[Edaily Reporter Han Kwangbeom ] Naver (NAVER(035420)) demonstrated its global AI technological competitiveness by presenting a wide range of research achievements—spanning from the advancement of AI models to physical AI in the real world—at “ICML 2026” (International Conference on Machine Learning), the world’s most prestigious AI conference.
Team Naver participated in ICML 2026, held at COEX in Seoul from the 6th to the 11th, under the theme “Where AI Research Becomes Reality,” setting up a booth to showcase its major research achievements and numerous real-world service cases based on them.
The research presented by Team Naver at this conference consisted of concrete achievements aimed at applying full-stack AI technology to actual services and industrial settings, covering three major areas: △enhancing AI safety, △streamlining model and Agent AI operations, and △understanding 3D spaces and expanding into the physical world.
The most notable research was the “Red-Teaming” technology, which identifies vulnerabilities in large language models (LLMs) from an attacker’s perspective. “Stable-GFlowNet” is a technology that structurally resolves the issues of training instability and repetitive patterns inherent in existing methods; it was selected for “Spotlight,” a distinction awarded to approximately the top 2.2% of all accepted papers. It was evaluated as enabling the verification of attack vulnerabilities in LLMs through more robust and diverse scenarios before they are deployed in actual services.
In the field of model and agent operational efficiency, technologies for more efficiently merging and operating AI were unveiled. “SyMerge” is a model merging technique that integrates multiple models specialized for different tasks into a single model. Despite presenting a simple idea—adjusting just a single layer to create synergy between models—it achieved state-of-the-art performance on various benchmarks, including vision and natural language processing.
“FlowBot” is a technology that allows AI systems to autonomously determine the optimal sequence of tasks when collaborating to solve problems, without requiring humans to design each step individually; it has been recognized as a key research achievement in harness engineering. Additionally, a technique was introduced that efficiently boosts the performance of large language models (LLMs) in post-processing by grouping hundreds to thousands of different datasets according to their characteristics for separate training, followed by a single merge.
In the field of 3D spatial understanding, research was presented on reconstructing moving 3D scenes using only single-camera footage that is shaky or out of focus. While existing methods struggled to achieve precise reconstruction because motion caused the object’s motion and shape information to become intermixed, this research demonstrated exceptional reconstruction performance by estimating shapes based on motion trajectories.
Alongside the paper presentations, Team Naver highlighted the “Seoul World Model,” a virtual recreation of the city of Seoul. Jointly developed by Naver, Naver Labs, the Korea Advanced Institute of Science and Technology (KAIST), and Seoul National University, this model garnered significant attention as a core platform for physical AI—capable of simulating spatial data across the entire city of Seoul and directly applied to robot path planning and behavior learning.
ICML is considered one of the world’s top three artificial intelligence and machine learning conferences, alongside NeurIPS and ICLR, and was held in Korea for the first time this year.
