Tree — 05

causal/3

features/3

probing/1

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.

Seven entries. Two are stubs and are marked as such — read them as open questions rather than answers.

Entries

Activation Patching

Causal intervention swapping activations between clean and corrupt runs.

Attribution Patching

First-order approximation to patching via a single backward pass.

Crosscodersstub

Shared dictionaries fitted across layers or across model checkpoints.

Feature Absorptionstub

When one SAE latent swallows a more specific feature’s activation mass.

Project intermediate residual states through the unembedding matrix.

Restrict the intervention to a single computational path between components.

Sparse Autoencoders

Overcomplete dictionary learning on activations for monosemantic features.