Research map / Reinforcement learning / Visual and representation RL
Decision Transformers As Zero-Shot Learners via Text-Behavior Alignment
Xin Zhang, Jonathan Martinez, Yanhua Li, Yingxue Zhang
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 both TBC and TBM alignment objectives (ablation) | TG-DT · Ant-dir (zero-shot) | test return: 133.6 -> 383.4 | Improved |
| add both TBC and TBM alignment objectives (ablation) | TG-DT · Cheetah-dir (zero-shot) | test return: 859.4 -> 958.4 | Improved |
| add text-behavior contrastive loss (TBC) (ablation) | TG-DT · Ant-dir (zero-shot) | test return: 322.1 -> 383.4 | Improved |
| add text-behavior matching loss (TBM) (ablation) | TG-DT · Ant-dir (zero-shot) | test return: 298.7 -> 383.4 | Improved |
| add text-behavior matching loss (TBM) (ablation) | TG-DT · Cheetah-dir (zero-shot) | test return: 875.2 -> 958.4 | Improved |
| share 1 or 2 retrieved related-task trajectories for fine-tuning (K=1 or 2) (ablation) | TG-DT · Ant-dir and ML45 | test return | Improved |
| increase shared demonstration count K beyond a small value (ablation) | TG-DT · data sharing | test return | No clear effect |
| text-only zero-shot TG-DT versus DT baselines using test-time interaction | TG-DT · Medium datasets zero-shot (5 tasks) | test return: Cheetah-dir 549.9 vs 548.9 (PDT) | No clear effect |
| condition on text descriptions instead of language imitation without return conditioning (BC-Z) | TG-DT · ML45 zero-shot Medium | test return: 157.4 -> 309.6 | Improved |
| switch from Expert to Mixed data | TG-DT · Cheetah-dir zero-shot | test return: 958.4 -> 830.1 | Hurt |
| TG-DT versus DPDT on Mixed Cheetah-dir | TG-DT · Cheetah-dir Mixed zero-shot | test return: 861.2 -> 830.1 | Hurt |
Closest papers by meaning
- Text-to-Decision Agent: Offline Meta-Reinforcement Learning from Natural Language Supervision (NeurIPS 2025 · similarity 0.93)
- Meta-DT: Offline Meta-RL as Conditional Sequence Modeling with World Model Disentanglement (NeurIPS 2024 · similarity 0.91)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.91)
- Generalizable Task Representation Learning for Offline Meta-Reinforcement Learning with Data Limitations (AAAI 2024 · similarity 0.90)
- RLZero: Direct Policy Inference from Language Without In-Domain Supervision (NeurIPS 2025 · similarity 0.90)
- Text-Aware Diffusion for Policy Learning (NeurIPS 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.
