EvoMAS: Learning Execution-Time Workflows for Multi-Agent Systems
Researchers have introduced EvoMAS, a novel framework designed to enhance Large Language Model (LLM)-based multi-agent systems by enabling dynamic workflow construction during task execution. Unlike traditional static methods that fix workflows before execution, EvoMAS adapts to evolving subgoals and information needs in long-horizon tasks. It formulates workflow construction as a meta-level sequential decision problem, utilizing a Planner-Evaluator-Updater pipeline to maintain an explicit task state. A learned Workflow Adapter then instantiates stage-specific layered workflows from a candidate agent pool, trained via policy gradients with sparse terminal success signals. Experimental results on benchmarks such as GAIA, HLE, and DeepResearcher demonstrate that EvoMAS significantly outperforms single-agent baselines and existing automated design methods. The study highlights the complementary benefits of explicit task-state construction and adaptive workflow learning, noting that process rewards are particularly valuable in scenarios with extremely sparse terminal success. This advancement addresses critical limitations in current multi-agent coordination strategies, offering a more flexible and efficient approach to complex, multi-stage computational tasks.
Wire timeline
EvoMAS: Learning Execution-Time Workflows for Multi-Agent Systems
Researchers have introduced EvoMAS, a novel framework designed to enhance Large Language Model (LLM)-based multi-agent systems by enabling dynamic workflow construction during task execution. Unlike traditional static methods that fix workflows before execution, EvoMAS adapts to evolving subgoals and information needs in long-horizon tasks. It formulates workflow construction as a meta-level sequential decision problem, utilizing a Planner-Evaluator-Updater pipeline to maintain an explicit task state. A learned Workflow Adapter then instantiates stage-specific layered workflows from a candidate agent pool, trained via policy gradients with sparse terminal success signals. Experimental results on benchmarks such as GAIA, HLE, and DeepResearcher demonstrate that EvoMAS significantly outperforms single-agent baselines and existing automated design methods. The study highlights the complementary benefits of explicit task-state construction and adaptive workflow learning, noting that process rewards are particularly valuable in scenarios with extremely sparse terminal success. This advancement addresses critical limitations in current multi-agent coordination strategies, offering a more flexible and efficient approach to complex, multi-stage computational tasks.
cs.AI updates on arXiv.org