Model as a Computational Graph, Formally, we break down the residual stream as input to component 2 into the sum of the output of each previous component. We subtract off the corrupted value of component 1’s output and patch in the clean value of component 1’s output.
Direct Path Patching
Creator
Creator
Seonglae ChoCreated
Created
2024 Oct 14 1:11Editor
Editor
Seonglae ChoEdited
Edited
2025 Feb 1 22:9Refs
Refs