Saved in:
Bibliographic Details
Main Authors: Jacobi, Jonathan, Niv, Gal
Format: Preprint
Published: 2025
Subjects:
Online Access:https://arxiv.org/abs/2503.02078
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866916646775422976
author Jacobi, Jonathan
Niv, Gal
author_facet Jacobi, Jonathan
Niv, Gal
contents Understanding and interpreting the internal representations of large language models (LLMs) remains an open challenge. Patchscopes introduced a method for probing internal activations by patching them into new prompts, prompting models to self-explain their hidden representations. We introduce Superscopes, a technique that systematically amplifies superposed features in MLP outputs (multilayer perceptron) and hidden states before patching them into new contexts. Inspired by the "features as directions" perspective and the Classifier-Free Guidance (CFG) approach from diffusion models, Superscopes amplifies weak but meaningful features, enabling the interpretation of internal representations that previous methods failed to explain-all without requiring additional training. This approach provides new insights into how LLMs build context and represent complex concepts, further advancing mechanistic interpretability.
format Preprint
id arxiv_https___arxiv_org_abs_2503_02078
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Superscopes: Amplifying Internal Feature Representations for Language Model Interpretation
Jacobi, Jonathan
Niv, Gal
Computation and Language
Artificial Intelligence
Machine Learning
Understanding and interpreting the internal representations of large language models (LLMs) remains an open challenge. Patchscopes introduced a method for probing internal activations by patching them into new prompts, prompting models to self-explain their hidden representations. We introduce Superscopes, a technique that systematically amplifies superposed features in MLP outputs (multilayer perceptron) and hidden states before patching them into new contexts. Inspired by the "features as directions" perspective and the Classifier-Free Guidance (CFG) approach from diffusion models, Superscopes amplifies weak but meaningful features, enabling the interpretation of internal representations that previous methods failed to explain-all without requiring additional training. This approach provides new insights into how LLMs build context and represent complex concepts, further advancing mechanistic interpretability.
title Superscopes: Amplifying Internal Feature Representations for Language Model Interpretation
topic Computation and Language
Artificial Intelligence
Machine Learning
url https://arxiv.org/abs/2503.02078