HYPER-MACNET: Task-Aware Hypergraphs for Collective Intelligence in Multi-Agent LLMs

ACL ARR 2026 January Submission8612 Authors

Keywords: Multi-agent LLMs, Hypergraph-based Coordination, DAG, Complex Tasks

Abstract: Multi-agent LLM systems perform well on bounded tasks, but they often break down on cross-domain, long-horizon tasks that require many interdependent constraints to be satisfied together. A key difficulty is coordination: isolated pairwise discussions can yield locally correct plans that become inconsistent when merged. To address this, we present HYPER-MACNET, a multi-agent framework that organizes collaboration with task-aware hypergraphs. Instead of using a fixed communication graph, HYPER-MACNET decomposes each problem into a subtask dependency DAG and assigns each subtask to a hyperedge, which explicitly defines the responsible collaboration unit. It further performs mode-aware collaboration, assigns roles within each unit, and records intermediate artifacts on a global blackboard for dependency-consistent propagation and aggregation. On standard benchmarks, HYPER-MACNET achieves a 6.1\% relative gain over the strongest baseline and a 37.8\% gain over a vanilla baseline. On the CLM complex-task evaluation, it attains the best mean ranking, indicating more globally consistent coordination under coupled constraints.

Paper Type: Long

Research Area: AI/LLM Agents

Research Area Keywords: Language Modeling

Contribution Types: Model analysis & interpretability, NLP engineering experiment

Languages Studied: Python

Submission Number: 8612