Research map / Reinforcement learning / Offline RL
MetaDiffuser: Diffusion Model as Conditional Planner for Offline Meta-RL
Fei Ni, Jianye HAO, Yao Mu, Yifu Yuan, YAN ZHENG, Bin Wang, Zhixuan Liang
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 |
|---|---|---|---|
| Use task-conditioned diffusion planning instead of the strongest listed baseline | Offline meta-RL · Ant-Dir | average test return: 213.2 -> 247.7 | Improved |
| Use task-conditioned diffusion planning instead of the strongest listed baseline | Offline meta-RL · Walker-Param | average test return: 300.5 -> 368.3 | Improved |
| Use task-conditioned diffusion planning instead of the strongest listed baseline | Offline meta-RL · Point-Robot, Cheetah-Dir, Cheetah-Vel, Hopper-Param | average test return: -5.04 -> -4.48; 931.7 -> 936.2; -51.3 -> -45.9; 289.3 -> 356.4 | Improved |
| Add dynamics guidance alongside reward guidance (ablation) | MetaDiffuser · Ant-Dir, Cheetah-Vel, Walker-Param, Hopper-Param | average test return: 210.3 -> 247.7; -52.9 -> -45.9; 326.5 -> 368.3; 293.3 -> 356.8 | Improved |
| Add dynamics guidance alongside reward guidance (ablation) | MetaDiffuser · Point-Robot, Cheetah-Dir | average test return: -4.57 -> -4.48; 924.2 -> 936.2 | Mixed |
| Replace conditional diffusion with a conditioned VAE planner (ablation) | MetaDiffuser · MuJoCo and Point-Robot tasks | average test return: -4.48 -> -6.11; 247.7 -> 130.6; 936.2 -> 759.4; -45.9 -> -87.1; 368.3 -> 187.9; 356.4 -> 157.3 | Hurt |
| Remove task conditioning from the diffusion model (ablation) | MetaDiffuser · MuJoCo and Point-Robot tasks | average test return: -4.48 -> -7.48; 247.7 -> 53.2; 936.2 -> 614.7; -45.9 -> -121.7; 368.3 -> 135.2; 356.4 -> 103.5 | Hurt |
| Use CORRO task representations instead of the paper's context representation (ablation) | MetaDiffuser · MuJoCo and Point-Robot tasks | average test return | Mixed |
| Use ground-truth task parameters as context (ablation) | MetaDiffuser · MuJoCo and Point-Robot tasks | average test return | Improved |
| Use medium or random warm-start data instead of expert data (ablation) | MetaDiffuser · Six reported environments | average test return: Expert -> Random: -4.48 -> -4.61; 247.7 -> 213.8; 936.2 -> 916.7; -45.9 -> -55.8; 368.3 -> 341.6; 356.4 -> 319.6 | Hurt |
| Use medium or random warm-start data instead of expert data (ablation) | Prompt-DT · Ant-Dir, Cheetah-Vel, Walker-Param, Hopper-Param | average test return: 213.2 -> 40.1; -51.3 -> -208.4; 287.7 -> 64.7; 265.2 -> 82.6 | Hurt |
| Train the context encoder on mixed expert and random data instead of expert data (ablation) | MetaDiffuser · Ant-Dir, Cheetah-Vel, Walker-Param, Hopper-Param | average test return: 247.7 -> 258.4; -45.9 -> -43.4; 368.3 -> 381.5; 356.4 -> 375.2 | Improved |
Later papers that compared with this method
1 later accepted paper reports MetaDiffuser as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Ada-Diffuser: Latent-Aware Adaptive Diffusion for Decision-Making (ICLR 2026 · better on 1 of 1 shared result)
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Stitching Sub-trajectories with Conditional Diffusion Model for Goal-Conditioned Offline RL (AAAI 2024 · similarity 0.92)
- Task-Agnostic Pre-training and Task-Guided Fine-tuning for Versatile Diffusion Planner (ICML 2025 · similarity 0.90)
- Meta-DT: Offline Meta-RL as Conditional Sequence Modeling with World Model Disentanglement (NeurIPS 2024 · similarity 0.89)
- Simple Hierarchical Planning with Diffusion (ICLR 2024 · similarity 0.89)
- Prior-Guided Diffusion Planning for Offline Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- MetaCARD: Meta-Reinforcement Learning with Task Uncertainty Feedback via Decoupled Context-Aware Reward and Dynamics Components (AAAI 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.
