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

WorldLens: Full-Spectrum Evaluations of Driving World Models in Real World

arXiv 25.12 2025 50.5 benchmark

TLDR

WorldLens is a full-spectrum benchmark for evaluating driving world models across visual, geometric, physical, and behavioral fidelity.

Reasoning

The paper introduces a comprehensive evaluation framework with a dataset and agent, addressing a clear gap in unified assessment. Strengths include multi-dimensional coverage and human alignment; weaknesses are not evident from the abstract alone, but the benchmark's scalability and generalizability remain to be seen.

Read-first score

Read-first score 50.5, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 51.

Methodology quality 18%
90

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

Recency 6%
86.7

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

Topical relevance 29%
72.9

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%
46

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

Citation impact 18%
0

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

Citation velocity 12%
0

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: 672.

Keyword Scores

world model
10
generative world model
9
video world model
8
interactive world model
7
world dynamics prediction
7
world simulator
6
model-based reinforcement learning world model
4

Deep Analysis

Innovations

  • WorldLens benchmark covering five evaluation aspects: Generation, Reconstruction, Action-Following, Downstream Task, and Human Preference
  • WorldLens-26K dataset of human-annotated videos with numerical scores and textual rationales
  • WorldLens-Agent evaluation model distilled from human annotations for scalable, explainable scoring

Methodology

WorldLens is a full-spectrum benchmark that evaluates generative world models across five aspects: Generation, Reconstruction, Action-Following, Downstream Task, and Human Preference. To align objective metrics with human judgment, the authors construct WorldLens-26K, a large-scale dataset of human-annotated videos with numerical scores and textual rationales, and develop WorldLens-Agent, an evaluation model distilled from these annotations to enable scalable, explainable scoring.

Key Results

No existing world model excels universally across all dimensions: models with strong textures often violate physics, while geometry-stable ones lack behavioral fidelity.

Limitations

  • Benchmark is specific to driving world models, limiting generalizability to other domains
  • Human annotations may introduce subjective biases, and the distilled evaluation model may not perfectly capture all aspects of human judgment

Resources

PaperProject

Tags