Small-World Connectivity Correlates with LLM Reasoning Performance
October 9, 2026
Analysis of attention-head activation similarities reveals that a higher small-world index (SWI) serves as a structural signature of reasoning ability. Models exhibiting higher local clustering and shorter global paths in their functional graphs consistently demonstrate superior fluid reasoning performance.
HOW THIS AFFECTS YOU
●
researcherYou can use small-world index metrics to potentially predict reasoning capabilities during training checkpoints.