Research map / Multi-agent systems and games / Human-AI and goal-conditioned
WoW!: World Models in a Closed-Loop World
Jiahan Zhang, Muqing Jiang, Nanru Dai, TaiMing Lu, Arda Uzunoglu, Shunchi Zhang, Yana Wei, Jiahao Wang, Vishal M. Patel, Paul Pu Liang, Daniel Khashabi, Cheng Peng and 5 more
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add a proprietary video world model (Runway Gen4) to the VLM policy | VLM policy · Active Recognition (Matterport3D val, 551 episodes) | success rate (%): 50.27 -> 64.79 | Improved |
| post-train Wan2.1 on action-observation data (ablation) | Wan2.1 · Active Recognition (Matterport3D val, 551 episodes) | success rate (%): 58.26 -> 62.61 | Improved |
| post-train Wan2.1 on action-observation data (ablation) | Wan2.1 · Image-Goal Navigation (HM3D val, 144 episodes) | success rate (%): 38.19 -> 45.14 | Improved |
| post-train SVD on action-observation data (ablation) | SVD · Active Recognition (Matterport3D val, 551 episodes) | success rate (%): 57.71 -> 60.98 | Improved |
| post-train LTX-Video on action-observation data (ablation) | LTX-Video · A-EQA (184 questions) | answer score: 46.6 -> 48.6 | Improved |
| add post-trained SVD world model to VLM policy | VLM policy · Robotic manipulation (RLBench, 200 episodes) | success rate (%): 44.5 -> 46.5 | Improved |
| add post-trained SVD world model to 3D diffusion policy | 3D-DP · Robotic manipulation (RLBench, 200 episodes) | success rate (%): 24.0 -> 44.7 | Improved |
| add zero-shot Wan2.1 world model to VLM policy | VLM policy · A-EQA (184 questions) | answer score: 45.7 -> 45.7 | No clear effect |
| increase post-training data from 400 to 80K examples (ablation) | Wan2.1† · Active Recognition (Matterport3D val, 551 episodes) | success rate (%): 60.25% -> 63.34% | Improved |
| increase average world-model inference count from 3 to 11 per episode (ablation) | SVD† · Active Recognition (Matterport3D val, 551 episodes) | success rate (%): 53.36% -> 60.98% | Improved |
| use panorama instead of front-view input for post-training (ablation) | post-trained WMs · Image-Goal Navigation (HM3D val, 144 episodes) | success rate (%): 48.61 -> 45.14 | Mixed |
| use LPIPS-based revision instead of VLM-based revision (ablation) | SVD† · Image-Goal Navigation (HM3D val, 144 episodes) | success rate (%): 43.05 -> 47.92 | Improved |
Closest papers by meaning
- Learning 3D Persistent Embodied World Models (NeurIPS 2025 · similarity 0.90)
- WorldSimBench: Towards Video Generation Models as World Simulators (ICML 2025 · similarity 0.90)
- RIG: Synergizing Reasoning and Imagination in End-to-End Generalist Policy (ICLR 2026 · similarity 0.89)
- DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning (ICML 2025 · similarity 0.89)
- World–Value–Action Model: Implicit Planning for Vision–Language–Action Systems (NeurIPS 2026 · similarity 0.89)
- From Zero to Hero: Training-Free Custom Concept Spawning in World Models (NeurIPS 2026 · similarity 0.89)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
