prompt-refinement

PACE introduces a two-timescale framework for self-evolution of small language model agents, coordinating low-risk prompt refinement with higher-risk control-logic updates, achieving up to +9.2% relative improvement across benchmarks.