@inproceedings{datta-etal-2025-exploring,
title = "Exploring the Limits of Model Compression in {LLM}s: A Knowledge Distillation Study on {QA} Tasks",
author = "Datta, Joyeeta and
Doll, Niclas and
Ramadan, Qusai and
Boukhers, Zeyd",
editor = "B{\'e}chet, Fr{\'e}d{\'e}ric and
Lef{\`e}vre, Fabrice and
Asher, Nicholas and
Kim, Seokhwan and
Merlin, Teva",
booktitle = "Proceedings of the 26th Annual Meeting of the Special Interest Group on Discourse and Dialogue",
month = aug,
year = "2025",
address = "Avignon, France",
publisher = "Association for Computational Linguistics",
url = "https://aclanthology.org/2025.sigdial-1.39/",
pages = "477--483",
abstract = "Large Language Models (LLMs) have shown outstanding performance across a range of NLP tasks, but their computational demands hinder deployment in real-world, resource-constrained environments. This work investigates the extent to which LLMs can be compressed using knowledge distillation (KD) while maintaining strong performance on question answering (QA) tasks. We evaluate student models distilled from the Pythia and Qwen2.5 families on two QA benchmarks, SQuAD and MLQA, under zero-shot and one-shot prompting conditions. Results show that student models retain over 90{\%} of their teacher models' performance while reducing parameter counts by up to 57.1{\%}. Furthermore, one-shot prompting yields additional performance gains over zero-shot setups for both model families. These findings underscore the trade-off between model efficiency and task performance, demonstrating that KD, combined with minimal prompting, can yield compact yet capable QA systems suitable for real-world applications."
}<?xml version="1.0" encoding="UTF-8"?>
<modsCollection xmlns="http://www.loc.gov/mods/v3">
<mods ID="datta-etal-2025-exploring">
<titleInfo>
<title>Exploring the Limits of Model Compression in LLMs: A Knowledge Distillation Study on QA Tasks</title>
</titleInfo>
<name type="personal">
<namePart type="given">Joyeeta</namePart>
<namePart type="family">Datta</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Niclas</namePart>
<namePart type="family">Doll</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Qusai</namePart>
<namePart type="family">Ramadan</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Zeyd</namePart>
<namePart type="family">Boukhers</namePart>
<role>
<roleTerm authority="marcrelator" type="text">author</roleTerm>
</role>
</name>
<originInfo>
<dateIssued>2025-08</dateIssued>
</originInfo>
<typeOfResource>text</typeOfResource>
<relatedItem type="host">
<titleInfo>
<title>Proceedings of the 26th Annual Meeting of the Special Interest Group on Discourse and Dialogue</title>
</titleInfo>
<name type="personal">
<namePart type="given">Frédéric</namePart>
<namePart type="family">Béchet</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Fabrice</namePart>
<namePart type="family">Lefèvre</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Nicholas</namePart>
<namePart type="family">Asher</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Seokhwan</namePart>
<namePart type="family">Kim</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<name type="personal">
<namePart type="given">Teva</namePart>
<namePart type="family">Merlin</namePart>
<role>
<roleTerm authority="marcrelator" type="text">editor</roleTerm>
</role>
</name>
<originInfo>
<publisher>Association for Computational Linguistics</publisher>
<place>
<placeTerm type="text">Avignon, France</placeTerm>
</place>
</originInfo>
<genre authority="marcgt">conference publication</genre>
</relatedItem>
<abstract>Large Language Models (LLMs) have shown outstanding performance across a range of NLP tasks, but their computational demands hinder deployment in real-world, resource-constrained environments. This work investigates the extent to which LLMs can be compressed using knowledge distillation (KD) while maintaining strong performance on question answering (QA) tasks. We evaluate student models distilled from the Pythia and Qwen2.5 families on two QA benchmarks, SQuAD and MLQA, under zero-shot and one-shot prompting conditions. Results show that student models retain over 90% of their teacher models’ performance while reducing parameter counts by up to 57.1%. Furthermore, one-shot prompting yields additional performance gains over zero-shot setups for both model families. These findings underscore the trade-off between model efficiency and task performance, demonstrating that KD, combined with minimal prompting, can yield compact yet capable QA systems suitable for real-world applications.</abstract>
<identifier type="citekey">datta-etal-2025-exploring</identifier>
<location>
<url>https://aclanthology.org/2025.sigdial-1.39/</url>
</location>
<part>
<date>2025-08</date>
<extent unit="page">
<start>477</start>
<end>483</end>
</extent>
</part>
</mods>
</modsCollection>
%0 Conference Proceedings
%T Exploring the Limits of Model Compression in LLMs: A Knowledge Distillation Study on QA Tasks
%A Datta, Joyeeta
%A Doll, Niclas
%A Ramadan, Qusai
%A Boukhers, Zeyd
%Y Béchet, Frédéric
%Y Lefèvre, Fabrice
%Y Asher, Nicholas
%Y Kim, Seokhwan
%Y Merlin, Teva
%S Proceedings of the 26th Annual Meeting of the Special Interest Group on Discourse and Dialogue
%D 2025
%8 August
%I Association for Computational Linguistics
%C Avignon, France
%F datta-etal-2025-exploring
%X Large Language Models (LLMs) have shown outstanding performance across a range of NLP tasks, but their computational demands hinder deployment in real-world, resource-constrained environments. This work investigates the extent to which LLMs can be compressed using knowledge distillation (KD) while maintaining strong performance on question answering (QA) tasks. We evaluate student models distilled from the Pythia and Qwen2.5 families on two QA benchmarks, SQuAD and MLQA, under zero-shot and one-shot prompting conditions. Results show that student models retain over 90% of their teacher models’ performance while reducing parameter counts by up to 57.1%. Furthermore, one-shot prompting yields additional performance gains over zero-shot setups for both model families. These findings underscore the trade-off between model efficiency and task performance, demonstrating that KD, combined with minimal prompting, can yield compact yet capable QA systems suitable for real-world applications.
%U https://aclanthology.org/2025.sigdial-1.39/
%P 477-483
Markdown (Informal)
[Exploring the Limits of Model Compression in LLMs: A Knowledge Distillation Study on QA Tasks](https://aclanthology.org/2025.sigdial-1.39/) (Datta et al., SIGDIAL 2025)
ACL