@inproceedings{lago-etal-2026-synthetic,
title = "Synthetic Data Characterization via Training Dynamics",
author = "Lago, Irene and
Ezquerro, Ana and
Vilares, David",
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.1789/",
doi = "10.18653/2026.findings-emnlp.1789",
pages = "38265--38288",
ISBN = "979-8-89176-462-0",
abstract = "Interpreting properties of LLM-generated data is important for understanding its utility and limitations across learning tasks. In this work, we characterize synthetic data through sample-level learnability, studying variation among LLM families and scales, alongside human-written data as a reference. We generate synthetic datasets spanning single- and multi-label classification, labeling, and tree prediction tasks using LLMs from multiple families and scales. We then derive empirical data distributions from encoder training dynamics for both machine and organic data, and estimate the robustness of these distributions across encoders. Finally, we evaluate how data selection strategies based on these learnability signals affect both data sources differently."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="lago-etal-2026-synthetic">
<titleInfo>
<title>Synthetic Data Characterization via Training Dynamics</title>
</titleInfo>
<name type="personal">
<namePart type="given">Irene</namePart>
<namePart type="family">Lago</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Ana</namePart>
<namePart type="family">Ezquerro</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">David</namePart>
<namePart type="family">Vilares</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>Interpreting properties of LLM-generated data is important for understanding its utility and limitations across learning tasks. In this work, we characterize synthetic data through sample-level learnability, studying variation among LLM families and scales, alongside human-written data as a reference. We generate synthetic datasets spanning single- and multi-label classification, labeling, and tree prediction tasks using LLMs from multiple families and scales. We then derive empirical data distributions from encoder training dynamics for both machine and organic data, and estimate the robustness of these distributions across encoders. Finally, we evaluate how data selection strategies based on these learnability signals affect both data sources differently.</abstract>
<identifier type="citekey">lago-etal-2026-synthetic</identifier>
<identifier type="doi">10.18653/2026.findings-emnlp.1789</identifier>
<location>
<url>https://aclanthology.org/2026.findings-emnlp.1789/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>38265</start>
<end>38288</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Synthetic Data Characterization via Training Dynamics
%A Lago, Irene
%A Ezquerro, Ana
%A Vilares, David
%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 lago-etal-2026-synthetic
%X Interpreting properties of LLM-generated data is important for understanding its utility and limitations across learning tasks. In this work, we characterize synthetic data through sample-level learnability, studying variation among LLM families and scales, alongside human-written data as a reference. We generate synthetic datasets spanning single- and multi-label classification, labeling, and tree prediction tasks using LLMs from multiple families and scales. We then derive empirical data distributions from encoder training dynamics for both machine and organic data, and estimate the robustness of these distributions across encoders. Finally, we evaluate how data selection strategies based on these learnability signals affect both data sources differently.
%R 10.18653/2026.findings-emnlp.1789
%U https://aclanthology.org/2026.findings-emnlp.1789/
%U https://doi.org/10.18653/2026.findings-emnlp.1789
%P 38265-38288
Markdown (Informal)
[Synthetic Data Characterization via Training Dynamics](https://aclanthology.org/2026.findings-emnlp.1789/) (Lago et al., Findings 2026)
ACL