Steering the CensorShip: Uncovering Representation Vectors for LLM "Thought" Control

Fuente: arXiv
Guardado en:
Detalles Bibliográficos
Autores principales: Cyberey, Hannah, Evans, David
Formato: Preprint
Publicado: 2025
Materias:
Acceso en línea:
Etiquetas: Agregar Etiqueta
Sin Etiquetas, Sea el primero en etiquetar este registro!
_version_ 1866911101079257088
author Cyberey, Hannah
Evans, David
author_facet Cyberey, Hannah
Evans, David
contents Large language models (LLMs) have transformed the way we access information. These models are often tuned to refuse to comply with requests that are considered harmful and to produce responses that better align with the preferences of those who control the models. To understand how this "censorship" works. We use representation engineering techniques to study open-weights safety-tuned models. We present a method for finding a refusal--compliance vector that detects and controls the level of censorship in model outputs. We also analyze recent reasoning LLMs, distilled from DeepSeek-R1, and uncover an additional dimension of censorship through "thought suppression". We show a similar approach can be used to find a vector that suppresses the model's reasoning process, allowing us to remove censorship by applying the negative multiples of this vector. Our code is publicly available at: https://github.com/hannahxchen/llm-censorship-steering
format Preprint
id arxiv_https___arxiv_org_abs_2504_17130
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Steering the CensorShip: Uncovering Representation Vectors for LLM "Thought" Control
Cyberey, Hannah
Evans, David
Computation and Language
Cryptography and Security
Computers and Society
Large language models (LLMs) have transformed the way we access information. These models are often tuned to refuse to comply with requests that are considered harmful and to produce responses that better align with the preferences of those who control the models. To understand how this "censorship" works. We use representation engineering techniques to study open-weights safety-tuned models. We present a method for finding a refusal--compliance vector that detects and controls the level of censorship in model outputs. We also analyze recent reasoning LLMs, distilled from DeepSeek-R1, and uncover an additional dimension of censorship through "thought suppression". We show a similar approach can be used to find a vector that suppresses the model's reasoning process, allowing us to remove censorship by applying the negative multiples of this vector. Our code is publicly available at: https://github.com/hannahxchen/llm-censorship-steering
title Steering the CensorShip: Uncovering Representation Vectors for LLM "Thought" Control
topic Computation and Language
Cryptography and Security
Computers and Society
url https://arxiv.org/abs/2504.17130