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

World-in-Context: Interactive Video World Modeling with In-Context Conditioning

arXiv 2026 61.3 method

TLDR

World-in-Context unifies view, action, and camera controls via in-context tokens for interactive video world modeling, achieving strong navigation and perspective-switching performance on WBench.

Reasoning

The paper introduces a novel conditioning framework that disentangles controls and uses IT-RoPE for temporal continuity, with convincing benchmark results. Weaknesses include limited abstract detail about methodology and no explicit discussion of limitations or broader applicability.

Read-first score

Read-first score 61.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 52.

Recency 8%
100

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

Topical relevance 42%
74.3

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%
50

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

Reproducibility 25%
38

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=code

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 395.

Keyword Scores

world model
10
interactive world model
10
video world model
10
generative world model
8
world simulator
7
world dynamics prediction
6
model-based reinforcement learning world model
1

Deep Analysis

Innovations

  • In-Context Conditioning (ICC) that unifies view, action, and camera controls into explicit tokens in a shared self-attention space instead of specialized conditioning pathways.
  • Explicit view tokens that disentangle conflicting dynamics and enable emergent view switching from single-view data.
  • Complementary use of action and camera inputs, with action governing local responsiveness and camera governing global stability.
  • Interaction-Transport RoPE (IT-RoPE) that parameterizes temporal attention with accumulated control transitions to preserve temporal continuity across execution intervals.

Methodology

World-in-Context uses In-Context Conditioning to represent view, action, and camera controls as explicit tokens contextualized within a shared self-attention space, avoiding fragmented architectural conditioning pathways. It introduces Interaction-Transport RoPE to encode accumulated control transitions in temporal attention. The framework is evaluated on WBench using World-in-Context-5B against baselines including EchoWM 22B and existing 5B models.

Key Results

World-in-Context-5B achieves WBench navigation performance of 81.5, on par with 22B EchoWM (81.6), and a 93.8% perspective-switching success rate versus 18.8% for existing 5B baselines. Controlled analyses show compounding benefits with richer conditioning mixtures, favorable data efficiency, and applicability to causal rollouts.

Resources

Paper

Tags