Research map / Reinforcement learning / Policy gradient and actor-critic
Safe policy improvement
Improve over a baseline policy with guarantees using bootstrapping, robust regret, and conformal control.
27 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Policy gradient and actor-critic, keyword cluster “safe · constrained · constraints”
State History Tried 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.
9 papers in 2023–24 and 12 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic held steady, from 2.6% to 2.6%.
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 · Safe Policy Improvement by Minimizing Robust Baseline Regret (NeurIPS 2016 · 64 citations)
- 2018 · Safe Reinforcement Learning via Formal Methods: Toward Safe Control Through Proof and Learning (AAAI 2018 · 189 citations)
- 2019 · Safe Policy Improvement with Baseline Bootstrapping (ICML 2019 · 79 citations)
- 2020 · Towards Safe Policy Improvement for Non-Stationary MDPs (NeurIPS 2020 · 9 citations)
- 2021 · Infinite Time Horizon Safety of Bayesian Neural Networks (NeurIPS 2021 · 1 citation)
- 2022 · Safe Exploration for Efficient Policy Evaluation and Comparison (ICML 2022 · 1 citation)
- 2023 · Guarded Policy Optimization with Imperfect Online Demonstrations (ICLR 2023 · 4 citations)
- 2024 · Towards Safe Policy Learning under Partial Identifiability: A Causal Approach (AAAI 2024 · 6 citations)
- 2025 · Neural Control and Certificate Repair via Runtime Monitoring (AAAI 2025 · 3 citations)
- 2026 · Safe In-Context Reinforcement Learning (ICML 2026 · 0 citations)
Earliest papers here
- Safe Policy Improvement by Minimizing Robust Baseline Regret (NeurIPS 2016 · 64 citations)
- Safe Reinforcement Learning via Formal Methods: Toward Safe Control Through Proof and Learning (AAAI 2018 · 189 citations)
- Safe Policy Improvement with Baseline Bootstrapping (ICML 2019 · 79 citations)
What has been tried
17 papers here report 131 measured changes (a component added, removed or swapped, compared with a baseline): helped 76%, no clear effect 5%, hurt 11%, mixed 8%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Use the combined observational and experimental bounds l1, l2, and l3 rather than only the observational bound | Synthetic data · mean outcome | Helped | Towards Safe Policy Learning under Partial Identifiability: A Causal Approach (AAAI 2024 · 6 citations) | |
| increase cluster size | Atari Skiing · testing effort around unsafe regions | Hurt | Test Where Decisions Matter: Importance-driven Testing for Deep Reinforcement Learning (NeurIPS 2024 · 1 citation) | |
| compose separately learned policies for ¬C and ¬D | grid world without a pure path · chattering | Hurt | Run-Time Task Composition with Safety Semantics (ICML 2024) |
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 21 papers; from OpenAlex and the papers' affiliations).
- University of Oxford · 2 papers
Use it
Download BibTeX of 27 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
- Safety layers and shielding (85 papers)
- Constrained policy optimization (48 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)
- 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)
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).