Modular Cognitive Architecture Emerges in Large Language Models (pengrui-han.github.io)

Researchers from MIT have discovered that Large Language Models (LLMs) exhibit a modular cognitive architecture, akin to that of the human brain. Analyzing six instruction-tuned LLMs across 46 tasks in four cognitive domains—language, formal reasoning, social reasoning, and physical reasoning—they found that tasks supported by the same networks in humans activate overlapping sets of neurons in LLMs. This suggests that modularity may not be an exclusive feature of biological brains but rather a fundamental property of intelligent systems. This research is significant for the AI/ML community as it supports the notion that the structural organization of intelligence might transcend the methods used to create it. The team employed advanced techniques like attribution patching to quantify modular organization by evaluating the overlap of task-supporting neurons. The findings reveal a clear division between domains, noting that within-domain ablations lead to a substantial drop in model accuracy compared to cross-domain ones. Moreover, LLMs demonstrate a similar efficiency in specialized neural activation as seen in human cognition—indicating that modular organization can emerge even under different optimization principles, challenging existing assumptions about cognitive architectures in intelligent systems.