WorldMemBench

A Comprehensive Benchmark for Memory of Video World Models

Nan Zhou1,3,*, Jianxiong Gao2,3,*, Qiuyu Wang3, Wen Wang1, Jiahao Wang4,3, Yufeng Yuan4,3, Shuaiting Li1,3, Hao Ouyang3, Yujun Shen3, Weiming Lu1, Jun Xiao1, Yueting Zhuang1, YongLiang Shen1,†

1 Zhejiang University   2 Fudan University   3 Ant Group   4 Nanjing University

Paper GitHub Leaderboard Datasets
260Cases
6,705Videos
20–35 sRollouts
9Models
Overview
01 / OVERVIEW

Overview

As video world models enable longer, more realistic interactive experiences, maintaining a world consistent with its history becomes essential. WorldMemBench evaluates world memory across 260 cases, using controlled camera departures and multi-revisit trajectories to measure static retention and dynamic state evolution. Across nine representative models, strong static retention does not imply reliable dynamic evolution, while later revisits expose degradation beyond the first return.

Overview of WorldMemBench case composition, multi-revisit evaluation, and capability profiles
02 / CONTRIBUTIONS

Contributions

01

A benchmark for world memory

We introduce WorldMemBench, a comprehensive benchmark for the structured evaluation of memory capabilities in video world models.

02

A diagnostic evaluation protocol

We develop a multi-revisit evaluation protocol to assess static and dynamic memory alongside camera controllability.

03

An empirical characterization of world memory

We evaluate nine video world models, revealing distinct memory profiles, cumulative degradation across revisits, and sensitivity to exploration direction and context access.

03 / LEADERBOARD

Model results

We evaluate nine representative video world models across General Video Quality, Static Memory, Dynamic Memory, and Camera Following. The table reports aggregate results from the paper, including four Dynamic Memory sub-metrics.

04 / QUALITATIVE COMPARISON

Compare the same world on return.

Three matched cases across nine models compare the initial view (V) with three successive returns (R1–R3). Select a case, return, and models; each score summarizes one revisit segment.

Choose models
GENERATION PROMPT

05 / BENCHMARK COMPARISON

Benchmark coverage

Input:camera-driven · action-driven · language-driven
Coverage: explicit partial not reported
BenchmarkInputLength (s)Camera
control		Visual
quality		Long-horizon
revisit		State
evolution		Physical
consistency		Multi-
revisit
WorldScoreC·L~2–10✓✓————
MINDA~5–10✓✓◐———
WorldMarkC·A·L20–60✓✓✓———
WBenchC·A·L~10–45✓✓✓◐◐—
MBenchC·A·L~25✓—✓✓✓—
MemoBenchC·A·L~4–6✓✓—✓✓—
WorldRoamBenchC·A·L10–60✓✓✓◐◐—
WorldExamC·A·L~3–12✓✓◐◐◐—
R2M-BenchC·A~30——✓———
STEVO-BenchC·L~5–30——◐✓✓—
WRBenchC·A·L~4–10✓—◐✓◐—
LoopBenchC·A~3–24——◐——✓
WorldMemBenchC·A·L~20–35✓✓✓✓✓✓
06 / CITATION

Citation

If you find WorldMemBench useful for your research, please consider citing our paper.

@misc{zhou2026worldmembenchcomprehensivebenchmarkmemory,
  title={WorldMemBench: A Comprehensive Benchmark for Memory of Video World Models},
  author={Nan Zhou and Jianxiong Gao and Qiuyu Wang and Wen Wang and
          Jiahao Wang and Yufeng Yuan and Shuaiting Li and Hao Ouyang and
          Yujun Shen and Weiming Lu and Jun Xiao and Yueting Zhuang and
          Yongliang Shen},
  year={2026},
  eprint={2610.19063},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2610.19063},
}