Dongyoon Han

Retrieval can replace costly fine-tuning in VLA models, enabling seamless adaptation to new tasks with just a single trained policy.

LLMs generate significantly better code when prompted with coherent natural language narratives that structure the task overview, constraints, and test cases, outperforming simple rephrasings by 18.7% on average.

Imagine a world model that doesn't just dream up environments, but flawlessly renders a real city like Seoul, complete with text-prompted scenarios and diverse camera movements.