SK Telecom demonstrated its capabilities at a global conference with AI recommendation technology aimed at more precisely identifying customers' actual tastes.

SKT said on April 28 it presented a paper at ICLR 2026, a top AI conference, on an AI recommendation model that finds customers' "true preferences." ICLR is considered one of the world's three major AI conferences along with NeurIPS and ICML. The company said research presented there is often used as a benchmark for industrial applications and follow-on research and development.

The core of the research is SKT's proprietary Collision-Aware Preference Optimization (C-APO) technology. It is used to filter out conflicting preferences in customer behavior data to more accurately recommend products customers truly want.

The model defines long-term, repeatedly observed usage patterns as "consistent preferences" and temporary clicks or one-off usage as "superficial preferences." It is designed to raise training weights when the two types of preferences match, and to lower them when they conflict by treating them as noise.

The company said this can reduce the problem of distorted recommendations when, for example, a user who typically consumes romance content watches an action movie once. It is also expected to strengthen the basis for logical explanations of recommendation results and help improve the credibility of AI recommendation services.

SKT plans to apply the technology to an agent-based recommendation system in which a three-stage hierarchy works in coordination, moving from interpreting usage context to generating recommendation candidates and then verifying and adjusting results.