Monosemanticity
- Control
- Research term
Monosemanticity is how well a part of a model, such as a neuron or an extracted feature, fits one coherent description across the inputs that activate it.
Single neurons, the small units that each hold one number, often fail to fit one description. A sparse autoencoder, a second small network trained on the model's activity, re-expresses that activity as many more features, patterns that each switch on only now and then. Training them to fire seldom is meant to push each toward a single kind of thing, and often does.
Where you’ll meet it
You read an AI lab's research post about a feature inside its model labeled "Golden Gate Bridge." It switches on for the bridge's name in several languages and for photos of it. The post shows a dozen of the examples where it fires most strongly.
What to do
Labels are usually written from those top examples. Look for how the feature behaves when it fires weakly, and how often it fires on things that aren't the bridge. Those show whether the label describes the feature's whole range or only the top of the list.
Easy to mix up
Polysemanticity may help the model, which can fit more ideas into the parts it has. Monosemanticity helps the people studying it, because one name then fits what a part does.
Check yourself
A feature fires on flattering replies and almost nothing else across a thousand samples. Turning it down leaves replies just as flattering. What does this evidence support?
Go deeperHow finely the features are cut decides what counts as one description
Anthropic's "Towards Monosemanticity" (2023) applied sparse autoencoders to a small one-layer model, and "Scaling Monosemanticity" (2024) applied them to a production model. The first paper found that asking for more features splits one broad feature into several finer ones, so what counts as one description depends on the size the researchers chose.
Researchers debate whether these features are the model's own units or a convenient slicing of its activity that other methods would cut differently, and no method yet shows that a set of them is complete. The features also rebuild the model's activity imperfectly, and the part they leave unexplained can matter for the behavior under study.