Why it matters

Causal tracing localizes knowledge. Understanding shapes editing methods.

Advertisement

The architecture

Sweep patches across (layer, position).

Heatmap of effect.

Identify critical states.

Causal tracing structureSweep (layer, pos)systematic patchesEffect heatmaplogit differencesLocalizecritical statesMeng et al. 2022 ROME/MEMIT edit models via localized knowledge
Causal tracing.
Advertisement

How it works end to end

Meng et al. ROME + MEMIT.

Factual recall localized to mid-layer MLPs.

Enables targeted editing.