Encodes yes/no, on/off, and mutually exclusive binary choices.
What It Does
Boolean.Binary neurons specialize in strict two-option scenarios — contexts where exactly two outcomes are possible and one must be selected. They activate on conditional structures ('either...or', 'if not A then B'), boolean logic ('true or false', 'yes or no'), and factual yes/no questions. They are the closest analogue to literal binary digits in human-readable text.
How It Behaves
Binary neurons are found disproportionately in the late layers, where the model is committing to a specific output choice. They co-activate with Entropy.Confidence neurons in well-defined binary scenarios (the model is certain which option applies) and with Entropy.Ambiguity when the binary choice is genuinely uncertain. Their late-layer concentration means they are particularly important to monitor for models used in decision-making applications.
Research Example
In OLMo 3 7B, Boolean.Binary neurons are strongly active during responses to yes/no medical questions — 'Is ibuprofen safe during pregnancy?' — and their response signature differs measurably between confident rejections ('No, avoid NSAIDs') and hedged responses ('Consult your doctor'). This makes them a key signal for detecting model overconfidence in high-stakes binary decisions.