Research map / Reinforcement learning / Visual and representation RL
Explore to Generalize in Zero-Shot RL
Ev Zisselman, Itai Lavie, Daniel Soudry, Aviv Tamar
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 a trained max-entropy exploration policy when the reward-policy ensemble disagrees, rather than random actions (ablation) | ExpGen · Maze | test score: 6.2 ± 0.2 -> 8.2 ± 0.1 | Improved |
| Increase ensemble size from 4 to 10 (ablation) | ExpGen · Maze | test score: 8.02±0.06 -> 8.22±0.11 | Improved |
| Use L0 rather than L2 state similarity (ablation) | ExpGen · Maze | test score: 8.3 ± 0.2 -> 8.2 ± 0.1 | No clear effect |
| Use L0 rather than L2 state similarity (ablation) | ExpGen · Heist | test score: 7.4 ± 0.2 -> 7.4 ± 0.1 | No clear effect |
| Combine intrinsic exploration reward with extrinsic reward in one weighted reward instead of using ExpGen's ensemble switching (ablation) | PPO on Maze · Maze | test score | Hurt |
| Add a GRU memory unit to the max-entropy policy (ablation) | Max-entropy policy · Maze, Jumper, and Miner | max-entropy performance | Improved |
| Add a GRU to an extrinsic-reward policy (ablation) | PPO · Maze and Jumper with 200 training levels | extrinsic-reward performance | Hurt |
| Train for maximum-entropy exploration rather than extrinsic reward | Policy trained on ProcGen · Maze, Jumper, and Miner with 200 training levels | generalization gap: 1.7% vs 38.8%; 8.5% vs 27.5%; 4.3% vs 13.1% | Improved |
| Combine ExpGen with IDAAC rather than using IDAAC alone | IDAAC · ProcGen test suite | aggregate performance: 0.6 | Improved |
Closest papers by meaning
- Horizon Generalization in Reinforcement Learning (ICLR 2025 · similarity 0.89)
- How Ensembles of Distilled Policies Improve Generalisation in Reinforcement Learning (NeurIPS 2025 · similarity 0.87)
- Zero Shot Generalization of Vision-Based RL Without Data Augmentation (ICML 2025 · similarity 0.87)
- VIGOR: Zero-Shot Visual Generalization via Latent-Space Consistency in Model-Based Reinforcement Learning (NeurIPS 2026 · similarity 0.87)
- Model-Based Transfer Learning for Contextual Reinforcement Learning (NeurIPS 2024 · similarity 0.87)
- Provable Zero-Shot Generalization in Offline Reinforcement Learning (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.
