@inproceedings{xu-etal-2026-training,
title = "When Does Training Data Matter? Dynamic Data Attribution via Window-level Influence Estimation",
author = "Xu, Jie and
Wu, Zihan and
Pan, Wenbo and
Ming, Yulong and
Jia, Xiaohua",
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.2256/",
pages = "51298--51316",
ISBN = "979-8-89176-459-0",
abstract = "Data attribution methods often compress the effect of each training data point into a single score, computed either at a final checkpoint or by aggregating quantities along the training trajectory. This static view hides how the contribution of a training data point changes during training. To overcome this limitation, we propose the Window-level Influence Estimator (WIE), a framework for estimating the contribution of a training data point zⱼ over a user-specified training window [t{\ensuremath{_1}}, t{\ensuremath{_2}}]. WIE models a leave-one-out-style counterfactual under a first-order linearization of SGD. An influence propagator P[t] = I {\ensuremath{-}} {\ensuremath{\eta}}{\ensuremath{_t}}H[t] at each step models how parameter deviation is amplified or dampened by the loss landscape. Empirically, the window-level estimator reveals four distinct influence patterns: Early Influencer, Late Bloomer, Stable, Highly Fluctuating. Across multiple benchmarks, WIE improves mislabeled-data detection over the considered static and trajectory-based baselines. On BERT-IMDB with 40{\%} label noise, WIE outperforms these baselines across the evaluated inspection ratios. On ViT-EMNIST, WIE reaches 70{\%} recall for mislabeled data while inspecting up to 7.5{\texttimes} fewer candidates than TracIn requires to reach the same recall."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="xu-etal-2026-training">
<titleInfo>
<title>When Does Training Data Matter? Dynamic Data Attribution via Window-level Influence Estimation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Jie</namePart>
<namePart type="family">Xu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zihan</namePart>
<namePart type="family">Wu</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Wenbo</namePart>
<namePart type="family">Pan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Yulong</namePart>
<namePart type="family">Ming</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Xiaohua</namePart>
<namePart type="family">Jia</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>Data attribution methods often compress the effect of each training data point into a single score, computed either at a final checkpoint or by aggregating quantities along the training trajectory. This static view hides how the contribution of a training data point changes during training. To overcome this limitation, we propose the Window-level Influence Estimator (WIE), a framework for estimating the contribution of a training data point zⱼ over a user-specified training window [t\ensuremath₁, t\ensuremath₂]. WIE models a leave-one-out-style counterfactual under a first-order linearization of SGD. An influence propagator P[t] = I \ensuremath- \ensuremathη\ensuremath_tH[t] at each step models how parameter deviation is amplified or dampened by the loss landscape. Empirically, the window-level estimator reveals four distinct influence patterns: Early Influencer, Late Bloomer, Stable, Highly Fluctuating. Across multiple benchmarks, WIE improves mislabeled-data detection over the considered static and trajectory-based baselines. On BERT-IMDB with 40% label noise, WIE outperforms these baselines across the evaluated inspection ratios. On ViT-EMNIST, WIE reaches 70% recall for mislabeled data while inspecting up to 7.5× fewer candidates than TracIn requires to reach the same recall.</abstract>
<identifier type="citekey">xu-etal-2026-training</identifier>
<location>
<url>https://aclanthology.org/2026.emnlp-main.2256/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>51298</start>
<end>51316</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T When Does Training Data Matter? Dynamic Data Attribution via Window-level Influence Estimation
%A Xu, Jie
%A Wu, Zihan
%A Pan, Wenbo
%A Ming, Yulong
%A Jia, Xiaohua
%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 xu-etal-2026-training
%X Data attribution methods often compress the effect of each training data point into a single score, computed either at a final checkpoint or by aggregating quantities along the training trajectory. This static view hides how the contribution of a training data point changes during training. To overcome this limitation, we propose the Window-level Influence Estimator (WIE), a framework for estimating the contribution of a training data point zⱼ over a user-specified training window [t\ensuremath₁, t\ensuremath₂]. WIE models a leave-one-out-style counterfactual under a first-order linearization of SGD. An influence propagator P[t] = I \ensuremath- \ensuremathη\ensuremath_tH[t] at each step models how parameter deviation is amplified or dampened by the loss landscape. Empirically, the window-level estimator reveals four distinct influence patterns: Early Influencer, Late Bloomer, Stable, Highly Fluctuating. Across multiple benchmarks, WIE improves mislabeled-data detection over the considered static and trajectory-based baselines. On BERT-IMDB with 40% label noise, WIE outperforms these baselines across the evaluated inspection ratios. On ViT-EMNIST, WIE reaches 70% recall for mislabeled data while inspecting up to 7.5× fewer candidates than TracIn requires to reach the same recall.
%U https://aclanthology.org/2026.emnlp-main.2256/
%P 51298-51316
Markdown (Informal)
[When Does Training Data Matter? Dynamic Data Attribution via Window-level Influence Estimation](https://aclanthology.org/2026.emnlp-main.2256/) (Xu et al., EMNLP 2026)
ACL