Research map / Reinforcement learning / Visual and representation RL
Scalable RL systems and environments
Scale RL training and benchmark environments.
179 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Visual and representation RL, keyword cluster “search · decision · systems”
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.
49 papers in 2023–24 and 47 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 8.6% to 6.8%.
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 · Graying the black box: Understanding DQNs (ICML 2016 · 56 citations)
- 2017 · Imagination-Augmented Agents for Deep Reinforcement Learning (NeurIPS 2017 · 209 citations)
- 2018 · Rainbow: Combining Improvements in Deep Reinforcement Learning (AAAI 2018 · 1,797 citations)
- 2019 · Multi-Task Deep Reinforcement Learning with PopArt (AAAI 2019 · 200 citations)
- 2020 · Agent57: Outperforming the Atari Human Benchmark (ICML 2020 · 226 citations)
- 2021 · DeepSynth: Automata Synthesis for Automatic Task Segmentation in Deep Reinforcement Learning (AAAI 2021 · 36 citations)
- 2022 · Multi-Sacle Dynamic Coding Improved Spiking Actor Network for Reinforcement Learning (AAAI 2022 · 34 citations)
- 2023 · CleanRL: High-quality Single-file Implementations of Deep Reinforcement Learning Algorithms (ICML 2023 · 86 citations)
- 2024 · DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning (NeurIPS 2024 · 15 citations)
- 2025 · APIRL: Deep Reinforcement Learning for REST API Fuzzing (AAAI 2025 · 4 citations)
- 2026 · DRMD: Deep Reinforcement Learning for Malware Detection Under Concept Drift (AAAI 2026 · 2 citations)
Earliest papers here
- Graying the black box: Understanding DQNs (ICML 2016 · 56 citations)
- Imagination-Augmented Agents for Deep Reinforcement Learning (NeurIPS 2017 · 209 citations)
- Distral: Robust multitask reinforcement learning (NeurIPS 2017 · 175 citations)
What has been tried
144 papers here report 1,357 measured changes (a component added, removed or swapped, compared with a baseline): helped 63%, no clear effect 12%, hurt 20%, 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 |
|---|---|---|---|---|
| Use cross-entropy rather than regression for value-function training | AitW device-control tasks · performance | around 12% | Helped | DigiRL: Training In-The-Wild Device-Control Agents with Autonomous Reinforcement Learning (NeurIPS 2024 · 15 citations) |
| use dynamic DRL-based actor scheduling instead of 32 actors every round | Hopper-v3 · training cost | 1.7 ± 0.3 vs 1.2 ± 0.2 | Helped | Cheaper and Faster: Distributed Deep Reinforcement Learning with Serverless Computing (AAAI 2024 · 10 citations) |
| Add attention during GNN aggregation | Centipede-{4,6} → Centipede-16 · episode rewards | Helped | A Transfer Approach Using Graph Neural Networks in Deep Reinforcement Learning (AAAI 2024 · 8 citations) | |
| Add the Bias Factor Loss (BFL) to BiCoGAN | CP-Noisy: HD, source and target domains · model bias | 84% (on the source domain), 43% (on the target domain 1), and 56% (on the target domain 2) | Helped | ACAMDA: Improving Data Efficiency in Reinforcement Learning through Guided Counterfactual Data Augmentation (AAAI 2024 · 7 citations) |
| add advantage per-minibatch zero-mean normalization | Five Brax MuJoCo environments · per-environment tuned normalized score | Helped | A Method for Evaluating Hyperparameter Sensitivity in Reinforcement Learning (NeurIPS 2024 · 7 citations) | |
| select the best checkpoint using generalization-environment performance | Same Game 5x5c3s2 generalization · success rate | 9.4\% -> 18.1\% | Helped | PUZZLES: A Benchmark for Neural Algorithmic Reasoning (NeurIPS 2024 · 6 citations) |
| Add object-detection noise (5% misdetection probability and Gaussian noise with 3-pixel standard deviation) | Atari games, especially Kangaroo · human-normalized score | Mixed | Interpretable Concept Bottlenecks to Align Reinforcement Learning Agents (NeurIPS 2024 · 5 citations) | |
| remove transformer embeddings | All APIs · Coverage | 95.6% -> 87.2% | Hurt | APIRL: Deep Reinforcement Learning for REST API Fuzzing (AAAI 2025 · 4 citations) |
| Use Robust PLR rather than vanilla PLR | Craftax-1B · evaluation reward | Hurt | Craftax: A Lightning-Fast Benchmark for Open-Ended Reinforcement Learning (ICML 2024 · 4 citations) | |
| Increase the classical DNN width from 16 to 64 | BeamManagement6G · sample complexity | Helped | Benchmarking Quantum Reinforcement Learning (ICML 2025 · 3 citations) | |
| Use CoT prompting instead of non-CoT prompting | miniF2F-test · pass rate | 57.4\% -> 60.2\% | Helped | DeepSeek-Prover-V1.5: Harnessing Proof Assistant Feedback for Reinforcement Learning and Monte-Carlo Tree Search (ICLR 2025 · 3 citations) |
| Train with four generators instead of only the uniform generator | VLSI TSP test set (1,020 instances) · mean tour length | 6.090 ;(± 0.029) -> 5.815 ;(± 0.025) | Helped | Jumanji: a Diverse Suite of Scalable Reinforcement Learning Environments in JAX (ICLR 2024 · 3 citations) |
What reviewers will expect
Datasets and baselines named in the 27 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- MuJoCo · 3 of 27 papers
- Meta-World · 2 of 27 papers
- MineDojo tasks · 2 of 27 papers
Baselines
- DreamerV3 · 3 of 27 papers
- SAC · 3 of 27 papers
- PPO · 2 of 27 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 96 papers; from OpenAlex and the papers' affiliations).
- Google · 9 papers
- Google DeepMind · 7 papers
- University of Montreal · 6 papers
- Tsinghua University · 4 papers
- Imperial College London · 3 papers
- University College London · 3 papers
- Shanghai Jiao Tong University · 3 papers
- Nanyang Technological University · 3 papers
- The Chinese University of Hong Kong · 3 papers
- Université de Montréal · 3 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
- In-context and adaptive RL (202 papers)
- Probabilistic and autonomous RL (177 papers)
- RL for combinatorial routing (158 papers)
- RL with tree search for discovery (121 papers)
- Inferring decision-making from behavior (93 papers)
- Curriculum and environment design (81 papers)
Elsewhere in the topic
- Learned state and action representations (148 papers)
- Dexterous manipulation transfer (141 papers)
- Skill learning and diffusion policies (86 papers)
- Robust visual encoders for RL (86 papers)
- Novelty and impact-driven exploration (85 papers)
- Hierarchical skills and priors (84 papers)
- Latent-space world models (83 papers)
- Goal-conditioned policy learning (66 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).