PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives

arXiv 2026 47.9 benchmark

TLDR

Introduces PlayWorld, a benchmark with 171 scenarios and agent players to evaluate video world models on long-horizon interactive objectives across geometry, interaction, and evolution dimensions.

Reasoning

The paper addresses a clear evaluation gap by using agent players for cross-model comparison and proposes a multi-dimensional benchmark tested on nine world models. However, the abstract is truncated and lacks detailed quantitative results or explicit limitations, making full assessment difficult.

Read-first score

Read-first score 47.9, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 47.

Recency 6%

100

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

Methodology quality 18%

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

Topical relevance 29%

67.1

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

Reproducibility 18%

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

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

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 581.

Keyword Scores

world model

video world model

interactive world model

world dynamics prediction

world simulator

generative world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Using multi-modal Agent Players to interact with world models for long-horizon objective evaluation, overcoming fixed action-conditioned comparison issues
Introducing PlayWorld benchmark with 171 scenarios each having a specified objective
Defining four core evaluation dimensions: geometry consistency, interaction fidelity, out-of-sight evolution, and insight evolution
Incorporating basic ability metrics for video quality and controllability

Methodology

The benchmark employs multi-modal Agent Players that interact with world models to pursue long-horizon objectives. It provides 171 scenarios, each with a specified objective. Models are evaluated on four core dimensions (geometry consistency, interaction fidelity, out-of-sight evolution, insight evolution) and basic video quality/controllability metrics, tested on nine state-of-the-art world models.

Key Results

Current world models are unreliable on long-horizon interactive objectives, particularly struggling with spatial consistency and persistent state evolution.

Resources

Paper