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

In three linesMetaEvo introduces a two-stage framework for continual LLM agent evolution. It combines preference-based optimization to enhance principle abstraction, then accumulates and reuses these principles in a modular agent architecture. Results on reasoning benchmarks show consistent improvements across iterations without performance plateaus.

Summary generated by Claude — human-verified