A recent preprint examines whether large language model agents can improve by learning from their own experience, a process called self-evolution. The authors note that existing approaches typically rely on successful trajectories or external evaluation to guide improvement, which may limit how agents learn from mistakes.

The study's central finding, reflected in its title, is that self-evolving agents can repair specific failures they have encountered, but they struggle to generalize from those repairs to new, related tasks. In other words, the agents become better at avoiding the exact same error, yet do not reliably transfer that knowledge to different but similar challenges.

This distinction matters for building agents that truly improve over time. If learning only patches individual failures, then self-evolution offers limited benefit in open-ended environments where novel problems are common. The authors suggest that current self-evolution strategies may need to focus more on extracting generalizable lessons rather than memorizing fixes for particular cases.