Working draft
The English PDF can be cited as a dated working survey draft, not as a peer-reviewed field standard.
Paper / English mirror
Self-Evolving AI Survey: what can be cited today?
This is the English evidence and citation page for the working survey draft. It gives readers a narrow answer first: cite the paper as a dated working draft, cite the definition and five-loop framework with the current evidence boundary, and treat scores, counts, and framework rankings as provisional snapshots.
Citation boundary
Use the draft, but do not overstate it.
Stable core
The safest claims are the definition checklist, the five evolution loops, and the evidence-boundary vocabulary.
Provisional metrics
Method-family shares, framework scores, corpus counts, and Evolve-AGI worksheets are dated snapshots and need source-level audits before strong citation.
Chinese parity
The Chinese reading PDF and survey pages are useful companions, but chapter-level EN-CN mapping still needs cleanup before parity can be claimed.
Evidence snapshot
The corpus is large; the public conclusions remain gated.
| Evidence surface | Count | How to read it |
|---|---|---|
| Survey-display papers | 196 | 196-paper display snapshot used by the current figure atlas. |
| GitHub raw captures | 686 | Raw capture scope for repository discovery, not a quality ranking. |
| Project/model-card analyses | 300 | Teaching and review layer for project evidence. |
| Public project reports | 491 | Review-gated evidence pages; individual public-copy review is still pending. |
| Paper report files | 2 | Public paper report files used as evidence entry points. |
Reading path
Read the framework before the long corpus.
Start with the definition page, then the five-loop page, then the code-evolution benchmark matrix. Those three pages teach the claim test before sending readers into project reports, growth pilots, and report backlogs.