Automatically Identifying Local and Global Circuits with Linear Computation Graphs

TL;DR

This paper presents a novel pipeline using Sparse Autoencoders and Transcoders to identify local and global circuits in neural networks through linear computation graphs, enabling detailed mechanistic interpretability without linear approximation.

Contribution

Introduces a new circuit discovery pipeline with SAEs and Transcoders that produces strictly linear computation graphs for neural models, enhancing interpretability.

Findings

Analyzed GPT-2 Small circuits including bracket, induction, and indirect object identification.

Revealed new insights underlying existing circuit discoveries.

Demonstrated scalable application with Hierarchical Attribution.

Abstract

Circuit analysis of any certain model behavior is a central task in mechanistic interpretability. We introduce our circuit discovery pipeline with Sparse Autoencoders (SAEs) and a variant called Transcoders. With these two modules inserted into the model, the model's computation graph with respect to OV and MLP circuits becomes strictly linear. Our methods do not require linear approximation to compute the causal effect of each node. This fine-grained graph identifies both end-to-end and local circuits accounting for either logits or intermediate features. We can scalably apply this pipeline with a technique called Hierarchical Attribution. We analyze three kinds of circuits in GPT-2 Small: bracket, induction, and Indirect Object Identification circuits. Our results reveal new findings underlying existing discoveries.