Paper 2511.09057
PAN: A World Model for General, Interactable, and Long-Horizon World Simulation
- Published
- Nov 2025
- Research lab
- Independent
- Citations
- 33
- GitHub
- Not linked
01 In brief
Summary
PAN is a general, interactable, and long-horizon world model that predicts future world states via video simulation conditioned on history and natural language actions.
It uses the Generative Latent Prediction (GLP) architecture, combining an autoregressive LLM-based latent dynamics backbone (Qwen2.5-VL-7B) with a video diffusion decoder (Wan2.1-T2V-14B) enhanced by Causal Swin-DPM for smooth long-term simulation.
Trained on large-scale video-action pairs, PAN supports open-domain, action-conditioned simulation.
Experiments show PAN achieves state-of-the-art performance among open-source models and is competitive with closed-source ones: 70.3% accuracy on agent simulation, 47.0% on environment simulation, 53.6% on transition smoothness, 64.1% on simulation consistency, and 56.1% on step-wise simulation.
In planning tasks, PAN improves task success by 26.7% (open-ended) and 23.4% (structured) over a VLM-only baseline.
The model is trained in two stages: module-wise training (5 epochs, 960 H200 GPUs) and joint training (early stopping after 1 epoch).
Inference uses autoregressive rollouts with classifier-free guidance and SageAttention2++ for acceleration.
02 From the paper
Abstract
A world model enables an intelligent agent to imagine, predict, and reason about how the world evolves in response to its actions, and accordingly to plan and strategize. While recent video generation models produce realistic visual sequences, they typically operate in the prompt-to-full-video manner without causal control, interactivity, or long-horizon consistency required for purposeful reasoning. Existing world modeling efforts, on the other hand, often focus on restricted domains (e.g., physical, game, or 3D-scene dynamics) with limited depth and controllability, and struggle to generalize across diverse environments and interaction formats. In this work, we introduce PAN, a general, interactable, and long-horizon world model that predicts future world states through high-quality video simulation conditioned on history and natural language actions. PAN employs the Generative Latent Prediction (GLP) architecture that combines an autoregressive latent dynamics backbone based on a large language model (LLM), which grounds simulation in extensive text-based knowledge and enables conditioning on language-specified actions, with a video diffusion decoder that reconstructs perceptually detailed and temporally coherent visual observations, to achieve a unification between latent space reasoning (imagination) and realizable world dynamics (reality). Trained on large-scale video-action pairs spanning diverse domains, PAN supports open-domain, action-conditioned simulation with coherent, long-term dynamics. Extensive experiments show that PAN achieves strong performance in action-conditioned world simulation, long-horizon forecasting, and simulative reasoning compared to other video generators and world models, taking a step towards general world models that enable predictive simulation of future world states for reasoning and acting.