Team Naver made a strong impression at ICML 2026, held at COEX in Seoul. The team unveiled a coherent suite of research spanning AI model advancement to technologies that interact with the physical world. Their work emphasizes safety, operational efficiency, and 3D spatial understanding.
Security of LLMs: red‑teaming and Stable‑GFlowNet
Security for large language models (LLMs) was a central theme. Team Naver showcased red‑teaming techniques that simulate adversarial attacks to uncover vulnerabilities before production deployment. A notable contribution, Stable‑GFlowNet, addresses training instabilities and the repetition of problematic patterns, enabling more diverse and realistic attack scenarios. Tools like this strengthen the ability to test models under conditions that better reflect real‑world adversarial pressure.
Operational efficiency: Simerge, FlowBot and model fusion
The team presented several advances focused on making models and agents more efficient in practice. Simerge enables merging multiple specialist models by tuning only a single layer, yielding significant improvements on vision and NLP benchmarks. FlowBot automates the search for optimal scheduling when multiple AIs must collaborate, reducing reliance on manual orchestration. Team Naver also demonstrated methods to improve LLM post‑processing by clustering thousands of datasets by feature and consolidating models through a single, efficient fusion step.
3D understanding and dynamic scene reconstruction
On 3D perception, Team Naver introduced a method that reconstructs moving three‑dimensional scenes from blurry or shaky monocular video sequences. Unlike conventional approaches where motion and shape are conflated, the new technique estimates geometry from motion trajectories, markedly improving reconstruction fidelity. This advance is critical for robotic perception and virtual‑reality applications that require accurate scene understanding under realistic camera motion.
Seoul World Model: a virtual city for research and robotics
Team Naver also unveiled the Seoul World Model, a virtual replica of Seoul developed in collaboration with universities and research centers. This physical‑AI platform simulates spatial data across the metropolis, providing a rich training ground for trajectory learning, robotic behaviors, and validation of agents in realistic urban environments.
Impact and outlook
The work presented reflects a pragmatic ambition: moving theoretical advances toward industrial deployment. Emphasizing safety through red‑teaming, efficiency via model fusion, and extension into the physical world with urban simulation, Team Naver outlines a coherent roadmap for applied AI. In the medium term, these technologies could accelerate integration of LLMs and autonomous agents into real‑world services while reducing operational risk and enhancing the spatial perception capabilities of embedded systems.
