Ant Group's Robbyant Unveils LingBot-World-Infinity Causal Model

Ant Group's embodied-intelligence unit, Robbyant, has released LingBot-World-Infinity (LingBot-World 2.0), a causal video generation model that functions as an interactive world simulator. This new model aims to address two significant challenges in generative AI: long-horizon drift and interactive latency. LingBot-World-Infinity generates video frame by frame, conditioned on a stream of user actions, with each state depending on past frames and current input. The research team formalized this process as a causal factorization: p_θ(x_1:T | a_1:T) = Π_t p_θ(x_t | x_<t, a_≤t), where x_t represents the visual state at time t and a_t combines a camera pose and a text prompt. Camera pose is integrated using Plücker embeddings via adaptive layer normalization (AdaLN), while text prompts are processed chunk-wise through cross-attention.

Robbyant claims four key upgrades over its predecessor, LingBot-World. These include an unbounded interaction horizon with consistent output quality, a distilled real-time variant capable of driving 720p video streams at 60 frames per second, an expanded action space encompassing activities like attacking, archery, spell-casting, and shooting, and an agentic harness that pairs a pilot agent with a director agent. The primary LingBot-World-Infinity model features 14 billion parameters, with a more lightweight 1.3 billion parameter counterpart described as deployable on a single GPU.

The core architectural innovation is the Mixture of Bidirectional and Autoregressive (MoBA) Attention Mask, which addresses the issue of drift. Traditional autoregressive video training employs a teacher forcing mask where each noisy frame attends to itself and its clean context. The research team identified a failure mode in this approach: as the context grows, the model relies on it excessively rather than accurately predicting future frames, leading to overfitting and visual quality degradation. The MoBA mask enhances the teacher forcing mask by appending a bidirectional full-attention block, which acts as a regularizer and improves the model's ability to handle flexible-length generation.