Automated Circuit Discovery (ACDC) — lends its ranking logic to → Automatic feature circuit detection

Automated Circuit Discovery is a borrowed algorithm; feature-circuit-detection is the paper’s own, and the two share a family resemblance the paper never states outright. Section 4.1 explicitly imports ACDC’s Algorithm 4.1 (Conmy et al., 2023) to select which dictionary features to patch for the IOI task: treating every feature as a node in a flat computational graph, ablating it, and ordering features by how much removing each one moves the model’s output KL divergence away from the target (sparse-autoencoders, §"4.1 Adapting activation patching to dictionary features", p. 6). Section 5.3’s feature-circuit-detection procedure, used to trace which upstream features cause the closing-parenthesis feature to fire, runs a strikingly similar routine one level down: for each candidate feature in the previous layer, ablate it, rerun the model, and rank previous-layer features by how much their removal decreases the target feature’s own activation, optionally recursing into further layers (sparse-autoencoders, §"5.3 Intermediate features: dictionary features allow automatic circuit detection", p. 7-8). Neither ACDC nor Conmy et al. is cited again in Section 5.3, so the paper presents feature-circuit-detection as a fresh, self-contained method rather than an explicit reapplication. But the underlying move — ablate a candidate cause, measure the drop in a target metric, rank and recurse — is the same one Section 4.1 imports wholesale, run here on a single feature’s activation strength instead of a whole task’s KL divergence.