Encodes variation and inconsistency across contexts.
What It Does
Entropy.Variability neurons activate on contexts where outcomes, behaviors, or values vary across instances or conditions: explicit variability markers ('results vary', 'outcomes differ', 'performance fluctuates'), conditional variation ('it depends on X', 'varies by context'), and distributional spread ('some models perform well; others do not'). They encode the concept of inconsistency or distributed outcomes.
How It Behaves
Variability neurons are extremely rare — tied with Probabilistic for the smallest element in our corpus — and show a middle-layer concentration. Despite their rarity, they have high mean firing magnitude, suggesting they carry strong signal when they fire. They interact with Entropy.Probabilistic neurons (variability is related to probability distributions) and Number.Statistic neurons (statistical spread is one form of variability).
Research Example
In Qwen3 1.7B, Entropy.Variability neurons activate strongly on 'individual results may vary' (product disclaimer language), 'model performance varies across tasks' (AI benchmark contexts), and 'symptoms differ between patients' (medical contexts) — three completely different domains but the same Variability neuron pattern. The domain-generality of this element confirms it encodes a semantic concept (variability) rather than domain-specific patterns.