When Do Multi-Agent Systems Help? An Information Bottleneck Perspective
Wendi Yu, Lianhao Zhou, Xiangjue Dong, Sai Sudarshan Barath, Declan Staunton, Byung-Jun Yoon, Xiaoning Qian, James Caverlee, Shuiwang Ji
Read on arXiv →
Key claim
Multi-agent systems optimize information transfer under constraints.
In plain English
Multi-agent systems (MAS) are becoming popular for complex tasks, but their advantages over single-agent systems (SAS) are not well understood. Current approaches often overlook how communication constraints affect performance. This paper introduces an information bottleneck perspective, clarifying when MAS can outperform SAS based on the efficiency of information transfer. Builders should pay attention to these insights to optimize their multi-agent designs, especially in scenarios with limited communication.
Introduces a new perspective on the trade-offs in multi-agent systems.
Supported by controlled experiments across multiple benchmarks and model scales.
Deep reliability assessment
The methodology supports the claim that MAS can outperform SAS under certain conditions, particularly when relay bandwidth is limited and context reduction is beneficial. However, it overclaims by suggesting MAS is generally advantageous without fully accounting for scenarios where stronger models may not benefit from context compression.
Reproducibility
Yes, the paper provides a GitHub repository for the code: https://github.com/divelab/MAS-SAS.
Key figure
Figure 1 illustrates the predicted MAS gain by relay regime, showing that MAS helps when context reduction outweighs capability-weighted relay loss, with stronger models gaining less from compression.