LingBot-World 2.0 Released: Focusing on Hour-Long Stable Interaction

Ant Group's embodied AI company introduced LingBot-World 2.0, an open-source interactive world model. Unlike previous models that focus on impressive initial visuals, this project tackles the critical issue of autoregressive drift and error accumulation that causes worlds to collapse over time. According to multiple authors, the model maintains 720p, 60fps interaction for up to an hour, shifting the evaluation metric from single-frame beauty to long-term stability.

Key Details and Open Source Release

The currently released model is the 14B causal-fast version, with a 1.3B lightweight version described in the paper for single-GPU deployment. A highlighted stress test result is a continuous 60-minute rollout covering 20 scenes without significant degradation. Alongside the paper, the project released weights, a GitHub repository, and an interactive demo, making it a comprehensive open-source release.

System Design

LingBot-World 2.0 features an interactive agentic harness. Its architecture includes a VLM Brain to read scenes and propose context-relevant events, a pilot agent to control character behavior, and a director agent to continuously add new elements to the environment. The focus is on keeping the world actively running during interaction rather than merely generating frames.

Limitations and Discussion Focus

The authors noted that the paper explicitly acknowledges the lack of true long-term memory; content exceeding the context window is regenerated rather than remembered. Commenters emphasized that the core value lies in reducing long-horizon error accumulation through training. The key metric for future world models might simply be keeping the world intact, rather than just looking realistic for the first few minutes.

2026-07-13 ~ 2026-07-15 · 13 related posts