⌁ Awesome World Model Hub Papers · Datasets · Projects
中文 GitHub
← Back to papers

WoW: Towards a World omniscient World model Through Embodied Interaction

arXiv 25.9 2025 68.6 method, application

TLDR

WoW is a 14B-parameter generative world model trained on robot interactions, achieving state-of-the-art physical causality and object permanence via embodied learning.

Reasoning

The paper presents a novel approach grounding world models in embodied interaction, with large-scale training and a new benchmark. Strengths include strong empirical results and a closed-loop action framework; weaknesses are acknowledged stochastic instabilities and potential limitations in generalizing beyond robot interaction data.

Read-first score

Read-first score 68.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 57.

Recency 8%
86.7

Uses a gentle age decay so recent papers surface without erasing older foundations. 2025

Topical relevance 42%
81.4

Uses existing LLM keyword relevance scores normalized to 0-100. world model,world simulator,generative world model,interactive world model,video world model,world dynamics prediction,model-based reinforcement learning world model

Methodology quality 25%
80

Screens visible abstract and analysis fields for experiment, dataset, baseline, metric, and limitation evidence. markers=benchmark,evaluation,metric

Reproducibility 25%
30

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Field roles

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 157.

Keyword Scores

world model
10
generative world model
9
interactive world model
9
video world model
8
world dynamics prediction
8
world simulator
7
model-based reinforcement learning world model
6

Deep Analysis

Innovations

  • Training a 14-billion-parameter generative world model on 2 million real-world robot interaction trajectories to ground physical intuition in causally rich interactions.
  • SOPHIA, a method using vision-language model agents to evaluate DiT-generated video output and iteratively refine language instructions to constrain outputs toward physical realism.
  • Co-training an Inverse Dynamics Model that translates refined plans into executable robotic actions, closing the imagination-to-action loop.
  • WoWBench, a new benchmark focused on physical consistency and causal reasoning in video.

Methodology

WoW is a 14-billion-parameter generative world model trained on 2 million robot interaction trajectories. It uses a Diffusion Transformer (DiT) for video generation, with a vision-language model agent (SOPHIA) that evaluates outputs and iteratively refines language instructions to improve physical realism. An Inverse Dynamics Model is co-trained to convert refined plans into robotic actions. The model is evaluated on the newly introduced WoWBench benchmark using both human and autonomous evaluation metrics.

Key Results

WoW achieves state-of-the-art performance on WoWBench for physical consistency and causal reasoning, demonstrating strong ability in physical causality, collision dynamics, and object permanence.

Limitations

  • The model's understanding of physics is a probabilistic distribution of plausible outcomes, leading to stochastic instabilities.
  • The model exhibits physical hallucinations, indicating imperfect grounding in physical causality.

Resources

Paper

Tags