Research map / Reinforcement learning / Offline RL
Offline RL sample complexity theory
Prove sample-complexity bounds for offline RL under limited data coverage.
156 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Offline RL, keyword cluster “offline reinforcement · offline rl · conservative”
State History Benchmarks Tried Checklist Preprints Who Export
Where it stands
Still moving The best reported result is still improving on 3 of 6 tracked benchmarks.
53 papers in 2023–24 and 66 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 22.1% to 26.7%.
Open questions the data points to
- D4RL Medium HalfCheetah (normalized score): best went from 47.1 (up to 2024) to 54.16 in 2025 — Energy-Weighted Flow Matching for Offline Reinforcement Learning (ICLR 2025)
- D4RL Medium Hopper (normalized score): best went from 57.1 (up to 2024) to 100.8 in 2025 — Learning on One Mode: Addressing Multi-modality in Offline Reinforcement Learning (ICLR 2025 · 1 citation)
- D4RL Medium-Replay Hopper (normalized score): best went from 86.2 (up to 2024) to 101.25 in 2025 — Energy-Weighted Flow Matching for Offline Reinforcement Learning (ICLR 2025)
What looks settled
- D4RL Medium-Replay HalfCheetah (normalized score): best up to 2024 57.6, best in 2025–26 48.8 from 2 papers, at most 0.2% better, if at all.
No longer reported by papers here: halfcheetah-medium-expert (normalized return, last in 2024); hopper-medium (normalized return, last in 2024).
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
- 2018 · An Optimal Online Method of Selecting Source Policies for Reinforcement Learning (AAAI 2018 · 30 citations)
- 2020 · An Optimistic Perspective on Offline Reinforcement Learning (ICML 2020 · 95 citations)
- 2021 · Uncertainty Weighted Actor-Critic for Offline Reinforcement Learning (ICML 2021 · 45 citations)
- 2022 · A Policy-Guided Imitation Approach for Offline Reinforcement Learning (NeurIPS 2022 · 10 citations)
- 2023 · Towards Safe Mechanical Ventilation Treatment Using Deep Offline Reinforcement Learning (AAAI 2023 · 16 citations)
- 2024 · Diffusion-DICE: In-Sample Diffusion Guidance for Offline Reinforcement Learning (NeurIPS 2024 · 4 citations)
- 2025 · Any-step Dynamics Model Improves Future Predictions for Online and Offline Reinforcement Learning (ICLR 2025 · 1 citation)
- 2026 · Trajectory-Level Data Augmentation for Offline Reinforcement Learning (ICML 2026 · 0 citations)
Earliest papers here
- An Optimal Online Method of Selecting Source Policies for Reinforcement Learning (AAAI 2018 · 30 citations)
- An Optimistic Perspective on Offline Reinforcement Learning (ICML 2020 · 95 citations)
- Keep Doing What Worked: Behavior Modelling Priors for Offline Reinforcement Learning (ICLR 2020 · 51 citations)
Benchmark records
The best number each year that beat every earlier one, among papers of this approach reporting their own method. Numbers are extracted by AI from each paper's text; check the paper before citing.
D4RL Medium HalfCheetah · normalized score (higher is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 47.1 | Model-Based Offline Weighted Policy Optimization (Student Abstract) (AAAI 2023 · 0 citations) |
| 2025 | 54.16 | Energy-Weighted Flow Matching for Offline Reinforcement Learning (ICLR 2025) |
D4RL Medium Hopper · normalized score (higher is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 57.1 | Offline Reinforcement Learning via High-Fidelity Generative Behavior Modeling (ICLR 2023 · 6 citations) |
| 2025 | 100.8 | Learning on One Mode: Addressing Multi-modality in Offline Reinforcement Learning (ICLR 2025 · 1 citation) |
D4RL Medium-Replay HalfCheetah · normalized score (higher is better)
4 papers, 2023–2025 · flat since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 57.6 | Model-Based Offline Weighted Policy Optimization (Student Abstract) (AAAI 2023 · 0 citations) |
D4RL Medium-Replay Hopper · normalized score (higher is better)
4 papers, 2023–2025 · still improving
| Year | Best so far | Set by |
|---|---|---|
| 2023 | 86.2 | Offline Reinforcement Learning via High-Fidelity Generative Behavior Modeling (ICLR 2023 · 6 citations) |
| 2025 | 101.25 | Energy-Weighted Flow Matching for Offline Reinforcement Learning (ICLR 2025) |
halfcheetah-medium-expert · normalized return (higher is better)
4 papers, 2021–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 93.5 | Offline RL Without Off-Policy Evaluation (NeurIPS 2021 · 16 citations) |
| 2023 | 99.1 | Recovering from Out-of-sample States via Inverse Dynamics in Offline Reinforcement Learning (NeurIPS 2023 · 3 citations) |
hopper-medium · normalized return (higher is better)
4 papers, 2021–2024 · not reported since 2024
| Year | Best so far | Set by |
|---|---|---|
| 2021 | 83.3 | Offline RL Without Off-Policy Evaluation (NeurIPS 2021 · 16 citations) |
