EvoMAS: A Runtime Workflow Learning Framework for Multi-Agent Systems

EvoMAS is a runtime workflow learning framework targeting the limitations of static coordination strategies in multi-agent systems. Its core lies in constructing dynamic task states via the Planner-Evaluator-Updater pipeline and using a learned workflow adapter to instantiate phase-specific hierarchical workflows, so as to cope with the evolution of subgoals, intermediate evidence, and information needs in long-horizon tasks.

Multi-agent systems based on large language models show significant potential in complex tasks, but most methods adopt a one-time paradigm: optimizing/selecting workflows before execution and using them fixed throughout. This static strategy is insufficient for long-horizon tasks, as their subgoals, intermediate evidence, and information needs evolve with execution phases.

EvoMAS frames workflow construction as a meta-level sequential decision problem. Its core innovations include: 1. Planner-Evaluator-Updater pipeline: The Planner analyzes the current state to plan actions, the Evaluator assesses the suitability of candidate agents, and the Updater updates the state based on feedback; 2. Learned workflow adapter: Instantiates phase-specific hierarchical workflows from a candidate pool, dynamically adjusting agent combinations and strategies; 3. Training mechanism: The adapter is trained using policy gradients, with terminal task success as the supervision signal, analyzing the role of process rewards under extremely sparse rewards.

Evaluations on benchmarks like GAIA, HLE, and DeepResearcher show: EvoMAS outperforms single-agent baselines and recent automated multi-agent workflow design methods; explicit task state construction and learned workflow adaptation have complementary advantages; process rewards are most useful when terminal success is extremely sparse. Qualitative cases indicate that the system can identify phase transition nodes, adjust agent division of labor, and reconfigure workflows based on new information.

EvoMAS enables the transition of multi-agent systems from static design to dynamic adaptation. Its core contributions are: 1. Runtime workflow construction, breaking through predefined limitations; 2. Explicit state modeling, supporting more intelligent decision-making; 3. End-to-end learning, learning complex coordination strategies from sparse rewards.

EvoMAS is suitable for open-domain research tasks (multi-round information collection and reasoning), complex problem solving (subgoals refined with exploration), and dynamic environment interaction (external changes requiring strategy adjustments). This framework provides a new paradigm for flexible and powerful multi-agent systems, promoting the practical deployment of autonomous AI agents.