Research map / Reinforcement learning / Visual and representation RL
Cross-Episodic Curriculum for Transformer Agents
Lucy Xiaoyang Shi, Yunfan Jiang, Jake Grigsby, Linxi Fan, Yuke Zhu
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 cross-episodic attention (ablation) | CEC Transformer agent · DMLab Goal Maze | success rate: 35.0 -> 65.2 | Improved |
| add cross-episodic attention (ablation) | CEC Transformer agent · DMLab Watermaze | success rate: 20.0 -> 50.9 | Improved |
| add cross-episodic attention (ablation) | CEC Transformer agent · DMLab Irreversible Path | success rate: 3.8 -> 38.2 | Improved |
| add cross-episodic attention (ablation) | CEC Transformer agent · RoboMimic Lift | success rate: 75.9 -> 100.0 | Improved |
| add cross-episodic attention (ablation) | CEC Transformer agent · RoboMimic Can | success rate: 99.3 -> 100.0 | No clear effect |
| use fine-grained instead of coarser task-difficulty curriculum (ablation) | CEC (Task Difficulty) · DMLab levels | relative performance | Improved |
| use a moderate context length instead of too short or too long (ablation) | CEC (Task Difficulty), Auto · DMLab levels | relative performance | Mixed |
| cross-episodic curriculum with task-difficulty vs single-difficulty AT | CEC (Task Difficulty), Auto · DMLab (avg of 3 levels) | average score: 42.7 -> 51.4 | Improved |
| cross-episodic curriculum vs DT with mixed difficulty data | CEC (Task Difficulty), Auto · DMLab (avg of 3 levels) | average score: 35.3 -> 51.4 | Improved |
| fixed task sequencing instead of automatic promotion | CEC (Task Difficulty) · DMLab (avg of 3 levels) | average score: 51.4 -> 54.4 | Improved |
| learning-progress curriculum instead of task-difficulty curriculum | CEC · DMLab (avg of 3 levels) | average score: 51.4 -> 32.4 | Hurt |
| CEC vs BC on expert data only | CEC (Task Difficulty), Auto · DMLab (avg of 3 levels) | average score: 14.2 -> 51.4 | Improved |
Closest papers by meaning
- Transformers Provably Implement In-Context Reinforcement Learning with Policy Improvement (NeurIPS 2026 · similarity 0.84)
- PEAC: Unsupervised Pre-training for Cross-Embodiment Reinforcement Learning (NeurIPS 2024 · similarity 0.84)
- Transformers Can Learn Temporal Difference Methods for In-Context Reinforcement Learning (ICLR 2025 · similarity 0.84)
- Diffusion-based Curriculum Reinforcement Learning (NeurIPS 2024 · similarity 0.84)
- RLIF: Interactive Imitation Learning as Reinforcement Learning (ICLR 2024 · similarity 0.84)
- TOP-ERL: Transformer-based Off-Policy Episodic Reinforcement Learning (ICLR 2025 · similarity 0.84)
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.
