Position: Agentic Evolution is the Path to Evolving LLMs

As Large Language Models (LLMs) move from controlled training environments to open-ended real-world applications, a critical limitation arises: static training cannot keep pace with continuous environmental changes. This paper argues for a new scaling axis—evolution—to bridge this 'train-deploy gap.' It proposes agentic evolution as the future of LLM adaptation, elevating evolution from a fixed pipeline to an autonomous evolver agent. The framework, A-Evolve, treats deployment-time improvement as a deliberate, goal-directed optimization over persistent system state, positing an evolution-scaling hypothesis where adaptation capacity scales with allocated compute, leading to sustained, open-ended adaptation in real-world scenarios.