Through considerable effort and intuition, several recent works have reverse-engineered
nontrivial behaviors oftransformer models. This paper systematizes the mechanistic …

Despite rapid adoption and deployment of large language models (LLMs), the internal
computations of these models remain opaque and poorly understood. In this work, we seek …

This work identifies 18 foundational challenges in assuring the alignment and safety of large
language models (LLMs). These challenges are organized into three different categories …

Understanding AI systems' inner workings is critical for ensuring value alignment and safety.
This review explores mechanistic interpretability: reverse-engineering the computational …

Mechanistic interpretability seeks to understand the internal mechanisms of machine
learning models, where localization--identifying the important model components--is a key …

Mechanistic interpretability seeks to understand the neural mechanisms that enable specific
behaviors in Large Language Models (LLMs) by leveraging causality-based methods. While …

A common method to study deep learning systems is to use simplified model
representations—for example, using singular value decomposition to visualize the model's …

In this work we present successor heads: attention heads that increment tokens with a
natural ordering, such as numbers, months, and days. For example, successor heads …

Recent work in mechanistic interpretability has shown that behaviors in language models
can be successfully reverse-engineered through circuit analysis. A common criticism …

Mechanistic interpretability (MI) is an emerging sub-field of interpretability that seeks to
understand a neural network model by reverse-engineering its internal computations …