ReActNet: From Learning Communication Graphs to Reasoning-Actuated Dynamic Orchestrations in Multi-Agent Systems

Recent advances in multi-agent systems have explored structuring communication through predefined or learned interaction topologies. While these approaches improve coordination, they treat communication structure as the primary object of optimization and largely overlook the role of reasoning dynamics during multi-round interaction. In this work, we revisit multi-agent orchestration from a different perspective: instead of learning a fixed communication topology, we propose to directly control the reasoning process. We introduce a reasoning-driven framework in which a large language model (LLM) acts as a controller that dynamically constructs and updates a temporal communication graph throughout the interaction process. Given a task and a set of role-specialized agents, the controller generates a sequence of communication graphs, monitors intermediate reasoning states, and adaptively coordinates information flow across agents over multiple rounds. Each agent updates its reasoning by aggregating its own previous state with messages from its dynamically assigned neighbors, while a final aggregator synthesizes the collective reasoning into the final answer. This design unifies communication structure and communication semantics within a single closed-loop system, eliminating the need for explicit topology learning or reinforcement learning-based optimization. Empirically, our approach consistently outperforms prior methods across a diverse set of benchmarks covering knowledge reasoning, mathematical problem solving, and code generation. Notably, our method achieves higher accuracy while maintaining comparable inference cost, demonstrating that dynamic reasoning control can be both more effective and efficient than learned communication structures. Our results suggest a fundamental shift in multi-agent system design: rather than optimizing communication topology as a static object, it is more effective to treat coordination as an adaptive reasoning process governed by a central controller.