Evolutionary Ensemble of Agents: A Decentralized Framework for Algorithmic Discovery
Researchers Zongmin Yu and Liu Yang have introduced the Evolutionary Ensemble (EvE), a decentralized framework designed to organize existing coding agents into a live, co-evolving system for algorithmic discovery. Published on arXiv in May 2026, this study shifts focus from optimizing Large Language Models (LLMs) themselves to evolving the cumulative guidance and skills that dictate agent behaviors. EvE maintains two co-evolving populations: functional code solvers and agent guidance states. The system evaluates agents through synchronous races, updating their empirical Elo ratings based on marginal gains contributed to the current solver state. When applied to In-Context Operator Networks (ICON), EvE autonomously discovered a robust rescale-then-interpolate mechanism, enabling reliable example-count generalization. Controlled ablations demonstrated that stage-dependent agent adaptation is essential for navigating complex codebases, avoiding the phase mismatch issues seen in fixed-agent variants. This research highlights that organizing agents into a self-revising ensemble is key to breaking through static performance ceilings in artificial intelligence and neural computing.
Wire timeline
Evolutionary Ensemble of Agents: A Decentralized Framework for Algorithmic Discovery
Researchers Zongmin Yu and Liu Yang have introduced the Evolutionary Ensemble (EvE), a decentralized framework designed to organize existing coding agents into a live, co-evolving system for algorithmic discovery. Published on arXiv in May 2026, this study shifts focus from optimizing Large Language Models (LLMs) themselves to evolving the cumulative guidance and skills that dictate agent behaviors. EvE maintains two co-evolving populations: functional code solvers and agent guidance states. The system evaluates agents through synchronous races, updating their empirical Elo ratings based on marginal gains contributed to the current solver state. When applied to In-Context Operator Networks (ICON), EvE autonomously discovered a robust rescale-then-interpolate mechanism, enabling reliable example-count generalization. Controlled ablations demonstrated that stage-dependent agent adaptation is essential for navigating complex codebases, avoiding the phase mismatch issues seen in fixed-agent variants. This research highlights that organizing agents into a self-revising ensemble is key to breaking through static performance ceilings in artificial intelligence and neural computing.
cs.AI updates on arXiv.org