Activations

Activation-based evidence comes from the model’s internal state during computation: residual-stream vectors, attention patterns, MLP outputs, and any derived quantity (probes, SAE features, logit lens). This is the most heavily populated evidence family in current mechanistic interpretability — most published circuit-discovery and feature-analysis work lives here.

Observational

Methods that read internal state without modifying it:

• Probing — training a classifier on intermediate representations to detect whether information is linearly (or nonlinearly) decodable at a given layer. Establishes that information is present, not that it is used.
• CKA / representational similarity — centered kernel alignment and related measures compare representation geometry across layers, models, or conditions.
• Mutual information / PID — information-theoretic measures quantify how much of the input or output is captured by an intermediate representation. Partial information decomposition separates redundant, unique, and synergistic contributions.
• Logit lens / tuned lens — projecting intermediate representations through the unembedding matrix to read what the model “would predict” at each layer.
• SAE feature activations — running a sparse autoencoder on activations and reading which features fire. Observational when the features are read; interventional when they are clamped or steered.
• Attention pattern analysis — reading which positions attend to which, though attention weights alone are not causal evidence of information flow.

The probe wars

The question “do probes find features the model actually uses?” is a construct-level question (C4 Discriminant validity). A probe that achieves high accuracy on a classification task demonstrates that the information is linearly accessible in the representation. Whether the model’s own computation accesses that information requires interventional evidence — typically activation patching or DAS-IIA. The observational finding motivates the interventional test; neither replaces the other.

Interventional

Methods that modify internal state and observe consequences:

• Activation patching — replacing one activation with another (from a different input, a mean, or zero) and measuring the effect on output. The workhorse of causal circuit discovery.
• DAS-IIA (Distributed Alignment Search – Interchange Intervention Accuracy) — learning a linear subspace and intervening on the projection of activations onto that subspace.
• Ablation — zeroing, mean-ablating, or resampling activations at specific sites. Establishes necessity when behavior degrades; combined with sufficiency tests, supports causal claims.
• Steering — adding a direction vector to activations at inference time to shift model behavior. Establishes that a direction has a causal effect, though the mechanism by which the shift works may be unclear.
• SAE feature clamping — setting specific SAE features to fixed values and observing behavioral effects.

The activation-patching ecosystem

Most published causal evidence in mechanistic interpretability comes from variants of activation patching: path patching, causal tracing, causal scrubbing, ACDC, EAP, and attribution patching all belong to this cell. Different variants make different assumptions about what constitutes a “clean” counterfactual, how to handle indirect effects, and what granularity to patch at. Disagreements between methods (E1 Intervention reach) are common and informative.

Relevant criteria

Activation evidence is most directly relevant to:

Criterion	How activation evidence bears on it
I1 Necessity	Ablation → behavior degrades
I2 Sufficiency	Activation patching from clean → corrupt restores behavior
I4 Specificity	Same intervention on matched control task → no effect
E1 Intervention reach	Multiple activation-level methods agree
C3 Convergent validity	Activation evidence converges with weight or behavioral evidence