LG Uplus announced that a paper on its in‑house generative AI technology, ixi‑GEN, has been accepted to EMNLP 2025, one of the world’s leading conferences in natural language processing (NLP).
EMNLP (Empirical Methods in Natural Language Processing), organized under the Association for Computational Linguistics (ACL), is among the top three global NLP conferences and rigorously evaluates the latest AI research from major tech companies and research institutions.
The submitted paper, titled “ixi‑GEN: Efficient Industrial sLLMs through Domain Adaptive Continual Pretraining,” proposes a new approach that improves both efficiency and quality of small language models (sLLMs). The method—Domain Adaptive Continual Pretraining (DACP)—was recognized for enabling sLLMs to continually learn industry‑specific data while preserving general language capabilities.
Traditionally, sLLMs have faced a trade‑off: fine‑tuning for a specific industry harms general language understanding, while preserving general performance reduces industry relevance. LG Uplus addressed this by designing DACP to balance learning from industry and general data. Applied results showed substantial performance gains in telecommunications and finance compared with existing models, demonstrating that high practical quality is achievable in industry settings without resorting to massive AI models.
The paper’s acceptance is significant both academically and commercially, signaling that LG Uplus’ research meets rigorous scientific standards and offers industrial value. DACP‑based small models could enable companies that cannot operate very large models to deploy industry‑customized AI services without heavy infrastructure and cost burdens. The approach is expected to be useful across data‑intensive and security‑sensitive sectors such as telecom, finance, education, and manufacturing.
LG Uplus has been advancing and validating ixi‑GEN, its proprietary sLLM, for practical use. This year, ixi‑GEN has been used in internal AI platforms and collaborations with industry and education partners to develop specialized sLLMs. In March, ixi‑GEN participated in a Red Team Challenge during MWC25 to validate its reliability. LG Uplus plans to continue researching and applying diverse techniques to further enhance ixi‑GEN.
Han Young‑seop, head of AI Tech Lab at LG Uplus, said, “EMNLP is one of the most watched conferences by AI researchers worldwide, and this acceptance demonstrates that LG Uplus’ industrial AI technology has received global academic validation. We will continue practical AI research to solve real industrial problems and strengthen Korean‑style AI competitiveness.”
