Publication:

Evolutionary Optimization of Instruction-Tuned Large Language Model Inference Parameters for Spanish Text Simplification

Date

 
cris.virtual.department#PLACEHOLDER_PARENT_METADATA_VALUE#
cris.virtual.orcid0000-0002-5145-872X
cris.virtualsource.departmenta6e15b57-cd91-48e5-9630-242b1b7129de
cris.virtualsource.orcida6e15b57-cd91-48e5-9630-242b1b7129de
dc.contributor.authorPerez-Rojas, Nelson
dc.contributor.authorXie-Li, Danny
dc.contributor.authorSolis, Martin
dc.contributor.authorCalderon-Ramirez, Saul
dc.contributor.authorRojas Gonzalez, Sebastian
dc.contributor.authorOrtiz-Ruiz, Victor
dc.date.accessioned2026-11-17T10:29:15Z
dc.date.available2026-11-17T10:29:15Z
dc.date.createdwos2026
dc.date.issued2026
dc.description.abstractLarge Language Models (LLM) have become valuable tools for automatic text simplification, yet their output quality is highly sensitive to inference-time parameters such as temperature, repetition penalty, top-p, and top-k. These parameters are typically set heuristically rather than systematically optimized. In this work, we apply Covariance Matrix Adaptation Evolution Strategy (CMA-ES) to identify high-performing inference configurations for an instruction-tuned Large Language Model (LLM) fine-tuned for Spanish Text Simplification (TS) on the Financial Education Corpus IN SpAnish (FEINA) benchmark. Using System output Against References and against the Input sentence (SARI) as the optimization objective and Sentence Bidirectional Encoder Representations from Transformers (SBERT) as a monitoring metric, CMA-ES converges in 9 of 25 generations, improving test-set SARI from 36.57 (default) to 38.87 (+6.3%) while increasing SBERT from 0.85 to 0.88. Grid search analysis reveals that repetition penalty and temperature are the dominant factors influencing simplification quality, while top-k and maximum tokens have negligible effects. The strong positive correlation between SARI and SBERT across all 2,500 evaluated configurations provides empirical support for the single-objective approach within the studied setting. Our framework offers a replicable methodology for inference parameter optimization applicable to Spanish and other under-resourced languages in Latin American Natural Language Processing (NLP) contexts.
dc.description.wosFundingTextThe authors express their gratitude to the Escuela de Ciencias del Lenguaje and the Vicerrectoria de Investigacion y Extension at the Instituto Tecnologico de Costa Rica (TEC). This research was conducted under the framework of project ID 1470018, titled "Evaluacion de modelos de inteligencia artificial en simplificacion automatica de textos en espanol para personas con discapacidades visuales." Additionally, the authors acknowledge the computational resources provided through a machine allocation on the Kabre supercomputer at the Costa Rica National High Technology Center and the Scientific Cluster of the University of Costa Rica, which contributed to the completion of this work.
dc.identifier.doi10.1145/3795101.3814720
dc.identifier.isbn979-8-4007-2488-6
dc.identifier.urihttps://imec-publications.be/handle/20.500.12860/60834
dc.language.isoeng
dc.provenance.editstepusergreet.vanhoof@imec.be
dc.publisherASSOC COMPUTING MACHINERY
dc.source.beginpage1465
dc.source.conferenceGenetic and Evolutionary Computation Conference Companion - GECCO
dc.source.conferencedate2026-07-13
dc.source.conferencelocationSan Jose, Costa Rica
dc.source.endpage1473
dc.source.journalPROCEEDINGS OF THE 2026 GENETIC AND EVOLUTIONARY COMPUTATION CONFERENCE, GECCO 2026 COMPANION
dc.source.numberofpages9
dc.title

Evolutionary Optimization of Instruction-Tuned Large Language Model Inference Parameters for Spanish Text Simplification

dc.typeProceedings paper
dspace.entity.typePublication
imec.internal.crawledAt2026-11-14
imec.internal.sourcecrawler
imec.internal.wosCreatedAt2026-11-14
Files

Original bundle

Name:
3795101.3814720.pdf
Size:
768.17 KB
Format:
Adobe Portable Document Format
Description:
Published
Publication available in collections: