We are open-sourcing WorldCompass, an RL post-training framework specifically…

Tencent's WorldCompass trains interactive video models to respond more faithfully to movement commands. Its training diagram shows several possible continuations generated from the same preceding footage, then scored for action-following and visual quality. Those scores guide further training. The released framework makes this adjustment process available to developers working with their own data and rewards.

This extends the control problem in WorldPlay: a generated scene needs to follow the requested movement as the interaction continues. Tencent presents WorldCompass as a way to improve that behavior through additional training.