<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="habibi-etal-2026-symbolic">
    <titleInfo>
        <title>Symbolic Mechanistic Data Attribution: Tracing Training Influence to Learned Behavioral Policies</title>
    </titleInfo>
    <name type="personal">
        <namePart type="given">Reza</namePart>
        <namePart type="family">Habibi</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Darian</namePart>
        <namePart type="family">Lee</namePart>
        <role>
            <roleTerm authority="marcrelator" type="text">author</roleTerm>
        </role>
    </name>
    <name type="personal">
        <namePart type="given">Magy</namePart>
        <namePart type="given">Seif</namePart>
        <namePart type="family">El-Nasr</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>Findings of the Association for Computational Linguistics: EMNLP 2026</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">979-8-89176-462-0</identifier>
    </relatedItem>
    <abstract>While existing data attribution methods can identify which training examples build specific mechanistic circuits, they cannot explain how training data shapes the behavioral decisions a model learns. We introduce Symbolic Mechanistic Data Attribution (SMDA), a framework that attributes training pairs to the interpretable symbolic policies governing model behavior. SMDA fits a closed-form Ridge regression over sparse autoencoder (SAE) features to model a behavior, then decomposes how each supervised fine-tuning (SFT) example shifts that policy through feature-activation (\ensuremathΔX) and output-probability (\ensuremathΔY) pathways. We distill a symbolic policy for refusal in Llama-3.2-3B-Instruct and analyze 200 SFT pairs. Our analysis reveals (1) the symbolic policy’s coefficients expose systematic gaps in the base model’s safety behavior, such as religious stereotyping; (2) per-feature \ensuremathΔX/\ensuremathΔY decomposition mechanistically explains why harmful and harmless pairs exert qualitatively different influences on certain features; and (3) individual pairs routinely exhibit cross-feature interference, allowing SMDA to identify pairs whose dominant effect falls on unintended features. Retraining without such pairs improved accuracy by up to 4.5 percentage points over full fine-tuning. These results show that combining mechanistic interpretability with data attribution yields a diagnostic tool both more fine-grained than black-box influence functions and more scalable than manual circuit analysis.</abstract>
    <identifier type="citekey">habibi-etal-2026-symbolic</identifier>
    <location>
        <url>https://aclanthology.org/2026.findings-emnlp.1671/</url>
    </location>
    <part>
        <date>2026-10</date>
        <extent unit="page">
            <start>35629</start>
            <end>35661</end>
        </extent>
    </part>
</mods>
</modsCollection>

%0 Conference Proceedings
%T Symbolic Mechanistic Data Attribution: Tracing Training Influence to Learned Behavioral Policies
%A Habibi, Reza
%A Lee, Darian
%A El-Nasr, Magy Seif
%Y Lee, Hung-Yi
%Y Dev, Sunipa
%Y Li, Jessy
%Y Elliott, Desmond
%S Findings of the Association for Computational Linguistics: EMNLP 2026
%D 2026
%8 October
%I Association for Computational Linguistics
%C Budapest, Hungary
%@ 979-8-89176-462-0
%F habibi-etal-2026-symbolic
%X While existing data attribution methods can identify which training examples build specific mechanistic circuits, they cannot explain how training data shapes the behavioral decisions a model learns. We introduce Symbolic Mechanistic Data Attribution (SMDA), a framework that attributes training pairs to the interpretable symbolic policies governing model behavior. SMDA fits a closed-form Ridge regression over sparse autoencoder (SAE) features to model a behavior, then decomposes how each supervised fine-tuning (SFT) example shifts that policy through feature-activation (\ensuremathΔX) and output-probability (\ensuremathΔY) pathways. We distill a symbolic policy for refusal in Llama-3.2-3B-Instruct and analyze 200 SFT pairs. Our analysis reveals (1) the symbolic policy’s coefficients expose systematic gaps in the base model’s safety behavior, such as religious stereotyping; (2) per-feature \ensuremathΔX/\ensuremathΔY decomposition mechanistically explains why harmful and harmless pairs exert qualitatively different influences on certain features; and (3) individual pairs routinely exhibit cross-feature interference, allowing SMDA to identify pairs whose dominant effect falls on unintended features. Retraining without such pairs improved accuracy by up to 4.5 percentage points over full fine-tuning. These results show that combining mechanistic interpretability with data attribution yields a diagnostic tool both more fine-grained than black-box influence functions and more scalable than manual circuit analysis.
%U https://aclanthology.org/2026.findings-emnlp.1671/
%P 35629-35661

Markdown (Informal)

[Symbolic Mechanistic Data Attribution: Tracing Training Influence to Learned Behavioral Policies](https://aclanthology.org/2026.findings-emnlp.1671/) (Habibi et al., Findings 2026)

• Symbolic Mechanistic Data Attribution: Tracing Training Influence to Learned Behavioral Policies (Habibi et al., Findings 2026)

ACL

• Reza Habibi, Darian Lee, and Magy Seif El-Nasr. 2026. Symbolic Mechanistic Data Attribution: Tracing Training Influence to Learned Behavioral Policies. In Findings of the Association for Computational Linguistics: EMNLP 2026, pages 35629–35661, Budapest, Hungary. Association for Computational Linguistics.