Video = World + Event Stream

arXiv 2026 39.3 method, system

TLDR

Reframes video as world plus event stream for real-time interactive prediction, applied to full-duplex audio-visual interaction.

Reasoning

Strengths: Novel conceptual framing of video as world + event stream, enabling general-purpose pretraining for real-time tasks. Weaknesses: Abstract lacks detailed methodology, evaluation metrics, or comparison to baselines; claims are high-level without empirical evidence beyond latency numbers.

Read-first score

Read-first score 39.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 46.

Recency 6%

100

Uses a gentle age decay so recent papers surface without erasing older foundations. 2026

Topical relevance 29%

65.7

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 18%

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

Reproducibility 18%

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

Citation impact 18%

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. cited_by_count=0

Citation velocity 12%

Citation velocity estimates citations per publication-year to reduce old-paper bias. velocity=0.00

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 279.

Keyword Scores

video world model

world dynamics prediction

world model

interactive world model

generative world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

Reframing video as a world (persistent context) plus an event stream (dynamic changes)
General-purpose pretraining task: given a world and incoming input, predict how the world moves, changes, and responds in real time
Specialization of the pretrained competence to real-time full-duplex audio-visual interaction
Vision-language-action-like mapping from multimodal user input to speech and behavior actions

Methodology

The model is a native-streaming interaction model that decomposes video into a stable world context and a dynamic event stream. It is pretrained on large amounts of real video to predict real-time world changes, then specialized to real-time audio-visual interaction by mapping multimodal input to language-form speech and behavior actions.

Key Results

The model preserves the Wan-Streamer v0.2 operating point: 640x368 video at 25 FPS, a 160 ms streaming unit, ~200 ms model-side latency, and ~550 ms total interaction latency under a 350 ms network budget.

Resources

Paper