic organization similar to human cognition.
In addition to behavioral analysis, the study included comparisons with human neuroimaging data to examine whether the internal representations of the models aligned with brain activity.
A remarkable correspondence was detected in regions such as the facial fusiform area, the retrosonial cortex or the parahipocampal area, all involved in the processing of visual and conceptual categories.
Although models do not identically reproduce human thought, their internal structures reflect key aspects of conceptual knowledgewhich reinforces its potential as a basis for building AI systems with more human cognitive architecture.
Recent Chinese language models such as Deepseek, Doubao or Yuanbao have caused a great media and economical stir in recent months for being able to rival in their abilities with their US competitors at a lower price.
However, some experts in the sector have been skeptical of the sudden irruption of so many similar services simultaneously in China and there are doubts about the real global expansion capacity of the Chinese ‘chatbots’ for the censorship exerted by the authorities.