SCIENCE UNDER AUTOMATION
AlphaEvolve technical report
Evolutionary code search couples model-generated candidates to executable evaluators.
What this result supports
Human objectives, evaluators and integration remain; selected successes lack an exhaustive success denominator.
Evidence record: preprint. Checked: 2026-10-20.
Where this evidence is used
Algorithm design
- Formulate objective and evaluator: Humans supply the optimization target and evaluation environment.
- Search for executable improvements: Executable feedback supports iterative algorithm search.
- Check correctness and performance: Automated tests cover the supplied evaluator, not arbitrary deployment conditions.
- Choose the abstraction and constraints: No representative autonomous problem-formulation evaluation.
- Develop an algorithm and proof: Search successes are narrower than unrestricted algorithm design.
- Establish generalization and integrate: Independent end-to-end integration is not evaluated.