Kun Zhan

No current LLM-based agent can reliably avoid executing unsafe actions when using third-party skills, with a staggering 17% failure rate even in optimal conditions.

Future feature foresight and sparse point tracking together can transform how VLA models navigate complex environments, leading to unprecedented performance in visuomotor tasks.

Combining language supervision with future latent alignment in VLA models leads to unprecedented stability and transfer performance across diverse robotic tasks.

Efficient dense 3D reconstruction can now serve as a scalable foundation for next-generation autonomous driving, achieving real-time performance without sacrificing fidelity.

Achieving up to 46% token compression without sacrificing accuracy, HMPO revolutionizes the efficiency of chain-of-thought reasoning in large language models.

RL fine-tuning unlocks a 6x performance gain for in-place trajectory editing in autonomous driving, demonstrating the power of aligning diffusion planners with reinforcement learning.

LinkVLA tackles the language-action misalignment problem in autonomous driving by unifying language and action tokens in a shared space, leading to faster and more accurate instruction following.

By explicitly encoding 3D geometry, GeoDrive achieves more realistic and controllable autonomous driving scene modeling, outperforming prior world models in action accuracy and spatial awareness.