Research map / Architectures and efficiency / Mixture of experts
One Model for All Tasks: Leveraging Efficient World Models in Multi-Task Planning
Yuan Pu, Yazhe Niu, Jia Tang, Junyu Xiong, Shuai Hu, Hongsheng Li
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 |
|---|---|---|---|
| replace dense FFN with sparse MoE backbone in UniZero (ablation) | UniZero · Atari8 multitask | return / plasticity metrics | Improved |
| add explicit learnable task embedding (ablation) | UniZero · Atari8 multitask | return | No clear effect |
| replace ResNet encoder with ViT (ablation) | UniZero · Atari8 multitask | return | No clear effect |
| replace LayerNorm with SimNorm (ablation) | UniZero · Atari8 multitask | return | Mixed |
| add MoCo gradient correction (ablation) | UniZero · multitask suite | return and training time: approximately 40% | Hurt |
| use frozen DINOv2 encoder instead of from-scratch ViT (ablation) | ScaleZero · Atari | return | No clear effect |
| train multitask ScaleZero instead of multitask UniZero | UniZero (MT) · Atari 100k (26 games) | human-normalized mean: 0.31 -> 0.39 | Improved |
| multitask ScaleZero vs 26 single-task UniZero | UniZero (ST) · Atari 100k (26 games) | human-normalized median: 0.21 -> 0.16 | Hurt |
| multitask ScaleZero vs 18 single-task UniZero | UniZero (ST) · DMControl (18 tasks) | mean return: 787.2 -> 769.7 | Hurt |
| multitask ScaleZero vs 18 single-task UniZero | UniZero (ST) · DMControl (18 tasks) | median return: 875.1 -> 887.3 | Improved |
| multitask ScaleZero vs single-task UniZero | UniZero (ST) · Jericho Detective | average return: 295 -> 280 | Hurt |
| add Dynamic Parameter Scaling (LoRA stages) | ScaleZero · DMControl (18 tasks) | environment interactions: 71.5% | Improved |
Closest papers by meaning
- Mastering Massive Multi-Task Reinforcement Learning via Mixture-of-Expert Decision Transformer (ICML 2025 · similarity 0.89)
- Representation Learning Enables Scalable Multitask Deep Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- PWM: Policy Learning with Multi-Task World Models (ICLR 2025 · similarity 0.87)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.86)
- Mixture-of-Experts Meets In-Context Reinforcement Learning (NeurIPS 2025 · similarity 0.86)
- Task-Agnostic Pre-training and Task-Guided Fine-tuning for Versatile Diffusion Planner (ICML 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.
