PhyGround: Benchmarking Physical Reasoning in Generative World Models

arXiv 2026 59.6 benchmark

TLDR

PhyGround benchmarks physical reasoning in generative world models via 250 prompts, 13 physical laws, and a large-scale human study with automated evaluator.

Reasoning

The paper's strength lies in its rigorous benchmark design with per-law diagnostics and a large-scale, quality-controlled human study (459 annotators, high split-half reliability). Weaknesses include potential limited coverage of physical laws and lack of explicit discussion on generalizability beyond the selected laws.

Read-first score

Read-first score 59.6, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 46.

Recency 6%

100

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

Citation impact 18%

75.1

Uses OpenAlex-shaped citation metadata as a bibliometric attention signal, separate from paper quality. citation_normalized_percentile=0.75132085

Methodology quality 18%

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

Topical relevance 29%

65.7

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=checkpoint,code,github

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: 337.

Keyword Scores

world model

generative world model

video world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Criteria-grounded benchmark with 250 curated prompts, each augmented with an expected physical outcome
Taxonomy of 13 physical laws across solid-body mechanics, fluid dynamics, and optics, operationalized through observable sub-questions for per-law diagnostics
Large-scale, quality-controlled human study grounded on social science lab experiment design, with 459 annotators, 5,796 complete annotations, and over 37.4K fine-grained labels
PhyJudge-9B, an open physics-specialized VLM judge with substantially lower aggregate relative bias (3.3%) compared to Gemini-3.1-Pro (16.6%)

Methodology

PhyGround consists of 250 curated prompts, each paired with an expected physical outcome, and a taxonomy of 13 physical laws (solid-body mechanics, fluid dynamics, optics) operationalized through observable sub-questions. Eight modern video generation models are evaluated via a large-scale human study with 459 annotators providing 5,796 complete annotations and over 37.4K fine-grained labels, followed by quality control that yields high split-half model-ranking correlations (Spearman's rho > 0.90). An open physics-specialized VLM judge, PhyJudge-9B, is released for reproducible automated evaluation.

Key Results

The human annotations after quality control exhibit high split-half model-ranking correlations (Spearman's rho > 0.90). PhyJudge-9B achieves substantially lower aggregate relative bias (3.3%) than Gemini-3.1-Pro (16.6%).

Resources

Paper