EvoMAS : Heuristics in the Loop—Evolving Smarter Agentic Workflows

ICLR 2026 Conference Submission18149 Authors

Keywords: Multi-Agent Systems, Evolutionary Optimization, Agentic Workflow, Large Language Models, Curriculum Learning, Meta-Control Mechanism

Abstract:

The rapid development of Large Language Models has driven Multi-Agent Systems (MAS) growth, but constructing efficient MAS requires labor-intensive manual design. Current automation methods generate templated agents, use monolithic optimization, and ignore task complexity gradients. This paper presents Evolutionary MAS (\textbf{EvoMAS}), a biologically-inspired framework that systematically addresses these limitations through three interconnected dimensions: (1) \textbf{dynamic and diverse evolutionary strategies} with six biologically-inspired operators (3 exploration, 3 exploitation) and adaptive strategy selection; (2) \textbf{role-level evolution} that dynamically optimizes agent specialization and collaboration patterns; and (3) \textbf{curriculum-guided evolution} partitioning tasks by difficulty levels and evolving sequentially from simple to complex with cross-stage stability constraints. Additionally, to resolve the contradiction between the inefficiency of pure evolutionary methods and the limited flexibility of manual design, we developed the \textbf{"Cyber Creator"}, a meta-control system combining dynamic rule formulation with reflective updates. Experimental evaluations demonstrate that EvoMAS consistently outperforms existing methods across multiple domains while maintaining cost efficiency, with agent roles dynamically evolving from homogeneous actors to specialized reasoning ensembles. Codes are available at \href{https://anonymous.4open.science/r/EvoMAS-DEF4}
{EvoMAS}.

Primary Area: applications to robotics, autonomy, planning

Submission Number: 18149