Reasoning transfer for an extremely low-resource and endangered language: Bridging languages through sample-efficient language understanding

Simple item page
dc.contributor.authorTran, Khanh-Tungen
dc.contributor.authorO'Sullivan, Barryen
dc.contributor.authorNguyen, Hoang D.en
dc.contributor.editorNguyen, Hoang D.en
dc.contributor.funderResearch Irelanden
dc.contributor.funderEuropean Regional Development Funden
dc.date.accessioned2025-12-18T15:23:12Z
dc.date.available2025-12-18T15:23:12Z
dc.date.issued2025en
dc.description.abstractRecent advances have enabled Large Language Models (LLMs) to tackle reasoning tasks by generating chain-of-thought (CoT) rationales, yet these gains have largely applied to high-resource languages, leaving low-resource languages behind. In this work, we first investigate CoT techniques in extremely low-resource scenarios through previous prompting, model-editing, and fine-tuning approaches. We introduce English-Pivoted CoT Training, leveraging the insight that LLMs internally operate in a latent space aligned toward the dominant language. Given input in a low-resource language, we perform supervised fine-tuning to generate CoT in English and output the final response in the target language. Across mathematical reasoning benchmarks, our approach outperforms other baselines with up to 28.33% improvement in low-resource scenarios. Our analysis and additional experiments, including Mixed-Language CoT and Two-Stage Training, show that explicitly separating language understanding from reasoning enhances cross-lingual reasoning abilities. To facilitate future work, we also release LC2024, the first benchmark for mathematical tasks in Irish, an extremely low-resource and endangered language. Our results and resources highlight a practical pathway to multilingual reasoning without extensive retraining in every extremely low-resource language, despite data scarcity.en
dc.description.sponsorshipResearch Ireland (12-RC-2289-P2)en
dc.description.statusPeer revieweden
dc.description.versionAccepted Versionen
dc.format.mimetypeapplication/pdfen
dc.identifier.citationTran, K.-T., O'Sullivan, B. and Nguyen, H. D. (2025) 'Reasoning transfer for an extremely low-resource and endangered language: Bridging languages through sample-efficient language understanding', 39th Annual AAAI Conference on Artificial Intelligence, Philadelphia, Pennsylvania, USA, 25 February - 4 March 2025.en
dc.identifier.endpage10en
dc.identifier.startpage1en
dc.identifier.urihttps://hdl.handle.net/10468/18355
dc.language.isoenen
dc.relation.ispartof39th Annual AAAI Conference on Artificial Intelligence, Philadelphia, Pennsylvania, USA, 25 February - 4 March 2025.en
dc.relation.projectinfo:eu-repo/grantAgreement/SFI/Centres for Research Training (CRT) Programme/18/CRT/6223/IE/SFI Centre for Research Training in Artificial Intelligence/en
dc.rights© 2025, the Authors. For the purpose of Open Access, a CC BY licence applies to any Author Accepted Manuscript from this submission.en
dc.rights.urihttps://creativecommons.org/licenses/by/4.0/en
dc.subjectReasoning transferen
dc.subjectLarge Language Models (LLMs)en
dc.subjectChain-of-thought (CoT)en
dc.titleReasoning transfer for an extremely low-resource and endangered language: Bridging languages through sample-efficient language understandingen
dc.typeConference itemen
Files
Original bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
2026_AAAI_Reason_Transfering.pdf
Size:
1.08 MB
Format:
Adobe Portable Document Format
Description:
Accepted Version
Download
License bundle
Now showing 1 - 1 of 1
Thumbnail Image
Name:
license.txt
Size:
2.71 KB
Format:
Item-specific license agreed upon to submission
Description:
Download
Collections
Computer Science - Conference Items
ADVANCE CRT - Conference Items
Insight Centre for Data Analytics - Conference Items