MetaEvo: A Meta-Optimization Framework for Experience-Driven Agent Evolution

Bowen Ren, Heyan Huang, Yinghao Li, Yang Gao

Keywords: Agent System, Meta-Optimization, Continual Evolution, Principle Abstraction

Abstract: Existing methods for experience-driven agent evolution often generate revision principles that lack actionable guidance, as they overlook the inherent limitations of Large Language Models (LLMs) in abstracting knowledge and self-correcting. To address this gap, we introduce MetaEvo, a novel framework that reframes agent evolution as a meta-optimization task. Instead of learning directly from experience, our core idea is to first enhance the agent’s intrinsic meta-ability—its capacity to learn how to effectively revise and improve itself. MetaEvo operationalizes this concept through a three-stage pipeline powered by a modular agent system. First, a meta-optimization stage explicitly trains the model on abstracting high-quality principles. These principles are then accumulated in a curated memory and subsequently retrieved by an execution module to guide generation on new tasks. Extensive experiments across a diverse suite of mathematical reasoning and multi-task benchmarks demonstrate that MetaEvo consistently and significantly improves model performance, enabling more robust and generalizable reasoning behaviors.

Primary Area: optimization

Code Of Ethics: I acknowledge that I and all co-authors of this work have read and commit to adhering to the ICLR Code of Ethics.

Submission Guidelines: I certify that this submission complies with the submission instructions as described on https://iclr.cc/Conferences/2026/AuthorGuide.

Anonymous Url: I certify that there is no URL (e.g., github page) that could be used to find authors’ identity.

No Acknowledgement Section: I certify that there is no acknowledgement section in this submission for double blind review.

Submission Number: 12370