SK Telecom Presents AI Reasoning Model at ICLR 2026 to Identify Customers’ True Preferences
Seoul, Korea, Apr 28 – SK Telecom (NYSE:SKM, hereinafter referred to as “SKT”) announced today that its new AI recommendation model, which more accurately identifies customers’ true preferences, was presented at a leading AI conference, the International Conference on Learning Representations (ICLR) 2026.
ICLR, along with the Conference on Neural Information Processing Systems (NeurIPS) and the International Conference on Machine Learning (ICML), is regarded as one of the world’s top three AI conferences. Technologies presented at these events see industrial application within a few years or become benchmarks for academic research and product development, making them key indicators of the future direction of AI.
The research paper, titled “More Than What Was Chosen: LLM-Based Explainable Recommendation Beyond Noisy User Preferences,” centers on SKT’s proprietary Conflict-Aware Direct Preference Optimization (C-APO) technology. The technology filters out conflicting preference signals, enabling more accurate recommendations of products that customers genuinely desire.
To achieve this, SKT introduced the concept of “coherent preference” — reflecting recurring patterns and contextual coherence across a customer’s overall behavior. Through this approach, the model distinguishes coherent preferences from revealed preferences derived from single actions, such as accidental clicks or one-time, unintentional choices.
This approach increases the learning weight when the two preference signals align, reinforcing the recommendation. When the signals conflict, the model treats the discrepancy as noise and lowers the learning weight, enabling it to identify the most appropriate recommendations.
Conventional AI recommendation systems have struggled to interpret conflicting preference signals — for example, a customer who typically watches romance films may receive only action recommendations after watching even a single action movie based on a friend’s suggestion. SKT’s new recommendation model, however, filters out such inconsistencies to identify the customer’s true preferences.
SKT plans to apply this research to future agent-based customer recommendation systems that integrate AI for interpreting usage context, generating recommendation candidates, and refining results through an orchestration layer. The company aims to deliver more advanced and personalized recommendations tailored to each customer’s needs and behavior.
“The technology lays the foundation for delivering personalized recommendations by better understanding customers’ true preferences,” said Seok Ji-hwan, Head of AI/DT Data Office at SK Telecom. “The presentation at ICLR 2026 marks a milestone in SKT’s AI transformation to enhance customer experience, and we will continue to advance AI technologies that customers can benefit from.”
