<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="bach-etal-2026-continual">
    <titleInfo>
        <title>Continual Safety Alignment via Gradient-Based Sample Selection</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Thong</namePart>
        <namePart type="family">Bach</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Dung</namePart>
        <namePart type="family">Nguyen</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Thao</namePart>
        <namePart type="given">Minh</namePart>
        <namePart type="family">Le</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Truyen</namePart>
        <namePart type="family">Tran</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <originInfo>
        <dateIssued>2026-07</dateIssued>
    </originInfo>
    <typeOfResource>text</typeOfResource>
    <relatedItem type="host">
        <titleInfo>
            <title>Findings of the Association for Computational Linguistics: ACL 2026</title>
        </titleInfo>
        <name type="personal">
            <namePart type="given">Maria</namePart>
            <namePart type="family">Liakata</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Viviane</namePart>
            <namePart type="given">P</namePart>
            <namePart type="family">Moreira</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">Jiajun</namePart>
            <namePart type="family">Zhang</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <name type="personal">
            <namePart type="given">David</namePart>
            <namePart type="family">Jurgens</namePart>
            <role>
                <roleTerm authority="marcrelator" type="text">editor</roleTerm>
            </role>
        </name>
        <originInfo>
            <publisher>Association for Computational Linguistics</publisher>
            <place>
                <placeTerm type="text">San Diego, California, United States</placeTerm>
            </place>
        </originInfo>
        <genre authority="marcgt">conference publication</genre>
        <identifier type="isbn">979-8-89176-395-1</identifier>
    </relatedItem>
    <abstract>Large language models require continuous adaptation to new tasks while preserving safety alignment. However, fine-tuning on even benign data often compromises safety behaviors. We investigate which training samples cause alignment drift through a data-centric lens. Our experiments show samples contribute unequally: high-gradient samples cause greater safety degradation and drive models toward pretrained distributions, while moderate-gradient samples enable task learning with minimal alignment loss. This connects to the elasticity phenomenon—high-gradient samples activate the reversion force pulling models toward pretrained behavior. We propose gradient-based sample selection that filters high-gradient samples during fine-tuning. Across multiple model families on continual domain tasks, our method substantially improves alignment preservation while maintaining competitive task performance, without requiring curated safe data or architectural modifications.</abstract>
    <identifier type="citekey">bach-etal-2026-continual</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-acl.942/</url>
    </location>
    <part>
        <date>2026-07</date>
        <extent unit="page">
            <start>18870</start>
            <end>18887</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Continual Safety Alignment via Gradient-Based Sample Selection
%A Bach, Thong
%A Nguyen, Dung
%A Le, Thao Minh
%A Tran, Truyen
%Y Liakata, Maria
%Y Moreira, Viviane P.
%Y Zhang, Jiajun
%Y Jurgens, David
%S Findings of the Association for Computational Linguistics: ACL 2026
%D 2026
%8 July
%I Association for Computational Linguistics
%C San Diego, California, United States
%@ 979-8-89176-395-1
%F bach-etal-2026-continual
%X Large language models require continuous adaptation to new tasks while preserving safety alignment. However, fine-tuning on even benign data often compromises safety behaviors. We investigate which training samples cause alignment drift through a data-centric lens. Our experiments show samples contribute unequally: high-gradient samples cause greater safety degradation and drive models toward pretrained distributions, while moderate-gradient samples enable task learning with minimal alignment loss. This connects to the elasticity phenomenon—high-gradient samples activate the reversion force pulling models toward pretrained behavior. We propose gradient-based sample selection that filters high-gradient samples during fine-tuning. Across multiple model families on continual domain tasks, our method substantially improves alignment preservation while maintaining competitive task performance, without requiring curated safe data or architectural modifications.
%U https://aclanthology.org/2026.findings-acl.942/
%P 18870-18887

Markdown (Informal)

[Continual Safety Alignment via Gradient-Based Sample Selection](https://aclanthology.org/2026.findings-acl.942/) (Bach et al., Findings 2026)

• Continual Safety Alignment via Gradient-Based Sample Selection (Bach et al., Findings 2026)

ACL

• Thong Bach, Dung Nguyen, Thao Minh Le, and Truyen Tran. 2026. Continual Safety Alignment via Gradient-Based Sample Selection. In Findings of the Association for Computational Linguistics: ACL 2026, pages 18870–18887, San Diego, California, United States. Association for Computational Linguistics.