AlphaEvolve — generator, evaluator, archive
LLM generator
Gemini Flash (many)
Gemini Pro (hard cases)
prompt = parent(s) +
top variants + task + eval sig
program archive
MAP-elites grid / islands
92.1
89.3
94.0
85.8
91.7
88.4
78.6
82.2
87.0
cell key = feature descriptor
(depth, complexity, footprint, ...)
each cell keeps its best
evaluator
correctness + benchmark
deterministic, fast
held-out inputs generated
at evaluation time
variants
variant
score
parent sample
documented wins (DeepMind, June 2025)
- 48-mul 4x4 complex matmul (Strassen 1969 bound 49)
- Google Borg scheduling heuristic, ~0.7% compute recovered in prod
- 32.5% FlashAttention kernel speedup · Gemini training throughput gains
reward hacking is bounded only by evaluator rigor. pick domains where the evaluator is not the weak link.