Research map / Reinforcement learning / Visual and representation RL
Learned state and action representations
Learn compact representations of states, actions, or contexts to speed up and generalize RL.
148 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Visual and representation RL, keyword cluster “representation · latent · generalization”
State History Tried Checklist Preprints Who Export
Where it stands
Not enough benchmark data Fewer than four papers report the same benchmark, so there is no record to follow.
39 papers in 2023–24 and 40 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 6.8% to 5.7%.
How this is decided
A benchmark is tracked when at least four papers of this approach report their own method on the same dataset and metric in at least two different years. Its record is the best number reported up to 2024 compared with the best reported in 2025–26: still improving means at least 1% better, flat means at most 0.2% better with at least two papers from 2025–26 trying. Numbers on a different scale from the rest (a fraction among percentages) and far outliers are left out. Results come from different papers with different setups, so they are signals, not head-to-head comparisons. The share of papers is counted at ICML, NeurIPS, ICLR, CVPR and AAAI.
How it got here
Most cited paper of each year
- 2016 · Control of Memory, Active Perception, and Action in Minecraft (ICML 2016 · 164 citations)
- 2018 · Recurrent World Models Facilitate Policy Evolution (NeurIPS 2018 · 452 citations)
- 2019 · Combined Reinforcement Learning via Abstract Representations (AAAI 2019 · 73 citations)
- 2020 · CURL: Contrastive Unsupervised Representations for Reinforcement Learning (ICML 2020 · 215 citations)
- 2021 · Data-Efficient Reinforcement Learning with Self-Predictive Representations (ICLR 2021 · 47 citations)
- 2022 · Pre-Trained Language Models for Interactive Decision-Making (NeurIPS 2022 · 53 citations)
- 2023 · Frustratingly Easy Regularization on Representation Can Boost Deep Reinforcement Learning (CVPR 2023 · 9 citations)
- 2024 · Jointly Training and Pruning CNNs via Learnable Agent Guidance and Alignment (CVPR 2024 · 20 citations)
- 2025 · Empirical Design in Reinforcement Learning (ICML 2025 · 7 citations)
- 2026 · From Representation to Action: A Unified Laplacian Framework for Spatial Representation and Path Planning (ICML 2026 · 0 citations)
Earliest papers here
- Control of Memory, Active Perception, and Action in Minecraft (ICML 2016 · 164 citations)
- Recurrent World Models Facilitate Policy Evolution (NeurIPS 2018 · 452 citations)
- Learning Structured Representation for Text Classification via Reinforcement Learning (AAAI 2018 · 154 citations)
What has been tried
113 papers here report 1,096 measured changes (a component added, removed or swapped, compared with a baseline): helped 61%, no clear effect 11%, hurt 23%, mixed 5%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| remove epoch embeddings and recurrent environment representation | CIFAR-10, ResNet-56 · ∆-Acc | +0.45% -> −0.03% | Hurt | Jointly Training and Pruning CNNs via Learnable Agent Guidance and Alignment (CVPR 2024 · 20 citations) |
| increase training dataset size | PushT · success rate | 0.08 -> 0.92 | Helped | DINO-WM: World Models on Pre-trained Visual Features enable Zero-shot Planning (ICML 2025 · 5 citations) |
| Use implicit rather than explicit dynamics models | Pick, Stack, Series, and DMC tasks · sample efficiency | Helped | Building Minimal and Reusable Causal State Abstractions for Reinforcement Learning (AAAI 2024 · 2 citations) | |
| Add the Last Transition Label to CTL+C-PREP | GN, MP, and PD; six EL/CM context settings · ttt_auc | Helped | Contextual Pre-planning on Reward Machine Abstractions for Enhanced Transfer in Deep Reinforcement Learning (AAAI 2024 · 2 citations) | |
| use only the current-step reward instead of accumulated Q-value rewards (γ=0) | FashionMNIST and CIFAR10 · accuracy | 75.63±0.11 vs 78.23±0.13; 41.91±0.15 vs 46.46±0.15 | Hurt | Improving Unsupervised Hierarchical Representation with Reinforcement Learning (CVPR 2024 · 2 citations) |
| remove chronological embedding ψ | DM_Control distraction, cheetah-run and walker-walk · score | Hurt | State Chrono Representation for Enhancing Generalization in Reinforcement Learning (NeurIPS 2024 · 2 citations) | |
| Replace basis-function reward reconstruction with an MLP conditioned on the representation | Ms. Pacman · success rate | Hurt | Zero-Shot Reinforcement Learning via Function Encoders (ICML 2024 · 1 citation) | |
| Add LESR state representation | Mujoco (mean, 1M steps) · RPI | 21% -> 29% | Helped | LLM-Empowered State Representation for Reinforcement Learning (ICML 2024 · 1 citation) |
| remove decoder gradients | ImageNet-1k validation, pools 1, 5, and 10 · million steps to 80% success | 0.42 0.06 -> 1.13 0.12; 1.23 0.20 -> 1.79 0.61; 1.44 0.58 -> 2.57 0.91 | Hurt | Sliding Puzzles Gym: A Scalable Benchmark for State Representation in Visual Reinforcement Learning (ICML 2025 · 0 citations) |
| Down-scale Gaussian component variances compared with direct GMM sampling | Kitchen · Task success | 7% | Helped | Quantization-Free Autoregressive Action Transformer (NeurIPS 2025 · 0 citations) |
| Add static/dynamic state factorization | Libero, compositional generalization · success rate | 0.64 -> 0.70 | Helped | Learning Interactive World Model for Object-Centric Reinforcement Learning (NeurIPS 2025 · 0 citations) |
| Replace relative trajectory actions with absolute trajectory actions | MimicGen, four-task ablation · mean success rate | 67.0 -> 55.9; reduces performance by 11.1% | Hurt | A Practical Guide for Incorporating Symmetry in Diffusion Policy (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 21 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Franka Kitchen · 2 of 21 papers
- Meta-World · 2 of 21 papers
- Atari · 2 of 21 papers
Baselines
- DQN · 4 of 21 papers
- DreamerV3 · 4 of 21 papers
- TD-MPC2 · 3 of 21 papers
- PPO · 3 of 21 papers
- SPR · 3 of 21 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 79 papers; from OpenAlex and the papers' affiliations).
- Google DeepMind · 12 papers
- Google · 9 papers
- Tsinghua University · 5 papers
- Carnegie Mellon University · 4 papers
- University of California, Berkeley · 4 papers
- University of Texas at Austin · 3 papers
- McGill University · 3 papers
- Tianjin University · 3 papers
- Meta · 3 papers
- University of Science and Technology of China · 2 papers
Use it
Download BibTeX of 40 key papersOpen the topic in the interactive map
Get alerts. Describe your idea in scime atlas and save it (free scime account): you get a weekly email when new preprints land close to it.
Related approaches
Same keyword cluster
- Robust visual encoders for RL (86 papers)
- Latent-space world models (83 papers)
- Successor features for transfer (12 papers)
Elsewhere in the topic
- In-context and adaptive RL (202 papers)
- Scalable RL systems and environments (179 papers)
- Probabilistic and autonomous RL (177 papers)
- RL for combinatorial routing (158 papers)
- Dexterous manipulation transfer (141 papers)
- RL with tree search for discovery (121 papers)
- Inferring decision-making from behavior (93 papers)
- Skill learning and diffusion policies (86 papers)
Benchmark numbers, ablations and baseline numbers are extracted by AI from each paper's full text and can be wrong; the paper is the authority. Data: results index of 2026-11-02, citation counts from OpenAlex (2026-10-09).