LG Uplus said on the 23rd that a paper related to its in-house developed Generative AI technology "ixi-GEN" was accepted by EMNLP 2025, one of the world's top natural language processing (NLP) conferences.
EMNLP (Empirical Methods in Natural Language Processing) is one of the world's top three NLP conferences organized under the Association for Computational Linguistics (ACL), rigorously reviewing and evaluating the latest AI research results from global big tech corporations and research institutions.
The paper submitted by LG Uplus, titled "ixi-GEN: Efficient Industrial sLLMs through Domain Adaptive Continual Pretraining," presents a new approach that simultaneously improves the efficiency and quality of small language models (sLLMs). By leveraging a "domain-adaptive continual pretraining (DACP)" technique, it devises a method to continually learn industrial data while maintaining general language capability.
Existing sLLM models faced limitations where industrially specialized training reduced their understanding of general language, or maintaining general performance lowered industrial suitability. In response, LG Uplus proposed a method to learn both in balance, significantly improving performance over existing models in the telecommunications and finance sectors.
The acceptance of this paper is significant in that LG Uplus has been recognized for high completeness both academically and industrially. Even corporations that find it difficult to operate ultra-large models are expected to implement industry-tailored AI services without expense and infrastructure burdens by utilizing DACP-based small models.
LG Uplus plans to continue diverse technology research and applications to further advance the performance of ixi-GEN.