Interpretability

What Interpretability is

Interpretability aims to identify circuits, features and representations inside a model, so behaviour can be explained and predicted rather than only observed.

How it works

Techniques include probing classifiers, activation patching, sparse autoencoders that decompose activations into interpretable features, and causal interventions on internal components.

Why it matters

It offers the possibility of detecting deception, bias or unsafe capability before it manifests in outputs, which black-box evaluation cannot do.

Common uses

• →Safety auditing
• →Debugging unexpected behaviour
• →Steering model behaviour by feature