Summary of Recent Research on Multi-Agent AI Systems

# Multi-Agent AI Systems

Multi-agent AI systems involve multiple intelligent agents interacting or working collaboratively to solve problems, make decisions, or achieve goals. The research in this area explores communication mechanisms, coordination, cooperation strategies, and control of distributed agent networks, both in simulated environments and real-world applications.

## Summary of Work

1. **A Survey of Multi-Agent Deep Reinforcement Learning with Communication (2022)**  
   This work surveys multi-agent deep reinforcement learning (MADRL) focusing on communication among agents to coordinate behavior, broaden environment views, and improve collaboration. It proposes nine dimensions to analyze communication approaches in MADRL, identifies research trends, and suggests future directions by exploring possible communication design combinations.

2. **A Methodology to Engineer and Validate Dynamic Multi-level Multi-agent Based Simulations (2013)**  
   The paper presents a modeling and simulation methodology using the IRM4MLS meta-model for dynamic multi-level agent-based simulations. It supports representing complex systems across multiple scales and dynamically manages agent activation and aggregation to optimize computing resources without information loss.

3. **Augmenting the Action Space with Conventions to Improve Multi-agent Cooperation in Hanabi (2024)**  
   This research focuses on multi-agent reinforcement learning in the cooperative card game Hanabi, emphasizing the use of implicit knowledge sharing through "conventions," which help agents convey ideas based on mutually agreed rules. The approach improves cooperation performance in both self-play and cross-play settings, reducing complexity compared to prior methods.

4. **Learning Distributed Stabilizing Controllers for Multi-Agent Systems (2021)**  
   This study proposes model-free reinforcement learning algorithms to achieve distributed stabilization of heterogeneous multi-agent systems. It introduces two algorithms solving centralized and distributed LQR problems and proves convergence under certain conditions, demonstrated by simulation.

5. **MAEBE: Multi-Agent Emergent Behavior Framework (2025 in preview)**  
   MAEBE framework evaluates safety risks in multi-agent AI ensembles, highlighting emergent group dynamics that differ from single-agent behavior. It reveals how LLM moral preferences can shift in interactive multi-agent settings due to phenomena like peer pressure, underscoring new AI safety and alignment challenges.

## Papers

1. [A Survey of Multi-Agent Deep Reinforcement Learning with Communication](https://arxiv.org/pdf/2203.08975v2) - Zhu et al., 2022  
2. [A Methodology to Engineer and Validate Dynamic Multi-level Multi-agent Based Simulations](https://arxiv.org/pdf/1311.5108v1) - Soyez et al., 2013  
3. [Augmenting the Action Space with Conventions to Improve Multi-agent Cooperation in Hanabi](https://arxiv.org/pdf/2412.06333v3) - Bredell et al., 2024  
4. [Learning Distributed Stabilizing Controllers for Multi-Agent Systems](https://arxiv.org/pdf/2103.04480v1) - Jing et al., 2021  
5. [MAEBE: Multi-Agent Emergent Behavior Framework](https://arxiv.org/pdf/2506.03053v2) - Erisken et al., 2025