@inproceedings{fotouhi-etal-2026-black,
title = "Black-Box Dataset-Level Source Ranking via Context Mixture Factorization",
author = "Fotouhi, Milad and
Bahadori, Mohammad Taha and
Arabshahi, Payman and
Feyisetan, Oluwaseyi",
editor = "Lee, Hung-Yi and
Dev, Sunipa and
Li, Jessy and
Elliott, Desmond",
booktitle = "Findings of the {A}ssociation for {C}omputational {L}inguistics: {EMNLP} 2026",
month = oct,
year = "2026",
address = "Budapest, Hungary",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2026.findings-emnlp.1587/",
doi = "10.18653/2026.findings-emnlp.1587",
pages = "33781--33801",
ISBN = "979-8-89176-462-0",
abstract = "Training data attribution (TDA) for large language models is dominated by gradient- and retraining-based methods that require model internals, ruling out the most widely deployed closed-weight systems. We study TDA in the strictly black-box setting, observing only input output behavior accessed through retrieval- augmented prompting. We introduce Context Mixture Factorization (CMF), which models the LLM{'}s output distribution as a mixture of a base component and dataset-specific components and recovers per-dataset contribution weights by regularized matrix factorization via alternating projected least squares. CMF explicitly decomposes background ({\ensuremath{\pi}}0) and dataset- specific influence and requires only K+1 prompts per query for K candidate datasets. As a comparison method we also formulate the Shapley Context Method (SCM), a residual- similarity attribution score aggregated via Shapley values. We evaluate four instruction-tuned LLMs (GPT-4, Bloomz, Mistral-7B, Phi-3mini) on BoolQ, a synthetic FakeQ variant, and a post-cutoff Olympic2024 dataset against three baselines (TRAK, ContextCite, retrieval cosine similarity). CMF and SCM recover the ordering predicted by our exposure-proxy setup (BoolQ{\ensuremath{>}}FakeQ{\ensuremath{>}}Olympic2024); ContextCite shows the opposite ordering, attributing most to Olympic2024 because the model relies on context most when it has no parametric knowledge {---} supporting the claim that span-level context attribution and dataset-level exposure attribution answer different questions. A synthetic identifiability experiment, controlled fine- tuning sweeps, and a multi-checkpoint unlearning case study provide supporting evidence that the metrics behave as exposure-sensitive source-ranking proxies."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="fotouhi-etal-2026-black">
<titleInfo>
<title>Black-Box Dataset-Level Source Ranking via Context Mixture Factorization</title>
</titleInfo>
<name type="personal">
<namePart type="given">Milad</namePart>
<namePart type="family">Fotouhi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Mohammad</namePart>
<namePart type="given">Taha</namePart>
<namePart type="family">Bahadori</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Payman</namePart>
<namePart type="family">Arabshahi</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Oluwaseyi</namePart>
<namePart type="family">Feyisetan</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>Training data attribution (TDA) for large language models is dominated by gradient- and retraining-based methods that require model internals, ruling out the most widely deployed closed-weight systems. We study TDA in the strictly black-box setting, observing only input output behavior accessed through retrieval- augmented prompting. We introduce Context Mixture Factorization (CMF), which models the LLM’s output distribution as a mixture of a base component and dataset-specific components and recovers per-dataset contribution weights by regularized matrix factorization via alternating projected least squares. CMF explicitly decomposes background (\ensuremathπ0) and dataset- specific influence and requires only K+1 prompts per query for K candidate datasets. As a comparison method we also formulate the Shapley Context Method (SCM), a residual- similarity attribution score aggregated via Shapley values. We evaluate four instruction-tuned LLMs (GPT-4, Bloomz, Mistral-7B, Phi-3mini) on BoolQ, a synthetic FakeQ variant, and a post-cutoff Olympic2024 dataset against three baselines (TRAK, ContextCite, retrieval cosine similarity). CMF and SCM recover the ordering predicted by our exposure-proxy setup (BoolQ\ensuremath>FakeQ\ensuremath>Olympic2024); ContextCite shows the opposite ordering, attributing most to Olympic2024 because the model relies on context most when it has no parametric knowledge — supporting the claim that span-level context attribution and dataset-level exposure attribution answer different questions. A synthetic identifiability experiment, controlled fine- tuning sweeps, and a multi-checkpoint unlearning case study provide supporting evidence that the metrics behave as exposure-sensitive source-ranking proxies.</abstract>
<identifier type="citekey">fotouhi-etal-2026-black</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1587</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1587/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>33781</start>
<end>33801</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Black-Box Dataset-Level Source Ranking via Context Mixture Factorization
%A Fotouhi, Milad
%A Bahadori, Mohammad Taha
%A Arabshahi, Payman
%A Feyisetan, Oluwaseyi
%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 fotouhi-etal-2026-black
%X Training data attribution (TDA) for large language models is dominated by gradient- and retraining-based methods that require model internals, ruling out the most widely deployed closed-weight systems. We study TDA in the strictly black-box setting, observing only input output behavior accessed through retrieval- augmented prompting. We introduce Context Mixture Factorization (CMF), which models the LLM’s output distribution as a mixture of a base component and dataset-specific components and recovers per-dataset contribution weights by regularized matrix factorization via alternating projected least squares. CMF explicitly decomposes background (\ensuremathπ0) and dataset- specific influence and requires only K+1 prompts per query for K candidate datasets. As a comparison method we also formulate the Shapley Context Method (SCM), a residual- similarity attribution score aggregated via Shapley values. We evaluate four instruction-tuned LLMs (GPT-4, Bloomz, Mistral-7B, Phi-3mini) on BoolQ, a synthetic FakeQ variant, and a post-cutoff Olympic2024 dataset against three baselines (TRAK, ContextCite, retrieval cosine similarity). CMF and SCM recover the ordering predicted by our exposure-proxy setup (BoolQ\ensuremath>FakeQ\ensuremath>Olympic2024); ContextCite shows the opposite ordering, attributing most to Olympic2024 because the model relies on context most when it has no parametric knowledge — supporting the claim that span-level context attribution and dataset-level exposure attribution answer different questions. A synthetic identifiability experiment, controlled fine- tuning sweeps, and a multi-checkpoint unlearning case study provide supporting evidence that the metrics behave as exposure-sensitive source-ranking proxies.
%R 10.18653/2026.findings-emnlp.1587
%U https://aclanthology.org/2026.findings-emnlp.1587/
%U https://doi.org/10.18653/2026.findings-emnlp.1587
%P 33781-33801
Markdown (Informal)
[Black-Box Dataset-Level Source Ranking via Context Mixture Factorization](https://aclanthology.org/2026.findings-emnlp.1587/) (Fotouhi et al., Findings 2026)
ACL