SK Telecom earned recognition for research on an AI model that understands personal behavioural context and explains why it makes recommendations, securing a slot at a leading AI conference.

SKT said on Jan. 25 that a research paper on its AI recommendation model was invited to the international AI conference AAAI 2026, being held at Singapore Expo, and that it gave an on-stage presentation on Jan. 24.

AAAI is assessed as among the world’s top AI conferences, along with NeurIPS, ICML and ICLR. About 24,000 papers were submitted to AAAI this year. While the number rose from a year earlier, the acceptance rate fell to 18 percent, intensifying competition.

SKT’s paper was selected for an on-site presentation, given to about the top 4 percent of submitted papers. A poster presentation was also held as part of the conference programme.

The paper covers research into One Model version 4.0, an LLM-based AI recommendation model developed in-house by SKT. It also introduced a method for applying inference-capability reinforcement learning, which has drawn attention among AI researchers worldwide, in a real service environment.

One Model version 4.0 is an AI recommendation model that comprehensively understands various behavioural data, including customer clicks, usage history and interests. It then generates in natural language what products and services to recommend, why the recommendation was made and how to structure marketing messages delivered to customers.

SKT explained that version 4.0 in particular introduced inference-capability reinforcement learning, in which AI improves its own decision-making process, sharply boosting the persuasiveness and accuracy of recommendations. Inference-capability reinforcement learning is a learning method in which AI develops thinking ability through a process of generating multiple answers on its own and being evaluated.

Version 4.0 also advanced the model by applying preference scores, known as collaborative filtering scores, used in existing recommendation methods during training. SKT said this enables it to reflect both LLM-based inference capability and users’ consistent preferences.

Since One Model’s first commercial deployment with version 1.0 in 2023, the company has steadily improved recommendation performance and training efficiency. Version 4.0 is currently applied to plan recommendations, T World, T Membership and the T Direct Shop. It delivered results including up to a doubling of customer responses such as click-through rates compared with earlier versions. SKT plans to expand its application within the year to a range of products and services including T Universe.

Seok Ji-hwan (석지환), head of AT/DT data at SKT, said, "By achieving meaningful results at leading conferences in succession, we have once again proven SKT’s AI capabilities." He added, "We will continue practical AI research in which research results are directly connected to real products and services."