Research map / Reinforcement learning / Offline RL
Recurrent Action Transformer with Memory
Egor Cherepanov, Aleksei Staroverov, Alexey Kovalev, Aleksandr Panov
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 Memory Retention Valve (cross-attention) to RATE (ablation) | RATE · T-Maze L=900 | success rate: 0.61 -> 0.90 | Improved |
| MRV-CA-1 (updated memory queries incoming) instead of MRV-CA-2 (ablation) | RATE · T-Maze | success rate: 0.90 -> 0.49 at 900 | Hurt |
| gating (MRV-G), GRU, or LSTM memory update instead of cross-attention MRV (ablation) | RATE · T-Maze L=900 | success rate: 0.90 -> 0.65 (G), 0.55 (GRU), 0.47 (LSTM) | Hurt |
| corrupt memory embeddings with noise at inference (ablation) | RATE · T-Maze | success rate: drops to 50% | Hurt |
| noise cached hidden states in ViZDoom (ablation) | RATE · ViZDoom-Two-Colors | return | Hurt |
| add recurrent memory (RATE) over Decision Transformer | DT · T-Maze T=480 | success rate: 0.50 -> 0.90 | Improved |
| add recurrent memory (RATE) over Decision Transformer | DT · ViZDoom-Two-Colors | total reward: 24.8 -> 41.5 | Improved |
| add memory on POPGym memory tasks | DT · POPGym Memory (33) | aggregated average return: -3.5 -> 0.5 | Improved |
| add memory on POPGym reactive tasks | DT · POPGym Reactive (15) | aggregated average return: 9.3 vs 9.1 | No clear effect |
| oracle memory (OracleDT) vs learned memory (ablation) | RATE · ViZDoom-Two-Colors | total reward: 41.5 -> 56.5 | Improved |
| RATE on MuJoCo versus DT | DT · D4RL MuJoCo average | normalized score: 74.7 -> 78.5 | Improved |
| RATE on Atari SeaQuest versus DT | DT · Atari SeaQuest | raw score: 1129.3 -> 1037.9 | Hurt |
Closest papers by meaning
- Memo: Training Memory-Efficient Embodied Agents with Reinforcement Learning (NeurIPS 2025 · similarity 0.90)
- Recurrent Reinforcement Learning with Memoroids (NeurIPS 2024 · similarity 0.88)
- Decision SpikeFormer: Spike-Driven Transformer for Decision Making (CVPR 2025 · similarity 0.88)
- Rethinking Decision Transformer via Hierarchical Reinforcement Learning (ICML 2024 · similarity 0.88)
- Efficient Recurrent Off-Policy RL Requires a Context-Encoder-Specific Learning Rate (NeurIPS 2024 · similarity 0.88)
- Decision ConvFormer: Local Filtering in MetaFormer is Sufficient for Decision Making (ICLR 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.
