LaCoMSA: Language-Consistency Multilingual Self-Alignment with latent representation rewarding

Simple item page
dc.contributor.authorTran, Khanh Tung
dc.contributor.authorO’Sullivan, Barry
dc.contributor.authorNguyen, Hoang D.
dc.contributor.editorDemberg, Vera
dc.contributor.editorInui, Kentaro
dc.contributor.editorMarquez Villodre, Lluis
dc.date.accessioned2026-07-08T10:50:01Z
dc.date.available2026-07-08T10:50:01Z
dc.date.issued2026-03-29
dc.description.abstractLarge Language Models (LLMs) have achieved impressive performance yet remain inconsistent across languages, often defaulting to high-resource outputs such as English. Existing multilingual alignment methods mitigate these issues through preference optimization but rely on external supervision, such as translation systems or English-biased signal. We propose Multilingual Self-Alignment (MSA), a targeted preference optimization framework that leverages an LLM’s own latent representations as intrinsic supervision signals, rewarding lower-resource language outputs based on their alignment with high-resource (English) counterparts in the “semantic hub”. We further introduce Language-Consistency MSA (LaCoMSA), which augments MSA with a final-layer language-consistency factor to prevent off-target generation. Integrated with Direct Preference Optimization, LaCoMSA improves a Llama 3 8B-based model multilingual win rates by up to 6.8% absolute (55.0% relatively) on X-AlpacaEval and achieves consistent gains across benchmarks and models. Our findings demonstrate that LaCoMSA can serve as an effective and scalable mechanism, opening a new venue toward multilingual self-alignment.en
dc.description.versionPublished Version
dc.format.extent15
dc.format.mimetypeapplication/pdfen
dc.identifier.authororcidTran, Khanh Tung
dc.identifier.authororcidO’Sullivan, Barry§0000-0002-0090-2085
dc.identifier.authororcidNguyen, Hoang D.§0000-0003-2541-3269
dc.identifier.authororcidDemberg, Vera
dc.identifier.authororcidInui, Kentaro
dc.identifier.authororcidMarquez Villodre, Lluis
dc.identifier.citationTran, K T, O’Sullivan, B & Nguyen, H D 2026, LaCoMSA: Language-Consistency Multilingual Self-Alignment with latent representation rewarding. in V Demberg, K Inui & L Marquez Villodre (eds), Proceedings of the 19th Conference of the European Chapter of the Association for Computational Linguistics (Volume 1: Long Papers) : Rabat, Morocco, 24–29 March 2026. EACL 2026 - 19th Conference of the European Chapter of the Association for Computational Linguistics, Proceedings of the Conference, Vol. 1 - (Long Papers), vol. 1, Association for Computational Linguistics (ACL), pp. 4839-4853, 19th Conference of the European Chapter of the Association for Computational Linguistics, EACL 2026, Rabat, Morocco, 24/03/26. https://doi.org/10.18653/v1/2026.eacl-long.224
dc.identifier.doi10.18653/v1/2026.eacl-long.224
dc.identifier.endpage4853
dc.identifier.isbn9798891763807
dc.identifier.otherORCID: /0000-0003-2541-3269/work/220111399
dc.identifier.startpage4839
dc.identifier.urihttps://hdl.handle.net/10468/19019
dc.language.isoen
dc.publisherAssociation for Computational Linguistics (ACL)
dc.relation.ispartofseriesEACL 2026 - 19th Conference of the European Chapter of the Association for Computational Linguistics, Proceedings of the Conference, Vol. 1 - (Long Papers)
dc.rights© 2026, Association for Computational Linguistics.
dc.rights.accessrightsopen access
dc.rights.licensenameAttribution 4.0 International
dc.rights.urihttps://creativecommons.org/licenses/by/4.0/
dc.statusPeer reviewed
dc.subjectMultilingual Self-Alignment
dc.subjectLarge Language Models (LLMs)
dc.subject[ComputerScience]
dc.titleLaCoMSA: Language-Consistency Multilingual Self-Alignment with latent representation rewardingen
dc.typeConference item
Files
Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
2026.eacl-long.224.pdf
Size:
2.48 MB
Format:
Adobe Portable Document Format
Download
Collections
School of Computer Science and Information Technology