@inproceedings{li-lee-2026-generalizability,
title = "Generalizability Theory for Evaluating Fine-Tuned {LLM}s in Automated Item Generation",
author = "Li, Zhifei and
Lee, Won-Chan",
editor = "Wilson, Joshua and
Ormerod, Christopher and
Beiting-Parrish, Magdalen",
booktitle = "Proceedings of the Artificial Intelligence in Measurement and Education Conference ({AIME}-Con): Coordinated Session Papers",
month = oct,
year = "2026",
address = "Wyndham Grand Pittsburgh Downtown, Pittsburgh, Pennsylvania, United States",
publisher = "National Council on Measurement in Education (NCME)",
url = "https://aclanthology.org/2026.aimecon-sessions.7/",
pages = "59--69",
ISBN = "979-8-9983004-2-4",
abstract = "Automated item generation with large language models (LLMs) is typically evaluated using aggregate accuracy metrics that conflate the quality of the source passage with noise in- troduced by the generation process itself. We address this gap by embedding a prospective i : (p$\times$s$\times$t) Generalizability Theory (G- theory) design into the evaluation of a QLoRA fine-tuned Qwen2.5-7B-Instruct model on the SciQ corpus. Passages (p) serve as the object of measurement; random seeds (s) and prompt templates (t) are fully crossed facets; items are nested within each (p,s,t) cell. Across 9,000 observations and five binary quality met- rics, we find that 77{--}81{\%} of total variance is attributable to the passage, seed and template main effects are negligible ($\leq$0.02{\%}), and G- coefficients (E $\rho$2) uniformly exceed 0.97 un- der the observed design. D-study projections show that a single seed and template already achieves E $\rho$2 = 0.87, while the observed de- sign (ns = 3, nt = 3, ni = 2) reaches 0.98. Code, data, and R analysis scripts are released to support reproducible psychometric evalua- tion of future item-generation systems."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="li-lee-2026-generalizability">
<titleInfo>
<title>Generalizability Theory for Evaluating Fine-Tuned LLMs in Automated Item Generation</title>
</titleInfo>
<name type="personal">
<namePart type="given">Zhifei</namePart>
<namePart type="family">Li</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Won-Chan</namePart>
<namePart type="family">Lee</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 Artificial Intelligence in Measurement and Education Conference (AIME-Con): Coordinated Session Papers</title>
</titleInfo>
<name type="personal">
<namePart type="given">Joshua</namePart>
<namePart type="family">Wilson</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Christopher</namePart>
<namePart type="family">Ormerod</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Magdalen</namePart>
<namePart type="family">Beiting-Parrish</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>National Council on Measurement in Education (NCME)</publisher>
<place>
<placeTerm type="text">Wyndham Grand Pittsburgh Downtown, Pittsburgh, Pennsylvania, United States</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
<identifier type="isbn">979-8-9983004-2-4</identifier>
</relatedItem>
<abstract>Automated item generation with large language models (LLMs) is typically evaluated using aggregate accuracy metrics that conflate the quality of the source passage with noise in- troduced by the generation process itself. We address this gap by embedding a prospective i : (p\timess\timest) Generalizability Theory (G- theory) design into the evaluation of a QLoRA fine-tuned Qwen2.5-7B-Instruct model on the SciQ corpus. Passages (p) serve as the object of measurement; random seeds (s) and prompt templates (t) are fully crossed facets; items are nested within each (p,s,t) cell. Across 9,000 observations and five binary quality met- rics, we find that 77–81% of total variance is attributable to the passage, seed and template main effects are negligible (łeq0.02%), and G- coefficients (E ρ2) uniformly exceed 0.97 un- der the observed design. D-study projections show that a single seed and template already achieves E ρ2 = 0.87, while the observed de- sign (ns = 3, nt = 3, ni = 2) reaches 0.98. Code, data, and R analysis scripts are released to support reproducible psychometric evalua- tion of future item-generation systems.</abstract>
<identifier type="citekey">li-lee-2026-generalizability</identifier>
<location>
<url>https://aclanthology.org/2026.aimecon-sessions.7/</url>
</location>
<part>
<date>2026-10</date>
<extent unit="page">
<start>59</start>
<end>69</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Generalizability Theory for Evaluating Fine-Tuned LLMs in Automated Item Generation
%A Li, Zhifei
%A Lee, Won-Chan
%Y Wilson, Joshua
%Y Ormerod, Christopher
%Y Beiting-Parrish, Magdalen
%S Proceedings of the Artificial Intelligence in Measurement and Education Conference (AIME-Con): Coordinated Session Papers
%D 2026
%8 October
%I National Council on Measurement in Education (NCME)
%C Wyndham Grand Pittsburgh Downtown, Pittsburgh, Pennsylvania, United States
%@ 979-8-9983004-2-4
%F li-lee-2026-generalizability
%X Automated item generation with large language models (LLMs) is typically evaluated using aggregate accuracy metrics that conflate the quality of the source passage with noise in- troduced by the generation process itself. We address this gap by embedding a prospective i : (p\timess\timest) Generalizability Theory (G- theory) design into the evaluation of a QLoRA fine-tuned Qwen2.5-7B-Instruct model on the SciQ corpus. Passages (p) serve as the object of measurement; random seeds (s) and prompt templates (t) are fully crossed facets; items are nested within each (p,s,t) cell. Across 9,000 observations and five binary quality met- rics, we find that 77–81% of total variance is attributable to the passage, seed and template main effects are negligible (łeq0.02%), and G- coefficients (E ρ2) uniformly exceed 0.97 un- der the observed design. D-study projections show that a single seed and template already achieves E ρ2 = 0.87, while the observed de- sign (ns = 3, nt = 3, ni = 2) reaches 0.98. Code, data, and R analysis scripts are released to support reproducible psychometric evalua- tion of future item-generation systems.
%U https://aclanthology.org/2026.aimecon-sessions.7/
%P 59-69
Markdown (Informal)
[Generalizability Theory for Evaluating Fine-Tuned LLMs in Automated Item Generation](https://aclanthology.org/2026.aimecon-sessions.7/) (Li & Lee, AIME-Con 2026)
ACL