Interpretability in the Wild: a Circuit for Indirect Object Identification in GPT-2 small

TL;DR

This paper presents a detailed mechanistic explanation of how GPT-2 small performs indirect object identification, involving 26 attention heads, and demonstrates the feasibility of understanding complex behaviors in language models.

Contribution

It introduces a comprehensive, causal intervention-based interpretability analysis of GPT-2 small's behavior on a natural language task, bridging the gap between simple and complex model interpretability.

Findings

26 attention heads grouped into 7 classes explain IOI behavior

Explanation is supported by faithfulness, completeness, and minimality criteria

Work demonstrates the feasibility of mechanistic understanding in language models

Abstract

Research in mechanistic interpretability seeks to explain behaviors of machine learning models in terms of their internal components. However, most previous work either focuses on simple behaviors in small models, or describes complicated behaviors in larger models with broad strokes. In this work, we bridge this gap by presenting an explanation for how GPT-2 small performs a natural language task called indirect object identification (IOI). Our explanation encompasses 26 attention heads grouped into 7 main classes, which we discovered using a combination of interpretability approaches relying on causal interventions. To our knowledge, this investigation is the largest end-to-end attempt at reverse-engineering a natural behavior "in the wild" in a language model. We evaluate the reliability of our explanation using three quantitative criteria--faithfulness, completeness and minimality. Though these criteria support our explanation, they also point to remaining gaps in our understanding. Our work provides evidence that a mechanistic understanding of large ML models is feasible, opening opportunities to scale our understanding to both larger models and more complex tasks.