Publication:

Multiobjective Evolutionary Calibration of Uncertainty Quantification for Large Language Models

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.authorMora-Cross, Maria
dc.contributor.authorCalderón-Ramírez, Saul
dc.contributor.authorRojas Gonzalez, Sebastian
dc.date.accessioned2026-11-17T10:29:12Z
dc.date.available2026-11-17T10:29:12Z
dc.date.createdwos2026
dc.date.issued2026
dc.description.abstractLarge language models (LLMs) are increasingly used in high impact applications such as clinical decision support, where fluent outputs can hide critical errors. Uncertainty quantification (UQ) should discriminate correct from incorrect answers and remain informative under dataset shift. We formulate UQ design as a multiobjective evolutionary optimization problem over unsupervised estimators derived from hidden state embeddings. Candidate estimators rely on clustering algorithms followed by a kNN based uncertainty signal. We evaluate Llama 3.1 8B, and DeepSeek 7B on MedQuAD, with a suite of out of distribution datasets (OOD). Objective functions minimize Pearson correlation between UQ and text quality metrics (BERTScore F1 and ROUGE L). Using RVEA, we analyze the search dynamics and the resulting Pareto front of the trade off between semantic and lexical quality. Preliminary results show competitive AUROC and AURC, and improved AUROC relative to a Monte Carlo Dropout (MCD) baseline, highlighting the potential of evolutionary multiobjective optimization as a robust and scalable strategy for building reliable, risk aware generative AI systems.
dc.description.wosFundingTextComputational support was provided by the Kabre supercomputer at the National High Technology Center of Costa Rica (CeNAT). Sebastian Rojas Gonzalez acknowledges support by FWO Belgium (grant 1216021N).
dc.identifier.doi10.1145/3795101.3805422
dc.identifier.isbn979-8-4007-2488-6
dc.identifier.urihttps://imec-publications.be/handle/20.500.12860/60833
dc.language.isoeng
dc.provenance.editstepusergreet.vanhoof@imec.be
dc.publisherASSOC COMPUTING MACHINERY
dc.source.beginpage213
dc.source.conferenceGenetic and Evolutionary Computation Conference Companion - GECCO
dc.source.conferencedate2026-07-13
dc.source.conferencelocationSan Jose, Costa Rica
dc.source.endpage216
dc.source.journalPROCEEDINGS OF THE 2026 GENETIC AND EVOLUTIONARY COMPUTATION CONFERENCE, GECCO 2026 COMPANION
dc.source.numberofpages4
dc.title

Multiobjective Evolutionary Calibration of Uncertainty Quantification for Large Language Models

dc.typeProceedings paper
dspace.entity.typePublication
imec.internal.crawledAt2026-08-14
imec.internal.sourcecrawler
imec.internal.wosCreatedAt2026-11-14
Files

Original bundle

Name:
3795101.3805422.pdf
Size:
949.67 KB
Format:
Adobe Portable Document Format
Description:
Published
Publication available in collections: