Successor Heads: Recurring, Interpretable Attention Heads In The Wild

Published: 16 Jan 2024, Last Modified: 14 Mar 2024ICLR 2024 posterEveryoneRevisionsBibTeX
Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.
Keywords: Interpretability, Large Language Models, Natural Language Processing, Science of Deep Learning, Representation Learning
Submission Guidelines: I certify that this submission complies with the submission instructions as described on
TL;DR: We find "Successor Heads": a type of attention head that transforms numeric representations in language models. We find successor heads across language models of different sizes and find the abstract number representations that make them work.
Abstract: In this work we describe successor heads: attention heads that increment tokens with a natural ordering, such as numbers, months, and days. For example, successor heads increment 'Monday' into 'Tuesday'. We explain the successor head behavior with an approach rooted in mechanistic interpretability, the field that aims to explain how models complete tasks in human-understandable terms. Existing research in this area has struggled to find recurring, mechanistically interpretable large language model (LLM) components beyond small toy models. Further, existing results have led to very little insight to explain the internals of the larger models that are used in practice. In this paper, we analyze the behavior of successor heads in LLMs and find that they implement abstract representations that are common to different architectures. Successor heads form in LLMs with as few as 31 million parameters, and at least as many as 12 billion parameters, such as GPT-2, Pythia, and Llama-2. We find a set of 'mod 10' features that underlie how successor heads increment in LLMs across different architectures and sizes. We perform vector arithmetic with these features to edit head behavior and provide insights into numeric representations within LLMs. Additionally, we study the behavior of successor heads on natural language data, where we find that successor heads are important for achieving a low loss on examples involving succession, and also identify interpretable polysemanticity in a Pythia successor head.
Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors' identity.
No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.
Primary Area: visualization or interpretation of learned representations
Submission Number: 3486