Borrowed causal tools, refitted to features
None of the causal-proof machinery in this paper was invented from scratch. It is older circuit-analysis tooling, pointed at a new kind of unit and sharpened along the way.
These four connections show that refitting in action: a causal-mediation technique repurposed from swapping components to swapping learned feature directions, tested against a task whose correct answer was already known from earlier work, an automated ranking method borrowed to trace which features cause which, and an even finer edit that stops precisely at a feature's own activation threshold rather than subtracting its full strength.
Four edges show the paper's causal machinery arriving from elsewhere and being refitted rather than invented. Activation patching is adapted into dictionary feature patching, keeping Vig et al.'s counterfactual-transplant logic but changing the unit of intervention from components like attention heads to individual learned directions. Dictionary feature patching is tested against the IOI task specifically because that task's circuit is already published, giving the paper external ground truth an internally-defined score like autointerpretability cannot supply. Automated circuit discovery lends its ranking logic to feature circuit detection: ablate a candidate cause, measure the drop, rank and recurse, the same routine run one level down, from ACDC's whole-task KL divergence to a single feature's activation strength. And less-than-rank-one ablation edits with finer granularity than dictionary feature patching, stopping exactly at the hyperplane boundary where a feature turns off rather than subtracting its full magnitude, sharpening the borrowed apparatus below where any baseline direction could even be defined.