LingBot World, an Open-Weight Model with Long-Term Rollout Stability

LingBot World, an open-weight interactive world model based on Causal DiT that generates frames in real time according to user input, has been released.

This model applies the following core techniques to prevent the drift phenomenon that occurs during long-term generation.

• MoBA (Mixed Bidirectional/Autoregressive) attention mask: Mixes bidirectional and autoregressive attention to mitigate the problem of excessive dependence on recent frames.
• Dynamic KV-cache scheduling: Designed to maintain computational efficiency even during long rollout processes.
• Consistency Distillation: Secures stability through distribution matching distillation based on long self-rollout trajectories, rather than simple teacher-forced frames.

Actual test results showed that the model maintains stable performance without visual degradation even during 60 minutes of continuous rollout. However, the model's persistence is limited to appearance, and areas that fall outside the context window are newly generated upon revisiting. The model weights have been released under the CC-BY-NC-SA license.