Sparse Feature Circuits: Discovering and Editing Interpretable Causal Graphs in Language Models

TL;DR

This paper presents methods for discovering sparse, interpretable subnetworks within language models that reveal causal mechanisms, enabling better understanding and manipulation of model behavior.

Contribution

It introduces sparse feature circuits and an unsupervised pipeline for scalable interpretability of language models, improving understanding and downstream task performance.

Findings

Sparse feature circuits are interpretable and causally implicated in model behavior.

The SHIFT method improves classifier generalization by ablating task-irrelevant features.

Thousands of circuits can be discovered automatically, enhancing interpretability.

Abstract

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.