VideoVerse: How Far is Your T2V Generator from a World Model?

arXiv 25.10 2025 38.3 benchmark

TLDR

VideoVerse benchmark evaluates T2V models on temporal causality and world knowledge, revealing gaps in world model capabilities.

Reasoning

Strengths include a comprehensive benchmark with 300 prompts and 793 evaluation questions, using human-aligned QA to assess world knowledge and temporal causality. Weaknesses are the focus solely on text-to-video generation, lacking interactive or RL-based world model evaluation.

Read-first score

Read-first score 38.3, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 37.

Recency 6%

86.7

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

Methodology quality 18%

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

Topical relevance 29%

52.9

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=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

FrontierMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 207.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduction of VideoVerse benchmark that evaluates T2V models on complex temporal causality and world knowledge
Design of ten evaluation dimensions covering dynamic and static properties
Development of a human preference-aligned QA-based evaluation pipeline using modern vision-language models

Methodology

VideoVerse collects representative videos across diverse domains, extracts event-level descriptions with inherent temporal causality, and rewrites them into text-to-video prompts by independent annotators. For each prompt, ten evaluation dimensions covering dynamic and static properties are designed, resulting in 300 prompts, 815 events, and 793 evaluation questions. A human preference-aligned QA-based evaluation pipeline using modern vision-language models is developed to systematically benchmark leading open- and closed-source T2V systems.

Key Results

The benchmark reveals a significant gap between current T2V models and desired world modeling abilities, particularly in understanding complex temporal causality and world knowledge.

Resources

Paper