Meet LingBot-World-Infinity: An Open Causal World Model With An Agentic Harness
Robbyant, Ant Group's embodied-intelligence division, has published LingBot-World-Infinity - also referred to as LingBot-World 2.0 - a 14B-parameter causal video model intended to serve as a persistent, interactive world simulator. The release marks a notable step in open world modeling research, even if the accompanying code and evaluation artifacts are more limited than the paper might suggest.
The central technical contribution is the Mixture of Bidirectional and Autoregressive (MoBA) attention mask. Most video generation models commit to either fully bidirectional attention - which processes all tokens simultaneously - or strictly autoregressive attention, which generates tokens one by one in sequence. MoBA blends both within the same architecture, giving the model the contextual richness of bidirectional processing while retaining the causal structure needed for continuous, interactive generation. This is combined with distribution matching distillation applied across long self-rollout trajectories, a training technique aimed squarely at long-horizon drift - the gradual degradation of texture fidelity and geometric consistency that accumulates over extended generation runs.
Wrapping the core generator is a Director-Pilot agentic harness. The Director is a vision-language model (VLM) responsible for proposing high-level events and narrative direction, while the Pilot is a Diffusion Transformer that renders those proposals frame by frame. This separation of planning and rendering is a practical architectural choice that allows each component to be updated or swapped independently. The researchers report a single uninterrupted 60-minute session spanning 20 scenarios as a demonstration of the system's stability under extended rollout.
The practical release, however, is narrower than the paper's scope. What is publicly available amounts to one model checkpoint, a 480P reference inference script, no deployment or training code, and no formal quantitative benchmark results - making independent evaluation difficult at this stage. The model is released under a CC BY-NC-SA 4.0 license, which restricts commercial use. For researchers interested in long-horizon video generation and interactive world modeling, the architectural ideas around MoBA and distillation-based drift correction are worth tracking, but those hoping to build on or benchmark the system directly will need to wait for a more complete release.
