Research map / Optimization / Optimizers
Parallel Stochastic Gradient-Based Planning for World Models
Michael Psenka, Michael Rabbat, Aditi S. Krishnapriyan, Yann LeCun, Amir Bar
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 periodic full-rollout GD sync steps (ablation) | GRASP · Push-T H=40 | accuracy (%): 1.6 -> 59.0 | Improved |
| use 25 sync steps rather than 10 (ablation) | GRASP · Push-T H=40 | accuracy (%): 48.0 -> 59.0 | Improved |
| increase sync steps from 25 to 50 (ablation) | GRASP · Push-T H=40 | accuracy (%): 59.0 -> 58.0 | No clear effect |
| add Langevin state noise (sigma 0.5 vs none) (ablation) | GRASP · Push-T H=40 | accuracy (%): 54.8 -> 59.0 | Improved |
| raise noise from sigma 0.5 to 1.0 (ablation) | GRASP · Push-T H=40 | accuracy (%): 59.0 -> 50.6 | Hurt |
| stop gradients through world model state inputs (ablation) | GRASP · Push-T H=40 | accuracy (%): 46.6 -> 59.0 | Improved |
| GRASP versus CEM and GD at long horizon | GRASP · Push-T H=60 | success rate (%): 7.2 (CEM), 16.4 (GD) -> 26.2 | Improved |
| GRASP versus CEM at short horizon H=40 | GRASP · Push-T H=40 | success rate (%): 61.4 (CEM) vs 59.0 | No clear effect |
| GRASP compared to baselines at short horizons | GRASP · Push-T, PointMaze, WallSingle H=10-30 | success rate (%): Push-T H=20: 96.0 (CEM) vs 89.2 | No clear effect |
| replace GD sync with small CEM sync | GRASP on le-wm PushT · le-wm PushT H=75 | success rate: 26% (GD) -> 46% | Improved |
| apply GRASP to le-wm Reacher | GRASP · le-wm Reacher H=75 | success rate: 42% (GD) -> 70% | Improved |
| linearly decay state noise and goal coefficient between syncs | GRASP · jepa-wm Metaworld H=30 | success rate: 12% (CEM) -> 19% | Improved |
Closest papers by meaning
- Sparse Planning in Visual World Models via Cost Gradients (NeurIPS 2026 · similarity 0.89)
- From Forecasting to Planning: Policy World Model for Collaborative State-Action Prediction (NeurIPS 2025 · similarity 0.88)
- HaM-World: Soft-Hamiltonian World Models with Selective Memory for Planning (NeurIPS 2026 · similarity 0.87)
- PWM: Policy Learning with Multi-Task World Models (ICLR 2025 · similarity 0.87)
- DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning (ICML 2025 · similarity 0.87)
- Predictive but Not Plannable: RC-aux for Latent World Models (NeurIPS 2026 · 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.
