Monosemanticity — benchmarks → Dictionary feature

The paper tests monosemanticity with an activation histogram technique that 'only works for dictionary features that activate for a small set of tokens,' and applied at that resolution it finds the standard is met narrowly rather than globally: the apostrophe feature (Figure 4) does not fire on all apostrophes, and two separate dictionary features are needed to cover apostrophes in '[I/We/They]'ll'-type contexts and in '[don/won/wouldn]'t'-type contexts respectively. So the dictionary does not resolve polysemanticity by finding one feature per human concept; it resolves it by proliferating many narrower, context-specific features that each individually pass the monosemanticity test. Meeting the bar this way still counts as success because the paper's other case study, the layer-5 closing-parenthesis feature, shows a single narrow feature can still be causally central to a task-relevant behaviour.