The paper 'CAGE: A Framework for Culturally Adaptive Red-Teaming Benchmark Generation' developed by the AI Safety team of Selectstar (CEO Se-yeop Kim) has been accepted for the main conference of ICLR 2026 to be held in Brazil in April.
ICLR is a top-tier international conference in the fields of AI and machine learning. This year, only the top 28% of approximately 19,000 papers were accepted. SelectStar papers are selected for the main track, demonstrating international recognition for their originality and technical sophistication. The research was conducted entirely by the company's internal staff, without any external agencies.
CAGE technology automatically generates red-teaming data to verify the safety of AI models, reflecting the cultural and legal environments of each country. Unlike existing methods that rely on translations of data primarily from English-speaking countries, CAGE generates localized attack questions through "Semantic Mold," testing the defense rate of AI models and effectively detecting potential risks. It has also demonstrated excellent performance in data-poor language regions, such as Cambodian.
The paper also unveiled "KoRSET," a Korean-style safety benchmark. KoRSET detects AI model vulnerabilities more effectively than existing simple translation datasets, demonstrating its performance optimized for safety verification based on Korean culture.
CAGE technology is also being applied in industrial settings, where it is being utilized to identify model vulnerabilities and improve operational efficiency in large-scale AI projects with which SelectStar collaborates. Kim Min-woo, AI Safety Team Leader, stated, "The adoption of ICLR demonstrates SelectStar's unrivaled position as a leading AI technology company."
Based on this achievement, SelectStar plans to expand its reliability assessment solution to industries requiring high levels of security, such as finance and the public sector. The paper is scheduled to be released on the open-source platform Arxiv in March.
- See more related articles
You must be logged in to post a comment.