Unraveling the cognitive patterns of Large Language Models through module communities

Kushal Raj Bhandari, Pin-Yu Chen, Jianxi Gao

Track: Extended Abstract Track

Keywords: Large Language Models, Network Community Structure, Cognitive Skills, AI interpretability

TL;DR: The paper maps cognitive skills and datasets to LLM modules, showing skills spread across communities; fine-tuning influences weights but not accuracy, suggesting LLMs work best as interconnected systems, not isolated parts.

Abstract: Large Language Models (LLMs) have transformed science, engineering, and society through applications from discovery and diagnostics to chatbots. Yet their mechanisms remain hidden within billions of parameters, making their architecture and cognitive processes difficult to grasp. We address this by drawing on biological cognition and introducing a network-based framework linking cognitive skills, LLM architectures, and datasets. The skill distribution in module communities shows that, while LLMs lack the highly localized specialization of some biological systems, they form distinct module clusters whose skill patterns partly mirror the distributed organization of avian and small mammalian brains. Our results highlight that skill acquisition in LLMs benefits from dynamic, cross-regional interactions and plasticity. Integrating cognitive science and machine learning, this framework offers new interpretability insights and suggests fine-tuning strategies should emphasize distributed learning over rigid modular approaches.

Submission Number: 53