Research map / Reinforcement learning / Policy gradient and actor-critic
Variance-reduced policy gradients
Reduce gradient variance with off-policy estimators, factorized baselines, and risk-sensitive formulations.
79 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Policy gradient and actor-critic, keyword cluster “policy gradient · gradients · variance”
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.
12 papers in 2023–24 and 21 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 3.5% to 4.5%.
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
- 2017 · Cold-Start Reinforcement Learning with Softmax Policy Gradient (NeurIPS 2017 · 21 citations)
- 2018 · Distributed Distributional Deterministic Policy Gradients (ICLR 2018 · 320 citations)
- 2019 · Neural Trust Region/Proximal Policy Optimization Attains Globally Optimal Policy (NeurIPS 2019 · 52 citations)
- 2020 · Variational Policy Gradient Method for Reinforcement Learning with General Utilities (NeurIPS 2020 · 44 citations)
- 2021 · Task-Agnostic Exploration via Policy Gradient of a Non-Parametric State Entropy Estimate (AAAI 2021 · 23 citations)
- 2022 · Policy Gradient Method For Robust Reinforcement Learning (ICML 2022 · 11 citations)
- 2023 · An Alternative to Variance: Gini Deviation for Risk-averse Policy Gradient (NeurIPS 2023 · 3 citations)
- 2024 · Optimistic Policy Gradient in Multi-Player Markov Games with a Single Controller: Convergence beyond the Minty Property (AAAI 2024 · 4 citations)
- 2025 · Improving Value Estimation Critically Enhances Vanilla Policy Gradient (ICML 2025 · 0 citations)
- 2026 · Boosting CVaR Policy Optimization with Quantile Gradients (ICML 2026 · 0 citations)
Earliest papers here
- Cold-Start Reinforcement Learning with Softmax Policy Gradient (NeurIPS 2017 · 21 citations)
- Improving Policy Gradient by Exploring Under-appreciated Rewards (ICLR 2017 · 5 citations)
- Distributed Distributional Deterministic Policy Gradients (ICLR 2018 · 320 citations)
What has been tried
57 papers here report 442 measured changes (a component added, removed or swapped, compared with a baseline): helped 63%, no clear effect 11%, hurt 19%, 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 |
|---|---|---|---|---|
| Increase VPG value updates per iteration (up to 50 or more) | HalfCheetah, Hopper, Walker, Ant, and Humanoid · cumulative reward | 50 | Helped | Improving Value Estimation Critically Enhances Vanilla Policy Gradient (ICML 2025 · 0 citations) |
| Use E-SoftTreeMax instead of C-SoftTreeMax | Atari games · reward | No clear effect | Policy Gradient with Tree Expansion (ICML 2025 · 0 citations) | |
| Set k_opt to match k_eval rather than use untransformed rewards (k_opt=1) | MATH · pass@4 | 30.73 ± 0.59 -> 34.93 ± 0.63 | Helped | Pass@K Policy Optimization: Solving Harder Reinforcement Learning Problems (NeurIPS 2025 · 0 citations) |
| Use hard clipping instead of smooth clipping | Offline finite-horizon MDP · plug-in occupancy-gradient error | O(1) | Hurt | Occupancy-based Policy Gradient: Estimation, Convergence, and Optimality (NeurIPS 2024 · 0 citations) |
| Use Nyström instead of random spectral dynamic embedding features | Simple Push (σ = 0.1/0.01) · winning rate | 49.69%/49.95% | No clear effect | Solving Zero-Sum Markov Games with Continuous State via Spectral Dynamic Embedding (NeurIPS 2024 · 0 citations) |
| Use Adam-CBO instead of M-CBO, adding adaptive momentum and timestep adjustment | LQG control, including the 16-dimensional setting · value function accuracy | Helped | Consensus Based Stochastic Optimal Control (ICML 2025) | |
| Add predictive tail probability reweighting instead of using the naïve NCVaR procedure | Continuous Blackjack, θ = 16.83 · gradient estimate variance | 97.59 -> 33.66 | Helped | Risk-Sensitive Policy Optimization via Predictive CVaR Policy Gradient (ICML 2024) |
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
- MATH · 2 of 12 papers
Baselines
- PPO · 5 of 12 papers
- REINFORCE · 2 of 12 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 33 papers; from OpenAlex and the papers' affiliations).
- Purdue University · 3 papers
- University of Michigan - Ann Arbor · 2 papers
- University of Oxford · 2 papers
- Google · 2 papers
- Google DeepMind · 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
- Trajectory reuse in policy gradients (68 papers)
- Alternative policy update rules (50 papers)
- Policy gradients for special settings (45 papers)
Elsewhere in the topic
- Model-based RL with guarantees (245 papers)
- Regularized scalable deep value learning (216 papers)
- Temporal-difference and discounting theory (143 papers)
- Tree search and policy optimization (109 papers)
- Safety layers and shielding (85 papers)
- Actor-critic theory (76 papers)
- Off-policy actor-critic with expressive policies (71 papers)
- Distributional value learning (60 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).