Search papers, labs, and topics across Lattice.
1
0
1
137
Mechanistic interpretability gets a formal footing: "Certified Circuits" uses data subsampling to find provably stable sub-networks, boosting accuracy by up to 91% while using 45% fewer neurons.
