Position: Agentic Evolution is the Path to Evolving LLMs
Minhua Lin, Hanqing Lu, Zhan Shi, Bing He, Rui Mao, Zhiwei Zhang, Zongyu Wu, Xianfeng Tang, Hui Liu, Zhenwei Dai, Xiang Zhang, Suhang Wang, Benoit Dumoulin, Jian Pei
TL;DR
The paper argues that static training cannot keep pace with deployment-time environment changes and proposes agentic evolution as a scalable, autonomous mechanism for LLM adaptation. It introduces A-Evolve, a framework that treats deployment-time improvement as a governed, cross-episode optimization over persistent artifacts $\pi_S$ and model parameters $\pi_\theta$, driven by an explicit evolver agent. The Evolution-Scaling Hypothesis posits that adaptation capability scales with evolution-time compute, and the authors provide formalism, a concrete architecture (Persistent Artifact State, Solve--Evolve loop, Evolver), and empirical evidence from AppWorld showing improved task success and robustness with agentic evolution. The work outlines practical implications for durability, governance, and privacy, while acknowledging risks and proposing verification gates to mitigate undesired drift. Overall, the results support a shift from reactive inference-time thinking to proactive, artifact-driven, scalable deployment-time adaptation that can progressively raise the frontier of open-ended AI capabilities.
Abstract
As Large Language Models (LLMs) move from curated training sets into open-ended real-world environments, a fundamental limitation emerges: static training cannot keep pace with continual deployment environment change. Scaling training-time and inference-time compute improves static capability but does not close this train-deploy gap. We argue that addressing this limitation requires a new scaling axis-evolution. Existing deployment-time adaptation methods, whether parametric fine-tuning or heuristic memory accumulation, lack the strategic agency needed to diagnose failures and produce durable improvements. Our position is that agentic evolution represents the inevitable future of LLM adaptation, elevating evolution itself from a fixed pipeline to an autonomous evolver agent. We instantiate this vision in a general framework, A-Evolve, which treats deployment-time improvement as a deliberate, goal-directed optimization process over persistent system state. We further propose the evolution-scaling hypothesis: the capacity for adaptation scales with the compute allocated to evolution, positioning agentic evolution as a scalable path toward sustained, open-ended adaptation in the real world.