ExpertLens: Activation steering features are highly interpretable

Fuente: arXiv
Gespeichert in:
Bibliographische Detailangaben
Hauptverfasser: Fedzechkina, Masha, Gualdoni, Eleonora, Williamson, Sinead, Metcalf, Katherine, Seto, Skyler, Theobald, Barry-John
Format: Preprint
Veröffentlicht: 2025
Schlagworte:
Online-Zugang:
Tags: Tag hinzufügen
Keine Tags, Fügen Sie den ersten Tag hinzu!
_version_ 1866909885087612928
author Fedzechkina, Masha
Gualdoni, Eleonora
Williamson, Sinead
Metcalf, Katherine
Seto, Skyler
Theobald, Barry-John
author_facet Fedzechkina, Masha
Gualdoni, Eleonora
Williamson, Sinead
Metcalf, Katherine
Seto, Skyler
Theobald, Barry-John
contents Activation steering methods in large language models (LLMs) have emerged as an effective way to perform targeted updates to enhance generated language without requiring large amounts of adaptation data. We ask whether the features discovered by activation steering methods are interpretable. We identify neurons responsible for specific concepts (e.g., ``cat'') using the ``finding experts'' method from research on activation steering and show that the ExpertLens, i.e., inspection of these neurons provides insights about model representation. We find that ExpertLens representations are stable across models and datasets and closely align with human representations inferred from behavioral data, matching inter-human alignment levels. ExpertLens significantly outperforms the alignment captured by word/sentence embeddings. By reconstructing human concept organization through ExpertLens, we show that it enables a granular view of LLM concept representation. Our findings suggest that ExpertLens is a flexible and lightweight approach for capturing and analyzing model representations.
format Preprint
id arxiv_https___arxiv_org_abs_2502_15090
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle ExpertLens: Activation steering features are highly interpretable
Fedzechkina, Masha
Gualdoni, Eleonora
Williamson, Sinead
Metcalf, Katherine
Seto, Skyler
Theobald, Barry-John
Computation and Language
Artificial Intelligence
Activation steering methods in large language models (LLMs) have emerged as an effective way to perform targeted updates to enhance generated language without requiring large amounts of adaptation data. We ask whether the features discovered by activation steering methods are interpretable. We identify neurons responsible for specific concepts (e.g., ``cat'') using the ``finding experts'' method from research on activation steering and show that the ExpertLens, i.e., inspection of these neurons provides insights about model representation. We find that ExpertLens representations are stable across models and datasets and closely align with human representations inferred from behavioral data, matching inter-human alignment levels. ExpertLens significantly outperforms the alignment captured by word/sentence embeddings. By reconstructing human concept organization through ExpertLens, we show that it enables a granular view of LLM concept representation. Our findings suggest that ExpertLens is a flexible and lightweight approach for capturing and analyzing model representations.
title ExpertLens: Activation steering features are highly interpretable
topic Computation and Language
Artificial Intelligence
url https://arxiv.org/abs/2502.15090