Modular Cognitive Architecture Emerges in Large Language Models
LLM circuits separated by cognitive domain in a pattern the authors say mirrors human brain networks.
The paper tests 46 tasks across language, formal reasoning, social reasoning, and physical reasoning. Its circuit analyses found overlapping neurons for tasks tied to the same human brain network, and more distinct neurons for tasks tied to different networks. The authors argue this convergence suggests modularity may be a broader feature of intelligent systems, not just biological brains. HF Daily Papers' note
The paper tests 46 tasks across language, formal reasoning, social reasoning, and physical reasoning. Its circuit analyses found overlapping neurons for tasks tied to the same human brain network, and more distinct neurons for tasks tied to different networks. The authors argue this convergence suggests modularity may be a broader feature of intelligent systems, not just biological brains. HF Daily Papers' note
score 6