A new study published on ArXiv (cs.AI) investigates whether functional specialization—a hallmark of the human brain—is a fundamental principle of intelligence or a biological accident. While the human brain utilizes distinct networks for language, formal reasoning, social cognition, and physical understanding, the research tested whether similar patterns emerge in Large Language Models (LLMs).
Methodology and Circuit Analysis
Researchers conducted circuit analyses across N=46 tasks spanning four primary cognitive domains:
- Language
- Formal reasoning
- Social reasoning (reasoning about other minds)
- Physical reasoning (reasoning about the physical world)
The findings indicate that LLMs develop a modular architecture that mirrors human brain organization. Specifically, tasks that draw on the same network in humans recruit overlapping neurons in LLMs, whereas tasks drawing on different networks recruit distinct neurons.
Convergent Emergence of Intelligence
The fact that two systems created through vastly different optimization processes—biological evolution and neural network training—arrive at a similar modular organization suggests that such structure may be essential for intelligence. This convergent emergence implies that modularity is not a coincidence but potentially a fundamental property of intelligent systems.
