Research map / Reinforcement learning / Offline RL
Diffusion models for offline decision making
Use diffusion sampling and representation learning for offline and offline-to-online decisions.
82 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Offline RL, keyword cluster “offline online · transformer · online reinforcement”
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.
29 papers in 2023–24 and 41 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 12.1% to 16.6%.
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
- 2018 · Optimizing Interventions via Offline Policy Evaluation: Studies in Citizen Science (AAAI 2018 · 15 citations)
- 2019 · A Model-Based Reinforcement Learning with Adversarial Training for Online Recommendation (NeurIPS 2019)
- 2020 · Reinforcement-Learning Based Portfolio Management with Augmented Asset Movement Prediction States (AAAI 2020 · 168 citations)
- 2021 · A General Offline Reinforcement Learning Framework for Interactive Recommendation (AAAI 2021 · 63 citations)
- 2022 · Robust Task Representations for Offline Meta-Reinforcement Learning via Contrastive Learning (ICML 2022 · 10 citations)
- 2023 · Adversarial Counterfactual Environment Model Learning (NeurIPS 2023 · 6 citations)
- 2024 · NondBREM: Nondeterministic Offline Reinforcement Learning for Large-Scale Order Dispatching (AAAI 2024 · 6 citations)
- 2025 · Active Reinforcement Learning Strategies for Offline Policy Improvement (AAAI 2025 · 1 citation)
- 2026 · Fast and Highly Expressive Policy Learning for Offline Reinforcement Learning via Bootstrapped Flow Q-Learning (ICML 2026 · 0 citations)
Earliest papers here
- Optimizing Interventions via Offline Policy Evaluation: Studies in Citizen Science (AAAI 2018 · 15 citations)
- A Model-Based Reinforcement Learning with Adversarial Training for Online Recommendation (NeurIPS 2019)
- Reinforcement-Learning Based Portfolio Management with Augmented Asset Movement Prediction States (AAAI 2020 · 168 citations)
What has been tried
53 papers here report 518 measured changes (a component added, removed or swapped, compared with a baseline): helped 64%, no clear effect 8%, hurt 21%, 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 the REM multi-head value network | Ride-hailing simulator, two-week evaluation · dispatch score | 2005392.15 -> 2043643.57 | Helped | NondBREM: Nondeterministic Offline Reinforcement Learning for Large-Scale Order Dispatching (AAAI 2024 · 6 citations) |
| remove loss scaling | D4RL offline RL tasks · normalized score | by 37.8\% on average | Hurt | Consistency Models as a Rich and Efficient Policy Class for Reinforcement Learning (ICLR 2024 · 2 citations) |
| Ensemble behavior scores from multiple diffusion times rather than using only the near-zero time score | D4RL tasks · performance | Helped | Score Regularized Policy Optimization through Diffusion Behavior (ICLR 2024 · 2 citations) | |
| Add active initial state selection to random exploration | maze2d-large-easy and maze2d-large-hard · mean normalized score | 45.5 -> 88.1; 25.0 -> 74.6 | Helped | Active Reinforcement Learning Strategies for Offline Policy Improvement (AAAI 2025 · 1 citation) |
| Add distance metric learning to the context encoder | ID and OOD MuJoCo tasks · adaptation performance | Helped | Entropy Regularized Task Representation Learning for Offline Meta-Reinforcement Learning (AAAI 2025 · 1 citation) | |
| Use action chunking without a behavior constraint | OGBench overall (25 tasks), online · success rate | 67 -> 61 | Hurt | Reinforcement Learning with Action Chunking (NeurIPS 2025 · 1 citation) |
| replace transition generation with trajectory generation | Locomotion · normalized return | 93.5 -> 97.1 | Helped | ATraDiff: Accelerating Online Reinforcement Learning with Imaginary Trajectories (ICML 2024 · 1 citation) |
| replace the local-Lipschitz modified MSE dynamics loss with standard MSE | point-maze goal-reaching tasks · success rate | Hurt | MGDA: Model-based Goal Data Augmentation for Offline Goal-conditioned Weighted Supervised Learning (AAAI 2025 · 0 citations) | |
| Remove conservative regularization (λ=0; β=10) | Alg0, unseen P_bbob · Perf. | 9.889E-01 ±7.780E-03 -> 9.833E-01 ±1.424E-02 | Hurt | Meta-Black-Box-Optimization through Offline Q-function Learning (ICML 2025 · 0 citations) |
| Replace aGRAPA betting with online Newton step betting | Instruction induction tasks · TPR | No clear effect | Adaptive Learn-then-Test: Statistically Valid and Efficient Hyperparameter Selection (ICML 2025 · 0 citations) | |
| Replace one-step guidance with backpropagation through time | 50 state-based OGBench tasks · score | Hurt | Flow Q-Learning (ICML 2025 · 0 citations) | |
| Use Neighborhood Preference Normalization instead of Preference Predicted Normalization | Complete Amateur Hopper · HV | 1.94±.01 -> 1.98±.02 | Helped | MODULI: Unlocking Preference Generalization via Diffusion Models for Offline Multi-Objective Reinforcement Learning (ICML 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 22 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 Maze2D · 3 of 22 papers
- D4RL AntMaze · 3 of 22 papers
- D4RL · 2 of 22 papers
- D4RL MuJoCo · 2 of 22 papers
- D4RL Kitchen · 2 of 22 papers
Baselines
- IQL · 6 of 22 papers
- IDQL · 4 of 22 papers
- Diffuser · 4 of 22 papers
- BC · 3 of 22 papers
- SRPO · 3 of 22 papers
- IFQL · 3 of 22 papers
- DQL · 3 of 22 papers
- AdaptDiffuser · 3 of 22 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 70 papers; from OpenAlex and the papers' affiliations).
- Tsinghua University · 8 papers
- Nanjing University · 6 papers
- Microsoft · 3 papers
- Tianjin University · 3 papers
- University of California, Berkeley · 3 papers
- Westlake University · 2 papers
- University of Illinois, Urbana Champaign · 2 papers
- Stanford University · 2 papers
- Google · 2 papers
- Tencent · 2 papers
Use it
Download BibTeX of 34 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
- Offline-to-online fine-tuning (99 papers)
- Decision Transformer sequence models (40 papers)
- Offline policy evaluation and selection (8 papers)
Elsewhere in the topic
- Offline RL sample complexity theory (156 papers)
- Simple practical offline RL (144 papers)
- Pessimism and uncertainty in offline RL (59 papers)
- Behavior regularization for offline RL (35 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).