Source
Conmy, Mavor-Parker, Lynch, Heimersheim and Garriga-Alonso (2023), Towards automated circuit discovery for mechanistic interpretability, NeurIPS 2023, LIT-1547, read in full in NOTE-1170: §§3–4, Appendices C–J, Tables 2 and 3.
What was actually shown
The method. ACDC visits the nodes of a fixed computational graph from the output back and, for each incoming edge, removes it if replacing it by its value on a corrupted prompt raises the KL divergence from the full model by less than τ. Sweeping τ gives an ROC curve against the edges of a circuit reported by earlier work.
The comparison (Table 2, corrupted activations, edge level):
| task | ACDC | HISP | SP | metric |
|---|---|---|---|---|
| Docstring | 0.982 | 0.805 | 0.937 | KL |
| Greater-Than | 0.853 | 0.693 | 0.806 | KL |
| IOI | 0.869 | 0.789 | 0.823 | KL |
| tracr-xproportion | 0.679 | 0.679 | 0.525 | L2 |
| tracr-reverse | 0.200 | 0.577 | 0.193 | L2 |
With zero activations (Table 3) ACDC scores 1.000 on both tracr programs.
What could have come out otherwise. A gradient-based mask (SP) or a first-order score (HISP) could have ranked the reported edges better than greedy patching on the language-model tasks; neither did.
What this does not say
- Not that ACDC finds the circuit. The ground truths are partial: the reported IOI circuit reaches a logit difference of 3.24 against the model’s 4.11, and the Greater-Than circuit 72% against 84%.
- Not robust across choices. Optimizing the task metric instead of KL drops ACDC’s edge AUC to 0.461 on Greater-Than and 0.589 on IOI; zero activations make Docstring far worse; parent order mattered on induction.
- Not that ACDC is best on tracr. The paper’s §4.1 says ACDC leads on tracr-reverse; Table 2 gives 0.200 against HISP’s 0.577 with corrupted activations (NOTE-1170, Limitations).
- Not complete. At sparse thresholds negative name movers are missed, and an OR gate keeps one input (THEORY-966).
- Not about later methods. Edge attribution patching, which later work reports as better, is not compared here.
Connections
- THEORY-997. Circuits compared across fine-tuning are found with this family of methods; their dependence on threshold, metric and order is a reason a changed circuit need not be a changed model.
- QUESTION-275. A circuit-only measurement that could be run at every checkpoint of a small model, with those choices held fixed.