Definition
Plain language
When a single internal unit in a model cleanly stands for just one human-understandable concept.
As stated in the literature
The property that a representational unit responds to exactly one interpretable concept, the goal of sparse-autoencoder feature extraction; contrasted with the polysemantic neurons of raw transformers and named in the Scaling Monosemanticity work.
Also called: monosemantic
Why it matters: Getting internal units to each mean one thing is the goal that would make a model's reasoning genuinely readable to humans.
For example, instead of one internal unit reacting to bridges, Korean text, and lawsuits all at once, a clean unit would respond to just bridges.