Team Naver said on Sunday it participated in the International Conference on Machine Learning (ICML) 2026 and presented research results spanning areas from advanced AI models to physical AI. ICML is considered one of the world's three major AI and machine-learning conferences along with NeurIPS and ICLR. It was held at Seoul's COEX from July 6 to 11.

Team Naver opened a booth under the theme "Where AI Research Becomes Reality" and showcased research results and real service cases.

The presentations were divided into three areas: strengthening AI safety, improving the operational efficiency of models and agents, and 3D spatial understanding and expansion into the physical world.

The most noted technology in Team Naver's research was "red-teaming". It is a technique for identifying vulnerabilities in large language models (LLMs) from an attacker's perspective. Team Naver structurally addressed training instability and repeated similar-pattern issues in existing approaches through "Stable-GFlowNet", a generative model training method. The paper was selected as a "Spotlight", placing it in the top 2.2 percent of accepted papers. It can verify attack vulnerabilities across various scenarios before deploying an LLM in real services.

In the area of improving the operational efficiency of models and agents, it disclosed a technology that integrates multiple models. "SyMerge" is a model-merging technique that combines multiple models specialized for different tasks into one. By adjusting only a single layer, it draws synergy across models and secured performance on various benchmarks such as vision and natural language processing.

"FlowBot" is a technology in which AI finds the order of tasks on its own when multiple AIs cooperate. It also introduced a technique that boosts LLM post-training performance by dividing hundreds to thousands of datasets into groups for split training and then merging them at once.

In 3D spatial understanding, it presented research that reconstructs a moving three-dimensional scene from single-camera footage that is shaky or out of focus. The research used a method that estimates shape based on motion trajectories.

Team Naver also introduced the "Seoul World Model", a virtual recreation of Seoul. The model was jointly developed by Naver, Naver Labs, Korea Advanced Institute of Science and Technology, and Seoul National University. It is a physical AI platform that can simulate spatial data across Seoul and use it to train robots in routes and actions.