Research map / Reinforcement learning / Policy gradient and actor-critic
Regularized scalable deep value learning
Stabilize deep value-based RL through regularization, overfitting control, and scaling.
216 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.
46 papers in 2023–24 and 89 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 13.4% to 19.2%.
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 · Deep Reinforcement Learning with Double Q-Learning (AAAI 2016 · 6,323 citations)
- 2017 · Hybrid Reward Architecture for Reinforcement Learning (NeurIPS 2017 · 185 citations)
- 2018 · Temporal Difference Models: Model-Free Deep RL for Model-Based Control (ICLR 2018 · 152 citations)
- 2019 · Off-Policy Deep Reinforcement Learning without Exploration (ICML 2019 · 272 citations)
- 2020 · Revisiting Fundamentals of Experience Replay (ICML 2020 · 79 citations)
- 2021 · DeepTrader: A Deep Reinforcement Learning Approach for Risk-Return Balanced Portfolio Management with Market Conditions Embedding (AAAI 2021 · 139 citations)
- 2022 · Sample Efficient Deep Reinforcement Learning via Uncertainty Estimation (ICLR 2022 · 15 citations)
- 2023 · Deep Reinforcement Learning for Cost-Effective Medical Diagnosis (ICLR 2023 · 14 citations)
- 2024 · Diffusion-based Reinforcement Learning via Q-weighted Variational Policy Optimization (NeurIPS 2024 · 17 citations)
- 2025 · Proxy Target: Bridging the Gap Between Discrete Spiking Neural Networks and Continuous Control (NeurIPS 2025 · 2 citations)
- 2026 · Beyond Distributions: Geometric Action Control for Continuous Reinforcement Learning (ICLR 2026 · 1 citation)
Earliest papers here
- Deep Reinforcement Learning with Double Q-Learning (AAAI 2016 · 6,323 citations)
- Dueling Network Architectures for Deep Reinforcement Learning (ICML 2016 · 1,777 citations)
- Deep Exploration via Bootstrapped DQN (NeurIPS 2016 · 451 citations)
What has been tried
149 papers here report 1,526 measured changes (a component added, removed or swapped, compared with a baseline): helped 61%, no clear effect 12%, 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 |
|---|---|---|---|---|
| Add diffusion entropy regularization | Ant-v3 · episodic reward | Helped | Diffusion-based Reinforcement Learning via Q-weighted Variational Policy Optimization (NeurIPS 2024 · 17 citations) | |
| jointly optimize task updating and offloading rather than offloading only | convergence evaluation · average AoI | 31.3% | Helped | Fractional Deep Reinforcement Learning for Age-Minimal Mobile Edge Computing (AAAI 2024 · 10 citations) |
| Replace β=0.5 with pink noise (β=1) | 16 benchmark environments (averaged) · performance (standardized area under learning curve) | Hurt | Colored Noise in PPO: Improved Exploration and Performance through Correlated Action Sampling (AAAI 2024 · 6 citations) | |
| add the third loss term L_C3/L_C3' to enforce tightness constraints | CartPole, B1, MountainCar, and B2 under perturbations · certified-bound tightness | Helped | Robustness Verification of Deep Reinforcement Learning Based Control Systems Using Reward Martingales (AAAI 2024 · 5 citations) | |
| Use the proxy target framework instead of vanilla Actor-Critic | MuJoCo environments (average) · average performance gain (APG) | 32.15% | Helped | Proxy Target: Bridging the Gap Between Discrete Spiking Neural Networks and Continuous Control (NeurIPS 2025 · 2 citations) |
| Increase the number of optimization epochs per rollout | ALE and MuJoCo · policy feature rank | Hurt | No Representation, No Trust: Connecting Representation, Collapse, and Trust Issues in PPO (NeurIPS 2024 · 2 citations) | |
| add policy churn reduction | Antmaze D4RL tasks · normalized score | 77.00 ± 5.52 -> 84.44 ± 3.19; 54.25 ± 5.54 -> 62.50 ± 3.75; 65.75 ± 11.71 -> 72.50 ± 2.92; 73.75 | Mixed | Improving Deep Reinforcement Learning by Reducing the Chain Effect of Value and Policy Churn (NeurIPS 2024 · 2 citations) |
| Add a batch size selected for the compute budget instead of using a constant fixed batch size | 4-task average · data efficiency ratio | 1.03 -> 1.00 | Helped | Compute-Optimal Scaling for Value-Based Deep RL (NeurIPS 2025 · 1 citation) |
| Add compression loss with coefficient ν=0.01 | Ant · convergence speed and final performance | Helped | GenPO: Generative Diffusion Models Meet On-Policy Reinforcement Learning (NeurIPS 2025 · 1 citation) | |
| Remove weight normalization | MyoSuite · return or success rate | Hurt | Scaling Off-Policy Reinforcement Learning with Batch and Weight Normalization (NeurIPS 2025 · 1 citation) | |
| add LayerNorm | Atari-10 · performance | Helped | Simplifying Deep Temporal Difference Learning (ICLR 2025 · 1 citation) | |
| Add entropy-augmented returns to actor training | HandFlip · episode return | Helped | Stabilizing Reinforcement Learning in Differentiable Multiphysics Simulation (ICLR 2025 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 55 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- DeepMind Control Suite · 7 of 55 papers
- HumanoidBench · 5 of 55 papers
- MuJoCo · 3 of 55 papers
- MyoSuite · 3 of 55 papers
- IsaacGym · 2 of 55 papers
- Atari-10 · 2 of 55 papers
- Ant-v4 · 2 of 55 papers
- Hopper-v4 · 2 of 55 papers
Baselines
- SAC · 13 of 55 papers
- PPO · 10 of 55 papers
- BRO · 5 of 55 papers
- DQN · 5 of 55 papers
- QSM · 4 of 55 papers
- QVPO · 4 of 55 papers
- CrossQ · 4 of 55 papers
- DDPG · 4 of 55 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 135 papers; from OpenAlex and the papers' affiliations).
- Google · 6 papers
- Université de Montréal · 6 papers
- University of Alberta · 5 papers
- Hong Kong University of Science and Technology · 5 papers
- Google DeepMind · 5 papers
- Zhejiang University · 4 papers
- University of Oxford · 4 papers
- University of California, Berkeley · 4 papers
- University of Montreal · 4 papers
- Tsinghua University · 4 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
- Model-based RL with guarantees (245 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).