Overview
AlphaEvolve turns language-model code proposals into cumulative search: humans define the target, representation, evaluator, and invariants; the system mutates executable programs, measures them, and retains useful variation. Across matrix algorithms, open mathematical constructions, and production systems, it shows how feedback can yield real improvements while leaving optimality, generalization, safety, and broader judgment outside any single score.
Learning outcomes
- AlphaEvolve’s discoveries come from coupling model-generated program mutations with executable evaluation and population memory.
- The chosen representation—solution, constructor, search strategy, or bounded subdecision—determines what improvement means.
- A better score can establish progress under tested conditions without proving optimality, generalization, or end-to-end value.
- Human formulation, protected invariants, affordable evaluation, and post-search validation remain essential to trustworthy use.
Keywords
Preview every chapter · Unlock full text, audio, and video from chapter 4
PlansFAQ
What is AlphaEvolve?
AlphaEvolve is a coding agent that evolves executable programs through language-model mutations, machine evaluation, and population-based retention. Humans define the target and boundaries, so it is constrained search rather than an autonomous scientific oracle.
How does AlphaEvolve turn code generation into discovery?
It turns code generation into discovery by executing proposed program changes, scoring them against a user-defined evaluator, and feeding retained results into later mutations. This cumulative loop can search algorithms, constructions, heuristics, and production decisions instead of treating each response as a one-shot suggestion.
What kinds of problems fit AlphaEvolve?
Problems fit AlphaEvolve when a target artifact can be executed, the objective is measurable and sufficiently aligned, essential invariants can be protected, and evaluation is affordable. The reported cases include algorithms, mathematical constructions, scheduling, kernels, circuits, and compiler-generated code.
What can AlphaEvolve’s results actually prove?
They can support a measured improvement or a checked construction within the stated evaluator, conditions, and scope. They do not automatically prove global optimality, performance on all workloads, end-to-end gains, or universal correctness.