When Do Multi-Agent Systems Help? An Information Bottleneck Perspective
Multi-agent systems help when their relay messages compress context without dropping task-relevant information.
The paper frames the split between single-agent and multi-agent setups as an information bottleneck: one shared reasoning trace versus isolated local contexts connected by bounded messages. It argues that if relays had infinite bandwidth, a multi-agent system could simply simulate a single agent, so the real difference appears only when communication is compressed. Across 18 controlled experiments on five benchmarks and three model scales, the authors report that MAS gains are strongest when relays are near-sufficient, especially for weaker models. When relay messages lose important information, the gains shrink or reverse, particularly for stronger models. ArXiv · AI/CL/LG's note
The paper frames the split between single-agent and multi-agent setups as an information bottleneck: one shared reasoning trace versus isolated local contexts connected by bounded messages. It argues that if relays had infinite bandwidth, a multi-agent system could simply simulate a single agent, so the real difference appears only when communication is compressed. Across 18 controlled experiments on five benchmarks and three model scales, the authors report that MAS gains are strongest when relays are near-sufficient, especially for weaker models. When relay messages lose important information, the gains shrink or reverse, particularly for stronger models. ArXiv · AI/CL/LG's note
score 5