When Multi-Agent Systems Help

Summary & Methodology Analysis

The paper investigates the performance differences between multi-agent systems and single-agent systems powered by large language models, addressing why multi-agent advantages vary inconsistently across settings. To explore this, the researchers establish a structured multi-agent workflow. First, a planner decomposes the instance X into an ordered sequence of sub-instances X_1 through X_n. Each basic agent then operates as a worker on its assigned sub-instance X_i, executing specific tasks within that scoped boundary.

Following the worker execution phase, the architecture handles inter-agent communication. Workers communicate sequentially through compressed relay messages defined as m_i = Agg_i(M_i) with m_0 = ≰, where worker i initializes its context as M_0^{(i)} = (X_i, o_0^{(i)}) ⊕ m_{i-1}. Finally, a system-level aggregator combines all terminal memories into the final output y_MAS = Agg_MAS(M_1, ..., M_n). This design allows information to flow through the agent network while managing context size via compression at each relay step.

The empirical evaluation spans multiple environments, specifically ALFWorld, WebShop, WorkBench, WideSearch, and TravelPlanner, using models such as Qwen2.5-7B-Instruct, GPT-4o-mini, Qwen3.5-27B, and Qwen3.5-27B-AWQ-4bit. The system demonstrates measurable improvements, such as outperforming single-agent contextflow on ALFWorld by +0.194 for Qwen2.5-7B, +0.157 for GPT-4o-mini, and +0.023 for Qwen3.5-27B. The paper does not specify any particular limitations.