@inproceedings{abu-shairah-etal-2026-turning,
title = "Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection",
author = "Abu Shairah, Harethah and
Hammoud, Hasan Abed Al Kader and
Turkiyyah, George and
Ghanem, Bernard",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.emnlp-main.75/",
pages = "1722--1738",
ISBN = "979-8-89176-459-0",
abstract = "Safety alignment in Large Language Models (LLMs) often involves mediating internal representations to refuse harmful requests. Recent research has demonstrated that these safety mechanisms can be bypassed by ablating or removing specific representational directions within the model. In this paper, we propose the opposite approach: Rank-One Safety Injection (ROSI), a white-box method that amplifies a model{'}s safety alignment by permanently steering its activations toward the refusal-mediating subspace. ROSI operates as a simple, fine-tuning-free rank-one weight modification applied to all residual stream write matrices. The required safety direction can be computed from a small set of harmful and harmless instruction pairs. We show that ROSI consistently increases safety refusal rates while preserving the utility of many models across diverse architectures. Furthermore, we show that ROSI can also re-align ``uncensored'' models by amplifying their own latent safety directions, demonstrating its utility as an effective last-mile safety procedure. Our results show that targeted, interpretable weight steering is a cheap and effective mechanism to improve LLM safety, complementing more resource-intensive fine-tuning paradigms."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="abu-shairah-etal-2026-turning">
<titleInfo>
<title>Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection</title>
</titleInfo>
<name type="personal">
<namePart type="given">Harethah</namePart>
<namePart type="family">Abu Shairah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Hasan</namePart>
<namePart type="given">Abed</namePart>
<namePart type="given">Al</namePart>
<namePart type="given">Kader</namePart>
<namePart type="family">Hammoud</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">George</namePart>
<namePart type="family">Turkiyyah</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Bernard</namePart>
<namePart type="family">Ghanem</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2026-10</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing</title>
</titleInfo>
<name type="personal">
<namePart type="given">Hung-Yi</namePart>
<namePart type="family">Lee</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Sunipa</namePart>
<namePart type="family">Dev</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Jessy</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Desmond</namePart>
<namePart type="family">Elliott</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Budapest, Hungary</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-89176-459-0</identifier>
</relatedItem>
<abstract>Safety alignment in Large Language Models (LLMs) often involves mediating internal representations to refuse harmful requests. Recent research has demonstrated that these safety mechanisms can be bypassed by ablating or removing specific representational directions within the model. In this paper, we propose the opposite approach: Rank-One Safety Injection (ROSI), a white-box method that amplifies a model’s safety alignment by permanently steering its activations toward the refusal-mediating subspace. ROSI operates as a simple, fine-tuning-free rank-one weight modification applied to all residual stream write matrices. The required safety direction can be computed from a small set of harmful and harmless instruction pairs. We show that ROSI consistently increases safety refusal rates while preserving the utility of many models across diverse architectures. Furthermore, we show that ROSI can also re-align “uncensored” models by amplifying their own latent safety directions, demonstrating its utility as an effective last-mile safety procedure. Our results show that targeted, interpretable weight steering is a cheap and effective mechanism to improve LLM safety, complementing more resource-intensive fine-tuning paradigms.</abstract>
<identifier type="citekey">abu-shairah-etal-2026-turning</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.75/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>1722</start>
<end>1738</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection
%A Abu Shairah, Harethah
%A Hammoud, Hasan Abed Al Kader
%A Turkiyyah, George
%A Ghanem, Bernard
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-459-0
%F abu-shairah-etal-2026-turning
%X Safety alignment in Large Language Models (LLMs) often involves mediating internal representations to refuse harmful requests. Recent research has demonstrated that these safety mechanisms can be bypassed by ablating or removing specific representational directions within the model. In this paper, we propose the opposite approach: Rank-One Safety Injection (ROSI), a white-box method that amplifies a model’s safety alignment by permanently steering its activations toward the refusal-mediating subspace. ROSI operates as a simple, fine-tuning-free rank-one weight modification applied to all residual stream write matrices. The required safety direction can be computed from a small set of harmful and harmless instruction pairs. We show that ROSI consistently increases safety refusal rates while preserving the utility of many models across diverse architectures. Furthermore, we show that ROSI can also re-align “uncensored” models by amplifying their own latent safety directions, demonstrating its utility as an effective last-mile safety procedure. Our results show that targeted, interpretable weight steering is a cheap and effective mechanism to improve LLM safety, complementing more resource-intensive fine-tuning paradigms.
%U https://aclanthology.org/2026.emnlp-main.75/
%P 1722-1738
Markdown (Informal)
[Turning the Spell Around: Lightweight Alignment Amplification via Rank-One Safety Injection](https://aclanthology.org/2026.emnlp-main.75/) (Abu Shairah et al., EMNLP 2026)
ACL