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

TL;DR

Patchscopes is a comprehensive framework that enables detailed inspection and explanation of hidden representations in language models, unifying existing methods and addressing their limitations.

Contribution

The paper introduces Patchscopes, a unifying framework for interpreting LLM internal representations, improving upon prior methods and enabling new interpretability applications.

Findings

Patchscopes unifies various interpretability techniques.

It mitigates shortcomings of previous methods, such as early layer inspection issues.

Enables new applications like cross-model explanations and multihop reasoning correction.

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.