ROWBench: Do Video Models Render What the Program Specifies?

arXiv 2026 41.4 benchmark

TLDR

Introduces PROWBench, a benchmark with 170 programmatic episodes and 600 proxy videos evaluating whether video models visually adhere to fine-grained program-specified world events.

Reasoning

The benchmark addresses an important gap by using replayable world records and VLM-based metrics to test fidelity to program-specified interactions. However, the evaluation relies on synthetic/proxy videos rather than real-world data, and there is an inconsistency between the title's ROWBench and the abstract's PROWBench.

Read-first score

Read-first score 41.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 45.

Methodology quality 15%

100

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

Recency 5%

100

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

Topical relevance 25%

64.3

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 15%

Screens links and visible text for paper, code, dataset, artifact, and repository signals. pdf=True; code=False; dataset=False; markers=none

Graph prestige 15%

5.7

PageRank-style prestige over the local citation graph built from available OpenAlex references. local_pagerank=0.0016

Citation impact 15%

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

Citation velocity 10%

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: 647.

Keyword Scores

world model

video world model

interactive world model

generative world model

world simulator

world dynamics prediction

model-based reinforcement learning world model

Deep Analysis

Innovations

Introduces PROWBench, a benchmark for evaluating visual adherence to explicit, fine-grained program-specified world events in programmable world models.
Provides 170 programmatically constructed episodes and 600 proxy videos covering diverse scenes and interactions.
Logs entity states and timestamped events, including events outside the camera's field of view, as replayable world records with synchronized views and proxy representations.
Uses an extensible framework that constructs scenes, controls behaviors, and renders each camera view in multiple representations such as coarse 3D and bounding boxes.
Supports first- and third-person perspectives, with synchronized multi-view observations for a subset of episodes.
Defines evaluation dimensions of entity control and long-horizon memory, plus two VLM-based metrics: Logic-Render Alignment and Interaction Success Rate.

Methodology

PROWBench programmatically constructs episodes and logs entity states and timestamped events as replayable world records, including off-camera events. From these records it renders synchronized camera views and proxy representations such as coarse 3D and bounding boxes. Generated videos are then checked against the observable consequences of program execution using entity control, long-horizon memory, and the VLM-based Logic-Render Alignment and Interaction Success Rate metrics.

Key Results

The title and abstract do not report explicit quantitative experimental results; they describe the benchmark construction, coverage, and evaluation metrics.

Limitations

Synchronized multi-view observations are available only for a subset of episodes.

Resources

Paper