Modular Cognitive Architecture Emerges in Large Language Models

Language models may develop groups of internal units that specialize in different kinds of tasks. Across 46 tasks involving language, formal reasoning, understanding other minds and reasoning about the physical world, researchers report that related tasks recruited overlapping units, while tasks from different domains recruited distinct ones. Their abstract compares that pattern with functional specialization in the human brain.

The finding suggests a way to investigate how different capabilities are organized inside a model. The authors propose that modularity could be a general organizing principle of intelligence; the reported similarity does not establish that every intelligent system must use this arrangement.