PACE: Two-Timescale Self-Evolution for Small Language Model Agents

In three linesPACE is a self-evolution framework for small language model agents (4B–14B parameters). It coordinates prompt refinement with control-logic updates via held-out validation, without requiring frontier models. Across 12 backbone–benchmark combinations, PACE improves vanilla SLM agents by +9.2% and single-mode evolution baselines by +5.4%.

Summary generated by Claude — human-verified