DySCo: New Sparse Consensus Cuts LLM Multi-Agent Costs by Avoiding Full Broadcast
Fully connected multi-agent debate wastes tokens – DySCo picks only high-value exchanges.
Large language model-driven multi-agent systems excel at complex reasoning through deliberation and cross-validation, but existing frameworks use fully connected topologies where every agent talks to every other agent each round. This causes messages, token costs, and latency to grow quadratically with the number of agents. Fixed sparse topologies save overhead but cannot adapt to different tasks or reasoning states, often losing critical corrections or wasting resources on low-value chatter. Researchers from the paper (arXiv:2606.01828) introduce DySCo (Dynamic Sparse Consensus), which replaces universal broadcasting with on-demand communication. In each reasoning round, DySCo estimates the value of each potential communication edge based on three factors: agent reliability (past correctness), answer divergence (how much agents disagree), and task relevance. It then selects a small set of high-value edges under a budget constraint, exchanges messages only along those edges, and aggregates answers using dynamic trust weights. The system also terminates discussion early once consensus stabilizes, further reducing waste.
DySCo’s communication complexity is analyzed and shown to achieve near-linear scaling with the number of agents, compared to the quadratic cost of full connectivity. The authors also provide formal analysis of consensus stability. In empirical evaluations on mathematical reasoning (GSM8K, MATH), logical reasoning (LogiQA), and factual question-answering (TriviaQA), DySCo matched or exceeded the accuracy of fully connected baselines while consuming significantly fewer tokens and reducing end-to-end latency. For example, with 8 agents, DySCo used roughly 60% fewer messages without any drop in correctness. This approach makes multi-agent debate practical for real-time applications where token and latency budgets are tight, such as interactive assistants, live decision support, and large-scale simulation. DySCo represents a step toward efficient, adaptive collaboration among AI agents that doesn’t compromise on the benefits of cross-validation.
- DySCo replaces fully connected multi-agent debate with on-demand, trust-aware edge selection, reducing messages from O(N²) to near-linear.
- It estimates edge value each round based on agent reliability, answer divergence, and task relevance, then selects top edges under a budget.
- Early termination when consensus stabilizes further reduces token consumption without sacrificing accuracy on math, logic, and QA benchmarks.
Why It Matters
Makes LLM multi-agent systems practical for real-time applications by slashing communication costs without losing accuracy.