Research map / Image generation / Fast diffusion sampling
Model-Based Diffusion Sampling for Predictive Control in Offline Decision Making
Haldun Balim, Na Li, Yilun Du
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 |
|---|---|---|---|
| alternate dynamics-model and planner denoising steps instead of planner alone (ablation) | MPDiffuser · D4RL medium-replay (3 tasks) | average normalized return: 69.3 -> 74.4 | Improved |
| give planner twice the denoising steps (200 vs 100) (ablation) | Planner · D4RL medium-replay (3 tasks) | average normalized return: 69.3 -> 68.8 | No clear effect |
| alternate planner and dynamics updates instead of combining their scores (ablation) | MPDiffuser · D4RL Hopper | average normalized return: 98.2 -> 102.4 | Improved |
| add a ranker selecting among 64 samples (ablation) | MPDiffuser · D4RL Locomotion average | normalized return: 84.9 -> 85.1 | Improved |
| add ranker (ablation) | MPDiffuser · D4RL Kitchen partial | normalized return: 67.9 -> 73.8 | Improved |
| add cheap random trajectories for dynamics model training (ablation) | MPDiffuser · FetchPickAndPlace | success rate: 0.75 -> 0.86 | Improved |
| add random trajectories for dynamics model training (ablation) | D-MPC · FetchPickAndPlace | success rate: 0.60 -> 0.60 | No clear effect |
| fine-tune only the dynamics model on 100 play episodes under defect | MPDiffuser · Walker2d-medium with ankle defect | normalized return: 58.6 -> 66.4 | Improved |
| add dynamics model to planner under defect (ablation) | Walker2d-medium with ankle defect · Walker2d-medium defect, post fine-tuning | normalized return: 56.0 -> 66.4 | Improved |
| increase number of ranked samples to 64 without dynamics feasibility (ablation) | Planner · Pendulum with velocity constraint | success rate: 89 -> 66 (4 to 64 samples) | Hurt |
| condition dynamics model on task information (ablation) | MPDiffuser · D4RL medium-replay (3 tasks) | average normalized return: 71.1 -> 74.4 | Improved |
| use acausal instead of causal denoisers (ablation) | MPDiffuser · D4RL medium-replay (3 tasks) | average normalized return: 69.0 -> 74.4 | Improved |
Closest papers by meaning
- Model-based Diffusion for Trajectory Optimization (NeurIPS 2024 · similarity 0.91)
- Generative Trajectory Stitching through Diffusion Composition (NeurIPS 2025 · similarity 0.90)
- DiffuserLite: Towards Real-time Diffusion Planning (NeurIPS 2024 · similarity 0.89)
- Prior-Guided Diffusion Planning for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- DiRecT: Safe Diffusion-Based Planning via Receding-Horizon Denoising (NeurIPS 2026 · similarity 0.89)
- Simple Hierarchical Planning with Diffusion (ICLR 2024 · 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.
