<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lin-etal-2026-prism">
    <titleInfo>
        <title>PRISM: Preference-aware Influence function based Data Selection Method for Efficient Fine-Tuning</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Qihao</namePart>
        <namePart type="family">Lin</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Guanxu</namePart>
        <namePart type="family">Chen</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Dongrui</namePart>
        <namePart type="family">Liu</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Jing</namePart>
        <namePart type="family">Shao</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>As LLMs scale, improving training efficiency relies on effective data utilization. Data selection mitigates this issue by allocating limited data budget to examples that optimally facilitate target behavior. Most existing approaches define target behavior via a set of target examples and score candidate training examples based on estimated influence on the target set. However, such methods treat target examples as equally important, ignoring their relevance to model optimization. Target examples that align closely with the model’s inherent behavior deliver stronger supervisory signals, whereas discrepant examples yield weak local guidance. We propose PRISM, a Preference-aware Influence function based Data Selection Method. It uses model preference to weight examples and build a preference-aware target direction. PRISM evaluates candidate training examples according to influence on this direction, and prioritizes data budget allocation to examples that drive the model to match target behavior. Theoretical analysis shows that preference weighting yields a superior first-order direction for boosting target preference, relative to equal aggregation. Experiments across model architectures and scales show that PRISM improves data selection in both capability improvement and reverse repair. The results validate that accurate characterization of target behavior serves as the core of cost-effective data selection.</abstract>
    <identifier type="citekey">lin-etal-2026-prism</identifier>
    <location>
        <url>https://aclanthology.org/2026.emnlp-main.2026/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>46150</start>
            <end>46169</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T PRISM: Preference-aware Influence function based Data Selection Method for Efficient Fine-Tuning
%A Lin, Qihao
%A Chen, Guanxu
%A Liu, Dongrui
%A Shao, Jing
%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 lin-etal-2026-prism
%X As LLMs scale, improving training efficiency relies on effective data utilization. Data selection mitigates this issue by allocating limited data budget to examples that optimally facilitate target behavior. Most existing approaches define target behavior via a set of target examples and score candidate training examples based on estimated influence on the target set. However, such methods treat target examples as equally important, ignoring their relevance to model optimization. Target examples that align closely with the model’s inherent behavior deliver stronger supervisory signals, whereas discrepant examples yield weak local guidance. We propose PRISM, a Preference-aware Influence function based Data Selection Method. It uses model preference to weight examples and build a preference-aware target direction. PRISM evaluates candidate training examples according to influence on this direction, and prioritizes data budget allocation to examples that drive the model to match target behavior. Theoretical analysis shows that preference weighting yields a superior first-order direction for boosting target preference, relative to equal aggregation. Experiments across model architectures and scales show that PRISM improves data selection in both capability improvement and reverse repair. The results validate that accurate characterization of target behavior serves as the core of cost-effective data selection.
%U https://aclanthology.org/2026.emnlp-main.2026/
%P 46150-46169

Markdown (Informal)

[PRISM: Preference-aware Influence function based Data Selection Method for Efficient Fine-Tuning](https://aclanthology.org/2026.emnlp-main.2026/) (Lin et al., EMNLP 2026)

• PRISM: Preference-aware Influence function based Data Selection Method for Efficient Fine-Tuning (Lin et al., EMNLP 2026)

ACL

• Qihao Lin, Guanxu Chen, Dongrui Liu, and Jing Shao. 2026. PRISM: Preference-aware Influence function based Data Selection Method for Efficient Fine-Tuning. In Proceedings of the 2026 Conference on Empirical Methods in Natural Language Processing, pages 46150–46169, Budapest, Hungary. Association for Computational Linguistics.