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

Hannah Cyberey, David Evans
4/23/2025
cs.CLcs.CRcs.CY

Official code: hannahxchen/llm-censorship-steering (12 stars, last commit 2025-12-22).

Abstract

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

View on arXivView PDF

Code Implementations(1)

hannahxchen/llm-censorship-steeringOfficial100%

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

121PythonFeb 28, 202510 months ago

Similar Papers

LLM Censorship: A Machine Learning Challenge or a Computer Security Problem?

89% match2023

Are LLMs Good Safety Agents or a Propaganda Engine?

86% match2025

Refusal Steering: Fine-grained Control over LLM Refusal Behaviour for Sensitive Topics

86% match2025

An Embarrassingly Simple Defense Against LLM Abliteration Attacks

85% match2025

Refuse Whenever You Feel Unsafe: Improving Safety in LLMs via Decoupled Refusal Training

84% match1 repo722024

Information Suppression in Large Language Models: Auditing, Quantifying, and Characterizing Censorship in DeepSeek

84% match2025

Cite this paper

@article{cyberey2025steering,
  title  = {Steering the CensorShip: Uncovering Representation Vectors for LLM "Thought" Control},
  author = {Hannah Cyberey and David Evans},
  year   = {2025},
  eprint = {2504.17130},
  archivePrefix = {arXiv},
  url    = {http://arxiv.org/abs/2504.17130v3}
}

Add a badge to your README

Shows whether official code is linked for this paper.

Code of Paper badge
[![Code of Paper](https://codeofpaper.com/badge/2504.17130.svg)](https://codeofpaper.com/paper/2504.17130)

Discussion