Yanyong Zhang

PPO-Clip's stationarity and critic tracking are jointly bounded by a common amplification factor dependent on algorithmic parameters, revealing a unified error mechanism.

Achieving a 3.36-fold speedup in action generation, DriftingVLA eliminates the need for iterative refinement while maintaining high performance in robot control tasks.

Generation-first training eliminates latent collapse, achieving unprecedented stability and performance in latent generative modeling.