Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models

Asma Ghandeharioun, Avi Caciularu, Adam Pearce, Lucas Dixon, Mor Geva

Research output: Contribution to journalConference articlepeer-review

Abstract

Understanding the internal representations of large language models (LLMs) can help explain models' behavior and verify their alignment with human values. Given the capabilities of LLMs in generating human-understandable text, we propose leveraging the model itself to explain its internal representations in natural language. We introduce a framework called Patchscopes and show how it can be used to answer a wide range of questions about an LLM's computation. We show that many prior interpretability methods based on projecting representations into the vocabulary space and intervening on the LLM computation can be viewed as instances of this framework. Moreover, several of their shortcomings such as failure in inspecting early layers or lack of expressivity can be mitigated by Patchscopes. Beyond unifying prior inspection techniques, Patchscopes also opens up new possibilities such as using a more capable model to explain the representations of a smaller model, and multihop reasoning error correction.

Original languageEnglish
Pages (from-to)15466-15490
Number of pages25
JournalProceedings of Machine Learning Research
Volume235
StatePublished - 2024
Externally publishedYes
Event41st International Conference on Machine Learning, ICML 2024 - Vienna, Austria
Duration: 21 Jul 202427 Jul 2024

Bibliographical note

Publisher Copyright:
Copyright 2024 by the author(s)

Fingerprint

Dive into the research topics of 'Patchscopes: A Unifying Framework for Inspecting Hidden Representations of Language Models'. Together they form a unique fingerprint.

Cite this