Research map / Reinforcement learning / Visual and representation RL
Transformers are Meta-Reinforcement Learners
Luckeciano C Melo
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 |
|---|---|---|---|
| remove T-Fixup initialization (ablation) | TrMRL · HalfCheetahVel | meta-training return | Hurt |
| remove T-Fixup initialization (ablation) | TrMRL · MetaWorld ML1 Reach-v2 | success rate | No clear effect |
| longer working-memory sequence (up to N=50) (ablation) | TrMRL · MetaWorld ML1 Reach-v2 | sample efficiency | Improved |
| longer working-memory sequence beyond N=5 (ablation) | TrMRL · HalfCheetahVel | sample efficiency | Hurt |
| increase transformer layers (ablation) | TrMRL · MetaWorld ML1 Reach-v2 | meta-training performance | Improved |
| increase layers from 1 to 4 or 8 (ablation) | TrMRL · HalfCheetahVel | sample efficiency | Improved |
| increase layers to 12 (ablation) | TrMRL · HalfCheetahVel | sample efficiency | Hurt |
| more attention heads (ablation) | TrMRL · HalfCheetahVel | meta-training performance | Improved |
| more attention heads (ablation) | TrMRL · MetaWorld ML1 Reach-v2 | meta-training performance | No clear effect |
| TrMRL vs PEARL on out-of-distribution target velocities [3,4] | PEARL/VariBAD/RL2/MAML · HalfCheetahVel OOD | return | Improved |
| TrMRL instead of VariBAD in high-ambiguity task | VariBAD · MetaWorld Push-v2 | meta-training success rate | Hurt |
| TrMRL vs RL2 on ML45-Train | RL2 · MetaWorld ML45-Train | meta-training success rate | Improved |
Closest papers by meaning
- Meta-DT: Offline Meta-RL as Conditional Sequence Modeling with World Model Disentanglement (NeurIPS 2024 · similarity 0.89)
- AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · similarity 0.89)
- Transformers Provably Implement In-Context Reinforcement Learning with Policy Improvement (NeurIPS 2026 · similarity 0.89)
- DynaMITE-RL: A Dynamic Model for Improved Temporal Meta-Reinforcement Learning (NeurIPS 2024 · similarity 0.88)
- MetaCARD: Meta-Reinforcement Learning with Task Uncertainty Feedback via Decoupled Context-Aware Reward and Dynamics Components (AAAI 2024 · similarity 0.88)
- Decoupling Meta-Reinforcement Learning with Gaussian Task Contexts and Skills (AAAI 2024 · similarity 0.88)
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.
