Research map / Reinforcement learning / Visual and representation RL
Latent-space world models
Learn latent dynamics models and temporal-consistency objectives for model-based and transfer RL.
83 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.
32 papers in 2023–24 and 19 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 5.6% to 2.7%.
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 · Benchmarking Deep Reinforcement Learning for Continuous Control (ICML 2016 · 949 citations)
- 2017 · Value Prediction Network (NeurIPS 2017 · 40 citations)
- 2018 · Feature Engineering for Predictive Modeling Using Reinforcement Learning (AAAI 2018 · 203 citations)
- 2019 · SOLAR: Deep Structured Representations for Model-Based Reinforcement Learning (ICML 2019 · 132 citations)
- 2020 · Predictive Information Accelerates Learning in RL (NeurIPS 2020 · 26 citations)
- 2021 · Reinforcement Learning with Prototypical Representations (ICML 2021 · 39 citations)
- 2022 · AdaRL: What, Where, and How to Adapt in Transfer Reinforcement Learning (ICLR 2022 · 6 citations)
- 2023 · ERL-Re$^2$: Efficient Evolutionary Reinforcement Learning with Shared State Representation and Individual Policy Representation (ICLR 2023 · 16 citations)
- 2024 · Predictive auxiliary objectives in deep RL mimic learning in the brain (ICLR 2024 · 5 citations)
- 2025 · Stealing That Free Lunch: Exposing the Limits of Dyna-Style Reinforcement Learning (ICML 2025 · 1 citation)
- 2026 · Ego-Foresight: Self-supervised Learning of Agent-Aware Representations for Improved RL (ICLR 2026 · 0 citations)
Earliest papers here
- Benchmarking Deep Reinforcement Learning for Continuous Control (ICML 2016 · 949 citations)
- Value Prediction Network (NeurIPS 2017 · 40 citations)
- Feature Engineering for Predictive Modeling Using Reinforcement Learning (AAAI 2018 · 203 citations)
What has been tried
70 papers here report 730 measured changes (a component added, removed or swapped, compared with a baseline): helped 60%, no clear effect 11%, hurt 22%, mixed 7%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use longer-horizon prediction (larger γ) rather than one-step prediction | Goal transfer in gridworld · episode score | Helped | Predictive auxiliary objectives in deep RL mimic learning in the brain (ICLR 2024 · 5 citations) | |
| Enforce expert diversity with Gram-Schmidt orthogonalization instead of using a mixture of experts without orthogonalization | MetaWorld MT10-rand and MT50-rand · success rate and sample efficiency | Helped | Multi-Task Reinforcement Learning with Mixture of Orthogonal Experts (ICLR 2024 · 3 citations) | |
| Increase the synthetic-to-real data ratio | DMC hopper-stand and humanoid-stand · episodic return | Hurt | Stealing That Free Lunch: Exposing the Limits of Dyna-Style Reinforcement Learning (ICML 2025 · 1 citation) | |
| integrate dynamic modulation into the RSSM sequence model | Boxing, Krull, Pong, and Road Runner · performance | Helped | DyMoDreamer: World Modeling with Dynamic Modulation (NeurIPS 2025 · 1 citation) | |
| Add counterfactual data augmentation | Ant · average return | 6231±81 -> 6418±81 | Helped | Causal Information Prioritization for Efficient Reinforcement Learning (ICLR 2025 · 1 citation) |
| Decrease the augmented replay ratio by generating more augmented transitions per observed transition | Panda tasks · data efficiency | Helped | Understanding when Dynamics-Invariant Data Augmentations Benefit Model-free Reinforcement Learning Updates (ICLR 2024 · 1 citation) | |
| Remove the representation-mismatch reward penalty (set β=0) | halfcheetah (broken back thigh) and walker (no right thigh) · final return | Hurt | Cross-Domain Policy Adaptation by Capturing Representation Mismatch (ICML 2024 · 1 citation) | |
| Increase the reward loss coefficient | Meta-world tasks · sample efficiency | Helped | HarmonyDream: Task Harmonization Inside World Models (ICML 2024 · 1 citation) | |
| Use limb-based state representation without augmentation (rho_aug=0%) | 6 tasks shown in the first 2 rows · data efficiency | Helped | Reinforcement Learning with Euclidean Data Augmentation for State-Based Continuous Control (NeurIPS 2024 · 1 citation) | |
| Add segmentation-based gradient aggregation | Cheetah Run, unaltered and Reafferent Control · score | 742.1 ± 79.7 -> 712.3 ± 32.3; 188.4 ± 9.4 -> 383.1 ± 23.8 | Mixed | Policy-shaped prediction: avoiding distractions in model-based reinforcement learning (NeurIPS 2024 · 1 citation) |
| Remove the optimism term from the self-predictive objective | Combination lock MDP · self-predictive loss | No clear effect | Reinforcement Learning Under Latent Dynamics: Toward Statistical and Algorithmic Modularity (NeurIPS 2024 · 1 citation) | |
| remove action prediction model KL regularization (MTC-NoA) | Cheetah Run; action noise and mass changes · robustness | Hurt | Maximum Total Correlation Reinforcement Learning (ICML 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 12 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Meta-World · 3 of 12 papers
- DeepMind Control Suite · 2 of 12 papers
- Atari 100k · 2 of 12 papers
Baselines
- SAC · 5 of 12 papers
- DreamerV3 · 5 of 12 papers
- TD-MPC2 · 4 of 12 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 51 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 7 papers
- Tianjin University · 5 papers
- Google · 5 papers
- Tianjin Unibersity · 3 papers
- University of California, San Diego · 3 papers
- Meta · 3 papers
- Huawei · 3 papers
- NetEase · 2 papers
- Stanford University · 2 papers
- Carnegie Mellon University · 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
- Learned state and action representations (148 papers)
- Robust visual encoders for RL (86 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).