Is Your Driving World Model an All-Around Player?

arXiv 2026 63.6 benchmark, application

TLDR

Introduces WorldLens, a unified benchmark evaluating driving world models across pixel quality, geometry, closed-loop planning, and human perception.

Reasoning

The paper addresses a critical gap in evaluating driving world models beyond visual quality, proposing a comprehensive benchmark with human annotations and an auto-evaluator. Strengths include multi-dimensional assessment and real-world data; weaknesses are domain specificity and lack of detailed model performance breakdown.

Read-first score

Read-first score 63.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 52.

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,dataset,evaluation,metric

Citation impact 18%

75.2

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.75159998

Topical relevance 29%

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

Reproducibility 18%

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

Citation velocity 12%

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

Field roles

FoundationFrontierBridgeMethodology anchor

Rank sensitivity

Stability: volatile; rank range: 364.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

interactive world model

world simulator

model-based reinforcement learning world model

Deep Analysis

Innovations

WorldLens: a unified benchmark measuring world-model fidelity across five complementary aspects and 24 standardized dimensions, covering pixel quality, 4D geometry, closed-loop driving, and human perceptual alignment.
WorldLens-26K: a 26,808-entry human-annotated preference dataset pairing numerical scores with textual rationales.
WorldLens-Agent: a vision-language evaluator distilled from human judgments for scalable, explainable auto-assessment.

Methodology

The paper introduces WorldLens, a benchmark that evaluates driving world models across five aspects (pixel quality, 4D geometry, closed-loop driving, human perceptual alignment, and additional dimensions) totaling 24 standardized metrics. Six representative models are assessed using this benchmark, and a human-annotated preference dataset (WorldLens-26K) is collected to bridge algorithmic metrics with human perception. A vision-language agent (WorldLens-Agent) is then distilled from these annotations to enable automated evaluation.

Key Results

No existing model dominates across all axes: texture-rich models violate geometry, geometry-aware models lack behavioral fidelity, and even the strongest performers achieve only 2–3 out of 10 on human realism ratings.

Limitations

The benchmark is limited to six representative models, which may not cover the full diversity of driving world models.
Human annotations in WorldLens-26K may contain subjectivity and bias, potentially affecting the distilled agent's alignment.
The auto-evaluator (WorldLens-Agent) may not fully capture all nuances of human perception, especially in edge cases.

Resources

Paper