experience-driven

MetaEvo proposes a two-stage framework for continual evolution of LLM-based agents, using preference-based optimization to enhance principle abstraction and modular architecture for experience reuse, outperforming strong baselines on reasoning benchmarks.