Interpretability
Methods for reading a trained network rather than training a better one: sparse dictionaries over activations, causal interventions on the residual stream, and the circuit-level claims those techniques license.
Nine entries across four branches. Two are stubs and are marked as such — read them as open questions rather than answers.
Features
features/ ›decomposing activations into things that can be named
05.1O(L·d·m)05.2—05.3O(d·m)
Crosscodersstubpromising
Shared dictionaries fitted across layers or across model checkpoints.
Feature Absorptionstubpromising
When one SAE latent swallows a more specific feature’s activation mass.
Sparse Autoencoderscommon
Overcomplete dictionary learning on activations for monosemantic features.
Circuits
circuits/ ›named mechanisms, traced end to end
Causal methods
causal/ ›intervening on activations to find what carries the behaviour
05.5O(L·n)05.6O(L·n)05.7O(L²·n)05.8O(L·d)
Activation Patchingstandard
Causal intervention swapping activations between clean and corrupt runs.
Attribution Patchingcommon
First-order approximation to patching via a single backward pass.
Path Patchingcommon
Restrict the intervention to a single computational path between components.
Steering Vectorscommon
Add a direction to the residual stream at inference to shift behaviour.
Probing
probing/ ›reading a model with a simpler model, and its limits