On March 10, Tencent's Hunyuan announced via its official social media account that its 3D team has open-sourced WorldCompass, the industry's first reinforcement learning post-training framework designed for world models. This framework serves as the official reinforcement learning extension module for the previously released Hunyuan World Model 1.5, enabling more accurate interactions and an improved user experience. According to the announcement, WorldCompass is a reinforcement learning post-training framework specifically tailored for long-sequence, interactive world models. By incorporating reinforcement learning mechanisms, it directly guides the model to more accurately follow user instructions when exploring virtual worlds while maintaining long-term visual consistency. Experimental results indicate that WorldCompass significantly enhances the interaction accuracy and visual fidelity of state-of-the-art open-source world models such as WorldPlay, with interaction accuracy increasing by nearly 35% in complex composite action scenarios. Technically, WorldCompass introduces three core innovations by restructuring traditional reinforcement learning frameworks to suit autoregressive world generation: first, slice-level sampling, a fine-grained sampling strategy for autoregressive world generation; second, a 3D reward function designed to prevent reward hacking; and third, an efficient reinforcement learning optimization algorithm that ensures more stable and faster training.