[Edaily Reporter Shin Yeong-bin ] Naver (NAVER(035420)) is expanding its reach in the global AI research competition. As the competition in hyper-large AI technology expands beyond model performance to include ensuring safety and reliability, interest in Naver’s cutting-edge AI research achievements is growing.

According to the information technology (IT) industry on the 9th, a joint research team from Naver AI Lab and KAIST was selected for the Spotlight Session at the ICML 2026 main conference with their paper titled “Stable ZipFlowNet: A Diverse and Stable LLM Red-Team Technique Based on Contrastive Trajectory Balance.”

ICML is regarded as one of the world’s most prestigious AI and machine learning conferences, alongside NeurIPS and ICLR. This year’s event was held at COEX in Seoul from the 6th to the 11th.

The Spotlight session is a presentation format reserved for research that garners significant academic attention even among accepted papers. At this year’s ICML 2026, 6,352 papers were accepted out of 23,918 submitted for review, and only 536 of those were selected for the Spotlight session—representing approximately 2.2% of all reviewed papers.

This research addresses automated red-teaming technology designed to identify security vulnerabilities in large language models (LLMs) prior to deployment. Red-teaming is a safety verification process that involves proactively identifying attack prompts that could induce an AI model to generate harmful or inappropriate responses, and using these findings to enhance the model’s defensive capabilities.

As generative AI services spread into real-world applications such as search, business, education, and finance, the ability to comprehensively identify model vulnerabilities prior to deployment has emerged as a critical challenge.

The research team addressed issues with existing methods based on the Generative Flow Network (ZFlowNet), which tended to become unstable during training or repeatedly generate only a few similar attacks while attempting to identify various attack types.

The key achievement is that this approach enables the detection of a wider variety of vulnerability types while maintaining attack success rates. According to the paper, the method proposed by the research team increased the number of unique attack types by approximately seven times compared to existing GFN, while maintaining an attack success rate of around 92%.

Industry observers view this achievement as a testament to the global competitiveness of Naver’s AI research. It is particularly significant that research成果 in the field of AI safety has been recognized at an international conference, especially as the focus of generative AI competition is broadening to include not only model size and performance but also safety, efficiency, and practical applicability in real-world services.

At ICML 2026, Naver showcased AI research achievements not only from Naver AI Lab but also from Team Naver, including Naver Cloud. Naver Cloud presented research on vision and foundation models, while Naver AI Lab unveiled the “Seoul World Model,” a study related to world models and physical AI.

The Seoul World Model is a world simulation model research project based on the real-world metropolis of Seoul, with participation from Naver AI Lab, KAIST, and the Seoul National University AI Research Institute, among others.

These achievements at ICML 2026 are part of a continuing trend in which Naver has steadily increased its presence on the global AI research stage. Last year, Naver also operated an on-site booth at NeurIPS 2025 to showcase research papers and technology demos.

Based on the list of published papers, Naver has been expanding its presence on the global AI research stage by successively presenting research on super-large AI efficiency, generative AI safety, vision and multimodal AI, and physical AI at major AI and machine learning conferences such as Neureps, ICLR, and ICML.