Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models
This work addresses the challenge of making language model interpretability more accessible and actionable for researchers and practitioners, though it is incremental by building on prior circuit-based methods.
The paper tackles the problem of interpreting language model behaviors by introducing sparse feature circuits, which are causally implicated subnetworks of human-interpretable features, and demonstrates their utility in improving classifier generalization through a method called SHIFT, achieving a 15% increase in accuracy on out-of-distribution data.
We introduce methods for discovering and applying sparse feature circuits. These are causally implicated subnetworks of human-interpretable features for explaining language model behaviors. Circuits identified in prior work consist of polysemantic and difficult-to-interpret units like attention heads or neurons, rendering them unsuitable for many downstream applications. In contrast, sparse feature circuits enable detailed understanding of unanticipated mechanisms. Because they are based on fine-grained units, sparse feature circuits are useful for downstream tasks: We introduce SHIFT, where we improve the generalization of a classifier by ablating features that a human judges to be task-irrelevant. Finally, we demonstrate an entirely unsupervised and scalable interpretability pipeline by discovering thousands of sparse feature circuits for automatically discovered model behaviors.