Do Self-Evolving Agents Forget? Capability Degradation and Preservation in Lifelong LLM Agent Adaptation
A new research paper submitted to arXiv investigates the stability of self-evolving Large Language Model (LLM) agents. The study reveals that while recent advances allow agents to autonomously refine workflows, accumulate skills, and self-train, this self-evolution is often non-monotonic. The authors identify a phenomenon termed 'capability erosion under self-evolution,' where adapting to new task distributions progressively degrades previously acquired capabilities across workflow, skill, model, and memory channels. To address this, the researchers propose 'Capability-Preserving Evolution' (CPE), a stabilization principle designed to constrain destructive capability drift during continual adaptation. Experimental results demonstrate that CPE consistently improves retained capability stability while maintaining adaptation performance. For instance, in workflow evolution using GPT-5.1 optimization, CPE increased retained simple-task performance from 41.8% to 52.8%, alongside stronger complex-task adaptation. The findings emphasize that stable long-horizon self-evolving agents must explicitly preserve learned capabilities rather than solely focusing on acquiring new ones. This work contributes significantly to the field of artificial intelligence by offering a solution to catastrophic forgetting in autonomous agent systems.
Wire timeline
Do Self-Evolving Agents Forget? Capability Degradation and Preservation in Lifelong LLM Agent Adaptation
A new research paper submitted to arXiv investigates the stability of self-evolving Large Language Model (LLM) agents. The study reveals that while recent advances allow agents to autonomously refine workflows, accumulate skills, and self-train, this self-evolution is often non-monotonic. The authors identify a phenomenon termed 'capability erosion under self-evolution,' where adapting to new task distributions progressively degrades previously acquired capabilities across workflow, skill, model, and memory channels. To address this, the researchers propose 'Capability-Preserving Evolution' (CPE), a stabilization principle designed to constrain destructive capability drift during continual adaptation. Experimental results demonstrate that CPE consistently improves retained capability stability while maintaining adaptation performance. For instance, in workflow evolution using GPT-5.1 optimization, CPE increased retained simple-task performance from 41.8% to 52.8%, alongside stronger complex-task adaptation. The findings emphasize that stable long-horizon self-evolving agents must explicitly preserve learned capabilities rather than solely focusing on acquiring new ones. This work contributes significantly to the field of artificial intelligence by offering a solution to catastrophic forgetting in autonomous agent systems.
cs.AI updates on arXiv.org