WorldBench: Disambiguating Physics for Diagnostic Evaluation of World Modelsl

arXiv 26.1 2026 58.4 benchmark

TLDR

WorldBench is a video benchmark for disentangled evaluation of world models' understanding of individual physics concepts, revealing failures in SOTA models.

Reasoning

The paper introduces a novel benchmark that isolates physical concepts for diagnostic evaluation, addressing entanglement in existing benchmarks. Its strength lies in providing nuanced failure analysis, but it is limited to video-based world models and does not cover interactive or reinforcement learning contexts.

Read-first score

Read-first score 58.4, weighted from topical fit, citation, graph, method, reproducibility, and recency signals. Original total remains 43.

Recency 8%

100

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

Topical relevance 42%

61.4

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

Methodology quality 25%

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

Reproducibility 25%

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

Field roles

Frontier

Rank sensitivity

Stability: volatile; rank range: 347.

Keyword Scores

world model

video world model

generative world model

world dynamics prediction

world simulator

interactive world model

model-based reinforcement learning world model

Deep Analysis

Innovations

Concept-specific, disentangled evaluation of physical understanding in world models
Two-level benchmark design: intuitive physical understanding (e.g., object permanence, scale/perspective) and low-level physical constants/material properties (e.g., friction coefficients, fluid viscosity)

Methodology

WorldBench is a video-based benchmark that isolates a single physical concept or law per test, enabling diagnostic evaluation. It comprises two levels: one for intuitive physics and one for low-level constants. State-of-the-art video-based world models are evaluated on this benchmark to assess their physical fidelity.

Key Results

All tested world models exhibit specific patterns of failure in particular physics concepts and lack the physical consistency required for reliable real-world interactions.

Resources

Paper Project