Research map / Reinforcement learning / Visual and representation RL
Task-inference meta-RL
Meta-train context encoders and task selection so agents infer new tasks and adapt quickly.
57 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2018–2026) · an approach in Visual and representation RL, keyword cluster “meta · adaptation · context”
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.
16 papers in 2023–24 and 10 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 2.8% to 1.4%.
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 · Meta-Reinforcement Learning of Structured Exploration Strategies (NeurIPS 2018 · 174 citations)
- 2019 · Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables (ICML 2019 · 214 citations)
- 2020 · Improving Generalization in Meta Reinforcement Learning using Learned Objectives (ICLR 2020 · 54 citations)
- 2021 · Towards Effective Context for Meta-Reinforcement Learning: an Approach based on Contrastive Learning (AAAI 2021 · 27 citations)
- 2022 · Adapt to Environment Sudden Changes by Learning a Context Sensitive Policy (AAAI 2022 · 28 citations)
- 2023 · Meta-Reinforcement Learning Based on Self-Supervised Task Representation Learning (AAAI 2023 · 13 citations)
- 2024 · Meta-Reinforcement Learning via Exploratory Task Clustering (AAAI 2024 · 7 citations)
- 2025 · Meta-World+: An Improved, Standardized, RL Benchmark (NeurIPS 2025 · 5 citations)
- 2026 · Self-Improving Skill Learning for Robust Skill-based Meta-Reinforcement Learning (ICLR 2026 · 0 citations)
Earliest papers here
- Meta-Reinforcement Learning of Structured Exploration Strategies (NeurIPS 2018 · 174 citations)
- Efficient Off-Policy Meta-Reinforcement Learning via Probabilistic Context Variables (ICML 2019 · 214 citations)
- Guided Meta-Policy Search (NeurIPS 2019 · 32 citations)
What has been tried
50 papers here report 428 measured changes (a component added, removed or swapped, compared with a baseline): helped 66%, no clear effect 12%, hurt 18%, mixed 4%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Set the cluster number to 2 instead of 4 | Ant-Goal · return | -125.4±5.1 -> -132.3±7.6 | Hurt | Meta-Reinforcement Learning via Exploratory Task Clustering (AAAI 2024 · 7 citations) |
| add the FOCAL contrastive objective to reconstruction-only UNICORN-SS-0 | Ant-Dir · average return | 220±16 -> 267±14 (IID); 200±9 -> 236±18 (OOD) | Helped | Towards an Information Theoretic Framework of Context-Based Offline Meta-Reinforcement Learning (NeurIPS 2024 · 6 citations) |
| add task representation to the transformer input | Point-Robot, Cheetah-Dir, Ant-Dir, Medium · test return | Point-Robot: -16.44 -> -10.18; Cheetah-Dir: 551.30 -> 608.18; Ant-Dir: 290.44 -> 412.00 | Helped | Meta-DT: Offline Meta-RL as Conditional Sequence Modeling with World Model Disentanglement (NeurIPS 2024 · 6 citations) |
| remove the linear feature baseline | ML10 V2 · success rate | 30.1 ± 11.45 -> 2.00 ± 3.46 | Hurt | Meta-World+: An Improved, Standardized, RL Benchmark (NeurIPS 2025 · 5 citations) |
| add LCARD to context encoder training | Plate-Slide · adaptation performance | Helped | MetaCARD: Meta-Reinforcement Learning with Task Uncertainty Feedback via Decoupled Context-Aware Reward and Dynamics Components (AAAI 2024 · 4 citations) | |
| Remove contrastive learning | Kitchen manipulation · sample efficiency and performance | Hurt | Decoupling Meta-Reinforcement Learning with Gaussian Task Contexts and Skills (AAAI 2024 · 3 citations) | |
| Add local reconstruction instead of global reconstruction | Rooms-4 · total return | 14.1 +- 3.0 -> 88.1+-39.2 | Helped | MAMBA: an Effective World Model Approach for Meta-Reinforcement Learning (ICLR 2024 · 2 citations) |
| add trajectory encoder pre-training with the world model | Point-Robot, Mixed · return | -7.7 -> -7.2 | Helped | Text-to-Decision Agent: Offline Meta-Reinforcement Learning from Natural Language Supervision (NeurIPS 2025 · 1 citation) |
| Reduce negative sample space via SaNCE (current-task sampling) vs InfoNCE across all tasks | Sample-limited analysis (buffer size and contrastive batch size, SlimHumanoid) · average return | No clear effect | Skill-aware Mutual Information Optimisation for Zero-shot Generalisation in Reinforcement Learning (NeurIPS 2024 · 1 citation) | |
| Replace Q-dependent actor and critic objectives with scale-independent classification updates | Meta-World ML45 · success rate | Helped | AMAGO-2: Breaking the Multi-Task Barrier in Meta-Reinforcement Learning with Transformers (NeurIPS 2024 · 1 citation) | |
| Add training proportion as an optimizer input | Breakout · mean final return | Helped | Can Learned Optimization Make Reinforcement Learning Less Difficult? (NeurIPS 2024 · 1 citation) | |
| Add model-based planning to the trained dynamics model | Darkroom and Reach-v2 · performance | Helped | Distilling Reinforcement Learning Algorithms for In-Context Model-Based Planning (ICLR 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 5 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Meta-World ML10 · 2 of 5 papers
Baselines
- RL2 · 3 of 5 papers
- MAML · 2 of 5 papers
- AD · 2 of 5 papers
- PEARL · 2 of 5 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 26 papers; from OpenAlex and the papers' affiliations).
- University of Oxford · 3 papers
- Meta · 2 papers
- Technion · 2 papers
- Technion - Israel Institute of Technology · 2 papers
- Wayve · 2 papers
- University of Texas at Austin · 2 papers
- The Chinese University of Hong Kong · 2 papers
- Nanjing University · 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
- Task structure and curricula for meta-RL (44 papers)
Elsewhere in the topic
- In-context and adaptive RL (202 papers)
- Scalable RL systems and environments (179 papers)
- Probabilistic and autonomous RL (177 papers)
- RL for combinatorial routing (158 papers)
- Learned state and action representations (148 papers)
- Dexterous manipulation transfer (141 papers)
- RL with tree search for discovery (121 papers)
- Inferring decision-making from behavior (93 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).