nucleation

LIT-1547: Towards Automated Circuit Discovery for Mechanistic Interpretability

1 min read

Properties5
StatusActive — read 2026-10-26 (NOTE-1170) in full, from arXiv v4, the NeurIPS 2023 version; worth reading for the workflow it names behind manual circuit discovery, for ACDC as that workflow's patching step made into an algorithm, and for the first scoring of circuit-finding methods against circuits found by hand (THEORY-970). Read it knowing that the hand-found circuits are themselves partial, that the results move with the metric, the ablation and the order of visiting edges, that sparse runs miss negative components, that an OR gate keeps one input (THEORY-966), and that §4.1 claims a tracr-reverse lead its own Table 2 does not show.
Version2
Tagsrepresentation-learning, causality, anthology-candidate
Cited as “paper” byNOTE-1170 — Conmy, Mavor-Parker, Lynch, Heimersheim and Garriga-Alonso — Towards automated circuit discovery (ACDC)
Cited as “source” byTHEORY-966 — Greedy edge pruning of the ACDC kind, which removes each edge whose removal changes the output metric by less than a threshold given the edges already removed, keeps only one input of a redundant OR gate, and which input it keeps depends on the order in which edges are visited, so when a model has redundant sufficient paths the circuit it returns is one of several sufficient subgraphs, THEORY-970 — On three language-model tasks whose circuits earlier manual work reported, indirect object identification and Greater-Than in GPT-2 Small and Python docstrings in a four-layer attention-only model, ACDC run with resampled corrupted activations against KL divergence ranked the reported edges with areas under the ROC curve of 0.85 to 0.98, above head-importance pruning and subnetwork probing on all three; on two compiled tracr programs it ranked them poorly with corrupted activations, 0.68 and 0.20, and exactly with zero activations

Graph View

Conmy, A., Mavor-Parker, A. N., Lynch, A., Heimersheim, S. & Garriga-Alonso, A. (2023). Towards automated circuit discovery for mechanistic interpretability. Advances in Neural Information Processing Systems 36 (NeurIPS 2023). arXiv:2304.14997.

Read in full on 2026-10-26 (NOTE-1170). Two accounts were drawn from it: THEORY-970 and THEORY-966.

Why this is here

The owner asked on 2026-10-26 how the record stood on neural circuit theory; the answer found it held almost none of the core circuits literature, and the owner said “go ahead and register and read those”. ACDC is the reference algorithm for finding a circuit rather than testing one, and the ancestor of the edge-attribution methods behind LIT-1539 (THEORY-997) and benchmarked by MIB (LIT-1550), which the record held without their source. It is the search counterpart of causal scrubbing (LIT-1549), runs on the induction setting of LIT-1538 and QUESTION-275, and its OR-gate case bears on THEORY-802’s several correct explanations.

Backlinks