Research map / Multi-agent systems and games / Cooperative multi-agent RL
Learned inter-agent communication
Learns when, what, and how to communicate among agents under bandwidth and delay limits.
58 accepted papers at ICML, NeurIPS, ICLR, CVPR and AAAI (2016–2026) · an approach in Cooperative multi-agent RL, keyword cluster “reinforcement marl · communication · marl algorithms”
State History Tried Baselines 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.
15 papers in 2023–24 and 17 in 2025–26 (at ICML, NeurIPS, ICLR, CVPR and AAAI); its share of the topic shrank, from 10.6% to 7.5%.
Open questions the data points to
- Papers report IC3Net on Traffic Junction Hard (success rate) anywhere from 40.47 to 73.26: comparisons against it depend on whose number you take.
- Papers report No communication on Traffic Junction hard (success rate) anywhere from 48.1 to 74.1: comparisons against it depend on whose number you take.
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 · Learning to Communicate with Deep Multi-Agent Reinforcement Learning (NeurIPS 2016 · 1,057 citations)
- 2018 · Learning Attentional Communication for Multi-Agent Cooperation (NeurIPS 2018 · 303 citations)
- 2019 · Learning to Teach in Cooperative Multiagent Reinforcement Learning (AAAI 2019 · 101 citations)
- 2020 · Learning Agent Communication under Limited Bandwidth by Message Pruning (AAAI 2020 · 80 citations)
- 2021 · Communication in Multi-Agent Reinforcement Learning: Intention Sharing (ICLR 2021 · 30 citations)
- 2022 · Efficient Multi-agent Communication via Self-supervised Information Aggregation (NeurIPS 2022 · 7 citations)
- 2023 · DACOM: Learning Delay-Aware Communication for Multi-Agent Reinforcement Learning (AAAI 2023 · 20 citations)
- 2024 · T2MAC: Targeted and Trusted Multi-Agent Communication through Selective Engagement and Evidence-Driven Integration (AAAI 2024 · 18 citations)
- 2025 · CoDe: Communication Delay-Tolerant Multi-Agent Collaboration via Dual Alignment of Intent and Timeliness (AAAI 2025 · 7 citations)
- 2026 · Communication-efficient Multi-Agent Reinforcement Learning with Spatiotemporal Information Hub (AAAI 2026 · 3 citations)
Earliest papers here
- Learning to Communicate with Deep Multi-Agent Reinforcement Learning (NeurIPS 2016 · 1,057 citations)
- ConTaCT: Deciding to Communicate during Time-Critical Collaborative Tasks in Unknown, Deterministic Domains (AAAI 2016 · 33 citations)
- Strengthening Agents Strategic Ability with Communication (AAAI 2016 · 1 citation)
What has been tried
40 papers here report 392 measured changes (a component added, removed or swapped, compared with a baseline): helped 70%, no clear effect 8%, hurt 19%, mixed 3%. Papers mostly publish changes that helped, so "helped" is over-represented.
Ablations in recent papers
| What was changed | Measured on | Result | Effect | Paper |
|---|---|---|---|---|
| Replace the categorical distribution with a Dirichlet distribution | 1o_10b_vs_1r · Win rate | Helped | T2MAC: Targeted and Trusted Multi-Agent Communication through Selective Engagement and Evidence-Driven Integration (AAAI 2024 · 18 citations) | |
| add graph information bottleneck optimization | MMM2 (SMAC) · win rate | 60.07±5.75 -> 82.87±2.93 | Helped | Learning Efficient and Robust Multi-Agent Communication via Graph Information Bottleneck (AAAI 2024 · 11 citations) |
| Set communication regularization coefficient η to 0.001 rather than the other tested coefficients | Predator-Prey · reward | Helped | PMAC: Personalized Multi-Agent Communication (AAAI 2024 · 8 citations) | |
| Use RGMComm with 2 message labels | Predator-Prey · normalized mean episode reward | Helped | RGMComm: Return Gap Minimization via Discrete Communications in Multi-Agent Reinforcement Learning (AAAI 2024 · 8 citations) | |
| replace TARMAC message aggregation with the decomposable policy net | Food Collector, Predator Prey, Treasure Hunt under attacks · average timesteps | Helped | Robust Communicative Multi-Agent Reinforcement Learning with Active Defense (AAAI 2024 · 8 citations) | |
