Publication:

From Intent to Manifest: Evaluation of LLM-Driven Generation of Custom Resources in Kubernetes

 
cris.virtual.department#PLACEHOLDER_PARENT_METADATA_VALUE#
cris.virtual.department#PLACEHOLDER_PARENT_METADATA_VALUE#
cris.virtual.orcid0000-0003-4824-1199
cris.virtual.orcid0000-0002-6276-2057
cris.virtualsource.department505a9fa2-2261-4859-8c77-73c2ba21244c
cris.virtualsource.department5fc1041b-34c9-4bdb-ba41-1a986f0c4c25
cris.virtualsource.orcid505a9fa2-2261-4859-8c77-73c2ba21244c
cris.virtualsource.orcid5fc1041b-34c9-4bdb-ba41-1a986f0c4c25
dc.contributor.authorFerrari, Lucia
dc.contributor.authorZaccarini, Mattia
dc.contributor.authorPereira dos Santos, José Pedro
dc.contributor.authorPoltronieri, Filippo
dc.contributor.authorTortonesi, Mauro
dc.contributor.authorStefanelli, Cesare
dc.contributor.authorDe Turck, Filip
dc.date.accessioned2026-11-17T10:39:02Z
dc.date.available2026-11-17T10:39:02Z
dc.date.createdwos2026
dc.date.issued2026
dc.description.abstractLarge Language Models (LLMs) have recently emerged as powerful enablers of intent-driven networking and automated service management in 5G and beyond. The aim of this work is to use LLMs to translate high-level intents into reconfiguration commands, operational support actions, or deployment decisions. Recent works have highlighted the feasibility of LLMdriven management and configuration, but these approaches operate primarily at the control and management plane level. The translation of orchestration intents into valid, system-specific configuration files remains an open challenge, particularly in cloudnative environments where declarative manifests are essential for automation. Therefore, this paper investigates the use of LLMs for generating Diktyo AppGroup Custom Resources (CRs) in Kubernetes (K8s) environments. Diktyo enables network-aware placement via grouping of workloads in K8s, but its adoption is hindered by the complexity and verbosity of the AppGroup configuration schema. Our approach leverages natural language and structured intents provided by cloud developers or operators, which are translated by an LLM into syntactically correct and meaningful AppGroups specifications. Results show that LLMs can bridge the gap between user-level orchestration intents and deployable AppGroup specifications, ensuring schema compliance and improving deployment efficiency. Models such as qwen2.5-coder:14b and gpt-oss:20b can correctly generate AppGroup CRs in most cases, achieving between 93 −100% accuracy, significantly reducing the cognitive load for developers. This work advances the vision of an autonomous, intentdriven cloud-native orchestration, by extending the paradigm into network-aware orchestration in K8s ecosystems.
dc.description.wosFundingTextThis work has been partially supported by the Spoke 1 "FutureHPC & BigData" of the Italian Research Center on HighPerformance Computing, Big Data and Quantum Computing (ICSC) funded by MUR Missione 4 -Next Generation EU (NGEU). Jose Santos is funded by the Research Foundation Flanders (FWO), grant number 1253226N.
dc.identifier.doi10.1109/noms69089.2026.11668123
dc.identifier.isbn979-8-3315-9269-1
dc.identifier.issn1542-1201
dc.identifier.urihttps://imec-publications.be/handle/20.500.12860/60835
dc.language.isoeng
dc.provenance.editstepusergreet.vanhoof@imec.be
dc.publisherIEEE
dc.relation.ispartofseriesIEEE IFIP Network Operations and Management Symposium
dc.source.beginpage1
dc.source.conferenceIEEE Network Operations and Management Symposium - NOMS
dc.source.conferencedate2026-05-18
dc.source.conferencelocationRome, Italy
dc.source.endpage9
dc.source.journalNOMS 2026-2026 IEEE NETWORK OPERATIONS AND MANAGEMENT SYMPOSIUM
dc.source.numberofpages9
dc.title

From Intent to Manifest: Evaluation of LLM-Driven Generation of Custom Resources in Kubernetes

dc.typeProceedings paper
dspace.entity.typePublication
imec.internal.crawledAt2026-09-03
imec.internal.sourcecrawler
imec.internal.wosCreatedAt2026-11-11
Files

Original bundle

Name:
9087.pdf
Size:
822.65 KB
Format:
Adobe Portable Document Format
Description:
Accepted
Publication available in collections: