Toss said on December 3 that a paper had been accepted to NeurIPS 2025, a leading global artificial intelligence conference. It said the study was led by researcher Jinwoo Lee of the Toss Face Modeling Team in collaboration with the Vision Lab at Seoul National University.

NeurIPS is one of the most influential conferences in machine learning and neural information processing, and its paper acceptance rate is about 20 percent. The event runs from December 2 to 7 at the San Diego Convention Center, where researchers share recent AI findings.

The accepted work is Federated Local Prior Alignment, or FedLPA. Toss said it helps train AI models in countries where privacy rules prevent data from being moved to a central server. It said the method addresses limits in existing federated learning systems, which see sharp performance drops when data characteristics differ by country or user group or when new data types appear.

The team combined Infomap-based local clustering, which groups data with similar traits such as country or user segment, with Local Prior Alignment, a technique that aligns predictions to improve training stability.

Toss said this allows each device to identify and use its own data structure and supports accurate discovery of new categories in settings where it is difficult to know category types or data distribution. It said the method demonstrated performance in Generalized Category Discovery.

A Toss official said the acceptance has significant meaning as it marks the first formal recognition of the company’s AI capability at a global conference. The official said Toss will continue research on technologies that can be applied to real services to provide more precise AI-based services while protecting privacy.