Self-evolving agents have become a focal point in AI research, yet most current approaches equate evolution with mere memory storage. OPD-Evolver challenges this assumption by proposing a slow-fast co-evolution framework where agents learn how to evolve from their experiences rather than just accumulating them. The fast loop involves agents interacting with a four-level memory hierarchy during test-time, while the slow loop distills these interactions into improved evolution strategies. This on-policy distillation approach ensures that agents continuously refine their learning processes based on real-time feedback. The collaboration between LV-NUS Lab, Fudan University, Peking University, and ByteDance brings together diverse expertise in reinforcement learning and agent design. Early results suggest significant improvements in agent adaptability and problem-solving capabilities across various benchmarks. For developers working on autonomous agents, this framework offers a new perspective on how to design systems that genuinely improve over time. The paper is available on arXiv and represents a meaningful contribution to the growing field of agent self-improvement.
OPD-Evolver introduces a novel slow-fast co-evolution framework for self-evolving agents, addressing the limitation that storing experiences is not the same as learning to evolve from them. The approach uses on-policy distillation with a four-level memory hierarchy to enhance agent capabilities. This research from LV-NUS Lab, Fudan, PKU, and ByteDance represents a significant step in agent self-improvement methodologies.