Activation patching – Definition & Meaning

Definition of activation patching

: a mechanistic-interpretability technique that replaces selected internal activations from one model run with activations from another and measures how the intervention changes the output
- Activation patching identified the attention head whose state carried the subject's name to the final prediction.
- The researchers patched clean activations into a corrupted run to test which layers restored the correct answer.

Origin & history

first known use2022

The intervention has roots in causal mediation and interchange-intervention research. It became common in transformer interpretability in 2022 through work on factual recall and circuit analysis, and the label activation patching was standardized in later mechanistic-interpretability tutorials and studies.

Test yourself

Which of these is the meaning of activation patching?

Related terms

activation steering a method for changing an AI model's behavior during inference by modifying selected internal… interpretability research into understanding what is happening inside a model, such as which internal features… sparse autoencoder an autoencoder trained so that only a small fraction of its learned features activate for any… neural network a computing system made of many simple connected units ("neurons") arranged in layers, which… black box a system whose inputs and outputs can be seen but whose internal reasoning cannot easily be… large language model a very large neural network trained on huge amounts of text to predict the next token, which…

Cite this entry

"Activation patching." AI Dictionary, Dadgogo, https://dadgogo.com/dictionary/activation-patching/. Accessed 8 Oct. 2026.