Research map / Reinforcement learning / Visual and representation RL
ICR-RL: Deep Reinforcement Learning via In-Context-Regression
David Schiff, Ofir Lindenbaum, Yonathan Efroni
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 ICR-RL (TabPFN in-context Q regression) instead of PPO or TRPO | ICR-RL with TabPFN · CartPole-v1, MountainCar-v0, Acrobot-v1 (250 episodes) | reward per episode | Improved |
| replace latest-episodes context truncation with naive de-duplication once context is full (ablation) | ICR-RL · continued learning after context budget is full | reward per episode | Improved |
| use embeddings de-duplication or reward-variance heuristics instead of naive de-duplication (ablation) | ICR-RL context engineering · continued learning after context is full | reward per episode | Hurt |
| use latest-episodes FIFO truncation vs stale context (ablation) | ICR-RL context engineering · continued learning after context is full | reward per episode | Hurt |
| use context engineering to shrink context to 512 samples (ablation) | ICR-RL · context size study | reward per episode: 512 samples on par with context size 8096 | No clear effect |
| apply proper context engineering method | ICR-RL · runtime | normalized runtime: nearly 40% reduction | Improved |
| swap TabPFN backbone for TabICL (ablation) | ICR-RL · sparse-reward CartPole | reward per episode | Improved |
| raise percentile refit threshold to 0.9 or 0.95 (vs 0.4-0.7) (ablation) | ICR-RL · MountainCar-v0 (300 episodes, 5 seeds) | convergence: 0.9 and 0.95 nearly identical top performance | Improved |
| increase context budget B from 256 up to 8192 (ablation) | ICR-RL · MountainCar-v0 (300 episodes, 5 seeds) | convergence | Improved |
Closest papers by meaning
- Towards General-Purpose Model-Free Reinforcement Learning (ICLR 2025 · similarity 0.89)
- XLand-100B: A Large-Scale Multi-Task Dataset for In-Context Reinforcement Learning (ICLR 2025 · similarity 0.89)
- Model-Based Transfer Learning for Contextual Reinforcement Learning (NeurIPS 2024 · similarity 0.89)
- Towards Provable Emergence of In-Context Reinforcement Learning (NeurIPS 2025 · similarity 0.88)
- Vintix: Action Model via In-Context Reinforcement Learning (ICML 2025 · similarity 0.88)
- Dual RL: Unification and New Methods for Reinforcement and Imitation Learning (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.
