LingBot-World 2.0: Ant Group's Open Causal World Model

Ant Group's Robbyant team has released LingBot-World-Infinity (LingBot-World 2.0), a 14B-parameter open causal video world model built on Wan2.2. It's trained with a Mixture of Bidirectional and Autoregressive (MoBA) attention mask and then distilled into a few-step, real-time generator, with no post-hoc drift correction applied anywhere in the pipeline.

The team notes that pure teacher forcing causes the model to lean on context rather than genuinely predicting frames as context grows; MoBA counters this by appending a bidirectional full-attention block as a regularizer. Its leak-free cross-attention design has autoregressive rows attend to background and prior chunk prompts in a lower-triangular fashion, while bidirectional rows see a single global prompt. Distillation via DMD is also run over long self-rollout trajectories rather than teacher-forced states, optimizing the student on the error distribution it actually induces.

The system's agentic layer, called Director-Pilot, has a VLM propose event cards while a DiT-based generator renders the physical dynamics; an optional Mode B adds a SAM tracking loop for object-centric interaction. The team demonstrated a single 60-minute uninterrupted session spanning 20 distinct scenarios with no perceptible decay.

For engineers, the significance lies in addressing long-horizon texture and geometry drift at the architectural level rather than through post-hoc filtering — a chronic weakness of interactive world models. With open weights, code, and paper available, the harness is directly testable for robotics simulation, agentic planning, and interactive world-modeling research.