Chinese scientists have discovered that multimodal large language models (LLMs) can spontaneously develop object concept representations mirroring human cognition, according to a groundbreaking study published in Nature Machine Intelligence. This breakthrough offers new pathways for understanding artificial intelligence's cognitive potential and designing systems with human-like reasoning.
Led by researcher He Huiguang of the Chinese Academy of Sciences' Institute of Automation, the study reveals how AI models process multidimensional concepts – such as a dog's physical traits, cultural symbolism, or emotional associations – much like the human brain. By integrating computational modeling, behavioral experiments, and neuroimaging, researchers mapped how LLMs correlate with neural activity patterns in category-selective brain regions.
"This isn't just about recognizing shapes or colors," explained He. "It's about bridging sensory input and abstract meaning – a cornerstone of human intelligence." The team found that 66 dimensions extracted from LLM behavioral data strongly aligned with human neural responses, with multimodal models outperforming text-only counterparts in consistency tests.
While humans blend visual and semantic cues when conceptualizing objects, the study notes LLMs prioritize semantic labels and abstract relationships. This divergence highlights both the promise and limitations of current AI systems in replicating biological cognition.
The findings hold implications for global tech developers, cognitive scientists, and policymakers navigating AI's evolution. For investors and businesses, it signals potential advancements in adaptive AI for healthcare, education, and cross-cultural applications across Asian markets.
