Research map / Reinforcement learning / Policy gradient and actor-critic
Model-based RL with guarantees
Analyze and design model-based RL algorithms with provable robustness and sample efficiency.
245 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Policy gradient and actor-critic, keyword cluster “value · td · distributional”
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.
55 papers in 2023–24 and 85 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 16.0% to 18.4%.
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 · Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization (ICML 2016 · 591 citations)
- 2017 · Learning to Optimize (ICLR 2017 · 153 citations)
- 2018 · Model-Ensemble Trust-Region Policy Optimization (ICLR 2018 · 261 citations)
- 2019 · Algorithmic Framework for Model-based Deep Reinforcement Learning with Theoretical Guarantees (ICLR 2019 · 127 citations)
- 2020 · Adaptive Trust Region Policy Optimization: Global Convergence and Faster Rates for Regularized MDPs (AAAI 2020 · 94 citations)
- 2021 · Robust Reinforcement Learning: A Case Study in Linear Quadratic Regulation (AAAI 2021 · 39 citations)
- 2022 · Maximum Entropy RL (Provably) Solves Some Robust RL Problems (ICLR 2022 · 28 citations)
- 2023 · Policy Evaluation and Temporal-Difference Learning in Continuous Time and Space: A Martingale Approach (ICML 2023 · 38 citations)
- 2024 · TD-MPC2: Scalable, Robust World Models for Continuous Control (ICLR 2024 · 8 citations)
- 2025 · Flow-Based Policy for Online Reinforcement Learning (NeurIPS 2025 · 1 citation)
- 2026 · CHDP: Cooperative Hybrid Diffusion Policies for Reinforcement Learning in Parameterized Action Space (AAAI 2026 · 2 citations)
Earliest papers here
- Guided Cost Learning: Deep Inverse Optimal Control via Policy Optimization (ICML 2016 · 591 citations)
- Increasing the Action Gap: New Operators for Reinforcement Learning (AAAI 2016 · 123 citations)
- Safe and Efficient Off-Policy Reinforcement Learning (NeurIPS 2016 · 91 citations)
What has been tried
167 papers here report 1,412 measured changes (a component added, removed or swapped, compared with a baseline): helped 60%, no clear effect 13%, hurt 20%, 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 |
|---|---|---|---|---|
| Normalize task embeddings to maximum L2 norm 1 | 80-task dataset · normalized score | Helped | TD-MPC2: Scalable, Robust World Models for Continuous Control (ICLR 2024 · 8 citations) | |
| Add the non-crossing quantile regressor | Toy environment, all evaluated risk measures · MSE actions | Helped | Risk-Conditioned Reinforcement Learning: A Generalized Approach for Adapting to Varying Risk Measures (AAAI 2024 · 5 citations) | |
| add Pareto adaptive fine-tuning | Walker2d · SP | 0.180 ± 0.096 -> 0.014 ± 0.006 | Helped | PA2D-MORL: Pareto Ascent Directional Decomposition Based Multi-Objective Reinforcement Learning (AAAI 2024 · 4 citations) |
| add Wasserstein-2 policy constraint toward the advantage-weighted behavior-optimal policy | DMControl and HumanoidBench · return | Helped | Flow-Based Policy for Online Reinforcement Learning (NeurIPS 2025 · 1 citation) | |
| use an environment-state-only value estimator instead of estimating value from the state and denoised action | Robomimic Can and Square · fine-tuning performance | Helped | Diffusion Policy Policy Optimization (ICLR 2025 · 1 citation) | |
| Add cross-trial memory to MC-DML | Zork1, Deephome, Detective, and Ztuu · game score | 38.33 ± 2.89 -> 48.66 ± 1.89; 62.66 ± 0.94 -> 67 ± 1.41; 326.67 ± 4.71 -> 346.67 ± 9.43; 20.66 ± 0.47 -> 23.67 ± 1.9 | Helped | Monte Carlo Planning with Large Language Model for Text-Based Game Agents (ICLR 2025 · 1 citation) |
| Increase aleatoric network capacity | Search and rescue, zero-shot setting · return | Mixed | Bayesian Exploration Networks (ICML 2024 · 1 citation) | |
| Add learned weight updates instead of keeping uniformly initialized weights | Traffic light control · Minimum return across four dimensions | Helped | The Max-Min Formulation of Multi-Objective Reinforcement Learning: From Theory to a Model-Free Algorithm (ICML 2024 · 1 citation) | |
| auto-tune the entropy coefficient | critical control environments · reward | Mixed | When Maximum Entropy Misleads Policy Optimization (ICML 2025 · 0 citations) | |
| Increase test uncertainty radius from R_test=0.01 to R_test=0.1 | Recycling Robot · robust value | 134.45 to 55.17 | Hurt | Pessimism Principle Can Be Effective: Towards a Framework for Zero-Shot Transfer Reinforcement Learning (ICML 2025 · 0 citations) |
| Replace ERAM's uniform-reference adversary regularization with ARAM's adaptive correlation-based regularizer | Base-4 traffic signal control · max-min return | -1387 -> -1160 | Helped | Multi-Objective Reinforcement Learning with Max-Min Criterion: A Game-Theoretic Approach (NeurIPS 2025 · 0 citations) |
| use state entropy regularization instead of policy entropy regularization | robustness across rollout budgets · robust performance | Mixed | State Entropy Regularization for Robust Reinforcement Learning (NeurIPS 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 46 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- HumanoidBench · 3 of 46 papers
- Acrobot · 2 of 46 papers
- DMControl · 2 of 46 papers
- Robomimic · 2 of 46 papers
- OGBench · 2 of 46 papers
Baselines
- SAC · 9 of 46 papers
- PPO · 7 of 46 papers
- TD-MPC2 · 4 of 46 papers
- DreamerV3 · 3 of 46 papers
- DDPG · 2 of 46 papers
- TD3 · 2 of 46 papers
- QVPO · 2 of 46 papers
- Vtrace · 2 of 46 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 140 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 11 papers
- Carnegie Mellon University · 6 papers
- Microsoft · 5 papers
- Massachusetts Institute of Technology · 4 papers
- University of Maryland, College Park · 4 papers
- NVIDIA · 3 papers
- University of Maryland · 3 papers
- University of Illinois, Urbana Champaign · 3 papers
- Korea Advanced Institute of Science and Technology · 3 papers
- Technion · 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
- Regularized scalable deep value learning (216 papers)
- Temporal-difference and discounting theory (143 papers)
- Tree search and policy optimization (109 papers)
- Distributional value learning (60 papers)
Elsewhere in the topic
- Safety layers and shielding (85 papers)
- Variance-reduced policy gradients (79 papers)
- Actor-critic theory (76 papers)
- Off-policy actor-critic with expressive policies (71 papers)
- Trajectory reuse in policy gradients (68 papers)
- Importance-weighted OPE estimators (57 papers)
- Alternative policy update rules (50 papers)
- Constrained policy optimization (48 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).