Team Naver participated in the global artificial intelligence (AI) conference “ICML 2026,” held at COEX in Seoul from the 6th to the 11th, showcasing a range of AI technologies. Visitors gather at the Naver booth set up at the venue. Photo courtesy of Team Naver
Team Naver has showcased research achievements spanning from AI model advancement to physical AI that interacts with the real world at ICML 2026 (the International Conference on Machine Learning), the world’s most authoritative artificial intelligence (AI) conference.
Team Naver said on the 13th that it participated in ICML 2026, held from the 6th to the 11th at COEX in Seoul, and presented numerous major research achievements along with cases of applying them to actual services. The research unveiled at this conference consists of concrete achievements aimed at integrating full-stack AI technology into actual services and industrial sites, and is broadly organized 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 notable achievement was red-teaming technology, which detects vulnerabilities in large language models (LLMs) from an attacker’s perspective. “Stable-GFlowNet,” announced by Team Naver, is a technology that structurally resolves the training instability and repetitive similar-pattern problems that were limitations of existing methods. The research was selected for a “Spotlight,” which is given to only the top approximately 2.2% of all accepted papers. Through this, it is evaluated that attack vulnerabilities can be verified with more powerful and diverse scenarios before LLMs are deployed to actual services.
In the area of improving the operational efficiency of models and agents, technologies for efficient integration and operation of AI came to the fore. “Simerge” is a model merging technique that combines multiple models specialized for different tasks into one. While taking a concise approach that draws out synergy between models by adjusting only a single layer, it achieved top-level performance across various benchmarks including vision and natural language processing.
“FlowBot” is a technology that allows AI to find the optimal sequence on its own, without humans having to design the task order one by one, in the process of multiple AIs collaborating to solve problems. This was evaluated as a key research achievement in harness engineering. In addition, a technology was also introduced that efficiently boosts LLM post-processing performance with just a single merge, after dividing hundreds to thousands of different datasets into groups by their characteristics and training them separately.
In the area of 3D spatial understanding, research was presented on restoring moving three-dimensional scenes using only single-camera footage that is shaky or out of focus. Existing methods made precise restoration difficult because the object’s motion and shape information became mixed together due to movement, but this research demonstrated outstanding restoration performance by introducing a method that estimates shape based on motion trajectory.
Beyond presenting papers, Team Naver also intensively introduced the “Seoul World Model,” which reproduces the actual city of Seoul in virtual space. Jointly developed by Naver, Naver Labs, the Korea Advanced Institute of Science and Technology (KAIST), and Seoul National University, this model is a core physical AI platform that simulates spatial data across all of Seoul and can be directly used for learning robot paths and behaviors.