interpretability
fact
neutral
Raw per-text vectors exhibit no natural cluster granularity, whereas benchmark-pooled capability vectors show an interior clustering optimum at a small number of clusters on all 12 evaluated models
Computation and Language01 Aug 2026