Hypothesis Testing the Circuit Hypothesis in LLMs

TL;DR

This paper develops hypothesis tests to evaluate whether small subnetworks, or circuits, within large language models are responsible for their capabilities, and applies these tests to both synthetic and discovered circuits.

Contribution

It formalizes criteria and creates a suite of tests for assessing circuits in LLMs, and introduces the circuitry software package for empirical circuit analysis.

Findings

Synthetic circuits align with idealized properties.

Discovered circuits satisfy criteria to varying degrees.

Software facilitates future circuit studies.

Abstract

Large language models (LLMs) demonstrate surprising capabilities, but we do not understand how they are implemented. One hypothesis suggests that these capabilities are primarily executed by small subnetworks within the LLM, known as circuits. But how can we evaluate this hypothesis? In this paper, we formalize a set of criteria that a circuit is hypothesized to meet and develop a suite of hypothesis tests to evaluate how well circuits satisfy them. The criteria focus on the extent to which the LLM's behavior is preserved, the degree of localization of this behavior, and whether the circuit is minimal. We apply these tests to six circuits described in the research literature. We find that synthetic circuits -- circuits that are hard-coded in the model -- align with the idealized properties. Circuits discovered in Transformer models satisfy the criteria to varying degrees. To facilitate future empirical studies of circuits, we created the \textit{circuitry} package, a wrapper around the \textit{TransformerLens} library, which abstracts away lower-level manipulations of hooks and activations. The software is available at \url{https://github.com/blei-lab/circuitry}.