<?xml version="1.0" encoding="utf-8"?>
<export-example>
  <doc>
    <id>8053</id>
    <completedYear/>
    <publishedYear>2025</publishedYear>
    <thesisYearAccepted/>
    <language>eng</language>
    <pageFirst/>
    <pageLast/>
    <pageNumber>8</pageNumber>
    <edition/>
    <issue/>
    <volume/>
    <type>conferenceobject</type>
    <publisherName/>
    <publisherPlace/>
    <creatingCorporation/>
    <contributingCorporation/>
    <belongsToBibliography>0</belongsToBibliography>
    <completedDate>2025-04-27</completedDate>
    <publishedDate>--</publishedDate>
    <thesisDateAccepted>--</thesisDateAccepted>
    <title language="eng">Controlled Diversity: Length-optimized Natural Language Generation</title>
    <abstract language="eng">LLMs are not generally able to adjust the length of their outputs based on strict length requirements, a capability that would improve their usefulness in applications that require adherence to diverse user and system requirements. We present an approach to train LLMs to acquire this capability by augmenting existing data and applying existing fine-tuning techniques, which we compare based on the trained models’ adherence to the length requirement and overall response quality relative to the baseline model. Our results demonstrate that these techniques can be successfully applied to train LLMs to adhere to length requirements, with the trained models generating texts which better align to the length requirements. Our results indicate that our method may change the response quality when using training data that was not generated by the baseline model. This allows simultaneous alignment to another training objective in certain scenarios, but is undesirable otherwise. Training on a dataset containing the model’s own responses eliminates this issue.</abstract>
    <parentTitle language="eng">Proceedings of the ISCA/ITG Workshop on Diversity in Large Speech and Language Models, February 20, 2025 in Berlin, Germany</parentTitle>
    <identifier type="arxiv">arXiv:2502.19347</identifier>
    <identifier type="doi">10.48550/arXiv.2502.19347</identifier>
    <identifier type="urn">urn:nbn:de:bvb:898-opus4-80534</identifier>
    <enrichment key="BegutachtungStatus">peer-reviewed</enrichment>
    <enrichment key="opus.source">publish</enrichment>
    <licence>Creative Commons - CC BY - Namensnennung 4.0 International</licence>
    <author>Diana Marie Schenke</author>
    <author>Timo Baumann</author>
    <collection role="institutes" number="FakIM">Fakultät Informatik und Mathematik</collection>
    <collection role="othforschungsschwerpunkt" number="16315">Information und Kommunikation</collection>
    <collection role="DFGFachsystematik" number="1">Ingenieurwissenschaften</collection>
    <thesisPublisher>Ostbayerische Technische Hochschule Regensburg</thesisPublisher>
    <file>https://opus4.kobv.de/opus4-oth-regensburg/files/8053/2502.19347v1.pdf</file>
  </doc>
</export-example>
