Encodes statistical measures, averages, and probability estimates.
What It Does
Number.Statistic neurons activate on statistical and analytical quantitative contexts: averages ('the mean age', 'on average'), statistical findings ('studies show', 'research indicates', 'data suggests'), probability expressions ('likely', '80 percent chance', 'statistically significant'), and measurement terms ('accuracy', 'precision', 'rate'). They encode the epistemic status of numbers — not just a quantity but a quantity derived from analysis.
How It Behaves
Statistic neurons concentrate in the middle layers, reflecting the inferential nature of statistical reasoning. They often co-activate with Entropy neurons — statistical statements are inherently about uncertainty — and with Number.Fraction neurons for percentage-form statistics. They are more active in Qwen3 and OLMo models (which have stronger scientific text in training data) than in older models like GPT-2.
Research Example
In Llama 3.1 8B, Number.Statistic neurons are active on 'GPT-4 achieves 90 percent accuracy on the benchmark' in a way that differs from 'the class scored 90 percent on the test' — the former triggers stronger Statistic neuron firing due to the AI performance evaluation context. This specialization is one reason why models are inconsistent about citing statistical claims accurately: the Statistic neuron cluster is doing different work than Cardinal neurons, and errors at either stage produce different failure modes.