Research map / Vision-language and multimodal / Robot action and navigation
Sparse Imagination for Efficient Visual World Model Planning
Junha Chun, Youngjoon Jeong, Taesup Kim
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 |
|---|---|---|---|
| drop 50% of patch tokens during planning (ablation) | DINO-WM planner · PushT | planning time (s/iter): 173 -> 82 | Improved |
| drop 50% of patch tokens during planning (ablation) | DINO-WM planner · PushT | success rate (%): 75.0 -> 70.0 | Mixed |
| drop 10-50% tokens (ablation) | DINO-WM planner · average of 6 environments | success rate (%): 69.8 -> 69.7 | No clear effect |
| drop 90% of tokens (ablation) | DINO-WM planner · average of 6 environments | success rate (%): 69.8 -> 46.7 | Hurt |
| use sparse patch tokens instead of CLS-only token | CLS-Token world model · Granular | success rate (%): 20.0 -> 85.0 | Improved |
| add sparse imagination planning to SmolVLA at 50% drop | SmolVLA · LIBERO-10 | success rate (%): 29 -> 33 | Improved |
| use 50% drop instead of full patches (ablation) | VLA-guided planner · LIBERO-10 | time per episode (s): 53.4 -> 29.7 | Improved |
| add sparse imagination planning to SmolVLA | SmolVLA · Meta-World (all) | success rate (%): 41.87 -> 47.73 | Improved |
| add sparse imagination planning on real robot | SmolVLA · Real-world PickPlace | success rate (%): 60 -> 80 | Improved |
| train with randomized grouped attention (ablation) | sparse-token world model · 50% token drop prediction | normalized L2 error: 0.036 -> 0.016 | Improved |
| use attention-encoder importance selection instead of random (ablation) | sparse imagination · Wall vision-only | success rate (%): 58.3 -> 21.7 | Hurt |
| use sophisticated token selection (LTRP, STAR, attention) instead of random (ablation) | sparse imagination · average over 6 envs | success rate (%): 66.7 -> 59.5-64.0 | No clear effect |
Closest papers by meaning
- Sparse Planning in Visual World Models via Cost Gradients (NeurIPS 2026 · similarity 0.90)
- DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning (ICML 2025 · similarity 0.87)
- FLIP: Flow-Centric Generative Planning as General-Purpose Manipulation World Model (ICLR 2025 · similarity 0.87)
- World-aware Planning Narratives Enhance Large Vision-Language Model Planner (NeurIPS 2025 · similarity 0.87)
- From Forecasting to Planning: Policy World Model for Collaborative State-Action Prediction (NeurIPS 2025 · similarity 0.87)
- Open-World Reinforcement Learning over Long Short-Term Imagination (ICLR 2025 · similarity 0.86)
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.
