Capability Erosion in Self-Evolving LLM Agents and a Stabilization Method
A new arXiv preprint (2605.09315) reveals that self-evolving LLM agents suffer from capability erosion: adapting to new tasks degrades previously acquired skills across workflow, skill, model, and memory evolution. The authors propose Capability-Preserving Evolution (CPE), a stabilization principle that constrains destructive drift. In workflow evolution, CPE improved retained simple-task performance while preserving adaptation.
Key facts
- Self-evolving LLM agents can autonomously refine workflows, accumulate skills, self-train, and maintain persistent memory.
- Self-evolution is often non-monotonic, causing capability erosion across all major evolution channels.
- Capability erosion occurs in workflow, skill, model, and memory evolution.
- CPE (Capability-Preserving Evolution) is proposed as a general stabilization principle.
- CPE improves retained capability stability while preserving adaptation performance.
- In workflow evolution, CPE improved retained simple-task performance.
- The paper is a preprint on arXiv with ID 2605.09315.
- The research addresses a fundamental challenge in lifelong LLM agent adaptation.
Entities
Institutions
- arXiv