On March 10, 2026, the Tencent Hunyuan 3D team made a groundbreaking announcement—the open-sourcing of WorldCompass, the inaugural reinforcement learning post-training framework specifically crafted for world models. This innovative framework is meticulously designed to cater to long-sequence, interactive world models. It harnesses the power of reinforcement learning mechanisms to meticulously guide models in navigating and exploring the world in strict accordance with user instructions, all while preserving long-sequence visual consistency.

Experimental results vividly demonstrate that WorldCompass substantially elevates both the interaction accuracy and visual fidelity of the state-of-the-art (SOTA) open-source world model, WorldPlay. This enhancement is particularly pronounced in complex combined action scenarios, where interaction accuracy witnesses a remarkable improvement of nearly 35%.