Title: Dynamic Trust-Aware Sparse Communication Topology for LLM-Based Multi-Agent Consensus

Large language model-powered multi-agent systems improve the robustness of intricate reasoning processes by leveraging multi-turn deliberation, specialized roles, and mutual validation. Nevertheless, current frameworks for agent debate and collaboration predominantly rely on fully connected communication networks. This approach results in a quadratic escalation of message volume, token consumption, and end-to-end latency as the agent count increases. While static sparse topologies can mitigate these costs, they lack the flexibility to adjust communication links based on specific task requirements or evolving reasoning states. Consequently, such rigid structures often retain low-value interactions while potentially discarding vital information necessary for error correction.

To overcome these limitations, this study introduces DySCo (Dynamic Sparse Consensus), a mechanism for dynamic, trust-aware sparse consensus. During each reasoning iteration, DySCo evaluates the utility of communication links by considering agent reliability, the divergence of answers, and task relevance. Under defined budget constraints, it identifies and selects a limited set of high-value edges for message exchange. The system then synthesizes responses from various agents using dynamic trust weights and halts the discussion prematurely once consensus reaches a stable state. By substituting universal broadcasting with on-demand communication, DySCo significantly lowers communication overhead while maintaining crucial cross-validation data. The paper also provides analyses of communication complexity and consensus stability, alongside performance evaluations of DySCo across tasks involving mathematical reasoning, logical deduction, and factual question-answering.

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC