Research map / Multi-agent systems and games / Cooperative multi-agent RL
Offline and networked MARL
Learns multi-agent policies from fixed datasets or with networked, scalable value decomposition.
65 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2017–2026) · an approach in Cooperative multi-agent RL, keyword cluster “decentralized · value · joint”
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 30 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic grew, from 11.3% to 13.2%.
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 · Stabilising Experience Replay for Deep Multi-Agent Reinforcement Learning (ICML 2017 · 420 citations)
- 2018 · Fully Decentralized Multi-Agent Reinforcement Learning with Networked Agents (ICML 2018 · 327 citations)
- 2019 · Value Propagation for Decentralized Networked Deep Multi-agent Reinforcement Learning (NeurIPS 2019 · 33 citations)
- 2020 · Q-value Path Decomposition for Deep Multiagent Reinforcement Learning (ICML 2020 · 26 citations)
- 2021 · Believe What You See: Implicit Constraint Approach for Offline Multi-Agent Reinforcement Learning (NeurIPS 2021 · 18 citations)
- 2022 · Trust Region Policy Optimisation in Multi-Agent Reinforcement Learning (ICLR 2022 · 82 citations)
- 2023 · Offline Multi-Agent Reinforcement Learning with Implicit Global-to-Local Value Regularization (NeurIPS 2023 · 10 citations)
- 2024 · STAS: Spatial-Temporal Return Decomposition for Solving Sparse Rewards Problems in Multi-agent Reinforcement Learning (AAAI 2024 · 9 citations)
- 2025 · Offline Multi-Agent Reinforcement Learning via In-Sample Sequential Policy Optimization (AAAI 2025 · 5 citations)
- 2026 · Value Aggregation with Uncertainty in Online Decentralized MARL (ICML 2026 · 0 citations)
Earliest papers here
- Stabilising Experience Replay for Deep Multi-Agent Reinforcement Learning (ICML 2017 · 420 citations)
- Fully Decentralized Multi-Agent Reinforcement Learning with Networked Agents (ICML 2018 · 327 citations)
- Value Propagation for Decentralized Networked Deep Multi-agent Reinforcement Learning (NeurIPS 2019 · 33 citations)
What has been tried
46 papers here report 399 measured changes (a component added, removed or swapped, compared with a baseline): helped 62%, no clear effect 12%, hurt 23%, 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 |
|---|---|---|---|---|
| use 1 Shapley-value approximation sample instead of more samples | Predator-prey (3 agents) · average reward | Hurt | STAS: Spatial-Temporal Return Decomposition for Solving Sparse Rewards Problems in Multi-agent Reinforcement Learning (AAAI 2024 · 9 citations) | |
| use inter-agent attention rather than independent U-Nets | MPE tasks, especially World · average normalized score | Helped | MADiff: Offline Multi-agent Learning with Diffusion Models (NeurIPS 2024 · 8 citations) | |
| add policy entropy | M-NE game, imbalanced · convergence to global optimum | Helped | Offline Multi-Agent Reinforcement Learning via In-Sample Sequential Policy Optimization (AAAI 2025 · 5 citations) | |
| increase κ from 1 to at least 3 | Safe ManyAgent Ant (6×1), Safe Ant (8×1), and Safe Coupled HalfCheetah (12×1) · average episode reward | Helped | Scalable Constrained Policy Optimization for Safe Multi-agent Reinforcement Learning (NeurIPS 2024 · 4 citations) | |
| use adaptive rather than uniform grouping | Gather · total reward | 11.6% | Helped | GTDE: Grouped Training with Decentralized Execution for Multi-agent Actor-Critic (AAAI 2025 · 3 citations) |
| Replace the learned mask with random masking | Monaco and NewYork51 · performance | Hurt | Bayesian Ego-graph inference for Networked Multi-Agent Reinforcement Learning (NeurIPS 2025 · 2 citations) | |
| Use simultaneous rather than sequential agent updates | Training schemes · performance | No clear effect | Backpropagation Through Agents (AAAI 2024 · 1 citation) | |
| Use RBF kernel parametrization instead of discretization | Aggregation d=4 · training performance | Helped | Learning Decentralized Partially Observable Mean Field Control for Artificial Collective Behavior (ICLR 2024 · 1 citation) | |
| increase the subsampling parameter k | Constrained Exploration (n=8) · discounted cumulative reward | Helped | Mean-Field Sampling for Cooperative Multi-Agent Reinforcement Learning (NeurIPS 2025 · 0 citations) | |
| remove the hierarchy and use a flat MAPPO policy | MetaDrive environments · success rate | 0.99 (±0.01) -> 0 | Hurt | HMARL-CBF – Hierarchical Multi-Agent Reinforcement Learning with Control Barrier Functions for Safety-Critical Autonomous Systems (NeurIPS 2025 · 0 citations) |
| use stochastic rather than deterministic rollouts to learn the DGCBF | unspecified comparison · cost | Hurt | Discrete GCBF Proximal Policy Optimization for Multi-agent Safe Optimal Control (ICLR 2025 · 0 citations) | |
| Set regularization parameter α around 10 rather than other tested values | SMACv2 and MaMuJoCo · win rate and returns | Helped | ComaDICE: Offline Cooperative Multi-Agent Reinforcement Learning with Stationary Distribution Shift Regularization (ICLR 2025 · 0 citations) |
What reviewers will expect
Datasets and baselines named in the 19 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- SMACv1 offline datasets (MAPPO-generated) · 2 of 19 papers
Baselines
- BC · 4 of 19 papers
- BCQ · 2 of 19 papers
- CQL · 2 of 19 papers
- AlberDICE · 2 of 19 papers
- IPPO · 2 of 19 papers
- OMAR · 2 of 19 papers
- OMIGA · 2 of 19 papers
- InforMARL · 2 of 19 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 46 papers; from OpenAlex and the papers' affiliations).
- Peking University · 4 papers
- Zhejiang University · 2 papers
- The Chinese University of Hong Kong · 2 papers
- King's College London · 2 papers
- Nanyang Technological University · 2 papers
- Sun Yat-sen University · 2 papers
- Nanjing University · 2 papers
- Tencent · 2 papers
- University of South Carolina · 2 papers
- Tsinghua University · 2 papers
Use it
Download BibTeX of 37 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
- Decentralized actor-critic MARL (77 papers)
- Coordination graphs (50 papers)
- Value factorization (39 papers)
Elsewhere in the topic
- Cooperative MARL algorithms (149 papers)
- MARL benchmarks and grouping (121 papers)
- Learned inter-agent communication (58 papers)
- Intrinsic-reward exploration for MARL (34 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).