Research map / Reinforcement learning / Visual and representation RL
RePreM: Representation Pre-training with Masked Model for Reinforcement Learning
Yuanying Cai, Chuheng Zhang, Wei Shen, Xuyun Zhang, Wenjie Ruan, Longbo Huang
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 |
|---|---|---|---|
| Pre-train representations with masked trajectory prediction | Rainbow · Atari-100k, Mixed (3M) | Median HNS: 0.548 -> 0.717 | Improved |
| Pre-train representations with masked trajectory prediction | Rainbow · Atari-100k, Mixed (3M) | Mean HNS: 0.858 -> 1.298 | Improved |
| Increase pre-training dataset quality from Random to Weak to Mixed | RePreM · Atari-100k, 10M samples | Median HNS: 0.190 -> 0.774 -> 0.953 | Improved |
| Increase Mixed pre-training data from 3M to 5M to 10M samples | RePreM · Atari-100k | Median HNS: 0.717 -> 0.788 -> 0.953 | Improved |
| Increase the state encoder from Small to Medium to Large | RePreM · Atari-100k, Mixed (10M) | Median HNS: 0.835 -> 0.953 -> 1.085 | Improved |
| Use RePreM representations with SAC | SAC · DMControl-100k, Walker-Walk | Score: 42$\pm$12 -> 633$\pm$52 | Improved |
| Use RePreM representations with SAC | SAC · DMControl-100k, Quadruped-Run | Score: 93$\pm$36 -> 579$\pm$85 | Improved |
| Fine-tune the state encoder during downstream RL (ablation) | RePreM · Atari-100k, Mixed (3M) | Median HNS: 0.717 -> 0.512 | Hurt |
| Use shorter sequences, T=10 instead of T=50 (ablation) | RePreM · Atari-100k, Mixed (3M) | Median HNS: 0.717 -> 0.322 | Hurt |
| Use the BERT masking scheme rather than replacing every masked element with a fixed token (ablation) | RePreM · Atari-100k, Mixed (3M) | Median HNS: 0.588 -> 0.717 | Improved |
| Use GTrXL instead of the original transformer block (ablation) | RePreM · Atari-100k, Mixed (3M) | Median HNS: 0.709 -> 0.717 | Improved |
| Use SimCLR loss instead of ReLIC loss (ablation) | RePreM · Atari-100k, Mixed (3M) | Median HNS: 0.715 -> 0.717 | No clear effect |
Later papers that compared with this method
1 later accepted paper reports RePreM as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Improving and Accelerating Offline RL in Large Discrete Action Spaces with Structured Policy Initialization (ICLR 2026 · better on 1 of 1 shared result)
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Rainbow — Rainbow: Combining Improvements in Deep Reinforcement Learning (AAAI 2018) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- The Surprising Ineffectiveness of Pre-Trained Visual Representations for Model-Based Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
- Efficient Recurrent Off-Policy RL Requires a Context-Encoder-Specific Learning Rate (NeurIPS 2024 · similarity 0.86)
- M^3PC: Test-time Model Predictive Control using Pretrained Masked Trajectory Model (ICLR 2025 · similarity 0.86)
- You Only Need Minimal RLVR Training: Extrapolating LLMs via Rank-1 Trajectories (NeurIPS 2026 · similarity 0.86)
- Reinforcement Learning for Reasoning in Large Language Models with One Training Example (NeurIPS 2025 · similarity 0.86)
- Are Expressive Models Truly Necessary for Offline RL? (AAAI 2025 · similarity 0.86)
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.
