World Model Benchmarks · 2018–2026

A Survey of
World Model Benchmarks

An evaluation-centric view of what capability is evaluated, how the model is evaluated, which metrics are used, and where the data come from.

Zheng Zhang, Zhixiang Guo, Yuan Si, Siyuan Liang, Shunyu Liu, Weibo Gao, Song Wang, Leszek Rutkowski, Giuseppe Valenzise, Ming-Hsuan Yang, Lin William Cong, M. Jamal Deen, Sally Cripps, Dacheng Tao

Benchmark list Explore benchmarks GitHub

Final manuscript snapshot · 102 benchmarks · 85 cross-category · corpus checked August 31, 2026 · coding finalized September 15, 2026

102representative benchmarks

85cross-category benchmarks

7evaluation targets

2 / 2protocol classes / metric levels

4evaluation data sources

Overview

Evaluation claims need four coordinates.

Two benchmarks can report strong performance while supporting very different claims. The survey separates capability target, evaluation protocol, metric family, and data source so results can be interpreted on comparable terms.

Evaluation Target

Specifies the property of world-state evolution or downstream use being characterized.

Evaluation Protocol

Distinguishes open-loop evaluation from feedback-driven closed-loop interaction.

Evaluation Metrics

Separates measurements of direct outputs from the performance or reliability of downstream functions. Closed-loop collection alone does not make a metric downstream.

Evaluation Data

Identifies whether benchmark cases come from real collection, simulation, curation, or hybrid construction.

Taxonomy

Target × Protocol × Metrics × Data

Click any taxonomy item or matrix cell to open the corresponding subset in the benchmark explorer.

D1 · Evaluation Target

What capability is evaluated?

D2 · Evaluation Protocol

How is the model evaluated?

D3 · Evaluation Metrics

Which metrics are used?

D4 · Evaluation Data

Where do the data come from?

Cross-dimensional map

Evaluation Target × Protocol

Counts use only tracks relevant to each target. OL+CL is included in both columns; categories can overlap.

Cross-dimensional map

Evaluation Target × Metrics

Only tracks relevant to each target are counted. P+O appears in both columns.

Manuscript Table 4

Target-specific protocol and evidence coverage

Each benchmark is counted once per row, using only tracks relevant to that target or role. Includes CL and Includes O are separate marginal counts over records relevant to each target or role.

Evaluation target	n	OL only	Includes CL	P only	Includes O
Control Fidelity & Interactive Dynamics	55	51 (92.7%)	4 (7.3%)	51 (92.7%)	4 (7.3%)
Pre-specified Control Fidelity	40	39 (97.5%)	1 (2.5%)	38 (95.0%)	2 (5.0%)
Interactive Action Fidelity	15	12 (80.0%)	3 (20.0%)	13 (86.7%)	2 (13.3%)
Functional Utility	13	2 (15.4%)	11 (84.6%)	0 (0.0%)	13 (100.0%)

Timeline

From “Does the generated world look reasonable?” to “Is it really useful to the agent?”

The corpus traces a shift from visual and temporal quality, physical plausibility, and causal and counterfactual reasoning toward spatial and state consistency, long-horizon memory and state persistence, control fidelity and interactive dynamics, and functional utility.

Evaluation target counts by release window

Counts overlap because cross-category benchmarks appear in more than one target.

Four-phase conclusions

Early benchmarks mainly focus on physical plausibility and causal reasoning.

10 cumulative benchmarks by 2023, concentrated in physical and causal evaluation.
From 2024 onward, control fidelity & interactive dynamics and functional utility begin to emerge.

10 new benchmarks in 2024, including broader control and functional-utility evaluation.
In 2025, benchmark growth accelerates, with broader coverage of spatial consistency, physical plausibility, and control fidelity.

30 new benchmarks in 2025, raising the cumulative corpus to 50.
By 2026, long-horizon memory & persistence and functional utility become prominent frontier directions.

52 new benchmarks in 2026, bringing the corpus to 102.

Benchmark Explorer

Browse the latest paper snapshot.

Search and filter the 102 representative benchmarks coded in Supplementary Figure S1 and Tables S1–S7 of the final manuscript.

Target, protocol, metrics, and data filters must match within the same task-supported combination. Selections within a filter group use OR; selections across groups use AND. Expand a card for its exact target/role combinations.

102 results

△ cross-category in the survey

Community

Help keep the benchmark map current.

Suggest a missing benchmark, report a metadata issue, or discuss a taxonomy placement through the project repository.

Suggest a benchmark

Share a paper, benchmark suite, project page, and your suggested Target–Protocol–Metrics–Data coding.

Open a suggestion ↗

Correct metadata

Flag an incorrect year, domain, protocol, metric level, data source, title, category, or link.

Report a correction ↗

Reuse the taxonomy

Download a normalized, machine-readable snapshot aligned with the latest manuscript.

Citation

Cite the survey.

The author list below follows the final manuscript; the publication identifier can be updated when available.

@article{world_model_benchmarks_survey_2026,
  title   = {A Survey of World Model Benchmarks},
  author  = {Zheng Zhang and Zhixiang Guo and Yuan Si and Siyuan Liang and Shunyu Liu and Weibo Gao and Song Wang and Leszek Rutkowski and Giuseppe Valenzise and Ming-Hsuan Yang and Lin William Cong and M. Jamal Deen and Sally Cripps and Dacheng Tao},
  year    = {2026},
  note    = {Manuscript}
}