| 2022 | 94.2 | Koopman Q-learning: Offline Reinforcement Learning via Symmetries of Dynamics (ICML 2022 · 1 citation) |
| 2023 | 100.5 | Behavior Estimation from Multi-Source Data for Offline Reinforcement Learning (AAAI 2023 · 1 citation) |
What has been tried
97 papers here report 885 measured changes (a component added, removed or swapped, compared with a baseline): helped 61%, no clear effect 9%, hurt 20%, mixed 9%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Add the guide stage before selecting candidate actions | D4RL evaluation · normalized score | Helped | Diffusion-DICE: In-Sample Diffusion Guidance for Offline Reinforcement Learning (NeurIPS 2024 · 4 citations) | |
| adapt temporal distance using ICM intrinsic rewards | 4 Walker tasks · computational efficiency at 100K steps | 1.25 × | Helped | CUDC: A Curiosity-Driven Unsupervised Data Collection Method with Adaptive Temporal Distances for Offline Reinforcement Learning (AAAI 2024 · 3 citations) |
| add dynamically trajectory-return-weighted Q-aid to RCSL | MuJoCo medium · normalized return | 74.7 -> 81.2 | Helped | Adaptive $Q$-Aid for Conditional Supervised Learning in Offline Reinforcement Learning (NeurIPS 2024 · 3 citations) |
| learn a policy with delphic-uncertainty pessimism rather than behaviour cloning | Real-world medical dataset · policy action discrepancy from the data policy | Mixed | Delphic Offline Reinforcement Learning under Nonidentifiable Hidden Confounding (ICLR 2024 · 2 citations) | |
| Disable state normalization | NetworkGym average across datasets · return per step | TD3+BC: 0.511 ± 0.047 -> 0.849 ± 0.045; IQL: 0.779 ± 0.047 -> 0.848 ± 0.042; CQL: 0.749 ± 0.047 -> 0.779 ± 0.056; BC: 0. | Mixed | NetworkGym: Reinforcement Learning Environments for Multi-Access Traffic Management in Network Simulation (NeurIPS 2024 · 2 citations) |
| Replace continuous advantage conditioning with a fixed ξ=1 for all samples | Gym random-expert halfcheetah · test return | 90.3 -> 5.8 | Hurt | A2PO: Towards Effective Offline Reinforcement Learning from an Advantage-aware Perspective (NeurIPS 2024 · 2 citations) |
| Vary the maximum backtracking length | hopper-medium-v2 and walker2d-medium-v2 · performance | No clear effect | Any-step Dynamics Model Improves Future Predictions for Online and Offline Reinforcement Learning (ICLR 2025 · 1 citation) | |
| Increase offline RL training from 100K to 200K steps | Downstream tasks · offline RL performance | No clear effect | Behavioral Entropy-Guided Dataset Generation for Offline Reinforcement Learning (ICLR 2025 · 1 citation) | |
| Set the single-step accuracy quantile ζ1 to 0.97 rather than 0.99 | Hopper · learning stability | Hurt | On Rollouts in Model-Based Reinforcement Learning (ICLR 2025 · 1 citation) | |
| Vary the number of Gaussian mixture components M | D4RL medium-replay · normalized score | Helped | Learning on One Mode: Addressing Multi-modality in Offline Reinforcement Learning (ICLR 2025 · 1 citation) | |
| Use a large modern architecture instead of a large simple architecture | MOOD mixed-objective datasets · normalized return | 373.6 -> 370.0 | No clear effect | Simple Ingredients for Offline Reinforcement Learning (ICML 2024 · 1 citation) |
| Add the step-estimation module rather than fixing stitching length at 40 | Hopper · normalized score | Helped | DiffStitch: Boosting Offline Reinforcement Learning with Diffusion-based Trajectory Stitching (ICML 2024 · 1 citation) |
What reviewers will expect
Datasets and baselines named in the 36 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- D4RL AntMaze · 4 of 36 papers
- D4RL HalfCheetah · 4 of 36 papers
- D4RL Hopper · 4 of 36 papers
- D4RL Walker2d · 4 of 36 papers
- OGBench · 3 of 36 papers
- D4RL Adroit · 2 of 36 papers
- D4RL Maze2D · 2 of 36 papers
- D4RL Kitchen · 2 of 36 papers
Baselines
- CQL · 9 of 36 papers
- IQL · 6 of 36 papers
- TD3+BC · 5 of 36 papers
- BC · 4 of 36 papers
- EDAC · 3 of 36 papers
- SPOT · 3 of 36 papers
- ReBRAC · 3 of 36 papers
- MOPO · 3 of 36 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 119 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 9 papers
- UC Santa Barbara · 4 papers
- Mila · 4 papers
- Princeton University · 4 papers
- Google · 4 papers
- Meta · 4 papers
- University of Texas at Austin · 3 papers
- Microsoft · 3 papers
- Beijing Institute of Technology · 3 papers
- Shanghai Jiao Tong University · 3 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
- Simple practical offline RL (144 papers)
- Pessimism and uncertainty in offline RL (59 papers)
- Behavior regularization for offline RL (35 papers)
Elsewhere in the topic
- Offline-to-online fine-tuning (99 papers)
- Diffusion models for offline decision making (82 papers)
- Decision Transformer sequence models (40 papers)
- Offline policy evaluation and selection (8 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).