Spectral-Aware Sparse Communication and Entropy-Balanced Tasking in Multi-Agent Systems

Yanli Wang, Yaowenqi Liu, Kaidi Xu, Haohan Wang

Keywords: Multi-Agent Systems, Reinforcement Learning, LLM Agents, Spectral Sparsification, Algebraic Connectivity, Mutual Information, Entropy-Regularized Assignment, Graph Laplacian

TL;DR: We prune redundant links via an information-theoretic priority and enforce connectivity with a guarded repair, then balance workload using an entropy-regularized assignment with an MDP allocator.

Abstract: Multi-agent systems (MAS) face scalability constraints stemming from dense messaging—raising bandwidth and energy—and from imbalanced tasking that produces bottlenecks, especially under non-stationary, LLM-driven workloads. We introduce a unified framework that prunes redundant links using an information-theoretic priority and enforces connectivity via a spectral $\lambda_2$ guard, and balances workload with an entropy-regularized assignment under capacity constraints; an MDP allocator adapts thresholds and repairs to system drift. We prove that the $\lambda_2$-guarded repair preserves connectivity and, under standard spectral envelope assumptions, the sparsified graph approximates dense-graph dynamics; we analyze discrete-time stability via Jury/Nyquist bounds. Across GSM8K, MMLU, and SMACv2, the method improves over a dense complete-graph baseline by +6.12\%/+5.59\%/+4.76\%, reduces active links by 28\%, and shows the strongest robustness under 30\% edge drops -- all while keeping per-iteration communication proportional to the number of active edges. These results indicate a practical route to communication-efficient, entropy-balanced coordination for LLM-augmented MAS and cooperative control.

Primary Area: reinforcement learning

Submission Number: 10958