| add intent alignment | SMAC 1o_10b_vs_1r, time-varying delay N(3,2) · win rate | 35.6 -> 62.1 | Helped | CoDe: Communication Delay-Tolerant Multi-Agent Collaboration via Dual Alignment of Intent and Timeliness (AAAI 2025 · 7 citations) |
| Replace the graph coarsening network with a QMIX hybrid network | one hard and two super hard SMAC scenarios · performance | No clear effect | Bridging Training and Execution via Dynamic Directed Graph-Based Communication in Cooperative Multi-Agent Systems (AAAI 2025 · 5 citations) | |
| Increase communication graph sparsity | SMAC tasks · win rate | Helped | Learning Multi-Agent Communication from Graph Modeling Perspective (ICLR 2024 · 4 citations) | |
| Determine decision priority with SeqComm instead of random priority at each timestep | SMACv2, all maps · win rate | Helped | Multi-Agent Coordination via Multi-Level Communication (NeurIPS 2024 · 3 citations) | |
| Add lossless-message MI maximization | Simple_Tag, Heavy DBC · average episode cumulative reward | 68.7 -> 81.7 | Helped | Multi-Agent Reinforcement Learning with Communication-Constrained Priors (NeurIPS 2025 · 0 citations) |
| Remove memory blocks from message generators | MAgent AdversarialPursuit · task performance | Hurt | Exponential Topology-enabled Scalable Communication in Multi-agent Reinforcement Learning (ICLR 2025 · 0 citations) | |
| Use hat mapping rather than the other tested ICP variants | Hanabi 4-player full version · average score | Helped | Learning to Communicate Through Implicit Communication Channels (ICLR 2025 · 0 citations) |
Baseline numbers that disagree
The same baseline on the same benchmark and metric, reported more than 5% apart by different papers here. Training budgets, data splits and reimplementations differ, so state which number you compare with.
| Baseline | Benchmark | Metric | Reported | Papers |
|---|---|---|---|---|
| IC3Net | Traffic Junction Hard | success rate | 40.47 – 73.26 | 2 papers, e.g. NeurIPS 2020 and AAAI 2024 |
| No communication | Traffic Junction hard | success rate | 48.1 – 74.1 | 2 papers, e.g. NeurIPS 2020 and ICML 2019 |
| CommNet | Traffic Junction Hard | success rate | 51.56 – 78.9 | 2 papers, e.g. AAAI 2024 and ICML 2019 |
| IC3Net | Traffic Junction Medium | success rate | 78.08 – 87.83 | 2 papers, e.g. NeurIPS 2020 and AAAI 2024 |
What reviewers will expect
Datasets and baselines named in the 10 papers of this approach from 2025–2026 that have a full-text summary. Leaving out the common ones is a frequent review question.
Datasets
- Hallway · 3 of 10 papers
- SMAC · 2 of 10 papers
Baselines
- QMIX · 4 of 10 papers
- MAIC · 3 of 10 papers
- TarMAC · 2 of 10 papers
- NDQ · 2 of 10 papers
- TMC · 2 of 10 papers
Recent preprints
Loading recent arXiv preprints…
Who publishes on it
Institutions on the most papers of this approach since 2023 (of 32 papers; from OpenAlex and the papers' affiliations).
- China University of Mining and Technology · 2 papers
- Tianjin University · 2 papers
- Mine Digitization Engineering Research Center of Ministry of Education of the People’s Republic of China · 2 papers
- Carnegie Mellon University · 2 papers
- Tsinghua University · 2 papers
- University of Chinese Academy of Sciences · 2 papers
- Chinese Academy of Sciences · 2 papers
- Shandong University · 2 papers
- Dyson Robot Learning Lab · 2 papers
- University of Oxford · 2 papers
Use it
Download BibTeX of 38 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
- Cooperative MARL algorithms (149 papers)
- MARL benchmarks and grouping (121 papers)
- Intrinsic-reward exploration for MARL (34 papers)
Elsewhere in the topic
- Decentralized actor-critic MARL (77 papers)
- Offline and networked MARL (65 papers)
- Coordination graphs (50 papers)
- Value factorization (39 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).