Video World Models · 2026-10-30
WorldProbe: When Video World Models Fall Apart under Long-Horizon Spatial Exploration
Minqi Wang, Jie Peng, Hejun Dong, Ruchang Yao, Puyi Wang, Siyuan Huang, Jihai Zhang, Yafu Li, Qi Dou, Xiaoye Qu, Yu Cheng
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
WorldProbe is a diagnostic benchmark for video world models under long-horizon spatial exploration. It uses a goal-conditioned multimodal agent to adapt camera actions online during generation, with 164 test cases across four probing families. Across 10 state-of-the-art models, the authors run 1,600+ rollouts totaling ~800 minutes with horizons up to 180s, finding local quality degradation and long-range inconsistencies, including weak long-range memory and evolution.
Why it matters
The authors argue existing benchmarks under-test persistent world consistency because they use short rollouts, fixed camera trajectories, and mostly local metrics. WorldProbe provides a closed-loop protocol and evaluation dimensions (visual/content quality, long-range memory, evolution fidelity) designed to expose long-range failures that short-horizon benchmarks miss.
Method
- Closed-loop probing: a goal-conditioned multimodal agent selects discrete camera actions at each step based on task spec, recent frames, and auxiliary spatial states; generation proceeds in chunks with feedback.
- Four scene-adaptive probing families (geometry-constrained approach, open-ended traversal, cross-view inspection, successive region exploration) with task prompts and online adaptation of camera motion.
- Evaluation uses local measures (no-reference visual IQA; VLM-rated content/relations) and long-range metrics for memory and evolution, including pose-sequence alignment (ViPE) and VLM scoring across revisited multi-view anchors.
Limitation
The authors report that “In extended rollouts, the probing loop remains stable for up to 5 minutes and ends only when the generated video itself collapses,” implying evaluation is limited by generation collapse.
Abstract (from arXiv)
Video world models generate future visual observations from camera or action inputs, with the goal of synthesizing persistent 3D worlds that can be explored across time and space. Existing benchmarks typically evaluate short interactions or rely on camera trajectories fixed before generation. The performance of video world models under long-horizon exploration with extensive camera motion remains under-explored. We introduce WorldProbe, a diagnostic benchmark that uses a goal-conditioned multimodal agent to adapt camera actions online during generation. WorldProbe comprises 164 test cases across four probing families. Across 10 state-of-the-art models, we generate more than 1,600 rollouts totaling approximately 800 minutes, with horizons up to 180 seconds. Our evaluation across four dimensions reveals both the degradation of local measures and the emergence of long-range inconsistencies. We find that recent models sustain visual and content quality for longer, while long-range memory and evolution remain weak. These results expose a gap between locally stable generation and persistent world consistency that is largely missed by existing benchmarks. Code and benchmark resources are available at https://github.com/Untitled-W/WorldProbe.
Related papers
- WorldMemBench: A Comprehensive Benchmark for Memory of Video World Models
- Orbia: Do Generated Videos Form a Consistent 3D World?
- StateCraft: Augmenting Video World Models with Abstract-State Reasoning
- ActionRoPE: action as coordinates for spatially consistent control in planar-camera game world models
- World-T3: Test-Time Training as Persistent Memory for Real-Time Video World Models
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First announced by arXiv: 2026-10-30 · Indexed by Pub-AI: 2026-10-30 UTC
WorldProbe: When Video World Models Fall Apart under Long-Horizon Spatial Exploration
First announced by arXiv: 2026-10-30 · Original record ↗
- arxiv
- 2610.39721
Explicit source versions (1)
- v1 · Date not supplied
- Preprint available
Source date: 2026-10-30 · First detected: 2026-10-30 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2610.39721v1; arxiv:announce_type=new; item pubDate (official metadata)
Abstract · arxiv
Video world models generate future visual observations from camera or action inputs, with the goal of synthesizing persistent 3D worlds that can be explored across time and space. Existing benchmarks typically evaluate short interactions or rely on camera trajectories fixed before generation. The performance of video world models under long-horizon exploration with extensive camera motion remains under-explored. We introduce WorldProbe, a diagnostic benchmark that uses a goal-conditioned multimodal agent to adapt camera actions online during generation. WorldProbe comprises 164 test cases across four probing families. Across 10 state-of-the-art models, we generate more than 1,600 rollouts totaling approximately 800 minutes, with horizons up to 180 seconds. Our evaluation across four dimensions reveals both the degradation of local measures and the emergence of long-range inconsistencies. We find that recent models sustain visual and content quality for longer, while long-range memory and evolution remain weak. These results expose a gap between locally stable generation and persistent world consistency that is largely missed by existing benchmarks. Code and benchmark resources are available at https://github.com/Untitled-W/WorldProbe.