WorldMemBench from Zhejiang, Fudan and Ant tests memory across 9 video world models
jiqizhixin · x · 2026-10-30
WorldMemBench, a new benchmark from Zhejiang University, Fudan University, Ant Group and Nanjing University, evaluates memory in video world models across 260 cases in 23 categories, testing nine models including HY-WorldPlay, SANA-WM, Matrix-Game, YUME and LingBot World.
Key design points:
- Leave, then return: controlled camera paths revisit scenes; static tests check whether identity, appearance and geometry survive the gap, dynamic tests check whether changes persist and events continue coherently.
- Memory vs. visual polish: 200 static and 60 dynamic cases include video-quality and camera-following diagnostics to separate forgetting from a model failing to follow the requested view.
- Repeated revisits expose weaknesses: average results decline across three revisits on all seven static-memory metrics, showing current models struggle to maintain world state over time.
Evaluation dimensions span general video quality, static memory (human ID, object appearance/geometry, environment reconstruction), dynamic memory (return progress, persistence/reset, causal physics) and camera-following accuracy.
More from Multimodal
- EveryFrame launches on Bittensor subnet 117: 25 video/image models behind one API from $0.04 per clip — markjeffrey · 2026-10-30
- Pika H3 Flex tested: character swapping and motion transfer with no LoRA needed — linoy_tsaban · 2026-10-30
- PrinGS anonymizes faces in Gaussian Splatting scenes without original images, ICVGIP 2026 — ponguru · 2026-10-30
- Zero-setup in-browser face swap workflow: no GPU or node graphs needed — Vast-Sympathy-5015 · 2026-10-30
- DeepGHS teams with Noob's Laxhar Lab on next-gen open-source anime base model — bdsqlsz · 2026-10-30
- Fully AI-generated Eurodance music video "There Is No Anti-Slopcore Division" drops — SammaVaco · 2026-10-30