Attribution Patching Outperforms Automated Circuit Discovery
This is an incremental improvement for researchers in interpretability, offering a more efficient method for identifying subnetworks in large models.
The paper tackled the problem of automated circuit discovery in neural networks by proposing attribution patching, which outperforms existing methods with greater AUC in circuit recovery across tasks.
Automated interpretability research has recently attracted attention as a potential research direction that could scale explanations of neural network behavior to large models. Existing automated circuit discovery work applies activation patching to identify subnetworks responsible for solving specific tasks (circuits). In this work, we show that a simple method based on attribution patching outperforms all existing methods while requiring just two forward passes and a backward pass. We apply a linear approximation to activation patching to estimate the importance of each edge in the computational subgraph. Using this approximation, we prune the least important edges of the network. We survey the performance and limitations of this method, finding that averaged over all tasks our method has greater AUC from circuit recovery than other methods.