Research map / Reinforcement learning / Visual and representation RL
ELMUR: External Layer Memory with Update/Rewrite for Long-Horizon RL
Egor Cherepanov, 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 |
|---|---|---|---|
| use LRU memory update instead of removing it (ablation) | ELMUR · MIKASA-Robo RememberColor3-v0 | normalized score: 0.43 -> 1.00 | Improved |
| use layer-local memory instead of memory shared across layers (ablation) | ELMUR · MIKASA-Robo RememberColor3-v0 | normalized score: 0.45 -> 1.00 | Improved |
| add relative bias to token-memory cross-attention (ablation) | ELMUR · MIKASA-Robo RememberColor3-v0 | normalized score: 0.95 -> 1.00 | Improved |
| add relative bias and LRU together (ablation) | ELMUR · MIKASA-Robo RememberColor3-v0 | normalized score: 0.22 -> 1.00 | Improved |
| replace MoE-FFN with MLP-FFN (ablation) | ELMUR · MIKASA-Robo RememberColor3-v0 | normalized score: 1.00 -> 1.00 | No clear effect |
| memory slots M at least the number of needed segments N (ablation) | ELMUR · MIKASA-Robo RememberColor3-v0 | success rate | Improved |
| intermediate LRU blending lambda about 0.4 to 0.6 (ablation) | ELMUR · MIKASA-Robo RememberColor3-v0 | success rate | Hurt |
| larger memory initialization scale sigma (ablation) | ELMUR · MIKASA-Robo RememberColor3-v0 | success rate | Improved |
| external layer memory with LRU instead of RATE on TakeItBack | ELMUR · MIKASA-Robo TakeItBack-v0 | success rate: 0.42 -> 0.78 | Improved |
| external layer memory with LRU instead of RATE on POPGym puzzles | ELMUR · POPGym puzzle tasks (33) | aggregated return: 0.45 -> 1.2 | Improved |
| external memory on reactive POPGym tasks | ELMUR · POPGym reactive tasks (15) | aggregated return: 9.3 (DT) vs 9.2 | No clear effect |
| condition ELMUR on returns-to-go and actions rather than observations only (ablation) | ELMUR · D4RL MuJoCo | normalized score: average 76.4, about 2% higher | No clear effect |
Closest papers by meaning
- Memo: Training Memory-Efficient Embodied Agents with Reinforcement Learning (NeurIPS 2025 · similarity 0.89)
- SAM2Act: Integrating Visual Foundation Model with A Memory Architecture for Robotic Manipulation (ICML 2025 · similarity 0.88)
- AMA-Bench: Evaluating Long-Horizon Memory for Agentic Applications (ICML 2026 · similarity 0.87)
- An Evolved Universal Transformer Memory (ICLR 2025 · similarity 0.87)
- Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents (NeurIPS 2026 · similarity 0.87)
- A Large Recurrent Action Model: xLSTM enables Fast Inference for Robotics Tasks (ICML 2025 · 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.
