Jianhua Han

Unlock multimodal interleaved generation in existing vision-language models without large interleaved datasets using a novel reinforcement learning approach with hybrid rewards.

Forget monolithic action decoders: AtomicVLA's skill-guided mixture-of-experts unlocks significant gains in long-horizon robotic manipulation and continual learning.

RL's inherent resilience to catastrophic forgetting can be harnessed to improve continual learning in GUI agents, outperforming SFT alone.