Research map / Reinforcement learning / Offline RL
Prompting Decision Transformer for Few-Shot Policy Generalization
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 a trajectory prompt instead of relying on reward-to-go in recent history (ablation) | Prompt-DT versus MT-ORL · MuJoCo few-shot tasks | average episode accumulated reward | Helped |
| Remove reward-to-go tokens from the history input (ablation) | Prompt-DT (Prompt-MT-BC variant) · Cheetah-dir, Cheetah-vel, Ant-dir, Meta-World reach-v2 | average episode accumulated reward | No clear effect |
| Remove reward-to-go tokens from the history input (ablation) | Prompt-DT (Prompt-MT-BC variant) · Dial | average episode accumulated reward | Hurt |
| Use no prompt augmentation and finetune on target-task data (ablation) | Prompt-DT versus MT-BC-Finetune · MuJoCo few-shot tasks | average episode accumulated reward | Hurt |
| Increase prompt length from K*=2 to K*=5 (ablation) | Prompt-DT · Cheetah-dir | episode return: 926.46 $\pm$ 2.87 -> \textbf{927.20 $\pm$ 18.02} | No clear effect |
| Increase prompt length from K*=2 to K*=40 (ablation) | Prompt-DT · Cheetah-vel | episode return: -45.26 $\pm$ 2.87 -> \textbf{-34.43 $\pm$ 2.33} | Helped |
| Increase prompt length from K*=2 to K*=40 (ablation) | Prompt-DT · Ant-dir | episode return: \textbf{409.81 $\pm$ 9.69} -> 323.83 $\pm $ 9.33 | Hurt |
| Use longer prompts sampled from multiple trajectories (ablation) | Prompt-DT · Ant-dir | episode return: 409.81 $\pm$ 9.69; 367.12 $\pm$ 10.50; 382.94 $\pm$ 25.21; 323.83 $\pm $ 9.33 | Hurt |
| Provide random rather than expert trajectory prompts (ablation) | Prompt-DT trained with expert data · Cheetah-vel | episode return | Hurt |
| Train with random rather than expert data (ablation) | Prompt-DT · Cheetah-vel | episode return | Hurt |
| Increase MT-BC-Finetune adaptation batch size to 1280 and use 100 finetune steps (ablation) | MT-BC-Finetune · Cheetah-dir, Cheetah-vel, Ant-dir | converged performance: 930; -32; 470 | Helped |
| Increase MACAW adaptation batch size (ablation) | MACAW · Cheetah-dir, Cheetah-vel, Ant-dir | performance | Helped |
Closest papers by meaning
- Prompt Tuning Decision Transformers with Structured and Scalable Bandits (NeurIPS 2025 · similarity 0.92)
- Decomposed Prompt Decision Transformer for Efficient Unseen Task Generalization (NeurIPS 2024 · similarity 0.89)
- Rethinking Decision Transformer via Hierarchical Reinforcement Learning (ICML 2024 · similarity 0.89)
- Meta-DT: Offline Meta-RL as Conditional Sequence Modeling with World Model Disentanglement (NeurIPS 2024 · similarity 0.88)
- Task-Adaptive Prompted Transformer for Cross-Domain Few-Shot Learning (AAAI 2024 · similarity 0.88)
- DaCe-DT: Data-Centric Offline Multi-Task Reinforcement Learning via Adaptive Prompts and Trajectory Correction for Heterogeneous Tasks (NeurIPS 2026 · similarity 0.87)
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.