Reproduce “Certified Circuits: Stability Guarantees for Mechanistic Circuits” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Alaa Anani, Tobias Lorenz, Bernt Schiele, Mario Fritz, Jonas Fischer
Certified Circuits reruns circuit-finding on randomly thinned datasets, certifying consistently selected components and abstaining on uncertain ones. Its guarantee covers non-abstained decisions under bounded dataset edits, not semantic correctness or distribution-shift generalization. For ResNet-101 circuits from ImageNet, evaluated on ImageNet-A, peak cACC is 0.94 versus 0.60, with 15% fewer neurons; this is a circuit-level result, not a full-model improvement.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.