AI understands concrete, but fails abstract: study proposes teaching machines like the human brain

The paradox of size in artificial intelligence models

A surprising discovery from Zhejiang University is challenging fundamental assumptions about how we train artificial intelligence. Chinese researchers have identified that, although increasing the number of parameters in large language models significantly improves the recognition of concrete concepts, this same expansion weakens the ability of systems to understand abstract concepts.

The study exposes a critical limitation in the current approach of "the bigger, the better" that dominates the development of AI. While industry celebrates increasingly robust models, research points to a growing gap between how machines and humans organize knowledge.

How the human brain processes abstract artificial intelligence

The key to understanding this difference lies in the way the human brain processes information. Human neurons do not treat all concepts in the same way: there is a sophisticated hierarchy that balances concrete and abstract organically.

When we think of chair, we activate neural networks different from when we consider justice or freedom. The brain has developed, over millions of years of evolution, specialized structures to deal with different types of abstractions. Current AI models, on the other hand, apply relatively uniform processing to all types of concepts.

The revolutionary proposal: brain signals to teach abstract artificial intelligence

The University of Zhejiang team proposes an innovative approach: using real brain signals as a guide to optimize AI model architecture. Instead of simply increasing parameters indiscriminately, researchers suggest mapping how the brain allocates cognitive resources and replicates this distribution into artificial neural networks.

This methodology could create more efficient models that not only recognize patterns in concrete data but also develop true capacity for abstract reasoning — an essential ability to solve complex problems and make decisions in new contexts.

Implications for the Future of Artificial Intelligence

The findings have profound implications for the development of general artificial intelligence (AGI). Many experts argue that true intelligence requires not only information processing but conceptual understanding at multiple levels of abstraction.

If larger models sacrifice abstract thought by concrete recognition, simple scalability will not lead us to AGI. The research suggests that we need to fundamentally rethink how we structure neural networks, possibly incorporating principles of cognitive neuroscience from early design.

Next steps in abstract artificial intelligence research

Researchers now work on prototypes that implement these ideas, using neuroimaging data to inform model architectures. The goal is to create systems that combine AI processing speed with the conceptual flexibility of human thinking.

This line of research represents a fascinating convergence between neuroscience and machine learning, suggesting that the path to smarter AI may not be in making larger models, but in making them more like real brains.