[Edaily Reporter Kim Hyun-ah ] SelectStar, a company specializing in AI data and reliability verification, announced on the 6th that three of its research papers on AI safety have been accepted for presentation at a workshop of the International Conference on Machine Learning (ICML) 2026, a world-renowned artificial intelligence (AI) conference.

Inc.

ICML is regarded as one of the world’s most prestigious AI conferences, alongside the International Conference on Learning Representations (ICLR) and the Conference on Neural Information Processing Systems (NeurIPS). This year’s event is being held at COEX in Seoul from the 6th to the 11th.

The accepted papers include one from the AIWILD (AI in the Wild) workshop, which addresses the safety and reliability of Agent AI, and two from the Culture x AI workshop, which explores generative AI from a cultural perspective. All papers addressed evaluation and training data construction techniques aimed at enhancing the safety and reliability of large language models (LLMs).

SelectStar explained that through this research, it proposed a methodology that addresses the limitations of existing AI safety evaluation methods. Current AI safety evaluations have been criticized for often relying on large language models or simply translating English-language data, which fails to adequately reflect real-world service environments and cultural contexts.

The first paper, “Reliable to Expressive: A Curriculum for Rubric-Following Safety Judges,” proposed a “safety judge” (a safety evaluation model) to assess the risk of AI responses.

The research team stated that by applying curriculum learning—an approach that involves learning step-by-step, starting with easy tasks—they demonstrated that consistent safety evaluation performance can be maintained even in new industries such as finance. The company expects this technology to be utilized for fast and accurate AI safety verification even within internal closed-network environments.

The second paper, “Culturally-Adapted Red-Teaming Across East and Southeast Asian Contexts,” presented red-teaming methodologies tailored to specific cultural contexts (a verification technique that intentionally identifies vulnerabilities and risks in AI).

By analyzing cases in Korean, Japanese, Thai, and Khmer using CAGE—Selectstar’s proprietary framework—the researchers confirmed that simply translating English-language benchmarks (performance evaluation standards) makes it difficult to fully capture country-specific risk factors. The research team proposed the need to build evaluation datasets that reflect cultural specificities and real-world usage environments.

The third paper, “Korean Culture into LLM Alignment: Toward Cultural Coherence,” introduced a methodology for building safety alignment data (a technique for training AI to provide responses that align with human values and norms) that reflects Korean social norms, laws, institutions, and linguistic characteristics.

SelectStar explained that this technology enhances the performance of AI responses tailored to Korean culture while minimizing any degradation in model performance, and that it is currently being utilized in the development of the company’s proprietary AI foundation model. SelectStar is active as a member of the “ SKTelecom(017670) ” consortium.

Kim Se-yeop, CEO of SelectStar, stated, “The acceptance of this ICML workshop marks global recognition of our core technical capabilities in the field of AI safety,” adding, “Having previously demonstrated our research competitiveness at ACL 2026 (Association for Computational Linguistics) and the ICLR 2026 main conference, we will continue to develop technologies to build a trustworthy AI ecosystem.”

Meanwhile, SelectStar will also participate in the “3rd AI Safety Policy Research Seminar” hosted by the Artificial Intelligence Safety Institute (AISI) on the same day. ScaleAI will also participate as a speaker at this event. The two companies plan to present on the limitations of translation-based AI safety evaluation and strategies for building culture-aware LLM benchmarks.