Activation Patching
A standard causal intervention method in mechanistic interpretability research. It operates by replacing activations at specific layer positions from a clean forward pass into a corrupted forward pass, then observing changes in output to identify causally critical positions for information processing within a model's internal computations.
Core Facts
Timeline (last 90 days)
The intervention experiment copied the transient state at the answer position in a number prompt to another number prompt at five different depths, observing which prompt the final animal score followed; this method is essentially an Activation Patching technique.
All Facts (3)
AUC值域为0.5到1.0，0.5代表随机猜测、1.0代表完美预测，不受样本类别不平衡影响
75%UnverifiedActivation patching was widely popularized by Meng et al. in the 2022 ROME paper; its principle is to transplant the activation values of a certain layer from one run into another run to observe output changes, thereby performing causal inference.
70%UnverifiedThe intervention experiment copied the transient state at the answer position in a number prompt to another number prompt at five different depths, observing which prompt the final animal score followed; this method is essentially an Activation Patching technique.
50%