When Do Multi-Agent Systems Help? An Information Bottleneck Perspective
- source_url: https://arxiv.org/abs/2607.16133
- source_type: paper
- platform: arxiv
- author: Wendi Yu, Lianhao Zhou, Xiangjue Dong, Sai Sudarshan Barath, Declan Staunton, Byung-Jun Yoon, Xiaoning Qian, James Caverlee, Shuiwang Ji
- original_date: 2026-07-17
- added_date: 2026-07-22
- category: agents
- tags: multi-agent, information-bottleneck, relay, mas-vs-sas, arxiv
- quality_score: 5
- arxiv_id: 2607.16133
- arxiv_categories: cs.LG, cs.AI
- pdf_url: https://arxiv.org/pdf/2607.16133v1
摘要（中文）
从信息瓶颈解释 MAS 相对 SAS 何时有利：SAS 共享完整推理轨迹上下文，MAS 用有界 relay 连接隔离局部上下文。无限带宽下 MAS 可模拟任意 SAS；真正优势出现在有界 relay——压缩可减冗余，也可能丢掉任务关键信息。用有效参数 β 刻画模型能力如何改变权衡。18 组对照实验（五基准×三模型尺度）显示：relay 近充分时 MAS 常有帮助（弱模型尤甚）；relay 丢信息时收益缩小甚至反转（强模型更能从冗余上下文抽取信息）。多 Agent 设计本质是信息瓶颈优化，不是默认更高级。
Summary (English)
Provides an information-bottleneck perspective on when multi-agent systems (MAS) help versus single-agent systems (SAS). A SAS accumulates a full reasoning trace in one shared context; a MAS uses isolated local contexts connected by bounded relay messages. Under infinite relay bandwidth any SAS can be simulated by a MAS; nontrivial MAS advantages arise under bounded relays, where compression trades reduced redundancy against loss of task-relevant information. Formalizes the trade-off with effective parameter β tied to model capability. Across 18 controlled experiments on five benchmarks and three model scales, MAS consistently helps when relays are near-sufficient (especially weaker models); gains shrink or reverse when relays lose information (especially stronger models). Multi-agent design is fundamentally information-bottleneck optimization.
One-liner
多 Agent 何时有用：看有界 relay 是压缩冗余还是丢掉任务关键信息。
Source body / metadata
arXiv abstract grounded intake for 2607.16133. PDF: https://arxiv.org/pdf/2607.16133v1
Provides an information-bottleneck perspective on when multi-agent systems (MAS) help versus single-agent systems (SAS). A SAS accumulates a full reasoning trace in one shared context; a MAS uses isolated local contexts connected by bounded relay messages. Under infinite relay bandwidth any SAS can be simulated by a MAS; nontrivial MAS advantages arise under bounded relays, where compression trades reduced redundancy against loss of task-relevant information. Formalizes the trade-off with effective parameter β tied to model capability. Across 18 controlled experiments on five benchmarks and three model scales, MAS consistently helps when relays are near-sufficient (especially weaker models); gains shrink or reverse when relays lose information (especially stronger models). Multi-agent design is fundamentally information-bottleneck optimization.
Obsidian evidence
- local_note: OpenClaw定时任务/研究写作流水线/2026-07-22-研究写作流水线.md
- intake_run: daily-intake-evening 2026-07-22