Tencent's MixEn open-source framework is the first reinforcement learning post-training framework targeting the world model.

On March 10th, Tencent's Mixed Reality 3D team open-sourced the first reinforcement learning framework WorldCompass for world models. It is designed specifically for long-temporal and interactive world models. If the world model is considered the engine, then WorldCompass is the accurate "compass", guiding the model on how to more accurately follow user commands to explore the world and maintain long-temporal visual consistency, through the introduction of reinforcement learning mechanisms.