Definition
What counts as a self-evolving AI agent?
Under this survey, strict self-evolution means feedback-driven retained change. A system must show what changed, why it changed, who verified it, and how the change survives or rolls back.
Strict survey gate
Six checks before calling a system self-evolving.
| Gate | Passes when | Common failure |
|---|---|---|
| Mutable object | Prompt, memory, tool policy, workflow, code, curriculum, weights, or population can change. | One-off reasoning inside a fixed workflow. |
| Feedback signal | The change is driven by benchmark, environment, user, test, verifier, or execution feedback. | The same model simply says it improved. |
| Candidate generation | The system participates in proposing candidate changes. | Humans rewrite the system offline while the agent only uses the result. |
| Independent verification | A test, hidden task, environment, or independent judge checks the candidate. | The proposer also approves the change with no outside check. |
| Retention | Approved changes affect later runs through memory, code, archive, weights, or configuration. | Nothing reusable survives the session. |
| Audit and rollback | Lineage, rejected candidates, costs, or rollback boundaries are visible. | Only a final score is reported. |
Evidence and limits
This definition is a review gate, not a universal law.
The checklist is synthesized from the survey draft, the Chinese definition page, and the project model-card workflow. Use it to reject weak claims first; do not treat it as a peer-reviewed formal definition until the paper citation map is complete.